Skip to content

Commit f2c41d4

Browse files
hyphyp
authored
Merge pull request #3074 from aschwaighofer/coro_async_cherry_picks_mangling_single_pred_phis
Cherry-pick coro async mangling and single predecessor phi fix
2 parents 9ad294d + 7438d49 commit f2c41d4

File tree

5 files changed

+273
-22
lines changed

5 files changed

+273
-22
lines changed

‎llvm/lib/Transforms/Coroutines/CoroFrame.cpp

Lines changed: 22 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1839,6 +1839,24 @@ static void rewritePHIsForCleanupPad(BasicBlock *CleanupPadBB,
18391839
}
18401840
}
18411841

1842+
static void cleanupSinglePredPHIs(Function &F) {
1843+
SmallVector<PHINode *, 32> Worklist;
1844+
for (auto &BB : F) {
1845+
for (auto &Phi : BB.phis()) {
1846+
if (Phi.getNumIncomingValues() == 1) {
1847+
Worklist.push_back(&Phi);
1848+
} else
1849+
break;
1850+
}
1851+
}
1852+
while (!Worklist.empty()) {
1853+
auto *Phi = Worklist.back();
1854+
Worklist.pop_back();
1855+
auto *OriginalValue = Phi->getIncomingValue(0);
1856+
Phi->replaceAllUsesWith(OriginalValue);
1857+
}
1858+
}
1859+
18421860
static void rewritePHIs(BasicBlock &BB) {
18431861
// For every incoming edge we will create a block holding all
18441862
// incoming values in a single PHI nodes.
@@ -2602,6 +2620,10 @@ void coro::buildCoroutineFrame(Function &F, Shape &Shape) {
26022620
}
26032621
}
26042622

2623+
// Later code makes structural assumptions about single predecessors phis e.g
2624+
// that they are not live accross a suspend point.
2625+
cleanupSinglePredPHIs(F);
2626+
26052627
// Transforms multi-edge PHI Nodes, so that any value feeding into a PHI will
26062628
// never has its definition separated from the PHI by the suspend point.
26072629
rewritePHIs(F);

‎llvm/lib/Transforms/Coroutines/CoroSplit.cpp

Lines changed: 38 additions & 12 deletions
Original file line numberDiff line numberDiff line change
@@ -765,8 +765,8 @@ Value *CoroCloner::deriveNewFramePointer() {
765765
// context header.
766766
case coro::ABI::Async: {
767767
auto *ActiveAsyncSuspend = cast<CoroSuspendAsyncInst>(ActiveSuspend);
768-
auto *CalleeContext =
769-
NewF->getArg(ActiveAsyncSuspend->getStorageArgumentIndex());
768+
auto ContextIdx = ActiveAsyncSuspend->getStorageArgumentIndex() & 0xff;
769+
auto *CalleeContext = NewF->getArg(ContextIdx);
770770
auto *FramePtrTy = Shape.FrameTy->getPointerTo();
771771
auto *ProjectionFunc =
772772
ActiveAsyncSuspend->getAsyncContextProjectionFunction();
@@ -827,6 +827,13 @@ static void addAsyncContextAttrs(AttributeList &Attrs, LLVMContext &Context,
827827
Attrs = Attrs.addParamAttributes(Context, ParamIndex, ParamAttrs);
828828
}
829829

