[KeyInstr][Clang] Add ApplyAtomGroup #134632
Conversation
llvmbot
added
clang
|
@llvm/pr-subscribers-clang @llvm/pr-subscribers-debuginfo Author: Orlando Cazalet-Hyams (OCHyams) ChangesThis is a scoped helper similar to ApplyDebugLocation that creates a new source A source atom is a source construct that is "interesting" for debug stepping This patch is part of a stack that teaches Clang to generate Key Instructions The Key Instructions project is introduced, including a "quick summary" section The feature is only functional in LLVM if LLVM is built with CMake flag The Clang-side work is demoed here: Full diff: https://github.com/llvm/llvm-project/pull/134632.diff 3 Files Affected:
diff --git a/clang/lib/CodeGen/CGDebugInfo.cpp b/clang/lib/CodeGen/CGDebugInfo.cpp
index 54025b767dc81..9fe1f24317b8a 100644
--- a/clang/lib/CodeGen/CGDebugInfo.cpp
+++ b/clang/lib/CodeGen/CGDebugInfo.cpp
@@ -43,6 +43,7 @@
#include "llvm/IR/Constants.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/DerivedTypes.h"
+#include "llvm/IR/Instruction.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/Metadata.h"
@@ -52,6 +53,7 @@
#include "llvm/Support/SHA1.h"
#include "llvm/Support/SHA256.h"
#include "llvm/Support/TimeProfiler.h"
+#include <cstdint>
#include <optional>
using namespace clang;
using namespace clang::CodeGen;
@@ -119,6 +121,114 @@ CGDebugInfo::~CGDebugInfo() {
"Region stack mismatch, stack not empty!");
}
+void CGDebugInfo::addInstSourceAtomMetadata(llvm::Instruction *I,
+ uint64_t Group, uint8_t Rank) {
+ if (!I->getDebugLoc() || Group == 0 || !I->getDebugLoc()->getLine())
+ return;
+
+ // Saturate the 3-bit rank.
+ Rank = std::min<uint8_t>(Rank, 7);
+
+ const llvm::DebugLoc &DL = I->getDebugLoc();
+
+ // Each instruction can only be attributed to one source atom (a limitation of
+ // the implementation). If this instruction is already part of a source atom,
+ // pick the group in which it has highest precedence (lowest rank).
+ if (DL.get()->getAtomGroup() && DL.get()->getAtomRank() &&
+ DL.get()->getAtomRank() < Rank) {
+ Group = DL.get()->getAtomGroup();
+ Rank = DL.get()->getAtomRank();
+ }
+
+ // Update the function-local watermark so we don't reuse this number for
+ // another atom.
+ KeyInstructionsInfo.HighestEmittedAtom =
+ std::max(Group, KeyInstructionsInfo.HighestEmittedAtom);
+
+ // Apply the new DILocation to the instruction.
+ llvm::DILocation *NewDL = llvm::DILocation::get(
+ I->getContext(), DL.getLine(), DL.getCol(), DL.getScope(),
+ DL.getInlinedAt(), DL.isImplicitCode(), Group, Rank);
+ I->setDebugLoc(NewDL);
+};
+
+void CGDebugInfo::addInstToCurrentSourceAtom(llvm::Instruction *KeyInstruction,
+ llvm::Value *Backup) {
+ if (!CGM.getCodeGenOpts().DebugKeyInstructions)
+ return;
+
+ uint64_t Group = KeyInstructionsInfo.CurrentAtom;
+ if (!Group)
+ return;
+
+ addInstSourceAtomMetadata(KeyInstruction, Group, /*Rank=*/1);
+
+ llvm::Instruction *BackupI =
+ llvm::dyn_cast_or_null<llvm::Instruction>(Backup);
+ if (!BackupI)
+ return;
+
+ // Add the backup instruction to the group.
+ addInstSourceAtomMetadata(BackupI, Group, /*Rank=*/2);
+
+ // Look through chains of casts too, as they're probably going to evaporate.
+ // FIXME: And other nops like zero length geps?
+ // FIXME: Should use Cast->isNoopCast()?
+ uint8_t Rank = 3;
+ while (auto *Cast = dyn_cast<llvm::CastInst>(BackupI)) {
+ BackupI = dyn_cast<llvm::Instruction>(Cast->getOperand(0));
+ if (!BackupI)
+ break;
+ addInstSourceAtomMetadata(BackupI, Group, Rank++);
+ }
+}
+
+void CGDebugInfo::addRetToOverrideOrNewSourceAtom(llvm::ReturnInst *Ret,
+ llvm::Value *Backup) {
+ if (KeyInstructionsInfo.RetAtomOverride) {
+ uint64_t CurrentAtom = KeyInstructionsInfo.CurrentAtom;
+ KeyInstructionsInfo.CurrentAtom = KeyInstructionsInfo.RetAtomOverride;
+ addInstToCurrentSourceAtom(Ret, Backup);
+ KeyInstructionsInfo.CurrentAtom = CurrentAtom;
+ KeyInstructionsInfo.RetAtomOverride = 0;
+ } else {
+ auto Grp = ApplyAtomGroup(this);
+ addInstToCurrentSourceAtom(Ret, Backup);
+ }
+}
+
+void CGDebugInfo::setRetInstSourceAtomOverride(uint64_t Group) {
+ assert(KeyInstructionsInfo.RetAtomOverride == 0);
+ KeyInstructionsInfo.RetAtomOverride = Group;
+}
+
+void CGDebugInfo::completeFunction() {
+ // Reset the atom group number tracker as the numbers are function-local.
