[KeyInstr][Clang] Ret atom

OCHyams · OCHyams · commit fc97db31e40a · 2025-06-03T09:24:52.000+01:00
This patch is part of a stack that teaches Clang to generate Key Instructions metadata for C and C++. The Key Instructions project is introduced, including a "quick summary" section at the top which adds context for this PR, here: https://discourse.llvm.org/t/rfc-improving-is-stmt-placement-for-better-interactive-debugging/82668 The feature is only functional in LLVM if LLVM is built with CMake flag LLVM_EXPERIMENTAL_KEY_INSTRUCTIONs. Eventually that flag will be removed. The Clang-side work is demoed here: #130943
diff --git a/clang/lib/CodeGen/CGCall.cpp b/clang/lib/CodeGen/CGCall.cpp
@@ -3938,7 +3938,8 @@ void CodeGenFunction::EmitFunctionEpilog(const CGFunctionInfo &FI,
 
   // Functions with no result always return void.
   if (!ReturnValue.isValid()) {
-    Builder.CreateRetVoid();
+    auto *I = Builder.CreateRetVoid();
+    addRetToOverrideOrNewSourceAtom(I, nullptr);
     return;
   }
 
@@ -4118,6 +4119,9 @@ void CodeGenFunction::EmitFunctionEpilog(const CGFunctionInfo &FI,
 
   if (RetDbgLoc)
     Ret->setDebugLoc(std::move(RetDbgLoc));
+
+  llvm::Value *Backup = RV ? Ret->getOperand(0) : nullptr;
+  addRetToOverrideOrNewSourceAtom(cast<llvm::ReturnInst>(Ret), Backup);
 }
 
 void CodeGenFunction::EmitReturnValueCheck(llvm::Value *RV) {
diff --git a/clang/lib/CodeGen/CGCleanup.cpp b/clang/lib/CodeGen/CGCleanup.cpp
@@ -1118,6 +1118,8 @@ void CodeGenFunction::EmitBranchThroughCleanup(JumpDest Dest) {
 
   // Create the branch.
   llvm::BranchInst *BI = Builder.CreateBr(Dest.getBlock());
+  // This is the primary instruction for this atom, acting in place of a ret.
+  addInstToCurrentSourceAtom(BI, nullptr);
 
   // Calculate the innermost active normal cleanup.
   EHScopeStack::stable_iterator
diff --git a/clang/lib/CodeGen/CGStmt.cpp b/clang/lib/CodeGen/CGStmt.cpp
@@ -1571,6 +1571,7 @@ static bool isSwiftAsyncCallee(const CallExpr *CE) {
 /// if the function returns void, or may be missing one if the function returns
 /// non-void.  Fun stuff :).
 void CodeGenFunction::EmitReturnStmt(const ReturnStmt &S) {
+  ApplyAtomGroup Grp(getDebugInfo());
   if (requiresReturnValueCheck()) {
     llvm::Constant *SLoc = EmitCheckSourceLocation(S.getBeginLoc());
     auto *SLocPtr =
@@ -1580,6 +1581,7 @@ void CodeGenFunction::EmitReturnStmt(const ReturnStmt &S) {
     CGM.getSanitizerMetadata()->disableSanitizerForGlobal(SLocPtr);
     assert(ReturnLocation.isValid() && "No valid return location");
     Builder.CreateStore(SLocPtr, ReturnLocation);
+    //*OCH?*//
   }
 
