[CodeExtractor] Preserve entire scope of !DIVariables when moving them

[fdeazeve]

Summary

When a dbg.value is moved into a new function, the corresponding
variable should have its entire scope chain reparented with the new
function. The current implementation drops the scope chain and replaces
it with a subprogram for the new function.

Depends on D139670

[fdeazeve]

Updated stack

[fdeazeve]

Fix scope

[fdeazeve]

rebased

[fdeazeve]

MLIR failures unrelated to this

llvm/lib/Transforms/Utils/CodeExtractor.cpp

static void fixupDebugInfoPostExtraction(Function &OldFunc, Function &NewFunc,
   // Debug intrinsics in the new function need to be updated in one of two
   // ways:
   //  1) They need to be deleted, because they describe a value in the old
   //     function.
   //  2) They need to point to fresh metadata, e.g. because they currently
   //     point to a variable in the wrong scope.
   SmallDenseMap<DINode *, DINode *> RemappedMetadata;
   SmallVector<Instruction *, 4> DebugIntrinsicsToDelete;
+  DenseMap<const MDNode *, MDNode *> Cache;
   for (Instruction &I : instructions(NewFunc)) {
     auto *DII = dyn_cast<DbgInfoIntrinsic>(&I);
     if (!DII)
       continue;
 
     // Point the intrinsic to a fresh label within the new function.
     if (auto *DLI = dyn_cast<DbgLabelInst>(&I)) {
       DILabel *OldLabel = DLI->getLabel();
if (any_of(DVI->location_ops(), IsInvalidLocation)) {
       DebugIntrinsicsToDelete.push_back(DVI);
       continue;
     }
     // If the variable was in the scope of the old function, i.e. it was not
     // inlined, point the intrinsic to a fresh variable within the new function.
     if (!DVI->getDebugLoc().getInlinedAt()) {
       DILocalVariable *OldVar = DVI->getVariable();
       DINode *&NewVar = RemappedMetadata[OldVar];
-      if (!NewVar)
+      if (!NewVar) {
+        DILocalScope *NewScope = DILocalScope::cloneScopeForSubprogram(
+            *OldVar->getScope(), *NewSP, Ctx, Cache);
         NewVar = DIB.createAutoVariable(
-            NewSP, OldVar->getName(), OldVar->getFile(), OldVar->getLine(),
+            NewScope, OldVar->getName(), OldVar->getFile(), OldVar->getLine(),
             OldVar->getType(), /*AlwaysPreserve=*/false, DINode::FlagZero,
             OldVar->getAlignInBits());
+      }
       DVI->setVariable(cast<DILocalVariable>(NewVar));
     }
   }
 
   for (auto *DII : DebugIntrinsicsToDelete)
     DII->eraseFromParent();
   DIB.finalizeSubprogram(NewSP);
 
   // Fix up the scope information attached to the line locations in the new
   // function.
-  DenseMap<const MDNode *, MDNode *> Cache;
   for (Instruction &I : instructions(NewFunc)) {
     if (const DebugLoc &DL = I.getDebugLoc())
       I.setDebugLoc(
           DebugLoc::replaceInlinedAtSubprogram(DL, *NewSP, Ctx, Cache));
 
     // Loop info metadata may contain line locations. Fix them up.
     auto updateLoopInfoLoc = [&Ctx, NewSP](Metadata *MD) -> Metadata * {
       if (auto *Loc = dyn_cast_or_null<DILocation>(MD))

llvm/test/Transforms/HotColdSplit/transfer-debug-info.ll

 
 ; CHECK-NEXT: call void @sink(i32 [[ADD1]]), !dbg [[LINE2:![0-9]+]]
 ; CHECK-NEXT: call void @sink(i32 [[ADD1]]), !dbg [[LINE3:![0-9]+]]
 
 ; CHECK-NEXT: call void @llvm.dbg.value(metadata i32 [[ADD1]]
 ; CHECK-SAME:      metadata [[VAR_FROM_INLINE_ME:![0-9]+]]
 ; CHECK-SAME:      !dbg [[LINE2]]
 
+; CHECK-NEXT: call void @llvm.dbg.value(metadata i32 [[ADD1]]
+; CHECK-SAME:      metadata [[VAR2:![0-9]+]]
+; CHECK-SAME:     !dbg [[LINE4:![0-9]+]]
+
+
 ; - The DISubprogram for @foo.cold.1 has an empty DISubroutineType
 ; CHECK: [[FILE:![0-9]+]] = !DIFile(filename: "<stdin>"
 ; CHECK: [[EMPTY_MD:![0-9]+]] = !{}
 ; CHECK: [[EMPTY_TYPE:![0-9]+]] = !DISubroutineType(types: [[EMPTY_MD]])
 ; CHECK: [[INLINE_ME_SCOPE:![0-9]+]] = distinct !DISubprogram(name: "inline_me"
 ; CHECK: [[NEWSCOPE:![0-9]+]] = distinct !DISubprogram(name: "foo.cold.1", linkageName: "foo.cold.1", scope: null, file: [[FILE]], type: [[EMPTY_TYPE]], spFlags: DISPFlagLocalToUnit | DISPFlagDefinition | DISPFlagOptimized
 
