Skip to content

Commit 78ede9f

Browse files
shiltianshiltian
committed
[Attributor] Add support for atomic operations in AAAddressSpace 
1 parent 109cd11 commit 78ede9f

File tree

3 files changed

+171
-4
lines changed

3 files changed

+171
-4
lines changed

‎llvm/lib/Target/AMDGPU/AMDGPUAttributor.cpp

Lines changed: 7 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -1084,10 +1084,15 @@ static bool runImpl(Module &M, AnalysisGetter &AG, TargetMachine &TM,
10841084
if (auto *LI = dyn_cast<LoadInst>(&I)) {
10851085
A.getOrCreateAAFor<AAAddressSpace>(
10861086
IRPosition::value(*LI->getPointerOperand()));
1087-
}
1088-
if (auto *SI = dyn_cast<StoreInst>(&I)) {
1087+
} else if (auto *SI = dyn_cast<StoreInst>(&I)) {
10891088
A.getOrCreateAAFor<AAAddressSpace>(
10901089
IRPosition::value(*SI->getPointerOperand()));
1090+
} else if (auto *RMW = dyn_cast<AtomicRMWInst>(&I)) {
1091+
A.getOrCreateAAFor<AAAddressSpace>(
1092+
IRPosition::value(*RMW->getPointerOperand()));
1093+
} else if (auto *CmpX = dyn_cast<AtomicCmpXchgInst>(&I)) {
1094+
A.getOrCreateAAFor<AAAddressSpace>(
1095+
IRPosition::value(*CmpX->getPointerOperand()));
10911096
}
10921097
}
10931098
}

‎llvm/lib/Transforms/IPO/AttributorAttributes.cpp

Lines changed: 7 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -12584,10 +12584,15 @@ struct AAAddressSpaceImpl : public AAAddressSpace {
1258412584
if (auto *LI = dyn_cast<LoadInst>(Inst)) {
1258512585
Changed |=
1258612586
makeChange(A, LI, U, OriginalValue, NewPtrTy, UseOriginalValue);
12587-
}
12588-
if (auto *SI = dyn_cast<StoreInst>(Inst)) {
12587+
} else if (auto *SI = dyn_cast<StoreInst>(Inst)) {
1258912588
Changed |=
1259012589
makeChange(A, SI, U, OriginalValue, NewPtrTy, UseOriginalValue);
12590+
} else if (auto *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
12591+
Changed |=
12592+
makeChange(A, RMW, U, OriginalValue, NewPtrTy, UseOriginalValue);
12593+
} else if (auto *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
12594+
Changed |=
12595+
makeChange(A, CmpX, U, OriginalValue, NewPtrTy, UseOriginalValue);
1259112596
}
1259212597
return true;
1259312598
};

‎llvm/test/CodeGen/AMDGPU/aa-as-infer.ll

Lines changed: 157 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -2,6 +2,7 @@
22
; RUN: opt -mtriple=amdgcn-amd-amdhsa -passes=amdgpu-attributor -S %s -o - | FileCheck %s
33

44
@g1 = protected addrspace(1) externally_initialized global i32 0, align 4
5+
@g2 = protected addrspace(1) externally_initialized global i32 0, align 4
56

