[DebugInfo] Update loop metadata in stripNonLineTableDebugInfo

[vsk]

Summary

Have stripNonLineTableDebugInfo() attach updated !llvm.loop metadata to
an instruction (instead of updating and then discarding the metadata).

This fixes "!dbg attachment points at wrong subprogram for function"
errors seen while archiving an iOS app.

It would be nice -- as a follow-up -- to catch this issue earlier,
perhaps by modifying the verifier to constrain where DILocations are
allowed. Any alternative suggestions appreciated.

rdar://61982466

[aprantl]

Thanks, this looks great!

llvm/lib/IR/DebugInfo.cpp

664

Should we keep a form of the old loop and assert that no other DILocations are hanging off not-yet-handled future metadata?

[vsk]

llvm/lib/IR/DebugInfo.cpp

664

There's a couple places where we'd need to add that assert, so I tried adding a verifier check instead (https://reviews.llvm.org/D79245). If you'd rather we just add a simple assert here that's fine too, lmk what you prefer.

llvm/lib/IR/DebugInfo.cpp

for (auto &BB : F) {
           Scope = remap(Scope);
           InlinedAt = remap(InlinedAt);
           return DebugLoc::get(DL.getLine(), DL.getCol(), Scope, InlinedAt);
         };
 
         if (I.getDebugLoc() != DebugLoc())
           I.setDebugLoc(remapDebugLoc(I.getDebugLoc()));
 
-        // Remap DILocations in untyped MDNodes (e.g., llvm.loop).
-        SmallVector<std::pair<unsigned, MDNode *>, 2> MDs;
-        I.getAllMetadata(MDs);
-        for (auto Attachment : MDs)
-          if (auto *T = dyn_cast_or_null<MDTuple>(Attachment.second))
-            for (unsigned N = 0; N < T->getNumOperands(); ++N)
-              if (auto *Loc = dyn_cast_or_null<DILocation>(T->getOperand(N)))
-                if (Loc != DebugLoc())
-                  T->replaceOperandWith(N, remapDebugLoc(Loc));
+        // Remap DILocations in llvm.loop attachments.
+        updateLoopMetadataDebugLocations(I, [&](const DILocation &Loc) {
+          return remapDebugLoc(&Loc).get();
+        });
       }
     }
   }
 
   // Create a new llvm.dbg.cu, which is equivalent to the one
   // -gline-tables-only would have created.
   for (auto &NMD : M.getNamedMDList()) {
     SmallVector<MDNode *, 8> Ops;

llvm/unittests/IR/DebugInfoTest.cpp

 //===- llvm/unittest/IR/DebugInfo.cpp - DebugInfo tests -------------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
+#include "llvm/IR/DebugInfo.h"
+#include "llvm/AsmParser/Parser.h"
 #include "llvm/IR/DebugInfoMetadata.h"
+#include "llvm/IR/LLVMContext.h"
+#include "llvm/IR/Module.h"
+#include "llvm/IR/Verifier.h"
+#include "llvm/Support/SourceMgr.h"
 #include "gtest/gtest.h"
 
 using namespace llvm;
 
+static std::unique_ptr<Module> parseIR(LLVMContext &C, const char *IR) {
+  SMDiagnostic Err;
+  std::unique_ptr<Module> Mod = parseAssemblyString(IR, Err, C);
+  if (!Mod)
+    Err.print("DebugInfoTest", errs());
+  return Mod;
+}
+
 namespace {
 
 TEST(DINodeTest, getFlag) {
   // Some valid flags.
   EXPECT_EQ(DINode::FlagPublic, DINode::getFlag("DIFlagPublic"));
   EXPECT_EQ(DINode::FlagProtected, DINode::getFlag("DIFlagProtected"));
   EXPECT_EQ(DINode::FlagPrivate, DINode::getFlag("DIFlagPrivate"));
   EXPECT_EQ(DINode::FlagVector, DINode::getFlag("DIFlagVector"));
CHECK_SPLIT(DINode::FlagRValueReference, {DINode::FlagRValueReference},
               DINode::FlagZero);
   DINode::DIFlags Flags[] = {DINode::FlagFwdDecl, DINode::FlagVector};
   CHECK_SPLIT(DINode::FlagFwdDecl | DINode::FlagVector, Flags,
               DINode::FlagZero);
   CHECK_SPLIT(DINode::FlagZero, {}, DINode::FlagZero);
 #undef CHECK_SPLIT
 }
 
+TEST(StripTest, LoopMetadata) {
+  LLVMContext C;
+  std::unique_ptr<Module> M = parseIR(C, R"(
+    define void @f() !dbg !5 {
+      ret void, !dbg !10, !llvm.loop !11
+    }
+
+    !llvm.dbg.cu = !{!0}
+    !llvm.debugify = !{!3, !3}
+    !llvm.module.flags = !{!4}
+
+    !0 = distinct !DICompileUnit(language: DW_LANG_C, file: !1, producer: "debugify", isOptimized: true, runtimeVersion: 0, emissionKind: FullDebug, enums: !2)
+    !1 = !DIFile(filename: "loop.ll", directory: "/")
+    !2 = !{}
+    !3 = !{i32 1}
+    !4 = !{i32 2, !"Debug Info Version", i32 3}
+    !5 = distinct !DISubprogram(name: "f", linkageName: "f", scope: null, file: !1, line: 1, type: !6, scopeLine: 1, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !0, retainedNodes: !7)
+    !6 = !DISubroutineType(types: !2)
+    !7 = !{!8}
+    !8 = !DILocalVariable(name: "1", scope: !5, file: !1, line: 1, type: !9)
+    !9 = !DIBasicType(name: "ty32", size: 32, encoding: DW_ATE_unsigned)
+    !10 = !DILocation(line: 1, column: 1, scope: !5)
+    !11 = distinct !{!11, !10, !10}
+)");
+
+  // Look up the debug info emission kind for the CU via the loop metadata
+  // attached to the terminator. If, when stripping non-line table debug info,
+  // we update the terminator's metadata correctly, we should be able to
+  // observe the change in emission kind for the CU.
+  auto getEmissionKind = [&]() {
+    Instruction &I = *M->getFunction("f")->getEntryBlock().getFirstNonPHI();
+    MDNode *LoopMD = I.getMetadata(LLVMContext::MD_loop);
+    return cast<DILocation>(LoopMD->getOperand(1))
+        ->getScope()
+        ->getSubprogram()
+        ->getUnit()
+        ->getEmissionKind();
+  };
+
+  EXPECT_EQ(getEmissionKind(), DICompileUnit::FullDebug);
+
+  bool Changed = stripNonLineTableDebugInfo(*M);
+  EXPECT_TRUE(Changed);
+
+  EXPECT_EQ(getEmissionKind(), DICompileUnit::LineTablesOnly);
+
+  bool BrokenDebugInfo = false;
+  bool HardError = verifyModule(*M, &errs(), &BrokenDebugInfo);
+  EXPECT_FALSE(HardError);
+  EXPECT_FALSE(BrokenDebugInfo);
+}
+
 } // end namespace