xref: /llvm-project/llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-atomic-cmpxchg-local.mir (revision 9e9907f1cfa424366fba58d9520f9305b537cec9) 
1# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
2# RUN: llc -mtriple=amdgcn -mcpu=tahiti -run-pass=instruction-select -verify-machineinstrs  -global-isel-abort=0 -o - %s | FileCheck -check-prefix=GFX6 %s
3# RUN: llc -mtriple=amdgcn -mcpu=hawaii -run-pass=instruction-select -verify-machineinstrs  -global-isel-abort=0 -o - %s | FileCheck -check-prefix=GFX7 %s
4# RUN: llc -mtriple=amdgcn -mcpu=fiji -run-pass=instruction-select -verify-machineinstrs  -global-isel-abort=0 -o - %s | FileCheck -check-prefix=GFX7 %s
5# RUN: llc -mtriple=amdgcn -mcpu=gfx900 -run-pass=instruction-select -verify-machineinstrs -global-isel-abort=0 -o - %s | FileCheck -check-prefix=GFX9 %s
6# RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -run-pass=instruction-select -verify-machineinstrs -global-isel-abort=0 -o - %s | FileCheck -check-prefix=GFX9 %s
7# RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -run-pass=instruction-select -verify-machineinstrs -global-isel-abort=0 -o - %s | FileCheck -check-prefix=GFX11 %s
8
9
10---
11name:            atomic_cmpxchg_s32_local
12legalized:       true
13regBankSelected: true
14tracksRegLiveness: true
15body:             |
16  bb.0:
17    liveins: $vgpr0, $vgpr1, $vgpr2
18
19    ; GFX6-LABEL: name: atomic_cmpxchg_s32_local
20    ; GFX6: liveins: $vgpr0, $vgpr1, $vgpr2
21    ; GFX6-NEXT: {{  $}}
22    ; GFX6-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
23    ; GFX6-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
24    ; GFX6-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
25    ; GFX6-NEXT: $m0 = S_MOV_B32 -1
26    ; GFX6-NEXT: [[DS_CMPST_RTN_B32_:%[0-9]+]]:vgpr_32 = DS_CMPST_RTN_B32 [[COPY]], [[COPY1]], [[COPY2]], 0, 0, implicit $m0, implicit $exec :: (load store seq_cst (s32), addrspace 3)
27    ; GFX6-NEXT: $vgpr0 = COPY [[DS_CMPST_RTN_B32_]]
28    ; GFX7-LABEL: name: atomic_cmpxchg_s32_local
29    ; GFX7: liveins: $vgpr0, $vgpr1, $vgpr2
30    ; GFX7-NEXT: {{  $}}
31    ; GFX7-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
32    ; GFX7-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
33    ; GFX7-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
34    ; GFX7-NEXT: $m0 = S_MOV_B32 -1
35    ; GFX7-NEXT: [[DS_CMPST_RTN_B32_:%[0-9]+]]:vgpr_32 = DS_CMPST_RTN_B32 [[COPY]], [[COPY1]], [[COPY2]], 0, 0, implicit $m0, implicit $exec :: (load store seq_cst (s32), addrspace 3)
36    ; GFX7-NEXT: $vgpr0 = COPY [[DS_CMPST_RTN_B32_]]
