xref: /llvm-project/llvm/test/CodeGen/AMDGPU/snippet-copy-bundle-regression.mir (revision ac0f64f06d67a93817ccd9a3c529ad40920115c9) 
1# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 2
2# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 \
3# RUN:   -start-before=greedy,0 -stop-after=virtregrewriter,1 -verify-regalloc -o - %s | FileCheck %s
4
5# This hit an assert when identifying snippet copy bundles from
6# running off the end of the block due to using
7# MachineBasicBlock::iterator instead of
8# MachineBasicBlock::instr_iterator.
9
10--- |
11
12  define amdgpu_kernel void @kernel() #0 {
13  bb:
14    ret void
15  }
16
17  attributes #0 = { "amdgpu-calls" "amdgpu-waves-per-eu"="10,10" "target-cpu"="gfx900" }
18
19...
20---
21name:            kernel
22tracksRegLiveness: true
23frameInfo:
24  adjustsStack:    true
25  hasCalls:        true
26machineFunctionInfo:
27  isEntryFunction: true
28  scratchRSrcReg:  '$sgpr0_sgpr1_sgpr2_sgpr3'
29  stackPtrOffsetReg: '$sgpr32'
30  occupancy:       8
31body:             |
32  ; CHECK-LABEL: name: kernel
33  ; CHECK: bb.0:
34  ; CHECK-NEXT:   successors: %bb.2(0x40000000), %bb.1(0x40000000)
35  ; CHECK-NEXT:   liveins: $sgpr14, $sgpr15, $sgpr16, $vgpr0, $vgpr1, $vgpr2, $sgpr4_sgpr5, $sgpr6_sgpr7, $sgpr8_sgpr9, $sgpr10_sgpr11
36  ; CHECK-NEXT: {{  $}}
37  ; CHECK-NEXT:   renamable $sgpr34_sgpr35 = IMPLICIT_DEF
38  ; CHECK-NEXT:   dead [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
39  ; CHECK-NEXT:   renamable $sgpr41 = IMPLICIT_DEF
40  ; CHECK-NEXT:   renamable $sgpr38_sgpr39 = COPY undef $sgpr8_sgpr9
41  ; CHECK-NEXT:   renamable $sgpr36_sgpr37 = IMPLICIT_DEF
42  ; CHECK-NEXT:   renamable $sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX8_IMM renamable $sgpr38_sgpr39, 0, 0 :: (dereferenceable invariant load (s256), align 16, addrspace 4)
43  ; CHECK-NEXT:   dead renamable $sgpr4 = S_LOAD_DWORD_IMM renamable $sgpr38_sgpr39, 48, 0 :: (dereferenceable invariant load (s32), align 16, addrspace 4)
44  ; CHECK-NEXT:   dead renamable $sgpr4_sgpr5 = S_LOAD_DWORDX2_IMM renamable $sgpr44_sgpr45, 0, 0 :: (invariant load (s64), align 16, addrspace 4)
45  ; CHECK-NEXT:   ADJCALLSTACKUP 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
46  ; CHECK-NEXT:   $vgpr1 = COPY renamable $sgpr51
47  ; CHECK-NEXT:   dead $sgpr30_sgpr31 = SI_CALL undef renamable $sgpr4_sgpr5, 0, csr_amdgpu, implicit undef $sgpr15, implicit $vgpr31, implicit $sgpr0_sgpr1_sgpr2_sgpr3
48  ; CHECK-NEXT:   ADJCALLSTACKDOWN 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
49  ; CHECK-NEXT:   $vcc = COPY renamable $sgpr40_sgpr41
50  ; CHECK-NEXT:   S_CBRANCH_VCCZ %bb.2, implicit undef $vcc
51  ; CHECK-NEXT: {{  $}}
52  ; CHECK-NEXT: bb.1:
53  ; CHECK-NEXT:   successors: %bb.3(0x80000000)
54  ; CHECK-NEXT:   liveins: $sgpr34_sgpr35, $sgpr36_sgpr37, $sgpr38_sgpr39, $sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51:0x000000000000FC00
55  ; CHECK-NEXT: {{  $}}
56  ; CHECK-NEXT:   renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11 = S_LOAD_DWORDX8_IMM renamable $sgpr38_sgpr39, 56, 0 :: (dereferenceable invariant load (s256), align 8, addrspace 4)
57  ; CHECK-NEXT:   S_BRANCH %bb.3
58  ; CHECK-NEXT: {{  $}}
59  ; CHECK-NEXT: bb.2:
60  ; CHECK-NEXT:   successors: %bb.3(0x80000000)
61  ; CHECK-NEXT:   liveins: $sgpr34_sgpr35, $sgpr36_sgpr37, $sgpr38_sgpr39, $sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51:0x000000000000FC00
62  ; CHECK-NEXT: {{  $}}
63  ; CHECK-NEXT:   renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11 = S_LOAD_DWORDX8_IMM renamable $sgpr38_sgpr39, 56, 0 :: (dereferenceable invariant load (s256), align 8, addrspace 4)
64  ; CHECK-NEXT:   S_CMP_LG_U64 renamable $sgpr4_sgpr5, 0, implicit-def $scc
65  ; CHECK-NEXT: {{  $}}
66  ; CHECK-NEXT: bb.3:
67  ; CHECK-NEXT:   successors: %bb.5(0x40000000), %bb.4(0x40000000)