67
define internal void @volatile_load_store_as_0(ptr %p) {
78
; CHECK-LABEL: define internal void @volatile_load_store_as_0(
@@ -86,3 +87,159 @@ define void @call_volatile_load_store_as_4(ptr addrspace(4) %p1, ptr addrspace(4
8687
call void @volatile_load_store_as_1(ptr %p2.cast)
8788
ret void
8889
}
90+
91+
define internal void @can_infer_cmpxchg(ptr %word) {
92+
; CHECK-LABEL: define internal void @can_infer_cmpxchg(
93+
; CHECK-SAME: ptr [[WORD:%.*]]) #[[ATTR0]] {
94+
; CHECK-NEXT: [[TMP1:%.*]] = addrspacecast ptr [[WORD]] to ptr addrspace(1)
95+
; CHECK-NEXT: [[CMPXCHG_0:%.*]] = cmpxchg ptr addrspace(1) [[TMP1]], i32 0, i32 4 monotonic monotonic, align 4
96+
; CHECK-NEXT: [[TMP2:%.*]] = addrspacecast ptr [[WORD]] to ptr addrspace(1)
97+
; CHECK-NEXT: [[CMPXCHG_1:%.*]] = cmpxchg ptr addrspace(1) [[TMP2]], i32 0, i32 5 acq_rel monotonic, align 4
98+
; CHECK-NEXT: [[TMP3:%.*]] = addrspacecast ptr [[WORD]] to ptr addrspace(1)
99+
; CHECK-NEXT: [[CMPXCHG_2:%.*]] = cmpxchg ptr addrspace(1) [[TMP3]], i32 0, i32 6 acquire monotonic, align 4
100+
; CHECK-NEXT: [[TMP4:%.*]] = addrspacecast ptr [[WORD]] to ptr addrspace(1)
101+
; CHECK-NEXT: [[CMPXCHG_3:%.*]] = cmpxchg ptr addrspace(1) [[TMP4]], i32 0, i32 7 release monotonic, align 4
102+
; CHECK-NEXT: [[TMP5:%.*]] = addrspacecast ptr [[WORD]] to ptr addrspace(1)
103+
; CHECK-NEXT: [[CMPXCHG_4:%.*]] = cmpxchg ptr addrspace(1) [[TMP5]], i32 0, i32 8 seq_cst monotonic, align 4
104+
; CHECK-NEXT: [[TMP6:%.*]] = addrspacecast ptr [[WORD]] to ptr addrspace(1)
105+
; CHECK-NEXT: [[CMPXCHG_5:%.*]] = cmpxchg weak ptr addrspace(1) [[TMP6]], i32 0, i32 9 seq_cst monotonic, align 4
106+
; CHECK-NEXT: [[CMPXCHG_6:%.*]] = cmpxchg volatile ptr [[WORD]], i32 0, i32 10 seq_cst monotonic, align 4
107+
; CHECK-NEXT: [[CMPXCHG_7:%.*]] = cmpxchg weak volatile ptr [[WORD]], i32 0, i32 11 syncscope("singlethread") seq_cst monotonic, align 4
108+
; CHECK-NEXT: ret void
109+
;
110+
%cmpxchg.0 = cmpxchg ptr %word, i32 0, i32 4 monotonic monotonic, align 4
111+
%cmpxchg.1 = cmpxchg ptr %word, i32 0, i32 5 acq_rel monotonic, align 4
112+
%cmpxchg.2 = cmpxchg ptr %word, i32 0, i32 6 acquire monotonic, align 4
113+
%cmpxchg.3 = cmpxchg ptr %word, i32 0, i32 7 release monotonic, align 4
114+
%cmpxchg.4 = cmpxchg ptr %word, i32 0, i32 8 seq_cst monotonic, align 4
115+
%cmpxchg.5 = cmpxchg weak ptr %word, i32 0, i32 9 seq_cst monotonic, align 4
116+
%cmpxchg.6 = cmpxchg volatile ptr %word, i32 0, i32 10 seq_cst monotonic, align 4
117+
%cmpxchg.7 = cmpxchg weak volatile ptr %word, i32 0, i32 11 syncscope("singlethread") seq_cst monotonic, align 4
118+
ret void
119+
}
120+
121+
define internal void @can_not_infer_cmpxchg(ptr %word) {
122+
; CHECK-LABEL: define internal void @can_not_infer_cmpxchg(
123+
; CHECK-SAME: ptr [[WORD:%.*]]) #[[ATTR0]] {
124+
; CHECK-NEXT: [[CMPXCHG_0:%.*]] = cmpxchg ptr [[WORD]], i32 0, i32 4 monotonic monotonic, align 4
125+
