Skip to content

Commit d61d35a

Browse files
arsenmarsenm
committed
Handle amdgpu-unsafe-fp-atomics attr
1 parent 67dc8d3 commit d61d35a

13 files changed

+4793
-5804
lines changed

‎llvm/lib/Target/AMDGPU/SIISelLowering.cpp

Lines changed: 11 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -16082,6 +16082,13 @@ SITargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *RMW) const {
1608216082
if (AS == AMDGPUAS::LOCAL_ADDRESS && (Ty->isFloatTy() || Ty->isDoubleTy()))
1608316083
return AtomicExpansionKind::None;
1608416084

16085+
if (unsafeFPAtomicsDisabled(RMW->getFunction()))
16086+
return AtomicExpansionKind::CmpXChg;
16087+
16088+
// Always expand system scope fp atomics.
16089+
if (HasSystemScope)
16090+
return AtomicExpansionKind::CmpXChg;
16091+
1608516092
// For flat and global cases:
1608616093
// float, double in gfx7. Manual claims denormal support.
1608716094
// Removed in gfx8.
@@ -16094,14 +16101,14 @@ SITargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *RMW) const {
1609416101
// FIXME: Check scope and fine grained memory
1609516102
if (AS == AMDGPUAS::FLAT_ADDRESS) {
1609616103
if (Subtarget->hasAtomicFMinFMaxF32FlatInsts() && Ty->isFloatTy())
16097-
return AtomicExpansionKind::None;
16104+
return ReportUnsafeHWInst(AtomicExpansionKind::None);
1609816105
if (Subtarget->hasAtomicFMinFMaxF64FlatInsts() && Ty->isDoubleTy())
16099-
return AtomicExpansionKind::None;
16106+
return ReportUnsafeHWInst(AtomicExpansionKind::None);
1610016107
} else if (AMDGPU::isExtendedGlobalAddrSpace(AS)) {
1610116108
if (Subtarget->hasAtomicFMinFMaxF32GlobalInsts() && Ty->isFloatTy())
16102-
return AtomicExpansionKind::None;
16109+
return ReportUnsafeHWInst(AtomicExpansionKind::None);
1610316110
if (Subtarget->hasAtomicFMinFMaxF64GlobalInsts() && Ty->isDoubleTy())
16104-
return AtomicExpansionKind::None;
16111+
return ReportUnsafeHWInst(AtomicExpansionKind::None);
1610516112
}
1610616113

1610716114
return AtomicExpansionKind::CmpXChg;

‎llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmax.ll

Lines changed: 290 additions & 34 deletions
Large diffs are not rendered by default.

‎llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmin.ll

Lines changed: 290 additions & 34 deletions
Large diffs are not rendered by default.

‎llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmax.ll

Lines changed: 372 additions & 30 deletions
Large diffs are not rendered by default.

‎llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmin.ll

Lines changed: 372 additions & 30 deletions
Large diffs are not rendered by default.

‎llvm/test/CodeGen/AMDGPU/global_atomics_scan_fmax.ll

Lines changed: 1262 additions & 360 deletions
Large diffs are not rendered by default.

‎llvm/test/CodeGen/AMDGPU/global_atomics_scan_fmin.ll

Lines changed: 1262 additions & 360 deletions
Large diffs are not rendered by default.

‎llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-f32-agent.ll

Lines changed: 261 additions & 1589 deletions
Large diffs are not rendered by default.

‎llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-f32-system.ll

Lines changed: 261 additions & 1589 deletions
Large diffs are not rendered by default.

‎llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-f64-agent.ll

Lines changed: 128 additions & 792 deletions
Large diffs are not rendered by default.

‎llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-f64-system.ll

Lines changed: 128 additions & 792 deletions
Large diffs are not rendered by default.

