[clang] [clang] Avoid repeated per-function target setup (PR #241192)
via cfe-commits
cfe-commits at lists.llvm.org
Wed Nov 11 01:54:26 PST 2026
llvmorg-github-actions[bot] wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-clang-codegen
Author: Pengcheng Wang (wangpc-pp)
<details>
<summary>Changes</summary>
Only AArch64 consumes the per-function feature map when computing a
vscale range. Skip constructing and parsing the map for other targets,
including RISC-V.
Reuse the canonical vscale_range attribute on targets whose range does
not depend on per-function state. Keep AArch64 target features and
streaming modes on the existing dynamic path.
Three interleaved CTMark pairs reduce median compile time by 0.506%
and geomean by 0.691%.
Assisted-by: TRAE CLI (GPT-6)
---
Full diff: https://github.com/llvm/llvm-project/pull/241192.diff
3 Files Affected:
- (modified) clang/lib/CodeGen/CodeGenFunction.cpp (+23-18)
- (modified) clang/lib/CodeGen/CodeGenModule.cpp (+12)
- (modified) clang/lib/CodeGen/CodeGenModule.h (+3)
``````````diff
diff --git a/clang/lib/CodeGen/CodeGenFunction.cpp b/clang/lib/CodeGen/CodeGenFunction.cpp
index 33434526d8632..9adfcfb35b304 100644
--- a/clang/lib/CodeGen/CodeGenFunction.cpp
+++ b/clang/lib/CodeGen/CodeGenFunction.cpp
@@ -1130,25 +1130,30 @@ void CodeGenFunction::StartFunction(GlobalDecl GD, QualType RetTy,
Fn->removeFnAttr("zero-call-used-regs");
// Add vscale_range attribute if appropriate.
- llvm::StringMap<bool> FeatureMap;
- auto IsArmStreaming = TargetInfo::ArmStreamingKind::NotStreaming;
- if (FD) {
- getContext().getFunctionFeatureMap(FeatureMap, FD);
- if (const auto *T = FD->getType()->getAs<FunctionProtoType>())
- if (T->getAArch64SMEAttributes() &
- FunctionType::SME_PStateSMCompatibleMask)
- IsArmStreaming = TargetInfo::ArmStreamingKind::StreamingCompatible;
-
- if (IsArmStreamingFunction(FD, true))
- IsArmStreaming = TargetInfo::ArmStreamingKind::Streaming;
- }
- std::optional<std::pair<unsigned, unsigned>> VScaleRange =
- getContext().getTargetInfo().getVScaleRange(getLangOpts(), IsArmStreaming,
- &FeatureMap);
- if (VScaleRange) {
- CurFn->addFnAttr(llvm::Attribute::getWithVScaleRangeArgs(
- getLLVMContext(), VScaleRange->first, VScaleRange->second));
+ llvm::Attribute VScaleRangeAttr;
+ if (CGM.getTriple().isAArch64()) {
+ llvm::StringMap<bool> FeatureMap;
+ auto IsArmStreaming = TargetInfo::ArmStreamingKind::NotStreaming;
+ if (FD) {
+ getContext().getFunctionFeatureMap(FeatureMap, FD);
+ if (const auto *T = FD->getType()->getAs<FunctionProtoType>())
+ if (T->getAArch64SMEAttributes() &
+ FunctionType::SME_PStateSMCompatibleMask)
+ IsArmStreaming = TargetInfo::ArmStreamingKind::StreamingCompatible;
+
+ if (IsArmStreamingFunction(FD, true))
+ IsArmStreaming = TargetInfo::ArmStreamingKind::Streaming;
+ }
+ auto VScaleRange = getContext().getTargetInfo().getVScaleRange(
+ getLangOpts(), IsArmStreaming, FD ? &FeatureMap : nullptr);
+ if (VScaleRange)
+ VScaleRangeAttr = llvm::Attribute::getWithVScaleRangeArgs(
+ getLLVMContext(), VScaleRange->first, VScaleRange->second);
+ } else {
+ VScaleRangeAttr = CGM.getDefaultVScaleRange();
}
+ if (VScaleRangeAttr.isValid())
+ CurFn->addFnAttr(VScaleRangeAttr);
llvm::BasicBlock *EntryBB = createBasicBlock("entry", CurFn);
diff --git a/clang/lib/CodeGen/CodeGenModule.cpp b/clang/lib/CodeGen/CodeGenModule.cpp
index 3b8ec02a94e54..303e56f91cc24 100644
--- a/clang/lib/CodeGen/CodeGenModule.cpp
+++ b/clang/lib/CodeGen/CodeGenModule.cpp
@@ -722,6 +722,18 @@ CodeGenModule::CodeGenModule(ASTContext &C,
CodeGenModule::~CodeGenModule() {}
+llvm::Attribute CodeGenModule::getDefaultVScaleRange() {
+ if (!DefaultVScaleRange) {
+ auto Range = Target.getVScaleRange(
+ LangOpts, TargetInfo::ArmStreamingKind::NotStreaming);
+ DefaultVScaleRange.emplace(
+ Range ? llvm::Attribute::getWithVScaleRangeArgs(
+ getLLVMContext(), Range->first, Range->second)
+ : llvm::Attribute());
+ }
+ return *DefaultVScaleRange;
+}
+
void CodeGenModule::createObjCRuntime() {
// This is just isGNUFamily(), but we want to force implementors of
// new ABIs to decide how best to do this.
diff --git a/clang/lib/CodeGen/CodeGenModule.h b/clang/lib/CodeGen/CodeGenModule.h
index 92afa8e708e2a..0bbe76c03b2c8 100644
--- a/clang/lib/CodeGen/CodeGenModule.h
+++ b/clang/lib/CodeGen/CodeGenModule.h
@@ -366,6 +366,8 @@ class CodeGenModule : public CodeGenTypeCache {
/// An invalid attribute means that the defaults were computed but no
/// target-features attribute needs to be emitted.
std::optional<llvm::Attribute> DefaultTargetFeatures;
+ /// Canonical vscale range for targets without per-function ranges.
+ std::optional<llvm::Attribute> DefaultVScaleRange;
std::unique_ptr<CGCXXABI> ABI;
llvm::LLVMContext &VMContext;
std::string ModuleNameHash;
@@ -915,6 +917,7 @@ class CodeGenModule : public CodeGenTypeCache {
}
const TargetInfo &getTarget() const { return Target; }
const llvm::Triple &getTriple() const { return Target.getTriple(); }
+ llvm::Attribute getDefaultVScaleRange();
bool supportsCOMDAT() const;
void maybeSetTrivialComdat(const Decl &D, llvm::GlobalObject &GO);
``````````
</details>
https://github.com/llvm/llvm-project/pull/241192
More information about the cfe-commits
mailing list