[DebugInfo][IR] Verifier checks for the extraData

[laxmansole]

LLVM IR verifier checks for extraData in debug info metadata.

This is a follow-up PR based on discussions in #165023

[llvmbot]

@llvm/pr-subscribers-llvm-ir

Author: Laxman Sole (laxmansole)

Changes

LLVM IR verifier checks for extraData in debug info metadata.

This is a follow-up PR based on discussions in #165023

---

Full diff: https://github.com/llvm/llvm-project/pull/167971.diff

1 Files Affected:

• (modified) llvm/lib/IR/Verifier.cpp (+17)

diff --git a/llvm/lib/IR/Verifier.cpp b/llvm/lib/IR/Verifier.cpp
index fa18c3cd0f404..b01ca73aa3623 100644
--- a/llvm/lib/IR/Verifier.cpp
+++ b/llvm/lib/IR/Verifier.cpp
@@ -1320,6 +1320,23 @@ void Verifier::visitDIDerivedType(const DIDerivedType &N) {
   if (N.getTag() == dwarf::DW_TAG_ptr_to_member_type) {
     CheckDI(isType(N.getRawExtraData()), "invalid pointer to member type", &N,
             N.getRawExtraData());
+  } else if (N.getTag() == dwarf::DW_TAG_template_alias) {
+    CheckDI(isa<MDTuple>(N.getRawExtraData()), "invalid template parameters",
+            &N, N.getRawExtraData());
+  } else if (auto *ExtraData = N.getRawExtraData()) {
+    auto IsValidExtraData = [&]() {
+      if (isa<ConstantAsMetadata>(ExtraData) || isa<MDString>(ExtraData) ||
+          isa<DIObjCProperty>(ExtraData))
+        return true;
+      if (auto *Tuple = dyn_cast<MDTuple>(ExtraData))
+        return Tuple->getNumOperands() == 1 &&
+               isa<ConstantAsMetadata>(Tuple->getOperand(0));
+      return false;
+    };
+    CheckDI(IsValidExtraData(),
+            "extraData must be ConstantAsMetadata, MDString, DIObjCProperty, "
+            "or MDTuple with single ConstantAsMetadata operand",
+            &N, ExtraData);
   }
 
   if (N.getTag() == dwarf::DW_TAG_set_type) {

[laxmansole]

CC: @OCHyams @dzhidzhoev @ayermolo @enferex

[Copilot]

Pull Request Overview

This PR adds LLVM IR verifier checks for the extraData field in DIDerivedType debug info metadata. The changes ensure that extraData contains only valid metadata types based on the debug info tag, preventing invalid metadata from being created.

Key Changes:

• Added validation for DW_TAG_template_alias to ensure extraData is an MDTuple
• Added general validation for all other tags with extraData to restrict it to specific allowed metadata types

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The lambda does not validate that an MDTuple with a single operand actually contains a ConstantAsMetadata. If getOperand(0) returns null, the check will pass but may lead to undefined behavior. Add a null check before calling isa<ConstantAsMetadata>.

[ayermolo]

Test?

[dzhidzhoev]

We shouldn't crash if N.getRawExtraData() returns nullptr. (also relevant for the fragment reviewed by Copilot)

[dzhidzhoev]

Should we allow this only for DW_TAG_member, DW_TAG_variable, and DW_TAG_inheritance tags?

[dzhidzhoev]

From what I see in pre-merge check results, some tests fail with this patch. Probably, we should limit the scope of extraData check to specific tags (if the reason of failures is that the checks in this patch apply to tags other than DW_TAG_member, DW_TAG_variable, or DW_TAG_inheritance — the only ones that have MDTuple-in-extraData support implemented. Supporting extraData check for all possible tags is nice, but I don't insist on that if it gets too complicated, as it is not relevant to #165023).

[github-actions]

🐧 Linux x64 Test Results

• 186262 tests passed
• 4852 tests skipped

[dzhidzhoev]

LGTM

I think it's better to use [IR] tag instead of [LLVM IR] in the commit title, as it is more widely used if we look at git log.