brintos

brintos / llvm-project-archived public Read only

0
0
Text · 63.8 KiB · 9189f4f Raw
541 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py2; RUN: llc -mtriple=amdgcn -mcpu=gfx908 -amdgpu-atomic-optimizer-strategy=None -stop-after=amdgpu-isel -o - %s | FileCheck -check-prefix=GCN %s3; RUN: llc -mtriple=amdgcn -mcpu=gfx908 -enable-new-pm -amdgpu-atomic-optimizer-strategy=None -stop-after=amdgpu-isel -o - %s | FileCheck -check-prefix=GCN %s4 5define amdgpu_cs void @mmo_offsets0(ptr addrspace(6) inreg noalias align(16) dereferenceable(18446744073709551615) %arg0, i32 %arg1) {6  ; GCN-LABEL: name: mmo_offsets07  ; GCN: bb.0.bb.0:8  ; GCN-NEXT:   liveins: $sgpr0, $vgpr09  ; GCN-NEXT: {{  $}}10  ; GCN-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr011  ; GCN-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr012  ; GCN-NEXT:   [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 013  ; GCN-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[S_MOV_B32_]], %subreg.sub114  ; GCN-NEXT:   [[S_LOAD_DWORDX4_IMM:%[0-9]+]]:sgpr_128 = S_LOAD_DWORDX4_IMM [[REG_SEQUENCE]], 0, 0 :: (dereferenceable invariant load (s128) from %ir.arg0, addrspace 6)15  ; GCN-NEXT:   [[S_LOAD_DWORDX2_IMM:%[0-9]+]]:sreg_64_xexec = S_LOAD_DWORDX2_IMM [[REG_SEQUENCE]], 0, 0 :: (dereferenceable invariant load (s64) from %ir.arg0, align 16, addrspace 6)16  ; GCN-NEXT:   [[S_LOAD_DWORDX2_IMM1:%[0-9]+]]:sreg_64_xexec = S_LOAD_DWORDX2_IMM [[REG_SEQUENCE]], 8, 0 :: (dereferenceable invariant load (s64) from %ir.arg0 + 8, basealign 16, addrspace 6)17  ; GCN-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY [[S_LOAD_DWORDX2_IMM1]].sub118  ; GCN-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY [[S_LOAD_DWORDX2_IMM1]].sub019  ; GCN-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY [[S_LOAD_DWORDX2_IMM]].sub120  ; GCN-NEXT:   [[COPY5:%[0-9]+]]:sreg_32 = COPY [[S_LOAD_DWORDX2_IMM]].sub021  ; GCN-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE killed [[COPY5]], %subreg.sub0, killed [[COPY4]], %subreg.sub1, killed [[COPY3]], %subreg.sub2, killed [[COPY2]], %subreg.sub322  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_OFFSET:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_OFFSET [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 128, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)23  ; GCN-NEXT:   [[S_MOV_B32_1:%[0-9]+]]:sreg_32 = S_MOV_B32 6424  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_OFFSET1:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_OFFSET [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_1]], 64, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)25  ; GCN-NEXT:   [[S_MOV_B32_2:%[0-9]+]]:sreg_32 = S_MOV_B32 12826  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_OFFSET2:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_OFFSET [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_2]], 0, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)27  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_OFFEN:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_OFFEN [[COPY]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_2]], 0, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)28  ; GCN-NEXT:   [[COPY6:%[0-9]+]]:sreg_32 = COPY [[COPY]]29  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_OFFSET3:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_OFFSET [[S_LOAD_DWORDX4_IMM]], [[COPY6]], 128, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)30  ; GCN-NEXT:   INLINEASM &"", 1 /* sideeffect attdialect */31  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_OFFSET4:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_OFFSET [[REG_SEQUENCE1]], [[S_MOV_B32_]], 128, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)32  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_OFFSET5:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_OFFSET [[REG_SEQUENCE1]], [[S_MOV_B32_1]], 64, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)33  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_OFFSET6:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_OFFSET [[REG_SEQUENCE1]], [[S_MOV_B32_2]], 0, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)34  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_OFFEN1:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_OFFEN [[COPY]], [[REG_SEQUENCE1]], [[S_MOV_B32_2]], 0, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)35  ; GCN-NEXT:   [[COPY7:%[0-9]+]]:sreg_32 = COPY [[COPY]]36  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_OFFSET7:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_OFFSET [[REG_SEQUENCE1]], [[COPY7]], 128, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)37  ; GCN-NEXT:   INLINEASM &"", 1 /* sideeffect attdialect */38  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_OFFSET:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_OFFSET [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 144, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)39  ; GCN-NEXT:   [[S_MOV_B32_3:%[0-9]+]]:sreg_32 = S_MOV_B32 7240  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_OFFSET1:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_OFFSET [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_3]], 72, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)41  ; GCN-NEXT:   [[S_MOV_B32_4:%[0-9]+]]:sreg_32 = S_MOV_B32 14442  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_OFFSET2:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_OFFSET [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_4]], 0, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)43  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_OFFEN:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_OFFEN [[COPY]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_4]], 0, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)44  ; GCN-NEXT:   [[COPY8:%[0-9]+]]:sreg_32 = COPY [[COPY]]45  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_OFFSET3:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_OFFSET [[S_LOAD_DWORDX4_IMM]], [[COPY8]], 144, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)46  ; GCN-NEXT:   INLINEASM &"", 1 /* sideeffect attdialect */47  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_OFFSET4:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_OFFSET [[REG_SEQUENCE1]], [[S_MOV_B32_]], 144, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)48  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_OFFSET5:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_OFFSET [[REG_SEQUENCE1]], [[S_MOV_B32_3]], 72, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)49  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_OFFSET6:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_OFFSET [[REG_SEQUENCE1]], [[S_MOV_B32_4]], 0, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)50  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_OFFEN1:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_OFFEN [[COPY]], [[REG_SEQUENCE1]], [[S_MOV_B32_4]], 0, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)51  ; GCN-NEXT:   [[COPY9:%[0-9]+]]:sreg_32 = COPY [[COPY]]52  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_OFFSET7:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_OFFSET [[REG_SEQUENCE1]], [[COPY9]], 144, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)53  ; GCN-NEXT:   INLINEASM &"", 1 /* sideeffect attdialect */54  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_OFFSET [[COPY]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 160, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)55  ; GCN-NEXT:   [[S_MOV_B32_5:%[0-9]+]]:sreg_32 = S_MOV_B32 8056  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_OFFSET [[COPY]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_5]], 80, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)57  ; GCN-NEXT:   [[S_MOV_B32_6:%[0-9]+]]:sreg_32 = S_MOV_B32 16058  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_OFFSET [[COPY]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)59  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_OFFEN [[COPY]], [[COPY]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)60  ; GCN-NEXT:   [[COPY10:%[0-9]+]]:sreg_32 = COPY [[COPY]]61  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_OFFSET [[COPY]], [[S_LOAD_DWORDX4_IMM]], [[COPY10]], 160, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)62  ; GCN-NEXT:   INLINEASM &"", 1 /* sideeffect attdialect */63  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_OFFSET [[COPY]], [[REG_SEQUENCE1]], [[S_MOV_B32_]], 160, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)64  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_OFFSET [[COPY]], [[REG_SEQUENCE1]], [[S_MOV_B32_5]], 80, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)65  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_OFFSET [[COPY]], [[REG_SEQUENCE1]], [[S_MOV_B32_6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)66  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_OFFEN [[COPY]], [[COPY]], [[REG_SEQUENCE1]], [[S_MOV_B32_6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)67  ; GCN-NEXT:   [[COPY11:%[0-9]+]]:sreg_32 = COPY [[COPY]]68  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_OFFSET [[COPY]], [[REG_SEQUENCE1]], [[COPY11]], 160, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)69  ; GCN-NEXT:   INLINEASM &"", 1 /* sideeffect attdialect */70  ; GCN-NEXT:   [[REG_SEQUENCE2:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY]], %subreg.sub171  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_OFFSET [[REG_SEQUENCE2]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 176, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)72  ; GCN-NEXT:   [[S_MOV_B32_7:%[0-9]+]]:sreg_32 = S_MOV_B32 8873  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_OFFSET [[REG_SEQUENCE2]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_7]], 88, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)74  ; GCN-NEXT:   [[S_MOV_B32_8:%[0-9]+]]:sreg_32 = S_MOV_B32 17675  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_OFFSET [[REG_SEQUENCE2]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_8]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)76  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_OFFEN [[REG_SEQUENCE2]], [[COPY]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_8]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)77  ; GCN-NEXT:   [[COPY12:%[0-9]+]]:sreg_32 = COPY [[COPY]]78  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_OFFSET [[REG_SEQUENCE2]], [[S_LOAD_DWORDX4_IMM]], [[COPY12]], 176, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)79  ; GCN-NEXT:   INLINEASM &"", 1 /* sideeffect attdialect */80  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_OFFSET [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], [[S_MOV_B32_]], 176, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)81  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_OFFSET [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], [[S_MOV_B32_7]], 88, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)82  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_OFFSET [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], [[S_MOV_B32_8]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)83  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_OFFEN [[REG_SEQUENCE2]], [[COPY]], [[REG_SEQUENCE1]], [[S_MOV_B32_8]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)84  ; GCN-NEXT:   [[COPY13:%[0-9]+]]:sreg_32 = COPY [[COPY]]85  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_OFFSET [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], [[COPY13]], 176, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)86  ; GCN-NEXT:   INLINEASM &"", 1 /* sideeffect attdialect */87  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_OFFSET_exact killed [[BUFFER_LOAD_DWORDX4_OFFSET]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 192, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)88  ; GCN-NEXT:   [[S_MOV_B32_9:%[0-9]+]]:sreg_32 = S_MOV_B32 9689  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_OFFSET_exact killed [[BUFFER_LOAD_DWORDX4_OFFSET1]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_9]], 96, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)90  ; GCN-NEXT:   [[S_MOV_B32_10:%[0-9]+]]:sreg_32 = S_MOV_B32 19291  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_OFFSET_exact killed [[BUFFER_LOAD_DWORDX4_OFFSET2]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_10]], 0, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)92  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_OFFEN_exact killed [[BUFFER_LOAD_DWORDX4_OFFEN]], [[COPY]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_10]], 0, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)93  ; GCN-NEXT:   [[COPY14:%[0-9]+]]:sreg_32 = COPY [[COPY]]94  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_OFFSET_exact killed [[BUFFER_LOAD_DWORDX4_OFFSET3]], [[S_LOAD_DWORDX4_IMM]], [[COPY14]], 192, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)95  ; GCN-NEXT:   INLINEASM &"", 1 /* sideeffect attdialect */96  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_OFFSET_exact killed [[BUFFER_LOAD_DWORDX4_OFFSET4]], [[REG_SEQUENCE1]], [[S_MOV_B32_]], 192, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)97  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_OFFSET_exact killed [[BUFFER_LOAD_DWORDX4_OFFSET5]], [[REG_SEQUENCE1]], [[S_MOV_B32_9]], 96, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)98  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_OFFSET_exact killed [[BUFFER_LOAD_DWORDX4_OFFSET6]], [[REG_SEQUENCE1]], [[S_MOV_B32_10]], 0, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)99  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_OFFEN_exact killed [[BUFFER_LOAD_DWORDX4_OFFEN1]], [[COPY]], [[REG_SEQUENCE1]], [[S_MOV_B32_10]], 0, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)100  ; GCN-NEXT:   [[COPY15:%[0-9]+]]:sreg_32 = COPY [[COPY]]101  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_OFFSET_exact killed [[BUFFER_LOAD_DWORDX4_OFFSET7]], [[REG_SEQUENCE1]], [[COPY15]], 192, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)102  ; GCN-NEXT:   INLINEASM &"", 1 /* sideeffect attdialect */103  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_OFFSET_exact killed [[BUFFER_LOAD_FORMAT_XYZW_OFFSET]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 208, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)104  ; GCN-NEXT:   [[S_MOV_B32_11:%[0-9]+]]:sreg_32 = S_MOV_B32 104105  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_OFFSET_exact killed [[BUFFER_LOAD_FORMAT_XYZW_OFFSET1]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_11]], 104, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)106  ; GCN-NEXT:   [[S_MOV_B32_12:%[0-9]+]]:sreg_32 = S_MOV_B32 208107  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_OFFSET_exact killed [[BUFFER_LOAD_FORMAT_XYZW_OFFSET2]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_12]], 0, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)108  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_OFFEN_exact killed [[BUFFER_LOAD_FORMAT_XYZW_OFFEN]], [[COPY]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_12]], 0, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)109  ; GCN-NEXT:   [[COPY16:%[0-9]+]]:sreg_32 = COPY [[COPY]]110  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_OFFSET_exact killed [[BUFFER_LOAD_FORMAT_XYZW_OFFSET3]], [[S_LOAD_DWORDX4_IMM]], [[COPY16]], 208, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)111  ; GCN-NEXT:   INLINEASM &"", 1 /* sideeffect attdialect */112  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_OFFSET_exact killed [[BUFFER_LOAD_FORMAT_XYZW_OFFSET4]], [[REG_SEQUENCE1]], [[S_MOV_B32_]], 208, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)113  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_OFFSET_exact killed [[BUFFER_LOAD_FORMAT_XYZW_OFFSET5]], [[REG_SEQUENCE1]], [[S_MOV_B32_11]], 104, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)114  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_OFFSET_exact killed [[BUFFER_LOAD_FORMAT_XYZW_OFFSET6]], [[REG_SEQUENCE1]], [[S_MOV_B32_12]], 0, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)115  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_OFFEN_exact killed [[BUFFER_LOAD_FORMAT_XYZW_OFFEN1]], [[COPY]], [[REG_SEQUENCE1]], [[S_MOV_B32_12]], 0, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)116  ; GCN-NEXT:   [[COPY17:%[0-9]+]]:sreg_32 = COPY [[COPY]]117  