CodeGen: Fix CodeView crashes with empty llvm.dbg.cu #163286

arsenm · 2025-10-13T23:12:30Z

No description provided.

arsenm · 2025-10-13T23:12:45Z

CodeGen: Fix CodeView crashes with empty llvm.dbg.cu #163286 👈 (View in Graphite)
main

This stack of pull requests is managed by Graphite. Learn more about stacking.

llvmbot · 2025-10-13T23:16:27Z

@llvm/pr-subscribers-platform-windows

@llvm/pr-subscribers-debuginfo

Author: Matt Arsenault (arsenm)

Changes

Full diff: https://github.com/llvm/llvm-project/pull/163286.diff

2 Files Affected:

(modified) llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp (+15-6)
(added) llvm/test/DebugInfo/X86/codeview-empty-dbg-cu-crash.ll (+30)

diff --git a/llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp b/llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp
index e57ed24a45065..6cbed50e14ea9 100644
--- a/llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp
+++ b/llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp
@@ -625,10 +625,13 @@ void CodeViewDebug::beginModule(Module *M) {
   if (Asm->hasDebugInfo()) {
     Node = *M->debug_compile_units_begin();
   } else {
+    auto DebugCompileUnits = MMI->getModule()->debug_compile_units();
+    if (DebugCompileUnits.empty())
+      return;
+
     // When emitting only compiler information, we may have only NoDebug CUs,
     // which would be skipped by debug_compile_units_begin.
-    NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu");
-    Node = *CUs->operands().begin();
+    Node = *DebugCompileUnits.begin();
   }
   const auto *CU = cast<DICompileUnit>(Node);
   DISourceLanguageName Lang = CU->getSourceLanguage();
@@ -900,8 +903,11 @@ void CodeViewDebug::emitCompilerInformation() {
   OS.AddComment("CPUType");
   OS.emitInt16(static_cast<uint64_t>(TheCPU));
 
-  NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu");
-  const MDNode *Node = *CUs->operands().begin();
+  auto CUs = MMI->getModule()->debug_compile_units();
+  if (CUs.empty())
+    return;
+
+  const MDNode *Node = *CUs.begin();
   const auto *CU = cast<DICompileUnit>(Node);
 
   StringRef CompilerVersion = CU->getProducer();
@@ -948,8 +954,11 @@ void CodeViewDebug::emitBuildInfo() {
   // not clear if the compiler path should refer to the executable for the
   // frontend or the backend. Leave it blank for now.
   TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {};
-  NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu");
-  const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs.
+  auto CUs = MMI->getModule()->debug_compile_units();
+  if (CUs.empty())
+    return;
+
+  const MDNode *Node = *CUs.begin(); // FIXME: Multiple CUs.
   const auto *CU = cast<DICompileUnit>(Node);
   const DIFile *MainSourceFile = CU->getFile();
   BuildInfoArgs[BuildInfoRecord::CurrentDirectory] =
diff --git a/llvm/test/DebugInfo/X86/codeview-empty-dbg-cu-crash.ll b/llvm/test/DebugInfo/X86/codeview-empty-dbg-cu-crash.ll
new file mode 100644
index 0000000000000..395b99e7f90b6
--- /dev/null
+++ b/llvm/test/DebugInfo/X86/codeview-empty-dbg-cu-crash.ll
@@ -0,0 +1,30 @@
+; RUN: llc -mtriple=x86_64-pc-windows-msvc < %s | FileCheck %s
+
+; CHECK: .file	"<stdin>"
+; CHECK-NEXT: .section	.debug$S,"dr"
+; CHECK-NEXT: .p2align	2, 0x0
+; CHECK-NEXT: .long	4                               # Debug section magic
+; CHECK-NEXT: .long	241
+; CHECK-NEXT: .long	.Ltmp1-.Ltmp0                   # Subsection size
+; CHECK-NEXT: .Ltmp0:
+; CHECK-NEXT: .short	.Ltmp3-.Ltmp2                   # Record length
+; CHECK-NEXT: .Ltmp2:
+; CHECK-NEXT: .short	4353                            # Record kind: S_OBJNAME
+; CHECK-NEXT: .long	0                               # Signature
+; CHECK-NEXT: .byte	0                               # Object name
+; CHECK-NEXT: .p2align	2, 0x0
+; CHECK-NEXT: .Ltmp3:
+; CHECK-NEXT: .short	.Ltmp5-.Ltmp4                   # Record length
+; CHECK-NEXT: .Ltmp4:
+; CHECK-NEXT: .short	4412                            # Record kind: S_COMPILE3
+; CHECK-NEXT: .long	3                               # Flags and language
+; CHECK-NEXT: .short	208                             # CPUType
+; CHECK-NEXT: .Ltmp1:
+; CHECK-NEXT: .p2align	2, 0x0
+; CHECK-NEXT: .cv_filechecksums                       # File index to string table offset subsection
+; CHECK-NEXT: .cv_stringtable                         # String table
+
+!llvm.dbg.cu = !{}
+!llvm.module.flags = !{!0}
+
+!0 = !{i32 2, !"Debug Info Version", i32 3}

cjacek · 2025-10-14T12:14:57Z

llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp

  } else {
+    auto DebugCompileUnits = MMI->getModule()->debug_compile_units();
+    if (DebugCompileUnits.empty())
+      return;


Should we set Asm to null here like we do for nodebug CUs? This would skip much more code, including emitBuildInfo.

