[llvm] [AMDGPU] Fix LSR cost comparator that regressed GFX9+ occupancy (PR #204344)

Yuyang Zhang via llvm-commits llvm-commits at lists.llvm.org
Mon Jun 29 19:23:27 PDT 2026


https://github.com/yuyzhang512 updated https://github.com/llvm/llvm-project/pull/204344

>From 615c794c30036ffd5b12aa21458c0f6bcad2e935 Mon Sep 17 00:00:00 2001
From: yuyzhang512 <yuyzhang at amd.com>
Date: Fri, 19 Jun 2026 11:43:15 +0000
Subject: [PATCH] [AMDGPU] Fix LSR cost comparator that regressed GFX9+
 occupancy

#184138 made the GFX9+ LSR cost model lead with the per-iteration instruction
count and demote register count, so LSR adds induction variables / registers
to shave instructions. On VGPR-bound GFX9+ that lowers occupancy -- the
pa_prefill paged-attention kernel needed 266 VGPRs and spilled instead of 256,
which costs far more than the instructions saved.

Lead the comparator with register pressure (NumRegs, then AddRecCost), keep the
effective per-iteration instruction count just below, and return true from
isNumRegsMajorCostOfLSR (the pre-#184138 default) so register count is
consistently the major cost.
---
 .../AMDGPU/AMDGPUTargetTransformInfo.cpp      |    20 +-
 .../CodeGen/AMDGPU/cf-loop-on-constant.ll     |    41 +-
 llvm/test/CodeGen/AMDGPU/flat-saddr-load.ll   |    79 +-
 llvm/test/CodeGen/AMDGPU/global-saddr-load.ll |   146 +-
 .../AMDGPU/lsr-cost-model-occupancy.ll        |   375 +
 ...p-var-out-of-divergent-loop-swdev407790.ll |    22 +-
 ...ne-sink-temporal-divergence-swdev407790.ll |   192 +-
 .../CodeGen/AMDGPU/memintrinsic-unroll.ll     | 12896 ++++++++--------
 llvm/test/CodeGen/AMDGPU/memset-pattern.ll    |     4 +-
 llvm/test/CodeGen/AMDGPU/uniform-cfg.ll       |     7 +-
 llvm/test/CodeGen/AMDGPU/valu-i1.ll           |    15 +-
 .../waitcnt-loop-ds-prefetch-pattern.ll       |    37 +-
 .../LoopStrengthReduce/AMDGPU/atomics.ll      |    20 +-
 ...fferent-addrspace-addressing-mode-loops.ll |    22 +-
 14 files changed, 7132 insertions(+), 6744 deletions(-)
 create mode 100644 llvm/test/CodeGen/AMDGPU/lsr-cost-model-occupancy.ll

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
index 40227c496a589..7c40207ee53b4 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
@@ -1858,22 +1858,18 @@ InstructionCost GCNTTIImpl::getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
 
 bool GCNTTIImpl::isLSRCostLess(const TTI::LSRCost &A,
                                const TTI::LSRCost &B) const {
-  // Favor lower per-iteration work over preheader/setup costs.
-  // AMDGPU lacks rich addressing modes, so ScaleCost is folded into the
-  // effective instruction count (base+scale*index requires a separate ADD).
+  // Occupancy is VGPR-bound, so register pressure must dominate the LSR cost:
+  // NumRegs/AddRecCost outrank EffInsns (Insns + ScaleCost), so LSR does not add
+  // registers to save a few per-iteration instructions.
   unsigned EffInsnsA = A.Insns + A.ScaleCost;
   unsigned EffInsnsB = B.Insns + B.ScaleCost;
-
-  return std::tie(EffInsnsA, A.NumIVMuls, A.AddRecCost, A.NumBaseAdds,
-                  A.SetupCost, A.ImmCost, A.NumRegs) <
-         std::tie(EffInsnsB, B.NumIVMuls, B.AddRecCost, B.NumBaseAdds,
-                  B.SetupCost, B.ImmCost, B.NumRegs);
+  return std::tie(A.NumRegs, A.AddRecCost, EffInsnsA, A.NumIVMuls,
+                  A.NumBaseAdds, A.ImmCost, A.SetupCost) <
+         std::tie(B.NumRegs, B.AddRecCost, EffInsnsB, B.NumIVMuls,
+                  B.NumBaseAdds, B.ImmCost, B.SetupCost);
 }
 
-bool GCNTTIImpl::isNumRegsMajorCostOfLSR() const {
-  // isLSRCostLess de-prioritizes register count; keep consistent.
-  return false;
-}
+bool GCNTTIImpl::isNumRegsMajorCostOfLSR() const { return true; }
 