+ KeyInstructionsInfo.NextAtom = 1;
+ KeyInstructionsInfo.HighestEmittedAtom = 0;
+ KeyInstructionsInfo.CurrentAtom = 0;
+ KeyInstructionsInfo.RetAtomOverride = 0;
+}
+
+ApplyAtomGroup::ApplyAtomGroup(CGDebugInfo *DI) : DI(DI) {
+ if (!DI)
+ return;
+ OriginalAtom = DI->KeyInstructionsInfo.CurrentAtom;
+ DI->KeyInstructionsInfo.CurrentAtom = DI->KeyInstructionsInfo.NextAtom++;
+}
+
+ApplyAtomGroup::~ApplyAtomGroup() {
+ if (!DI)
+ return;
+
+ // We may not have used the group number at all.
+ DI->KeyInstructionsInfo.NextAtom =
+ std::min(DI->KeyInstructionsInfo.HighestEmittedAtom + 1,
+ DI->KeyInstructionsInfo.NextAtom);
+
+ DI->KeyInstructionsInfo.CurrentAtom = OriginalAtom;
+}
+
ApplyDebugLocation::ApplyDebugLocation(CodeGenFunction &CGF,
SourceLocation TemporaryLocation)
: CGF(&CGF) {
@@ -174,8 +284,15 @@ ApplyDebugLocation::ApplyDebugLocation(CodeGenFunction &CGF, llvm::DebugLoc Loc)
return;
}
OriginalLocation = CGF.Builder.getCurrentDebugLocation();
- if (Loc)
+ if (Loc) {
+ // Key Instructions: drop the atom group and rank to avoid accidentally
+ // propagating it around.
+ if (Loc->getAtomGroup())
+ Loc = llvm::DILocation::get(Loc->getContext(), Loc.getLine(),
+ Loc->getColumn(), Loc->getScope(),
+ Loc->getInlinedAt(), Loc.isImplicitCode());
CGF.Builder.SetCurrentDebugLocation(std::move(Loc));
+ }
}
ApplyDebugLocation::~ApplyDebugLocation() {
diff --git a/clang/lib/CodeGen/CGDebugInfo.h b/clang/lib/CodeGen/CGDebugInfo.h
index 9db5113fe5d8e..e948298d517da 100644
--- a/clang/lib/CodeGen/CGDebugInfo.h
+++ b/clang/lib/CodeGen/CGDebugInfo.h
@@ -58,6 +58,8 @@ class CGBlockInfo;
class CGDebugInfo {
friend class ApplyDebugLocation;
friend class SaveAndRestoreLocation;
+ friend class ApplyAtomGroup;
+
CodeGenModule &CGM;
const llvm::codegenoptions::DebugInfoKind DebugKind;
bool DebugTypeExtRefs;
@@ -179,6 +181,17 @@ class CGDebugInfo {
/// The key is coroutine real parameters, value is DIVariable in LLVM IR.
Param2DILocTy ParamDbgMappings;
+ /// Key Instructions bookkeeping.
+ /// Source atoms are identified by a {AtomGroup, InlinedAt} pair, meaning
+ /// AtomGroup numbers can be repeated across different functions.
+ struct {
+ uint64_t NextAtom = 1;
+ uint64_t HighestEmittedAtom = 0;
+ uint64_t CurrentAtom = 0;
+ uint64_t RetAtomOverride = 0;
+ } KeyInstructionsInfo;
+
+private:
/// Helper functions for getOrCreateType.
/// @{
/// Currently the checksum of an interface includes the number of
@@ -636,7 +649,30 @@ class CGDebugInfo {
StringRef Category,
StringRef FailureMsg);
+ /// Reset internal state.
+ void completeFunction();
+
+ /// Add \p KeyInstruction and an optional \p Backup instruction to the
+ /// current atom group, created using ApplyAtomGroup.
+ void addInstToCurrentSourceAtom(llvm::Instruction *KeyInstruction,
+ llvm::Value *Backup);
+
+ /// Add \p Ret and an optional \p Backup instruction to the
+ /// saved override used for some ret instructions if it exists, or a new atom.