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_OFFSET_exact killed [[BUFFER_LOAD_FORMAT_XYZW_OFFSET7]], [[REG_SEQUENCE1]], [[COPY17]], 208, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)118  ; GCN-NEXT:   INLINEASM &"", 1 /* sideeffect attdialect */119  ; GCN-NEXT:   [[COPY18:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]120  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_IDXEN:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_IDXEN [[COPY18]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 224, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)121  ; GCN-NEXT:   [[S_MOV_B32_13:%[0-9]+]]:sreg_32 = S_MOV_B32 112122  ; GCN-NEXT:   [[COPY19:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]123  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_IDXEN1:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_IDXEN [[COPY19]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_13]], 112, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)124  ; GCN-NEXT:   [[S_MOV_B32_14:%[0-9]+]]:sreg_32 = S_MOV_B32 224125  ; GCN-NEXT:   [[COPY20:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]126  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_IDXEN2:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_IDXEN [[COPY20]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_14]], 0, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)127  ; GCN-NEXT:   [[REG_SEQUENCE3:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[S_MOV_B32_]], %subreg.sub0, [[COPY]], %subreg.sub1128  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_BOTHEN:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_BOTHEN [[REG_SEQUENCE3]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_14]], 0, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)129  ; GCN-NEXT:   [[COPY21:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]130  ; GCN-NEXT:   [[COPY22:%[0-9]+]]:sreg_32 = COPY [[COPY]]131  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_IDXEN3:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_IDXEN [[COPY21]], [[S_LOAD_DWORDX4_IMM]], [[COPY22]], 224, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)132  ; GCN-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 1, implicit $exec133  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_IDXEN4:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_IDXEN [[V_MOV_B32_e32_]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 224, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)134  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_IDXEN5:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_IDXEN [[COPY]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 224, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)135  ; GCN-NEXT:   INLINEASM &"", 1 /* sideeffect attdialect */136  ; GCN-NEXT:   [[COPY23:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]137  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_IDXEN6:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_IDXEN [[COPY23]], [[REG_SEQUENCE1]], [[S_MOV_B32_]], 224, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)138  ; GCN-NEXT:   [[COPY24:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]139  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_IDXEN7:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_IDXEN [[COPY24]], [[REG_SEQUENCE1]], [[S_MOV_B32_13]], 112, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)140  ; GCN-NEXT:   [[COPY25:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]141  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_IDXEN8:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_IDXEN [[COPY25]], [[REG_SEQUENCE1]], [[S_MOV_B32_14]], 0, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)142  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_BOTHEN1:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_BOTHEN [[REG_SEQUENCE3]], [[REG_SEQUENCE1]], [[S_MOV_B32_14]], 0, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)143  ; GCN-NEXT:   [[COPY26:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]144  ; GCN-NEXT:   [[COPY27:%[0-9]+]]:sreg_32 = COPY [[COPY]]145  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_IDXEN9:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_IDXEN [[COPY26]], [[REG_SEQUENCE1]], [[COPY27]], 224, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)146  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_IDXEN10:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_IDXEN [[V_MOV_B32_e32_]], [[REG_SEQUENCE1]], [[S_MOV_B32_]], 224, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)147  ; GCN-NEXT:   [[BUFFER_LOAD_DWORDX4_IDXEN11:%[0-9]+]]:vreg_128 = BUFFER_LOAD_DWORDX4_IDXEN [[COPY]], [[REG_SEQUENCE1]], [[S_MOV_B32_]], 224, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)148  ; GCN-NEXT:   INLINEASM &"", 1 /* sideeffect attdialect */149  ; GCN-NEXT:   [[COPY28:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]150  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_IDXEN:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_IDXEN [[COPY28]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 240, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)151  ; GCN-NEXT:   [[S_MOV_B32_15:%[0-9]+]]:sreg_32 = S_MOV_B32 120152  ; GCN-NEXT:   [[COPY29:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]153  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_IDXEN1:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_IDXEN [[COPY29]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_15]], 120, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)154  ; GCN-NEXT:   [[S_MOV_B32_16:%[0-9]+]]:sreg_32 = S_MOV_B32 240155  ; GCN-NEXT:   [[COPY30:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]156  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_IDXEN2:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_IDXEN [[COPY30]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_16]], 0, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)157  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_BOTHEN:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_BOTHEN [[REG_SEQUENCE3]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_16]], 0, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)158  ; GCN-NEXT:   [[COPY31:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]159  ; GCN-NEXT:   [[COPY32:%[0-9]+]]:sreg_32 = COPY [[COPY]]160  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_IDXEN3:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_IDXEN [[COPY31]], [[S_LOAD_DWORDX4_IMM]], [[COPY32]], 240, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)161  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_IDXEN4:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_IDXEN [[V_MOV_B32_e32_]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 240, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)162  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_IDXEN5:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_IDXEN [[COPY]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 240, 0, 0, implicit $exec :: (dereferenceable load (s128), align 1, addrspace 8)163  ; GCN-NEXT:   INLINEASM &"", 1 /* sideeffect attdialect */164  ; GCN-NEXT:   [[COPY33:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]165  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_IDXEN6:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_IDXEN [[COPY33]], [[REG_SEQUENCE1]], [[S_MOV_B32_]], 240, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)166  ; GCN-NEXT:   [[COPY34:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]167  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_IDXEN7:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_IDXEN [[COPY34]], [[REG_SEQUENCE1]], [[S_MOV_B32_15]], 120, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)168  ; GCN-NEXT:   [[COPY35:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]169  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_IDXEN8:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_IDXEN [[COPY35]], [[REG_SEQUENCE1]], [[S_MOV_B32_16]], 0, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)170  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_BOTHEN1:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_BOTHEN [[REG_SEQUENCE3]], [[REG_SEQUENCE1]], [[S_MOV_B32_16]], 0, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)171  ; GCN-NEXT:   [[COPY36:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]172  ; GCN-NEXT:   [[COPY37:%[0-9]+]]:sreg_32 = COPY [[COPY]]173  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_IDXEN9:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_IDXEN [[COPY36]], [[REG_SEQUENCE1]], [[COPY37]], 240, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)174  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_IDXEN10:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_IDXEN [[V_MOV_B32_e32_]], [[REG_SEQUENCE1]], [[S_MOV_B32_]], 240, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)175  ; GCN-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_IDXEN11:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_IDXEN [[COPY]], [[REG_SEQUENCE1]], [[S_MOV_B32_]], 240, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.tmp1, align 1, addrspace 8)176  ; GCN-NEXT:   INLINEASM &"", 1 /* sideeffect attdialect */177  ; GCN-NEXT:   [[COPY38:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]178  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_IDXEN [[COPY]], [[COPY38]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 256, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)179  ; GCN-NEXT:   [[COPY39:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]180  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_IDXEN [[COPY]], [[COPY39]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_2]], 128, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)181  ; GCN-NEXT:   [[S_MOV_B32_17:%[0-9]+]]:sreg_32 = S_MOV_B32 256182  ; GCN-NEXT:   [[COPY40:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]183  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_IDXEN [[COPY]], [[COPY40]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_17]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)184  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_BOTHEN [[COPY]], [[REG_SEQUENCE3]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_17]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)185  ; GCN-NEXT:   [[COPY41:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]186  ; GCN-NEXT:   [[COPY42:%[0-9]+]]:sreg_32 = COPY [[COPY]]187  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_IDXEN [[COPY]], [[COPY41]], [[S_LOAD_DWORDX4_IMM]], [[COPY42]], 256, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)188  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_IDXEN [[COPY]], [[V_MOV_B32_e32_]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 256, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)189  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_IDXEN [[COPY]], [[COPY]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 256, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)190  ; GCN-NEXT:   INLINEASM &"", 1 /* sideeffect attdialect */191  ; GCN-NEXT:   [[COPY43:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]192  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_IDXEN [[COPY]], [[COPY43]], [[REG_SEQUENCE1]], [[S_MOV_B32_]], 256, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)193  ; GCN-NEXT:   [[COPY44:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]194  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_IDXEN [[COPY]], [[COPY44]], [[REG_SEQUENCE1]], [[S_MOV_B32_2]], 128, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)195  ; GCN-NEXT:   [[COPY45:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]196  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_IDXEN [[COPY]], [[COPY45]], [[REG_SEQUENCE1]], [[S_MOV_B32_17]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)197  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_BOTHEN [[COPY]], [[REG_SEQUENCE3]], [[REG_SEQUENCE1]], [[S_MOV_B32_17]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)198  ; GCN-NEXT:   [[COPY46:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]199  ; GCN-NEXT:   [[COPY47:%[0-9]+]]:sreg_32 = COPY [[COPY]]200  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_IDXEN [[COPY]], [[COPY46]], [[REG_SEQUENCE1]], [[COPY47]], 256, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)201  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_IDXEN [[COPY]], [[V_MOV_B32_e32_]], [[REG_SEQUENCE1]], [[S_MOV_B32_]], 256, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)202  ; GCN-NEXT:   BUFFER_ATOMIC_ADD_IDXEN [[COPY]], [[COPY]], [[REG_SEQUENCE1]], [[S_MOV_B32_]], 256, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)203  ; GCN-NEXT:   INLINEASM &"", 1 /* sideeffect attdialect */204  ; GCN-NEXT:   [[COPY48:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]205  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_IDXEN [[REG_SEQUENCE2]], [[COPY48]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 272, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)206  ; GCN-NEXT:   [[S_MOV_B32_18:%[0-9]+]]:sreg_32 = S_MOV_B32 136207  ; GCN-NEXT:   [[COPY49:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]208  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_IDXEN [[REG_SEQUENCE2]], [[COPY49]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_18]], 136, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)209  ; GCN-NEXT:   [[S_MOV_B32_19:%[0-9]+]]:sreg_32 = S_MOV_B32 272210  ; GCN-NEXT:   [[COPY50:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]211  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_IDXEN [[REG_SEQUENCE2]], [[COPY50]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_19]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)212  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_BOTHEN [[REG_SEQUENCE2]], [[REG_SEQUENCE3]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_19]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)213  ; GCN-NEXT:   [[COPY51:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]214  ; GCN-NEXT:   [[COPY52:%[0-9]+]]:sreg_32 = COPY [[COPY]]215  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_IDXEN [[REG_SEQUENCE2]], [[COPY51]], [[S_LOAD_DWORDX4_IMM]], [[COPY52]], 272, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)216  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_IDXEN [[REG_SEQUENCE2]], [[V_MOV_B32_e32_]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 272, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)217  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_IDXEN [[REG_SEQUENCE2]], [[COPY]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 272, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 8)218  ; GCN-NEXT:   INLINEASM &"", 1 /* sideeffect attdialect */219  ; GCN-NEXT:   [[COPY53:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]220  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_IDXEN [[REG_SEQUENCE2]], [[COPY53]], [[REG_SEQUENCE1]], [[S_MOV_B32_]], 272, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)221  ; GCN-NEXT:   [[COPY54:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]222  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_IDXEN [[REG_SEQUENCE2]], [[COPY54]], [[REG_SEQUENCE1]], [[S_MOV_B32_18]], 136, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)223  ; GCN-NEXT:   [[COPY55:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]224  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_IDXEN [[REG_SEQUENCE2]], [[COPY55]], [[REG_SEQUENCE1]], [[S_MOV_B32_19]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)225  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_BOTHEN [[REG_SEQUENCE2]], [[REG_SEQUENCE3]], [[REG_SEQUENCE1]], [[S_MOV_B32_19]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)226  ; GCN-NEXT:   [[COPY56:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]227  ; GCN-NEXT:   [[COPY57:%[0-9]+]]:sreg_32 = COPY [[COPY]]228  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_IDXEN [[REG_SEQUENCE2]], [[COPY56]], [[REG_SEQUENCE1]], [[COPY57]], 272, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)229  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_IDXEN [[REG_SEQUENCE2]], [[V_MOV_B32_e32_]], [[REG_SEQUENCE1]], [[S_MOV_B32_]], 272, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)230  ; GCN-NEXT:   BUFFER_ATOMIC_CMPSWAP_IDXEN [[REG_SEQUENCE2]], [[COPY]], [[REG_SEQUENCE1]], [[S_MOV_B32_]], 272, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.tmp1, align 