37    ; GFX9-LABEL: name: atomic_cmpxchg_s32_local
38    ; GFX9: liveins: $vgpr0, $vgpr1, $vgpr2
39    ; GFX9-NEXT: {{  $}}
40    ; GFX9-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
41    ; GFX9-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
42    ; GFX9-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
43    ; GFX9-NEXT: [[DS_CMPST_RTN_B32_gfx9_:%[0-9]+]]:vgpr_32 = DS_CMPST_RTN_B32_gfx9 [[COPY]], [[COPY1]], [[COPY2]], 0, 0, implicit $exec :: (load store seq_cst (s32), addrspace 3)
44    ; GFX9-NEXT: $vgpr0 = COPY [[DS_CMPST_RTN_B32_gfx9_]]
45    ; GFX11-LABEL: name: atomic_cmpxchg_s32_local
46    ; GFX11: liveins: $vgpr0, $vgpr1, $vgpr2
47    ; GFX11-NEXT: {{  $}}
48    ; GFX11-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
49    ; GFX11-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
50    ; GFX11-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
51    ; GFX11-NEXT: [[DS_CMPSTORE_RTN_B32_gfx9_:%[0-9]+]]:vgpr_32 = DS_CMPSTORE_RTN_B32_gfx9 [[COPY]], [[COPY2]], [[COPY1]], 0, 0, implicit $exec :: (load store seq_cst (s32), addrspace 3)
52    ; GFX11-NEXT: $vgpr0 = COPY [[DS_CMPSTORE_RTN_B32_gfx9_]]
53    %0:vgpr(p3) = COPY $vgpr0
54    %1:vgpr(s32) = COPY $vgpr1
55    %2:vgpr(s32) = COPY $vgpr2
56    %3:vgpr(s32) = G_ATOMIC_CMPXCHG %0, %1, %2 :: (load store seq_cst (s32), addrspace 3)
57    $vgpr0 = COPY %3
58
59...
60
61---
62name:            atomic_cmpxchg_s32_local_gep4
63legalized:       true
64regBankSelected: true
65tracksRegLiveness: true
66body:             |
67  bb.0:
68    liveins: $vgpr0, $vgpr1, $vgpr2
69
70    ; GFX6-LABEL: name: atomic_cmpxchg_s32_local_gep4
71    ; GFX6: liveins: $vgpr0, $vgpr1, $vgpr2
72    ; GFX6-NEXT: {{  $}}
73    ; GFX6-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
74    ; GFX6-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
75    ; GFX6-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
76    ; GFX6-NEXT: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 4, implicit $exec
77    ; GFX6-NEXT: %4:vgpr_32, dead %6:sreg_64_xexec = V_ADD_CO_U32_e64 [[COPY]], [[V_MOV_B32_e32_]], 0, implicit $exec
78    ; GFX6-NEXT: $m0 = S_MOV_B32 -1
79    ; GFX6-NEXT: [[DS_CMPST_RTN_B32_:%[0-9]+]]:vgpr_32 = DS_CMPST_RTN_B32 %4, [[COPY1]], [[COPY2]], 0, 0, implicit $m0, implicit $exec :: (load store seq_cst (s32), addrspace 3)
80    ; GFX6-NEXT: $vgpr0 = COPY [[DS_CMPST_RTN_B32_]]
81    ; GFX7-LABEL: name: atomic_cmpxchg_s32_local_gep4
82    ; GFX7: liveins: $vgpr0, $vgpr1, $vgpr2
83    ; GFX7-NEXT: {{  $}}
84    ; GFX7-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