 bool GCNTTIImpl::shouldDropLSRSolutionIfLessProfitable() const {
   // Prefer the baseline when LSR cannot clearly reduce per-iteration work.
diff --git a/llvm/test/CodeGen/AMDGPU/cf-loop-on-constant.ll b/llvm/test/CodeGen/AMDGPU/cf-loop-on-constant.ll
index 637b5f95a7b62..a49148c09bde1 100644
--- a/llvm/test/CodeGen/AMDGPU/cf-loop-on-constant.ll
+++ b/llvm/test/CodeGen/AMDGPU/cf-loop-on-constant.ll
@@ -11,18 +11,15 @@ define amdgpu_kernel void @test_loop(ptr addrspace(3) %ptr, i32 %n) nounwind {
 ; GCN-NEXT:    s_cbranch_scc1 .LBB0_3
 ; GCN-NEXT:  ; %bb.1: ; %for.body.preheader
 ; GCN-NEXT:    s_load_dword s0, s[4:5], 0x9
-; GCN-NEXT:    s_mov_b32 s1, 0
 ; GCN-NEXT:    s_and_b64 vcc, exec, -1
 ; GCN-NEXT:    s_mov_b32 m0, -1
+; GCN-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN-NEXT:    s_addk_i32 s0, 0x80
 ; GCN-NEXT:  .LBB0_2: ; %for.body
 ; GCN-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GCN-NEXT:    s_lshl_b32 s2, s1, 2
-; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_add_i32 s2, s0, s2
-; GCN-NEXT:    s_addk_i32 s2, 0x80
-; GCN-NEXT:    v_mov_b32_e32 v0, s2
+; GCN-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN-NEXT:    ds_read_b32 v1, v0
-; GCN-NEXT:    s_add_i32 s1, s1, 1
+; GCN-NEXT:    s_add_i32 s0, s0, 4
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN-NEXT:    v_add_f32_e32 v1, 1.0, v1
 ; GCN-NEXT:    ds_write_b32 v0, v1
@@ -116,18 +113,15 @@ define amdgpu_kernel void @loop_const_true(ptr addrspace(3) %ptr, i32 %n) nounwi
 ; GCN-LABEL: loop_const_true:
 ; GCN:       ; %bb.0: ; %entry
 ; GCN-NEXT:    s_load_dword s0, s[4:5], 0x9
-; GCN-NEXT:    s_mov_b32 s1, 0
 ; GCN-NEXT:    s_and_b64 vcc, exec, -1
 ; GCN-NEXT:    s_mov_b32 m0, -1
+; GCN-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN-NEXT:    s_addk_i32 s0, 0x80
 ; GCN-NEXT:  .LBB1_1: ; %for.body
 ; GCN-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GCN-NEXT:    s_lshl_b32 s2, s1, 2
-; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    s_add_i32 s2, s0, s2
-; GCN-NEXT:    s_addk_i32 s2, 0x80
-; GCN-NEXT:    v_mov_b32_e32 v0, s2
+; GCN-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN-NEXT:    ds_read_b32 v1, v0
-; GCN-NEXT:    s_add_i32 s1, s1, 1
+; GCN-NEXT:    s_add_i32 s0, s0, 4
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN-NEXT:    v_add_f32_e32 v1, 1.0, v1
 ; GCN-NEXT:    ds_write_b32 v0, v1
@@ -381,22 +375,19 @@ define amdgpu_kernel void @loop_arg_0(ptr addrspace(3) %ptr, i32 %n) nounwind {
 ; GCN-NEXT:    v_mov_b32_e32 v0, 0
 ; GCN-NEXT:    s_mov_b32 m0, -1
 ; GCN-NEXT:    ds_read_u8 v0, v0
-; GCN-NEXT:    s_load_dword s0, s[4:5], 0x9
+; GCN-NEXT:    s_load_dword s4, s[4:5], 0x9
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    v_readfirstlane_b32 s1, v0
-; GCN-NEXT:    s_bitcmp1_b32 s1, 0
-; GCN-NEXT:    s_cselect_b64 s[2:3], -1, 0
-; GCN-NEXT:    s_xor_b64 s[2:3], s[2:3], -1
-; GCN-NEXT:    s_mov_b32 s1, 0
+; GCN-NEXT:    v_readfirstlane_b32 s0, v0
+; GCN-NEXT:    s_bitcmp1_b32 s0, 0
+; GCN-NEXT:    s_cselect_b64 s[0:1], -1, 0
+; GCN-NEXT:    s_xor_b64 s[2:3], s[0:1], -1
+; GCN-NEXT:    s_add_i32 s0, s4, 0x80
 ; GCN-NEXT:    s_and_b64 vcc, exec, s[2:3]
 ; GCN-NEXT:  .LBB4_1: ; %for.body
 ; GCN-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GCN-NEXT:    s_lshl_b32 s2, s1, 2
-; GCN-NEXT:    s_add_i32 s2, s0, s2
-; GCN-NEXT:    s_addk_i32 s2, 0x80
-; GCN-NEXT:    v_mov_b32_e32 v0, s2
+; GCN-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN-NEXT:    ds_read_b32 v1, v0
-; GCN-NEXT:    s_add_i32 s1, s1, 1
+; GCN-NEXT:    s_add_i32 s0, s0, 4
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN-NEXT:    v_add_f32_e32 v1, 1.0, v1
 ; GCN-NEXT:    ds_write_b32 v0, v1
diff --git a/llvm/test/CodeGen/AMDGPU/flat-saddr-load.ll b/llvm/test/CodeGen/AMDGPU/flat-saddr-load.ll
index 35aa5e401afe6..ec5dccad997b3 100644
--- a/llvm/test/CodeGen/AMDGPU/flat-saddr-load.ll
+++ b/llvm/test/CodeGen/AMDGPU/flat-saddr-load.ll
@@ -2929,15 +2929,16 @@ define amdgpu_ps void @flat_addr_64bit_lsr_iv(ptr inreg %arg) {
 ; GFX1250-SDAG:       ; %bb.0: ; %bb
 ; GFX1250-SDAG-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
 ; GFX1250-SDAG-NEXT:    v_mov_b32_e32 v0, 0
-; GFX1250-SDAG-NEXT:    s_movk_i32 s0, 0x100
+; GFX1250-SDAG-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX1250-SDAG-NEXT:  .LBB116_1: ; %bb3
 ; GFX1250-SDAG-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX1250-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1250-SDAG-NEXT:    s_add_nc_u64 s[4:5], s[2:3], s[0:1]
+; GFX1250-SDAG-NEXT:    s_add_nc_u64 s[0:1], s[0:1], 4
 ; GFX1250-SDAG-NEXT:    s_wait_dscnt 0x0
-; GFX1250-SDAG-NEXT:    flat_load_b32 v1, v0, s[2:3] scope:SCOPE_SYS
+; GFX1250-SDAG-NEXT:    flat_load_b32 v1, v0, s[4:5] scope:SCOPE_SYS
 ; GFX1250-SDAG-NEXT:    s_wait_loadcnt 0x0
-; GFX1250-SDAG-NEXT:    s_add_co_i32 s0, s0, -1
-; GFX1250-SDAG-NEXT:    s_add_nc_u64 s[2:3], s[2:3], 4
-; GFX1250-SDAG-NEXT:    s_cmp_eq_u32 s0, 0
+; GFX1250-SDAG-NEXT:    s_cmp_eq_u32 s0, 0x400
 ; GFX1250-SDAG-NEXT:    s_cbranch_scc0 .LBB116_1
 ; GFX1250-SDAG-NEXT:  ; %bb.2: ; %bb2
 ; GFX1250-SDAG-NEXT:    s_endpgm
@@ -2946,16 +2947,17 @@ define amdgpu_ps void @flat_addr_64bit_lsr_iv(ptr inreg %arg) {
 ; GFX1250-GISEL:       ; %bb.0: ; %bb
 ; GFX1250-GISEL-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
 ; GFX1250-GISEL-NEXT:    v_mov_b32_e32 v0, 0
-; GFX1250-GISEL-NEXT:    s_movk_i32 s0, 0x100
+; GFX1250-GISEL-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX1250-GISEL-NEXT:  .LBB116_1: ; %bb3
 ; GFX1250-GISEL-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX1250-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1250-GISEL-NEXT:    s_add_co_u32 s4, s2, s0
+; GFX1250-GISEL-NEXT:    s_add_co_ci_u32 s5, s3, s1
+; GFX1250-GISEL-NEXT:    s_add_nc_u64 s[0:1], s[0:1], 4
 ; GFX1250-GISEL-NEXT:    s_wait_dscnt 0x0
-; GFX1250-GISEL-NEXT:    flat_load_b32 v1, v0, s[2:3] scope:SCOPE_SYS
+; GFX1250-GISEL-NEXT:    flat_load_b32 v1, v0, s[4:5] scope:SCOPE_SYS
 ; GFX1250-GISEL-NEXT:    s_wait_loadcnt 0x0
-; GFX1250-GISEL-NEXT:    s_add_co_i32 s0, s0, -1
-; GFX1250-GISEL-NEXT:    s_add_co_u32 s2, s2, 4
-; GFX1250-GISEL-NEXT:    s_add_co_ci_u32 s3, s3, 0
-; GFX1250-GISEL-NEXT:    s_cmp_eq_u32 s0, 0
+; GFX1250-GISEL-NEXT:    s_cmp_eq_u32 s0, 0x400
 ; GFX1250-GISEL-NEXT:    s_cbranch_scc0 .LBB116_1
 ; GFX1250-GISEL-NEXT:  ; %bb.2: ; %bb2
 ; GFX1250-GISEL-NEXT:    s_endpgm
@@ -2964,15 +2966,16 @@ define amdgpu_ps void @flat_addr_64bit_lsr_iv(ptr inreg %arg) {
 ; GFX1250-NOECC:       ; %bb.0: ; %bb
 ; GFX1250-NOECC-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
 ; GFX1250-NOECC-NEXT:    v_mov_b32_e32 v0, 0
-; GFX1250-NOECC-NEXT:    s_movk_i32 s0, 0x100
+; GFX1250-NOECC-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX1250-NOECC-NEXT:  .LBB116_1: ; %bb3
 ; GFX1250-NOECC-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX1250-NOECC-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1250-NOECC-NEXT:    s_add_nc_u64 s[4:5], s[2:3], s[0:1]
+; GFX1250-NOECC-NEXT:    s_add_nc_u64 s[0:1], s[0:1], 4
 ; GFX1250-NOECC-NEXT:    s_wait_dscnt 0x0
-; GFX1250-NOECC-NEXT:    flat_load_b32 v1, v0, s[2:3] scope:SCOPE_SYS
+; GFX1250-NOECC-NEXT:    flat_load_b32 v1, v0, s[4:5] scope:SCOPE_SYS
 ; GFX1250-NOECC-NEXT:    s_wait_loadcnt 0x0
-; GFX1250-NOECC-NEXT:    s_add_co_i32 s0, s0, -1
-; GFX1250-NOECC-NEXT:    s_add_nc_u64 s[2:3], s[2:3], 4
-; GFX1250-NOECC-NEXT:    s_cmp_eq_u32 s0, 0
+; GFX1250-NOECC-NEXT:    s_cmp_eq_u32 s0, 0x400
 ; GFX1250-NOECC-NEXT:    s_cbranch_scc0 .LBB116_1
 ; GFX1250-NOECC-NEXT:  ; %bb.2: ; %bb2
 ; GFX1250-NOECC-NEXT:    s_endpgm
@@ -2999,17 +3002,19 @@ define amdgpu_ps void @flat_addr_64bit_lsr_iv_multiload(ptr inreg %arg, ptr inre
 ; GFX1250-SDAG:       ; %bb.0: ; %bb
 ; GFX1250-SDAG-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
 ; GFX1250-SDAG-NEXT:    v_mov_b32_e32 v0, 0
-; GFX1250-SDAG-NEXT:    s_movk_i32 s0, 0x100
+; GFX1250-SDAG-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX1250-SDAG-NEXT:  .LBB117_1: ; %bb3
 ; GFX1250-SDAG-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX1250-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1250-SDAG-NEXT:    s_add_nc_u64 s[4:5], s[2:3], s[0:1]
+; GFX1250-SDAG-NEXT:    s_add_nc_u64 s[0:1], s[0:1], 4
 ; GFX1250-SDAG-NEXT:    s_wait_dscnt 0x0
-; GFX1250-SDAG-NEXT:    flat_load_b32 v1, v0, s[2:3] scope:SCOPE_SYS
+; GFX1250-SDAG-NEXT:    flat_load_b32 v1, v0, s[4:5] scope:SCOPE_SYS
 ; GFX1250-SDAG-NEXT:    s_wait_loadcnt_dscnt 0x0
-; GFX1250-SDAG-NEXT:    flat_load_b32 v1, v0, s[2:3] scope:SCOPE_SYS
+; GFX1250-SDAG-NEXT:    flat_load_b32 v1, v0, s[4:5] scope:SCOPE_SYS
 ; GFX1250-SDAG-NEXT:    s_wait_loadcnt 0x0
-; GFX1250-SDAG-NEXT:    s_add_co_i32 s0, s0, -1
-; GFX1250-SDAG-NEXT:    s_add_nc_u64 s[2:3], s[2:3], 4
-; GFX1250-SDAG-NEXT:    s_cmp_eq_u32 s0, 0
+; GFX1250-SDAG-NEXT:    s_cmp_eq_u32 s0, 0x400
+; GFX1250-SDAG-NEXT:    ; kill: killed $sgpr4_sgpr5
 ; GFX1250-SDAG-NEXT:    s_cbranch_scc0 .LBB117_1
 ; GFX1250-SDAG-NEXT:  ; %bb.2: ; %bb2
 ; GFX1250-SDAG-NEXT:    s_endpgm
@@ -3018,18 +3023,20 @@ define amdgpu_ps void @flat_addr_64bit_lsr_iv_multiload(ptr inreg %arg, ptr inre
 ; GFX1250-GISEL:       ; %bb.0: ; %bb
 ; GFX1250-GISEL-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
 ; GFX1250-GISEL-NEXT:    v_mov_b32_e32 v0, 0
-; GFX1250-GISEL-NEXT:    s_movk_i32 s0, 0x100
+; GFX1250-GISEL-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX1250-GISEL-NEXT:  .LBB117_1: ; %bb3
 ; GFX1250-GISEL-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX1250-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1250-GISEL-NEXT:    s_add_co_u32 s4, s2, s0
+; GFX1250-GISEL-NEXT:    s_add_co_ci_u32 s5, s3, s1
+; GFX1250-GISEL-NEXT:    s_add_nc_u64 s[0:1], s[0:1], 4
 ; GFX1250-GISEL-NEXT:    s_wait_dscnt 0x0
-; GFX1250-GISEL-NEXT:    flat_load_b32 v1, v0, s[2:3] scope:SCOPE_SYS
+; GFX1250-GISEL-NEXT:    flat_load_b32 v1, v0, s[4:5] scope:SCOPE_SYS
 ; GFX1250-GISEL-NEXT:    s_wait_loadcnt_dscnt 0x0
-; GFX1250-GISEL-NEXT:    flat_load_b32 v1, v0, s[2:3] scope:SCOPE_SYS
+; GFX1250-GISEL-NEXT:    flat_load_b32 v1, v0, s[4:5] scope:SCOPE_SYS
 ; GFX1250-GISEL-NEXT:    s_wait_loadcnt 0x0
-; GFX1250-GISEL-NEXT:    s_add_co_i32 s0, s0, -1
-; GFX1250-GISEL-NEXT:    s_add_co_u32 s2, s2, 4
-; GFX1250-GISEL-NEXT:    s_add_co_ci_u32 s3, s3, 0
-; GFX1250-GISEL-NEXT:    s_cmp_eq_u32 s0, 0
+; GFX1250-GISEL-NEXT:    s_cmp_eq_u32 s0, 0x400
+; GFX1250-GISEL-NEXT:    ; kill: killed $sgpr4 killed $sgpr5
 ; GFX1250-GISEL-NEXT:    s_cbranch_scc0 .LBB117_1
 ; GFX1250-GISEL-NEXT:  ; %bb.2: ; %bb2
 ; GFX1250-GISEL-NEXT:    s_endpgm
@@ -3038,17 +3045,19 @@ define amdgpu_ps void @flat_addr_64bit_lsr_iv_multiload(ptr inreg %arg, ptr inre
 ; GFX1250-NOECC:       ; %bb.0: ; %bb
 ; GFX1250-NOECC-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
 ; GFX1250-NOECC-NEXT:    v_mov_b32_e32 v0, 0
-; GFX1250-NOECC-NEXT:    s_movk_i32 s0, 0x100
+; GFX1250-NOECC-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX1250-NOECC-NEXT:  .LBB117_1: ; %bb3
 ; GFX1250-NOECC-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX1250-NOECC-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1250-NOECC-NEXT:    s_add_nc_u64 s[4:5], s[2:3], s[0:1]
+; GFX1250-NOECC-NEXT:    s_add_nc_u64 s[0:1], s[0:1], 4
 ; GFX1250-NOECC-NEXT:    s_wait_dscnt 0x0
-; GFX1250-NOECC-NEXT:    flat_load_b32 v1, v0, s[2:3] scope:SCOPE_SYS
+; GFX1250-NOECC-NEXT:    flat_load_b32 v1, v0, s[4:5] scope:SCOPE_SYS
 ; GFX1250-NOECC-NEXT:    s_wait_loadcnt_dscnt 0x0
-; GFX1250-NOECC-NEXT:    flat_load_b32 v1, v0, s[2:3] scope:SCOPE_SYS
+; GFX1250-NOECC-NEXT:    flat_load_b32 v1, v0, s[4:5] scope:SCOPE_SYS
 ; GFX1250-NOECC-NEXT:    s_wait_loadcnt 0x0
-; GFX1250-NOECC-NEXT:    s_add_co_i32 s0, s0, -1
-; GFX1250-NOECC-NEXT:    s_add_nc_u64 s[2:3], s[2:3], 4
-; GFX1250-NOECC-NEXT:    s_cmp_eq_u32 s0, 0
+; GFX1250-NOECC-NEXT:    s_cmp_eq_u32 s0, 0x400
+; GFX1250-NOECC-NEXT:    ; kill: killed $sgpr4_sgpr5
 ; GFX1250-NOECC-NEXT:    s_cbranch_scc0 .LBB117_1
 ; GFX1250-NOECC-NEXT:  ; %bb.2: ; %bb2
 ; GFX1250-NOECC-NEXT:    s_endpgm
diff --git a/llvm/test/CodeGen/AMDGPU/global-saddr-load.ll b/llvm/test/CodeGen/AMDGPU/global-saddr-load.ll
index a72306dcd381a..bd7505cc2b4d2 100644
--- a/llvm/test/CodeGen/AMDGPU/global-saddr-load.ll
+++ b/llvm/test/CodeGen/AMDGPU/global-saddr-load.ll
@@ -4720,16 +4720,17 @@ define amdgpu_ps float @global_load_saddr_i8_offset_or_i64_imm_offset_4160(ptr a
 define amdgpu_ps void @global_addr_64bit_lsr_iv(ptr addrspace(1) inreg %arg) {
 ; GFX9-LABEL: global_addr_64bit_lsr_iv:
 ; GFX9:       ; %bb.0: ; %bb
-; GFX9-NEXT:    s_movk_i32 s0, 0x100
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-NEXT:  .LBB132_1: ; %bb3
 ; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX9-NEXT:    global_load_dword v1, v0, s[2:3] glc
+; GFX9-NEXT:    s_add_u32 s4, s2, s0
+; GFX9-NEXT:    s_addc_u32 s5, s3, s1
+; GFX9-NEXT:    global_load_dword v1, v0, s[4:5] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    s_add_i32 s0, s0, -1
-; GFX9-NEXT:    s_add_u32 s2, s2, 4
-; GFX9-NEXT:    s_addc_u32 s3, s3, 0
-; GFX9-NEXT:    s_cmp_eq_u32 s0, 0
+; GFX9-NEXT:    s_add_u32 s0, s0, 4
+; GFX9-NEXT:    s_addc_u32 s1, s1, 0
+; GFX9-NEXT:    s_cmpk_eq_i32 s0, 0x400
 ; GFX9-NEXT:    s_cbranch_scc0 .LBB132_1
 ; GFX9-NEXT:  ; %bb.2: ; %bb2
 ; GFX9-NEXT:    s_endpgm
@@ -4737,16 +4738,17 @@ define amdgpu_ps void @global_addr_64bit_lsr_iv(ptr addrspace(1) inreg %arg) {
 ; GFX10-LABEL: global_addr_64bit_lsr_iv:
 ; GFX10:       ; %bb.0: ; %bb
 ; GFX10-NEXT:    v_mov_b32_e32 v0, 0
-; GFX10-NEXT:    s_movk_i32 s0, 0x100
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX10-NEXT:  .LBB132_1: ; %bb3
 ; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT:    global_load_dword v1, v0, s[2:3] glc dlc
-; GFX10-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-NEXT:    s_add_i32 s0, s0, -1
 ; GFX10-NEXT:    s_waitcnt_depctr depctr_vm_vsrc(0)
-; GFX10-NEXT:    s_add_u32 s2, s2, 4
-; GFX10-NEXT:    s_addc_u32 s3, s3, 0
-; GFX10-NEXT:    s_cmp_eq_u32 s0, 0
+; GFX10-NEXT:    s_add_u32 s4, s2, s0
+; GFX10-NEXT:    s_addc_u32 s5, s3, s1
+; GFX10-NEXT:    s_add_u32 s0, s0, 4
+; GFX10-NEXT:    global_load_dword v1, v0, s[4:5] glc dlc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-NEXT:    s_cmpk_eq_i32 s0, 0x400
 ; GFX10-NEXT:    s_cbranch_scc0 .LBB132_1
 ; GFX10-NEXT:  ; %bb.2: ; %bb2
 ; GFX10-NEXT:    s_endpgm
@@ -4754,15 +4756,17 @@ define amdgpu_ps void @global_addr_64bit_lsr_iv(ptr addrspace(1) inreg %arg) {
 ; GFX11-LABEL: global_addr_64bit_lsr_iv:
 ; GFX11:       ; %bb.0: ; %bb
 ; GFX11-NEXT:    v_mov_b32_e32 v0, 0
-; GFX11-NEXT:    s_movk_i32 s0, 0x100
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX11-NEXT:  .LBB132_1: ; %bb3
 ; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT:    global_load_b32 v1, v0, s[2:3] glc dlc
-; GFX11-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-NEXT:    s_add_i32 s0, s0, -1
-; GFX11-NEXT:    s_add_u32 s2, s2, 4
-; GFX11-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-NEXT:    s_cmp_eq_u32 s0, 0
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_add_u32 s4, s2, s0
+; GFX11-NEXT:    s_addc_u32 s5, s3, s1
+; GFX11-NEXT:    s_add_u32 s0, s0, 4
+; GFX11-NEXT:    global_load_b32 v1, v0, s[4:5] glc dlc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_addc_u32 s1, s1, 0
+; GFX11-NEXT:    s_cmpk_eq_i32 s0, 0x400
 ; GFX11-NEXT:    s_cbranch_scc0 .LBB132_1
 ; GFX11-NEXT:  ; %bb.2: ; %bb2
 ; GFX11-NEXT:    s_endpgm
@@ -4770,14 +4774,15 @@ define amdgpu_ps void @global_addr_64bit_lsr_iv(ptr addrspace(1) inreg %arg) {
 ; GFX12-SDAG-LABEL: global_addr_64bit_lsr_iv:
 ; GFX12-SDAG:       ; %bb.0: ; %bb
 ; GFX12-SDAG-NEXT:    v_mov_b32_e32 v0, 0
-; GFX12-SDAG-NEXT:    s_movk_i32 s0, 0x100
+; GFX12-SDAG-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX12-SDAG-NEXT:  .LBB132_1: ; %bb3
 ; GFX12-SDAG-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX12-SDAG-NEXT:    global_load_b32 v1, v0, s[2:3] scope:SCOPE_SYS
+; GFX12-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-SDAG-NEXT:    s_add_nc_u64 s[4:5], s[2:3], s[0:1]
+; GFX12-SDAG-NEXT:    s_add_nc_u64 s[0:1], s[0:1], 4
+; GFX12-SDAG-NEXT:    global_load_b32 v1, v0, s[4:5] scope:SCOPE_SYS
 ; GFX12-SDAG-NEXT:    s_wait_loadcnt 0x0
-; GFX12-SDAG-NEXT:    s_add_co_i32 s0, s0, -1
-; GFX12-SDAG-NEXT:    s_add_nc_u64 s[2:3], s[2:3], 4
-; GFX12-SDAG-NEXT:    s_cmp_eq_u32 s0, 0
+; GFX12-SDAG-NEXT:    s_cmp_eq_u32 s0, 0x400
 ; GFX12-SDAG-NEXT:    s_cbranch_scc0 .LBB132_1
 ; GFX12-SDAG-NEXT:  ; %bb.2: ; %bb2
 ; GFX12-SDAG-NEXT:    s_endpgm
@@ -4785,15 +4790,16 @@ define amdgpu_ps void @global_addr_64bit_lsr_iv(ptr addrspace(1) inreg %arg) {
 ; GFX12-GISEL-LABEL: global_addr_64bit_lsr_iv:
 ; GFX12-GISEL:       ; %bb.0: ; %bb
 ; GFX12-GISEL-NEXT:    v_mov_b32_e32 v0, 0
-; GFX12-GISEL-NEXT:    s_movk_i32 s0, 0x100
+; GFX12-GISEL-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX12-GISEL-NEXT:  .LBB132_1: ; %bb3
 ; GFX12-GISEL-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX12-GISEL-NEXT:    global_load_b32 v1, v0, s[2:3] scope:SCOPE_SYS
+; GFX12-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-GISEL-NEXT:    s_add_co_u32 s4, s2, s0
+; GFX12-GISEL-NEXT:    s_add_co_ci_u32 s5, s3, s1
+; GFX12-GISEL-NEXT:    s_add_nc_u64 s[0:1], s[0:1], 4
+; GFX12-GISEL-NEXT:    global_load_b32 v1, v0, s[4:5] scope:SCOPE_SYS
 ; GFX12-GISEL-NEXT:    s_wait_loadcnt 0x0
-; GFX12-GISEL-NEXT:    s_add_co_i32 s0, s0, -1
-; GFX12-GISEL-NEXT:    s_add_co_u32 s2, s2, 4
-; GFX12-GISEL-NEXT:    s_add_co_ci_u32 s3, s3, 0
-; GFX12-GISEL-NEXT:    s_cmp_eq_u32 s0, 0
+; GFX12-GISEL-NEXT:    s_cmp_eq_u32 s0, 0x400
 ; GFX12-GISEL-NEXT:    s_cbranch_scc0 .LBB132_1
 ; GFX12-GISEL-NEXT:  ; %bb.2: ; %bb2
 ; GFX12-GISEL-NEXT:    s_endpgm
@@ -4818,18 +4824,20 @@ bb3:                                              ; preds = %bb3, %bb
 define amdgpu_ps void @global_addr_64bit_lsr_iv_multiload(ptr addrspace(1) inreg %arg, ptr addrspace(1) inreg %arg.1) {
 ; GFX9-LABEL: global_addr_64bit_lsr_iv_multiload:
 ; GFX9:       ; %bb.0: ; %bb
-; GFX9-NEXT:    s_movk_i32 s0, 0x100
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX9-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-NEXT:  .LBB133_1: ; %bb3
 ; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX9-NEXT:    global_load_dword v1, v0, s[2:3] glc
+; GFX9-NEXT:    s_add_u32 s4, s2, s0
+; GFX9-NEXT:    s_addc_u32 s5, s3, s1
+; GFX9-NEXT:    global_load_dword v1, v0, s[4:5] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    global_load_dword v1, v0, s[2:3] glc
+; GFX9-NEXT:    global_load_dword v1, v0, s[4:5] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    s_add_i32 s0, s0, -1
-; GFX9-NEXT:    s_add_u32 s2, s2, 4
-; GFX9-NEXT:    s_addc_u32 s3, s3, 0
-; GFX9-NEXT:    s_cmp_eq_u32 s0, 0
+; GFX9-NEXT:    s_add_u32 s0, s0, 4
+; GFX9-NEXT:    s_addc_u32 s1, s1, 0
+; GFX9-NEXT:    s_cmpk_eq_i32 s0, 0x400
+; GFX9-NEXT:    ; kill: killed $sgpr4 killed $sgpr5
 ; GFX9-NEXT:    s_cbranch_scc0 .LBB133_1
 ; GFX9-NEXT:  ; %bb.2: ; %bb2
 ; GFX9-NEXT:    s_endpgm
@@ -4837,18 +4845,20 @@ define amdgpu_ps void @global_addr_64bit_lsr_iv_multiload(ptr addrspace(1) inreg
 ; GFX10-LABEL: global_addr_64bit_lsr_iv_multiload:
 ; GFX10:       ; %bb.0: ; %bb
 ; GFX10-NEXT:    v_mov_b32_e32 v0, 0
-; GFX10-NEXT:    s_movk_i32 s0, 0x100
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX10-NEXT:  .LBB133_1: ; %bb3
 ; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX10-NEXT:    global_load_dword v1, v0, s[2:3] glc dlc
+; GFX10-NEXT:    s_waitcnt_depctr depctr_vm_vsrc(0)
+; GFX10-NEXT:    s_add_u32 s4, s2, s0
+; GFX10-NEXT:    s_addc_u32 s5, s3, s1
+; GFX10-NEXT:    s_add_u32 s0, s0, 4
+; GFX10-NEXT:    global_load_dword v1, v0, s[4:5] glc dlc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-NEXT:    global_load_dword v1, v0, s[2:3] glc dlc
+; GFX10-NEXT:    global_load_dword v1, v0, s[4:5] glc dlc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-NEXT:    s_add_i32 s0, s0, -1
-; GFX10-NEXT:    s_waitcnt_depctr depctr_vm_vsrc(0)
-; GFX10-NEXT:    s_add_u32 s2, s2, 4
-; GFX10-NEXT:    s_addc_u32 s3, s3, 0
-; GFX10-NEXT:    s_cmp_eq_u32 s0, 0
+; GFX10-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-NEXT:    s_cmpk_eq_i32 s0, 0x400
+; GFX10-NEXT:    ; kill: killed $sgpr4 killed $sgpr5
 ; GFX10-NEXT:    s_cbranch_scc0 .LBB133_1
 ; GFX10-NEXT:  ; %bb.2: ; %bb2
 ; GFX10-NEXT:    s_endpgm
@@ -4856,17 +4866,19 @@ define amdgpu_ps void @global_addr_64bit_lsr_iv_multiload(ptr addrspace(1) inreg
 ; GFX11-LABEL: global_addr_64bit_lsr_iv_multiload:
 ; GFX11:       ; %bb.0: ; %bb
 ; GFX11-NEXT:    v_mov_b32_e32 v0, 0
-; GFX11-NEXT:    s_movk_i32 s0, 0x100
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX11-NEXT:  .LBB133_1: ; %bb3
 ; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT:    global_load_b32 v1, v0, s[2:3] glc dlc
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_add_u32 s4, s2, s0
+; GFX11-NEXT:    s_addc_u32 s5, s3, s1
+; GFX11-NEXT:    s_add_u32 s0, s0, 4
+; GFX11-NEXT:    global_load_b32 v1, v0, s[4:5] glc dlc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-NEXT:    global_load_b32 v1, v0, s[2:3] glc dlc
+; GFX11-NEXT:    global_load_b32 v1, v0, s[4:5] glc dlc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-NEXT:    s_add_i32 s0, s0, -1
-; GFX11-NEXT:    s_add_u32 s2, s2, 4
-; GFX11-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-NEXT:    s_cmp_eq_u32 s0, 0
+; GFX11-NEXT:    s_addc_u32 s1, s1, 0
+; GFX11-NEXT:    s_cmpk_eq_i32 s0, 0x400
 ; GFX11-NEXT:    s_cbranch_scc0 .LBB133_1
 ; GFX11-NEXT:  ; %bb.2: ; %bb2
 ; GFX11-NEXT:    s_endpgm
@@ -4874,16 +4886,17 @@ define amdgpu_ps void @global_addr_64bit_lsr_iv_multiload(ptr addrspace(1) inreg
 ; GFX12-SDAG-LABEL: global_addr_64bit_lsr_iv_multiload:
 ; GFX12-SDAG:       ; %bb.0: ; %bb
 ; GFX12-SDAG-NEXT:    v_mov_b32_e32 v0, 0
-; GFX12-SDAG-NEXT:    s_movk_i32 s0, 0x100
+; GFX12-SDAG-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX12-SDAG-NEXT:  .LBB133_1: ; %bb3
 ; GFX12-SDAG-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX12-SDAG-NEXT:    global_load_b32 v1, v0, s[2:3] scope:SCOPE_SYS
+; GFX12-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-SDAG-NEXT:    s_add_nc_u64 s[4:5], s[2:3], s[0:1]
+; GFX12-SDAG-NEXT:    s_add_nc_u64 s[0:1], s[0:1], 4
+; GFX12-SDAG-NEXT:    global_load_b32 v1, v0, s[4:5] scope:SCOPE_SYS
 ; GFX12-SDAG-NEXT:    s_wait_loadcnt 0x0
-; GFX12-SDAG-NEXT:    global_load_b32 v1, v0, s[2:3] scope:SCOPE_SYS
+; GFX12-SDAG-NEXT:    global_load_b32 v1, v0, s[4:5] scope:SCOPE_SYS
 ; GFX12-SDAG-NEXT:    s_wait_loadcnt 0x0
-; GFX12-SDAG-NEXT:    s_add_co_i32 s0, s0, -1
-; GFX12-SDAG-NEXT:    s_add_nc_u64 s[2:3], s[2:3], 4
-; GFX12-SDAG-NEXT:    s_cmp_eq_u32 s0, 0
+; GFX12-SDAG-NEXT:    s_cmp_eq_u32 s0, 0x400
 ; GFX12-SDAG-NEXT:    s_cbranch_scc0 .LBB133_1
 ; GFX12-SDAG-NEXT:  ; %bb.2: ; %bb2
 ; GFX12-SDAG-NEXT:    s_endpgm
@@ -4891,17 +4904,18 @@ define amdgpu_ps void @global_addr_64bit_lsr_iv_multiload(ptr addrspace(1) inreg
 ; GFX12-GISEL-LABEL: global_addr_64bit_lsr_iv_multiload:
 ; GFX12-GISEL:       ; %bb.0: ; %bb
 ; GFX12-GISEL-NEXT:    v_mov_b32_e32 v0, 0
-; GFX12-GISEL-NEXT:    s_movk_i32 s0, 0x100
+; GFX12-GISEL-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX12-GISEL-NEXT:  .LBB133_1: ; %bb3
 ; GFX12-GISEL-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX12-GISEL-NEXT:    global_load_b32 v1, v0, s[2:3] scope:SCOPE_SYS
+; GFX12-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-GISEL-NEXT:    s_add_co_u32 s4, s2, s0
+; GFX12-GISEL-NEXT:    s_add_co_ci_u32 s5, s3, s1
+; GFX12-GISEL-NEXT:    s_add_nc_u64 s[0:1], s[0:1], 4
+; GFX12-GISEL-NEXT:    global_load_b32 v1, v0, s[4:5] scope:SCOPE_SYS
 ; GFX12-GISEL-NEXT:    s_wait_loadcnt 0x0
-; GFX12-GISEL-NEXT:    global_load_b32 v1, v0, s[2:3] scope:SCOPE_SYS
+; GFX12-GISEL-NEXT:    global_load_b32 v1, v0, s[4:5] scope:SCOPE_SYS
 ; GFX12-GISEL-NEXT:    s_wait_loadcnt 0x0
-; GFX12-GISEL-NEXT:    s_add_co_i32 s0, s0, -1
-; GFX12-GISEL-NEXT:    s_add_co_u32 s2, s2, 4
-; GFX12-GISEL-NEXT:    s_add_co_ci_u32 s3, s3, 0
-; GFX12-GISEL-NEXT:    s_cmp_eq_u32 s0, 0
+; GFX12-GISEL-NEXT:    s_cmp_eq_u32 s0, 0x400
 ; GFX12-GISEL-NEXT:    s_cbranch_scc0 .LBB133_1
 ; GFX12-GISEL-NEXT:  ; %bb.2: ; %bb2
 ; GFX12-GISEL-NEXT:    s_endpgm
diff --git a/llvm/test/CodeGen/AMDGPU/lsr-cost-model-occupancy.ll b/llvm/test/CodeGen/AMDGPU/lsr-cost-model-occupancy.ll
new file mode 100644
index 0000000000000..360a5bce0e21a
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/lsr-cost-model-occupancy.ll
@@ -0,0 +1,375 @@
+; RUN: llc -O3 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx950 < %s | FileCheck %s
+
+; Reduced from the AITER Triton paged-attention _fwd_kernel regressed by the
+; GFX9+ LSR cost model (#184138). The old EffInsns-first comparator makes LSR
+; add IV registers (260 VGPRs, Occupancy 1); the NumRegs-first fix keeps
+; register pressure at 244 VGPRs, restoring Occupancy 2.
+
+; CHECK-LABEL: {{^}}lsr_cost_occupancy:
+; CHECK: ; Occupancy: 2
+
+target triple = "amdgcn-amd-amdhsa"
+
+define amdgpu_kernel void @lsr_cost_occupancy(i32 %arg2, ptr addrspace(3) %0, ptr addrspace(3) %1, ptr addrspace(3) %2, ptr addrspace(3) %3, ptr addrspace(3) %4, ptr addrspace(3) %5, ptr addrspace(3) %6, <8 x float> %i469, <4 x float> %i757, <2 x float> %i835, i1 %i1118) #0 {
+.lr.ph:
+  %i90 = load <8 x half>, ptr addrspace(3) null, align 16
+  %i93 = load <8 x half>, ptr addrspace(3) %0, align 16
+  %i96 = load <8 x half>, ptr addrspace(3) %1, align 16
+  %i99 = load <8 x half>, ptr addrspace(3) %2, align 16
+  %i102 = load <8 x half>, ptr addrspace(3) %3, align 16
+  %i105 = load <8 x half>, ptr addrspace(3) %4, align 16
+  %i108 = load <8 x half>, ptr addrspace(3) %5, align 16
+  %i111 = load <8 x half>, ptr addrspace(3) %6, align 16
+  %i = tail call i32 @llvm.amdgcn.workitem.id.x()
+  %i121 = insertelement <4 x i32> zeroinitializer, i32 %i, i64 0
+  %i1224 = shufflevector <4 x i32> %i121, <4 x i32> zeroinitializer, <4 x i32> zeroinitializer
+  %i128 = or <4 x i32> %i1224, splat (i32 1)
+  %i123 = or <4 x i32> %i1224, <i32 8, i32 9, i32 10, i32 11>
+  %i136 = extractelement <4 x i32> %i123, i64 1
+  %i413 = and i32 %i, 1
+  %i140 = extractelement <4 x i32> %i123, i64 3
+  %i124 = insertelement <8 x i32> zeroinitializer, i32 %i, i64 0
+  %i125 = shufflevector <8 x i32> %i124, <8 x i32> zeroinitializer, <8 x i32> zeroinitializer
+  %i126 = or <8 x i32> %i125, <i32 17, i32 18, i32 19, i32 24, i32 25, i32 26, i32 27, i32 32>
+  %i144 = extractelement <8 x i32> %i126, i64 1
+  %i146 = extractelement <8 x i32> %i126, i64 2
+  %i148 = extractelement <8 x i32> %i126, i64 3
+  %i150 = extractelement <8 x i32> %i126, i64 4
+  %i152 = extractelement <8 x i32> %i126, i64 5
+  %i154 = extractelement <8 x i32> %i126, i64 6
+  %i127 = or <8 x i32> %i125, <i32 34, i32 35, i32 40, i32 41, i32 42, i32 43, i32 48, i32 49>
+  %i160 = extractelement <8 x i32> %i127, i64 1
+  %i162 = extractelement <8 x i32> %i127, i64 2
+  %i164 = extractelement <8 x i32> %i127, i64 3
+  %i166 = extractelement <8 x i32> %i127, i64 4
+  %i168 = extractelement <8 x i32> %i127, i64 5
+  %i117 = lshr i32 %i, 3
+  br label %bb
+
+bb:                                               ; preds = %bb, %.lr.ph
+  %i472 = phi float [ 0.000000e+00, %.lr.ph ], [ %i929, %bb ]
+  %i473 = phi float [ 0.000000e+00, %.lr.ph ], [ 1.000000e+00, %bb ]
+  %i474 = phi i64 [ 0, %.lr.ph ], [ %i111731, %bb ]
+  %i475 = phi <2 x float> [ zeroinitializer, %.lr.ph ], [ %i1119, %bb ]
+  %i490 = phi <2 x float> [ zeroinitializer, %.lr.ph ], [ %i1134, %bb ]
+  %i25 = or disjoint i32 %i, 1
+  %i114 = zext i32 %i25 to i64
+  %i504 = add i64 %i114, %i474
+  %i20 = or disjoint i32 %arg2, %i
+  %i113 = zext i32 %i20 to i64
+  %i492 = or disjoint i64 %i474, %i113
+  %i497 = icmp slt i64 %i492, 0
+  %i503 = add i64 %i113, %i474
+  %i507 = trunc i64 %i503 to i32
+  %i498 = icmp slt i64 %i474, 1
+  %i508 = trunc i64 %i504 to i32
+  %i517 = shl i32 %i508, 3
+  %i518 = select i1 %i498, i32 %i517, i32 -2147483648
+  %i519 = tail call i64 @llvm.amdgcn.raw.ptr.buffer.load.i64(ptr addrspace(8) null, i32 %i518, i32 0, i32 0)
+  %i134 = extractelement <4 x i32> %i123, i64 0
+  %i135 = zext i32 %i134 to i64
+  %i530 = or disjoint i64 %i474, %i135
+  %i719 = insertelement <8 x i64> zeroinitializer, i64 %i530, i64 4
+  %i137 = zext i32 %i136 to i64
+  %i531 = or disjoint i64 %i474, %i137
+  %i720 = insertelement <8 x i64> %i719, i64 %i531, i64 5
+  %i139 = zext i32 %i413 to i64
+  %i532 = or disjoint i64 %i474, %i139
+  %i721 = insertelement <8 x i64> %i720, i64 %i532, i64 6
+  %i141 = zext i32 %i140 to i64
+  %i533 = or disjoint i64 %i474, %i141
+  %i722 = insertelement <8 x i64> %i721, i64 %i533, i64 7
+  %i723 = icmp slt <8 x i64> %i722, zeroinitializer
+  %i142 = extractelement <8 x i32> %i126, i64 0
+  %i143 = zext i32 %i142 to i64
+  %i537 = or disjoint i64 %i474, %i143
+  %i727 = insertelement <8 x i64> zeroinitializer, i64 %i537, i64 0
+  %i145 = zext i32 %i144 to i64
+  %i538 = or disjoint i64 %i474, %i145
+  %i728 = insertelement <8 x i64> %i727, i64 %i538, i64 1
+  %i147 = zext i32 %i146 to i64
+  %i539 = or disjoint i64 %i474, %i147
+  %i729 = insertelement <8 x i64> %i728, i64 %i539, i64 2
+  %i149 = zext i32 %i148 to i64
+  %i540 = or disjoint i64 %i474, %i149
+  %i730 = insertelement <8 x i64> %i729, i64 %i540, i64 3
+  %i151 = zext i32 %i150 to i64
+  %i541 = or disjoint i64 %i474, %i151
+  %i731 = insertelement <8 x i64> %i730, i64 %i541, i64 4
+  %i153 = zext i32 %i152 to i64
+  %i542 = or disjoint i64 %i474, %i153
+  %i732 = insertelement <8 x i64> %i731, i64 %i542, i64 5
+  %i155 = zext i32 %i154 to i64
+  %i543 = or disjoint i64 %i474, %i155
+  %i733 = insertelement <8 x i64> %i732, i64 %i543, i64 6
+  %i156 = extractelement <8 x i32> %i126, i64 7
+  %i157 = zext i32 %i156 to i64
+  %i544 = or disjoint i64 %i474, %i157
+  %i158 = extractelement <8 x i32> %i127, i64 0
+  %i159 = zext i32 %i158 to i64
+  %i545 = or disjoint i64 %i474, %i159
+  %i739 = insertelement <8 x i64> zeroinitializer, i64 %i545, i64 0
+  %i161 = zext i32 %i160 to i64
+  %i546 = or disjoint i64 %i474, %i161
+  %i740 = insertelement <8 x i64> %i739, i64 %i546, i64 1
+  %i163 = zext i32 %i162 to i64
+  %i547 = or disjoint i64 %i474, %i163
+  %i741 = insertelement <8 x i64> %i740, i64 %i547, i64 2
+  %i165 = zext i32 %i164 to i64
+  %i548 = or disjoint i64 %i474, %i165
+  %i742 = insertelement <8 x i64> %i741, i64 %i548, i64 3
+  %i167 = zext i32 %i166 to i64
+  %i549 = or disjoint i64 %i474, %i167
+  %i743 = insertelement <8 x i64> %i742, i64 %i549, i64 4
+  %i169 = zext i32 %i168 to i64
+  %i550 = or disjoint i64 %i474, %i169
+  %i744 = insertelement <8 x i64> %i743, i64 %i550, i64 5
+  %i171 = zext i32 %i117 to i64
+  %i551 = or disjoint i64 %i474, %i171
+  %i745 = insertelement <8 x i64> %i744, i64 %i551, i64 6
+  %i173 = zext i32 %i to i64
+  %i552 = or disjoint i64 %i474, %i173
+  %i534 = insertelement <4 x i64> zeroinitializer, i64 %i474, i64 0
+  %i53515 = shufflevector <4 x i64> %i534, <4 x i64> zeroinitializer, <4 x i32> zeroinitializer
+  %i174 = zext <4 x i32> %i121 to <4 x i64>
+  %i553 = or disjoint <4 x i64> %i53515, %i174
+  %i758 = fmul <4 x float> zeroinitializer, %i757
+  %i751 = icmp slt <4 x i64> %i553, zeroinitializer
+  %i753 = select <4 x i1> %i751, <4 x float> zeroinitializer, <4 x float> splat (float +qnan)
+  %i807 = extractelement <4 x float> %i753, i64 2
+  %i809 = extractelement <4 x float> %i753, i64 1
+  %i810 = extractelement <4 x float> %i758, i64 0
+  %i811 = fmul float %i809, %i810
+  %i818 = fmul float %i807, %i811
+  %i82220 = bitcast float %i818 to i32
+  %i823 = tail call { i32, i32 } @llvm.amdgcn.permlane32.swap(i32 %i82220, i32 0, i1 false, i1 false)
+  %i824 = extractvalue { i32, i32 } %i823, 0
+  %i826 = bitcast i32 %i824 to float
+  %i82821 = tail call float @llvm.maxnum.f32(float %i826, float 0.000000e+00)
+  %i725 = select <8 x i1> %i723, <8 x float> zeroinitializer, <8 x float> splat (float -inf)
+  %i72617 = fmul <8 x float> %i469, %i725
+  %i829 = shufflevector <8 x float> %i72617, <8 x float> zeroinitializer, <2 x i32> <i32 0, i32 1>
+  %i830 = insertelement <2 x float> zeroinitializer, float %i82821, i64 0
+  %i83122 = shufflevector <2 x float> %i830, <2 x float> zeroinitializer, <2 x i32> zeroinitializer
+  %i832 = fsub <2 x float> %i829, %i83122
+  %i833 = fmul <2 x float> %i832, zeroinitializer
+  %i834 = tail call <2 x float> @llvm.exp2.v2f32(<2 x float> %i833)
+  %i836 = fsub <2 x float> %i835, %i83122
+  %i837 = fmul <2 x float> %i836, zeroinitializer
+  %i838 = tail call <2 x float> @llvm.exp2.v2f32(<2 x float> %i837)
+  %i839 = shufflevector <8 x float> %i72617, <8 x float> zeroinitializer, <2 x i32> <i32 4, i32 5>
+  %i840 = fsub <2 x float> %i839, splat (float 1.000000e+00)
+  %i841 = fmul <2 x float> %i840, zeroinitializer
+  %i842 = tail call <2 x float> @llvm.exp2.v2f32(<2 x float> %i841)
+  %i843 = shufflevector <8 x float> %i72617, <8 x float> zeroinitializer, <2 x i32> <i32 6, i32 7>
+  %i844 = fsub <2 x float> %i843, %i83122
+  %i845 = fmul <2 x float> %i844, zeroinitializer
+  %i846 = tail call <2 x float> @llvm.exp2.v2f32(<2 x float> %i845)
+  %i734 = insertelement <8 x i64> %i733, i64 %i544, i64 7
+  %i735 = icmp slt <8 x i64> %i734, zeroinitializer
+  %i737 = select <8 x i1> %i735, <8 x float> zeroinitializer, <8 x float> splat (float 1.000000e+00)
+  %i848 = shufflevector <8 x float> zeroinitializer, <8 x float> %i737, <2 x i32> <i32 0, i32 8>
+  %i849 = fsub <2 x float> %i848, %i83122
+  %i850 = fmul <2 x float> %i849, zeroinitializer
+  %i851 = tail call <2 x float> @llvm.exp2.v2f32(<2 x float> %i850)
+  %i738 = fmul <8 x float> %i469, %i737
+  %i852 = shufflevector <8 x float> %i738, <8 x float> zeroinitializer, <2 x i32> <i32 1, i32 2>
+  %i853 = fsub <2 x float> %i852, splat (float 1.000000e+00)
+  %i854 = fmul <2 x float> %i853, zeroinitializer
+  %i855 = tail call <2 x float> @llvm.exp2.v2f32(<2 x float> %i854)
+  %i856 = shufflevector <8 x float> %i737, <8 x float> zeroinitializer, <2 x i32> <i32 3, i32 4>
+  %i857 = fsub <2 x float> %i856, %i83122
+  %i858 = fmul <2 x float> %i857, zeroinitializer
+  %i859 = tail call <2 x float> @llvm.exp2.v2f32(<2 x float> %i858)
+  %i860 = shufflevector <8 x float> %i738, <8 x float> zeroinitializer, <2 x i32> <i32 5, i32 6>
+  %i861 = fsub <2 x float> %i860, splat (float 1.000000e+00)
+  %i862 = fmul <2 x float> %i861, zeroinitializer
+  %i863 = tail call <2 x float> @llvm.exp2.v2f32(<2 x float> %i862)
+  %i847 = shufflevector <4 x float> %i758, <4 x float> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison>
+  %i864 = shufflevector <8 x float> %i738, <8 x float> %i847, <2 x i32> <i32 7, i32 9>
+  %i865 = fsub <2 x float> %i864, %i83122
+  %i866 = fmul <2 x float> %i865, zeroinitializer
+  %i867 = tail call <2 x float> @llvm.exp2.v2f32(<2 x float> %i866)
+  %i746 = insertelement <8 x i64> %i745, i64 %i552, i64 7
+  %i747 = icmp slt <8 x i64> %i746, zeroinitializer
+  %i749 = select <8 x i1> %i747, <8 x float> zeroinitializer, <8 x float> splat (float 1.000000e+00)
+  %i868 = shufflevector <8 x float> %i749, <8 x float> zeroinitializer, <2 x i32> <i32 0, i32 1>
+  %i869 = fsub <2 x float> %i868, %i83122
+  %i870 = fmul <2 x float> %i869, zeroinitializer
+  %i871 = tail call <2 x float> @llvm.exp2.v2f32(<2 x float> %i870)
+  %i872 = shufflevector <8 x float> %i749, <8 x float> zeroinitializer, <2 x i32> <i32 2, i32 3>
+  %i873 = fsub <2 x float> %i872, %i83122
+  %i874 = fmul <2 x float> %i873, zeroinitializer
+  %i875 = tail call <2 x float> @llvm.exp2.v2f32(<2 x float> %i874)
+  %i876 = shufflevector <8 x float> %i749, <8 x float> zeroinitializer, <2 x i32> <i32 4, i32 5>
+  %i877 = fsub <2 x float> %i876, %i83122
+  %i878 = fmul <2 x float> %i877, zeroinitializer
+  %i879 = tail call <2 x float> @llvm.exp2.v2f32(<2 x float> %i878)
+  %i880 = shufflevector <8 x float> %i749, <8 x float> zeroinitializer, <2 x i32> <i32 6, i32 7>
+  %i881 = fsub <2 x float> %i880, %i83122
+  %i882 = fmul <2 x float> %i881, zeroinitializer
+  %i883 = tail call <2 x float> @llvm.exp2.v2f32(<2 x float> %i882)
+  %i884 = shufflevector <4 x float> splat (float 1.000000e+00), <4 x float> %i753, <2 x i32> <i32 2, i32 4>
+  %i885 = fsub <2 x float> %i884, %i83122
+  %i886 = fmul <2 x float> %i885, zeroinitializer
+  %i887 = tail call <2 x float> @llvm.exp2.v2f32(<2 x float> %i886)
+  %i900 = fadd <2 x float> %i867, %i871
+  %i901 = fadd <2 x float> %i875, %i879
+  %i896 = fadd <2 x float> %i834, %i838
+  %i897 = fadd <2 x float> %i842, %i846
+  %i904 = fadd <2 x float> %i896, %i897
+  %i898 = fadd <2 x float> %i851, %i855
+  %i899 = fadd <2 x float> %i859, %i863
+  %i905 = fadd <2 x float> %i898, %i899
+  %i908 = fadd <2 x float> %i904, %i905
+  %i906 = fadd <2 x float> %i900, %i901
+  %i902 = fadd <2 x float> %i883, %i887
+  %i909 = fadd <2 x float> %i906, %i902
+  %i910 = fadd <2 x float> %i908, %i909
+  %foldExtExtBinop215 = fadd <2 x float> %i910, %i902
+  %bc = bitcast <2 x float> %foldExtExtBinop215 to <2 x i32>
+  %i91124 = extractelement <2 x i32> %bc, i64 0
+  %i912 = tail call { i32, i32 } @llvm.amdgcn.permlane32.swap(i32 %i91124, i32 0, i1 false, i1 false)
+  %i919 = insertelement <2 x float> zeroinitializer, float %i473, i64 0
+  %i920 = insertelement <2 x float> %i919, float %i82821, i64 1
+  %i913 = extractvalue { i32, i32 } %i912, 0
+  %i915 = bitcast i32 %i913 to float
+  %i924 = fmul <2 x float> %i920, zeroinitializer
+  %i926 = insertelement <2 x float> zeroinitializer, float %i472, i64 0
+  %i927 = insertelement <2 x float> %i926, float %i915, i64 1
+  %i925 = tail call <2 x float> @llvm.exp2.v2f32(<2 x float> %i924)
+  %i928 = fmul <2 x float> %i927, %i925
+  %shift217 = shufflevector <2 x float> %i928, <2 x float> zeroinitializer, <2 x i32> <i32 1, i32 poison>
+  %foldExtExtBinop218 = fadd <2 x float> %i928, %shift217
+  %i929 = extractelement <2 x float> %foldExtExtBinop218, i64 0
+  tail call void @llvm.amdgcn.s.waitcnt(i32 0)
+  %i954 = load <1 x float>, ptr addrspace(3) null, align 4
+  %i514 = shl i32 %i507, 3
+  %i515 = select i1 %i497, i32 %i514, i32 0
+  %i969 = tail call <4 x i32> @llvm.amdgcn.raw.ptr.buffer.load.v4i32(ptr addrspace(8) null, i32 %i515, i32 0, i32 0)
+  %i958 = trunc i64 %i519 to i32
+  %i973 = tail call <4 x i32> @llvm.amdgcn.raw.ptr.buffer.load.v4i32(ptr addrspace(8) null, i32 %i958, i32 0, i32 0)
+  %i97026 = bitcast <4 x i32> %i969 to <8 x half>
+  %i1055 = shufflevector <8 x half> %i97026, <8 x half> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %i430 = lshr i32 %i, 1
+  %i4329 = getelementptr i8, ptr addrspace(3) null, i32 %i430
+  store <4 x half> %i1055, ptr addrspace(3) %i4329, align 8
+  store <4 x half> splat (half +qnan), ptr addrspace(3) %1, align 8
+  store <4 x half> zeroinitializer, ptr addrspace(3) %0, align 8
+  %i436 = xor i32 %i430, 1
+  %i43712 = getelementptr i8, ptr addrspace(3) null, i32 %i436
+  store <4 x half> splat (half 1.000000e+00), ptr addrspace(3) %i43712, align 8
+  %i974 = bitcast <4 x i32> %i973 to <8 x half>
+  %i1060 = shufflevector <8 x half> %i974, <8 x half> zeroinitializer, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+  store <4 x half> %i1060, ptr addrspace(3) null, align 8
+  %i1094 = shufflevector <2 x float> zeroinitializer, <2 x float> %i490, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %i1096 = shufflevector <4 x float> zeroinitializer, <4 x float> %i1094, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %i1097 = shufflevector <8 x float> zeroinitializer, <8 x float> %i1096, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  %i1098 = shufflevector <1 x float> %i954, <1 x float> zeroinitializer, <16 x i32> zeroinitializer
+  %i1099 = fmul <16 x float> %i1097, %i1098
+  %i1100 = tail call <16 x float> @llvm.amdgcn.mfma.f32.32x32x16.f16(<8 x half> zeroinitializer, <8 x half> zeroinitializer, <16 x float> %i1099, i32 0, i32 0, i32 0)
+  %i1101 = tail call <16 x float> @llvm.amdgcn.mfma.f32.32x32x16.f16(<8 x half> zeroinitializer, <8 x half> zeroinitializer, <16 x float> %i1100, i32 0, i32 0, i32 0)
+  %i1104 = shufflevector <2 x float> %i475, <2 x float> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %i1108 = shufflevector <4 x float> %i1104, <4 x float> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %i1110 = shufflevector <8 x float> %i1108, <8 x float> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  %i1113 = tail call <16 x float> @llvm.amdgcn.mfma.f32.32x32x16.f16(<8 x half> zeroinitializer, <8 x half> zeroinitializer, <16 x float> %i1110, i32 0, i32 0, i32 0)
+  %i1114 = tail call <16 x float> @llvm.amdgcn.mfma.f32.32x32x16.f16(<8 x half> zeroinitializer, <8 x half> zeroinitializer, <16 x float> %i1113, i32 0, i32 0, i32 0)
+  %i1115 = tail call <16 x float> @llvm.amdgcn.mfma.f32.32x32x16.f16(<8 x half> zeroinitializer, <8 x half> zeroinitializer, <16 x float> %i1114, i32 0, i32 0, i32 0)
+  %i111731 = add i64 %i474, 1
+  %i1119 = shufflevector <16 x float> %i1115, <16 x float> zeroinitializer, <2 x i32> <i32 0, i32 1>
+  %i1134 = shufflevector <16 x float> %i1101, <16 x float> zeroinitializer, <2 x i32> <i32 14, i32 15>
+  br i1 %i1118, label %bb, label %._crit_edge
+
+._crit_edge:                                      ; preds = %bb
+  %i1388 = tail call <16 x float> @llvm.amdgcn.mfma.f32.32x32x16.f16(<8 x half> zeroinitializer, <8 x half> %i90, <16 x float> zeroinitializer, i32 0, i32 0, i32 0)
+  %i1389 = tail call <16 x float> @llvm.amdgcn.mfma.f32.32x32x16.f16(<8 x half> zeroinitializer, <8 x half> %i93, <16 x float> %i1388, i32 0, i32 0, i32 0)
+  %i1390 = tail call <16 x float> @llvm.amdgcn.mfma.f32.32x32x16.f16(<8 x half> zeroinitializer, <8 x half> %i96, <16 x float> %i1389, i32 0, i32 0, i32 0)
+  %i1391 = tail call <16 x float> @llvm.amdgcn.mfma.f32.32x32x16.f16(<8 x half> zeroinitializer, <8 x half> %i99, <16 x float> %i1390, i32 0, i32 0, i32 0)
+  %i1324 = shufflevector <4 x i32> %i123, <4 x i32> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison>
+  %i1325 = shufflevector <8 x i32> zeroinitializer, <8 x i32> %i1324, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 8, i32 9, i32 10, i32 11>
+  %i1392 = tail call <16 x float> @llvm.amdgcn.mfma.f32.32x32x16.f16(<8 x half> zeroinitializer, <8 x half> %i102, <16 x float> %i1391, i32 0, i32 0, i32 0)
+  %i1393 = tail call <16 x float> @llvm.amdgcn.mfma.f32.32x32x16.f16(<8 x half> zeroinitializer, <8 x half> %i105, <16 x float> %i1392, i32 0, i32 0, i32 0)
+  %i1394 = tail call <16 x float> @llvm.amdgcn.mfma.f32.32x32x16.f16(<8 x half> zeroinitializer, <8 x half> %i108, <16 x float> %i1393, i32 0, i32 0, i32 0)
+  %i1398 = icmp slt <8 x i32> zeroinitializer, %i1325
+  %i1399 = select <8 x i1> %i1398, <8 x float> splat (float 1.000000e+00), <8 x float> zeroinitializer
+  %i1402 = icmp slt <8 x i32> splat (i32 1), %i126
+  %i1403 = select <8 x i1> %i1402, <8 x float> splat (float 1.000000e+00), <8 x float> zeroinitializer
+  %i1441 = extractelement <8 x float> %i1403, i64 6
+  %i1442 = extractelement <8 x float> %i1403, i64 7
+  %i1406 = icmp slt <8 x i32> zeroinitializer, %i125
+  %i1407 = select <8 x i1> %i1406, <8 x float> splat (float 1.000000e+00), <8 x float> zeroinitializer
+  %i1456 = extractelement <8 x float> %i1407, i64 6
+  %i1457 = extractelement <8 x float> %i1407, i64 7
+  %i1410 = icmp slt <4 x i32> splat (i32 1), %i128
+  %i1395 = tail call <16 x float> @llvm.amdgcn.mfma.f32.32x32x16.f16(<8 x half> zeroinitializer, <8 x half> %i111, <16 x float> %i1394, i32 0, i32 0, i32 0)
+  %i1408 = shufflevector <16 x float> %i1395, <16 x float> zeroinitializer, <4 x i32> <i32 11, i32 12, i32 13, i32 14>
+  %i1411 = select <4 x i1> %i1410, <4 x float> zeroinitializer, <4 x float> %i1408
+  %i1461 = extractelement <4 x float> %i1411, i64 0
+  %i1462 = extractelement <4 x float> %i1411, i64 1
+  %i1431 = extractelement <8 x float> %i1403, i64 0
+  %i1432 = extractelement <8 x float> %i1403, i64 1
+  %i1433 = fmul float %i1431, %i1432
+  %i1434 = extractelement <8 x float> %i1403, i64 2
+  %i1435 = fmul float %i1433, %i1434
+  %i1436 = extractelement <8 x float> %i1403, i64 3
+  %i1437 = extractelement <8 x float> %i1403, i64 4
+  %i1438 = fmul float %i1436, %i1437
+  %i1439 = extractelement <8 x float> %i1403, i64 5
+  %i1440 = fmul float %i1438, %i1439
+  %i1446 = extractelement <8 x float> %i1407, i64 0
+  %i1447 = extractelement <8 x float> %i1407, i64 1
+  %i1448 = fmul float %i1446, %i1447
+  %i1449 = extractelement <8 x float> %i1407, i64 2
+  %i1450 = fmul float %i1448, %i1449
+  %i1451 = extractelement <8 x float> %i1407, i64 3
+  %i1452 = extractelement <8 x float> %i1407, i64 4
+  %i1453 = fmul float %i1451, %i1452
+  %i1454 = extractelement <8 x float> %i1407, i64 5
+  %i1455 = fmul float %i1453, %i1454
+  %i1422 = extractelement <8 x float> %i1399, i64 4
+  %i1424 = extractelement <8 x float> %i1399, i64 5
+  %i1425 = fmul float %i1422, %i1424
+  %i1426 = extractelement <8 x float> %i1399, i64 6
+  %i1427 = extractelement <8 x float> %i1399, i64 7
+  %i1428 = fmul float %i1426, %i1427
+  %i1470 = fmul float %i1425, %i1428
+  %i1471 = fmul float %i1435, %i1440
+  %i1443 = fmul float %i1441, %i1442
+  %i1472 = fmul float %i1471, %i1443
+  %i1476 = fmul float %i1470, %i1472
+  %i1473 = fmul float %i1450, %i1455
+  %i1458 = fmul float %i1456, %i1457
+  %i1474 = fmul float %i1473, %i1458
+  %i1477 = fmul float %i1476, %i1474
+  %i1463 = fmul float %i1461, %i1462
+  %i1464 = extractelement <4 x float> %i1411, i64 2
+  %i1465 = fmul float %i1463, %i1464
+  %i1478 = fmul float %i1477, %i1465
+  %i1610 = insertelement <1 x float> zeroinitializer, float %i1478, i64 0
+  store <1 x float> %i1610, ptr addrspace(3) null, align 4
+  ret void
+}
+
+declare noundef range(i32 0, 1024) i32 @llvm.amdgcn.workitem.id.x() #1
+
+declare <4 x i32> @llvm.amdgcn.raw.ptr.buffer.load.v4i32(ptr addrspace(8) readonly captures(none), i32, i32, i32 immarg) #2
+
+declare void @llvm.amdgcn.s.waitcnt(i32 immarg) #3
+
+declare { i32, i32 } @llvm.amdgcn.permlane32.swap(i32, i32, i1 immarg, i1 immarg) #4
+
+declare <16 x float> @llvm.amdgcn.mfma.f32.32x32x16.f16(<8 x half>, <8 x half>, <16 x float>, i32 immarg, i32 immarg, i32 immarg) #5
+
+declare float @llvm.maxnum.f32(float, float) #6
+
+declare i64 @llvm.amdgcn.raw.ptr.buffer.load.i64(ptr addrspace(8) readonly captures(none), i32, i32, i32 immarg) #2
+
+declare <2 x float> @llvm.exp2.v2f32(<2 x float>) #6
+
+attributes #0 = { nofree norecurse nounwind "amdgpu-agpr-alloc"="0" "amdgpu-cluster-dims"="1,1,1" "amdgpu-flat-work-group-size"="1,256" "amdgpu-no-cluster-id-x" "amdgpu-no-cluster-id-y" "amdgpu-no-cluster-id-z" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-flat-scratch-init" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "amdgpu-no-wwm" "amdgpu-waves-per-eu"="0, 0" "uniform-work-group-size" }
+attributes #1 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }
+attributes #2 = { nocallback nofree nosync nounwind willreturn memory(argmem: read) }
+attributes #3 = { nocallback nofree nounwind willreturn }
+attributes #4 = { convergent nocallback nofree nounwind willreturn memory(none) }
+attributes #5 = { convergent nocallback nocreateundeforpoison nofree nosync nounwind willreturn memory(none) }
+attributes #6 = { nocallback nocreateundeforpoison nofree nosync nounwind speculatable willreturn memory(none) }
diff --git a/llvm/test/CodeGen/AMDGPU/machine-sink-loop-var-out-of-divergent-loop-swdev407790.ll b/llvm/test/CodeGen/AMDGPU/machine-sink-loop-var-out-of-divergent-loop-swdev407790.ll
index dfd4870787e62..b36d3212b688a 100644
--- a/llvm/test/CodeGen/AMDGPU/machine-sink-loop-var-out-of-divergent-loop-swdev407790.ll
+++ b/llvm/test/CodeGen/AMDGPU/machine-sink-loop-var-out-of-divergent-loop-swdev407790.ll
@@ -19,10 +19,11 @@ define void @machinesink_loop_variable_out_of_divergent_loop(i32 %arg, i1 %cmp49
 ; CHECK-NEXT:  .LBB0_1: ; %Flow
 ; CHECK-NEXT:    ; in Loop: Header=BB0_3 Depth=1
 ; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s8
+; CHECK-NEXT:    v_add_nc_u32_e32 v3, -4, v3
 ; CHECK-NEXT:  .LBB0_2: ; %Flow1
 ; CHECK-NEXT:    ; in Loop: Header=BB0_3 Depth=1
 ; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s7
-; CHECK-NEXT:    v_cmp_ne_u32_e64 s5, 0, v3
+; CHECK-NEXT:    v_cmp_ne_u32_e64 s5, 0, v1
 ; CHECK-NEXT:    ;;#ASMSTART
 ; CHECK-NEXT:    ; j lastloop entry
 ; CHECK-NEXT:    ;;#ASMEND
@@ -32,41 +33,40 @@ define void @machinesink_loop_variable_out_of_divergent_loop(i32 %arg, i1 %cmp49
 ; CHECK-NEXT:  .LBB0_3: ; %for.body33
 ; CHECK-NEXT:    ; =>This Loop Header: Depth=1
 ; CHECK-NEXT:    ; Child Loop BB0_6 Depth 2
-; CHECK-NEXT:    v_mov_b32_e32 v1, 0
 ; CHECK-NEXT:    v_mov_b32_e32 v3, 0
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
 ; CHECK-NEXT:    s_and_saveexec_b32 s7, s4
 ; CHECK-NEXT:    s_cbranch_execz .LBB0_2
 ; CHECK-NEXT:  ; %bb.4: ; %for.body51.preheader
 ; CHECK-NEXT:    ; in Loop: Header=BB0_3 Depth=1
-; CHECK-NEXT:    v_mov_b32_e32 v4, v2
-; CHECK-NEXT:    s_mov_b32 s9, 4
 ; CHECK-NEXT:    s_mov_b32 s8, 0
+; CHECK-NEXT:    s_mov_b32 s9, 0
 ; CHECK-NEXT:    s_branch .LBB0_6
 ; CHECK-NEXT:    .p2align 6
 ; CHECK-NEXT:  .LBB0_5: ; %if.end118
 ; CHECK-NEXT:    ; in Loop: Header=BB0_6 Depth=2
 ; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s5
-; CHECK-NEXT:    v_add_nc_u32_e32 v4, 4, v1
 ; CHECK-NEXT:    s_add_i32 s9, s9, 4
 ; CHECK-NEXT:    ;;#ASMSTART
 ; CHECK-NEXT:    ; backedge
 ; CHECK-NEXT:    ;;#ASMEND
-; CHECK-NEXT:    v_cmp_ge_u32_e64 s5, v4, v0
+; CHECK-NEXT:    v_add_nc_u32_e32 v3, s9, v2
+; CHECK-NEXT:    v_cmp_ge_u32_e64 s5, v3, v0
 ; CHECK-NEXT:    s_or_b32 s8, s5, s8
 ; CHECK-NEXT:    s_andn2_b32 exec_lo, exec_lo, s8
 ; CHECK-NEXT:    s_cbranch_execz .LBB0_1
 ; CHECK-NEXT:  .LBB0_6: ; %for.body51
 ; CHECK-NEXT:    ; Parent Loop BB0_3 Depth=1
 ; CHECK-NEXT:    ; => This Inner Loop Header: Depth=2
-; CHECK-NEXT:    v_mov_b32_e32 v1, v4
-; CHECK-NEXT:    v_mov_b32_e32 v3, 1
+; CHECK-NEXT:    v_mov_b32_e32 v1, 1
 ; CHECK-NEXT:    s_and_saveexec_b32 s5, vcc_lo
 ; CHECK-NEXT:    s_cbranch_execz .LBB0_5
 ; CHECK-NEXT:  ; %bb.7: ; %if.then112
 ; CHECK-NEXT:    ; in Loop: Header=BB0_6 Depth=2
-; CHECK-NEXT:    v_mov_b32_e32 v3, 0
-; CHECK-NEXT:    v_mov_b32_e32 v4, s9
-; CHECK-NEXT:    ds_write_b32 v3, v4
+; CHECK-NEXT:    s_add_i32 s10, s9, 4
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
+; CHECK-NEXT:    v_mov_b32_e32 v3, s10
+; CHECK-NEXT:    ds_write_b32 v1, v3
 ; CHECK-NEXT:    s_branch .LBB0_5
 ; CHECK-NEXT:  .LBB0_8: ; %for.body159.preheader
 ; CHECK-NEXT:    s_inst_prefetch 0x2
diff --git a/llvm/test/CodeGen/AMDGPU/machine-sink-temporal-divergence-swdev407790.ll b/llvm/test/CodeGen/AMDGPU/machine-sink-temporal-divergence-swdev407790.ll
index bca93a1d91ef1..e2a3e5480702c 100644
--- a/llvm/test/CodeGen/AMDGPU/machine-sink-temporal-divergence-swdev407790.ll
+++ b/llvm/test/CodeGen/AMDGPU/machine-sink-temporal-divergence-swdev407790.ll
@@ -62,7 +62,7 @@ define protected amdgpu_kernel void @kernel_round1(ptr addrspace(1) nocapture no
 ; CHECK-NEXT:    s_mov_b64 s[36:37], s[10:11]
 ; CHECK-NEXT:    s_mov_b64 s[38:39], s[6:7]
 ; CHECK-NEXT:    s_mov_b64 s[48:49], s[4:5]
-; CHECK-NEXT:    v_mov_b32_e32 v46, 0
+; CHECK-NEXT:    v_mov_b32_e32 v45, 0
 ; CHECK-NEXT:    s_swappc_b64 s[30:31], s[16:17]
 ; CHECK-NEXT:    v_mov_b32_e32 v43, v0
 ; CHECK-NEXT:    v_mov_b32_e32 v31, v40
@@ -92,7 +92,7 @@ define protected amdgpu_kernel void @kernel_round1(ptr addrspace(1) nocapture no
 ; CHECK-NEXT:    s_mov_b32 s12, s51
 ; CHECK-NEXT:    s_mov_b32 s13, s50
 ; CHECK-NEXT:    s_mov_b32 s14, s33
-; CHECK-NEXT:    ds_write_b32 v46, v46 offset:15360
+; CHECK-NEXT:    ds_write_b32 v45, v45 offset:15360
 ; CHECK-NEXT:    s_swappc_b64 s[30:31], s[16:17]
 ; CHECK-NEXT:    v_lshrrev_b32_e32 v0, 1, v43
 ; CHECK-NEXT:    v_lshlrev_b32_e32 v1, 2, v43
@@ -120,66 +120,69 @@ define protected amdgpu_kernel void @kernel_round1(ptr addrspace(1) nocapture no
 ; CHECK-NEXT:    v_cmpx_ne_u32_e32 0, v42
 ; CHECK-NEXT:    s_cbranch_execz .LBB0_25
 ; CHECK-NEXT:  ; %bb.1: ; %.preheader5
-; CHECK-NEXT:    v_mul_lo_u32 v44, v41, 14
+; CHECK-NEXT:    v_mul_lo_u32 v0, v41, 14
 ; CHECK-NEXT:    s_mov_b32 s4, 0
 ; CHECK-NEXT:    s_mov_b32 s5, 0
-; CHECK-NEXT:    v_add_nc_u32_e32 v45, 0x3c04, v44
+; CHECK-NEXT:    v_add_nc_u32_e32 v44, 0x3c04, v0
 ; CHECK-NEXT:  .LBB0_2: ; =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    v_add_nc_u32_e32 v0, s5, v45
+; CHECK-NEXT:    v_add_nc_u32_e32 v1, s5, v44
 ; CHECK-NEXT:    s_add_i32 s5, s5, 1
 ; CHECK-NEXT:    v_cmp_eq_u32_e32 vcc_lo, s5, v42
-; CHECK-NEXT:    ds_write_b8 v0, v46
+; CHECK-NEXT:    ds_write_b8 v1, v45
 ; CHECK-NEXT:    s_or_b32 s4, vcc_lo, s4
 ; CHECK-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
 ; CHECK-NEXT:    s_cbranch_execnz .LBB0_2
 ; CHECK-NEXT:  ; %bb.3:
 ; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s4
-; CHECK-NEXT:    v_add_nc_u32_e32 v46, -1, v42
-; CHECK-NEXT:    s_mov_b32 s53, 0
-; CHECK-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v46
+; CHECK-NEXT:    v_add_nc_u32_e32 v45, -1, v42
+; CHECK-NEXT:    s_mov_b32 s55, 0
+; CHECK-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v45
 ; CHECK-NEXT:    s_and_b32 exec_lo, exec_lo, vcc_lo
 ; CHECK-NEXT:    s_cbranch_execz .LBB0_25
 ; CHECK-NEXT:  ; %bb.4:
 ; CHECK-NEXT:    v_lshlrev_b32_e32 v43, 10, v43
+; CHECK-NEXT:    v_add_nc_u32_e32 v46, 0x3c05, v0
 ; CHECK-NEXT:    v_mov_b32_e32 v47, 0
-; CHECK-NEXT:    s_mov_b32 s54, 0
+; CHECK-NEXT:    s_mov_b32 s53, 0
 ; CHECK-NEXT:  .LBB0_5: ; =>This Loop Header: Depth=1
 ; CHECK-NEXT:    ; Child Loop BB0_8 Depth 2
 ; CHECK-NEXT:    ; Child Loop BB0_20 Depth 2
-; CHECK-NEXT:    s_mov_b32 s4, s53
-; CHECK-NEXT:    s_lshl_b32 s5, s53, 5
-; CHECK-NEXT:    v_add_nc_u32_e32 v0, s4, v45
-; CHECK-NEXT:    s_add_i32 s53, s53, 1
-; CHECK-NEXT:    s_add_i32 s4, s4, 5
-; CHECK-NEXT:    v_or3_b32 v57, s5, v43, s53
-; CHECK-NEXT:    v_mov_b32_e32 v58, s53
+; CHECK-NEXT:    v_add_nc_u32_e32 v0, s55, v44
+; CHECK-NEXT:    s_lshl_b32 s4, s55, 5
+; CHECK-NEXT:    s_add_i32 s54, s55, 1
+; CHECK-NEXT:    s_add_i32 s5, s55, 5
+; CHECK-NEXT:    v_or3_b32 v57, s4, v43, s54
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
 ; CHECK-NEXT:    ds_read_u8 v56, v0
-; CHECK-NEXT:    s_mov_b32 s55, exec_lo
-; CHECK-NEXT:    v_cmpx_lt_u32_e32 s4, v42
+; CHECK-NEXT:    v_mov_b32_e32 v58, s54
+; CHECK-NEXT:    s_mov_b32 s68, exec_lo
+; CHECK-NEXT:    v_cmpx_lt_u32_e32 s5, v42
 ; CHECK-NEXT:    s_cbranch_execz .LBB0_17
 ; CHECK-NEXT:  ; %bb.6: ; %.preheader2
 ; CHECK-NEXT:    ; in Loop: Header=BB0_5 Depth=1
-; CHECK-NEXT:    s_mov_b32 s68, 0
-; CHECK-NEXT:    s_mov_b32 s69, s53
+; CHECK-NEXT:    s_mov_b32 s69, 0
+; CHECK-NEXT:    s_mov_b32 s80, 0
 ; CHECK-NEXT:    s_branch .LBB0_8
 ; CHECK-NEXT:  .LBB0_7: ; in Loop: Header=BB0_8 Depth=2
-; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s80
-; CHECK-NEXT:    s_add_i32 s4, s69, 4
-; CHECK-NEXT:    v_add_nc_u32_e32 v57, 4, v57
-; CHECK-NEXT:    v_cmp_ge_u32_e32 vcc_lo, s4, v42
-; CHECK-NEXT:    v_mov_b32_e32 v58, s69
-; CHECK-NEXT:    s_or_b32 s68, vcc_lo, s68
-; CHECK-NEXT:    s_andn2_b32 exec_lo, exec_lo, s68
+; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s81
+; CHECK-NEXT:    s_add_i32 s80, s80, 4
+; CHECK-NEXT:    s_add_i32 s4, s55, s80
+; CHECK-NEXT:    v_add_nc_u32_e32 v0, s80, v57
+; CHECK-NEXT:    s_add_i32 s5, s4, 5
+; CHECK-NEXT:    s_add_i32 s4, s4, 1
+; CHECK-NEXT:    v_cmp_ge_u32_e32 vcc_lo, s5, v42
+; CHECK-NEXT:    v_mov_b32_e32 v58, s4
+; CHECK-NEXT:    s_or_b32 s69, vcc_lo, s69
+; CHECK-NEXT:    s_andn2_b32 exec_lo, exec_lo, s69
 ; CHECK-NEXT:    s_cbranch_execz .LBB0_16
 ; CHECK-NEXT:  .LBB0_8: ; Parent Loop BB0_5 Depth=1
 ; CHECK-NEXT:    ; => This Inner Loop Header: Depth=2
-; CHECK-NEXT:    v_add_nc_u32_e32 v58, s69, v45
-; CHECK-NEXT:    s_mov_b32 s69, s4
-; CHECK-NEXT:    ds_read_u8 v0, v58
+; CHECK-NEXT:    v_add_nc_u32_e32 v59, s80, v46
+; CHECK-NEXT:    v_add_nc_u32_e32 v58, s80, v57
+; CHECK-NEXT:    ds_read_u8 v0, v59
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
-; CHECK-NEXT:    v_cmp_eq_u16_sdwa s5, v56, v0 src0_sel:BYTE_0 src1_sel:DWORD
-; CHECK-NEXT:    s_and_saveexec_b32 s80, s5
+; CHECK-NEXT:    v_cmp_eq_u16_sdwa s4, v56, v0 src0_sel:BYTE_0 src1_sel:DWORD
+; CHECK-NEXT:    s_and_saveexec_b32 s81, s4
 ; CHECK-NEXT:    s_cbranch_execz .LBB0_10
 ; CHECK-NEXT:  ; %bb.9: ; in Loop: Header=BB0_8 Depth=2
 ; CHECK-NEXT:    v_mov_b32_e32 v31, v40
@@ -199,13 +202,13 @@ define protected amdgpu_kernel void @kernel_round1(ptr addrspace(1) nocapture no
 ; CHECK-NEXT:    ; implicit-def: $sgpr15
 ; CHECK-NEXT:    s_swappc_b64 s[30:31], s[16:17]
 ; CHECK-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; CHECK-NEXT:    ds_write_b32 v0, v57
+; CHECK-NEXT:    ds_write_b32 v0, v58
 ; CHECK-NEXT:  .LBB0_10: ; in Loop: Header=BB0_8 Depth=2
-; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s80
-; CHECK-NEXT:    ds_read_u8 v0, v58 offset:1
+; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s81
+; CHECK-NEXT:    ds_read_u8 v0, v59 offset:1
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
 ; CHECK-NEXT:    v_cmp_eq_u16_sdwa s4, v56, v0 src0_sel:BYTE_0 src1_sel:DWORD
-; CHECK-NEXT:    s_and_saveexec_b32 s80, s4
+; CHECK-NEXT:    s_and_saveexec_b32 s81, s4
 ; CHECK-NEXT:    s_cbranch_execz .LBB0_12
 ; CHECK-NEXT:  ; %bb.11: ; in Loop: Header=BB0_8 Depth=2
 ; CHECK-NEXT:    v_mov_b32_e32 v31, v40
@@ -221,18 +224,18 @@ define protected amdgpu_kernel void @kernel_round1(ptr addrspace(1) nocapture no
 ; CHECK-NEXT:    s_mov_b32 s12, s51
 ; CHECK-NEXT:    s_mov_b32 s13, s50
 ; CHECK-NEXT:    s_mov_b32 s14, s33
-; CHECK-NEXT:    v_add_nc_u32_e32 v59, 1, v57
+; CHECK-NEXT:    v_add_nc_u32_e32 v60, 1, v58
 ; CHECK-NEXT:    v_add_nc_u32_e32 v47, 1, v47
 ; CHECK-NEXT:    ; implicit-def: $sgpr15
 ; CHECK-NEXT:    s_swappc_b64 s[30:31], s[16:17]
 ; CHECK-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; CHECK-NEXT:    ds_write_b32 v0, v59
+; CHECK-NEXT:    ds_write_b32 v0, v60
 ; CHECK-NEXT:  .LBB0_12: ; in Loop: Header=BB0_8 Depth=2
-; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s80
-; CHECK-NEXT:    ds_read_u8 v0, v58 offset:2
+; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s81
+; CHECK-NEXT:    ds_read_u8 v0, v59 offset:2
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
 ; CHECK-NEXT:    v_cmp_eq_u16_sdwa s4, v56, v0 src0_sel:BYTE_0 src1_sel:DWORD
-; CHECK-NEXT:    s_and_saveexec_b32 s80, s4
+; CHECK-NEXT:    s_and_saveexec_b32 s81, s4
 ; CHECK-NEXT:    s_cbranch_execz .LBB0_14
 ; CHECK-NEXT:  ; %bb.13: ; in Loop: Header=BB0_8 Depth=2
 ; CHECK-NEXT:    v_mov_b32_e32 v31, v40
@@ -248,18 +251,18 @@ define protected amdgpu_kernel void @kernel_round1(ptr addrspace(1) nocapture no
 ; CHECK-NEXT:    s_mov_b32 s12, s51
 ; CHECK-NEXT:    s_mov_b32 s13, s50
 ; CHECK-NEXT:    s_mov_b32 s14, s33
-; CHECK-NEXT:    v_add_nc_u32_e32 v59, 2, v57
+; CHECK-NEXT:    v_add_nc_u32_e32 v60, 2, v58
 ; CHECK-NEXT:    v_add_nc_u32_e32 v47, 1, v47
 ; CHECK-NEXT:    ; implicit-def: $sgpr15
 ; CHECK-NEXT:    s_swappc_b64 s[30:31], s[16:17]
 ; CHECK-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; CHECK-NEXT:    ds_write_b32 v0, v59
+; CHECK-NEXT:    ds_write_b32 v0, v60
 ; CHECK-NEXT:  .LBB0_14: ; in Loop: Header=BB0_8 Depth=2
-; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s80
-; CHECK-NEXT:    ds_read_u8 v0, v58 offset:3
+; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s81
+; CHECK-NEXT:    ds_read_u8 v0, v59 offset:3
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
 ; CHECK-NEXT:    v_cmp_eq_u16_sdwa s4, v56, v0 src0_sel:BYTE_0 src1_sel:DWORD
-; CHECK-NEXT:    s_and_saveexec_b32 s80, s4
+; CHECK-NEXT:    s_and_saveexec_b32 s81, s4
 ; CHECK-NEXT:    s_cbranch_execz .LBB0_7
 ; CHECK-NEXT:  ; %bb.15: ; in Loop: Header=BB0_8 Depth=2
 ; CHECK-NEXT:    v_mov_b32_e32 v31, v40
@@ -275,19 +278,20 @@ define protected amdgpu_kernel void @kernel_round1(ptr addrspace(1) nocapture no
 ; CHECK-NEXT:    s_mov_b32 s12, s51
 ; CHECK-NEXT:    s_mov_b32 s13, s50
 ; CHECK-NEXT:    s_mov_b32 s14, s33
-; CHECK-NEXT:    v_add_nc_u32_e32 v58, 3, v57
+; CHECK-NEXT:    v_add_nc_u32_e32 v58, 3, v58
 ; CHECK-NEXT:    v_add_nc_u32_e32 v47, 1, v47
 ; CHECK-NEXT:    ; implicit-def: $sgpr15
 ; CHECK-NEXT:    s_swappc_b64 s[30:31], s[16:17]
 ; CHECK-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; CHECK-NEXT:    ds_write_b32 v0, v58
 ; CHECK-NEXT:    s_branch .LBB0_7
-; CHECK-NEXT:  .LBB0_16: ; %Flow32
+; CHECK-NEXT:  .LBB0_16: ; %Flow41
 ; CHECK-NEXT:    ; in Loop: Header=BB0_5 Depth=1
-; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s68
-; CHECK-NEXT:  .LBB0_17: ; %Flow33
+; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s69
+; CHECK-NEXT:    v_mov_b32_e32 v57, v0
+; CHECK-NEXT:  .LBB0_17: ; %Flow42
 ; CHECK-NEXT:    ; in Loop: Header=BB0_5 Depth=1
-; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s55
+; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s68
 ; CHECK-NEXT:    s_mov_b32 s55, exec_lo
 ; CHECK-NEXT:    v_cmpx_lt_u32_e32 v58, v42
 ; CHECK-NEXT:    s_cbranch_execz .LBB0_23
@@ -308,7 +312,7 @@ define protected amdgpu_kernel void @kernel_round1(ptr addrspace(1) nocapture no
 ; CHECK-NEXT:  .LBB0_20: ; Parent Loop BB0_5 Depth=1
 ; CHECK-NEXT:    ; => This Inner Loop Header: Depth=2
 ; CHECK-NEXT:    v_add_nc_u32_e32 v0, v44, v58
-; CHECK-NEXT:    ds_read_u8 v0, v0 offset:15364
+; CHECK-NEXT:    ds_read_u8 v0, v0
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
 ; CHECK-NEXT:    v_cmp_eq_u16_sdwa s4, v56, v0 src0_sel:BYTE_0 src1_sel:DWORD
 ; CHECK-NEXT:    s_and_saveexec_b32 s69, s4
@@ -333,22 +337,24 @@ define protected amdgpu_kernel void @kernel_round1(ptr addrspace(1) nocapture no
 ; CHECK-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; CHECK-NEXT:    ds_write_b32 v0, v57
 ; CHECK-NEXT:    s_branch .LBB0_19
-; CHECK-NEXT:  .LBB0_22: ; %Flow30
+; CHECK-NEXT:  .LBB0_22: ; %Flow39
 ; CHECK-NEXT:    ; in Loop: Header=BB0_5 Depth=1
 ; CHECK-NEXT:    s_inst_prefetch 0x2
 ; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s68
-; CHECK-NEXT:  .LBB0_23: ; %Flow31
+; CHECK-NEXT:  .LBB0_23: ; %Flow40
 ; CHECK-NEXT:    ; in Loop: Header=BB0_5 Depth=1
 ; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s55
 ; CHECK-NEXT:  ; %bb.24: ; in Loop: Header=BB0_5 Depth=1
-; CHECK-NEXT:    v_cmp_ge_u32_e32 vcc_lo, s53, v46
+; CHECK-NEXT:    v_cmp_ge_u32_e32 vcc_lo, s54, v45
 ; CHECK-NEXT:    v_cmp_lt_u32_e64 s4, 59, v47
+; CHECK-NEXT:    v_add_nc_u32_e32 v46, 1, v46
+; CHECK-NEXT:    s_mov_b32 s55, s54
 ; CHECK-NEXT:    s_or_b32 s4, vcc_lo, s4
 ; CHECK-NEXT:    s_and_b32 s4, exec_lo, s4
-; CHECK-NEXT:    s_or_b32 s54, s4, s54
-; CHECK-NEXT:    s_andn2_b32 exec_lo, exec_lo, s54
+; CHECK-NEXT:    s_or_b32 s53, s4, s53
+; CHECK-NEXT:    s_andn2_b32 exec_lo, exec_lo, s53
 ; CHECK-NEXT:    s_cbranch_execnz .LBB0_5
-; CHECK-NEXT:  .LBB0_25: ; %Flow38
+; CHECK-NEXT:  .LBB0_25: ; %Flow47
 ; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s52
 ; CHECK-NEXT:    v_mov_b32_e32 v31, v40
 ; CHECK-NEXT:    v_mov_b32_e32 v0, 1
@@ -834,7 +840,7 @@ define protected amdgpu_kernel void @kernel_round1_short(ptr addrspace(1) nocapt
 ; CHECK-NEXT:    s_mov_b32 s14, s33
 ; CHECK-NEXT:    ; implicit-def: $sgpr15
 ; CHECK-NEXT:    s_swappc_b64 s[30:31], s[16:17]
-; CHECK-NEXT:    v_mul_lo_u32 v44, v0, 14
+; CHECK-NEXT:    v_mul_lo_u32 v46, v0, 14
 ; CHECK-NEXT:    v_mov_b32_e32 v31, v40
 ; CHECK-NEXT:    v_mov_b32_e32 v0, 1
 ; CHECK-NEXT:    s_getpc_b64 s[16:17]
@@ -848,7 +854,7 @@ define protected amdgpu_kernel void @kernel_round1_short(ptr addrspace(1) nocapt
 ; CHECK-NEXT:    s_mov_b32 s13, s50
 ; CHECK-NEXT:    s_mov_b32 s14, s33
 ; CHECK-NEXT:    ds_write_b32 v43, v43 offset:15360
-; CHECK-NEXT:    v_add_nc_u32_e32 v45, 0x3c04, v44
+; CHECK-NEXT:    v_add_nc_u32_e32 v44, 0x3c04, v46
 ; CHECK-NEXT:    s_swappc_b64 s[30:31], s[16:17]
 ; CHECK-NEXT:    v_lshrrev_b32_e32 v0, 1, v42
 ; CHECK-NEXT:    v_lshlrev_b32_e32 v1, 2, v42
@@ -873,49 +879,52 @@ define protected amdgpu_kernel void @kernel_round1_short(ptr addrspace(1) nocapt
 ; CHECK-NEXT:    s_swappc_b64 s[30:31], s[16:17]
 ; CHECK-NEXT:    v_mov_b32_e32 v41, v0
 ; CHECK-NEXT:    v_lshlrev_b32_e32 v42, 10, v42
-; CHECK-NEXT:    s_mov_b32 s53, 0
+; CHECK-NEXT:    s_mov_b32 s4, 0
 ; CHECK-NEXT:    s_mov_b32 s52, 0
-; CHECK-NEXT:    ds_write_b8 v44, v43 offset:15364
-; CHECK-NEXT:    v_add_nc_u32_e32 v46, -1, v41
+; CHECK-NEXT:    ds_write_b8 v46, v43 offset:15364
+; CHECK-NEXT:    v_add_nc_u32_e32 v45, -1, v41
 ; CHECK-NEXT:  .LBB1_1: ; %.37
 ; CHECK-NEXT:    ; =>This Loop Header: Depth=1
 ; CHECK-NEXT:    ; Child Loop BB1_3 Depth 2
 ; CHECK-NEXT:    ; Child Loop BB1_8 Depth 2
-; CHECK-NEXT:    s_mov_b32 s4, s53
-; CHECK-NEXT:    s_lshl_b32 s6, s53, 5
-; CHECK-NEXT:    v_add_nc_u32_e32 v0, s4, v45
-; CHECK-NEXT:    s_add_i32 s53, s53, 1
-; CHECK-NEXT:    s_add_i32 s5, s4, 5
-; CHECK-NEXT:    v_or3_b32 v56, s6, v42, s53
-; CHECK-NEXT:    v_mov_b32_e32 v57, s53
+; CHECK-NEXT:    v_add_nc_u32_e32 v0, s4, v44
+; CHECK-NEXT:    s_lshl_b32 s5, s4, 5
+; CHECK-NEXT:    s_add_i32 s53, s4, 1
+; CHECK-NEXT:    s_add_i32 s6, s4, 5
+; CHECK-NEXT:    v_or3_b32 v47, s5, v42, s53
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
-; CHECK-NEXT:    ds_read_u8 v47, v0
-; CHECK-NEXT:    s_mov_b32 s4, exec_lo
-; CHECK-NEXT:    v_cmpx_lt_u32_e32 s5, v41
+; CHECK-NEXT:    ds_read_u8 v46, v0
+; CHECK-NEXT:    v_mov_b32_e32 v56, s53
+; CHECK-NEXT:    s_mov_b32 s5, exec_lo
+; CHECK-NEXT:    v_cmpx_lt_u32_e32 s6, v41
 ; CHECK-NEXT:    s_cbranch_execz .LBB1_5
 ; CHECK-NEXT:  ; %bb.2: ; %.53.preheader
 ; CHECK-NEXT:    ; in Loop: Header=BB1_1 Depth=1
 ; CHECK-NEXT:    s_mov_b32 s6, 0
+; CHECK-NEXT:    s_mov_b32 s7, 0
 ; CHECK-NEXT:  .LBB1_3: ; %.53
 ; CHECK-NEXT:    ; Parent Loop BB1_1 Depth=1
 ; CHECK-NEXT:    ; => This Inner Loop Header: Depth=2
-; CHECK-NEXT:    s_add_i32 s7, s5, 4
+; CHECK-NEXT:    s_add_i32 s7, s7, 4
 ; CHECK-NEXT:    v_add_nc_u32_e32 v43, 1, v43
-; CHECK-NEXT:    v_cmp_ge_u32_e32 vcc_lo, s7, v41
-; CHECK-NEXT:    v_add_nc_u32_e32 v56, 4, v56
-; CHECK-NEXT:    v_mov_b32_e32 v57, s5
-; CHECK-NEXT:    s_mov_b32 s5, s7
+; CHECK-NEXT:    s_add_i32 s8, s4, s7
+; CHECK-NEXT:    v_add_nc_u32_e32 v0, s7, v47
+; CHECK-NEXT:    s_add_i32 s9, s8, 5
+; CHECK-NEXT:    s_add_i32 s8, s8, 1
+; CHECK-NEXT:    v_cmp_ge_u32_e32 vcc_lo, s9, v41
+; CHECK-NEXT:    v_mov_b32_e32 v56, s8
 ; CHECK-NEXT:    s_or_b32 s6, vcc_lo, s6
 ; CHECK-NEXT:    s_andn2_b32 exec_lo, exec_lo, s6
 ; CHECK-NEXT:    s_cbranch_execnz .LBB1_3
-; CHECK-NEXT:  ; %bb.4: ; %Flow4
+; CHECK-NEXT:  ; %bb.4: ; %Flow3
 ; CHECK-NEXT:    ; in Loop: Header=BB1_1 Depth=1
 ; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s6
-; CHECK-NEXT:  .LBB1_5: ; %Flow5
+; CHECK-NEXT:    v_mov_b32_e32 v47, v0
+; CHECK-NEXT:  .LBB1_5: ; %Flow4
 ; CHECK-NEXT:    ; in Loop: Header=BB1_1 Depth=1
-; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s4
+; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s5
 ; CHECK-NEXT:    s_mov_b32 s54, exec_lo
-; CHECK-NEXT:    v_cmpx_lt_u32_e32 v57, v41
+; CHECK-NEXT:    v_cmpx_lt_u32_e32 v56, v41
 ; CHECK-NEXT:    s_cbranch_execz .LBB1_11
 ; CHECK-NEXT:  ; %bb.6: ; %.103.preheader
 ; CHECK-NEXT:    ; in Loop: Header=BB1_1 Depth=1
@@ -926,19 +935,19 @@ define protected amdgpu_kernel void @kernel_round1_short(ptr addrspace(1) nocapt
 ; CHECK-NEXT:  .LBB1_7: ; %.114
 ; CHECK-NEXT:    ; in Loop: Header=BB1_8 Depth=2
 ; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s64
-; CHECK-NEXT:    v_add_nc_u32_e32 v57, 1, v57
 ; CHECK-NEXT:    v_add_nc_u32_e32 v56, 1, v56
-; CHECK-NEXT:    v_cmp_ge_u32_e32 vcc_lo, v57, v41
+; CHECK-NEXT:    v_add_nc_u32_e32 v47, 1, v47
+; CHECK-NEXT:    v_cmp_ge_u32_e32 vcc_lo, v56, v41
 ; CHECK-NEXT:    s_or_b32 s55, vcc_lo, s55
 ; CHECK-NEXT:    s_andn2_b32 exec_lo, exec_lo, s55
 ; CHECK-NEXT:    s_cbranch_execz .LBB1_10
 ; CHECK-NEXT:  .LBB1_8: ; %.103
 ; CHECK-NEXT:    ; Parent Loop BB1_1 Depth=1
 ; CHECK-NEXT:    ; => This Inner Loop Header: Depth=2
-; CHECK-NEXT:    v_add_nc_u32_e32 v0, v44, v57
-; CHECK-NEXT:    ds_read_u8 v0, v0 offset:15364
+; CHECK-NEXT:    v_add_nc_u32_e32 v0, v44, v56
+; CHECK-NEXT:    ds_read_u8 v0, v0
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
-; CHECK-NEXT:    v_cmp_eq_u16_sdwa s4, v47, v0 src0_sel:BYTE_0 src1_sel:DWORD
+; CHECK-NEXT:    v_cmp_eq_u16_sdwa s4, v46, v0 src0_sel:BYTE_0 src1_sel:DWORD
 ; CHECK-NEXT:    s_and_saveexec_b32 s64, s4
 ; CHECK-NEXT:    s_cbranch_execz .LBB1_7
 ; CHECK-NEXT:  ; %bb.9: ; %.110
@@ -960,22 +969,23 @@ define protected amdgpu_kernel void @kernel_round1_short(ptr addrspace(1) nocapt
 ; CHECK-NEXT:    ; implicit-def: $sgpr15
 ; CHECK-NEXT:    s_swappc_b64 s[30:31], s[16:17]
 ; CHECK-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; CHECK-NEXT:    ds_write_b32 v0, v56
+; CHECK-NEXT:    ds_write_b32 v0, v47
 ; CHECK-NEXT:    s_branch .LBB1_7
 ; CHECK-NEXT:  .LBB1_10: ; %Flow
 ; CHECK-NEXT:    ; in Loop: Header=BB1_1 Depth=1
 ; CHECK-NEXT:    s_inst_prefetch 0x2
 ; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s55
-; CHECK-NEXT:  .LBB1_11: ; %Flow3
+; CHECK-NEXT:  .LBB1_11: ; %Flow2
 ; CHECK-NEXT:    ; in Loop: Header=BB1_1 Depth=1
 ; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s54
 ; CHECK-NEXT:  ; %bb.12: ; %.32
 ; CHECK-NEXT:    ; in Loop: Header=BB1_1 Depth=1
-; CHECK-NEXT:    v_cmp_ge_u32_e32 vcc_lo, s53, v46
+; CHECK-NEXT:    v_cmp_ge_u32_e32 vcc_lo, s53, v45
 ; CHECK-NEXT:    v_cmp_lt_u32_e64 s4, 59, v43
 ; CHECK-NEXT:    s_or_b32 s4, vcc_lo, s4
 ; CHECK-NEXT:    s_and_b32 s4, exec_lo, s4
 ; CHECK-NEXT:    s_or_b32 s52, s4, s52
+; CHECK-NEXT:    s_mov_b32 s4, s53
 ; CHECK-NEXT:    s_andn2_b32 exec_lo, exec_lo, s52
 ; CHECK-NEXT:    s_cbranch_execnz .LBB1_1
 ; CHECK-NEXT:  ; %bb.13: ; %.119
diff --git a/llvm/test/CodeGen/AMDGPU/memintrinsic-unroll.ll b/llvm/test/CodeGen/AMDGPU/memintrinsic-unroll.ll
index a64db94eae5be..f92d3120c9f91 100644
--- a/llvm/test/CodeGen/AMDGPU/memintrinsic-unroll.ll
+++ b/llvm/test/CodeGen/AMDGPU/memintrinsic-unroll.ll
@@ -5395,90 +5395,89 @@ define void @memmove_p0_p0_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(0
 ; CHECK-NEXT:    s_xor_b32 s6, exec_lo, s4
 ; CHECK-NEXT:    s_cbranch_execz .LBB5_3
 ; CHECK-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader
-; CHECK-NEXT:    s_mov_b64 s[4:5], 0x800
+; CHECK-NEXT:    s_mov_b64 s[4:5], 0
 ; CHECK-NEXT:  .LBB5_2: ; %memmove_fwd_loop
 ; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    v_add_co_u32 v96, vcc_lo, v2, 48
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v97, null, 0, v3, vcc_lo
+; CHECK-NEXT:    v_add_co_u32 v52, vcc_lo, v2, s4
+; CHECK-NEXT:    v_add_co_ci_u32_e64 v53, null, s5, v3, vcc_lo
+; CHECK-NEXT:    v_add_co_u32 v96, vcc_lo, v52, 48
+; CHECK-NEXT:    v_add_co_ci_u32_e64 v97, null, 0, v53, vcc_lo
 ; CHECK-NEXT:    s_clause 0xf
-; CHECK-NEXT:    flat_load_dwordx4 v[4:7], v[2:3] offset:128
-; CHECK-NEXT:    flat_load_dwordx4 v[8:11], v[2:3] offset:64
-; CHECK-NEXT:    flat_load_dwordx4 v[12:15], v[2:3] offset:32
-; CHECK-NEXT:    flat_load_dwordx4 v[16:19], v[2:3] offset:16
+; CHECK-NEXT:    flat_load_dwordx4 v[4:7], v[52:53] offset:64
+; CHECK-NEXT:    flat_load_dwordx4 v[8:11], v[52:53] offset:32
+; CHECK-NEXT:    flat_load_dwordx4 v[12:15], v[52:53]
+; CHECK-NEXT:    flat_load_dwordx4 v[16:19], v[52:53] offset:16
 ; CHECK-NEXT:    flat_load_dwordx4 v[20:23], v[96:97] offset:176
 ; CHECK-NEXT:    flat_load_dwordx4 v[24:27], v[96:97] offset:192
 ; CHECK-NEXT:    flat_load_dwordx4 v[28:31], v[96:97] offset:144
 ; CHECK-NEXT:    flat_load_dwordx4 v[32:35], v[96:97] offset:160
 ; CHECK-NEXT:    flat_load_dwordx4 v[36:39], v[96:97] offset:112
 ; CHECK-NEXT:    flat_load_dwordx4 v[48:51], v[96:97] offset:128
-; CHECK-NEXT:    flat_load_dwordx4 v[52:55], v[96:97] offset:96
-; CHECK-NEXT:    flat_load_dwordx4 v[64:67], v[2:3]
+; CHECK-NEXT:    flat_load_dwordx4 v[52:55], v[52:53] offset:128
+; CHECK-NEXT:    flat_load_dwordx4 v[64:67], v[96:97] offset:96
 ; CHECK-NEXT:    flat_load_dwordx4 v[68:71], v[96:97] offset:48
 ; CHECK-NEXT:    flat_load_dwordx4 v[80:83], v[96:97] offset:64
 ; CHECK-NEXT:    flat_load_dwordx4 v[84:87], v[96:97] offset:32
 ; CHECK-NEXT:    flat_load_dwordx4 v[96:99], v[96:97]
-; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, 48
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, 0, v1, vcc_lo
-; CHECK-NEXT:    v_add_co_u32 v2, vcc_lo, 0x100, v2
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; CHECK-NEXT:    s_add_u32 s4, s4, 0xffffff00
-; CHECK-NEXT:    s_addc_u32 s5, s5, -1
+; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s4
+; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo
+; CHECK-NEXT:    s_add_u32 s4, s4, 0x100
+; CHECK-NEXT:    s_addc_u32 s5, s5, 0
+; CHECK-NEXT:    v_add_co_u32 v102, vcc_lo, v100, 48
+; CHECK-NEXT:    s_cmp_lg_u64 s[4:5], 0x800
+; CHECK-NEXT:    v_add_co_ci_u32_e64 v103, null, 0, v101, vcc_lo
 ; CHECK-NEXT:    s_waitcnt vmcnt(15) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[4:7] offset:128
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[4:7] offset:64
 ; CHECK-NEXT:    s_waitcnt vmcnt(14) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[8:11] offset:64
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[8:11] offset:32
 ; CHECK-NEXT:    s_waitcnt vmcnt(13) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[12:15] offset:32
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[12:15]
 ; CHECK-NEXT:    s_waitcnt vmcnt(12) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[16:19] offset:16
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[16:19] offset:16
 ; CHECK-NEXT:    s_waitcnt vmcnt(11) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[20:23] offset:176
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[20:23] offset:176
 ; CHECK-NEXT:    s_waitcnt vmcnt(10) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[24:27] offset:192
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[24:27] offset:192
 ; CHECK-NEXT:    s_waitcnt vmcnt(9) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[28:31] offset:144
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[28:31] offset:144
 ; CHECK-NEXT:    s_waitcnt vmcnt(8) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[32:35] offset:160
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[32:35] offset:160
 ; CHECK-NEXT:    s_waitcnt vmcnt(7) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[36:39] offset:112
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[36:39] offset:112
 ; CHECK-NEXT:    s_waitcnt vmcnt(6) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[48:51] offset:128
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[48:51] offset:128
 ; CHECK-NEXT:    s_waitcnt vmcnt(5) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[52:55] offset:96
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[52:55] offset:128
 ; CHECK-NEXT:    s_waitcnt vmcnt(4) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[64:67]
-; CHECK-NEXT:    v_add_co_u32 v0, vcc_lo, 0x100, v0
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; CHECK-NEXT:    s_cmp_lg_u64 s[4:5], 0
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[64:67] offset:96
 ; CHECK-NEXT:    s_waitcnt vmcnt(3) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[68:71] offset:48
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[68:71] offset:48
 ; CHECK-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[80:83] offset:64
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[80:83] offset:64
 ; CHECK-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[84:87] offset:32
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[84:87] offset:32
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[96:99]
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[96:99]
 ; CHECK-NEXT:    s_cbranch_scc1 .LBB5_2
-; CHECK-NEXT:  .LBB5_3: ; %Flow15
-; CHECK-NEXT:    s_andn2_saveexec_b32 s6, s6
+; CHECK-NEXT:  .LBB5_3: ; %Flow5
+; CHECK-NEXT:    s_andn2_saveexec_b32 s8, s6
 ; CHECK-NEXT:    s_cbranch_execz .LBB5_6
 ; CHECK-NEXT:  ; %bb.4: ; %memmove_bwd_loop.preheader
-; CHECK-NEXT:    v_add_co_u32 v0, vcc_lo, 0x700, v0
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; CHECK-NEXT:    v_add_co_u32 v2, vcc_lo, 0x700, v2
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; CHECK-NEXT:    s_movk_i32 s4, 0xf800
-; CHECK-NEXT:    s_mov_b32 s5, -1
+; CHECK-NEXT:    s_mov_b64 s[4:5], 0x700
+; CHECK-NEXT:    s_movk_i32 s6, 0xff00
+; CHECK-NEXT:    s_mov_b32 s7, -1
 ; CHECK-NEXT:  .LBB5_5: ; %memmove_bwd_loop
 ; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    v_add_co_u32 v96, vcc_lo, v2, 48
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v97, null, 0, v3, vcc_lo
+; CHECK-NEXT:    v_add_co_u32 v20, vcc_lo, v2, s4
+; CHECK-NEXT:    v_add_co_ci_u32_e64 v21, null, s5, v3, vcc_lo
+; CHECK-NEXT:    v_add_co_u32 v96, vcc_lo, v20, 48
+; CHECK-NEXT:    v_add_co_ci_u32_e64 v97, null, 0, v21, vcc_lo
 ; CHECK-NEXT:    s_clause 0xf
-; CHECK-NEXT:    flat_load_dwordx4 v[4:7], v[2:3] offset:128
-; CHECK-NEXT:    flat_load_dwordx4 v[8:11], v[2:3] offset:64
-; CHECK-NEXT:    flat_load_dwordx4 v[12:15], v[2:3] offset:32
-; CHECK-NEXT:    flat_load_dwordx4 v[16:19], v[2:3] offset:16
-; CHECK-NEXT:    flat_load_dwordx4 v[20:23], v[2:3]
+; CHECK-NEXT:    flat_load_dwordx4 v[4:7], v[20:21] offset:64
+; CHECK-NEXT:    flat_load_dwordx4 v[8:11], v[20:21] offset:32
+; CHECK-NEXT:    flat_load_dwordx4 v[12:15], v[20:21]
+; CHECK-NEXT:    flat_load_dwordx4 v[16:19], v[20:21] offset:16
+; CHECK-NEXT:    flat_load_dwordx4 v[20:23], v[20:21] offset:128
 ; CHECK-NEXT:    flat_load_dwordx4 v[24:27], v[96:97] offset:176
 ; CHECK-NEXT:    flat_load_dwordx4 v[28:31], v[96:97] offset:192
 ; CHECK-NEXT:    flat_load_dwordx4 v[32:35], v[96:97] offset:144
@@ -5490,400 +5489,401 @@ define void @memmove_p0_p0_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(0
 ; CHECK-NEXT:    flat_load_dwordx4 v[80:83], v[96:97] offset:64
 ; CHECK-NEXT:    flat_load_dwordx4 v[84:87], v[96:97] offset:32
 ; CHECK-NEXT:    flat_load_dwordx4 v[96:99], v[96:97]
-; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, 48
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, 0, v1, vcc_lo
-; CHECK-NEXT:    v_add_co_u32 v2, vcc_lo, 0xffffff00, v2
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v3, null, -1, v3, vcc_lo
-; CHECK-NEXT:    s_add_u32 s4, s4, 0x100
-; CHECK-NEXT:    s_addc_u32 s5, s5, 0
+; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s4
+; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo
+; CHECK-NEXT:    s_add_u32 s4, s4, 0xffffff00
+; CHECK-NEXT:    s_addc_u32 s5, s5, -1
+; CHECK-NEXT:    v_add_co_u32 v102, vcc_lo, v100, 48
+; CHECK-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]
+; CHECK-NEXT:    v_add_co_ci_u32_e64 v103, null, 0, v101, vcc_lo
 ; CHECK-NEXT:    s_waitcnt vmcnt(15) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[4:7] offset:128
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[4:7] offset:64
 ; CHECK-NEXT:    s_waitcnt vmcnt(14) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[8:11] offset:64
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[8:11] offset:32
 ; CHECK-NEXT:    s_waitcnt vmcnt(13) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[12:15] offset:32
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[12:15]
 ; CHECK-NEXT:    s_waitcnt vmcnt(12) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[16:19] offset:16
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[16:19] offset:16
 ; CHECK-NEXT:    s_waitcnt vmcnt(11) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[20:23]
-; CHECK-NEXT:    v_add_co_u32 v0, vcc_lo, 0xffffff00, v0
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo
-; CHECK-NEXT:    s_cmp_eq_u64 s[4:5], 0
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[20:23] offset:128
 ; CHECK-NEXT:    s_waitcnt vmcnt(10) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[24:27] offset:176
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[24:27] offset:176
 ; CHECK-NEXT:    s_waitcnt vmcnt(9) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[28:31] offset:192
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[28:31] offset:192
 ; CHECK-NEXT:    s_waitcnt vmcnt(8) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[32:35] offset:144
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[32:35] offset:144
 ; CHECK-NEXT:    s_waitcnt vmcnt(7) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[36:39] offset:160
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[36:39] offset:160
 ; CHECK-NEXT:    s_waitcnt vmcnt(6) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[48:51] offset:112
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[48:51] offset:112
 ; CHECK-NEXT:    s_waitcnt vmcnt(5) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[52:55] offset:128
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[52:55] offset:128
 ; CHECK-NEXT:    s_waitcnt vmcnt(4) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[64:67] offset:96
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[64:67] offset:96
 ; CHECK-NEXT:    s_waitcnt vmcnt(3) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[68:71] offset:48
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[68:71] offset:48
 ; CHECK-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[80:83] offset:64
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[80:83] offset:64
 ; CHECK-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[84:87] offset:32
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[84:87] offset:32
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[96:99]
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[96:99]
 ; CHECK-NEXT:    s_cbranch_scc0 .LBB5_5
-; CHECK-NEXT:  .LBB5_6: ; %Flow16
-; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s6
+; CHECK-NEXT:  .LBB5_6: ; %Flow6
+; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s8
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; ALIGNED-LABEL: memmove_p0_p0_sz2048:
 ; ALIGNED:       ; %bb.0: ; %entry
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v58, off, s[0:3], s32 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v72, off, s[0:3], s32 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_mov_b32 s4, exec_lo
 ; ALIGNED-NEXT:    v_cmpx_ge_u64_e32 v[2:3], v[0:1]
 ; ALIGNED-NEXT:    s_xor_b32 s6, exec_lo, s4
 ; ALIGNED-NEXT:    s_cbranch_execz .LBB5_3
 ; ALIGNED-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader
-; ALIGNED-NEXT:    s_mov_b64 s[4:5], 0x800
+; ALIGNED-NEXT:    s_mov_b64 s[4:5], 0
 ; ALIGNED-NEXT:  .LBB5_2: ; %memmove_fwd_loop
 ; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=1
-; ALIGNED-NEXT:    flat_load_dwordx4 v[20:23], v[2:3] offset:128
-; ALIGNED-NEXT:    v_add_co_u32 v36, vcc_lo, v2, 48
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v37, null, 0, v3, vcc_lo
+; ALIGNED-NEXT:    v_add_co_u32 v4, vcc_lo, v2, s4
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v5, null, s5, v3, vcc_lo
+; ALIGNED-NEXT:    v_add_co_u32 v24, vcc_lo, v4, 48
+; ALIGNED-NEXT:    flat_load_dwordx4 v[20:23], v[4:5] offset:128
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v25, null, 0, v5, vcc_lo
 ; ALIGNED-NEXT:    s_clause 0xe
-; ALIGNED-NEXT:    flat_load_dwordx4 v[24:27], v[2:3] offset:64
-; ALIGNED-NEXT:    flat_load_dwordx4 v[28:31], v[2:3] offset:16
-; ALIGNED-NEXT:    flat_load_dwordx4 v[32:35], v[2:3] offset:32
-; ALIGNED-NEXT:    flat_load_dwordx4 v[4:7], v[36:37]
-; ALIGNED-NEXT:    flat_load_dwordx4 v[8:11], v[36:37] offset:32
-; ALIGNED-NEXT:    flat_load_dwordx4 v[12:15], v[36:37] offset:48
-; ALIGNED-NEXT:    flat_load_dwordx4 v[16:19], v[36:37] offset:64
-; ALIGNED-NEXT:    flat_load_dwordx4 v[84:87], v[36:37] offset:96
-; ALIGNED-NEXT:    flat_load_dwordx4 v[80:83], v[36:37] offset:112
-; ALIGNED-NEXT:    flat_load_dwordx4 v[68:71], v[36:37] offset:128
-; ALIGNED-NEXT:    flat_load_dwordx4 v[64:67], v[36:37] offset:144
-; ALIGNED-NEXT:    flat_load_dwordx4 v[52:55], v[36:37] offset:160
-; ALIGNED-NEXT:    flat_load_dwordx4 v[48:51], v[36:37] offset:176
-; ALIGNED-NEXT:    flat_load_dwordx4 v[36:39], v[36:37] offset:192
-; ALIGNED-NEXT:    flat_load_dwordx4 v[100:103], v[2:3]
-; ALIGNED-NEXT:    s_add_u32 s4, s4, 0xffffff00
-; ALIGNED-NEXT:    s_addc_u32 s5, s5, -1
+; ALIGNED-NEXT:    flat_load_dwordx4 v[34:37], v[4:5]
+; ALIGNED-NEXT:    flat_load_dwordx4 v[30:33], v[4:5] offset:16
+; ALIGNED-NEXT:    flat_load_dwordx4 v[26:29], v[4:5] offset:32
+; ALIGNED-NEXT:    flat_load_dwordx4 v[48:51], v[4:5] offset:64
+; ALIGNED-NEXT:    flat_load_dwordx4 v[4:7], v[24:25]
+; ALIGNED-NEXT:    flat_load_dwordx4 v[8:11], v[24:25] offset:32
+; ALIGNED-NEXT:    flat_load_dwordx4 v[12:15], v[24:25] offset:48
+; ALIGNED-NEXT:    flat_load_dwordx4 v[16:19], v[24:25] offset:64
+; ALIGNED-NEXT:    flat_load_dwordx4 v[112:115], v[24:25] offset:96
+; ALIGNED-NEXT:    flat_load_dwordx4 v[82:85], v[24:25] offset:112
+; ALIGNED-NEXT:    flat_load_dwordx4 v[116:119], v[24:25] offset:128
+; ALIGNED-NEXT:    flat_load_dwordx4 v[66:69], v[24:25] offset:144
+; ALIGNED-NEXT:    flat_load_dwordx4 v[40:43], v[24:25] offset:160
+; ALIGNED-NEXT:    flat_load_dwordx4 v[52:55], v[24:25] offset:176
+; ALIGNED-NEXT:    flat_load_dwordx4 v[44:47], v[24:25] offset:192
+; ALIGNED-NEXT:    v_add_co_u32 v24, vcc_lo, v0, s4
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v25, null, s5, v1, vcc_lo
+; ALIGNED-NEXT:    s_add_u32 s4, s4, 0x100
+; ALIGNED-NEXT:    s_addc_u32 s5, s5, 0
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(15) lgkmcnt(15)
-; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:128
-; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:132
-; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:136
-; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:140
+; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:160
+; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:164
+; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:168
+; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:172
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v112, off, s[0:3], s32 offset:140
-; ALIGNED-NEXT:    buffer_load_dword v113, off, s[0:3], s32 offset:136
-; ALIGNED-NEXT:    buffer_load_dword v114, off, s[0:3], s32 offset:132
-; ALIGNED-NEXT:    buffer_load_dword v115, off, s[0:3], s32 offset:128
-; ALIGNED-NEXT:    v_add_co_u32 v22, vcc_lo, v0, 6
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v23, null, 0, v1, vcc_lo
-; ALIGNED-NEXT:    v_add_co_u32 v20, vcc_lo, v0, 3
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v21, null, 0, v1, vcc_lo
-; ALIGNED-NEXT:    v_add_co_u32 v2, vcc_lo, 0x100, v2
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; ALIGNED-NEXT:    s_cmp_lg_u64 s[4:5], 0
+; ALIGNED-NEXT:    buffer_load_dword v102, off, s[0:3], s32 offset:172
+; ALIGNED-NEXT:    buffer_load_dword v103, off, s[0:3], s32 offset:168
+; ALIGNED-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:164
+; ALIGNED-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:160
+; ALIGNED-NEXT:    v_add_co_u32 v22, vcc_lo, v24, 6
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v23, null, 0, v25, vcc_lo
+; ALIGNED-NEXT:    v_add_co_u32 v20, vcc_lo, v24, 3
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v21, null, 0, v25, vcc_lo
+; ALIGNED-NEXT:    s_cmp_lg_u64 s[4:5], 0x800
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v112 offset:136
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:137
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v102 offset:136
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:137
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v113 offset:132
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:133
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v103 offset:132
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:133
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v114 offset:128
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v114 offset:129
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v115 offset:124
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v115 offset:128
-; ALIGNED-NEXT:    s_waitcnt lgkmcnt(22)
-; ALIGNED-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:192
-; ALIGNED-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:196
-; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:200
-; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:204
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v56 offset:128
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v56 offset:129
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v57 offset:124
+; ALIGNED-NEXT:    flat_store_byte v[24:25], v57 offset:128
+; ALIGNED-NEXT:    s_waitcnt lgkmcnt(19)
+; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:224
+; ALIGNED-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:228
+; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:232
+; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:236
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v99, off, s[0:3], s32 offset:204
-; ALIGNED-NEXT:    buffer_load_dword v98, off, s[0:3], s32 offset:200
-; ALIGNED-NEXT:    buffer_load_dword v97, off, s[0:3], s32 offset:196
-; ALIGNED-NEXT:    buffer_load_dword v96, off, s[0:3], s32 offset:192
+; ALIGNED-NEXT:    buffer_load_dword v101, off, s[0:3], s32 offset:236
+; ALIGNED-NEXT:    buffer_load_dword v100, off, s[0:3], s32 offset:232
+; ALIGNED-NEXT:    buffer_load_dword v99, off, s[0:3], s32 offset:228
+; ALIGNED-NEXT:    buffer_load_dword v98, off, s[0:3], s32 offset:224
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v99 offset:72
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:73
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v101 offset:72
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:73
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v98 offset:68
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:69
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v100 offset:68
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:69
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v97 offset:64
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:65
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v99 offset:64
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:65
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v96 offset:60
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v96 offset:64
-; ALIGNED-NEXT:    s_waitcnt lgkmcnt(28)
-; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:288
-; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292
-; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:296
-; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:300
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v98 offset:60
+; ALIGNED-NEXT:    flat_store_byte v[24:25], v98 offset:64
+; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:320
+; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:324
+; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:328
+; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:332
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:296
-; ALIGNED-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:300
-; ALIGNED-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:288
-; ALIGNED-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:292
+; ALIGNED-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:328
+; ALIGNED-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:332
+; ALIGNED-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:320
+; ALIGNED-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:324
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v27 offset:36
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v27 offset:37
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v29 offset:36
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v29 offset:37
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v26 offset:40
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v26 offset:41
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v28 offset:40
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v28 offset:41
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v25 offset:28
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v25 offset:32
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v24 offset:32
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v24 offset:33
-; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:240
-; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:244
-; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:248
-; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:252
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v27 offset:28
+; ALIGNED-NEXT:    flat_store_byte v[24:25], v27 offset:32
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v26 offset:32
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v26 offset:33
+; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:272
+; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:276
+; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:280
+; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:252
-; ALIGNED-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:248
-; ALIGNED-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:244
-; ALIGNED-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:240
+; ALIGNED-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:284
+; ALIGNED-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:280
+; ALIGNED-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:276
+; ALIGNED-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:272
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v31 offset:24
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v31 offset:25
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v33 offset:24
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v33 offset:25
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v30 offset:20
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v30 offset:21
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v32 offset:20
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v32 offset:21
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v29 offset:16
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v29 offset:17
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v28 offset:12
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v28 offset:16
-; ALIGNED-NEXT:    s_waitcnt lgkmcnt(32)
-; ALIGNED-NEXT:    buffer_store_dword v100, off, s[0:3], s32 offset:256
-; ALIGNED-NEXT:    buffer_store_dword v101, off, s[0:3], s32 offset:260
-; ALIGNED-NEXT:    buffer_store_dword v102, off, s[0:3], s32 offset:264
-; ALIGNED-NEXT:    buffer_store_dword v103, off, s[0:3], s32 offset:268
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v31 offset:16
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v31 offset:17
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v30 offset:12
+; ALIGNED-NEXT:    flat_store_byte v[24:25], v30 offset:16
+; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:288
+; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:292
+; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:296
+; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:300
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:268
-; ALIGNED-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:264
-; ALIGNED-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:260
-; ALIGNED-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:256
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v112
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v112
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v113
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v113
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v114
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 8, v114
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v115
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v115
+; ALIGNED-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:300
+; ALIGNED-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:296
+; ALIGNED-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:292
+; ALIGNED-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:288
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v35 offset:8
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v35 offset:9
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v37 offset:8
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v37 offset:9
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v34 offset:4
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v34 offset:8
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v36 offset:4
+; ALIGNED-NEXT:    flat_store_byte v[24:25], v36 offset:8
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v33
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v33 offset:4
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[0:1], v32 offset:2
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v32
-; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:80
-; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:84
-; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:88
-; ALIGNED-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:92
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v35
+; ALIGNED-NEXT:    flat_store_byte v[24:25], v35 offset:4
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[24:25], v34 offset:2
+; ALIGNED-NEXT:    flat_store_byte v[24:25], v34
+; ALIGNED-NEXT:    s_waitcnt lgkmcnt(40)
+; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:112
+; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:116
+; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:120
+; ALIGNED-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:124
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:92
-; ALIGNED-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:88
-; ALIGNED-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:84
-; ALIGNED-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:80
+; ALIGNED-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:124
+; ALIGNED-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:120
+; ALIGNED-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:116
+; ALIGNED-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:112
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v39 offset:248
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:249
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v49 offset:248
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v49 offset:249
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v38 offset:244
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v38 offset:245
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v48 offset:244
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v48 offset:245
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v37 offset:240
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v37 offset:241
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v36 offset:236
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v36 offset:237
-; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:96
-; ALIGNED-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:100
-; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:104
-; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:108
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v39 offset:240
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:241
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v38 offset:236
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v38 offset:237
+; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:128
+; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:132
+; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:136
+; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:140
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:108
-; ALIGNED-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:104
-; ALIGNED-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:100
-; ALIGNED-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:96
+; ALIGNED-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:140
+; ALIGNED-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:136
+; ALIGNED-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:132
+; ALIGNED-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:128
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v51 offset:232
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v51 offset:233
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v53 offset:232
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v53 offset:233
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v50 offset:228
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v50 offset:229
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v52 offset:228
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v52 offset:229
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v49 offset:224
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v49 offset:225
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v48 offset:220
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v48 offset:221
-; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:48
-; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:52
-; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:56
-; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:60
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v51 offset:224
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v51 offset:225
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v50 offset:220
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v50 offset:221
+; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:80
+; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:84
+; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:88
+; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:92
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:60
-; ALIGNED-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:56
-; ALIGNED-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:52
-; ALIGNED-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:48
+; ALIGNED-NEXT:    buffer_load_dword v65, off, s[0:3], s32 offset:92
+; ALIGNED-NEXT:    buffer_load_dword v64, off, s[0:3], s32 offset:88
+; ALIGNED-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:84
+; ALIGNED-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:80
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v55 offset:216
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v55 offset:217
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v65 offset:216
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v65 offset:217
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v54 offset:212
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v54 offset:213
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v64 offset:212
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v64 offset:213
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v53 offset:208
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v53 offset:209
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v52 offset:204
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v52 offset:205
-; ALIGNED-NEXT:    buffer_store_dword v64, off, s[0:3], s32 offset:64
-; ALIGNED-NEXT:    buffer_store_dword v65, off, s[0:3], s32 offset:68
-; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:72
-; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:76
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v55 offset:208
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v55 offset:209
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v54 offset:204
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v54 offset:205
+; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:96
+; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:100
+; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:104
+; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:108
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v67, off, s[0:3], s32 offset:76
-; ALIGNED-NEXT:    buffer_load_dword v66, off, s[0:3], s32 offset:72
-; ALIGNED-NEXT:    buffer_load_dword v65, off, s[0:3], s32 offset:68
-; ALIGNED-NEXT:    buffer_load_dword v64, off, s[0:3], s32 offset:64
+; ALIGNED-NEXT:    buffer_load_dword v69, off, s[0:3], s32 offset:108
+; ALIGNED-NEXT:    buffer_load_dword v68, off, s[0:3], s32 offset:104
+; ALIGNED-NEXT:    buffer_load_dword v67, off, s[0:3], s32 offset:100
+; ALIGNED-NEXT:    buffer_load_dword v66, off, s[0:3], s32 offset:96
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v67 offset:200
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v67 offset:201
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v69 offset:200
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v69 offset:201
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v66 offset:196
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v66 offset:197
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v68 offset:196
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v68 offset:197
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v65 offset:192
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v65 offset:193
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v67 offset:192
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v67 offset:193
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v64 offset:188
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v64 offset:189
-; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:144
-; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:148
-; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:152
-; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:156
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v66 offset:188
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v66 offset:189
+; ALIGNED-NEXT:    buffer_store_dword v116, off, s[0:3], s32 offset:176
+; ALIGNED-NEXT:    buffer_store_dword v117, off, s[0:3], s32 offset:180
+; ALIGNED-NEXT:    buffer_store_dword v118, off, s[0:3], s32 offset:184
+; ALIGNED-NEXT:    buffer_store_dword v119, off, s[0:3], s32 offset:188
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v71, off, s[0:3], s32 offset:156
-; ALIGNED-NEXT:    buffer_load_dword v70, off, s[0:3], s32 offset:152
-; ALIGNED-NEXT:    buffer_load_dword v69, off, s[0:3], s32 offset:148
-; ALIGNED-NEXT:    buffer_load_dword v68, off, s[0:3], s32 offset:144
+; ALIGNED-NEXT:    buffer_load_dword v81, off, s[0:3], s32 offset:188
+; ALIGNED-NEXT:    buffer_load_dword v80, off, s[0:3], s32 offset:184
+; ALIGNED-NEXT:    buffer_load_dword v71, off, s[0:3], s32 offset:180
+; ALIGNED-NEXT:    buffer_load_dword v70, off, s[0:3], s32 offset:176
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v71 offset:184
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v71 offset:185
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v81 offset:184
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v81 offset:185
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v70 offset:180
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v70 offset:181
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v80 offset:180
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v80 offset:181
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v69 offset:176
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v69 offset:177
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v68 offset:172
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v68 offset:173
-; ALIGNED-NEXT:    buffer_store_dword v80, off, s[0:3], s32 offset:160
-; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:164
-; ALIGNED-NEXT:    buffer_store_dword v82, off, s[0:3], s32 offset:168
-; ALIGNED-NEXT:    buffer_store_dword v83, off, s[0:3], s32 offset:172
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v71 offset:176
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v71 offset:177
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v70 offset:172
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v70 offset:173
+; ALIGNED-NEXT:    buffer_store_dword v82, off, s[0:3], s32 offset:192
+; ALIGNED-NEXT:    buffer_store_dword v83, off, s[0:3], s32 offset:196
+; ALIGNED-NEXT:    buffer_store_dword v84, off, s[0:3], s32 offset:200
+; ALIGNED-NEXT:    buffer_store_dword v85, off, s[0:3], s32 offset:204
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v83, off, s[0:3], s32 offset:172
-; ALIGNED-NEXT:    buffer_load_dword v82, off, s[0:3], s32 offset:168
-; ALIGNED-NEXT:    buffer_load_dword v81, off, s[0:3], s32 offset:164
-; ALIGNED-NEXT:    buffer_load_dword v80, off, s[0:3], s32 offset:160
+; ALIGNED-NEXT:    buffer_load_dword v85, off, s[0:3], s32 offset:204
+; ALIGNED-NEXT:    buffer_load_dword v84, off, s[0:3], s32 offset:200
+; ALIGNED-NEXT:    buffer_load_dword v83, off, s[0:3], s32 offset:196
+; ALIGNED-NEXT:    buffer_load_dword v82, off, s[0:3], s32 offset:192
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v83 offset:168
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v83 offset:169
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v85 offset:168
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v85 offset:169
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v82 offset:164
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v82 offset:165
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v84 offset:164
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v84 offset:165
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v81 offset:160
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v81 offset:161
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v80 offset:156
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v80 offset:157
-; ALIGNED-NEXT:    buffer_store_dword v84, off, s[0:3], s32 offset:112
-; ALIGNED-NEXT:    buffer_store_dword v85, off, s[0:3], s32 offset:116
-; ALIGNED-NEXT:    buffer_store_dword v86, off, s[0:3], s32 offset:120
-; ALIGNED-NEXT:    buffer_store_dword v87, off, s[0:3], s32 offset:124
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v83 offset:160
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v83 offset:161
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v82 offset:156
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v82 offset:157
+; ALIGNED-NEXT:    buffer_store_dword v112, off, s[0:3], s32 offset:144
+; ALIGNED-NEXT:    buffer_store_dword v113, off, s[0:3], s32 offset:148
+; ALIGNED-NEXT:    buffer_store_dword v114, off, s[0:3], s32 offset:152
+; ALIGNED-NEXT:    buffer_store_dword v115, off, s[0:3], s32 offset:156
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v87, off, s[0:3], s32 offset:124
-; ALIGNED-NEXT:    buffer_load_dword v86, off, s[0:3], s32 offset:120
-; ALIGNED-NEXT:    buffer_load_dword v85, off, s[0:3], s32 offset:116
-; ALIGNED-NEXT:    buffer_load_dword v84, off, s[0:3], s32 offset:112
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:140
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:138
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:136
+; ALIGNED-NEXT:    buffer_load_dword v97, off, s[0:3], s32 offset:156
+; ALIGNED-NEXT:    buffer_load_dword v96, off, s[0:3], s32 offset:152
+; ALIGNED-NEXT:    buffer_load_dword v87, off, s[0:3], s32 offset:148
+; ALIGNED-NEXT:    buffer_load_dword v86, off, s[0:3], s32 offset:144
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v102
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 8, v102
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v103
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v103
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 8, v56
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:140
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v56
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:138
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v57
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v57
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:136
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:134
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:132
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:130
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v114 offset:128
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v114 offset:130
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:128
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v115 offset:126
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v87 offset:152
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v87 offset:153
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v97 offset:152
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:153
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v86 offset:148
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v86 offset:149
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v96 offset:148
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v96 offset:149
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v85 offset:144
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v85 offset:145
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v84 offset:140
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v84 offset:141
-; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:208
-; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:212
-; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:216
-; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:220
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v87 offset:144
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v87 offset:145
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v86 offset:140
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v86 offset:141
+; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:240
+; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:244
+; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:248
+; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:252
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:220
-; ALIGNED-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:216
-; ALIGNED-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:212
-; ALIGNED-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:208
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v99
+; ALIGNED-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:252
+; ALIGNED-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:248
+; ALIGNED-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:244
+; ALIGNED-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:240
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v101
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v101
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v100
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 8, v100
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v99
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v99
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v98
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v98
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 8, v98
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v97
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 8, v97
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v96
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 8, v96
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v32
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 8, v32
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v35
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 8, v35
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v34
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 8, v34
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v33
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 8, v33
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 8, v34
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v37
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 8, v37
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 24, v34
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v19 offset:120
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v19 offset:121
@@ -5896,15 +5896,15 @@ define void @memmove_p0_p0_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(0
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v16 offset:108
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v16 offset:109
-; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:224
-; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:228
-; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:232
-; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:236
+; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:256
+; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:260
+; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:264
+; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:268
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:236
-; ALIGNED-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:232
-; ALIGNED-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:228
-; ALIGNED-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:224
+; ALIGNED-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:268
+; ALIGNED-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:264
+; ALIGNED-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:260
+; ALIGNED-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:256
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v15 offset:104
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v15 offset:105
@@ -5917,23 +5917,23 @@ define void @memmove_p0_p0_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(0
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v12 offset:92
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v12 offset:93
-; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:176
-; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:180
-; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:184
-; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:188
+; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:208
+; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:212
+; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:216
+; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:220
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:188
-; ALIGNED-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:184
-; ALIGNED-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:180
-; ALIGNED-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:176
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:76
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:74
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:72
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:70
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:68
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:66
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:64
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v96 offset:62
+; ALIGNED-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:220
+; ALIGNED-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:216
+; ALIGNED-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:212
+; ALIGNED-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:208
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:76
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:74
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:72
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:70
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:68
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:66
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:64
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:62
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v11 offset:88
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v11 offset:89
@@ -5946,216 +5946,218 @@ define void @memmove_p0_p0_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(0
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v8 offset:76
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v8 offset:77
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:272
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:276
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:280
-; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:284
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:304
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:308
+; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:312
+; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:316
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:284
-; ALIGNED-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:280
-; ALIGNED-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:276
-; ALIGNED-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:272
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v24
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 8, v24
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 24, v27
+; ALIGNED-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:316
+; ALIGNED-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:312
+; ALIGNED-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:308
+; ALIGNED-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:304
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v26
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v29
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 8, v29
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v28
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 8, v28
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v27
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 8, v27
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 24, v26
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 8, v26
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v25
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 8, v25
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v31
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:36
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v49
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v33
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 8, v33
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:36
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v50
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 8, v50
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v32
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 8, v32
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v31
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 8, v31
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v30
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 8, v30
+; ALIGNED-NEXT:    flat_store_byte v[24:25], v114 offset:1
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v36
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 8, v36
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 24, v35
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 8, v35
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v49
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v49
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v30
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v24 offset:34
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v48
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:40
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v48
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 8, v48
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 8, v30
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v29
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 8, v29
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v28
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 8, v28
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v96 offset:40
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 24, v39
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v29 offset:38
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 24, v39
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 8, v39
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v27 offset:38
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 24, v38
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:44
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v38
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 8, v38
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:44
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 24, v37
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 8, v37
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v26 offset:42
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 24, v36
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 8, v36
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:32
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v51
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v28 offset:42
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 24, v53
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v53
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:32
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v52
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 8, v52
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v27 offset:30
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 24, v51
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v51
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v25 offset:30
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 24, v50
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 8, v50
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:28
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v55
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:228
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v11
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v26 offset:34
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 24, v65
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v65
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:28
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v64
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 8, v64
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v33 offset:26
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 24, v55
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v50 offset:222
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v8
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v8, 8, v8
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v55
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v11, 8, v11
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v31 offset:26
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 24, v54
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v49 offset:226
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 24, v10
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:24
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v54
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 8, v54
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v32 offset:1
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v10, 8, v10
-; ALIGNED-NEXT:    v_add_co_u32 v0, vcc_lo, 0x100, v0
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:24
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v53
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v24 offset:224
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v9
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v48 offset:222
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 24, v8
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v8, 8, v8
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v53
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v30 offset:22
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v52
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 8, v52
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:20
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v67
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v32 offset:22
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 24, v69
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v69
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:20
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v68
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 8, v68
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v31 offset:18
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 24, v67
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v67
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v29 offset:18
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 24, v66
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:16
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v66
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 8, v66
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:16
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v65
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v65
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v28 offset:14
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 24, v64
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 8, v64
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:12
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v71
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v30 offset:14
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v81
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v81
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v115 offset:12
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v80
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 8, v80
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v37 offset:10
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 24, v71
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v71
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v35 offset:10
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 24, v70
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v24 offset:8
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v70
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 8, v70
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:8
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v69
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v69
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v34 offset:6
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 24, v68
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 8, v68
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v114 offset:4
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v83
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v25 offset:6
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 24, v85
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 8, v85
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v36 offset:4
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 24, v84
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 8, v84
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v35 offset:2
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 24, v83
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v83
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v33 offset:2
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 24, v82
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v34
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 24, v82
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 8, v82
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v115
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v81
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v81
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 24, v80
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 8, v80
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v96 offset:252
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 24, v87
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v114 offset:252
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v97
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 8, v97
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v49 offset:250
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 24, v96
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 8, v96
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:248
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v87
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 8, v87
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:250
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 24, v86
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v48 offset:246
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 24, v86
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 8, v86
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v27 offset:248
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 24, v85
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 8, v85
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v38 offset:246
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 24, v84
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 8, v84
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:244
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 24, v19
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v29 offset:244
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 24, v19
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v19, 8, v19
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v37 offset:242
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 24, v18
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:242
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 24, v18
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v18, 8, v18
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v26 offset:240
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 24, v17
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:240
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v17
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v17, 8, v17
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v36 offset:238
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 24, v16
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v38 offset:238
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 24, v16
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v16, 8, v16
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:236
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v15
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v28 offset:236
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 24, v15
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v15, 8, v15
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v51 offset:234
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 24, v14
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v53 offset:234
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 24, v14
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v14, 8, v14
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v25 offset:232
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 24, v13
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:232
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v13
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v13, 8, v13
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v50 offset:230
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v12
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v52 offset:230
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v12
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v12, 8, v12
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v27 offset:228
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 24, v11
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v11, 8, v11
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v51 offset:226
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 24, v10
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v10, 8, v10
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:224
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v9
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v9, 8, v9
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:220
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v55 offset:218
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v31 offset:216
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v54 offset:214
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:212
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v53 offset:210
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v30 offset:208
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v52 offset:206
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:204
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v67 offset:202
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v29 offset:200
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v66 offset:198
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:196
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v65 offset:194
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v28 offset:192
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v64 offset:190
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:188
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v71 offset:186
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v35 offset:184
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v70 offset:182
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:180
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v69 offset:178
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v34 offset:176
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v68 offset:174
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v114 offset:172
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v83 offset:170
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v33 offset:168
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v82 offset:166
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v115 offset:164
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v81 offset:162
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v32 offset:160
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v80 offset:158
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v96 offset:156
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v87 offset:154
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:152
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v86 offset:150
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v27 offset:148
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v85 offset:146
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v38 offset:144
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v84 offset:142
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:124
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v26 offset:220
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v65 offset:218
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:216
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v64 offset:214
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v33 offset:212
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v55 offset:210
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:208
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v54 offset:206
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v32 offset:204
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v69 offset:202
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:200
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v68 offset:198
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v31 offset:196
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v67 offset:194
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:192
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v66 offset:190
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v30 offset:188
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v81 offset:186
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v115 offset:184
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v80 offset:182
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v37 offset:180
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v71 offset:178
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v24 offset:176
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v70 offset:174
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v25 offset:172
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v85 offset:170
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v36 offset:168
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v84 offset:166
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v35 offset:164
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v83 offset:162
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v34 offset:160
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v82 offset:158
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v114 offset:156
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:154
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v49 offset:152
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v96 offset:150
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:148
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v87 offset:146
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v48 offset:144
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v86 offset:142
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v29 offset:124
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v19 offset:122
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v37 offset:120
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:120
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v18 offset:118
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v26 offset:116
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:116
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v17 offset:114
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v36 offset:112
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v38 offset:112
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v16 offset:110
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:108
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v28 offset:108
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v15 offset:106
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v51 offset:104
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v53 offset:104
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v14 offset:102
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v25 offset:100
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:100
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v13 offset:98
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v50 offset:96
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v52 offset:96
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v12 offset:94
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:92
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v27 offset:92
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v11 offset:90
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v49 offset:88
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v51 offset:88
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v10 offset:86
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v24 offset:84
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:84
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v9 offset:82
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v48 offset:80
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v50 offset:80
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v8 offset:78
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v7 offset:56
@@ -6186,659 +6188,661 @@ define void @memmove_p0_p0_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(0
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v11 offset:48
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v4 offset:46
 ; ALIGNED-NEXT:    s_cbranch_scc1 .LBB5_2
-; ALIGNED-NEXT:  .LBB5_3: ; %Flow15
-; ALIGNED-NEXT:    s_andn2_saveexec_b32 s6, s6
+; ALIGNED-NEXT:  .LBB5_3: ; %Flow5
+; ALIGNED-NEXT:    s_andn2_saveexec_b32 s8, s6
 ; ALIGNED-NEXT:    s_cbranch_execz .LBB5_6
 ; ALIGNED-NEXT:  ; %bb.4: ; %memmove_bwd_loop.preheader
-; ALIGNED-NEXT:    v_add_co_u32 v16, vcc_lo, 0x700, v0
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v17, null, 0, v1, vcc_lo
-; ALIGNED-NEXT:    v_add_co_u32 v18, vcc_lo, 0x700, v2
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v19, null, 0, v3, vcc_lo
-; ALIGNED-NEXT:    s_movk_i32 s4, 0xf800
-; ALIGNED-NEXT:    s_mov_b32 s5, -1
+; ALIGNED-NEXT:    s_mov_b64 s[4:5], 0x700
+; ALIGNED-NEXT:    s_movk_i32 s6, 0xff00
+; ALIGNED-NEXT:    s_mov_b32 s7, -1
 ; ALIGNED-NEXT:  .LBB5_5: ; %memmove_bwd_loop
 ; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=1
-; ALIGNED-NEXT:    flat_load_dwordx4 v[20:23], v[18:19] offset:128
-; ALIGNED-NEXT:    v_add_co_u32 v37, vcc_lo, v18, 48
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v38, null, 0, v19, vcc_lo
+; ALIGNED-NEXT:    v_add_co_u32 v4, vcc_lo, v2, s4
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v5, null, s5, v3, vcc_lo
+; ALIGNED-NEXT:    v_add_co_u32 v24, vcc_lo, v4, 48
+; ALIGNED-NEXT:    flat_load_dwordx4 v[20:23], v[4:5] offset:128
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v25, null, 0, v5, vcc_lo
 ; ALIGNED-NEXT:    s_clause 0xe
-; ALIGNED-NEXT:    flat_load_dwordx4 v[24:27], v[18:19] offset:64
-; ALIGNED-NEXT:    flat_load_dwordx4 v[33:36], v[18:19] offset:16
-; ALIGNED-NEXT:    flat_load_dwordx4 v[28:31], v[18:19] offset:32
-; ALIGNED-NEXT:    flat_load_dwordx4 v[0:3], v[37:38]
-; ALIGNED-NEXT:    flat_load_dwordx4 v[4:7], v[37:38] offset:32
-; ALIGNED-NEXT:    flat_load_dwordx4 v[8:11], v[37:38] offset:48
-; ALIGNED-NEXT:    flat_load_dwordx4 v[12:15], v[37:38] offset:64
-; ALIGNED-NEXT:    flat_load_dwordx4 v[100:103], v[37:38] offset:96
-; ALIGNED-NEXT:    flat_load_dwordx4 v[112:115], v[37:38] offset:112
-; ALIGNED-NEXT:    flat_load_dwordx4 v[116:119], v[37:38] offset:128
-; ALIGNED-NEXT:    flat_load_dwordx4 v[39:42], v[37:38] offset:144
-; ALIGNED-NEXT:    flat_load_dwordx4 v[43:46], v[37:38] offset:160
-; ALIGNED-NEXT:    flat_load_dwordx4 v[52:55], v[37:38] offset:176
-; ALIGNED-NEXT:    flat_load_dwordx4 v[66:69], v[37:38] offset:192
-; ALIGNED-NEXT:    flat_load_dwordx4 v[80:83], v[18:19]
-; ALIGNED-NEXT:    s_add_u32 s4, s4, 0x100
-; ALIGNED-NEXT:    s_addc_u32 s5, s5, 0
+; ALIGNED-NEXT:    flat_load_dwordx4 v[26:29], v[4:5]
+; ALIGNED-NEXT:    flat_load_dwordx4 v[35:38], v[4:5] offset:16
+; ALIGNED-NEXT:    flat_load_dwordx4 v[30:33], v[4:5] offset:32
+; ALIGNED-NEXT:    flat_load_dwordx4 v[48:51], v[4:5] offset:64
+; ALIGNED-NEXT:    flat_load_dwordx4 v[4:7], v[24:25]
+; ALIGNED-NEXT:    flat_load_dwordx4 v[8:11], v[24:25] offset:32
+; ALIGNED-NEXT:    flat_load_dwordx4 v[12:15], v[24:25] offset:48
+; ALIGNED-NEXT:    flat_load_dwordx4 v[16:19], v[24:25] offset:64
+; ALIGNED-NEXT:    flat_load_dwordx4 v[112:115], v[24:25] offset:96
+; ALIGNED-NEXT:    flat_load_dwordx4 v[116:119], v[24:25] offset:112
+; ALIGNED-NEXT:    flat_load_dwordx4 v[40:43], v[24:25] offset:128
+; ALIGNED-NEXT:    flat_load_dwordx4 v[44:47], v[24:25] offset:144
+; ALIGNED-NEXT:    flat_load_dwordx4 v[55:58], v[24:25] offset:160
+; ALIGNED-NEXT:    flat_load_dwordx4 v[59:62], v[24:25] offset:176
+; ALIGNED-NEXT:    flat_load_dwordx4 v[68:71], v[24:25] offset:192
+; ALIGNED-NEXT:    v_add_co_u32 v24, vcc_lo, v0, s4
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v25, null, s5, v1, vcc_lo
+; ALIGNED-NEXT:    s_add_u32 s4, s4, 0xffffff00
+; ALIGNED-NEXT:    s_addc_u32 s5, s5, -1
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(15) lgkmcnt(15)
-; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:384
-; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:388
-; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:392
-; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:396
+; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:416
+; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:420
+; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:424
+; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:428
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:396
-; ALIGNED-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:392
-; ALIGNED-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:388
-; ALIGNED-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:384
-; ALIGNED-NEXT:    v_add_co_u32 v22, vcc_lo, v16, 6
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v23, null, 0, v17, vcc_lo
-; ALIGNED-NEXT:    v_add_co_u32 v20, vcc_lo, v16, 3
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v21, null, 0, v17, vcc_lo
-; ALIGNED-NEXT:    v_add_co_u32 v18, vcc_lo, 0xffffff00, v18
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v19, null, -1, v19, vcc_lo
-; ALIGNED-NEXT:    s_cmp_eq_u64 s[4:5], 0
+; ALIGNED-NEXT:    buffer_load_dword v102, off, s[0:3], s32 offset:428
+; ALIGNED-NEXT:    buffer_load_dword v103, off, s[0:3], s32 offset:424
+; ALIGNED-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:420
+; ALIGNED-NEXT:    buffer_load_dword v72, off, s[0:3], s32 offset:416
+; ALIGNED-NEXT:    v_add_co_u32 v22, vcc_lo, v24, 6
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v23, null, 0, v25, vcc_lo
+; ALIGNED-NEXT:    v_add_co_u32 v20, vcc_lo, v24, 3
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v21, null, 0, v25, vcc_lo
+; ALIGNED-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v47 offset:136
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v47 offset:137
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v102 offset:136
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:137
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v56 offset:132
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v56 offset:133
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v103 offset:132
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:133
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v57 offset:128
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v57 offset:129
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v58 offset:124
-; ALIGNED-NEXT:    flat_store_byte v[16:17], v58 offset:128
-; ALIGNED-NEXT:    s_waitcnt lgkmcnt(22)
-; ALIGNED-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:448
-; ALIGNED-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:452
-; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:456
-; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:460
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v63 offset:128
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v63 offset:129
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v72 offset:124
+; ALIGNED-NEXT:    flat_store_byte v[24:25], v72 offset:128
+; ALIGNED-NEXT:    s_waitcnt lgkmcnt(19)
+; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:480
+; ALIGNED-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:484
+; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:488
+; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:492
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v99, off, s[0:3], s32 offset:460
-; ALIGNED-NEXT:    buffer_load_dword v98, off, s[0:3], s32 offset:456
-; ALIGNED-NEXT:    buffer_load_dword v97, off, s[0:3], s32 offset:452
-; ALIGNED-NEXT:    buffer_load_dword v96, off, s[0:3], s32 offset:448
+; ALIGNED-NEXT:    buffer_load_dword v101, off, s[0:3], s32 offset:492
+; ALIGNED-NEXT:    buffer_load_dword v100, off, s[0:3], s32 offset:488