830+
static void addSwiftSelfAttrs(AttributeList &Attrs, LLVMContext &Context,
831+
unsigned ParamIndex) {
832+
AttrBuilder ParamAttrs;
833+
ParamAttrs.addAttribute(Attribute::SwiftSelf);
834+
Attrs = Attrs.addParamAttributes(Context, ParamIndex, ParamAttrs);
835+
}
836+
830837
/// Clone the body of the original function into a resume function of
831838
/// some sort.
832839
void CoroCloner::create() {
@@ -906,10 +913,21 @@ void CoroCloner::create() {
906913
Shape.FrameSize, Shape.FrameAlign);
907914
break;
908915
case coro::ABI::Async: {
916+
auto *ActiveAsyncSuspend = cast<CoroSuspendAsyncInst>(ActiveSuspend);
909917
if (OrigF.hasParamAttribute(Shape.AsyncLowering.ContextArgNo,
910918
Attribute::SwiftAsync)) {
911-
addAsyncContextAttrs(NewAttrs, Context, Shape.AsyncLowering.ContextArgNo);
919+
uint32_t ArgAttributeIndices =
920+
ActiveAsyncSuspend->getStorageArgumentIndex();
921+
auto ContextArgIndex = ArgAttributeIndices & 0xff;
922+
addAsyncContextAttrs(NewAttrs, Context, ContextArgIndex);
923+
924+
// `swiftasync` must preceed `swiftself` so 0 is not a valid index for
925+
// `swiftself`.
926+
auto SwiftSelfIndex = ArgAttributeIndices >> 8;
927+
if (SwiftSelfIndex)
928+
addSwiftSelfAttrs(NewAttrs, Context, SwiftSelfIndex);
912929
}
930+
913931
// Transfer the original function's attributes.
914932
auto FnAttrs = OrigF.getAttributes().getFnAttributes();
915933
NewAttrs =
@@ -949,16 +967,9 @@ void CoroCloner::create() {
949967
// followed by a return.
950968
// Don't change returns to unreachable because that will trip up the verifier.
951969
// These returns should be unreachable from the clone.
952-
case coro::ABI::Async: {
953-
auto *ActiveAsyncSuspend = cast<CoroSuspendAsyncInst>(ActiveSuspend);
954-
if (OrigF.hasParamAttribute(Shape.AsyncLowering.ContextArgNo,
955-
Attribute::SwiftAsync)) {
956-
auto ContextArgIndex = ActiveAsyncSuspend->getStorageArgumentIndex();
957-
addAsyncContextAttrs(NewAttrs, Context, ContextArgIndex);
958-
}
970+
case coro::ABI::Async:
959971
break;
960972
}
961-
}
962973

963974
NewF->setAttributes(NewAttrs);
964975
NewF->setCallingConv(Shape.getResumeFunctionCC());
@@ -1600,8 +1611,23 @@ static void splitAsyncCoroutine(Function &F, coro::Shape &Shape,
16001611
auto *Suspend = cast<CoroSuspendAsyncInst>(Shape.CoroSuspends[Idx]);
16011612

16021613
// Create the clone declaration.
1614+
auto ResumeNameSuffix = ".resume.";
1615+
auto ProjectionFunctionName =
1616+
Suspend->getAsyncContextProjectionFunction()->getName();
1617+
bool UseSwiftMangling = false;
1618+
if (ProjectionFunctionName.equals("__swift_async_resume_project_context")) {
1619+
ResumeNameSuffix = "TQ";
1620+
UseSwiftMangling = true;
1621+
} else if (ProjectionFunctionName.equals(
1622+
"__swift_async_resume_get_context")) {
1623+
ResumeNameSuffix = "TY";
1624+
UseSwiftMangling = true;
1625+
}
16031626
auto *Continuation = createCloneDeclaration(
1604-
F, Shape, ".resume." + Twine(Idx), NextF, Suspend);
1627+
F, Shape,
1628+
UseSwiftMangling ? ResumeNameSuffix + Twine(Idx) + "_"
1629+
: ResumeNameSuffix + Twine(Idx),
1630+
NextF, Suspend);
16051631
Clones.push_back(Continuation);
16061632

