[llvm] 5417232 - AMDGPU: Regenerate test checks

Matt Arsenault via llvm-commits llvm-commits at lists.llvm.org
Sat Nov 13 08:35:39 PST 2021


Author: Matt Arsenault
Date: 2021-11-13T11:35:35-05:00
New Revision: 54172326e09503e2669fd4ab0f0333f52c387bba

URL: https://github.com/llvm/llvm-project/commit/54172326e09503e2669fd4ab0f0333f52c387bba
DIFF: https://github.com/llvm/llvm-project/commit/54172326e09503e2669fd4ab0f0333f52c387bba.diff

LOG: AMDGPU: Regenerate test checks

Regenerate with -NEXT checks to make a future diff clearer.

Added: 
    

Modified: 
    llvm/test/CodeGen/AMDGPU/sgpr-spill.mir

Removed: 
    


################################################################################
diff  --git a/llvm/test/CodeGen/AMDGPU/sgpr-spill.mir b/llvm/test/CodeGen/AMDGPU/sgpr-spill.mir
index 430970d4168bb..6dfcfd035fec7 100644
--- a/llvm/test/CodeGen/AMDGPU/sgpr-spill.mir
+++ b/llvm/test/CodeGen/AMDGPU/sgpr-spill.mir
@@ -446,486 +446,489 @@ body:             |
     ; FLATSCR: $exec = S_MOV_B64 killed $sgpr64_sgpr65
     ; GCN64-MUBUF-LABEL: name: check_spill
     ; GCN64-MUBUF: liveins: $sgpr8, $sgpr4_sgpr5, $sgpr6_sgpr7, $sgpr11
-    ; GCN64-MUBUF: $sgpr33 = S_MOV_B32 0
-    ; GCN64-MUBUF: $sgpr28 = S_MOV_B32 &SCRATCH_RSRC_DWORD0, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
-    ; GCN64-MUBUF: $sgpr29 = S_MOV_B32 &SCRATCH_RSRC_DWORD1, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
-    ; GCN64-MUBUF: $sgpr30 = S_MOV_B32 4294967295, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
-    ; GCN64-MUBUF: $sgpr31 = S_MOV_B32 14680064, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
-    ; GCN64-MUBUF: $sgpr28 = S_ADD_U32 $sgpr28, $sgpr11, implicit-def $scc, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
-    ; GCN64-MUBUF: $sgpr29 = S_ADDC_U32 $sgpr29, 0, implicit-def $scc, implicit $scc, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
-    ; GCN64-MUBUF: renamable $sgpr12 = IMPLICIT_DEF
-    ; GCN64-MUBUF: $sgpr0_sgpr1 = S_MOV_B64 $exec
-    ; GCN64-MUBUF: $exec = S_MOV_B64 1, implicit-def $vgpr0
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 killed $sgpr12, 0, undef $vgpr0
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 4, 0, 0, 0, implicit $exec :: (store (s32) into %stack.0, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
-    ; GCN64-MUBUF: renamable $sgpr12 = IMPLICIT_DEF
-    ; GCN64-MUBUF: $sgpr0_sgpr1 = S_MOV_B64 $exec
-    ; GCN64-MUBUF: $exec = S_MOV_B64 1, implicit-def $vgpr0
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 4, 0, 0, 0, implicit $exec :: (store (s32) into %stack.0, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
-    ; GCN64-MUBUF: renamable $sgpr12_sgpr13 = IMPLICIT_DEF
-    ; GCN64-MUBUF: $sgpr0_sgpr1 = S_MOV_B64 $exec
-    ; GCN64-MUBUF: $exec = S_MOV_B64 3, implicit-def $vgpr0
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit killed $sgpr12_sgpr13
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 8, 0, 0, 0, implicit $exec :: (store (s32) into %stack.1, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
-    ; GCN64-MUBUF: renamable $sgpr12_sgpr13 = IMPLICIT_DEF
-    ; GCN64-MUBUF: $sgpr0_sgpr1 = S_MOV_B64 $exec
-    ; GCN64-MUBUF: $exec = S_MOV_B64 3, implicit-def $vgpr0
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 8, 0, 0, 0, implicit $exec :: (store (s32) into %stack.1, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
-    ; GCN64-MUBUF: renamable $sgpr12_sgpr13_sgpr14 = IMPLICIT_DEF
-    ; GCN64-MUBUF: $sgpr0_sgpr1 = S_MOV_B64 $exec
-    ; GCN64-MUBUF: $exec = S_MOV_B64 7, implicit-def $vgpr0
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 16, 0, 0, 0, implicit $exec :: (store (s32) into %stack.2, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
-    ; GCN64-MUBUF: renamable $sgpr12_sgpr13_sgpr14_sgpr15 = IMPLICIT_DEF
-    ; GCN64-MUBUF: $sgpr0_sgpr1 = S_MOV_B64 $exec
-    ; GCN64-MUBUF: $exec = S_MOV_B64 15, implicit-def $vgpr0
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 28, 0, 0, 0, implicit $exec :: (store (s32) into %stack.3, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
-    ; GCN64-MUBUF: renamable $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16 = IMPLICIT_DEF
-    ; GCN64-MUBUF: $sgpr0_sgpr1 = S_MOV_B64 $exec
-    ; GCN64-MUBUF: $exec = S_MOV_B64 31, implicit-def $vgpr0
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 44, 0, 0, 0, implicit $exec :: (store (s32) into %stack.4, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
-    ; GCN64-MUBUF: renamable $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19 = IMPLICIT_DEF
-    ; GCN64-MUBUF: $sgpr0_sgpr1 = S_MOV_B64 $exec
-    ; GCN64-MUBUF: $exec = S_MOV_B64 255, implicit-def $vgpr0
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr17, 5, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr18, 6, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr19, 7, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 64, 0, 0, 0, implicit $exec :: (store (s32) into %stack.5, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
-    ; GCN64-MUBUF: renamable $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27 = IMPLICIT_DEF
-    ; GCN64-MUBUF: $sgpr0_sgpr1 = S_MOV_B64 $exec
-    ; GCN64-MUBUF: $exec = S_MOV_B64 65535, implicit-def $vgpr0
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr17, 5, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr18, 6, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr19, 7, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr20, 8, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr21, 9, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr22, 10, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr23, 11, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr24, 12, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr25, 13, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr26, 14, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr27, 15, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 96, 0, 0, 0, implicit $exec :: (store (s32) into %stack.6, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
-    ; GCN64-MUBUF: renamable $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95 = IMPLICIT_DEF
-    ; GCN64-MUBUF: $sgpr0_sgpr1 = S_MOV_B64 $exec
-    ; GCN64-MUBUF: $exec = S_MOV_B64 4294967295, implicit-def $vgpr0
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr64, 0, undef $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr65, 1, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr66, 2, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr67, 3, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr68, 4, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr69, 5, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr70, 6, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr71, 7, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr72, 8, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr73, 9, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr74, 10, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr75, 11, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr76, 12, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr77, 13, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr78, 14, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr79, 15, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr80, 16, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr81, 17, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr82, 18, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr83, 19, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr84, 20, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr85, 21, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr86, 22, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr87, 23, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr88, 24, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr89, 25, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr90, 26, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr91, 27, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr92, 28, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr93, 29, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr94, 30, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr95, 31, $vgpr0, implicit killed $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 160, 0, 0, 0, implicit $exec :: (store (s32) into %stack.7, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
-    ; GCN64-MUBUF: renamable $sgpr12 = IMPLICIT_DEF
-    ; GCN64-MUBUF: $sgpr0_sgpr1 = S_MOV_B64 $exec
-    ; GCN64-MUBUF: $exec = S_MOV_B64 1, implicit-def $vgpr0
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0
-    ; GCN64-MUBUF: $sgpr2 = S_ADD_I32 $sgpr33, 262144, implicit-def $scc
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, killed $sgpr2, 0, 0, 0, 0, implicit $exec :: (store (s32) into %stack.8, align 4096, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
+    ; GCN64-MUBUF-NEXT: {{  $}}
+    ; GCN64-MUBUF-NEXT: $sgpr33 = S_MOV_B32 0
+    ; GCN64-MUBUF-NEXT: $sgpr28 = S_MOV_B32 &SCRATCH_RSRC_DWORD0, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
+    ; GCN64-MUBUF-NEXT: $sgpr29 = S_MOV_B32 &SCRATCH_RSRC_DWORD1, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
+    ; GCN64-MUBUF-NEXT: $sgpr30 = S_MOV_B32 4294967295, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
+    ; GCN64-MUBUF-NEXT: $sgpr31 = S_MOV_B32 14680064, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
+    ; GCN64-MUBUF-NEXT: $sgpr28 = S_ADD_U32 $sgpr28, $sgpr11, implicit-def $scc, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
+    ; GCN64-MUBUF-NEXT: $sgpr29 = S_ADDC_U32 $sgpr29, 0, implicit-def $scc, implicit $scc, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
+    ; GCN64-MUBUF-NEXT: renamable $sgpr12 = IMPLICIT_DEF
+    ; GCN64-MUBUF-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 1, implicit-def $vgpr0
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr12, 0, undef $vgpr0
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 4, 0, 0, 0, implicit $exec :: (store (s32) into %stack.0, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
+    ; GCN64-MUBUF-NEXT: renamable $sgpr12 = IMPLICIT_DEF
+    ; GCN64-MUBUF-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 1, implicit-def $vgpr0
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 4, 0, 0, 0, implicit $exec :: (store (s32) into %stack.0, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
+    ; GCN64-MUBUF-NEXT: renamable $sgpr12_sgpr13 = IMPLICIT_DEF
+    ; GCN64-MUBUF-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 3, implicit-def $vgpr0
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit killed $sgpr12_sgpr13
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 8, 0, 0, 0, implicit $exec :: (store (s32) into %stack.1, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
+    ; GCN64-MUBUF-NEXT: renamable $sgpr12_sgpr13 = IMPLICIT_DEF
+    ; GCN64-MUBUF-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 3, implicit-def $vgpr0
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 8, 0, 0, 0, implicit $exec :: (store (s32) into %stack.1, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
+    ; GCN64-MUBUF-NEXT: renamable $sgpr12_sgpr13_sgpr14 = IMPLICIT_DEF
+    ; GCN64-MUBUF-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 7, implicit-def $vgpr0
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 16, 0, 0, 0, implicit $exec :: (store (s32) into %stack.2, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
+    ; GCN64-MUBUF-NEXT: renamable $sgpr12_sgpr13_sgpr14_sgpr15 = IMPLICIT_DEF
+    ; GCN64-MUBUF-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 15, implicit-def $vgpr0