 ; - Line locations in @foo.cold.1 point to the new scope for @foo.cold.1
 ; CHECK: [[LINE1]] = !DILocation(line: 1, column: 1, scope: [[NEWSCOPE]])
 
 ; CHECK: [[LINE2]] =          !DILocation(line: 2, column: 2, scope: [[INLINE_ME_SCOPE]]
 ; CHECK-SAME:                            inlinedAt: [[LINE3]]
 ; CHECK: [[LINE3]] =          !DILocation(line: 3, column: 3, scope: [[INLINED_SCOPE1:![0-9]*]]
 ; CHECK: [[INLINED_SCOPE1]] = !DILexicalBlock(scope: [[INLINED_SCOPE2:![0-9]*]], file: [[FILE]], line: 4, column: 4)
 ; CHECK: [[INLINED_SCOPE2]] = !DILexicalBlock(scope: [[NEWSCOPE]], file: [[FILE]], line: 5, column: 5)
 
 ; CHECK: [[VAR_FROM_INLINE_ME]] = !DILocalVariable(name: "var_from_inline_me",
 ; CHECK-SAME:                                      scope: [[INLINE_ME_SCOPE]]
 
+; CHECK: [[VAR2]] = !DILocalVariable(name: "var_from_scope_in_foo",
+; CHECK-SAME:                        scope: [[NEWSCOPE2:![0-9]+]]
+; CHECK: [[NEWSCOPE2]] = !DILexicalBlock(scope: [[NEWSCOPE]], file: [[FILE]], line: 7, column: 7)
+; CHECK: [[LINE4]] =     !DILocation(line: 6, column: 6, scope: [[NEWSCOPE2]]
+
 define void @foo(i32 %arg1) !dbg !6 {
 entry:
   %var = add i32 0, 0, !dbg !11
   br i1 undef, label %if.then, label %if.end
 
 if.then:                                          ; preds = %entry
   ret void
 
 if.end:                                           ; preds = %entry
   call void @llvm.dbg.value(metadata i32 %arg1, metadata !9, metadata !DIExpression()), !dbg !11
   %add1 = add i32 %arg1, 1
   call void @sink(i32 %add1)
   call void @sink(i32 %add1), !dbg !11
   call void @llvm.dbg.value(metadata i32 %add1, metadata !9, metadata !DIExpression()), !dbg !11
   call void @llvm.dbg.value(metadata i32 %add1, metadata !9, metadata !DIExpression(DW_OP_constu, 1, DW_OP_plus, DW_OP_stack_value)), !dbg !11
   call void @sink(i32 %add1), !dbg !13 ; inlined from @inline_me
   call void @sink(i32 %add1), !dbg !14 ; not inlined, but inside some scope of foo
   call void @llvm.dbg.value(metadata i32 %add1, metadata !17, metadata !DIExpression()), !dbg !13 ; variable from @inline_me, should preserve scope in !17.
+  call void @llvm.dbg.value(metadata i32 %add1, metadata !18, metadata !DIExpression()), !dbg !19 ; variable not inlined, but inside some scope of foo
   ret void
 }
 
 declare void @llvm.dbg.value(metadata, metadata, metadata)
 
 declare void @sink(i32) cold
 
 define void @inline_me() !dbg !12{

 !10 = !DIBasicType(name: "ty32", size: 32, encoding: DW_ATE_unsigned)
 !11 = !DILocation(line: 1, column: 1, scope: !6)
 !12 = distinct !DISubprogram(name: "inline_me", linkageName: "inline_me", scope: null, file: !1, line: 1, type: !7, isLocal: false, isDefinition: true, scopeLine: 1, isOptimized: true, unit: !0, retainedNodes: !8)
 !13 = !DILocation(line: 2, column: 2, scope: !12, inlinedAt: !14)
 !14 = !DILocation(line: 3, column: 3, scope: !15)
 !15 = distinct !DILexicalBlock(scope: !16, file: !1, line: 4, column: 4)
 !16 = distinct !DILexicalBlock(scope: !6, file: !1, line: 5, column: 5)
 !17 = !DILocalVariable(name: "var_from_inline_me", scope: !12, file: !1, line: 1, type: !10)
+!18 = !DILocalVariable(name: "var_from_scope_in_foo", scope: !20, file: !1, line: 1, type: !10)
+!19 = !DILocation(line: 6, column: 6, scope: !20)
+!20 = distinct !DILexicalBlock(scope: !6, file: !1, line: 7, column: 7)