16071633
// Insert a branch to a new return block immediately before the suspend

‎llvm/test/Transforms/Coroutines/coro-async-phi.ll

Lines changed: 196 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,196 @@
1+
; RUN: opt < %s -enable-coroutines -O0 -S | FileCheck --check-prefixes=CHECK %s
2+
; RUN: opt < %s -enable-coroutines -passes='default<O0>' -S | FileCheck --check-prefixes=CHECK %s
3+
4+
%swift.async_func_pointer = type <{ i32, i32 }>
5+
%swift.context = type { %swift.context*, void (%swift.context*)*, i64 }
6+
%T10RR13AC = type <{ %swift.refcounted, %swift.defaultactor }>
7+
%swift.refcounted = type { %swift.type*, i64 }
8+
%swift.type = type { i64 }
9+
%swift.defaultactor = type { [10 x i8*] }
10+
%swift.bridge = type opaque
11+
%swift.error = type opaque
12+
%swift.executor = type {}
13+
14+
@repoTU = hidden global %swift.async_func_pointer <{ i32 trunc (i64 sub (i64 ptrtoint (void (%swift.context*, i64, i64, %T10RR13AC*)* @repo to i64), i64 ptrtoint (%swift.async_func_pointer* @repoTU to i64)) to i32), i32 20 }>, section "__TEXT,__const", align 8
15+
16+
declare void @use(i8*)
17+
18+
; This used to crash.
19+
; CHECK: repo
20+
define hidden swifttailcc void @repo(%swift.context* swiftasync %arg, i64 %arg1, i64 %arg2, %T10RR13AC* swiftself %arg3) #0 {
21+
entry:
22+
%i = alloca %swift.context*, align 8
23+
%i11 = call token @llvm.coro.id.async(i32 20, i32 16, i32 0, i8* bitcast (%swift.async_func_pointer* @repoTU to i8*))
24+
%i12 = call i8* @llvm.coro.begin(token %i11, i8* null)
25+
%i18 = call i8* @llvm.coro.async.resume()
26+
call void @use(i8* %i18)
27+
%i21 = call { i8* } (i32, i8*, i8*, ...) @llvm.coro.suspend.async.sl_p0i8s(i32 0, i8* %i18, i8* bitcast (i8* (i8*)* @__swift_async_resume_get_context to i8*), i8* bitcast (void (i8*, %swift.executor*, %swift.context*)* @__swift_suspend_point to i8*), i8* %i18, %swift.executor* null, %swift.context* null)
28+
%i22 = extractvalue { i8* } %i21, 0
29+
%i23 = call i8* @__swift_async_resume_get_context(i8* %i22)
30+
%i28 = icmp eq i64 %arg2, 0
31+
br i1 %i28, label %bb126, label %bb
32+
33+
bb: ; preds = %entry
34+
%i29 = inttoptr i64 %arg2 to %swift.bridge*
35+
br label %bb30
36+
37+
bb30: ; preds = %bb
38+
%i31 = phi i64 [ %arg1, %bb ]
39+
%i32 = phi %swift.bridge* [ %i29, %bb ]
40+
%i35 = ptrtoint %swift.bridge* %i32 to i64
41+
%i36 = bitcast %T10RR13AC* %arg3 to %swift.type**
42+
%i37 = load %swift.type*, %swift.type** %i36, align 8
43+
%i38 = bitcast %swift.type* %i37 to void (%swift.context*, i64, i64, %T10RR13AC*)**
44+
%i39 = getelementptr inbounds void (%swift.context*, i64, i64, %T10RR13AC*)*, void (%swift.context*, i64, i64, %T10RR13AC*)** %i38, i64 11
45+
%i40 = load void (%swift.context*, i64, i64, %T10RR13AC*)*, void (%swift.context*, i64, i64, %T10RR13AC*)** %i39, align 8
46+
%i41 = bitcast void (%swift.context*, i64, i64, %T10RR13AC*)* %i40 to %swift.async_func_pointer*
47+
%i42 = getelementptr inbounds %swift.async_func_pointer, %swift.async_func_pointer* %i41, i32 0, i32 0
48+
%i43 = load i32, i32* %i42, align 8
49+
%i44 = sext i32 %i43 to i64
50+
%i45 = ptrtoint i32* %i42 to i64
51+
%i46 = add i64 %i45, %i44
52+
%i47 = inttoptr i64 %i46 to i8*
53+
%i48 = bitcast i8* %i47 to void (%swift.context*, i64, i64, %T10RR13AC*)*
54+
%i52 = call swiftcc i8* @swift_task_alloc(i64 24) #1
55+
%i53 = bitcast i8* %i52 to <{ %swift.context*, void (%swift.context*)*, i32 }>*
56+
%i54 = load %swift.context*, %swift.context** %i, align 8
57+
%i55 = getelementptr inbounds <{ %swift.context*, void (%swift.context*)*, i32 }>, <{ %swift.context*, void (%swift.context*)*, i32 }>* %i53, i32 0, i32 0
58+
store %swift.context* %i54, %swift.context** %i55, align 8
59+
%i56 = call i8* @llvm.coro.async.resume()
60+
