[llvm] [AMDGPU] Fix incorrect stepping in gdb for amdgcn.end.cf intrinsic. (PR #83010)
Valery Pykhtin via llvm-commits
llvm-commits at lists.llvm.org
Tue Apr 30 07:25:34 PDT 2024
https://github.com/vpykhtin updated https://github.com/llvm/llvm-project/pull/83010
>From ea867d110dccf26bb0aa4bff8afe8d81d265e6c8 Mon Sep 17 00:00:00 2001
From: Valery Pykhtin <valery.pykhtin at gmail.com>
Date: Mon, 26 Feb 2024 15:31:59 +0100
Subject: [PATCH 1/2] [AMDGPU] Clear dbg info on CFG intrinsic due to incorrect
behaviour in gdb.
---
.../Target/AMDGPU/SIAnnotateControlFlow.cpp | 15 +++++---
.../CodeGen/AMDGPU/si-annotate-dbg-info.ll | 36 +++++++++----------
2 files changed, 29 insertions(+), 22 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/SIAnnotateControlFlow.cpp b/llvm/lib/Target/AMDGPU/SIAnnotateControlFlow.cpp
index 58214f30bb8d67..2ccdc496f36de8 100644
--- a/llvm/lib/Target/AMDGPU/SIAnnotateControlFlow.cpp
+++ b/llvm/lib/Target/AMDGPU/SIAnnotateControlFlow.cpp
@@ -207,6 +207,7 @@ bool SIAnnotateControlFlow::openIf(BranchInst *Term) {
return false;
IRBuilder<> IRB(Term);
+ IRB.SetCurrentDebugLocation(DebugLoc());
Value *IfCall = IRB.CreateCall(If, {Term->getCondition()});
Value *Cond = IRB.CreateExtractValue(IfCall, {0});
Value *Mask = IRB.CreateExtractValue(IfCall, {1});
@@ -222,6 +223,7 @@ bool SIAnnotateControlFlow::insertElse(BranchInst *Term) {
}
IRBuilder<> IRB(Term);
+ IRB.SetCurrentDebugLocation(DebugLoc());
Value *ElseCall = IRB.CreateCall(Else, {popSaved()});
Value *Cond = IRB.CreateExtractValue(ElseCall, {0});
Value *Mask = IRB.CreateExtractValue(ElseCall, {1});
@@ -235,7 +237,9 @@ Value *SIAnnotateControlFlow::handleLoopCondition(
Value *Cond, PHINode *Broken, llvm::Loop *L, BranchInst *Term) {
auto CreateBreak = [this, Cond, Broken](Instruction *I) -> CallInst * {
- return IRBuilder<>(I).CreateCall(IfBreak, {Cond, Broken});
+ IRBuilder<> IRB(I);
+ IRB.SetCurrentDebugLocation(DebugLoc());
+ return IRB.CreateCall(IfBreak, {Cond, Broken});
};
if (Instruction *Inst = dyn_cast<Instruction>(Cond)) {
@@ -296,7 +300,9 @@ bool SIAnnotateControlFlow::handleLoop(BranchInst *Term) {
Broken->addIncoming(PHIValue, Pred);
}
- CallInst *LoopCall = IRBuilder<>(Term).CreateCall(Loop, {Arg});
+ IRBuilder<> IRB(Term);
+ IRB.SetCurrentDebugLocation(DebugLoc());
+ CallInst *LoopCall = IRB.CreateCall(Loop, {Arg});
Term->setCondition(LoopCall);
push(Term->getSuccessor(0), Arg);
@@ -336,8 +342,9 @@ bool SIAnnotateControlFlow::closeControlFlow(BasicBlock *BB) {
// Split edge to make Def dominate Use
FirstInsertionPt = SplitEdge(DefBB, BB, DT, LI)->getFirstInsertionPt();
}