+; ALIGNED-NEXT:    buffer_load_dword v99, off, s[0:3], s32 offset:484
+; ALIGNED-NEXT:    buffer_load_dword v98, off, s[0:3], s32 offset:480
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v99 offset:72
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:73
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v101 offset:72
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:73
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v98 offset:68
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:69
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v100 offset:68
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:69
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v97 offset:64
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:65
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v96 offset:60
-; ALIGNED-NEXT:    flat_store_byte v[16:17], v96 offset:64
-; ALIGNED-NEXT:    s_waitcnt lgkmcnt(28)
-; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:544
-; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:548
-; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:552
-; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:556
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v99 offset:64
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:65
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v98 offset:60
+; ALIGNED-NEXT:    flat_store_byte v[24:25], v98 offset:64
+; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:576
+; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:580
+; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:584
+; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:588
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:552
-; ALIGNED-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:556
-; ALIGNED-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:544
-; ALIGNED-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:548
+; ALIGNED-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:584
+; ALIGNED-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:588
+; ALIGNED-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:576
+; ALIGNED-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:580
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v37 offset:36
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v37 offset:37
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v39 offset:36
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:37
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v32 offset:40
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v32 offset:41
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v34 offset:40
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v34 offset:41
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v31 offset:28
-; ALIGNED-NEXT:    flat_store_byte v[16:17], v31 offset:32
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v30 offset:32
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v30 offset:33
-; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:496
-; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:500
-; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:504
-; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:508
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v33 offset:28
+; ALIGNED-NEXT:    flat_store_byte v[24:25], v33 offset:32
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v32 offset:32
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v32 offset:33
+; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:528
+; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:532
+; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:536
+; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:540
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v65, off, s[0:3], s32 offset:508
-; ALIGNED-NEXT:    buffer_load_dword v64, off, s[0:3], s32 offset:504
-; ALIGNED-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:500
-; ALIGNED-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:496
+; ALIGNED-NEXT:    buffer_load_dword v67, off, s[0:3], s32 offset:540
+; ALIGNED-NEXT:    buffer_load_dword v66, off, s[0:3], s32 offset:536
+; ALIGNED-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:532
+; ALIGNED-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:528
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v65 offset:24
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v65 offset:25
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v67 offset:24
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v67 offset:25
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v64 offset:20
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v64 offset:21
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v66 offset:20
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v66 offset:21
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v51 offset:16
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v51 offset:17
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v50 offset:12
-; ALIGNED-NEXT:    flat_store_byte v[16:17], v50 offset:16
-; ALIGNED-NEXT:    s_waitcnt lgkmcnt(32)
-; ALIGNED-NEXT:    buffer_store_dword v80, off, s[0:3], s32 offset:512
-; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:516
-; ALIGNED-NEXT:    buffer_store_dword v82, off, s[0:3], s32 offset:520
-; ALIGNED-NEXT:    buffer_store_dword v83, off, s[0:3], s32 offset:524
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v53 offset:16
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v53 offset:17
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v52 offset:12
+; ALIGNED-NEXT:    flat_store_byte v[24:25], v52 offset:16
+; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:544
+; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:548
+; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:552
+; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:556
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v84, off, s[0:3], s32 offset:524
-; ALIGNED-NEXT:    buffer_load_dword v83, off, s[0:3], s32 offset:520
-; ALIGNED-NEXT:    buffer_load_dword v82, off, s[0:3], s32 offset:516
-; ALIGNED-NEXT:    buffer_load_dword v70, off, s[0:3], s32 offset:512
+; ALIGNED-NEXT:    buffer_load_dword v86, off, s[0:3], s32 offset:556
+; ALIGNED-NEXT:    buffer_load_dword v82, off, s[0:3], s32 offset:552
+; ALIGNED-NEXT:    buffer_load_dword v81, off, s[0:3], s32 offset:548
+; ALIGNED-NEXT:    buffer_load_dword v80, off, s[0:3], s32 offset:544
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v84 offset:8
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v84 offset:9
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v86 offset:8
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v86 offset:9
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v83 offset:4
-; ALIGNED-NEXT:    flat_store_byte v[16:17], v83 offset:8
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v82 offset:4
+; ALIGNED-NEXT:    flat_store_byte v[24:25], v82 offset:8
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v82
-; ALIGNED-NEXT:    flat_store_byte v[16:17], v82 offset:4
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[16:17], v70 offset:2
-; ALIGNED-NEXT:    flat_store_byte v[16:17], v70
-; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:336
-; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:340
-; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:344
-; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:348
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v81
+; ALIGNED-NEXT:    flat_store_byte v[24:25], v81 offset:4
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[24:25], v80 offset:2
+; ALIGNED-NEXT:    flat_store_byte v[24:25], v80
+; ALIGNED-NEXT:    s_waitcnt lgkmcnt(40)
+; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:368
+; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:372
+; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:376
+; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:380
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v87, off, s[0:3], s32 offset:348
-; ALIGNED-NEXT:    buffer_load_dword v85, off, s[0:3], s32 offset:344
-; ALIGNED-NEXT:    buffer_load_dword v71, off, s[0:3], s32 offset:340
-; ALIGNED-NEXT:    buffer_load_dword v66, off, s[0:3], s32 offset:336
+; ALIGNED-NEXT:    buffer_load_dword v97, off, s[0:3], s32 offset:380
+; ALIGNED-NEXT:    buffer_load_dword v87, off, s[0:3], s32 offset:376
+; ALIGNED-NEXT:    buffer_load_dword v83, off, s[0:3], s32 offset:372
+; ALIGNED-NEXT:    buffer_load_dword v68, off, s[0:3], s32 offset:368
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v87 offset:248
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v87 offset:249
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v97 offset:248
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:249
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v85 offset:244
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v85 offset:245
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v87 offset:244
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v87 offset:245
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v71 offset:240
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v71 offset:241
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v83 offset:240
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v83 offset:241
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v66 offset:236
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v66 offset:237
-; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:352
-; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:356
-; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:360
-; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:364
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v68 offset:236
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v68 offset:237
+; ALIGNED-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:384
+; ALIGNED-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:388
+; ALIGNED-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:392
+; ALIGNED-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v86, off, s[0:3], s32 offset:364
-; ALIGNED-NEXT:    buffer_load_dword v80, off, s[0:3], s32 offset:360
-; ALIGNED-NEXT:    buffer_load_dword v67, off, s[0:3], s32 offset:356
-; ALIGNED-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:352
+; ALIGNED-NEXT:    buffer_load_dword v96, off, s[0:3], s32 offset:396
+; ALIGNED-NEXT:    buffer_load_dword v84, off, s[0:3], s32 offset:392
+; ALIGNED-NEXT:    buffer_load_dword v69, off, s[0:3], s32 offset:388
+; ALIGNED-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:384
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v86 offset:232
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v86 offset:233
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v96 offset:232
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v96 offset:233
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v80 offset:228
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v80 offset:229
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v84 offset:228
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v84 offset:229
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v67 offset:224
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v67 offset:225
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v52 offset:220
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v52 offset:221
-; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:304
-; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:308
-; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:312
-; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:316
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v69 offset:224
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v69 offset:225
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v54 offset:220
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v54 offset:221
+; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:336
+; ALIGNED-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:340
+; ALIGNED-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:344
+; ALIGNED-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:348
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v81, off, s[0:3], s32 offset:316
-; ALIGNED-NEXT:    buffer_load_dword v68, off, s[0:3], s32 offset:312
-; ALIGNED-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:308
-; ALIGNED-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:304
+; ALIGNED-NEXT:    buffer_load_dword v85, off, s[0:3], s32 offset:348
+; ALIGNED-NEXT:    buffer_load_dword v70, off, s[0:3], s32 offset:344
+; ALIGNED-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:340
+; ALIGNED-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:336
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v81 offset:216
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v81 offset:217
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v85 offset:216
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v85 offset:217
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v68 offset:212
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v68 offset:213
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v70 offset:212
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v70 offset:213
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v53 offset:208
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v53 offset:209
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v38 offset:204
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v38 offset:205
-; ALIGNED-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:320
-; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:324
-; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:328
-; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:332
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v55 offset:208
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v55 offset:209
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v48 offset:204
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v48 offset:205
+; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:352
+; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:356
+; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:360
+; ALIGNED-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:364
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v69, off, s[0:3], s32 offset:332
-; ALIGNED-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:328
-; ALIGNED-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:324
-; ALIGNED-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:320
+; ALIGNED-NEXT:    buffer_load_dword v71, off, s[0:3], s32 offset:364
+; ALIGNED-NEXT:    buffer_load_dword v64, off, s[0:3], s32 offset:360
+; ALIGNED-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:356
+; ALIGNED-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:352
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v69 offset:200
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v69 offset:201
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v71 offset:200
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v71 offset:201
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v54 offset:196
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v54 offset:197
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v64 offset:196
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v64 offset:197
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v39 offset:192
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:193
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v33 offset:188
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v33 offset:189
-; ALIGNED-NEXT:    buffer_store_dword v116, off, s[0:3], s32 offset:400
-; ALIGNED-NEXT:    buffer_store_dword v117, off, s[0:3], s32 offset:404
-; ALIGNED-NEXT:    buffer_store_dword v118, off, s[0:3], s32 offset:408
-; ALIGNED-NEXT:    buffer_store_dword v119, off, s[0:3], s32 offset:412
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v49 offset:192
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v49 offset:193
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v35 offset:188
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v35 offset:189
+; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:432
+; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:436
+; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:440
+; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:444
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:412
-; ALIGNED-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:408
-; ALIGNED-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:404
-; ALIGNED-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:400
+; ALIGNED-NEXT:    buffer_load_dword v65, off, s[0:3], s32 offset:444
+; ALIGNED-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:440
+; ALIGNED-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:436
+; ALIGNED-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:432
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v55 offset:184
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v55 offset:185
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v65 offset:184
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v65 offset:185
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v48 offset:180
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v48 offset:181
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v50 offset:180
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v50 offset:181
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v34 offset:176
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v34 offset:177
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v27 offset:172
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v27 offset:173
-; ALIGNED-NEXT:    buffer_store_dword v112, off, s[0:3], s32 offset:416
-; ALIGNED-NEXT:    buffer_store_dword v113, off, s[0:3], s32 offset:420
-; ALIGNED-NEXT:    buffer_store_dword v114, off, s[0:3], s32 offset:424
-; ALIGNED-NEXT:    buffer_store_dword v115, off, s[0:3], s32 offset:428
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v36 offset:176
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v36 offset:177
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v29 offset:172
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v29 offset:173
+; ALIGNED-NEXT:    buffer_store_dword v116, off, s[0:3], s32 offset:448
+; ALIGNED-NEXT:    buffer_store_dword v117, off, s[0:3], s32 offset:452
+; ALIGNED-NEXT:    buffer_store_dword v118, off, s[0:3], s32 offset:456
+; ALIGNED-NEXT:    buffer_store_dword v119, off, s[0:3], s32 offset:460
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:428
-; ALIGNED-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:424
-; ALIGNED-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:420
-; ALIGNED-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:416
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v57
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 8, v57
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v58
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v58
+; ALIGNED-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:460
+; ALIGNED-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:456
+; ALIGNED-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:452
+; ALIGNED-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:448
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v49 offset:168
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v49 offset:169
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v51 offset:168
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v51 offset:169
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v35 offset:164
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v35 offset:165
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v37 offset:164
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v37 offset:165
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v28 offset:160
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v28 offset:161
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v25 offset:156
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v25 offset:157
-; ALIGNED-NEXT:    buffer_store_dword v100, off, s[0:3], s32 offset:368
-; ALIGNED-NEXT:    buffer_store_dword v101, off, s[0:3], s32 offset:372
-; ALIGNED-NEXT:    buffer_store_dword v102, off, s[0:3], s32 offset:376
-; ALIGNED-NEXT:    buffer_store_dword v103, off, s[0:3], s32 offset:380
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v30 offset:160
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v30 offset:161
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v27 offset:156
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v27 offset:157
+; ALIGNED-NEXT:    buffer_store_dword v112, off, s[0:3], s32 offset:400
+; ALIGNED-NEXT:    buffer_store_dword v113, off, s[0:3], s32 offset:404
+; ALIGNED-NEXT:    buffer_store_dword v114, off, s[0:3], s32 offset:408
+; ALIGNED-NEXT:    buffer_store_dword v115, off, s[0:3], s32 offset:412
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:380
-; ALIGNED-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:376
-; ALIGNED-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:372
-; ALIGNED-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:368
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v47
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v47
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v56
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v56
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:140
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:138
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:136
+; ALIGNED-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:412
+; ALIGNED-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:408
+; ALIGNED-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:404
+; ALIGNED-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:400
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v102
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 8, v102
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v103
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v103
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 8, v63
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:140
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v63
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:138
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v72
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v72
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:136
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:134
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:132
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:130
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v114 offset:128
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v114 offset:130
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:128
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v115 offset:126
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v36 offset:152
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v36 offset:153
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v38 offset:152
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v38 offset:153
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v29 offset:148
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v29 offset:149
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v31 offset:148
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v31 offset:149
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v26 offset:144
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v26 offset:145
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v24 offset:140
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v24 offset:141
-; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:464
-; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:468
-; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:472
-; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:476
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v28 offset:144
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v28 offset:145
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v26 offset:140
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v26 offset:141
+; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:496
+; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:500
+; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:504
+; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:508
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:476
-; ALIGNED-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:472
-; ALIGNED-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:468
-; ALIGNED-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:464
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v99
+; ALIGNED-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:508
+; ALIGNED-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:504
+; ALIGNED-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:500
+; ALIGNED-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:496
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v101
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v101
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v100
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 8, v100
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v99
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v99
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v98
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v98
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 8, v98
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v97
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 8, v97
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v96
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 8, v96
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v70
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 8, v70
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v84
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 8, v84
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v83
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v83
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v82
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 8, v82
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 8, v80
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v86
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 8, v86
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v80
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v15 offset:120
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v15 offset:121
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v19 offset:120
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v19 offset:121
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v14 offset:116
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v14 offset:117
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v18 offset:116
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v18 offset:117
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v13 offset:112
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v13 offset:113
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v12 offset:108
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v12 offset:109
-; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:480
-; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:484
-; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:488
-; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:492
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v17 offset:112
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v17 offset:113
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v16 offset:108
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v16 offset:109
+; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:512
+; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:516
+; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:520
+; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:524
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:492
-; ALIGNED-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:488
-; ALIGNED-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:484
-; ALIGNED-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:480
+; ALIGNED-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:524
+; ALIGNED-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:520
+; ALIGNED-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:516
+; ALIGNED-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:512
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v11 offset:104
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v11 offset:105
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v15 offset:104
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v15 offset:105
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v10 offset:100
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v10 offset:101
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v14 offset:100
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v14 offset:101
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v9 offset:96
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v9 offset:97
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v8 offset:92
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v8 offset:93
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:432
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:436
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:440
-; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:444
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v13 offset:96
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v13 offset:97
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v12 offset:92
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v12 offset:93
+; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:464
+; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:468
+; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:472
+; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:476
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:444
-; ALIGNED-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:440
-; ALIGNED-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:436
-; ALIGNED-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:432
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:76
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:74
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:72
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:70
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:68
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:66
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:64
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v96 offset:62
+; ALIGNED-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:476
+; ALIGNED-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:472
+; ALIGNED-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:468
+; ALIGNED-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:464
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:76
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:74
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:72
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:70
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:68
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:66
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:64
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:62
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v7 offset:88
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v7 offset:89
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v11 offset:88
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v11 offset:89
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v6 offset:84
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v6 offset:85
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v10 offset:84
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v10 offset:85
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v5 offset:80
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v5 offset:81
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v9 offset:80
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v9 offset:81
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v4 offset:76
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v4 offset:77
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:528
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:532
-; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:536
-; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:540
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v8 offset:76
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v8 offset:77
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:564
+; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:568
+; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:572
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:540
-; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:536
-; ALIGNED-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:532
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:528
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v30
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 8, v30
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 24, v37
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 8, v37
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 24, v32
+; ALIGNED-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:572
+; ALIGNED-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:568
+; ALIGNED-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:564
+; ALIGNED-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:560
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v32
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v39
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 8, v39
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v34
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 8, v34
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v33
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 8, v33
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 8, v32
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v31
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 8, v31
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v65
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:36
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v67
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v65
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v67
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v67
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v64
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v30 offset:34
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v52
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 8, v52
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 8, v64
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v51
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v51
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v50
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 8, v50
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v96 offset:40
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 24, v87
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 8, v87
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v37 offset:38
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 24, v85
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 8, v85
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:44
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 24, v71
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v71
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v32 offset:42
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 24, v66
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:36
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v54
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 8, v54
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v66
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 8, v66
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:32
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v86
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 8, v86
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v31 offset:30
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 24, v80
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 8, v80
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:28
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v81
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:228
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v7
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v53
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v53
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v52
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 8, v52
+; ALIGNED-NEXT:    flat_store_byte v[24:25], v114 offset:1
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v82
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 8, v82
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v81
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v81
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v7, 8, v7
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v65 offset:26
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 24, v68
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v67 offset:226
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 24, v6
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v97
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 8, v97
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:40
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v87
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 8, v87
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:38
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 24, v83
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v83
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:44
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v68
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 8, v68
-; ALIGNED-NEXT:    flat_store_byte v[16:17], v70 offset:1
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v6, 8, v6
-; ALIGNED-NEXT:    v_add_co_u32 v16, vcc_lo, 0xffffff00, v16
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:24
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v53
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v30 offset:224
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v5
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v52 offset:222
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v4
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v4, 8, v4
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v53
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v64 offset:22
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v38
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 8, v38
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:20
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v69
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v34 offset:42
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 24, v96
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 8, v96
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:32
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v84
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 8, v84
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v33 offset:30
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 24, v69
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v69
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v51 offset:18
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 24, v54
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 8, v54
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:16
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v39
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 8, v39
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v50 offset:14
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v33
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 8, v33
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:12
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v55
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v32 offset:34
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 24, v85
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 8, v85
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:28
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v70
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 8, v70
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v67 offset:26
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 24, v55
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v54 offset:222
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 24, v8
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v8, 8, v8
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v55
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v84 offset:10
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 24, v48
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:24
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v48
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 8, v48
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:8
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v34
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 8, v34
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v83 offset:6
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 24, v27
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 8, v27
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v114 offset:4
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v49
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v66 offset:22
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v71
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v71
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:20
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v64
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 8, v64
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v53 offset:18
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 24, v49
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v49
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v82 offset:2
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v35
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:16
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v35
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 8, v35
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v115
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v28
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 8, v28
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v25
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 8, v25
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v96 offset:252
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 24, v36
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v52 offset:14
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v65
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v65
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v115 offset:12
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v50
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 8, v50
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v86 offset:10
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 24, v36
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 8, v36
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v87 offset:250
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 24, v29
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v24 offset:8
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v29
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 8, v29
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v37 offset:248
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 24, v26
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v25 offset:6
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 24, v51
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v51
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v82 offset:4
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v37
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 8, v37
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v81 offset:2
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 24, v30
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 8, v30
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v80
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v27
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 8, v27
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v114 offset:252
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v38
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 8, v38
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:250
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 24, v31
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 8, v31
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:248
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v28
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 8, v28
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v87 offset:246
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 24, v26
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 8, v26
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v85 offset:246
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 24, v24
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 8, v24
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:244
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 24, v15
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:244
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 24, v19
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v19, 8, v19
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v83 offset:242
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 24, v18
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v18, 8, v18
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:240
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v17
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v17, 8, v17
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v68 offset:238
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v16
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v16, 8, v16
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v34 offset:236
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 24, v15
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v15, 8, v15
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v71 offset:242
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 24, v14
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v96 offset:234
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 24, v14
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v14, 8, v14
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v32 offset:240
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 24, v13
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:232
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v13
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v13, 8, v13
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v66 offset:238
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v12
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v84 offset:230
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 24, v12
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v12, 8, v12
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:236
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v11
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v33 offset:228
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 24, v11
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v11, 8, v11
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v86 offset:234
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 24, v10
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v69 offset:226
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 24, v10
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v10, 8, v10
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v31 offset:232
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 24, v9
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:224
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v9
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v9, 8, v9
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v80 offset:230
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v8
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v8, 8, v8
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v5, 8, v5
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:220
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v81 offset:218
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v65 offset:216
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v68 offset:214
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:212
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v53 offset:210
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v64 offset:208
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v38 offset:206
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:204
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v69 offset:202
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v51 offset:200
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v54 offset:198
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:196
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:194
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v50 offset:192
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v33 offset:190
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:188
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v55 offset:186
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v84 offset:184
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v48 offset:182
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:180
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v34 offset:178
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v83 offset:176
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v27 offset:174
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v114 offset:172
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v49 offset:170
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v32 offset:220
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v85 offset:218
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v102 offset:216
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v70 offset:214
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v67 offset:212
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v55 offset:210
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v103 offset:208
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v48 offset:206
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v66 offset:204
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v71 offset:202
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v112 offset:200
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v64 offset:198
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v53 offset:196
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v49 offset:194
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v113 offset:192
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v35 offset:190
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v52 offset:188
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v65 offset:186
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v115 offset:184
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v50 offset:182
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v86 offset:180
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v36 offset:178
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v24 offset:176
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v29 offset:174
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v25 offset:172
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v51 offset:170
 ; ALIGNED-NEXT:    flat_store_byte v[20:21], v82 offset:168
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v35 offset:166
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v115 offset:164
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v28 offset:162
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v70 offset:160
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v25 offset:158
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v96 offset:156
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v36 offset:154
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v87 offset:152
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v29 offset:150
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v37 offset:148
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v26 offset:146
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v85 offset:144
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v24 offset:142
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:124
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v15 offset:122
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v71 offset:120
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v14 offset:118
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v32 offset:116
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v13 offset:114
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v66 offset:112
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v12 offset:110
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:108
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v11 offset:106
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v86 offset:104
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v10 offset:102
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v31 offset:100
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v9 offset:98
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v80 offset:96
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v8 offset:94
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v17, null, -1, v17, vcc_lo
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:92
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v7 offset:90
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v67 offset:88
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v6 offset:86
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v30 offset:84
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v5 offset:82
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v52 offset:80
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v4 offset:78
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v37 offset:166
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v81 offset:164
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v30 offset:162
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v80 offset:160
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v27 offset:158
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v114 offset:156
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v38 offset:154
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v97 offset:152
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v31 offset:150
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v98 offset:148
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v28 offset:146
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v87 offset:144
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v26 offset:142
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v39 offset:124
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v19 offset:122
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v83 offset:120
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v18 offset:118
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v99 offset:116
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v17 offset:114
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v68 offset:112
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v16 offset:110
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v34 offset:108
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v15 offset:106
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v96 offset:104
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v14 offset:102
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v100 offset:100
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v13 offset:98
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v84 offset:96
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v12 offset:94
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v33 offset:92
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v11 offset:90
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v69 offset:88
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v10 offset:86
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v101 offset:84
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v9 offset:82
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v54 offset:80
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v8 offset:78
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v3 offset:56
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v3 offset:57
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v7 offset:56
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v7 offset:57
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v2 offset:52
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v2 offset:53
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v6 offset:52
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v6 offset:53
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v1 offset:48
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v1 offset:49
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v0 offset:44
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v0 offset:45
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v4, 24, v3
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v3, 8, v3
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v5, 24, v2
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v2, 8, v2
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v6, 24, v1
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v1, 8, v1
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v7, 24, v0
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v0, 8, v0
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v4 offset:60
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v3 offset:58
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v5 offset:56
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v2 offset:54
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v6 offset:52
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v1 offset:50
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v7 offset:48
-; ALIGNED-NEXT:    flat_store_byte v[20:21], v0 offset:46
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v5 offset:48
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v5 offset:49
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[22:23], v4 offset:44
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v4 offset:45
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v8, 24, v7
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v7, 8, v7
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v9, 24, v6
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v6, 8, v6
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v10, 24, v5
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v5, 8, v5
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v11, 24, v4
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v4, 8, v4
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v8 offset:60
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v7 offset:58
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v9 offset:56
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v6 offset:54
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v10 offset:52
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v5 offset:50
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v11 offset:48
+; ALIGNED-NEXT:    flat_store_byte v[20:21], v4 offset:46
 ; ALIGNED-NEXT:    s_cbranch_scc0 .LBB5_5
-; ALIGNED-NEXT:  .LBB5_6: ; %Flow16
-; ALIGNED-NEXT:    s_or_b32 exec_lo, exec_lo, s6
-; ALIGNED-NEXT:    s_clause 0xa ; 44-byte Folded Reload
-; ALIGNED-NEXT:    buffer_load_dword v58, off, s[0:3], s32
-; ALIGNED-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:4
-; ALIGNED-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:8
-; ALIGNED-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:12
-; ALIGNED-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:16
-; ALIGNED-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:20
-; ALIGNED-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:24
-; ALIGNED-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:28
-; ALIGNED-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:32
-; ALIGNED-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:36
-; ALIGNED-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:40
+; ALIGNED-NEXT:  .LBB5_6: ; %Flow6
+; ALIGNED-NEXT:    s_or_b32 exec_lo, exec_lo, s8
+; ALIGNED-NEXT:    s_clause 0x10 ; 68-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v72, off, s[0:3], s32
+; ALIGNED-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:4
+; ALIGNED-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:8
+; ALIGNED-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:12
+; ALIGNED-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:16
+; ALIGNED-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:20
+; ALIGNED-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:24
+; ALIGNED-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:28
+; ALIGNED-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:32
+; ALIGNED-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:36
+; ALIGNED-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:40
+; ALIGNED-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:44
+; ALIGNED-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:48
+; ALIGNED-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:52
+; ALIGNED-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:56
+; ALIGNED-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:60
+; ALIGNED-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:64
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; ALIGNED-NEXT:    s_setpc_b64 s[30:31]
 ;
@@ -6850,31 +6854,27 @@ define void @memmove_p0_p0_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(0
 ; UNROLL3-NEXT:    s_xor_b32 s6, exec_lo, s4
 ; UNROLL3-NEXT:    s_cbranch_execz .LBB5_4
 ; UNROLL3-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader
-; UNROLL3-NEXT:    v_mov_b32_e32 v5, v3
-; UNROLL3-NEXT:    v_mov_b32_e32 v7, v1
-; UNROLL3-NEXT:    v_mov_b32_e32 v4, v2
-; UNROLL3-NEXT:    v_mov_b32_e32 v6, v0
-; UNROLL3-NEXT:    s_mov_b64 s[4:5], 0x7e0
+; UNROLL3-NEXT:    s_mov_b64 s[4:5], 0
 ; UNROLL3-NEXT:    .p2align 6
 ; UNROLL3-NEXT:  .LBB5_2: ; %memmove_fwd_loop
 ; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; UNROLL3-NEXT:    v_add_co_u32 v12, vcc_lo, v2, s4
+; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v13, null, s5, v3, vcc_lo
+; UNROLL3-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s4
+; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo
 ; UNROLL3-NEXT:    s_clause 0x2
-; UNROLL3-NEXT:    flat_load_dwordx4 v[8:11], v[4:5]
-; UNROLL3-NEXT:    flat_load_dwordx4 v[12:15], v[4:5] offset:16
-; UNROLL3-NEXT:    flat_load_dwordx4 v[16:19], v[4:5] offset:32
-; UNROLL3-NEXT:    v_add_co_u32 v4, vcc_lo, v4, 48
-; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; UNROLL3-NEXT:    s_add_u32 s4, s4, 0xffffffd0
-; UNROLL3-NEXT:    s_addc_u32 s5, s5, -1
+; UNROLL3-NEXT:    flat_load_dwordx4 v[4:7], v[12:13]
+; UNROLL3-NEXT:    flat_load_dwordx4 v[8:11], v[12:13] offset:16
+; UNROLL3-NEXT:    flat_load_dwordx4 v[12:15], v[12:13] offset:32
+; UNROLL3-NEXT:    s_add_u32 s4, s4, 48
+; UNROLL3-NEXT:    s_addc_u32 s5, s5, 0
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(2)
-; UNROLL3-NEXT:    flat_store_dwordx4 v[6:7], v[8:11]
+; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[4:7]
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(2)
-; UNROLL3-NEXT:    flat_store_dwordx4 v[6:7], v[12:15] offset:16
+; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[8:11] offset:16
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(2)
-; UNROLL3-NEXT:    flat_store_dwordx4 v[6:7], v[16:19] offset:32
-; UNROLL3-NEXT:    v_add_co_u32 v6, vcc_lo, v6, 48
-; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; UNROLL3-NEXT:    s_cmp_lg_u64 s[4:5], 0
+; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[12:15] offset:32
+; UNROLL3-NEXT:    s_cmp_lg_u64 s[4:5], 0x7e0
 ; UNROLL3-NEXT:    s_cbranch_scc1 .LBB5_2
 ; UNROLL3-NEXT:  ; %bb.3: ; %memmove_fwd_residual
 ; UNROLL3-NEXT:    flat_load_dwordx4 v[4:7], v[2:3] offset:2016
@@ -6885,45 +6885,42 @@ define void @memmove_p0_p0_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(0
 ; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[2:5] offset:2032
 ; UNROLL3-NEXT:    ; implicit-def: $vgpr2_vgpr3
 ; UNROLL3-NEXT:    ; implicit-def: $vgpr0_vgpr1
-; UNROLL3-NEXT:  .LBB5_4: ; %Flow13
-; UNROLL3-NEXT:    s_andn2_saveexec_b32 s6, s6
+; UNROLL3-NEXT:  .LBB5_4: ; %Flow3
+; UNROLL3-NEXT:    s_andn2_saveexec_b32 s8, s6
 ; UNROLL3-NEXT:    s_cbranch_execz .LBB5_7
 ; UNROLL3-NEXT:  ; %bb.5: ; %memmove_bwd_residual
 ; UNROLL3-NEXT:    flat_load_dwordx4 v[4:7], v[2:3] offset:2032
-; UNROLL3-NEXT:    s_movk_i32 s4, 0xf820
-; UNROLL3-NEXT:    s_mov_b32 s5, -1
+; UNROLL3-NEXT:    s_mov_b64 s[4:5], 0x7b0
+; UNROLL3-NEXT:    s_movk_i32 s6, 0xffd0
+; UNROLL3-NEXT:    s_mov_b32 s7, -1
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[4:7] offset:2032
-; UNROLL3-NEXT:    flat_load_dwordx4 v[6:9], v[2:3] offset:2016
-; UNROLL3-NEXT:    v_add_co_u32 v4, vcc_lo, 0x7b0, v0
-; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v5, null, 0, v1, vcc_lo
-; UNROLL3-NEXT:    v_add_co_u32 v2, vcc_lo, 0x7b0, v2
-; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
+; UNROLL3-NEXT:    flat_load_dwordx4 v[4:7], v[2:3] offset:2016
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[6:9] offset:2016
+; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[4:7] offset:2016
 ; UNROLL3-NEXT:    .p2align 6
 ; UNROLL3-NEXT:  .LBB5_6: ; %memmove_bwd_loop
 ; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; UNROLL3-NEXT:    v_add_co_u32 v12, vcc_lo, v2, s4
+; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v13, null, s5, v3, vcc_lo
+; UNROLL3-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s4
+; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo
 ; UNROLL3-NEXT:    s_clause 0x2
-; UNROLL3-NEXT:    flat_load_dwordx4 v[6:9], v[2:3]
-; UNROLL3-NEXT:    flat_load_dwordx4 v[10:13], v[2:3] offset:16
-; UNROLL3-NEXT:    flat_load_dwordx4 v[14:17], v[2:3] offset:32
-; UNROLL3-NEXT:    v_add_co_u32 v2, vcc_lo, 0xffffffd0, v2
-; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v3, null, -1, v3, vcc_lo
-; UNROLL3-NEXT:    s_add_u32 s4, s4, 48
-; UNROLL3-NEXT:    s_addc_u32 s5, s5, 0
+; UNROLL3-NEXT:    flat_load_dwordx4 v[4:7], v[12:13]
+; UNROLL3-NEXT:    flat_load_dwordx4 v[8:11], v[12:13] offset:16
+; UNROLL3-NEXT:    flat_load_dwordx4 v[12:15], v[12:13] offset:32
+; UNROLL3-NEXT:    s_add_u32 s4, s4, 0xffffffd0
+; UNROLL3-NEXT:    s_addc_u32 s5, s5, -1
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(2)
-; UNROLL3-NEXT:    flat_store_dwordx4 v[4:5], v[6:9]
+; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[4:7]
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(2)
-; UNROLL3-NEXT:    flat_store_dwordx4 v[4:5], v[10:13] offset:16
+; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[8:11] offset:16
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(2)
-; UNROLL3-NEXT:    flat_store_dwordx4 v[4:5], v[14:17] offset:32
-; UNROLL3-NEXT:    v_add_co_u32 v4, vcc_lo, 0xffffffd0, v4
-; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v5, null, -1, v5, vcc_lo
-; UNROLL3-NEXT:    s_cmp_eq_u64 s[4:5], 0
+; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[12:15] offset:32
+; UNROLL3-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]
 ; UNROLL3-NEXT:    s_cbranch_scc0 .LBB5_6
-; UNROLL3-NEXT:  .LBB5_7: ; %Flow14
-; UNROLL3-NEXT:    s_or_b32 exec_lo, exec_lo, s6
+; UNROLL3-NEXT:  .LBB5_7: ; %Flow4
+; UNROLL3-NEXT:    s_or_b32 exec_lo, exec_lo, s8
 ; UNROLL3-NEXT:    s_waitcnt lgkmcnt(0)
 ; UNROLL3-NEXT:    s_setpc_b64 s[30:31]
 entry:
@@ -6940,759 +6937,756 @@ define void @memmove_p1_p1_sz2048(ptr addrspace(1) align 1 %dst, ptr addrspace(1
 ; CHECK-NEXT:    s_xor_b32 s6, exec_lo, s4
 ; CHECK-NEXT:    s_cbranch_execz .LBB6_3
 ; CHECK-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader
-; CHECK-NEXT:    s_mov_b64 s[4:5], 0x800
+; CHECK-NEXT:    s_mov_b64 s[4:5], 0
 ; CHECK-NEXT:  .LBB6_2: ; %memmove_fwd_loop
 ; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    v_add_co_u32 v96, vcc_lo, v2, s4
+; CHECK-NEXT:    v_add_co_ci_u32_e64 v97, null, s5, v3, vcc_lo
+; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s4
+; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo
 ; CHECK-NEXT:    s_clause 0xf
-; CHECK-NEXT:    global_load_dwordx4 v[4:7], v[2:3], off offset:224
-; CHECK-NEXT:    global_load_dwordx4 v[8:11], v[2:3], off offset:240
-; CHECK-NEXT:    global_load_dwordx4 v[12:15], v[2:3], off offset:192
-; CHECK-NEXT:    global_load_dwordx4 v[16:19], v[2:3], off offset:208
-; CHECK-NEXT:    global_load_dwordx4 v[20:23], v[2:3], off offset:160
-; CHECK-NEXT:    global_load_dwordx4 v[24:27], v[2:3], off offset:176
-; CHECK-NEXT:    global_load_dwordx4 v[28:31], v[2:3], off offset:128
-; CHECK-NEXT:    global_load_dwordx4 v[32:35], v[2:3], off offset:144
-; CHECK-NEXT:    global_load_dwordx4 v[36:39], v[2:3], off offset:96
-; CHECK-NEXT:    global_load_dwordx4 v[48:51], v[2:3], off offset:112
-; CHECK-NEXT:    global_load_dwordx4 v[52:55], v[2:3], off offset:64
-; CHECK-NEXT:    global_load_dwordx4 v[64:67], v[2:3], off offset:80
-; CHECK-NEXT:    global_load_dwordx4 v[68:71], v[2:3], off offset:32
-; CHECK-NEXT:    global_load_dwordx4 v[80:83], v[2:3], off offset:48
-; CHECK-NEXT:    global_load_dwordx4 v[84:87], v[2:3], off
-; CHECK-NEXT:    global_load_dwordx4 v[96:99], v[2:3], off offset:16
-; CHECK-NEXT:    v_add_co_u32 v2, vcc_lo, 0x100, v2
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; CHECK-NEXT:    s_add_u32 s4, s4, 0xffffff00
-; CHECK-NEXT:    s_addc_u32 s5, s5, -1
+; CHECK-NEXT:    global_load_dwordx4 v[4:7], v[96:97], off offset:224
+; CHECK-NEXT:    global_load_dwordx4 v[8:11], v[96:97], off offset:240
+; CHECK-NEXT:    global_load_dwordx4 v[12:15], v[96:97], off offset:192
+; CHECK-NEXT:    global_load_dwordx4 v[16:19], v[96:97], off offset:208
+; CHECK-NEXT:    global_load_dwordx4 v[20:23], v[96:97], off offset:160
+; CHECK-NEXT:    global_load_dwordx4 v[24:27], v[96:97], off offset:176
+; CHECK-NEXT:    global_load_dwordx4 v[28:31], v[96:97], off offset:128
+; CHECK-NEXT:    global_load_dwordx4 v[32:35], v[96:97], off offset:144
+; CHECK-NEXT:    global_load_dwordx4 v[36:39], v[96:97], off offset:96
+; CHECK-NEXT:    global_load_dwordx4 v[48:51], v[96:97], off offset:112
+; CHECK-NEXT:    global_load_dwordx4 v[52:55], v[96:97], off offset:64
+; CHECK-NEXT:    global_load_dwordx4 v[64:67], v[96:97], off offset:80
+; CHECK-NEXT:    global_load_dwordx4 v[68:71], v[96:97], off offset:32
+; CHECK-NEXT:    global_load_dwordx4 v[80:83], v[96:97], off offset:48
+; CHECK-NEXT:    global_load_dwordx4 v[84:87], v[96:97], off
+; CHECK-NEXT:    global_load_dwordx4 v[96:99], v[96:97], off offset:16
+; CHECK-NEXT:    s_add_u32 s4, s4, 0x100
+; CHECK-NEXT:    s_addc_u32 s5, s5, 0
 ; CHECK-NEXT:    s_waitcnt vmcnt(15)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[4:7], off offset:224
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[4:7], off offset:224
 ; CHECK-NEXT:    s_waitcnt vmcnt(14)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[8:11], off offset:240
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[8:11], off offset:240
 ; CHECK-NEXT:    s_waitcnt vmcnt(13)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[12:15], off offset:192
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[12:15], off offset:192
 ; CHECK-NEXT:    s_waitcnt vmcnt(12)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[16:19], off offset:208
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[16:19], off offset:208
 ; CHECK-NEXT:    s_waitcnt vmcnt(11)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[20:23], off offset:160
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[20:23], off offset:160
 ; CHECK-NEXT:    s_waitcnt vmcnt(10)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[24:27], off offset:176
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[24:27], off offset:176
 ; CHECK-NEXT:    s_waitcnt vmcnt(9)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[28:31], off offset:128
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[28:31], off offset:128
 ; CHECK-NEXT:    s_waitcnt vmcnt(8)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[32:35], off offset:144
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[32:35], off offset:144
 ; CHECK-NEXT:    s_waitcnt vmcnt(7)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[36:39], off offset:96
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[36:39], off offset:96
 ; CHECK-NEXT:    s_waitcnt vmcnt(6)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[48:51], off offset:112
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[48:51], off offset:112
 ; CHECK-NEXT:    s_waitcnt vmcnt(5)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[52:55], off offset:64
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[52:55], off offset:64
 ; CHECK-NEXT:    s_waitcnt vmcnt(4)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[64:67], off offset:80
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[64:67], off offset:80
 ; CHECK-NEXT:    s_waitcnt vmcnt(3)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[68:71], off offset:32
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[68:71], off offset:32
 ; CHECK-NEXT:    s_waitcnt vmcnt(2)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[80:83], off offset:48
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[80:83], off offset:48
 ; CHECK-NEXT:    s_waitcnt vmcnt(1)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[84:87], off
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[84:87], off
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[96:99], off offset:16
-; CHECK-NEXT:    v_add_co_u32 v0, vcc_lo, 0x100, v0
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; CHECK-NEXT:    s_cmp_lg_u64 s[4:5], 0
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[96:99], off offset:16
+; CHECK-NEXT:    s_cmp_lg_u64 s[4:5], 0x800
 ; CHECK-NEXT:    s_cbranch_scc1 .LBB6_2
-; CHECK-NEXT:  .LBB6_3: ; %Flow17
-; CHECK-NEXT:    s_andn2_saveexec_b32 s6, s6
+; CHECK-NEXT:  .LBB6_3: ; %Flow7
+; CHECK-NEXT:    s_andn2_saveexec_b32 s8, s6
 ; CHECK-NEXT:    s_cbranch_execz .LBB6_6
 ; CHECK-NEXT:  ; %bb.4: ; %memmove_bwd_loop.preheader
-; CHECK-NEXT:    v_add_co_u32 v0, vcc_lo, 0x700, v0
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; CHECK-NEXT:    v_add_co_u32 v2, vcc_lo, 0x700, v2
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; CHECK-NEXT:    s_movk_i32 s4, 0xf800
-; CHECK-NEXT:    s_mov_b32 s5, -1
+; CHECK-NEXT:    s_mov_b64 s[4:5], 0x700
+; CHECK-NEXT:    s_movk_i32 s6, 0xff00
+; CHECK-NEXT:    s_mov_b32 s7, -1
 ; CHECK-NEXT:  .LBB6_5: ; %memmove_bwd_loop
 ; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    v_add_co_u32 v96, vcc_lo, v2, s4
+; CHECK-NEXT:    v_add_co_ci_u32_e64 v97, null, s5, v3, vcc_lo
+; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s4
+; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo
 ; CHECK-NEXT:    s_clause 0xf
-; CHECK-NEXT:    global_load_dwordx4 v[4:7], v[2:3], off offset:224
-; CHECK-NEXT:    global_load_dwordx4 v[8:11], v[2:3], off offset:240
-; CHECK-NEXT:    global_load_dwordx4 v[12:15], v[2:3], off offset:192
-; CHECK-NEXT:    global_load_dwordx4 v[16:19], v[2:3], off offset:208
-; CHECK-NEXT:    global_load_dwordx4 v[20:23], v[2:3], off offset:160
-; CHECK-NEXT:    global_load_dwordx4 v[24:27], v[2:3], off offset:176
-; CHECK-NEXT:    global_load_dwordx4 v[28:31], v[2:3], off offset:128
-; CHECK-NEXT:    global_load_dwordx4 v[32:35], v[2:3], off offset:144
-; CHECK-NEXT:    global_load_dwordx4 v[36:39], v[2:3], off offset:96
-; CHECK-NEXT:    global_load_dwordx4 v[48:51], v[2:3], off offset:112
-; CHECK-NEXT:    global_load_dwordx4 v[52:55], v[2:3], off offset:64
-; CHECK-NEXT:    global_load_dwordx4 v[64:67], v[2:3], off offset:80
-; CHECK-NEXT:    global_load_dwordx4 v[68:71], v[2:3], off offset:32
-; CHECK-NEXT:    global_load_dwordx4 v[80:83], v[2:3], off offset:48
-; CHECK-NEXT:    global_load_dwordx4 v[84:87], v[2:3], off
-; CHECK-NEXT:    global_load_dwordx4 v[96:99], v[2:3], off offset:16
-; CHECK-NEXT:    v_add_co_u32 v2, vcc_lo, 0xffffff00, v2
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v3, null, -1, v3, vcc_lo
-; CHECK-NEXT:    s_add_u32 s4, s4, 0x100
-; CHECK-NEXT:    s_addc_u32 s5, s5, 0
+; CHECK-NEXT:    global_load_dwordx4 v[4:7], v[96:97], off offset:224
+; CHECK-NEXT:    global_load_dwordx4 v[8:11], v[96:97], off offset:240
+; CHECK-NEXT:    global_load_dwordx4 v[12:15], v[96:97], off offset:192
+; CHECK-NEXT:    global_load_dwordx4 v[16:19], v[96:97], off offset:208
+; CHECK-NEXT:    global_load_dwordx4 v[20:23], v[96:97], off offset:160
+; CHECK-NEXT:    global_load_dwordx4 v[24:27], v[96:97], off offset:176
+; CHECK-NEXT:    global_load_dwordx4 v[28:31], v[96:97], off offset:128
+; CHECK-NEXT:    global_load_dwordx4 v[32:35], v[96:97], off offset:144
+; CHECK-NEXT:    global_load_dwordx4 v[36:39], v[96:97], off offset:96
+; CHECK-NEXT:    global_load_dwordx4 v[48:51], v[96:97], off offset:112
+; CHECK-NEXT:    global_load_dwordx4 v[52:55], v[96:97], off offset:64
+; CHECK-NEXT:    global_load_dwordx4 v[64:67], v[96:97], off offset:80
+; CHECK-NEXT:    global_load_dwordx4 v[68:71], v[96:97], off offset:32
+; CHECK-NEXT:    global_load_dwordx4 v[80:83], v[96:97], off offset:48
+; CHECK-NEXT:    global_load_dwordx4 v[84:87], v[96:97], off
+; CHECK-NEXT:    global_load_dwordx4 v[96:99], v[96:97], off offset:16
+; CHECK-NEXT:    s_add_u32 s4, s4, 0xffffff00
+; CHECK-NEXT:    s_addc_u32 s5, s5, -1
 ; CHECK-NEXT:    s_waitcnt vmcnt(15)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[4:7], off offset:224
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[4:7], off offset:224
 ; CHECK-NEXT:    s_waitcnt vmcnt(14)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[8:11], off offset:240
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[8:11], off offset:240
 ; CHECK-NEXT:    s_waitcnt vmcnt(13)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[12:15], off offset:192
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[12:15], off offset:192
 ; CHECK-NEXT:    s_waitcnt vmcnt(12)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[16:19], off offset:208
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[16:19], off offset:208
 ; CHECK-NEXT:    s_waitcnt vmcnt(11)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[20:23], off offset:160
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[20:23], off offset:160
 ; CHECK-NEXT:    s_waitcnt vmcnt(10)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[24:27], off offset:176
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[24:27], off offset:176
 ; CHECK-NEXT:    s_waitcnt vmcnt(9)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[28:31], off offset:128
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[28:31], off offset:128
 ; CHECK-NEXT:    s_waitcnt vmcnt(8)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[32:35], off offset:144
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[32:35], off offset:144
 ; CHECK-NEXT:    s_waitcnt vmcnt(7)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[36:39], off offset:96
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[36:39], off offset:96
 ; CHECK-NEXT:    s_waitcnt vmcnt(6)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[48:51], off offset:112
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[48:51], off offset:112
 ; CHECK-NEXT:    s_waitcnt vmcnt(5)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[52:55], off offset:64
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[52:55], off offset:64
 ; CHECK-NEXT:    s_waitcnt vmcnt(4)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[64:67], off offset:80
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[64:67], off offset:80
 ; CHECK-NEXT:    s_waitcnt vmcnt(3)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[68:71], off offset:32
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[68:71], off offset:32
 ; CHECK-NEXT:    s_waitcnt vmcnt(2)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[80:83], off offset:48
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[80:83], off offset:48
 ; CHECK-NEXT:    s_waitcnt vmcnt(1)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[84:87], off
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[84:87], off
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
-; CHECK-NEXT:    global_store_dwordx4 v[0:1], v[96:99], off offset:16
-; CHECK-NEXT:    v_add_co_u32 v0, vcc_lo, 0xffffff00, v0
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo
-; CHECK-NEXT:    s_cmp_eq_u64 s[4:5], 0
+; CHECK-NEXT:    global_store_dwordx4 v[100:101], v[96:99], off offset:16
+; CHECK-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]
 ; CHECK-NEXT:    s_cbranch_scc0 .LBB6_5
-; CHECK-NEXT:  .LBB6_6: ; %Flow18
-; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s6
+; CHECK-NEXT:  .LBB6_6: ; %Flow8
+; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s8
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; ALIGNED-LABEL: memmove_p1_p1_sz2048:
 ; ALIGNED:       ; %bb.0: ; %entry
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v57, off, s[0:3], s32 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v47, off, s[0:3], s32 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_mov_b32 s4, exec_lo
 ; ALIGNED-NEXT:    v_cmpx_ge_u64_e32 v[2:3], v[0:1]
 ; ALIGNED-NEXT:    s_xor_b32 s6, exec_lo, s4
 ; ALIGNED-NEXT:    s_cbranch_execz .LBB6_3
 ; ALIGNED-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader
-; ALIGNED-NEXT:    s_mov_b64 s[4:5], 0x800
+; ALIGNED-NEXT:    s_mov_b64 s[4:5], 0
 ; ALIGNED-NEXT:  .LBB6_2: ; %memmove_fwd_loop
 ; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=1
+; ALIGNED-NEXT:    v_add_co_u32 v20, vcc_lo, v2, s4
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v21, null, s5, v3, vcc_lo
 ; ALIGNED-NEXT:    s_clause 0xf
-; ALIGNED-NEXT:    global_load_dwordx4 v[16:19], v[2:3], off offset:240
-; ALIGNED-NEXT:    global_load_dwordx4 v[20:23], v[2:3], off offset:224
-; ALIGNED-NEXT:    global_load_dwordx4 v[4:7], v[2:3], off
-; ALIGNED-NEXT:    global_load_dwordx4 v[8:11], v[2:3], off offset:16
-; ALIGNED-NEXT:    global_load_dwordx4 v[12:15], v[2:3], off offset:32
-; ALIGNED-NEXT:    global_load_dwordx4 v[96:99], v[2:3], off offset:48
-; ALIGNED-NEXT:    global_load_dwordx4 v[84:87], v[2:3], off offset:64
-; ALIGNED-NEXT:    global_load_dwordx4 v[80:83], v[2:3], off offset:80
-; ALIGNED-NEXT:    global_load_dwordx4 v[68:71], v[2:3], off offset:96
-; ALIGNED-NEXT:    global_load_dwordx4 v[64:67], v[2:3], off offset:112
-; ALIGNED-NEXT:    global_load_dwordx4 v[52:55], v[2:3], off offset:128
-; ALIGNED-NEXT:    global_load_dwordx4 v[48:51], v[2:3], off offset:144
-; ALIGNED-NEXT:    global_load_dwordx4 v[36:39], v[2:3], off offset:160
-; ALIGNED-NEXT:    global_load_dwordx4 v[32:35], v[2:3], off offset:176
-; ALIGNED-NEXT:    global_load_dwordx4 v[28:31], v[2:3], off offset:192
-; ALIGNED-NEXT:    global_load_dwordx4 v[24:27], v[2:3], off offset:208
-; ALIGNED-NEXT:    v_add_co_u32 v2, vcc_lo, 0x100, v2
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; ALIGNED-NEXT:    s_add_u32 s4, s4, 0xffffff00
-; ALIGNED-NEXT:    s_addc_u32 s5, s5, -1
+; ALIGNED-NEXT:    global_load_dwordx4 v[16:19], v[20:21], off offset:240
+; ALIGNED-NEXT:    global_load_dwordx4 v[22:25], v[20:21], off offset:224
+; ALIGNED-NEXT:    global_load_dwordx4 v[4:7], v[20:21], off
+; ALIGNED-NEXT:    global_load_dwordx4 v[8:11], v[20:21], off offset:16
+; ALIGNED-NEXT:    global_load_dwordx4 v[12:15], v[20:21], off offset:32
+; ALIGNED-NEXT:    global_load_dwordx4 v[98:101], v[20:21], off offset:48
+; ALIGNED-NEXT:    global_load_dwordx4 v[112:115], v[20:21], off offset:64
+; ALIGNED-NEXT:    global_load_dwordx4 v[82:85], v[20:21], off offset:80
+; ALIGNED-NEXT:    global_load_dwordx4 v[116:119], v[20:21], off offset:96
+; ALIGNED-NEXT:    global_load_dwordx4 v[66:69], v[20:21], off offset:112
+; ALIGNED-NEXT:    global_load_dwordx4 v[40:43], v[20:21], off offset:128
+; ALIGNED-NEXT:    global_load_dwordx4 v[50:53], v[20:21], off offset:144
+; ALIGNED-NEXT:    global_load_dwordx4 v[44:47], v[20:21], off offset:160
+; ALIGNED-NEXT:    global_load_dwordx4 v[34:37], v[20:21], off offset:176
+; ALIGNED-NEXT:    global_load_dwordx4 v[30:33], v[20:21], off offset:192
+; ALIGNED-NEXT:    global_load_dwordx4 v[26:29], v[20:21], off offset:208
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(15)
-; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:80
-; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:84
-; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:88
-; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:92
+; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:64
+; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:68
+; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:72
+; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:76
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:92
-; ALIGNED-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:88
-; ALIGNED-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:84
-; ALIGNED-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:80
-; ALIGNED-NEXT:    s_cmp_lg_u64 s[4:5], 0
+; ALIGNED-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:76
+; ALIGNED-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:72
+; ALIGNED-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:68
+; ALIGNED-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:64
+; ALIGNED-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s4
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo
+; ALIGNED-NEXT:    s_add_u32 s4, s4, 0x100
+; ALIGNED-NEXT:    s_addc_u32 s5, s5, 0
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v19, off offset:254
-; ALIGNED-NEXT:    global_store_byte v[0:1], v19, off offset:252
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v21, off offset:254
+; ALIGNED-NEXT:    global_store_byte v[16:17], v21, off offset:252
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v18, off offset:250
-; ALIGNED-NEXT:    global_store_byte v[0:1], v18, off offset:248
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v20, off offset:250
+; ALIGNED-NEXT:    global_store_byte v[16:17], v20, off offset:248
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v17, off offset:246
-; ALIGNED-NEXT:    global_store_byte v[0:1], v17, off offset:244
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v16, off offset:242
-; ALIGNED-NEXT:    global_store_byte v[0:1], v16, off offset:240
-; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:96
-; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:100
-; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:104
-; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:108
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v19, off offset:246
+; ALIGNED-NEXT:    global_store_byte v[16:17], v19, off offset:244
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v18, off offset:242
+; ALIGNED-NEXT:    global_store_byte v[16:17], v18, off offset:240
+; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:80
+; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:84
+; ALIGNED-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:88
+; ALIGNED-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:92
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:108
-; ALIGNED-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:104
-; ALIGNED-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:100
-; ALIGNED-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:96
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v19
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v19, 8, v19
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v18
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v18, 8, v18
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v17
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v17, 8, v17
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v16
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v16, 8, v16
+; ALIGNED-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:92
+; ALIGNED-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:88
+; ALIGNED-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:84
+; ALIGNED-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:80
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v21
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v21, 8, v21
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v20
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v20, 8, v20
+; ALIGNED-NEXT:    s_cmp_lg_u64 s[4:5], 0x800
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v23, off offset:238
-; ALIGNED-NEXT:    global_store_byte v[0:1], v23, off offset:236
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v25, off offset:238
+; ALIGNED-NEXT:    global_store_byte v[16:17], v25, off offset:236
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v22, off offset:234
-; ALIGNED-NEXT:    global_store_byte v[0:1], v22, off offset:232
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v24, off offset:234
+; ALIGNED-NEXT:    global_store_byte v[16:17], v24, off offset:232
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v21, off offset:230
-; ALIGNED-NEXT:    global_store_byte v[0:1], v21, off offset:228
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v20, off offset:226
-; ALIGNED-NEXT:    global_store_byte v[0:1], v20, off offset:224
-; ALIGNED-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:48
-; ALIGNED-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:52
-; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:56
-; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:60
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v23, off offset:230
+; ALIGNED-NEXT:    global_store_byte v[16:17], v23, off offset:228
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v22, off offset:226
+; ALIGNED-NEXT:    global_store_byte v[16:17], v22, off offset:224
+; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:32
+; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:36
+; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:40
+; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:44
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:60
-; ALIGNED-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:56
-; ALIGNED-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:52
-; ALIGNED-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:48
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v23
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v23, 8, v23
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v22
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v22, 8, v22
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v21
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v21, 8, v21
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v20
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v20, 8, v20
+; ALIGNED-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:44
+; ALIGNED-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:40
+; ALIGNED-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:36
+; ALIGNED-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:32
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v27, off offset:222
-; ALIGNED-NEXT:    global_store_byte v[0:1], v27, off offset:220
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v29, off offset:222
+; ALIGNED-NEXT:    global_store_byte v[16:17], v29, off offset:220
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v26, off offset:218
-; ALIGNED-NEXT:    global_store_byte v[0:1], v26, off offset:216
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v28, off offset:218
+; ALIGNED-NEXT:    global_store_byte v[16:17], v28, off offset:216
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v25, off offset:214
-; ALIGNED-NEXT:    global_store_byte v[0:1], v25, off offset:212
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v24, off offset:210
-; ALIGNED-NEXT:    global_store_byte v[0:1], v24, off offset:208
-; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:64
-; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:68
-; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:72
-; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:76
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v27, off offset:214
+; ALIGNED-NEXT:    global_store_byte v[16:17], v27, off offset:212
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v26, off offset:210
+; ALIGNED-NEXT:    global_store_byte v[16:17], v26, off offset:208
+; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:48
+; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:52
+; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:56
+; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:60
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:76
-; ALIGNED-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:72
-; ALIGNED-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:68
-; ALIGNED-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:64
+; ALIGNED-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:60
+; ALIGNED-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:56
+; ALIGNED-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:52
+; ALIGNED-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:48
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v31, off offset:206
-; ALIGNED-NEXT:    global_store_byte v[0:1], v31, off offset:204
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v33, off offset:206
+; ALIGNED-NEXT:    global_store_byte v[16:17], v33, off offset:204
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v30, off offset:202
-; ALIGNED-NEXT:    global_store_byte v[0:1], v30, off offset:200
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v32, off offset:202
+; ALIGNED-NEXT:    global_store_byte v[16:17], v32, off offset:200
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v29, off offset:198
-; ALIGNED-NEXT:    global_store_byte v[0:1], v29, off offset:196
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v28, off offset:194
-; ALIGNED-NEXT:    global_store_byte v[0:1], v28, off offset:192
-; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:144
-; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148
-; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:152
-; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:156
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v31, off offset:198
+; ALIGNED-NEXT:    global_store_byte v[16:17], v31, off offset:196
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v30, off offset:194
+; ALIGNED-NEXT:    global_store_byte v[16:17], v30, off offset:192
+; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:128
+; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:132
+; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:136
+; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:140
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:156
-; ALIGNED-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:152
-; ALIGNED-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:148
-; ALIGNED-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:144
+; ALIGNED-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:140
+; ALIGNED-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:136
+; ALIGNED-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:132
+; ALIGNED-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:128
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v35, off offset:190
-; ALIGNED-NEXT:    global_store_byte v[0:1], v35, off offset:188
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v37, off offset:190
+; ALIGNED-NEXT:    global_store_byte v[16:17], v37, off offset:188
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v34, off offset:186
-; ALIGNED-NEXT:    global_store_byte v[0:1], v34, off offset:184
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v36, off offset:186
+; ALIGNED-NEXT:    global_store_byte v[16:17], v36, off offset:184
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v33, off offset:182
-; ALIGNED-NEXT:    global_store_byte v[0:1], v33, off offset:180
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v32, off offset:178
-; ALIGNED-NEXT:    global_store_byte v[0:1], v32, off offset:176
-; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:160
-; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:164
-; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:168
-; ALIGNED-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:172
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v35, off offset:182
+; ALIGNED-NEXT:    global_store_byte v[16:17], v35, off offset:180
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v34, off offset:178
+; ALIGNED-NEXT:    global_store_byte v[16:17], v34, off offset:176
+; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:144
+; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:148
+; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:152
+; ALIGNED-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:156
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:172
-; ALIGNED-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:168
-; ALIGNED-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:164
-; ALIGNED-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:160
+; ALIGNED-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:156
+; ALIGNED-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:152
+; ALIGNED-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:148
+; ALIGNED-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:144
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v39, off offset:174
-; ALIGNED-NEXT:    global_store_byte v[0:1], v39, off offset:172
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v49, off offset:174
+; ALIGNED-NEXT:    global_store_byte v[16:17], v49, off offset:172
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v38, off offset:170
-; ALIGNED-NEXT:    global_store_byte v[0:1], v38, off offset:168
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v48, off offset:170
+; ALIGNED-NEXT:    global_store_byte v[16:17], v48, off offset:168
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v37, off offset:166
-; ALIGNED-NEXT:    global_store_byte v[0:1], v37, off offset:164
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v36, off offset:162
-; ALIGNED-NEXT:    global_store_byte v[0:1], v36, off offset:160
-; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:112
-; ALIGNED-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:116
-; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:120
-; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:124
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v39, off offset:166
+; ALIGNED-NEXT:    global_store_byte v[16:17], v39, off offset:164
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v38, off offset:162
+; ALIGNED-NEXT:    global_store_byte v[16:17], v38, off offset:160
+; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:96
+; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:100
+; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:104
+; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:108
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:124
-; ALIGNED-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:120
-; ALIGNED-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:116
-; ALIGNED-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:112
+; ALIGNED-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:108
+; ALIGNED-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:104
+; ALIGNED-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:100
+; ALIGNED-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:96
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v51, off offset:158
-; ALIGNED-NEXT:    global_store_byte v[0:1], v51, off offset:156
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v53, off offset:158
+; ALIGNED-NEXT:    global_store_byte v[16:17], v53, off offset:156
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v50, off offset:154
-; ALIGNED-NEXT:    global_store_byte v[0:1], v50, off offset:152
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v52, off offset:154
+; ALIGNED-NEXT:    global_store_byte v[16:17], v52, off offset:152
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v49, off offset:150
-; ALIGNED-NEXT:    global_store_byte v[0:1], v49, off offset:148
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v48, off offset:146
-; ALIGNED-NEXT:    global_store_byte v[0:1], v48, off offset:144
-; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:128
-; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:132
-; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:136
-; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:140
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v51, off offset:150
+; ALIGNED-NEXT:    global_store_byte v[16:17], v51, off offset:148
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v50, off offset:146
+; ALIGNED-NEXT:    global_store_byte v[16:17], v50, off offset:144
+; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:112
+; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:116
+; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:120
+; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:124
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:140
-; ALIGNED-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:136
-; ALIGNED-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:132
-; ALIGNED-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:128
+; ALIGNED-NEXT:    buffer_load_dword v65, off, s[0:3], s32 offset:124
+; ALIGNED-NEXT:    buffer_load_dword v64, off, s[0:3], s32 offset:120
+; ALIGNED-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:116
+; ALIGNED-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:112
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v55, off offset:142
-; ALIGNED-NEXT:    global_store_byte v[0:1], v55, off offset:140
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v65, off offset:142
+; ALIGNED-NEXT:    global_store_byte v[16:17], v65, off offset:140
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v54, off offset:138
-; ALIGNED-NEXT:    global_store_byte v[0:1], v54, off offset:136
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v64, off offset:138
+; ALIGNED-NEXT:    global_store_byte v[16:17], v64, off offset:136
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v53, off offset:134
-; ALIGNED-NEXT:    global_store_byte v[0:1], v53, off offset:132
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v52, off offset:130
-; ALIGNED-NEXT:    global_store_byte v[0:1], v52, off offset:128
-; ALIGNED-NEXT:    buffer_store_dword v64, off, s[0:3], s32 offset:208
-; ALIGNED-NEXT:    buffer_store_dword v65, off, s[0:3], s32 offset:212
-; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:216
-; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:220
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v55, off offset:134
+; ALIGNED-NEXT:    global_store_byte v[16:17], v55, off offset:132
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v54, off offset:130
+; ALIGNED-NEXT:    global_store_byte v[16:17], v54, off offset:128
+; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:192
+; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:196
+; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:200
+; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:204
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v67, off, s[0:3], s32 offset:220
-; ALIGNED-NEXT:    buffer_load_dword v66, off, s[0:3], s32 offset:216
-; ALIGNED-NEXT:    buffer_load_dword v65, off, s[0:3], s32 offset:212
-; ALIGNED-NEXT:    buffer_load_dword v64, off, s[0:3], s32 offset:208
+; ALIGNED-NEXT:    buffer_load_dword v69, off, s[0:3], s32 offset:204
+; ALIGNED-NEXT:    buffer_load_dword v68, off, s[0:3], s32 offset:200
+; ALIGNED-NEXT:    buffer_load_dword v67, off, s[0:3], s32 offset:196
+; ALIGNED-NEXT:    buffer_load_dword v66, off, s[0:3], s32 offset:192
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v67, off offset:126
-; ALIGNED-NEXT:    global_store_byte v[0:1], v67, off offset:124
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v69, off offset:126
+; ALIGNED-NEXT:    global_store_byte v[16:17], v69, off offset:124
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v66, off offset:122
-; ALIGNED-NEXT:    global_store_byte v[0:1], v66, off offset:120
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v68, off offset:122
+; ALIGNED-NEXT:    global_store_byte v[16:17], v68, off offset:120
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v65, off offset:118
-; ALIGNED-NEXT:    global_store_byte v[0:1], v65, off offset:116
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v64, off offset:114
-; ALIGNED-NEXT:    global_store_byte v[0:1], v64, off offset:112
-; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:224
-; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:228
-; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:232
-; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:236
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v67, off offset:118
+; ALIGNED-NEXT:    global_store_byte v[16:17], v67, off offset:116
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v66, off offset:114
+; ALIGNED-NEXT:    global_store_byte v[16:17], v66, off offset:112
+; ALIGNED-NEXT:    buffer_store_dword v116, off, s[0:3], s32 offset:208
+; ALIGNED-NEXT:    buffer_store_dword v117, off, s[0:3], s32 offset:212
+; ALIGNED-NEXT:    buffer_store_dword v118, off, s[0:3], s32 offset:216
+; ALIGNED-NEXT:    buffer_store_dword v119, off, s[0:3], s32 offset:220
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v71, off, s[0:3], s32 offset:236
-; ALIGNED-NEXT:    buffer_load_dword v70, off, s[0:3], s32 offset:232
-; ALIGNED-NEXT:    buffer_load_dword v69, off, s[0:3], s32 offset:228
-; ALIGNED-NEXT:    buffer_load_dword v68, off, s[0:3], s32 offset:224
+; ALIGNED-NEXT:    buffer_load_dword v81, off, s[0:3], s32 offset:220
+; ALIGNED-NEXT:    buffer_load_dword v80, off, s[0:3], s32 offset:216
+; ALIGNED-NEXT:    buffer_load_dword v71, off, s[0:3], s32 offset:212
+; ALIGNED-NEXT:    buffer_load_dword v70, off, s[0:3], s32 offset:208
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v71, off offset:110
-; ALIGNED-NEXT:    global_store_byte v[0:1], v71, off offset:108
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v81, off offset:110
+; ALIGNED-NEXT:    global_store_byte v[16:17], v81, off offset:108
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v70, off offset:106
-; ALIGNED-NEXT:    global_store_byte v[0:1], v70, off offset:104
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v80, off offset:106
+; ALIGNED-NEXT:    global_store_byte v[16:17], v80, off offset:104
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v69, off offset:102
-; ALIGNED-NEXT:    global_store_byte v[0:1], v69, off offset:100
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v68, off offset:98
-; ALIGNED-NEXT:    global_store_byte v[0:1], v68, off offset:96
-; ALIGNED-NEXT:    buffer_store_dword v80, off, s[0:3], s32 offset:176
-; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:180
-; ALIGNED-NEXT:    buffer_store_dword v82, off, s[0:3], s32 offset:184
-; ALIGNED-NEXT:    buffer_store_dword v83, off, s[0:3], s32 offset:188
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v71, off offset:102
+; ALIGNED-NEXT:    global_store_byte v[16:17], v71, off offset:100
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v70, off offset:98
+; ALIGNED-NEXT:    global_store_byte v[16:17], v70, off offset:96
+; ALIGNED-NEXT:    buffer_store_dword v82, off, s[0:3], s32 offset:160
+; ALIGNED-NEXT:    buffer_store_dword v83, off, s[0:3], s32 offset:164
+; ALIGNED-NEXT:    buffer_store_dword v84, off, s[0:3], s32 offset:168
+; ALIGNED-NEXT:    buffer_store_dword v85, off, s[0:3], s32 offset:172
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v83, off, s[0:3], s32 offset:188
-; ALIGNED-NEXT:    buffer_load_dword v82, off, s[0:3], s32 offset:184
-; ALIGNED-NEXT:    buffer_load_dword v81, off, s[0:3], s32 offset:180
-; ALIGNED-NEXT:    buffer_load_dword v80, off, s[0:3], s32 offset:176
+; ALIGNED-NEXT:    buffer_load_dword v85, off, s[0:3], s32 offset:172
+; ALIGNED-NEXT:    buffer_load_dword v84, off, s[0:3], s32 offset:168
+; ALIGNED-NEXT:    buffer_load_dword v83, off, s[0:3], s32 offset:164
+; ALIGNED-NEXT:    buffer_load_dword v82, off, s[0:3], s32 offset:160
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v83, off offset:94
-; ALIGNED-NEXT:    global_store_byte v[0:1], v83, off offset:92
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v85, off offset:94
+; ALIGNED-NEXT:    global_store_byte v[16:17], v85, off offset:92
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v82, off offset:90
-; ALIGNED-NEXT:    global_store_byte v[0:1], v82, off offset:88
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v84, off offset:90
+; ALIGNED-NEXT:    global_store_byte v[16:17], v84, off offset:88
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v81, off offset:86
-; ALIGNED-NEXT:    global_store_byte v[0:1], v81, off offset:84
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v80, off offset:82
-; ALIGNED-NEXT:    global_store_byte v[0:1], v80, off offset:80
-; ALIGNED-NEXT:    buffer_store_dword v84, off, s[0:3], s32 offset:192
-; ALIGNED-NEXT:    buffer_store_dword v85, off, s[0:3], s32 offset:196
-; ALIGNED-NEXT:    buffer_store_dword v86, off, s[0:3], s32 offset:200
-; ALIGNED-NEXT:    buffer_store_dword v87, off, s[0:3], s32 offset:204
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v83, off offset:86
+; ALIGNED-NEXT:    global_store_byte v[16:17], v83, off offset:84
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v82, off offset:82
+; ALIGNED-NEXT:    global_store_byte v[16:17], v82, off offset:80
+; ALIGNED-NEXT:    buffer_store_dword v112, off, s[0:3], s32 offset:176
+; ALIGNED-NEXT:    buffer_store_dword v113, off, s[0:3], s32 offset:180
+; ALIGNED-NEXT:    buffer_store_dword v114, off, s[0:3], s32 offset:184
+; ALIGNED-NEXT:    buffer_store_dword v115, off, s[0:3], s32 offset:188
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v87, off, s[0:3], s32 offset:204
-; ALIGNED-NEXT:    buffer_load_dword v86, off, s[0:3], s32 offset:200
-; ALIGNED-NEXT:    buffer_load_dword v85, off, s[0:3], s32 offset:196
-; ALIGNED-NEXT:    buffer_load_dword v84, off, s[0:3], s32 offset:192
+; ALIGNED-NEXT:    buffer_load_dword v97, off, s[0:3], s32 offset:188
+; ALIGNED-NEXT:    buffer_load_dword v96, off, s[0:3], s32 offset:184
+; ALIGNED-NEXT:    buffer_load_dword v87, off, s[0:3], s32 offset:180
+; ALIGNED-NEXT:    buffer_load_dword v86, off, s[0:3], s32 offset:176
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v19
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v19, 8, v19
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v18
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v18, 8, v18
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v25
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 8, v25
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v24
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 8, v24
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v87, off offset:78
-; ALIGNED-NEXT:    global_store_byte v[0:1], v87, off offset:76
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v97, off offset:78
+; ALIGNED-NEXT:    global_store_byte v[16:17], v97, off offset:76
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v86, off offset:74
-; ALIGNED-NEXT:    global_store_byte v[0:1], v86, off offset:72
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v96, off offset:74
+; ALIGNED-NEXT:    global_store_byte v[16:17], v96, off offset:72
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v85, off offset:70
-; ALIGNED-NEXT:    global_store_byte v[0:1], v85, off offset:68
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v84, off offset:66
-; ALIGNED-NEXT:    global_store_byte v[0:1], v84, off offset:64
-; ALIGNED-NEXT:    buffer_store_dword v96, off, s[0:3], s32 offset:272
-; ALIGNED-NEXT:    buffer_store_dword v97, off, s[0:3], s32 offset:276
-; ALIGNED-NEXT:    buffer_store_dword v98, off, s[0:3], s32 offset:280
-; ALIGNED-NEXT:    buffer_store_dword v99, off, s[0:3], s32 offset:284
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v87, off offset:70
+; ALIGNED-NEXT:    global_store_byte v[16:17], v87, off offset:68
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v86, off offset:66
+; ALIGNED-NEXT:    global_store_byte v[16:17], v86, off offset:64
+; ALIGNED-NEXT:    buffer_store_dword v98, off, s[0:3], s32 offset:256
+; ALIGNED-NEXT:    buffer_store_dword v99, off, s[0:3], s32 offset:260
+; ALIGNED-NEXT:    buffer_store_dword v100, off, s[0:3], s32 offset:264
+; ALIGNED-NEXT:    buffer_store_dword v101, off, s[0:3], s32 offset:268
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v99, off, s[0:3], s32 offset:284
-; ALIGNED-NEXT:    buffer_load_dword v98, off, s[0:3], s32 offset:280
-; ALIGNED-NEXT:    buffer_load_dword v97, off, s[0:3], s32 offset:276
-; ALIGNED-NEXT:    buffer_load_dword v96, off, s[0:3], s32 offset:272
+; ALIGNED-NEXT:    buffer_load_dword v101, off, s[0:3], s32 offset:268
+; ALIGNED-NEXT:    buffer_load_dword v100, off, s[0:3], s32 offset:264
+; ALIGNED-NEXT:    buffer_load_dword v99, off, s[0:3], s32 offset:260
+; ALIGNED-NEXT:    buffer_load_dword v98, off, s[0:3], s32 offset:256
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v99, off offset:62
-; ALIGNED-NEXT:    global_store_byte v[0:1], v99, off offset:60
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v101, off offset:62
+; ALIGNED-NEXT:    global_store_byte v[16:17], v101, off offset:60
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v98, off offset:58
-; ALIGNED-NEXT:    global_store_byte v[0:1], v98, off offset:56
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v100, off offset:58
+; ALIGNED-NEXT:    global_store_byte v[16:17], v100, off offset:56
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v97, off offset:54
-; ALIGNED-NEXT:    global_store_byte v[0:1], v97, off offset:52
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v96, off offset:50
-; ALIGNED-NEXT:    global_store_byte v[0:1], v96, off offset:48
-; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:288
-; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:292
-; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:296
-; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:300
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v99, off offset:54
+; ALIGNED-NEXT:    global_store_byte v[16:17], v99, off offset:52
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v98, off offset:50
+; ALIGNED-NEXT:    global_store_byte v[16:17], v98, off offset:48
+; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:272
+; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:276
+; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:280
+; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:284
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:296
-; ALIGNED-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:300
-; ALIGNED-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:288
-; ALIGNED-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:292
+; ALIGNED-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:280
+; ALIGNED-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:284
+; ALIGNED-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:272
+; ALIGNED-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:276
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v15, off offset:42
-; ALIGNED-NEXT:    global_store_byte v[0:1], v15, off offset:40
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v15, off offset:42
+; ALIGNED-NEXT:    global_store_byte v[16:17], v15, off offset:40
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v14, off offset:46
-; ALIGNED-NEXT:    global_store_byte v[0:1], v14, off offset:44
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v14, off offset:46
+; ALIGNED-NEXT:    global_store_byte v[16:17], v14, off offset:44
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v13, off offset:34
-; ALIGNED-NEXT:    global_store_byte v[0:1], v13, off offset:32
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v12, off offset:38
-; ALIGNED-NEXT:    global_store_byte v[0:1], v12, off offset:36
-; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:240
-; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:244
-; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:248
-; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:252
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v13, off offset:34
+; ALIGNED-NEXT:    global_store_byte v[16:17], v13, off offset:32
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v12, off offset:38
+; ALIGNED-NEXT:    global_store_byte v[16:17], v12, off offset:36
+; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:224
+; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:228
+; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:232
+; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:236
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:252
-; ALIGNED-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:248
-; ALIGNED-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:244
-; ALIGNED-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:240
+; ALIGNED-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:236
+; ALIGNED-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:232
+; ALIGNED-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:228
+; ALIGNED-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:224
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v11, off offset:30
-; ALIGNED-NEXT:    global_store_byte v[0:1], v11, off offset:28
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v11, off offset:30
+; ALIGNED-NEXT:    global_store_byte v[16:17], v11, off offset:28
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v10, off offset:26
-; ALIGNED-NEXT:    global_store_byte v[0:1], v10, off offset:24
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v10, off offset:26
+; ALIGNED-NEXT:    global_store_byte v[16:17], v10, off offset:24
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v9, off offset:22
-; ALIGNED-NEXT:    global_store_byte v[0:1], v9, off offset:20
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v8, off offset:18
-; ALIGNED-NEXT:    global_store_byte v[0:1], v8, off offset:16
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:256
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:260
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:264
-; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:268
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v9, off offset:22
+; ALIGNED-NEXT:    global_store_byte v[16:17], v9, off offset:20
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v8, off offset:18
+; ALIGNED-NEXT:    global_store_byte v[16:17], v8, off offset:16
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:240
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:244
+; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:248
+; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:252
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:268
-; ALIGNED-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:264
-; ALIGNED-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:260
-; ALIGNED-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:256
-; ALIGNED-NEXT:    global_store_byte v[0:1], v100, off offset:255
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v27
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 8, v27
-; ALIGNED-NEXT:    global_store_byte v[0:1], v19, off offset:253
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v19, 24, v26
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 8, v26
-; ALIGNED-NEXT:    global_store_byte v[0:1], v101, off offset:251
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v25
-; ALIGNED-NEXT:    global_store_byte v[0:1], v27, off offset:221
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 24, v50
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 8, v50
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 8, v25
-; ALIGNED-NEXT:    global_store_byte v[0:1], v18, off offset:249
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v18, 24, v24
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 8, v24
-; ALIGNED-NEXT:    global_store_byte v[0:1], v102, off offset:247
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v31
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 8, v31
-; ALIGNED-NEXT:    global_store_byte v[0:1], v17, off offset:245
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v17, 24, v30
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 8, v30
-; ALIGNED-NEXT:    global_store_byte v[0:1], v103, off offset:243
+; ALIGNED-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:252
+; ALIGNED-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:248
+; ALIGNED-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:244
+; ALIGNED-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:240
+; ALIGNED-NEXT:    global_store_byte v[16:17], v112, off offset:247
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v27
+; ALIGNED-NEXT:    global_store_byte v[16:17], v102, off offset:255
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v23
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v23, 8, v23
+; ALIGNED-NEXT:    global_store_byte v[16:17], v21, off offset:253
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v21, 24, v22
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v22, 8, v22
+; ALIGNED-NEXT:    global_store_byte v[16:17], v103, off offset:251
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v29
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 8, v29
-; ALIGNED-NEXT:    global_store_byte v[0:1], v16, off offset:241
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v16, 24, v28
+; ALIGNED-NEXT:    global_store_byte v[16:17], v20, off offset:249
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v20, 24, v28
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 8, v28
-; ALIGNED-NEXT:    global_store_byte v[0:1], v112, off offset:239
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v35
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 8, v35
-; ALIGNED-NEXT:    global_store_byte v[0:1], v23, off offset:237
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v23, 24, v34
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 8, v34
-; ALIGNED-NEXT:    global_store_byte v[0:1], v113, off offset:235
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 8, v27
+; ALIGNED-NEXT:    global_store_byte v[16:17], v19, off offset:245
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v19, 24, v26
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 8, v26
+; ALIGNED-NEXT:    global_store_byte v[16:17], v112, off offset:215
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v51
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v51
+; ALIGNED-NEXT:    global_store_byte v[16:17], v113, off offset:243
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v33
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 8, v33
-; ALIGNED-NEXT:    global_store_byte v[0:1], v22, off offset:233
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v22, 24, v32
+; ALIGNED-NEXT:    global_store_byte v[16:17], v18, off offset:241
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v18, 24, v32
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 8, v32
-; ALIGNED-NEXT:    global_store_byte v[0:1], v114, off offset:231
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v39
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 8, v39
-; ALIGNED-NEXT:    global_store_byte v[0:1], v21, off offset:229
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v21, 24, v38
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 8, v38
-; ALIGNED-NEXT:    global_store_byte v[0:1], v115, off offset:227
+; ALIGNED-NEXT:    global_store_byte v[16:17], v114, off offset:239
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v31
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 8, v31
+; ALIGNED-NEXT:    global_store_byte v[16:17], v25, off offset:237
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 24, v30
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 8, v30
+; ALIGNED-NEXT:    global_store_byte v[16:17], v115, off offset:235
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v37
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 8, v37
-; ALIGNED-NEXT:    global_store_byte v[0:1], v20, off offset:225
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v20, 24, v36
+; ALIGNED-NEXT:    global_store_byte v[16:17], v24, off offset:233
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v36
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 8, v36
-; ALIGNED-NEXT:    global_store_byte v[0:1], v100, off offset:223
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v51
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v51
-; ALIGNED-NEXT:    global_store_byte v[0:1], v19, off offset:219
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v19, 24, v49
+; ALIGNED-NEXT:    global_store_byte v[16:17], v102, off offset:231
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v35
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 8, v35
+; ALIGNED-NEXT:    global_store_byte v[16:17], v23, off offset:229
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v23, 24, v34
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 8, v34
+; ALIGNED-NEXT:    global_store_byte v[16:17], v21, off offset:227
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v21, 24, v49
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v49
-; ALIGNED-NEXT:    global_store_byte v[0:1], v26, off offset:217
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 24, v48
+; ALIGNED-NEXT:    global_store_byte v[16:17], v22, off offset:225
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v22, 24, v48
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 8, v48
-; ALIGNED-NEXT:    global_store_byte v[0:1], v101, off offset:215
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v55
-; ALIGNED-NEXT:    global_store_byte v[0:1], v50, off offset:153
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v8
+; ALIGNED-NEXT:    global_store_byte v[16:17], v103, off offset:223
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v39
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 8, v39
+; ALIGNED-NEXT:    global_store_byte v[16:17], v29, off offset:221
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 24, v38
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 8, v38
+; ALIGNED-NEXT:    global_store_byte v[16:17], v20, off offset:219
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v20, 24, v53
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v53
+; ALIGNED-NEXT:    global_store_byte v[16:17], v28, off offset:217
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 24, v52
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 8, v52
+; ALIGNED-NEXT:    global_store_byte v[16:17], v27, off offset:213
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 24, v50
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 8, v50
+; ALIGNED-NEXT:    global_store_byte v[16:17], v19, off offset:211
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v19, 24, v65
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v65
+; ALIGNED-NEXT:    global_store_byte v[16:17], v26, off offset:209
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 24, v64
+; ALIGNED-NEXT:    global_store_byte v[16:17], v51, off offset:149
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 24, v8
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v8, 8, v8
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 8, v64
+; ALIGNED-NEXT:    global_store_byte v[16:17], v113, off offset:207
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v55
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v55
-; ALIGNED-NEXT:    global_store_byte v[0:1], v25, off offset:213
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 24, v54
+; ALIGNED-NEXT:    global_store_byte v[16:17], v33, off offset:205
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 24, v54
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 8, v54
-; ALIGNED-NEXT:    global_store_byte v[0:1], v18, off offset:211
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v18, 24, v53
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v53
-; ALIGNED-NEXT:    global_store_byte v[0:1], v24, off offset:209
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v52
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 8, v52
-; ALIGNED-NEXT:    global_store_byte v[0:1], v102, off offset:207
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v67
+; ALIGNED-NEXT:    global_store_byte v[16:17], v18, off offset:203
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v18, 24, v69
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v69
+; ALIGNED-NEXT:    global_store_byte v[16:17], v32, off offset:201
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 24, v68
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 8, v68
+; ALIGNED-NEXT:    global_store_byte v[16:17], v114, off offset:199
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v67
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v67
-; ALIGNED-NEXT:    global_store_byte v[0:1], v31, off offset:205
+; ALIGNED-NEXT:    global_store_byte v[16:17], v31, off offset:197
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 24, v66
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 8, v66
-; ALIGNED-NEXT:    global_store_byte v[0:1], v17, off offset:203
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v17, 24, v65
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v65
-; ALIGNED-NEXT:    global_store_byte v[0:1], v30, off offset:201
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v64
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 8, v64
-; ALIGNED-NEXT:    global_store_byte v[0:1], v103, off offset:199
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v71
+; ALIGNED-NEXT:    global_store_byte v[16:17], v25, off offset:195
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 24, v81
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v81
+; ALIGNED-NEXT:    global_store_byte v[16:17], v30, off offset:193
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v80
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 8, v80
+; ALIGNED-NEXT:    global_store_byte v[16:17], v115, off offset:191
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v71
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v71
-; ALIGNED-NEXT:    global_store_byte v[0:1], v29, off offset:197
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 24, v70
+; ALIGNED-NEXT:    global_store_byte v[16:17], v37, off offset:189
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 24, v70
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 8, v70
-; ALIGNED-NEXT:    global_store_byte v[0:1], v16, off offset:195
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v16, 24, v69
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v69
-; ALIGNED-NEXT:    global_store_byte v[0:1], v28, off offset:193
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 24, v68
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 8, v68
-; ALIGNED-NEXT:    global_store_byte v[0:1], v112, off offset:191
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v83
+; ALIGNED-NEXT:    global_store_byte v[16:17], v24, off offset:187
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v85
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 8, v85
+; ALIGNED-NEXT:    global_store_byte v[16:17], v36, off offset:185
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 24, v84
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 8, v84
+; ALIGNED-NEXT:    global_store_byte v[16:17], v102, off offset:183
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v83
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v83
-; ALIGNED-NEXT:    global_store_byte v[0:1], v35, off offset:189
+; ALIGNED-NEXT:    global_store_byte v[16:17], v35, off offset:181
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 24, v82
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 8, v82
-; ALIGNED-NEXT:    global_store_byte v[0:1], v23, off offset:187
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v23, 24, v81
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v81
-; ALIGNED-NEXT:    global_store_byte v[0:1], v34, off offset:185
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 24, v80
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 8, v80
-; ALIGNED-NEXT:    global_store_byte v[0:1], v113, off offset:183
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v87
+; ALIGNED-NEXT:    global_store_byte v[16:17], v23, off offset:179
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v23, 24, v97
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 8, v97
+; ALIGNED-NEXT:    global_store_byte v[16:17], v34, off offset:177
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 24, v96
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 8, v96
+; ALIGNED-NEXT:    global_store_byte v[16:17], v21, off offset:175
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v21, 24, v87
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 8, v87
-; ALIGNED-NEXT:    global_store_byte v[0:1], v33, off offset:181
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 24, v86
+; ALIGNED-NEXT:    global_store_byte v[16:17], v49, off offset:173
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 24, v86
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 8, v86
-; ALIGNED-NEXT:    global_store_byte v[0:1], v22, off offset:179
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v22, 24, v85
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 8, v85
-; ALIGNED-NEXT:    global_store_byte v[0:1], v32, off offset:177
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 24, v84
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 8, v84
-; ALIGNED-NEXT:    global_store_byte v[0:1], v114, off offset:175
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v99
+; ALIGNED-NEXT:    global_store_byte v[16:17], v22, off offset:171
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v22, 24, v101
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v101
+; ALIGNED-NEXT:    global_store_byte v[16:17], v48, off offset:169
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 24, v100
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 8, v100
+; ALIGNED-NEXT:    global_store_byte v[16:17], v103, off offset:167
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v99
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v99
-; ALIGNED-NEXT:    global_store_byte v[0:1], v39, off offset:173
+; ALIGNED-NEXT:    global_store_byte v[16:17], v39, off offset:165
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 24, v98
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 8, v98
-; ALIGNED-NEXT:    global_store_byte v[0:1], v21, off offset:171
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v21, 24, v97
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 8, v97
-; ALIGNED-NEXT:    global_store_byte v[0:1], v38, off offset:169
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 24, v96
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 8, v96
-; ALIGNED-NEXT:    global_store_byte v[0:1], v115, off offset:167
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v15
+; ALIGNED-NEXT:    global_store_byte v[16:17], v29, off offset:163
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 24, v15
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v15, 8, v15
-; ALIGNED-NEXT:    global_store_byte v[0:1], v37, off offset:165
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 24, v14
+; ALIGNED-NEXT:    global_store_byte v[16:17], v38, off offset:161
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 24, v14
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v14, 8, v14
-; ALIGNED-NEXT:    global_store_byte v[0:1], v20, off offset:163
+; ALIGNED-NEXT:    global_store_byte v[16:17], v20, off offset:159
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v20, 24, v13
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v13, 8, v13
-; ALIGNED-NEXT:    global_store_byte v[0:1], v36, off offset:161
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 24, v12
+; ALIGNED-NEXT:    global_store_byte v[16:17], v53, off offset:157
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 24, v12
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v12, 8, v12
-; ALIGNED-NEXT:    global_store_byte v[0:1], v100, off offset:159
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v11
+; ALIGNED-NEXT:    global_store_byte v[16:17], v28, off offset:155
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 24, v11
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v11, 8, v11
-; ALIGNED-NEXT:    global_store_byte v[0:1], v51, off offset:157
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 24, v10
+; ALIGNED-NEXT:    global_store_byte v[16:17], v52, off offset:153
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v10
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v10, 8, v10
-; ALIGNED-NEXT:    global_store_byte v[0:1], v27, off offset:155
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 24, v9
+; ALIGNED-NEXT:    global_store_byte v[16:17], v112, off offset:151
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v9
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v9, 8, v9
-; ALIGNED-NEXT:    global_store_byte v[0:1], v19, off offset:151
-; ALIGNED-NEXT:    global_store_byte v[0:1], v49, off offset:149
-; ALIGNED-NEXT:    global_store_byte v[0:1], v26, off offset:147
-; ALIGNED-NEXT:    global_store_byte v[0:1], v48, off offset:145
-; ALIGNED-NEXT:    global_store_byte v[0:1], v101, off offset:143
-; ALIGNED-NEXT:    global_store_byte v[0:1], v55, off offset:141
-; ALIGNED-NEXT:    global_store_byte v[0:1], v25, off offset:139
-; ALIGNED-NEXT:    global_store_byte v[0:1], v54, off offset:137
-; ALIGNED-NEXT:    global_store_byte v[0:1], v18, off offset:135
-; ALIGNED-NEXT:    global_store_byte v[0:1], v53, off offset:133
-; ALIGNED-NEXT:    global_store_byte v[0:1], v24, off offset:131
-; ALIGNED-NEXT:    global_store_byte v[0:1], v52, off offset:129
-; ALIGNED-NEXT:    global_store_byte v[0:1], v102, off offset:127
-; ALIGNED-NEXT:    global_store_byte v[0:1], v67, off offset:125
-; ALIGNED-NEXT:    global_store_byte v[0:1], v31, off offset:123
-; ALIGNED-NEXT:    global_store_byte v[0:1], v66, off offset:121
-; ALIGNED-NEXT:    global_store_byte v[0:1], v17, off offset:119
-; ALIGNED-NEXT:    global_store_byte v[0:1], v65, off offset:117
-; ALIGNED-NEXT:    global_store_byte v[0:1], v30, off offset:115
-; ALIGNED-NEXT:    global_store_byte v[0:1], v64, off offset:113
-; ALIGNED-NEXT:    global_store_byte v[0:1], v103, off offset:111
-; ALIGNED-NEXT:    global_store_byte v[0:1], v71, off offset:109
-; ALIGNED-NEXT:    global_store_byte v[0:1], v29, off offset:107
-; ALIGNED-NEXT:    global_store_byte v[0:1], v70, off offset:105
-; ALIGNED-NEXT:    global_store_byte v[0:1], v16, off offset:103
-; ALIGNED-NEXT:    global_store_byte v[0:1], v69, off offset:101
-; ALIGNED-NEXT:    global_store_byte v[0:1], v28, off offset:99
-; ALIGNED-NEXT:    global_store_byte v[0:1], v68, off offset:97
-; ALIGNED-NEXT:    global_store_byte v[0:1], v112, off offset:95
-; ALIGNED-NEXT:    global_store_byte v[0:1], v83, off offset:93
-; ALIGNED-NEXT:    global_store_byte v[0:1], v35, off offset:91
-; ALIGNED-NEXT:    global_store_byte v[0:1], v82, off offset:89
-; ALIGNED-NEXT:    global_store_byte v[0:1], v23, off offset:87
-; ALIGNED-NEXT:    global_store_byte v[0:1], v81, off offset:85
-; ALIGNED-NEXT:    global_store_byte v[0:1], v34, off offset:83
-; ALIGNED-NEXT:    global_store_byte v[0:1], v80, off offset:81
-; ALIGNED-NEXT:    global_store_byte v[0:1], v113, off offset:79
-; ALIGNED-NEXT:    global_store_byte v[0:1], v87, off offset:77
-; ALIGNED-NEXT:    global_store_byte v[0:1], v33, off offset:75
-; ALIGNED-NEXT:    global_store_byte v[0:1], v86, off offset:73
-; ALIGNED-NEXT:    global_store_byte v[0:1], v22, off offset:71
-; ALIGNED-NEXT:    global_store_byte v[0:1], v85, off offset:69
-; ALIGNED-NEXT:    global_store_byte v[0:1], v32, off offset:67
-; ALIGNED-NEXT:    global_store_byte v[0:1], v84, off offset:65
-; ALIGNED-NEXT:    global_store_byte v[0:1], v114, off offset:63
-; ALIGNED-NEXT:    global_store_byte v[0:1], v99, off offset:61
-; ALIGNED-NEXT:    global_store_byte v[0:1], v39, off offset:59
-; ALIGNED-NEXT:    global_store_byte v[0:1], v98, off offset:57
-; ALIGNED-NEXT:    global_store_byte v[0:1], v21, off offset:55
-; ALIGNED-NEXT:    global_store_byte v[0:1], v97, off offset:53
-; ALIGNED-NEXT:    global_store_byte v[0:1], v38, off offset:51
-; ALIGNED-NEXT:    global_store_byte v[0:1], v96, off offset:49
-; ALIGNED-NEXT:    global_store_byte v[0:1], v115, off offset:43
-; ALIGNED-NEXT:    global_store_byte v[0:1], v15, off offset:41
-; ALIGNED-NEXT:    global_store_byte v[0:1], v37, off offset:47
-; ALIGNED-NEXT:    global_store_byte v[0:1], v14, off offset:45
-; ALIGNED-NEXT:    global_store_byte v[0:1], v20, off offset:35
-; ALIGNED-NEXT:    global_store_byte v[0:1], v13, off offset:33
-; ALIGNED-NEXT:    global_store_byte v[0:1], v36, off offset:39
-; ALIGNED-NEXT:    global_store_byte v[0:1], v12, off offset:37
-; ALIGNED-NEXT:    global_store_byte v[0:1], v100, off offset:31
-; ALIGNED-NEXT:    global_store_byte v[0:1], v11, off offset:29
-; ALIGNED-NEXT:    global_store_byte v[0:1], v51, off offset:27
-; ALIGNED-NEXT:    global_store_byte v[0:1], v10, off offset:25
-; ALIGNED-NEXT:    global_store_byte v[0:1], v27, off offset:23
-; ALIGNED-NEXT:    global_store_byte v[0:1], v9, off offset:21
-; ALIGNED-NEXT:    global_store_byte v[0:1], v50, off offset:19
-; ALIGNED-NEXT:    global_store_byte v[0:1], v8, off offset:17
+; ALIGNED-NEXT:    global_store_byte v[16:17], v27, off offset:147
+; ALIGNED-NEXT:    global_store_byte v[16:17], v50, off offset:145
+; ALIGNED-NEXT:    global_store_byte v[16:17], v19, off offset:143
+; ALIGNED-NEXT:    global_store_byte v[16:17], v65, off offset:141
+; ALIGNED-NEXT:    global_store_byte v[16:17], v26, off offset:139
+; ALIGNED-NEXT:    global_store_byte v[16:17], v64, off offset:137
+; ALIGNED-NEXT:    global_store_byte v[16:17], v113, off offset:135
+; ALIGNED-NEXT:    global_store_byte v[16:17], v55, off offset:133
+; ALIGNED-NEXT:    global_store_byte v[16:17], v33, off offset:131
+; ALIGNED-NEXT:    global_store_byte v[16:17], v54, off offset:129
+; ALIGNED-NEXT:    global_store_byte v[16:17], v18, off offset:127
+; ALIGNED-NEXT:    global_store_byte v[16:17], v69, off offset:125
+; ALIGNED-NEXT:    global_store_byte v[16:17], v32, off offset:123
+; ALIGNED-NEXT:    global_store_byte v[16:17], v68, off offset:121
+; ALIGNED-NEXT:    global_store_byte v[16:17], v114, off offset:119
+; ALIGNED-NEXT:    global_store_byte v[16:17], v67, off offset:117
+; ALIGNED-NEXT:    global_store_byte v[16:17], v31, off offset:115
+; ALIGNED-NEXT:    global_store_byte v[16:17], v66, off offset:113
+; ALIGNED-NEXT:    global_store_byte v[16:17], v25, off offset:111
+; ALIGNED-NEXT:    global_store_byte v[16:17], v81, off offset:109
+; ALIGNED-NEXT:    global_store_byte v[16:17], v30, off offset:107
+; ALIGNED-NEXT:    global_store_byte v[16:17], v80, off offset:105
+; ALIGNED-NEXT:    global_store_byte v[16:17], v115, off offset:103
+; ALIGNED-NEXT:    global_store_byte v[16:17], v71, off offset:101
+; ALIGNED-NEXT:    global_store_byte v[16:17], v37, off offset:99
+; ALIGNED-NEXT:    global_store_byte v[16:17], v70, off offset:97
+; ALIGNED-NEXT:    global_store_byte v[16:17], v24, off offset:95
+; ALIGNED-NEXT:    global_store_byte v[16:17], v85, off offset:93
+; ALIGNED-NEXT:    global_store_byte v[16:17], v36, off offset:91
+; ALIGNED-NEXT:    global_store_byte v[16:17], v84, off offset:89
+; ALIGNED-NEXT:    global_store_byte v[16:17], v102, off offset:87
+; ALIGNED-NEXT:    global_store_byte v[16:17], v83, off offset:85
+; ALIGNED-NEXT:    global_store_byte v[16:17], v35, off offset:83
+; ALIGNED-NEXT:    global_store_byte v[16:17], v82, off offset:81
+; ALIGNED-NEXT:    global_store_byte v[16:17], v23, off offset:79
+; ALIGNED-NEXT:    global_store_byte v[16:17], v97, off offset:77
+; ALIGNED-NEXT:    global_store_byte v[16:17], v34, off offset:75
+; ALIGNED-NEXT:    global_store_byte v[16:17], v96, off offset:73
+; ALIGNED-NEXT:    global_store_byte v[16:17], v21, off offset:71
+; ALIGNED-NEXT:    global_store_byte v[16:17], v87, off offset:69
+; ALIGNED-NEXT:    global_store_byte v[16:17], v49, off offset:67
+; ALIGNED-NEXT:    global_store_byte v[16:17], v86, off offset:65
+; ALIGNED-NEXT:    global_store_byte v[16:17], v22, off offset:63
+; ALIGNED-NEXT:    global_store_byte v[16:17], v101, off offset:61
+; ALIGNED-NEXT:    global_store_byte v[16:17], v48, off offset:59
+; ALIGNED-NEXT:    global_store_byte v[16:17], v100, off offset:57
+; ALIGNED-NEXT:    global_store_byte v[16:17], v103, off offset:55
+; ALIGNED-NEXT:    global_store_byte v[16:17], v99, off offset:53
+; ALIGNED-NEXT:    global_store_byte v[16:17], v39, off offset:51
+; ALIGNED-NEXT:    global_store_byte v[16:17], v98, off offset:49
+; ALIGNED-NEXT:    global_store_byte v[16:17], v29, off offset:43
+; ALIGNED-NEXT:    global_store_byte v[16:17], v15, off offset:41
+; ALIGNED-NEXT:    global_store_byte v[16:17], v38, off offset:47
+; ALIGNED-NEXT:    global_store_byte v[16:17], v14, off offset:45
+; ALIGNED-NEXT:    global_store_byte v[16:17], v20, off offset:35
+; ALIGNED-NEXT:    global_store_byte v[16:17], v13, off offset:33
+; ALIGNED-NEXT:    global_store_byte v[16:17], v53, off offset:39
+; ALIGNED-NEXT:    global_store_byte v[16:17], v12, off offset:37
+; ALIGNED-NEXT:    global_store_byte v[16:17], v28, off offset:31
+; ALIGNED-NEXT:    global_store_byte v[16:17], v11, off offset:29
+; ALIGNED-NEXT:    global_store_byte v[16:17], v52, off offset:27
+; ALIGNED-NEXT:    global_store_byte v[16:17], v10, off offset:25
+; ALIGNED-NEXT:    global_store_byte v[16:17], v112, off offset:23
+; ALIGNED-NEXT:    global_store_byte v[16:17], v9, off offset:21
+; ALIGNED-NEXT:    global_store_byte v[16:17], v51, off offset:19
+; ALIGNED-NEXT:    global_store_byte v[16:17], v8, off offset:17
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v7, off offset:14
-; ALIGNED-NEXT:    global_store_byte v[0:1], v7, off offset:12
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v7, off offset:14
+; ALIGNED-NEXT:    global_store_byte v[16:17], v7, off offset:12
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v6, off offset:10
-; ALIGNED-NEXT:    global_store_byte v[0:1], v6, off offset:8
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v6, off offset:10
+; ALIGNED-NEXT:    global_store_byte v[16:17], v6, off offset:8
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v5, off offset:6
-; ALIGNED-NEXT:    global_store_byte v[0:1], v5, off offset:4
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v5, off offset:6
+; ALIGNED-NEXT:    global_store_byte v[16:17], v5, off offset:4
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[0:1], v4, off offset:2
-; ALIGNED-NEXT:    global_store_byte v[0:1], v4, off
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v4, off offset:2
+; ALIGNED-NEXT:    global_store_byte v[16:17], v4, off
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v8, 24, v7
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v7, 8, v7
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v9, 24, v6
@@ -7701,362 +7695,376 @@ define void @memmove_p1_p1_sz2048(ptr addrspace(1) align 1 %dst, ptr addrspace(1
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v5, 8, v5
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v11, 24, v4
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v4, 8, v4
-; ALIGNED-NEXT:    global_store_byte v[0:1], v8, off offset:15
-; ALIGNED-NEXT:    global_store_byte v[0:1], v7, off offset:13
-; ALIGNED-NEXT:    global_store_byte v[0:1], v9, off offset:11
-; ALIGNED-NEXT:    global_store_byte v[0:1], v6, off offset:9
-; ALIGNED-NEXT:    global_store_byte v[0:1], v10, off offset:7
-; ALIGNED-NEXT:    global_store_byte v[0:1], v5, off offset:5
-; ALIGNED-NEXT:    global_store_byte v[0:1], v11, off offset:3
-; ALIGNED-NEXT:    global_store_byte v[0:1], v4, off offset:1
-; ALIGNED-NEXT:    v_add_co_u32 v0, vcc_lo, 0x100, v0
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
+; ALIGNED-NEXT:    global_store_byte v[16:17], v8, off offset:15
+; ALIGNED-NEXT:    global_store_byte v[16:17], v7, off offset:13
+; ALIGNED-NEXT:    global_store_byte v[16:17], v9, off offset:11
+; ALIGNED-NEXT:    global_store_byte v[16:17], v6, off offset:9
+; ALIGNED-NEXT:    global_store_byte v[16:17], v10, off offset:7
+; ALIGNED-NEXT:    global_store_byte v[16:17], v5, off offset:5
+; ALIGNED-NEXT:    global_store_byte v[16:17], v11, off offset:3
+; ALIGNED-NEXT:    global_store_byte v[16:17], v4, off offset:1
 ; ALIGNED-NEXT:    s_cbranch_scc1 .LBB6_2
-; ALIGNED-NEXT:  .LBB6_3: ; %Flow17
-; ALIGNED-NEXT:    s_andn2_saveexec_b32 s6, s6
+; ALIGNED-NEXT:  .LBB6_3: ; %Flow7
+; ALIGNED-NEXT:    s_andn2_saveexec_b32 s8, s6
 ; ALIGNED-NEXT:    s_cbranch_execz .LBB6_6
 ; ALIGNED-NEXT:  ; %bb.4: ; %memmove_bwd_loop.preheader
-; ALIGNED-NEXT:    v_add_co_u32 v12, vcc_lo, 0x700, v0
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v13, null, 0, v1, vcc_lo
-; ALIGNED-NEXT:    v_add_co_u32 v14, vcc_lo, 0x700, v2
