[clang] [llvm] Add option to generate additional debug info for expression dereferencing pointer to pointers. (PR #81545)
William Junda Huang via llvm-commits
llvm-commits at lists.llvm.org
Wed Mar 13 15:08:35 PDT 2024
https://github.com/huangjd updated https://github.com/llvm/llvm-project/pull/81545
>From f2c82758e1cba7773e41d941d2812c829c339675 Mon Sep 17 00:00:00 2001
From: William Huang <williamjhuang at google.com>
Date: Mon, 12 Feb 2024 02:27:13 -0500
Subject: [PATCH 1/3] Add option to generate additional info for expression
containing pointer of pointers.
Such expression does correspond to a variable in the source code thus
does not have a debug location. However the user may want to collect
sampling counter for memory accesses to analyze usage frequency of class
members. By enabling -fdebug_info_for_pointer_type a psuedo variable and
its debug info is generated in place whenever there's an intermediate
expression with pointer access.
---
clang/include/clang/Basic/DebugOptions.def | 4 ++
clang/include/clang/Driver/Options.td | 4 ++
clang/lib/CodeGen/CGDebugInfo.cpp | 16 +++++
clang/lib/CodeGen/CGDebugInfo.h | 6 ++
clang/lib/CodeGen/CGDecl.cpp | 4 ++
clang/lib/CodeGen/CGExpr.cpp | 79 ++++++++++++++++++++++
clang/lib/CodeGen/CodeGenFunction.h | 5 ++
clang/lib/Driver/ToolChains/Clang.cpp | 3 +
8 files changed, 121 insertions(+)
diff --git a/clang/include/clang/Basic/DebugOptions.def b/clang/include/clang/Basic/DebugOptions.def
index 7cd3edf08a17ea..6dd09f46842077 100644
--- a/clang/include/clang/Basic/DebugOptions.def
+++ b/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)
diff --git a/clang/include/clang/Driver/Options.td b/clang/include/clang/Driver/Options.td
index 7f4fa33748faca..96b22d3f7640dd 100644
--- a/clang/include/clang/Driver/Options.td
+++ b/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)">;
diff --git a/clang/lib/CodeGen/CGDebugInfo.cpp b/clang/lib/CodeGen/CGDebugInfo.cpp
index 0f3f684d61dc94..6ce40da22dc97d 100644
--- a/clang/lib/CodeGen/CGDebugInfo.cpp
+++ b/clang/lib/CodeGen/CGDebugInfo.cpp
@@ -5636,6 +5636,22 @@ 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());
+}
+
void CGDebugInfo::EmitGlobalAlias(const llvm::GlobalValue *GV,
const GlobalDecl GD) {
diff --git a/clang/lib/CodeGen/CGDebugInfo.h b/clang/lib/CodeGen/CGDebugInfo.h
index 7b60e94555d060..a2c484f50b2bc5 100644
--- a/clang/lib/CodeGen/CGDebugInfo.h
+++ b/clang/lib/CodeGen/CGDebugInfo.h
@@ -529,6 +529,12 @@ 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);
+
/// Emit information about global variable alias.
void EmitGlobalAlias(const llvm::GlobalValue *GV, const GlobalDecl Decl);
diff --git a/clang/lib/CodeGen/CGDecl.cpp b/clang/lib/CodeGen/CGDecl.cpp
index bbe14ef4c17244..5f7b2529179003 100644
--- a/clang/lib/CodeGen/CGDecl.cpp
+++ b/clang/lib/CodeGen/CGDecl.cpp
@@ -793,6 +793,10 @@ void CodeGenFunction::EmitScalarInit(const Expr *init, const ValueDecl *D,
Qualifiers::ObjCLifetime lifetime = lvalue.getObjCLifetime();
if (!lifetime) {
llvm::Value *value = EmitScalarExpr(init);
+ if (CGM.getCodeGenOpts().getDebugInfo() >
+ llvm::codegenoptions::DebugLineTablesOnly &&
+ CGM.getCodeGenOpts().DebugInfoForPointerType)
+ value = UnemitPseudoVariable(value);
if (capturedByInit)
drillIntoBlockVariable(*this, lvalue, cast<VarDecl>(D));
EmitNullabilityCheck(lvalue, value, init->getExprLoc());
diff --git a/clang/lib/CodeGen/CGExpr.cpp b/clang/lib/CodeGen/CGExpr.cpp
index c5f6b6d3a99f0b..b979c0830c5b34 100644
--- a/clang/lib/CodeGen/CGExpr.cpp
+++ b/clang/lib/CodeGen/CGExpr.cpp
@@ -951,6 +951,58 @@ 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 to %1.