- IRBuilder<>(FirstInsertionPt->getParent(), FirstInsertionPt)
- .CreateCall(EndCf, {Exec});
+ IRBuilder<> IRB(FirstInsertionPt->getParent(), FirstInsertionPt);
+ IRB.SetCurrentDebugLocation(DebugLoc());
+ IRB.CreateCall(EndCf, {Exec});
}
return true;
diff --git a/llvm/test/CodeGen/AMDGPU/si-annotate-dbg-info.ll b/llvm/test/CodeGen/AMDGPU/si-annotate-dbg-info.ll
index a7af02017001fb..41828d163fba72 100644
--- a/llvm/test/CodeGen/AMDGPU/si-annotate-dbg-info.ll
+++ b/llvm/test/CodeGen/AMDGPU/si-annotate-dbg-info.ll
@@ -7,15 +7,15 @@ define amdgpu_ps i32 @if_else(i32 %0) !dbg !5 {
; OPT-SAME: i32 [[TMP0:%.*]]) !dbg [[DBG5:![0-9]+]] {
; OPT-NEXT: [[C:%.*]] = icmp ne i32 [[TMP0]], 0, !dbg [[DBG13:![0-9]+]]
; OPT-NEXT: tail call void @llvm.dbg.value(metadata i1 [[C]], metadata [[META9:![0-9]+]], metadata !DIExpression()), !dbg [[DBG13]]
-; OPT-NEXT: [[TMP2:%.*]] = call { i1, i64 } @llvm.amdgcn.if.i64(i1 [[C]]), !dbg [[DBG14:![0-9]+]]
-; OPT-NEXT: [[TMP3:%.*]] = extractvalue { i1, i64 } [[TMP2]], 0, !dbg [[DBG14]]
-; OPT-NEXT: [[TMP4:%.*]] = extractvalue { i1, i64 } [[TMP2]], 1, !dbg [[DBG14]]
-; OPT-NEXT: br i1 [[TMP3]], label [[FALSE:%.*]], label [[FLOW:%.*]], !dbg [[DBG14]]
+; OPT-NEXT: [[TMP2:%.*]] = call { i1, i64 } @llvm.amdgcn.if.i64(i1 [[C]])
+; OPT-NEXT: [[TMP3:%.*]] = extractvalue { i1, i64 } [[TMP2]], 0
+; OPT-NEXT: [[TMP4:%.*]] = extractvalue { i1, i64 } [[TMP2]], 1
+; OPT-NEXT: br i1 [[TMP3]], label [[FALSE:%.*]], label [[FLOW:%.*]], !dbg [[DBG14:![0-9]+]]
; OPT: Flow:
; OPT-NEXT: [[TMP5:%.*]] = phi i32 [ 33, [[FALSE]] ], [ undef, [[TMP1:%.*]] ]
-; OPT-NEXT: [[TMP6:%.*]] = call { i1, i64 } @llvm.amdgcn.else.i64.i64(i64 [[TMP4]]), !dbg [[DBG14]]
-; OPT-NEXT: [[TMP7:%.*]] = extractvalue { i1, i64 } [[TMP6]], 0, !dbg [[DBG14]]
-; OPT-NEXT: [[TMP8:%.*]] = extractvalue { i1, i64 } [[TMP6]], 1, !dbg [[DBG14]]
+; OPT-NEXT: [[TMP6:%.*]] = call { i1, i64 } @llvm.amdgcn.else.i64.i64(i64 [[TMP4]])
+; OPT-NEXT: [[TMP7:%.*]] = extractvalue { i1, i64 } [[TMP6]], 0
+; OPT-NEXT: [[TMP8:%.*]] = extractvalue { i1, i64 } [[TMP6]], 1
; OPT-NEXT: br i1 [[TMP7]], label [[TRUE:%.*]], label [[EXIT:%.*]], !dbg [[DBG14]]
; OPT: true:
; OPT-NEXT: br label [[EXIT]], !dbg [[DBG15:![0-9]+]]
@@ -23,9 +23,9 @@ define amdgpu_ps i32 @if_else(i32 %0) !dbg !5 {
; OPT-NEXT: br label [[FLOW]], !dbg [[DBG16:![0-9]+]]
; OPT: exit:
; OPT-NEXT: [[RET:%.*]] = phi i32 [ [[TMP5]], [[FLOW]] ], [ 42, [[TRUE]] ], !dbg [[DBG17:![0-9]+]]
-; OPT-NEXT: call void @llvm.amdgcn.end.cf.i64(i64 [[TMP8]]), !dbg [[DBG18:![0-9]+]]
+; OPT-NEXT: call void @llvm.amdgcn.end.cf.i64(i64 [[TMP8]])