   // Returning from an outlined SEH helper is UB, and we already warn on it.
@@ -1646,16 +1648,19 @@ void CodeGenFunction::EmitReturnStmt(const ReturnStmt &S) {
     // If this function returns a reference, take the address of the expression
     // rather than the value.
     RValue Result = EmitReferenceBindingToExpr(RV);
-    Builder.CreateStore(Result.getScalarVal(), ReturnValue);
+    auto *I = Builder.CreateStore(Result.getScalarVal(), ReturnValue);
+    addInstToCurrentSourceAtom(I, I->getValueOperand());
   } else {
     switch (getEvaluationKind(RV->getType())) {
     case TEK_Scalar: {
       llvm::Value *Ret = EmitScalarExpr(RV);
-      if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect)
+      if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect) {
         EmitStoreOfScalar(Ret, MakeAddrLValue(ReturnValue, RV->getType()),
                           /*isInit*/ true);
-      else
-        Builder.CreateStore(Ret, ReturnValue);
+      } else {
+        auto *I = Builder.CreateStore(Ret, ReturnValue);
+        addInstToCurrentSourceAtom(I, I->getValueOperand());
+      }
       break;
     }
     case TEK_Complex:
diff --git a/clang/lib/CodeGen/CodeGenFunction.cpp b/clang/lib/CodeGen/CodeGenFunction.cpp
@@ -341,6 +341,15 @@ llvm::DebugLoc CodeGenFunction::EmitReturnBlock() {
       // later by the actual 'ret' instruction.
       llvm::DebugLoc Loc = BI->getDebugLoc();
       Builder.SetInsertPoint(BI->getParent());
+
+      // Key Instructions: If there's only one `ret` then we want to put the
+      // instruction in the same source atom group as the store to the ret-value
+      // alloca and unconditional `br` to the return block that we're about to
+      // delete. It all comes from the same source (`return (value)`).
+      if (auto *DI = getDebugInfo(); DI && BI->getDebugLoc())
+        DI->setRetInstSourceAtomOverride(
+            BI->getDebugLoc().get()->getAtomGroup());
+
       BI->eraseFromParent();
       delete ReturnBlock.getBlock();
       ReturnBlock = JumpDest();
@@ -1549,6 +1558,12 @@ void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn,
       Bypasses.Init(CGM, Body);
   }
 
+  // Finalize function debug info on exit.
+  auto Cleanup = llvm::make_scope_exit([this] {
+    if (CGDebugInfo *DI = getDebugInfo())
+      DI->completeFunction();
+  });
+
   // Emit the standard function prologue.
   StartFunction(GD, ResTy, Fn, FnInfo, Args, Loc, BodyRange.getBegin());
 