‎llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-rmw-fmax.ll

Lines changed: 78 additions & 95 deletions
Original file line numberDiff line numberDiff line change
@@ -3,50 +3,42 @@
33
; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -passes=atomic-expand %s | FileCheck -check-prefixes=GCN,GFX9 %s
44

55
define float @test_atomicrmw_fmax_f32_flat(ptr %ptr, float %value) {
6-
; GFX7-LABEL: @test_atomicrmw_fmax_f32_flat(
7-
; GFX7-NEXT: [[RES:%.*]] = atomicrmw fmax ptr [[PTR:%.*]], float [[VALUE:%.*]] seq_cst, align 4
8-
; GFX7-NEXT: ret float [[RES]]
9-
;
10-
; GFX9-LABEL: @test_atomicrmw_fmax_f32_flat(
11-
; GFX9-NEXT: [[TMP1:%.*]] = load float, ptr [[PTR:%.*]], align 4
12-
; GFX9-NEXT: br label [[ATOMICRMW_START:%.*]]
13-
; GFX9: atomicrmw.start:
14-
; GFX9-NEXT: [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
15-
; GFX9-NEXT: [[TMP2:%.*]] = call float @llvm.maxnum.f32(float [[LOADED]], float [[VALUE:%.*]])
16-
; GFX9-NEXT: [[TMP3:%.*]] = bitcast float [[TMP2]] to i32
17-
; GFX9-NEXT: [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
18-
; GFX9-NEXT: [[TMP5:%.*]] = cmpxchg ptr [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
19-
; GFX9-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
20-
; GFX9-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
21-
; GFX9-NEXT: [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
22-
; GFX9-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
23-
; GFX9: atomicrmw.end:
24-
; GFX9-NEXT: ret float [[TMP6]]
6+
; GCN-LABEL: @test_atomicrmw_fmax_f32_flat(
7+
; GCN-NEXT: [[TMP1:%.*]] = load float, ptr [[PTR:%.*]], align 4
8+
; GCN-NEXT: br label [[ATOMICRMW_START:%.*]]
9+
; GCN: atomicrmw.start:
10+
; GCN-NEXT: [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
11+
; GCN-NEXT: [[TMP2:%.*]] = call float @llvm.maxnum.f32(float [[LOADED]], float [[VALUE:%.*]])
12+
; GCN-NEXT: [[TMP3:%.*]] = bitcast float [[TMP2]] to i32
13+
; GCN-NEXT: [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
14+
; GCN-NEXT: [[TMP5:%.*]] = cmpxchg ptr [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
15+
; GCN-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
16+
; GCN-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
17+
; GCN-NEXT: [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
18+
; GCN-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
19+
; GCN: atomicrmw.end:
20+
; GCN-NEXT: ret float [[TMP6]]
2521
;
2622
%res = atomicrmw fmax ptr %ptr, float %value seq_cst
2723
ret float %res
2824
}
2925

3026
define float @test_atomicrmw_fmax_f32_global(ptr addrspace(1) %ptr, float %value) {
31-
; GFX7-LABEL: @test_atomicrmw_fmax_f32_global(
32-
; GFX7-NEXT: [[RES:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR:%.*]], float [[VALUE:%.*]] seq_cst, align 4
33-
; GFX7-NEXT: ret float [[RES]]
34-
;
35-
; GFX9-LABEL: @test_atomicrmw_fmax_f32_global(
36-
; GFX9-NEXT: [[TMP1:%.*]] = load float, ptr addrspace(1) [[PTR:%.*]], align 4
37-
; GFX9-NEXT: br label [[ATOMICRMW_START:%.*]]
38-
; GFX9: atomicrmw.start:
39-
; GFX9-NEXT: [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
40-
; GFX9-NEXT: [[TMP2:%.*]] = call float @llvm.maxnum.f32(float [[LOADED]], float [[VALUE:%.*]])
41-
; GFX9-NEXT: [[TMP3:%.*]] = bitcast float [[TMP2]] to i32
42-
; GFX9-NEXT: [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
43-
; GFX9-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
44-
; GFX9-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
45-
; GFX9-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
46-
; GFX9-NEXT: [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
47-
; GFX9-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
48-
; GFX9: atomicrmw.end:
49-
; GFX9-NEXT: ret float [[TMP6]]
27+
; GCN-LABEL: @test_atomicrmw_fmax_f32_global(
28+
; GCN-NEXT: [[TMP1:%.*]] = load float, ptr addrspace(1) [[PTR:%.*]], align 4
29+
; GCN-NEXT: br label [[ATOMICRMW_START:%.*]]
30+
; GCN: atomicrmw.start:
31+
; GCN-NEXT: [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
32+
; GCN-NEXT: [[TMP2:%.*]] = call float @llvm.maxnum.f32(float [[LOADED]], float [[VALUE:%.*]])
33+
; GCN-NEXT: [[TMP3:%.*]] = bitcast float [[TMP2]] to i32
34+
; GCN-NEXT: [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
35+
; GCN-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
36+
; GCN-NEXT: [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
37+
; GCN-NEXT: [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
38+
; GCN-NEXT: [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
39+
; GCN-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
40+
; GCN: atomicrmw.end:
41+
; GCN-NEXT: ret float [[TMP6]]
5042
;
5143
%res = atomicrmw fmax ptr addrspace(1) %ptr, float %value seq_cst
5244
ret float %res
@@ -195,50 +187,42 @@ define half @test_atomicrmw_fmax_f16_local(ptr addrspace(3) %ptr, half %value) {
195187
}
196188