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 28, 0, 0, 0, implicit $exec :: (store (s32) into %stack.3, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
+    ; GCN64-MUBUF-NEXT: renamable $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16 = IMPLICIT_DEF
+    ; GCN64-MUBUF-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 31, implicit-def $vgpr0
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 44, 0, 0, 0, implicit $exec :: (store (s32) into %stack.4, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
+    ; GCN64-MUBUF-NEXT: renamable $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19 = IMPLICIT_DEF
+    ; GCN64-MUBUF-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 255, implicit-def $vgpr0
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr17, 5, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr18, 6, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr19, 7, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 64, 0, 0, 0, implicit $exec :: (store (s32) into %stack.5, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
+    ; GCN64-MUBUF-NEXT: renamable $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27 = IMPLICIT_DEF
+    ; GCN64-MUBUF-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 65535, implicit-def $vgpr0
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr17, 5, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr18, 6, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr19, 7, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr20, 8, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr21, 9, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr22, 10, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr23, 11, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr24, 12, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr25, 13, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr26, 14, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr27, 15, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 96, 0, 0, 0, implicit $exec :: (store (s32) into %stack.6, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
+    ; GCN64-MUBUF-NEXT: renamable $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95 = IMPLICIT_DEF
+    ; GCN64-MUBUF-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 4294967295, implicit-def $vgpr0
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr64, 0, undef $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr65, 1, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr66, 2, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr67, 3, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr68, 4, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr69, 5, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr70, 6, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr71, 7, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr72, 8, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr73, 9, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr74, 10, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr75, 11, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr76, 12, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr77, 13, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr78, 14, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr79, 15, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr80, 16, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr81, 17, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr82, 18, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr83, 19, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr84, 20, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr85, 21, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr86, 22, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr87, 23, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr88, 24, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr89, 25, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr90, 26, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr91, 27, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr92, 28, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr93, 29, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr94, 30, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr95, 31, $vgpr0, implicit killed $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 160, 0, 0, 0, implicit $exec :: (store (s32) into %stack.7, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
+    ; GCN64-MUBUF-NEXT: renamable $sgpr12 = IMPLICIT_DEF
+    ; GCN64-MUBUF-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 1, implicit-def $vgpr0
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0
+    ; GCN64-MUBUF-NEXT: $sgpr2 = S_ADD_I32 $sgpr33, 262144, implicit-def $scc
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, killed $sgpr2, 0, 0, 0, 0, implicit $exec :: (store (s32) into %stack.8, align 4096, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
     ; GCN32-MUBUF-LABEL: name: check_spill
     ; GCN32-MUBUF: liveins: $sgpr8, $sgpr4_sgpr5, $sgpr6_sgpr7, $sgpr11
-    ; GCN32-MUBUF: $sgpr33 = S_MOV_B32 0
-    ; GCN32-MUBUF: $sgpr96 = S_MOV_B32 &SCRATCH_RSRC_DWORD0, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
-    ; GCN32-MUBUF: $sgpr97 = S_MOV_B32 &SCRATCH_RSRC_DWORD1, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
-    ; GCN32-MUBUF: $sgpr98 = S_MOV_B32 4294967295, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
-    ; GCN32-MUBUF: $sgpr99 = S_MOV_B32 834756608, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
-    ; GCN32-MUBUF: $sgpr96 = S_ADD_U32 $sgpr96, $sgpr11, implicit-def $scc, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
-    ; GCN32-MUBUF: $sgpr97 = S_ADDC_U32 $sgpr97, 0, implicit-def $scc, implicit $scc, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
-    ; GCN32-MUBUF: renamable $sgpr12 = IMPLICIT_DEF
-    ; GCN32-MUBUF: $sgpr0 = S_MOV_B32 $exec_lo
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 1, implicit-def $vgpr0
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 killed $sgpr12, 0, undef $vgpr0
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 4, 0, 0, 0, implicit $exec :: (store (s32) into %stack.0, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
-    ; GCN32-MUBUF: renamable $sgpr12 = IMPLICIT_DEF
-    ; GCN32-MUBUF: $sgpr0 = S_MOV_B32 $exec_lo
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 1, implicit-def $vgpr0
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 4, 0, 0, 0, implicit $exec :: (store (s32) into %stack.0, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
-    ; GCN32-MUBUF: renamable $sgpr12_sgpr13 = IMPLICIT_DEF
-    ; GCN32-MUBUF: $sgpr0 = S_MOV_B32 $exec_lo
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 3, implicit-def $vgpr0
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit killed $sgpr12_sgpr13
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 8, 0, 0, 0, implicit $exec :: (store (s32) into %stack.1, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
-    ; GCN32-MUBUF: renamable $sgpr12_sgpr13 = IMPLICIT_DEF
-    ; GCN32-MUBUF: $sgpr0 = S_MOV_B32 $exec_lo
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 3, implicit-def $vgpr0
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 8, 0, 0, 0, implicit $exec :: (store (s32) into %stack.1, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
-    ; GCN32-MUBUF: renamable $sgpr12_sgpr13_sgpr14 = IMPLICIT_DEF
-    ; GCN32-MUBUF: $sgpr0 = S_MOV_B32 $exec_lo
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 7, implicit-def $vgpr0
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 16, 0, 0, 0, implicit $exec :: (store (s32) into %stack.2, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
-    ; GCN32-MUBUF: renamable $sgpr12_sgpr13_sgpr14_sgpr15 = IMPLICIT_DEF
-    ; GCN32-MUBUF: $sgpr0 = S_MOV_B32 $exec_lo
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 15, implicit-def $vgpr0
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 28, 0, 0, 0, implicit $exec :: (store (s32) into %stack.3, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
-    ; GCN32-MUBUF: renamable $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16 = IMPLICIT_DEF
-    ; GCN32-MUBUF: $sgpr0 = S_MOV_B32 $exec_lo
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 31, implicit-def $vgpr0
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 44, 0, 0, 0, implicit $exec :: (store (s32) into %stack.4, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
-    ; GCN32-MUBUF: renamable $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19 = IMPLICIT_DEF
-    ; GCN32-MUBUF: $sgpr0 = S_MOV_B32 $exec_lo
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 255, implicit-def $vgpr0
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr17, 5, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr18, 6, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr19, 7, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 64, 0, 0, 0, implicit $exec :: (store (s32) into %stack.5, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
-    ; GCN32-MUBUF: renamable $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27 = IMPLICIT_DEF
-    ; GCN32-MUBUF: $sgpr0 = S_MOV_B32 $exec_lo
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 65535, implicit-def $vgpr0
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr17, 5, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr18, 6, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr19, 7, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr20, 8, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr21, 9, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr22, 10, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr23, 11, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr24, 12, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr25, 13, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr26, 14, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr27, 15, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 96, 0, 0, 0, implicit $exec :: (store (s32) into %stack.6, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
-    ; GCN32-MUBUF: renamable $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95 = IMPLICIT_DEF
-    ; GCN32-MUBUF: $sgpr0 = S_MOV_B32 $exec_lo
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 4294967295, implicit-def $vgpr0
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr64, 0, undef $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr65, 1, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr66, 2, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr67, 3, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr68, 4, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr69, 5, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr70, 6, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr71, 7, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr72, 8, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr73, 9, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr74, 10, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr75, 11, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr76, 12, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr77, 13, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr78, 14, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr79, 15, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr80, 16, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr81, 17, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr82, 18, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr83, 19, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr84, 20, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr85, 21, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr86, 22, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr87, 23, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr88, 24, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr89, 25, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr90, 26, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr91, 27, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr92, 28, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr93, 29, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr94, 30, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr95, 31, $vgpr0, implicit killed $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 160, 0, 0, 0, implicit $exec :: (store (s32) into %stack.7, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