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v15, null, 0, v3, vcc_lo
-; ALIGNED-NEXT:    s_movk_i32 s4, 0xf800
-; ALIGNED-NEXT:    s_mov_b32 s5, -1
+; ALIGNED-NEXT:    s_mov_b64 s[4:5], 0x700
+; ALIGNED-NEXT:    s_movk_i32 s6, 0xff00
+; ALIGNED-NEXT:    s_mov_b32 s7, -1
 ; ALIGNED-NEXT:  .LBB6_5: ; %memmove_bwd_loop
 ; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=1
+; ALIGNED-NEXT:    v_add_co_u32 v24, vcc_lo, v2, s4
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v25, null, s5, v3, vcc_lo
 ; ALIGNED-NEXT:    s_clause 0xf
-; ALIGNED-NEXT:    global_load_dwordx4 v[16:19], v[14:15], off offset:240
-; ALIGNED-NEXT:    global_load_dwordx4 v[20:23], v[14:15], off offset:224
-; ALIGNED-NEXT:    global_load_dwordx4 v[0:3], v[14:15], off
-; ALIGNED-NEXT:    global_load_dwordx4 v[4:7], v[14:15], off offset:16
-; ALIGNED-NEXT:    global_load_dwordx4 v[8:11], v[14:15], off offset:32
-; ALIGNED-NEXT:    global_load_dwordx4 v[100:103], v[14:15], off offset:48
-; ALIGNED-NEXT:    global_load_dwordx4 v[112:115], v[14:15], off offset:64
-; ALIGNED-NEXT:    global_load_dwordx4 v[116:119], v[14:15], off offset:80
-; ALIGNED-NEXT:    global_load_dwordx4 v[40:43], v[14:15], off offset:96
-; ALIGNED-NEXT:    global_load_dwordx4 v[28:31], v[14:15], off offset:112
-; ALIGNED-NEXT:    global_load_dwordx4 v[34:37], v[14:15], off offset:128
-; ALIGNED-NEXT:    global_load_dwordx4 v[44:47], v[14:15], off offset:144
-; ALIGNED-NEXT:    global_load_dwordx4 v[54:57], v[14:15], off offset:160
-; ALIGNED-NEXT:    global_load_dwordx4 v[68:71], v[14:15], off offset:176
-; ALIGNED-NEXT:    global_load_dwordx4 v[82:85], v[14:15], off offset:192
-; ALIGNED-NEXT:    global_load_dwordx4 v[64:67], v[14:15], off offset:208
-; ALIGNED-NEXT:    v_add_co_u32 v14, vcc_lo, 0xffffff00, v14
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v15, null, -1, v15, vcc_lo
-; ALIGNED-NEXT:    s_add_u32 s4, s4, 0x100
-; ALIGNED-NEXT:    s_addc_u32 s5, s5, 0
+; ALIGNED-NEXT:    global_load_dwordx4 v[16:19], v[24:25], off offset:240
+; ALIGNED-NEXT:    global_load_dwordx4 v[20:23], v[24:25], off offset:224
+; ALIGNED-NEXT:    global_load_dwordx4 v[4:7], v[24:25], off
+; ALIGNED-NEXT:    global_load_dwordx4 v[8:11], v[24:25], off offset:16
+; ALIGNED-NEXT:    global_load_dwordx4 v[12:15], v[24:25], off offset:32
+; ALIGNED-NEXT:    global_load_dwordx4 v[112:115], v[24:25], off offset:48
+; ALIGNED-NEXT:    global_load_dwordx4 v[116:119], v[24:25], off offset:64
+; ALIGNED-NEXT:    global_load_dwordx4 v[40:43], v[24:25], off offset:80
+; ALIGNED-NEXT:    global_load_dwordx4 v[26:29], v[24:25], off offset:96
+; ALIGNED-NEXT:    global_load_dwordx4 v[32:35], v[24:25], off offset:112
+; ALIGNED-NEXT:    global_load_dwordx4 v[44:47], v[24:25], off offset:128
+; ALIGNED-NEXT:    global_load_dwordx4 v[52:55], v[24:25], off offset:144
+; ALIGNED-NEXT:    global_load_dwordx4 v[66:69], v[24:25], off offset:160
+; ALIGNED-NEXT:    global_load_dwordx4 v[81:84], v[24:25], off offset:176
+; ALIGNED-NEXT:    global_load_dwordx4 v[96:99], v[24:25], off offset:192
+; ALIGNED-NEXT:    global_load_dwordx4 v[100:103], v[24:25], off offset:208
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(15)
-; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:336
-; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:340
-; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:344
-; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:348
+; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:320
+; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:324
+; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:328
+; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:332
+; ALIGNED-NEXT:    s_clause 0x3
+; ALIGNED-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:332
+; ALIGNED-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:328
+; ALIGNED-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:324
+; ALIGNED-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:320
+; ALIGNED-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s4
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo
+; ALIGNED-NEXT:    s_add_u32 s4, s4, 0xffffff00
+; ALIGNED-NEXT:    s_addc_u32 s5, s5, -1
+; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v31, off offset:254
+; ALIGNED-NEXT:    global_store_byte v[16:17], v31, off offset:252
+; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v30, off offset:250
+; ALIGNED-NEXT:    global_store_byte v[16:17], v30, off offset:248
+; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v25, off offset:246
+; ALIGNED-NEXT:    global_store_byte v[16:17], v25, off offset:244
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v24, off offset:242
+; ALIGNED-NEXT:    global_store_byte v[16:17], v24, off offset:240
+; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:336
+; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:340
+; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:344
+; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:348
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:348
-; ALIGNED-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:344
-; ALIGNED-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:340
-; ALIGNED-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:336
-; ALIGNED-NEXT:    s_cmp_eq_u64 s[4:5], 0
+; ALIGNED-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:348
+; ALIGNED-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:344
+; ALIGNED-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:340
+; ALIGNED-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:336
+; ALIGNED-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v33, off offset:254
-; ALIGNED-NEXT:    global_store_byte v[12:13], v33, off offset:252
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v51, off offset:238
+; ALIGNED-NEXT:    global_store_byte v[16:17], v51, off offset:236
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v32, off offset:250
-; ALIGNED-NEXT:    global_store_byte v[12:13], v32, off offset:248
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v50, off offset:234
+; ALIGNED-NEXT:    global_store_byte v[16:17], v50, off offset:232
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v27, off offset:246
-; ALIGNED-NEXT:    global_store_byte v[12:13], v27, off offset:244
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v26, off offset:242
-; ALIGNED-NEXT:    global_store_byte v[12:13], v26, off offset:240
-; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:352
-; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:356
-; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:360
-; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:364
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v49, off offset:230
+; ALIGNED-NEXT:    global_store_byte v[16:17], v49, off offset:228
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v36, off offset:226
+; ALIGNED-NEXT:    global_store_byte v[16:17], v36, off offset:224
+; ALIGNED-NEXT:    buffer_store_dword v100, off, s[0:3], s32 offset:288
+; ALIGNED-NEXT:    buffer_store_dword v101, off, s[0:3], s32 offset:292
+; ALIGNED-NEXT:    buffer_store_dword v102, off, s[0:3], s32 offset:296
+; ALIGNED-NEXT:    buffer_store_dword v103, off, s[0:3], s32 offset:300
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:364
-; ALIGNED-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:360
-; ALIGNED-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:356
-; ALIGNED-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:352
+; ALIGNED-NEXT:    buffer_load_dword v71, off, s[0:3], s32 offset:300
+; ALIGNED-NEXT:    buffer_load_dword v70, off, s[0:3], s32 offset:296
+; ALIGNED-NEXT:    buffer_load_dword v65, off, s[0:3], s32 offset:292
+; ALIGNED-NEXT:    buffer_load_dword v64, off, s[0:3], s32 offset:288
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v31
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 8, v31
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v30
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 8, v30
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v53, off offset:238
-; ALIGNED-NEXT:    global_store_byte v[12:13], v53, off offset:236
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v71, off offset:222
+; ALIGNED-NEXT:    global_store_byte v[16:17], v71, off offset:220
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v52, off offset:234
-; ALIGNED-NEXT:    global_store_byte v[12:13], v52, off offset:232
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v70, off offset:218
+; ALIGNED-NEXT:    global_store_byte v[16:17], v70, off offset:216
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v51, off offset:230
-; ALIGNED-NEXT:    global_store_byte v[12:13], v51, off offset:228
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v50, off offset:226
-; ALIGNED-NEXT:    global_store_byte v[12:13], v50, off offset:224
-; ALIGNED-NEXT:    buffer_store_dword v64, off, s[0:3], s32 offset:304
-; ALIGNED-NEXT:    buffer_store_dword v65, off, s[0:3], s32 offset:308
-; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:312
-; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:316
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v65, off offset:214
+; ALIGNED-NEXT:    global_store_byte v[16:17], v65, off offset:212
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v64, off offset:210
+; ALIGNED-NEXT:    global_store_byte v[16:17], v64, off offset:208
+; ALIGNED-NEXT:    buffer_store_dword v96, off, s[0:3], s32 offset:304
+; ALIGNED-NEXT:    buffer_store_dword v97, off, s[0:3], s32 offset:308
+; ALIGNED-NEXT:    buffer_store_dword v98, off, s[0:3], s32 offset:312
+; ALIGNED-NEXT:    buffer_store_dword v99, off, s[0:3], s32 offset:316
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v81, off, s[0:3], s32 offset:316
-; ALIGNED-NEXT:    buffer_load_dword v80, off, s[0:3], s32 offset:312
-; ALIGNED-NEXT:    buffer_load_dword v67, off, s[0:3], s32 offset:308
-; ALIGNED-NEXT:    buffer_load_dword v66, off, s[0:3], s32 offset:304
+; ALIGNED-NEXT:    buffer_load_dword v87, off, s[0:3], s32 offset:316
+; ALIGNED-NEXT:    buffer_load_dword v86, off, s[0:3], s32 offset:312
+; ALIGNED-NEXT:    buffer_load_dword v85, off, s[0:3], s32 offset:308
+; ALIGNED-NEXT:    buffer_load_dword v80, off, s[0:3], s32 offset:304
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v81, off offset:222
-; ALIGNED-NEXT:    global_store_byte v[12:13], v81, off offset:220
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v87, off offset:206
+; ALIGNED-NEXT:    global_store_byte v[16:17], v87, off offset:204
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v80, off offset:218
-; ALIGNED-NEXT:    global_store_byte v[12:13], v80, off offset:216
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v86, off offset:202
+; ALIGNED-NEXT:    global_store_byte v[16:17], v86, off offset:200
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v67, off offset:214
-; ALIGNED-NEXT:    global_store_byte v[12:13], v67, off offset:212
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v66, off offset:210
-; ALIGNED-NEXT:    global_store_byte v[12:13], v66, off offset:208
-; ALIGNED-NEXT:    buffer_store_dword v82, off, s[0:3], s32 offset:320
-; ALIGNED-NEXT:    buffer_store_dword v83, off, s[0:3], s32 offset:324
-; ALIGNED-NEXT:    buffer_store_dword v84, off, s[0:3], s32 offset:328
-; ALIGNED-NEXT:    buffer_store_dword v85, off, s[0:3], s32 offset:332
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v85, off offset:198
+; ALIGNED-NEXT:    global_store_byte v[16:17], v85, off offset:196
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v80, off offset:194
+; ALIGNED-NEXT:    global_store_byte v[16:17], v80, off offset:192
+; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:384
+; ALIGNED-NEXT:    buffer_store_dword v82, off, s[0:3], s32 offset:388
+; ALIGNED-NEXT:    buffer_store_dword v83, off, s[0:3], s32 offset:392
+; ALIGNED-NEXT:    buffer_store_dword v84, off, s[0:3], s32 offset:396
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v98, off, s[0:3], s32 offset:332
-; ALIGNED-NEXT:    buffer_load_dword v87, off, s[0:3], s32 offset:328
-; ALIGNED-NEXT:    buffer_load_dword v86, off, s[0:3], s32 offset:324
-; ALIGNED-NEXT:    buffer_load_dword v82, off, s[0:3], s32 offset:320
+; ALIGNED-NEXT:    buffer_load_dword v101, off, s[0:3], s32 offset:396
+; ALIGNED-NEXT:    buffer_load_dword v99, off, s[0:3], s32 offset:392
+; ALIGNED-NEXT:    buffer_load_dword v96, off, s[0:3], s32 offset:388
+; ALIGNED-NEXT:    buffer_load_dword v81, off, s[0:3], s32 offset:384
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v98, off offset:206
-; ALIGNED-NEXT:    global_store_byte v[12:13], v98, off offset:204
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v101, off offset:190
+; ALIGNED-NEXT:    global_store_byte v[16:17], v101, off offset:188
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v87, off offset:202
-; ALIGNED-NEXT:    global_store_byte v[12:13], v87, off offset:200
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v99, off offset:186
+; ALIGNED-NEXT:    global_store_byte v[16:17], v99, off offset:184
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v86, off offset:198
-; ALIGNED-NEXT:    global_store_byte v[12:13], v86, off offset:196
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v96, off offset:182
+; ALIGNED-NEXT:    global_store_byte v[16:17], v96, off offset:180
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v82, off offset:194
-; ALIGNED-NEXT:    global_store_byte v[12:13], v82, off offset:192
-; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:400
-; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:404
-; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:408
-; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:412
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v81, off offset:178
+; ALIGNED-NEXT:    global_store_byte v[16:17], v81, off offset:176
+; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:400
+; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:404
+; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:408
+; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:412
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v99, off, s[0:3], s32 offset:412
-; ALIGNED-NEXT:    buffer_load_dword v96, off, s[0:3], s32 offset:408
-; ALIGNED-NEXT:    buffer_load_dword v83, off, s[0:3], s32 offset:404
-; ALIGNED-NEXT:    buffer_load_dword v68, off, s[0:3], s32 offset:400
+; ALIGNED-NEXT:    buffer_load_dword v100, off, s[0:3], s32 offset:412
+; ALIGNED-NEXT:    buffer_load_dword v97, off, s[0:3], s32 offset:408
+; ALIGNED-NEXT:    buffer_load_dword v82, off, s[0:3], s32 offset:404
+; ALIGNED-NEXT:    buffer_load_dword v66, off, s[0:3], s32 offset:400
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v99, off offset:190
-; ALIGNED-NEXT:    global_store_byte v[12:13], v99, off offset:188
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v100, off offset:174
+; ALIGNED-NEXT:    global_store_byte v[16:17], v100, off offset:172
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v96, off offset:186
-; ALIGNED-NEXT:    global_store_byte v[12:13], v96, off offset:184
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v97, off offset:170
+; ALIGNED-NEXT:    global_store_byte v[16:17], v97, off offset:168
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v83, off offset:182
-; ALIGNED-NEXT:    global_store_byte v[12:13], v83, off offset:180
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v68, off offset:178
-; ALIGNED-NEXT:    global_store_byte v[12:13], v68, off offset:176
-; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:416
-; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:420
-; ALIGNED-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:424
-; ALIGNED-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:428
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v82, off offset:166
+; ALIGNED-NEXT:    global_store_byte v[16:17], v82, off offset:164
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v66, off offset:162
+; ALIGNED-NEXT:    global_store_byte v[16:17], v66, off offset:160
+; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:352
+; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:356
+; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:360
+; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:364
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v97, off, s[0:3], s32 offset:428
-; ALIGNED-NEXT:    buffer_load_dword v84, off, s[0:3], s32 offset:424
-; ALIGNED-NEXT:    buffer_load_dword v69, off, s[0:3], s32 offset:420
-; ALIGNED-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:416
+; ALIGNED-NEXT:    buffer_load_dword v98, off, s[0:3], s32 offset:364
+; ALIGNED-NEXT:    buffer_load_dword v83, off, s[0:3], s32 offset:360
+; ALIGNED-NEXT:    buffer_load_dword v67, off, s[0:3], s32 offset:356
+; ALIGNED-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:352
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v97, off offset:174
-; ALIGNED-NEXT:    global_store_byte v[12:13], v97, off offset:172
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v98, off offset:158
+; ALIGNED-NEXT:    global_store_byte v[16:17], v98, off offset:156
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v84, off offset:170
-; ALIGNED-NEXT:    global_store_byte v[12:13], v84, off offset:168
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v83, off offset:154
+; ALIGNED-NEXT:    global_store_byte v[16:17], v83, off offset:152
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v69, off offset:166
-; ALIGNED-NEXT:    global_store_byte v[12:13], v69, off offset:164
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v67, off offset:150
+; ALIGNED-NEXT:    global_store_byte v[16:17], v67, off offset:148
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v54, off offset:162
-; ALIGNED-NEXT:    global_store_byte v[12:13], v54, off offset:160
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v52, off offset:146
+; ALIGNED-NEXT:    global_store_byte v[16:17], v52, off offset:144
 ; ALIGNED-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:368
 ; ALIGNED-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:372
 ; ALIGNED-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:376
 ; ALIGNED-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:380
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v85, off, s[0:3], s32 offset:380
-; ALIGNED-NEXT:    buffer_load_dword v70, off, s[0:3], s32 offset:376
-; ALIGNED-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:372
-; ALIGNED-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:368
+; ALIGNED-NEXT:    buffer_load_dword v84, off, s[0:3], s32 offset:380
+; ALIGNED-NEXT:    buffer_load_dword v68, off, s[0:3], s32 offset:376
+; ALIGNED-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:372
+; ALIGNED-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:368
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v85, off offset:158
-; ALIGNED-NEXT:    global_store_byte v[12:13], v85, off offset:156
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v84, off offset:142
+; ALIGNED-NEXT:    global_store_byte v[16:17], v84, off offset:140
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v70, off offset:154
-; ALIGNED-NEXT:    global_store_byte v[12:13], v70, off offset:152
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v68, off offset:138
+; ALIGNED-NEXT:    global_store_byte v[16:17], v68, off offset:136
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v55, off offset:150
-; ALIGNED-NEXT:    global_store_byte v[12:13], v55, off offset:148
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v38, off offset:146
-; ALIGNED-NEXT:    global_store_byte v[12:13], v38, off offset:144
-; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:384
-; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:388
-; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:392
-; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:396
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v53, off offset:134
+; ALIGNED-NEXT:    global_store_byte v[16:17], v53, off offset:132
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v37, off offset:130
+; ALIGNED-NEXT:    global_store_byte v[16:17], v37, off offset:128
+; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:448
+; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:452
+; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:456
+; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:460
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v71, off, s[0:3], s32 offset:396
-; ALIGNED-NEXT:    buffer_load_dword v64, off, s[0:3], s32 offset:392
-; ALIGNED-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:388
-; ALIGNED-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:384
+; ALIGNED-NEXT:    buffer_load_dword v69, off, s[0:3], s32 offset:460
+; ALIGNED-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:456
+; ALIGNED-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:452
+; ALIGNED-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:448
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v71, off offset:142
-; ALIGNED-NEXT:    global_store_byte v[12:13], v71, off offset:140
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v69, off offset:126
+; ALIGNED-NEXT:    global_store_byte v[16:17], v69, off offset:124
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v64, off offset:138
-; ALIGNED-NEXT:    global_store_byte v[12:13], v64, off offset:136
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v54, off offset:122
+; ALIGNED-NEXT:    global_store_byte v[16:17], v54, off offset:120
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v39, off offset:134
-; ALIGNED-NEXT:    global_store_byte v[12:13], v39, off offset:132
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v34, off offset:130
-; ALIGNED-NEXT:    global_store_byte v[12:13], v34, off offset:128
-; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:464
-; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:468
-; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:472
-; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:476
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v38, off offset:118
+; ALIGNED-NEXT:    global_store_byte v[16:17], v38, off offset:116
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v32, off offset:114
+; ALIGNED-NEXT:    global_store_byte v[16:17], v32, off offset:112
+; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:464
+; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:468
+; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:472
+; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:476
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v65, off, s[0:3], s32 offset:476
-; ALIGNED-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:472
-; ALIGNED-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:468
-; ALIGNED-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:464
+; ALIGNED-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:476
+; ALIGNED-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:472
+; ALIGNED-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:468
+; ALIGNED-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:464
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v65, off offset:126
-; ALIGNED-NEXT:    global_store_byte v[12:13], v65, off offset:124
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v55, off offset:110
+; ALIGNED-NEXT:    global_store_byte v[16:17], v55, off offset:108
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v48, off offset:122
-; ALIGNED-NEXT:    global_store_byte v[12:13], v48, off offset:120
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v39, off offset:106
+; ALIGNED-NEXT:    global_store_byte v[16:17], v39, off offset:104
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v35, off offset:118
-; ALIGNED-NEXT:    global_store_byte v[12:13], v35, off offset:116
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v28, off offset:114
-; ALIGNED-NEXT:    global_store_byte v[12:13], v28, off offset:112
-; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:480
-; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:484
-; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:488
-; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:492
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v33, off offset:102
+; ALIGNED-NEXT:    global_store_byte v[16:17], v33, off offset:100
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v26, off offset:98
+; ALIGNED-NEXT:    global_store_byte v[16:17], v26, off offset:96
+; ALIGNED-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:416
+; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:420
+; ALIGNED-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:424
+; ALIGNED-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:428
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:492
-; ALIGNED-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:488
-; ALIGNED-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:484
-; ALIGNED-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:480
+; ALIGNED-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:428
+; ALIGNED-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:424
+; ALIGNED-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:420
+; ALIGNED-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:416
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v49, off offset:110
-; ALIGNED-NEXT:    global_store_byte v[12:13], v49, off offset:108
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v48, off offset:94
+; ALIGNED-NEXT:    global_store_byte v[16:17], v48, off offset:92
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v36, off offset:106
-; ALIGNED-NEXT:    global_store_byte v[12:13], v36, off offset:104
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v34, off offset:90
+; ALIGNED-NEXT:    global_store_byte v[16:17], v34, off offset:88
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v29, off offset:102
-; ALIGNED-NEXT:    global_store_byte v[12:13], v29, off offset:100
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v27, off offset:86
+; ALIGNED-NEXT:    global_store_byte v[16:17], v27, off offset:84
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v22, off offset:98
-; ALIGNED-NEXT:    global_store_byte v[12:13], v22, off offset:96
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v21, off offset:82
+; ALIGNED-NEXT:    global_store_byte v[16:17], v21, off offset:80
 ; ALIGNED-NEXT:    buffer_store_dword v116, off, s[0:3], s32 offset:432
 ; ALIGNED-NEXT:    buffer_store_dword v117, off, s[0:3], s32 offset:436
 ; ALIGNED-NEXT:    buffer_store_dword v118, off, s[0:3], s32 offset:440
 ; ALIGNED-NEXT:    buffer_store_dword v119, off, s[0:3], s32 offset:444
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:444
-; ALIGNED-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:440
-; ALIGNED-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:436
+; ALIGNED-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:444
+; ALIGNED-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:440
+; ALIGNED-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:436
 ; ALIGNED-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:432
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v37, off offset:94
-; ALIGNED-NEXT:    global_store_byte v[12:13], v37, off offset:92
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v35, off offset:78
+; ALIGNED-NEXT:    global_store_byte v[16:17], v35, off offset:76
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v30, off offset:90
-; ALIGNED-NEXT:    global_store_byte v[12:13], v30, off offset:88
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v28, off offset:74
+; ALIGNED-NEXT:    global_store_byte v[16:17], v28, off offset:72
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v23, off offset:86
-; ALIGNED-NEXT:    global_store_byte v[12:13], v23, off offset:84
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v19, off offset:82
-; ALIGNED-NEXT:    global_store_byte v[12:13], v19, off offset:80
-; ALIGNED-NEXT:    buffer_store_dword v112, off, s[0:3], s32 offset:448
-; ALIGNED-NEXT:    buffer_store_dword v113, off, s[0:3], s32 offset:452
-; ALIGNED-NEXT:    buffer_store_dword v114, off, s[0:3], s32 offset:456
-; ALIGNED-NEXT:    buffer_store_dword v115, off, s[0:3], s32 offset:460
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v22, off offset:70
+; ALIGNED-NEXT:    global_store_byte v[16:17], v22, off offset:68
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v19, off offset:66
+; ALIGNED-NEXT:    global_store_byte v[16:17], v19, off offset:64
+; ALIGNED-NEXT:    buffer_store_dword v112, off, s[0:3], s32 offset:512
+; ALIGNED-NEXT:    buffer_store_dword v113, off, s[0:3], s32 offset:516
+; ALIGNED-NEXT:    buffer_store_dword v114, off, s[0:3], s32 offset:520
+; ALIGNED-NEXT:    buffer_store_dword v115, off, s[0:3], s32 offset:524
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:460
-; ALIGNED-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:456
-; ALIGNED-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:452
-; ALIGNED-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:448
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v53
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v53
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v52
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 8, v52
+; ALIGNED-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:524
+; ALIGNED-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:520
+; ALIGNED-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:516
+; ALIGNED-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:512
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v25
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 8, v25
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v24
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 8, v24
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v51
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v51
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v50
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 8, v50
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v31, off offset:78
-; ALIGNED-NEXT:    global_store_byte v[12:13], v31, off offset:76
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v29, off offset:62
+; ALIGNED-NEXT:    global_store_byte v[16:17], v29, off offset:60
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v24, off offset:74
-; ALIGNED-NEXT:    global_store_byte v[12:13], v24, off offset:72
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v23, off offset:58
+; ALIGNED-NEXT:    global_store_byte v[16:17], v23, off offset:56
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v20, off offset:70
-; ALIGNED-NEXT:    global_store_byte v[12:13], v20, off offset:68
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v17, off offset:66
-; ALIGNED-NEXT:    global_store_byte v[12:13], v17, off offset:64
-; ALIGNED-NEXT:    buffer_store_dword v100, off, s[0:3], s32 offset:528
-; ALIGNED-NEXT:    buffer_store_dword v101, off, s[0:3], s32 offset:532
-; ALIGNED-NEXT:    buffer_store_dword v102, off, s[0:3], s32 offset:536
-; ALIGNED-NEXT:    buffer_store_dword v103, off, s[0:3], s32 offset:540
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v20, off offset:54
+; ALIGNED-NEXT:    global_store_byte v[16:17], v20, off offset:52
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v18, off offset:50
+; ALIGNED-NEXT:    global_store_byte v[16:17], v18, off offset:48
+; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:528
+; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:532
+; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:536
+; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:540
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:540
-; ALIGNED-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:536
-; ALIGNED-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:532
-; ALIGNED-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:528
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v33
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 8, v33
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v32
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 8, v32
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v27
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 8, v27
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v26
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 8, v26
+; ALIGNED-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:536
+; ALIGNED-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:540
+; ALIGNED-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:528
+; ALIGNED-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:532
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v25, off offset:62
-; ALIGNED-NEXT:    global_store_byte v[12:13], v25, off offset:60
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v15, off offset:42
+; ALIGNED-NEXT:    global_store_byte v[16:17], v15, off offset:40
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v21, off offset:58
-; ALIGNED-NEXT:    global_store_byte v[12:13], v21, off offset:56
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v14, off offset:46
+; ALIGNED-NEXT:    global_store_byte v[16:17], v14, off offset:44
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v18, off offset:54
-; ALIGNED-NEXT:    global_store_byte v[12:13], v18, off offset:52
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v16, off offset:50
-; ALIGNED-NEXT:    global_store_byte v[12:13], v16, off offset:48
-; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:544
-; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:548
-; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:552
-; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:556
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v13, off offset:34
+; ALIGNED-NEXT:    global_store_byte v[16:17], v13, off offset:32
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v12, off offset:38
+; ALIGNED-NEXT:    global_store_byte v[16:17], v12, off offset:36
+; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:480
+; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:484
+; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:488
+; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:492
 ; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:552
-; ALIGNED-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:556
-; ALIGNED-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:544
-; ALIGNED-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:548
+; ALIGNED-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:492
+; ALIGNED-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:488
+; ALIGNED-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:484
+; ALIGNED-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:480
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v11, off offset:42
-; ALIGNED-NEXT:    global_store_byte v[12:13], v11, off offset:40
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v11, off offset:30
+; ALIGNED-NEXT:    global_store_byte v[16:17], v11, off offset:28
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v10, off offset:46
-; ALIGNED-NEXT:    global_store_byte v[12:13], v10, off offset:44
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v10, off offset:26
+; ALIGNED-NEXT:    global_store_byte v[16:17], v10, off offset:24
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v9, off offset:34
-; ALIGNED-NEXT:    global_store_byte v[12:13], v9, off offset:32
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v9, off offset:22
+; ALIGNED-NEXT:    global_store_byte v[16:17], v9, off offset:20
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v8, off offset:38
-; ALIGNED-NEXT:    global_store_byte v[12:13], v8, off offset:36
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v8, off offset:18
+; ALIGNED-NEXT:    global_store_byte v[16:17], v8, off offset:16
 ; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:496
 ; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:500
 ; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:504
@@ -8066,295 +8074,274 @@ define void @memmove_p1_p1_sz2048(ptr addrspace(1) align 1 %dst, ptr addrspace(1
 ; ALIGNED-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:504
 ; ALIGNED-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:500
 ; ALIGNED-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:496
-; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v7, off offset:30
-; ALIGNED-NEXT:    global_store_byte v[12:13], v7, off offset:28
-; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v6, off offset:26
-; ALIGNED-NEXT:    global_store_byte v[12:13], v6, off offset:24
-; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v5, off offset:22
-; ALIGNED-NEXT:    global_store_byte v[12:13], v5, off offset:20
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v4, off offset:18
-; ALIGNED-NEXT:    global_store_byte v[12:13], v4, off offset:16
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:516
-; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:520
-; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:524
-; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:524
-; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:520
-; ALIGNED-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:516
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:512
-; ALIGNED-NEXT:    global_store_byte v[12:13], v100, off offset:255
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v81
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v81
-; ALIGNED-NEXT:    global_store_byte v[12:13], v33, off offset:253
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 24, v80
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 8, v80
-; ALIGNED-NEXT:    global_store_byte v[12:13], v101, off offset:251
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v67
-; ALIGNED-NEXT:    global_store_byte v[12:13], v81, off offset:221
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 24, v70
+; ALIGNED-NEXT:    global_store_byte v[16:17], v112, off offset:247
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v65
+; ALIGNED-NEXT:    global_store_byte v[16:17], v102, off offset:255
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v49
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v49
+; ALIGNED-NEXT:    global_store_byte v[16:17], v31, off offset:253
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 24, v36
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 8, v36
+; ALIGNED-NEXT:    global_store_byte v[16:17], v103, off offset:251
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v71
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v71
+; ALIGNED-NEXT:    global_store_byte v[16:17], v30, off offset:249
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v70
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 8, v70
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v65
+; ALIGNED-NEXT:    global_store_byte v[16:17], v25, off offset:245
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 24, v64
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 8, v64
+; ALIGNED-NEXT:    global_store_byte v[16:17], v112, off offset:215
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v67
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v67
-; ALIGNED-NEXT:    global_store_byte v[12:13], v32, off offset:249
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 24, v66
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 8, v66
-; ALIGNED-NEXT:    global_store_byte v[12:13], v102, off offset:247
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v98
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 8, v98
-; ALIGNED-NEXT:    global_store_byte v[12:13], v27, off offset:245
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 24, v87
+; ALIGNED-NEXT:    global_store_byte v[16:17], v113, off offset:243
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v87
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 8, v87
-; ALIGNED-NEXT:    global_store_byte v[12:13], v103, off offset:243
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v86
+; ALIGNED-NEXT:    global_store_byte v[16:17], v24, off offset:241
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v86
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 8, v86
-; ALIGNED-NEXT:    global_store_byte v[12:13], v26, off offset:241
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 24, v82
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 8, v82
-; ALIGNED-NEXT:    global_store_byte v[12:13], v112, off offset:239
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v99
+; ALIGNED-NEXT:    global_store_byte v[16:17], v114, off offset:239
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v85
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 8, v85
+; ALIGNED-NEXT:    global_store_byte v[16:17], v51, off offset:237
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 24, v80
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 8, v80
+; ALIGNED-NEXT:    global_store_byte v[16:17], v115, off offset:235
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v101
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v101
+; ALIGNED-NEXT:    global_store_byte v[16:17], v50, off offset:233
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v99
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v99
-; ALIGNED-NEXT:    global_store_byte v[12:13], v53, off offset:237
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 24, v96
+; ALIGNED-NEXT:    global_store_byte v[16:17], v102, off offset:231
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v96
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 8, v96
-; ALIGNED-NEXT:    global_store_byte v[12:13], v113, off offset:235
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v83
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v83
-; ALIGNED-NEXT:    global_store_byte v[12:13], v52, off offset:233
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v68
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 8, v68
-; ALIGNED-NEXT:    global_store_byte v[12:13], v114, off offset:231
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v97
+; ALIGNED-NEXT:    global_store_byte v[16:17], v49, off offset:229
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 24, v81
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v81
+; ALIGNED-NEXT:    global_store_byte v[16:17], v31, off offset:227
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 24, v100
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 8, v100
+; ALIGNED-NEXT:    global_store_byte v[16:17], v36, off offset:225
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 24, v97
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 8, v97
-; ALIGNED-NEXT:    global_store_byte v[12:13], v51, off offset:229
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 24, v84
+; ALIGNED-NEXT:    global_store_byte v[16:17], v103, off offset:223
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v82
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 8, v82
+; ALIGNED-NEXT:    global_store_byte v[16:17], v71, off offset:221
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 24, v66
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 8, v66
+; ALIGNED-NEXT:    global_store_byte v[16:17], v30, off offset:219
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v98
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 8, v98
+; ALIGNED-NEXT:    global_store_byte v[16:17], v70, off offset:217
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v83
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v83
+; ALIGNED-NEXT:    global_store_byte v[16:17], v65, off offset:213
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 24, v52
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 8, v52
+; ALIGNED-NEXT:    global_store_byte v[16:17], v25, off offset:211
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 24, v84
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 8, v84
-; ALIGNED-NEXT:    global_store_byte v[12:13], v115, off offset:227
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v69
+; ALIGNED-NEXT:    global_store_byte v[16:17], v64, off offset:209
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v68
+; ALIGNED-NEXT:    global_store_byte v[16:17], v67, off offset:149
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 24, v8
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v8, 8, v8
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 8, v68
+; ALIGNED-NEXT:    global_store_byte v[16:17], v113, off offset:207
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v53
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v53
+; ALIGNED-NEXT:    global_store_byte v[16:17], v87, off offset:205
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 24, v37
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 8, v37
+; ALIGNED-NEXT:    global_store_byte v[16:17], v24, off offset:203
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v69
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v69
-; ALIGNED-NEXT:    global_store_byte v[12:13], v50, off offset:225
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v54
+; ALIGNED-NEXT:    global_store_byte v[16:17], v86, off offset:201
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 24, v54
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 8, v54
-; ALIGNED-NEXT:    global_store_byte v[12:13], v100, off offset:223
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v85
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 8, v85
-; ALIGNED-NEXT:    global_store_byte v[12:13], v33, off offset:219
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 24, v55
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v55
-; ALIGNED-NEXT:    global_store_byte v[12:13], v80, off offset:217
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v38
+; ALIGNED-NEXT:    global_store_byte v[16:17], v114, off offset:199
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v38
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 8, v38
-; ALIGNED-NEXT:    global_store_byte v[12:13], v101, off offset:215
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v71
-; ALIGNED-NEXT:    global_store_byte v[12:13], v70, off offset:153
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v4
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v4, 8, v4
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v71
-; ALIGNED-NEXT:    global_store_byte v[12:13], v67, off offset:213
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 24, v64
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 8, v64
-; ALIGNED-NEXT:    global_store_byte v[12:13], v32, off offset:211
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 24, v39
+; ALIGNED-NEXT:    global_store_byte v[16:17], v85, off offset:197
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 24, v32
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 8, v32
+; ALIGNED-NEXT:    global_store_byte v[16:17], v51, off offset:195
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 24, v55
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v55
+; ALIGNED-NEXT:    global_store_byte v[16:17], v80, off offset:193
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v39
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 8, v39
-; ALIGNED-NEXT:    global_store_byte v[12:13], v66, off offset:209
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v34
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 8, v34
-; ALIGNED-NEXT:    global_store_byte v[12:13], v102, off offset:207
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v65
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v65
-; ALIGNED-NEXT:    global_store_byte v[12:13], v98, off offset:205
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v48
+; ALIGNED-NEXT:    global_store_byte v[16:17], v115, off offset:191
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v33
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 8, v33
+; ALIGNED-NEXT:    global_store_byte v[16:17], v101, off offset:189
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v26
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 8, v26
+; ALIGNED-NEXT:    global_store_byte v[16:17], v50, off offset:187
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v48
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 8, v48
-; ALIGNED-NEXT:    global_store_byte v[12:13], v27, off offset:203
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 24, v35
+; ALIGNED-NEXT:    global_store_byte v[16:17], v99, off offset:185
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v34
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 8, v34
+; ALIGNED-NEXT:    global_store_byte v[16:17], v102, off offset:183
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v27
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 8, v27
+; ALIGNED-NEXT:    global_store_byte v[16:17], v96, off offset:181
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 24, v21
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v21, 8, v21
+; ALIGNED-NEXT:    global_store_byte v[16:17], v49, off offset:179
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 24, v35
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 8, v35
-; ALIGNED-NEXT:    global_store_byte v[12:13], v87, off offset:201
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 24, v28
+; ALIGNED-NEXT:    global_store_byte v[16:17], v81, off offset:177
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 24, v28
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 8, v28
-; ALIGNED-NEXT:    global_store_byte v[12:13], v103, off offset:199
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v49
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v49
-; ALIGNED-NEXT:    global_store_byte v[12:13], v86, off offset:197
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 24, v36
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 8, v36
-; ALIGNED-NEXT:    global_store_byte v[12:13], v26, off offset:195
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 24, v29
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 8, v29
-; ALIGNED-NEXT:    global_store_byte v[12:13], v82, off offset:193
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v22
+; ALIGNED-NEXT:    global_store_byte v[16:17], v31, off offset:175
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 24, v22
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v22, 8, v22
-; ALIGNED-NEXT:    global_store_byte v[12:13], v112, off offset:191
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v37
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 8, v37
-; ALIGNED-NEXT:    global_store_byte v[12:13], v99, off offset:189
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 24, v30
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 8, v30
-; ALIGNED-NEXT:    global_store_byte v[12:13], v53, off offset:187
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 24, v23
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v23, 8, v23
-; ALIGNED-NEXT:    global_store_byte v[12:13], v96, off offset:185
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v96, 24, v19
+; ALIGNED-NEXT:    global_store_byte v[16:17], v100, off offset:173
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v19
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v19, 8, v19
-; ALIGNED-NEXT:    global_store_byte v[12:13], v113, off offset:183
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v31
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 8, v31
-; ALIGNED-NEXT:    global_store_byte v[12:13], v83, off offset:181
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 24, v24
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 8, v24
-; ALIGNED-NEXT:    global_store_byte v[12:13], v52, off offset:179
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v20
+; ALIGNED-NEXT:    global_store_byte v[16:17], v36, off offset:171
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 24, v29
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 8, v29
+; ALIGNED-NEXT:    global_store_byte v[16:17], v97, off offset:169
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 24, v23
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v23, 8, v23
+; ALIGNED-NEXT:    global_store_byte v[16:17], v103, off offset:167
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v20
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v20, 8, v20
-; ALIGNED-NEXT:    global_store_byte v[12:13], v68, off offset:177
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v17
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v17, 8, v17
-; ALIGNED-NEXT:    global_store_byte v[12:13], v114, off offset:175
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v25
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 8, v25
-; ALIGNED-NEXT:    global_store_byte v[12:13], v97, off offset:173
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v97, 24, v21
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v21, 8, v21
-; ALIGNED-NEXT:    global_store_byte v[12:13], v51, off offset:171
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 24, v18
+; ALIGNED-NEXT:    global_store_byte v[16:17], v82, off offset:165
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v18
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v18, 8, v18
-; ALIGNED-NEXT:    global_store_byte v[12:13], v84, off offset:169
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 24, v16
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v16, 8, v16
-; ALIGNED-NEXT:    global_store_byte v[12:13], v115, off offset:167
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v11
+; ALIGNED-NEXT:    global_store_byte v[16:17], v71, off offset:163
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 24, v15
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v15, 8, v15
+; ALIGNED-NEXT:    global_store_byte v[16:17], v66, off offset:161
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v14
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v14, 8, v14
+; ALIGNED-NEXT:    global_store_byte v[16:17], v30, off offset:159
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v13
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v13, 8, v13
+; ALIGNED-NEXT:    global_store_byte v[16:17], v98, off offset:157
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v12
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v12, 8, v12
+; ALIGNED-NEXT:    global_store_byte v[16:17], v70, off offset:155
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v11
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v11, 8, v11
-; ALIGNED-NEXT:    global_store_byte v[12:13], v69, off offset:165
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 24, v10
+; ALIGNED-NEXT:    global_store_byte v[16:17], v83, off offset:153
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 24, v10
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v10, 8, v10
-; ALIGNED-NEXT:    global_store_byte v[12:13], v50, off offset:163
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v9
+; ALIGNED-NEXT:    global_store_byte v[16:17], v112, off offset:151
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 24, v9
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v9, 8, v9
-; ALIGNED-NEXT:    global_store_byte v[12:13], v54, off offset:161
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 24, v8
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v8, 8, v8
-; ALIGNED-NEXT:    global_store_byte v[12:13], v100, off offset:159
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v7
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v7, 8, v7
-; ALIGNED-NEXT:    global_store_byte v[12:13], v85, off offset:157
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 24, v6
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v6, 8, v6
-; ALIGNED-NEXT:    global_store_byte v[12:13], v81, off offset:155
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 24, v5
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v5, 8, v5
-; ALIGNED-NEXT:    global_store_byte v[12:13], v33, off offset:151
-; ALIGNED-NEXT:    global_store_byte v[12:13], v55, off offset:149
-; ALIGNED-NEXT:    global_store_byte v[12:13], v80, off offset:147
-; ALIGNED-NEXT:    global_store_byte v[12:13], v38, off offset:145
-; ALIGNED-NEXT:    global_store_byte v[12:13], v101, off offset:143
-; ALIGNED-NEXT:    global_store_byte v[12:13], v71, off offset:141
-; ALIGNED-NEXT:    global_store_byte v[12:13], v67, off offset:139
-; ALIGNED-NEXT:    global_store_byte v[12:13], v64, off offset:137
-; ALIGNED-NEXT:    global_store_byte v[12:13], v32, off offset:135
-; ALIGNED-NEXT:    global_store_byte v[12:13], v39, off offset:133
-; ALIGNED-NEXT:    global_store_byte v[12:13], v66, off offset:131
-; ALIGNED-NEXT:    global_store_byte v[12:13], v34, off offset:129
-; ALIGNED-NEXT:    global_store_byte v[12:13], v102, off offset:127
-; ALIGNED-NEXT:    global_store_byte v[12:13], v65, off offset:125
-; ALIGNED-NEXT:    global_store_byte v[12:13], v98, off offset:123
-; ALIGNED-NEXT:    global_store_byte v[12:13], v48, off offset:121
-; ALIGNED-NEXT:    global_store_byte v[12:13], v27, off offset:119
-; ALIGNED-NEXT:    global_store_byte v[12:13], v35, off offset:117
-; ALIGNED-NEXT:    global_store_byte v[12:13], v87, off offset:115
-; ALIGNED-NEXT:    global_store_byte v[12:13], v28, off offset:113
-; ALIGNED-NEXT:    global_store_byte v[12:13], v103, off offset:111
-; ALIGNED-NEXT:    global_store_byte v[12:13], v49, off offset:109
-; ALIGNED-NEXT:    global_store_byte v[12:13], v86, off offset:107
-; ALIGNED-NEXT:    global_store_byte v[12:13], v36, off offset:105
-; ALIGNED-NEXT:    global_store_byte v[12:13], v26, off offset:103
-; ALIGNED-NEXT:    global_store_byte v[12:13], v29, off offset:101
-; ALIGNED-NEXT:    global_store_byte v[12:13], v82, off offset:99
-; ALIGNED-NEXT:    global_store_byte v[12:13], v22, off offset:97
-; ALIGNED-NEXT:    global_store_byte v[12:13], v112, off offset:95
-; ALIGNED-NEXT:    global_store_byte v[12:13], v37, off offset:93
-; ALIGNED-NEXT:    global_store_byte v[12:13], v99, off offset:91
-; ALIGNED-NEXT:    global_store_byte v[12:13], v30, off offset:89
-; ALIGNED-NEXT:    global_store_byte v[12:13], v53, off offset:87
-; ALIGNED-NEXT:    global_store_byte v[12:13], v23, off offset:85
-; ALIGNED-NEXT:    global_store_byte v[12:13], v96, off offset:83
-; ALIGNED-NEXT:    global_store_byte v[12:13], v19, off offset:81
-; ALIGNED-NEXT:    global_store_byte v[12:13], v113, off offset:79
-; ALIGNED-NEXT:    global_store_byte v[12:13], v31, off offset:77
-; ALIGNED-NEXT:    global_store_byte v[12:13], v83, off offset:75
-; ALIGNED-NEXT:    global_store_byte v[12:13], v24, off offset:73
-; ALIGNED-NEXT:    global_store_byte v[12:13], v52, off offset:71
-; ALIGNED-NEXT:    global_store_byte v[12:13], v20, off offset:69
-; ALIGNED-NEXT:    global_store_byte v[12:13], v68, off offset:67
-; ALIGNED-NEXT:    global_store_byte v[12:13], v17, off offset:65
-; ALIGNED-NEXT:    global_store_byte v[12:13], v114, off offset:63
-; ALIGNED-NEXT:    global_store_byte v[12:13], v25, off offset:61
-; ALIGNED-NEXT:    global_store_byte v[12:13], v97, off offset:59
-; ALIGNED-NEXT:    global_store_byte v[12:13], v21, off offset:57
-; ALIGNED-NEXT:    global_store_byte v[12:13], v51, off offset:55
-; ALIGNED-NEXT:    global_store_byte v[12:13], v18, off offset:53
-; ALIGNED-NEXT:    global_store_byte v[12:13], v84, off offset:51
-; ALIGNED-NEXT:    global_store_byte v[12:13], v16, off offset:49
-; ALIGNED-NEXT:    global_store_byte v[12:13], v115, off offset:43
-; ALIGNED-NEXT:    global_store_byte v[12:13], v11, off offset:41
-; ALIGNED-NEXT:    global_store_byte v[12:13], v69, off offset:47
-; ALIGNED-NEXT:    global_store_byte v[12:13], v10, off offset:45
-; ALIGNED-NEXT:    global_store_byte v[12:13], v50, off offset:35
-; ALIGNED-NEXT:    global_store_byte v[12:13], v9, off offset:33
-; ALIGNED-NEXT:    global_store_byte v[12:13], v54, off offset:39
-; ALIGNED-NEXT:    global_store_byte v[12:13], v8, off offset:37
-; ALIGNED-NEXT:    global_store_byte v[12:13], v100, off offset:31
-; ALIGNED-NEXT:    global_store_byte v[12:13], v7, off offset:29
-; ALIGNED-NEXT:    global_store_byte v[12:13], v85, off offset:27
-; ALIGNED-NEXT:    global_store_byte v[12:13], v6, off offset:25
-; ALIGNED-NEXT:    global_store_byte v[12:13], v81, off offset:23
-; ALIGNED-NEXT:    global_store_byte v[12:13], v5, off offset:21
-; ALIGNED-NEXT:    global_store_byte v[12:13], v70, off offset:19
-; ALIGNED-NEXT:    global_store_byte v[12:13], v4, off offset:17
+; ALIGNED-NEXT:    global_store_byte v[16:17], v65, off offset:147
+; ALIGNED-NEXT:    global_store_byte v[16:17], v52, off offset:145
+; ALIGNED-NEXT:    global_store_byte v[16:17], v25, off offset:143
+; ALIGNED-NEXT:    global_store_byte v[16:17], v84, off offset:141
+; ALIGNED-NEXT:    global_store_byte v[16:17], v64, off offset:139
+; ALIGNED-NEXT:    global_store_byte v[16:17], v68, off offset:137
+; ALIGNED-NEXT:    global_store_byte v[16:17], v113, off offset:135
+; ALIGNED-NEXT:    global_store_byte v[16:17], v53, off offset:133
+; ALIGNED-NEXT:    global_store_byte v[16:17], v87, off offset:131
+; ALIGNED-NEXT:    global_store_byte v[16:17], v37, off offset:129
+; ALIGNED-NEXT:    global_store_byte v[16:17], v24, off offset:127
+; ALIGNED-NEXT:    global_store_byte v[16:17], v69, off offset:125
+; ALIGNED-NEXT:    global_store_byte v[16:17], v86, off offset:123
+; ALIGNED-NEXT:    global_store_byte v[16:17], v54, off offset:121
+; ALIGNED-NEXT:    global_store_byte v[16:17], v114, off offset:119
+; ALIGNED-NEXT:    global_store_byte v[16:17], v38, off offset:117
+; ALIGNED-NEXT:    global_store_byte v[16:17], v85, off offset:115
+; ALIGNED-NEXT:    global_store_byte v[16:17], v32, off offset:113
+; ALIGNED-NEXT:    global_store_byte v[16:17], v51, off offset:111
+; ALIGNED-NEXT:    global_store_byte v[16:17], v55, off offset:109
+; ALIGNED-NEXT:    global_store_byte v[16:17], v80, off offset:107
+; ALIGNED-NEXT:    global_store_byte v[16:17], v39, off offset:105
+; ALIGNED-NEXT:    global_store_byte v[16:17], v115, off offset:103
+; ALIGNED-NEXT:    global_store_byte v[16:17], v33, off offset:101
+; ALIGNED-NEXT:    global_store_byte v[16:17], v101, off offset:99
+; ALIGNED-NEXT:    global_store_byte v[16:17], v26, off offset:97
+; ALIGNED-NEXT:    global_store_byte v[16:17], v50, off offset:95
+; ALIGNED-NEXT:    global_store_byte v[16:17], v48, off offset:93
+; ALIGNED-NEXT:    global_store_byte v[16:17], v99, off offset:91
+; ALIGNED-NEXT:    global_store_byte v[16:17], v34, off offset:89
+; ALIGNED-NEXT:    global_store_byte v[16:17], v102, off offset:87
+; ALIGNED-NEXT:    global_store_byte v[16:17], v27, off offset:85
+; ALIGNED-NEXT:    global_store_byte v[16:17], v96, off offset:83
+; ALIGNED-NEXT:    global_store_byte v[16:17], v21, off offset:81
+; ALIGNED-NEXT:    global_store_byte v[16:17], v49, off offset:79
+; ALIGNED-NEXT:    global_store_byte v[16:17], v35, off offset:77
+; ALIGNED-NEXT:    global_store_byte v[16:17], v81, off offset:75
+; ALIGNED-NEXT:    global_store_byte v[16:17], v28, off offset:73
+; ALIGNED-NEXT:    global_store_byte v[16:17], v31, off offset:71
+; ALIGNED-NEXT:    global_store_byte v[16:17], v22, off offset:69
+; ALIGNED-NEXT:    global_store_byte v[16:17], v100, off offset:67
+; ALIGNED-NEXT:    global_store_byte v[16:17], v19, off offset:65
+; ALIGNED-NEXT:    global_store_byte v[16:17], v36, off offset:63
+; ALIGNED-NEXT:    global_store_byte v[16:17], v29, off offset:61
+; ALIGNED-NEXT:    global_store_byte v[16:17], v97, off offset:59
+; ALIGNED-NEXT:    global_store_byte v[16:17], v23, off offset:57
+; ALIGNED-NEXT:    global_store_byte v[16:17], v103, off offset:55
+; ALIGNED-NEXT:    global_store_byte v[16:17], v20, off offset:53
+; ALIGNED-NEXT:    global_store_byte v[16:17], v82, off offset:51
+; ALIGNED-NEXT:    global_store_byte v[16:17], v18, off offset:49
+; ALIGNED-NEXT:    global_store_byte v[16:17], v71, off offset:43
+; ALIGNED-NEXT:    global_store_byte v[16:17], v15, off offset:41
+; ALIGNED-NEXT:    global_store_byte v[16:17], v66, off offset:47
+; ALIGNED-NEXT:    global_store_byte v[16:17], v14, off offset:45
+; ALIGNED-NEXT:    global_store_byte v[16:17], v30, off offset:35
+; ALIGNED-NEXT:    global_store_byte v[16:17], v13, off offset:33
+; ALIGNED-NEXT:    global_store_byte v[16:17], v98, off offset:39
+; ALIGNED-NEXT:    global_store_byte v[16:17], v12, off offset:37
+; ALIGNED-NEXT:    global_store_byte v[16:17], v70, off offset:31
+; ALIGNED-NEXT:    global_store_byte v[16:17], v11, off offset:29
+; ALIGNED-NEXT:    global_store_byte v[16:17], v83, off offset:27
+; ALIGNED-NEXT:    global_store_byte v[16:17], v10, off offset:25
+; ALIGNED-NEXT:    global_store_byte v[16:17], v112, off offset:23
+; ALIGNED-NEXT:    global_store_byte v[16:17], v9, off offset:21
+; ALIGNED-NEXT:    global_store_byte v[16:17], v67, off offset:19
+; ALIGNED-NEXT:    global_store_byte v[16:17], v8, off offset:17
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v3, off offset:14
-; ALIGNED-NEXT:    global_store_byte v[12:13], v3, off offset:12
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v7, off offset:14
+; ALIGNED-NEXT:    global_store_byte v[16:17], v7, off offset:12
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v2, off offset:10
-; ALIGNED-NEXT:    global_store_byte v[12:13], v2, off offset:8
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v6, off offset:10
+; ALIGNED-NEXT:    global_store_byte v[16:17], v6, off offset:8
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v1, off offset:6
-; ALIGNED-NEXT:    global_store_byte v[12:13], v1, off offset:4
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    global_store_byte_d16_hi v[12:13], v0, off offset:2
-; ALIGNED-NEXT:    global_store_byte v[12:13], v0, off
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v4, 24, v3
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v3, 8, v3
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v5, 24, v2
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v2, 8, v2
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v6, 24, v1
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v1, 8, v1
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v7, 24, v0
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v0, 8, v0
-; ALIGNED-NEXT:    global_store_byte v[12:13], v4, off offset:15
-; ALIGNED-NEXT:    global_store_byte v[12:13], v3, off offset:13
-; ALIGNED-NEXT:    global_store_byte v[12:13], v5, off offset:11
-; ALIGNED-NEXT:    global_store_byte v[12:13], v2, off offset:9
-; ALIGNED-NEXT:    global_store_byte v[12:13], v6, off offset:7
-; ALIGNED-NEXT:    global_store_byte v[12:13], v1, off offset:5
-; ALIGNED-NEXT:    global_store_byte v[12:13], v7, off offset:3
-; ALIGNED-NEXT:    global_store_byte v[12:13], v0, off offset:1
-; ALIGNED-NEXT:    v_add_co_u32 v12, vcc_lo, 0xffffff00, v12
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v13, null, -1, v13, vcc_lo
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v5, off offset:6
+; ALIGNED-NEXT:    global_store_byte v[16:17], v5, off offset:4
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    global_store_byte_d16_hi v[16:17], v4, off offset:2
+; ALIGNED-NEXT:    global_store_byte v[16:17], v4, off
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v8, 24, v7
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v7, 8, v7
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v9, 24, v6
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v6, 8, v6
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v10, 24, v5
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v5, 8, v5
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v11, 24, v4
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v4, 8, v4
+; ALIGNED-NEXT:    global_store_byte v[16:17], v8, off offset:15
+; ALIGNED-NEXT:    global_store_byte v[16:17], v7, off offset:13
+; ALIGNED-NEXT:    global_store_byte v[16:17], v9, off offset:11
+; ALIGNED-NEXT:    global_store_byte v[16:17], v6, off offset:9
+; ALIGNED-NEXT:    global_store_byte v[16:17], v10, off offset:7
+; ALIGNED-NEXT:    global_store_byte v[16:17], v5, off offset:5
+; ALIGNED-NEXT:    global_store_byte v[16:17], v11, off offset:3
+; ALIGNED-NEXT:    global_store_byte v[16:17], v4, off offset:1
 ; ALIGNED-NEXT:    s_cbranch_scc0 .LBB6_5
-; ALIGNED-NEXT:  .LBB6_6: ; %Flow18
-; ALIGNED-NEXT:    s_or_b32 exec_lo, exec_lo, s6
-; ALIGNED-NEXT:    s_clause 0x9 ; 40-byte Folded Reload
-; ALIGNED-NEXT:    buffer_load_dword v57, off, s[0:3], s32
-; ALIGNED-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:4
-; ALIGNED-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:8
-; ALIGNED-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:12
-; ALIGNED-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:16
-; ALIGNED-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:20
-; ALIGNED-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:24
-; ALIGNED-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:28
-; ALIGNED-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:32
-; ALIGNED-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:36
+; ALIGNED-NEXT:  .LBB6_6: ; %Flow8
+; ALIGNED-NEXT:    s_or_b32 exec_lo, exec_lo, s8
+; ALIGNED-NEXT:    s_clause 0x7 ; 32-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v47, off, s[0:3], s32
+; ALIGNED-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:4
+; ALIGNED-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:8
+; ALIGNED-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:12
+; ALIGNED-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:16
+; ALIGNED-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:20
+; ALIGNED-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:24
+; ALIGNED-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:28
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    s_setpc_b64 s[30:31]
 ;
@@ -8366,31 +8353,27 @@ define void @memmove_p1_p1_sz2048(ptr addrspace(1) align 1 %dst, ptr addrspace(1
 ; UNROLL3-NEXT:    s_xor_b32 s6, exec_lo, s4
 ; UNROLL3-NEXT:    s_cbranch_execz .LBB6_4
 ; UNROLL3-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader
-; UNROLL3-NEXT:    v_mov_b32_e32 v5, v3
-; UNROLL3-NEXT:    v_mov_b32_e32 v7, v1
-; UNROLL3-NEXT:    v_mov_b32_e32 v4, v2
-; UNROLL3-NEXT:    v_mov_b32_e32 v6, v0
-; UNROLL3-NEXT:    s_mov_b64 s[4:5], 0x7e0
+; UNROLL3-NEXT:    s_mov_b64 s[4:5], 0
 ; UNROLL3-NEXT:    .p2align 6
 ; UNROLL3-NEXT:  .LBB6_2: ; %memmove_fwd_loop
 ; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; UNROLL3-NEXT:    v_add_co_u32 v12, vcc_lo, v2, s4
+; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v13, null, s5, v3, vcc_lo
+; UNROLL3-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s4
+; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo
 ; UNROLL3-NEXT:    s_clause 0x2
-; UNROLL3-NEXT:    global_load_dwordx4 v[8:11], v[4:5], off
-; UNROLL3-NEXT:    global_load_dwordx4 v[12:15], v[4:5], off offset:16
-; UNROLL3-NEXT:    global_load_dwordx4 v[16:19], v[4:5], off offset:32
-; UNROLL3-NEXT:    v_add_co_u32 v4, vcc_lo, v4, 48
-; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; UNROLL3-NEXT:    s_add_u32 s4, s4, 0xffffffd0
-; UNROLL3-NEXT:    s_addc_u32 s5, s5, -1
+; UNROLL3-NEXT:    global_load_dwordx4 v[4:7], v[12:13], off
+; UNROLL3-NEXT:    global_load_dwordx4 v[8:11], v[12:13], off offset:16
+; UNROLL3-NEXT:    global_load_dwordx4 v[12:15], v[12:13], off offset:32
+; UNROLL3-NEXT:    s_add_u32 s4, s4, 48
+; UNROLL3-NEXT:    s_addc_u32 s5, s5, 0
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(2)
-; UNROLL3-NEXT:    global_store_dwordx4 v[6:7], v[8:11], off
+; UNROLL3-NEXT:    global_store_dwordx4 v[16:17], v[4:7], off
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(1)
-; UNROLL3-NEXT:    global_store_dwordx4 v[6:7], v[12:15], off offset:16
+; UNROLL3-NEXT:    global_store_dwordx4 v[16:17], v[8:11], off offset:16
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(0)
-; UNROLL3-NEXT:    global_store_dwordx4 v[6:7], v[16:19], off offset:32
-; UNROLL3-NEXT:    v_add_co_u32 v6, vcc_lo, v6, 48
-; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; UNROLL3-NEXT:    s_cmp_lg_u64 s[4:5], 0
+; UNROLL3-NEXT:    global_store_dwordx4 v[16:17], v[12:15], off offset:32
+; UNROLL3-NEXT:    s_cmp_lg_u64 s[4:5], 0x7e0
 ; UNROLL3-NEXT:    s_cbranch_scc1 .LBB6_2
 ; UNROLL3-NEXT:  ; %bb.3: ; %memmove_fwd_residual
 ; UNROLL3-NEXT:    global_load_dwordx4 v[4:7], v[2:3], off offset:2016
@@ -8401,45 +8384,42 @@ define void @memmove_p1_p1_sz2048(ptr addrspace(1) align 1 %dst, ptr addrspace(1
 ; UNROLL3-NEXT:    global_store_dwordx4 v[0:1], v[2:5], off offset:2032
 ; UNROLL3-NEXT:    ; implicit-def: $vgpr2_vgpr3
 ; UNROLL3-NEXT:    ; implicit-def: $vgpr0_vgpr1
-; UNROLL3-NEXT:  .LBB6_4: ; %Flow15
-; UNROLL3-NEXT:    s_andn2_saveexec_b32 s6, s6
+; UNROLL3-NEXT:  .LBB6_4: ; %Flow5
+; UNROLL3-NEXT:    s_andn2_saveexec_b32 s8, s6
 ; UNROLL3-NEXT:    s_cbranch_execz .LBB6_7
 ; UNROLL3-NEXT:  ; %bb.5: ; %memmove_bwd_residual
 ; UNROLL3-NEXT:    global_load_dwordx4 v[4:7], v[2:3], off offset:2032
-; UNROLL3-NEXT:    s_movk_i32 s4, 0xf820
-; UNROLL3-NEXT:    s_mov_b32 s5, -1
+; UNROLL3-NEXT:    s_mov_b64 s[4:5], 0x7b0
+; UNROLL3-NEXT:    s_movk_i32 s6, 0xffd0
+; UNROLL3-NEXT:    s_mov_b32 s7, -1
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(0)
 ; UNROLL3-NEXT:    global_store_dwordx4 v[0:1], v[4:7], off offset:2032
-; UNROLL3-NEXT:    global_load_dwordx4 v[6:9], v[2:3], off offset:2016
-; UNROLL3-NEXT:    v_add_co_u32 v4, vcc_lo, 0x7b0, v0
-; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v5, null, 0, v1, vcc_lo
-; UNROLL3-NEXT:    v_add_co_u32 v2, vcc_lo, 0x7b0, v2
-; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
+; UNROLL3-NEXT:    global_load_dwordx4 v[4:7], v[2:3], off offset:2016
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(0)
-; UNROLL3-NEXT:    global_store_dwordx4 v[0:1], v[6:9], off offset:2016
+; UNROLL3-NEXT:    global_store_dwordx4 v[0:1], v[4:7], off offset:2016
 ; UNROLL3-NEXT:    .p2align 6
 ; UNROLL3-NEXT:  .LBB6_6: ; %memmove_bwd_loop
 ; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; UNROLL3-NEXT:    v_add_co_u32 v12, vcc_lo, v2, s4
+; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v13, null, s5, v3, vcc_lo
+; UNROLL3-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s4
+; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo
 ; UNROLL3-NEXT:    s_clause 0x2
-; UNROLL3-NEXT:    global_load_dwordx4 v[6:9], v[2:3], off
-; UNROLL3-NEXT:    global_load_dwordx4 v[10:13], v[2:3], off offset:16
-; UNROLL3-NEXT:    global_load_dwordx4 v[14:17], v[2:3], off offset:32
-; UNROLL3-NEXT:    v_add_co_u32 v2, vcc_lo, 0xffffffd0, v2
-; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v3, null, -1, v3, vcc_lo
-; UNROLL3-NEXT:    s_add_u32 s4, s4, 48
-; UNROLL3-NEXT:    s_addc_u32 s5, s5, 0
+; UNROLL3-NEXT:    global_load_dwordx4 v[4:7], v[12:13], off
+; UNROLL3-NEXT:    global_load_dwordx4 v[8:11], v[12:13], off offset:16
+; UNROLL3-NEXT:    global_load_dwordx4 v[12:15], v[12:13], off offset:32
+; UNROLL3-NEXT:    s_add_u32 s4, s4, 0xffffffd0
+; UNROLL3-NEXT:    s_addc_u32 s5, s5, -1
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(2)
-; UNROLL3-NEXT:    global_store_dwordx4 v[4:5], v[6:9], off
+; UNROLL3-NEXT:    global_store_dwordx4 v[16:17], v[4:7], off
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(1)
-; UNROLL3-NEXT:    global_store_dwordx4 v[4:5], v[10:13], off offset:16
+; UNROLL3-NEXT:    global_store_dwordx4 v[16:17], v[8:11], off offset:16
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(0)
-; UNROLL3-NEXT:    global_store_dwordx4 v[4:5], v[14:17], off offset:32
-; UNROLL3-NEXT:    v_add_co_u32 v4, vcc_lo, 0xffffffd0, v4
-; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v5, null, -1, v5, vcc_lo
-; UNROLL3-NEXT:    s_cmp_eq_u64 s[4:5], 0
+; UNROLL3-NEXT:    global_store_dwordx4 v[16:17], v[12:15], off offset:32
+; UNROLL3-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]
 ; UNROLL3-NEXT:    s_cbranch_scc0 .LBB6_6
-; UNROLL3-NEXT:  .LBB6_7: ; %Flow16
-; UNROLL3-NEXT:    s_or_b32 exec_lo, exec_lo, s6
+; UNROLL3-NEXT:  .LBB6_7: ; %Flow6
+; UNROLL3-NEXT:    s_or_b32 exec_lo, exec_lo, s8
 ; UNROLL3-NEXT:    s_setpc_b64 s[30:31]
 entry:
   tail call void @llvm.memmove.p1.p1.i64(ptr addrspace(1) noundef nonnull align 1 %dst, ptr addrspace(1) noundef nonnull align 1 %src, i64 2048, i1 false)
@@ -8455,139 +8435,138 @@ define void @memmove_p0_p4_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(4
 ; CHECK-NEXT:    s_xor_b32 s6, exec_lo, s4
 ; CHECK-NEXT:    s_cbranch_execz .LBB7_3
 ; CHECK-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader
-; CHECK-NEXT:    s_mov_b64 s[4:5], 0x800
+; CHECK-NEXT:    s_mov_b64 s[4:5], 0
 ; CHECK-NEXT:  .LBB7_2: ; %memmove_fwd_loop
 ; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    v_add_co_u32 v96, vcc_lo, v2, s4
+; CHECK-NEXT:    v_add_co_ci_u32_e64 v97, null, s5, v3, vcc_lo
+; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s4
+; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo
 ; CHECK-NEXT:    s_clause 0xf
-; CHECK-NEXT:    global_load_dwordx4 v[4:7], v[2:3], off offset:240
-; CHECK-NEXT:    global_load_dwordx4 v[8:11], v[2:3], off offset:224
-; CHECK-NEXT:    global_load_dwordx4 v[12:15], v[2:3], off offset:208
-; CHECK-NEXT:    global_load_dwordx4 v[16:19], v[2:3], off offset:192
-; CHECK-NEXT:    global_load_dwordx4 v[20:23], v[2:3], off offset:176
-; CHECK-NEXT:    global_load_dwordx4 v[24:27], v[2:3], off offset:160
-; CHECK-NEXT:    global_load_dwordx4 v[28:31], v[2:3], off offset:144
-; CHECK-NEXT:    global_load_dwordx4 v[32:35], v[2:3], off offset:128
-; CHECK-NEXT:    global_load_dwordx4 v[36:39], v[2:3], off offset:64
-; CHECK-NEXT:    global_load_dwordx4 v[48:51], v[2:3], off offset:32
-; CHECK-NEXT:    global_load_dwordx4 v[52:55], v[2:3], off offset:16
-; CHECK-NEXT:    global_load_dwordx4 v[64:67], v[2:3], off
-; CHECK-NEXT:    global_load_dwordx4 v[68:71], v[2:3], off offset:112
-; CHECK-NEXT:    global_load_dwordx4 v[80:83], v[2:3], off offset:96
-; CHECK-NEXT:    global_load_dwordx4 v[84:87], v[2:3], off offset:80
-; CHECK-NEXT:    global_load_dwordx4 v[96:99], v[2:3], off offset:48
-; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, 48
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, 0, v1, vcc_lo
-; CHECK-NEXT:    v_add_co_u32 v2, vcc_lo, 0x100, v2
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; CHECK-NEXT:    s_add_u32 s4, s4, 0xffffff00
-; CHECK-NEXT:    s_addc_u32 s5, s5, -1
+; CHECK-NEXT:    global_load_dwordx4 v[4:7], v[96:97], off offset:240
+; CHECK-NEXT:    global_load_dwordx4 v[8:11], v[96:97], off offset:224
+; CHECK-NEXT:    global_load_dwordx4 v[12:15], v[96:97], off offset:208
+; CHECK-NEXT:    global_load_dwordx4 v[16:19], v[96:97], off offset:192
+; CHECK-NEXT:    global_load_dwordx4 v[20:23], v[96:97], off offset:176
+; CHECK-NEXT:    global_load_dwordx4 v[24:27], v[96:97], off offset:160
+; CHECK-NEXT:    global_load_dwordx4 v[28:31], v[96:97], off offset:144
+; CHECK-NEXT:    global_load_dwordx4 v[32:35], v[96:97], off offset:128
+; CHECK-NEXT:    global_load_dwordx4 v[36:39], v[96:97], off offset:112
+; CHECK-NEXT:    global_load_dwordx4 v[48:51], v[96:97], off offset:96
+; CHECK-NEXT:    global_load_dwordx4 v[52:55], v[96:97], off offset:80
+; CHECK-NEXT:    global_load_dwordx4 v[64:67], v[96:97], off offset:64
+; CHECK-NEXT:    global_load_dwordx4 v[68:71], v[96:97], off offset:48
+; CHECK-NEXT:    global_load_dwordx4 v[80:83], v[96:97], off offset:32
+; CHECK-NEXT:    global_load_dwordx4 v[84:87], v[96:97], off offset:16
+; CHECK-NEXT:    global_load_dwordx4 v[96:99], v[96:97], off
+; CHECK-NEXT:    v_add_co_u32 v102, vcc_lo, v100, 48
+; CHECK-NEXT:    s_add_u32 s4, s4, 0x100
+; CHECK-NEXT:    v_add_co_ci_u32_e64 v103, null, 0, v101, vcc_lo
+; CHECK-NEXT:    s_addc_u32 s5, s5, 0
 ; CHECK-NEXT:    s_waitcnt vmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[4:7] offset:192
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[4:7] offset:192
 ; CHECK-NEXT:    s_waitcnt vmcnt(14)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[8:11] offset:176
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[8:11] offset:176
 ; CHECK-NEXT:    s_waitcnt vmcnt(13)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[12:15] offset:160
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[12:15] offset:160
 ; CHECK-NEXT:    s_waitcnt vmcnt(12)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[16:19] offset:144
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[16:19] offset:144
 ; CHECK-NEXT:    s_waitcnt vmcnt(11)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[20:23] offset:128
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[20:23] offset:128
 ; CHECK-NEXT:    s_waitcnt vmcnt(10)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[24:27] offset:112
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[24:27] offset:112
 ; CHECK-NEXT:    s_waitcnt vmcnt(9)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[28:31] offset:96
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[28:31] offset:96
 ; CHECK-NEXT:    s_waitcnt vmcnt(8)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[32:35] offset:128
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[32:35] offset:128
 ; CHECK-NEXT:    s_waitcnt vmcnt(7)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[36:39] offset:64
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[36:39] offset:64
 ; CHECK-NEXT:    s_waitcnt vmcnt(6)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[48:51] offset:32
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[48:51] offset:48
 ; CHECK-NEXT:    s_waitcnt vmcnt(5)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[52:55] offset:16
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[52:55] offset:32
 ; CHECK-NEXT:    s_waitcnt vmcnt(4)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[64:67]
-; CHECK-NEXT:    v_add_co_u32 v0, vcc_lo, 0x100, v0
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; CHECK-NEXT:    s_cmp_lg_u64 s[4:5], 0
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[64:67] offset:64
 ; CHECK-NEXT:    s_waitcnt vmcnt(3)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[68:71] offset:64
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[68:71]
 ; CHECK-NEXT:    s_waitcnt vmcnt(2)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[80:83] offset:48
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[80:83] offset:32
 ; CHECK-NEXT:    s_waitcnt vmcnt(1)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[84:87] offset:32
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[84:87] offset:16
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[96:99]
+; CHECK-NEXT:    s_cmp_lg_u64 s[4:5], 0x800
 ; CHECK-NEXT:    s_cbranch_scc1 .LBB7_2
-; CHECK-NEXT:  .LBB7_3: ; %Flow16
-; CHECK-NEXT:    s_andn2_saveexec_b32 s6, s6
+; CHECK-NEXT:  .LBB7_3: ; %Flow6
+; CHECK-NEXT:    s_andn2_saveexec_b32 s8, s6
 ; CHECK-NEXT:    s_cbranch_execz .LBB7_6
 ; CHECK-NEXT:  ; %bb.4: ; %memmove_bwd_loop.preheader
-; CHECK-NEXT:    v_add_co_u32 v0, vcc_lo, 0x700, v0
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; CHECK-NEXT:    s_movk_i32 s4, 0xf800
-; CHECK-NEXT:    s_mov_b32 s5, -1
+; CHECK-NEXT:    s_mov_b64 s[4:5], 0x700
+; CHECK-NEXT:    s_movk_i32 s6, 0xff00
+; CHECK-NEXT:    s_mov_b32 s7, -1
 ; CHECK-NEXT:  .LBB7_5: ; %memmove_bwd_loop
 ; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    v_add_co_u32 v96, vcc_lo, v2, s4
+; CHECK-NEXT:    v_add_co_ci_u32_e64 v97, null, s5, v3, vcc_lo
+; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s4
+; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo
 ; CHECK-NEXT:    s_clause 0xf
-; CHECK-NEXT:    global_load_dwordx4 v[4:7], v[2:3], off offset:2032
-; CHECK-NEXT:    global_load_dwordx4 v[8:11], v[2:3], off offset:2016
-; CHECK-NEXT:    global_load_dwordx4 v[12:15], v[2:3], off offset:2000
-; CHECK-NEXT:    global_load_dwordx4 v[16:19], v[2:3], off offset:1984
-; CHECK-NEXT:    global_load_dwordx4 v[20:23], v[2:3], off offset:1968
-; CHECK-NEXT:    global_load_dwordx4 v[24:27], v[2:3], off offset:1952
-; CHECK-NEXT:    global_load_dwordx4 v[28:31], v[2:3], off offset:1936
-; CHECK-NEXT:    global_load_dwordx4 v[32:35], v[2:3], off offset:1920
-; CHECK-NEXT:    global_load_dwordx4 v[36:39], v[2:3], off offset:1856
-; CHECK-NEXT:    global_load_dwordx4 v[48:51], v[2:3], off offset:1824
-; CHECK-NEXT:    global_load_dwordx4 v[52:55], v[2:3], off offset:1808
-; CHECK-NEXT:    global_load_dwordx4 v[64:67], v[2:3], off offset:1792
-; CHECK-NEXT:    global_load_dwordx4 v[68:71], v[2:3], off offset:1904
-; CHECK-NEXT:    global_load_dwordx4 v[80:83], v[2:3], off offset:1888
-; CHECK-NEXT:    global_load_dwordx4 v[84:87], v[2:3], off offset:1872
-; CHECK-NEXT:    global_load_dwordx4 v[96:99], v[2:3], off offset:1840
-; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, 48
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, 0, v1, vcc_lo
-; CHECK-NEXT:    v_add_co_u32 v2, vcc_lo, 0xffffff00, v2
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v3, null, -1, v3, vcc_lo
-; CHECK-NEXT:    s_add_u32 s4, s4, 0x100
-; CHECK-NEXT:    s_addc_u32 s5, s5, 0
+; CHECK-NEXT:    global_load_dwordx4 v[4:7], v[96:97], off offset:240
+; CHECK-NEXT:    global_load_dwordx4 v[8:11], v[96:97], off offset:224
+; CHECK-NEXT:    global_load_dwordx4 v[12:15], v[96:97], off offset:208
+; CHECK-NEXT:    global_load_dwordx4 v[16:19], v[96:97], off offset:192
+; CHECK-NEXT:    global_load_dwordx4 v[20:23], v[96:97], off offset:176
+; CHECK-NEXT:    global_load_dwordx4 v[24:27], v[96:97], off offset:160
+; CHECK-NEXT:    global_load_dwordx4 v[28:31], v[96:97], off offset:144
+; CHECK-NEXT:    global_load_dwordx4 v[32:35], v[96:97], off offset:128
+; CHECK-NEXT:    global_load_dwordx4 v[36:39], v[96:97], off offset:112
+; CHECK-NEXT:    global_load_dwordx4 v[48:51], v[96:97], off offset:96
+; CHECK-NEXT:    global_load_dwordx4 v[52:55], v[96:97], off offset:80
+; CHECK-NEXT:    global_load_dwordx4 v[64:67], v[96:97], off offset:64
+; CHECK-NEXT:    global_load_dwordx4 v[68:71], v[96:97], off offset:48
+; CHECK-NEXT:    global_load_dwordx4 v[80:83], v[96:97], off offset:32
+; CHECK-NEXT:    global_load_dwordx4 v[84:87], v[96:97], off offset:16
+; CHECK-NEXT:    global_load_dwordx4 v[96:99], v[96:97], off
+; CHECK-NEXT:    v_add_co_u32 v102, vcc_lo, v100, 48
+; CHECK-NEXT:    s_add_u32 s4, s4, 0xffffff00
+; CHECK-NEXT:    v_add_co_ci_u32_e64 v103, null, 0, v101, vcc_lo
+; CHECK-NEXT:    s_addc_u32 s5, s5, -1
 ; CHECK-NEXT:    s_waitcnt vmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[4:7] offset:192
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[4:7] offset:192
 ; CHECK-NEXT:    s_waitcnt vmcnt(14)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[8:11] offset:176
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[8:11] offset:176
 ; CHECK-NEXT:    s_waitcnt vmcnt(13)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[12:15] offset:160
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[12:15] offset:160
 ; CHECK-NEXT:    s_waitcnt vmcnt(12)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[16:19] offset:144
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[16:19] offset:144
 ; CHECK-NEXT:    s_waitcnt vmcnt(11)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[20:23] offset:128
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[20:23] offset:128
 ; CHECK-NEXT:    s_waitcnt vmcnt(10)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[24:27] offset:112
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[24:27] offset:112
 ; CHECK-NEXT:    s_waitcnt vmcnt(9)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[28:31] offset:96
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[28:31] offset:96
 ; CHECK-NEXT:    s_waitcnt vmcnt(8)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[32:35] offset:128
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[32:35] offset:128
 ; CHECK-NEXT:    s_waitcnt vmcnt(7)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[36:39] offset:64
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[36:39] offset:64
 ; CHECK-NEXT:    s_waitcnt vmcnt(6)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[48:51] offset:32
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[48:51] offset:48
 ; CHECK-NEXT:    s_waitcnt vmcnt(5)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[52:55] offset:16
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[52:55] offset:32
 ; CHECK-NEXT:    s_waitcnt vmcnt(4)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[64:67]
-; CHECK-NEXT:    v_add_co_u32 v0, vcc_lo, 0xffffff00, v0
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo
-; CHECK-NEXT:    s_cmp_eq_u64 s[4:5], 0
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[64:67] offset:64
 ; CHECK-NEXT:    s_waitcnt vmcnt(3)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[68:71] offset:64
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[68:71]
 ; CHECK-NEXT:    s_waitcnt vmcnt(2)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[80:83] offset:48
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[80:83] offset:32
 ; CHECK-NEXT:    s_waitcnt vmcnt(1)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[84:87] offset:32
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[84:87] offset:16
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[96:99]
+; CHECK-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]
 ; CHECK-NEXT:    s_cbranch_scc0 .LBB7_5
-; CHECK-NEXT:  .LBB7_6: ; %Flow17
-; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s6
+; CHECK-NEXT:  .LBB7_6: ; %Flow7
+; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s8
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
 ;
@@ -8599,1004 +8578,1005 @@ define void @memmove_p0_p4_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(4
 ; ALIGNED-NEXT:    s_xor_b32 s6, exec_lo, s4
 ; ALIGNED-NEXT:    s_cbranch_execz .LBB7_3
 ; ALIGNED-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader
-; ALIGNED-NEXT:    s_mov_b64 s[4:5], 0x800
+; ALIGNED-NEXT:    s_mov_b64 s[4:5], 0
 ; ALIGNED-NEXT:  .LBB7_2: ; %memmove_fwd_loop
 ; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=1
+; ALIGNED-NEXT:    v_add_co_u32 v4, vcc_lo, v2, s4
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v5, null, s5, v3, vcc_lo
+; ALIGNED-NEXT:    v_add_co_u32 v84, vcc_lo, v0, s4
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v85, null, s5, v1, vcc_lo
 ; ALIGNED-NEXT:    s_clause 0xf
-; ALIGNED-NEXT:    global_load_dwordx4 v[112:115], v[2:3], off offset:240
-; ALIGNED-NEXT:    global_load_dwordx4 v[84:87], v[2:3], off offset:224
-; ALIGNED-NEXT:    global_load_dwordx4 v[80:83], v[2:3], off offset:208
-; ALIGNED-NEXT:    global_load_dwordx4 v[68:71], v[2:3], off offset:192
-; ALIGNED-NEXT:    global_load_dwordx4 v[64:67], v[2:3], off offset:176
-; ALIGNED-NEXT:    global_load_dwordx4 v[52:55], v[2:3], off offset:160
-; ALIGNED-NEXT:    global_load_dwordx4 v[48:51], v[2:3], off offset:144
-; ALIGNED-NEXT:    global_load_dwordx4 v[36:39], v[2:3], off offset:128
-; ALIGNED-NEXT:    global_load_dwordx4 v[32:35], v[2:3], off offset:112
-; ALIGNED-NEXT:    global_load_dwordx4 v[28:31], v[2:3], off offset:96
-; ALIGNED-NEXT:    global_load_dwordx4 v[24:27], v[2:3], off offset:80
-; ALIGNED-NEXT:    global_load_dwordx4 v[20:23], v[2:3], off offset:64
-; ALIGNED-NEXT:    global_load_dwordx4 v[16:19], v[2:3], off offset:48
-; ALIGNED-NEXT:    global_load_dwordx4 v[12:15], v[2:3], off offset:32
-; ALIGNED-NEXT:    global_load_dwordx4 v[4:7], v[2:3], off
-; ALIGNED-NEXT:    global_load_dwordx4 v[8:11], v[2:3], off offset:16
-; ALIGNED-NEXT:    v_add_co_u32 v96, vcc_lo, v0, 6
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v97, null, 0, v1, vcc_lo
-; ALIGNED-NEXT:    v_add_co_u32 v98, vcc_lo, v0, 3
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v99, null, 0, v1, vcc_lo
-; ALIGNED-NEXT:    v_add_co_u32 v2, vcc_lo, 0x100, v2
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
-; ALIGNED-NEXT:    s_add_u32 s4, s4, 0xffffff00
-; ALIGNED-NEXT:    s_addc_u32 s5, s5, -1
+; ALIGNED-NEXT:    global_load_dwordx4 v[98:101], v[4:5], off offset:240
+; ALIGNED-NEXT:    global_load_dwordx4 v[112:115], v[4:5], off offset:224
+; ALIGNED-NEXT:    global_load_dwordx4 v[80:83], v[4:5], off offset:208
+; ALIGNED-NEXT:    global_load_dwordx4 v[68:71], v[4:5], off offset:192
+; ALIGNED-NEXT:    global_load_dwordx4 v[64:67], v[4:5], off offset:176
+; ALIGNED-NEXT:    global_load_dwordx4 v[52:55], v[4:5], off offset:160
+; ALIGNED-NEXT:    global_load_dwordx4 v[48:51], v[4:5], off offset:144
+; ALIGNED-NEXT:    global_load_dwordx4 v[36:39], v[4:5], off offset:128
+; ALIGNED-NEXT:    global_load_dwordx4 v[32:35], v[4:5], off offset:112
+; ALIGNED-NEXT:    global_load_dwordx4 v[28:31], v[4:5], off offset:96
+; ALIGNED-NEXT:    global_load_dwordx4 v[24:27], v[4:5], off offset:80
+; ALIGNED-NEXT:    global_load_dwordx4 v[20:23], v[4:5], off offset:64
+; ALIGNED-NEXT:    global_load_dwordx4 v[16:19], v[4:5], off offset:48
+; ALIGNED-NEXT:    global_load_dwordx4 v[12:15], v[4:5], off offset:32
+; ALIGNED-NEXT:    global_load_dwordx4 v[8:11], v[4:5], off offset:16
+; ALIGNED-NEXT:    global_load_dwordx4 v[4:7], v[4:5], off
+; ALIGNED-NEXT:    v_add_co_u32 v86, vcc_lo, v84, 6
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v87, null, 0, v85, vcc_lo
+; ALIGNED-NEXT:    v_add_co_u32 v96, vcc_lo, v84, 3
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v97, null, 0, v85, vcc_lo
+; ALIGNED-NEXT:    s_add_u32 s4, s4, 0x100
+; ALIGNED-NEXT:    s_addc_u32 s5, s5, 0
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(15)
-; ALIGNED-NEXT:    buffer_store_dword v114, off, s[0:3], s32 offset:168
-; ALIGNED-NEXT:    buffer_store_dword v115, off, s[0:3], s32 offset:172
-; ALIGNED-NEXT:    buffer_store_dword v113, off, s[0:3], s32 offset:164
-; ALIGNED-NEXT:    buffer_store_dword v112, off, s[0:3], s32 offset:160
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v114 offset:244
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v114
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v114
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v115 offset:248
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v115
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v115 offset:249
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v115
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v114 offset:245
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v113 offset:240
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v113 offset:241
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v112 offset:236
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v112 offset:237
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v113
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 8, v113
+; ALIGNED-NEXT:    buffer_store_dword v100, off, s[0:3], s32 offset:168
+; ALIGNED-NEXT:    buffer_store_dword v101, off, s[0:3], s32 offset:172
+; ALIGNED-NEXT:    buffer_store_dword v99, off, s[0:3], s32 offset:164
+; ALIGNED-NEXT:    buffer_store_dword v98, off, s[0:3], s32 offset:160
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v100 offset:244
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v100
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v100
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v101 offset:248
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v101 offset:249
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v100 offset:245
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v99 offset:240
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v99 offset:241
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v98 offset:236
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v98 offset:237
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v101
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v101
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v116, 24, v99
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v117, 8, v99
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v118, 24, v98
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v119, 8, v98
+; ALIGNED-NEXT:    s_waitcnt vmcnt(14)
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v114
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v114
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v102 offset:248
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v103 offset:246
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v100 offset:252
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v101 offset:250
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v116 offset:244
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v115
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v117 offset:242
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v115
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v118 offset:240
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v113
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v119 offset:238
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v113
+; ALIGNED-NEXT:    buffer_store_dword v114, off, s[0:3], s32 offset:184
+; ALIGNED-NEXT:    buffer_store_dword v115, off, s[0:3], s32 offset:188
+; ALIGNED-NEXT:    buffer_store_dword v113, off, s[0:3], s32 offset:180
+; ALIGNED-NEXT:    buffer_store_dword v112, off, s[0:3], s32 offset:176
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v114 offset:228
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v115 offset:232
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v115 offset:233
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v114 offset:229
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v113 offset:224
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v113 offset:225
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v112 offset:220
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v112 offset:221
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v112
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v112
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v100 offset:248
-; ALIGNED-NEXT:    s_waitcnt vmcnt(14)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v86
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v101 offset:246
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v86
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v102 offset:252
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v87
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v103 offset:250
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v87
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v115 offset:244
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v85
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v114 offset:242
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 8, v85
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v113 offset:240
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v84
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v112 offset:238
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v84
-; ALIGNED-NEXT:    buffer_store_dword v86, off, s[0:3], s32 offset:184
-; ALIGNED-NEXT:    buffer_store_dword v87, off, s[0:3], s32 offset:188
-; ALIGNED-NEXT:    buffer_store_dword v85, off, s[0:3], s32 offset:180
-; ALIGNED-NEXT:    buffer_store_dword v84, off, s[0:3], s32 offset:176
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v86 offset:228
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v87 offset:232
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v87 offset:233
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v86 offset:229
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v85 offset:224
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v85 offset:225
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v84 offset:220
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v84 offset:221
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(13)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 24, v82
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 8, v82
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 24, v83
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 8, v83
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v100 offset:232
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v82
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v82
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v98 offset:232
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v83
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v99 offset:230
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v83
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v100 offset:236
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v81
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v101 offset:230
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v101 offset:234
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v81
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v102 offset:236
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v102 offset:228
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v80
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v103 offset:234
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v103 offset:226
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v80
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v115 offset:228
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v113 offset:224
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(12)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v70
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v114 offset:226
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 8, v70
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v113 offset:224
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v71
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v112 offset:222
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v71
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v70
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v112 offset:222
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v70
 ; ALIGNED-NEXT:    buffer_store_dword v82, off, s[0:3], s32 offset:136
 ; ALIGNED-NEXT:    buffer_store_dword v83, off, s[0:3], s32 offset:140
 ; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:132
 ; ALIGNED-NEXT:    buffer_store_dword v80, off, s[0:3], s32 offset:128
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v82 offset:212
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v83 offset:216
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v83 offset:217
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v82 offset:213
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v81 offset:208
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v81 offset:209
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v80 offset:204
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v80 offset:205
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v69
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v69
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v68
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v68
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v84 offset:216
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v82 offset:212
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v83 offset:216
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v83 offset:217
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v82 offset:213
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v81 offset:208
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v81 offset:209
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v80 offset:204
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v80 offset:205
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v71
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v71
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v69
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v69
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v114 offset:216
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v68
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v115 offset:214
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v68
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v98 offset:220
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(11)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 24, v66
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v85 offset:214
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 8, v66
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v86 offset:220
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 24, v67
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v87 offset:218
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 8, v67
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v100 offset:212
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v65
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v101 offset:210
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v65
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v102 offset:208
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v64
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v103 offset:206
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v64
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v66
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v99 offset:218
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v66
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v100 offset:212
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v67
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v101 offset:210
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v67
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v102 offset:208
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v65
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v103 offset:206
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v65
 ; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:152
 ; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:156
 ; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:148
 ; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:144
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v70 offset:196
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v71 offset:200
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v71 offset:201
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v70 offset:197
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v69 offset:192
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v69 offset:193
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v68 offset:188
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v68 offset:189
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v70 offset:196
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v71 offset:200
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v71 offset:201
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v70 offset:197
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v69 offset:192
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v69 offset:193
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v68 offset:188
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v68 offset:189
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v64
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v64
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(10)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v54
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v54
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v55
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v55
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v115 offset:200
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v52
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v114 offset:198
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 8, v52
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v113 offset:204
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v53
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v112 offset:202
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v53
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v80 offset:196
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v54
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v54
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v113 offset:200
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v55
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v112 offset:198
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v55
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v80 offset:204
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v52
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v81 offset:202
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v52
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v82 offset:196
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v53
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v83 offset:194
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v53
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v114 offset:192
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(9)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v50
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v81 offset:194
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v50
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v82 offset:192
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v51
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v83 offset:190
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v51
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v50
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v115 offset:190
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v50
 ; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:232
 ; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:236
 ; ALIGNED-NEXT:    buffer_store_dword v65, off, s[0:3], s32 offset:228
 ; ALIGNED-NEXT:    buffer_store_dword v64, off, s[0:3], s32 offset:224
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v66 offset:180
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v67 offset:184
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v67 offset:185
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v66 offset:181
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v65 offset:176
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v65 offset:177
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v64 offset:172
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v64 offset:173
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v49
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v49
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v48
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v48
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v84 offset:184
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v66 offset:180
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v67 offset:184
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v67 offset:185
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v66 offset:181
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v65 offset:176
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v65 offset:177
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v64 offset:172
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v64 offset:173
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v51
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v51
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v49
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v49
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v98 offset:184
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v48
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v99 offset:182
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v48
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v100 offset:188
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 24, v38
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v85 offset:182
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 8, v38
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v86 offset:188
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 24, v39
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v87 offset:186
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 8, v39
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v100 offset:180
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v37
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v101 offset:178
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v37
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v102 offset:176
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v36
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v103 offset:174
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v36
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v38
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v101 offset:186
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v38
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v102 offset:180
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v39
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v103 offset:178
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v39
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v68 offset:176
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v37
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v69 offset:174
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v37
 ; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:252
 ; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:248
 ; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:244
 ; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:240
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v54 offset:164
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v54 offset:165
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v55 offset:168
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v55 offset:169
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v52 offset:156
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v52 offset:157
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v53 offset:160
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v53 offset:161
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v54 offset:164
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v54 offset:165
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v55 offset:168
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v55 offset:169
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v52 offset:156
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v52 offset:157
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v53 offset:160
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v53 offset:161
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v36
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v36
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v34
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v34
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 24, v35
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v35
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v68 offset:168
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v33
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v69 offset:166
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v33
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v70 offset:172
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v32
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v71 offset:170
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v32
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v115 offset:160
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v114 offset:158
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v113 offset:164
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v112 offset:162
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 24, v34
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v34
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v70 offset:168
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v35
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v71 offset:166
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v35
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v113 offset:172
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v33
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v112 offset:170
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v33
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v80 offset:160
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v32
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v81 offset:158
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v32
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v82 offset:164
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v83 offset:162
 ; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:200
 ; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:204
 ; ALIGNED-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:196
 ; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:192
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v50 offset:148
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v51 offset:152
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v51 offset:153
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v50 offset:149
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v49 offset:144
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v49 offset:145
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v48 offset:140
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v48 offset:141
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v114 offset:152
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 24, v29
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v29
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v28
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v30
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 8, v30
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v31
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v31
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v28
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v80 offset:152
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v28
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v115 offset:150
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v28
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v30
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v30
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v50 offset:148
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v51 offset:152
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v51 offset:153
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v50 offset:149
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v49 offset:144
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v49 offset:145
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v48 offset:140
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v48 offset:141
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 24, v31
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v31
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v29
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v29
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v64 offset:156
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(5)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v26
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v81 offset:150
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v26
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v82 offset:156
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v27
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v83 offset:154
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v27
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v64 offset:148
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v65 offset:146
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v66 offset:144
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v67 offset:142
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v26
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v65 offset:154
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v26
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v66 offset:148
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v27
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v67 offset:146
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v27
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v98 offset:144
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v25
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v99 offset:142
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v25
 ; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:216
 ; ALIGNED-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:220
 ; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:212
 ; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:208
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v38 offset:132
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v39 offset:136
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v39 offset:137
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v38 offset:133
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v37 offset:128
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v37 offset:129
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v36 offset:124
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v36 offset:128
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v38 offset:132
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v39 offset:136
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v39 offset:137
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v38 offset:133
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v37 offset:128
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v37 offset:129
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v36 offset:124
+; ALIGNED-NEXT:    flat_store_byte v[84:85], v36 offset:128
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(4)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 24, v22
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v84 offset:136
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v85 offset:134
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v86 offset:140
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v87 offset:138
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v100 offset:132
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 24, v22
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v100 offset:136
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v101 offset:134
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v102 offset:140
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v103 offset:138
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v68 offset:132
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v69 offset:130
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v52 offset:128
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v18
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v101 offset:130
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v102 offset:128
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v103 offset:126
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v18
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v53 offset:126
 ; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:40
 ; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:44
 ; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:36
 ; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:32
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v34 offset:116
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v35 offset:120
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v35 offset:121
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v34 offset:117
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v33 offset:112
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v33 offset:113
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v32 offset:108
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v32 offset:109
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v52 offset:120
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v34 offset:116
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v35 offset:120
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v35 offset:121
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v34 offset:117
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v33 offset:112
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v33 offset:113
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v32 offset:108
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v32 offset:109
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v54 offset:120
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v55 offset:118
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v70 offset:124
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v14
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v53 offset:118
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v54 offset:124
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v55 offset:122
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v68 offset:116
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v69 offset:114
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v70 offset:112
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v71 offset:110
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v14
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v71 offset:122
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v113 offset:116
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v112 offset:114
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v80 offset:112
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v81 offset:110
 ; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:56
 ; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:60
 ; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:52
 ; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:48
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v30 offset:100
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v31 offset:104
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v31 offset:105
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v30 offset:101
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v29 offset:96
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v29 offset:97
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v28 offset:92
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v28 offset:93
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 24, v10
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v48 offset:100
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 24, v6
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 8, v22
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v18
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v14
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 8, v10
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v49 offset:98
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v6
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v25
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 24, v23
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v19
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 24, v15
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v11
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v50 offset:96
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v7
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v25
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v24
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v24
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 8, v23
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 24, v21
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 8, v21
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 24, v20
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 8, v20
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v19
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 24, v17
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 8, v17
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 24, v16
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 8, v16
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v15
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v13
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v13
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v12
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v12
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 8, v11
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v115 offset:104
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v9
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v114 offset:102
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 8, v9
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v113 offset:108
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v8
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v112 offset:106
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v8
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v51 offset:94
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v7
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v30 offset:100
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v31 offset:104
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v31 offset:105
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v30 offset:101
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v29 offset:96
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v29 offset:97
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v28 offset:92
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v28 offset:93
+; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v10
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v114 offset:96
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v6
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 8, v22
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v18
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v14
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 8, v10
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v115 offset:94
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v6
+; ALIGNED-NEXT:    s_cmp_lg_u64 s[4:5], 0x800
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v23
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 24, v19
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v15
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v82 offset:104
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v11
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v83 offset:102
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v48 offset:108
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v49 offset:106
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v50 offset:100
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v51 offset:98
 ; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:8
 ; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:12
 ; ALIGNED-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:4
 ; ALIGNED-NEXT:    buffer_store_dword v24, off, s[0:3], s32
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v80 offset:88
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v81 offset:86
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v82 offset:92
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v83 offset:90
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v4
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v26 offset:84
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v27 offset:88
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v27 offset:89
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v26 offset:85
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v64 offset:84
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v25 offset:80
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v65 offset:82
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v25 offset:81
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v66 offset:80
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v24 offset:76
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v67 offset:78
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v24 offset:77
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v64 offset:88
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v7
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 24, v24
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 8, v24
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v23
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v21
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v21
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v20
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v20
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 8, v19
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 24, v17
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 8, v17
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 24, v16
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v16
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v15
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v13
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v13
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 24, v12
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 8, v12
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v11
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 24, v9
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v9
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v8
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v8
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v65 offset:86
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v7
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v66 offset:92
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v5
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v67 offset:90
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v5
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v98 offset:84
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v4
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v99 offset:82
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v4
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v26 offset:84
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v27 offset:88
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v27 offset:89
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v26 offset:85
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v25 offset:80
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v25 offset:81
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v36 offset:80
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v24 offset:76
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v37 offset:78
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v24 offset:77
 ; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:24
 ; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:28
 ; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:20
 ; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:16
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v22 offset:68
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v36 offset:72
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v37 offset:70
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v23 offset:72
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v38 offset:76
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v23 offset:73
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v39 offset:74
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v22 offset:69
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v84 offset:68
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v21 offset:64
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v85 offset:66
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v21 offset:65
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v86 offset:64
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v20 offset:60
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v87 offset:62
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v20 offset:64
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v22 offset:68
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v38 offset:72
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v39 offset:70
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v23 offset:72
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v100 offset:76
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v23 offset:73
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v101 offset:74
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v22 offset:69
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v102 offset:68
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v21 offset:64
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v103 offset:66
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v21 offset:65
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v68 offset:64
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v20 offset:60
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v69 offset:62
+; ALIGNED-NEXT:    flat_store_byte v[84:85], v20 offset:64
 ; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:104
 ; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:108
 ; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:100
 ; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:96
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v18 offset:52
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v100 offset:56
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v101 offset:54
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v19 offset:56
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v102 offset:60
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v19 offset:57
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v103 offset:58
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v18 offset:53
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v32 offset:52
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v17 offset:48
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v33 offset:50
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v17 offset:49
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v34 offset:48
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v16 offset:44
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v35 offset:46
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v16 offset:45
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v18 offset:52
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v52 offset:56
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v53 offset:54
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v19 offset:56
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v32 offset:60
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v19 offset:57
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v33 offset:58
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v18 offset:53
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v34 offset:52
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v17 offset:48
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v35 offset:50
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v17 offset:49
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v54 offset:48
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v16 offset:44
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v55 offset:46
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v16 offset:45
 ; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:120
 ; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:124
 ; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:116
 ; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:112
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v14 offset:36
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v52 offset:40
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v53 offset:38
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v15 offset:40
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v54 offset:44
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v15 offset:41
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v55 offset:42
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v14 offset:37
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v68 offset:36
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v13 offset:32
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v69 offset:34
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v13 offset:33
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v70 offset:32
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v12 offset:28
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v71 offset:30
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v12 offset:32
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v14 offset:36
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v70 offset:40
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v71 offset:38
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v15 offset:40
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v113 offset:44
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v15 offset:41
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v112 offset:42
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v14 offset:37
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v80 offset:36
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v13 offset:32
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v81 offset:34
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v13 offset:33
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v28 offset:32
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v12 offset:28
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v29 offset:30
+; ALIGNED-NEXT:    flat_store_byte v[84:85], v12 offset:32
 ; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:72
 ; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:76
 ; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:68
 ; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:64
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v10 offset:20
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v28 offset:24
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v29 offset:22
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v11 offset:24
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v30 offset:28
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v11 offset:25
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v31 offset:26
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v10 offset:21
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v115 offset:20
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v9 offset:16
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v114 offset:18
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v9 offset:17
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v113 offset:16
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v8 offset:12
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v112 offset:14
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v8 offset:16
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v10 offset:20
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v30 offset:24
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v31 offset:22
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v11 offset:24
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v82 offset:28
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v11 offset:25
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v83 offset:26
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v10 offset:21
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v48 offset:20
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v9 offset:16
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v49 offset:18
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v9 offset:17
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v50 offset:16
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v8 offset:12
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v51 offset:14
+; ALIGNED-NEXT:    flat_store_byte v[84:85], v8 offset:16
 ; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:88
 ; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:92
 ; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:84
 ; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:80
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v6 offset:4
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v48 offset:8
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v49 offset:6
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v7 offset:8
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v50 offset:12
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v7 offset:9
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v51 offset:10
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v6 offset:8
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v5 offset:4
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[0:1], v4 offset:2
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v83 offset:1
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v4
-; ALIGNED-NEXT:    v_add_co_u32 v0, vcc_lo, 0x100, v0
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v5
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v5
-; ALIGNED-NEXT:    s_cmp_lg_u64 s[4:5], 0
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v4
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v80 offset:4
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v5
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v81 offset:2
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v82
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v6 offset:4
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v114 offset:8
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v115 offset:6
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v7 offset:8
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v64 offset:12
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v7 offset:9
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v65 offset:10
+; ALIGNED-NEXT:    flat_store_byte v[84:85], v6 offset:8
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v66 offset:4
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[86:87], v5
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v67 offset:2
+; ALIGNED-NEXT:    flat_store_byte v[84:85], v5 offset:4
+; ALIGNED-NEXT:    flat_store_byte v[96:97], v98
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[84:85], v4 offset:2
+; ALIGNED-NEXT:    flat_store_byte v[84:85], v99 offset:1
+; ALIGNED-NEXT:    flat_store_byte v[84:85], v4
 ; ALIGNED-NEXT:    s_cbranch_scc1 .LBB7_2
-; ALIGNED-NEXT:  .LBB7_3: ; %Flow16
-; ALIGNED-NEXT:    s_andn2_saveexec_b32 s6, s6
+; ALIGNED-NEXT:  .LBB7_3: ; %Flow6
+; ALIGNED-NEXT:    s_andn2_saveexec_b32 s8, s6
 ; ALIGNED-NEXT:    s_cbranch_execz .LBB7_6
 ; ALIGNED-NEXT:  ; %bb.4: ; %memmove_bwd_loop.preheader
-; ALIGNED-NEXT:    v_add_co_u32 v0, vcc_lo, 0x700, v0
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; ALIGNED-NEXT:    s_movk_i32 s4, 0xf800
-; ALIGNED-NEXT:    s_mov_b32 s5, -1
+; ALIGNED-NEXT:    s_mov_b64 s[4:5], 0x700
+; ALIGNED-NEXT:    s_movk_i32 s6, 0xff00
+; ALIGNED-NEXT:    s_mov_b32 s7, -1
 ; ALIGNED-NEXT:  .LBB7_5: ; %memmove_bwd_loop
 ; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=1
+; ALIGNED-NEXT:    v_add_co_u32 v4, vcc_lo, v2, s4
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v5, null, s5, v3, vcc_lo
+; ALIGNED-NEXT:    v_add_co_u32 v84, vcc_lo, v0, s4
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v85, null, s5, v1, vcc_lo
 ; ALIGNED-NEXT:    s_clause 0xf
-; ALIGNED-NEXT:    global_load_dwordx4 v[100:103], v[2:3], off offset:2032
-; ALIGNED-NEXT:    global_load_dwordx4 v[84:87], v[2:3], off offset:2016
-; ALIGNED-NEXT:    global_load_dwordx4 v[80:83], v[2:3], off offset:2000
-; ALIGNED-NEXT:    global_load_dwordx4 v[68:71], v[2:3], off offset:1984
-; ALIGNED-NEXT:    global_load_dwordx4 v[64:67], v[2:3], off offset:1968
-; ALIGNED-NEXT:    global_load_dwordx4 v[52:55], v[2:3], off offset:1952
-; ALIGNED-NEXT:    global_load_dwordx4 v[48:51], v[2:3], off offset:1936
-; ALIGNED-NEXT:    global_load_dwordx4 v[36:39], v[2:3], off offset:1920
-; ALIGNED-NEXT:    global_load_dwordx4 v[32:35], v[2:3], off offset:1904
-; ALIGNED-NEXT:    global_load_dwordx4 v[28:31], v[2:3], off offset:1888
-; ALIGNED-NEXT:    global_load_dwordx4 v[24:27], v[2:3], off offset:1872
-; ALIGNED-NEXT:    global_load_dwordx4 v[20:23], v[2:3], off offset:1856
-; ALIGNED-NEXT:    global_load_dwordx4 v[16:19], v[2:3], off offset:1840
-; ALIGNED-NEXT:    global_load_dwordx4 v[12:15], v[2:3], off offset:1824
-; ALIGNED-NEXT:    global_load_dwordx4 v[4:7], v[2:3], off offset:1792
-; ALIGNED-NEXT:    global_load_dwordx4 v[8:11], v[2:3], off offset:1808
-; ALIGNED-NEXT:    v_add_co_u32 v96, vcc_lo, v0, 6
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v97, null, 0, v1, vcc_lo
-; ALIGNED-NEXT:    v_add_co_u32 v98, vcc_lo, v0, 3
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v99, null, 0, v1, vcc_lo
-; ALIGNED-NEXT:    v_add_co_u32 v2, vcc_lo, 0xffffff00, v2
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v3, null, -1, v3, vcc_lo
-; ALIGNED-NEXT:    s_add_u32 s4, s4, 0x100
-; ALIGNED-NEXT:    s_addc_u32 s5, s5, 0
+; ALIGNED-NEXT:    global_load_dwordx4 v[98:101], v[4:5], off offset:240
+; ALIGNED-NEXT:    global_load_dwordx4 v[112:115], v[4:5], off offset:224
+; ALIGNED-NEXT:    global_load_dwordx4 v[80:83], v[4:5], off offset:208
+; ALIGNED-NEXT:    global_load_dwordx4 v[68:71], v[4:5], off offset:192
+; ALIGNED-NEXT:    global_load_dwordx4 v[64:67], v[4:5], off offset:176
+; ALIGNED-NEXT:    global_load_dwordx4 v[52:55], v[4:5], off offset:160
+; ALIGNED-NEXT:    global_load_dwordx4 v[48:51], v[4:5], off offset:144
+; ALIGNED-NEXT:    global_load_dwordx4 v[36:39], v[4:5], off offset:128
+; ALIGNED-NEXT:    global_load_dwordx4 v[32:35], v[4:5], off offset:112
+; ALIGNED-NEXT:    global_load_dwordx4 v[28:31], v[4:5], off offset:96
+; ALIGNED-NEXT:    global_load_dwordx4 v[24:27], v[4:5], off offset:80
+; ALIGNED-NEXT:    global_load_dwordx4 v[20:23], v[4:5], off offset:64
+; ALIGNED-NEXT:    global_load_dwordx4 v[16:19], v[4:5], off offset:48
+; ALIGNED-NEXT:    global_load_dwordx4 v[12:15], v[4:5], off offset:32
+; ALIGNED-NEXT:    global_load_dwordx4 v[8:11], v[4:5], off offset:16
+; ALIGNED-NEXT:    global_load_dwordx4 v[4:7], v[4:5], off
+; ALIGNED-NEXT:    v_add_co_u32 v96, vcc_lo, v84, 6
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v97, null, 0, v85, vcc_lo