1, addrspace 8)231  ; GCN-NEXT:   INLINEASM &"", 1 /* sideeffect attdialect */232  ; GCN-NEXT:   [[COPY58:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]233  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_IDXEN_exact killed [[BUFFER_LOAD_DWORDX4_IDXEN]], [[COPY58]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 288, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)234  ; GCN-NEXT:   [[COPY59:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]235  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_IDXEN_exact killed [[BUFFER_LOAD_DWORDX4_IDXEN1]], [[COPY59]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_4]], 144, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)236  ; GCN-NEXT:   [[S_MOV_B32_20:%[0-9]+]]:sreg_32 = S_MOV_B32 288237  ; GCN-NEXT:   [[COPY60:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]238  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_IDXEN_exact killed [[BUFFER_LOAD_DWORDX4_IDXEN2]], [[COPY60]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_20]], 0, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)239  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_BOTHEN_exact killed [[BUFFER_LOAD_DWORDX4_BOTHEN]], [[REG_SEQUENCE3]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_20]], 0, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)240  ; GCN-NEXT:   [[COPY61:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]241  ; GCN-NEXT:   [[COPY62:%[0-9]+]]:sreg_32 = COPY [[COPY]]242  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_IDXEN_exact killed [[BUFFER_LOAD_DWORDX4_IDXEN3]], [[COPY61]], [[S_LOAD_DWORDX4_IMM]], [[COPY62]], 288, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)243  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_IDXEN_exact killed [[BUFFER_LOAD_DWORDX4_IDXEN4]], [[V_MOV_B32_e32_]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 288, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)244  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_IDXEN_exact killed [[BUFFER_LOAD_DWORDX4_IDXEN5]], [[COPY]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 288, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)245  ; GCN-NEXT:   INLINEASM &"", 1 /* sideeffect attdialect */246  ; GCN-NEXT:   [[COPY63:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]247  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_IDXEN_exact killed [[BUFFER_LOAD_DWORDX4_IDXEN6]], [[COPY63]], [[REG_SEQUENCE1]], [[S_MOV_B32_]], 288, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)248  ; GCN-NEXT:   [[COPY64:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]249  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_IDXEN_exact killed [[BUFFER_LOAD_DWORDX4_IDXEN7]], [[COPY64]], [[REG_SEQUENCE1]], [[S_MOV_B32_4]], 144, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)250  ; GCN-NEXT:   [[COPY65:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]251  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_IDXEN_exact killed [[BUFFER_LOAD_DWORDX4_IDXEN8]], [[COPY65]], [[REG_SEQUENCE1]], [[S_MOV_B32_20]], 0, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)252  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_BOTHEN_exact killed [[BUFFER_LOAD_DWORDX4_BOTHEN1]], [[REG_SEQUENCE3]], [[REG_SEQUENCE1]], [[S_MOV_B32_20]], 0, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)253  ; GCN-NEXT:   [[COPY66:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]254  ; GCN-NEXT:   [[COPY67:%[0-9]+]]:sreg_32 = COPY [[COPY]]255  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_IDXEN_exact killed [[BUFFER_LOAD_DWORDX4_IDXEN9]], [[COPY66]], [[REG_SEQUENCE1]], [[COPY67]], 288, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)256  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_IDXEN_exact killed [[BUFFER_LOAD_DWORDX4_IDXEN10]], [[V_MOV_B32_e32_]], [[REG_SEQUENCE1]], [[S_MOV_B32_]], 288, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)257  ; GCN-NEXT:   BUFFER_STORE_DWORDX4_IDXEN_exact killed [[BUFFER_LOAD_DWORDX4_IDXEN11]], [[COPY]], [[REG_SEQUENCE1]], [[S_MOV_B32_]], 288, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)258  ; GCN-NEXT:   INLINEASM &"", 1 /* sideeffect attdialect */259  ; GCN-NEXT:   [[COPY68:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]260  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_IDXEN_exact killed [[BUFFER_LOAD_FORMAT_XYZW_IDXEN]], [[COPY68]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 304, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)261  ; GCN-NEXT:   [[S_MOV_B32_21:%[0-9]+]]:sreg_32 = S_MOV_B32 152262  ; GCN-NEXT:   [[COPY69:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]263  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_IDXEN_exact killed [[BUFFER_LOAD_FORMAT_XYZW_IDXEN1]], [[COPY69]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_21]], 152, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)264  ; GCN-NEXT:   [[S_MOV_B32_22:%[0-9]+]]:sreg_32 = S_MOV_B32 304265  ; GCN-NEXT:   [[COPY70:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]266  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_IDXEN_exact killed [[BUFFER_LOAD_FORMAT_XYZW_IDXEN2]], [[COPY70]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_22]], 0, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)267  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_BOTHEN_exact killed [[BUFFER_LOAD_FORMAT_XYZW_BOTHEN]], [[REG_SEQUENCE3]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_22]], 0, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)268  ; GCN-NEXT:   [[COPY71:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]269  ; GCN-NEXT:   [[COPY72:%[0-9]+]]:sreg_32 = COPY [[COPY]]270  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_IDXEN_exact killed [[BUFFER_LOAD_FORMAT_XYZW_IDXEN3]], [[COPY71]], [[S_LOAD_DWORDX4_IMM]], [[COPY72]], 304, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)271  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_IDXEN_exact killed [[BUFFER_LOAD_FORMAT_XYZW_IDXEN4]], [[V_MOV_B32_e32_]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 304, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)272  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_IDXEN_exact killed [[BUFFER_LOAD_FORMAT_XYZW_IDXEN5]], [[COPY]], [[S_LOAD_DWORDX4_IMM]], [[S_MOV_B32_]], 304, 0, 0, implicit $exec :: (dereferenceable store (s128), align 1, addrspace 8)273  ; GCN-NEXT:   [[COPY73:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]274  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_IDXEN_exact killed [[BUFFER_LOAD_FORMAT_XYZW_IDXEN6]], [[COPY73]], [[REG_SEQUENCE1]], [[S_MOV_B32_]], 304, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)275  ; GCN-NEXT:   [[COPY74:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]276  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_IDXEN_exact killed [[BUFFER_LOAD_FORMAT_XYZW_IDXEN7]], [[COPY74]], [[REG_SEQUENCE1]], [[S_MOV_B32_21]], 152, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)277  ; GCN-NEXT:   [[COPY75:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]278  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_IDXEN_exact killed [[BUFFER_LOAD_FORMAT_XYZW_IDXEN8]], [[COPY75]], [[REG_SEQUENCE1]], [[S_MOV_B32_22]], 0, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)279  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_BOTHEN_exact killed [[BUFFER_LOAD_FORMAT_XYZW_BOTHEN1]], [[REG_SEQUENCE3]], [[REG_SEQUENCE1]], [[S_MOV_B32_22]], 0, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)280  ; GCN-NEXT:   [[COPY76:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]281  ; GCN-NEXT:   [[COPY77:%[0-9]+]]:sreg_32 = COPY [[COPY]]282  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_IDXEN_exact killed [[BUFFER_LOAD_FORMAT_XYZW_IDXEN9]], [[COPY76]], [[REG_SEQUENCE1]], [[COPY77]], 304, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)283  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_IDXEN_exact killed [[BUFFER_LOAD_FORMAT_XYZW_IDXEN10]], [[V_MOV_B32_e32_]], [[REG_SEQUENCE1]], [[S_MOV_B32_]], 304, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)284  ; GCN-NEXT:   BUFFER_STORE_FORMAT_XYZW_IDXEN_exact killed [[BUFFER_LOAD_FORMAT_XYZW_IDXEN11]], [[COPY]], [[REG_SEQUENCE1]], [[S_MOV_B32_]], 304, 0, 0, implicit $exec :: (dereferenceable store (s128) into %ir.tmp1, align 1, addrspace 8)285  ; GCN-NEXT:   S_ENDPGM 0286bb.0:287  %tmp0 = load <4 x i32>, ptr addrspace(6) %arg0, align 16, !invariant.load !0288  %tmp1 = load ptr addrspace(8), ptr addrspace(6) %arg0, align 16, !invariant.load !0289 290  %raw_buffer0 = call <4 x float> @llvm.amdgcn.raw.buffer.load.v4f32(<4 x i32> %tmp0, i32 128, i32 0, i32 0) #0291  %raw_buffer1 = call <4 x float> @llvm.amdgcn.raw.buffer.load.v4f32(<4 x i32> %tmp0, i32 64, i32 64, i32 0) #0292  %raw_buffer2 = call <4 x float> @llvm.amdgcn.raw.buffer.load.v4f32(<4 x i32> %tmp0, i32 0, i32 128, i32 0) #0293  %raw_buffer3 = call <4 x float> @llvm.amdgcn.raw.buffer.load.v4f32(<4 x i32> %tmp0, i32 %arg1, i32 128, i32 0) #0294  %raw_buffer4 = call <4 x float> @llvm.amdgcn.raw.buffer.load.v4f32(<4 x i32> %tmp0, i32 128, i32 %arg1, i32 0) #0295 296  call void asm sideeffect "", "" ()297 298  %raw_ptr_buffer0 = call <4 x float> @llvm.amdgcn.raw.ptr.buffer.load.v4f32(ptr addrspace(8) %tmp1, i32 128, i32 0, i32 0) #3299  %raw_ptr_buffer1 = call <4 x float> @llvm.amdgcn.raw.ptr.buffer.load.v4f32(ptr addrspace(8) %tmp1, i32 64, i32 64, i32 0) #3300  %raw_ptr_buffer2 = call <4 x float> @llvm.amdgcn.raw.ptr.buffer.load.v4f32(ptr addrspace(8) %tmp1, i32 0, i32 128, i32 0) #3301  %raw_ptr_buffer3 = call <4 x float> @llvm.amdgcn.raw.ptr.buffer.load.v4f32(ptr addrspace(8) %tmp1, i32 %arg1, i32 128, i32 0) #3302  %raw_ptr_buffer4 = call <4 x float> @llvm.amdgcn.raw.ptr.buffer.load.v4f32(ptr addrspace(8) %tmp1, i32 128, i32 %arg1, i32 0) #3303 304  call void asm sideeffect "", "" ()305 306  %raw_buffer_format0 = call <4 x float> @llvm.amdgcn.raw.buffer.load.format.v4f32(<4 x i32> %tmp0, i32 144, i32 0, i32 0) #0307  %raw_buffer_format1 = call <4 x float> @llvm.amdgcn.raw.buffer.load.format.v4f32(<4 x i32> %tmp0, i32 72, i32 72, i32 0) #0308  %raw_buffer_format2 = call <4 x float> @llvm.amdgcn.raw.buffer.load.format.v4f32(<4 x i32> %tmp0, i32 0, i32 144, i32 0) #0309  %raw_buffer_format3 = call <4 x float> @llvm.amdgcn.raw.buffer.load.format.v4f32(<4 x i32> %tmp0, i32 %arg1, i32 144, i32 0) #0310  %raw_buffer_format4 = call <4 x float> @llvm.amdgcn.raw.buffer.load.format.v4f32(<4 x i32> %tmp0, i32 144, i32 %arg1, i32 0) #0311 312  call void asm sideeffect "", "" ()313 314  %raw_buffer_format_ptr0 = call <4 x float> @llvm.amdgcn.raw.ptr.buffer.load.format.v4f32(ptr addrspace(8) %tmp1, i32 144, i32 0, i32 0) #3315  %raw_buffer_format_ptr1 = call <4 x float> @llvm.amdgcn.raw.ptr.buffer.load.format.v4f32(ptr addrspace(8) %tmp1, i32 72, i32 72, i32 0) #3316  %raw_buffer_format_ptr2 = call <4 x float> @llvm.amdgcn.raw.ptr.buffer.load.format.v4f32(ptr addrspace(8) %tmp1, i32 0, i32 144, i32 0) #3317  %raw_buffer_format_ptr3 = call <4 x float> @llvm.amdgcn.raw.ptr.buffer.load.format.v4f32(ptr addrspace(8) %tmp1, i32 %arg1, i32 144, i32 0) #3318  %raw_buffer_format_ptr4 = call <4 x float> @llvm.amdgcn.raw.ptr.buffer.load.format.v4f32(ptr addrspace(8) %tmp1, i32 144, i32 %arg1, i32 0) #3319 320  call void asm sideeffect "", "" ()321 322  %raw_atomic_add0 = call i32 @llvm.amdgcn.raw.buffer.atomic.add.i32(i32 %arg1, <4 x i32> %tmp0, i32 160, i32 0, i32 0) #2323  %raw_atomic_add1 = call i32 @llvm.amdgcn.raw.buffer.atomic.add.i32(i32 %arg1, <4 x i32> %tmp0, i32 80, i32 80, i32 0) #2324  %raw_atomic_add2 = call i32 @llvm.amdgcn.raw.buffer.atomic.add.i32(i32 %arg1, <4 x i32> %tmp0, i32 0, i32 160, i32 0) #2325  %raw_atomic_add3 = call i32 @llvm.amdgcn.raw.buffer.atomic.add.i32(i32 %arg1, <4 x i32> %tmp0, i32 %arg1, i32 160, i32 0) #2326  %raw_atomic_add4 = call i32 @llvm.amdgcn.raw.buffer.atomic.add.i32(i32 %arg1, <4 x i32> %tmp0, i32 160, i32 %arg1, i32 0) #2327 328  call void asm sideeffect "", "" ()329 330  %raw_ptr_atomic_add0 = call i32 @llvm.amdgcn.raw.ptr.buffer.atomic.add.i32(i32 %arg1, ptr addrspace(8) %tmp1, i32 160, i32 0, i32 0) #5331  %raw_ptr_atomic_add1 = call i32 @llvm.amdgcn.raw.ptr.buffer.atomic.add.i32(i32 %arg1, ptr addrspace(8) %tmp1, i32 80, i32 80, i32 0) #5332  %raw_ptr_atomic_add2 = call i32 @llvm.amdgcn.raw.ptr.buffer.atomic.add.i32(i32 %arg1, ptr addrspace(8) %tmp1, i32 0, i32 160, i32 0) #5333  %raw_ptr_atomic_add3 = call i32 @llvm.amdgcn.raw.ptr.buffer.atomic.add.i32(i32 %arg1, ptr addrspace(8) %tmp1, i32 %arg1, i32 160, i32 0) #5334  %raw_ptr_atomic_add4 = call i32 @llvm.amdgcn.raw.ptr.buffer.atomic.add.i32(i32 %arg1, ptr addrspace(8) %tmp1, i32 160, i32 %arg1, i32 0) #5335 336  call void asm sideeffect "", "" ()337 338  %raw_atomic_cmpswap0 = call i32 @llvm.amdgcn.raw.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, <4 x i32> %tmp0, i32 176, i32 0, i32 0) #2339  %raw_atomic_cmpswap1 = call i32 @llvm.amdgcn.raw.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, <4 x i32> %tmp0, i32 88, i32 88, i32 0) #2340  %raw_atomic_cmpswap2 = call i32 @llvm.amdgcn.raw.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, <4 x i32> %tmp0, i32 0, i32 176, i32 0) #2341  %raw_atomic_cmpswap3 = call i32 @llvm.amdgcn.raw.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, <4 x i32> %tmp0, i32 %arg1, i32 176, i32 0) #2342  %raw_atomic_cmpswap4 = call i32 @llvm.amdgcn.raw.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, <4 x i32> %tmp0, i32 176, i32 %arg1, i32 0) #2343 344  call void asm sideeffect "", "" ()345 346  %raw_ptr_atomic_cmpswap0 = call i32 @llvm.amdgcn.raw.ptr.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, ptr addrspace(8) %tmp1, i32 176, i32 0, i32 0) #5347  %raw_ptr_atomic_cmpswap1 = call i32 @llvm.amdgcn.raw.ptr.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, ptr addrspace(8) %tmp1, i32 88, i32 88, i32 0) #5348  %raw_ptr_atomic_cmpswap2 = call i32 @llvm.amdgcn.raw.ptr.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, ptr addrspace(8) %tmp1, i32 0, i32 176, i32 0) #5349  %raw_ptr_atomic_cmpswap3 = call i32 @llvm.amdgcn.raw.ptr.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, ptr addrspace(8) %tmp1, i32 %arg1, i32 176, i32 0) #5350  %raw_ptr_atomic_cmpswap4 = call i32 @llvm.amdgcn.raw.ptr.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, ptr addrspace(8) %tmp1, i32 176, i32 %arg1, i32 0) #5351 352  call void asm sideeffect "", "" ()353 354  call void @llvm.amdgcn.raw.buffer.store.v4f32(<4 x float> %raw_buffer0, <4 x i32> %tmp0, i32 192, i32 0, i32 0) #2355  call void @llvm.amdgcn.raw.buffer.store.v4f32(<4 x float> %raw_buffer1, <4 x i32> %tmp0, i32 96, i32 96, i32 0) #2356  call void @llvm.amdgcn.raw.buffer.store.v4f32(<4 x float> %raw_buffer2, <4 x i32> %tmp0, i32 0, i32 192, i32 0) #2357  call void @llvm.amdgcn.raw.buffer.store.v4f32(<4 x float> %raw_buffer3, <4 x i32> %tmp0, i32 %arg1, i32 192, i32 0) #2358  call void @llvm.amdgcn.raw.buffer.store.v4f32(<4 x float> %raw_buffer4, <4 x i32> %tmp0, i32 192, i32 %arg1, i32 0) #2359 360  call void asm sideeffect "", "" ()361 362  call void @llvm.amdgcn.raw.ptr.buffer.store.v4f32(<4 x float> %raw_ptr_buffer0, ptr addrspace(8) %tmp1, i32 192, i32 0, i32 0) #5363  call void @llvm.amdgcn.raw.ptr.buffer.store.v4f32(<4 x float> %raw_ptr_buffer1, ptr addrspace(8) %tmp1, i32 96, i32 96, i32 0) #5364  call void @llvm.amdgcn.raw.ptr.buffer.store.v4f32(<4 x float> %raw_ptr_buffer2, ptr addrspace(8) %tmp1, i32 0, i32 192, i32 0) #5365  call void @llvm.amdgcn.raw.ptr.buffer.store.v4f32(<4 x float> %raw_ptr_buffer3, ptr addrspace(8) %tmp1, i32 %arg1, i32 192, i32 0) #5366  call void @llvm.amdgcn.raw.ptr.buffer.store.v4f32(<4 x float> %raw_ptr_buffer4, ptr addrspace(8) %tmp1, i32 192, i32 %arg1, i32 0) #5367 368  call void asm sideeffect "", "" ()369 370  call void @llvm.amdgcn.raw.buffer.store.format.v4f32(<4 x float> %raw_buffer_format0, <4 x i32> %tmp0, i32 208, i32 0, i32 0) #2371  call void @llvm.amdgcn.raw.buffer.store.format.v4f32(<4 x float> %raw_buffer_format1, <4 x i32> %tmp0, i32 104, i32 104, i32 0) #2372  call void @llvm.amdgcn.raw.buffer.store.format.v4f32(<4 x float> %raw_buffer_format2, <4 x i32> %tmp0, i32 0, i32 208, i32 0) #2373  call void @llvm.amdgcn.raw.buffer.store.format.v4f32(<4 x float> %raw_buffer_format3, <4 x i32> %tmp0, i32 %arg1, i32 208, i32 0) #2374  call void @llvm.amdgcn.raw.buffer.store.format.v4f32(<4 x float> %raw_buffer_format4, <4 x i32> %tmp0, i32 208, i32 %arg1, i32 0) #2375 376  call void asm sideeffect "", "" ()377 378  call void @llvm.amdgcn.raw.ptr.buffer.store.format.v4f32(<4 