+llvm::Value *CodeGenFunction::UnemitPseudoVariable(llvm::Value *V) {
+ if (!getDebugInfo())
+ return V;
+
+ if (llvm::LoadInst *Load = dyn_cast<llvm::LoadInst>(V)) {
+ llvm::Value *PseudoVar = Load->getPointerOperand();
+ if (llvm::StoreInst *Store =
+ dyn_cast<llvm::StoreInst>(Load->getPrevNode())) {
+ if (Store->getPointerOperand() != PseudoVar)
+ return V;
+ llvm::Value *OriginalValue = Store->getValueOperand();
+ if (llvm::CallInst *DbgCall =
+ dyn_cast<llvm::CallInst>(Store->getPrevNode())) {
+ if (DbgCall->getCalledFunction() !=
+ llvm::Intrinsic::getDeclaration(&CGM.getModule(),
+ llvm::Intrinsic::dbg_declare) ||
+ DbgCall->getNumOperands() != 4)
+ return V;
+ for (int i = 0; i < 3; i++) {
+ if (!isa<llvm::MetadataAsValue>(DbgCall->getArgOperand(i)))
+ return V;
+ }
+ if (llvm::MetadataAsValue *Metadata =
+ dyn_cast<llvm::MetadataAsValue>(DbgCall->getOperand(0))) {
+ if (llvm::ValueAsMetadata *Value =
+ dyn_cast<llvm::ValueAsMetadata>(Metadata->getMetadata())) {
+ if (Value->getValue() != PseudoVar)
+ return V;
+ if (llvm::AllocaInst *Alloca =
+ dyn_cast<llvm::AllocaInst>(DbgCall->getPrevNode())) {
+ V->replaceAllUsesWith(OriginalValue);
+ Load->eraseFromParent();
+ Store->eraseFromParent();
+ DbgCall->eraseFromParent();
+ Alloca->eraseFromParent();
+ return OriginalValue;
+ }
+ }
+ }
+ }
+ }
+ }
+ return V;
+}
+
namespace {
/// \p StructAccessBase returns the base \p Expr of a field access. It returns
@@ -2015,6 +2067,29 @@ 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);
+ }
+ }
+
return EmitFromMemory(Load, Ty);
}
@@ -5569,6 +5644,10 @@ LValue CodeGenFunction::EmitBinaryOperatorLValue(const BinaryOperator *E) {
}
RValue RV = EmitAnyExpr(E->getRHS());
+ if (CGM.getCodeGenOpts().getDebugInfo() > llvm::codegenoptions::DebugLineTablesOnly &&
+ CGM.getCodeGenOpts().DebugInfoForPointerType)
+ 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());
diff --git a/clang/lib/CodeGen/CodeGenFunction.h b/clang/lib/CodeGen/CodeGenFunction.h
index 143ad64e8816b1..36a572ace2ef64 100644
--- a/clang/lib/CodeGen/CodeGenFunction.h
+++ b/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
diff --git a/clang/lib/Driver/ToolChains/Clang.cpp b/clang/lib/Driver/ToolChains/Clang.cpp
index bcba7cbbdb58c2..7882c4f1225f1f 100644
--- a/clang/lib/Driver/ToolChains/Clang.cpp
