[DEBUGINFO] Propagate debug metadata for sext SDNode. #135971

apeskov · 2025-04-16T14:34:11Z

In some cases of chained sext operators the debug metadata can be missed. This patch propagates proper metadata to resulting node.

Particular case of issue is NVPTX codegen for function with bool local variable:

void test(int i) {
  bool xyz = i == 0;
  foo(i);
}

llvmbot · 2025-04-16T14:34:45Z

@llvm/pr-subscribers-llvm-selectiondag

@llvm/pr-subscribers-backend-nvptx

Author: Alexander Peskov (apeskov)

Changes

In some cases of chained sext operators the debug metadata can be missed. This patch propagates proper metadata to resulting node.

Particular case of issue is NVPTX codegen for function with bool local variable:

void test(int i) {
  bool xyz = i == 0;
  foo(i);
}

Full diff: https://github.com/llvm/llvm-project/pull/135971.diff

2 Files Affected:

(modified) llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp (+3-1)
(added) llvm/test/DebugInfo/NVPTX/debug-bool-var.ll (+43)

diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 8682c40898046..c5825a64c1c18 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -6244,7 +6244,9 @@ SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
       SDNodeFlags Flags;
       if (OpOpcode == ISD::ZERO_EXTEND)
         Flags.setNonNeg(N1->getFlags().hasNonNeg());
-      return getNode(OpOpcode, DL, VT, N1.getOperand(0), Flags);
+      SDValue NewVal = getNode(OpOpcode, DL, VT, N1.getOperand(0), Flags);
+      transferDbgValues(N1, NewVal);
+      return NewVal;
     }
     if (N1.isUndef())
       // sext(undef) = 0, because the top bits will all be the same.
diff --git a/llvm/test/DebugInfo/NVPTX/debug-bool-var.ll b/llvm/test/DebugInfo/NVPTX/debug-bool-var.ll
new file mode 100644
index 0000000000000..7e50f1ea18fea
--- /dev/null
+++ b/llvm/test/DebugInfo/NVPTX/debug-bool-var.ll
@@ -0,0 +1,43 @@
+; RUN: llc < %s -mtriple=nvptx64-nvidia-cuda | FileCheck %s
+
+declare void @foo(i32)
+
+define void @test1(i32 noundef %gid) !dbg !3 {
+entry:
+  ;
+  ; verify that debug info exists for "xyz" variable
+  ;
+  ; CHECK-LABEL:    DW_TAG_variable
+  ; CHECK:      .b8 120     // DW_AT_name
+  ; CHECK-NEXT: .b8 121
+  ; CHECK-NEXT: .b8 122
+  ; CHECK-NEXT: .b8 0
+  ; CHECK-NEXT: .b8 1       // DW_AT_decl_file
+  ; CHECK-NEXT: .b8 6       // DW_AT_decl_line
+  ;
+  %cmp = icmp eq i32 %gid, 0, !dbg !12
+  %conv = zext i1 %cmp to i32, !dbg !12
+  %conv1 = trunc i32 %conv to i8, !dbg !12
+    #dbg_value(i8 %conv1, !10, !DIExpression(), !13)
+  %conv3 = sext i8 %conv1 to i32
+  call void @foo(i32 %conv3)
+  ret void
+}
+
+!llvm.dbg.cu = !{!0}
+!llvm.module.flags = !{!2}
+
+!0 = distinct !DICompileUnit(language: DW_LANG_C_plus_plus, file: !1, producer: "", isOptimized: false, runtimeVersion: 0, emissionKind: FullDebug)
+!1 = !DIFile(filename: "test.cu", directory: "/source/dir")
+!2 = !{i32 1, !"Debug Info Version", i32 3}
+!3 = distinct !DISubprogram(name: "test1", linkageName: "_test1i", scope: !1, file: !1, line: 5, type: !4, scopeLine: 5, unit: !0, retainedNodes: !8)
+!4 = !DISubroutineType(types: !5)
+!5 = !{!6, !7}
+!6 = !DIBasicType(tag: DW_TAG_unspecified_type, name: "void")
+!7 = !DIBasicType(name: "int", size: 32, encoding: DW_ATE_signed)
+!8 = !{}
+!9 = distinct !DILexicalBlock(scope: !3, file: !1, line: 5, column: 30)
+!10 = !DILocalVariable(name: "xyz", scope: !9, file: !1, line: 6, type: !11)
+!11 = !DIBasicType(name: "bool", size: 8, encoding: DW_ATE_boolean)
+!12 = !DILocation(line: 1, column: 3, scope: !9)
+!13 = !DILocation(line: 2, scope: !9)

llvmbot · 2025-04-16T14:34:46Z

@llvm/pr-subscribers-debuginfo

Author: Alexander Peskov (apeskov)

Changes

In some cases of chained sext operators the debug metadata can be missed. This patch propagates proper metadata to resulting node.