x float> %raw_buffer_format_ptr0, ptr addrspace(8) %tmp1, i32 208, i32 0, i32 0) #4379  call void @llvm.amdgcn.raw.ptr.buffer.store.format.v4f32(<4 x float> %raw_buffer_format_ptr1, ptr addrspace(8) %tmp1, i32 104, i32 104, i32 0) #4380  call void @llvm.amdgcn.raw.ptr.buffer.store.format.v4f32(<4 x float> %raw_buffer_format_ptr2, ptr addrspace(8) %tmp1, i32 0, i32 208, i32 0) #4381  call void @llvm.amdgcn.raw.ptr.buffer.store.format.v4f32(<4 x float> %raw_buffer_format_ptr3, ptr addrspace(8) %tmp1, i32 %arg1, i32 208, i32 0) #4382  call void @llvm.amdgcn.raw.ptr.buffer.store.format.v4f32(<4 x float> %raw_buffer_format_ptr4, ptr addrspace(8) %tmp1, i32 208, i32 %arg1, i32 0) #4383 384  call void asm sideeffect "", "" ()385 386  %struct_buffer0 = call <4 x float> @llvm.amdgcn.struct.buffer.load.v4f32(<4 x i32> %tmp0, i32 0, i32 224, i32 0, i32 0) #0387  %struct_buffer1 = call <4 x float> @llvm.amdgcn.struct.buffer.load.v4f32(<4 x i32> %tmp0, i32 0, i32 112, i32 112, i32 0) #0388  %struct_buffer2 = call <4 x float> @llvm.amdgcn.struct.buffer.load.v4f32(<4 x i32> %tmp0, i32 0, i32 0, i32 224, i32 0) #0389  %struct_buffer3 = call <4 x float> @llvm.amdgcn.struct.buffer.load.v4f32(<4 x i32> %tmp0, i32 0, i32 %arg1, i32 224, i32 0) #0390  %struct_buffer4 = call <4 x float> @llvm.amdgcn.struct.buffer.load.v4f32(<4 x i32> %tmp0, i32 0, i32 224, i32 %arg1, i32 0) #0391  %struct_buffer5 = call <4 x float> @llvm.amdgcn.struct.buffer.load.v4f32(<4 x i32> %tmp0, i32 1, i32 224, i32 0, i32 0) #0392  %struct_buffer6 = call <4 x float> @llvm.amdgcn.struct.buffer.load.v4f32(<4 x i32> %tmp0, i32 %arg1, i32 224, i32 0, i32 0) #0393 394  call void asm sideeffect "", "" ()395 396  %struct_ptr_buffer0 = call <4 x float> @llvm.amdgcn.struct.ptr.buffer.load.v4f32(ptr addrspace(8) %tmp1, i32 0, i32 224, i32 0, i32 0) #3397  %struct_ptr_buffer1 = call <4 x float> @llvm.amdgcn.struct.ptr.buffer.load.v4f32(ptr addrspace(8) %tmp1, i32 0, i32 112, i32 112, i32 0) #3398  %struct_ptr_buffer2 = call <4 x float> @llvm.amdgcn.struct.ptr.buffer.load.v4f32(ptr addrspace(8) %tmp1, i32 0, i32 0, i32 224, i32 0) #3399  %struct_ptr_buffer3 = call <4 x float> @llvm.amdgcn.struct.ptr.buffer.load.v4f32(ptr addrspace(8) %tmp1, i32 0, i32 %arg1, i32 224, i32 0) #3400  %struct_ptr_buffer4 = call <4 x float> @llvm.amdgcn.struct.ptr.buffer.load.v4f32(ptr addrspace(8) %tmp1, i32 0, i32 224, i32 %arg1, i32 0) #3401  %struct_ptr_buffer5 = call <4 x float> @llvm.amdgcn.struct.ptr.buffer.load.v4f32(ptr addrspace(8) %tmp1, i32 1, i32 224, i32 0, i32 0) #3402  %struct_ptr_buffer6 = call <4 x float> @llvm.amdgcn.struct.ptr.buffer.load.v4f32(ptr addrspace(8) %tmp1, i32 %arg1, i32 224, i32 0, i32 0) #3403 404  call void asm sideeffect "", "" ()405 406  %struct_buffer_format0 = call <4 x float> @llvm.amdgcn.struct.buffer.load.format.v4f32(<4 x i32> %tmp0, i32 0, i32 240, i32 0, i32 0) #0407  %struct_buffer_format1 = call <4 x float> @llvm.amdgcn.struct.buffer.load.format.v4f32(<4 x i32> %tmp0, i32 0, i32 120, i32 120, i32 0) #0408  %struct_buffer_format2 = call <4 x float> @llvm.amdgcn.struct.buffer.load.format.v4f32(<4 x i32> %tmp0, i32 0, i32 0, i32 240, i32 0) #0409  %struct_buffer_format3 = call <4 x float> @llvm.amdgcn.struct.buffer.load.format.v4f32(<4 x i32> %tmp0, i32 0, i32 %arg1, i32 240, i32 0) #0410  %struct_buffer_format4 = call <4 x float> @llvm.amdgcn.struct.buffer.load.format.v4f32(<4 x i32> %tmp0, i32 0, i32 240, i32 %arg1, i32 0) #0411  %struct_buffer_format5 = call <4 x float> @llvm.amdgcn.struct.buffer.load.format.v4f32(<4 x i32> %tmp0, i32 1, i32 240, i32 0, i32 0) #0412  %struct_buffer_format6 = call <4 x float> @llvm.amdgcn.struct.buffer.load.format.v4f32(<4 x i32> %tmp0, i32 %arg1, i32 240, i32 0, i32 0) #0413 414  call void asm sideeffect "", "" ()415 416  %struct_buffer_format_ptr0 = call <4 x float> @llvm.amdgcn.struct.ptr.buffer.load.format.v4f32(ptr addrspace(8) %tmp1, i32 0, i32 240, i32 0, i32 0) #3417  %struct_buffer_format_ptr1 = call <4 x float> @llvm.amdgcn.struct.ptr.buffer.load.format.v4f32(ptr addrspace(8) %tmp1, i32 0, i32 120, i32 120, i32 0) #3418  %struct_buffer_format_ptr2 = call <4 x float> @llvm.amdgcn.struct.ptr.buffer.load.format.v4f32(ptr addrspace(8) %tmp1, i32 0, i32 0, i32 240, i32 0) #3419  %struct_buffer_format_ptr3 = call <4 x float> @llvm.amdgcn.struct.ptr.buffer.load.format.v4f32(ptr addrspace(8) %tmp1, i32 0, i32 %arg1, i32 240, i32 0) #3420  %struct_buffer_format_ptr4 = call <4 x float> @llvm.amdgcn.struct.ptr.buffer.load.format.v4f32(ptr addrspace(8) %tmp1, i32 0, i32 240, i32 %arg1, i32 0) #3421  %struct_buffer_format_ptr5 = call <4 x float> @llvm.amdgcn.struct.ptr.buffer.load.format.v4f32(ptr addrspace(8) %tmp1, i32 1, i32 240, i32 0, i32 0) #3422  %struct_buffer_format_ptr6 = call <4 x float> @llvm.amdgcn.struct.ptr.buffer.load.format.v4f32(ptr addrspace(8) %tmp1, i32 %arg1, i32 240, i32 0, i32 0) #3423 424  call void asm sideeffect "", "" ()425 426  %struct_atomic_add0 = call i32 @llvm.amdgcn.struct.buffer.atomic.add.i32(i32 %arg1, <4 x i32> %tmp0, i32 0, i32 256, i32 0, i32 0) #2427  %struct_atomic_add1 = call i32 @llvm.amdgcn.struct.buffer.atomic.add.i32(i32 %arg1, <4 x i32> %tmp0, i32 0, i32 128, i32 128, i32 0) #2428  %struct_atomic_add2 = call i32 @llvm.amdgcn.struct.buffer.atomic.add.i32(i32 %arg1, <4 x i32> %tmp0, i32 0, i32 0, i32 256, i32 0) #2429  %struct_atomic_add3 = call i32 @llvm.amdgcn.struct.buffer.atomic.add.i32(i32 %arg1, <4 x i32> %tmp0, i32 0, i32 %arg1, i32 256, i32 0) #2430  %struct_atomic_add4 = call i32 @llvm.amdgcn.struct.buffer.atomic.add.i32(i32 %arg1, <4 x i32> %tmp0, i32 0, i32 256, i32 %arg1, i32 0) #2431  %struct_atomic_add5 = call i32 @llvm.amdgcn.struct.buffer.atomic.add.i32(i32 %arg1, <4 x i32> %tmp0, i32 1, i32 256, i32 0, i32 0) #2432  %struct_atomic_add6 = call i32 @llvm.amdgcn.struct.buffer.atomic.add.i32(i32 %arg1, <4 x i32> %tmp0, i32 %arg1, i32 256, i32 0, i32 0) #2433 434  call void asm sideeffect "", "" ()435 436  %struct_atomic_add_ptr0 = call i32 @llvm.amdgcn.struct.ptr.buffer.atomic.add.i32(i32 %arg1, ptr addrspace(8) %tmp1, i32 0, i32 256, i32 0, i32 0) #5437  %struct_atomic_add_ptr1 = call i32 @llvm.amdgcn.struct.ptr.buffer.atomic.add.i32(i32 %arg1, ptr addrspace(8) %tmp1, i32 0, i32 128, i32 128, i32 0) #5438  %struct_atomic_add_ptr2 = call i32 @llvm.amdgcn.struct.ptr.buffer.atomic.add.i32(i32 %arg1, ptr addrspace(8) %tmp1, i32 0, i32 0, i32 256, i32 0) #5439  %struct_atomic_add_ptr3 = call i32 @llvm.amdgcn.struct.ptr.buffer.atomic.add.i32(i32 %arg1, ptr addrspace(8) %tmp1, i32 0, i32 %arg1, i32 256, i32 0) #5440  %struct_atomic_add_ptr4 = call i32 @llvm.amdgcn.struct.ptr.buffer.atomic.add.i32(i32 %arg1, ptr addrspace(8) %tmp1, i32 0, i32 256, i32 %arg1, i32 0) #5441  %struct_atomic_add_ptr5 = call i32 @llvm.amdgcn.struct.ptr.buffer.atomic.add.i32(i32 %arg1, ptr addrspace(8) %tmp1, i32 1, i32 256, i32 0, i32 0) #5442  %struct_atomic_add_ptr6 = call i32 @llvm.amdgcn.struct.ptr.buffer.atomic.add.i32(i32 %arg1, ptr addrspace(8) %tmp1, i32 %arg1, i32 256, i32 0, i32 0) #5443 444  call void asm sideeffect "", "" ()445 446  %struct_atomic_cmpswap0 = call i32 @llvm.amdgcn.struct.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, <4 x i32> %tmp0, i32 0, i32 272, i32 0, i32 0) #2447  %struct_atomic_cmpswap1 = call i32 @llvm.amdgcn.struct.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, <4 x i32> %tmp0, i32 0, i32 136, i32 136, i32 0) #2448  %struct_atomic_cmpswap2 = call i32 @llvm.amdgcn.struct.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, <4 x i32> %tmp0, i32 0, i32 0, i32 272, i32 0) #2449  %struct_atomic_cmpswap3 = call i32 @llvm.amdgcn.struct.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, <4 x i32> %tmp0, i32 0, i32 %arg1, i32 272, i32 0) #2450  %struct_atomic_cmpswap4 = call i32 @llvm.amdgcn.struct.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, <4 x i32> %tmp0, i32 0, i32 272, i32 %arg1, i32 0) #2451  %struct_atomic_cmpswap5 = call i32 @llvm.amdgcn.struct.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, <4 x i32> %tmp0, i32 1, i32 272, i32 0, i32 0) #2452  %struct_atomic_cmpswap6 = call i32 @llvm.amdgcn.struct.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, <4 x i32> %tmp0, i32 %arg1, i32 272, i32 0, i32 0) #2453 454  call void asm sideeffect "", "" ()455 456  %struct_atomic_cmpswap_ptr0 = call i32 @llvm.amdgcn.struct.ptr.