+++ b/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);
>From 95235b97774f8700ee9108e057ac360e86c79597 Mon Sep 17 00:00:00 2001
From: William Huang <williamjhuang at google.com>
Date: Wed, 6 Mar 2024 05:59:07 -0500
Subject: [PATCH 2/3] Code cleanup
---
clang/lib/CodeGen/CGDebugInfo.cpp | 9 ++++
clang/lib/CodeGen/CGDebugInfo.h | 7 +++
clang/lib/CodeGen/CGDecl.cpp | 5 +-
clang/lib/CodeGen/CGExpr.cpp | 80 ++++++++++++++-----------------
4 files changed, 52 insertions(+), 49 deletions(-)
diff --git a/clang/lib/CodeGen/CGDebugInfo.cpp b/clang/lib/CodeGen/CGDebugInfo.cpp
index 6ce40da22dc97d..7aa96332e495b3 100644
--- a/clang/lib/CodeGen/CGDebugInfo.cpp
+++ b/clang/lib/CodeGen/CGDebugInfo.cpp
@@ -5652,6 +5652,15 @@ void CGDebugInfo::EmitPseudoVariable(llvm::AllocaInst *Alloca, QualType Ty,
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) {
diff --git a/clang/lib/CodeGen/CGDebugInfo.h b/clang/lib/CodeGen/CGDebugInfo.h
index a2c484f50b2bc5..047095073c6965 100644
--- a/clang/lib/CodeGen/CGDebugInfo.h
+++ b/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
@@ -535,6 +538,10 @@ class CGDebugInfo {
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);
diff --git a/clang/lib/CodeGen/CGDecl.cpp b/clang/lib/CodeGen/CGDecl.cpp
index 5f7b2529179003..04cd0eaddb83ad 100644
--- a/clang/lib/CodeGen/CGDecl.cpp
+++ b/clang/lib/CodeGen/CGDecl.cpp
@@ -793,10 +793,7 @@ void CodeGenFunction::EmitScalarInit(const Expr *init, const ValueDecl *D,
Qualifiers::ObjCLifetime lifetime = lvalue.getObjCLifetime();
if (!lifetime) {
llvm::Value *value = EmitScalarExpr(init);
- if (CGM.getCodeGenOpts().getDebugInfo() >
- llvm::codegenoptions::DebugLineTablesOnly &&
- CGM.getCodeGenOpts().DebugInfoForPointerType)
- value = UnemitPseudoVariable(value);
+ value = UnemitPseudoVariable(value);
if (capturedByInit)
drillIntoBlockVariable(*this, lvalue, cast<VarDecl>(D));
EmitNullabilityCheck(lvalue, value, init->getExprLoc());
diff --git a/clang/lib/CodeGen/CGExpr.cpp b/clang/lib/CodeGen/CGExpr.cpp
index b979c0830c5b34..43c962da7a3e3e 100644
--- a/clang/lib/CodeGen/CGExpr.cpp
+++ b/clang/lib/CodeGen/CGExpr.cpp
@@ -957,45 +957,26 @@ static llvm::Value *getArrayIndexingBound(CodeGenFunction &CGF,
/// 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 to %1.
+/// To undo, we detect and remove this sequence, and replace %2 back with %1.