85    ; GFX7-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
86    ; GFX7-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
87    ; GFX7-NEXT: $m0 = S_MOV_B32 -1
88    ; GFX7-NEXT: [[DS_CMPST_RTN_B32_:%[0-9]+]]:vgpr_32 = DS_CMPST_RTN_B32 [[COPY]], [[COPY1]], [[COPY2]], 4, 0, implicit $m0, implicit $exec :: (load store seq_cst (s32), addrspace 3)
89    ; GFX7-NEXT: $vgpr0 = COPY [[DS_CMPST_RTN_B32_]]
90    ; GFX9-LABEL: name: atomic_cmpxchg_s32_local_gep4
91    ; GFX9: liveins: $vgpr0, $vgpr1, $vgpr2
92    ; GFX9-NEXT: {{  $}}
93    ; GFX9-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
94    ; GFX9-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
95    ; GFX9-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
96    ; GFX9-NEXT: [[DS_CMPST_RTN_B32_gfx9_:%[0-9]+]]:vgpr_32 = DS_CMPST_RTN_B32_gfx9 [[COPY]], [[COPY1]], [[COPY2]], 4, 0, implicit $exec :: (load store seq_cst (s32), addrspace 3)
97    ; GFX9-NEXT: $vgpr0 = COPY [[DS_CMPST_RTN_B32_gfx9_]]
98    ; GFX11-LABEL: name: atomic_cmpxchg_s32_local_gep4
99    ; GFX11: liveins: $vgpr0, $vgpr1, $vgpr2
100    ; GFX11-NEXT: {{  $}}
101    ; GFX11-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
102    ; GFX11-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
103    ; GFX11-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
104    ; GFX11-NEXT: [[DS_CMPSTORE_RTN_B32_gfx9_:%[0-9]+]]:vgpr_32 = DS_CMPSTORE_RTN_B32_gfx9 [[COPY]], [[COPY2]], [[COPY1]], 4, 0, implicit $exec :: (load store seq_cst (s32), addrspace 3)
105    ; GFX11-NEXT: $vgpr0 = COPY [[DS_CMPSTORE_RTN_B32_gfx9_]]
106    %0:vgpr(p3) = COPY $vgpr0
107    %1:vgpr(s32) = COPY $vgpr1
108    %2:vgpr(s32) = COPY $vgpr2
109    %3:vgpr(s32) = G_CONSTANT i32 4
110    %4:vgpr(p3) = G_PTR_ADD %0, %3
111    %5:vgpr(s32) = G_ATOMIC_CMPXCHG %4, %1, %2 :: (load store seq_cst (s32), addrspace 3)
112    $vgpr0 = COPY %5
113
114...
115
116---
117name:            atomic_cmpxchg_s64_local
118legalized:       true
119regBankSelected: true
120tracksRegLiveness: true
121body:             |
122  bb.0:
123    liveins: $vgpr0, $vgpr1_vgpr2, $vgpr3_vgpr4
124
125    ; GFX6-LABEL: name: atomic_cmpxchg_s64_local
126    ; GFX6: liveins: $vgpr0, $vgpr1_vgpr2, $vgpr3_vgpr4
127    ; GFX6-NEXT: {{  $}}
128    ; GFX6-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
129    ; GFX6-NEXT: [[COPY1:%[0-9]+]]:vreg_64 = COPY $vgpr1_vgpr2
130    ; GFX6-NEXT: [[COPY2:%[0-9]+]]:vreg_64 = COPY $vgpr3_vgpr4
131    ; GFX6-NEXT: $m0 = S_MOV_B32 -1
132    ; GFX6-NEXT: [[DS_CMPST_RTN_B64_:%[0-9]+]]:vreg_64 = DS_CMPST_RTN_B64 [[COPY]], [[COPY1]], [[COPY2]], 0, 0, implicit $m0, implicit $exec :: (load store seq_cst (s64), addrspace 3)