Particular case of issue is NVPTX codegen for function with bool local variable:

void test(int i) {
  bool xyz = i == 0;
  foo(i);
}

Full diff: https://github.com/llvm/llvm-project/pull/135971.diff

2 Files Affected:

(modified) llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp (+3-1)
(added) llvm/test/DebugInfo/NVPTX/debug-bool-var.ll (+43)

diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 8682c40898046..c5825a64c1c18 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -6244,7 +6244,9 @@ SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT,
       SDNodeFlags Flags;
       if (OpOpcode == ISD::ZERO_EXTEND)
         Flags.setNonNeg(N1->getFlags().hasNonNeg());
-      return getNode(OpOpcode, DL, VT, N1.getOperand(0), Flags);
+      SDValue NewVal = getNode(OpOpcode, DL, VT, N1.getOperand(0), Flags);
+      transferDbgValues(N1, NewVal);
+      return NewVal;
     }
     if (N1.isUndef())
       // sext(undef) = 0, because the top bits will all be the same.
diff --git a/llvm/test/DebugInfo/NVPTX/debug-bool-var.ll b/llvm/test/DebugInfo/NVPTX/debug-bool-var.ll
new file mode 100644
index 0000000000000..7e50f1ea18fea
--- /dev/null
+++ b/llvm/test/DebugInfo/NVPTX/debug-bool-var.ll
@@ -0,0 +1,43 @@
+; RUN: llc < %s -mtriple=nvptx64-nvidia-cuda | FileCheck %s
+
+declare void @foo(i32)
+
+define void @test1(i32 noundef %gid) !dbg !3 {
+entry:
+  ;
+  ; verify that debug info exists for "xyz" variable
+  ;
+  ; CHECK-LABEL:    DW_TAG_variable
+  ; CHECK:      .b8 120     // DW_AT_name
+  ; CHECK-NEXT: .b8 121
+  ; CHECK-NEXT: .b8 122
+  ; CHECK-NEXT: .b8 0
+  ; CHECK-NEXT: .b8 1       // DW_AT_decl_file
+  ; CHECK-NEXT: .b8 6       // DW_AT_decl_line
+  ;
+  %cmp = icmp eq i32 %gid, 0, !dbg !12
+  %conv = zext i1 %cmp to i32, !dbg !12
+  %conv1 = trunc i32 %conv to i8, !dbg !12
+    #dbg_value(i8 %conv1, !10, !DIExpression(), !13)
+  %conv3 = sext i8 %conv1 to i32
+  call void @foo(i32 %conv3)
+  ret void
+}
+
+!llvm.dbg.cu = !{!0}
+!llvm.module.flags = !{!2}
+
+!0 = distinct !DICompileUnit(language: DW_LANG_C_plus_plus, file: !1, producer: "", isOptimized: false, runtimeVersion: 0, emissionKind: FullDebug)
+!1 = !DIFile(filename: "test.cu", directory: "/source/dir")
+!2 = !{i32 1, !"Debug Info Version", i32 3}
+!3 = distinct !DISubprogram(name: "test1", linkageName: "_test1i", scope: !1, file: !1, line: 5, type: !4, scopeLine: 5, unit: !0, retainedNodes: !8)
+!4 = !DISubroutineType(types: !5)
+!5 = !{!6, !7}
+!6 = !DIBasicType(tag: DW_TAG_unspecified_type, name: "void")
+!7 = !DIBasicType(name: "int", size: 32, encoding: DW_ATE_signed)
+!8 = !{}
+!9 = distinct !DILexicalBlock(scope: !3, file: !1, line: 5, column: 30)
+!10 = !DILocalVariable(name: "xyz", scope: !9, file: !1, line: 6, type: !11)
+!11 = !DIBasicType(name: "bool", size: 8, encoding: DW_ATE_boolean)
+!12 = !DILocation(line: 1, column: 3, scope: !9)
+!13 = !DILocation(line: 2, scope: !9)

OCHyams

I think this LGTM, just one nit on the test from me.

Unrelated to your change - does this problem occur a few lines down in the if (OpOpcode == ISD::ZERO_EXTEND) { // (zext (zext x)) -> (zext x) block too?

OCHyams

Actually - could this create incorrect locations, if the sddbgvalue was attached to a zext but with this change gets moved onto a sext?

github-actions · 2025-04-17T12:54:29Z

✅ With the latest revision this PR passed the C/C++ code formatter.

apeskov · 2025-04-17T13:31:03Z

does this problem occur a few lines down?

Yes, it does. I extended test to cover zext-zext case as well. I checkd, frontend generates similar kind of IR in case of unsigned char local variable.

just one nit on the test from me.

I don't see any nit comments about test itself. Perhaps you forgot press "submit" button for review.

OCHyams

Yes, it does. I extended test to cover zext-zext case as well. I checkd, frontend generates similar kind of IR in case of unsigned char local variable.

Thanks!

I don't see any nit comments about test itself. Perhaps you forgot press "submit" button for review.