+ void addRetToOverrideOrNewSourceAtom(llvm::ReturnInst *Ret,
+ llvm::Value *Backup);
+
+ /// Set an atom group override for use in addRetToOverrideOrNewSourceAtom.
+ void setRetInstSourceAtomOverride(uint64_t Group);
+
private:
+ /// Amend \p I's DebugLoc with \p Group (its source atom group) and \p
+ /// Rank (lower nonzero rank is higher precedence). Does nothing if \p I
+ /// has no DebugLoc, and chooses the atom group in which the instruction
+ /// has the highest precedence if it's already in one.
+ void addInstSourceAtomMetadata(llvm::Instruction *I, uint64_t Group,
+ uint8_t Rank);
+
/// Emit call to llvm.dbg.declare for a variable declaration.
/// Returns a pointer to the DILocalVariable associated with the
/// llvm.dbg.declare, or nullptr otherwise.
@@ -853,6 +889,20 @@ class CGDebugInfo {
}
};
+/// A scoped helper to set the current source atom group for
+/// CGDebugInfo::addInstToCurrentSourceAtom. A source atom is a source construct
+/// that is "interesting" for debug stepping purposes. We use an atom group
+/// number to track the instruction(s) that implement the functionality for the
+/// atom, plus backup instructions/source locations.
+class ApplyAtomGroup {
+ uint64_t OriginalAtom = 0;
+ CGDebugInfo *DI = nullptr;
+
+public:
+ ApplyAtomGroup(CGDebugInfo *DI);
+ ~ApplyAtomGroup();
+};
+
/// A scoped helper to set the current debug location to the specified
/// location or preferred location of the specified Expr.
class ApplyDebugLocation {
diff --git a/clang/lib/CodeGen/CodeGenFunction.h b/clang/lib/CodeGen/CodeGenFunction.h
index ca00a0e8c6cf4..0c0019fbea5ea 100644
--- a/clang/lib/CodeGen/CodeGenFunction.h
+++ b/clang/lib/CodeGen/CodeGenFunction.h
@@ -1761,6 +1761,20 @@ class CodeGenFunction : public CodeGenTypeCache {
/// recently incremented counter.
uint64_t getCurrentProfileCount() { return PGO.getCurrentRegionCount(); }
+ /// See CGDebugInfo::addInstToCurrentSourceAtom.
+ void addInstToCurrentSourceAtom(llvm::Instruction *KeyInstruction,
+ llvm::Value *Backup) {
+ if (CGDebugInfo *DI = getDebugInfo())
+ DI->addInstToCurrentSourceAtom(KeyInstruction, Backup);
+ }
+
+ /// See CGDebugInfo::addRetToOverrideOrNewSourceAtom.
+ void addRetToOverrideOrNewSourceAtom(llvm::ReturnInst *Ret,
+ llvm::Value *Backup) {
+ if (CGDebugInfo *DI = getDebugInfo())
+ DI->addRetToOverrideOrNewSourceAtom(Ret, Backup);
+ }
+
private:
/// SwitchInsn - This is nearest current switch instruction. It is null if
/// current context is not in a switch.
|
There was a problem hiding this comment.
This all seems fine -- I guess the plumbing here has to get in without a test, before then later real changes come in and can be tested.
The "Override" vs new-source-atom distinction seems a little clunky, although I haven't read how it's used to get the full context. IMO it's worth putting thought into a better name: can we pick an abstract name describing the purpose ("New key operation"?) where the override thing is just an implementation feature?
OCHyams
force-pushed
the
users/OCHyams/ki-clang-flag
branch
from
7132dd3 to
5c9d5e0
Compare
Hmm you're right it's not very clear. The "override" situation is needed because of how returns are handled in Clang (ret atoms review - #134652), which is that multiple returns are emitted as branches to a single return-block. Those branches get the source location info for the return, and the actual ret is associated with the closing brace. However, if there's only one pred to the return block it's folded into it, and the return takes the source location of the branch. I can't remember exactly why the atom application had to be structured in this slightly convoluted way (should've written better comments!). I'll see if it can be simplified a bit, and if not I'll rename the functions and add better comments. |
OCHyams
force-pushed
the
users/OCHyams/ki-clang-scoped-setter
branch
from
856f99e to
0544aed
Compare
This is a scoped helper similar to ApplyDebugLocation that creates a new source atom group which instructions can be added to. A source atom is a source construct that is "interesting" for debug stepping purposes. We use an atom group number to track the instruction(s) that implement the functionality for the atom, plus backup instructions/source locations. --- 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
OCHyams
force-pushed
the
users/OCHyams/ki-clang-scoped-setter
branch
from
0544aed to
8e7f643
Compare
|
I jumped forward to the How does this look now? |
This is a scoped helper similar to ApplyDebugLocation that creates a new source
atom group which instructions can be added to.
A source atom is a source construct that is "interesting" for debug stepping
purposes. We use an atom group number to track the instruction(s) that implement
the functionality for the atom, plus backup instructions/source locations.
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