call void @use(i8* %i56)
61+
%i57 = bitcast i8* %i56 to void (%swift.context*)*
62+
%i58 = getelementptr inbounds <{ %swift.context*, void (%swift.context*)*, i32 }>, <{ %swift.context*, void (%swift.context*)*, i32 }>* %i53, i32 0, i32 1
63+
store void (%swift.context*)* %i57, void (%swift.context*)** %i58, align 8
64+
%i59 = bitcast i8* %i52 to %swift.context*
65+
%i60 = bitcast void (%swift.context*, i64, i64, %T10RR13AC*)* %i48 to i8*
66+
%i61 = call { i8*, %swift.error* } (i32, i8*, i8*, ...) @llvm.coro.suspend.async.sl_p0i8p0s_swift.errorss(i32 256, i8* %i56, i8* bitcast (i8* (i8*)* @__swift_async_resume_project_context to i8*), i8* bitcast (void (i8*, %swift.context*, i64, i64, %T10RR13AC*)* @__swift_suspend_dispatch_4 to i8*), i8* %i60, %swift.context* %i59, i64 %i31, i64 0, %T10RR13AC* %arg3)
67+
%i62 = extractvalue { i8*, %swift.error* } %i61, 0
68+
%i63 = call i8* @__swift_async_resume_project_context(i8* %i62)
69+
%i64 = bitcast i8* %i63 to %swift.context*
70+
store %swift.context* %i64, %swift.context** %i, align 8
71+
%i65 = extractvalue { i8*, %swift.error* } %i61, 1
72+
call swiftcc void @swift_task_dealloc(i8* %i52) #1
73+
br i1 %i28, label %bb126, label %bb68
74+
75+
bb68: ; preds = %bb30
76+
%i69 = call i8* @llvm.coro.async.resume()
77+
call void @use(i8* %i69)
78+
%i70 = load %swift.context*, %swift.context** %i, align 8
79+
%i71 = call { i8* } (i32, i8*, i8*, ...) @llvm.coro.suspend.async.sl_p0i8s(i32 0, i8* %i69, i8* bitcast (i8* (i8*)* @__swift_async_resume_get_context to i8*), i8* bitcast (void (i8*, %swift.executor*, %swift.context*)* @__swift_suspend_point to i8*), i8* %i69, %swift.executor* null, %swift.context* %i70)
80+
%i77 = ptrtoint %swift.bridge* %i32 to i64
81+
%i78 = bitcast %T10RR13AC* %arg3 to %swift.type**
82+
%i79 = load %swift.type*, %swift.type** %i78, align 8
83+
%i80 = bitcast %swift.type* %i79 to void (%swift.context*, i64, i64, %T10RR13AC*)**
84+
%i81 = getelementptr inbounds void (%swift.context*, i64, i64, %T10RR13AC*)*, void (%swift.context*, i64, i64, %T10RR13AC*)** %i80, i64 11
85+
%i82 = load void (%swift.context*, i64, i64, %T10RR13AC*)*, void (%swift.context*, i64, i64, %T10RR13AC*)** %i81, align 8
86+
%i83 = bitcast void (%swift.context*, i64, i64, %T10RR13AC*)* %i82 to %swift.async_func_pointer*
87+
%i84 = getelementptr inbounds %swift.async_func_pointer, %swift.async_func_pointer* %i83, i32 0, i32 0
88+
%i85 = load i32, i32* %i84, align 8
89+
%i86 = sext i32 %i85 to i64
90+
%i87 = ptrtoint i32* %i84 to i64
91+
%i88 = add i64 %i87, %i86
92+
%i89 = inttoptr i64 %i88 to i8*
93+
%i90 = bitcast i8* %i89 to void (%swift.context*, i64, i64, %T10RR13AC*)*
94+
%i94 = call swiftcc i8* @swift_task_alloc(i64 24) #1
95+
%i98 = call i8* @llvm.coro.async.resume()
96+
call void @use(i8* %i98)
97+
%i102 = bitcast void (%swift.context*, i64, i64, %T10RR13AC*)* %i90 to i8*
98+
%i103 = call { i8*, %swift.error* } (i32, i8*, i8*, ...) @llvm.coro.suspend.async.sl_p0i8p0s_swift.errorss(i32 256, i8* %i98, i8* bitcast (i8* (i8*)* @__swift_async_resume_project_context to i8*), i8* bitcast (void (i8*, %swift.context*, i64, i64, %T10RR13AC*)* @__swift_suspend_dispatch_4.1 to i8*), i8* %i102, %swift.context* null, i64 %i31, i64 0, %T10RR13AC* %arg3)
99+
call swiftcc void @swift_task_dealloc(i8* %i94) #1
100+
br label %bb126
101+
102+
bb126:
103+
%i162 = call i1 (i8*, i1, ...) @llvm.coro.end.async(i8* %i12, i1 false, void (i8*, %swift.context*, %swift.error*)* @__swift_suspend_dispatch_2, i8* bitcast (void (%swift.context*, %swift.error*)* @doIt to i8*), %swift.context* null, %swift.error* null)
104+
unreachable
105+
}
106+
107+
; Function Attrs: nounwind
108+
declare token @llvm.coro.id.async(i32, i32, i32, i8*) #1
109+
110+
; Function Attrs: nounwind