+; ALIGNED-NEXT:    v_add_co_u32 v86, vcc_lo, v84, 3
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v87, null, 0, v85, vcc_lo
+; ALIGNED-NEXT:    s_add_u32 s4, s4, 0xffffff00
+; ALIGNED-NEXT:    s_addc_u32 s5, s5, -1
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(15)
-; ALIGNED-NEXT:    buffer_store_dword v102, off, s[0:3], s32 offset:296
-; ALIGNED-NEXT:    buffer_store_dword v103, off, s[0:3], s32 offset:300
-; ALIGNED-NEXT:    buffer_store_dword v101, off, s[0:3], s32 offset:292
-; ALIGNED-NEXT:    buffer_store_dword v100, off, s[0:3], s32 offset:288
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v102
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 8, v102
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v103
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v103
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v102 offset:244
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v103 offset:248
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v103 offset:249
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v102 offset:245
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v101 offset:240
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v101 offset:241
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v100 offset:236
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v100 offset:237
+; ALIGNED-NEXT:    buffer_store_dword v100, off, s[0:3], s32 offset:424
+; ALIGNED-NEXT:    buffer_store_dword v101, off, s[0:3], s32 offset:428
+; ALIGNED-NEXT:    buffer_store_dword v99, off, s[0:3], s32 offset:420
+; ALIGNED-NEXT:    buffer_store_dword v98, off, s[0:3], s32 offset:416
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v100 offset:244
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v101 offset:248
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v101 offset:249
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v100 offset:245
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v99 offset:240
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v99 offset:241
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v98 offset:236
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v98 offset:237
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v100
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v101
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 8, v101
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v100
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v101
 ; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 8, v100
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v115 offset:248
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v116, 24, v99
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v117, 8, v99
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v118, 24, v98
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v119, 8, v98
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v102 offset:248
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v100 offset:246
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v103 offset:252
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(14)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v86
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v114 offset:246
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 8, v86
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v113 offset:252
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v87
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v112 offset:250
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v87
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v103 offset:244
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v85
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v102 offset:242
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 8, v85
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v101 offset:240
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v84
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v100 offset:238
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 8, v84
-; ALIGNED-NEXT:    buffer_store_dword v86, off, s[0:3], s32 offset:312
-; ALIGNED-NEXT:    buffer_store_dword v87, off, s[0:3], s32 offset:316
-; ALIGNED-NEXT:    buffer_store_dword v85, off, s[0:3], s32 offset:308
-; ALIGNED-NEXT:    buffer_store_dword v84, off, s[0:3], s32 offset:304
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v86 offset:228
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v87 offset:232
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v87 offset:233
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v86 offset:229
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v85 offset:224
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v85 offset:225
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v84 offset:220
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v84 offset:221
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v114
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v101 offset:250
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v114
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v115
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v115
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v113
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v113
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v116 offset:244
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v117 offset:242
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v118 offset:240
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v119 offset:238
+; ALIGNED-NEXT:    buffer_store_dword v114, off, s[0:3], s32 offset:440
+; ALIGNED-NEXT:    buffer_store_dword v115, off, s[0:3], s32 offset:444
+; ALIGNED-NEXT:    buffer_store_dword v113, off, s[0:3], s32 offset:436
+; ALIGNED-NEXT:    buffer_store_dword v112, off, s[0:3], s32 offset:432
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v114 offset:228
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v115 offset:232
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v115 offset:233
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v114 offset:229
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v113 offset:224
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v113 offset:225
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v112 offset:220
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v112 offset:221
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v112
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v112
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(13)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 24, v82
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 8, v82
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 24, v83
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 8, v83
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v115 offset:232
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v81
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v114 offset:230
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 8, v81
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v113 offset:236
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v80
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v112 offset:234
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v80
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v103 offset:228
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v82
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v82
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v98 offset:232
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v83
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v99 offset:230
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v83
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v100 offset:236
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v81
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v101 offset:234
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v81
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v102 offset:228
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v80
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v103 offset:226
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v80
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v113 offset:224
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(12)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v70
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v102 offset:226
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 8, v70
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v101 offset:224
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v71
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v100 offset:222
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 8, v71
-; ALIGNED-NEXT:    buffer_store_dword v82, off, s[0:3], s32 offset:264
-; ALIGNED-NEXT:    buffer_store_dword v83, off, s[0:3], s32 offset:268
-; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:260
-; ALIGNED-NEXT:    buffer_store_dword v80, off, s[0:3], s32 offset:256
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v70
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v112 offset:222
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v70
+; ALIGNED-NEXT:    buffer_store_dword v82, off, s[0:3], s32 offset:392
+; ALIGNED-NEXT:    buffer_store_dword v83, off, s[0:3], s32 offset:396
+; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:388
+; ALIGNED-NEXT:    buffer_store_dword v80, off, s[0:3], s32 offset:384
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v82 offset:212
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v83 offset:216
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v83 offset:217
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v82 offset:213
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v83 offset:217
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v82 offset:213
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v81 offset:208
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v81 offset:209
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v81 offset:209
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v80 offset:204
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v80 offset:205
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v69
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v69
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v68
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v68
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v84 offset:216
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v80 offset:205
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v71
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v71
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v69
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v69
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v114 offset:216
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v68
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v115 offset:214
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v68
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v98 offset:220
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(11)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 24, v66
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v85 offset:214
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 8, v66
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v86 offset:220
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 24, v67
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v87 offset:218
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 8, v67
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v115 offset:212
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v65
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v114 offset:210
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 8, v65
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v113 offset:208
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v64
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v112 offset:206
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v64
-; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:280
-; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:284
-; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:276
-; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:272
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v66
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v99 offset:218
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v66
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v100 offset:212
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v67
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v101 offset:210
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v67
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v102 offset:208
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v65
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v103 offset:206
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v65
+; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:408
+; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:412
+; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:404
+; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:400
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v70 offset:196
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v71 offset:200
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v71 offset:201
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v70 offset:197
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v71 offset:201
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v70 offset:197
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v69 offset:192
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v69 offset:193
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v69 offset:193
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v68 offset:188
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v68 offset:189
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v68 offset:189
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v64
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v64
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(10)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v54
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v54
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v55
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v55
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v103 offset:200
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v53
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v102 offset:198
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 8, v53
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v101 offset:204
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v52
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v100 offset:202
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 8, v52
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v80 offset:196
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v54
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v54
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v113 offset:200
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v55
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v112 offset:198
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v55
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v80 offset:204
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v52
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v81 offset:202
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v52
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v82 offset:196
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v53
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v83 offset:194
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v53
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v114 offset:192
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(9)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v50
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v81 offset:194
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v50
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v82 offset:192
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v51
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v83 offset:190
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v51
-; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:360
-; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:364
-; ALIGNED-NEXT:    buffer_store_dword v65, off, s[0:3], s32 offset:356
-; ALIGNED-NEXT:    buffer_store_dword v64, off, s[0:3], s32 offset:352
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v50
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v115 offset:190
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v50
+; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:488
+; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:492
+; ALIGNED-NEXT:    buffer_store_dword v65, off, s[0:3], s32 offset:484
+; ALIGNED-NEXT:    buffer_store_dword v64, off, s[0:3], s32 offset:480
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v66 offset:180
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v67 offset:184
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v67 offset:185
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v66 offset:181
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v67 offset:185
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v66 offset:181
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v65 offset:176
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v65 offset:177
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v65 offset:177
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v64 offset:172
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v64 offset:173
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v49
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v49
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v48
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v48
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v84 offset:184
-; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 24, v38
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v85 offset:182
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 8, v38
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v86 offset:188
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 24, v39
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v87 offset:186
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 8, v39
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v115 offset:180
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v37
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v114 offset:178
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 8, v37
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v113 offset:176
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v36
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v112 offset:174
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v36
-; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:376
-; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:380
-; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:372
-; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:368
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v64 offset:173
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v51
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v51
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v49
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v49
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v98 offset:184
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v48
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v99 offset:182
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v48
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v100 offset:188
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v101 offset:186
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v102 offset:180
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v103 offset:178
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v68 offset:176
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v69 offset:174
+; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:508
+; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:504
+; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:500
+; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:496
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v54 offset:164
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v54 offset:165
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v55 offset:168
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v55 offset:169
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v54 offset:165
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v53 offset:160
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v53 offset:161
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v55 offset:169
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v52 offset:156
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v52 offset:157
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v52 offset:157
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v53 offset:160
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v53 offset:161
+; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v36
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v38
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v38
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v39
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v39
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v37
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v37
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v36
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v34
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v34
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 24, v35
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v35
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v68 offset:168
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v33
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v69 offset:166
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v33
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v70 offset:172
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v32
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v71 offset:170
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v32
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v103 offset:164
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v102 offset:162
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v101 offset:160
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v100 offset:158
-; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:328
-; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:332
-; ALIGNED-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:324
-; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:320
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 24, v34
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v34
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v70 offset:168
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v35
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v71 offset:166
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v35
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v113 offset:172
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v33
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v112 offset:170
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v33
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v80 offset:160
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v32
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v81 offset:158
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v32
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v82 offset:164
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v83 offset:162
+; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:456
+; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:460
+; ALIGNED-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:452
+; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:448
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v114 offset:152
+; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v28
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v115 offset:150
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v28
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v30
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v30
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v50 offset:148
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v51 offset:152
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v51 offset:153
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v50 offset:149
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v51 offset:153
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v50 offset:149
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v49 offset:144
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v49 offset:145
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v49 offset:145
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v48 offset:140
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v48 offset:141
-; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 24, v29
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v29
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v28
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v30
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 8, v30
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v31
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 8, v31
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v28
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v80 offset:152
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v48 offset:141
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 24, v31
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v31
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v29
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v29
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v64 offset:156
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(5)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v26
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v81 offset:150
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v82 offset:156
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v83 offset:154
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v64 offset:148
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v65 offset:146
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v66 offset:144
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v67 offset:142
-; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:344
-; ALIGNED-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:348
-; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:340
-; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:336
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v26
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v65 offset:154
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v26
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v66 offset:148
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v27
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v67 offset:146
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v98 offset:144
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v99 offset:142
+; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:472
+; ALIGNED-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:476
+; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:468
+; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:464
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v52 offset:128
+; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v18
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v27
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v38 offset:132
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v39 offset:136
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v39 offset:137
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v38 offset:133
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v39 offset:137
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v38 offset:133
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v37 offset:128
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v37 offset:129
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v37 offset:129
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v36 offset:124
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v36 offset:128
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v84 offset:136
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v85 offset:134
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v86 offset:140
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v87 offset:138
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v115 offset:132
-; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 24, v18
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v114 offset:130
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v113 offset:128
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v112 offset:126
-; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:424
-; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:428
-; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:420
-; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:416
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v52 offset:120
-; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v52, 24, v14
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v26
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 8, v18
+; ALIGNED-NEXT:    flat_store_byte v[84:85], v36 offset:128
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v100 offset:136
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v101 offset:134
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v102 offset:140
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v103 offset:138
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v68 offset:132
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v69 offset:130
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v53 offset:126
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v18
+; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:296
+; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:300
+; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292
+; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:288
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v34 offset:116
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v35 offset:120
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v35 offset:121
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v34 offset:117
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v35 offset:121
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v34 offset:117
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v33 offset:112
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v33 offset:113
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v33 offset:113
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v32 offset:108
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v32 offset:109
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v53 offset:118
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v54 offset:124
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v55 offset:122
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v68 offset:116
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v69 offset:114
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v70 offset:112
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v71 offset:110
-; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:440
-; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:444
-; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:436
-; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:432
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v32 offset:109
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v54 offset:120
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v55 offset:118
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v70 offset:124
+; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v14
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v71 offset:122
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v113 offset:116
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v112 offset:114
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v80 offset:112
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v81 offset:110
+; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:312
+; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:316
+; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:308
+; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:304
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v30 offset:100
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v31 offset:104
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v31 offset:105
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v30 offset:101
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v31 offset:105
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v30 offset:101
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v29 offset:96
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v29 offset:97
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v29 offset:97
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v28 offset:92
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v28 offset:93
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 24, v10
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v48 offset:100
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 24, v6
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v27
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v19
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v53, 8, v14
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 8, v10
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v49 offset:98
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v6
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v27
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v25
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 24, v15
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v11
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v50 offset:96
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v7
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v25
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v66, 24, v24
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v67, 8, v24
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 24, v22
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 8, v22
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 24, v23
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 8, v23
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v84, 24, v21
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v85, 8, v21
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v86, 24, v20
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v87, 8, v20
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v19
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 24, v17
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 8, v17
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 24, v16
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 8, v16
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v15
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v12
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v12
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v70, 24, v13
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v13
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 8, v11
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v103 offset:104
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 24, v9
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v102 offset:102
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 8, v9
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v101 offset:108
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 24, v8
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v100 offset:106
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 8, v8
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v51 offset:94
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v7
-; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:392
-; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:396
-; ALIGNED-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:388
-; ALIGNED-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:384
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v28 offset:93
+; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v30, 24, v10
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v114 offset:96
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v114, 24, v6
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v98, 24, v25
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v32, 24, v19
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v71, 8, v14
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v31, 8, v10
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v115 offset:94
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v115, 8, v6
+; ALIGNED-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v99, 8, v25
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v33, 8, v19
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v36, 24, v24
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v37, 8, v24
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v34, 24, v17
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v113, 24, v15
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v82 offset:104
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v82, 24, v11
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v83 offset:102
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v48 offset:108
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v49 offset:106
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v50 offset:100
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v51 offset:98
+; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:264
+; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:268
+; ALIGNED-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:260
+; ALIGNED-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:256
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v26 offset:84
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v27 offset:88
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v27 offset:89
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v26 offset:85
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v27 offset:89
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v26 offset:85
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v25 offset:80
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v25 offset:81
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v25 offset:81
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v24 offset:76
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v24 offset:77
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 8, v4
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v80 offset:88
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v81 offset:86
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v82 offset:92
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v83 offset:90
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v64 offset:84
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v65 offset:82
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v66 offset:80
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v67 offset:78
-; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:408
-; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:412
-; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:404
-; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:400
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v24 offset:77
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v7
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v38, 24, v22
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v39, 8, v22
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v100, 24, v23
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v101, 8, v23
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v102, 24, v21
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v103, 8, v21
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v68, 24, v20
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v69, 8, v20
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v35, 8, v17
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v54, 24, v16
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v55, 8, v16
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v112, 8, v15
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v80, 24, v13
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v81, 8, v13
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v28, 24, v12
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v29, 8, v12
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v83, 8, v11
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v48, 24, v9
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v49, 8, v9
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v50, 24, v8
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v51, 8, v8
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 8, v7
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 24, v5
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v27, 8, v5
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v64 offset:88
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v64, 24, v4
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v65 offset:86
+; ALIGNED-NEXT:    v_lshrrev_b32_e32 v65, 8, v4
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v66 offset:92
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v67 offset:90
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v98 offset:84
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v99 offset:82
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v36 offset:80
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v37 offset:78
+; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:280
+; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:284
+; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:276
+; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:272
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v22 offset:68
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v23 offset:72
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v23 offset:73
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v22 offset:69
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v23 offset:73
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v22 offset:69
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v21 offset:64
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v21 offset:65
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v21 offset:65
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v20 offset:60
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v20 offset:64
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v36 offset:72
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v37 offset:70
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v38 offset:76
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v39 offset:74
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v84 offset:68
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v85 offset:66
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v86 offset:64
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v87 offset:62
-; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:488
-; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:492
-; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:484
-; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:480
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v115 offset:56
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v114 offset:54
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v113 offset:60
+; ALIGNED-NEXT:    flat_store_byte v[84:85], v20 offset:64
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v38 offset:72
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v39 offset:70
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v100 offset:76
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v101 offset:74
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v102 offset:68
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v103 offset:66
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v68 offset:64
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v69 offset:62
+; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:360
+; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:364
+; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:356
+; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:352
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v52 offset:56
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v53 offset:54
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v32 offset:60
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v33 offset:58
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v34 offset:52
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v18 offset:52
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v19 offset:56
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v19 offset:57
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v112 offset:58
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v18 offset:53
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v32 offset:52
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v19 offset:57
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v18 offset:53
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v17 offset:48
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v33 offset:50
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v17 offset:49
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v34 offset:48
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v35 offset:50
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v17 offset:49
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v54 offset:48
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v16 offset:44
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v35 offset:46
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v16 offset:45
-; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:508
-; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:504
-; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:500
-; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:496
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v52 offset:40
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v55 offset:46
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v16 offset:45
+; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:376
+; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:380
+; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:372
+; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:368
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v14 offset:36
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v53 offset:38
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v14 offset:37
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v54 offset:44
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v70 offset:40
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v71 offset:38
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v15 offset:40
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v55 offset:42
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v15 offset:41
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v68 offset:32
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v12 offset:28
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v69 offset:30
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v12 offset:32
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v70 offset:36
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v113 offset:44
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v15 offset:41
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v112 offset:42
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v14 offset:37
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v80 offset:36
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v13 offset:32
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v71 offset:34
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v13 offset:33
-; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:456
-; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:460
-; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:452
-; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:448
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v81 offset:34
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v13 offset:33
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v28 offset:32
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v12 offset:28
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v29 offset:30
+; ALIGNED-NEXT:    flat_store_byte v[84:85], v12 offset:32
+; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:328
+; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:332
+; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:324
+; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:320
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v10 offset:20
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v28 offset:24
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v29 offset:22
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v30 offset:24
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v31 offset:22
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v11 offset:24
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v30 offset:28
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v11 offset:25
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v31 offset:26
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v10 offset:21
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v103 offset:20
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v82 offset:28
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v11 offset:25
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v83 offset:26
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v10 offset:21
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v48 offset:20
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v9 offset:16
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v102 offset:18
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v9 offset:17
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v101 offset:16
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v49 offset:18
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v9 offset:17
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v50 offset:16
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v8 offset:12
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v100 offset:14
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v8 offset:16
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:472
-; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:476
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:468
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:464
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v51 offset:14
+; ALIGNED-NEXT:    flat_store_byte v[84:85], v8 offset:16
+; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:344
+; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:348
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:340
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:336
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v6 offset:4
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v48 offset:8
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v49 offset:6
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v114 offset:8
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v115 offset:6
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v7 offset:8
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v50 offset:12
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v7 offset:9
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v51 offset:10
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v6 offset:8
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v5 offset:4
-; ALIGNED-NEXT:    flat_store_byte_d16_hi v[0:1], v4 offset:2
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v27 offset:1
-; ALIGNED-NEXT:    flat_store_byte v[0:1], v4
-; ALIGNED-NEXT:    v_add_co_u32 v0, vcc_lo, 0xffffff00, v0
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v24, 24, v5
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v25, 8, v5
-; ALIGNED-NEXT:    s_cmp_eq_u64 s[4:5], 0
-; ALIGNED-NEXT:    v_lshrrev_b32_e32 v26, 24, v4
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v24 offset:4
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v24 offset:12
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v7 offset:9
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v25 offset:10
+; ALIGNED-NEXT:    flat_store_byte v[84:85], v6 offset:8
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v26 offset:4
 ; ALIGNED-NEXT:    flat_store_byte_d16_hi v[96:97], v5
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v25 offset:2
-; ALIGNED-NEXT:    flat_store_byte v[98:99], v26
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v27 offset:2
+; ALIGNED-NEXT:    flat_store_byte v[84:85], v5 offset:4
+; ALIGNED-NEXT:    flat_store_byte v[86:87], v64
+; ALIGNED-NEXT:    flat_store_byte_d16_hi v[84:85], v4 offset:2
+; ALIGNED-NEXT:    flat_store_byte v[84:85], v65 offset:1
+; ALIGNED-NEXT:    flat_store_byte v[84:85], v4
 ; ALIGNED-NEXT:    s_cbranch_scc0 .LBB7_5
-; ALIGNED-NEXT:  .LBB7_6: ; %Flow17
-; ALIGNED-NEXT:    s_or_b32 exec_lo, exec_lo, s6
+; ALIGNED-NEXT:  .LBB7_6: ; %Flow7
+; ALIGNED-NEXT:    s_or_b32 exec_lo, exec_lo, s8
 ; ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
 ; ALIGNED-NEXT:    s_setpc_b64 s[30:31]
 ;
@@ -9608,31 +9588,27 @@ define void @memmove_p0_p4_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(4
 ; UNROLL3-NEXT:    s_xor_b32 s6, exec_lo, s4
 ; UNROLL3-NEXT:    s_cbranch_execz .LBB7_4
 ; UNROLL3-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader
-; UNROLL3-NEXT:    v_mov_b32_e32 v5, v3
-; UNROLL3-NEXT:    v_mov_b32_e32 v7, v1
-; UNROLL3-NEXT:    v_mov_b32_e32 v4, v2
-; UNROLL3-NEXT:    v_mov_b32_e32 v6, v0
-; UNROLL3-NEXT:    s_mov_b64 s[4:5], 0x7e0
+; UNROLL3-NEXT:    s_mov_b64 s[4:5], 0
 ; UNROLL3-NEXT:    .p2align 6
 ; UNROLL3-NEXT:  .LBB7_2: ; %memmove_fwd_loop
 ; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; UNROLL3-NEXT:    v_add_co_u32 v12, vcc_lo, v2, s4
+; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v13, null, s5, v3, vcc_lo
+; UNROLL3-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s4
+; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo
 ; UNROLL3-NEXT:    s_clause 0x2
-; UNROLL3-NEXT:    global_load_dwordx4 v[8:11], v[4:5], off offset:16
-; UNROLL3-NEXT:    global_load_dwordx4 v[12:15], v[4:5], off
-; UNROLL3-NEXT:    global_load_dwordx4 v[16:19], v[4:5], off offset:32
-; UNROLL3-NEXT:    v_add_co_u32 v4, vcc_lo, v4, 48
-; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
-; UNROLL3-NEXT:    s_add_u32 s4, s4, 0xffffffd0
-; UNROLL3-NEXT:    s_addc_u32 s5, s5, -1
+; UNROLL3-NEXT:    global_load_dwordx4 v[4:7], v[12:13], off offset:16
+; UNROLL3-NEXT:    global_load_dwordx4 v[8:11], v[12:13], off
+; UNROLL3-NEXT:    global_load_dwordx4 v[12:15], v[12:13], off offset:32
+; UNROLL3-NEXT:    s_add_u32 s4, s4, 48
+; UNROLL3-NEXT:    s_addc_u32 s5, s5, 0
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(2)
-; UNROLL3-NEXT:    flat_store_dwordx4 v[6:7], v[8:11] offset:16
+; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[4:7] offset:16
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(1)
-; UNROLL3-NEXT:    flat_store_dwordx4 v[6:7], v[12:15]
+; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[8:11]
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(0)
-; UNROLL3-NEXT:    flat_store_dwordx4 v[6:7], v[16:19] offset:32
-; UNROLL3-NEXT:    v_add_co_u32 v6, vcc_lo, v6, 48
-; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
-; UNROLL3-NEXT:    s_cmp_lg_u64 s[4:5], 0
+; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[12:15] offset:32
+; UNROLL3-NEXT:    s_cmp_lg_u64 s[4:5], 0x7e0
 ; UNROLL3-NEXT:    s_cbranch_scc1 .LBB7_2
 ; UNROLL3-NEXT:  ; %bb.3: ; %memmove_fwd_residual
 ; UNROLL3-NEXT:    s_clause 0x1
@@ -9644,44 +9620,43 @@ define void @memmove_p0_p4_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(4
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(0)
 ; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[8:11] offset:2032
 ; UNROLL3-NEXT:    ; implicit-def: $vgpr0_vgpr1
-; UNROLL3-NEXT:  .LBB7_4: ; %Flow14
-; UNROLL3-NEXT:    s_andn2_saveexec_b32 s6, s6
+; UNROLL3-NEXT:  .LBB7_4: ; %Flow4
+; UNROLL3-NEXT:    s_andn2_saveexec_b32 s8, s6
 ; UNROLL3-NEXT:    s_cbranch_execz .LBB7_7
 ; UNROLL3-NEXT:  ; %bb.5: ; %memmove_bwd_residual
 ; UNROLL3-NEXT:    s_clause 0x1
-; UNROLL3-NEXT:    global_load_dwordx4 v[6:9], v[2:3], off offset:2032
-; UNROLL3-NEXT:    global_load_dwordx4 v[10:13], v[2:3], off offset:2016
-; UNROLL3-NEXT:    v_add_co_u32 v4, vcc_lo, 0x7b0, v0
-; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v5, null, 0, v1, vcc_lo
-; UNROLL3-NEXT:    s_movk_i32 s4, 0xf820
-; UNROLL3-NEXT:    s_mov_b32 s5, -1
+; UNROLL3-NEXT:    global_load_dwordx4 v[4:7], v[2:3], off offset:2032
+; UNROLL3-NEXT:    global_load_dwordx4 v[8:11], v[2:3], off offset:2016
+; UNROLL3-NEXT:    s_mov_b64 s[4:5], 0x7b0
+; UNROLL3-NEXT:    s_movk_i32 s6, 0xffd0
+; UNROLL3-NEXT:    s_mov_b32 s7, -1
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(1)
-; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[6:9] offset:2032
+; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[4:7] offset:2032
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(0)
-; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[10:13] offset:2016
+; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[8:11] offset:2016
 ; UNROLL3-NEXT:    .p2align 6
 ; UNROLL3-NEXT:  .LBB7_6: ; %memmove_bwd_loop
 ; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; UNROLL3-NEXT:    v_add_co_u32 v12, vcc_lo, v2, s4
+; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v13, null, s5, v3, vcc_lo
+; UNROLL3-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s4
+; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo
 ; UNROLL3-NEXT:    s_clause 0x2
-; UNROLL3-NEXT:    global_load_dwordx4 v[6:9], v[2:3], off offset:1984
-; UNROLL3-NEXT:    global_load_dwordx4 v[10:13], v[2:3], off offset:1968
-; UNROLL3-NEXT:    global_load_dwordx4 v[14:17], v[2:3], off offset:2000
-; UNROLL3-NEXT:    v_add_co_u32 v2, vcc_lo, 0xffffffd0, v2
-; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v3, null, -1, v3, vcc_lo
-; UNROLL3-NEXT:    s_add_u32 s4, s4, 48
-; UNROLL3-NEXT:    s_addc_u32 s5, s5, 0
+; UNROLL3-NEXT:    global_load_dwordx4 v[4:7], v[12:13], off offset:16
+; UNROLL3-NEXT:    global_load_dwordx4 v[8:11], v[12:13], off
+; UNROLL3-NEXT:    global_load_dwordx4 v[12:15], v[12:13], off offset:32
+; UNROLL3-NEXT:    s_add_u32 s4, s4, 0xffffffd0
+; UNROLL3-NEXT:    s_addc_u32 s5, s5, -1
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(2)
-; UNROLL3-NEXT:    flat_store_dwordx4 v[4:5], v[6:9] offset:16
+; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[4:7] offset:16
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(1)
-; UNROLL3-NEXT:    flat_store_dwordx4 v[4:5], v[10:13]
+; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[8:11]
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(0)
-; UNROLL3-NEXT:    flat_store_dwordx4 v[4:5], v[14:17] offset:32
-; UNROLL3-NEXT:    v_add_co_u32 v4, vcc_lo, 0xffffffd0, v4
-; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v5, null, -1, v5, vcc_lo
-; UNROLL3-NEXT:    s_cmp_eq_u64 s[4:5], 0
+; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[12:15] offset:32
+; UNROLL3-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]
 ; UNROLL3-NEXT:    s_cbranch_scc0 .LBB7_6
-; UNROLL3-NEXT:  .LBB7_7: ; %Flow15
-; UNROLL3-NEXT:    s_or_b32 exec_lo, exec_lo, s6
+; UNROLL3-NEXT:  .LBB7_7: ; %Flow5
+; UNROLL3-NEXT:    s_or_b32 exec_lo, exec_lo, s8
 ; UNROLL3-NEXT:    s_waitcnt lgkmcnt(0)
 ; UNROLL3-NEXT:    s_setpc_b64 s[30:31]
 entry:
@@ -9903,205 +9878,207 @@ define void @memmove_p5_p5_sz2048(ptr addrspace(5) align 1 %dst, ptr addrspace(5
 ; CHECK-NEXT:    s_andn2_saveexec_b32 s6, s6
 ; CHECK-NEXT:    s_cbranch_execz .LBB8_6
 ; CHECK-NEXT:  ; %bb.4: ; %memmove_bwd_loop.preheader
+; CHECK-NEXT:    v_add_nc_u32_e32 v0, 0x700, v0
+; CHECK-NEXT:    v_add_nc_u32_e32 v1, 0x700, v1
 ; CHECK-NEXT:    s_movk_i32 s4, 0xf800
 ; CHECK-NEXT:    s_mov_b32 s5, -1
 ; CHECK-NEXT:  .LBB8_5: ; %memmove_bwd_loop
 ; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; CHECK-NEXT:    s_clause 0x3e
-; CHECK-NEXT:    buffer_load_dword v2, v1, s[0:3], 0 offen offset:2044
-; CHECK-NEXT:    buffer_load_dword v3, v1, s[0:3], 0 offen offset:2040
-; CHECK-NEXT:    buffer_load_dword v4, v1, s[0:3], 0 offen offset:2036
-; CHECK-NEXT:    buffer_load_dword v5, v1, s[0:3], 0 offen offset:2032
-; CHECK-NEXT:    buffer_load_dword v6, v1, s[0:3], 0 offen offset:2028
-; CHECK-NEXT:    buffer_load_dword v7, v1, s[0:3], 0 offen offset:2024
-; CHECK-NEXT:    buffer_load_dword v8, v1, s[0:3], 0 offen offset:2020
-; CHECK-NEXT:    buffer_load_dword v9, v1, s[0:3], 0 offen offset:2016
-; CHECK-NEXT:    buffer_load_dword v10, v1, s[0:3], 0 offen offset:2012
-; CHECK-NEXT:    buffer_load_dword v11, v1, s[0:3], 0 offen offset:2008
-; CHECK-NEXT:    buffer_load_dword v12, v1, s[0:3], 0 offen offset:2004
-; CHECK-NEXT:    buffer_load_dword v13, v1, s[0:3], 0 offen offset:2000
-; CHECK-NEXT:    buffer_load_dword v14, v1, s[0:3], 0 offen offset:1996
-; CHECK-NEXT:    buffer_load_dword v15, v1, s[0:3], 0 offen offset:1992
-; CHECK-NEXT:    buffer_load_dword v16, v1, s[0:3], 0 offen offset:1988
-; CHECK-NEXT:    buffer_load_dword v17, v1, s[0:3], 0 offen offset:1984
-; CHECK-NEXT:    buffer_load_dword v18, v1, s[0:3], 0 offen offset:1980
-; CHECK-NEXT:    buffer_load_dword v19, v1, s[0:3], 0 offen offset:1976
-; CHECK-NEXT:    buffer_load_dword v20, v1, s[0:3], 0 offen offset:1972
-; CHECK-NEXT:    buffer_load_dword v21, v1, s[0:3], 0 offen offset:1968
-; CHECK-NEXT:    buffer_load_dword v22, v1, s[0:3], 0 offen offset:1964
-; CHECK-NEXT:    buffer_load_dword v23, v1, s[0:3], 0 offen offset:1960
-; CHECK-NEXT:    buffer_load_dword v24, v1, s[0:3], 0 offen offset:1956
-; CHECK-NEXT:    buffer_load_dword v25, v1, s[0:3], 0 offen offset:1952
-; CHECK-NEXT:    buffer_load_dword v26, v1, s[0:3], 0 offen offset:1948
-; CHECK-NEXT:    buffer_load_dword v27, v1, s[0:3], 0 offen offset:1944
-; CHECK-NEXT:    buffer_load_dword v28, v1, s[0:3], 0 offen offset:1940
-; CHECK-NEXT:    buffer_load_dword v29, v1, s[0:3], 0 offen offset:1936
-; CHECK-NEXT:    buffer_load_dword v30, v1, s[0:3], 0 offen offset:1932
-; CHECK-NEXT:    buffer_load_dword v31, v1, s[0:3], 0 offen offset:1928
-; CHECK-NEXT:    buffer_load_dword v32, v1, s[0:3], 0 offen offset:1924
-; CHECK-NEXT:    buffer_load_dword v33, v1, s[0:3], 0 offen offset:1920
-; CHECK-NEXT:    buffer_load_dword v34, v1, s[0:3], 0 offen offset:1916
-; CHECK-NEXT:    buffer_load_dword v35, v1, s[0:3], 0 offen offset:1912
-; CHECK-NEXT:    buffer_load_dword v36, v1, s[0:3], 0 offen offset:1908
-; CHECK-NEXT:    buffer_load_dword v37, v1, s[0:3], 0 offen offset:1904
-; CHECK-NEXT:    buffer_load_dword v38, v1, s[0:3], 0 offen offset:1900
-; CHECK-NEXT:    buffer_load_dword v39, v1, s[0:3], 0 offen offset:1896
-; CHECK-NEXT:    buffer_load_dword v48, v1, s[0:3], 0 offen offset:1892
-; CHECK-NEXT:    buffer_load_dword v49, v1, s[0:3], 0 offen offset:1888
-; CHECK-NEXT:    buffer_load_dword v50, v1, s[0:3], 0 offen offset:1884
-; CHECK-NEXT:    buffer_load_dword v51, v1, s[0:3], 0 offen offset:1880
-; CHECK-NEXT:    buffer_load_dword v52, v1, s[0:3], 0 offen offset:1876
-; CHECK-NEXT:    buffer_load_dword v53, v1, s[0:3], 0 offen offset:1872
-; CHECK-NEXT:    buffer_load_dword v54, v1, s[0:3], 0 offen offset:1868
-; CHECK-NEXT:    buffer_load_dword v55, v1, s[0:3], 0 offen offset:1864
-; CHECK-NEXT:    buffer_load_dword v64, v1, s[0:3], 0 offen offset:1860
-; CHECK-NEXT:    buffer_load_dword v65, v1, s[0:3], 0 offen offset:1856
-; CHECK-NEXT:    buffer_load_dword v66, v1, s[0:3], 0 offen offset:1852
-; CHECK-NEXT:    buffer_load_dword v67, v1, s[0:3], 0 offen offset:1848
-; CHECK-NEXT:    buffer_load_dword v68, v1, s[0:3], 0 offen offset:1844
-; CHECK-NEXT:    buffer_load_dword v69, v1, s[0:3], 0 offen offset:1840
-; CHECK-NEXT:    buffer_load_dword v70, v1, s[0:3], 0 offen offset:1836
-; CHECK-NEXT:    buffer_load_dword v71, v1, s[0:3], 0 offen offset:1832
-; CHECK-NEXT:    buffer_load_dword v80, v1, s[0:3], 0 offen offset:1828
-; CHECK-NEXT:    buffer_load_dword v81, v1, s[0:3], 0 offen offset:1824
-; CHECK-NEXT:    buffer_load_dword v82, v1, s[0:3], 0 offen offset:1820
-; CHECK-NEXT:    buffer_load_dword v83, v1, s[0:3], 0 offen offset:1816
-; CHECK-NEXT:    buffer_load_dword v84, v1, s[0:3], 0 offen offset:1812
-; CHECK-NEXT:    buffer_load_dword v85, v1, s[0:3], 0 offen offset:1808
-; CHECK-NEXT:    buffer_load_dword v86, v1, s[0:3], 0 offen offset:1804
-; CHECK-NEXT:    buffer_load_dword v87, v1, s[0:3], 0 offen offset:1800
-; CHECK-NEXT:    buffer_load_dword v96, v1, s[0:3], 0 offen offset:1796
-; CHECK-NEXT:    buffer_load_dword v97, v1, s[0:3], 0 offen offset:1792
+; CHECK-NEXT:    buffer_load_dword v2, v1, s[0:3], 0 offen offset:252
+; CHECK-NEXT:    buffer_load_dword v3, v1, s[0:3], 0 offen offset:248
+; CHECK-NEXT:    buffer_load_dword v4, v1, s[0:3], 0 offen offset:244
+; CHECK-NEXT:    buffer_load_dword v5, v1, s[0:3], 0 offen offset:240
+; CHECK-NEXT:    buffer_load_dword v6, v1, s[0:3], 0 offen offset:236
+; CHECK-NEXT:    buffer_load_dword v7, v1, s[0:3], 0 offen offset:232
+; CHECK-NEXT:    buffer_load_dword v8, v1, s[0:3], 0 offen offset:228
+; CHECK-NEXT:    buffer_load_dword v9, v1, s[0:3], 0 offen offset:224
+; CHECK-NEXT:    buffer_load_dword v10, v1, s[0:3], 0 offen offset:220
+; CHECK-NEXT:    buffer_load_dword v11, v1, s[0:3], 0 offen offset:216
+; CHECK-NEXT:    buffer_load_dword v12, v1, s[0:3], 0 offen offset:212
+; CHECK-NEXT:    buffer_load_dword v13, v1, s[0:3], 0 offen offset:208
+; CHECK-NEXT:    buffer_load_dword v14, v1, s[0:3], 0 offen offset:204
+; CHECK-NEXT:    buffer_load_dword v15, v1, s[0:3], 0 offen offset:200
+; CHECK-NEXT:    buffer_load_dword v16, v1, s[0:3], 0 offen offset:196
+; CHECK-NEXT:    buffer_load_dword v17, v1, s[0:3], 0 offen offset:192
+; CHECK-NEXT:    buffer_load_dword v18, v1, s[0:3], 0 offen offset:188
+; CHECK-NEXT:    buffer_load_dword v19, v1, s[0:3], 0 offen offset:184
+; CHECK-NEXT:    buffer_load_dword v20, v1, s[0:3], 0 offen offset:180
+; CHECK-NEXT:    buffer_load_dword v21, v1, s[0:3], 0 offen offset:176
+; CHECK-NEXT:    buffer_load_dword v22, v1, s[0:3], 0 offen offset:172
+; CHECK-NEXT:    buffer_load_dword v23, v1, s[0:3], 0 offen offset:168
+; CHECK-NEXT:    buffer_load_dword v24, v1, s[0:3], 0 offen offset:164
+; CHECK-NEXT:    buffer_load_dword v25, v1, s[0:3], 0 offen offset:160
+; CHECK-NEXT:    buffer_load_dword v26, v1, s[0:3], 0 offen offset:156
+; CHECK-NEXT:    buffer_load_dword v27, v1, s[0:3], 0 offen offset:152
+; CHECK-NEXT:    buffer_load_dword v28, v1, s[0:3], 0 offen offset:148
+; CHECK-NEXT:    buffer_load_dword v29, v1, s[0:3], 0 offen offset:144
+; CHECK-NEXT:    buffer_load_dword v30, v1, s[0:3], 0 offen offset:140
+; CHECK-NEXT:    buffer_load_dword v31, v1, s[0:3], 0 offen offset:136
+; CHECK-NEXT:    buffer_load_dword v32, v1, s[0:3], 0 offen offset:132
+; CHECK-NEXT:    buffer_load_dword v33, v1, s[0:3], 0 offen offset:128
+; CHECK-NEXT:    buffer_load_dword v34, v1, s[0:3], 0 offen offset:124
+; CHECK-NEXT:    buffer_load_dword v35, v1, s[0:3], 0 offen offset:120
+; CHECK-NEXT:    buffer_load_dword v36, v1, s[0:3], 0 offen offset:116
+; CHECK-NEXT:    buffer_load_dword v37, v1, s[0:3], 0 offen offset:112
+; CHECK-NEXT:    buffer_load_dword v38, v1, s[0:3], 0 offen offset:108
+; CHECK-NEXT:    buffer_load_dword v39, v1, s[0:3], 0 offen offset:104
+; CHECK-NEXT:    buffer_load_dword v48, v1, s[0:3], 0 offen offset:100
+; CHECK-NEXT:    buffer_load_dword v49, v1, s[0:3], 0 offen offset:96
+; CHECK-NEXT:    buffer_load_dword v50, v1, s[0:3], 0 offen offset:92
+; CHECK-NEXT:    buffer_load_dword v51, v1, s[0:3], 0 offen offset:88
+; CHECK-NEXT:    buffer_load_dword v52, v1, s[0:3], 0 offen offset:84
+; CHECK-NEXT:    buffer_load_dword v53, v1, s[0:3], 0 offen offset:80
+; CHECK-NEXT:    buffer_load_dword v54, v1, s[0:3], 0 offen offset:76
+; CHECK-NEXT:    buffer_load_dword v55, v1, s[0:3], 0 offen offset:72
+; CHECK-NEXT:    buffer_load_dword v64, v1, s[0:3], 0 offen offset:68
+; CHECK-NEXT:    buffer_load_dword v65, v1, s[0:3], 0 offen offset:64
+; CHECK-NEXT:    buffer_load_dword v66, v1, s[0:3], 0 offen offset:60
+; CHECK-NEXT:    buffer_load_dword v67, v1, s[0:3], 0 offen offset:56
+; CHECK-NEXT:    buffer_load_dword v68, v1, s[0:3], 0 offen offset:52
+; CHECK-NEXT:    buffer_load_dword v69, v1, s[0:3], 0 offen offset:48
+; CHECK-NEXT:    buffer_load_dword v70, v1, s[0:3], 0 offen offset:44
+; CHECK-NEXT:    buffer_load_dword v71, v1, s[0:3], 0 offen offset:40
+; CHECK-NEXT:    buffer_load_dword v80, v1, s[0:3], 0 offen offset:36
+; CHECK-NEXT:    buffer_load_dword v81, v1, s[0:3], 0 offen offset:32
+; CHECK-NEXT:    buffer_load_dword v82, v1, s[0:3], 0 offen offset:28
+; CHECK-NEXT:    buffer_load_dword v83, v1, s[0:3], 0 offen offset:24
+; CHECK-NEXT:    buffer_load_dword v84, v1, s[0:3], 0 offen offset:20
+; CHECK-NEXT:    buffer_load_dword v85, v1, s[0:3], 0 offen offset:16
+; CHECK-NEXT:    buffer_load_dword v86, v1, s[0:3], 0 offen offset:12
+; CHECK-NEXT:    buffer_load_dword v87, v1, s[0:3], 0 offen offset:8
+; CHECK-NEXT:    buffer_load_dword v96, v1, s[0:3], 0 offen offset:4
+; CHECK-NEXT:    buffer_load_dword v97, v1, s[0:3], 0 offen
 ; CHECK-NEXT:    v_add_nc_u32_e32 v1, 0xffffff00, v1
 ; CHECK-NEXT:    s_add_u32 s4, s4, 0x100
 ; CHECK-NEXT:    s_addc_u32 s5, s5, 0
 ; CHECK-NEXT:    s_waitcnt vmcnt(62)
-; CHECK-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:2044
-; CHECK-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:2040
+; CHECK-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:252
+; CHECK-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:248
 ; CHECK-NEXT:    s_waitcnt vmcnt(61)
-; CHECK-NEXT:    buffer_store_dword v4, v0, s[0:3], 0 offen offset:2036
+; CHECK-NEXT:    buffer_store_dword v4, v0, s[0:3], 0 offen offset:244
 ; CHECK-NEXT:    s_waitcnt vmcnt(60)
-; CHECK-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:2032
+; CHECK-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:240
 ; CHECK-NEXT:    s_waitcnt vmcnt(59)
-; CHECK-NEXT:    buffer_store_dword v6, v0, s[0:3], 0 offen offset:2028
+; CHECK-NEXT:    buffer_store_dword v6, v0, s[0:3], 0 offen offset:236
 ; CHECK-NEXT:    s_waitcnt vmcnt(58)
-; CHECK-NEXT:    buffer_store_dword v7, v0, s[0:3], 0 offen offset:2024
+; CHECK-NEXT:    buffer_store_dword v7, v0, s[0:3], 0 offen offset:232
 ; CHECK-NEXT:    s_waitcnt vmcnt(57)
-; CHECK-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:2020
+; CHECK-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:228
 ; CHECK-NEXT:    s_waitcnt vmcnt(56)
-; CHECK-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:2016
+; CHECK-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:224
 ; CHECK-NEXT:    s_waitcnt vmcnt(55)
-; CHECK-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:2012
+; CHECK-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:220
 ; CHECK-NEXT:    s_waitcnt vmcnt(54)
-; CHECK-NEXT:    buffer_store_dword v11, v0, s[0:3], 0 offen offset:2008
+; CHECK-NEXT:    buffer_store_dword v11, v0, s[0:3], 0 offen offset:216
 ; CHECK-NEXT:    s_waitcnt vmcnt(53)
-; CHECK-NEXT:    buffer_store_dword v12, v0, s[0:3], 0 offen offset:2004
+; CHECK-NEXT:    buffer_store_dword v12, v0, s[0:3], 0 offen offset:212
 ; CHECK-NEXT:    s_waitcnt vmcnt(52)
-; CHECK-NEXT:    buffer_store_dword v13, v0, s[0:3], 0 offen offset:2000
+; CHECK-NEXT:    buffer_store_dword v13, v0, s[0:3], 0 offen offset:208
 ; CHECK-NEXT:    s_waitcnt vmcnt(51)
-; CHECK-NEXT:    buffer_store_dword v14, v0, s[0:3], 0 offen offset:1996
+; CHECK-NEXT:    buffer_store_dword v14, v0, s[0:3], 0 offen offset:204
 ; CHECK-NEXT:    s_waitcnt vmcnt(50)
-; CHECK-NEXT:    buffer_store_dword v15, v0, s[0:3], 0 offen offset:1992
+; CHECK-NEXT:    buffer_store_dword v15, v0, s[0:3], 0 offen offset:200
 ; CHECK-NEXT:    s_waitcnt vmcnt(49)
-; CHECK-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:1988
+; CHECK-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:196
 ; CHECK-NEXT:    s_waitcnt vmcnt(48)
-; CHECK-NEXT:    buffer_store_dword v17, v0, s[0:3], 0 offen offset:1984
+; CHECK-NEXT:    buffer_store_dword v17, v0, s[0:3], 0 offen offset:192
 ; CHECK-NEXT:    s_waitcnt vmcnt(47)
-; CHECK-NEXT:    buffer_store_dword v18, v0, s[0:3], 0 offen offset:1980
+; CHECK-NEXT:    buffer_store_dword v18, v0, s[0:3], 0 offen offset:188
 ; CHECK-NEXT:    s_waitcnt vmcnt(46)
-; CHECK-NEXT:    buffer_store_dword v19, v0, s[0:3], 0 offen offset:1976
+; CHECK-NEXT:    buffer_store_dword v19, v0, s[0:3], 0 offen offset:184
 ; CHECK-NEXT:    s_waitcnt vmcnt(45)
-; CHECK-NEXT:    buffer_store_dword v20, v0, s[0:3], 0 offen offset:1972
+; CHECK-NEXT:    buffer_store_dword v20, v0, s[0:3], 0 offen offset:180
 ; CHECK-NEXT:    s_waitcnt vmcnt(44)
-; CHECK-NEXT:    buffer_store_dword v21, v0, s[0:3], 0 offen offset:1968
+; CHECK-NEXT:    buffer_store_dword v21, v0, s[0:3], 0 offen offset:176
 ; CHECK-NEXT:    s_waitcnt vmcnt(43)
-; CHECK-NEXT:    buffer_store_dword v22, v0, s[0:3], 0 offen offset:1964
+; CHECK-NEXT:    buffer_store_dword v22, v0, s[0:3], 0 offen offset:172
 ; CHECK-NEXT:    s_waitcnt vmcnt(42)
-; CHECK-NEXT:    buffer_store_dword v23, v0, s[0:3], 0 offen offset:1960
+; CHECK-NEXT:    buffer_store_dword v23, v0, s[0:3], 0 offen offset:168
 ; CHECK-NEXT:    s_waitcnt vmcnt(41)
-; CHECK-NEXT:    buffer_store_dword v24, v0, s[0:3], 0 offen offset:1956
+; CHECK-NEXT:    buffer_store_dword v24, v0, s[0:3], 0 offen offset:164
 ; CHECK-NEXT:    s_waitcnt vmcnt(40)
-; CHECK-NEXT:    buffer_store_dword v25, v0, s[0:3], 0 offen offset:1952
+; CHECK-NEXT:    buffer_store_dword v25, v0, s[0:3], 0 offen offset:160
 ; CHECK-NEXT:    s_waitcnt vmcnt(39)
-; CHECK-NEXT:    buffer_store_dword v26, v0, s[0:3], 0 offen offset:1948
+; CHECK-NEXT:    buffer_store_dword v26, v0, s[0:3], 0 offen offset:156
 ; CHECK-NEXT:    s_waitcnt vmcnt(38)
-; CHECK-NEXT:    buffer_store_dword v27, v0, s[0:3], 0 offen offset:1944
+; CHECK-NEXT:    buffer_store_dword v27, v0, s[0:3], 0 offen offset:152
 ; CHECK-NEXT:    s_waitcnt vmcnt(37)
-; CHECK-NEXT:    buffer_store_dword v28, v0, s[0:3], 0 offen offset:1940
+; CHECK-NEXT:    buffer_store_dword v28, v0, s[0:3], 0 offen offset:148
 ; CHECK-NEXT:    s_waitcnt vmcnt(36)
-; CHECK-NEXT:    buffer_store_dword v29, v0, s[0:3], 0 offen offset:1936
+; CHECK-NEXT:    buffer_store_dword v29, v0, s[0:3], 0 offen offset:144
 ; CHECK-NEXT:    s_waitcnt vmcnt(35)
-; CHECK-NEXT:    buffer_store_dword v30, v0, s[0:3], 0 offen offset:1932
+; CHECK-NEXT:    buffer_store_dword v30, v0, s[0:3], 0 offen offset:140
 ; CHECK-NEXT:    s_waitcnt vmcnt(34)
-; CHECK-NEXT:    buffer_store_dword v31, v0, s[0:3], 0 offen offset:1928
+; CHECK-NEXT:    buffer_store_dword v31, v0, s[0:3], 0 offen offset:136
 ; CHECK-NEXT:    s_waitcnt vmcnt(33)
-; CHECK-NEXT:    buffer_store_dword v32, v0, s[0:3], 0 offen offset:1924
+; CHECK-NEXT:    buffer_store_dword v32, v0, s[0:3], 0 offen offset:132
 ; CHECK-NEXT:    s_waitcnt vmcnt(32)
-; CHECK-NEXT:    buffer_store_dword v33, v0, s[0:3], 0 offen offset:1920
+; CHECK-NEXT:    buffer_store_dword v33, v0, s[0:3], 0 offen offset:128
 ; CHECK-NEXT:    s_waitcnt vmcnt(31)
-; CHECK-NEXT:    buffer_store_dword v34, v0, s[0:3], 0 offen offset:1916
+; CHECK-NEXT:    buffer_store_dword v34, v0, s[0:3], 0 offen offset:124
 ; CHECK-NEXT:    s_waitcnt vmcnt(30)
-; CHECK-NEXT:    buffer_store_dword v35, v0, s[0:3], 0 offen offset:1912
+; CHECK-NEXT:    buffer_store_dword v35, v0, s[0:3], 0 offen offset:120
 ; CHECK-NEXT:    s_waitcnt vmcnt(29)
-; CHECK-NEXT:    buffer_store_dword v36, v0, s[0:3], 0 offen offset:1908
+; CHECK-NEXT:    buffer_store_dword v36, v0, s[0:3], 0 offen offset:116
 ; CHECK-NEXT:    s_waitcnt vmcnt(28)
-; CHECK-NEXT:    buffer_store_dword v37, v0, s[0:3], 0 offen offset:1904
+; CHECK-NEXT:    buffer_store_dword v37, v0, s[0:3], 0 offen offset:112
 ; CHECK-NEXT:    s_waitcnt vmcnt(27)
-; CHECK-NEXT:    buffer_store_dword v38, v0, s[0:3], 0 offen offset:1900
+; CHECK-NEXT:    buffer_store_dword v38, v0, s[0:3], 0 offen offset:108
 ; CHECK-NEXT:    s_waitcnt vmcnt(26)
-; CHECK-NEXT:    buffer_store_dword v39, v0, s[0:3], 0 offen offset:1896
+; CHECK-NEXT:    buffer_store_dword v39, v0, s[0:3], 0 offen offset:104
 ; CHECK-NEXT:    s_waitcnt vmcnt(25)
-; CHECK-NEXT:    buffer_store_dword v48, v0, s[0:3], 0 offen offset:1892
+; CHECK-NEXT:    buffer_store_dword v48, v0, s[0:3], 0 offen offset:100
 ; CHECK-NEXT:    s_waitcnt vmcnt(24)
-; CHECK-NEXT:    buffer_store_dword v49, v0, s[0:3], 0 offen offset:1888
+; CHECK-NEXT:    buffer_store_dword v49, v0, s[0:3], 0 offen offset:96
 ; CHECK-NEXT:    s_waitcnt vmcnt(23)
-; CHECK-NEXT:    buffer_store_dword v50, v0, s[0:3], 0 offen offset:1884
+; CHECK-NEXT:    buffer_store_dword v50, v0, s[0:3], 0 offen offset:92
 ; CHECK-NEXT:    s_waitcnt vmcnt(22)
-; CHECK-NEXT:    buffer_store_dword v51, v0, s[0:3], 0 offen offset:1880
+; CHECK-NEXT:    buffer_store_dword v51, v0, s[0:3], 0 offen offset:88
 ; CHECK-NEXT:    s_waitcnt vmcnt(21)
-; CHECK-NEXT:    buffer_store_dword v52, v0, s[0:3], 0 offen offset:1876
+; CHECK-NEXT:    buffer_store_dword v52, v0, s[0:3], 0 offen offset:84
 ; CHECK-NEXT:    s_waitcnt vmcnt(20)
-; CHECK-NEXT:    buffer_store_dword v53, v0, s[0:3], 0 offen offset:1872
+; CHECK-NEXT:    buffer_store_dword v53, v0, s[0:3], 0 offen offset:80
 ; CHECK-NEXT:    s_waitcnt vmcnt(19)
-; CHECK-NEXT:    buffer_store_dword v54, v0, s[0:3], 0 offen offset:1868
+; CHECK-NEXT:    buffer_store_dword v54, v0, s[0:3], 0 offen offset:76
 ; CHECK-NEXT:    s_waitcnt vmcnt(18)
-; CHECK-NEXT:    buffer_store_dword v55, v0, s[0:3], 0 offen offset:1864
+; CHECK-NEXT:    buffer_store_dword v55, v0, s[0:3], 0 offen offset:72
 ; CHECK-NEXT:    s_waitcnt vmcnt(17)
-; CHECK-NEXT:    buffer_store_dword v64, v0, s[0:3], 0 offen offset:1860
+; CHECK-NEXT:    buffer_store_dword v64, v0, s[0:3], 0 offen offset:68
 ; CHECK-NEXT:    s_waitcnt vmcnt(16)
-; CHECK-NEXT:    buffer_store_dword v65, v0, s[0:3], 0 offen offset:1856
+; CHECK-NEXT:    buffer_store_dword v65, v0, s[0:3], 0 offen offset:64
 ; CHECK-NEXT:    s_waitcnt vmcnt(15)
-; CHECK-NEXT:    buffer_store_dword v66, v0, s[0:3], 0 offen offset:1852
+; CHECK-NEXT:    buffer_store_dword v66, v0, s[0:3], 0 offen offset:60
 ; CHECK-NEXT:    s_waitcnt vmcnt(14)
-; CHECK-NEXT:    buffer_store_dword v67, v0, s[0:3], 0 offen offset:1848
+; CHECK-NEXT:    buffer_store_dword v67, v0, s[0:3], 0 offen offset:56
 ; CHECK-NEXT:    s_waitcnt vmcnt(13)
-; CHECK-NEXT:    buffer_store_dword v68, v0, s[0:3], 0 offen offset:1844
+; CHECK-NEXT:    buffer_store_dword v68, v0, s[0:3], 0 offen offset:52
 ; CHECK-NEXT:    s_waitcnt vmcnt(12)
-; CHECK-NEXT:    buffer_store_dword v69, v0, s[0:3], 0 offen offset:1840
+; CHECK-NEXT:    buffer_store_dword v69, v0, s[0:3], 0 offen offset:48
 ; CHECK-NEXT:    s_waitcnt vmcnt(11)
-; CHECK-NEXT:    buffer_store_dword v70, v0, s[0:3], 0 offen offset:1836
+; CHECK-NEXT:    buffer_store_dword v70, v0, s[0:3], 0 offen offset:44
 ; CHECK-NEXT:    s_waitcnt vmcnt(10)
-; CHECK-NEXT:    buffer_store_dword v71, v0, s[0:3], 0 offen offset:1832
+; CHECK-NEXT:    buffer_store_dword v71, v0, s[0:3], 0 offen offset:40
 ; CHECK-NEXT:    s_waitcnt vmcnt(9)
-; CHECK-NEXT:    buffer_store_dword v80, v0, s[0:3], 0 offen offset:1828
+; CHECK-NEXT:    buffer_store_dword v80, v0, s[0:3], 0 offen offset:36
 ; CHECK-NEXT:    s_waitcnt vmcnt(8)
-; CHECK-NEXT:    buffer_store_dword v81, v0, s[0:3], 0 offen offset:1824
+; CHECK-NEXT:    buffer_store_dword v81, v0, s[0:3], 0 offen offset:32
 ; CHECK-NEXT:    s_waitcnt vmcnt(7)
-; CHECK-NEXT:    buffer_store_dword v82, v0, s[0:3], 0 offen offset:1820
+; CHECK-NEXT:    buffer_store_dword v82, v0, s[0:3], 0 offen offset:28
 ; CHECK-NEXT:    s_waitcnt vmcnt(6)
-; CHECK-NEXT:    buffer_store_dword v83, v0, s[0:3], 0 offen offset:1816
+; CHECK-NEXT:    buffer_store_dword v83, v0, s[0:3], 0 offen offset:24
 ; CHECK-NEXT:    s_waitcnt vmcnt(5)
-; CHECK-NEXT:    buffer_store_dword v84, v0, s[0:3], 0 offen offset:1812
+; CHECK-NEXT:    buffer_store_dword v84, v0, s[0:3], 0 offen offset:20
 ; CHECK-NEXT:    s_waitcnt vmcnt(4)
-; CHECK-NEXT:    buffer_store_dword v85, v0, s[0:3], 0 offen offset:1808
+; CHECK-NEXT:    buffer_store_dword v85, v0, s[0:3], 0 offen offset:16
 ; CHECK-NEXT:    s_waitcnt vmcnt(3)
-; CHECK-NEXT:    buffer_store_dword v86, v0, s[0:3], 0 offen offset:1804
+; CHECK-NEXT:    buffer_store_dword v86, v0, s[0:3], 0 offen offset:12
 ; CHECK-NEXT:    s_waitcnt vmcnt(2)
-; CHECK-NEXT:    buffer_store_dword v87, v0, s[0:3], 0 offen offset:1800
+; CHECK-NEXT:    buffer_store_dword v87, v0, s[0:3], 0 offen offset:8
 ; CHECK-NEXT:    s_waitcnt vmcnt(1)
-; CHECK-NEXT:    buffer_store_dword v96, v0, s[0:3], 0 offen offset:1796
+; CHECK-NEXT:    buffer_store_dword v96, v0, s[0:3], 0 offen offset:4
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
-; CHECK-NEXT:    buffer_store_dword v97, v0, s[0:3], 0 offen offset:1792
+; CHECK-NEXT:    buffer_store_dword v97, v0, s[0:3], 0 offen
 ; CHECK-NEXT:    v_add_nc_u32_e32 v0, 0xffffff00, v0
 ; CHECK-NEXT:    s_cmp_eq_u64 s[4:5], 0
 ; CHECK-NEXT:    s_cbranch_scc0 .LBB8_5
@@ -11217,1053 +11194,1055 @@ define void @memmove_p5_p5_sz2048(ptr addrspace(5) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    s_andn2_saveexec_b32 s6, s6
 ; ALIGNED-NEXT:    s_cbranch_execz .LBB8_6
 ; ALIGNED-NEXT:  ; %bb.4: ; %memmove_bwd_loop.preheader
+; ALIGNED-NEXT:    v_add_nc_u32_e32 v0, 0x700, v0
+; ALIGNED-NEXT:    v_add_nc_u32_e32 v1, 0x700, v1
 ; ALIGNED-NEXT:    s_movk_i32 s4, 0xf800
 ; ALIGNED-NEXT:    s_mov_b32 s5, -1
 ; ALIGNED-NEXT:  .LBB8_5: ; %memmove_bwd_loop
 ; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=1
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1792
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen
 ; ALIGNED-NEXT:    s_add_u32 s4, s4, 0x100
 ; ALIGNED-NEXT:    s_addc_u32 s5, s5, 0
 ; ALIGNED-NEXT:    s_cmp_eq_u64 s[4:5], 0
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1793
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1794
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:2
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1795
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:3
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1796
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:4
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1797
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:5
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1798
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:6
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1799
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:7
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1800
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:8
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1801
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:9
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1802
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:10
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1803
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:11
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1804
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:12
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_clause 0x34
-; ALIGNED-NEXT:    buffer_load_ubyte v116, v1, s[0:3], 0 offen offset:2047
-; ALIGNED-NEXT:    buffer_load_ubyte v117, v1, s[0:3], 0 offen offset:2046
-; ALIGNED-NEXT:    buffer_load_ubyte v118, v1, s[0:3], 0 offen offset:2045
-; ALIGNED-NEXT:    buffer_load_ubyte v119, v1, s[0:3], 0 offen offset:2044
-; ALIGNED-NEXT:    buffer_load_ubyte v40, v1, s[0:3], 0 offen offset:2043
-; ALIGNED-NEXT:    buffer_load_ubyte v41, v1, s[0:3], 0 offen offset:2042
-; ALIGNED-NEXT:    buffer_load_ubyte v42, v1, s[0:3], 0 offen offset:2041
-; ALIGNED-NEXT:    buffer_load_ubyte v43, v1, s[0:3], 0 offen offset:2040
-; ALIGNED-NEXT:    buffer_load_ubyte v44, v1, s[0:3], 0 offen offset:2039
-; ALIGNED-NEXT:    buffer_load_ubyte v45, v1, s[0:3], 0 offen offset:2038
-; ALIGNED-NEXT:    buffer_load_ubyte v46, v1, s[0:3], 0 offen offset:2037
-; ALIGNED-NEXT:    buffer_load_ubyte v47, v1, s[0:3], 0 offen offset:2036
-; ALIGNED-NEXT:    buffer_load_ubyte v56, v1, s[0:3], 0 offen offset:2035
-; ALIGNED-NEXT:    buffer_load_ubyte v57, v1, s[0:3], 0 offen offset:2034
-; ALIGNED-NEXT:    buffer_load_ubyte v58, v1, s[0:3], 0 offen offset:2033
-; ALIGNED-NEXT:    buffer_load_ubyte v59, v1, s[0:3], 0 offen offset:2032
-; ALIGNED-NEXT:    buffer_load_ubyte v60, v1, s[0:3], 0 offen offset:2031
-; ALIGNED-NEXT:    buffer_load_ubyte v61, v1, s[0:3], 0 offen offset:2030
-; ALIGNED-NEXT:    buffer_load_ubyte v62, v1, s[0:3], 0 offen offset:2029
-; ALIGNED-NEXT:    buffer_load_ubyte v63, v1, s[0:3], 0 offen offset:2028
-; ALIGNED-NEXT:    buffer_load_ubyte v72, v1, s[0:3], 0 offen offset:2027
-; ALIGNED-NEXT:    buffer_load_ubyte v73, v1, s[0:3], 0 offen offset:2026
-; ALIGNED-NEXT:    buffer_load_ubyte v74, v1, s[0:3], 0 offen offset:2025
-; ALIGNED-NEXT:    buffer_load_ubyte v75, v1, s[0:3], 0 offen offset:2024
-; ALIGNED-NEXT:    buffer_load_ubyte v76, v1, s[0:3], 0 offen offset:2023
-; ALIGNED-NEXT:    buffer_load_ubyte v77, v1, s[0:3], 0 offen offset:2022
-; ALIGNED-NEXT:    buffer_load_ubyte v78, v1, s[0:3], 0 offen offset:2021
-; ALIGNED-NEXT:    buffer_load_ubyte v79, v1, s[0:3], 0 offen offset:2020
-; ALIGNED-NEXT:    buffer_load_ubyte v88, v1, s[0:3], 0 offen offset:2019
-; ALIGNED-NEXT:    buffer_load_ubyte v89, v1, s[0:3], 0 offen offset:2018
-; ALIGNED-NEXT:    buffer_load_ubyte v90, v1, s[0:3], 0 offen offset:2017
-; ALIGNED-NEXT:    buffer_load_ubyte v91, v1, s[0:3], 0 offen offset:2016
-; ALIGNED-NEXT:    buffer_load_ubyte v92, v1, s[0:3], 0 offen offset:2015
-; ALIGNED-NEXT:    buffer_load_ubyte v93, v1, s[0:3], 0 offen offset:2014
-; ALIGNED-NEXT:    buffer_load_ubyte v94, v1, s[0:3], 0 offen offset:2013
-; ALIGNED-NEXT:    buffer_load_ubyte v95, v1, s[0:3], 0 offen offset:2012
-; ALIGNED-NEXT:    buffer_load_ubyte v104, v1, s[0:3], 0 offen offset:2011
-; ALIGNED-NEXT:    buffer_load_ubyte v105, v1, s[0:3], 0 offen offset:2010
-; ALIGNED-NEXT:    buffer_load_ubyte v106, v1, s[0:3], 0 offen offset:2009
-; ALIGNED-NEXT:    buffer_load_ubyte v107, v1, s[0:3], 0 offen offset:2008
-; ALIGNED-NEXT:    buffer_load_ubyte v108, v1, s[0:3], 0 offen offset:2007
-; ALIGNED-NEXT:    buffer_load_ubyte v109, v1, s[0:3], 0 offen offset:2006
-; ALIGNED-NEXT:    buffer_load_ubyte v110, v1, s[0:3], 0 offen offset:2005
-; ALIGNED-NEXT:    buffer_load_ubyte v111, v1, s[0:3], 0 offen offset:2004
-; ALIGNED-NEXT:    buffer_load_ubyte v120, v1, s[0:3], 0 offen offset:2003
-; ALIGNED-NEXT:    buffer_load_ubyte v121, v1, s[0:3], 0 offen offset:2002
-; ALIGNED-NEXT:    buffer_load_ubyte v122, v1, s[0:3], 0 offen offset:2001
-; ALIGNED-NEXT:    buffer_load_ubyte v123, v1, s[0:3], 0 offen offset:2000
-; ALIGNED-NEXT:    buffer_load_ubyte v124, v1, s[0:3], 0 offen offset:1999
-; ALIGNED-NEXT:    buffer_load_ubyte v125, v1, s[0:3], 0 offen offset:1998
-; ALIGNED-NEXT:    buffer_load_ubyte v126, v1, s[0:3], 0 offen offset:1997
-; ALIGNED-NEXT:    buffer_load_ubyte v127, v1, s[0:3], 0 offen offset:1996
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1995
+; ALIGNED-NEXT:    buffer_load_ubyte v116, v1, s[0:3], 0 offen offset:255
+; ALIGNED-NEXT:    buffer_load_ubyte v117, v1, s[0:3], 0 offen offset:254
+; ALIGNED-NEXT:    buffer_load_ubyte v118, v1, s[0:3], 0 offen offset:253
+; ALIGNED-NEXT:    buffer_load_ubyte v119, v1, s[0:3], 0 offen offset:252
+; ALIGNED-NEXT:    buffer_load_ubyte v40, v1, s[0:3], 0 offen offset:251
+; ALIGNED-NEXT:    buffer_load_ubyte v41, v1, s[0:3], 0 offen offset:250
+; ALIGNED-NEXT:    buffer_load_ubyte v42, v1, s[0:3], 0 offen offset:249
+; ALIGNED-NEXT:    buffer_load_ubyte v43, v1, s[0:3], 0 offen offset:248
+; ALIGNED-NEXT:    buffer_load_ubyte v44, v1, s[0:3], 0 offen offset:247
+; ALIGNED-NEXT:    buffer_load_ubyte v45, v1, s[0:3], 0 offen offset:246
+; ALIGNED-NEXT:    buffer_load_ubyte v46, v1, s[0:3], 0 offen offset:245
+; ALIGNED-NEXT:    buffer_load_ubyte v47, v1, s[0:3], 0 offen offset:244
+; ALIGNED-NEXT:    buffer_load_ubyte v56, v1, s[0:3], 0 offen offset:243
+; ALIGNED-NEXT:    buffer_load_ubyte v57, v1, s[0:3], 0 offen offset:242
+; ALIGNED-NEXT:    buffer_load_ubyte v58, v1, s[0:3], 0 offen offset:241
+; ALIGNED-NEXT:    buffer_load_ubyte v59, v1, s[0:3], 0 offen offset:240
+; ALIGNED-NEXT:    buffer_load_ubyte v60, v1, s[0:3], 0 offen offset:239
+; ALIGNED-NEXT:    buffer_load_ubyte v61, v1, s[0:3], 0 offen offset:238
+; ALIGNED-NEXT:    buffer_load_ubyte v62, v1, s[0:3], 0 offen offset:237
+; ALIGNED-NEXT:    buffer_load_ubyte v63, v1, s[0:3], 0 offen offset:236
+; ALIGNED-NEXT:    buffer_load_ubyte v72, v1, s[0:3], 0 offen offset:235
+; ALIGNED-NEXT:    buffer_load_ubyte v73, v1, s[0:3], 0 offen offset:234
+; ALIGNED-NEXT:    buffer_load_ubyte v74, v1, s[0:3], 0 offen offset:233
+; ALIGNED-NEXT:    buffer_load_ubyte v75, v1, s[0:3], 0 offen offset:232
+; ALIGNED-NEXT:    buffer_load_ubyte v76, v1, s[0:3], 0 offen offset:231
+; ALIGNED-NEXT:    buffer_load_ubyte v77, v1, s[0:3], 0 offen offset:230
+; ALIGNED-NEXT:    buffer_load_ubyte v78, v1, s[0:3], 0 offen offset:229
+; ALIGNED-NEXT:    buffer_load_ubyte v79, v1, s[0:3], 0 offen offset:228
+; ALIGNED-NEXT:    buffer_load_ubyte v88, v1, s[0:3], 0 offen offset:227
+; ALIGNED-NEXT:    buffer_load_ubyte v89, v1, s[0:3], 0 offen offset:226
+; ALIGNED-NEXT:    buffer_load_ubyte v90, v1, s[0:3], 0 offen offset:225
+; ALIGNED-NEXT:    buffer_load_ubyte v91, v1, s[0:3], 0 offen offset:224
+; ALIGNED-NEXT:    buffer_load_ubyte v92, v1, s[0:3], 0 offen offset:223
+; ALIGNED-NEXT:    buffer_load_ubyte v93, v1, s[0:3], 0 offen offset:222
+; ALIGNED-NEXT:    buffer_load_ubyte v94, v1, s[0:3], 0 offen offset:221
+; ALIGNED-NEXT:    buffer_load_ubyte v95, v1, s[0:3], 0 offen offset:220
+; ALIGNED-NEXT:    buffer_load_ubyte v104, v1, s[0:3], 0 offen offset:219
+; ALIGNED-NEXT:    buffer_load_ubyte v105, v1, s[0:3], 0 offen offset:218
+; ALIGNED-NEXT:    buffer_load_ubyte v106, v1, s[0:3], 0 offen offset:217
+; ALIGNED-NEXT:    buffer_load_ubyte v107, v1, s[0:3], 0 offen offset:216
+; ALIGNED-NEXT:    buffer_load_ubyte v108, v1, s[0:3], 0 offen offset:215
+; ALIGNED-NEXT:    buffer_load_ubyte v109, v1, s[0:3], 0 offen offset:214
+; ALIGNED-NEXT:    buffer_load_ubyte v110, v1, s[0:3], 0 offen offset:213
+; ALIGNED-NEXT:    buffer_load_ubyte v111, v1, s[0:3], 0 offen offset:212
+; ALIGNED-NEXT:    buffer_load_ubyte v120, v1, s[0:3], 0 offen offset:211
+; ALIGNED-NEXT:    buffer_load_ubyte v121, v1, s[0:3], 0 offen offset:210
+; ALIGNED-NEXT:    buffer_load_ubyte v122, v1, s[0:3], 0 offen offset:209
+; ALIGNED-NEXT:    buffer_load_ubyte v123, v1, s[0:3], 0 offen offset:208
+; ALIGNED-NEXT:    buffer_load_ubyte v124, v1, s[0:3], 0 offen offset:207
+; ALIGNED-NEXT:    buffer_load_ubyte v125, v1, s[0:3], 0 offen offset:206
+; ALIGNED-NEXT:    buffer_load_ubyte v126, v1, s[0:3], 0 offen offset:205
+; ALIGNED-NEXT:    buffer_load_ubyte v127, v1, s[0:3], 0 offen offset:204
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:203
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_clause 0x3e
-; ALIGNED-NEXT:    buffer_load_ubyte v3, v1, s[0:3], 0 offen offset:1994
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v1, s[0:3], 0 offen offset:1993
-; ALIGNED-NEXT:    buffer_load_ubyte v5, v1, s[0:3], 0 offen offset:1992
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v1, s[0:3], 0 offen offset:1991
-; ALIGNED-NEXT:    buffer_load_ubyte v7, v1, s[0:3], 0 offen offset:1990
-; ALIGNED-NEXT:    buffer_load_ubyte v8, v1, s[0:3], 0 offen offset:1989
-; ALIGNED-NEXT:    buffer_load_ubyte v9, v1, s[0:3], 0 offen offset:1988
-; ALIGNED-NEXT:    buffer_load_ubyte v10, v1, s[0:3], 0 offen offset:1987
-; ALIGNED-NEXT:    buffer_load_ubyte v11, v1, s[0:3], 0 offen offset:1986
-; ALIGNED-NEXT:    buffer_load_ubyte v12, v1, s[0:3], 0 offen offset:1985
-; ALIGNED-NEXT:    buffer_load_ubyte v13, v1, s[0:3], 0 offen offset:1984
-; ALIGNED-NEXT:    buffer_load_ubyte v14, v1, s[0:3], 0 offen offset:1983
-; ALIGNED-NEXT:    buffer_load_ubyte v15, v1, s[0:3], 0 offen offset:1982
-; ALIGNED-NEXT:    buffer_load_ubyte v16, v1, s[0:3], 0 offen offset:1981
-; ALIGNED-NEXT:    buffer_load_ubyte v17, v1, s[0:3], 0 offen offset:1980
-; ALIGNED-NEXT:    buffer_load_ubyte v18, v1, s[0:3], 0 offen offset:1979
-; ALIGNED-NEXT:    buffer_load_ubyte v19, v1, s[0:3], 0 offen offset:1978
-; ALIGNED-NEXT:    buffer_load_ubyte v20, v1, s[0:3], 0 offen offset:1977
-; ALIGNED-NEXT:    buffer_load_ubyte v21, v1, s[0:3], 0 offen offset:1976
-; ALIGNED-NEXT:    buffer_load_ubyte v22, v1, s[0:3], 0 offen offset:1975
-; ALIGNED-NEXT:    buffer_load_ubyte v23, v1, s[0:3], 0 offen offset:1974
-; ALIGNED-NEXT:    buffer_load_ubyte v24, v1, s[0:3], 0 offen offset:1973
-; ALIGNED-NEXT:    buffer_load_ubyte v25, v1, s[0:3], 0 offen offset:1972
-; ALIGNED-NEXT:    buffer_load_ubyte v26, v1, s[0:3], 0 offen offset:1971
-; ALIGNED-NEXT:    buffer_load_ubyte v27, v1, s[0:3], 0 offen offset:1970
-; ALIGNED-NEXT:    buffer_load_ubyte v28, v1, s[0:3], 0 offen offset:1969
-; ALIGNED-NEXT:    buffer_load_ubyte v29, v1, s[0:3], 0 offen offset:1968
-; ALIGNED-NEXT:    buffer_load_ubyte v30, v1, s[0:3], 0 offen offset:1967
-; ALIGNED-NEXT:    buffer_load_ubyte v31, v1, s[0:3], 0 offen offset:1966
-; ALIGNED-NEXT:    buffer_load_ubyte v32, v1, s[0:3], 0 offen offset:1965
-; ALIGNED-NEXT:    buffer_load_ubyte v33, v1, s[0:3], 0 offen offset:1964
-; ALIGNED-NEXT:    buffer_load_ubyte v34, v1, s[0:3], 0 offen offset:1963
-; ALIGNED-NEXT:    buffer_load_ubyte v35, v1, s[0:3], 0 offen offset:1962
-; ALIGNED-NEXT:    buffer_load_ubyte v36, v1, s[0:3], 0 offen offset:1961
-; ALIGNED-NEXT:    buffer_load_ubyte v37, v1, s[0:3], 0 offen offset:1960
-; ALIGNED-NEXT:    buffer_load_ubyte v38, v1, s[0:3], 0 offen offset:1959
-; ALIGNED-NEXT:    buffer_load_ubyte v39, v1, s[0:3], 0 offen offset:1958
-; ALIGNED-NEXT:    buffer_load_ubyte v48, v1, s[0:3], 0 offen offset:1957
-; ALIGNED-NEXT:    buffer_load_ubyte v49, v1, s[0:3], 0 offen offset:1956
-; ALIGNED-NEXT:    buffer_load_ubyte v50, v1, s[0:3], 0 offen offset:1955
-; ALIGNED-NEXT:    buffer_load_ubyte v51, v1, s[0:3], 0 offen offset:1954
-; ALIGNED-NEXT:    buffer_load_ubyte v52, v1, s[0:3], 0 offen offset:1953
-; ALIGNED-NEXT:    buffer_load_ubyte v53, v1, s[0:3], 0 offen offset:1952
-; ALIGNED-NEXT:    buffer_load_ubyte v54, v1, s[0:3], 0 offen offset:1951
-; ALIGNED-NEXT:    buffer_load_ubyte v55, v1, s[0:3], 0 offen offset:1950
-; ALIGNED-NEXT:    buffer_load_ubyte v64, v1, s[0:3], 0 offen offset:1949
-; ALIGNED-NEXT:    buffer_load_ubyte v65, v1, s[0:3], 0 offen offset:1948
-; ALIGNED-NEXT:    buffer_load_ubyte v66, v1, s[0:3], 0 offen offset:1947
-; ALIGNED-NEXT:    buffer_load_ubyte v67, v1, s[0:3], 0 offen offset:1946
-; ALIGNED-NEXT:    buffer_load_ubyte v68, v1, s[0:3], 0 offen offset:1945
-; ALIGNED-NEXT:    buffer_load_ubyte v69, v1, s[0:3], 0 offen offset:1944
-; ALIGNED-NEXT:    buffer_load_ubyte v70, v1, s[0:3], 0 offen offset:1943
-; ALIGNED-NEXT:    buffer_load_ubyte v71, v1, s[0:3], 0 offen offset:1942
-; ALIGNED-NEXT:    buffer_load_ubyte v80, v1, s[0:3], 0 offen offset:1941
-; ALIGNED-NEXT:    buffer_load_ubyte v81, v1, s[0:3], 0 offen offset:1940
-; ALIGNED-NEXT:    buffer_load_ubyte v82, v1, s[0:3], 0 offen offset:1939
-; ALIGNED-NEXT:    buffer_load_ubyte v83, v1, s[0:3], 0 offen offset:1938
-; ALIGNED-NEXT:    buffer_load_ubyte v84, v1, s[0:3], 0 offen offset:1937
-; ALIGNED-NEXT:    buffer_load_ubyte v85, v1, s[0:3], 0 offen offset:1936
-; ALIGNED-NEXT:    buffer_load_ubyte v86, v1, s[0:3], 0 offen offset:1935
-; ALIGNED-NEXT:    buffer_load_ubyte v87, v1, s[0:3], 0 offen offset:1934
-; ALIGNED-NEXT:    buffer_load_ubyte v96, v1, s[0:3], 0 offen offset:1933
-; ALIGNED-NEXT:    buffer_load_ubyte v97, v1, s[0:3], 0 offen offset:1932
+; ALIGNED-NEXT:    buffer_load_ubyte v3, v1, s[0:3], 0 offen offset:202
+; ALIGNED-NEXT:    buffer_load_ubyte v4, v1, s[0:3], 0 offen offset:201
+; ALIGNED-NEXT:    buffer_load_ubyte v5, v1, s[0:3], 0 offen offset:200
+; ALIGNED-NEXT:    buffer_load_ubyte v6, v1, s[0:3], 0 offen offset:199
+; ALIGNED-NEXT:    buffer_load_ubyte v7, v1, s[0:3], 0 offen offset:198
+; ALIGNED-NEXT:    buffer_load_ubyte v8, v1, s[0:3], 0 offen offset:197
+; ALIGNED-NEXT:    buffer_load_ubyte v9, v1, s[0:3], 0 offen offset:196
+; ALIGNED-NEXT:    buffer_load_ubyte v10, v1, s[0:3], 0 offen offset:195
+; ALIGNED-NEXT:    buffer_load_ubyte v11, v1, s[0:3], 0 offen offset:194
+; ALIGNED-NEXT:    buffer_load_ubyte v12, v1, s[0:3], 0 offen offset:193
+; ALIGNED-NEXT:    buffer_load_ubyte v13, v1, s[0:3], 0 offen offset:192
+; ALIGNED-NEXT:    buffer_load_ubyte v14, v1, s[0:3], 0 offen offset:191
+; ALIGNED-NEXT:    buffer_load_ubyte v15, v1, s[0:3], 0 offen offset:190
+; ALIGNED-NEXT:    buffer_load_ubyte v16, v1, s[0:3], 0 offen offset:189
+; ALIGNED-NEXT:    buffer_load_ubyte v17, v1, s[0:3], 0 offen offset:188
+; ALIGNED-NEXT:    buffer_load_ubyte v18, v1, s[0:3], 0 offen offset:187
+; ALIGNED-NEXT:    buffer_load_ubyte v19, v1, s[0:3], 0 offen offset:186
+; ALIGNED-NEXT:    buffer_load_ubyte v20, v1, s[0:3], 0 offen offset:185
+; ALIGNED-NEXT:    buffer_load_ubyte v21, v1, s[0:3], 0 offen offset:184
+; ALIGNED-NEXT:    buffer_load_ubyte v22, v1, s[0:3], 0 offen offset:183
+; ALIGNED-NEXT:    buffer_load_ubyte v23, v1, s[0:3], 0 offen offset:182
+; ALIGNED-NEXT:    buffer_load_ubyte v24, v1, s[0:3], 0 offen offset:181
+; ALIGNED-NEXT:    buffer_load_ubyte v25, v1, s[0:3], 0 offen offset:180
+; ALIGNED-NEXT:    buffer_load_ubyte v26, v1, s[0:3], 0 offen offset:179
+; ALIGNED-NEXT:    buffer_load_ubyte v27, v1, s[0:3], 0 offen offset:178
+; ALIGNED-NEXT:    buffer_load_ubyte v28, v1, s[0:3], 0 offen offset:177
+; ALIGNED-NEXT:    buffer_load_ubyte v29, v1, s[0:3], 0 offen offset:176
+; ALIGNED-NEXT:    buffer_load_ubyte v30, v1, s[0:3], 0 offen offset:175
+; ALIGNED-NEXT:    buffer_load_ubyte v31, v1, s[0:3], 0 offen offset:174
+; ALIGNED-NEXT:    buffer_load_ubyte v32, v1, s[0:3], 0 offen offset:173
+; ALIGNED-NEXT:    buffer_load_ubyte v33, v1, s[0:3], 0 offen offset:172
+; ALIGNED-NEXT:    buffer_load_ubyte v34, v1, s[0:3], 0 offen offset:171
+; ALIGNED-NEXT:    buffer_load_ubyte v35, v1, s[0:3], 0 offen offset:170
+; ALIGNED-NEXT:    buffer_load_ubyte v36, v1, s[0:3], 0 offen offset:169
+; ALIGNED-NEXT:    buffer_load_ubyte v37, v1, s[0:3], 0 offen offset:168
+; ALIGNED-NEXT:    buffer_load_ubyte v38, v1, s[0:3], 0 offen offset:167
+; ALIGNED-NEXT:    buffer_load_ubyte v39, v1, s[0:3], 0 offen offset:166
+; ALIGNED-NEXT:    buffer_load_ubyte v48, v1, s[0:3], 0 offen offset:165
+; ALIGNED-NEXT:    buffer_load_ubyte v49, v1, s[0:3], 0 offen offset:164
+; ALIGNED-NEXT:    buffer_load_ubyte v50, v1, s[0:3], 0 offen offset:163
+; ALIGNED-NEXT:    buffer_load_ubyte v51, v1, s[0:3], 0 offen offset:162
+; ALIGNED-NEXT:    buffer_load_ubyte v52, v1, s[0:3], 0 offen offset:161
+; ALIGNED-NEXT:    buffer_load_ubyte v53, v1, s[0:3], 0 offen offset:160
+; ALIGNED-NEXT:    buffer_load_ubyte v54, v1, s[0:3], 0 offen offset:159
+; ALIGNED-NEXT:    buffer_load_ubyte v55, v1, s[0:3], 0 offen offset:158
+; ALIGNED-NEXT:    buffer_load_ubyte v64, v1, s[0:3], 0 offen offset:157
+; ALIGNED-NEXT:    buffer_load_ubyte v65, v1, s[0:3], 0 offen offset:156
+; ALIGNED-NEXT:    buffer_load_ubyte v66, v1, s[0:3], 0 offen offset:155
+; ALIGNED-NEXT:    buffer_load_ubyte v67, v1, s[0:3], 0 offen offset:154
+; ALIGNED-NEXT:    buffer_load_ubyte v68, v1, s[0:3], 0 offen offset:153
+; ALIGNED-NEXT:    buffer_load_ubyte v69, v1, s[0:3], 0 offen offset:152
+; ALIGNED-NEXT:    buffer_load_ubyte v70, v1, s[0:3], 0 offen offset:151
+; ALIGNED-NEXT:    buffer_load_ubyte v71, v1, s[0:3], 0 offen offset:150
+; ALIGNED-NEXT:    buffer_load_ubyte v80, v1, s[0:3], 0 offen offset:149
+; ALIGNED-NEXT:    buffer_load_ubyte v81, v1, s[0:3], 0 offen offset:148
+; ALIGNED-NEXT:    buffer_load_ubyte v82, v1, s[0:3], 0 offen offset:147
+; ALIGNED-NEXT:    buffer_load_ubyte v83, v1, s[0:3], 0 offen offset:146
+; ALIGNED-NEXT:    buffer_load_ubyte v84, v1, s[0:3], 0 offen offset:145
+; ALIGNED-NEXT:    buffer_load_ubyte v85, v1, s[0:3], 0 offen offset:144
+; ALIGNED-NEXT:    buffer_load_ubyte v86, v1, s[0:3], 0 offen offset:143
+; ALIGNED-NEXT:    buffer_load_ubyte v87, v1, s[0:3], 0 offen offset:142
+; ALIGNED-NEXT:    buffer_load_ubyte v96, v1, s[0:3], 0 offen offset:141
+; ALIGNED-NEXT:    buffer_load_ubyte v97, v1, s[0:3], 0 offen offset:140
 ; ALIGNED-NEXT:    s_clause 0xa
-; ALIGNED-NEXT:    buffer_load_ubyte v98, v1, s[0:3], 0 offen offset:1931
-; ALIGNED-NEXT:    buffer_load_ubyte v99, v1, s[0:3], 0 offen offset:1930
-; ALIGNED-NEXT:    buffer_load_ubyte v100, v1, s[0:3], 0 offen offset:1929
-; ALIGNED-NEXT:    buffer_load_ubyte v101, v1, s[0:3], 0 offen offset:1928
-; ALIGNED-NEXT:    buffer_load_ubyte v102, v1, s[0:3], 0 offen offset:1927
-; ALIGNED-NEXT:    buffer_load_ubyte v103, v1, s[0:3], 0 offen offset:1926
-; ALIGNED-NEXT:    buffer_load_ubyte v112, v1, s[0:3], 0 offen offset:1925
-; ALIGNED-NEXT:    buffer_load_ubyte v113, v1, s[0:3], 0 offen offset:1924
-; ALIGNED-NEXT:    buffer_load_ubyte v114, v1, s[0:3], 0 offen offset:1923
-; ALIGNED-NEXT:    buffer_load_ubyte v115, v1, s[0:3], 0 offen offset:1922
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1921
+; ALIGNED-NEXT:    buffer_load_ubyte v98, v1, s[0:3], 0 offen offset:139
+; ALIGNED-NEXT:    buffer_load_ubyte v99, v1, s[0:3], 0 offen offset:138
+; ALIGNED-NEXT:    buffer_load_ubyte v100, v1, s[0:3], 0 offen offset:137
+; ALIGNED-NEXT:    buffer_load_ubyte v101, v1, s[0:3], 0 offen offset:136
+; ALIGNED-NEXT:    buffer_load_ubyte v102, v1, s[0:3], 0 offen offset:135
+; ALIGNED-NEXT:    buffer_load_ubyte v103, v1, s[0:3], 0 offen offset:134
+; ALIGNED-NEXT:    buffer_load_ubyte v112, v1, s[0:3], 0 offen offset:133
+; ALIGNED-NEXT:    buffer_load_ubyte v113, v1, s[0:3], 0 offen offset:132
+; ALIGNED-NEXT:    buffer_load_ubyte v114, v1, s[0:3], 0 offen offset:131
+; ALIGNED-NEXT:    buffer_load_ubyte v115, v1, s[0:3], 0 offen offset:130
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:129
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1920
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:128
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1919
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:127
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1918
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:126
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1917
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:125
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1916
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:124
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1915
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:123
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1914
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:122
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1913
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:121
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1912
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:120
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1911
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:119
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1910
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:118
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1909
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:117
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1908
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:116
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1907
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:115
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1906
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:114
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1905
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:113
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1904
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:112
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1903
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:111
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1902
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:110
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1901
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:109
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1900
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:108
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1899
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:107
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1898
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:106
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1897
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:105
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1896
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:104
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1895
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:103
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1894
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:102
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1893
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:101
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1892
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:100
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1891
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:99
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1890
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:98
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1889
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:97
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1888
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:96
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1887
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:95
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1886
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:94
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1885
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:93
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1884
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:92
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1883
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:91
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1882
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:90
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1881
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:89
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1880
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:88
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1879
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:87
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1878
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:86
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1877
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:85
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1876
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:84
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1875
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:83
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1874
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:82
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1873
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:81
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1872
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:80
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1871
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:79
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1870
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:78
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1869
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:77
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1868
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:76
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1867
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:75
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1866
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:74
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1865
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:73
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1864
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:72
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1863
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:71
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1862
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:70
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1861
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:69
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1860
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:68
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1859
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:67
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1858
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:66
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1857
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:65
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1856
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:64
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1855
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:63
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1854
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:62
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1853
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:61
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1852
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:60
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1851
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:59
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1850
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:58
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1849
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:57
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1848
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:56
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1847
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:55
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1846
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:54
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1845
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:53
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1844
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:52
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1843
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:51
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1842
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:50
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1841
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:49
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1840
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:48
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1839
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:47
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1838
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:46
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1837
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:45
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1836
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:44
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1835
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:43
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1834
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:42
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1833
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:41
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1832
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:40
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1831
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:39
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1830
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:38
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1829
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:37
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1828
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:36
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1827
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:35
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1826
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:34
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1825
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:33
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1824
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:32
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1823
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:31
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1822
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:30
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1821
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:29
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1820
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:28
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1819
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:27
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1818
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:26
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1817
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:25
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1816
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:24
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1815
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:23
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1814
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:22
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1813
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:21
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1812
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:20
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1811
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:19
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1810
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:18
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1809
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:17
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1808
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:16
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1807
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:15
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1806
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:14
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:1805
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v1, s[0:3], 0 offen offset:13
 ; ALIGNED-NEXT:    v_add_nc_u32_e32 v1, 0xffffff00, v1
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_byte v116, v0, s[0:3], 0 offen offset:2047
-; ALIGNED-NEXT:    buffer_store_byte v117, v0, s[0:3], 0 offen offset:2046
-; ALIGNED-NEXT:    buffer_store_byte v118, v0, s[0:3], 0 offen offset:2045
-; ALIGNED-NEXT:    buffer_store_byte v119, v0, s[0:3], 0 offen offset:2044
-; ALIGNED-NEXT:    buffer_store_byte v40, v0, s[0:3], 0 offen offset:2043
-; ALIGNED-NEXT:    buffer_store_byte v41, v0, s[0:3], 0 offen offset:2042
-; ALIGNED-NEXT:    buffer_store_byte v42, v0, s[0:3], 0 offen offset:2041
-; ALIGNED-NEXT:    buffer_store_byte v43, v0, s[0:3], 0 offen offset:2040
-; ALIGNED-NEXT:    buffer_store_byte v44, v0, s[0:3], 0 offen offset:2039
-; ALIGNED-NEXT:    buffer_store_byte v45, v0, s[0:3], 0 offen offset:2038
-; ALIGNED-NEXT:    buffer_store_byte v46, v0, s[0:3], 0 offen offset:2037
-; ALIGNED-NEXT:    buffer_store_byte v47, v0, s[0:3], 0 offen offset:2036
-; ALIGNED-NEXT:    buffer_store_byte v56, v0, s[0:3], 0 offen offset:2035
-; ALIGNED-NEXT:    buffer_store_byte v57, v0, s[0:3], 0 offen offset:2034
-; ALIGNED-NEXT:    buffer_store_byte v58, v0, s[0:3], 0 offen offset:2033
-; ALIGNED-NEXT:    buffer_store_byte v59, v0, s[0:3], 0 offen offset:2032
-; ALIGNED-NEXT:    buffer_store_byte v60, v0, s[0:3], 0 offen offset:2031
-; ALIGNED-NEXT:    buffer_store_byte v61, v0, s[0:3], 0 offen offset:2030
-; ALIGNED-NEXT:    buffer_store_byte v62, v0, s[0:3], 0 offen offset:2029
-; ALIGNED-NEXT:    buffer_store_byte v63, v0, s[0:3], 0 offen offset:2028
-; ALIGNED-NEXT:    buffer_store_byte v72, v0, s[0:3], 0 offen offset:2027
-; ALIGNED-NEXT:    buffer_store_byte v73, v0, s[0:3], 0 offen offset:2026
-; ALIGNED-NEXT:    buffer_store_byte v74, v0, s[0:3], 0 offen offset:2025
-; ALIGNED-NEXT:    buffer_store_byte v75, v0, s[0:3], 0 offen offset:2024
-; ALIGNED-NEXT:    buffer_store_byte v76, v0, s[0:3], 0 offen offset:2023
-; ALIGNED-NEXT:    buffer_store_byte v77, v0, s[0:3], 0 offen offset:2022
-; ALIGNED-NEXT:    buffer_store_byte v78, v0, s[0:3], 0 offen offset:2021
-; ALIGNED-NEXT:    buffer_store_byte v79, v0, s[0:3], 0 offen offset:2020
-; ALIGNED-NEXT:    buffer_store_byte v88, v0, s[0:3], 0 offen offset:2019
-; ALIGNED-NEXT:    buffer_store_byte v89, v0, s[0:3], 0 offen offset:2018
-; ALIGNED-NEXT:    buffer_store_byte v90, v0, s[0:3], 0 offen offset:2017
-; ALIGNED-NEXT:    buffer_store_byte v91, v0, s[0:3], 0 offen offset:2016
-; ALIGNED-NEXT:    buffer_store_byte v92, v0, s[0:3], 0 offen offset:2015
-; ALIGNED-NEXT:    buffer_store_byte v93, v0, s[0:3], 0 offen offset:2014
-; ALIGNED-NEXT:    buffer_store_byte v94, v0, s[0:3], 0 offen offset:2013
-; ALIGNED-NEXT:    buffer_store_byte v95, v0, s[0:3], 0 offen offset:2012
-; ALIGNED-NEXT:    buffer_store_byte v104, v0, s[0:3], 0 offen offset:2011
-; ALIGNED-NEXT:    buffer_store_byte v105, v0, s[0:3], 0 offen offset:2010
-; ALIGNED-NEXT:    buffer_store_byte v106, v0, s[0:3], 0 offen offset:2009
-; ALIGNED-NEXT:    buffer_store_byte v107, v0, s[0:3], 0 offen offset:2008
-; ALIGNED-NEXT:    buffer_store_byte v108, v0, s[0:3], 0 offen offset:2007
-; ALIGNED-NEXT:    buffer_store_byte v109, v0, s[0:3], 0 offen offset:2006
-; ALIGNED-NEXT:    buffer_store_byte v110, v0, s[0:3], 0 offen offset:2005
-; ALIGNED-NEXT:    buffer_store_byte v111, v0, s[0:3], 0 offen offset:2004
-; ALIGNED-NEXT:    buffer_store_byte v120, v0, s[0:3], 0 offen offset:2003
-; ALIGNED-NEXT:    buffer_store_byte v121, v0, s[0:3], 0 offen offset:2002
-; ALIGNED-NEXT:    buffer_store_byte v122, v0, s[0:3], 0 offen offset:2001
-; ALIGNED-NEXT:    buffer_store_byte v123, v0, s[0:3], 0 offen offset:2000
-; ALIGNED-NEXT:    buffer_store_byte v124, v0, s[0:3], 0 offen offset:1999
-; ALIGNED-NEXT:    buffer_store_byte v125, v0, s[0:3], 0 offen offset:1998
-; ALIGNED-NEXT:    buffer_store_byte v126, v0, s[0:3], 0 offen offset:1997
-; ALIGNED-NEXT:    buffer_store_byte v127, v0, s[0:3], 0 offen offset:1996
+; ALIGNED-NEXT:    buffer_store_byte v116, v0, s[0:3], 0 offen offset:255
+; ALIGNED-NEXT:    buffer_store_byte v117, v0, s[0:3], 0 offen offset:254
+; ALIGNED-NEXT:    buffer_store_byte v118, v0, s[0:3], 0 offen offset:253
+; ALIGNED-NEXT:    buffer_store_byte v119, v0, s[0:3], 0 offen offset:252
+; ALIGNED-NEXT:    buffer_store_byte v40, v0, s[0:3], 0 offen offset:251
+; ALIGNED-NEXT:    buffer_store_byte v41, v0, s[0:3], 0 offen offset:250
+; ALIGNED-NEXT:    buffer_store_byte v42, v0, s[0:3], 0 offen offset:249
+; ALIGNED-NEXT:    buffer_store_byte v43, v0, s[0:3], 0 offen offset:248
+; ALIGNED-NEXT:    buffer_store_byte v44, v0, s[0:3], 0 offen offset:247
+; ALIGNED-NEXT:    buffer_store_byte v45, v0, s[0:3], 0 offen offset:246
+; ALIGNED-NEXT:    buffer_store_byte v46, v0, s[0:3], 0 offen offset:245
+; ALIGNED-NEXT:    buffer_store_byte v47, v0, s[0:3], 0 offen offset:244
+; ALIGNED-NEXT:    buffer_store_byte v56, v0, s[0:3], 0 offen offset:243
+; ALIGNED-NEXT:    buffer_store_byte v57, v0, s[0:3], 0 offen offset:242
+; ALIGNED-NEXT:    buffer_store_byte v58, v0, s[0:3], 0 offen offset:241
+; ALIGNED-NEXT:    buffer_store_byte v59, v0, s[0:3], 0 offen offset:240
+; ALIGNED-NEXT:    buffer_store_byte v60, v0, s[0:3], 0 offen offset:239
+; ALIGNED-NEXT:    buffer_store_byte v61, v0, s[0:3], 0 offen offset:238
+; ALIGNED-NEXT:    buffer_store_byte v62, v0, s[0:3], 0 offen offset:237
+; ALIGNED-NEXT:    buffer_store_byte v63, v0, s[0:3], 0 offen offset:236
+; ALIGNED-NEXT:    buffer_store_byte v72, v0, s[0:3], 0 offen offset:235
+; ALIGNED-NEXT:    buffer_store_byte v73, v0, s[0:3], 0 offen offset:234
+; ALIGNED-NEXT:    buffer_store_byte v74, v0, s[0:3], 0 offen offset:233
+; ALIGNED-NEXT:    buffer_store_byte v75, v0, s[0:3], 0 offen offset:232
+; ALIGNED-NEXT:    buffer_store_byte v76, v0, s[0:3], 0 offen offset:231
+; ALIGNED-NEXT:    buffer_store_byte v77, v0, s[0:3], 0 offen offset:230
+; ALIGNED-NEXT:    buffer_store_byte v78, v0, s[0:3], 0 offen offset:229
+; ALIGNED-NEXT:    buffer_store_byte v79, v0, s[0:3], 0 offen offset:228
+; ALIGNED-NEXT:    buffer_store_byte v88, v0, s[0:3], 0 offen offset:227
+; ALIGNED-NEXT:    buffer_store_byte v89, v0, s[0:3], 0 offen offset:226
+; ALIGNED-NEXT:    buffer_store_byte v90, v0, s[0:3], 0 offen offset:225
+; ALIGNED-NEXT:    buffer_store_byte v91, v0, s[0:3], 0 offen offset:224
+; ALIGNED-NEXT:    buffer_store_byte v92, v0, s[0:3], 0 offen offset:223
+; ALIGNED-NEXT:    buffer_store_byte v93, v0, s[0:3], 0 offen offset:222
+; ALIGNED-NEXT:    buffer_store_byte v94, v0, s[0:3], 0 offen offset:221
+; ALIGNED-NEXT:    buffer_store_byte v95, v0, s[0:3], 0 offen offset:220
+; ALIGNED-NEXT:    buffer_store_byte v104, v0, s[0:3], 0 offen offset:219
+; ALIGNED-NEXT:    buffer_store_byte v105, v0, s[0:3], 0 offen offset:218
+; ALIGNED-NEXT:    buffer_store_byte v106, v0, s[0:3], 0 offen offset:217
+; ALIGNED-NEXT:    buffer_store_byte v107, v0, s[0:3], 0 offen offset:216
+; ALIGNED-NEXT:    buffer_store_byte v108, v0, s[0:3], 0 offen offset:215
+; ALIGNED-NEXT:    buffer_store_byte v109, v0, s[0:3], 0 offen offset:214
+; ALIGNED-NEXT:    buffer_store_byte v110, v0, s[0:3], 0 offen offset:213
+; ALIGNED-NEXT:    buffer_store_byte v111, v0, s[0:3], 0 offen offset:212
+; ALIGNED-NEXT:    buffer_store_byte v120, v0, s[0:3], 0 offen offset:211
+; ALIGNED-NEXT:    buffer_store_byte v121, v0, s[0:3], 0 offen offset:210
+; ALIGNED-NEXT:    buffer_store_byte v122, v0, s[0:3], 0 offen offset:209
+; ALIGNED-NEXT:    buffer_store_byte v123, v0, s[0:3], 0 offen offset:208
+; ALIGNED-NEXT:    buffer_store_byte v124, v0, s[0:3], 0 offen offset:207
+; ALIGNED-NEXT:    buffer_store_byte v125, v0, s[0:3], 0 offen offset:206
+; ALIGNED-NEXT:    buffer_store_byte v126, v0, s[0:3], 0 offen offset:205
+; ALIGNED-NEXT:    buffer_store_byte v127, v0, s[0:3], 0 offen offset:204
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1995
-; ALIGNED-NEXT:    buffer_store_byte v3, v0, s[0:3], 0 offen offset:1994
-; ALIGNED-NEXT:    buffer_store_byte v4, v0, s[0:3], 0 offen offset:1993
-; ALIGNED-NEXT:    buffer_store_byte v5, v0, s[0:3], 0 offen offset:1992
-; ALIGNED-NEXT:    buffer_store_byte v6, v0, s[0:3], 0 offen offset:1991
-; ALIGNED-NEXT:    buffer_store_byte v7, v0, s[0:3], 0 offen offset:1990
-; ALIGNED-NEXT:    buffer_store_byte v8, v0, s[0:3], 0 offen offset:1989
-; ALIGNED-NEXT:    buffer_store_byte v9, v0, s[0:3], 0 offen offset:1988
-; ALIGNED-NEXT:    buffer_store_byte v10, v0, s[0:3], 0 offen offset:1987
-; ALIGNED-NEXT:    buffer_store_byte v11, v0, s[0:3], 0 offen offset:1986
-; ALIGNED-NEXT:    buffer_store_byte v12, v0, s[0:3], 0 offen offset:1985
-; ALIGNED-NEXT:    buffer_store_byte v13, v0, s[0:3], 0 offen offset:1984
-; ALIGNED-NEXT:    buffer_store_byte v14, v0, s[0:3], 0 offen offset:1983
-; ALIGNED-NEXT:    buffer_store_byte v15, v0, s[0:3], 0 offen offset:1982
-; ALIGNED-NEXT:    buffer_store_byte v16, v0, s[0:3], 0 offen offset:1981
-; ALIGNED-NEXT:    buffer_store_byte v17, v0, s[0:3], 0 offen offset:1980
-; ALIGNED-NEXT:    buffer_store_byte v18, v0, s[0:3], 0 offen offset:1979
-; ALIGNED-NEXT:    buffer_store_byte v19, v0, s[0:3], 0 offen offset:1978
-; ALIGNED-NEXT:    buffer_store_byte v20, v0, s[0:3], 0 offen offset:1977
-; ALIGNED-NEXT:    buffer_store_byte v21, v0, s[0:3], 0 offen offset:1976
-; ALIGNED-NEXT:    buffer_store_byte v22, v0, s[0:3], 0 offen offset:1975
-; ALIGNED-NEXT:    buffer_store_byte v23, v0, s[0:3], 0 offen offset:1974
-; ALIGNED-NEXT:    buffer_store_byte v24, v0, s[0:3], 0 offen offset:1973
-; ALIGNED-NEXT:    buffer_store_byte v25, v0, s[0:3], 0 offen offset:1972
-; ALIGNED-NEXT:    buffer_store_byte v26, v0, s[0:3], 0 offen offset:1971
-; ALIGNED-NEXT:    buffer_store_byte v27, v0, s[0:3], 0 offen offset:1970
-; ALIGNED-NEXT:    buffer_store_byte v28, v0, s[0:3], 0 offen offset:1969
-; ALIGNED-NEXT:    buffer_store_byte v29, v0, s[0:3], 0 offen offset:1968
-; ALIGNED-NEXT:    buffer_store_byte v30, v0, s[0:3], 0 offen offset:1967
-; ALIGNED-NEXT:    buffer_store_byte v31, v0, s[0:3], 0 offen offset:1966
-; ALIGNED-NEXT:    buffer_store_byte v32, v0, s[0:3], 0 offen offset:1965
-; ALIGNED-NEXT:    buffer_store_byte v33, v0, s[0:3], 0 offen offset:1964
-; ALIGNED-NEXT:    buffer_store_byte v34, v0, s[0:3], 0 offen offset:1963
-; ALIGNED-NEXT:    buffer_store_byte v35, v0, s[0:3], 0 offen offset:1962
-; ALIGNED-NEXT:    buffer_store_byte v36, v0, s[0:3], 0 offen offset:1961
-; ALIGNED-NEXT:    buffer_store_byte v37, v0, s[0:3], 0 offen offset:1960
-; ALIGNED-NEXT:    buffer_store_byte v38, v0, s[0:3], 0 offen offset:1959
-; ALIGNED-NEXT:    buffer_store_byte v39, v0, s[0:3], 0 offen offset:1958
-; ALIGNED-NEXT:    buffer_store_byte v48, v0, s[0:3], 0 offen offset:1957
-; ALIGNED-NEXT:    buffer_store_byte v49, v0, s[0:3], 0 offen offset:1956
-; ALIGNED-NEXT:    buffer_store_byte v50, v0, s[0:3], 0 offen offset:1955
-; ALIGNED-NEXT:    buffer_store_byte v51, v0, s[0:3], 0 offen offset:1954
-; ALIGNED-NEXT:    buffer_store_byte v52, v0, s[0:3], 0 offen offset:1953
-; ALIGNED-NEXT:    buffer_store_byte v53, v0, s[0:3], 0 offen offset:1952
-; ALIGNED-NEXT:    buffer_store_byte v54, v0, s[0:3], 0 offen offset:1951
-; ALIGNED-NEXT:    buffer_store_byte v55, v0, s[0:3], 0 offen offset:1950
-; ALIGNED-NEXT:    buffer_store_byte v64, v0, s[0:3], 0 offen offset:1949
-; ALIGNED-NEXT:    buffer_store_byte v65, v0, s[0:3], 0 offen offset:1948
-; ALIGNED-NEXT:    buffer_store_byte v66, v0, s[0:3], 0 offen offset:1947
-; ALIGNED-NEXT:    buffer_store_byte v67, v0, s[0:3], 0 offen offset:1946
-; ALIGNED-NEXT:    buffer_store_byte v68, v0, s[0:3], 0 offen offset:1945
-; ALIGNED-NEXT:    buffer_store_byte v69, v0, s[0:3], 0 offen offset:1944
-; ALIGNED-NEXT:    buffer_store_byte v70, v0, s[0:3], 0 offen offset:1943
-; ALIGNED-NEXT:    buffer_store_byte v71, v0, s[0:3], 0 offen offset:1942
-; ALIGNED-NEXT:    buffer_store_byte v80, v0, s[0:3], 0 offen offset:1941
-; ALIGNED-NEXT:    buffer_store_byte v81, v0, s[0:3], 0 offen offset:1940
-; ALIGNED-NEXT:    buffer_store_byte v82, v0, s[0:3], 0 offen offset:1939
-; ALIGNED-NEXT:    buffer_store_byte v83, v0, s[0:3], 0 offen offset:1938
-; ALIGNED-NEXT:    buffer_store_byte v84, v0, s[0:3], 0 offen offset:1937
-; ALIGNED-NEXT:    buffer_store_byte v85, v0, s[0:3], 0 offen offset:1936
-; ALIGNED-NEXT:    buffer_store_byte v86, v0, s[0:3], 0 offen offset:1935
-; ALIGNED-NEXT:    buffer_store_byte v87, v0, s[0:3], 0 offen offset:1934
-; ALIGNED-NEXT:    buffer_store_byte v96, v0, s[0:3], 0 offen offset:1933
-; ALIGNED-NEXT:    buffer_store_byte v97, v0, s[0:3], 0 offen offset:1932
-; ALIGNED-NEXT:    buffer_store_byte v98, v0, s[0:3], 0 offen offset:1931
-; ALIGNED-NEXT:    buffer_store_byte v99, v0, s[0:3], 0 offen offset:1930
-; ALIGNED-NEXT:    buffer_store_byte v100, v0, s[0:3], 0 offen offset:1929
-; ALIGNED-NEXT:    buffer_store_byte v101, v0, s[0:3], 0 offen offset:1928
-; ALIGNED-NEXT:    buffer_store_byte v102, v0, s[0:3], 0 offen offset:1927
-; ALIGNED-NEXT:    buffer_store_byte v103, v0, s[0:3], 0 offen offset:1926
-; ALIGNED-NEXT:    buffer_store_byte v112, v0, s[0:3], 0 offen offset:1925
-; ALIGNED-NEXT:    buffer_store_byte v113, v0, s[0:3], 0 offen offset:1924
-; ALIGNED-NEXT:    buffer_store_byte v114, v0, s[0:3], 0 offen offset:1923
-; ALIGNED-NEXT:    buffer_store_byte v115, v0, s[0:3], 0 offen offset:1922
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:203
+; ALIGNED-NEXT:    buffer_store_byte v3, v0, s[0:3], 0 offen offset:202
+; ALIGNED-NEXT:    buffer_store_byte v4, v0, s[0:3], 0 offen offset:201
+; ALIGNED-NEXT:    buffer_store_byte v5, v0, s[0:3], 0 offen offset:200
+; ALIGNED-NEXT:    buffer_store_byte v6, v0, s[0:3], 0 offen offset:199
+; ALIGNED-NEXT:    buffer_store_byte v7, v0, s[0:3], 0 offen offset:198
+; ALIGNED-NEXT:    buffer_store_byte v8, v0, s[0:3], 0 offen offset:197
+; ALIGNED-NEXT:    buffer_store_byte v9, v0, s[0:3], 0 offen offset:196
+; ALIGNED-NEXT:    buffer_store_byte v10, v0, s[0:3], 0 offen offset:195
+; ALIGNED-NEXT:    buffer_store_byte v11, v0, s[0:3], 0 offen offset:194
+; ALIGNED-NEXT:    buffer_store_byte v12, v0, s[0:3], 0 offen offset:193
+; ALIGNED-NEXT:    buffer_store_byte v13, v0, s[0:3], 0 offen offset:192
+; ALIGNED-NEXT:    buffer_store_byte v14, v0, s[0:3], 0 offen offset:191
+; ALIGNED-NEXT:    buffer_store_byte v15, v0, s[0:3], 0 offen offset:190
+; ALIGNED-NEXT:    buffer_store_byte v16, v0, s[0:3], 0 offen offset:189
+; ALIGNED-NEXT:    buffer_store_byte v17, v0, s[0:3], 0 offen offset:188
+; ALIGNED-NEXT:    buffer_store_byte v18, v0, s[0:3], 0 offen offset:187
+; ALIGNED-NEXT:    buffer_store_byte v19, v0, s[0:3], 0 offen offset:186
+; ALIGNED-NEXT:    buffer_store_byte v20, v0, s[0:3], 0 offen offset:185
+; ALIGNED-NEXT:    buffer_store_byte v21, v0, s[0:3], 0 offen offset:184
+; ALIGNED-NEXT:    buffer_store_byte v22, v0, s[0:3], 0 offen offset:183
+; ALIGNED-NEXT:    buffer_store_byte v23, v0, s[0:3], 0 offen offset:182
+; ALIGNED-NEXT:    buffer_store_byte v24, v0, s[0:3], 0 offen offset:181
+; ALIGNED-NEXT:    buffer_store_byte v25, v0, s[0:3], 0 offen offset:180
+; ALIGNED-NEXT:    buffer_store_byte v26, v0, s[0:3], 0 offen offset:179
+; ALIGNED-NEXT:    buffer_store_byte v27, v0, s[0:3], 0 offen offset:178
+; ALIGNED-NEXT:    buffer_store_byte v28, v0, s[0:3], 0 offen offset:177
+; ALIGNED-NEXT:    buffer_store_byte v29, v0, s[0:3], 0 offen offset:176
+; ALIGNED-NEXT:    buffer_store_byte v30, v0, s[0:3], 0 offen offset:175
+; ALIGNED-NEXT:    buffer_store_byte v31, v0, s[0:3], 0 offen offset:174
+; ALIGNED-NEXT:    buffer_store_byte v32, v0, s[0:3], 0 offen offset:173
+; ALIGNED-NEXT:    buffer_store_byte v33, v0, s[0:3], 0 offen offset:172
+; ALIGNED-NEXT:    buffer_store_byte v34, v0, s[0:3], 0 offen offset:171
+; ALIGNED-NEXT:    buffer_store_byte v35, v0, s[0:3], 0 offen offset:170
+; ALIGNED-NEXT:    buffer_store_byte v36, v0, s[0:3], 0 offen offset:169
+; ALIGNED-NEXT:    buffer_store_byte v37, v0, s[0:3], 0 offen offset:168
+; ALIGNED-NEXT:    buffer_store_byte v38, v0, s[0:3], 0 offen offset:167
+; ALIGNED-NEXT:    buffer_store_byte v39, v0, s[0:3], 0 offen offset:166
+; ALIGNED-NEXT:    buffer_store_byte v48, v0, s[0:3], 0 offen offset:165
+; ALIGNED-NEXT:    buffer_store_byte v49, v0, s[0:3], 0 offen offset:164
+; ALIGNED-NEXT:    buffer_store_byte v50, v0, s[0:3], 0 offen offset:163
+; ALIGNED-NEXT:    buffer_store_byte v51, v0, s[0:3], 0 offen offset:162
+; ALIGNED-NEXT:    buffer_store_byte v52, v0, s[0:3], 0 offen offset:161
+; ALIGNED-NEXT:    buffer_store_byte v53, v0, s[0:3], 0 offen offset:160
+; ALIGNED-NEXT:    buffer_store_byte v54, v0, s[0:3], 0 offen offset:159
+; ALIGNED-NEXT:    buffer_store_byte v55, v0, s[0:3], 0 offen offset:158
+; ALIGNED-NEXT:    buffer_store_byte v64, v0, s[0:3], 0 offen offset:157
+; ALIGNED-NEXT:    buffer_store_byte v65, v0, s[0:3], 0 offen offset:156
+; ALIGNED-NEXT:    buffer_store_byte v66, v0, s[0:3], 0 offen offset:155
+; ALIGNED-NEXT:    buffer_store_byte v67, v0, s[0:3], 0 offen offset:154
+; ALIGNED-NEXT:    buffer_store_byte v68, v0, s[0:3], 0 offen offset:153
+; ALIGNED-NEXT:    buffer_store_byte v69, v0, s[0:3], 0 offen offset:152
+; ALIGNED-NEXT:    buffer_store_byte v70, v0, s[0:3], 0 offen offset:151
+; ALIGNED-NEXT:    buffer_store_byte v71, v0, s[0:3], 0 offen offset:150
+; ALIGNED-NEXT:    buffer_store_byte v80, v0, s[0:3], 0 offen offset:149
+; ALIGNED-NEXT:    buffer_store_byte v81, v0, s[0:3], 0 offen offset:148
+; ALIGNED-NEXT:    buffer_store_byte v82, v0, s[0:3], 0 offen offset:147
+; ALIGNED-NEXT:    buffer_store_byte v83, v0, s[0:3], 0 offen offset:146
+; ALIGNED-NEXT:    buffer_store_byte v84, v0, s[0:3], 0 offen offset:145
+; ALIGNED-NEXT:    buffer_store_byte v85, v0, s[0:3], 0 offen offset:144
+; ALIGNED-NEXT:    buffer_store_byte v86, v0, s[0:3], 0 offen offset:143
+; ALIGNED-NEXT:    buffer_store_byte v87, v0, s[0:3], 0 offen offset:142
+; ALIGNED-NEXT:    buffer_store_byte v96, v0, s[0:3], 0 offen offset:141
+; ALIGNED-NEXT:    buffer_store_byte v97, v0, s[0:3], 0 offen offset:140
+; ALIGNED-NEXT:    buffer_store_byte v98, v0, s[0:3], 0 offen offset:139
+; ALIGNED-NEXT:    buffer_store_byte v99, v0, s[0:3], 0 offen offset:138
+; ALIGNED-NEXT:    buffer_store_byte v100, v0, s[0:3], 0 offen offset:137
+; ALIGNED-NEXT:    buffer_store_byte v101, v0, s[0:3], 0 offen offset:136
+; ALIGNED-NEXT:    buffer_store_byte v102, v0, s[0:3], 0 offen offset:135
+; ALIGNED-NEXT:    buffer_store_byte v103, v0, s[0:3], 0 offen offset:134
+; ALIGNED-NEXT:    buffer_store_byte v112, v0, s[0:3], 0 offen offset:133
+; ALIGNED-NEXT:    buffer_store_byte v113, v0, s[0:3], 0 offen offset:132
+; ALIGNED-NEXT:    buffer_store_byte v114, v0, s[0:3], 0 offen offset:131
+; ALIGNED-NEXT:    buffer_store_byte v115, v0, s[0:3], 0 offen offset:130
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1921
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:129
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1920
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:128
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1919
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:127
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1918
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:126
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1917
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:125
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1916
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:124
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1915
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:123
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1914
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:122
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1913
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:121
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1912
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:120
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1911
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:119
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1910
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:118
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1909
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:117
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1908
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:116
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1907
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:115
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1906
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:114
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1905
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:113
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1904
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:112
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1903
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:111
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1902
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:110
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1901
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:109
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1900
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:108
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1899
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:107
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1898
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:106
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1897
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:105
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1896
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:104
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1895
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:103
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1894
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:102
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1893
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:101
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1892
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:100
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1891
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:99
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1890
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:98
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1889
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:97
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1888
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:96
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1887
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:95
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1886
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:94
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1885
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:93
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1884
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:92
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1883
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:91
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1882
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:90
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1881
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:89
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1880
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:88
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1879
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:87
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1878
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:86
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1877
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:85
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1876
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:84
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1875
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:83
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1874
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:82
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1873
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:81
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1872
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:80
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1871
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:79
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1870
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:78
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1869
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:77
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1868
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:76
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1867
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:75
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1866
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:74
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1865
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:73
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1864
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:72
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1863
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:71
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1862
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:70
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1861
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:69
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1860
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:68
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1859
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:67
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1858
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:66
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1857
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:65
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1856
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:64
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1855
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:63
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1854
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:62
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1853
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:61
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1852
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:60
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1851
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:59
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1850
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:58
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1849
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:57
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1848
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:56
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1847
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:55
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1846
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:54
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1845
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:53
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1844
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:52
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1843
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:51
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1842
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:50
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1841
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:49
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1840
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:48
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1839
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:47
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1838
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:46
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1837
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:45
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1836
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:44
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1835
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:43
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1834
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:42
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1833
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:41
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1832
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:40
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1831
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:39
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1830
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:38
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1829
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:37
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1828
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:36
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1827
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:35
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1826
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:34
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1825
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:33
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1824
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:32
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1823
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:31
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1822
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:30
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1821
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:29
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1820
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:28
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1819
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:27
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1818
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:26
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1817
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:25
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1816
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:24
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1815
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:23
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1814
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:22
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1813
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:21
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1812
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:20
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1811
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:19
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1810
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:18
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1809
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:17
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1808
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:16
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1807
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:15
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1806
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:14
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1805
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:13
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1804
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:12
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1803
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:11
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1802
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:10
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1801
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:9
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1800
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:8
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1799
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:7
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1798
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:6
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1797
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:5
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1796
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:4
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1795
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:3
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1794
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:2
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1793
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1
 ; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen offset:1792
+; ALIGNED-NEXT:    buffer_store_byte v2, v0, s[0:3], 0 offen
 ; ALIGNED-NEXT:    v_add_nc_u32_e32 v0, 0xffffff00, v0
 ; ALIGNED-NEXT:    s_cbranch_scc0 .LBB8_5
 ; ALIGNED-NEXT:  .LBB8_6: ; %Flow19
@@ -12425,61 +12404,63 @@ define void @memmove_p5_p5_sz2048(ptr addrspace(5) align 1 %dst, ptr addrspace(5
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(0)
 ; UNROLL3-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:2032
 ; UNROLL3-NEXT:    s_clause 0x3
-; UNROLL3-NEXT:    buffer_load_dword v2, v1, s[0:3], 0 offen offset:2028
-; UNROLL3-NEXT:    buffer_load_dword v3, v1, s[0:3], 0 offen offset:2024