; OPT-NEXT: tail call void @llvm.dbg.value(metadata i32 [[RET]], metadata [[META11:![0-9]+]], metadata !DIExpression()), !dbg [[DBG17]]
-; OPT-NEXT: ret i32 [[RET]], !dbg [[DBG18]]
+; OPT-NEXT: ret i32 [[RET]], !dbg [[DBG18:![0-9]+]]
;
%c = icmp eq i32 %0, 0, !dbg !13
tail call void @llvm.dbg.value(metadata i1 %c, metadata !9, metadata !DIExpression()), !dbg !13
@@ -54,10 +54,10 @@ define amdgpu_ps void @loop_if_break(i32 %n) !dbg !19 {
; OPT-NEXT: tail call void @llvm.dbg.value(metadata i32 [[I]], metadata [[META21:![0-9]+]], metadata !DIExpression()), !dbg [[DBG25]]
; OPT-NEXT: [[C:%.*]] = icmp ugt i32 [[I]], 0, !dbg [[DBG26:![0-9]+]]
; OPT-NEXT: tail call void @llvm.dbg.value(metadata i1 [[C]], metadata [[META22:![0-9]+]], metadata !DIExpression()), !dbg [[DBG26]]
-; OPT-NEXT: [[TMP0:%.*]] = call { i1, i64 } @llvm.amdgcn.if.i64(i1 [[C]]), !dbg [[DBG27:![0-9]+]]
-; OPT-NEXT: [[TMP1:%.*]] = extractvalue { i1, i64 } [[TMP0]], 0, !dbg [[DBG27]]
-; OPT-NEXT: [[TMP2:%.*]] = extractvalue { i1, i64 } [[TMP0]], 1, !dbg [[DBG27]]
-; OPT-NEXT: br i1 [[TMP1]], label [[LOOP_BODY:%.*]], label [[FLOW]], !dbg [[DBG27]]
+; OPT-NEXT: [[TMP0:%.*]] = call { i1, i64 } @llvm.amdgcn.if.i64(i1 [[C]])
+; OPT-NEXT: [[TMP1:%.*]] = extractvalue { i1, i64 } [[TMP0]], 0
+; OPT-NEXT: [[TMP2:%.*]] = extractvalue { i1, i64 } [[TMP0]], 1
+; OPT-NEXT: br i1 [[TMP1]], label [[LOOP_BODY:%.*]], label [[FLOW]], !dbg [[DBG27:![0-9]+]]
; OPT: loop_body:
; OPT-NEXT: [[I_NEXT:%.*]] = sub i32 [[I]], 1, !dbg [[DBG28:![0-9]+]]
; OPT-NEXT: tail call void @llvm.dbg.value(metadata i32 [[I_NEXT]], metadata [[META23:![0-9]+]], metadata !DIExpression()), !dbg [[DBG28]]
@@ -65,13 +65,13 @@ define amdgpu_ps void @loop_if_break(i32 %n) !dbg !19 {
; OPT: Flow:
; OPT-NEXT: [[TMP3]] = phi i32 [ [[I_NEXT]], [[LOOP_BODY]] ], [ undef, [[LOOP]] ]
; OPT-NEXT: [[TMP4:%.*]] = phi i1 [ false, [[LOOP_BODY]] ], [ true, [[LOOP]] ]
-; OPT-NEXT: call void @llvm.amdgcn.end.cf.i64(i64 [[TMP2]]), !dbg [[DBG27]]
-; OPT-NEXT: [[TMP5]] = call i64 @llvm.amdgcn.if.break.i64(i1 [[TMP4]], i64 [[PHI_BROKEN]]), !dbg [[DBG27]]
-; OPT-NEXT: [[TMP6:%.*]] = call i1 @llvm.amdgcn.loop.i64(i64 [[TMP5]]), !dbg [[DBG27]]
+; OPT-NEXT: call void @llvm.amdgcn.end.cf.i64(i64 [[TMP2]])
+; OPT-NEXT: [[TMP5]] = call i64 @llvm.amdgcn.if.break.i64(i1 [[TMP4]], i64 [[PHI_BROKEN]])
+; OPT-NEXT: [[TMP6:%.*]] = call i1 @llvm.amdgcn.loop.i64(i64 [[TMP5]])
; OPT-NEXT: br i1 [[TMP6]], label [[EXIT:%.*]], label [[LOOP]], !dbg [[DBG27]]
; OPT: exit:
-; OPT-NEXT: call void @llvm.amdgcn.end.cf.i64(i64 [[TMP5]]), !dbg [[DBG30:![0-9]+]]
-; OPT-NEXT: ret void, !dbg [[DBG30]]
+; OPT-NEXT: call void @llvm.amdgcn.end.cf.i64(i64 [[TMP5]])