llvm::Value *CodeGenFunction::UnemitPseudoVariable(llvm::Value *V) {
- if (!getDebugInfo())
- return V;
-
- if (llvm::LoadInst *Load = dyn_cast<llvm::LoadInst>(V)) {
- llvm::Value *PseudoVar = Load->getPointerOperand();
- if (llvm::StoreInst *Store =
- dyn_cast<llvm::StoreInst>(Load->getPrevNode())) {
- if (Store->getPointerOperand() != PseudoVar)
- return V;
- llvm::Value *OriginalValue = Store->getValueOperand();
- if (llvm::CallInst *DbgCall =
- dyn_cast<llvm::CallInst>(Store->getPrevNode())) {
- if (DbgCall->getCalledFunction() !=
- llvm::Intrinsic::getDeclaration(&CGM.getModule(),
- llvm::Intrinsic::dbg_declare) ||
- DbgCall->getNumOperands() != 4)
- return V;
- for (int i = 0; i < 3; i++) {
- if (!isa<llvm::MetadataAsValue>(DbgCall->getArgOperand(i)))
- return V;
- }
- if (llvm::MetadataAsValue *Metadata =
- dyn_cast<llvm::MetadataAsValue>(DbgCall->getOperand(0))) {
- if (llvm::ValueAsMetadata *Value =
- dyn_cast<llvm::ValueAsMetadata>(Metadata->getMetadata())) {
- if (Value->getValue() != PseudoVar)
- return V;
- if (llvm::AllocaInst *Alloca =
- dyn_cast<llvm::AllocaInst>(DbgCall->getPrevNode())) {
- V->replaceAllUsesWith(OriginalValue);
- Load->eraseFromParent();
- Store->eraseFromParent();
- DbgCall->eraseFromParent();
- Alloca->eraseFromParent();
- return OriginalValue;
- }
+ 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;
}
}
}
@@ -2074,20 +2055,31 @@ llvm::Value *CodeGenFunction::EmitLoadOfScalar(Address Addr, bool Volatile,
// 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 &&
+ if (CGM.getCodeGenOpts().getDebugInfo() >
+ llvm::codegenoptions::DebugLineTablesOnly &&
CGM.getCodeGenOpts().DebugInfoForPointerType) {
- if (CGDebugInfo *DI = getDebugInfo())
+ 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())));
+ 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);
@@ -5644,10 +5636,8 @@ LValue CodeGenFunction::EmitBinaryOperatorLValue(const BinaryOperator *E) {
}
RValue RV = EmitAnyExpr(E->getRHS());
- if (CGM.getCodeGenOpts().getDebugInfo() > llvm::codegenoptions::DebugLineTablesOnly &&
- CGM.getCodeGenOpts().DebugInfoForPointerType)
- if (isa<DeclRefExpr>(E->getLHS()) && RV.isScalar())
- RV = RValue::get(UnemitPseudoVariable(RV.getScalarVal()));
+ 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());
>From 37d3c1b74d1931819de211bbdb9ff07bf7c092d7 Mon Sep 17 00:00:00 2001
From: William Huang <williamjhuang at google.com>
Date: Wed, 13 Mar 2024 18:07:42 -0400
Subject: [PATCH 3/3] Rewrite code to emit debug info for pointer instead of
pointee, as requested by kernel dev.
---
clang/lib/CodeGen/CGDebugInfo.cpp | 68 ++++++++++++++++++++++------
clang/lib/CodeGen/CGDebugInfo.h | 17 +++----
clang/lib/CodeGen/CGDecl.cpp | 1 -
clang/lib/CodeGen/CGExpr.cpp | 69 -----------------------------
clang/lib/CodeGen/CGExprScalar.cpp | 19 +++++++-
clang/lib/CodeGen/CodeGenFunction.h | 5 ---
llvm/include/llvm/IR/DIBuilder.h | 2 +
7 files changed, 79 insertions(+), 102 deletions(-)
diff --git a/clang/lib/CodeGen/CGDebugInfo.cpp b/clang/lib/CodeGen/CGDebugInfo.cpp
index 7aa96332e495b3..1aaed9e9f15eb7 100644
--- a/clang/lib/CodeGen/CGDebugInfo.cpp
+++ b/clang/lib/CodeGen/CGDebugInfo.cpp
@@ -5636,29 +5636,69 @@ void CGDebugInfo::EmitExternalVariable(llvm::GlobalVariable *Var,
Var->addDebugInfo(GVE);
}
-void CGDebugInfo::EmitPseudoVariable(llvm::AllocaInst *Alloca, QualType Ty,
+void CGDebugInfo::EmitPseudoVariable(CGBuilderTy &Builder,
+ llvm::Instruction *Value, QualType Ty,
SourceLocation Loc) {
+ // Only when -g2 or above is specified, debug info for variables will be
+ // generated.