-; UNROLL3-NEXT:    buffer_load_dword v4, v1, s[0:3], 0 offen offset:2020
-; UNROLL3-NEXT:    buffer_load_dword v5, v1, s[0:3], 0 offen offset:2016
+; UNROLL3-NEXT:    buffer_load_dword v3, v1, s[0:3], 0 offen offset:2028
+; UNROLL3-NEXT:    buffer_load_dword v4, v1, s[0:3], 0 offen offset:2024
+; UNROLL3-NEXT:    buffer_load_dword v5, v1, s[0:3], 0 offen offset:2020
+; UNROLL3-NEXT:    buffer_load_dword v6, v1, s[0:3], 0 offen offset:2016
+; UNROLL3-NEXT:    v_add_nc_u32_e32 v2, 0x7b0, v0
+; UNROLL3-NEXT:    v_add_nc_u32_e32 v1, 0x7b0, v1
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(3)
-; UNROLL3-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:2028
+; UNROLL3-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:2028
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(2)
-; UNROLL3-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:2024
+; UNROLL3-NEXT:    buffer_store_dword v4, v0, s[0:3], 0 offen offset:2024
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(1)
-; UNROLL3-NEXT:    buffer_store_dword v4, v0, s[0:3], 0 offen offset:2020
+; UNROLL3-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:2020
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(0)
-; UNROLL3-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:2016
+; UNROLL3-NEXT:    buffer_store_dword v6, v0, s[0:3], 0 offen offset:2016
 ; UNROLL3-NEXT:  .LBB8_6: ; %memmove_bwd_loop
 ; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; UNROLL3-NEXT:    s_clause 0xb
-; UNROLL3-NEXT:    buffer_load_dword v2, v1, s[0:3], 0 offen offset:2012
-; UNROLL3-NEXT:    buffer_load_dword v3, v1, s[0:3], 0 offen offset:2008
-; UNROLL3-NEXT:    buffer_load_dword v4, v1, s[0:3], 0 offen offset:2004
-; UNROLL3-NEXT:    buffer_load_dword v5, v1, s[0:3], 0 offen offset:2000
-; UNROLL3-NEXT:    buffer_load_dword v6, v1, s[0:3], 0 offen offset:1996
-; UNROLL3-NEXT:    buffer_load_dword v7, v1, s[0:3], 0 offen offset:1992
-; UNROLL3-NEXT:    buffer_load_dword v8, v1, s[0:3], 0 offen offset:1988
-; UNROLL3-NEXT:    buffer_load_dword v9, v1, s[0:3], 0 offen offset:1984
-; UNROLL3-NEXT:    buffer_load_dword v10, v1, s[0:3], 0 offen offset:1980
-; UNROLL3-NEXT:    buffer_load_dword v11, v1, s[0:3], 0 offen offset:1976
-; UNROLL3-NEXT:    buffer_load_dword v12, v1, s[0:3], 0 offen offset:1972
-; UNROLL3-NEXT:    buffer_load_dword v13, v1, s[0:3], 0 offen offset:1968
+; UNROLL3-NEXT:    buffer_load_dword v0, v1, s[0:3], 0 offen offset:44
+; UNROLL3-NEXT:    buffer_load_dword v3, v1, s[0:3], 0 offen offset:40
+; UNROLL3-NEXT:    buffer_load_dword v4, v1, s[0:3], 0 offen offset:36
+; UNROLL3-NEXT:    buffer_load_dword v5, v1, s[0:3], 0 offen offset:32
+; UNROLL3-NEXT:    buffer_load_dword v6, v1, s[0:3], 0 offen offset:28
+; UNROLL3-NEXT:    buffer_load_dword v7, v1, s[0:3], 0 offen offset:24
+; UNROLL3-NEXT:    buffer_load_dword v8, v1, s[0:3], 0 offen offset:20
+; UNROLL3-NEXT:    buffer_load_dword v9, v1, s[0:3], 0 offen offset:16
+; UNROLL3-NEXT:    buffer_load_dword v10, v1, s[0:3], 0 offen offset:12
+; UNROLL3-NEXT:    buffer_load_dword v11, v1, s[0:3], 0 offen offset:8
+; UNROLL3-NEXT:    buffer_load_dword v12, v1, s[0:3], 0 offen offset:4
+; UNROLL3-NEXT:    buffer_load_dword v13, v1, s[0:3], 0 offen
 ; UNROLL3-NEXT:    v_subrev_nc_u32_e32 v1, 48, v1
 ; UNROLL3-NEXT:    s_add_u32 s4, s4, 48
 ; UNROLL3-NEXT:    s_addc_u32 s5, s5, 0
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(11)
-; UNROLL3-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:2012
+; UNROLL3-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen offset:44
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(10)
-; UNROLL3-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:2008
+; UNROLL3-NEXT:    buffer_store_dword v3, v2, s[0:3], 0 offen offset:40
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(9)
-; UNROLL3-NEXT:    buffer_store_dword v4, v0, s[0:3], 0 offen offset:2004
+; UNROLL3-NEXT:    buffer_store_dword v4, v2, s[0:3], 0 offen offset:36
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(8)
-; UNROLL3-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:2000
+; UNROLL3-NEXT:    buffer_store_dword v5, v2, s[0:3], 0 offen offset:32
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(7)
-; UNROLL3-NEXT:    buffer_store_dword v6, v0, s[0:3], 0 offen offset:1996
+; UNROLL3-NEXT:    buffer_store_dword v6, v2, s[0:3], 0 offen offset:28
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(6)
-; UNROLL3-NEXT:    buffer_store_dword v7, v0, s[0:3], 0 offen offset:1992
+; UNROLL3-NEXT:    buffer_store_dword v7, v2, s[0:3], 0 offen offset:24
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(5)
-; UNROLL3-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:1988
+; UNROLL3-NEXT:    buffer_store_dword v8, v2, s[0:3], 0 offen offset:20
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(4)
-; UNROLL3-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:1984
+; UNROLL3-NEXT:    buffer_store_dword v9, v2, s[0:3], 0 offen offset:16
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(3)
-; UNROLL3-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:1980
+; UNROLL3-NEXT:    buffer_store_dword v10, v2, s[0:3], 0 offen offset:12
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(2)
-; UNROLL3-NEXT:    buffer_store_dword v11, v0, s[0:3], 0 offen offset:1976
+; UNROLL3-NEXT:    buffer_store_dword v11, v2, s[0:3], 0 offen offset:8
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(1)
-; UNROLL3-NEXT:    buffer_store_dword v12, v0, s[0:3], 0 offen offset:1972
+; UNROLL3-NEXT:    buffer_store_dword v12, v2, s[0:3], 0 offen offset:4
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(0)
-; UNROLL3-NEXT:    buffer_store_dword v13, v0, s[0:3], 0 offen offset:1968
-; UNROLL3-NEXT:    v_subrev_nc_u32_e32 v0, 48, v0
+; UNROLL3-NEXT:    buffer_store_dword v13, v2, s[0:3], 0 offen
+; UNROLL3-NEXT:    v_subrev_nc_u32_e32 v2, 48, v2
 ; UNROLL3-NEXT:    s_cmp_eq_u64 s[4:5], 0
 ; UNROLL3-NEXT:    s_cbranch_scc0 .LBB8_6
 ; UNROLL3-NEXT:  .LBB8_7: ; %Flow17
@@ -12495,225 +12476,222 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; CHECK:       ; %bb.0: ; %entry
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    v_cmp_ne_u64_e32 vcc_lo, 0, v[0:1]
-; CHECK-NEXT:    s_mov_b32 s4, exec_lo
+; CHECK-NEXT:    s_mov_b64 s[4:5], 0
+; CHECK-NEXT:    s_mov_b32 s6, exec_lo
 ; CHECK-NEXT:    v_cndmask_b32_e32 v3, -1, v0, vcc_lo
 ; CHECK-NEXT:    v_cmpx_ge_u32_e32 v2, v3
-; CHECK-NEXT:    s_xor_b32 s6, exec_lo, s4
-; CHECK-NEXT:    s_cbranch_execz .LBB9_3
-; CHECK-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader
-; CHECK-NEXT:    s_mov_b64 s[4:5], 0x800
-; CHECK-NEXT:  .LBB9_2: ; %memmove_fwd_loop
+; CHECK-NEXT:    s_xor_b32 s6, exec_lo, s6
+; CHECK-NEXT:    s_cbranch_execz .LBB9_2
+; CHECK-NEXT:  .LBB9_1: ; %memmove_fwd_loop
 ; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; CHECK-NEXT:    s_clause 0x3e
-; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen
-; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:4
-; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:8
-; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:12
-; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:16
-; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:20
-; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:24
-; CHECK-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:28
-; CHECK-NEXT:    buffer_load_dword v11, v2, s[0:3], 0 offen offset:32
-; CHECK-NEXT:    buffer_load_dword v12, v2, s[0:3], 0 offen offset:36
-; CHECK-NEXT:    buffer_load_dword v13, v2, s[0:3], 0 offen offset:40
-; CHECK-NEXT:    buffer_load_dword v14, v2, s[0:3], 0 offen offset:44
-; CHECK-NEXT:    buffer_load_dword v18, v2, s[0:3], 0 offen offset:76
-; CHECK-NEXT:    buffer_load_dword v22, v2, s[0:3], 0 offen offset:92
-; CHECK-NEXT:    buffer_load_dword v21, v2, s[0:3], 0 offen offset:88
-; CHECK-NEXT:    buffer_load_dword v20, v2, s[0:3], 0 offen offset:84
-; CHECK-NEXT:    buffer_load_dword v19, v2, s[0:3], 0 offen offset:80
-; CHECK-NEXT:    buffer_load_dword v17, v2, s[0:3], 0 offen offset:72
-; CHECK-NEXT:    buffer_load_dword v16, v2, s[0:3], 0 offen offset:68
-; CHECK-NEXT:    buffer_load_dword v15, v2, s[0:3], 0 offen offset:64
-; CHECK-NEXT:    buffer_load_dword v26, v2, s[0:3], 0 offen offset:172
-; CHECK-NEXT:    buffer_load_dword v30, v2, s[0:3], 0 offen offset:188
-; CHECK-NEXT:    buffer_load_dword v29, v2, s[0:3], 0 offen offset:184
-; CHECK-NEXT:    buffer_load_dword v28, v2, s[0:3], 0 offen offset:180
-; CHECK-NEXT:    buffer_load_dword v27, v2, s[0:3], 0 offen offset:176
-; CHECK-NEXT:    buffer_load_dword v25, v2, s[0:3], 0 offen offset:168
-; CHECK-NEXT:    buffer_load_dword v24, v2, s[0:3], 0 offen offset:164
-; CHECK-NEXT:    buffer_load_dword v23, v2, s[0:3], 0 offen offset:160
-; CHECK-NEXT:    buffer_load_dword v34, v2, s[0:3], 0 offen offset:204
-; CHECK-NEXT:    buffer_load_dword v38, v2, s[0:3], 0 offen offset:220
-; CHECK-NEXT:    buffer_load_dword v37, v2, s[0:3], 0 offen offset:216
-; CHECK-NEXT:    buffer_load_dword v36, v2, s[0:3], 0 offen offset:212
-; CHECK-NEXT:    buffer_load_dword v35, v2, s[0:3], 0 offen offset:208
-; CHECK-NEXT:    buffer_load_dword v33, v2, s[0:3], 0 offen offset:200
-; CHECK-NEXT:    buffer_load_dword v32, v2, s[0:3], 0 offen offset:196
-; CHECK-NEXT:    buffer_load_dword v31, v2, s[0:3], 0 offen offset:192
-; CHECK-NEXT:    buffer_load_dword v51, v2, s[0:3], 0 offen offset:236
-; CHECK-NEXT:    buffer_load_dword v55, v2, s[0:3], 0 offen offset:252
-; CHECK-NEXT:    buffer_load_dword v54, v2, s[0:3], 0 offen offset:248
-; CHECK-NEXT:    buffer_load_dword v53, v2, s[0:3], 0 offen offset:244
-; CHECK-NEXT:    buffer_load_dword v52, v2, s[0:3], 0 offen offset:240
-; CHECK-NEXT:    buffer_load_dword v50, v2, s[0:3], 0 offen offset:232
-; CHECK-NEXT:    buffer_load_dword v49, v2, s[0:3], 0 offen offset:228
-; CHECK-NEXT:    buffer_load_dword v48, v2, s[0:3], 0 offen offset:224
-; CHECK-NEXT:    buffer_load_dword v67, v2, s[0:3], 0 offen offset:140
+; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:32
+; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:36
+; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:40
+; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:44
+; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:48
+; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:52
+; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:56
+; CHECK-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:60
+; CHECK-NEXT:    buffer_load_dword v14, v2, s[0:3], 0 offen offset:76
+; CHECK-NEXT:    buffer_load_dword v18, v2, s[0:3], 0 offen offset:92
+; CHECK-NEXT:    buffer_load_dword v17, v2, s[0:3], 0 offen offset:88
+; CHECK-NEXT:    buffer_load_dword v16, v2, s[0:3], 0 offen offset:84
+; CHECK-NEXT:    buffer_load_dword v15, v2, s[0:3], 0 offen offset:80
+; CHECK-NEXT:    buffer_load_dword v13, v2, s[0:3], 0 offen offset:72
+; CHECK-NEXT:    buffer_load_dword v12, v2, s[0:3], 0 offen offset:68
+; CHECK-NEXT:    buffer_load_dword v11, v2, s[0:3], 0 offen offset:64
+; CHECK-NEXT:    buffer_load_dword v22, v2, s[0:3], 0 offen offset:108
+; CHECK-NEXT:    buffer_load_dword v26, v2, s[0:3], 0 offen offset:124
+; CHECK-NEXT:    buffer_load_dword v25, v2, s[0:3], 0 offen offset:120
+; CHECK-NEXT:    buffer_load_dword v24, v2, s[0:3], 0 offen offset:116
+; CHECK-NEXT:    buffer_load_dword v23, v2, s[0:3], 0 offen offset:112
+; CHECK-NEXT:    buffer_load_dword v21, v2, s[0:3], 0 offen offset:104
+; CHECK-NEXT:    buffer_load_dword v20, v2, s[0:3], 0 offen offset:100
+; CHECK-NEXT:    buffer_load_dword v19, v2, s[0:3], 0 offen offset:96
+; CHECK-NEXT:    buffer_load_dword v30, v2, s[0:3], 0 offen offset:172
+; CHECK-NEXT:    buffer_load_dword v34, v2, s[0:3], 0 offen offset:188
+; CHECK-NEXT:    buffer_load_dword v33, v2, s[0:3], 0 offen offset:184
+; CHECK-NEXT:    buffer_load_dword v32, v2, s[0:3], 0 offen offset:180
+; CHECK-NEXT:    buffer_load_dword v31, v2, s[0:3], 0 offen offset:176
+; CHECK-NEXT:    buffer_load_dword v29, v2, s[0:3], 0 offen offset:168
+; CHECK-NEXT:    buffer_load_dword v28, v2, s[0:3], 0 offen offset:164
+; CHECK-NEXT:    buffer_load_dword v27, v2, s[0:3], 0 offen offset:160
+; CHECK-NEXT:    buffer_load_dword v38, v2, s[0:3], 0 offen offset:204
+; CHECK-NEXT:    buffer_load_dword v51, v2, s[0:3], 0 offen offset:220
+; CHECK-NEXT:    buffer_load_dword v50, v2, s[0:3], 0 offen offset:216
+; CHECK-NEXT:    buffer_load_dword v49, v2, s[0:3], 0 offen offset:212
+; CHECK-NEXT:    buffer_load_dword v48, v2, s[0:3], 0 offen offset:208
+; CHECK-NEXT:    buffer_load_dword v37, v2, s[0:3], 0 offen offset:200
+; CHECK-NEXT:    buffer_load_dword v36, v2, s[0:3], 0 offen offset:196
+; CHECK-NEXT:    buffer_load_dword v35, v2, s[0:3], 0 offen offset:192
+; CHECK-NEXT:    buffer_load_dword v55, v2, s[0:3], 0 offen offset:236
+; CHECK-NEXT:    buffer_load_dword v67, v2, s[0:3], 0 offen offset:252
+; CHECK-NEXT:    buffer_load_dword v66, v2, s[0:3], 0 offen offset:248
+; CHECK-NEXT:    buffer_load_dword v65, v2, s[0:3], 0 offen offset:244
+; CHECK-NEXT:    buffer_load_dword v64, v2, s[0:3], 0 offen offset:240
+; CHECK-NEXT:    buffer_load_dword v54, v2, s[0:3], 0 offen offset:232
+; CHECK-NEXT:    buffer_load_dword v53, v2, s[0:3], 0 offen offset:228
+; CHECK-NEXT:    buffer_load_dword v52, v2, s[0:3], 0 offen offset:224
+; CHECK-NEXT:    buffer_load_dword v71, v2, s[0:3], 0 offen offset:140
+; CHECK-NEXT:    buffer_load_dword v83, v2, s[0:3], 0 offen offset:156
+; CHECK-NEXT:    buffer_load_dword v82, v2, s[0:3], 0 offen offset:152
+; CHECK-NEXT:    buffer_load_dword v81, v2, s[0:3], 0 offen offset:148
+; CHECK-NEXT:    buffer_load_dword v80, v2, s[0:3], 0 offen offset:144
+; CHECK-NEXT:    buffer_load_dword v70, v2, s[0:3], 0 offen offset:136
+; CHECK-NEXT:    buffer_load_dword v69, v2, s[0:3], 0 offen offset:132
+; CHECK-NEXT:    buffer_load_dword v68, v2, s[0:3], 0 offen offset:128
+; CHECK-NEXT:    buffer_load_dword v84, v2, s[0:3], 0 offen
+; CHECK-NEXT:    buffer_load_dword v85, v2, s[0:3], 0 offen offset:4
+; CHECK-NEXT:    buffer_load_dword v86, v2, s[0:3], 0 offen offset:8
+; CHECK-NEXT:    buffer_load_dword v96, v2, s[0:3], 0 offen offset:16
+; CHECK-NEXT:    buffer_load_dword v97, v2, s[0:3], 0 offen offset:20
+; CHECK-NEXT:    buffer_load_dword v98, v2, s[0:3], 0 offen offset:24
+; CHECK-NEXT:    buffer_load_dword v99, v2, s[0:3], 0 offen offset:28
+; CHECK-NEXT:    buffer_load_dword v87, v2, s[0:3], 0 offen offset:12
+; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s4
+; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo
+; CHECK-NEXT:    s_add_u32 s4, s4, 0x100
+; CHECK-NEXT:    v_add_co_u32 v102, vcc_lo, v100, 48
+; CHECK-NEXT:    v_add_nc_u32_e32 v2, 0x100, v2
+; CHECK-NEXT:    v_add_co_ci_u32_e64 v103, null, 0, v101, vcc_lo
+; CHECK-NEXT:    s_addc_u32 s5, s5, 0
+; CHECK-NEXT:    s_waitcnt vmcnt(19)
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[64:67] offset:192
+; CHECK-NEXT:    s_waitcnt vmcnt(16)
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[52:55] offset:176
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[48:51] offset:160
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[35:38] offset:144
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[31:34] offset:128
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[27:30] offset:112
+; CHECK-NEXT:    s_waitcnt vmcnt(11)
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[80:83] offset:96
+; CHECK-NEXT:    s_waitcnt vmcnt(8)
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[68:71] offset:128
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[23:26] offset:64
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[19:22] offset:48
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[15:18] offset:32
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[11:14] offset:64
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[7:10]
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[3:6] offset:32
+; CHECK-NEXT:    s_waitcnt vmcnt(1)
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[96:99] offset:16
+; CHECK-NEXT:    s_waitcnt vmcnt(0)
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[84:87]
+; CHECK-NEXT:    s_cmp_lg_u64 s[4:5], 0x800
+; CHECK-NEXT:    s_cbranch_scc1 .LBB9_1
+; CHECK-NEXT:  .LBB9_2: ; %Flow10
+; CHECK-NEXT:    s_andn2_saveexec_b32 s8, s6
+; CHECK-NEXT:    s_cbranch_execz .LBB9_5
+; CHECK-NEXT:  ; %bb.3: ; %memmove_bwd_loop.preheader
+; CHECK-NEXT:    v_add_nc_u32_e32 v2, 0x700, v2
+; CHECK-NEXT:    s_mov_b64 s[4:5], 0x700
+; CHECK-NEXT:    s_movk_i32 s6, 0xff00
+; CHECK-NEXT:    s_mov_b32 s7, -1
+; CHECK-NEXT:  .LBB9_4: ; %memmove_bwd_loop
+; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    s_clause 0x3e
+; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:32
+; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:36
+; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:40
+; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:44
+; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:48
+; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:52
+; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:56
+; CHECK-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:60
+; CHECK-NEXT:    buffer_load_dword v14, v2, s[0:3], 0 offen offset:76
+; CHECK-NEXT:    buffer_load_dword v18, v2, s[0:3], 0 offen offset:92
+; CHECK-NEXT:    buffer_load_dword v17, v2, s[0:3], 0 offen offset:88
+; CHECK-NEXT:    buffer_load_dword v16, v2, s[0:3], 0 offen offset:84
+; CHECK-NEXT:    buffer_load_dword v15, v2, s[0:3], 0 offen offset:80
+; CHECK-NEXT:    buffer_load_dword v13, v2, s[0:3], 0 offen offset:72
+; CHECK-NEXT:    buffer_load_dword v12, v2, s[0:3], 0 offen offset:68
+; CHECK-NEXT:    buffer_load_dword v11, v2, s[0:3], 0 offen offset:64
+; CHECK-NEXT:    buffer_load_dword v22, v2, s[0:3], 0 offen offset:108
+; CHECK-NEXT:    buffer_load_dword v26, v2, s[0:3], 0 offen offset:124
+; CHECK-NEXT:    buffer_load_dword v25, v2, s[0:3], 0 offen offset:120
+; CHECK-NEXT:    buffer_load_dword v24, v2, s[0:3], 0 offen offset:116
+; CHECK-NEXT:    buffer_load_dword v23, v2, s[0:3], 0 offen offset:112
+; CHECK-NEXT:    buffer_load_dword v21, v2, s[0:3], 0 offen offset:104
+; CHECK-NEXT:    buffer_load_dword v20, v2, s[0:3], 0 offen offset:100
+; CHECK-NEXT:    buffer_load_dword v19, v2, s[0:3], 0 offen offset:96
+; CHECK-NEXT:    buffer_load_dword v30, v2, s[0:3], 0 offen offset:236
+; CHECK-NEXT:    buffer_load_dword v34, v2, s[0:3], 0 offen offset:252
+; CHECK-NEXT:    buffer_load_dword v33, v2, s[0:3], 0 offen offset:248
+; CHECK-NEXT:    buffer_load_dword v32, v2, s[0:3], 0 offen offset:244
+; CHECK-NEXT:    buffer_load_dword v31, v2, s[0:3], 0 offen offset:240
+; CHECK-NEXT:    buffer_load_dword v29, v2, s[0:3], 0 offen offset:232
+; CHECK-NEXT:    buffer_load_dword v28, v2, s[0:3], 0 offen offset:228
+; CHECK-NEXT:    buffer_load_dword v27, v2, s[0:3], 0 offen offset:224
+; CHECK-NEXT:    buffer_load_dword v38, v2, s[0:3], 0 offen offset:204
+; CHECK-NEXT:    buffer_load_dword v51, v2, s[0:3], 0 offen offset:220
+; CHECK-NEXT:    buffer_load_dword v50, v2, s[0:3], 0 offen offset:216
+; CHECK-NEXT:    buffer_load_dword v49, v2, s[0:3], 0 offen offset:212
+; CHECK-NEXT:    buffer_load_dword v48, v2, s[0:3], 0 offen offset:208
+; CHECK-NEXT:    buffer_load_dword v37, v2, s[0:3], 0 offen offset:200
+; CHECK-NEXT:    buffer_load_dword v36, v2, s[0:3], 0 offen offset:196
+; CHECK-NEXT:    buffer_load_dword v35, v2, s[0:3], 0 offen offset:192
+; CHECK-NEXT:    buffer_load_dword v55, v2, s[0:3], 0 offen offset:172
+; CHECK-NEXT:    buffer_load_dword v67, v2, s[0:3], 0 offen offset:188
+; CHECK-NEXT:    buffer_load_dword v66, v2, s[0:3], 0 offen offset:184
+; CHECK-NEXT:    buffer_load_dword v65, v2, s[0:3], 0 offen offset:180
+; CHECK-NEXT:    buffer_load_dword v64, v2, s[0:3], 0 offen offset:176
+; CHECK-NEXT:    buffer_load_dword v54, v2, s[0:3], 0 offen offset:168
+; CHECK-NEXT:    buffer_load_dword v53, v2, s[0:3], 0 offen offset:164
+; CHECK-NEXT:    buffer_load_dword v52, v2, s[0:3], 0 offen offset:160
 ; CHECK-NEXT:    buffer_load_dword v71, v2, s[0:3], 0 offen offset:156
 ; CHECK-NEXT:    buffer_load_dword v70, v2, s[0:3], 0 offen offset:152
 ; CHECK-NEXT:    buffer_load_dword v69, v2, s[0:3], 0 offen offset:148
 ; CHECK-NEXT:    buffer_load_dword v68, v2, s[0:3], 0 offen offset:144
-; CHECK-NEXT:    buffer_load_dword v66, v2, s[0:3], 0 offen offset:136
-; CHECK-NEXT:    buffer_load_dword v65, v2, s[0:3], 0 offen offset:132
-; CHECK-NEXT:    buffer_load_dword v64, v2, s[0:3], 0 offen offset:128
-; CHECK-NEXT:    buffer_load_dword v83, v2, s[0:3], 0 offen offset:108
-; CHECK-NEXT:    buffer_load_dword v87, v2, s[0:3], 0 offen offset:124
-; CHECK-NEXT:    buffer_load_dword v86, v2, s[0:3], 0 offen offset:120
-; CHECK-NEXT:    buffer_load_dword v85, v2, s[0:3], 0 offen offset:116
-; CHECK-NEXT:    buffer_load_dword v84, v2, s[0:3], 0 offen offset:112
-; CHECK-NEXT:    buffer_load_dword v82, v2, s[0:3], 0 offen offset:104
-; CHECK-NEXT:    buffer_load_dword v81, v2, s[0:3], 0 offen offset:100
-; CHECK-NEXT:    buffer_load_dword v80, v2, s[0:3], 0 offen offset:96
-; CHECK-NEXT:    buffer_load_dword v96, v2, s[0:3], 0 offen offset:48
-; CHECK-NEXT:    buffer_load_dword v97, v2, s[0:3], 0 offen offset:52
-; CHECK-NEXT:    buffer_load_dword v98, v2, s[0:3], 0 offen offset:56
-; CHECK-NEXT:    buffer_load_dword v99, v2, s[0:3], 0 offen offset:60
-; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, 48
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, 0, v1, vcc_lo
-; CHECK-NEXT:    v_add_nc_u32_e32 v2, 0x100, v2
+; CHECK-NEXT:    buffer_load_dword v83, v2, s[0:3], 0 offen offset:140
+; CHECK-NEXT:    buffer_load_dword v82, v2, s[0:3], 0 offen offset:136
+; CHECK-NEXT:    buffer_load_dword v81, v2, s[0:3], 0 offen offset:132
+; CHECK-NEXT:    buffer_load_dword v80, v2, s[0:3], 0 offen offset:128
+; CHECK-NEXT:    buffer_load_dword v84, v2, s[0:3], 0 offen
+; CHECK-NEXT:    buffer_load_dword v85, v2, s[0:3], 0 offen offset:4
+; CHECK-NEXT:    buffer_load_dword v86, v2, s[0:3], 0 offen offset:8
+; CHECK-NEXT:    buffer_load_dword v96, v2, s[0:3], 0 offen offset:16
+; CHECK-NEXT:    buffer_load_dword v97, v2, s[0:3], 0 offen offset:20
+; CHECK-NEXT:    buffer_load_dword v98, v2, s[0:3], 0 offen offset:24
+; CHECK-NEXT:    buffer_load_dword v99, v2, s[0:3], 0 offen offset:28
+; CHECK-NEXT:    buffer_load_dword v87, v2, s[0:3], 0 offen offset:12
+; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, s4
+; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, s5, v1, vcc_lo
+; CHECK-NEXT:    v_add_nc_u32_e32 v2, 0xffffff00, v2
+; CHECK-NEXT:    v_add_co_u32 v102, vcc_lo, v100, 48
 ; CHECK-NEXT:    s_add_u32 s4, s4, 0xffffff00
+; CHECK-NEXT:    v_add_co_ci_u32_e64 v103, null, 0, v101, vcc_lo
 ; CHECK-NEXT:    s_addc_u32 s5, s5, -1
-; CHECK-NEXT:    s_waitcnt vmcnt(23)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[52:55] offset:192
-; CHECK-NEXT:    s_waitcnt vmcnt(20)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[48:51] offset:176
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[35:38] offset:160
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[31:34] offset:144
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[27:30] offset:128
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[23:26] offset:112
-; CHECK-NEXT:    s_waitcnt vmcnt(15)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[68:71] offset:96
-; CHECK-NEXT:    s_waitcnt vmcnt(12)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[64:67] offset:128
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[15:18] offset:64
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[11:14] offset:32
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[7:10] offset:16
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[3:6]
-; CHECK-NEXT:    v_add_co_u32 v0, vcc_lo, 0x100, v0
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; CHECK-NEXT:    s_cmp_lg_u64 s[4:5], 0
-; CHECK-NEXT:    s_waitcnt vmcnt(7)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[84:87] offset:64
-; CHECK-NEXT:    s_waitcnt vmcnt(4)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[80:83] offset:48
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[19:22] offset:32
-; CHECK-NEXT:    s_waitcnt vmcnt(0)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[96:99]
-; CHECK-NEXT:    s_cbranch_scc1 .LBB9_2
-; CHECK-NEXT:  .LBB9_3: ; %Flow16
-; CHECK-NEXT:    s_andn2_saveexec_b32 s6, s6
-; CHECK-NEXT:    s_cbranch_execz .LBB9_6
-; CHECK-NEXT:  ; %bb.4: ; %memmove_bwd_loop.preheader
-; CHECK-NEXT:    v_add_co_u32 v0, vcc_lo, 0x700, v0
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
-; CHECK-NEXT:    s_movk_i32 s4, 0xf800
-; CHECK-NEXT:    s_mov_b32 s5, -1
-; CHECK-NEXT:  .LBB9_5: ; %memmove_bwd_loop
-; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    s_clause 0x3e
-; CHECK-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:1792
-; CHECK-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:1796
-; CHECK-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:1800
-; CHECK-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:1804
-; CHECK-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:1808
-; CHECK-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:1812
-; CHECK-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:1816
-; CHECK-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:1820
-; CHECK-NEXT:    buffer_load_dword v11, v2, s[0:3], 0 offen offset:1824
-; CHECK-NEXT:    buffer_load_dword v12, v2, s[0:3], 0 offen offset:1828
-; CHECK-NEXT:    buffer_load_dword v13, v2, s[0:3], 0 offen offset:1832
-; CHECK-NEXT:    buffer_load_dword v14, v2, s[0:3], 0 offen offset:1836
-; CHECK-NEXT:    buffer_load_dword v18, v2, s[0:3], 0 offen offset:1868
-; CHECK-NEXT:    buffer_load_dword v22, v2, s[0:3], 0 offen offset:1884
-; CHECK-NEXT:    buffer_load_dword v21, v2, s[0:3], 0 offen offset:1880
-; CHECK-NEXT:    buffer_load_dword v20, v2, s[0:3], 0 offen offset:1876
-; CHECK-NEXT:    buffer_load_dword v19, v2, s[0:3], 0 offen offset:1872
-; CHECK-NEXT:    buffer_load_dword v17, v2, s[0:3], 0 offen offset:1864
-; CHECK-NEXT:    buffer_load_dword v16, v2, s[0:3], 0 offen offset:1860
-; CHECK-NEXT:    buffer_load_dword v15, v2, s[0:3], 0 offen offset:1856
-; CHECK-NEXT:    buffer_load_dword v26, v2, s[0:3], 0 offen offset:2028
-; CHECK-NEXT:    buffer_load_dword v30, v2, s[0:3], 0 offen offset:2044
-; CHECK-NEXT:    buffer_load_dword v29, v2, s[0:3], 0 offen offset:2040
-; CHECK-NEXT:    buffer_load_dword v28, v2, s[0:3], 0 offen offset:2036
-; CHECK-NEXT:    buffer_load_dword v27, v2, s[0:3], 0 offen offset:2032
-; CHECK-NEXT:    buffer_load_dword v25, v2, s[0:3], 0 offen offset:2024
-; CHECK-NEXT:    buffer_load_dword v24, v2, s[0:3], 0 offen offset:2020
-; CHECK-NEXT:    buffer_load_dword v23, v2, s[0:3], 0 offen offset:2016
-; CHECK-NEXT:    buffer_load_dword v34, v2, s[0:3], 0 offen offset:1996
-; CHECK-NEXT:    buffer_load_dword v38, v2, s[0:3], 0 offen offset:2012
-; CHECK-NEXT:    buffer_load_dword v37, v2, s[0:3], 0 offen offset:2008
-; CHECK-NEXT:    buffer_load_dword v36, v2, s[0:3], 0 offen offset:2004
-; CHECK-NEXT:    buffer_load_dword v35, v2, s[0:3], 0 offen offset:2000
-; CHECK-NEXT:    buffer_load_dword v33, v2, s[0:3], 0 offen offset:1992
-; CHECK-NEXT:    buffer_load_dword v32, v2, s[0:3], 0 offen offset:1988
-; CHECK-NEXT:    buffer_load_dword v31, v2, s[0:3], 0 offen offset:1984
-; CHECK-NEXT:    buffer_load_dword v51, v2, s[0:3], 0 offen offset:1964
-; CHECK-NEXT:    buffer_load_dword v55, v2, s[0:3], 0 offen offset:1980
-; CHECK-NEXT:    buffer_load_dword v54, v2, s[0:3], 0 offen offset:1976
-; CHECK-NEXT:    buffer_load_dword v53, v2, s[0:3], 0 offen offset:1972
-; CHECK-NEXT:    buffer_load_dword v52, v2, s[0:3], 0 offen offset:1968
-; CHECK-NEXT:    buffer_load_dword v50, v2, s[0:3], 0 offen offset:1960
-; CHECK-NEXT:    buffer_load_dword v49, v2, s[0:3], 0 offen offset:1956
-; CHECK-NEXT:    buffer_load_dword v48, v2, s[0:3], 0 offen offset:1952
-; CHECK-NEXT:    buffer_load_dword v67, v2, s[0:3], 0 offen offset:1948
-; CHECK-NEXT:    buffer_load_dword v66, v2, s[0:3], 0 offen offset:1944
-; CHECK-NEXT:    buffer_load_dword v65, v2, s[0:3], 0 offen offset:1940
-; CHECK-NEXT:    buffer_load_dword v64, v2, s[0:3], 0 offen offset:1936
-; CHECK-NEXT:    buffer_load_dword v71, v2, s[0:3], 0 offen offset:1932
-; CHECK-NEXT:    buffer_load_dword v70, v2, s[0:3], 0 offen offset:1928
-; CHECK-NEXT:    buffer_load_dword v69, v2, s[0:3], 0 offen offset:1924
-; CHECK-NEXT:    buffer_load_dword v68, v2, s[0:3], 0 offen offset:1920
-; CHECK-NEXT:    buffer_load_dword v83, v2, s[0:3], 0 offen offset:1916
-; CHECK-NEXT:    buffer_load_dword v82, v2, s[0:3], 0 offen offset:1912
-; CHECK-NEXT:    buffer_load_dword v81, v2, s[0:3], 0 offen offset:1908
-; CHECK-NEXT:    buffer_load_dword v80, v2, s[0:3], 0 offen offset:1904
-; CHECK-NEXT:    buffer_load_dword v87, v2, s[0:3], 0 offen offset:1900
-; CHECK-NEXT:    buffer_load_dword v86, v2, s[0:3], 0 offen offset:1896
-; CHECK-NEXT:    buffer_load_dword v85, v2, s[0:3], 0 offen offset:1892
-; CHECK-NEXT:    buffer_load_dword v84, v2, s[0:3], 0 offen offset:1888
-; CHECK-NEXT:    buffer_load_dword v96, v2, s[0:3], 0 offen offset:1840
-; CHECK-NEXT:    buffer_load_dword v97, v2, s[0:3], 0 offen offset:1844
-; CHECK-NEXT:    buffer_load_dword v98, v2, s[0:3], 0 offen offset:1848
-; CHECK-NEXT:    buffer_load_dword v99, v2, s[0:3], 0 offen offset:1852
-; CHECK-NEXT:    v_add_co_u32 v100, vcc_lo, v0, 48
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v101, null, 0, v1, vcc_lo
-; CHECK-NEXT:    v_add_nc_u32_e32 v2, 0xffffff00, v2
-; CHECK-NEXT:    s_add_u32 s4, s4, 0x100
-; CHECK-NEXT:    s_addc_u32 s5, s5, 0
-; CHECK-NEXT:    s_waitcnt vmcnt(39)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[27:30] offset:192
-; CHECK-NEXT:    s_waitcnt vmcnt(36)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[23:26] offset:176
-; CHECK-NEXT:    s_waitcnt vmcnt(31)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[35:38] offset:160
-; CHECK-NEXT:    s_waitcnt vmcnt(28)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[31:34] offset:144
-; CHECK-NEXT:    s_waitcnt vmcnt(23)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[52:55] offset:128
-; CHECK-NEXT:    s_waitcnt vmcnt(20)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[48:51] offset:112
+; CHECK-NEXT:    s_waitcnt vmcnt(35)
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[31:34] offset:192
+; CHECK-NEXT:    s_waitcnt vmcnt(32)
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[27:30] offset:176
+; CHECK-NEXT:    s_waitcnt vmcnt(27)
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[48:51] offset:160
+; CHECK-NEXT:    s_waitcnt vmcnt(24)
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[35:38] offset:144
+; CHECK-NEXT:    s_waitcnt vmcnt(19)
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[64:67] offset:128
 ; CHECK-NEXT:    s_waitcnt vmcnt(16)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[64:67] offset:96
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[52:55] offset:112
 ; CHECK-NEXT:    s_waitcnt vmcnt(12)
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[68:71] offset:128
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[15:18] offset:64
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[11:14] offset:32
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[7:10] offset:16
-; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[3:6]
-; CHECK-NEXT:    v_add_co_u32 v0, vcc_lo, 0xffffff00, v0
-; CHECK-NEXT:    v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo
-; CHECK-NEXT:    s_cmp_eq_u64 s[4:5], 0
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[68:71] offset:96
 ; CHECK-NEXT:    s_waitcnt vmcnt(8)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[80:83] offset:64
-; CHECK-NEXT:    s_waitcnt vmcnt(4)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[84:87] offset:48
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[19:22] offset:32
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[80:83] offset:128
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[23:26] offset:64
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[19:22] offset:48
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[15:18] offset:32
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[11:14] offset:64
+; CHECK-NEXT:    flat_store_dwordx4 v[102:103], v[7:10]
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[3:6] offset:32
+; CHECK-NEXT:    s_waitcnt vmcnt(1)
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[96:99] offset:16
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
-; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[96:99]
-; CHECK-NEXT:    s_cbranch_scc0 .LBB9_5
-; CHECK-NEXT:  .LBB9_6: ; %Flow17
-; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s6
+; CHECK-NEXT:    flat_store_dwordx4 v[100:101], v[84:87]
+; CHECK-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]
+; CHECK-NEXT:    s_cbranch_scc0 .LBB9_4
+; CHECK-NEXT:  .LBB9_5: ; %Flow11
+; CHECK-NEXT:    s_or_b32 exec_lo, exec_lo, s8
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
 ;
@@ -12768,881 +12746,898 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_store_dword v125, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v126, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v127, off, s[0:3], s32 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_mov_b32_e32 v88, v1
-; ALIGNED-NEXT:    v_mov_b32_e32 v87, v0
-; ALIGNED-NEXT:    s_mov_b32 s4, exec_lo
-; ALIGNED-NEXT:    v_cmp_ne_u64_e32 vcc_lo, 0, v[87:88]
-; ALIGNED-NEXT:    v_cndmask_b32_e32 v3, -1, v87, vcc_lo
-; ALIGNED-NEXT:    v_cmpx_ge_u32_e32 v2, v3
-; ALIGNED-NEXT:    s_xor_b32 s6, exec_lo, s4
-; ALIGNED-NEXT:    s_cbranch_execz .LBB9_3
-; ALIGNED-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader
-; ALIGNED-NEXT:    s_mov_b64 s[4:5], 0x800
-; ALIGNED-NEXT:  .LBB9_2: ; %memmove_fwd_loop
+; ALIGNED-NEXT:    v_cmp_ne_u64_e32 vcc_lo, 0, v[0:1]
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_mov_b64 s[4:5], 0
+; ALIGNED-NEXT:    s_mov_b32 s6, exec_lo
+; ALIGNED-NEXT:    v_cndmask_b32_e32 v0, -1, v0, vcc_lo
+; ALIGNED-NEXT:    v_cmpx_ge_u32_e32 v2, v0
+; ALIGNED-NEXT:    s_xor_b32 s6, exec_lo, s6
+; ALIGNED-NEXT:    s_cbranch_execz .LBB9_2
+; ALIGNED-NEXT:  .LBB9_1: ; %memmove_fwd_loop
 ; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; ALIGNED-NEXT:    s_clause 0x3a
 ; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:20
 ; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:21
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:22
-; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:23
-; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:24
-; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:25
-; ALIGNED-NEXT:    buffer_load_ubyte v12, v2, s[0:3], 0 offen offset:26
-; ALIGNED-NEXT:    buffer_load_ubyte v127, v2, s[0:3], 0 offen offset:19
-; ALIGNED-NEXT:    buffer_load_ubyte v13, v2, s[0:3], 0 offen offset:27
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:28
-; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:29
-; ALIGNED-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:30
-; ALIGNED-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:31
-; ALIGNED-NEXT:    buffer_load_ubyte v15, v2, s[0:3], 0 offen offset:32
-; ALIGNED-NEXT:    buffer_load_ubyte v16, v2, s[0:3], 0 offen offset:33
-; ALIGNED-NEXT:    buffer_load_ubyte v18, v2, s[0:3], 0 offen offset:34
-; ALIGNED-NEXT:    buffer_load_ubyte v20, v2, s[0:3], 0 offen offset:35
-; ALIGNED-NEXT:    buffer_load_ubyte v14, v2, s[0:3], 0 offen offset:36
-; ALIGNED-NEXT:    buffer_load_ubyte v17, v2, s[0:3], 0 offen offset:37
-; ALIGNED-NEXT:    buffer_load_ubyte v19, v2, s[0:3], 0 offen offset:38
-; ALIGNED-NEXT:    buffer_load_ubyte v21, v2, s[0:3], 0 offen offset:39
-; ALIGNED-NEXT:    buffer_load_ubyte v23, v2, s[0:3], 0 offen offset:40
-; ALIGNED-NEXT:    buffer_load_ubyte v24, v2, s[0:3], 0 offen offset:41
-; ALIGNED-NEXT:    buffer_load_ubyte v26, v2, s[0:3], 0 offen offset:42
-; ALIGNED-NEXT:    buffer_load_ubyte v28, v2, s[0:3], 0 offen offset:43
-; ALIGNED-NEXT:    buffer_load_ubyte v22, v2, s[0:3], 0 offen offset:44
-; ALIGNED-NEXT:    buffer_load_ubyte v25, v2, s[0:3], 0 offen offset:45
-; ALIGNED-NEXT:    buffer_load_ubyte v27, v2, s[0:3], 0 offen offset:46
-; ALIGNED-NEXT:    buffer_load_ubyte v29, v2, s[0:3], 0 offen offset:47
-; ALIGNED-NEXT:    buffer_load_ubyte v30, v2, s[0:3], 0 offen offset:48
-; ALIGNED-NEXT:    buffer_load_ubyte v31, v2, s[0:3], 0 offen offset:49
-; ALIGNED-NEXT:    buffer_load_ubyte v33, v2, s[0:3], 0 offen offset:50
-; ALIGNED-NEXT:    buffer_load_ubyte v34, v2, s[0:3], 0 offen offset:51
-; ALIGNED-NEXT:    buffer_load_ubyte v32, v2, s[0:3], 0 offen offset:52
-; ALIGNED-NEXT:    buffer_load_ubyte v37, v2, s[0:3], 0 offen offset:53
+; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:22
+; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:23
+; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:24
+; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:25
+; ALIGNED-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:26
+; ALIGNED-NEXT:    buffer_load_ubyte v126, v2, s[0:3], 0 offen offset:19
+; ALIGNED-NEXT:    buffer_load_ubyte v12, v2, s[0:3], 0 offen offset:27
+; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:28
+; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:29
+; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:30
+; ALIGNED-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:31
+; ALIGNED-NEXT:    buffer_load_ubyte v14, v2, s[0:3], 0 offen offset:32
+; ALIGNED-NEXT:    buffer_load_ubyte v15, v2, s[0:3], 0 offen offset:33
+; ALIGNED-NEXT:    buffer_load_ubyte v17, v2, s[0:3], 0 offen offset:34
+; ALIGNED-NEXT:    buffer_load_ubyte v19, v2, s[0:3], 0 offen offset:35
+; ALIGNED-NEXT:    buffer_load_ubyte v13, v2, s[0:3], 0 offen offset:36
+; ALIGNED-NEXT:    buffer_load_ubyte v16, v2, s[0:3], 0 offen offset:37
+; ALIGNED-NEXT:    buffer_load_ubyte v18, v2, s[0:3], 0 offen offset:38
+; ALIGNED-NEXT:    buffer_load_ubyte v20, v2, s[0:3], 0 offen offset:39
+; ALIGNED-NEXT:    buffer_load_ubyte v22, v2, s[0:3], 0 offen offset:40
+; ALIGNED-NEXT:    buffer_load_ubyte v23, v2, s[0:3], 0 offen offset:41
+; ALIGNED-NEXT:    buffer_load_ubyte v25, v2, s[0:3], 0 offen offset:42
+; ALIGNED-NEXT:    buffer_load_ubyte v27, v2, s[0:3], 0 offen offset:43
+; ALIGNED-NEXT:    buffer_load_ubyte v21, v2, s[0:3], 0 offen offset:44
+; ALIGNED-NEXT:    buffer_load_ubyte v24, v2, s[0:3], 0 offen offset:45
+; ALIGNED-NEXT:    buffer_load_ubyte v26, v2, s[0:3], 0 offen offset:46
+; ALIGNED-NEXT:    buffer_load_ubyte v28, v2, s[0:3], 0 offen offset:47
+; ALIGNED-NEXT:    buffer_load_ubyte v29, v2, s[0:3], 0 offen offset:48
+; ALIGNED-NEXT:    buffer_load_ubyte v30, v2, s[0:3], 0 offen offset:49
+; ALIGNED-NEXT:    buffer_load_ubyte v32, v2, s[0:3], 0 offen offset:50
+; ALIGNED-NEXT:    buffer_load_ubyte v33, v2, s[0:3], 0 offen offset:51
+; ALIGNED-NEXT:    buffer_load_ubyte v31, v2, s[0:3], 0 offen offset:52
+; ALIGNED-NEXT:    buffer_load_ubyte v34, v2, s[0:3], 0 offen offset:53
 ; ALIGNED-NEXT:    buffer_load_ubyte v35, v2, s[0:3], 0 offen offset:54
 ; ALIGNED-NEXT:    buffer_load_ubyte v36, v2, s[0:3], 0 offen offset:55
-; ALIGNED-NEXT:    buffer_load_ubyte v48, v2, s[0:3], 0 offen offset:56
-; ALIGNED-NEXT:    buffer_load_ubyte v51, v2, s[0:3], 0 offen offset:57
-; ALIGNED-NEXT:    buffer_load_ubyte v53, v2, s[0:3], 0 offen offset:58
-; ALIGNED-NEXT:    buffer_load_ubyte v38, v2, s[0:3], 0 offen offset:60
-; ALIGNED-NEXT:    buffer_load_ubyte v50, v2, s[0:3], 0 offen offset:61
+; ALIGNED-NEXT:    buffer_load_ubyte v38, v2, s[0:3], 0 offen offset:56
+; ALIGNED-NEXT:    buffer_load_ubyte v50, v2, s[0:3], 0 offen offset:57
+; ALIGNED-NEXT:    buffer_load_ubyte v52, v2, s[0:3], 0 offen offset:58
+; ALIGNED-NEXT:    buffer_load_ubyte v37, v2, s[0:3], 0 offen offset:60
+; ALIGNED-NEXT:    buffer_load_ubyte v49, v2, s[0:3], 0 offen offset:61
 ; ALIGNED-NEXT:    buffer_load_ubyte v39, v2, s[0:3], 0 offen offset:62
-; ALIGNED-NEXT:    buffer_load_ubyte v49, v2, s[0:3], 0 offen offset:63
-; ALIGNED-NEXT:    buffer_load_ubyte v54, v2, s[0:3], 0 offen offset:64
-; ALIGNED-NEXT:    buffer_load_ubyte v55, v2, s[0:3], 0 offen offset:65
-; ALIGNED-NEXT:    buffer_load_ubyte v66, v2, s[0:3], 0 offen offset:66
-; ALIGNED-NEXT:    buffer_load_ubyte v52, v2, s[0:3], 0 offen offset:59
-; ALIGNED-NEXT:    buffer_load_ubyte v64, v2, s[0:3], 0 offen offset:67
-; ALIGNED-NEXT:    buffer_load_ubyte v65, v2, s[0:3], 0 offen offset:68
-; ALIGNED-NEXT:    buffer_load_ubyte v67, v2, s[0:3], 0 offen offset:69
-; ALIGNED-NEXT:    buffer_load_ubyte v68, v2, s[0:3], 0 offen offset:70
-; ALIGNED-NEXT:    buffer_load_ubyte v69, v2, s[0:3], 0 offen offset:71
-; ALIGNED-NEXT:    buffer_load_ubyte v81, v2, s[0:3], 0 offen offset:72
-; ALIGNED-NEXT:    buffer_load_ubyte v82, v2, s[0:3], 0 offen offset:73
-; ALIGNED-NEXT:    buffer_load_ubyte v84, v2, s[0:3], 0 offen offset:74
-; ALIGNED-NEXT:    buffer_load_ubyte v70, v2, s[0:3], 0 offen offset:76
-; ALIGNED-NEXT:    buffer_load_ubyte v71, v2, s[0:3], 0 offen offset:77
-; ALIGNED-NEXT:    buffer_load_ubyte v80, v2, s[0:3], 0 offen offset:78
-; ALIGNED-NEXT:    s_add_u32 s4, s4, 0xffffff00
-; ALIGNED-NEXT:    s_addc_u32 s5, s5, -1
-; ALIGNED-NEXT:    s_cmp_lg_u64 s[4:5], 0
+; ALIGNED-NEXT:    buffer_load_ubyte v48, v2, s[0:3], 0 offen offset:63
+; ALIGNED-NEXT:    buffer_load_ubyte v53, v2, s[0:3], 0 offen offset:64
+; ALIGNED-NEXT:    buffer_load_ubyte v54, v2, s[0:3], 0 offen offset:65
+; ALIGNED-NEXT:    buffer_load_ubyte v65, v2, s[0:3], 0 offen offset:66
+; ALIGNED-NEXT:    buffer_load_ubyte v51, v2, s[0:3], 0 offen offset:59
+; ALIGNED-NEXT:    buffer_load_ubyte v55, v2, s[0:3], 0 offen offset:67
+; ALIGNED-NEXT:    buffer_load_ubyte v64, v2, s[0:3], 0 offen offset:68
+; ALIGNED-NEXT:    buffer_load_ubyte v66, v2, s[0:3], 0 offen offset:69
+; ALIGNED-NEXT:    buffer_load_ubyte v67, v2, s[0:3], 0 offen offset:70
+; ALIGNED-NEXT:    buffer_load_ubyte v68, v2, s[0:3], 0 offen offset:71
+; ALIGNED-NEXT:    buffer_load_ubyte v80, v2, s[0:3], 0 offen offset:72
+; ALIGNED-NEXT:    buffer_load_ubyte v81, v2, s[0:3], 0 offen offset:73
+; ALIGNED-NEXT:    buffer_load_ubyte v83, v2, s[0:3], 0 offen offset:74
+; ALIGNED-NEXT:    buffer_load_ubyte v69, v2, s[0:3], 0 offen offset:76
+; ALIGNED-NEXT:    buffer_load_ubyte v70, v2, s[0:3], 0 offen offset:77
+; ALIGNED-NEXT:    buffer_load_ubyte v71, v2, s[0:3], 0 offen offset:78
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(58)
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(57)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v1, 8, v0
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(56)
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(55)
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v5, 8, v4
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(54)
-; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(53)
-; ALIGNED-NEXT:    v_lshl_or_b32 v7, v9, 8, v7
-; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
+; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(52)
+; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v0
+; ALIGNED-NEXT:    s_waitcnt vmcnt(50)
+; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(49)
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(48)
-; ALIGNED-NEXT:    v_lshl_or_b32 v5, v8, 8, v6
-; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(47)
+; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(46)
-; ALIGNED-NEXT:    v_lshl_or_b32 v6, v11, 8, v10
-; ALIGNED-NEXT:    v_lshl_or_b32 v8, v13, 8, v12
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v3
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v7, 8, v5
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v10, 8, v9
+; ALIGNED-NEXT:    v_lshl_or_b32 v5, v8, 8, v6
+; ALIGNED-NEXT:    v_lshl_or_b32 v6, v12, 8, v11
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(44)
-; ALIGNED-NEXT:    v_lshl_or_b32 v9, v16, 8, v15
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v6, 16, v5
+; ALIGNED-NEXT:    v_lshl_or_b32 v7, v15, 8, v14
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(42)
-; ALIGNED-NEXT:    v_lshl_or_b32 v10, v20, 8, v18
-; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v8, v19, 8, v17
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(40)
-; ALIGNED-NEXT:    v_lshl_or_b32 v11, v17, 8, v14
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v8, 16, v7
+; ALIGNED-NEXT:    v_lshl_or_b32 v9, v16, 8, v13
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(38)
-; ALIGNED-NEXT:    v_lshl_or_b32 v12, v21, 8, v19
-; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(36)
-; ALIGNED-NEXT:    v_lshl_or_b32 v13, v24, 8, v23
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v10, 16, v9
+; ALIGNED-NEXT:    v_lshl_or_b32 v10, v20, 8, v18
+; ALIGNED-NEXT:    s_waitcnt vmcnt(36)
+; ALIGNED-NEXT:    v_lshl_or_b32 v11, v23, 8, v22
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(34)
-; ALIGNED-NEXT:    v_lshl_or_b32 v14, v28, 8, v26
-; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v12, v27, 8, v25
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 16, v3
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v6, 16, v5
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v8, 16, v7
+; ALIGNED-NEXT:    v_lshl_or_b32 v5, v10, 16, v9
+; ALIGNED-NEXT:    v_lshl_or_b32 v6, v12, 16, v11
+; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(32)
-; ALIGNED-NEXT:    v_lshl_or_b32 v15, v25, 8, v22
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v12, 16, v11
+; ALIGNED-NEXT:    v_lshl_or_b32 v13, v24, 8, v21
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(30)
-; ALIGNED-NEXT:    v_lshl_or_b32 v16, v29, 8, v27
+; ALIGNED-NEXT:    v_lshl_or_b32 v14, v28, 8, v26
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(28)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v31, 8, v30
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v30, 8, v29
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(26)
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v34, 8, v33
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v33, 8, v32
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(24)
-; ALIGNED-NEXT:    v_lshl_or_b32 v5, v37, 8, v32
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v14, 16, v13
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v34, 8, v31
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(22)
-; ALIGNED-NEXT:    v_lshl_or_b32 v6, v36, 8, v35
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v36, 8, v35
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(17)
-; ALIGNED-NEXT:    v_lshl_or_b32 v7, v50, 8, v38
+; ALIGNED-NEXT:    v_lshl_or_b32 v5, v49, 8, v37
+; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(15)
-; ALIGNED-NEXT:    v_lshl_or_b32 v8, v49, 8, v39
-; ALIGNED-NEXT:    v_lshl_or_b32 v9, v51, 8, v48
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v16, 16, v15
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v6, v48, 8, v39
+; ALIGNED-NEXT:    v_lshl_or_b32 v7, v14, 16, v13
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 16, v3
+; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:81
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v6, 16, v5
+; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v7, v50, 8, v38
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(12)
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v51, 8, v52
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v54, 8, v53
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:928 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(11)
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v55, 8, v65
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v0, 16, v7
+; ALIGNED-NEXT:    s_clause 0x3
+; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:82
+; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:83
+; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:90
+; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:111
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 16, v1
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(13)
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v66, 8, v64
+; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:80
+; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:964 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(12)
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v68, 8, v67
+; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    s_clause 0x1
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:79
+; ALIGNED-NEXT:    buffer_load_ubyte v82, v2, s[0:3], 0 offen offset:75
 ; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    s_waitcnt vmcnt(11)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v52, 8, v53
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v55, 8, v54
-; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v6, 16, v5
-; ALIGNED-NEXT:    s_waitcnt vmcnt(10)
-; ALIGNED-NEXT:    v_lshl_or_b32 v5, v64, 8, v66
-; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:972 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(9)
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v70, 8, v69
+; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v8, 16, v7
-; ALIGNED-NEXT:    s_clause 0x4
-; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:83
-; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:90
-; ALIGNED-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:112
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:80
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:82
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v3, 16, v9
-; ALIGNED-NEXT:    s_waitcnt vmcnt(13)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v67, 8, v65
 ; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:928 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v5, 16, v4
-; ALIGNED-NEXT:    s_waitcnt vmcnt(11)
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v69, 8, v68
-; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:81
-; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:956 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:964 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_clause 0x1
-; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:79
-; ALIGNED-NEXT:    buffer_load_ubyte v83, v2, s[0:3], 0 offen offset:75
-; ALIGNED-NEXT:    s_waitcnt vmcnt(9)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v71, 8, v70
-; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:940 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v64, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v65, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:948 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:952 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:960 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:968 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v65, off, s[0:3], s32 offset:948 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:940 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v64, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:952 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:956 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:960 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:968 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:976 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
-; ALIGNED-NEXT:    buffer_store_dword v80, off, s[0:3], s32 offset:972 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:980 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v82, off, s[0:3], s32 offset:984 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v84, off, s[0:3], s32 offset:992 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v127, off, s[0:3], s32 offset:1452 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:980 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v80, off, s[0:3], s32 offset:988 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:992 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v83, off, s[0:3], s32 offset:1000 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v126, off, s[0:3], s32 offset:1456 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_clause 0x4
-; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:91
-; ALIGNED-NEXT:    buffer_load_ubyte v12, v2, s[0:3], 0 offen offset:113
-; ALIGNED-NEXT:    buffer_load_ubyte v13, v2, s[0:3], 0 offen offset:114
-; ALIGNED-NEXT:    buffer_load_ubyte v14, v2, s[0:3], 0 offen offset:115
+; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:91
+; ALIGNED-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:113
+; ALIGNED-NEXT:    buffer_load_ubyte v12, v2, s[0:3], 0 offen offset:114
 ; ALIGNED-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:116
+; ALIGNED-NEXT:    buffer_load_ubyte v13, v2, s[0:3], 0 offen offset:115
+; ALIGNED-NEXT:    s_waitcnt vmcnt(12)
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1016 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(11)
-; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1072 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(10)
-; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:1164 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1024 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(9)
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1004 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1076 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1016 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:1144 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1008 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1012 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v0, 8, v80
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:976 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:984 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v1, 8, v71
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(5)
-; ALIGNED-NEXT:    buffer_store_dword v83, off, s[0:3], s32 offset:988 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v82, 8, v81
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v83, 8, v84
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:996 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v5, 8, v1
+; ALIGNED-NEXT:    buffer_store_dword v82, off, s[0:3], s32 offset:996 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v82, 8, v83
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1004 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v81, 8, v80
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1008 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 8, v3
 ; ALIGNED-NEXT:    s_clause 0x1
+; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:86
+; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:88
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v6, 8, v5
+; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:87
+; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1020 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:89
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:85
-; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:88
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v7, 8, v6
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1000 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:87
-; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1012 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:89
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:86
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1036 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1044 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:84
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(10)
-; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:1068 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1072 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(9)
-; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:1168 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:1176 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
-; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:1176 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:1184 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
-; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:1172 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:1192 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
-; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:1184 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:1180 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(5)
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1024 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1036 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(4)
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1060 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1064 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1032 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1040 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1064 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1068 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1028 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v6, 8, v4
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:95
-; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v1, 8, v0
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1020 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1032 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1028 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v5, 8, v3
+; ALIGNED-NEXT:    s_clause 0x1
+; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:94
+; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:95
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:93
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:94
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1052 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1080 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:92
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1056 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1056 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1044 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1048 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v6, 8, v4
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:103
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1060 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v1, 8, v0
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1040 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:100
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v7, 8, v5
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v9, 8, v8
-; ALIGNED-NEXT:    s_clause 0x2
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1052 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1048 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v5, 8, v3
+; ALIGNED-NEXT:    s_clause 0x1
 ; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:98
-; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:105
-; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:107
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1076 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    s_clause 0x2
-; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:97
+; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:100
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v8, 8, v7
+; ALIGNED-NEXT:    s_clause 0x1
+; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:107
+; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:106
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1084 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v6, 8, v4
+; ALIGNED-NEXT:    s_clause 0x1
 ; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:99
-; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:106
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1080 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:105
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:97
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1088 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:96
-; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1112 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1100 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1104 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1096 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1108 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(5)
-; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1148 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1160 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(4)
-; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:1152 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1164 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1088 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1092 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v4, 8, v5
-; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:104
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1100 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1156 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1156 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v3, 8, v0
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1084 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:102
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1116 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1096 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1092 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v5
+; ALIGNED-NEXT:    s_clause 0x1
+; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:103
+; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:104
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:102
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1124 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:101
+; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1120 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1140 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1148 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1108 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v6, 8, v4
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:111
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1116 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v5, 8, v1
+; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:110
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v0, 8, v1
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1104 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1112 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v0, 8, v3
+; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:112
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:109
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:110
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1120 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1128 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:108
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1136 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1140 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1128 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1132 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v6, 8, v4
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:120
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1172 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v1, 8, v0
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1124 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:118
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v7, 8, v5
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v9, 8, v8
-; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:121
-; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:122
-; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:123
-; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:124
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1144 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v12, 8, v11
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v14, 8, v13
-; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:128
-; ALIGNED-NEXT:    buffer_load_ubyte v12, v2, s[0:3], 0 offen offset:129
-; ALIGNED-NEXT:    buffer_load_ubyte v13, v2, s[0:3], 0 offen offset:130
-; ALIGNED-NEXT:    buffer_load_ubyte v14, v2, s[0:3], 0 offen offset:131
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1160 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1136 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1132 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v9, 8, v5
+; ALIGNED-NEXT:    s_clause 0x1
+; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:120
+; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:132
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v8, 8, v7
+; ALIGNED-NEXT:    s_clause 0x1
+; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:123
+; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:124
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1152 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v6, 8, v4
+; ALIGNED-NEXT:    s_clause 0x1
 ; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:119
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1180 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:121
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v13, 8, v12
+; ALIGNED-NEXT:    s_clause 0x1
+; ALIGNED-NEXT:    buffer_load_ubyte v12, v2, s[0:3], 0 offen offset:130
+; ALIGNED-NEXT:    buffer_load_ubyte v13, v2, s[0:3], 0 offen offset:131
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1168 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v11, 8, v3
+; ALIGNED-NEXT:    s_clause 0x1
+; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:122
+; ALIGNED-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:129
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:118
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1188 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:117
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(11)
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1220 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1228 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(10)
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1192 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:1272 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(9)
-; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1228 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1240 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
-; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1236 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1212 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
-; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:1232 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1204 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1204 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1196 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v4, 8, v1
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:126
-; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v0, 8, v10
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1188 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:132
-; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:1244 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:1248 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:127
-; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:1256 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:1252 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1200 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:125
+; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1236 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(5)
+; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:1264 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(4)
+; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:1260 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1212 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1244 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:1264 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:1256 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1216 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v4, 8, v1
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:134
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1200 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v1
+; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:127
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v0, 8, v5
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1196 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v0, 8, v10
+; ALIGNED-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:128
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:126
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1208 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v7, 8, v6
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v9, 8, v8
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:136
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1224 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v12, 8, v11
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v14, 8, v13
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1240 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:135
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1260 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:133
+; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:125
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1272 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1224 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1300 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:1252 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1276 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v4, 8, v1
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1220 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v1
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v0, 8, v10
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1216 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v0, 8, v8
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v7, 8, v3
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1232 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v6, 8, v5
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v13, 8, v12
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1248 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v11, 8, v10
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1268 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_clause 0x6
-; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:137
-; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:138
-; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:139
-; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:140
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:142
-; ALIGNED-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:144
-; ALIGNED-NEXT:    buffer_load_ubyte v12, v2, s[0:3], 0 offen offset:145
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:143
-; ALIGNED-NEXT:    buffer_load_ubyte v13, v2, s[0:3], 0 offen offset:146
-; ALIGNED-NEXT:    buffer_load_ubyte v14, v2, s[0:3], 0 offen offset:147
-; ALIGNED-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:148
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1280 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:133
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1276 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_clause 0x7
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:134
+; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:135
+; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:136
+; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:137
+; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:138
+; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:139
+; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:140
+; ALIGNED-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:144
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v0, 8, v9
 ; ALIGNED-NEXT:    s_clause 0x3e
-; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:141
-; ALIGNED-NEXT:    buffer_load_ubyte v123, v2, s[0:3], 0 offen offset:160
-; ALIGNED-NEXT:    buffer_load_ubyte v105, v2, s[0:3], 0 offen offset:161
-; ALIGNED-NEXT:    buffer_load_ubyte v95, v2, s[0:3], 0 offen offset:162
-; ALIGNED-NEXT:    buffer_load_ubyte v97, v2, s[0:3], 0 offen offset:163
-; ALIGNED-NEXT:    buffer_load_ubyte v107, v2, s[0:3], 0 offen offset:164
-; ALIGNED-NEXT:    buffer_load_ubyte v57, v2, s[0:3], 0 offen offset:165
-; ALIGNED-NEXT:    buffer_load_ubyte v59, v2, s[0:3], 0 offen offset:166
-; ALIGNED-NEXT:    buffer_load_ubyte v47, v2, s[0:3], 0 offen offset:167
-; ALIGNED-NEXT:    buffer_load_ubyte v45, v2, s[0:3], 0 offen offset:168
-; ALIGNED-NEXT:    buffer_load_ubyte v43, v2, s[0:3], 0 offen offset:169
-; ALIGNED-NEXT:    buffer_load_ubyte v115, v2, s[0:3], 0 offen offset:170
-; ALIGNED-NEXT:    buffer_load_ubyte v117, v2, s[0:3], 0 offen offset:171
-; ALIGNED-NEXT:    buffer_load_ubyte v113, v2, s[0:3], 0 offen offset:172
-; ALIGNED-NEXT:    buffer_load_ubyte v100, v2, s[0:3], 0 offen offset:173
-; ALIGNED-NEXT:    buffer_load_ubyte v102, v2, s[0:3], 0 offen offset:174
-; ALIGNED-NEXT:    buffer_load_ubyte v86, v2, s[0:3], 0 offen offset:175
-; ALIGNED-NEXT:    buffer_load_ubyte v85, v2, s[0:3], 0 offen offset:176
-; ALIGNED-NEXT:    buffer_load_ubyte v83, v2, s[0:3], 0 offen offset:177
-; ALIGNED-NEXT:    buffer_load_ubyte v81, v2, s[0:3], 0 offen offset:178
-; ALIGNED-NEXT:    buffer_load_ubyte v82, v2, s[0:3], 0 offen offset:179
-; ALIGNED-NEXT:    buffer_load_ubyte v84, v2, s[0:3], 0 offen offset:180
-; ALIGNED-NEXT:    buffer_load_ubyte v69, v2, s[0:3], 0 offen offset:181
-; ALIGNED-NEXT:    buffer_load_ubyte v70, v2, s[0:3], 0 offen offset:182
-; ALIGNED-NEXT:    buffer_load_ubyte v68, v2, s[0:3], 0 offen offset:183
-; ALIGNED-NEXT:    buffer_load_ubyte v67, v2, s[0:3], 0 offen offset:184
-; ALIGNED-NEXT:    buffer_load_ubyte v66, v2, s[0:3], 0 offen offset:185
-; ALIGNED-NEXT:    buffer_load_ubyte v65, v2, s[0:3], 0 offen offset:186
-; ALIGNED-NEXT:    buffer_load_ubyte v64, v2, s[0:3], 0 offen offset:187
-; ALIGNED-NEXT:    buffer_load_ubyte v55, v2, s[0:3], 0 offen offset:188
-; ALIGNED-NEXT:    buffer_load_ubyte v53, v2, s[0:3], 0 offen offset:189
-; ALIGNED-NEXT:    buffer_load_ubyte v54, v2, s[0:3], 0 offen offset:190
-; ALIGNED-NEXT:    buffer_load_ubyte v52, v2, s[0:3], 0 offen offset:191
-; ALIGNED-NEXT:    buffer_load_ubyte v51, v2, s[0:3], 0 offen offset:192
-; ALIGNED-NEXT:    buffer_load_ubyte v48, v2, s[0:3], 0 offen offset:193
-; ALIGNED-NEXT:    buffer_load_ubyte v39, v2, s[0:3], 0 offen offset:194
-; ALIGNED-NEXT:    buffer_load_ubyte v38, v2, s[0:3], 0 offen offset:195
-; ALIGNED-NEXT:    buffer_load_ubyte v49, v2, s[0:3], 0 offen offset:196
-; ALIGNED-NEXT:    buffer_load_ubyte v35, v2, s[0:3], 0 offen offset:197
-; ALIGNED-NEXT:    buffer_load_ubyte v36, v2, s[0:3], 0 offen offset:198
-; ALIGNED-NEXT:    buffer_load_ubyte v34, v2, s[0:3], 0 offen offset:199
-; ALIGNED-NEXT:    buffer_load_ubyte v33, v2, s[0:3], 0 offen offset:200
-; ALIGNED-NEXT:    buffer_load_ubyte v31, v2, s[0:3], 0 offen offset:201
-; ALIGNED-NEXT:    buffer_load_ubyte v29, v2, s[0:3], 0 offen offset:202
-; ALIGNED-NEXT:    buffer_load_ubyte v30, v2, s[0:3], 0 offen offset:203
-; ALIGNED-NEXT:    buffer_load_ubyte v28, v2, s[0:3], 0 offen offset:204
-; ALIGNED-NEXT:    buffer_load_ubyte v25, v2, s[0:3], 0 offen offset:205
-; ALIGNED-NEXT:    buffer_load_ubyte v26, v2, s[0:3], 0 offen offset:206
-; ALIGNED-NEXT:    buffer_load_ubyte v24, v2, s[0:3], 0 offen offset:207
-; ALIGNED-NEXT:    buffer_load_ubyte v23, v2, s[0:3], 0 offen offset:208
-; ALIGNED-NEXT:    buffer_load_ubyte v21, v2, s[0:3], 0 offen offset:209
-; ALIGNED-NEXT:    buffer_load_ubyte v22, v2, s[0:3], 0 offen offset:210
-; ALIGNED-NEXT:    buffer_load_ubyte v20, v2, s[0:3], 0 offen offset:211
-; ALIGNED-NEXT:    buffer_load_ubyte v18, v2, s[0:3], 0 offen offset:212
-; ALIGNED-NEXT:    buffer_load_ubyte v122, v2, s[0:3], 0 offen offset:229
-; ALIGNED-NEXT:    buffer_load_ubyte v120, v2, s[0:3], 0 offen offset:230
-; ALIGNED-NEXT:    buffer_load_ubyte v110, v2, s[0:3], 0 offen offset:231
-; ALIGNED-NEXT:    buffer_load_ubyte v108, v2, s[0:3], 0 offen offset:232
-; ALIGNED-NEXT:    buffer_load_ubyte v106, v2, s[0:3], 0 offen offset:233
-; ALIGNED-NEXT:    buffer_load_ubyte v104, v2, s[0:3], 0 offen offset:234
-; ALIGNED-NEXT:    buffer_load_ubyte v93, v2, s[0:3], 0 offen offset:235
-; ALIGNED-NEXT:    buffer_load_ubyte v94, v2, s[0:3], 0 offen offset:236
-; ALIGNED-NEXT:    buffer_load_ubyte v78, v2, s[0:3], 0 offen offset:237
-; ALIGNED-NEXT:    s_clause 0x17
-; ALIGNED-NEXT:    buffer_load_ubyte v89, v2, s[0:3], 0 offen offset:238
-; ALIGNED-NEXT:    buffer_load_ubyte v90, v2, s[0:3], 0 offen offset:239
-; ALIGNED-NEXT:    buffer_load_ubyte v91, v2, s[0:3], 0 offen offset:240
-; ALIGNED-NEXT:    buffer_load_ubyte v75, v2, s[0:3], 0 offen offset:241
-; ALIGNED-NEXT:    buffer_load_ubyte v76, v2, s[0:3], 0 offen offset:242
-; ALIGNED-NEXT:    buffer_load_ubyte v72, v2, s[0:3], 0 offen offset:243
-; ALIGNED-NEXT:    buffer_load_ubyte v73, v2, s[0:3], 0 offen offset:244
-; ALIGNED-NEXT:    buffer_load_ubyte v58, v2, s[0:3], 0 offen offset:245
-; ALIGNED-NEXT:    buffer_load_ubyte v56, v2, s[0:3], 0 offen offset:246
-; ALIGNED-NEXT:    buffer_load_ubyte v46, v2, s[0:3], 0 offen offset:247
-; ALIGNED-NEXT:    buffer_load_ubyte v44, v2, s[0:3], 0 offen offset:248
-; ALIGNED-NEXT:    buffer_load_ubyte v42, v2, s[0:3], 0 offen offset:249
-; ALIGNED-NEXT:    buffer_load_ubyte v40, v2, s[0:3], 0 offen offset:250
-; ALIGNED-NEXT:    buffer_load_ubyte v118, v2, s[0:3], 0 offen offset:251
-; ALIGNED-NEXT:    buffer_load_ubyte v116, v2, s[0:3], 0 offen offset:252
-; ALIGNED-NEXT:    buffer_load_ubyte v114, v2, s[0:3], 0 offen offset:253
-; ALIGNED-NEXT:    buffer_load_ubyte v112, v2, s[0:3], 0 offen offset:254
-; ALIGNED-NEXT:    buffer_load_ubyte v101, v2, s[0:3], 0 offen offset:255
-; ALIGNED-NEXT:    buffer_load_ubyte v96, v2, s[0:3], 0 offen
-; ALIGNED-NEXT:    buffer_load_ubyte v61, v2, s[0:3], 0 offen offset:3
-; ALIGNED-NEXT:    buffer_load_ubyte v92, v2, s[0:3], 0 offen offset:4
-; ALIGNED-NEXT:    buffer_load_ubyte v63, v2, s[0:3], 0 offen offset:5
-; ALIGNED-NEXT:    buffer_load_ubyte v77, v2, s[0:3], 0 offen offset:6
-; ALIGNED-NEXT:    buffer_load_ubyte v79, v2, s[0:3], 0 offen offset:7
+; ALIGNED-NEXT:    buffer_load_ubyte v125, v2, s[0:3], 0 offen offset:159
+; ALIGNED-NEXT:    buffer_load_ubyte v124, v2, s[0:3], 0 offen offset:160
+; ALIGNED-NEXT:    buffer_load_ubyte v120, v2, s[0:3], 0 offen offset:161
+; ALIGNED-NEXT:    buffer_load_ubyte v109, v2, s[0:3], 0 offen offset:162
+; ALIGNED-NEXT:    buffer_load_ubyte v108, v2, s[0:3], 0 offen offset:163
+; ALIGNED-NEXT:    buffer_load_ubyte v121, v2, s[0:3], 0 offen offset:164
+; ALIGNED-NEXT:    buffer_load_ubyte v93, v2, s[0:3], 0 offen offset:165
+; ALIGNED-NEXT:    buffer_load_ubyte v94, v2, s[0:3], 0 offen offset:166
+; ALIGNED-NEXT:    buffer_load_ubyte v92, v2, s[0:3], 0 offen offset:167
+; ALIGNED-NEXT:    buffer_load_ubyte v90, v2, s[0:3], 0 offen offset:168
+; ALIGNED-NEXT:    buffer_load_ubyte v78, v2, s[0:3], 0 offen offset:169
+; ALIGNED-NEXT:    buffer_load_ubyte v77, v2, s[0:3], 0 offen offset:170
+; ALIGNED-NEXT:    buffer_load_ubyte v76, v2, s[0:3], 0 offen offset:171
+; ALIGNED-NEXT:    buffer_load_ubyte v75, v2, s[0:3], 0 offen offset:172
+; ALIGNED-NEXT:    buffer_load_ubyte v73, v2, s[0:3], 0 offen offset:173
+; ALIGNED-NEXT:    buffer_load_ubyte v74, v2, s[0:3], 0 offen offset:174
+; ALIGNED-NEXT:    buffer_load_ubyte v72, v2, s[0:3], 0 offen offset:175
+; ALIGNED-NEXT:    buffer_load_ubyte v63, v2, s[0:3], 0 offen offset:176
+; ALIGNED-NEXT:    buffer_load_ubyte v61, v2, s[0:3], 0 offen offset:177
+; ALIGNED-NEXT:    buffer_load_ubyte v59, v2, s[0:3], 0 offen offset:178
+; ALIGNED-NEXT:    buffer_load_ubyte v60, v2, s[0:3], 0 offen offset:179
+; ALIGNED-NEXT:    buffer_load_ubyte v62, v2, s[0:3], 0 offen offset:180
+; ALIGNED-NEXT:    buffer_load_ubyte v47, v2, s[0:3], 0 offen offset:181
+; ALIGNED-NEXT:    buffer_load_ubyte v56, v2, s[0:3], 0 offen offset:182
+; ALIGNED-NEXT:    buffer_load_ubyte v46, v2, s[0:3], 0 offen offset:183
+; ALIGNED-NEXT:    buffer_load_ubyte v45, v2, s[0:3], 0 offen offset:184
+; ALIGNED-NEXT:    buffer_load_ubyte v44, v2, s[0:3], 0 offen offset:185
+; ALIGNED-NEXT:    buffer_load_ubyte v43, v2, s[0:3], 0 offen offset:186
+; ALIGNED-NEXT:    buffer_load_ubyte v42, v2, s[0:3], 0 offen offset:187
+; ALIGNED-NEXT:    buffer_load_ubyte v41, v2, s[0:3], 0 offen offset:188
+; ALIGNED-NEXT:    buffer_load_ubyte v119, v2, s[0:3], 0 offen offset:189
+; ALIGNED-NEXT:    buffer_load_ubyte v40, v2, s[0:3], 0 offen offset:190
+; ALIGNED-NEXT:    buffer_load_ubyte v118, v2, s[0:3], 0 offen offset:191
+; ALIGNED-NEXT:    buffer_load_ubyte v117, v2, s[0:3], 0 offen offset:192
+; ALIGNED-NEXT:    buffer_load_ubyte v114, v2, s[0:3], 0 offen offset:193
+; ALIGNED-NEXT:    buffer_load_ubyte v113, v2, s[0:3], 0 offen offset:194
+; ALIGNED-NEXT:    buffer_load_ubyte v112, v2, s[0:3], 0 offen offset:195
+; ALIGNED-NEXT:    buffer_load_ubyte v115, v2, s[0:3], 0 offen offset:196
+; ALIGNED-NEXT:    buffer_load_ubyte v101, v2, s[0:3], 0 offen offset:197
+; ALIGNED-NEXT:    buffer_load_ubyte v102, v2, s[0:3], 0 offen offset:198
+; ALIGNED-NEXT:    buffer_load_ubyte v100, v2, s[0:3], 0 offen offset:199
+; ALIGNED-NEXT:    buffer_load_ubyte v99, v2, s[0:3], 0 offen offset:200
+; ALIGNED-NEXT:    buffer_load_ubyte v97, v2, s[0:3], 0 offen offset:201
+; ALIGNED-NEXT:    buffer_load_ubyte v87, v2, s[0:3], 0 offen offset:202
+; ALIGNED-NEXT:    buffer_load_ubyte v96, v2, s[0:3], 0 offen offset:203
+; ALIGNED-NEXT:    buffer_load_ubyte v86, v2, s[0:3], 0 offen offset:204
+; ALIGNED-NEXT:    buffer_load_ubyte v83, v2, s[0:3], 0 offen offset:205
+; ALIGNED-NEXT:    buffer_load_ubyte v84, v2, s[0:3], 0 offen offset:206
+; ALIGNED-NEXT:    buffer_load_ubyte v82, v2, s[0:3], 0 offen offset:207
+; ALIGNED-NEXT:    buffer_load_ubyte v81, v2, s[0:3], 0 offen offset:208
+; ALIGNED-NEXT:    buffer_load_ubyte v71, v2, s[0:3], 0 offen offset:209
+; ALIGNED-NEXT:    buffer_load_ubyte v80, v2, s[0:3], 0 offen offset:210
+; ALIGNED-NEXT:    buffer_load_ubyte v70, v2, s[0:3], 0 offen offset:211
+; ALIGNED-NEXT:    buffer_load_ubyte v68, v2, s[0:3], 0 offen offset:212
+; ALIGNED-NEXT:    buffer_load_ubyte v53, v2, s[0:3], 0 offen offset:213
+; ALIGNED-NEXT:    buffer_load_ubyte v64, v2, s[0:3], 0 offen offset:214
+; ALIGNED-NEXT:    buffer_load_ubyte v52, v2, s[0:3], 0 offen offset:215
+; ALIGNED-NEXT:    buffer_load_ubyte v67, v2, s[0:3], 0 offen offset:216
+; ALIGNED-NEXT:    buffer_load_ubyte v65, v2, s[0:3], 0 offen offset:217
+; ALIGNED-NEXT:    buffer_load_ubyte v55, v2, s[0:3], 0 offen offset:218
+; ALIGNED-NEXT:    buffer_load_ubyte v54, v2, s[0:3], 0 offen offset:219
+; ALIGNED-NEXT:    buffer_load_ubyte v66, v2, s[0:3], 0 offen offset:220
+; ALIGNED-NEXT:    buffer_load_ubyte v49, v2, s[0:3], 0 offen offset:221
+; ALIGNED-NEXT:    s_clause 0x20
+; ALIGNED-NEXT:    buffer_load_ubyte v50, v2, s[0:3], 0 offen offset:222
+; ALIGNED-NEXT:    buffer_load_ubyte v48, v2, s[0:3], 0 offen offset:223
+; ALIGNED-NEXT:    buffer_load_ubyte v39, v2, s[0:3], 0 offen offset:224
+; ALIGNED-NEXT:    buffer_load_ubyte v36, v2, s[0:3], 0 offen offset:225
+; ALIGNED-NEXT:    buffer_load_ubyte v35, v2, s[0:3], 0 offen offset:226
+; ALIGNED-NEXT:    buffer_load_ubyte v34, v2, s[0:3], 0 offen offset:227
+; ALIGNED-NEXT:    buffer_load_ubyte v37, v2, s[0:3], 0 offen offset:228
+; ALIGNED-NEXT:    buffer_load_ubyte v31, v2, s[0:3], 0 offen offset:229
+; ALIGNED-NEXT:    buffer_load_ubyte v32, v2, s[0:3], 0 offen offset:230
+; ALIGNED-NEXT:    buffer_load_ubyte v30, v2, s[0:3], 0 offen offset:231
+; ALIGNED-NEXT:    buffer_load_ubyte v29, v2, s[0:3], 0 offen offset:232
+; ALIGNED-NEXT:    buffer_load_ubyte v28, v2, s[0:3], 0 offen offset:233
+; ALIGNED-NEXT:    buffer_load_ubyte v27, v2, s[0:3], 0 offen offset:234
+; ALIGNED-NEXT:    buffer_load_ubyte v26, v2, s[0:3], 0 offen offset:235
+; ALIGNED-NEXT:    buffer_load_ubyte v25, v2, s[0:3], 0 offen offset:236
+; ALIGNED-NEXT:    buffer_load_ubyte v23, v2, s[0:3], 0 offen offset:237
+; ALIGNED-NEXT:    buffer_load_ubyte v24, v2, s[0:3], 0 offen offset:238
+; ALIGNED-NEXT:    buffer_load_ubyte v22, v2, s[0:3], 0 offen offset:239
+; ALIGNED-NEXT:    buffer_load_ubyte v21, v2, s[0:3], 0 offen offset:240
+; ALIGNED-NEXT:    buffer_load_ubyte v19, v2, s[0:3], 0 offen offset:241
+; ALIGNED-NEXT:    buffer_load_ubyte v17, v2, s[0:3], 0 offen offset:242
+; ALIGNED-NEXT:    buffer_load_ubyte v18, v2, s[0:3], 0 offen offset:243
+; ALIGNED-NEXT:    buffer_load_ubyte v20, v2, s[0:3], 0 offen offset:244
+; ALIGNED-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:145
+; ALIGNED-NEXT:    buffer_load_ubyte v12, v2, s[0:3], 0 offen offset:146
+; ALIGNED-NEXT:    buffer_load_ubyte v13, v2, s[0:3], 0 offen offset:147
+; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:148
+; ALIGNED-NEXT:    buffer_load_ubyte v105, v2, s[0:3], 0 offen offset:2
+; ALIGNED-NEXT:    buffer_load_ubyte v88, v2, s[0:3], 0 offen offset:3
+; ALIGNED-NEXT:    buffer_load_ubyte v91, v2, s[0:3], 0 offen offset:4
+; ALIGNED-NEXT:    buffer_load_ubyte v95, v2, s[0:3], 0 offen offset:5
+; ALIGNED-NEXT:    buffer_load_ubyte v104, v2, s[0:3], 0 offen offset:6
+; ALIGNED-NEXT:    buffer_load_ubyte v106, v2, s[0:3], 0 offen offset:7
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(62)
-; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1308 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1316 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:1312 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1284 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1292 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:1324 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:1328 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1296 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v4, 8, v1
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:150
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v0, 8, v5
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1280 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v1
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1284 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:143
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1308 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1316 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:142
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1324 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1320 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1292 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1288 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:156
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v7, 8, v6
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v9, 8, v8
+; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:141
+; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:1332 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(12)
+; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:1336 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(11)
+; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:1344 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(10)
+; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:1340 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(9)
+; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:1352 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
+; ALIGNED-NEXT:    buffer_store_dword v105, off, s[0:3], s32 offset:1384 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
+; ALIGNED-NEXT:    buffer_store_dword v88, off, s[0:3], s32 offset:1376 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v88, v88, 8, v105
+; ALIGNED-NEXT:    s_clause 0x4
+; ALIGNED-NEXT:    buffer_load_ubyte v123, v2, s[0:3], 0 offen offset:12
+; ALIGNED-NEXT:    buffer_load_ubyte v111, v2, s[0:3], 0 offen offset:13
+; ALIGNED-NEXT:    buffer_load_ubyte v122, v2, s[0:3], 0 offen offset:14
+; ALIGNED-NEXT:    buffer_load_ubyte v110, v2, s[0:3], 0 offen offset:15
+; ALIGNED-NEXT:    buffer_load_ubyte v105, v2, s[0:3], 0 offen offset:11
+; ALIGNED-NEXT:    s_waitcnt vmcnt(10)
+; ALIGNED-NEXT:    buffer_store_dword v95, off, s[0:3], s32 offset:1392 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(9)
+; ALIGNED-NEXT:    buffer_store_dword v104, off, s[0:3], s32 offset:1396 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
+; ALIGNED-NEXT:    buffer_store_dword v106, off, s[0:3], s32 offset:1400 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v91, off, s[0:3], s32 offset:1380 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1304 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1300 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v1
+; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:151
+; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1296 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v0, 8, v8
+; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:156
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v7, 8, v3
 ; ALIGNED-NEXT:    s_clause 0x1
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:152
-; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:154
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1304 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_load_ubyte v3, v2, s[0:3], 0 offen offset:154
+; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:155
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1312 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v6, 8, v5
 ; ALIGNED-NEXT:    s_clause 0x1
-; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:153
-; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:155
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v12, 8, v11
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v14, 8, v13
-; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:1336 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:1332 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1320 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:1344 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:151
-; ALIGNED-NEXT:    s_waitcnt vmcnt(21)
-; ALIGNED-NEXT:    v_lshl_or_b32 v103, v46, 8, v56
-; ALIGNED-NEXT:    s_waitcnt vmcnt(13)
-; ALIGNED-NEXT:    v_lshl_or_b32 v119, v101, 8, v112
-; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
-; ALIGNED-NEXT:    buffer_store_dword v79, off, s[0:3], s32 offset:1380 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1340 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:152
+; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:153
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v13, 8, v12
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1328 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v11, 8, v10
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:150
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1348 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:149
-; ALIGNED-NEXT:    buffer_store_dword v92, off, s[0:3], s32 offset:1388 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:1372 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v77, off, s[0:3], s32 offset:1376 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:1360 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1352 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1364 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1400 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1404 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(5)
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1420 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1436 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(4)
-; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1436 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1432 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1428 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1416 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:1432 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1356 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v4, 8, v1
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:158
+; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1428 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v0, 8, v10
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1348 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1360 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v1
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v82, 8, v84
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1356 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v0, 8, v9
 ; ALIGNED-NEXT:    s_clause 0x7
-; ALIGNED-NEXT:    buffer_load_ubyte v13, v2, s[0:3], 0 offen offset:213
-; ALIGNED-NEXT:    buffer_load_ubyte v14, v2, s[0:3], 0 offen offset:214
-; ALIGNED-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:215
-; ALIGNED-NEXT:    buffer_load_ubyte v17, v2, s[0:3], 0 offen offset:216
-; ALIGNED-NEXT:    buffer_load_ubyte v15, v2, s[0:3], 0 offen offset:217
-; ALIGNED-NEXT:    buffer_load_ubyte v12, v2, s[0:3], 0 offen offset:218
-; ALIGNED-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:219
-; ALIGNED-NEXT:    buffer_load_ubyte v16, v2, s[0:3], 0 offen offset:220
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:159
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1392 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:157
-; ALIGNED-NEXT:    s_waitcnt vmcnt(10)
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1408 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_load_ubyte v15, v2, s[0:3], 0 offen offset:245
+; ALIGNED-NEXT:    buffer_load_ubyte v16, v2, s[0:3], 0 offen offset:246
+; ALIGNED-NEXT:    buffer_load_ubyte v14, v2, s[0:3], 0 offen offset:247
+; ALIGNED-NEXT:    buffer_load_ubyte v13, v2, s[0:3], 0 offen offset:248
+; ALIGNED-NEXT:    buffer_load_ubyte v12, v2, s[0:3], 0 offen offset:249
+; ALIGNED-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:250
+; ALIGNED-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:251
+; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:252
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    s_clause 0x1
+; ALIGNED-NEXT:    buffer_load_ubyte v127, v2, s[0:3], 0 offen offset:157
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:158
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1388 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
+; ALIGNED-NEXT:    v_lshl_or_b32 v79, v10, 8, v11
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1412 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v4, 8, v1
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v0, 8, v5
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1404 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v7, 8, v6
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v9, 8, v8
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v127, 8, v8
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1412 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v125, 8, v1
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v7, 8, v3
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v83, 8, v86
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1424 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v105, 8, v123
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v97, 8, v95
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v6, 8, v5
+; ALIGNED-NEXT:    v_lshl_or_b32 v58, v4, 16, v3
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v97, 8, v99
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v96, 8, v87
+; ALIGNED-NEXT:    v_lshl_or_b32 v5, v14, 8, v16
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v108, 8, v109
+; ALIGNED-NEXT:    v_lshl_or_b32 v57, v4, 16, v3
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v53, 8, v68
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1440 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v57, 8, v107
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v47, 8, v59
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1456 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v100, 8, v113
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v86, 8, v102
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v120, 8, v124
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v52, 8, v64
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v92, 8, v94
+; ALIGNED-NEXT:    v_lshl_or_b32 v116, v4, 16, v3
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v65, 8, v67
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v54, 8, v55
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1452 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v93, 8, v121
+; ALIGNED-NEXT:    v_lshl_or_b32 v103, v4, 16, v3
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v49, 8, v66
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v72, 8, v74
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v48, 8, v50
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1460 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v43, 8, v45
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v117, 8, v115
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v73, 8, v75
+; ALIGNED-NEXT:    v_lshl_or_b32 v98, v4, 16, v3
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v71, 8, v81
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v70, 8, v80
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v76, 8, v77
+; ALIGNED-NEXT:    v_lshl_or_b32 v85, v4, 16, v3
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v36, 8, v39
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1464 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v83, 8, v85
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v82, 8, v81
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v78, 8, v90
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v34, 8, v35
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v60, 8, v59
+; ALIGNED-NEXT:    v_lshl_or_b32 v69, v4, 16, v3
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v31, 8, v37
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v30, 8, v32
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1468 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v69, 8, v84
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v68, 8, v70
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v61, 8, v63
+; ALIGNED-NEXT:    v_lshl_or_b32 v51, v4, 16, v3
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v23, 8, v25
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v46, 8, v56
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v22, 8, v24
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1472 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v53, 8, v55
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v52, 8, v54
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v47, 8, v62
+; ALIGNED-NEXT:    v_lshl_or_b32 v38, v4, 16, v3
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v28, 8, v29
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v26, 8, v27
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v118, 8, v40
+; ALIGNED-NEXT:    v_lshl_or_b32 v33, v4, 16, v3
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v19, 8, v21
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1476 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v66, 8, v67
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v64, 8, v65
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v119, 8, v41
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v18, 8, v17
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v42, 8, v43
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v4, 16, v3
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v15, 8, v20
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1480 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v48, 8, v51
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v38, 8, v39
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v44, 8, v45
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v5, 16, v4
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v112, 8, v113
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1484 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_clause 0x7
-; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:221
-; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:222
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:223
-; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:224
-; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:225
-; ALIGNED-NEXT:    buffer_load_ubyte v126, v2, s[0:3], 0 offen offset:226
-; ALIGNED-NEXT:    buffer_load_ubyte v124, v2, s[0:3], 0 offen offset:227
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:228
-; ALIGNED-NEXT:    v_lshl_or_b32 v74, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v35, 8, v49
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v34, 8, v36
-; ALIGNED-NEXT:    v_lshl_or_b32 v41, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v25, 8, v28
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v24, 8, v26
-; ALIGNED-NEXT:    v_lshl_or_b32 v80, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v31, 8, v33
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v30, 8, v29
-; ALIGNED-NEXT:    v_lshl_or_b32 v71, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v13, 8, v18
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v11, 8, v14
-; ALIGNED-NEXT:    v_lshl_or_b32 v50, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v15, 8, v17
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v10, 8, v12
-; ALIGNED-NEXT:    v_lshl_or_b32 v37, v4, 16, v3
-; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v7, 8, v16
-; ALIGNED-NEXT:    s_waitcnt vmcnt(5)
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v6, 8, v8
-; ALIGNED-NEXT:    v_lshl_or_b32 v32, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v21, 8, v23
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v20, 8, v22
-; ALIGNED-NEXT:    v_lshl_or_b32 v27, v4, 16, v3
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v114, 8, v117
+; ALIGNED-NEXT:    v_lshl_or_b32 v107, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v101, 8, v115
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v100, 8, v102
+; ALIGNED-NEXT:    v_lshl_or_b32 v89, v1, 16, v0
+; ALIGNED-NEXT:    s_clause 0x3
+; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:253
+; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:254
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:255
+; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v0, 8, v5
+; ALIGNED-NEXT:    v_lshl_or_b32 v5, v7, 8, v9
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v124, 8, v126
-; ALIGNED-NEXT:    v_lshl_or_b32 v19, v4, 16, v3
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v122, 8, v1
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v110, 8, v120
-; ALIGNED-NEXT:    v_lshl_or_b32 v9, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v78, 8, v94
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v90, 8, v89
-; ALIGNED-NEXT:    v_lshl_or_b32 v62, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v106, 8, v108
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v93, 8, v104
-; ALIGNED-NEXT:    v_lshl_or_b32 v60, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v75, 8, v91
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v72, 8, v76
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v58, 8, v73
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v103, 16, v4
-; ALIGNED-NEXT:    v_lshl_or_b32 v103, v114, 8, v116
-; ALIGNED-NEXT:    v_lshl_or_b32 v98, v119, 16, v103
-; ALIGNED-NEXT:    v_lshl_or_b32 v103, v42, 8, v44
-; ALIGNED-NEXT:    v_lshl_or_b32 v119, v118, 8, v40
-; ALIGNED-NEXT:    v_lshl_or_b32 v99, v119, 16, v103
-; ALIGNED-NEXT:    s_clause 0x1
-; ALIGNED-NEXT:    buffer_load_ubyte v103, v2, s[0:3], 0 offen offset:1
-; ALIGNED-NEXT:    buffer_load_ubyte v119, v2, s[0:3], 0 offen offset:2
-; ALIGNED-NEXT:    buffer_store_dword v96, off, s[0:3], s32 offset:1364 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v6, v1, 8, v8
+; ALIGNED-NEXT:    v_lshl_or_b32 v5, v6, 16, v5
+; ALIGNED-NEXT:    v_lshl_or_b32 v6, v12, 8, v13
+; ALIGNED-NEXT:    v_lshl_or_b32 v6, v79, 16, v6
+; ALIGNED-NEXT:    buffer_load_ubyte v79, v2, s[0:3], 0 offen offset:1
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    buffer_store_dword v103, off, s[0:3], s32 offset:1368 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1368 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_dword v119, off, s[0:3], s32 offset:1384 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v103, v103, 8, v96
-; ALIGNED-NEXT:    v_lshl_or_b32 v119, v61, 8, v119
-; ALIGNED-NEXT:    v_lshl_or_b32 v96, v119, 16, v103
-; ALIGNED-NEXT:    v_lshl_or_b32 v119, v79, 8, v77
-; ALIGNED-NEXT:    s_clause 0x4
-; ALIGNED-NEXT:    buffer_load_ubyte v125, v2, s[0:3], 0 offen offset:12
-; ALIGNED-NEXT:    buffer_load_ubyte v109, v2, s[0:3], 0 offen offset:13
-; ALIGNED-NEXT:    buffer_load_ubyte v121, v2, s[0:3], 0 offen offset:14
-; ALIGNED-NEXT:    buffer_load_ubyte v111, v2, s[0:3], 0 offen offset:15
-; ALIGNED-NEXT:    buffer_load_ubyte v79, v2, s[0:3], 0 offen offset:11
-; ALIGNED-NEXT:    v_lshl_or_b32 v103, v63, 8, v92
+; ALIGNED-NEXT:    buffer_store_dword v79, off, s[0:3], s32 offset:1372 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v79, v79, 8, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v88, 16, v79
+; ALIGNED-NEXT:    v_lshl_or_b32 v79, v95, 8, v91
+; ALIGNED-NEXT:    v_lshl_or_b32 v88, v106, 8, v104
 ; ALIGNED-NEXT:    s_clause 0x2
-; ALIGNED-NEXT:    buffer_load_ubyte v92, v2, s[0:3], 0 offen offset:8
-; ALIGNED-NEXT:    buffer_load_ubyte v77, v2, s[0:3], 0 offen offset:9
-; ALIGNED-NEXT:    buffer_load_ubyte v63, v2, s[0:3], 0 offen offset:10
-; ALIGNED-NEXT:    buffer_store_dword v96, off, s[0:3], s32 offset:1396 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v96, v119, 16, v103
-; ALIGNED-NEXT:    buffer_store_dword v96, off, s[0:3], s32 offset:1416 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
-; ALIGNED-NEXT:    v_lshl_or_b32 v103, v109, 8, v125
-; ALIGNED-NEXT:    s_waitcnt vmcnt(4)
-; ALIGNED-NEXT:    v_lshl_or_b32 v119, v111, 8, v121
-; ALIGNED-NEXT:    v_lshl_or_b32 v96, v119, 16, v103
+; ALIGNED-NEXT:    buffer_load_ubyte v106, v2, s[0:3], 0 offen offset:8
+; ALIGNED-NEXT:    buffer_load_ubyte v104, v2, s[0:3], 0 offen offset:9
+; ALIGNED-NEXT:    buffer_load_ubyte v95, v2, s[0:3], 0 offen offset:10
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1408 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v88, 16, v79
+; ALIGNED-NEXT:    v_lshl_or_b32 v79, v111, 8, v123
+; ALIGNED-NEXT:    v_lshl_or_b32 v88, v110, 8, v122
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1420 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v88, 16, v79
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1444 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v103, v77, 8, v92
+; ALIGNED-NEXT:    v_lshl_or_b32 v79, v104, 8, v106
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    v_lshl_or_b32 v119, v79, 8, v63
-; ALIGNED-NEXT:    buffer_store_dword v96, off, s[0:3], s32 offset:1444 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v96, v119, 16, v103
+; ALIGNED-NEXT:    v_lshl_or_b32 v88, v105, 8, v95
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v88, 16, v79
 ; ALIGNED-NEXT:    s_clause 0x2
-; ALIGNED-NEXT:    buffer_load_ubyte v119, v2, s[0:3], 0 offen offset:18
-; ALIGNED-NEXT:    buffer_load_ubyte v61, v2, s[0:3], 0 offen offset:16
-; ALIGNED-NEXT:    buffer_load_ubyte v103, v2, s[0:3], 0 offen offset:17
+; ALIGNED-NEXT:    buffer_load_ubyte v88, v2, s[0:3], 0 offen offset:18
+; ALIGNED-NEXT:    buffer_load_ubyte v91, v2, s[0:3], 0 offen offset:16
+; ALIGNED-NEXT:    buffer_load_ubyte v79, v2, s[0:3], 0 offen offset:17
 ; ALIGNED-NEXT:    v_add_nc_u32_e32 v2, 0x100, v2
-; ALIGNED-NEXT:    buffer_store_dword v96, off, s[0:3], s32 offset:1448 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v99, off, s[0:3], s32 offset:232
-; ALIGNED-NEXT:    buffer_store_dword v98, off, s[0:3], s32 offset:236
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1448 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:232
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:236
 ; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:228
 ; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:224
-; ALIGNED-NEXT:    v_add_co_u32 v3, vcc_lo, v87, 3
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v4, null, 0, v88, vcc_lo
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v40 offset:247
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v118 offset:248
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v42 offset:246
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v101 offset:252
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v114 offset:250
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v112 offset:251
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v116 offset:249
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v44 offset:245
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v76 offset:239
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v72 offset:240
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v75 offset:238
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v46 offset:244
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v58 offset:242
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v56 offset:243
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v73 offset:241
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v91 offset:237
-; ALIGNED-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:248
-; ALIGNED-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:252
-; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:244
-; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:240
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v104 offset:231
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v93 offset:232
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v106 offset:230
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v90 offset:236
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v78 offset:234
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v89 offset:235
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v94 offset:233
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v108 offset:229
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v126 offset:223
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v124 offset:224
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:222
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v110 offset:228
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v122 offset:226
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v120 offset:227
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v1 offset:225
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v5 offset:221
-; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:192
-; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:204
-; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:200
-; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:196
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v21 offset:206
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v20 offset:208
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v22 offset:207
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v13 offset:210
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v11 offset:212
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v14 offset:211
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v18 offset:209
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v12 offset:215
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v10 offset:216
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v15 offset:214
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v6 offset:220
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v7 offset:218
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v8 offset:219
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v16 offset:217
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v17 offset:213
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v23 offset:205
-; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:216
-; ALIGNED-NEXT:    buffer_store_dword v80, off, s[0:3], s32 offset:220
-; ALIGNED-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:212
-; ALIGNED-NEXT:    buffer_store_dword v74, off, s[0:3], s32 offset:208
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1484 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v29 offset:199
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v30 offset:200
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v31 offset:198
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v24 offset:204
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v25 offset:202
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v26 offset:203
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v28 offset:201
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v33 offset:197
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v39 offset:191
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v38 offset:192
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v48 offset:190
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v34 offset:196
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v35 offset:194
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v36 offset:195
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v49 offset:193
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v51 offset:189
-; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    v_lshl_or_b32 v96, v127, 8, v119
+; ALIGNED-NEXT:    s_clause 0x1 ; 8-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:704
+; ALIGNED-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:708
+; ALIGNED-NEXT:    s_waitcnt vmcnt(4)
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v126, 8, v88
+; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
+; ALIGNED-NEXT:    v_lshl_or_b32 v126, v79, 8, v91
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v127, v103, 8, v61
-; ALIGNED-NEXT:    v_lshl_or_b32 v96, v96, 16, v127
+; ALIGNED-NEXT:    v_add_co_u32 v3, vcc_lo, v3, s4
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v4, null, s5, v4, vcc_lo
+; ALIGNED-NEXT:    v_lshl_or_b32 v126, v0, 16, v126
+; ALIGNED-NEXT:    v_add_co_u32 v5, vcc_lo, v3, 3
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v6, null, 0, v4, vcc_lo
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v11 offset:247
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v10 offset:248
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v12 offset:246
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v1 offset:252
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v7 offset:250
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v8 offset:251
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v9 offset:249
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v13 offset:245
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v17 offset:239
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v18 offset:240
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v19 offset:238
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v14 offset:244
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v15 offset:242
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v16 offset:243
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v20 offset:241
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v21 offset:237
+; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248
+; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:252
+; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:244
+; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:240
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v27 offset:231
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v26 offset:232
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v28 offset:230
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v22 offset:236
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v23 offset:234
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v24 offset:235
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v25 offset:233
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v29 offset:229
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v35 offset:223
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v34 offset:224
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v36 offset:222
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v30 offset:228
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v31 offset:226
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v32 offset:227
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v37 offset:225
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v39 offset:221
+; ALIGNED-NEXT:    buffer_store_dword v85, off, s[0:3], s32 offset:192
+; ALIGNED-NEXT:    buffer_store_dword v98, off, s[0:3], s32 offset:204
+; ALIGNED-NEXT:    buffer_store_dword v103, off, s[0:3], s32 offset:200
+; ALIGNED-NEXT:    buffer_store_dword v116, off, s[0:3], s32 offset:196
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v53 offset:210
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v52 offset:212
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v71 offset:206
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v70 offset:208
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v80 offset:207
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v64 offset:211
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v68 offset:209
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v55 offset:215
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v54 offset:216
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v65 offset:214
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v48 offset:220
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v49 offset:218
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v50 offset:219
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v66 offset:217
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v67 offset:213
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v81 offset:205
+; ALIGNED-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:216
+; ALIGNED-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:220
+; ALIGNED-NEXT:    buffer_store_dword v89, off, s[0:3], s32 offset:212
+; ALIGNED-NEXT:    buffer_store_dword v107, off, s[0:3], s32 offset:208
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1484 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v87 offset:199
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v96 offset:200
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v97 offset:198
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v82 offset:204
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v83 offset:202
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v84 offset:203
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v86 offset:201
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v99 offset:197
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v113 offset:191
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v112 offset:192
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v114 offset:190
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v100 offset:196
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v101 offset:194
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v102 offset:195
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v115 offset:193
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v117 offset:189
+; ALIGNED-NEXT:    s_add_u32 s4, s4, 0x100
+; ALIGNED-NEXT:    s_addc_u32 s5, s5, 0
+; ALIGNED-NEXT:    s_cmp_lg_u64 s[4:5], 0x800
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:296
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1480 ; 4-byte Folded Reload
@@ -13655,22 +13650,22 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:288
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1468 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v65 offset:183
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v64 offset:184
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v66 offset:182
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v52 offset:188
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v53 offset:186
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v54 offset:187
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v55 offset:185
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v67 offset:181
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v81 offset:175
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v82 offset:176
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v83 offset:174
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v68 offset:180
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v69 offset:178
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v70 offset:179
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v84 offset:177
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v85 offset:173
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v43 offset:183
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v42 offset:184
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v44 offset:182
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v118 offset:188
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v119 offset:186
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v40 offset:187
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v41 offset:185
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v45 offset:181
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v59 offset:175
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v60 offset:176
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v61 offset:174
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v46 offset:180
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v47 offset:178
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v56 offset:179
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v62 offset:177
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v63 offset:173
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:312
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1464 ; 4-byte Folded Reload
@@ -13679,1478 +13674,1477 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1460 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:308
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1456 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1452 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:304
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1440 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v115 offset:167
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v117 offset:168
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v43 offset:166
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v86 offset:172
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v100 offset:170
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v102 offset:171
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v113 offset:169
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v45 offset:165
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v95 offset:159
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v97 offset:160
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v105 offset:158
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v47 offset:164
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v57 offset:162
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v59 offset:163
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v107 offset:161
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v123 offset:157
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v77 offset:167
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v76 offset:168
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v78 offset:166
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v72 offset:172
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v73 offset:170
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v74 offset:171
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v75 offset:169
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v90 offset:165
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v109 offset:159
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v108 offset:160
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v120 offset:158
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v92 offset:164
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v93 offset:162
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v94 offset:163
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v121 offset:161
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v124 offset:157
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:264
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1424 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:268
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1392 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1388 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:260
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1340 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1348 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:256
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1436 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:151
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:151
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1432 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:152
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:152
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1428 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:150
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:150
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v125 offset:156
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v127 offset:154
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1412 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:156
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:155
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1404 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:154
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1408 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:153
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1416 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:155
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1400 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:149
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1344 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:153
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1420 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:143
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1340 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:149
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:144
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1336 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:143
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1332 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:144
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1328 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:142
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1364 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:142
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:148
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1356 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:148
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1348 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:146
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1360 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:146
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:147
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1352 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:147
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1344 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:145
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1324 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:145
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1332 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:141
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1320 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:141
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1328 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:280
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1304 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1312 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:284
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1280 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1288 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:276
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1260 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1268 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:272
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1316 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1324 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:135
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1312 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:135
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1320 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:136
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1308 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:136
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1316 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:134
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1304 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:134
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:140
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1296 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:140
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1288 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:138
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1300 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:138
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:139
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1292 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:139
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1284 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:137
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1308 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:137
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1300 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:133
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1264 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:133
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1256 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:127
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1260 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:127
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1252 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:128
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1256 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:128
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1248 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:126
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1284 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:126
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:132
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1276 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:132
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1268 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:130
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1280 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:130
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:131
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1272 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:131
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1264 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:129
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1244 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:129
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1252 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[87:88], v0 offset:128
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1240 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:128
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1248 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:360
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1224 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1232 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:364
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1200 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1208 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:356
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1180 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1188 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:352
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1236 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1244 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:119
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1232 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:119
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1240 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:120
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1228 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:120
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1236 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:118
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1224 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:118
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:124
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1216 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:124
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1208 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:122
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1220 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:122
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:123
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1212 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:123
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1204 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:121
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1228 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:121
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1220 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:117
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1184 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:117
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1176 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:111
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1180 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:111
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1172 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:112
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1176 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:112
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1168 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:110
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1204 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:110
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:116
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1196 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:116
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1188 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:114
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1200 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:114
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:115
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1192 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:115
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1184 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:113
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1164 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:113
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1172 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:109
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1160 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:109
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1168 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:376
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1144 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1152 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:380
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1120 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1128 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:372
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1116 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1124 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:368
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1156 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1164 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:103
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1152 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:103
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1160 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:104
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1148 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:104
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1156 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:102
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:102
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1144 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:108
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1136 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:108
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1128 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:106
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1140 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:106
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:107
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1132 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:107
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1124 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:105
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1148 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:105
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1140 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:101
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1104 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:101
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1096 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:95
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1100 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:95
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1092 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:96
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1096 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:96
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1088 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:94
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1120 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:94
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:100
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1112 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:100
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1104 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:98
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1116 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:98
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:99
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1108 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:99
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1100 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:97
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1084 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:97
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1092 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:93
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1080 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:93
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1088 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:328
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1076 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1084 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:332
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1052 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1080 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:324
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1036 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1044 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:320
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1076 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:87
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1072 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:87
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:88
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1068 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:88
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1064 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:86
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1060 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:86
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1056 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:92
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1052 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:92
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1044 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:90
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1056 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:90
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:91
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1048 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:91
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1040 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:89
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1064 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:89
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1060 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:85
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1024 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:85
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1016 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:79
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1020 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:79
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1012 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:80
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1016 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:80
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1008 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:78
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1040 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:78
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:84
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1032 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:84
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1024 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:82
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1036 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:82
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:83
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1028 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:83
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1020 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:81
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1004 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:81
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1012 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:77
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1000 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:77
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1008 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:344
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:996 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1004 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:348
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:340
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:336
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:992 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1000 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:71
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:988 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:71
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:996 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:72
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:992 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:72
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:70
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:984 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:70
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:76
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:976 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:76
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:968 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:74
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:980 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:74
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:75
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:968 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:75
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:73
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:988 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:73
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:980 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:69
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:69
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:63
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:63
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:64
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:64
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:62
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:62
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:68
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:68
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:66
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:67
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:944 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:66
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:65
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:67
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:64
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:65
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[87:88], v0 offset:64
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:424
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:424
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:428
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:428
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:416
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:58
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:55
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:416
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:58
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:56
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:55
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:54
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:56
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:60
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:54
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:60
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:59
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:57
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:59
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:53
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:57
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:53
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:864 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:50
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:50
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:47
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:47
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:48
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:48
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:46
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:46
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:52
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:52
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:51
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:51
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:49
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:49
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:45
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:45
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:444
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:440
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:436
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:432
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:40
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:39
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:38
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:37
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:40
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:44