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, ptr addrspace(8) %tmp1, i32 0, i32 272, i32 0, i32 0) #5457  %struct_atomic_cmpswap_ptr1 = call i32 @llvm.amdgcn.struct.ptr.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, ptr addrspace(8) %tmp1, i32 0, i32 136, i32 136, i32 0) #5458  %struct_atomic_cmpswap_ptr2 = call i32 @llvm.amdgcn.struct.ptr.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, ptr addrspace(8) %tmp1, i32 0, i32 0, i32 272, i32 0) #5459  %struct_atomic_cmpswap_ptr3 = call i32 @llvm.amdgcn.struct.ptr.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, ptr addrspace(8) %tmp1, i32 0, i32 %arg1, i32 272, i32 0) #5460  %struct_atomic_cmpswap_ptr4 = call i32 @llvm.amdgcn.struct.ptr.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, ptr addrspace(8) %tmp1, i32 0, i32 272, i32 %arg1, i32 0) #5461  %struct_atomic_cmpswap_ptr5 = call i32 @llvm.amdgcn.struct.ptr.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, ptr addrspace(8) %tmp1, i32 1, i32 272, i32 0, i32 0) #5462  %struct_atomic_cmpswap_ptr6 = call i32 @llvm.amdgcn.struct.ptr.buffer.atomic.cmpswap.i32(i32 %arg1, i32 %arg1, ptr addrspace(8) %tmp1, i32 %arg1, i32 272, i32 0, i32 0) #5463 464  call void asm sideeffect "", "" ()465 466  call void @llvm.amdgcn.struct.buffer.store.v4f32(<4 x float> %struct_buffer0, <4 x i32> %tmp0, i32 0, i32 288, i32 0, i32 0) #2467  call void @llvm.amdgcn.struct.buffer.store.v4f32(<4 x float> %struct_buffer1, <4 x i32> %tmp0, i32 0, i32 144, i32 144, i32 0) #2468  call void @llvm.amdgcn.struct.buffer.store.v4f32(<4 x float> %struct_buffer2, <4 x i32> %tmp0, i32 0, i32 0, i32 288, i32 0) #2469  call void @llvm.amdgcn.struct.buffer.store.v4f32(<4 x float> %struct_buffer3, <4 x i32> %tmp0, i32 0, i32 %arg1, i32 288, i32 0) #2470  call void @llvm.amdgcn.struct.buffer.store.v4f32(<4 x float> %struct_buffer4, <4 x i32> %tmp0, i32 0, i32 288, i32 %arg1, i32 0) #2471  call void @llvm.amdgcn.struct.buffer.store.v4f32(<4 x float> %struct_buffer5, <4 x i32> %tmp0, i32 1, i32 288, i32 0, i32 0) #2472  call void @llvm.amdgcn.struct.buffer.store.v4f32(<4 x float> %struct_buffer6, <4 x i32> %tmp0, i32 %arg1, i32 288, i32 0, i32 0) #2473 474  call void asm sideeffect "", "" ()475 476  call void @llvm.amdgcn.struct.ptr.buffer.store.v4f32(<4 x float> %struct_ptr_buffer0, ptr addrspace(8) %tmp1, i32 0, i32 288, i32 0, i32 0) #4477  call void @llvm.amdgcn.struct.ptr.buffer.store.v4f32(<4 x float> %struct_ptr_buffer1, ptr addrspace(8) %tmp1, i32 0, i32 144, i32 144, i32 0) #4478  call void @llvm.amdgcn.struct.ptr.buffer.store.v4f32(<4 x float> %struct_ptr_buffer2, ptr addrspace(8) %tmp1, i32 0, i32 0, i32 288, i32 0) #4479  call void @llvm.amdgcn.struct.ptr.buffer.store.v4f32(<4 x float> %struct_ptr_buffer3, ptr addrspace(8) %tmp1, i32 0, i32 %arg1, i32 288, i32 0) #4480  call void @llvm.amdgcn.struct.ptr.buffer.store.v4f32(<4 x float> %struct_ptr_buffer4, ptr addrspace(8) %tmp1, i32 0, i32 288, i32 %arg1, i32 0) #4481  call void @llvm.amdgcn.struct.ptr.buffer.store.v4f32(<4 x float> %struct_ptr_buffer5, ptr addrspace(8) %tmp1, i32 1, i32 288, i32 0, i32 0) #4482  call void @llvm.amdgcn.struct.ptr.buffer.store.v4f32(<4 x float> %struct_ptr_buffer6, ptr addrspace(8) %tmp1, i32 %arg1, i32 288, i32 0, i32 0) #4483 484  call void asm sideeffect "", "" ()485 486  call void @llvm.amdgcn.struct.buffer.store.format.v4f32(<4 x float> %struct_buffer_format0, <4 x i32> %tmp0, i32 0, i32 304, i32 0, i32 0) #2487  call void @llvm.amdgcn.struct.buffer.store.format.v4f32(<4 x float> %struct_buffer_format1, <4 x i32> %tmp0, i32 0, i32 152, i32 152, i32 0) #2488  call void @llvm.amdgcn.struct.buffer.store.format.v4f32(<4 x float> %struct_buffer_format2, <4 x i32> %tmp0, i32 0, i32 0, i32 304, i32 0) #2489  call void @llvm.amdgcn.struct.buffer.store.format.v4f32(<4 x float> %struct_buffer_format3, <4 x i32> %tmp0, i32 0, i32 %arg1, i32 304, i32 0) #2490  call void @llvm.amdgcn.struct.buffer.store.format.v4f32(<4 x float> %struct_buffer_format4, <4 x i32> %tmp0, i32 0, i32 304, i32 %arg1, i32 0) #2491  call void @llvm.amdgcn.struct.buffer.store.format.v4f32(<4 x float> %struct_buffer_format5, <4 x i32> %tmp0, i32 1, i32 304, i32 0, i32 0) #2492  call void @llvm.amdgcn.struct.buffer.store.format.v4f32(<4 x float> %struct_buffer_format6, <4 x i32> %tmp0, i32 %arg1, i32 304, i32 0, i32 0) #2493 494  call void @llvm.amdgcn.struct.ptr.buffer.store.format.v4f32(<4 x float> %struct_buffer_format_ptr0, ptr addrspace(8) %tmp1, i32 0, i32 304, i32 0, i32 0) #4495  call void @llvm.amdgcn.struct.ptr.buffer.store.format.v4f32(<4 x float> %struct_buffer_format_ptr1, ptr addrspace(8) %tmp1, i32 0, i32 152, i32 152, i32 0) #4496  call void @llvm.amdgcn.struct.ptr.buffer.store.format.v4f32(<4 x float> %struct_buffer_format_ptr2, ptr addrspace(8) %tmp1, i32 0, i32 0, i32 304, i32 0) #4497  call void @llvm.amdgcn.struct.ptr.buffer.store.format.v4f32(<4 x float> %struct_buffer_format_ptr3, ptr addrspace(8) %tmp1, i32 0, i32 %arg1, i32 304, i32 0) #4498  call void @llvm.amdgcn.struct.ptr.buffer.store.format.v4f32(<4 x float> %struct_buffer_format_ptr4, ptr addrspace(8) %tmp1, i32 0, i32 304, i32 %arg1, i32 0) #4499  call void @llvm.amdgcn.struct.ptr.buffer.store.format.v4f32(<4 x float> %struct_buffer_format_ptr5, ptr addrspace(8) %tmp1, i32 1, i32 304, i32 0, i32 0) #4500  call void @llvm.amdgcn.struct.ptr.buffer.store.format.v4f32(<4 x float> %struct_buffer_format_ptr6, ptr addrspace(8) %tmp1, i32 %arg1, i32 304, i32 0, i32 0) #4501 502  ret void503}504 505declare <4 x float> @llvm.amdgcn.raw.buffer.load.v4f32(<4 x i32>, i32, i32, i32) #0506declare <4 x float> @llvm.amdgcn.raw.buffer.load.format.v4f32(<4 x i32>, i32, i32, i32) #0507declare i32 @llvm.amdgcn.raw.buffer.atomic.add.i32(i32, <4 x i32>, i32, i32, i32) #2508declare i32 @llvm.amdgcn.raw.buffer.atomic.cmpswap.i32(i32, i32, <4 x i32>, i32, i32, i32) #2509declare void @llvm.amdgcn.raw.buffer.store.v4f32(<4 x float>, <4 x i32>, i32, i32, i32) #2510declare void @llvm.amdgcn.raw.buffer.store.format.v4f32(<4 x float>, <4 x i32>, i32, i32, i32) #2511declare <4 x float> @llvm.amdgcn.raw.ptr.buffer.load.v4f32(ptr addrspace(8), i32, i32, i32) #3512declare <4 x float> @llvm.amdgcn.raw.ptr.buffer.load.format.v4f32(ptr addrspace(8), i32, i32, i32) #3513declare i32 @llvm.amdgcn.raw.ptr.buffer.atomic.add.i32(i32, ptr addrspace(8), i32, i32, i32) #5514declare i32 @llvm.amdgcn.raw.ptr.buffer.atomic.cmpswap.i32(i32, i32, ptr addrspace(8), i32, i32, i32) #5515declare void @llvm.amdgcn.raw.ptr.buffer.store.v4f32(<4 x float>, ptr addrspace(8), i32, i32, i32) #4516declare void @llvm.amdgcn.raw.ptr.buffer.store.format.v4f32(<4 x float>, ptr addrspace(8), i32, i32, i32) #4517 518 519declare <4 x float> @llvm.amdgcn.struct.buffer.load.v4f32(<4 x i32>, i32, i32, i32, i32) #0520declare <4 x float> @llvm.amdgcn.struct.buffer.load.format.v4f32(<4 x i32>, i32, i32, i32, i32) #0521declare i32 @llvm.amdgcn.struct.buffer.atomic.add.i32(i32, <4 x i32>, i32, i32, i32, i32) #2522declare i32 @llvm.amdgcn.struct.buffer.atomic.cmpswap.i32(i32, i32, <4 x i32>, i32, i32, i32, i32) #2523declare void @llvm.amdgcn.struct.buffer.store.v4f32(<4 x float>, <4 x i32>, i32, i32, i32, i32) #2524declare void @llvm.amdgcn.struct.buffer.store.format.v4f32(<4 x float>, <4 x i32>, i32, i32, i32, i32) #2525 526declare <4 x float> @llvm.amdgcn.struct.ptr.buffer.load.v4f32(ptr addrspace(8), i32, i32, i32, i32) #3527declare <4 x float> @llvm.amdgcn.struct.ptr.buffer.load.format.v4f32(ptr addrspace(8), i32, i32, i32, i32) #3528declare i32 @llvm.amdgcn.struct.ptr.buffer.atomic.add.i32(i32, ptr addrspace(8), i32, i32, i32, i32) #5529declare i32 @llvm.amdgcn.struct.ptr.buffer.atomic.cmpswap.i32(i32, i32, ptr addrspace(8), i32, i32, i32, i32) #5530declare void @llvm.amdgcn.struct.ptr.buffer.store.v4f32(<4 x float>, ptr addrspace(8), i32, i32, i32, i32) #4531declare void @llvm.amdgcn.struct.ptr.buffer.store.format.v4f32(<4 x float>, ptr addrspace(8), i32, i32, i32, i32) #4532 533 534attributes #0 = { nounwind readonly }535attributes #2 = { nounwind }536attributes #3 = { nounwind memory(argmem: read) }537attributes #4 = { nounwind memory(argmem: write) }538attributes #5 = { nounwind memory(argmem: readwrite) }539 540!0 = !{}541