68  ; CHECK-NEXT:   liveins: $sgpr34_sgpr35, $sgpr36_sgpr37, $sgpr38_sgpr39, $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11:0x00000000000003F0, $sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51:0x000000000000FC00
69  ; CHECK-NEXT: {{  $}}
70  ; CHECK-NEXT:   S_CBRANCH_VCCZ %bb.5, implicit undef $vcc
71  ; CHECK-NEXT: {{  $}}
72  ; CHECK-NEXT: bb.4:
73  ; CHECK-NEXT:   successors: %bb.5(0x80000000)
74  ; CHECK-NEXT:   liveins: $sgpr34_sgpr35, $sgpr36_sgpr37, $sgpr38_sgpr39, $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11:0x00000000000003F0, $sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51:0x000000000000FC00
75  ; CHECK-NEXT: {{  $}}
76  ; CHECK-NEXT:   S_CMP_EQ_U32 renamable $sgpr8, 0, implicit-def $scc
77  ; CHECK-NEXT: {{  $}}
78  ; CHECK-NEXT: bb.5:
79  ; CHECK-NEXT:   liveins: $sgpr34_sgpr35, $sgpr36_sgpr37, $sgpr38_sgpr39, $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11:0x00000000000000F0, $sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51:0x000000000000FC00
80  ; CHECK-NEXT: {{  $}}
81  ; CHECK-NEXT:   dead renamable $sgpr4_sgpr5 = S_LOAD_DWORDX2_IMM killed renamable $sgpr38_sgpr39, 40, 0 :: (dereferenceable invariant load (s64), addrspace 4)
82  ; CHECK-NEXT:   GLOBAL_STORE_DWORD_SADDR undef [[DEF]], undef [[DEF]], killed renamable $sgpr6_sgpr7, 0, 0, implicit $exec :: (store (s32), addrspace 1)
83  ; CHECK-NEXT:   GLOBAL_STORE_DWORD_SADDR undef [[DEF]], undef [[DEF]], renamable $sgpr50_sgpr51, 0, 0, implicit $exec :: (store (s32), addrspace 1)
84  ; CHECK-NEXT:   dead [[COPY:%[0-9]+]]:vgpr_32 = COPY killed renamable $sgpr49
85  ; CHECK-NEXT:   ADJCALLSTACKUP 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
86  ; CHECK-NEXT:   $sgpr6_sgpr7 = COPY killed renamable $sgpr36_sgpr37
87  ; CHECK-NEXT:   $sgpr10_sgpr11 = COPY killed renamable $sgpr34_sgpr35
88  ; CHECK-NEXT:   ADJCALLSTACKDOWN 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
89  ; CHECK-NEXT:   S_ENDPGM 0
90  bb.0:
91    liveins: $vgpr0, $vgpr1, $vgpr2, $sgpr4_sgpr5, $sgpr6_sgpr7, $sgpr8_sgpr9, $sgpr10_sgpr11, $sgpr14, $sgpr15, $sgpr16
92
93    %0:sgpr_64 = IMPLICIT_DEF
94    %1:vgpr_32 = IMPLICIT_DEF
95    undef %2.sub1:sreg_64 = IMPLICIT_DEF
96    %3:sgpr_64 = COPY undef $sgpr8_sgpr9
97    %4:sgpr_64 = IMPLICIT_DEF
98    %5:sgpr_256 = S_LOAD_DWORDX8_IMM %3, 0, 0 :: (dereferenceable invariant load (s256), align 16, addrspace 4)
99    %6:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM %3, 48, 0 :: (dereferenceable invariant load (s32), align 16, addrspace 4)
100    %7:sgpr_256 = S_LOAD_DWORDX8_IMM %3, 56, 0 :: (dereferenceable invariant load (s256), align 8, addrspace 4)
101    %8:sreg_64_xexec = S_LOAD_DWORDX2_IMM %5.sub0_sub1, 0, 0 :: (invariant load (s64), align 16, addrspace 4)
102    ADJCALLSTACKUP 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
103    $vgpr1 = COPY %5.sub7
104    dead $sgpr30_sgpr31 = SI_CALL undef %8, 0, csr_amdgpu, implicit undef $sgpr15, implicit $vgpr31, implicit $sgpr0_sgpr1_sgpr2_sgpr3
105    ADJCALLSTACKDOWN 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
106    $vcc = COPY %2
107    S_CBRANCH_VCCZ %bb.2, implicit undef $vcc
108
109  bb.1:
110    S_BRANCH %bb.3
111
112  bb.2:
113    S_CMP_LG_U64 %7.sub0_sub1, 0, implicit-def $scc
114
115  bb.3:
116    S_CBRANCH_VCCZ %bb.5, implicit undef $vcc
117
118  bb.4:
119    S_CMP_EQ_U32 %7.sub4, 0, implicit-def $scc
120
121  bb.5:
122    %9:sreg_64_xexec = S_LOAD_DWORDX2_IMM %3, 40, 0 :: (dereferenceable invariant load (s64), addrspace 4)
123    GLOBAL_STORE_DWORD_SADDR undef %1, undef %1, %7.sub2_sub3, 0, 0, implicit $exec :: (store (s32), addrspace 1)
124    GLOBAL_STORE_DWORD_SADDR undef %1, undef %1, %5.sub6_sub7, 0, 0, implicit $exec :: (store (s32), addrspace 1)
125    %10:vgpr_32 = COPY %5.sub5
126    ADJCALLSTACKUP 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
127    $sgpr6_sgpr7 = COPY %4
128    $sgpr10_sgpr11 = COPY %0
129    ADJCALLSTACKDOWN 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
130    S_ENDPGM 0
131
132...
133