Add option to generate additional debug info for expression dereferencing pointer to pointers.

clang/include/clang/Basic/DebugOptions.def

@@ -129,6 +129,10 @@ DEBUGOPT(CodeViewCommandLine, 1, 0)
 /// Whether emit extra debug info for sample pgo profile collection.
 DEBUGOPT(DebugInfoForProfiling, 1, 0)
 
+/// Whether to generate pseudo variables and their debug info for intermediate
+/// pointer accesses.
+DEBUGOPT(DebugInfoForPointerType, 1, 0)
+
 /// Whether to emit .debug_gnu_pubnames section instead of .debug_pubnames.
 DEBUGOPT(DebugNameTable, 2, 0)
 

clang/include/clang/Driver/Options.td

@@ -1675,6 +1675,10 @@ defm debug_info_for_profiling : BoolFOption<"debug-info-for-profiling",
   PosFlag<SetTrue, [], [ClangOption, CC1Option],
           "Emit extra debug info to make sample profile more accurate">,
   NegFlag<SetFalse>>;
+def fdebug_info_for_pointer_type : Flag<["-"], "fdebug-info-for-pointer-type">,
+  Group<f_Group>, Visibility<[ClangOption, CC1Option]>,
+  HelpText<"Generate pseudo variables and their debug info for intermediate pointer accesses">,
+  MarshallingInfoFlag<CodeGenOpts<"DebugInfoForPointerType">>;
 def fprofile_instr_generate : Flag<["-"], "fprofile-instr-generate">,
     Group<f_Group>, Visibility<[ClangOption, CLOption]>,
     HelpText<"Generate instrumented code to collect execution counts into default.profraw file (overridden by '=' form of option or LLVM_PROFILE_FILE env var)">;

clang/lib/CodeGen/CGDebugInfo.cpp

@@ -5636,6 +5636,31 @@ void CGDebugInfo::EmitExternalVariable(llvm::GlobalVariable *Var,
   Var->addDebugInfo(GVE);
 }
 
+void CGDebugInfo::EmitPseudoVariable(llvm::AllocaInst *Alloca, QualType Ty,
+                                     SourceLocation Loc) {
+  llvm::DIFile *Unit = getOrCreateFile(Loc);
+  unsigned Line = getLineNumber(Loc);
+  unsigned Column = getColumnNumber(Loc);
+  llvm::DILocalVariable *D = DBuilder.createAutoVariable(
+      LexicalBlockStack.back(), Alloca->getName(), getOrCreateFile(Loc), Line,
+      getOrCreateType(Ty, Unit));
+  llvm::DILocation *DIL =
+      llvm::DILocation::get(CGM.getLLVMContext(), Line, Column,
+                            LexicalBlockStack.back(), CurInlinedAt);
+  SmallVector<uint64_t> Expr;
+  DBuilder.insertDeclare(Alloca, D, DBuilder.createExpression(Expr), DIL,
+                         Alloca->getParent());
+}
+
+llvm::MDNode *CGDebugInfo::GetPseudoVariableAnnotation() {
+  if (!PseudoVariableAnnotation)
+    PseudoVariableAnnotation =
+        llvm::MDNode::get(CGM.getLLVMContext(),
+                          llvm::MDString::get(CGM.getLLVMContext(),
+                                              "fdebug-info-for-pointer-type"));
+  return PseudoVariableAnnotation;
+}
+
 void CGDebugInfo::EmitGlobalAlias(const llvm::GlobalValue *GV,
                                   const GlobalDecl GD) {
 

clang/lib/CodeGen/CGDebugInfo.h

@@ -170,6 +170,9 @@ class CGDebugInfo {
   /// The key is coroutine real parameters, value is DIVariable in LLVM IR.
   Param2DILocTy ParamDbgMappings;
 