133    ; GFX6-NEXT: $vgpr0_vgpr1 = COPY [[DS_CMPST_RTN_B64_]]
134    ; GFX7-LABEL: name: atomic_cmpxchg_s64_local
135    ; GFX7: liveins: $vgpr0, $vgpr1_vgpr2, $vgpr3_vgpr4
136    ; GFX7-NEXT: {{  $}}
137    ; GFX7-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
138    ; GFX7-NEXT: [[COPY1:%[0-9]+]]:vreg_64 = COPY $vgpr1_vgpr2
139    ; GFX7-NEXT: [[COPY2:%[0-9]+]]:vreg_64 = COPY $vgpr3_vgpr4
140    ; GFX7-NEXT: $m0 = S_MOV_B32 -1
141    ; GFX7-NEXT: [[DS_CMPST_RTN_B64_:%[0-9]+]]:vreg_64 = DS_CMPST_RTN_B64 [[COPY]], [[COPY1]], [[COPY2]], 0, 0, implicit $m0, implicit $exec :: (load store seq_cst (s64), addrspace 3)
142    ; GFX7-NEXT: $vgpr0_vgpr1 = COPY [[DS_CMPST_RTN_B64_]]
143    ; GFX9-LABEL: name: atomic_cmpxchg_s64_local
144    ; GFX9: liveins: $vgpr0, $vgpr1_vgpr2, $vgpr3_vgpr4
145    ; GFX9-NEXT: {{  $}}
146    ; GFX9-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
147    ; GFX9-NEXT: [[COPY1:%[0-9]+]]:vreg_64 = COPY $vgpr1_vgpr2
148    ; GFX9-NEXT: [[COPY2:%[0-9]+]]:vreg_64 = COPY $vgpr3_vgpr4
149    ; GFX9-NEXT: [[DS_CMPST_RTN_B64_gfx9_:%[0-9]+]]:vreg_64 = DS_CMPST_RTN_B64_gfx9 [[COPY]], [[COPY1]], [[COPY2]], 0, 0, implicit $exec :: (load store seq_cst (s64), addrspace 3)
150    ; GFX9-NEXT: $vgpr0_vgpr1 = COPY [[DS_CMPST_RTN_B64_gfx9_]]
151    ; GFX11-LABEL: name: atomic_cmpxchg_s64_local
152    ; GFX11: liveins: $vgpr0, $vgpr1_vgpr2, $vgpr3_vgpr4
153    ; GFX11-NEXT: {{  $}}
154    ; GFX11-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
155    ; GFX11-NEXT: [[COPY1:%[0-9]+]]:vreg_64 = COPY $vgpr1_vgpr2
156    ; GFX11-NEXT: [[COPY2:%[0-9]+]]:vreg_64 = COPY $vgpr3_vgpr4
157    ; GFX11-NEXT: [[DS_CMPSTORE_RTN_B64_gfx9_:%[0-9]+]]:vreg_64 = DS_CMPSTORE_RTN_B64_gfx9 [[COPY]], [[COPY2]], [[COPY1]], 0, 0, implicit $exec :: (load store seq_cst (s64), addrspace 3)
158    ; GFX11-NEXT: $vgpr0_vgpr1 = COPY [[DS_CMPSTORE_RTN_B64_gfx9_]]
159    %0:vgpr(p3) = COPY $vgpr0
160    %1:vgpr(s64) = COPY $vgpr1_vgpr2
161    %2:vgpr(s64) = COPY $vgpr3_vgpr4
162    %3:vgpr(s64) = G_ATOMIC_CMPXCHG %0, %1, %2 :: (load store seq_cst (s64), addrspace 3)
163    $vgpr0_vgpr1 = COPY %3
164
165...
166
167---
168name:            atomic_cmpxchg_s64_local_gep4
169legalized:       true
170regBankSelected: true
171tracksRegLiveness: true
172body:             |
173  bb.0:
174    liveins: $vgpr0, $vgpr1_vgpr2, $vgpr3_vgpr4
175
176    ; GFX6-LABEL: name: atomic_cmpxchg_s64_local_gep4
177    ; GFX6: liveins: $vgpr0, $vgpr1_vgpr2, $vgpr3_vgpr4
178    ; GFX6-NEXT: {{  $}}
179    ; GFX6-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