diff --git a/clang/test/DebugInfo/KeyInstructions/return-va-arg.c b/clang/test/DebugInfo/KeyInstructions/return-va-arg.c
@@ -0,0 +1,25 @@
+// RUN: %clang -gkey-instructions -gno-column-info -x c++ %s -gmlt -S -emit-llvm -o - \
+// RUN: | FileCheck %s --implicit-check-not atomGroup --implicit-check-not atomRank
+
+// RUN: %clang -gkey-instructions -gno-column-info -x c %s -gmlt -S -emit-llvm -o - \
+// RUN: | FileCheck %s --implicit-check-not atomGroup --implicit-check-not atomRank
+
+typedef struct {
+  struct{} a;
+  double b;
+} s1;
+
+s1 f(int z, ...) {
+  __builtin_va_list list;
+  __builtin_va_start(list, z);
+// CHECK: vaarg.end:
+// CHECK-NEXT: %vaarg.addr = phi ptr
+// CHECK-NEXT: call void @llvm.memcpy{{.*}}, !dbg [[G1R1:!.*]]
+// CHECK-NEXT: {{.*}} = getelementptr{{.*}}
+// CHECK-NEXT: [[LOAD:%.*]] = load double{{.*}}, !dbg [[G1R2:!.*]]
+// CHECK-NEXT: ret double [[LOAD]], !dbg [[G1R1]]
+  return __builtin_va_arg(list, s1);
+}
+
+// CHECK: [[G1R1]] = !DILocation({{.*}}, atomGroup: 1, atomRank: 1)
+// CHECK: [[G1R2]] = !DILocation({{.*}}, atomGroup: 1, atomRank: 2)
diff --git a/clang/test/DebugInfo/KeyInstructions/return.c b/clang/test/DebugInfo/KeyInstructions/return.c
@@ -0,0 +1,90 @@
+// RUN: %clang -gkey-instructions -gno-column-info -x c++ %s -gmlt -S -emit-llvm -o - \
+// RUN: | FileCheck %s --check-prefixes=CHECK,CHECK-CXX
+
+// RUN: %clang -gkey-instructions -gno-column-info -x c %s -gmlt -S -emit-llvm -o - \
+// RUN: | FileCheck %s
+
+// Check the stores to `retval` allocas and branches to `return` block are in
+// the same atom group. They are both rank 1, which could in theory introduce
+// an extra step in some optimized code. This low risk currently feels an
+// acceptable for keeping the code a bit simpler (as opposed to adding
+// scaffolding to make the store rank 2).
+
+// Also check that in the case of a single return (no control flow) the
+// return instruction inherits the atom group of the branch to the return
+// block when the blocks get folded togather.
+
+#ifdef __cplusplus
+#define nomangle extern "C"
+#else
+#define nomangle
+#endif
+
+int g;
+nomangle float a() {
+// CHECK: float @a()
+  if (g)
+// CHECK: if.then:
+// CHECK-NEXT: %1 = load i32, ptr @g{{.*}}, !dbg [[G2R3:!.*]]
+// CHECK-NEXT: %conv = sitofp i32 %1 to float{{.*}}, !dbg [[G2R2:!.*]]
+// CHECK-NEXT: store float %conv, ptr %retval{{.*}}, !dbg [[G2R1:!.*]]
+// CHECK-NEXT: br label %return{{.*}}, !dbg [[G2R1]]
+    return g;
+// CHECK: if.end:
+// CHECK-NEXT: store float 1.000000e+00, ptr %retval{{.*}}, !dbg [[G3R1:!.*]]
+// CHECK-NEXT: br label %return, !dbg [[G3R1]]
+
+// CHECK: return:
+// CHECK-NEXT:  %2 = load float, ptr %retval{{.*}}, !dbg [[G4R2:!.*]]
+// CHECK-NEXT:  ret float %2{{.*}}, !dbg [[G4R1:!.*]]
+  return 1;
+}
+
+// CHECK: void @b()
+// CHECK: ret void{{.*}}, !dbg [[B_G1R1:!.*]]
+nomangle void b() { return; }
+
+// CHECK: i32 @c()
+// CHECK: %add = add{{.*}}, !dbg [[C_G1R2:!.*]]
+// CHECK: ret i32 %add{{.*}}, !dbg [[C_G1R1:!.*]]
+nomangle int c() { return g + 1; }
+
+// NOTE: (return) (g = 1) are two separate atoms.
+// CHECK: i32 @d()
+// CHECK: store{{.*}}, !dbg [[D_G2R1:!.*]]
+// CHECK: ret i32 1{{.*}}, !dbg [[D_G1R1:!.*]]
+nomangle int d() { return g = 1; }
+
+// The implicit return here get the line number of the closing brace; make it
+// key to match existing behaviour.
+// CHECK: ret void, !dbg [[E_G1R1:!.*]]
+nomangle void e() {}
+
+#ifdef __cplusplus
+// CHECK-CXX: ptr @_Z1fRi
+int &f(int &r) {
+// Include ctrl-flow to stop ret value store being elided.
+    if (r)
+// CHECK-CXX: if.then:
+// CHECK-CXX-NEXT: %2 = load ptr, ptr %r.addr{{.*}}, !dbg [[F_G2R2:!.*]]
+// CHECK-CXX-NEXT: store ptr %2, ptr %retval{{.*}}, !dbg [[F_G2R1:!.*]]
+// CHECK-CXX-NEXT: br label %return, !dbg [[F_G2R1:!.*]]
+    return r;
+  return g;
+}
+#endif
+
+// CHECK: [[G2R3]] = !DILocation({{.*}}, atomGroup: 2, atomRank: 3)
+// CHECK: [[G2R2]] = !DILocation({{.*}}, atomGroup: 2, atomRank: 2)
+// CHECK: [[G2R1]] = !DILocation({{.*}}, atomGroup: 2, atomRank: 1)
+// CHECK: [[G3R1]] = !DILocation({{.*}}, atomGroup: 3, atomRank: 1)
+// CHECK: [[G4R2]] = !DILocation({{.*}}, atomGroup: 4, atomRank: 2)
+// CHECK: [[G4R1]] = !DILocation({{.*}}, atomGroup: 4, atomRank: 1)
+// CHECK: [[B_G1R1]] = !DILocation({{.*}}, atomGroup: 1, atomRank: 1)
+// CHECK: [[C_G1R2]] = !DILocation({{.*}}, atomGroup: 1, atomRank: 2)
+// CHECK: [[C_G1R1]] = !DILocation({{.*}}, atomGroup: 1, atomRank: 1)
+// CHECK: [[D_G2R1]] = !DILocation({{.*}}, atomGroup: 2, atomRank: 1)
+// CHECK: [[D_G1R1]] = !DILocation({{.*}}, atomGroup: 1, atomRank: 1)
+// CHECK: [[E_G1R1]] = !DILocation({{.*}}, atomGroup: 1, atomRank: 1)
+// CHECK-CXX: [[F_G2R2]] = !DILocation({{.*}}, atomGroup: 2, atomRank: 2)
+// CHECK-CXX: [[F_G2R1]] = !DILocation({{.*}}, atomGroup: 2, atomRank: 1)