Oops. Probably - I've added it now.

Actually - could this create incorrect locations, if the sddbgvalue was attached to a zext but with this change gets moved onto a sext?

Looking again at my question - the new SDNode is created with the OpOpcode type, so that's not a concern. We only go sext (sext x) ->sext x or sext (zext x) ->zext x, never sext (zext x) -> sext x. Right?

OCHyams · 2025-04-17T17:05:33Z

llvm/test/DebugInfo/NVPTX/debug-bool-var.ll

@@ -0,0 +1,84 @@
+; RUN: llc < %s -mtriple=nvptx64-nvidia-cuda | FileCheck %s


Could we use -print-after=finalize-isel to reduce the surface area being tested (and it'd make the output slightly easier to read IMO)?

Done. Test is reworded with "finalize-isel".

During type promotion debug info is messed for bool vars.

Signed-off-by: Alexander Peskov <[email protected]>

apeskov · 2025-05-02T09:24:56Z

We only go sext (sext x) ->sext x or sext (zext x) ->zext x, never sext (zext x) -> sext x. Right?

You are right, sext (zext x) -> sext x is not the case. In other words, debug data is only transferred between nodes with identical opcode.

OCHyams

Great, this LGTM (with a final nit, but feel free to merge once that's fixed). Thanks!

llvm/test/DebugInfo/NVPTX/debug-bool-var.ll

In some cases of chained `sext` operators the debug metadata can be missed. This patch propagates proper metadata to resulting node. Particular case of issue is NVPTX codegen for function with bool local variable: ``` void test(int i) { bool xyz = i == 0; foo(i); } ``` --------- Signed-off-by: Alexander Peskov <[email protected]>

llvmbot added debuginfo backend:NVPTX llvm:SelectionDAG SelectionDAGISel as well labels Apr 16, 2025

OCHyams approved these changes Apr 16, 2025

View reviewed changes

apeskov force-pushed the ap/debug-info-propagation-ext-node branch from 3e325b9 to 0f7477b Compare April 16, 2025 16:36

OCHyams reviewed Apr 16, 2025

View reviewed changes

apeskov force-pushed the ap/debug-info-propagation-ext-node branch from 0f7477b to 0f34707 Compare April 17, 2025 12:50

OCHyams reviewed Apr 17, 2025

View reviewed changes

apeskov added 5 commits May 2, 2025 13:01

[DEBUGINFO] Fix missed debug info for bool local var

d5670ff

During type promotion debug info is messed for bool vars.

Plus zext-zext case

eb8da2a

Signed-off-by: Alexander Peskov <[email protected]>

fix lint

a6e9bbf

minor

72932c8

reword test with -print-after=finalize-isel

d61d118

apeskov force-pushed the ap/debug-info-propagation-ext-node branch from 54d98c9 to d61d118 Compare May 2, 2025 09:11

OCHyams approved these changes May 2, 2025

View reviewed changes

llvm/test/DebugInfo/NVPTX/debug-bool-var.ll Outdated Show resolved Hide resolved

min

e3adc2b

enferex merged commit cffb8ae into llvm:main May 2, 2025
11 checks passed

apeskov deleted the ap/debug-info-propagation-ext-node branch May 15, 2025 12:27

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[DEBUGINFO] Propagate debug metadata for sext SDNode. #135971

[DEBUGINFO] Propagate debug metadata for sext SDNode. #135971

apeskov commented Apr 16, 2025

llvmbot commented Apr 16, 2025 •

edited

Loading

llvmbot commented Apr 16, 2025

OCHyams left a comment

OCHyams left a comment

github-actions bot commented Apr 17, 2025 •

edited

Loading

apeskov commented Apr 17, 2025

OCHyams left a comment

OCHyams Apr 17, 2025

apeskov May 2, 2025

apeskov commented May 2, 2025

OCHyams left a comment

		@@ -0,0 +1,84 @@
		; RUN: llc < %s -mtriple=nvptx64-nvidia-cuda \| FileCheck %s

[DEBUGINFO] Propagate debug metadata for sext SDNode. #135971

[DEBUGINFO] Propagate debug metadata for sext SDNode. #135971

Conversation

apeskov commented Apr 16, 2025

llvmbot commented Apr 16, 2025 • edited Loading

llvmbot commented Apr 16, 2025

OCHyams left a comment

Choose a reason for hiding this comment

OCHyams left a comment

Choose a reason for hiding this comment

github-actions bot commented Apr 17, 2025 • edited Loading

apeskov commented Apr 17, 2025

OCHyams left a comment

Choose a reason for hiding this comment

OCHyams Apr 17, 2025

Choose a reason for hiding this comment

apeskov May 2, 2025

Choose a reason for hiding this comment

apeskov commented May 2, 2025

OCHyams left a comment

Choose a reason for hiding this comment

llvmbot commented Apr 16, 2025 •

edited

Loading

github-actions bot commented Apr 17, 2025 •

edited

Loading