-    ; GCN32-MUBUF: renamable $sgpr12 = IMPLICIT_DEF
-    ; GCN32-MUBUF: $sgpr0 = S_MOV_B32 $exec_lo
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 1, implicit-def $vgpr0
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0
-    ; GCN32-MUBUF: $sgpr1 = S_ADD_I32 $sgpr33, 131072, implicit-def $scc
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, killed $sgpr1, 0, 0, 0, 0, implicit $exec :: (store (s32) into %stack.8, align 4096, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
+    ; GCN32-MUBUF-NEXT: {{  $}}
+    ; GCN32-MUBUF-NEXT: $sgpr33 = S_MOV_B32 0
+    ; GCN32-MUBUF-NEXT: $sgpr96 = S_MOV_B32 &SCRATCH_RSRC_DWORD0, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
+    ; GCN32-MUBUF-NEXT: $sgpr97 = S_MOV_B32 &SCRATCH_RSRC_DWORD1, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
+    ; GCN32-MUBUF-NEXT: $sgpr98 = S_MOV_B32 4294967295, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
+    ; GCN32-MUBUF-NEXT: $sgpr99 = S_MOV_B32 834756608, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
+    ; GCN32-MUBUF-NEXT: $sgpr96 = S_ADD_U32 $sgpr96, $sgpr11, implicit-def $scc, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
+    ; GCN32-MUBUF-NEXT: $sgpr97 = S_ADDC_U32 $sgpr97, 0, implicit-def $scc, implicit $scc, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
+    ; GCN32-MUBUF-NEXT: renamable $sgpr12 = IMPLICIT_DEF
+    ; GCN32-MUBUF-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 1, implicit-def $vgpr0
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr12, 0, undef $vgpr0
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 4, 0, 0, 0, implicit $exec :: (store (s32) into %stack.0, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
+    ; GCN32-MUBUF-NEXT: renamable $sgpr12 = IMPLICIT_DEF
+    ; GCN32-MUBUF-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 1, implicit-def $vgpr0
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 4, 0, 0, 0, implicit $exec :: (store (s32) into %stack.0, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
+    ; GCN32-MUBUF-NEXT: renamable $sgpr12_sgpr13 = IMPLICIT_DEF
+    ; GCN32-MUBUF-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 3, implicit-def $vgpr0
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit killed $sgpr12_sgpr13
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 8, 0, 0, 0, implicit $exec :: (store (s32) into %stack.1, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
+    ; GCN32-MUBUF-NEXT: renamable $sgpr12_sgpr13 = IMPLICIT_DEF
+    ; GCN32-MUBUF-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 3, implicit-def $vgpr0
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 8, 0, 0, 0, implicit $exec :: (store (s32) into %stack.1, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
+    ; GCN32-MUBUF-NEXT: renamable $sgpr12_sgpr13_sgpr14 = IMPLICIT_DEF
+    ; GCN32-MUBUF-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 7, implicit-def $vgpr0
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 16, 0, 0, 0, implicit $exec :: (store (s32) into %stack.2, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
+    ; GCN32-MUBUF-NEXT: renamable $sgpr12_sgpr13_sgpr14_sgpr15 = IMPLICIT_DEF
+    ; GCN32-MUBUF-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 15, implicit-def $vgpr0
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 28, 0, 0, 0, implicit $exec :: (store (s32) into %stack.3, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
+    ; GCN32-MUBUF-NEXT: renamable $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16 = IMPLICIT_DEF
+    ; GCN32-MUBUF-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 31, implicit-def $vgpr0
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 44, 0, 0, 0, implicit $exec :: (store (s32) into %stack.4, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
+    ; GCN32-MUBUF-NEXT: renamable $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19 = IMPLICIT_DEF
+    ; GCN32-MUBUF-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 255, implicit-def $vgpr0
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr17, 5, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr18, 6, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr19, 7, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 64, 0, 0, 0, implicit $exec :: (store (s32) into %stack.5, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
+    ; GCN32-MUBUF-NEXT: renamable $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27 = IMPLICIT_DEF
+    ; GCN32-MUBUF-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 65535, implicit-def $vgpr0
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr17, 5, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr18, 6, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr19, 7, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr20, 8, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr21, 9, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr22, 10, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr23, 11, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr24, 12, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr25, 13, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr26, 14, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr27, 15, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 96, 0, 0, 0, implicit $exec :: (store (s32) into %stack.6, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
+    ; GCN32-MUBUF-NEXT: renamable $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95 = IMPLICIT_DEF
+    ; GCN32-MUBUF-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 4294967295, implicit-def $vgpr0
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr64, 0, undef $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr65, 1, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr66, 2, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr67, 3, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr68, 4, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr69, 5, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr70, 6, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr71, 7, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr72, 8, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr73, 9, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr74, 10, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr75, 11, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr76, 12, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr77, 13, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr78, 14, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr79, 15, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr80, 16, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr81, 17, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr82, 18, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr83, 19, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr84, 20, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr85, 21, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr86, 22, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr87, 23, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr88, 24, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr89, 25, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr90, 26, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr91, 27, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr92, 28, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr93, 29, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr94, 30, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr95, 31, $vgpr0, implicit killed $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 160, 0, 0, 0, implicit $exec :: (store (s32) into %stack.7, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
+    ; GCN32-MUBUF-NEXT: renamable $sgpr12 = IMPLICIT_DEF
+    ; GCN32-MUBUF-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 1, implicit-def $vgpr0
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0
+    ; GCN32-MUBUF-NEXT: $sgpr1 = S_ADD_I32 $sgpr33, 131072, implicit-def $scc
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, killed $sgpr1, 0, 0, 0, 0, implicit $exec :: (store (s32) into %stack.8, align 4096, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
     ; GCN64-FLATSCR-LABEL: name: check_spill
     ; GCN64-FLATSCR: liveins: $sgpr8, $sgpr4_sgpr5, $sgpr6_sgpr7, $sgpr11, $sgpr0_sgpr1
-    ; GCN64-FLATSCR: $sgpr33 = S_MOV_B32 0
-    ; GCN64-FLATSCR: $flat_scr_lo = S_ADD_U32 $sgpr0, $sgpr11, implicit-def $scc
-    ; GCN64-FLATSCR: $flat_scr_hi = S_ADDC_U32 $sgpr1, 0, implicit-def $scc, implicit $scc
-    ; GCN64-FLATSCR: renamable $sgpr12 = IMPLICIT_DEF
-    ; GCN64-FLATSCR: $sgpr2_sgpr3 = S_MOV_B64 $exec
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 1, implicit-def $vgpr0
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 killed $sgpr12, 0, undef $vgpr0
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
-    ; GCN64-FLATSCR: renamable $sgpr12 = IMPLICIT_DEF
-    ; GCN64-FLATSCR: $sgpr2_sgpr3 = S_MOV_B64 $exec
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 1, implicit-def $vgpr0
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
-    ; GCN64-FLATSCR: renamable $sgpr12_sgpr13 = IMPLICIT_DEF
-    ; GCN64-FLATSCR: $sgpr2_sgpr3 = S_MOV_B64 $exec
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 3, implicit-def $vgpr0
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit killed $sgpr12_sgpr13
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.1, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
-    ; GCN64-FLATSCR: renamable $sgpr12_sgpr13 = IMPLICIT_DEF
-    ; GCN64-FLATSCR: $sgpr2_sgpr3 = S_MOV_B64 $exec
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 3, implicit-def $vgpr0
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.1, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
-    ; GCN64-FLATSCR: renamable $sgpr12_sgpr13_sgpr14 = IMPLICIT_DEF
-    ; GCN64-FLATSCR: $sgpr2_sgpr3 = S_MOV_B64 $exec
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 7, implicit-def $vgpr0
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 16, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.2, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
-    ; GCN64-FLATSCR: renamable $sgpr12_sgpr13_sgpr14_sgpr15 = IMPLICIT_DEF
-    ; GCN64-FLATSCR: $sgpr2_sgpr3 = S_MOV_B64 $exec
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 15, implicit-def $vgpr0
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 28, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.3, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
-    ; GCN64-FLATSCR: renamable $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16 = IMPLICIT_DEF
-    ; GCN64-FLATSCR: $sgpr2_sgpr3 = S_MOV_B64 $exec
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 31, implicit-def $vgpr0