+; OPT-NEXT: ret void, !dbg [[DBG30:![0-9]+]]
;
entry:
br label %loop, !dbg !24
>From 63ec93f93b55ee5bfb6806a89cd2c5c629042938 Mon Sep 17 00:00:00 2001
From: Valery Pykhtin <valery.pykhtin at gmail.com>
Date: Mon, 26 Feb 2024 17:44:23 +0100
Subject: [PATCH 2/2] minimize fix
---
.../Target/AMDGPU/SIAnnotateControlFlow.cpp | 13 ++++------
.../CodeGen/AMDGPU/si-annotate-dbg-info.ll | 26 +++++++++----------
2 files changed, 18 insertions(+), 21 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/SIAnnotateControlFlow.cpp b/llvm/lib/Target/AMDGPU/SIAnnotateControlFlow.cpp
index 2ccdc496f36de8..08e1d6b87b0df0 100644
--- a/llvm/lib/Target/AMDGPU/SIAnnotateControlFlow.cpp
+++ b/llvm/lib/Target/AMDGPU/SIAnnotateControlFlow.cpp
@@ -207,7 +207,6 @@ bool SIAnnotateControlFlow::openIf(BranchInst *Term) {
return false;
IRBuilder<> IRB(Term);
- IRB.SetCurrentDebugLocation(DebugLoc());
Value *IfCall = IRB.CreateCall(If, {Term->getCondition()});
Value *Cond = IRB.CreateExtractValue(IfCall, {0});
Value *Mask = IRB.CreateExtractValue(IfCall, {1});
@@ -223,7 +222,6 @@ bool SIAnnotateControlFlow::insertElse(BranchInst *Term) {
}
IRBuilder<> IRB(Term);
- IRB.SetCurrentDebugLocation(DebugLoc());
Value *ElseCall = IRB.CreateCall(Else, {popSaved()});
Value *Cond = IRB.CreateExtractValue(ElseCall, {0});
Value *Mask = IRB.CreateExtractValue(ElseCall, {1});
@@ -237,9 +235,7 @@ Value *SIAnnotateControlFlow::handleLoopCondition(
Value *Cond, PHINode *Broken, llvm::Loop *L, BranchInst *Term) {
auto CreateBreak = [this, Cond, Broken](Instruction *I) -> CallInst * {
- IRBuilder<> IRB(I);
- IRB.SetCurrentDebugLocation(DebugLoc());
- return IRB.CreateCall(IfBreak, {Cond, Broken});
+ return IRBuilder<>(I).CreateCall(IfBreak, {Cond, Broken});
};
if (Instruction *Inst = dyn_cast<Instruction>(Cond)) {
@@ -300,9 +296,7 @@ bool SIAnnotateControlFlow::handleLoop(BranchInst *Term) {
Broken->addIncoming(PHIValue, Pred);
}
- IRBuilder<> IRB(Term);
- IRB.SetCurrentDebugLocation(DebugLoc());
- CallInst *LoopCall = IRB.CreateCall(Loop, {Arg});
+ CallInst *LoopCall = IRBuilder<>(Term).CreateCall(Loop, {Arg});
Term->setCondition(LoopCall);
push(Term->getSuccessor(0), Arg);
@@ -343,6 +337,9 @@ bool SIAnnotateControlFlow::closeControlFlow(BasicBlock *BB) {
FirstInsertionPt = SplitEdge(DefBB, BB, DT, LI)->getFirstInsertionPt();
}
IRBuilder<> IRB(FirstInsertionPt->getParent(), FirstInsertionPt);
+ // TODO: StructurizeCFG 'Flow' blocks have debug locations from the
+ // condition, for now just avoid copying these DebugLocs so that stepping
+ // out of the then/else block in a debugger doesn't step to the condition.