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:39
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:43
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:38
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:42
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:37
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:44
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:43
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:41
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:42
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:32
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:41
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:31
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:30
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:32
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:36
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:31
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:35
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:30
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:34
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[87:88], v0 offset:32
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:36
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:35
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:34
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:33
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:33
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:392
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:396
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:388
-; ALIGNED-NEXT:    buffer_store_dword v96, off, s[0:3], s32 offset:384
+; ALIGNED-NEXT:    buffer_store_dword v126, off, s[0:3], s32 offset:384
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:23
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:24
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:23
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:22
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:24
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:22
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:28
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:28
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:26
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:27
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:26
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:25
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:27
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:21
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v88 offset:15
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1456 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:25
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:16
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v79 offset:14
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:21
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v119 offset:15
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1452 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:16
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v103 offset:14
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:20
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:20
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:18
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:18
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:19
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:19
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:17
-; ALIGNED-NEXT:    flat_store_byte v[87:88], v61 offset:16
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:17
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v91 offset:16
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1448 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:408
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1444 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1416 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1420 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:404
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1396 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1408 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:400
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1384 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v63 offset:7
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v79 offset:8
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v109 offset:10
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v77 offset:6
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v111 offset:12
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v121 offset:11
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v125 offset:9
-; ALIGNED-NEXT:    flat_store_byte v[87:88], v92 offset:8
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[87:88], v0 offset:2
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1368 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v95 offset:7
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v105 offset:8
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v111 offset:10
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v104 offset:6
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v110 offset:12
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v122 offset:11
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v123 offset:9
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v106 offset:8
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[87:88], v0 offset:1
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1388 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:2
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1376 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[87:88], v0 offset:4
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1364 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1372 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[87:88], v0
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1360 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    v_add_co_u32 v87, vcc_lo, 0x100, v87
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v88, null, 0, v88, vcc_lo
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:1
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1400 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:4
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1392 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:2
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1396 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[5:6], v0 offset:3
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1380 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:4
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1372 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:2
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1376 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1368 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:3
-; ALIGNED-NEXT:    s_cbranch_scc1 .LBB9_2
-; ALIGNED-NEXT:  .LBB9_3: ; %Flow16
-; ALIGNED-NEXT:    s_andn2_saveexec_b32 s6, s6
-; ALIGNED-NEXT:    s_cbranch_execz .LBB9_6
-; ALIGNED-NEXT:  ; %bb.4: ; %memmove_bwd_loop.preheader
-; ALIGNED-NEXT:    v_add_co_u32 v103, vcc_lo, 0x700, v87
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v104, null, 0, v88, vcc_lo
-; ALIGNED-NEXT:    s_movk_i32 s4, 0xf800
-; ALIGNED-NEXT:    s_mov_b32 s5, -1
-; ALIGNED-NEXT:  .LBB9_5: ; %memmove_bwd_loop
+; ALIGNED-NEXT:    flat_store_byte v[3:4], v0
+; ALIGNED-NEXT:    s_cbranch_scc1 .LBB9_1
+; ALIGNED-NEXT:  .LBB9_2: ; %Flow10
+; ALIGNED-NEXT:    s_andn2_saveexec_b32 s8, s6
+; ALIGNED-NEXT:    s_cbranch_execz .LBB9_5
+; ALIGNED-NEXT:  ; %bb.3: ; %memmove_bwd_loop.preheader
+; ALIGNED-NEXT:    v_add_nc_u32_e32 v6, 0x700, v2
+; ALIGNED-NEXT:    s_mov_b64 s[4:5], 0x700
+; ALIGNED-NEXT:    s_movk_i32 s6, 0xff00
+; ALIGNED-NEXT:    s_mov_b32 s7, -1
+; ALIGNED-NEXT:  .LBB9_4: ; %memmove_bwd_loop
 ; ALIGNED-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; ALIGNED-NEXT:    s_clause 0x3b
-; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1812
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1813
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:1814
-; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:1815
-; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:1816
-; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:1817
-; ALIGNED-NEXT:    buffer_load_ubyte v12, v2, s[0:3], 0 offen offset:1818
-; ALIGNED-NEXT:    buffer_load_ubyte v112, v2, s[0:3], 0 offen offset:1811
-; ALIGNED-NEXT:    buffer_load_ubyte v13, v2, s[0:3], 0 offen offset:1819
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:1820
-; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:1821
-; ALIGNED-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:1822
-; ALIGNED-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:1823
-; ALIGNED-NEXT:    buffer_load_ubyte v15, v2, s[0:3], 0 offen offset:1824
-; ALIGNED-NEXT:    buffer_load_ubyte v16, v2, s[0:3], 0 offen offset:1825
-; ALIGNED-NEXT:    buffer_load_ubyte v18, v2, s[0:3], 0 offen offset:1826
-; ALIGNED-NEXT:    buffer_load_ubyte v20, v2, s[0:3], 0 offen offset:1827
-; ALIGNED-NEXT:    buffer_load_ubyte v14, v2, s[0:3], 0 offen offset:1828
-; ALIGNED-NEXT:    buffer_load_ubyte v19, v2, s[0:3], 0 offen offset:1829
-; ALIGNED-NEXT:    buffer_load_ubyte v21, v2, s[0:3], 0 offen offset:1830
-; ALIGNED-NEXT:    buffer_load_ubyte v22, v2, s[0:3], 0 offen offset:1831
-; ALIGNED-NEXT:    buffer_load_ubyte v23, v2, s[0:3], 0 offen offset:1832
-; ALIGNED-NEXT:    buffer_load_ubyte v24, v2, s[0:3], 0 offen offset:1833
-; ALIGNED-NEXT:    buffer_load_ubyte v26, v2, s[0:3], 0 offen offset:1834
-; ALIGNED-NEXT:    buffer_load_ubyte v30, v2, s[0:3], 0 offen offset:1835
-; ALIGNED-NEXT:    buffer_load_ubyte v25, v2, s[0:3], 0 offen offset:1836
-; ALIGNED-NEXT:    buffer_load_ubyte v27, v2, s[0:3], 0 offen offset:1837
-; ALIGNED-NEXT:    buffer_load_ubyte v28, v2, s[0:3], 0 offen offset:1838
-; ALIGNED-NEXT:    buffer_load_ubyte v29, v2, s[0:3], 0 offen offset:1839
-; ALIGNED-NEXT:    buffer_load_ubyte v31, v2, s[0:3], 0 offen offset:1840
-; ALIGNED-NEXT:    buffer_load_ubyte v32, v2, s[0:3], 0 offen offset:1841
-; ALIGNED-NEXT:    buffer_load_ubyte v34, v2, s[0:3], 0 offen offset:1842
-; ALIGNED-NEXT:    buffer_load_ubyte v35, v2, s[0:3], 0 offen offset:1843
-; ALIGNED-NEXT:    buffer_load_ubyte v33, v2, s[0:3], 0 offen offset:1844
-; ALIGNED-NEXT:    buffer_load_ubyte v38, v2, s[0:3], 0 offen offset:1845
-; ALIGNED-NEXT:    buffer_load_ubyte v36, v2, s[0:3], 0 offen offset:1846
-; ALIGNED-NEXT:    buffer_load_ubyte v37, v2, s[0:3], 0 offen offset:1847
-; ALIGNED-NEXT:    buffer_load_ubyte v48, v2, s[0:3], 0 offen offset:1848
-; ALIGNED-NEXT:    buffer_load_ubyte v51, v2, s[0:3], 0 offen offset:1849
-; ALIGNED-NEXT:    buffer_load_ubyte v52, v2, s[0:3], 0 offen offset:1850
-; ALIGNED-NEXT:    buffer_load_ubyte v39, v2, s[0:3], 0 offen offset:1852
-; ALIGNED-NEXT:    buffer_load_ubyte v50, v2, s[0:3], 0 offen offset:1853
-; ALIGNED-NEXT:    buffer_load_ubyte v17, v2, s[0:3], 0 offen offset:1854
-; ALIGNED-NEXT:    buffer_load_ubyte v49, v2, s[0:3], 0 offen offset:1855
-; ALIGNED-NEXT:    buffer_load_ubyte v54, v2, s[0:3], 0 offen offset:1856
-; ALIGNED-NEXT:    buffer_load_ubyte v55, v2, s[0:3], 0 offen offset:1857
-; ALIGNED-NEXT:    buffer_load_ubyte v64, v2, s[0:3], 0 offen offset:1858
-; ALIGNED-NEXT:    buffer_load_ubyte v53, v2, s[0:3], 0 offen offset:1851
-; ALIGNED-NEXT:    buffer_load_ubyte v65, v2, s[0:3], 0 offen offset:1859
-; ALIGNED-NEXT:    buffer_load_ubyte v66, v2, s[0:3], 0 offen offset:1860
-; ALIGNED-NEXT:    buffer_load_ubyte v67, v2, s[0:3], 0 offen offset:1861
-; ALIGNED-NEXT:    buffer_load_ubyte v68, v2, s[0:3], 0 offen offset:1862
-; ALIGNED-NEXT:    buffer_load_ubyte v69, v2, s[0:3], 0 offen offset:1863
-; ALIGNED-NEXT:    buffer_load_ubyte v70, v2, s[0:3], 0 offen offset:1864
-; ALIGNED-NEXT:    buffer_load_ubyte v71, v2, s[0:3], 0 offen offset:1865
-; ALIGNED-NEXT:    buffer_load_ubyte v80, v2, s[0:3], 0 offen offset:1866
-; ALIGNED-NEXT:    buffer_load_ubyte v81, v2, s[0:3], 0 offen offset:1867
-; ALIGNED-NEXT:    buffer_load_ubyte v82, v2, s[0:3], 0 offen offset:1868
-; ALIGNED-NEXT:    buffer_load_ubyte v83, v2, s[0:3], 0 offen offset:1869
-; ALIGNED-NEXT:    buffer_load_ubyte v84, v2, s[0:3], 0 offen offset:1870
-; ALIGNED-NEXT:    s_add_u32 s4, s4, 0x100
-; ALIGNED-NEXT:    s_addc_u32 s5, s5, 0
-; ALIGNED-NEXT:    s_cmp_eq_u64 s[4:5], 0
+; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:20
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:21
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:22
+; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:23
+; ALIGNED-NEXT:    buffer_load_ubyte v5, v6, s[0:3], 0 offen offset:24
+; ALIGNED-NEXT:    buffer_load_ubyte v8, v6, s[0:3], 0 offen offset:25
+; ALIGNED-NEXT:    buffer_load_ubyte v11, v6, s[0:3], 0 offen offset:26
+; ALIGNED-NEXT:    buffer_load_ubyte v127, v6, s[0:3], 0 offen offset:19
+; ALIGNED-NEXT:    buffer_load_ubyte v12, v6, s[0:3], 0 offen offset:27
+; ALIGNED-NEXT:    buffer_load_ubyte v4, v6, s[0:3], 0 offen offset:28
+; ALIGNED-NEXT:    buffer_load_ubyte v7, v6, s[0:3], 0 offen offset:29
+; ALIGNED-NEXT:    buffer_load_ubyte v9, v6, s[0:3], 0 offen offset:30
+; ALIGNED-NEXT:    buffer_load_ubyte v10, v6, s[0:3], 0 offen offset:31
+; ALIGNED-NEXT:    buffer_load_ubyte v14, v6, s[0:3], 0 offen offset:32
+; ALIGNED-NEXT:    buffer_load_ubyte v15, v6, s[0:3], 0 offen offset:33
+; ALIGNED-NEXT:    buffer_load_ubyte v19, v6, s[0:3], 0 offen offset:34
+; ALIGNED-NEXT:    buffer_load_ubyte v20, v6, s[0:3], 0 offen offset:35
+; ALIGNED-NEXT:    buffer_load_ubyte v13, v6, s[0:3], 0 offen offset:36
+; ALIGNED-NEXT:    buffer_load_ubyte v16, v6, s[0:3], 0 offen offset:37
+; ALIGNED-NEXT:    buffer_load_ubyte v17, v6, s[0:3], 0 offen offset:38
+; ALIGNED-NEXT:    buffer_load_ubyte v18, v6, s[0:3], 0 offen offset:39
+; ALIGNED-NEXT:    buffer_load_ubyte v22, v6, s[0:3], 0 offen offset:40
+; ALIGNED-NEXT:    buffer_load_ubyte v27, v6, s[0:3], 0 offen offset:41
+; ALIGNED-NEXT:    buffer_load_ubyte v28, v6, s[0:3], 0 offen offset:42
+; ALIGNED-NEXT:    buffer_load_ubyte v29, v6, s[0:3], 0 offen offset:43
+; ALIGNED-NEXT:    buffer_load_ubyte v23, v6, s[0:3], 0 offen offset:44
+; ALIGNED-NEXT:    buffer_load_ubyte v24, v6, s[0:3], 0 offen offset:45
+; ALIGNED-NEXT:    buffer_load_ubyte v25, v6, s[0:3], 0 offen offset:46
+; ALIGNED-NEXT:    buffer_load_ubyte v26, v6, s[0:3], 0 offen offset:47
+; ALIGNED-NEXT:    buffer_load_ubyte v30, v6, s[0:3], 0 offen offset:48
+; ALIGNED-NEXT:    buffer_load_ubyte v31, v6, s[0:3], 0 offen offset:49
+; ALIGNED-NEXT:    buffer_load_ubyte v33, v6, s[0:3], 0 offen offset:50
+; ALIGNED-NEXT:    buffer_load_ubyte v34, v6, s[0:3], 0 offen offset:51
+; ALIGNED-NEXT:    buffer_load_ubyte v32, v6, s[0:3], 0 offen offset:52
+; ALIGNED-NEXT:    buffer_load_ubyte v37, v6, s[0:3], 0 offen offset:53
+; ALIGNED-NEXT:    buffer_load_ubyte v35, v6, s[0:3], 0 offen offset:54
+; ALIGNED-NEXT:    buffer_load_ubyte v36, v6, s[0:3], 0 offen offset:55
+; ALIGNED-NEXT:    buffer_load_ubyte v38, v6, s[0:3], 0 offen offset:56
+; ALIGNED-NEXT:    buffer_load_ubyte v50, v6, s[0:3], 0 offen offset:57
+; ALIGNED-NEXT:    buffer_load_ubyte v51, v6, s[0:3], 0 offen offset:58
+; ALIGNED-NEXT:    buffer_load_ubyte v21, v6, s[0:3], 0 offen offset:60
+; ALIGNED-NEXT:    buffer_load_ubyte v49, v6, s[0:3], 0 offen offset:61
+; ALIGNED-NEXT:    buffer_load_ubyte v39, v6, s[0:3], 0 offen offset:62
+; ALIGNED-NEXT:    buffer_load_ubyte v48, v6, s[0:3], 0 offen offset:63
+; ALIGNED-NEXT:    buffer_load_ubyte v53, v6, s[0:3], 0 offen offset:64
+; ALIGNED-NEXT:    buffer_load_ubyte v54, v6, s[0:3], 0 offen offset:65
+; ALIGNED-NEXT:    buffer_load_ubyte v55, v6, s[0:3], 0 offen offset:66
+; ALIGNED-NEXT:    buffer_load_ubyte v52, v6, s[0:3], 0 offen offset:59
+; ALIGNED-NEXT:    buffer_load_ubyte v64, v6, s[0:3], 0 offen offset:67
+; ALIGNED-NEXT:    buffer_load_ubyte v65, v6, s[0:3], 0 offen offset:68
+; ALIGNED-NEXT:    buffer_load_ubyte v66, v6, s[0:3], 0 offen offset:69
+; ALIGNED-NEXT:    buffer_load_ubyte v67, v6, s[0:3], 0 offen offset:70
+; ALIGNED-NEXT:    buffer_load_ubyte v68, v6, s[0:3], 0 offen offset:71
+; ALIGNED-NEXT:    buffer_load_ubyte v69, v6, s[0:3], 0 offen offset:72
+; ALIGNED-NEXT:    buffer_load_ubyte v70, v6, s[0:3], 0 offen offset:73
+; ALIGNED-NEXT:    buffer_load_ubyte v71, v6, s[0:3], 0 offen offset:74
+; ALIGNED-NEXT:    buffer_load_ubyte v80, v6, s[0:3], 0 offen offset:75
+; ALIGNED-NEXT:    buffer_load_ubyte v81, v6, s[0:3], 0 offen offset:76
+; ALIGNED-NEXT:    buffer_load_ubyte v82, v6, s[0:3], 0 offen offset:77
+; ALIGNED-NEXT:    buffer_load_ubyte v83, v6, s[0:3], 0 offen offset:78
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(59)
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(58)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v1, 8, v0
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(57)
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(56)
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v5, 8, v4
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(55)
-; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(54)
-; ALIGNED-NEXT:    v_lshl_or_b32 v7, v9, 8, v7
-; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
+; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(53)
+; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v0
+; ALIGNED-NEXT:    s_waitcnt vmcnt(51)
+; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(50)
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(49)
-; ALIGNED-NEXT:    v_lshl_or_b32 v5, v8, 8, v6
-; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(48)
+; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(47)
-; ALIGNED-NEXT:    v_lshl_or_b32 v6, v11, 8, v10
-; ALIGNED-NEXT:    v_lshl_or_b32 v8, v13, 8, v12
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 8, v2
+; ALIGNED-NEXT:    v_lshl_or_b32 v2, v7, 8, v4
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v10, 8, v9
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v8, 8, v5
+; ALIGNED-NEXT:    v_lshl_or_b32 v5, v12, 8, v11
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(45)
-; ALIGNED-NEXT:    v_lshl_or_b32 v9, v16, 8, v15
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v6, 16, v5
+; ALIGNED-NEXT:    v_lshl_or_b32 v7, v15, 8, v14
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(43)
-; ALIGNED-NEXT:    v_lshl_or_b32 v10, v20, 8, v18
-; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v8, v20, 8, v19
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(41)
-; ALIGNED-NEXT:    v_lshl_or_b32 v11, v19, 8, v14
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v8, 16, v7
+; ALIGNED-NEXT:    v_lshl_or_b32 v9, v16, 8, v13
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(39)
-; ALIGNED-NEXT:    v_lshl_or_b32 v12, v22, 8, v21
-; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v10, v18, 8, v17
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(37)
-; ALIGNED-NEXT:    v_lshl_or_b32 v13, v24, 8, v23
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v10, 16, v9
+; ALIGNED-NEXT:    v_lshl_or_b32 v11, v27, 8, v22
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(35)
-; ALIGNED-NEXT:    v_lshl_or_b32 v14, v30, 8, v26
-; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v12, v29, 8, v28
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
+; ALIGNED-NEXT:    v_lshl_or_b32 v2, v5, 16, v4
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v8, 16, v7
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v10, 16, v9
+; ALIGNED-NEXT:    v_lshl_or_b32 v5, v12, 16, v11
+; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(33)
-; ALIGNED-NEXT:    v_lshl_or_b32 v15, v27, 8, v25
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v12, 16, v11
+; ALIGNED-NEXT:    v_lshl_or_b32 v13, v24, 8, v23
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(31)
-; ALIGNED-NEXT:    v_lshl_or_b32 v16, v29, 8, v28
+; ALIGNED-NEXT:    v_lshl_or_b32 v14, v26, 8, v25
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(29)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v32, 8, v31
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v31, 8, v30
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(27)
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v35, 8, v34
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v34, 8, v33
+; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(25)
-; ALIGNED-NEXT:    v_lshl_or_b32 v5, v38, 8, v33
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v14, 16, v13
+; ALIGNED-NEXT:    v_lshl_or_b32 v2, v37, 8, v32
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(23)
-; ALIGNED-NEXT:    v_lshl_or_b32 v6, v37, 8, v36
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v36, 8, v35
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(18)
-; ALIGNED-NEXT:    v_lshl_or_b32 v7, v50, 8, v39
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v49, 8, v21
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(16)
-; ALIGNED-NEXT:    v_lshl_or_b32 v8, v49, 8, v17
-; ALIGNED-NEXT:    v_lshl_or_b32 v9, v51, 8, v48
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v16, 16, v15
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
+; ALIGNED-NEXT:    v_lshl_or_b32 v5, v48, 8, v39
+; ALIGNED-NEXT:    v_lshl_or_b32 v7, v14, 16, v13
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
+; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:80
+; ALIGNED-NEXT:    v_lshl_or_b32 v2, v5, 16, v4
+; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v7, v50, 8, v38
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(13)
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v52, 8, v51
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v54, 8, v53
+; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(12)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v53, 8, v52
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v55, 8, v54
-; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v6, 16, v5
-; ALIGNED-NEXT:    s_waitcnt vmcnt(11)
-; ALIGNED-NEXT:    v_lshl_or_b32 v5, v65, 8, v64
+; ALIGNED-NEXT:    v_lshl_or_b32 v2, v64, 8, v55
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v0, 16, v7
+; ALIGNED-NEXT:    s_clause 0x3
+; ALIGNED-NEXT:    buffer_load_ubyte v4, v6, s[0:3], 0 offen offset:82
+; ALIGNED-NEXT:    buffer_load_ubyte v5, v6, s[0:3], 0 offen offset:83
+; ALIGNED-NEXT:    buffer_load_ubyte v8, v6, s[0:3], 0 offen offset:90
+; ALIGNED-NEXT:    buffer_load_ubyte v9, v6, s[0:3], 0 offen offset:108
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(14)
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v66, 8, v65
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:81
+; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:964 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(13)
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v68, 8, v67
+; ALIGNED-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:79
+; ALIGNED-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:968 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v82, 8, v81
+; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v8, 16, v7
-; ALIGNED-NEXT:    s_clause 0x4
-; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:1875
-; ALIGNED-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:1899
-; ALIGNED-NEXT:    buffer_load_ubyte v14, v2, s[0:3], 0 offen offset:1907
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1872
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:1874
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v3, 16, v9
-; ALIGNED-NEXT:    s_waitcnt vmcnt(14)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v67, 8, v66
-; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:1882
-; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v5, 16, v4
-; ALIGNED-NEXT:    s_waitcnt vmcnt(13)
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v69, 8, v68
-; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:1873
-; ALIGNED-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:956 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:960 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1871
-; ALIGNED-NEXT:    s_waitcnt vmcnt(9)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v83, 8, v82
-; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:928 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:928 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:956 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v64, off, s[0:3], s32 offset:948 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v65, off, s[0:3], s32 offset:940 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:952 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:980 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:988 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v80, off, s[0:3], s32 offset:996 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:992 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v82, off, s[0:3], s32 offset:964 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v83, off, s[0:3], s32 offset:968 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
-; ALIGNED-NEXT:    buffer_store_dword v84, off, s[0:3], s32 offset:972 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v112, off, s[0:3], s32 offset:1464 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:1883
-; ALIGNED-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:1904
-; ALIGNED-NEXT:    buffer_load_ubyte v12, v2, s[0:3], 0 offen offset:1905
-; ALIGNED-NEXT:    buffer_load_ubyte v13, v2, s[0:3], 0 offen offset:1906
+; ALIGNED-NEXT:    buffer_store_dword v66, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v67, off, s[0:3], s32 offset:952 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v68, off, s[0:3], s32 offset:960 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:988 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v70, off, s[0:3], s32 offset:996 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v71, off, s[0:3], s32 offset:1004 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v80, off, s[0:3], s32 offset:1000 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v81, off, s[0:3], s32 offset:972 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v82, off, s[0:3], s32 offset:976 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
+; ALIGNED-NEXT:    buffer_store_dword v83, off, s[0:3], s32 offset:980 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v127, off, s[0:3], s32 offset:1464 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_clause 0x4
+; ALIGNED-NEXT:    buffer_load_ubyte v7, v6, s[0:3], 0 offen offset:91
+; ALIGNED-NEXT:    buffer_load_ubyte v10, v6, s[0:3], 0 offen offset:112
+; ALIGNED-NEXT:    buffer_load_ubyte v11, v6, s[0:3], 0 offen offset:113
+; ALIGNED-NEXT:    buffer_load_ubyte v13, v6, s[0:3], 0 offen offset:115
+; ALIGNED-NEXT:    buffer_load_ubyte v12, v6, s[0:3], 0 offen offset:114
+; ALIGNED-NEXT:    s_waitcnt vmcnt(11)
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1012 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(10)
-; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:1152 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(9)
-; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:1172 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1024 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1004 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1084 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1016 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:1132 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
-; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:1076 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1016 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(5)
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1008 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(4)
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v0, 8, v84
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:976 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v71, 8, v70
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v81, 8, v80
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:984 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v5, 8, v1
-; ALIGNED-NEXT:    s_clause 0x1
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1877
-; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:1879
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v7, 8, v6
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1000 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1012 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:984 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v1, 8, v83
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v80, 8, v71
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:992 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v70, 8, v69
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1008 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v2, 8, v3
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:86
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v5, 8, v4
 ; ALIGNED-NEXT:    s_clause 0x1
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:1880
-; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:1881
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:1878
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1020 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1876
+; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:87
+; ALIGNED-NEXT:    buffer_load_ubyte v4, v6, s[0:3], 0 offen offset:88
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1020 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_load_ubyte v5, v6, s[0:3], 0 offen offset:89
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:85
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1028 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:84
+; ALIGNED-NEXT:    s_waitcnt vmcnt(10)
+; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1080 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(9)
-; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1072 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:1172 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
-; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:1164 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:1176 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
-; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:1168 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:1180 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
-; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:1176 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:1184 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(5)
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1028 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1040 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(4)
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1036 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1044 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1060 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1068 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1068 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1076 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1032 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v5, 8, v4
-; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:1887
-; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v1, 8, v0
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1024 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1885
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:1886
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1052 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1884
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1036 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1032 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 8, v2
+; ALIGNED-NEXT:    s_clause 0x1
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:94
+; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:95
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:93
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1060 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:92
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1056 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1056 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1044 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1064 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1048 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v5, 8, v4
-; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:1892
-; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v1, 8, v0
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1040 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1889
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v7, 8, v6
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v8, 8, v9
-; ALIGNED-NEXT:    s_clause 0x2
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:1890
-; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:1896
-; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:1898
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1064 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1052 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1048 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 8, v2
 ; ALIGNED-NEXT:    s_clause 0x1
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:1891
-; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:1897
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1080 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1888
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:99
+; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:100
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v7, 8, v8
+; ALIGNED-NEXT:    s_clause 0x1
+; ALIGNED-NEXT:    buffer_load_ubyte v8, v6, s[0:3], 0 offen offset:107
+; ALIGNED-NEXT:    buffer_load_ubyte v7, v6, s[0:3], 0 offen offset:105
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1072 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v5, 8, v4
+; ALIGNED-NEXT:    s_clause 0x1
+; ALIGNED-NEXT:    buffer_load_ubyte v4, v6, s[0:3], 0 offen offset:98
+; ALIGNED-NEXT:    buffer_load_ubyte v5, v6, s[0:3], 0 offen offset:104
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:97
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1088 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:96
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1104 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1100 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1088 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1112 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(5)
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1096 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(4)
-; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1140 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1160 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:1156 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1104 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1092 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v4, 8, v6
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:1900
-; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1148 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1148 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v1, 8, v0
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1084 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1894
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:1895
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1100 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1893
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1096 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1092 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 8, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v2, 8, v4
+; ALIGNED-NEXT:    s_clause 0x1
+; ALIGNED-NEXT:    buffer_load_ubyte v4, v6, s[0:3], 0 offen offset:103
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:106
+; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1156 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:102
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1108 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:101
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1124 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1124 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1112 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1164 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1116 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v4, 8, v1
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1902
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1120 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v4, 8, v1
+; ALIGNED-NEXT:    buffer_load_ubyte v4, v6, s[0:3], 0 offen offset:116
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v0, 8, v5
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1108 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:1908
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:1903
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1120 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1901
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1116 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v0, 8, v3
+; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:111
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:110
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1128 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:109
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1132 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1192 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1184 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1144 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1136 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v4, 8, v1
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1910
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1140 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 8, v1
+; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:119
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v0, 8, v6
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1128 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:1916
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v8, 8, v7
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v10, 8, v9
-; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:1912
-; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:1913
-; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:1914
-; ALIGNED-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:1915
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1144 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v12, 8, v11
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v14, 8, v13
-; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:1920
-; ALIGNED-NEXT:    buffer_load_ubyte v12, v2, s[0:3], 0 offen offset:1921
-; ALIGNED-NEXT:    buffer_load_ubyte v13, v2, s[0:3], 0 offen offset:1922
-; ALIGNED-NEXT:    buffer_load_ubyte v14, v2, s[0:3], 0 offen offset:1923
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1160 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:1911
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1180 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1909
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1136 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v0, 8, v9
+; ALIGNED-NEXT:    buffer_load_ubyte v9, v6, s[0:3], 0 offen offset:124
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v8, 8, v2
+; ALIGNED-NEXT:    s_clause 0x1
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:122
+; ALIGNED-NEXT:    buffer_load_ubyte v8, v6, s[0:3], 0 offen offset:123
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1152 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v7, 8, v5
+; ALIGNED-NEXT:    s_clause 0x1
+; ALIGNED-NEXT:    buffer_load_ubyte v5, v6, s[0:3], 0 offen offset:120
+; ALIGNED-NEXT:    buffer_load_ubyte v7, v6, s[0:3], 0 offen offset:121
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v13, 8, v12
+; ALIGNED-NEXT:    s_clause 0x1
+; ALIGNED-NEXT:    buffer_load_ubyte v12, v6, s[0:3], 0 offen offset:130
+; ALIGNED-NEXT:    buffer_load_ubyte v13, v6, s[0:3], 0 offen offset:131
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1168 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v11, 8, v10
+; ALIGNED-NEXT:    s_clause 0x1
+; ALIGNED-NEXT:    buffer_load_ubyte v10, v6, s[0:3], 0 offen offset:128
+; ALIGNED-NEXT:    buffer_load_ubyte v11, v6, s[0:3], 0 offen offset:129
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:118
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1188 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:117
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(11)
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1192 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1204 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(10)
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1204 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:1212 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(9)
-; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1220 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1244 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
-; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1228 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1240 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
-; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:1236 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1228 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
-; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:1232 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1236 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1196 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v4, 8, v1
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1918
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1200 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 8, v1
+; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:127
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v0, 8, v5
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1188 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:1924
-; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:1244 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:1248 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:1919
-; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:1256 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:1252 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1200 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1917
-; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1212 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1264 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1216 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v4, 8, v1
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1926
-; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v0, 8, v6
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1196 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v0, 8, v4
+; ALIGNED-NEXT:    buffer_load_ubyte v4, v6, s[0:3], 0 offen offset:132
+; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:1252 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:1256 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:1264 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:126
+; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:1260 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1208 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v8, 8, v7
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v10, 8, v9
-; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:1928
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1224 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v12, 8, v11
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v14, 8, v13
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1240 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:1927
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1260 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1925
+; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:125
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1272 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1224 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1300 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1272 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1276 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v4, 8, v1
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1220 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 8, v1
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v0, 8, v5
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1216 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v0, 8, v9
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v8, 8, v2
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1232 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v7, 8, v5
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v13, 8, v12
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1248 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v11, 8, v10
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1268 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_clause 0x6
-; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:1929
-; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:1930
-; ALIGNED-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:1931
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:1932
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1934
-; ALIGNED-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:1936
-; ALIGNED-NEXT:    buffer_load_ubyte v12, v2, s[0:3], 0 offen offset:1937
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    s_clause 0x3
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:1935
-; ALIGNED-NEXT:    buffer_load_ubyte v13, v2, s[0:3], 0 offen offset:1938
-; ALIGNED-NEXT:    buffer_load_ubyte v14, v2, s[0:3], 0 offen offset:1939
-; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:1940
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1280 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_clause 0x3e
-; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1933
-; ALIGNED-NEXT:    buffer_load_ubyte v107, v2, s[0:3], 0 offen offset:1957
-; ALIGNED-NEXT:    buffer_load_ubyte v95, v2, s[0:3], 0 offen offset:1958
-; ALIGNED-NEXT:    buffer_load_ubyte v93, v2, s[0:3], 0 offen offset:1959
-; ALIGNED-NEXT:    buffer_load_ubyte v91, v2, s[0:3], 0 offen offset:1960
-; ALIGNED-NEXT:    buffer_load_ubyte v89, v2, s[0:3], 0 offen offset:1961
-; ALIGNED-NEXT:    buffer_load_ubyte v75, v2, s[0:3], 0 offen offset:1962
-; ALIGNED-NEXT:    buffer_load_ubyte v79, v2, s[0:3], 0 offen offset:1963
-; ALIGNED-NEXT:    buffer_load_ubyte v63, v2, s[0:3], 0 offen offset:1964
-; ALIGNED-NEXT:    buffer_load_ubyte v47, v2, s[0:3], 0 offen offset:1965
-; ALIGNED-NEXT:    buffer_load_ubyte v45, v2, s[0:3], 0 offen offset:1966
-; ALIGNED-NEXT:    buffer_load_ubyte v42, v2, s[0:3], 0 offen offset:1967
-; ALIGNED-NEXT:    buffer_load_ubyte v115, v2, s[0:3], 0 offen offset:1968
-; ALIGNED-NEXT:    buffer_load_ubyte v100, v2, s[0:3], 0 offen offset:1969
-; ALIGNED-NEXT:    buffer_load_ubyte v99, v2, s[0:3], 0 offen offset:1970
-; ALIGNED-NEXT:    buffer_load_ubyte v98, v2, s[0:3], 0 offen offset:1971
-; ALIGNED-NEXT:    buffer_load_ubyte v101, v2, s[0:3], 0 offen offset:1972
-; ALIGNED-NEXT:    buffer_load_ubyte v83, v2, s[0:3], 0 offen offset:1973
-; ALIGNED-NEXT:    buffer_load_ubyte v84, v2, s[0:3], 0 offen offset:1974
-; ALIGNED-NEXT:    buffer_load_ubyte v82, v2, s[0:3], 0 offen offset:1975
-; ALIGNED-NEXT:    buffer_load_ubyte v81, v2, s[0:3], 0 offen offset:1976
-; ALIGNED-NEXT:    buffer_load_ubyte v80, v2, s[0:3], 0 offen offset:1977
-; ALIGNED-NEXT:    buffer_load_ubyte v71, v2, s[0:3], 0 offen offset:1978
-; ALIGNED-NEXT:    buffer_load_ubyte v70, v2, s[0:3], 0 offen offset:1979
-; ALIGNED-NEXT:    buffer_load_ubyte v69, v2, s[0:3], 0 offen offset:1980
-; ALIGNED-NEXT:    buffer_load_ubyte v67, v2, s[0:3], 0 offen offset:1981
-; ALIGNED-NEXT:    buffer_load_ubyte v68, v2, s[0:3], 0 offen offset:1982
-; ALIGNED-NEXT:    buffer_load_ubyte v66, v2, s[0:3], 0 offen offset:1983
-; ALIGNED-NEXT:    buffer_load_ubyte v65, v2, s[0:3], 0 offen offset:1984
-; ALIGNED-NEXT:    buffer_load_ubyte v54, v2, s[0:3], 0 offen offset:1985
-; ALIGNED-NEXT:    buffer_load_ubyte v53, v2, s[0:3], 0 offen offset:1986
-; ALIGNED-NEXT:    buffer_load_ubyte v52, v2, s[0:3], 0 offen offset:1987
-; ALIGNED-NEXT:    buffer_load_ubyte v55, v2, s[0:3], 0 offen offset:1988
-; ALIGNED-NEXT:    buffer_load_ubyte v49, v2, s[0:3], 0 offen offset:1989
-; ALIGNED-NEXT:    buffer_load_ubyte v50, v2, s[0:3], 0 offen offset:1990
-; ALIGNED-NEXT:    buffer_load_ubyte v48, v2, s[0:3], 0 offen offset:1991
-; ALIGNED-NEXT:    buffer_load_ubyte v39, v2, s[0:3], 0 offen offset:1992
-; ALIGNED-NEXT:    buffer_load_ubyte v37, v2, s[0:3], 0 offen offset:1993
-; ALIGNED-NEXT:    buffer_load_ubyte v35, v2, s[0:3], 0 offen offset:1994
-; ALIGNED-NEXT:    buffer_load_ubyte v36, v2, s[0:3], 0 offen offset:1995
-; ALIGNED-NEXT:    buffer_load_ubyte v34, v2, s[0:3], 0 offen offset:1996
-; ALIGNED-NEXT:    buffer_load_ubyte v31, v2, s[0:3], 0 offen offset:1997
-; ALIGNED-NEXT:    buffer_load_ubyte v32, v2, s[0:3], 0 offen offset:1998
-; ALIGNED-NEXT:    buffer_load_ubyte v30, v2, s[0:3], 0 offen offset:1999
-; ALIGNED-NEXT:    buffer_load_ubyte v29, v2, s[0:3], 0 offen offset:2000
-; ALIGNED-NEXT:    buffer_load_ubyte v27, v2, s[0:3], 0 offen offset:2001
-; ALIGNED-NEXT:    buffer_load_ubyte v28, v2, s[0:3], 0 offen offset:2002
-; ALIGNED-NEXT:    buffer_load_ubyte v26, v2, s[0:3], 0 offen offset:2003
-; ALIGNED-NEXT:    buffer_load_ubyte v24, v2, s[0:3], 0 offen offset:2004
-; ALIGNED-NEXT:    buffer_load_ubyte v19, v2, s[0:3], 0 offen offset:2005
-; ALIGNED-NEXT:    buffer_load_ubyte v20, v2, s[0:3], 0 offen offset:2006
-; ALIGNED-NEXT:    buffer_load_ubyte v17, v2, s[0:3], 0 offen offset:2007
-; ALIGNED-NEXT:    buffer_load_ubyte v23, v2, s[0:3], 0 offen offset:2008
-; ALIGNED-NEXT:    buffer_load_ubyte v21, v2, s[0:3], 0 offen offset:2009
-; ALIGNED-NEXT:    buffer_load_ubyte v18, v2, s[0:3], 0 offen offset:2010
-; ALIGNED-NEXT:    buffer_load_ubyte v16, v2, s[0:3], 0 offen offset:2011
-; ALIGNED-NEXT:    buffer_load_ubyte v22, v2, s[0:3], 0 offen offset:2012
-; ALIGNED-NEXT:    buffer_load_ubyte v105, v2, s[0:3], 0 offen offset:2029
-; ALIGNED-NEXT:    buffer_load_ubyte v106, v2, s[0:3], 0 offen offset:2030
-; ALIGNED-NEXT:    buffer_load_ubyte v94, v2, s[0:3], 0 offen offset:2031
-; ALIGNED-NEXT:    buffer_load_ubyte v92, v2, s[0:3], 0 offen offset:2032
-; ALIGNED-NEXT:    buffer_load_ubyte v90, v2, s[0:3], 0 offen offset:2033
-; ALIGNED-NEXT:    buffer_load_ubyte v88, v2, s[0:3], 0 offen offset:2034
-; ALIGNED-NEXT:    s_clause 0x1b
-; ALIGNED-NEXT:    buffer_load_ubyte v78, v2, s[0:3], 0 offen offset:2035
-; ALIGNED-NEXT:    buffer_load_ubyte v76, v2, s[0:3], 0 offen offset:2036
-; ALIGNED-NEXT:    buffer_load_ubyte v74, v2, s[0:3], 0 offen offset:2037
-; ALIGNED-NEXT:    buffer_load_ubyte v72, v2, s[0:3], 0 offen offset:2038
-; ALIGNED-NEXT:    buffer_load_ubyte v61, v2, s[0:3], 0 offen offset:2039
-; ALIGNED-NEXT:    buffer_load_ubyte v62, v2, s[0:3], 0 offen offset:2040
-; ALIGNED-NEXT:    buffer_load_ubyte v59, v2, s[0:3], 0 offen offset:2041
-; ALIGNED-NEXT:    buffer_load_ubyte v57, v2, s[0:3], 0 offen offset:2042
-; ALIGNED-NEXT:    buffer_load_ubyte v46, v2, s[0:3], 0 offen offset:2043
-; ALIGNED-NEXT:    buffer_load_ubyte v43, v2, s[0:3], 0 offen offset:2044
-; ALIGNED-NEXT:    buffer_load_ubyte v44, v2, s[0:3], 0 offen offset:2045
-; ALIGNED-NEXT:    buffer_load_ubyte v41, v2, s[0:3], 0 offen offset:2046
-; ALIGNED-NEXT:    buffer_load_ubyte v118, v2, s[0:3], 0 offen offset:2047
-; ALIGNED-NEXT:    buffer_load_ubyte v121, v2, s[0:3], 0 offen offset:1793
-; ALIGNED-NEXT:    buffer_load_ubyte v109, v2, s[0:3], 0 offen offset:1794
-; ALIGNED-NEXT:    buffer_load_ubyte v125, v2, s[0:3], 0 offen offset:1795
-; ALIGNED-NEXT:    buffer_load_ubyte v113, v2, s[0:3], 0 offen offset:1796
-; ALIGNED-NEXT:    buffer_load_ubyte v97, v2, s[0:3], 0 offen offset:1797
-; ALIGNED-NEXT:    buffer_load_ubyte v127, v2, s[0:3], 0 offen offset:1798
-; ALIGNED-NEXT:    buffer_load_ubyte v111, v2, s[0:3], 0 offen offset:1799
-; ALIGNED-NEXT:    buffer_load_ubyte v73, v2, s[0:3], 0 offen offset:1804
-; ALIGNED-NEXT:    buffer_load_ubyte v58, v2, s[0:3], 0 offen offset:1805
-; ALIGNED-NEXT:    buffer_load_ubyte v60, v2, s[0:3], 0 offen offset:1806
-; ALIGNED-NEXT:    buffer_load_ubyte v56, v2, s[0:3], 0 offen offset:1807
-; ALIGNED-NEXT:    buffer_load_ubyte v119, v2, s[0:3], 0 offen offset:1803
-; ALIGNED-NEXT:    buffer_load_ubyte v40, v2, s[0:3], 0 offen offset:1800
-; ALIGNED-NEXT:    buffer_load_ubyte v117, v2, s[0:3], 0 offen offset:1801
-; ALIGNED-NEXT:    buffer_load_ubyte v102, v2, s[0:3], 0 offen offset:1802
+; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:133
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:134
+; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:135
+; ALIGNED-NEXT:    buffer_load_ubyte v5, v6, s[0:3], 0 offen offset:136
+; ALIGNED-NEXT:    buffer_load_ubyte v7, v6, s[0:3], 0 offen offset:137
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:138
+; ALIGNED-NEXT:    buffer_load_ubyte v8, v6, s[0:3], 0 offen offset:139
+; ALIGNED-NEXT:    buffer_load_ubyte v9, v6, s[0:3], 0 offen offset:140
+; ALIGNED-NEXT:    buffer_load_ubyte v10, v6, s[0:3], 0 offen offset:144
+; ALIGNED-NEXT:    buffer_load_ubyte v123, v6, s[0:3], 0 offen offset:165
+; ALIGNED-NEXT:    buffer_load_ubyte v122, v6, s[0:3], 0 offen offset:166
+; ALIGNED-NEXT:    buffer_load_ubyte v109, v6, s[0:3], 0 offen offset:167
+; ALIGNED-NEXT:    buffer_load_ubyte v107, v6, s[0:3], 0 offen offset:168
+; ALIGNED-NEXT:    buffer_load_ubyte v106, v6, s[0:3], 0 offen offset:169
+; ALIGNED-NEXT:    buffer_load_ubyte v95, v6, s[0:3], 0 offen offset:170
+; ALIGNED-NEXT:    buffer_load_ubyte v104, v6, s[0:3], 0 offen offset:171
+; ALIGNED-NEXT:    buffer_load_ubyte v94, v6, s[0:3], 0 offen offset:172
+; ALIGNED-NEXT:    buffer_load_ubyte v91, v6, s[0:3], 0 offen offset:173
+; ALIGNED-NEXT:    buffer_load_ubyte v92, v6, s[0:3], 0 offen offset:174
+; ALIGNED-NEXT:    buffer_load_ubyte v88, v6, s[0:3], 0 offen offset:175
+; ALIGNED-NEXT:    buffer_load_ubyte v77, v6, s[0:3], 0 offen offset:176
+; ALIGNED-NEXT:    buffer_load_ubyte v73, v6, s[0:3], 0 offen offset:177
+; ALIGNED-NEXT:    buffer_load_ubyte v63, v6, s[0:3], 0 offen offset:178
+; ALIGNED-NEXT:    buffer_load_ubyte v62, v6, s[0:3], 0 offen offset:179
+; ALIGNED-NEXT:    buffer_load_ubyte v74, v6, s[0:3], 0 offen offset:180
+; ALIGNED-NEXT:    buffer_load_ubyte v47, v6, s[0:3], 0 offen offset:181
+; ALIGNED-NEXT:    buffer_load_ubyte v58, v6, s[0:3], 0 offen offset:182
+; ALIGNED-NEXT:    buffer_load_ubyte v46, v6, s[0:3], 0 offen offset:183
+; ALIGNED-NEXT:    buffer_load_ubyte v45, v6, s[0:3], 0 offen offset:184
+; ALIGNED-NEXT:    buffer_load_ubyte v44, v6, s[0:3], 0 offen offset:185
+; ALIGNED-NEXT:    buffer_load_ubyte v43, v6, s[0:3], 0 offen offset:186
+; ALIGNED-NEXT:    buffer_load_ubyte v42, v6, s[0:3], 0 offen offset:187
+; ALIGNED-NEXT:    buffer_load_ubyte v41, v6, s[0:3], 0 offen offset:188
+; ALIGNED-NEXT:    buffer_load_ubyte v119, v6, s[0:3], 0 offen offset:189
+; ALIGNED-NEXT:    buffer_load_ubyte v40, v6, s[0:3], 0 offen offset:190
+; ALIGNED-NEXT:    buffer_load_ubyte v118, v6, s[0:3], 0 offen offset:191
+; ALIGNED-NEXT:    buffer_load_ubyte v117, v6, s[0:3], 0 offen offset:192
+; ALIGNED-NEXT:    buffer_load_ubyte v114, v6, s[0:3], 0 offen offset:193
+; ALIGNED-NEXT:    buffer_load_ubyte v113, v6, s[0:3], 0 offen offset:194
+; ALIGNED-NEXT:    buffer_load_ubyte v112, v6, s[0:3], 0 offen offset:195
+; ALIGNED-NEXT:    buffer_load_ubyte v115, v6, s[0:3], 0 offen offset:196
+; ALIGNED-NEXT:    buffer_load_ubyte v101, v6, s[0:3], 0 offen offset:197
+; ALIGNED-NEXT:    buffer_load_ubyte v102, v6, s[0:3], 0 offen offset:198
+; ALIGNED-NEXT:    buffer_load_ubyte v100, v6, s[0:3], 0 offen offset:199
+; ALIGNED-NEXT:    buffer_load_ubyte v99, v6, s[0:3], 0 offen offset:200
+; ALIGNED-NEXT:    buffer_load_ubyte v97, v6, s[0:3], 0 offen offset:201
+; ALIGNED-NEXT:    buffer_load_ubyte v87, v6, s[0:3], 0 offen offset:202
+; ALIGNED-NEXT:    buffer_load_ubyte v96, v6, s[0:3], 0 offen offset:203
+; ALIGNED-NEXT:    buffer_load_ubyte v86, v6, s[0:3], 0 offen offset:204
+; ALIGNED-NEXT:    buffer_load_ubyte v83, v6, s[0:3], 0 offen offset:205
+; ALIGNED-NEXT:    buffer_load_ubyte v84, v6, s[0:3], 0 offen offset:206
+; ALIGNED-NEXT:    buffer_load_ubyte v82, v6, s[0:3], 0 offen offset:207
+; ALIGNED-NEXT:    buffer_load_ubyte v81, v6, s[0:3], 0 offen offset:208
+; ALIGNED-NEXT:    buffer_load_ubyte v71, v6, s[0:3], 0 offen offset:209
+; ALIGNED-NEXT:    buffer_load_ubyte v80, v6, s[0:3], 0 offen offset:210
+; ALIGNED-NEXT:    buffer_load_ubyte v70, v6, s[0:3], 0 offen offset:211
+; ALIGNED-NEXT:    buffer_load_ubyte v68, v6, s[0:3], 0 offen offset:212
+; ALIGNED-NEXT:    buffer_load_ubyte v53, v6, s[0:3], 0 offen offset:213
+; ALIGNED-NEXT:    buffer_load_ubyte v64, v6, s[0:3], 0 offen offset:214
+; ALIGNED-NEXT:    buffer_load_ubyte v52, v6, s[0:3], 0 offen offset:215
+; ALIGNED-NEXT:    buffer_load_ubyte v67, v6, s[0:3], 0 offen offset:216
+; ALIGNED-NEXT:    buffer_load_ubyte v65, v6, s[0:3], 0 offen offset:217
+; ALIGNED-NEXT:    buffer_load_ubyte v55, v6, s[0:3], 0 offen offset:218
+; ALIGNED-NEXT:    s_clause 0x2c
+; ALIGNED-NEXT:    buffer_load_ubyte v54, v6, s[0:3], 0 offen offset:219
+; ALIGNED-NEXT:    buffer_load_ubyte v66, v6, s[0:3], 0 offen offset:220
+; ALIGNED-NEXT:    buffer_load_ubyte v49, v6, s[0:3], 0 offen offset:221
+; ALIGNED-NEXT:    buffer_load_ubyte v50, v6, s[0:3], 0 offen offset:222
+; ALIGNED-NEXT:    buffer_load_ubyte v48, v6, s[0:3], 0 offen offset:223
+; ALIGNED-NEXT:    buffer_load_ubyte v39, v6, s[0:3], 0 offen offset:224
+; ALIGNED-NEXT:    buffer_load_ubyte v36, v6, s[0:3], 0 offen offset:225
+; ALIGNED-NEXT:    buffer_load_ubyte v35, v6, s[0:3], 0 offen offset:226
+; ALIGNED-NEXT:    buffer_load_ubyte v34, v6, s[0:3], 0 offen offset:227
+; ALIGNED-NEXT:    buffer_load_ubyte v37, v6, s[0:3], 0 offen offset:228
+; ALIGNED-NEXT:    buffer_load_ubyte v31, v6, s[0:3], 0 offen offset:229
+; ALIGNED-NEXT:    buffer_load_ubyte v32, v6, s[0:3], 0 offen offset:230
+; ALIGNED-NEXT:    buffer_load_ubyte v30, v6, s[0:3], 0 offen offset:231
+; ALIGNED-NEXT:    buffer_load_ubyte v29, v6, s[0:3], 0 offen offset:232
+; ALIGNED-NEXT:    buffer_load_ubyte v28, v6, s[0:3], 0 offen offset:233
+; ALIGNED-NEXT:    buffer_load_ubyte v27, v6, s[0:3], 0 offen offset:234
+; ALIGNED-NEXT:    buffer_load_ubyte v26, v6, s[0:3], 0 offen offset:235
+; ALIGNED-NEXT:    buffer_load_ubyte v25, v6, s[0:3], 0 offen offset:236
+; ALIGNED-NEXT:    buffer_load_ubyte v23, v6, s[0:3], 0 offen offset:237
+; ALIGNED-NEXT:    buffer_load_ubyte v24, v6, s[0:3], 0 offen offset:238
+; ALIGNED-NEXT:    buffer_load_ubyte v22, v6, s[0:3], 0 offen offset:239
+; ALIGNED-NEXT:    buffer_load_ubyte v21, v6, s[0:3], 0 offen offset:240
+; ALIGNED-NEXT:    buffer_load_ubyte v19, v6, s[0:3], 0 offen offset:241
+; ALIGNED-NEXT:    buffer_load_ubyte v17, v6, s[0:3], 0 offen offset:242
+; ALIGNED-NEXT:    buffer_load_ubyte v18, v6, s[0:3], 0 offen offset:243
+; ALIGNED-NEXT:    buffer_load_ubyte v20, v6, s[0:3], 0 offen offset:244
+; ALIGNED-NEXT:    buffer_load_ubyte v126, v6, s[0:3], 0 offen
+; ALIGNED-NEXT:    buffer_load_ubyte v124, v6, s[0:3], 0 offen offset:1
+; ALIGNED-NEXT:    buffer_load_ubyte v111, v6, s[0:3], 0 offen offset:2
+; ALIGNED-NEXT:    buffer_load_ubyte v110, v6, s[0:3], 0 offen offset:3
+; ALIGNED-NEXT:    buffer_load_ubyte v125, v6, s[0:3], 0 offen offset:4
+; ALIGNED-NEXT:    buffer_load_ubyte v121, v6, s[0:3], 0 offen offset:5
+; ALIGNED-NEXT:    buffer_load_ubyte v120, v6, s[0:3], 0 offen offset:6
+; ALIGNED-NEXT:    buffer_load_ubyte v108, v6, s[0:3], 0 offen offset:7
+; ALIGNED-NEXT:    buffer_load_ubyte v93, v6, s[0:3], 0 offen offset:12
+; ALIGNED-NEXT:    buffer_load_ubyte v89, v6, s[0:3], 0 offen offset:13
+; ALIGNED-NEXT:    buffer_load_ubyte v90, v6, s[0:3], 0 offen offset:14
+; ALIGNED-NEXT:    buffer_load_ubyte v79, v6, s[0:3], 0 offen offset:15
+; ALIGNED-NEXT:    buffer_load_ubyte v76, v6, s[0:3], 0 offen offset:11
+; ALIGNED-NEXT:    buffer_load_ubyte v78, v6, s[0:3], 0 offen offset:8
+; ALIGNED-NEXT:    buffer_load_ubyte v75, v6, s[0:3], 0 offen offset:9
+; ALIGNED-NEXT:    buffer_load_ubyte v72, v6, s[0:3], 0 offen offset:10
+; ALIGNED-NEXT:    buffer_load_ubyte v11, v6, s[0:3], 0 offen offset:145
+; ALIGNED-NEXT:    buffer_load_ubyte v12, v6, s[0:3], 0 offen offset:146
+; ALIGNED-NEXT:    buffer_load_ubyte v13, v6, s[0:3], 0 offen offset:147
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(62)
-; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1308 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:1316 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:1312 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1284 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1292 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:1324 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:1328 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1296 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v4, 8, v1
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1942
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v0, 8, v6
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1276 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1280 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v0, 8, v4
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 8, v1
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1284 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_clause 0x1
+; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:143
+; ALIGNED-NEXT:    buffer_load_ubyte v4, v6, s[0:3], 0 offen offset:148
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1308 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:142
+; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1316 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1324 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1320 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1288 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:1948
-; ALIGNED-NEXT:    s_waitcnt vmcnt(25)
-; ALIGNED-NEXT:    v_lshl_or_b32 v86, v61, 8, v72
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v8, 8, v7
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v10, 8, v9
+; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:141
+; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:1292 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:1332 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(19)
+; ALIGNED-NEXT:    v_lshl_or_b32 v57, v110, 8, v111
+; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
+; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:1336 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(5)
+; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:1344 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(4)
+; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:1340 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1304 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1352 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1300 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 8, v1
+; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:151
+; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1296 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v0, 8, v9
+; ALIGNED-NEXT:    buffer_load_ubyte v9, v6, s[0:3], 0 offen offset:156
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v8, 8, v2
 ; ALIGNED-NEXT:    s_clause 0x1
-; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:1944
-; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:1946
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1304 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_load_ubyte v2, v6, s[0:3], 0 offen offset:154
+; ALIGNED-NEXT:    buffer_load_ubyte v8, v6, s[0:3], 0 offen offset:155
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1312 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v7, 8, v5
 ; ALIGNED-NEXT:    s_clause 0x1
-; ALIGNED-NEXT:    buffer_load_ubyte v8, v2, s[0:3], 0 offen offset:1945
-; ALIGNED-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:1947
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v12, 8, v11
-; ALIGNED-NEXT:    buffer_load_ubyte v12, v2, s[0:3], 0 offen offset:1953
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v14, 8, v13
-; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:1332 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1320 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_load_ubyte v5, v6, s[0:3], 0 offen offset:152
+; ALIGNED-NEXT:    buffer_load_ubyte v7, v6, s[0:3], 0 offen offset:153
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v13, 8, v12
 ; ALIGNED-NEXT:    s_clause 0x1
-; ALIGNED-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:1952
-; ALIGNED-NEXT:    buffer_load_ubyte v14, v2, s[0:3], 0 offen offset:1955
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:1943
-; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:1336 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1344 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(25)
-; ALIGNED-NEXT:    v_lshl_or_b32 v87, v118, 8, v41
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1340 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1941
-; ALIGNED-NEXT:    s_waitcnt vmcnt(21)
-; ALIGNED-NEXT:    buffer_store_dword v97, off, s[0:3], s32 offset:1424 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v13, v2, s[0:3], 0 offen offset:1954
+; ALIGNED-NEXT:    buffer_load_ubyte v12, v6, s[0:3], 0 offen offset:162
+; ALIGNED-NEXT:    buffer_load_ubyte v13, v6, s[0:3], 0 offen offset:163
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1328 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v11, 8, v10
+; ALIGNED-NEXT:    s_clause 0x1
+; ALIGNED-NEXT:    buffer_load_ubyte v10, v6, s[0:3], 0 offen offset:160
+; ALIGNED-NEXT:    buffer_load_ubyte v11, v6, s[0:3], 0 offen offset:161
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:150
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1348 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:149
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(11)
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1352 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1364 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(10)
-; ALIGNED-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:1364 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:1372 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(9)
-; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1380 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:1404 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
-; ALIGNED-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:1396 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1400 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
-; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:1388 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1388 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(6)
-; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:1392 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(5)
-; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:1408 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(4)
-; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:1404 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:1412 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1356 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v4, 8, v1
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:1950
-; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v0, 8, v5
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1348 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v5, v2, s[0:3], 0 offen offset:1956
-; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:1416 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    buffer_load_ubyte v4, v2, s[0:3], 0 offen offset:1951
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1360 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:1949
+; ALIGNED-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:1396 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1360 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 8, v1
+; ALIGNED-NEXT:    buffer_load_ubyte v3, v6, s[0:3], 0 offen offset:159
+; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1356 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v0, 8, v4
+; ALIGNED-NEXT:    buffer_load_ubyte v4, v6, s[0:3], 0 offen offset:164
+; ALIGNED-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:1412 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:1416 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:1424 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:158
+; ALIGNED-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:1420 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1368 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_load_ubyte v0, v6, s[0:3], 0 offen offset:157
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1372 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:1384 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:1432 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1432 ; 4-byte Folded Spill
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:1376 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v4, 8, v1
+; ALIGNED-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:1380 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v3, 8, v1
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v82, 8, v84
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v0, 8, v6
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1368 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v8, 8, v7
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v10, 8, v9
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1384 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v12, 8, v11
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v14, 8, v13
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1376 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v0, 8, v9
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v8, 8, v2
+; ALIGNED-NEXT:    v_lshl_or_b32 v2, v83, 8, v86
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1392 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v7, 8, v5
+; ALIGNED-NEXT:    v_lshl_or_b32 v61, v3, 16, v2
+; ALIGNED-NEXT:    v_lshl_or_b32 v2, v97, 8, v99
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v96, 8, v87
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v13, 8, v12
+; ALIGNED-NEXT:    v_lshl_or_b32 v59, v3, 16, v2
+; ALIGNED-NEXT:    v_lshl_or_b32 v2, v53, 8, v68
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1408 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v11, 8, v10
 ; ALIGNED-NEXT:    s_clause 0x7
-; ALIGNED-NEXT:    buffer_load_ubyte v13, v2, s[0:3], 0 offen offset:2013
-; ALIGNED-NEXT:    buffer_load_ubyte v14, v2, s[0:3], 0 offen offset:2014
-; ALIGNED-NEXT:    buffer_load_ubyte v12, v2, s[0:3], 0 offen offset:2015
-; ALIGNED-NEXT:    buffer_load_ubyte v11, v2, s[0:3], 0 offen offset:2016
-; ALIGNED-NEXT:    buffer_load_ubyte v9, v2, s[0:3], 0 offen offset:2017
-; ALIGNED-NEXT:    buffer_load_ubyte v7, v2, s[0:3], 0 offen offset:2018
-; ALIGNED-NEXT:    buffer_load_ubyte v6, v2, s[0:3], 0 offen offset:2019
-; ALIGNED-NEXT:    buffer_load_ubyte v10, v2, s[0:3], 0 offen offset:2020
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1400 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v107, 8, v5
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v93, 8, v95
+; ALIGNED-NEXT:    buffer_load_ubyte v15, v6, s[0:3], 0 offen offset:245
+; ALIGNED-NEXT:    buffer_load_ubyte v16, v6, s[0:3], 0 offen offset:246
+; ALIGNED-NEXT:    buffer_load_ubyte v14, v6, s[0:3], 0 offen offset:247
+; ALIGNED-NEXT:    buffer_load_ubyte v13, v6, s[0:3], 0 offen offset:248
+; ALIGNED-NEXT:    buffer_load_ubyte v12, v6, s[0:3], 0 offen offset:249
+; ALIGNED-NEXT:    buffer_load_ubyte v11, v6, s[0:3], 0 offen offset:250
+; ALIGNED-NEXT:    buffer_load_ubyte v10, v6, s[0:3], 0 offen offset:251
+; ALIGNED-NEXT:    buffer_load_ubyte v9, v6, s[0:3], 0 offen offset:252
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v52, 8, v64
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v109, 8, v122
+; ALIGNED-NEXT:    v_lshl_or_b32 v116, v3, 16, v2
+; ALIGNED-NEXT:    v_lshl_or_b32 v2, v65, 8, v67
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v54, 8, v55
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1428 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v47, 8, v63
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v42, 8, v45
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1444 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v89, 8, v91
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v79, 8, v75
-; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1456 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v100, 8, v115
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v98, 8, v99
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v123, 8, v4
+; ALIGNED-NEXT:    v_lshl_or_b32 v103, v3, 16, v2
+; ALIGNED-NEXT:    v_lshl_or_b32 v2, v49, 8, v66
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v88, 8, v92
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v48, 8, v50
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1440 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v91, 8, v94
+; ALIGNED-NEXT:    v_lshl_or_b32 v98, v3, 16, v2
+; ALIGNED-NEXT:    v_lshl_or_b32 v2, v71, 8, v81
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v70, 8, v80
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v104, 8, v95
+; ALIGNED-NEXT:    v_lshl_or_b32 v85, v3, 16, v2
+; ALIGNED-NEXT:    v_lshl_or_b32 v2, v36, 8, v39
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1452 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v106, 8, v107
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v34, 8, v35
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v62, 8, v63
+; ALIGNED-NEXT:    v_lshl_or_b32 v69, v3, 16, v2
+; ALIGNED-NEXT:    v_lshl_or_b32 v2, v31, 8, v37
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v30, 8, v32
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1460 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v83, 8, v101
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v82, 8, v84
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v73, 8, v77
+; ALIGNED-NEXT:    v_lshl_or_b32 v51, v3, 16, v2
+; ALIGNED-NEXT:    v_lshl_or_b32 v2, v23, 8, v25
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v46, 8, v58
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v22, 8, v24
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1468 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v67, 8, v69
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v66, 8, v68
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v47, 8, v74
+; ALIGNED-NEXT:    v_lshl_or_b32 v38, v3, 16, v2
+; ALIGNED-NEXT:    v_lshl_or_b32 v2, v28, 8, v29
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v26, 8, v27
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v118, 8, v40
+; ALIGNED-NEXT:    v_lshl_or_b32 v33, v3, 16, v2
+; ALIGNED-NEXT:    v_lshl_or_b32 v2, v19, 8, v21
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1472 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v80, 8, v81
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v70, 8, v71
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v119, 8, v41
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v18, 8, v17
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v42, 8, v43
+; ALIGNED-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1476 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v54, 8, v65
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v52, 8, v53
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v44, 8, v45
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v112, 8, v113
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1480 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v0, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v49, 8, v55
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v48, 8, v50
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v114, 8, v117
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; ALIGNED-NEXT:    v_lshl_or_b32 v1, v100, 8, v102
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1484 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    s_clause 0x7
-; ALIGNED-NEXT:    buffer_load_ubyte v0, v2, s[0:3], 0 offen offset:2021
-; ALIGNED-NEXT:    buffer_load_ubyte v1, v2, s[0:3], 0 offen offset:2022
-; ALIGNED-NEXT:    buffer_load_ubyte v126, v2, s[0:3], 0 offen offset:2023
-; ALIGNED-NEXT:    buffer_load_ubyte v124, v2, s[0:3], 0 offen offset:2024
-; ALIGNED-NEXT:    buffer_load_ubyte v122, v2, s[0:3], 0 offen offset:2025
-; ALIGNED-NEXT:    buffer_load_ubyte v120, v2, s[0:3], 0 offen offset:2026
-; ALIGNED-NEXT:    buffer_load_ubyte v110, v2, s[0:3], 0 offen offset:2027
-; ALIGNED-NEXT:    buffer_load_ubyte v108, v2, s[0:3], 0 offen offset:2028
-; ALIGNED-NEXT:    v_lshl_or_b32 v77, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v31, 8, v34
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v30, 8, v32
-; ALIGNED-NEXT:    v_lshl_or_b32 v96, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v37, 8, v39
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v36, 8, v35
-; ALIGNED-NEXT:    v_lshl_or_b32 v85, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v19, 8, v24
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v17, 8, v20
-; ALIGNED-NEXT:    v_lshl_or_b32 v64, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v21, 8, v23
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v16, 8, v18
-; ALIGNED-NEXT:    v_lshl_or_b32 v51, v4, 16, v3
-; ALIGNED-NEXT:    s_waitcnt vmcnt(15)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v13, 8, v22
-; ALIGNED-NEXT:    s_waitcnt vmcnt(13)
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v12, 8, v14
-; ALIGNED-NEXT:    v_lshl_or_b32 v38, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v27, 8, v29
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v26, 8, v28
-; ALIGNED-NEXT:    v_lshl_or_b32 v33, v4, 16, v3
-; ALIGNED-NEXT:    s_waitcnt vmcnt(11)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v9, 8, v11
-; ALIGNED-NEXT:    s_waitcnt vmcnt(9)
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v6, 8, v7
-; ALIGNED-NEXT:    v_lshl_or_b32 v25, v4, 16, v3
-; ALIGNED-NEXT:    s_waitcnt vmcnt(7)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v0, 8, v10
-; ALIGNED-NEXT:    s_waitcnt vmcnt(5)
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v126, 8, v1
-; ALIGNED-NEXT:    v_lshl_or_b32 v15, v4, 16, v3
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v105, 8, v108
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v94, 8, v106
-; ALIGNED-NEXT:    v_lshl_or_b32 v8, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v122, 8, v124
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v110, 8, v120
-; ALIGNED-NEXT:    v_lshl_or_b32 v5, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v3, v90, 8, v92
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v78, 8, v88
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v101, 8, v115
+; ALIGNED-NEXT:    v_lshl_or_b32 v105, v1, 16, v0
+; ALIGNED-NEXT:    s_clause 0x2
+; ALIGNED-NEXT:    buffer_load_ubyte v7, v6, s[0:3], 0 offen offset:253
+; ALIGNED-NEXT:    buffer_load_ubyte v8, v6, s[0:3], 0 offen offset:254
+; ALIGNED-NEXT:    buffer_load_ubyte v1, v6, s[0:3], 0 offen offset:255
+; ALIGNED-NEXT:    s_waitcnt vmcnt(10)
+; ALIGNED-NEXT:    v_lshl_or_b32 v3, v15, 8, v20
+; ALIGNED-NEXT:    s_waitcnt vmcnt(8)
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v14, 8, v16
+; ALIGNED-NEXT:    s_waitcnt vmcnt(4)
+; ALIGNED-NEXT:    v_lshl_or_b32 v56, v10, 8, v11
 ; ALIGNED-NEXT:    v_lshl_or_b32 v3, v4, 16, v3
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v74, 8, v76
-; ALIGNED-NEXT:    v_lshl_or_b32 v4, v86, 16, v4
-; ALIGNED-NEXT:    v_lshl_or_b32 v86, v44, 8, v43
-; ALIGNED-NEXT:    v_lshl_or_b32 v116, v87, 16, v86
-; ALIGNED-NEXT:    v_lshl_or_b32 v86, v59, 8, v62
-; ALIGNED-NEXT:    v_lshl_or_b32 v87, v46, 8, v57
-; ALIGNED-NEXT:    v_lshl_or_b32 v114, v87, 16, v86
-; ALIGNED-NEXT:    buffer_load_ubyte v86, v2, s[0:3], 0 offen offset:1792
-; ALIGNED-NEXT:    v_lshl_or_b32 v87, v125, 8, v109
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    buffer_store_dword v86, off, s[0:3], s32 offset:1420 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v86, v121, 8, v86
-; ALIGNED-NEXT:    v_lshl_or_b32 v86, v87, 16, v86
-; ALIGNED-NEXT:    v_lshl_or_b32 v87, v111, 8, v127
-; ALIGNED-NEXT:    buffer_store_dword v86, off, s[0:3], s32 offset:1436 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v86, v97, 8, v113
-; ALIGNED-NEXT:    v_lshl_or_b32 v86, v87, 16, v86
-; ALIGNED-NEXT:    v_lshl_or_b32 v87, v56, 8, v60
-; ALIGNED-NEXT:    buffer_store_dword v86, off, s[0:3], s32 offset:1440 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v86, v58, 8, v73
-; ALIGNED-NEXT:    v_lshl_or_b32 v86, v87, 16, v86
-; ALIGNED-NEXT:    v_lshl_or_b32 v87, v119, 8, v102
-; ALIGNED-NEXT:    buffer_store_dword v86, off, s[0:3], s32 offset:1448 ; 4-byte Folded Spill
-; ALIGNED-NEXT:    v_lshl_or_b32 v86, v117, 8, v40
-; ALIGNED-NEXT:    v_lshl_or_b32 v86, v87, 16, v86
-; ALIGNED-NEXT:    buffer_store_dword v86, off, s[0:3], s32 offset:1452 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v7, 8, v9
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    v_lshl_or_b32 v5, v1, 8, v8
+; ALIGNED-NEXT:    v_lshl_or_b32 v4, v5, 16, v4
+; ALIGNED-NEXT:    v_lshl_or_b32 v5, v12, 8, v13
+; ALIGNED-NEXT:    v_lshl_or_b32 v5, v56, 16, v5
+; ALIGNED-NEXT:    v_lshl_or_b32 v56, v124, 8, v126
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v57, 16, v56
+; ALIGNED-NEXT:    v_lshl_or_b32 v56, v121, 8, v125
+; ALIGNED-NEXT:    v_lshl_or_b32 v57, v108, 8, v120
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1436 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v57, 16, v56
+; ALIGNED-NEXT:    v_lshl_or_b32 v56, v89, 8, v93
+; ALIGNED-NEXT:    v_lshl_or_b32 v57, v79, 8, v90
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1444 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v57, 16, v56
+; ALIGNED-NEXT:    v_lshl_or_b32 v56, v75, 8, v78
+; ALIGNED-NEXT:    v_lshl_or_b32 v57, v76, 8, v72
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1448 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v57, 16, v56
 ; ALIGNED-NEXT:    s_clause 0x2
-; ALIGNED-NEXT:    buffer_load_ubyte v86, v2, s[0:3], 0 offen offset:1810
-; ALIGNED-NEXT:    buffer_load_ubyte v97, v2, s[0:3], 0 offen offset:1808
-; ALIGNED-NEXT:    buffer_load_ubyte v87, v2, s[0:3], 0 offen offset:1809
-; ALIGNED-NEXT:    buffer_store_dword v114, off, s[0:3], s32 offset:488
-; ALIGNED-NEXT:    buffer_store_dword v116, off, s[0:3], s32 offset:492
-; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:484
-; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:480
-; ALIGNED-NEXT:    v_add_co_u32 v3, vcc_lo, v103, 3
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v4, null, 0, v104, vcc_lo
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v57 offset:247
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v46 offset:248
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v59 offset:246
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v118 offset:252
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v44 offset:250
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v41 offset:251
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v43 offset:249
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v62 offset:245
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v88 offset:239
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v78 offset:240
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v90 offset:238
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v61 offset:244
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v74 offset:242
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v72 offset:243
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v76 offset:241
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v92 offset:237
-; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:504
-; ALIGNED-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:508
-; ALIGNED-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:500
-; ALIGNED-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:496
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v120 offset:231
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v110 offset:232
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v122 offset:230
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v94 offset:236
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v105 offset:234
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v106 offset:235
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v108 offset:233
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v124 offset:229
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v7 offset:223
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v6 offset:224
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v9 offset:222
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v126 offset:228
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:226
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v1 offset:227
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v10 offset:225
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v11 offset:221
-; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:448
-; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:460
-; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:456
-; ALIGNED-NEXT:    buffer_store_dword v64, off, s[0:3], s32 offset:452
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v27 offset:206
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v26 offset:208
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v28 offset:207
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v19 offset:210
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v17 offset:212
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v20 offset:211
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v24 offset:209
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v18 offset:215
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v16 offset:216
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v21 offset:214
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v12 offset:220
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v13 offset:218
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v14 offset:219
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v22 offset:217
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v23 offset:213
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v29 offset:205
-; ALIGNED-NEXT:    buffer_store_dword v85, off, s[0:3], s32 offset:472
-; ALIGNED-NEXT:    buffer_store_dword v96, off, s[0:3], s32 offset:476
-; ALIGNED-NEXT:    buffer_store_dword v77, off, s[0:3], s32 offset:468
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1484 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    v_add_nc_u32_e32 v2, 0xffffff00, v2
-; ALIGNED-NEXT:    s_waitcnt vmcnt(3)
-; ALIGNED-NEXT:    v_lshl_or_b32 v112, v112, 8, v86
+; ALIGNED-NEXT:    buffer_load_ubyte v56, v6, s[0:3], 0 offen offset:18
+; ALIGNED-NEXT:    buffer_load_ubyte v60, v6, s[0:3], 0 offen offset:16
+; ALIGNED-NEXT:    buffer_load_ubyte v57, v6, s[0:3], 0 offen offset:17
+; ALIGNED-NEXT:    v_add_nc_u32_e32 v6, 0xffffff00, v6
+; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:1456 ; 4-byte Folded Spill
+; ALIGNED-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:488
+; ALIGNED-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:492
+; ALIGNED-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:484
+; ALIGNED-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:480
+; ALIGNED-NEXT:    s_clause 0x1 ; 8-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:704
+; ALIGNED-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:708
+; ALIGNED-NEXT:    s_waitcnt vmcnt(4)
+; ALIGNED-NEXT:    v_lshl_or_b32 v0, v127, 8, v56
+; ALIGNED-NEXT:    s_waitcnt vmcnt(2)
+; ALIGNED-NEXT:    v_lshl_or_b32 v127, v57, 8, v60
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(1)
-; ALIGNED-NEXT:    v_lshl_or_b32 v123, v87, 8, v97
-; ALIGNED-NEXT:    v_lshl_or_b32 v112, v112, 16, v123
+; ALIGNED-NEXT:    v_add_co_u32 v2, vcc_lo, v2, s4
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v3, null, s5, v3, vcc_lo
+; ALIGNED-NEXT:    v_lshl_or_b32 v127, v0, 16, v127
+; ALIGNED-NEXT:    v_add_co_u32 v4, vcc_lo, v2, 3
+; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v5, null, 0, v3, vcc_lo
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v11 offset:247
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v10 offset:248
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v12 offset:246
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v1 offset:252
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v7 offset:250
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v8 offset:251
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v9 offset:249
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v13 offset:245
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v17 offset:239
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v18 offset:240
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v19 offset:238
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v14 offset:244
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v15 offset:242
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v16 offset:243
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v20 offset:241
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v21 offset:237
+; ALIGNED-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:504
+; ALIGNED-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:508
+; ALIGNED-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:500
+; ALIGNED-NEXT:    buffer_store_dword v69, off, s[0:3], s32 offset:496
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v27 offset:231
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v26 offset:232
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v28 offset:230
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v22 offset:236
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v23 offset:234
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v24 offset:235
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v25 offset:233
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v29 offset:229
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v35 offset:223
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v34 offset:224
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v36 offset:222
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v30 offset:228
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v31 offset:226
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v32 offset:227
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v37 offset:225
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v39 offset:221
+; ALIGNED-NEXT:    buffer_store_dword v85, off, s[0:3], s32 offset:448
+; ALIGNED-NEXT:    buffer_store_dword v98, off, s[0:3], s32 offset:460
+; ALIGNED-NEXT:    buffer_store_dword v103, off, s[0:3], s32 offset:456
+; ALIGNED-NEXT:    buffer_store_dword v116, off, s[0:3], s32 offset:452
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v53 offset:210
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v52 offset:212
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v71 offset:206
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v70 offset:208
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v80 offset:207
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v64 offset:211
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v68 offset:209
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v55 offset:215
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v54 offset:216
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v65 offset:214
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v48 offset:220
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v49 offset:218
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v50 offset:219
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v66 offset:217
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v67 offset:213
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v81 offset:205
+; ALIGNED-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:472
+; ALIGNED-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:476
+; ALIGNED-NEXT:    buffer_store_dword v105, off, s[0:3], s32 offset:468
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1484 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_add_u32 s4, s4, 0xffffff00
+; ALIGNED-NEXT:    s_addc_u32 s5, s5, -1
+; ALIGNED-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:464
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1480 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v35 offset:199
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v36 offset:200
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v37 offset:198
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v30 offset:204
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v31 offset:202
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v32 offset:203
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v34 offset:201
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v39 offset:197
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v53 offset:191
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v52 offset:192
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v54 offset:190
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v48 offset:196
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v49 offset:194
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v50 offset:195
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v55 offset:193
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v65 offset:189
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v87 offset:199
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v96 offset:200
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v97 offset:198
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v82 offset:204
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v83 offset:202
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v84 offset:203
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v86 offset:201
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v99 offset:197
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v113 offset:191
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v112 offset:192
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v114 offset:190
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v100 offset:196
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v101 offset:194
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v102 offset:195
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v115 offset:193
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v117 offset:189
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1476 ; 4-byte Folded Reload
@@ -15163,628 +15157,622 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1460 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v71 offset:183
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v70 offset:184
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v80 offset:182
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v66 offset:188
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v67 offset:186
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v68 offset:187
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v69 offset:185
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v81 offset:181
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v99 offset:175
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v98 offset:176
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v100 offset:174
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v82 offset:180
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v83 offset:178
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v84 offset:179
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v101 offset:177
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v115 offset:173
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v43 offset:183
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v42 offset:184
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v44 offset:182
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v118 offset:188
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v119 offset:186
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v40 offset:187
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v41 offset:185
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v45 offset:181
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v63 offset:175
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v62 offset:176
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v73 offset:174
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v46 offset:180
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v47 offset:178
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v58 offset:179
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v74 offset:177
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v77 offset:173
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1456 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1452 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1444 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1440 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1428 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:560
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1416 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v75 offset:167
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v79 offset:168
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v89 offset:166
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v42 offset:172
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v47 offset:170
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v45 offset:171
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v63 offset:169
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v91 offset:165
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:159
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1412 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1424 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v95 offset:167
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v104 offset:168
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v106 offset:166
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v88 offset:172
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v91 offset:170
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v92 offset:171
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v94 offset:169
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v107 offset:165
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:159
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1420 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:160
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1408 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:160
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1416 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:158
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v93 offset:164
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v107 offset:162
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v95 offset:163
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:158
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v109 offset:164
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v123 offset:162
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v122 offset:163
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1432 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:161
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1404 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:161
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1412 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:157
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1400 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:157
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1408 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1384 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1392 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:524
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1360 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1368 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1340 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1348 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1396 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1404 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:151
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1392 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:151
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1400 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:152
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1388 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:152
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1396 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:150
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1384 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:150
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:156
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1376 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:156
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1368 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:154
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1380 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:154
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:155
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1372 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:155
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1364 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:153
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1388 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:153
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1380 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:149
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1344 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:149
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1336 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:143
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1340 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:143
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1332 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:144
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1336 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:144
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1328 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:142
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1364 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:142
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:148
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1356 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:148
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1348 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:146
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1360 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:146
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:147
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1352 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:147
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1344 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:145
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1324 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:145
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1332 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:141
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1320 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:141
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1328 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:536
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1304 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1312 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:540
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1280 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1288 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1260 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1268 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:528
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1316 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1324 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:135
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1312 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:135
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1320 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:136
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1308 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:136
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1316 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:134
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1304 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:134
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:140
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1296 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:140
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1288 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:138
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1300 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:138
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:139
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1292 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:139
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1284 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:137
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1308 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:137
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1300 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:133
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1264 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:133
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1256 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:127
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1260 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:127
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1252 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:128
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1256 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:128
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1248 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:126
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1284 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:126
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:132
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1276 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:132
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1268 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:130
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1280 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:130
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:131
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1272 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:131
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1264 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:129
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1244 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:129
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1252 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[103:104], v0 offset:128
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1240 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[2:3], v0 offset:128
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1248 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1224 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1232 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1200 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1208 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1180 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1188 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:608
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1236 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1244 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:119
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1232 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:119
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1240 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:120
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1228 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:120
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1236 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:118
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1224 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:118
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:124
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1216 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:124
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1208 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:122
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1220 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:122
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:123
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1212 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:123
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1204 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:121
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1228 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:121
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1220 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:117
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1184 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:117
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1176 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:111
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1180 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:111
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1172 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:112
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1176 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:112
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1168 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:110
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1204 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:110
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:116
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1196 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:116
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1188 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:114
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1200 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:114
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:115
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1192 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:115
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1184 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:113
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1164 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:113
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1172 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:109
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1160 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:109
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1168 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:632
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1144 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1152 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1120 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1128 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1100 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1108 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1156 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1164 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:103
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1152 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:103
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1160 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:104
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1148 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:104
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1156 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:102
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:102
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1144 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:108
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1136 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:108
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1128 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:106
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1140 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:106
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:107
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1132 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:107
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1124 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:105
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1148 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:105
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1140 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:101
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1104 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:101
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1096 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:95
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1100 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:95
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1092 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:96
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1096 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:96
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1088 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:94
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1124 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:94
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:100
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1116 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:100
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1108 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:98
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1120 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:98
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:99
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1112 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:99
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1104 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:97
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1084 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:97
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1092 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:93
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1080 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:93
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1088 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1064 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1072 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:588
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1052 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1060 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1020 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1028 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1084 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:87
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1080 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:88
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1076 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:87
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1072 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:86
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1064 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:88
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1068 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:92
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1052 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:86
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:90
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1056 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:92
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1044 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:90
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:91
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1048 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:91
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1040 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:89
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1068 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:89
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1060 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:85
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1024 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:85
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1016 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:79
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1020 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:79
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1012 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:80
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1016 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:80
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1008 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:78
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1044 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:78
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:84
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1036 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:84
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1028 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:82
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1040 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:82
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:83
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1032 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:83
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1024 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:81
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1004 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:81
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1012 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:77
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1000 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:77
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1008 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:984 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:992 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:968 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:592
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:996 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1004 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:71
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:992 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:71
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1000 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:72
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:988 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:72
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:996 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:70
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:70
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:984 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:76
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:976 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:76
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:968 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:74
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:980 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:74
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:75
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:972 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:75
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:73
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:988 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:73
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:980 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:69
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:69
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:63
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:63
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:64
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:64
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:62
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:68
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:944 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:62
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:66
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:68
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:67
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:66
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:944 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:65
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:67
+; ALIGNED-NEXT:    flat_store_byte v[2:3], v0 offset:64
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:65
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[103:104], v0 offset:64
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:58
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:55
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:56
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:54
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:58
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:60
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:55
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:59
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:56
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:57
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:54
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:53
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:60
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:50
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:59
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:47
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:57
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:48
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:53
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:46
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:50
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:864 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:52
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:47
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:51
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:48
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:49
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:46
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:45
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:52
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:51
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:49
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:45
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:40
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:39
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:40
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:38
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:37
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:39
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:44
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:43
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:38
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:42
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:37
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:44
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:43
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:41
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:42
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:32
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:41
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:32
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:31
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:31
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:30
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:30
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[103:104], v0 offset:32
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[2:3], v0 offset:32
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:36
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:36
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:35
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:35
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:34
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:34
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:33
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:33
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:644
-; ALIGNED-NEXT:    buffer_store_dword v112, off, s[0:3], s32 offset:640
+; ALIGNED-NEXT:    buffer_store_dword v127, off, s[0:3], s32 offset:640
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:23
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:24
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:23
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:22
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:24
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:22
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:28
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:28
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:26
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:27
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:26
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:25
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:27
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:21
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v56 offset:15
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1464 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:25
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:16
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v57 offset:14
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:21
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v86 offset:15
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1464 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:16
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v87 offset:14
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:20
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:20
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:18
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:18
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:19
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:19
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:17
-; ALIGNED-NEXT:    flat_store_byte v[103:104], v97 offset:16
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1452 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v0 offset:17
+; ALIGNED-NEXT:    flat_store_byte v[2:3], v60 offset:16
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1456 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1448 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1440 ; 4-byte Folded Reload
+; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1444 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660
 ; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1436 ; 4-byte Folded Reload
 ; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; ALIGNED-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1420 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v102 offset:7
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v119 offset:8
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v58 offset:10
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v117 offset:6
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v56 offset:12
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v60 offset:11
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v73 offset:9
-; ALIGNED-NEXT:    flat_store_byte v[103:104], v40 offset:8
-; ALIGNED-NEXT:    flat_store_byte v[103:104], v109 offset:2
-; ALIGNED-NEXT:    flat_store_byte v[103:104], v121 offset:1
-; ALIGNED-NEXT:    flat_store_byte v[103:104], v113 offset:4
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v125
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v111 offset:4
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[103:104], v0
-; ALIGNED-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:1424 ; 4-byte Folded Reload
-; ALIGNED-NEXT:    v_add_co_u32 v103, vcc_lo, 0xffffff00, v103
-; ALIGNED-NEXT:    v_add_co_ci_u32_e64 v104, null, -1, v104, vcc_lo
-; ALIGNED-NEXT:    s_waitcnt vmcnt(0)
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v0 offset:2
-; ALIGNED-NEXT:    flat_store_byte v[3:4], v127 offset:3
-; ALIGNED-NEXT:    s_cbranch_scc0 .LBB9_5
-; ALIGNED-NEXT:  .LBB9_6: ; %Flow17
-; ALIGNED-NEXT:    s_or_b32 exec_lo, exec_lo, s6
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v72 offset:7
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v76 offset:8
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v89 offset:10
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v75 offset:6
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v79 offset:12
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v90 offset:11
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v93 offset:9
+; ALIGNED-NEXT:    flat_store_byte v[2:3], v78 offset:8
+; ALIGNED-NEXT:    flat_store_byte v[2:3], v111 offset:2
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v110
+; ALIGNED-NEXT:    flat_store_byte v[2:3], v124 offset:1
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v108 offset:4
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v121 offset:2
+; ALIGNED-NEXT:    flat_store_byte v[4:5], v120 offset:3
+; ALIGNED-NEXT:    flat_store_byte v[2:3], v125 offset:4
+; ALIGNED-NEXT:    flat_store_byte v[2:3], v126
+; ALIGNED-NEXT:    s_cbranch_scc0 .LBB9_4
+; ALIGNED-NEXT:  .LBB9_5: ; %Flow11
+; ALIGNED-NEXT:    s_or_b32 exec_lo, exec_lo, s8
 ; ALIGNED-NEXT:    s_clause 0x2f ; 192-byte Folded Reload
 ; ALIGNED-NEXT:    buffer_load_dword v127, off, s[0:3], s32
 ; ALIGNED-NEXT:    buffer_load_dword v126, off, s[0:3], s32 offset:4
@@ -15841,44 +15829,42 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; UNROLL3:       ; %bb.0: ; %entry
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; UNROLL3-NEXT:    v_cmp_ne_u64_e32 vcc_lo, 0, v[0:1]
-; UNROLL3-NEXT:    s_mov_b32 s4, exec_lo
+; UNROLL3-NEXT:    s_mov_b64 s[4:5], 0
+; UNROLL3-NEXT:    s_mov_b32 s6, exec_lo
 ; UNROLL3-NEXT:    v_cndmask_b32_e32 v3, -1, v0, vcc_lo
 ; UNROLL3-NEXT:    v_cmpx_ge_u32_e32 v2, v3
-; UNROLL3-NEXT:    s_xor_b32 s6, exec_lo, s4
+; UNROLL3-NEXT:    s_xor_b32 s6, exec_lo, s6
 ; UNROLL3-NEXT:    s_cbranch_execz .LBB9_4
 ; UNROLL3-NEXT:  ; %bb.1: ; %memmove_fwd_loop.preheader
-; UNROLL3-NEXT:    v_mov_b32_e32 v4, v1
-; UNROLL3-NEXT:    v_mov_b32_e32 v3, v0
-; UNROLL3-NEXT:    v_mov_b32_e32 v5, v2
-; UNROLL3-NEXT:    s_mov_b64 s[4:5], 0x7e0
+; UNROLL3-NEXT:    v_mov_b32_e32 v3, v2
 ; UNROLL3-NEXT:    s_inst_prefetch 0x1
 ; UNROLL3-NEXT:    .p2align 6
 ; UNROLL3-NEXT:  .LBB9_2: ; %memmove_fwd_loop
 ; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; UNROLL3-NEXT:    s_clause 0xb
-; UNROLL3-NEXT:    buffer_load_dword v6, v5, s[0:3], 0 offen
-; UNROLL3-NEXT:    buffer_load_dword v7, v5, s[0:3], 0 offen offset:4
-; UNROLL3-NEXT:    buffer_load_dword v8, v5, s[0:3], 0 offen offset:8
-; UNROLL3-NEXT:    buffer_load_dword v9, v5, s[0:3], 0 offen offset:12
-; UNROLL3-NEXT:    buffer_load_dword v10, v5, s[0:3], 0 offen offset:16
-; UNROLL3-NEXT:    buffer_load_dword v11, v5, s[0:3], 0 offen offset:20
-; UNROLL3-NEXT:    buffer_load_dword v12, v5, s[0:3], 0 offen offset:24
-; UNROLL3-NEXT:    buffer_load_dword v13, v5, s[0:3], 0 offen offset:28
-; UNROLL3-NEXT:    buffer_load_dword v14, v5, s[0:3], 0 offen offset:32
-; UNROLL3-NEXT:    buffer_load_dword v15, v5, s[0:3], 0 offen offset:36
-; UNROLL3-NEXT:    buffer_load_dword v16, v5, s[0:3], 0 offen offset:40
-; UNROLL3-NEXT:    buffer_load_dword v17, v5, s[0:3], 0 offen offset:44
-; UNROLL3-NEXT:    v_add_nc_u32_e32 v5, 48, v5
-; UNROLL3-NEXT:    s_add_u32 s4, s4, 0xffffffd0
-; UNROLL3-NEXT:    s_addc_u32 s5, s5, -1
+; UNROLL3-NEXT:    buffer_load_dword v4, v3, s[0:3], 0 offen
+; UNROLL3-NEXT:    buffer_load_dword v5, v3, s[0:3], 0 offen offset:4
+; UNROLL3-NEXT:    buffer_load_dword v6, v3, s[0:3], 0 offen offset:8
+; UNROLL3-NEXT:    buffer_load_dword v7, v3, s[0:3], 0 offen offset:12
+; UNROLL3-NEXT:    buffer_load_dword v8, v3, s[0:3], 0 offen offset:16
+; UNROLL3-NEXT:    buffer_load_dword v9, v3, s[0:3], 0 offen offset:20
+; UNROLL3-NEXT:    buffer_load_dword v10, v3, s[0:3], 0 offen offset:24
+; UNROLL3-NEXT:    buffer_load_dword v11, v3, s[0:3], 0 offen offset:28
+; UNROLL3-NEXT:    buffer_load_dword v12, v3, s[0:3], 0 offen offset:32
+; UNROLL3-NEXT:    buffer_load_dword v13, v3, s[0:3], 0 offen offset:36
+; UNROLL3-NEXT:    buffer_load_dword v14, v3, s[0:3], 0 offen offset:40
+; UNROLL3-NEXT:    buffer_load_dword v15, v3, s[0:3], 0 offen offset:44
+; UNROLL3-NEXT:    v_add_co_u32 v16, vcc_lo, v0, s4
+; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v17, null, s5, v1, vcc_lo
+; UNROLL3-NEXT:    s_add_u32 s4, s4, 48
+; UNROLL3-NEXT:    v_add_nc_u32_e32 v3, 48, v3
+; UNROLL3-NEXT:    s_addc_u32 s5, s5, 0
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(4)
-; UNROLL3-NEXT:    flat_store_dwordx4 v[3:4], v[10:13] offset:16
-; UNROLL3-NEXT:    flat_store_dwordx4 v[3:4], v[6:9]
+; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[8:11] offset:16
+; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[4:7]
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(0)
-; UNROLL3-NEXT:    flat_store_dwordx4 v[3:4], v[14:17] offset:32
-; UNROLL3-NEXT:    v_add_co_u32 v3, vcc_lo, v3, 48
-; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v4, null, 0, v4, vcc_lo
-; UNROLL3-NEXT:    s_cmp_lg_u64 s[4:5], 0
+; UNROLL3-NEXT:    flat_store_dwordx4 v[16:17], v[12:15] offset:32
+; UNROLL3-NEXT:    s_cmp_lg_u64 s[4:5], 0x7e0
 ; UNROLL3-NEXT:    s_cbranch_scc1 .LBB9_2
 ; UNROLL3-NEXT:  ; %bb.3: ; %memmove_fwd_residual
 ; UNROLL3-NEXT:    s_inst_prefetch 0x2
@@ -15898,8 +15884,8 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(0)
 ; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[3:6] offset:2032
 ; UNROLL3-NEXT:    ; implicit-def: $vgpr0_vgpr1
-; UNROLL3-NEXT:  .LBB9_4: ; %Flow14
-; UNROLL3-NEXT:    s_andn2_saveexec_b32 s6, s6
+; UNROLL3-NEXT:  .LBB9_4: ; %Flow8
+; UNROLL3-NEXT:    s_andn2_saveexec_b32 s8, s6
 ; UNROLL3-NEXT:    s_cbranch_execz .LBB9_7
 ; UNROLL3-NEXT:  ; %bb.5: ; %memmove_bwd_residual
 ; UNROLL3-NEXT:    s_clause 0x3
@@ -15907,51 +15893,51 @@ define void @memmove_p0_p5_sz2048(ptr addrspace(0) align 1 %dst, ptr addrspace(5
 ; UNROLL3-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:2036
 ; UNROLL3-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:2040
 ; UNROLL3-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:2044
-; UNROLL3-NEXT:    s_movk_i32 s4, 0xf820
-; UNROLL3-NEXT:    s_mov_b32 s5, -1
+; UNROLL3-NEXT:    s_mov_b64 s[4:5], 0x7b0
+; UNROLL3-NEXT:    s_movk_i32 s6, 0xffd0
+; UNROLL3-NEXT:    s_mov_b32 s7, -1
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(0)
 ; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[3:6] offset:2032
 ; UNROLL3-NEXT:    s_clause 0x3
-; UNROLL3-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:2016
-; UNROLL3-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:2020
-; UNROLL3-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:2024
-; UNROLL3-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:2028
-; UNROLL3-NEXT:    v_add_co_u32 v3, vcc_lo, 0x7b0, v0
-; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v4, null, 0, v1, vcc_lo
+; UNROLL3-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen offset:2016
+; UNROLL3-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:2020
+; UNROLL3-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:2024
+; UNROLL3-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:2028
+; UNROLL3-NEXT:    v_add_nc_u32_e32 v2, 0x7b0, v2
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(0)
-; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[5:8] offset:2016
+; UNROLL3-NEXT:    flat_store_dwordx4 v[0:1], v[3:6] offset:2016
 ; UNROLL3-NEXT:    s_inst_prefetch 0x1
 ; UNROLL3-NEXT:    .p2align 6
 ; UNROLL3-NEXT:  .LBB9_6: ; %memmove_bwd_loop
 ; UNROLL3-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; UNROLL3-NEXT:    s_clause 0xb
-; UNROLL3-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:1968
-; UNROLL3-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:1972
-; UNROLL3-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:1976
-; UNROLL3-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:1980
-; UNROLL3-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:1984
-; UNROLL3-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:1988
-; UNROLL3-NEXT:    buffer_load_dword v11, v2, s[0:3], 0 offen offset:1992
-; UNROLL3-NEXT:    buffer_load_dword v12, v2, s[0:3], 0 offen offset:1996
-; UNROLL3-NEXT:    buffer_load_dword v13, v2, s[0:3], 0 offen offset:2000
-; UNROLL3-NEXT:    buffer_load_dword v14, v2, s[0:3], 0 offen offset:2004
-; UNROLL3-NEXT:    buffer_load_dword v15, v2, s[0:3], 0 offen offset:2008
-; UNROLL3-NEXT:    buffer_load_dword v16, v2, s[0:3], 0 offen offset:2012
+; UNROLL3-NEXT:    buffer_load_dword v3, v2, s[0:3], 0 offen
+; UNROLL3-NEXT:    buffer_load_dword v4, v2, s[0:3], 0 offen offset:4
+; UNROLL3-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen offset:8
+; UNROLL3-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:12
+; UNROLL3-NEXT:    buffer_load_dword v7, v2, s[0:3], 0 offen offset:16
+; UNROLL3-NEXT:    buffer_load_dword v8, v2, s[0:3], 0 offen offset:20
+; UNROLL3-NEXT:    buffer_load_dword v9, v2, s[0:3], 0 offen offset:24
+; UNROLL3-NEXT:    buffer_load_dword v10, v2, s[0:3], 0 offen offset:28
+; UNROLL3-NEXT:    buffer_load_dword v11, v2, s[0:3], 0 offen offset:32
+; UNROLL3-NEXT:    buffer_load_dword v12, v2, s[0:3], 0 offen offset:36
+; UNROLL3-NEXT:    buffer_load_dword v13, v2, s[0:3], 0 offen offset:40
+; UNROLL3-NEXT:    buffer_load_dword v14, v2, s[0:3], 0 offen offset:44
+; UNROLL3-NEXT:    v_add_co_u32 v15, vcc_lo, v0, s4
+; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v16, null, s5, v1, vcc_lo
 ; UNROLL3-NEXT:    v_subrev_nc_u32_e32 v2, 48, v2
-; UNROLL3-NEXT:    s_add_u32 s4, s4, 48
-; UNROLL3-NEXT:    s_addc_u32 s5, s5, 0
+; UNROLL3-NEXT:    s_add_u32 s4, s4, 0xffffffd0
+; UNROLL3-NEXT:    s_addc_u32 s5, s5, -1
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(4)
-; UNROLL3-NEXT:    flat_store_dwordx4 v[3:4], v[9:12] offset:16
-; UNROLL3-NEXT:    flat_store_dwordx4 v[3:4], v[5:8]
+; UNROLL3-NEXT:    flat_store_dwordx4 v[15:16], v[7:10] offset:16
+; UNROLL3-NEXT:    flat_store_dwordx4 v[15:16], v[3:6]
 ; UNROLL3-NEXT:    s_waitcnt vmcnt(0)
-; UNROLL3-NEXT:    flat_store_dwordx4 v[3:4], v[13:16] offset:32
-; UNROLL3-NEXT:    v_add_co_u32 v3, vcc_lo, 0xffffffd0, v3
-; UNROLL3-NEXT:    v_add_co_ci_u32_e64 v4, null, -1, v4, vcc_lo
-; UNROLL3-NEXT:    s_cmp_eq_u64 s[4:5], 0
+; UNROLL3-NEXT:    flat_store_dwordx4 v[15:16], v[11:14] offset:32
+; UNROLL3-NEXT:    s_cmp_eq_u64 s[4:5], s[6:7]
 ; UNROLL3-NEXT:    s_cbranch_scc0 .LBB9_6
-; UNROLL3-NEXT:  .LBB9_7: ; %Flow15
+; UNROLL3-NEXT:  .LBB9_7: ; %Flow9
 ; UNROLL3-NEXT:    s_inst_prefetch 0x2
-; UNROLL3-NEXT:    s_or_b32 exec_lo, exec_lo, s6
+; UNROLL3-NEXT:    s_or_b32 exec_lo, exec_lo, s8
 ; UNROLL3-NEXT:    s_waitcnt lgkmcnt(0)
 ; UNROLL3-NEXT:    s_setpc_b64 s[30:31]
 entry:
diff --git a/llvm/test/CodeGen/AMDGPU/memset-pattern.ll b/llvm/test/CodeGen/AMDGPU/memset-pattern.ll
index d388ae2020da8..488402ac085ca 100644
--- a/llvm/test/CodeGen/AMDGPU/memset-pattern.ll
+++ b/llvm/test/CodeGen/AMDGPU/memset-pattern.ll
@@ -1049,10 +1049,9 @@ define void @memset_pattern_i64_as7_len33_dynval(ptr addrspace(7) inreg align 16
 ; GFX942-GISEL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX942-GISEL-NEXT:    v_mov_b32_e32 v4, v1
 ; GFX942-GISEL-NEXT:    v_mov_b32_e32 v5, v2
-; GFX942-GISEL-NEXT:    v_add_u32_e32 v0, s16, v0
+; GFX942-GISEL-NEXT:    v_add_u32_e32 v1, s16, v0
 ; GFX942-GISEL-NEXT:    s_mov_b64 s[4:5], 0
 ; GFX942-GISEL-NEXT:    v_mov_b64_e32 v[6:7], v[4:5]
-; GFX942-GISEL-NEXT:    v_mov_b32_e32 v1, v0
 ; GFX942-GISEL-NEXT:  .LBB14_1: ; %memset.pattern-expansion-main-body
 ; GFX942-GISEL-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GFX942-GISEL-NEXT:    s_add_u32 s4, s4, 32
@@ -1080,6 +1079,7 @@ define void @memset_pattern_i64_as7_len33_dynval(ptr addrspace(7) inreg align 16
 ; GFX942-GISEL-NEXT:    s_cmp_lg_u32 s6, 0
 ; GFX942-GISEL-NEXT:    s_cbranch_scc1 .LBB14_1
 ; GFX942-GISEL-NEXT:  ; %bb.2: ; %memset.pattern-expansion-residual-body.preheader
+; GFX942-GISEL-NEXT:    v_add_u32_e32 v0, s16, v0
 ; GFX942-GISEL-NEXT:    buffer_store_dwordx2 v[4:5], v0, s[0:3], 0 offen offset:256
 ; GFX942-GISEL-NEXT:    s_waitcnt vmcnt(0)
 ; GFX942-GISEL-NEXT:    s_setpc_b64 s[30:31]
diff --git a/llvm/test/CodeGen/AMDGPU/uniform-cfg.ll b/llvm/test/CodeGen/AMDGPU/uniform-cfg.ll
index d6495648a594f..5108159e7a847 100644
--- a/llvm/test/CodeGen/AMDGPU/uniform-cfg.ll
+++ b/llvm/test/CodeGen/AMDGPU/uniform-cfg.ll
@@ -1140,6 +1140,7 @@ define void @move_to_valu_vgpr_operand_phi(ptr addrspace(3) %out) {
 ; SI-LABEL: move_to_valu_vgpr_operand_phi:
 ; SI:       ; %bb.0: ; %bb0
 ; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SI-NEXT:    v_add_i32_e32 v0, vcc, 28, v0
 ; SI-NEXT:    v_mov_b32_e32 v1, 1
 ; SI-NEXT:    s_and_b64 vcc, exec, 0
 ; SI-NEXT:    s_mov_b32 m0, -1
@@ -1156,8 +1157,7 @@ define void @move_to_valu_vgpr_operand_phi(ptr addrspace(3) %out) {
 ; SI-NEXT:    s_cbranch_scc1 .LBB20_1
 ; SI-NEXT:  ; %bb.3: ; %bb2
 ; SI-NEXT:    ; in Loop: Header=BB20_2 Depth=1
-; SI-NEXT:    v_add_i32_e64 v2, s[4:5], 28, v0
-; SI-NEXT:    ds_write_b32 v2, v1
+; SI-NEXT:    ds_write_b32 v0, v1
 ; SI-NEXT:    s_mov_b64 vcc, vcc
 ; SI-NEXT:    s_cbranch_vccz .LBB20_1
 ; SI-NEXT:  ; %bb.4: ; %DummyReturnBlock
@@ -1167,6 +1167,7 @@ define void @move_to_valu_vgpr_operand_phi(ptr addrspace(3) %out) {
 ; VI-LABEL: move_to_valu_vgpr_operand_phi:
 ; VI:       ; %bb.0: ; %bb0
 ; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_add_u32_e32 v0, vcc, 28, v0
 ; VI-NEXT:    v_mov_b32_e32 v1, 1
 ; VI-NEXT:    s_and_b64 vcc, exec, 0
 ; VI-NEXT:    s_mov_b32 m0, -1
@@ -1183,7 +1184,7 @@ define void @move_to_valu_vgpr_operand_phi(ptr addrspace(3) %out) {
 ; VI-NEXT:    s_cbranch_scc1 .LBB20_1
 ; VI-NEXT:  ; %bb.3: ; %bb2
 ; VI-NEXT:    ; in Loop: Header=BB20_2 Depth=1
-; VI-NEXT:    ds_write_b32 v0, v1 offset:28
+; VI-NEXT:    ds_write_b32 v0, v1
 ; VI-NEXT:    s_mov_b64 vcc, vcc
 ; VI-NEXT:    s_cbranch_vccz .LBB20_1
 ; VI-NEXT:  ; %bb.4: ; %DummyReturnBlock
diff --git a/llvm/test/CodeGen/AMDGPU/valu-i1.ll b/llvm/test/CodeGen/AMDGPU/valu-i1.ll
index 9df052d6910e8..f6203b32269c1 100644
--- a/llvm/test/CodeGen/AMDGPU/valu-i1.ll
+++ b/llvm/test/CodeGen/AMDGPU/valu-i1.ll
@@ -248,28 +248,25 @@ define amdgpu_kernel void @simple_test_v_loop(ptr addrspace(1) %dst, ptr addrspa
 ; SI-NEXT:  ; %bb.1: ; %loop.preheader
 ; SI-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x9
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b64 s[0:1], 0
 ; SI-NEXT:    s_mov_b32 s6, -1
-; SI-NEXT:    s_mov_b32 s3, s7
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    v_mov_b32_e32 v1, s9
 ; SI-NEXT:    v_add_i32_e32 v0, vcc, s8, v0
 ; SI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; SI-NEXT:    s_mov_b32 s8, 64
 ; SI-NEXT:    s_mov_b32 s4, s10
 ; SI-NEXT:    s_mov_b32 s5, s11
-; SI-NEXT:    s_mov_b32 s0, s2
-; SI-NEXT:    s_mov_b32 s1, s2
+; SI-NEXT:    s_mov_b32 s7, s3
 ; SI-NEXT:  .LBB4_2: ; %loop
 ; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; SI-NEXT:    s_waitcnt expcnt(0)
 ; SI-NEXT:    buffer_load_dword v2, off, s[4:7], 0
-; SI-NEXT:    s_add_i32 s8, s8, -1
-; SI-NEXT:    s_cmp_lg_u32 s8, 0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
-; SI-NEXT:    v_add_i32_e32 v0, vcc, 4, v0
-; SI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; SI-NEXT:    s_add_u32 s0, s0, 4
+; SI-NEXT:    s_addc_u32 s1, s1, 0
+; SI-NEXT:    s_cmpk_lg_i32 s0, 0x100
 ; SI-NEXT:    s_cbranch_scc1 .LBB4_2
 ; SI-NEXT:  .LBB4_3: ; %exit
 ; SI-NEXT:    s_endpgm
diff --git a/llvm/test/CodeGen/AMDGPU/waitcnt-loop-ds-prefetch-pattern.ll b/llvm/test/CodeGen/AMDGPU/waitcnt-loop-ds-prefetch-pattern.ll
index 7f98f8cf8de06..f4c8bee32b645 100644
--- a/llvm/test/CodeGen/AMDGPU/waitcnt-loop-ds-prefetch-pattern.ll
+++ b/llvm/test/CodeGen/AMDGPU/waitcnt-loop-ds-prefetch-pattern.ll
@@ -12,39 +12,40 @@ define amdgpu_kernel void @ds_prefetch_pattern(ptr addrspace(3) %lds, ptr addrsp
 ; CHECK-NEXT:    s_load_b32 s1, s[4:5], 0x0 nv
 ; CHECK-NEXT:    s_load_b32 s0, s[4:5], 0x10 nv
 ; CHECK-NEXT:    v_and_b32_e32 v12, 0x3ff, v0
-; CHECK-NEXT:    v_mov_b32_e32 v4, 0
-; CHECK-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; CHECK-NEXT:    v_dual_mov_b32 v5, v4 :: v_dual_mov_b32 v6, v4
-; CHECK-NEXT:    v_mov_b32_e32 v7, v4
 ; CHECK-NEXT:    s_wait_kmcnt 0x0
-; CHECK-NEXT:    v_lshl_add_u32 v13, v12, 8, s1
+; CHECK-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; CHECK-NEXT:    v_lshl_add_u32 v1, v12, 8, s1
+; CHECK-NEXT:    v_mov_b32_e32 v0, 0
 ; CHECK-NEXT:    s_mov_b32 s1, 0
-; CHECK-NEXT:    ds_load_b128 v[8:11], v13
-; CHECK-NEXT:    ds_load_b128 v[0:3], v13 offset:16
+; CHECK-NEXT:    ds_load_b128 v[4:7], v1
+; CHECK-NEXT:    ds_load_b128 v[8:11], v1 offset:16
+; CHECK-NEXT:    v_dual_add_nc_u32 v13, 32, v1 :: v_dual_mov_b32 v1, v0
+; CHECK-NEXT:    v_dual_mov_b32 v2, v0 :: v_dual_mov_b32 v3, v0
 ; CHECK-NEXT:    s_wait_dscnt 0x0
 ; CHECK-NEXT:  .LBB0_1: ; %loop
 ; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; CHECK-NEXT:    s_barrier_signal -1
-; CHECK-NEXT:    s_add_co_i32 s1, s1, 1
 ; CHECK-NEXT:    s_wait_dscnt 0x1
-; CHECK-NEXT:    v_pk_add_f32 v[6:7], v[6:7], v[10:11]
-; CHECK-NEXT:    v_pk_add_f32 v[4:5], v[4:5], v[8:9]
-; CHECK-NEXT:    v_lshl_add_u32 v14, s1, 5, v13
+; CHECK-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; CHECK-NEXT:    v_pk_add_f32 v[2:3], v[2:3], v[6:7]
+; CHECK-NEXT:    v_pk_add_f32 v[0:1], v[0:1], v[4:5]
+; CHECK-NEXT:    s_add_co_i32 s1, s1, 1
 ; CHECK-NEXT:    s_cmp_lt_i32 s1, s0
 ; CHECK-NEXT:    s_wait_dscnt 0x0
-; CHECK-NEXT:    v_pk_add_f32 v[6:7], v[6:7], v[2:3]
-; CHECK-NEXT:    v_pk_add_f32 v[4:5], v[4:5], v[0:1]
+; CHECK-NEXT:    v_pk_add_f32 v[2:3], v[2:3], v[10:11]
+; CHECK-NEXT:    v_pk_add_f32 v[0:1], v[0:1], v[8:9]
 ; CHECK-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; CHECK-NEXT:    v_pk_add_f32 v[6:7], v[6:7], v[6:7]
-; CHECK-NEXT:    v_pk_add_f32 v[4:5], v[4:5], v[4:5]
+; CHECK-NEXT:    v_pk_add_f32 v[2:3], v[2:3], v[2:3]
+; CHECK-NEXT:    v_pk_add_f32 v[0:1], v[0:1], v[0:1]
 ; CHECK-NEXT:    s_barrier_wait -1
-; CHECK-NEXT:    ds_load_b128 v[8:11], v14
-; CHECK-NEXT:    ds_load_b128 v[0:3], v14 offset:16
+; CHECK-NEXT:    ds_load_b128 v[4:7], v13
+; CHECK-NEXT:    ds_load_b128 v[8:11], v13 offset:16
+; CHECK-NEXT:    v_add_nc_u32_e32 v13, 32, v13
 ; CHECK-NEXT:    s_cbranch_scc1 .LBB0_1
 ; CHECK-NEXT:  ; %bb.2: ; %exit
 ; CHECK-NEXT:    s_load_b64 s[0:1], s[4:5], 0x8 nv
 ; CHECK-NEXT:    s_wait_kmcnt 0x0
-; CHECK-NEXT:    global_store_b128 v12, v[4:7], s[0:1] scale_offset
+; CHECK-NEXT:    global_store_b128 v12, v[0:3], s[0:1] scale_offset
 ; CHECK-NEXT:    s_endpgm
 entry:
   %tid = call i32 @llvm.amdgcn.workitem.id.x()
diff --git a/llvm/test/Transforms/LoopStrengthReduce/AMDGPU/atomics.ll b/llvm/test/Transforms/LoopStrengthReduce/AMDGPU/atomics.ll
index 33afe40efb380..f2c00b095495c 100644
--- a/llvm/test/Transforms/LoopStrengthReduce/AMDGPU/atomics.ll
+++ b/llvm/test/Transforms/LoopStrengthReduce/AMDGPU/atomics.ll
@@ -4,15 +4,17 @@
 