+  /// Cached object for GetPseudoVariableAnnotation().
+  llvm::MDNode *PseudoVariableAnnotation = nullptr;
+
   /// Helper functions for getOrCreateType.
   /// @{
   /// Currently the checksum of an interface includes the number of
@@ -529,6 +532,16 @@ class CGDebugInfo {
   /// Emit information about an external variable.
   void EmitExternalVariable(llvm::GlobalVariable *GV, const VarDecl *Decl);
 
+  /// Emit debug information for a pseudo variable assigned to the value of an
+  /// intermediate expression, so that a performance counter can track the usage
+  /// of a specific expression of interest.
+  void EmitPseudoVariable(llvm::AllocaInst *Alloca, QualType Ty,
+                          SourceLocation Loc);
+
+  /// Get the special annotation tag that indicates the instruction is
+  /// associated with EmitPseudoVariable.
+  llvm::MDNode *GetPseudoVariableAnnotation();
+
   /// Emit information about global variable alias.
   void EmitGlobalAlias(const llvm::GlobalValue *GV, const GlobalDecl Decl);
 

clang/lib/CodeGen/CGDecl.cpp

@@ -793,6 +793,7 @@ void CodeGenFunction::EmitScalarInit(const Expr *init, const ValueDecl *D,
   Qualifiers::ObjCLifetime lifetime = lvalue.getObjCLifetime();
   if (!lifetime) {
     llvm::Value *value = EmitScalarExpr(init);
+    value = UnemitPseudoVariable(value);
     if (capturedByInit)
       drillIntoBlockVariable(*this, lvalue, cast<VarDecl>(D));
     EmitNullabilityCheck(lvalue, value, init->getExprLoc());

clang/lib/CodeGen/CGExpr.cpp

@@ -951,6 +951,39 @@ static llvm::Value *getArrayIndexingBound(CodeGenFunction &CGF,
   return nullptr;
 }
 
+/// When a pseudo variable is created for %1, it generates these instructions
+/// in sequence and return %2:
+/// %pseudo = alloca Ty
+/// call void @llvm.dbg.declare(metadata ptr %pseudo, metadata, metadata)
+/// store Ty %1, ptr %pseudo
+/// %2 = load ptr, ptr %pseudo
+/// To undo, we detect and remove this sequence, and replace %2 back with %1.
+llvm::Value *CodeGenFunction::UnemitPseudoVariable(llvm::Value *V) {
+  if (CGDebugInfo *DI = getDebugInfo()) {
+    if (llvm::LoadInst *Load = dyn_cast<llvm::LoadInst>(V)) {
+      if (llvm::MDNode *Tag =
+              Load->getMetadata(llvm::LLVMContext::MD_annotation)) {
+        if (Tag == DI->GetPseudoVariableAnnotation()) {
+          llvm::Value *PseudoVar = Load->getPointerOperand();
+          llvm::AllocaInst *Alloca = dyn_cast<llvm::AllocaInst>(PseudoVar);
+          llvm::StoreInst *Store =
+              dyn_cast_if_present<llvm::StoreInst>(Load->getPrevNode());
+          assert(Store && Store->getPointerOperand() == PseudoVar);
+          llvm::Value *OriginalValue = Store->getValueOperand();
+          V->replaceAllUsesWith(OriginalValue);
+          assert(Store->getPrevNode()->getPrevNode() == PseudoVar);
+          auto It = Load->getIterator();
+          for (int i = 0; i < 4; i++) {
+            (It--)->eraseFromParent();
+          }
+          return OriginalValue;
+        }
+      }
+    }
+  }
+  return V;
+}
+
 namespace {
 
 /// \p StructAccessBase returns the base \p Expr of a field access. It returns
@@ -2015,6 +2048,40 @@ llvm::Value *CodeGenFunction::EmitLoadOfScalar(Address Addr, bool Volatile,
                         llvm::MDNode::get(getLLVMContext(), std::nullopt));
     }
 