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 44, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.4, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
-    ; GCN64-FLATSCR: renamable $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19 = IMPLICIT_DEF
-    ; GCN64-FLATSCR: $sgpr2_sgpr3 = S_MOV_B64 $exec
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 255, implicit-def $vgpr0
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr17, 5, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr18, 6, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr19, 7, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 64, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.5, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
-    ; GCN64-FLATSCR: renamable $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27 = IMPLICIT_DEF
-    ; GCN64-FLATSCR: $sgpr2_sgpr3 = S_MOV_B64 $exec
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 65535, implicit-def $vgpr0
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr17, 5, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr18, 6, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr19, 7, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr20, 8, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr21, 9, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr22, 10, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr23, 11, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr24, 12, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr25, 13, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr26, 14, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr27, 15, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 96, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.6, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
-    ; GCN64-FLATSCR: renamable $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95 = IMPLICIT_DEF
-    ; GCN64-FLATSCR: $sgpr2_sgpr3 = S_MOV_B64 $exec
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 4294967295, implicit-def $vgpr0
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr64, 0, undef $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr65, 1, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr66, 2, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr67, 3, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr68, 4, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr69, 5, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr70, 6, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr71, 7, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr72, 8, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr73, 9, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr74, 10, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr75, 11, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr76, 12, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr77, 13, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr78, 14, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr79, 15, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr80, 16, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr81, 17, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr82, 18, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr83, 19, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr84, 20, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr85, 21, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr86, 22, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr87, 23, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr88, 24, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr89, 25, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr90, 26, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr91, 27, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr92, 28, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr93, 29, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr94, 30, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr95, 31, $vgpr0, implicit killed $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 160, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.7, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
-    ; GCN64-FLATSCR: renamable $sgpr12 = IMPLICIT_DEF
-    ; GCN64-FLATSCR: $sgpr2_sgpr3 = S_MOV_B64 $exec
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 1, implicit-def $vgpr0
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0
-    ; GCN64-FLATSCR: $sgpr9 = S_ADD_I32 $sgpr33, 4096, implicit-def $scc
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, killed $sgpr9, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.8, align 4096, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
+    ; GCN64-FLATSCR-NEXT: {{  $}}
+    ; GCN64-FLATSCR-NEXT: $sgpr33 = S_MOV_B32 0
+    ; GCN64-FLATSCR-NEXT: $flat_scr_lo = S_ADD_U32 $sgpr0, $sgpr11, implicit-def $scc
+    ; GCN64-FLATSCR-NEXT: $flat_scr_hi = S_ADDC_U32 $sgpr1, 0, implicit-def $scc, implicit $scc
+    ; GCN64-FLATSCR-NEXT: renamable $sgpr12 = IMPLICIT_DEF
+    ; GCN64-FLATSCR-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 1, implicit-def $vgpr0
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr12, 0, undef $vgpr0
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
+    ; GCN64-FLATSCR-NEXT: renamable $sgpr12 = IMPLICIT_DEF
+    ; GCN64-FLATSCR-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 1, implicit-def $vgpr0
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
+    ; GCN64-FLATSCR-NEXT: renamable $sgpr12_sgpr13 = IMPLICIT_DEF
+    ; GCN64-FLATSCR-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 3, implicit-def $vgpr0
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit killed $sgpr12_sgpr13
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.1, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
+    ; GCN64-FLATSCR-NEXT: renamable $sgpr12_sgpr13 = IMPLICIT_DEF
+    ; GCN64-FLATSCR-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 3, implicit-def $vgpr0
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.1, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
+    ; GCN64-FLATSCR-NEXT: renamable $sgpr12_sgpr13_sgpr14 = IMPLICIT_DEF
+    ; GCN64-FLATSCR-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 7, implicit-def $vgpr0
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 16, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.2, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
+    ; GCN64-FLATSCR-NEXT: renamable $sgpr12_sgpr13_sgpr14_sgpr15 = IMPLICIT_DEF
+    ; GCN64-FLATSCR-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 15, implicit-def $vgpr0
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 28, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.3, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
+    ; GCN64-FLATSCR-NEXT: renamable $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16 = IMPLICIT_DEF
+    ; GCN64-FLATSCR-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 31, implicit-def $vgpr0
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 44, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.4, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
+    ; GCN64-FLATSCR-NEXT: renamable $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19 = IMPLICIT_DEF
+    ; GCN64-FLATSCR-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 255, implicit-def $vgpr0
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr17, 5, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr18, 6, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr19, 7, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 64, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.5, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
+    ; GCN64-FLATSCR-NEXT: renamable $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27 = IMPLICIT_DEF
+    ; GCN64-FLATSCR-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 65535, implicit-def $vgpr0
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr17, 5, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr18, 6, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr19, 7, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr20, 8, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr21, 9, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr22, 10, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr23, 11, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr24, 12, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr25, 13, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr26, 14, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr27, 15, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 96, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.6, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
+    ; GCN64-FLATSCR-NEXT: renamable $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95 = IMPLICIT_DEF
+    ; GCN64-FLATSCR-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 4294967295, implicit-def $vgpr0
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr64, 0, undef $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr65, 1, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr66, 2, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr67, 3, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr68, 4, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr69, 5, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr70, 6, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr71, 7, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr72, 8, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr73, 9, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr74, 10, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr75, 11, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr76, 12, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr77, 13, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr78, 14, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr79, 15, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr80, 16, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr81, 17, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr82, 18, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr83, 19, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr84, 20, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr85, 21, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr86, 22, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr87, 23, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr88, 24, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr89, 25, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr90, 26, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr91, 27, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr92, 28, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr93, 29, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr94, 30, $vgpr0, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr95, 31, $vgpr0, implicit killed $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 160, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.7, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
+    ; GCN64-FLATSCR-NEXT: renamable $sgpr12 = IMPLICIT_DEF
+    ; GCN64-FLATSCR-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 1, implicit-def $vgpr0
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0
+    ; GCN64-FLATSCR-NEXT: $sgpr9 = S_ADD_I32 $sgpr33, 4096, implicit-def $scc
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, killed $sgpr9, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.8, align 4096, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
     renamable $sgpr12 = IMPLICIT_DEF
