Add debug location verification for !llvm.loop attachments

[aprantl]

Summary

This patch teaches the Verifier how to detect broken !llvm.loop attachments as discussed in https://reviews.llvm.org/D60831. This allows LLVM to warn and strip out the broken debug info before attempting an LTO compilation with input generated by LLVM predating https://reviews.llvm.org/rL361149.

rdar://problem/51631158

[Orlando]

LGTM. Thanks for doing this!

llvm/lib/IR/Verifier.cpp

2380 ↗

(On Diff #205378)

The loop metadata can contain nodes which are not DILocations but the check at the start of VisitDebugLoc() will catch this, so this seems fine.

llvm/trunk/lib/IR/Verifier.cpp

if (!HasDebugInfo)
     return;
 
   // Check that all !dbg attachments lead to back to N (or, at least, another
   // subprogram that describes the same function).
   //
   // FIXME: Check this incrementally while visiting !dbg attachments.
   // FIXME: Only check when N is the canonical subprogram for F.
   SmallPtrSet<const MDNode *, 32> Seen;
-  for (auto &BB : F)
-    for (auto &I : BB) {
-      // Be careful about using DILocation here since we might be dealing with
-      // broken code (this is the Verifier after all).
-      DILocation *DL =
-          dyn_cast_or_null<DILocation>(I.getDebugLoc().getAsMDNode());
-      if (!DL)
-        continue;
-      if (!Seen.insert(DL).second)
-        continue;
+  auto VisitDebugLoc = [&](const Instruction &I, const MDNode *Node) {
+    const DILocation *DL = dyn_cast_or_null<DILocation>(Node);
+      return;
+      return;
 
-      Metadata *Parent = DL->getRawScope();
-      AssertDI(Parent && isa<DILocalScope>(Parent),
-               "DILocation's scope must be a DILocalScope", N, &F, &I, DL,
-               Parent);
-      DILocalScope *Scope = DL->getInlinedAtScope();
-      if (Scope && !Seen.insert(Scope).second)
-        continue;
+      return;
 
-      DISubprogram *SP = Scope ? Scope->getSubprogram() : nullptr;
 
-      // Scope and SP could be the same MDNode and we don't want to skip
-      // validation in that case
-      if (SP && ((Scope != SP) && !Seen.insert(SP).second))
-        continue;
+      return;
 
-      // FIXME: Once N is canonical, check "SP == &N".
-      AssertDI(SP->describes(&F),
-               "!dbg attachment points at wrong subprogram for function", N, &F,
-               &I, DL, Scope, SP);
+  };
+  for (auto &BB : F)
+    for (auto &I : BB) {
+      VisitDebugLoc(I, I.getDebugLoc().getAsMDNode());
+      // The llvm.loop annotations also contain two DILocations.
+      if (auto MD = I.getMetadata(LLVMContext::MD_loop))
+        for (unsigned i = 1; i < MD->getNumOperands(); ++i)
+          VisitDebugLoc(I, dyn_cast_or_null<MDNode>(MD->getOperand(i)));
+      if (BrokenDebugInfo)
+        return;
     }
 }
 
 // verifyBasicBlock - Verify that a basic block is well formed...
 //
 void Verifier::visitBasicBlock(BasicBlock &BB) {
   InstsInThisBlock.clear();
 

llvm/trunk/test/Verifier/llvm.loop.ll

+; RUN: llvm-as -disable-output < %s -o /dev/null 2>&1 | FileCheck %s
+
+define i32 @foo() !dbg !4 {
+entry:
+  ret i32 0, !dbg !6
+}
+
+define i32 @bar() !dbg !5 {
+entry:
+; CHECK: !dbg attachment points at wrong subprogram for function
+  ret i32 0, !dbg !10, !llvm.loop !9
+}
+
+; CHECK: warning: ignoring invalid debug info
+!llvm.dbg.cu = !{!0}
+!llvm.module.flags = !{!7, !8}
+
+!0 = distinct !DICompileUnit(language: DW_LANG_C99, file: !1, producer: "clang", isOptimized: false, runtimeVersion: 0, emissionKind: FullDebug, enums: !2, retainedTypes: !2, globals: !2, imports: !2)
+!1 = !DIFile(filename: "dwarf-test.c", directory: "test")
+!2 = !{}
+!4 = distinct !DISubprogram(name: "foo", scope: !0, isDefinition: true, unit: !0)
+!5 = distinct !DISubprogram(name: "bar", scope: !0, isDefinition: true, unit: !0)
+!6 = !DILocation(line: 7, scope: !4)
+!7 = !{i32 2, !"Dwarf Version", i32 3}
+!8 = !{i32 1, !"Debug Info Version", i32 3}
+!9 = !{!9, !10, !11}
+!10 = !DILocation(line: 1, scope: !5)
+!11 = !DILocation(line: 1, scope: !4)