197189
define double @test_atomicrmw_fmax_f64_flat(ptr %ptr, double %value) {
198-
; GFX7-LABEL: @test_atomicrmw_fmax_f64_flat(
199-
; GFX7-NEXT: [[RES:%.*]] = atomicrmw fmax ptr [[PTR:%.*]], double [[VALUE:%.*]] seq_cst, align 8
200-
; GFX7-NEXT: ret double [[RES]]
201-
;
202-
; GFX9-LABEL: @test_atomicrmw_fmax_f64_flat(
203-
; GFX9-NEXT: [[TMP1:%.*]] = load double, ptr [[PTR:%.*]], align 8
204-
; GFX9-NEXT: br label [[ATOMICRMW_START:%.*]]
205-
; GFX9: atomicrmw.start:
206-
; GFX9-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
207-
; GFX9-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE:%.*]])
208-
; GFX9-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i64
209-
; GFX9-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
210-
; GFX9-NEXT: [[TMP5:%.*]] = cmpxchg ptr [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
211-
; GFX9-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
212-
; GFX9-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
213-
; GFX9-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
214-
; GFX9-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
215-
; GFX9: atomicrmw.end:
216-
; GFX9-NEXT: ret double [[TMP6]]
190+
; GCN-LABEL: @test_atomicrmw_fmax_f64_flat(
191+
; GCN-NEXT: [[TMP1:%.*]] = load double, ptr [[PTR:%.*]], align 8
192+
; GCN-NEXT: br label [[ATOMICRMW_START:%.*]]
193+
; GCN: atomicrmw.start:
194+
; GCN-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
195+
; GCN-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE:%.*]])
196+
; GCN-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i64
197+
; GCN-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
198+
; GCN-NEXT: [[TMP5:%.*]] = cmpxchg ptr [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
199+
; GCN-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
200+
; GCN-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
201+
; GCN-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
202+
; GCN-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
203+
; GCN: atomicrmw.end:
204+
; GCN-NEXT: ret double [[TMP6]]
217205
;
218206
%res = atomicrmw fmax ptr %ptr, double %value seq_cst
219207
ret double %res
220208
}
221209