111+
declare i8* @llvm.coro.begin(token, i8* writeonly) #1
112+
113+
; Function Attrs: argmemonly nofree nosync nounwind willreturn writeonly
114+
declare void @llvm.memset.p0i8.i64(i8* nocapture writeonly, i8, i64, i1 immarg) #3
115+
116+
; Function Attrs: nounwind
117+
declare i8* @llvm.coro.async.resume() #1
118+
119+
; Function Attrs: noinline
120+
define linkonce_odr hidden i8* @__swift_async_resume_get_context(i8* %arg) #4 {
121+
entry:
122+
ret i8* %arg
123+
}
124+
125+
; Function Attrs: nounwind
126+
declare extern_weak swifttailcc void @swift_task_switch(%swift.context*, i8*, %swift.executor*) #1
127+
128+
; Function Attrs: nounwind
129+
define internal swifttailcc void @__swift_suspend_point(i8* %arg, %swift.executor* %arg1, %swift.context* %arg2) #1 {
130+
entry:
131+
musttail call swifttailcc void @swift_task_switch(%swift.context* swiftasync %arg2, i8* %arg, %swift.executor* %arg1) #1
132+
ret void
133+
}
134+
135+
; Function Attrs: nounwind
136+
declare { i8* } @llvm.coro.suspend.async.sl_p0i8s(i32, i8*, i8*, ...) #1
137+
138+
; Function Attrs: nounwind
139+
declare i1 @llvm.coro.end.async(i8*, i1, ...) #1
140+
141+
; Function Attrs: argmemonly nounwind
142+
declare extern_weak swiftcc i8* @swift_task_alloc(i64) #5
143+
144+
; Function Attrs: nounwind readnone
145+
declare i8** @llvm.swift.async.context.addr() #6
146+
147+
; Function Attrs: alwaysinline nounwind
148+
define linkonce_odr hidden i8* @__swift_async_resume_project_context(i8* %arg) #7 {
149+
entry:
150+
%i = bitcast i8* %arg to i8**
151+
%i1 = load i8*, i8** %i, align 8
152+
%i2 = call i8** @llvm.swift.async.context.addr()
153+
store i8* %i1, i8** %i2, align 8
154+
ret i8* %i1
155+
}
156+
157+
; Function Attrs: nounwind
158+
declare { i8*, %swift.error* } @llvm.coro.suspend.async.sl_p0i8p0s_swift.errorss(i32, i8*, i8*, ...) #1
159+
160+
; Function Attrs: argmemonly nounwind
161+
declare extern_weak swiftcc void @swift_task_dealloc(i8*) #5
162+
163+
; Function Attrs: nounwind
164+
define internal swifttailcc void @__swift_suspend_dispatch_4(i8* %arg, %swift.context* %arg1, i64 %arg2, i64 %arg3, %T10RR13AC* %arg4) #1 {
165+
entry:
166+
%i = bitcast i8* %arg to void (%swift.context*, i64, i64, %T10RR13AC*)*
167+
musttail call swifttailcc void %i(%swift.context* swiftasync %arg1, i64 %arg2, i64 %arg3, %T10RR13AC* swiftself %arg4)
168+
ret void
169+
}
170+
171+
declare swifttailcc void @doIt(%swift.context* swiftasync %arg1, %swift.error* swiftself %arg2)
172+
173+
; Function Attrs: nounwind
174+
define internal swifttailcc void @__swift_suspend_dispatch_2(i8* %arg, %swift.context* %arg1, %swift.error* %arg2) #1 {
175+
entry:
176+
%i = bitcast i8* %arg to void (%swift.context*, %swift.error*)*
177+
musttail call swifttailcc void %i(%swift.context* swiftasync %arg1, %swift.error* swiftself %arg2)
178+
ret void
179+
}
180+
181+
; Function Attrs: nounwind
182+
define internal swifttailcc void @__swift_suspend_dispatch_4.1(i8* %arg, %swift.context* %arg1, i64 %arg2, i64 %arg3, %T10RR13AC* %arg4) #1 {
183+
entry:
184+
%i = bitcast i8* %arg to void (%swift.context*, i64, i64, %T10RR13AC*)*
185+
musttail call swifttailcc void %i(%swift.context* swiftasync %arg1, i64 %arg2, i64 %arg3, %T10RR13AC* swiftself %arg4)
186+
ret void
187+
}
188+
189+
attributes #0 = { "frame-pointer"="all" }
190+
attributes #1 = { nounwind }
191+
attributes #2 = { argmemonly nofree nosync nounwind willreturn }
192+
attributes #3 = { argmemonly nofree nosync nounwind willreturn writeonly }
193+
attributes #4 = { noinline "frame-pointer"="all" }
194+
attributes #5 = { argmemonly nounwind }
195+
attributes #6 = { nounwind readnone }
196+
attributes #7 = { alwaysinline nounwind "frame-pointer"="all" }

0 commit comments

Comments
 (0)