; CHECK-NEXT: [[CMPXCHG_1:%.*]] = cmpxchg ptr [[WORD]], i32 0, i32 5 acq_rel monotonic, align 4
126+
; CHECK-NEXT: [[CMPXCHG_2:%.*]] = cmpxchg ptr [[WORD]], i32 0, i32 6 acquire monotonic, align 4
127+
; CHECK-NEXT: [[CMPXCHG_3:%.*]] = cmpxchg ptr [[WORD]], i32 0, i32 7 release monotonic, align 4
128+
; CHECK-NEXT: [[CMPXCHG_4:%.*]] = cmpxchg ptr [[WORD]], i32 0, i32 8 seq_cst monotonic, align 4
129+
; CHECK-NEXT: [[CMPXCHG_5:%.*]] = cmpxchg weak ptr [[WORD]], i32 0, i32 9 seq_cst monotonic, align 4
130+
; CHECK-NEXT: [[CMPXCHG_6:%.*]] = cmpxchg volatile ptr [[WORD]], i32 0, i32 10 seq_cst monotonic, align 4
131+
; CHECK-NEXT: [[CMPXCHG_7:%.*]] = cmpxchg weak volatile ptr [[WORD]], i32 0, i32 11 syncscope("singlethread") seq_cst monotonic, align 4
132+
; CHECK-NEXT: ret void
133+
;
134+
%cmpxchg.0 = cmpxchg ptr %word, i32 0, i32 4 monotonic monotonic, align 4
135+
%cmpxchg.1 = cmpxchg ptr %word, i32 0, i32 5 acq_rel monotonic, align 4
136+
%cmpxchg.2 = cmpxchg ptr %word, i32 0, i32 6 acquire monotonic, align 4
137+
%cmpxchg.3 = cmpxchg ptr %word, i32 0, i32 7 release monotonic, align 4
138+
%cmpxchg.4 = cmpxchg ptr %word, i32 0, i32 8 seq_cst monotonic, align 4
139+
%cmpxchg.5 = cmpxchg weak ptr %word, i32 0, i32 9 seq_cst monotonic, align 4
140+
%cmpxchg.6 = cmpxchg volatile ptr %word, i32 0, i32 10 seq_cst monotonic, align 4
141+
%cmpxchg.7 = cmpxchg weak volatile ptr %word, i32 0, i32 11 syncscope("singlethread") seq_cst monotonic, align 4
142+
ret void
143+
}
144+
145+
define internal void @can_infer_atomicrmw(ptr %word) {
146+
; CHECK-LABEL: define internal void @can_infer_atomicrmw(
147+
; CHECK-SAME: ptr [[WORD:%.*]]) #[[ATTR0]] {
148+
; CHECK-NEXT: [[TMP1:%.*]] = addrspacecast ptr [[WORD]] to ptr addrspace(1)
149+
; CHECK-NEXT: [[ATOMICRMW_XCHG:%.*]] = atomicrmw xchg ptr addrspace(1) [[TMP1]], i32 12 monotonic, align 4
150+
; CHECK-NEXT: [[TMP2:%.*]] = addrspacecast ptr [[WORD]] to ptr addrspace(1)
151+
; CHECK-NEXT: [[ATOMICRMW_ADD:%.*]] = atomicrmw add ptr addrspace(1) [[TMP2]], i32 13 monotonic, align 4
152+
; CHECK-NEXT: [[TMP3:%.*]] = addrspacecast ptr [[WORD]] to ptr addrspace(1)
153+
; CHECK-NEXT: [[ATOMICRMW_SUB:%.*]] = atomicrmw sub ptr addrspace(1) [[TMP3]], i32 14 monotonic, align 4
154+
; CHECK-NEXT: [[TMP4:%.*]] = addrspacecast ptr [[WORD]] to ptr addrspace(1)
155+
; CHECK-NEXT: [[ATOMICRMW_AND:%.*]] = atomicrmw and ptr addrspace(1) [[TMP4]], i32 15 monotonic, align 4
156+
; CHECK-NEXT: [[TMP5:%.*]] = addrspacecast ptr [[WORD]] to ptr addrspace(1)
157+
; CHECK-NEXT: [[ATOMICRMW_NAND:%.*]] = atomicrmw nand ptr addrspace(1) [[TMP5]], i32 16 monotonic, align 4
158+
; CHECK-NEXT: [[TMP6:%.*]] = addrspacecast ptr [[WORD]] to ptr addrspace(1)
159+
; CHECK-NEXT: [[ATOMICRMW_OR:%.*]] = atomicrmw or ptr addrspace(1) [[TMP6]], i32 17 monotonic, align 4
160+
; CHECK-NEXT: [[TMP7:%.*]] = addrspacecast ptr [[WORD]] to ptr addrspace(1)
161+
; CHECK-NEXT: [[ATOMICRMW_XOR:%.*]] = atomicrmw xor ptr addrspace(1) [[TMP7]], i32 18 monotonic, align 4