     SI_SPILL_S32_SAVE killed $sgpr12, %stack.0, implicit $exec, implicit $sgpr96_sgpr97_sgpr98_sgpr99, implicit $sgpr32
 
@@ -1000,408 +1003,411 @@ body:             |
 
     ; GCN64-MUBUF-LABEL: name: check_reload
     ; GCN64-MUBUF: liveins: $sgpr8, $sgpr4_sgpr5, $sgpr6_sgpr7, $sgpr11
-    ; GCN64-MUBUF: $sgpr33 = S_MOV_B32 0
-    ; GCN64-MUBUF: $sgpr28 = S_MOV_B32 &SCRATCH_RSRC_DWORD0, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
-    ; GCN64-MUBUF: $sgpr29 = S_MOV_B32 &SCRATCH_RSRC_DWORD1, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
-    ; GCN64-MUBUF: $sgpr30 = S_MOV_B32 4294967295, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
-    ; GCN64-MUBUF: $sgpr31 = S_MOV_B32 14680064, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
-    ; GCN64-MUBUF: $sgpr28 = S_ADD_U32 $sgpr28, $sgpr11, implicit-def $scc, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
-    ; GCN64-MUBUF: $sgpr29 = S_ADDC_U32 $sgpr29, 0, implicit-def $scc, implicit $scc, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
-    ; GCN64-MUBUF: $sgpr0_sgpr1 = S_MOV_B64 $exec
-    ; GCN64-MUBUF: $exec = S_MOV_B64 1, implicit-def $vgpr0
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 4, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
-    ; GCN64-MUBUF: $sgpr12 = V_READLANE_B32 killed $vgpr0, 0
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
-    ; GCN64-MUBUF: $sgpr0_sgpr1 = S_MOV_B64 $exec
-    ; GCN64-MUBUF: $exec = S_MOV_B64 3, implicit-def $vgpr0
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 8, 0, 0, 0, implicit $exec :: (load (s32) from %stack.1, addrspace 5)
-    ; GCN64-MUBUF: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13
-    ; GCN64-MUBUF: $sgpr13 = V_READLANE_B32 killed $vgpr0, 1
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
-    ; GCN64-MUBUF: $sgpr0_sgpr1 = S_MOV_B64 $exec
-    ; GCN64-MUBUF: $exec = S_MOV_B64 7, implicit-def $vgpr0
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 16, 0, 0, 0, implicit $exec :: (load (s32) from %stack.2, addrspace 5)
-    ; GCN64-MUBUF: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14
-    ; GCN64-MUBUF: $sgpr13 = V_READLANE_B32 $vgpr0, 1
-    ; GCN64-MUBUF: $sgpr14 = V_READLANE_B32 killed $vgpr0, 2
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
-    ; GCN64-MUBUF: $sgpr0_sgpr1 = S_MOV_B64 $exec
-    ; GCN64-MUBUF: $exec = S_MOV_B64 15, implicit-def $vgpr0
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 28, 0, 0, 0, implicit $exec :: (load (s32) from %stack.3, addrspace 5)
-    ; GCN64-MUBUF: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15
-    ; GCN64-MUBUF: $sgpr13 = V_READLANE_B32 $vgpr0, 1
-    ; GCN64-MUBUF: $sgpr14 = V_READLANE_B32 $vgpr0, 2
-    ; GCN64-MUBUF: $sgpr15 = V_READLANE_B32 killed $vgpr0, 3
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
-    ; GCN64-MUBUF: $sgpr0_sgpr1 = S_MOV_B64 $exec
-    ; GCN64-MUBUF: $exec = S_MOV_B64 31, implicit-def $vgpr0
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 44, 0, 0, 0, implicit $exec :: (load (s32) from %stack.4, addrspace 5)
-    ; GCN64-MUBUF: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN64-MUBUF: $sgpr13 = V_READLANE_B32 $vgpr0, 1
-    ; GCN64-MUBUF: $sgpr14 = V_READLANE_B32 $vgpr0, 2
-    ; GCN64-MUBUF: $sgpr15 = V_READLANE_B32 $vgpr0, 3
-    ; GCN64-MUBUF: $sgpr16 = V_READLANE_B32 killed $vgpr0, 4
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
-    ; GCN64-MUBUF: $sgpr0_sgpr1 = S_MOV_B64 $exec
-    ; GCN64-MUBUF: $exec = S_MOV_B64 255, implicit-def $vgpr0
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 64, 0, 0, 0, implicit $exec :: (load (s32) from %stack.5, addrspace 5)
-    ; GCN64-MUBUF: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-MUBUF: $sgpr13 = V_READLANE_B32 $vgpr0, 1
-    ; GCN64-MUBUF: $sgpr14 = V_READLANE_B32 $vgpr0, 2
-    ; GCN64-MUBUF: $sgpr15 = V_READLANE_B32 $vgpr0, 3
-    ; GCN64-MUBUF: $sgpr16 = V_READLANE_B32 $vgpr0, 4
-    ; GCN64-MUBUF: $sgpr17 = V_READLANE_B32 $vgpr0, 5
-    ; GCN64-MUBUF: $sgpr18 = V_READLANE_B32 $vgpr0, 6
-    ; GCN64-MUBUF: $sgpr19 = V_READLANE_B32 killed $vgpr0, 7
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
-    ; GCN64-MUBUF: $sgpr0_sgpr1 = S_MOV_B64 $exec
-    ; GCN64-MUBUF: $exec = S_MOV_B64 65535, implicit-def $vgpr0
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 96, 0, 0, 0, implicit $exec :: (load (s32) from %stack.6, addrspace 5)
-    ; GCN64-MUBUF: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF: $sgpr13 = V_READLANE_B32 $vgpr0, 1
-    ; GCN64-MUBUF: $sgpr14 = V_READLANE_B32 $vgpr0, 2
-    ; GCN64-MUBUF: $sgpr15 = V_READLANE_B32 $vgpr0, 3
-    ; GCN64-MUBUF: $sgpr16 = V_READLANE_B32 $vgpr0, 4
-    ; GCN64-MUBUF: $sgpr17 = V_READLANE_B32 $vgpr0, 5
-    ; GCN64-MUBUF: $sgpr18 = V_READLANE_B32 $vgpr0, 6
-    ; GCN64-MUBUF: $sgpr19 = V_READLANE_B32 $vgpr0, 7
-    ; GCN64-MUBUF: $sgpr20 = V_READLANE_B32 $vgpr0, 8
-    ; GCN64-MUBUF: $sgpr21 = V_READLANE_B32 $vgpr0, 9
-    ; GCN64-MUBUF: $sgpr22 = V_READLANE_B32 $vgpr0, 10
-    ; GCN64-MUBUF: $sgpr23 = V_READLANE_B32 $vgpr0, 11
-    ; GCN64-MUBUF: $sgpr24 = V_READLANE_B32 $vgpr0, 12
-    ; GCN64-MUBUF: $sgpr25 = V_READLANE_B32 $vgpr0, 13
-    ; GCN64-MUBUF: $sgpr26 = V_READLANE_B32 $vgpr0, 14
-    ; GCN64-MUBUF: $sgpr27 = V_READLANE_B32 killed $vgpr0, 15
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
-    ; GCN64-MUBUF: $sgpr0_sgpr1 = S_MOV_B64 $exec
-    ; GCN64-MUBUF: $exec = S_MOV_B64 4294967295, implicit-def $vgpr0
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 160, 0, 0, 0, implicit $exec :: (load (s32) from %stack.7, addrspace 5)
-    ; GCN64-MUBUF: $sgpr64 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF: $sgpr65 = V_READLANE_B32 $vgpr0, 1
-    ; GCN64-MUBUF: $sgpr66 = V_READLANE_B32 $vgpr0, 2
-    ; GCN64-MUBUF: $sgpr67 = V_READLANE_B32 $vgpr0, 3
-    ; GCN64-MUBUF: $sgpr68 = V_READLANE_B32 $vgpr0, 4
-    ; GCN64-MUBUF: $sgpr69 = V_READLANE_B32 $vgpr0, 5
-    ; GCN64-MUBUF: $sgpr70 = V_READLANE_B32 $vgpr0, 6
-    ; GCN64-MUBUF: $sgpr71 = V_READLANE_B32 $vgpr0, 7
-    ; GCN64-MUBUF: $sgpr72 = V_READLANE_B32 $vgpr0, 8
-    ; GCN64-MUBUF: $sgpr73 = V_READLANE_B32 $vgpr0, 9
-    ; GCN64-MUBUF: $sgpr74 = V_READLANE_B32 $vgpr0, 10
-    ; GCN64-MUBUF: $sgpr75 = V_READLANE_B32 $vgpr0, 11
-    ; GCN64-MUBUF: $sgpr76 = V_READLANE_B32 $vgpr0, 12
-    ; GCN64-MUBUF: $sgpr77 = V_READLANE_B32 $vgpr0, 13
-    ; GCN64-MUBUF: $sgpr78 = V_READLANE_B32 $vgpr0, 14
-    ; GCN64-MUBUF: $sgpr79 = V_READLANE_B32 $vgpr0, 15
-    ; GCN64-MUBUF: $sgpr80 = V_READLANE_B32 $vgpr0, 16
-    ; GCN64-MUBUF: $sgpr81 = V_READLANE_B32 $vgpr0, 17
-    ; GCN64-MUBUF: $sgpr82 = V_READLANE_B32 $vgpr0, 18
-    ; GCN64-MUBUF: $sgpr83 = V_READLANE_B32 $vgpr0, 19
-    ; GCN64-MUBUF: $sgpr84 = V_READLANE_B32 $vgpr0, 20
-    ; GCN64-MUBUF: $sgpr85 = V_READLANE_B32 $vgpr0, 21
-    ; GCN64-MUBUF: $sgpr86 = V_READLANE_B32 $vgpr0, 22
-    ; GCN64-MUBUF: $sgpr87 = V_READLANE_B32 $vgpr0, 23
-    ; GCN64-MUBUF: $sgpr88 = V_READLANE_B32 $vgpr0, 24
-    ; GCN64-MUBUF: $sgpr89 = V_READLANE_B32 $vgpr0, 25
-    ; GCN64-MUBUF: $sgpr90 = V_READLANE_B32 $vgpr0, 26
-    ; GCN64-MUBUF: $sgpr91 = V_READLANE_B32 $vgpr0, 27
-    ; GCN64-MUBUF: $sgpr92 = V_READLANE_B32 $vgpr0, 28
-    ; GCN64-MUBUF: $sgpr93 = V_READLANE_B32 $vgpr0, 29
-    ; GCN64-MUBUF: $sgpr94 = V_READLANE_B32 $vgpr0, 30
-    ; GCN64-MUBUF: $sgpr95 = V_READLANE_B32 killed $vgpr0, 31
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
-    ; GCN64-MUBUF: $sgpr0_sgpr1 = S_MOV_B64 $exec
-    ; GCN64-MUBUF: $exec = S_MOV_B64 1, implicit-def $vgpr0
-    ; GCN64-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $sgpr2 = S_ADD_I32 $sgpr33, 262144, implicit-def $scc
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, killed $sgpr2, 0, 0, 0, 0, implicit $exec :: (load (s32) from %stack.8, align 4096, addrspace 5)
-    ; GCN64-MUBUF: $sgpr12 = V_READLANE_B32 killed $vgpr0, 0
-    ; GCN64-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
+    ; GCN64-MUBUF-NEXT: {{  $}}
+    ; GCN64-MUBUF-NEXT: $sgpr33 = S_MOV_B32 0
+    ; GCN64-MUBUF-NEXT: $sgpr28 = S_MOV_B32 &SCRATCH_RSRC_DWORD0, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
+    ; GCN64-MUBUF-NEXT: $sgpr29 = S_MOV_B32 &SCRATCH_RSRC_DWORD1, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
+    ; GCN64-MUBUF-NEXT: $sgpr30 = S_MOV_B32 4294967295, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
+    ; GCN64-MUBUF-NEXT: $sgpr31 = S_MOV_B32 14680064, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
+    ; GCN64-MUBUF-NEXT: $sgpr28 = S_ADD_U32 $sgpr28, $sgpr11, implicit-def $scc, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
+    ; GCN64-MUBUF-NEXT: $sgpr29 = S_ADDC_U32 $sgpr29, 0, implicit-def $scc, implicit $scc, implicit-def $sgpr28_sgpr29_sgpr30_sgpr31
+    ; GCN64-MUBUF-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 1, implicit-def $vgpr0
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 4, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $sgpr12 = V_READLANE_B32 killed $vgpr0, 0
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
+    ; GCN64-MUBUF-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 3, implicit-def $vgpr0
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 8, 0, 0, 0, implicit $exec :: (load (s32) from %stack.1, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13
+    ; GCN64-MUBUF-NEXT: $sgpr13 = V_READLANE_B32 killed $vgpr0, 1
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
+    ; GCN64-MUBUF-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 7, implicit-def $vgpr0
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 16, 0, 0, 0, implicit $exec :: (load (s32) from %stack.2, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14
+    ; GCN64-MUBUF-NEXT: $sgpr13 = V_READLANE_B32 $vgpr0, 1
+    ; GCN64-MUBUF-NEXT: $sgpr14 = V_READLANE_B32 killed $vgpr0, 2
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
+    ; GCN64-MUBUF-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 15, implicit-def $vgpr0
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 28, 0, 0, 0, implicit $exec :: (load (s32) from %stack.3, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15
+    ; GCN64-MUBUF-NEXT: $sgpr13 = V_READLANE_B32 $vgpr0, 1
+    ; GCN64-MUBUF-NEXT: $sgpr14 = V_READLANE_B32 $vgpr0, 2
+    ; GCN64-MUBUF-NEXT: $sgpr15 = V_READLANE_B32 killed $vgpr0, 3
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
+    ; GCN64-MUBUF-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 31, implicit-def $vgpr0
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 44, 0, 0, 0, implicit $exec :: (load (s32) from %stack.4, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
+    ; GCN64-MUBUF-NEXT: $sgpr13 = V_READLANE_B32 $vgpr0, 1
+    ; GCN64-MUBUF-NEXT: $sgpr14 = V_READLANE_B32 $vgpr0, 2
+    ; GCN64-MUBUF-NEXT: $sgpr15 = V_READLANE_B32 $vgpr0, 3
+    ; GCN64-MUBUF-NEXT: $sgpr16 = V_READLANE_B32 killed $vgpr0, 4
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
+    ; GCN64-MUBUF-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 255, implicit-def $vgpr0
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 64, 0, 0, 0, implicit $exec :: (load (s32) from %stack.5, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN64-MUBUF-NEXT: $sgpr13 = V_READLANE_B32 $vgpr0, 1
+    ; GCN64-MUBUF-NEXT: $sgpr14 = V_READLANE_B32 $vgpr0, 2
+    ; GCN64-MUBUF-NEXT: $sgpr15 = V_READLANE_B32 $vgpr0, 3
+    ; GCN64-MUBUF-NEXT: $sgpr16 = V_READLANE_B32 $vgpr0, 4
+    ; GCN64-MUBUF-NEXT: $sgpr17 = V_READLANE_B32 $vgpr0, 5
+    ; GCN64-MUBUF-NEXT: $sgpr18 = V_READLANE_B32 $vgpr0, 6
+    ; GCN64-MUBUF-NEXT: $sgpr19 = V_READLANE_B32 killed $vgpr0, 7
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