 ; OPT-LABEL: @test_local_atomicrmw_addressing_loop_uniform_index_max_offset_i32(
 
+; OPT: .lr.ph.preheader:
+; OPT: %scevgep2 = getelementptr i8, ptr addrspace(3) %arg1, i32 65532
+; OPT: br label %.lr.ph
 ; OPT: .lr.ph:
-; OPT: %lsr.iv2 = phi ptr addrspace(3) [ %scevgep3, %.lr.ph ], [ %arg1, %.lr.ph.preheader ]
+; OPT: %lsr.iv3 = phi ptr addrspace(3) [ %scevgep4, %.lr.ph ], [ %scevgep2, %.lr.ph.preheader ]
 ; OPT: %lsr.iv1 = phi ptr addrspace(3) [ %scevgep, %.lr.ph ], [ %arg0, %.lr.ph.preheader ]
 ; OPT: %lsr.iv = phi i32 [ %lsr.iv.next, %.lr.ph ], [ %n, %.lr.ph.preheader ]
-; OPT: %scevgep4 = getelementptr i8, ptr addrspace(3) %lsr.iv2, i32 65532
-; OPT: %tmp4 = atomicrmw add ptr addrspace(3) %scevgep4, i32 undef seq_cst, align 4
+; OPT: %tmp4 = atomicrmw add ptr addrspace(3) %lsr.iv3, i32 undef seq_cst, align 4
 ; OPT: %tmp7 = atomicrmw add ptr addrspace(3) %lsr.iv1, i32 undef seq_cst, align 4
 ; OPT: %0 = atomicrmw add ptr addrspace(3) %lsr.iv1, i32 %tmp8 seq_cst, align 4
-; OPT: %scevgep3 = getelementptr i8, ptr addrspace(3) %lsr.iv2, i32 4
+; OPT: %scevgep4 = getelementptr i8, ptr addrspace(3) %lsr.iv3, i32 4
 ; OPT: br i1 %exitcond
 define amdgpu_kernel void @test_local_atomicrmw_addressing_loop_uniform_index_max_offset_i32(ptr addrspace(3) noalias nocapture %arg0, ptr addrspace(3) noalias nocapture readonly %arg1, i32 %n) #0 {
 bb:
@@ -44,13 +46,15 @@ bb:
 
 ; OPT-LABEL: test_local_cmpxchg_addressing_loop_uniform_index_max_offset_i32(
 
+; OPT: .lr.ph.preheader:
+; OPT: %scevgep2 = getelementptr i8, ptr addrspace(3)  %arg1, i32 65532
+; OPT: br label %.lr.ph
 ; OPT: .lr.ph:
-; OPT: %lsr.iv2 = phi ptr addrspace(3) [ %scevgep3, %.lr.ph ], [ %arg1, %.lr.ph.preheader ]
+; OPT: %lsr.iv3 = phi ptr addrspace(3) [ %scevgep4, %.lr.ph ], [ %scevgep2, %.lr.ph.preheader ]
 ; OPT: %lsr.iv1 = phi ptr addrspace(3) [ %scevgep, %.lr.ph ], [ %arg0, %.lr.ph.preheader ]
 ; OPT: %lsr.iv = phi i32 [ %lsr.iv.next, %.lr.ph ], [ %n, %.lr.ph.preheader ]
-; OPT: %scevgep4 = getelementptr i8, ptr addrspace(3) %lsr.iv2, i32 65532
-; OPT: %tmp4 = cmpxchg ptr addrspace(3) %scevgep4, i32 undef, i32 undef seq_cst monotonic, align 4
-; OPT: %scevgep3 = getelementptr i8, ptr addrspace(3) %lsr.iv2, i32 4
+; OPT: %tmp4 =  cmpxchg ptr addrspace(3) %lsr.iv3, i32 undef, i32 undef seq_cst monotonic, align 4
+; OPT: %scevgep4 = getelementptr i8, ptr addrspace(3) %lsr.iv3, i32 4
 define amdgpu_kernel void @test_local_cmpxchg_addressing_loop_uniform_index_max_offset_i32(ptr addrspace(3) noalias nocapture %arg0, ptr addrspace(3) noalias nocapture readonly %arg1, i32 %n) #0 {
 bb:
   %tmp = icmp sgt i32 %n, 0
diff --git a/llvm/test/Transforms/LoopStrengthReduce/AMDGPU/different-addrspace-addressing-mode-loops.ll b/llvm/test/Transforms/LoopStrengthReduce/AMDGPU/different-addrspace-addressing-mode-loops.ll
index 124b7c1885649..25ded4bc64c6f 100644
--- a/llvm/test/Transforms/LoopStrengthReduce/AMDGPU/different-addrspace-addressing-mode-loops.ll
+++ b/llvm/test/Transforms/LoopStrengthReduce/AMDGPU/different-addrspace-addressing-mode-loops.ll
@@ -4,11 +4,13 @@
 ; spaces are correctly handled.
 
 ; OPT-LABEL: @test_global_addressing_loop_uniform_index_max_offset_i32(
+; OPT: .lr.ph.preheader:
+; OPT: %scevgep2 = getelementptr i8, ptr addrspace(1) %arg1, i64 4095
+; OPT: br label %.lr.ph
 ; OPT: {{^}}.lr.ph:
-; OPT: %lsr.iv2 = phi ptr addrspace(1) [ %scevgep3, %.lr.ph ], [ %arg1, %.lr.ph.preheader ]
-; OPT: %scevgep4 = getelementptr i8, ptr addrspace(1) %lsr.iv2, i64 4095
-; OPT: load i8, ptr addrspace(1) %scevgep4, align 1
-; OPT: %scevgep3 = getelementptr i8, ptr addrspace(1) %lsr.iv2, i64 1
+; OPT: %lsr.iv3 = phi ptr addrspace(1) [ %scevgep4, %.lr.ph ], [ %scevgep2, %.lr.ph.preheader ]
+; OPT: load i8, ptr addrspace(1) %lsr.iv3, align 1
+; OPT: %scevgep4 = getelementptr i8, ptr addrspace(1) %lsr.iv3, i64 1
 define amdgpu_kernel void @test_global_addressing_loop_uniform_index_max_offset_i32(ptr addrspace(1) noalias nocapture %arg0, ptr addrspace(1) noalias nocapture readonly %arg1, i32 %n) #0 {
 bb:
   %tmp = icmp sgt i32 %n, 0
@@ -78,11 +80,13 @@ bb:
 }
 
 ; OPT-LABEL: @test_local_addressing_loop_uniform_index_max_offset_i32(
-; OPT: {{^}}.lr.ph:
-; OPT: %lsr.iv2 = phi ptr addrspace(3) [ %scevgep3, %.lr.ph ], [ %arg1, %.lr.ph.preheader ]
-; OPT: %scevgep4 = getelementptr i8, ptr addrspace(3) %lsr.iv2, i32 65535
-; OPT: %tmp4 = load i8, ptr addrspace(3) %scevgep4, align 1
-; OPT: %scevgep3 = getelementptr i8, ptr addrspace(3) %lsr.iv2, i32 1
+; OPT: .lr.ph.preheader:
+; OPT: %scevgep2 = getelementptr i8, ptr addrspace(3) %arg1, i32 65535
+; OPT: br label %.lr.ph
+; OPT: {{^}}.lr.ph
+; OPT: %lsr.iv3 = phi ptr addrspace(3) [ %scevgep4, %.lr.ph ], [ %scevgep2, %.lr.ph.preheader ]
+; OPT: %tmp4 = load i8, ptr addrspace(3) %lsr.iv3, align 1
+; OPT: %scevgep4 = getelementptr i8, ptr addrspace(3) %lsr.iv3, i32 1
 define amdgpu_kernel void @test_local_addressing_loop_uniform_index_max_offset_i32(ptr addrspace(1) noalias nocapture %arg0, ptr addrspace(3) noalias nocapture readonly %arg1, i32 %n) #0 {
 bb:
   %tmp = icmp sgt i32 %n, 0



More information about the llvm-commits mailing list