162+
; CHECK-NEXT: [[TMP8:%.*]] = addrspacecast ptr [[WORD]] to ptr addrspace(1)
163+
; CHECK-NEXT: [[ATOMICRMW_MAX:%.*]] = atomicrmw max ptr addrspace(1) [[TMP8]], i32 19 monotonic, align 4
164+
; CHECK-NEXT: [[ATOMICRMW_MIN:%.*]] = atomicrmw volatile min ptr [[WORD]], i32 20 monotonic, align 4
165+
; CHECK-NEXT: [[TMP10:%.*]] = addrspacecast ptr [[WORD]] to ptr addrspace(1)
166+
; CHECK-NEXT: [[ATOMICRMW_UMAX:%.*]] = atomicrmw umax ptr addrspace(1) [[TMP10]], i32 21 syncscope("singlethread") monotonic, align 4
167+
; CHECK-NEXT: [[ATOMICRMW_UMIN:%.*]] = atomicrmw volatile umin ptr [[WORD]], i32 22 syncscope("singlethread") monotonic, align 4
168+
; CHECK-NEXT: ret void
169+
;
170+
%atomicrmw.xchg = atomicrmw xchg ptr %word, i32 12 monotonic, align 4
171+
%atomicrmw.add = atomicrmw add ptr %word, i32 13 monotonic, align 4
172+
%atomicrmw.sub = atomicrmw sub ptr %word, i32 14 monotonic, align 4
173+
%atomicrmw.and = atomicrmw and ptr %word, i32 15 monotonic, align 4
174+
%atomicrmw.nand = atomicrmw nand ptr %word, i32 16 monotonic, align 4
175+
%atomicrmw.or = atomicrmw or ptr %word, i32 17 monotonic, align 4
176+
%atomicrmw.xor = atomicrmw xor ptr %word, i32 18 monotonic, align 4
177+
%atomicrmw.max = atomicrmw max ptr %word, i32 19 monotonic, align 4
178+
%atomicrmw.min = atomicrmw volatile min ptr %word, i32 20 monotonic, align 4
179+
%atomicrmw.umax = atomicrmw umax ptr %word, i32 21 syncscope("singlethread") monotonic, align 4
180+
%atomicrmw.umin = atomicrmw volatile umin ptr %word, i32 22 syncscope("singlethread") monotonic, align 4
181+
ret void
182+
}
183+
184+
define internal void @can_not_infer_atomicrmw(ptr %word) {
185+
; CHECK-LABEL: define internal void @can_not_infer_atomicrmw(
186+
; CHECK-SAME: ptr [[WORD:%.*]]) #[[ATTR0]] {
187+
; CHECK-NEXT: [[ATOMICRMW_XCHG:%.*]] = atomicrmw xchg ptr [[WORD]], i32 12 monotonic, align 4
188+
; CHECK-NEXT: [[ATOMICRMW_ADD:%.*]] = atomicrmw add ptr [[WORD]], i32 13 monotonic, align 4
189+
; CHECK-NEXT: [[ATOMICRMW_SUB:%.*]] = atomicrmw sub ptr [[WORD]], i32 14 monotonic, align 4
190+
; CHECK-NEXT: [[ATOMICRMW_AND:%.*]] = atomicrmw and ptr [[WORD]], i32 15 monotonic, align 4
191+
; CHECK-NEXT: [[ATOMICRMW_NAND:%.*]] = atomicrmw nand ptr [[WORD]], i32 16 monotonic, align 4
192+
; CHECK-NEXT: [[ATOMICRMW_OR:%.*]] = atomicrmw or ptr [[WORD]], i32 17 monotonic, align 4
193+
; CHECK-NEXT: [[ATOMICRMW_XOR:%.*]] = atomicrmw xor ptr [[WORD]], i32 18 monotonic, align 4
194+
; CHECK-NEXT: [[ATOMICRMW_MAX:%.*]] = atomicrmw max ptr [[WORD]], i32 19 monotonic, align 4
195+
; CHECK-NEXT: [[ATOMICRMW_MIN:%.*]] = atomicrmw volatile min ptr [[WORD]], i32 20 monotonic, align 4
196+
; CHECK-NEXT: [[ATOMICRMW_UMAX:%.*]] = atomicrmw umax ptr [[WORD]], i32 21 syncscope("singlethread") monotonic, align 4
197+
; CHECK-NEXT: [[ATOMICRMW_UMIN:%.*]] = atomicrmw volatile umin ptr [[WORD]], i32 22 syncscope("singlethread") monotonic, align 4