+  // if -g2 or above and -fdebug-info-for-pointer-type are enabled, emit
+  // additional debug info for loads in an intermediate expression, which allows
+  // a performance counter to deduce the type of the value being loaded, even if
+  // it does not correspond to a variable in the source code.
+  // Since there is no variable correspond to an intermediate expression, we
+  // create a pseudo variable for it and emit its debug info, as if the
+  // expression were written in SSA form.
+  if (CGM.getCodeGenOpts().getDebugInfo() >
+          llvm::codegenoptions::DebugLineTablesOnly &&
+      CGM.getCodeGenOpts().DebugInfoForPointerType) {
+    if (CGDebugInfo *DI = getDebugInfo()) {
+      // We only generate this debug info if loading from GEP, not from other
+      // cases such as loading a function argument.
+      if (isa<llvm::GetElementPtrInst>(Load->getOperand(0))) {
+        const llvm::DebugLoc &DebugLoc = Load->getDebugLoc();
+        llvm::AllocaInst *PseudoVar =
+            Builder.CreateAlloca(Load->getType(), nullptr,
+                                 Twine("pseudo_")
+                                     .concat(Twine(DebugLoc.getLine()))
+                                     .concat("_")
+                                     .concat(Twine(DebugLoc.getCol())));
+        DI->EmitPseudoVariable(PseudoVar, Ty, Loc);
+        Address PseudoVarAddr(PseudoVar, Load->getType(), Addr.getAlignment());
+        Builder.CreateStore(Load, PseudoVarAddr);
+        Load = Builder.CreateLoad(PseudoVarAddr);
+        // Set a special metadata tag to this instruction, in the case we need
+        // to revert it because there is already a destination variable for the
+        // load.
+        Load->setMetadata(llvm::LLVMContext::MD_annotation,
+                          DI->GetPseudoVariableAnnotation());
+      }
+    }
+  }
+
   return EmitFromMemory(Load, Ty);
 }
 
@@ -5569,6 +5636,8 @@ LValue CodeGenFunction::EmitBinaryOperatorLValue(const BinaryOperator *E) {
     }
 
     RValue RV = EmitAnyExpr(E->getRHS());
+    if (isa<DeclRefExpr>(E->getLHS()) && RV.isScalar())
+      RV = RValue::get(UnemitPseudoVariable(RV.getScalarVal()));
     LValue LV = EmitCheckedLValue(E->getLHS(), TCK_Store);
     if (RV.isScalar())
       EmitNullabilityCheck(LV, RV.getScalarVal(), E->getExprLoc());

clang/lib/CodeGen/CodeGenFunction.h

@@ -3104,6 +3104,11 @@ class CodeGenFunction : public CodeGenTypeCache {
   /// Get the record field index as represented in debug info.
   unsigned getDebugInfoFIndex(const RecordDecl *Rec, unsigned FieldIndex);
 
+  /// When the result of EmitLoadOfScalar is immediately assigned to a declared
+  /// variable, the pseudo variable emitted for it (when the flag
+  /// -fdebug-info-for-pointer-type is specified) should be undone since there
+  /// is already a debug value emitted for the declared variable.
+  llvm::Value *UnemitPseudoVariable(llvm::Value *V);
 
   //===--------------------------------------------------------------------===//
   //                            Declaration Emission

clang/lib/Driver/ToolChains/Clang.cpp

@@ -4256,6 +4256,9 @@ renderDebugOptions(const ToolChain &TC, const Driver &D, const llvm::Triple &T,
   // decision should be made in the driver as well though.
   llvm::DebuggerKind DebuggerTuning = TC.getDefaultDebuggerTuning();
 
+  if (Args.hasArg(options::OPT_fdebug_info_for_pointer_type))
+    CmdArgs.push_back("-fdebug-info-for-pointer-type");
+
   bool SplitDWARFInlining =
       Args.hasFlag(options::OPT_fsplit_dwarf_inlining,
                    options::OPT_fno_split_dwarf_inlining, false);