180    ; GFX6-NEXT: [[COPY1:%[0-9]+]]:vreg_64 = COPY $vgpr1_vgpr2
181    ; GFX6-NEXT: [[COPY2:%[0-9]+]]:vreg_64 = COPY $vgpr3_vgpr4
182    ; GFX6-NEXT: $m0 = S_MOV_B32 -1
183    ; GFX6-NEXT: [[DS_CMPST_RTN_B64_:%[0-9]+]]:vreg_64 = DS_CMPST_RTN_B64 [[COPY]], [[COPY1]], [[COPY2]], 0, 0, implicit $m0, implicit $exec :: (load store seq_cst (s64), addrspace 3)
184    ; GFX6-NEXT: $vgpr0_vgpr1 = COPY [[DS_CMPST_RTN_B64_]]
185    ; GFX7-LABEL: name: atomic_cmpxchg_s64_local_gep4
186    ; GFX7: liveins: $vgpr0, $vgpr1_vgpr2, $vgpr3_vgpr4
187    ; GFX7-NEXT: {{  $}}
188    ; GFX7-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
189    ; GFX7-NEXT: [[COPY1:%[0-9]+]]:vreg_64 = COPY $vgpr1_vgpr2
190    ; GFX7-NEXT: [[COPY2:%[0-9]+]]:vreg_64 = COPY $vgpr3_vgpr4
191    ; GFX7-NEXT: $m0 = S_MOV_B32 -1
192    ; GFX7-NEXT: [[DS_CMPST_RTN_B64_:%[0-9]+]]:vreg_64 = DS_CMPST_RTN_B64 [[COPY]], [[COPY1]], [[COPY2]], 0, 0, implicit $m0, implicit $exec :: (load store seq_cst (s64), addrspace 3)
193    ; GFX7-NEXT: $vgpr0_vgpr1 = COPY [[DS_CMPST_RTN_B64_]]
194    ; GFX9-LABEL: name: atomic_cmpxchg_s64_local_gep4
195    ; GFX9: liveins: $vgpr0, $vgpr1_vgpr2, $vgpr3_vgpr4
196    ; GFX9-NEXT: {{  $}}
197    ; GFX9-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
198    ; GFX9-NEXT: [[COPY1:%[0-9]+]]:vreg_64 = COPY $vgpr1_vgpr2
199    ; GFX9-NEXT: [[COPY2:%[0-9]+]]:vreg_64 = COPY $vgpr3_vgpr4
200    ; GFX9-NEXT: [[DS_CMPST_RTN_B64_gfx9_:%[0-9]+]]:vreg_64 = DS_CMPST_RTN_B64_gfx9 [[COPY]], [[COPY1]], [[COPY2]], 0, 0, implicit $exec :: (load store seq_cst (s64), addrspace 3)
201    ; GFX9-NEXT: $vgpr0_vgpr1 = COPY [[DS_CMPST_RTN_B64_gfx9_]]
202    ; GFX11-LABEL: name: atomic_cmpxchg_s64_local_gep4
203    ; GFX11: liveins: $vgpr0, $vgpr1_vgpr2, $vgpr3_vgpr4
204    ; GFX11-NEXT: {{  $}}
205    ; GFX11-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
206    ; GFX11-NEXT: [[COPY1:%[0-9]+]]:vreg_64 = COPY $vgpr1_vgpr2
207    ; GFX11-NEXT: [[COPY2:%[0-9]+]]:vreg_64 = COPY $vgpr3_vgpr4
208    ; GFX11-NEXT: [[DS_CMPSTORE_RTN_B64_gfx9_:%[0-9]+]]:vreg_64 = DS_CMPSTORE_RTN_B64_gfx9 [[COPY]], [[COPY2]], [[COPY1]], 0, 0, implicit $exec :: (load store seq_cst (s64), addrspace 3)
209    ; GFX11-NEXT: $vgpr0_vgpr1 = COPY [[DS_CMPSTORE_RTN_B64_gfx9_]]
210    %0:vgpr(p3) = COPY $vgpr0
211    %1:vgpr(s64) = COPY $vgpr1_vgpr2
212    %2:vgpr(s64) = COPY $vgpr3_vgpr4
213    %3:vgpr(s32) = G_CONSTANT i32 4
214    %4:vgpr(p3) = G_PTR_ADD %0, %3
215    %5:vgpr(s64) = G_ATOMIC_CMPXCHG %0, %1, %2 :: (load store seq_cst (s64), addrspace 3)
216    $vgpr0_vgpr1 = COPY %5
217
218...
219