I wasn't sure why this was doing that, or how to test it

cjacek · 2025-10-14T12:22:20Z

llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp

+  auto CUs = MMI->getModule()->debug_compile_units();
+  if (CUs.empty())
+    return;


I'm not sure about emitting incomplete information, maybe we could default to version 0 or something similar?

dwblaikie

Perhaps this could follow a code path more like the DwarfDebug handling - up in AsmPrinter::doInitialization, the DwarfDebug pass is only added if hasDebugInfo is true, which is based on DbgInfoAvailable which is based on !debug_compile_units.empty - whereas CodeViewDebug is added if the raw llvm.dbg.cu node exists. Seems like Rather than checking the raw node exists, we should check hasDebugInfo in a broader conditional that covers both CV and DWARF.

dwblaikie · 2025-10-14T15:58:42Z

Perhaps this could follow a code path more like the DwarfDebug handling - up in AsmPrinter::doInitialization, the DwarfDebug pass is only added if hasDebugInfo is true, which is based on DbgInfoAvailable which is based on !debug_compile_units.empty - whereas CodeViewDebug is added if the raw llvm.dbg.cu node exists. Seems like Rather than checking the raw node exists, we should check hasDebugInfo in a broader conditional that covers both CV and DWARF.

Hmm, nope, guess not - read further and there's a comment that says "On Windows targets, emit minimal CodeView compiler info even when debug info is disabled"... Not sure what to do with that, in the context of there then being no CUs to base that debug info on.

Maybe check what happens in the more normal case, where there's no llvm.dbg.cu node, and try to collapse this non-empty/at-least-one-cu logic so it shares that codepath as much as possible?

arsenm · 2025-11-11T22:51:39Z

ping

of skipping entries that are there.

rnk · 2025-11-12T04:50:40Z

llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp

+      Asm = nullptr;
+      return;
+    }
+


Sorry to code golf this, but I think this would be cleaner if we took the CU variable below and stored it in CodeViewDebug (call it TheCU or something) and reused it in the endModule invocation. It goes right next to the CompilerInfoAsm state.

Part of what's going on here is that we never came up with a proper strategy for handling full LTO, which is what usually creates a multi-CU situation. The emergent behavior is probably fine, and maybe the best we could do, which is just using the global module information from the first compilation unit in the LTO unit. People probably already rely on that.

cjacek

LGTM, thanks!

juliannagele · 2025-11-14T17:37:07Z

The test added here seems to be failing on the stage2 green dragon bot: https://green.lab.llvm.org/job/llvm.org/job/clang-stage2-Rthinlto/1366/testReport/junit/LLVM/DebugInfo_X86/codeview_empty_dbg_cu_crash_ll/
Could you have a look?

arsenm · 2025-11-14T19:51:43Z

The test added here seems to be failing on the stage2 green dragon bot: https://green.lab.llvm.org/job/llvm.org/job/clang-stage2-Rthinlto/1366/testReport/junit/LLVM/DebugInfo_X86/codeview_empty_dbg_cu_crash_ll/ Could you have a look?

#168111

arsenm added debuginfo platform:windows labels Oct 13, 2025 — with Graphite App

arsenm requested review from adrian-prantl, cjacek, dpaoliello, dwblaikie and rnk October 13, 2025 23:16

arsenm marked this pull request as ready for review October 13, 2025 23:22

llvmbot added the llvm:codegen label Oct 13, 2025

cjacek reviewed Oct 14, 2025

View reviewed changes

dwblaikie reviewed Oct 14, 2025

View reviewed changes

arsenm force-pushed the users/arsenm/codegen/codeview-fix-crash-empty-dbg-cus branch 2 times, most recently from fddd368 to 9c5ab4d Compare November 4, 2025 06:16

arsenm added 3 commits November 11, 2025 20:46

CodeGen: Fix CodeView crashes with empty llvm.dbg.cu

633af6e

Emit 0 instead of incomplete information

aadcb99

Redo this. The compile unit iterator has the unhelpful behavior

e10eb45

of skipping entries that are there.

rnk reviewed Nov 12, 2025

View reviewed changes

Store TheCU

f4ee03f

arsenm force-pushed the users/arsenm/codegen/codeview-fix-crash-empty-dbg-cus branch from 9c5ab4d to f4ee03f Compare November 12, 2025 05:25

cjacek approved these changes Nov 12, 2025

View reviewed changes

arsenm merged commit 2e489f7 into main Nov 12, 2025
10 checks passed

arsenm deleted the users/arsenm/codegen/codeview-fix-crash-empty-dbg-cus branch November 12, 2025 22:59

git-crd pushed a commit to git-crd/crd-llvm-project that referenced this pull request Nov 13, 2025

CodeGen: Fix CodeView crashes with empty llvm.dbg.cu (llvm#163286)

03ed866

CodeGen: Fix CodeView crashes with empty llvm.dbg.cu #163286

CodeGen: Fix CodeView crashes with empty llvm.dbg.cu #163286

Uh oh!

Conversation

arsenm commented Oct 13, 2025

Uh oh!

arsenm commented Oct 13, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvmbot commented Oct 13, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

cjacek Oct 14, 2025

Choose a reason for hiding this comment

Uh oh!

arsenm Oct 14, 2025

Choose a reason for hiding this comment

Uh oh!

cjacek Oct 14, 2025

Choose a reason for hiding this comment

Uh oh!

dwblaikie left a comment

Choose a reason for hiding this comment

Uh oh!

dwblaikie commented Oct 14, 2025

Uh oh!

arsenm commented Nov 11, 2025

Uh oh!

rnk Nov 12, 2025

Choose a reason for hiding this comment

Uh oh!

cjacek left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

juliannagele commented Nov 14, 2025

Uh oh!

arsenm commented Nov 14, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

arsenm commented Oct 13, 2025 •

edited

Loading

llvmbot commented Oct 13, 2025 •

edited

Loading