+    ; GCN64-MUBUF-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 65535, implicit-def $vgpr0
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 96, 0, 0, 0, implicit $exec :: (load (s32) from %stack.6, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-MUBUF-NEXT: $sgpr13 = V_READLANE_B32 $vgpr0, 1
+    ; GCN64-MUBUF-NEXT: $sgpr14 = V_READLANE_B32 $vgpr0, 2
+    ; GCN64-MUBUF-NEXT: $sgpr15 = V_READLANE_B32 $vgpr0, 3
+    ; GCN64-MUBUF-NEXT: $sgpr16 = V_READLANE_B32 $vgpr0, 4
+    ; GCN64-MUBUF-NEXT: $sgpr17 = V_READLANE_B32 $vgpr0, 5
+    ; GCN64-MUBUF-NEXT: $sgpr18 = V_READLANE_B32 $vgpr0, 6
+    ; GCN64-MUBUF-NEXT: $sgpr19 = V_READLANE_B32 $vgpr0, 7
+    ; GCN64-MUBUF-NEXT: $sgpr20 = V_READLANE_B32 $vgpr0, 8
+    ; GCN64-MUBUF-NEXT: $sgpr21 = V_READLANE_B32 $vgpr0, 9
+    ; GCN64-MUBUF-NEXT: $sgpr22 = V_READLANE_B32 $vgpr0, 10
+    ; GCN64-MUBUF-NEXT: $sgpr23 = V_READLANE_B32 $vgpr0, 11
+    ; GCN64-MUBUF-NEXT: $sgpr24 = V_READLANE_B32 $vgpr0, 12
+    ; GCN64-MUBUF-NEXT: $sgpr25 = V_READLANE_B32 $vgpr0, 13
+    ; GCN64-MUBUF-NEXT: $sgpr26 = V_READLANE_B32 $vgpr0, 14
+    ; GCN64-MUBUF-NEXT: $sgpr27 = V_READLANE_B32 killed $vgpr0, 15
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
+    ; GCN64-MUBUF-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 4294967295, implicit-def $vgpr0
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 160, 0, 0, 0, implicit $exec :: (load (s32) from %stack.7, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $sgpr64 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $sgpr65 = V_READLANE_B32 $vgpr0, 1
+    ; GCN64-MUBUF-NEXT: $sgpr66 = V_READLANE_B32 $vgpr0, 2
+    ; GCN64-MUBUF-NEXT: $sgpr67 = V_READLANE_B32 $vgpr0, 3
+    ; GCN64-MUBUF-NEXT: $sgpr68 = V_READLANE_B32 $vgpr0, 4
+    ; GCN64-MUBUF-NEXT: $sgpr69 = V_READLANE_B32 $vgpr0, 5
+    ; GCN64-MUBUF-NEXT: $sgpr70 = V_READLANE_B32 $vgpr0, 6
+    ; GCN64-MUBUF-NEXT: $sgpr71 = V_READLANE_B32 $vgpr0, 7
+    ; GCN64-MUBUF-NEXT: $sgpr72 = V_READLANE_B32 $vgpr0, 8
+    ; GCN64-MUBUF-NEXT: $sgpr73 = V_READLANE_B32 $vgpr0, 9
+    ; GCN64-MUBUF-NEXT: $sgpr74 = V_READLANE_B32 $vgpr0, 10
+    ; GCN64-MUBUF-NEXT: $sgpr75 = V_READLANE_B32 $vgpr0, 11
+    ; GCN64-MUBUF-NEXT: $sgpr76 = V_READLANE_B32 $vgpr0, 12
+    ; GCN64-MUBUF-NEXT: $sgpr77 = V_READLANE_B32 $vgpr0, 13
+    ; GCN64-MUBUF-NEXT: $sgpr78 = V_READLANE_B32 $vgpr0, 14
+    ; GCN64-MUBUF-NEXT: $sgpr79 = V_READLANE_B32 $vgpr0, 15
+    ; GCN64-MUBUF-NEXT: $sgpr80 = V_READLANE_B32 $vgpr0, 16
+    ; GCN64-MUBUF-NEXT: $sgpr81 = V_READLANE_B32 $vgpr0, 17
+    ; GCN64-MUBUF-NEXT: $sgpr82 = V_READLANE_B32 $vgpr0, 18
+    ; GCN64-MUBUF-NEXT: $sgpr83 = V_READLANE_B32 $vgpr0, 19
+    ; GCN64-MUBUF-NEXT: $sgpr84 = V_READLANE_B32 $vgpr0, 20
+    ; GCN64-MUBUF-NEXT: $sgpr85 = V_READLANE_B32 $vgpr0, 21
+    ; GCN64-MUBUF-NEXT: $sgpr86 = V_READLANE_B32 $vgpr0, 22
+    ; GCN64-MUBUF-NEXT: $sgpr87 = V_READLANE_B32 $vgpr0, 23
+    ; GCN64-MUBUF-NEXT: $sgpr88 = V_READLANE_B32 $vgpr0, 24
+    ; GCN64-MUBUF-NEXT: $sgpr89 = V_READLANE_B32 $vgpr0, 25
+    ; GCN64-MUBUF-NEXT: $sgpr90 = V_READLANE_B32 $vgpr0, 26
+    ; GCN64-MUBUF-NEXT: $sgpr91 = V_READLANE_B32 $vgpr0, 27
+    ; GCN64-MUBUF-NEXT: $sgpr92 = V_READLANE_B32 $vgpr0, 28
+    ; GCN64-MUBUF-NEXT: $sgpr93 = V_READLANE_B32 $vgpr0, 29
+    ; GCN64-MUBUF-NEXT: $sgpr94 = V_READLANE_B32 $vgpr0, 30
+    ; GCN64-MUBUF-NEXT: $sgpr95 = V_READLANE_B32 killed $vgpr0, 31
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
+    ; GCN64-MUBUF-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 1, implicit-def $vgpr0
+    ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $sgpr2 = S_ADD_I32 $sgpr33, 262144, implicit-def $scc
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, killed $sgpr2, 0, 0, 0, 0, implicit $exec :: (load (s32) from %stack.8, align 4096, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $sgpr12 = V_READLANE_B32 killed $vgpr0, 0
+    ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1, implicit killed $vgpr0
     ; GCN32-MUBUF-LABEL: name: check_reload
     ; GCN32-MUBUF: liveins: $sgpr8, $sgpr4_sgpr5, $sgpr6_sgpr7, $sgpr11
-    ; GCN32-MUBUF: $sgpr33 = S_MOV_B32 0
-    ; GCN32-MUBUF: $sgpr96 = S_MOV_B32 &SCRATCH_RSRC_DWORD0, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
-    ; GCN32-MUBUF: $sgpr97 = S_MOV_B32 &SCRATCH_RSRC_DWORD1, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
-    ; GCN32-MUBUF: $sgpr98 = S_MOV_B32 4294967295, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
-    ; GCN32-MUBUF: $sgpr99 = S_MOV_B32 834756608, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
-    ; GCN32-MUBUF: $sgpr96 = S_ADD_U32 $sgpr96, $sgpr11, implicit-def $scc, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
-    ; GCN32-MUBUF: $sgpr97 = S_ADDC_U32 $sgpr97, 0, implicit-def $scc, implicit $scc, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
-    ; GCN32-MUBUF: $sgpr0 = S_MOV_B32 $exec_lo
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 1, implicit-def $vgpr0
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 4, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
-    ; GCN32-MUBUF: $sgpr12 = V_READLANE_B32 killed $vgpr0, 0
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
-    ; GCN32-MUBUF: $sgpr0 = S_MOV_B32 $exec_lo
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 3, implicit-def $vgpr0
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 8, 0, 0, 0, implicit $exec :: (load (s32) from %stack.1, addrspace 5)
-    ; GCN32-MUBUF: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13
-    ; GCN32-MUBUF: $sgpr13 = V_READLANE_B32 killed $vgpr0, 1
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
-    ; GCN32-MUBUF: $sgpr0 = S_MOV_B32 $exec_lo
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 7, implicit-def $vgpr0
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 16, 0, 0, 0, implicit $exec :: (load (s32) from %stack.2, addrspace 5)
-    ; GCN32-MUBUF: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14
-    ; GCN32-MUBUF: $sgpr13 = V_READLANE_B32 $vgpr0, 1
-    ; GCN32-MUBUF: $sgpr14 = V_READLANE_B32 killed $vgpr0, 2
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
-    ; GCN32-MUBUF: $sgpr0 = S_MOV_B32 $exec_lo
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 15, implicit-def $vgpr0
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 28, 0, 0, 0, implicit $exec :: (load (s32) from %stack.3, addrspace 5)
-    ; GCN32-MUBUF: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15
-    ; GCN32-MUBUF: $sgpr13 = V_READLANE_B32 $vgpr0, 1
-    ; GCN32-MUBUF: $sgpr14 = V_READLANE_B32 $vgpr0, 2
-    ; GCN32-MUBUF: $sgpr15 = V_READLANE_B32 killed $vgpr0, 3
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
-    ; GCN32-MUBUF: $sgpr0 = S_MOV_B32 $exec_lo
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 31, implicit-def $vgpr0
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 44, 0, 0, 0, implicit $exec :: (load (s32) from %stack.4, addrspace 5)
-    ; GCN32-MUBUF: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN32-MUBUF: $sgpr13 = V_READLANE_B32 $vgpr0, 1
-    ; GCN32-MUBUF: $sgpr14 = V_READLANE_B32 $vgpr0, 2
-    ; GCN32-MUBUF: $sgpr15 = V_READLANE_B32 $vgpr0, 3
-    ; GCN32-MUBUF: $sgpr16 = V_READLANE_B32 killed $vgpr0, 4
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
-    ; GCN32-MUBUF: $sgpr0 = S_MOV_B32 $exec_lo
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 255, implicit-def $vgpr0
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 64, 0, 0, 0, implicit $exec :: (load (s32) from %stack.5, addrspace 5)
-    ; GCN32-MUBUF: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN32-MUBUF: $sgpr13 = V_READLANE_B32 $vgpr0, 1
-    ; GCN32-MUBUF: $sgpr14 = V_READLANE_B32 $vgpr0, 2
-    ; GCN32-MUBUF: $sgpr15 = V_READLANE_B32 $vgpr0, 3
-    ; GCN32-MUBUF: $sgpr16 = V_READLANE_B32 $vgpr0, 4
-    ; GCN32-MUBUF: $sgpr17 = V_READLANE_B32 $vgpr0, 5
-    ; GCN32-MUBUF: $sgpr18 = V_READLANE_B32 $vgpr0, 6
-    ; GCN32-MUBUF: $sgpr19 = V_READLANE_B32 killed $vgpr0, 7
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
-    ; GCN32-MUBUF: $sgpr0 = S_MOV_B32 $exec_lo
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 65535, implicit-def $vgpr0
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 96, 0, 0, 0, implicit $exec :: (load (s32) from %stack.6, addrspace 5)
-    ; GCN32-MUBUF: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF: $sgpr13 = V_READLANE_B32 $vgpr0, 1
-    ; GCN32-MUBUF: $sgpr14 = V_READLANE_B32 $vgpr0, 2
-    ; GCN32-MUBUF: $sgpr15 = V_READLANE_B32 $vgpr0, 3
-    ; GCN32-MUBUF: $sgpr16 = V_READLANE_B32 $vgpr0, 4
-    ; GCN32-MUBUF: $sgpr17 = V_READLANE_B32 $vgpr0, 5
-    ; GCN32-MUBUF: $sgpr18 = V_READLANE_B32 $vgpr0, 6
-    ; GCN32-MUBUF: $sgpr19 = V_READLANE_B32 $vgpr0, 7
-    ; GCN32-MUBUF: $sgpr20 = V_READLANE_B32 $vgpr0, 8
-    ; GCN32-MUBUF: $sgpr21 = V_READLANE_B32 $vgpr0, 9
-    ; GCN32-MUBUF: $sgpr22 = V_READLANE_B32 $vgpr0, 10
-    ; GCN32-MUBUF: $sgpr23 = V_READLANE_B32 $vgpr0, 11
-    ; GCN32-MUBUF: $sgpr24 = V_READLANE_B32 $vgpr0, 12
-    ; GCN32-MUBUF: $sgpr25 = V_READLANE_B32 $vgpr0, 13
-    ; GCN32-MUBUF: $sgpr26 = V_READLANE_B32 $vgpr0, 14
-    ; GCN32-MUBUF: $sgpr27 = V_READLANE_B32 killed $vgpr0, 15
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
-    ; GCN32-MUBUF: $sgpr0 = S_MOV_B32 $exec_lo
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 4294967295, implicit-def $vgpr0
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 160, 0, 0, 0, implicit $exec :: (load (s32) from %stack.7, addrspace 5)
-    ; GCN32-MUBUF: $sgpr64 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF: $sgpr65 = V_READLANE_B32 $vgpr0, 1
-    ; GCN32-MUBUF: $sgpr66 = V_READLANE_B32 $vgpr0, 2
-    ; GCN32-MUBUF: $sgpr67 = V_READLANE_B32 $vgpr0, 3
-    ; GCN32-MUBUF: $sgpr68 = V_READLANE_B32 $vgpr0, 4
-    ; GCN32-MUBUF: $sgpr69 = V_READLANE_B32 $vgpr0, 5
-    ; GCN32-MUBUF: $sgpr70 = V_READLANE_B32 $vgpr0, 6
-    ; GCN32-MUBUF: $sgpr71 = V_READLANE_B32 $vgpr0, 7
-    ; GCN32-MUBUF: $sgpr72 = V_READLANE_B32 $vgpr0, 8
-    ; GCN32-MUBUF: $sgpr73 = V_READLANE_B32 $vgpr0, 9
-    ; GCN32-MUBUF: $sgpr74 = V_READLANE_B32 $vgpr0, 10
-    ; GCN32-MUBUF: $sgpr75 = V_READLANE_B32 $vgpr0, 11
-    ; GCN32-MUBUF: $sgpr76 = V_READLANE_B32 $vgpr0, 12
-    ; GCN32-MUBUF: $sgpr77 = V_READLANE_B32 $vgpr0, 13
-    ; GCN32-MUBUF: $sgpr78 = V_READLANE_B32 $vgpr0, 14
-    ; GCN32-MUBUF: $sgpr79 = V_READLANE_B32 $vgpr0, 15
-    ; GCN32-MUBUF: $sgpr80 = V_READLANE_B32 $vgpr0, 16
-    ; GCN32-MUBUF: $sgpr81 = V_READLANE_B32 $vgpr0, 17
-    ; GCN32-MUBUF: $sgpr82 = V_READLANE_B32 $vgpr0, 18
-    ; GCN32-MUBUF: $sgpr83 = V_READLANE_B32 $vgpr0, 19
-    ; GCN32-MUBUF: $sgpr84 = V_READLANE_B32 $vgpr0, 20
-    ; GCN32-MUBUF: $sgpr85 = V_READLANE_B32 $vgpr0, 21
-    ; GCN32-MUBUF: $sgpr86 = V_READLANE_B32 $vgpr0, 22
-    ; GCN32-MUBUF: $sgpr87 = V_READLANE_B32 $vgpr0, 23
-    ; GCN32-MUBUF: $sgpr88 = V_READLANE_B32 $vgpr0, 24
-    ; GCN32-MUBUF: $sgpr89 = V_READLANE_B32 $vgpr0, 25
-    ; GCN32-MUBUF: $sgpr90 = V_READLANE_B32 $vgpr0, 26
-    ; GCN32-MUBUF: $sgpr91 = V_READLANE_B32 $vgpr0, 27
-    ; GCN32-MUBUF: $sgpr92 = V_READLANE_B32 $vgpr0, 28
-    ; GCN32-MUBUF: $sgpr93 = V_READLANE_B32 $vgpr0, 29
-    ; GCN32-MUBUF: $sgpr94 = V_READLANE_B32 $vgpr0, 30
-    ; GCN32-MUBUF: $sgpr95 = V_READLANE_B32 killed $vgpr0, 31
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
-    ; GCN32-MUBUF: $sgpr0 = S_MOV_B32 $exec_lo
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 1, implicit-def $vgpr0
-    ; GCN32-MUBUF: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $sgpr1 = S_ADD_I32 $sgpr33, 131072, implicit-def $scc
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, killed $sgpr1, 0, 0, 0, 0, implicit $exec :: (load (s32) from %stack.8, align 4096, addrspace 5)