IRB.SetCurrentDebugLocation(DebugLoc());
IRB.CreateCall(EndCf, {Exec});
}
diff --git a/llvm/test/CodeGen/AMDGPU/si-annotate-dbg-info.ll b/llvm/test/CodeGen/AMDGPU/si-annotate-dbg-info.ll
index 41828d163fba72..a7b4eee84cb9e4 100644
--- a/llvm/test/CodeGen/AMDGPU/si-annotate-dbg-info.ll
+++ b/llvm/test/CodeGen/AMDGPU/si-annotate-dbg-info.ll
@@ -7,15 +7,15 @@ define amdgpu_ps i32 @if_else(i32 %0) !dbg !5 {
; OPT-SAME: i32 [[TMP0:%.*]]) !dbg [[DBG5:![0-9]+]] {
; OPT-NEXT: [[C:%.*]] = icmp ne i32 [[TMP0]], 0, !dbg [[DBG13:![0-9]+]]
; OPT-NEXT: tail call void @llvm.dbg.value(metadata i1 [[C]], metadata [[META9:![0-9]+]], metadata !DIExpression()), !dbg [[DBG13]]
-; OPT-NEXT: [[TMP2:%.*]] = call { i1, i64 } @llvm.amdgcn.if.i64(i1 [[C]])
-; OPT-NEXT: [[TMP3:%.*]] = extractvalue { i1, i64 } [[TMP2]], 0
-; OPT-NEXT: [[TMP4:%.*]] = extractvalue { i1, i64 } [[TMP2]], 1
-; OPT-NEXT: br i1 [[TMP3]], label [[FALSE:%.*]], label [[FLOW:%.*]], !dbg [[DBG14:![0-9]+]]
+; OPT-NEXT: [[TMP2:%.*]] = call { i1, i64 } @llvm.amdgcn.if.i64(i1 [[C]]), !dbg [[DBG14:![0-9]+]]
+; OPT-NEXT: [[TMP3:%.*]] = extractvalue { i1, i64 } [[TMP2]], 0, !dbg [[DBG14]]
+; OPT-NEXT: [[TMP4:%.*]] = extractvalue { i1, i64 } [[TMP2]], 1, !dbg [[DBG14]]
+; OPT-NEXT: br i1 [[TMP3]], label [[FALSE:%.*]], label [[FLOW:%.*]], !dbg [[DBG14]]
; OPT: Flow:
; OPT-NEXT: [[TMP5:%.*]] = phi i32 [ 33, [[FALSE]] ], [ undef, [[TMP1:%.*]] ]
-; OPT-NEXT: [[TMP6:%.*]] = call { i1, i64 } @llvm.amdgcn.else.i64.i64(i64 [[TMP4]])
-; OPT-NEXT: [[TMP7:%.*]] = extractvalue { i1, i64 } [[TMP6]], 0
-; OPT-NEXT: [[TMP8:%.*]] = extractvalue { i1, i64 } [[TMP6]], 1
+; OPT-NEXT: [[TMP6:%.*]] = call { i1, i64 } @llvm.amdgcn.else.i64.i64(i64 [[TMP4]]), !dbg [[DBG14]]
+; OPT-NEXT: [[TMP7:%.*]] = extractvalue { i1, i64 } [[TMP6]], 0, !dbg [[DBG14]]
+; OPT-NEXT: [[TMP8:%.*]] = extractvalue { i1, i64 } [[TMP6]], 1, !dbg [[DBG14]]
; OPT-NEXT: br i1 [[TMP7]], label [[TRUE:%.*]], label [[EXIT:%.*]], !dbg [[DBG14]]
; OPT: true:
; OPT-NEXT: br label [[EXIT]], !dbg [[DBG15:![0-9]+]]
@@ -54,10 +54,10 @@ define amdgpu_ps void @loop_if_break(i32 %n) !dbg !19 {