+ if (CGM.getCodeGenOpts().getDebugInfo() <=
+ llvm::codegenoptions::DebugLineTablesOnly)
+ return;
+
llvm::DIFile *Unit = getOrCreateFile(Loc);
+ llvm::DIType *Type = getOrCreateType(Ty, Unit);
+
+ // Check if Value is already a declared variable and has debug info, in this
+ // case we have nothing to do. Clang emits declared variable as alloca, and
+ // it is loaded upon use, so we identify such pattern here.
+ if (llvm::LoadInst *Load = dyn_cast<llvm::LoadInst>(Value)) {
+ llvm::Value *Var = Load->getPointerOperand();
+ if (llvm::Metadata *MDValue = llvm::ValueAsMetadata::getIfExists(Var)) {
+ if (llvm::Value *DbgValue = llvm::MetadataAsValue::getIfExists(
+ CGM.getLLVMContext(), MDValue)) {
+ for (llvm::User *U : DbgValue->users()) {
+ if (llvm::CallInst *DbgDeclare = dyn_cast<llvm::CallInst>(U)) {
+ if (DbgDeclare->getCalledFunction() == DBuilder.GetDeclareFn() &&
+ DbgDeclare->getArgOperand(0) == DbgValue) {
+ // There can be implicit type cast applied on a variable if it is
+ // an opaque ptr, in this case its debug info may not match the
+ // actual type of object being used as in the next instruction, so
+ // we will need to emit a pseudo variable for type-casted value.
+ llvm::DILocalVariable *MDNode = dyn_cast<llvm::DILocalVariable>(
+ dyn_cast<llvm::MetadataAsValue>(DbgDeclare->getOperand(1))
+ ->getMetadata());
+ if (MDNode->getType() == Type)
+ return;
+ }
+ }
+ }
+ }
+ }
+ }
+
+ // Insert a sequence of instructions to materialize Value on the stack.
+ auto SaveInsertionPoint = Builder.saveIP();
+ Builder.SetInsertPoint(++(Value->getIterator()));
+ llvm::AllocaInst *PseudoVar = Builder.CreateAlloca(Value->getType());
+ Address PseudoVarAddr(PseudoVar, Value->getType(),
+ CharUnits::fromQuantity(PseudoVar->getAlign()));
+ llvm::LoadInst *Load = Builder.CreateLoad(PseudoVarAddr);
+ Value->replaceAllUsesWith(Load);
+ Builder.SetInsertPoint(Load);
+ Builder.CreateStore(Value, PseudoVarAddr);
+
+ // Emit debug info for materialized Value.
unsigned Line = getLineNumber(Loc);
unsigned Column = getColumnNumber(Loc);
llvm::DILocalVariable *D = DBuilder.createAutoVariable(
- LexicalBlockStack.back(), Alloca->getName(), getOrCreateFile(Loc), Line,
- getOrCreateType(Ty, Unit));
+ LexicalBlockStack.back(), "pseudo_var", Unit, Line, Type);
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;
+ DBuilder.insertDeclare(PseudoVar, D, DBuilder.createExpression(Expr), DIL,
+ Load);
+ Builder.restoreIP(SaveInsertionPoint);
}
void CGDebugInfo::EmitGlobalAlias(const llvm::GlobalValue *GV,
diff --git a/clang/lib/CodeGen/CGDebugInfo.h b/clang/lib/CodeGen/CGDebugInfo.h
index 047095073c6965..6452861a1be104 100644
--- a/clang/lib/CodeGen/CGDebugInfo.h
+++ b/clang/lib/CodeGen/CGDebugInfo.h
@@ -170,9 +170,6 @@ 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
@@ -532,15 +529,11 @@ 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 a pseudo variable and debug info for an intermediate value if it does
+ /// not correspond to a variable in the source code, so that a profiler can
+ /// track more accurate usage of certain instructions of interest.