-    ; GCN32-MUBUF: $sgpr12 = V_READLANE_B32 killed $vgpr0, 0
-    ; GCN32-MUBUF: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
+    ; GCN32-MUBUF-NEXT: {{  $}}
+    ; GCN32-MUBUF-NEXT: $sgpr33 = S_MOV_B32 0
+    ; GCN32-MUBUF-NEXT: $sgpr96 = S_MOV_B32 &SCRATCH_RSRC_DWORD0, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
+    ; GCN32-MUBUF-NEXT: $sgpr97 = S_MOV_B32 &SCRATCH_RSRC_DWORD1, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
+    ; GCN32-MUBUF-NEXT: $sgpr98 = S_MOV_B32 4294967295, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
+    ; GCN32-MUBUF-NEXT: $sgpr99 = S_MOV_B32 834756608, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
+    ; GCN32-MUBUF-NEXT: $sgpr96 = S_ADD_U32 $sgpr96, $sgpr11, implicit-def $scc, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
+    ; GCN32-MUBUF-NEXT: $sgpr97 = S_ADDC_U32 $sgpr97, 0, implicit-def $scc, implicit $scc, implicit-def $sgpr96_sgpr97_sgpr98_sgpr99
+    ; GCN32-MUBUF-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 1, implicit-def $vgpr0
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 4, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $sgpr12 = V_READLANE_B32 killed $vgpr0, 0
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
+    ; GCN32-MUBUF-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 3, implicit-def $vgpr0
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 8, 0, 0, 0, implicit $exec :: (load (s32) from %stack.1, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13
+    ; GCN32-MUBUF-NEXT: $sgpr13 = V_READLANE_B32 killed $vgpr0, 1
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
+    ; GCN32-MUBUF-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 7, implicit-def $vgpr0
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 16, 0, 0, 0, implicit $exec :: (load (s32) from %stack.2, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14
+    ; GCN32-MUBUF-NEXT: $sgpr13 = V_READLANE_B32 $vgpr0, 1
+    ; GCN32-MUBUF-NEXT: $sgpr14 = V_READLANE_B32 killed $vgpr0, 2
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
+    ; GCN32-MUBUF-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 15, implicit-def $vgpr0
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 28, 0, 0, 0, implicit $exec :: (load (s32) from %stack.3, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15
+    ; GCN32-MUBUF-NEXT: $sgpr13 = V_READLANE_B32 $vgpr0, 1
+    ; GCN32-MUBUF-NEXT: $sgpr14 = V_READLANE_B32 $vgpr0, 2
+    ; GCN32-MUBUF-NEXT: $sgpr15 = V_READLANE_B32 killed $vgpr0, 3
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
+    ; GCN32-MUBUF-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 31, implicit-def $vgpr0
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 44, 0, 0, 0, implicit $exec :: (load (s32) from %stack.4, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
+    ; GCN32-MUBUF-NEXT: $sgpr13 = V_READLANE_B32 $vgpr0, 1
+    ; GCN32-MUBUF-NEXT: $sgpr14 = V_READLANE_B32 $vgpr0, 2
+    ; GCN32-MUBUF-NEXT: $sgpr15 = V_READLANE_B32 $vgpr0, 3
+    ; GCN32-MUBUF-NEXT: $sgpr16 = V_READLANE_B32 killed $vgpr0, 4
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
+    ; GCN32-MUBUF-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 255, implicit-def $vgpr0
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 64, 0, 0, 0, implicit $exec :: (load (s32) from %stack.5, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN32-MUBUF-NEXT: $sgpr13 = V_READLANE_B32 $vgpr0, 1
+    ; GCN32-MUBUF-NEXT: $sgpr14 = V_READLANE_B32 $vgpr0, 2
+    ; GCN32-MUBUF-NEXT: $sgpr15 = V_READLANE_B32 $vgpr0, 3
+    ; GCN32-MUBUF-NEXT: $sgpr16 = V_READLANE_B32 $vgpr0, 4
+    ; GCN32-MUBUF-NEXT: $sgpr17 = V_READLANE_B32 $vgpr0, 5
+    ; GCN32-MUBUF-NEXT: $sgpr18 = V_READLANE_B32 $vgpr0, 6
+    ; GCN32-MUBUF-NEXT: $sgpr19 = V_READLANE_B32 killed $vgpr0, 7
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
+    ; GCN32-MUBUF-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 65535, implicit-def $vgpr0
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 96, 0, 0, 0, implicit $exec :: (load (s32) from %stack.6, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN32-MUBUF-NEXT: $sgpr13 = V_READLANE_B32 $vgpr0, 1
+    ; GCN32-MUBUF-NEXT: $sgpr14 = V_READLANE_B32 $vgpr0, 2
+    ; GCN32-MUBUF-NEXT: $sgpr15 = V_READLANE_B32 $vgpr0, 3
+    ; GCN32-MUBUF-NEXT: $sgpr16 = V_READLANE_B32 $vgpr0, 4
+    ; GCN32-MUBUF-NEXT: $sgpr17 = V_READLANE_B32 $vgpr0, 5
+    ; GCN32-MUBUF-NEXT: $sgpr18 = V_READLANE_B32 $vgpr0, 6
+    ; GCN32-MUBUF-NEXT: $sgpr19 = V_READLANE_B32 $vgpr0, 7
+    ; GCN32-MUBUF-NEXT: $sgpr20 = V_READLANE_B32 $vgpr0, 8
+    ; GCN32-MUBUF-NEXT: $sgpr21 = V_READLANE_B32 $vgpr0, 9
+    ; GCN32-MUBUF-NEXT: $sgpr22 = V_READLANE_B32 $vgpr0, 10
+    ; GCN32-MUBUF-NEXT: $sgpr23 = V_READLANE_B32 $vgpr0, 11
+    ; GCN32-MUBUF-NEXT: $sgpr24 = V_READLANE_B32 $vgpr0, 12
+    ; GCN32-MUBUF-NEXT: $sgpr25 = V_READLANE_B32 $vgpr0, 13
+    ; GCN32-MUBUF-NEXT: $sgpr26 = V_READLANE_B32 $vgpr0, 14
+    ; GCN32-MUBUF-NEXT: $sgpr27 = V_READLANE_B32 killed $vgpr0, 15
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
+    ; GCN32-MUBUF-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 4294967295, implicit-def $vgpr0
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 160, 0, 0, 0, implicit $exec :: (load (s32) from %stack.7, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $sgpr64 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $sgpr65 = V_READLANE_B32 $vgpr0, 1
+    ; GCN32-MUBUF-NEXT: $sgpr66 = V_READLANE_B32 $vgpr0, 2
+    ; GCN32-MUBUF-NEXT: $sgpr67 = V_READLANE_B32 $vgpr0, 3
+    ; GCN32-MUBUF-NEXT: $sgpr68 = V_READLANE_B32 $vgpr0, 4
+    ; GCN32-MUBUF-NEXT: $sgpr69 = V_READLANE_B32 $vgpr0, 5
+    ; GCN32-MUBUF-NEXT: $sgpr70 = V_READLANE_B32 $vgpr0, 6
+    ; GCN32-MUBUF-NEXT: $sgpr71 = V_READLANE_B32 $vgpr0, 7
+    ; GCN32-MUBUF-NEXT: $sgpr72 = V_READLANE_B32 $vgpr0, 8
+    ; GCN32-MUBUF-NEXT: $sgpr73 = V_READLANE_B32 $vgpr0, 9
+    ; GCN32-MUBUF-NEXT: $sgpr74 = V_READLANE_B32 $vgpr0, 10
+    ; GCN32-MUBUF-NEXT: $sgpr75 = V_READLANE_B32 $vgpr0, 11
+    ; GCN32-MUBUF-NEXT: $sgpr76 = V_READLANE_B32 $vgpr0, 12
+    ; GCN32-MUBUF-NEXT: $sgpr77 = V_READLANE_B32 $vgpr0, 13
+    ; GCN32-MUBUF-NEXT: $sgpr78 = V_READLANE_B32 $vgpr0, 14
+    ; GCN32-MUBUF-NEXT: $sgpr79 = V_READLANE_B32 $vgpr0, 15
+    ; GCN32-MUBUF-NEXT: $sgpr80 = V_READLANE_B32 $vgpr0, 16
+    ; GCN32-MUBUF-NEXT: $sgpr81 = V_READLANE_B32 $vgpr0, 17
+    ; GCN32-MUBUF-NEXT: $sgpr82 = V_READLANE_B32 $vgpr0, 18
+    ; GCN32-MUBUF-NEXT: $sgpr83 = V_READLANE_B32 $vgpr0, 19
+    ; GCN32-MUBUF-NEXT: $sgpr84 = V_READLANE_B32 $vgpr0, 20
+    ; GCN32-MUBUF-NEXT: $sgpr85 = V_READLANE_B32 $vgpr0, 21
+    ; GCN32-MUBUF-NEXT: $sgpr86 = V_READLANE_B32 $vgpr0, 22
+    ; GCN32-MUBUF-NEXT: $sgpr87 = V_READLANE_B32 $vgpr0, 23
+    ; GCN32-MUBUF-NEXT: $sgpr88 = V_READLANE_B32 $vgpr0, 24
+    ; GCN32-MUBUF-NEXT: $sgpr89 = V_READLANE_B32 $vgpr0, 25
+    ; GCN32-MUBUF-NEXT: $sgpr90 = V_READLANE_B32 $vgpr0, 26
+    ; GCN32-MUBUF-NEXT: $sgpr91 = V_READLANE_B32 $vgpr0, 27
+    ; GCN32-MUBUF-NEXT: $sgpr92 = V_READLANE_B32 $vgpr0, 28
+    ; GCN32-MUBUF-NEXT: $sgpr93 = V_READLANE_B32 $vgpr0, 29
+    ; GCN32-MUBUF-NEXT: $sgpr94 = V_READLANE_B32 $vgpr0, 30
+    ; GCN32-MUBUF-NEXT: $sgpr95 = V_READLANE_B32 killed $vgpr0, 31
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
+    ; GCN32-MUBUF-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 1, implicit-def $vgpr0
+    ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $sgpr1 = S_ADD_I32 $sgpr33, 131072, implicit-def $scc
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, killed $sgpr1, 0, 0, 0, 0, implicit $exec :: (load (s32) from %stack.8, align 4096, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $sgpr12 = V_READLANE_B32 killed $vgpr0, 0
+    ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 killed $sgpr0, implicit killed $vgpr0
     ; GCN64-FLATSCR-LABEL: name: check_reload
     ; GCN64-FLATSCR: liveins: $sgpr8, $sgpr4_sgpr5, $sgpr6_sgpr7, $sgpr11, $sgpr0_sgpr1
-    ; GCN64-FLATSCR: $sgpr33 = S_MOV_B32 0
-    ; GCN64-FLATSCR: $flat_scr_lo = S_ADD_U32 $sgpr0, $sgpr11, implicit-def $scc
-    ; GCN64-FLATSCR: $flat_scr_hi = S_ADDC_U32 $sgpr1, 0, implicit-def $scc, implicit $scc
-    ; GCN64-FLATSCR: $sgpr2_sgpr3 = S_MOV_B64 $exec
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 1, implicit-def $vgpr0
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 4, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0, addrspace 5)
-    ; GCN64-FLATSCR: $sgpr12 = V_READLANE_B32 killed $vgpr0, 0
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
-    ; GCN64-FLATSCR: $sgpr2_sgpr3 = S_MOV_B64 $exec
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 3, implicit-def $vgpr0
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 8, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.1, addrspace 5)
-    ; GCN64-FLATSCR: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13
-    ; GCN64-FLATSCR: $sgpr13 = V_READLANE_B32 killed $vgpr0, 1
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
-    ; GCN64-FLATSCR: $sgpr2_sgpr3 = S_MOV_B64 $exec
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 7, implicit-def $vgpr0
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 16, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.2, addrspace 5)
-    ; GCN64-FLATSCR: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14
-    ; GCN64-FLATSCR: $sgpr13 = V_READLANE_B32 $vgpr0, 1
-    ; GCN64-FLATSCR: $sgpr14 = V_READLANE_B32 killed $vgpr0, 2
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
-    ; GCN64-FLATSCR: $sgpr2_sgpr3 = S_MOV_B64 $exec
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 15, implicit-def $vgpr0
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 28, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.3, addrspace 5)
-    ; GCN64-FLATSCR: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15
-    ; GCN64-FLATSCR: $sgpr13 = V_READLANE_B32 $vgpr0, 1
-    ; GCN64-FLATSCR: $sgpr14 = V_READLANE_B32 $vgpr0, 2
-    ; GCN64-FLATSCR: $sgpr15 = V_READLANE_B32 killed $vgpr0, 3
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
-    ; GCN64-FLATSCR: $sgpr2_sgpr3 = S_MOV_B64 $exec
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 31, implicit-def $vgpr0
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 44, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.4, addrspace 5)
-    ; GCN64-FLATSCR: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN64-FLATSCR: $sgpr13 = V_READLANE_B32 $vgpr0, 1
-    ; GCN64-FLATSCR: $sgpr14 = V_READLANE_B32 $vgpr0, 2
-    ; GCN64-FLATSCR: $sgpr15 = V_READLANE_B32 $vgpr0, 3
-    ; GCN64-FLATSCR: $sgpr16 = V_READLANE_B32 killed $vgpr0, 4
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
-    ; GCN64-FLATSCR: $sgpr2_sgpr3 = S_MOV_B64 $exec
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 255, implicit-def $vgpr0
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 64, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.5, addrspace 5)
-    ; GCN64-FLATSCR: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-FLATSCR: $sgpr13 = V_READLANE_B32 $vgpr0, 1
-    ; GCN64-FLATSCR: $sgpr14 = V_READLANE_B32 $vgpr0, 2
-    ; GCN64-FLATSCR: $sgpr15 = V_READLANE_B32 $vgpr0, 3
-    ; GCN64-FLATSCR: $sgpr16 = V_READLANE_B32 $vgpr0, 4
-    ; GCN64-FLATSCR: $sgpr17 = V_READLANE_B32 $vgpr0, 5
-    ; GCN64-FLATSCR: $sgpr18 = V_READLANE_B32 $vgpr0, 6
-    ; GCN64-FLATSCR: $sgpr19 = V_READLANE_B32 killed $vgpr0, 7
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
-    ; GCN64-FLATSCR: $sgpr2_sgpr3 = S_MOV_B64 $exec