; OPT-NEXT: tail call void @llvm.dbg.value(metadata i32 [[I]], metadata [[META21:![0-9]+]], metadata !DIExpression()), !dbg [[DBG25]]
; OPT-NEXT: [[C:%.*]] = icmp ugt i32 [[I]], 0, !dbg [[DBG26:![0-9]+]]
; OPT-NEXT: tail call void @llvm.dbg.value(metadata i1 [[C]], metadata [[META22:![0-9]+]], metadata !DIExpression()), !dbg [[DBG26]]
-; OPT-NEXT: [[TMP0:%.*]] = call { i1, i64 } @llvm.amdgcn.if.i64(i1 [[C]])
-; OPT-NEXT: [[TMP1:%.*]] = extractvalue { i1, i64 } [[TMP0]], 0
-; OPT-NEXT: [[TMP2:%.*]] = extractvalue { i1, i64 } [[TMP0]], 1
-; OPT-NEXT: br i1 [[TMP1]], label [[LOOP_BODY:%.*]], label [[FLOW]], !dbg [[DBG27:![0-9]+]]
+; OPT-NEXT: [[TMP0:%.*]] = call { i1, i64 } @llvm.amdgcn.if.i64(i1 [[C]]), !dbg [[DBG27:![0-9]+]]
+; OPT-NEXT: [[TMP1:%.*]] = extractvalue { i1, i64 } [[TMP0]], 0, !dbg [[DBG27]]
+; OPT-NEXT: [[TMP2:%.*]] = extractvalue { i1, i64 } [[TMP0]], 1, !dbg [[DBG27]]
+; OPT-NEXT: br i1 [[TMP1]], label [[LOOP_BODY:%.*]], label [[FLOW]], !dbg [[DBG27]]
; OPT: loop_body:
; OPT-NEXT: [[I_NEXT:%.*]] = sub i32 [[I]], 1, !dbg [[DBG28:![0-9]+]]
; OPT-NEXT: tail call void @llvm.dbg.value(metadata i32 [[I_NEXT]], metadata [[META23:![0-9]+]], metadata !DIExpression()), !dbg [[DBG28]]
@@ -66,8 +66,8 @@ define amdgpu_ps void @loop_if_break(i32 %n) !dbg !19 {
; OPT-NEXT: [[TMP3]] = phi i32 [ [[I_NEXT]], [[LOOP_BODY]] ], [ undef, [[LOOP]] ]
; OPT-NEXT: [[TMP4:%.*]] = phi i1 [ false, [[LOOP_BODY]] ], [ true, [[LOOP]] ]
; OPT-NEXT: call void @llvm.amdgcn.end.cf.i64(i64 [[TMP2]])
-; OPT-NEXT: [[TMP5]] = call i64 @llvm.amdgcn.if.break.i64(i1 [[TMP4]], i64 [[PHI_BROKEN]])
-; OPT-NEXT: [[TMP6:%.*]] = call i1 @llvm.amdgcn.loop.i64(i64 [[TMP5]])
+; OPT-NEXT: [[TMP5]] = call i64 @llvm.amdgcn.if.break.i64(i1 [[TMP4]], i64 [[PHI_BROKEN]]), !dbg [[DBG27]]
+; OPT-NEXT: [[TMP6:%.*]] = call i1 @llvm.amdgcn.loop.i64(i64 [[TMP5]]), !dbg [[DBG27]]
; OPT-NEXT: br i1 [[TMP6]], label [[EXIT:%.*]], label [[LOOP]], !dbg [[DBG27]]
; OPT: exit:
; OPT-NEXT: call void @llvm.amdgcn.end.cf.i64(i64 [[TMP5]])
More information about the llvm-commits
mailing list