+ void EmitPseudoVariable(CGBuilderTy &Builder, llvm::Instruction *Value,
+ QualType Ty, SourceLocation Loc);
/// Emit information about global variable alias.
void EmitGlobalAlias(const llvm::GlobalValue *GV, const GlobalDecl Decl);
diff --git a/clang/lib/CodeGen/CGDecl.cpp b/clang/lib/CodeGen/CGDecl.cpp
index 04cd0eaddb83ad..bbe14ef4c17244 100644
--- a/clang/lib/CodeGen/CGDecl.cpp
+++ b/clang/lib/CodeGen/CGDecl.cpp
@@ -793,7 +793,6 @@ 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());
diff --git a/clang/lib/CodeGen/CGExpr.cpp b/clang/lib/CodeGen/CGExpr.cpp
index 43c962da7a3e3e..c5f6b6d3a99f0b 100644
--- a/clang/lib/CodeGen/CGExpr.cpp
+++ b/clang/lib/CodeGen/CGExpr.cpp
@@ -951,39 +951,6 @@ 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
@@ -2048,40 +2015,6 @@ 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);
}
@@ -5636,8 +5569,6 @@ 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());
diff --git a/clang/lib/CodeGen/CGExprScalar.cpp b/clang/lib/CodeGen/CGExprScalar.cpp
index 181b15e9c7d0a7..0d64607bac399a 100644
--- a/clang/lib/CodeGen/CGExprScalar.cpp
+++ b/clang/lib/CodeGen/CGExprScalar.cpp
@@ -1787,7 +1787,24 @@ Value *ScalarExprEmitter::VisitMemberExpr(MemberExpr *E) {
}
}
- return EmitLoadOfLValue(E);
+ llvm::Value *Result = EmitLoadOfLValue(E);
+
+ // If -fdebug-info-for-pointer-type is specified, emit a pseudo variable and
+ // its debug info for the pointer, even if there is no variable associated
+ // with the pointer's expression.
+ if (CGF.CGM.getCodeGenOpts().DebugInfoForPointerType && CGF.getDebugInfo()) {
+ if (llvm::LoadInst *Load = dyn_cast<llvm::LoadInst>(Result)) {
+ if (llvm::GetElementPtrInst *GEP =
+ dyn_cast<llvm::GetElementPtrInst>(Load->getPointerOperand())) {
+ if (llvm::Instruction *Pointer =
+ dyn_cast<llvm::Instruction>(GEP->getPointerOperand())) {
+ CGF.getDebugInfo()->EmitPseudoVariable(
+ Builder, Pointer, E->getBase()->getType(), E->getExprLoc());
+ }
+ }
+ }
+ }
+ return Result;
}
Value *ScalarExprEmitter::VisitArraySubscriptExpr(ArraySubscriptExpr *E) {
diff --git a/clang/lib/CodeGen/CodeGenFunction.h b/clang/lib/CodeGen/CodeGenFunction.h
index 36a572ace2ef64..143ad64e8816b1 100644
--- a/clang/lib/CodeGen/CodeGenFunction.h
+++ b/clang/lib/CodeGen/CodeGenFunction.h
@@ -3104,11 +3104,6 @@ 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
diff --git a/llvm/include/llvm/IR/DIBuilder.h b/llvm/include/llvm/IR/DIBuilder.h
index edec161b397155..e420f73e152907 100644
--- a/llvm/include/llvm/IR/DIBuilder.h
+++ b/llvm/include/llvm/IR/DIBuilder.h
@@ -1024,6 +1024,8 @@ namespace llvm {
N->replaceAllUsesWith(Replacement);
return Replacement;
}
+
+ Function *GetDeclareFn() { return DeclareFn; }
};
// Create wrappers for C Binding types (see CBindingWrapping.h).
More information about the llvm-commits
mailing list