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 65535, implicit-def $vgpr0
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 96, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.6, addrspace 5)
-    ; GCN64-FLATSCR: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR: $sgpr13 = V_READLANE_B32 $vgpr0, 1
-    ; GCN64-FLATSCR: $sgpr14 = V_READLANE_B32 $vgpr0, 2
-    ; GCN64-FLATSCR: $sgpr15 = V_READLANE_B32 $vgpr0, 3
-    ; GCN64-FLATSCR: $sgpr16 = V_READLANE_B32 $vgpr0, 4
-    ; GCN64-FLATSCR: $sgpr17 = V_READLANE_B32 $vgpr0, 5
-    ; GCN64-FLATSCR: $sgpr18 = V_READLANE_B32 $vgpr0, 6
-    ; GCN64-FLATSCR: $sgpr19 = V_READLANE_B32 $vgpr0, 7
-    ; GCN64-FLATSCR: $sgpr20 = V_READLANE_B32 $vgpr0, 8
-    ; GCN64-FLATSCR: $sgpr21 = V_READLANE_B32 $vgpr0, 9
-    ; GCN64-FLATSCR: $sgpr22 = V_READLANE_B32 $vgpr0, 10
-    ; GCN64-FLATSCR: $sgpr23 = V_READLANE_B32 $vgpr0, 11
-    ; GCN64-FLATSCR: $sgpr24 = V_READLANE_B32 $vgpr0, 12
-    ; GCN64-FLATSCR: $sgpr25 = V_READLANE_B32 $vgpr0, 13
-    ; GCN64-FLATSCR: $sgpr26 = V_READLANE_B32 $vgpr0, 14
-    ; GCN64-FLATSCR: $sgpr27 = V_READLANE_B32 killed $vgpr0, 15
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
-    ; GCN64-FLATSCR: $sgpr2_sgpr3 = S_MOV_B64 $exec
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 4294967295, implicit-def $vgpr0
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 160, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.7, addrspace 5)
-    ; GCN64-FLATSCR: $sgpr64 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR: $sgpr65 = V_READLANE_B32 $vgpr0, 1
-    ; GCN64-FLATSCR: $sgpr66 = V_READLANE_B32 $vgpr0, 2
-    ; GCN64-FLATSCR: $sgpr67 = V_READLANE_B32 $vgpr0, 3
-    ; GCN64-FLATSCR: $sgpr68 = V_READLANE_B32 $vgpr0, 4
-    ; GCN64-FLATSCR: $sgpr69 = V_READLANE_B32 $vgpr0, 5
-    ; GCN64-FLATSCR: $sgpr70 = V_READLANE_B32 $vgpr0, 6
-    ; GCN64-FLATSCR: $sgpr71 = V_READLANE_B32 $vgpr0, 7
-    ; GCN64-FLATSCR: $sgpr72 = V_READLANE_B32 $vgpr0, 8
-    ; GCN64-FLATSCR: $sgpr73 = V_READLANE_B32 $vgpr0, 9
-    ; GCN64-FLATSCR: $sgpr74 = V_READLANE_B32 $vgpr0, 10
-    ; GCN64-FLATSCR: $sgpr75 = V_READLANE_B32 $vgpr0, 11
-    ; GCN64-FLATSCR: $sgpr76 = V_READLANE_B32 $vgpr0, 12
-    ; GCN64-FLATSCR: $sgpr77 = V_READLANE_B32 $vgpr0, 13
-    ; GCN64-FLATSCR: $sgpr78 = V_READLANE_B32 $vgpr0, 14
-    ; GCN64-FLATSCR: $sgpr79 = V_READLANE_B32 $vgpr0, 15
-    ; GCN64-FLATSCR: $sgpr80 = V_READLANE_B32 $vgpr0, 16
-    ; GCN64-FLATSCR: $sgpr81 = V_READLANE_B32 $vgpr0, 17
-    ; GCN64-FLATSCR: $sgpr82 = V_READLANE_B32 $vgpr0, 18
-    ; GCN64-FLATSCR: $sgpr83 = V_READLANE_B32 $vgpr0, 19
-    ; GCN64-FLATSCR: $sgpr84 = V_READLANE_B32 $vgpr0, 20
-    ; GCN64-FLATSCR: $sgpr85 = V_READLANE_B32 $vgpr0, 21
-    ; GCN64-FLATSCR: $sgpr86 = V_READLANE_B32 $vgpr0, 22
-    ; GCN64-FLATSCR: $sgpr87 = V_READLANE_B32 $vgpr0, 23
-    ; GCN64-FLATSCR: $sgpr88 = V_READLANE_B32 $vgpr0, 24
-    ; GCN64-FLATSCR: $sgpr89 = V_READLANE_B32 $vgpr0, 25
-    ; GCN64-FLATSCR: $sgpr90 = V_READLANE_B32 $vgpr0, 26
-    ; GCN64-FLATSCR: $sgpr91 = V_READLANE_B32 $vgpr0, 27
-    ; GCN64-FLATSCR: $sgpr92 = V_READLANE_B32 $vgpr0, 28
-    ; GCN64-FLATSCR: $sgpr93 = V_READLANE_B32 $vgpr0, 29
-    ; GCN64-FLATSCR: $sgpr94 = V_READLANE_B32 $vgpr0, 30
-    ; GCN64-FLATSCR: $sgpr95 = V_READLANE_B32 killed $vgpr0, 31
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
-    ; GCN64-FLATSCR: $sgpr2_sgpr3 = S_MOV_B64 $exec
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 1, implicit-def $vgpr0
-    ; GCN64-FLATSCR: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $sgpr9 = S_ADD_I32 $sgpr33, 4096, implicit-def $scc
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR killed $sgpr9, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.8, align 4096, addrspace 5)
-    ; GCN64-FLATSCR: $sgpr12 = V_READLANE_B32 killed $vgpr0, 0
-    ; GCN64-FLATSCR: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
+    ; GCN64-FLATSCR-NEXT: {{  $}}
+    ; GCN64-FLATSCR-NEXT: $sgpr33 = S_MOV_B32 0
+    ; GCN64-FLATSCR-NEXT: $flat_scr_lo = S_ADD_U32 $sgpr0, $sgpr11, implicit-def $scc
+    ; GCN64-FLATSCR-NEXT: $flat_scr_hi = S_ADDC_U32 $sgpr1, 0, implicit-def $scc, implicit $scc
+    ; GCN64-FLATSCR-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 1, implicit-def $vgpr0
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 4, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $sgpr12 = V_READLANE_B32 killed $vgpr0, 0
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
+    ; GCN64-FLATSCR-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 3, implicit-def $vgpr0
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 8, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.1, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13
+    ; GCN64-FLATSCR-NEXT: $sgpr13 = V_READLANE_B32 killed $vgpr0, 1
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
+    ; GCN64-FLATSCR-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 7, implicit-def $vgpr0
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 16, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.2, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14
+    ; GCN64-FLATSCR-NEXT: $sgpr13 = V_READLANE_B32 $vgpr0, 1
+    ; GCN64-FLATSCR-NEXT: $sgpr14 = V_READLANE_B32 killed $vgpr0, 2
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
+    ; GCN64-FLATSCR-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 15, implicit-def $vgpr0
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 28, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.3, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15
+    ; GCN64-FLATSCR-NEXT: $sgpr13 = V_READLANE_B32 $vgpr0, 1
+    ; GCN64-FLATSCR-NEXT: $sgpr14 = V_READLANE_B32 $vgpr0, 2
+    ; GCN64-FLATSCR-NEXT: $sgpr15 = V_READLANE_B32 killed $vgpr0, 3
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
+    ; GCN64-FLATSCR-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 31, implicit-def $vgpr0
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 44, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.4, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
+    ; GCN64-FLATSCR-NEXT: $sgpr13 = V_READLANE_B32 $vgpr0, 1
+    ; GCN64-FLATSCR-NEXT: $sgpr14 = V_READLANE_B32 $vgpr0, 2
+    ; GCN64-FLATSCR-NEXT: $sgpr15 = V_READLANE_B32 $vgpr0, 3
+    ; GCN64-FLATSCR-NEXT: $sgpr16 = V_READLANE_B32 killed $vgpr0, 4
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
+    ; GCN64-FLATSCR-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 255, implicit-def $vgpr0
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 64, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.5, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN64-FLATSCR-NEXT: $sgpr13 = V_READLANE_B32 $vgpr0, 1
+    ; GCN64-FLATSCR-NEXT: $sgpr14 = V_READLANE_B32 $vgpr0, 2
+    ; GCN64-FLATSCR-NEXT: $sgpr15 = V_READLANE_B32 $vgpr0, 3
+    ; GCN64-FLATSCR-NEXT: $sgpr16 = V_READLANE_B32 $vgpr0, 4
+    ; GCN64-FLATSCR-NEXT: $sgpr17 = V_READLANE_B32 $vgpr0, 5
+    ; GCN64-FLATSCR-NEXT: $sgpr18 = V_READLANE_B32 $vgpr0, 6
+    ; GCN64-FLATSCR-NEXT: $sgpr19 = V_READLANE_B32 killed $vgpr0, 7
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
+    ; GCN64-FLATSCR-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 65535, implicit-def $vgpr0
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 96, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.6, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $sgpr12 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-FLATSCR-NEXT: $sgpr13 = V_READLANE_B32 $vgpr0, 1
+    ; GCN64-FLATSCR-NEXT: $sgpr14 = V_READLANE_B32 $vgpr0, 2
+    ; GCN64-FLATSCR-NEXT: $sgpr15 = V_READLANE_B32 $vgpr0, 3
+    ; GCN64-FLATSCR-NEXT: $sgpr16 = V_READLANE_B32 $vgpr0, 4
+    ; GCN64-FLATSCR-NEXT: $sgpr17 = V_READLANE_B32 $vgpr0, 5
+    ; GCN64-FLATSCR-NEXT: $sgpr18 = V_READLANE_B32 $vgpr0, 6
+    ; GCN64-FLATSCR-NEXT: $sgpr19 = V_READLANE_B32 $vgpr0, 7
+    ; GCN64-FLATSCR-NEXT: $sgpr20 = V_READLANE_B32 $vgpr0, 8
+    ; GCN64-FLATSCR-NEXT: $sgpr21 = V_READLANE_B32 $vgpr0, 9
+    ; GCN64-FLATSCR-NEXT: $sgpr22 = V_READLANE_B32 $vgpr0, 10
+    ; GCN64-FLATSCR-NEXT: $sgpr23 = V_READLANE_B32 $vgpr0, 11
+    ; GCN64-FLATSCR-NEXT: $sgpr24 = V_READLANE_B32 $vgpr0, 12
+    ; GCN64-FLATSCR-NEXT: $sgpr25 = V_READLANE_B32 $vgpr0, 13
+    ; GCN64-FLATSCR-NEXT: $sgpr26 = V_READLANE_B32 $vgpr0, 14
+    ; GCN64-FLATSCR-NEXT: $sgpr27 = V_READLANE_B32 killed $vgpr0, 15
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
+    ; GCN64-FLATSCR-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 4294967295, implicit-def $vgpr0
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 160, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.7, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $sgpr64 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $sgpr65 = V_READLANE_B32 $vgpr0, 1
+    ; GCN64-FLATSCR-NEXT: $sgpr66 = V_READLANE_B32 $vgpr0, 2
+    ; GCN64-FLATSCR-NEXT: $sgpr67 = V_READLANE_B32 $vgpr0, 3
+    ; GCN64-FLATSCR-NEXT: $sgpr68 = V_READLANE_B32 $vgpr0, 4
+    ; GCN64-FLATSCR-NEXT: $sgpr69 = V_READLANE_B32 $vgpr0, 5
+    ; GCN64-FLATSCR-NEXT: $sgpr70 = V_READLANE_B32 $vgpr0, 6
+    ; GCN64-FLATSCR-NEXT: $sgpr71 = V_READLANE_B32 $vgpr0, 7
+    ; GCN64-FLATSCR-NEXT: $sgpr72 = V_READLANE_B32 $vgpr0, 8
+    ; GCN64-FLATSCR-NEXT: $sgpr73 = V_READLANE_B32 $vgpr0, 9
+    ; GCN64-FLATSCR-NEXT: $sgpr74 = V_READLANE_B32 $vgpr0, 10
+    ; GCN64-FLATSCR-NEXT: $sgpr75 = V_READLANE_B32 $vgpr0, 11
+    ; GCN64-FLATSCR-NEXT: $sgpr76 = V_READLANE_B32 $vgpr0, 12
+    ; GCN64-FLATSCR-NEXT: $sgpr77 = V_READLANE_B32 $vgpr0, 13
+    ; GCN64-FLATSCR-NEXT: $sgpr78 = V_READLANE_B32 $vgpr0, 14
+    ; GCN64-FLATSCR-NEXT: $sgpr79 = V_READLANE_B32 $vgpr0, 15
+    ; GCN64-FLATSCR-NEXT: $sgpr80 = V_READLANE_B32 $vgpr0, 16
+    ; GCN64-FLATSCR-NEXT: $sgpr81 = V_READLANE_B32 $vgpr0, 17
+    ; GCN64-FLATSCR-NEXT: $sgpr82 = V_READLANE_B32 $vgpr0, 18
+    ; GCN64-FLATSCR-NEXT: $sgpr83 = V_READLANE_B32 $vgpr0, 19
+    ; GCN64-FLATSCR-NEXT: $sgpr84 = V_READLANE_B32 $vgpr0, 20
+    ; GCN64-FLATSCR-NEXT: $sgpr85 = V_READLANE_B32 $vgpr0, 21
+    ; GCN64-FLATSCR-NEXT: $sgpr86 = V_READLANE_B32 $vgpr0, 22
+    ; GCN64-FLATSCR-NEXT: $sgpr87 = V_READLANE_B32 $vgpr0, 23
+    ; GCN64-FLATSCR-NEXT: $sgpr88 = V_READLANE_B32 $vgpr0, 24
+    ; GCN64-FLATSCR-NEXT: $sgpr89 = V_READLANE_B32 $vgpr0, 25
+    ; GCN64-FLATSCR-NEXT: $sgpr90 = V_READLANE_B32 $vgpr0, 26
+    ; GCN64-FLATSCR-NEXT: $sgpr91 = V_READLANE_B32 $vgpr0, 27
+    ; GCN64-FLATSCR-NEXT: $sgpr92 = V_READLANE_B32 $vgpr0, 28
+    ; GCN64-FLATSCR-NEXT: $sgpr93 = V_READLANE_B32 $vgpr0, 29
+    ; GCN64-FLATSCR-NEXT: $sgpr94 = V_READLANE_B32 $vgpr0, 30
+    ; GCN64-FLATSCR-NEXT: $sgpr95 = V_READLANE_B32 killed $vgpr0, 31
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
+    ; GCN64-FLATSCR-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 1, implicit-def $vgpr0
+    ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $sgpr9 = S_ADD_I32 $sgpr33, 4096, implicit-def $scc
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR killed $sgpr9, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.8, align 4096, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $sgpr12 = V_READLANE_B32 killed $vgpr0, 0
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
+    ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 killed $sgpr2_sgpr3, implicit killed $vgpr0
     renamable $sgpr12 = SI_SPILL_S32_RESTORE %stack.0, implicit $exec, implicit $sgpr96_sgpr97_sgpr98_sgpr99, implicit $sgpr32
 
     renamable $sgpr12_sgpr13 = SI_SPILL_S64_RESTORE %stack.1, implicit $exec, implicit $sgpr96_sgpr97_sgpr98_sgpr99, implicit $sgpr32


        


More information about the llvm-commits mailing list