198+
; CHECK-NEXT: ret void
199+
;
200+
%atomicrmw.xchg = atomicrmw xchg ptr %word, i32 12 monotonic, align 4
201+
%atomicrmw.add = atomicrmw add ptr %word, i32 13 monotonic, align 4
202+
%atomicrmw.sub = atomicrmw sub ptr %word, i32 14 monotonic, align 4
203+
%atomicrmw.and = atomicrmw and ptr %word, i32 15 monotonic, align 4
204+
%atomicrmw.nand = atomicrmw nand ptr %word, i32 16 monotonic, align 4
205+
%atomicrmw.or = atomicrmw or ptr %word, i32 17 monotonic, align 4
206+
%atomicrmw.xor = atomicrmw xor ptr %word, i32 18 monotonic, align 4
207+
%atomicrmw.max = atomicrmw max ptr %word, i32 19 monotonic, align 4
208+
%atomicrmw.min = atomicrmw volatile min ptr %word, i32 20 monotonic, align 4
209+
%atomicrmw.umax = atomicrmw umax ptr %word, i32 21 syncscope("singlethread") monotonic, align 4
210+
%atomicrmw.umin = atomicrmw volatile umin ptr %word, i32 22 syncscope("singlethread") monotonic, align 4
211+
ret void
212+
}
213+
214+
define void @foo(ptr addrspace(3) %val) {
215+
; CHECK-LABEL: define void @foo(
216+
; CHECK-SAME: ptr addrspace(3) [[VAL:%.*]]) #[[ATTR1:[0-9]+]] {
217+
; CHECK-NEXT: [[VAL_CAST:%.*]] = addrspacecast ptr addrspace(3) [[VAL]] to ptr
218+
; CHECK-NEXT: call void @can_infer_cmpxchg(ptr addrspacecast (ptr addrspace(1) @g1 to ptr))
219+
; CHECK-NEXT: call void @can_infer_cmpxchg(ptr addrspacecast (ptr addrspace(1) @g2 to ptr))
220+
; CHECK-NEXT: call void @can_not_infer_cmpxchg(ptr addrspacecast (ptr addrspace(1) @g1 to ptr))
221+
; CHECK-NEXT: call void @can_not_infer_cmpxchg(ptr addrspacecast (ptr addrspace(1) @g2 to ptr))
222+
; CHECK-NEXT: call void @can_not_infer_cmpxchg(ptr [[VAL_CAST]])
223+
; CHECK-NEXT: call void @can_infer_atomicrmw(ptr addrspacecast (ptr addrspace(1) @g1 to ptr))
224+
; CHECK-NEXT: call void @can_infer_atomicrmw(ptr addrspacecast (ptr addrspace(1) @g2 to ptr))
225+
; CHECK-NEXT: call void @can_not_infer_atomicrmw(ptr addrspacecast (ptr addrspace(1) @g1 to ptr))
226+
; CHECK-NEXT: call void @can_not_infer_atomicrmw(ptr addrspacecast (ptr addrspace(1) @g2 to ptr))
227+
; CHECK-NEXT: call void @can_not_infer_atomicrmw(ptr [[VAL_CAST]])
228+
; CHECK-NEXT: ret void
229+
;
230+
%g1.cast = addrspacecast ptr addrspace(1) @g1 to ptr
231+
%g2.cast = addrspacecast ptr addrspace(1) @g2 to ptr
232+
%val.cast = addrspacecast ptr addrspace(3) %val to ptr
233+
call void @can_infer_cmpxchg(ptr %g1.cast)
234+
call void @can_infer_cmpxchg(ptr %g2.cast)
235+
call void @can_not_infer_cmpxchg(ptr %g1.cast)
236+
call void @can_not_infer_cmpxchg(ptr %g2.cast)
237+
call void @can_not_infer_cmpxchg(ptr %val.cast)
238+
call void @can_infer_atomicrmw(ptr %g1.cast)
239+
call void @can_infer_atomicrmw(ptr %g2.cast)
240+
call void @can_not_infer_atomicrmw(ptr %g1.cast)
241+
call void @can_not_infer_atomicrmw(ptr %g2.cast)
242+
call void @can_not_infer_atomicrmw(ptr %val.cast)
243+
ret void
244+
}
245+

0 commit comments

Comments
 (0)