222210
define double @test_atomicrmw_fmax_f64_global(ptr addrspace(1) %ptr, double %value) {
223-
; GFX7-LABEL: @test_atomicrmw_fmax_f64_global(
224-
; GFX7-NEXT: [[RES:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]] seq_cst, align 8
225-
; GFX7-NEXT: ret double [[RES]]
226-
;
227-
; GFX9-LABEL: @test_atomicrmw_fmax_f64_global(
228-
; GFX9-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR:%.*]], align 8
229-
; GFX9-NEXT: br label [[ATOMICRMW_START:%.*]]
230-
; GFX9: atomicrmw.start:
231-
; GFX9-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
232-
; GFX9-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE:%.*]])
233-
; GFX9-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i64
234-
; GFX9-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
235-
; GFX9-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
236-
; GFX9-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
237-
; GFX9-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
238-
; GFX9-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
239-
; GFX9-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
240-
; GFX9: atomicrmw.end:
241-
; GFX9-NEXT: ret double [[TMP6]]
211+
; GCN-LABEL: @test_atomicrmw_fmax_f64_global(
212+
; GCN-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR:%.*]], align 8
213+
; GCN-NEXT: br label [[ATOMICRMW_START:%.*]]
214+
; GCN: atomicrmw.start:
215+
; GCN-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
216+
; GCN-NEXT: [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE:%.*]])
217+
; GCN-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i64
218+
; GCN-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
219+
; GCN-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
220+
; GCN-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
221+
; GCN-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
222+
; GCN-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
223+
; GCN-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
224+
; GCN: atomicrmw.end:
225+
; GCN-NEXT: ret double [[TMP6]]
242226
;
243227
%res = atomicrmw fmax ptr addrspace(1) %ptr, double %value seq_cst
244228
ret double %res
@@ -254,26 +238,25 @@ define double @test_atomicrmw_fmax_f64_local(ptr addrspace(3) %ptr, double %valu
254238
}
255239

256240
define double @test_atomicrmw_fmax_f64_global_strictfp(ptr addrspace(1) %ptr, double %value) strictfp {
257-
; GFX7-LABEL: @test_atomicrmw_fmax_f64_global_strictfp(
258-
; GFX7-NEXT: [[RES:%.*]] = atomicrmw fmax ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]] seq_cst, align 8
259-
; GFX7-NEXT: ret double [[RES]]
260-
;
261-
; GFX9-LABEL: @test_atomicrmw_fmax_f64_global_strictfp(
262-
; GFX9-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR:%.*]], align 8
263-
; GFX9-NEXT: br label [[ATOMICRMW_START:%.*]]
264-
; GFX9: atomicrmw.start:
265-
; GFX9-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
266-
; GFX9-NEXT: [[TMP2:%.*]] = call double @llvm.experimental.constrained.maxnum.f64(double [[LOADED]], double [[VALUE:%.*]], metadata !"fpexcept.strict") #[[ATTR4:[0-9]+]]
267-
; GFX9-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i64
268-
; GFX9-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
269-
; GFX9-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
270-
; GFX9-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
271-
; GFX9-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
272-
; GFX9-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
273-
; GFX9-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
274-
; GFX9: atomicrmw.end:
275-
; GFX9-NEXT: ret double [[TMP6]]
241+
; GCN-LABEL: @test_atomicrmw_fmax_f64_global_strictfp(
242+
; GCN-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR:%.*]], align 8
243+
; GCN-NEXT: br label [[ATOMICRMW_START:%.*]]
244+
; GCN: atomicrmw.start:
245+
; GCN-NEXT: [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
246+
; GCN-NEXT: [[TMP2:%.*]] = call double @llvm.experimental.constrained.maxnum.f64(double [[LOADED]], double [[VALUE:%.*]], metadata !"fpexcept.strict") #[[ATTR4:[0-9]+]]
247+
; GCN-NEXT: [[TMP3:%.*]] = bitcast double [[TMP2]] to i64
248+
; GCN-NEXT: [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
249+
; GCN-NEXT: [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
250+
; GCN-NEXT: [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
251+
; GCN-NEXT: [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
252+
; GCN-NEXT: [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
253+
; GCN-NEXT: br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
254+
; GCN: atomicrmw.end:
255+
; GCN-NEXT: ret double [[TMP6]]
276256
;
277257
%res = atomicrmw fmax ptr addrspace(1) %ptr, double %value seq_cst
278258
ret double %res
279259
}
260+
;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
261+
; GFX7: {{.*}}
262+
; GFX9: {{.*}}

0 commit comments

Comments
 (0)