[flang-commits] [flang] [flang] Set fast math related function attributes for -Ofast/-ffast-math (PR #79301)
Alex Bradbury via flang-commits
flang-commits at lists.llvm.org
Tue Jan 30 09:15:27 PST 2024
https://github.com/asb updated https://github.com/llvm/llvm-project/pull/79301
>From 6047b5a534d552a54f25bc6dc5f42b3ed5039dd4 Mon Sep 17 00:00:00 2001
From: Alex Bradbury <asb at igalia.com>
Date: Mon, 29 Jan 2024 11:25:56 +0000
Subject: [PATCH 1/3] [flang] Set fast math related function attributes for
-Ofast/-ffast-math
The implemented logic matches the logic used for Clang in emitting these
attributes. Although it's hoped that function attributes won't be needed
in the future (vs using fast math flags in individual IR instructions),
there are codegen differences currently with/without these attributes,
as can be seen in issues like #79257 or by hacking Clang to avoid
producing these attributes and observing codegen changes.
---
.../flang/Optimizer/Transforms/Passes.h | 4 ++-
.../flang/Optimizer/Transforms/Passes.td | 15 +++++++++++
flang/include/flang/Tools/CLOptions.inc | 13 ++++++---
flang/include/flang/Tools/CrossToolHelpers.h | 20 +++++++++++++-
flang/lib/Frontend/FrontendActions.cpp | 3 ++-
.../lib/Optimizer/Transforms/FunctionAttr.cpp | 27 +++++++++++++++++--
flang/test/Driver/func-attr-fast-math.f90 | 20 ++++++++++++++
7 files changed, 94 insertions(+), 8 deletions(-)
create mode 100644 flang/test/Driver/func-attr-fast-math.f90
diff --git a/flang/include/flang/Optimizer/Transforms/Passes.h b/flang/include/flang/Optimizer/Transforms/Passes.h
index 6970da8698ae8..d33623c690489 100644
--- a/flang/include/flang/Optimizer/Transforms/Passes.h
+++ b/flang/include/flang/Optimizer/Transforms/Passes.h
@@ -91,7 +91,9 @@ struct FunctionAttrTypes {
std::unique_ptr<mlir::Pass> createFunctionAttrPass();
std::unique_ptr<mlir::Pass>
-createFunctionAttrPass(FunctionAttrTypes &functionAttr);
+createFunctionAttrPass(FunctionAttrTypes &functionAttr, bool noInfsFPMath,
+ bool noNaNsFPMath, bool approxFuncFPMath,
+ bool noSignedZerosFPMath, bool unsafeFPMath);
// declarative passes
#define GEN_PASS_REGISTRATION
diff --git a/flang/include/flang/Optimizer/Transforms/Passes.td b/flang/include/flang/Optimizer/Transforms/Passes.td
index e3c45d41f04cc..397985383b226 100644
--- a/flang/include/flang/Optimizer/Transforms/Passes.td
+++ b/flang/include/flang/Optimizer/Transforms/Passes.td
@@ -366,6 +366,21 @@ def FunctionAttr : Pass<"function-attr", "mlir::func::FuncOp"> {
"mlir::LLVM::framePointerKind::FramePointerKind",
/*default=*/"mlir::LLVM::framePointerKind::FramePointerKind{}",
"frame pointer">,
+ Option<"noInfsFPMath", "no-infs-fp-math",
+ "bool", /*default=*/"false",
+ "Set the no-infs-fp-math attribute on functions in the module.">,
+ Option<"noNaNsFPMath", "no-nans-fp-math",
+ "bool", /*default=*/"false",
+ "Set the no-nans-fp-math attribute on functions in the module.">,
+ Option<"approxFuncFPMath", "approx-func-fp-math",
+ "bool", /*default=*/"false",
+ "Set the approx-func-fp-math attribute on functions in the module.">,
+ Option<"noSignedZerosFPMath", "no-signed-zeros-fp-math",
+ "bool", /*default=*/"false",
+ "Set the no-signed-zeros-fp-math attribute on functions in the module.">,
+ Option<"unsafeFPMath", "unsafe-fp-math",
+ "bool", /*default=*/"false",
+ "Set the unsafe-fp-math attribute on functions in the module.">,
];
let constructor = "::fir::createFunctionAttrPass()";
}
diff --git a/flang/include/flang/Tools/CLOptions.inc b/flang/include/flang/Tools/CLOptions.inc
index 96d3869cd0939..166ba5f58774d 100644
--- a/flang/include/flang/Tools/CLOptions.inc
+++ b/flang/include/flang/Tools/CLOptions.inc
@@ -315,15 +315,22 @@ inline void createDefaultFIRCodeGenPassPipeline(
// Add function attributes
fir::FunctionAttrTypes functionAttrs;
- if (config.FramePointerKind != llvm::FramePointerKind::None) {
+ if (config.FramePointerKind != llvm::FramePointerKind::None ||
+ config.NoInfsFPMath || config.NoNaNsFPMath || config.ApproxFuncFPMath ||
+ config.NoSignedZerosFPMath || config.UnsafeFPMath) {
if (config.FramePointerKind == llvm::FramePointerKind::NonLeaf)
functionAttrs.framePointerKind =
mlir::LLVM::framePointerKind::FramePointerKind::NonLeaf;
- else
+ else if (config.FramePointerKind == llvm::FramePointerKind::All)
functionAttrs.framePointerKind =
mlir::LLVM::framePointerKind::FramePointerKind::All;
+ else
+ functionAttrs.framePointerKind =
+ mlir::LLVM::framePointerKind::FramePointerKind::None;
- pm.addPass(fir::createFunctionAttrPass(functionAttrs));
+ pm.addPass(fir::createFunctionAttrPass(functionAttrs, config.NoInfsFPMath,
+ config.NoNaNsFPMath, config.ApproxFuncFPMath,
+ config.NoSignedZerosFPMath, config.UnsafeFPMath));
}
fir::addFIRToLLVMPass(pm, config);
diff --git a/flang/include/flang/Tools/CrossToolHelpers.h b/flang/include/flang/Tools/CrossToolHelpers.h
index b61224ff4f1b3..0f39bdb86102f 100644
--- a/flang/include/flang/Tools/CrossToolHelpers.h
+++ b/flang/include/flang/Tools/CrossToolHelpers.h
@@ -13,6 +13,7 @@
#ifndef FORTRAN_TOOLS_CROSS_TOOL_HELPERS_H
#define FORTRAN_TOOLS_CROSS_TOOL_HELPERS_H
+#include "flang/Common/MathOptionsBase.h"
#include "flang/Frontend/CodeGenOptions.h"
#include "flang/Frontend/LangOptions.h"
#include <cstdint>
@@ -28,7 +29,8 @@ struct MLIRToLLVMPassPipelineConfig {
OptLevel = level;
}
explicit MLIRToLLVMPassPipelineConfig(llvm::OptimizationLevel level,
- const Fortran::frontend::CodeGenOptions &opts) {
+ const Fortran::frontend::CodeGenOptions &opts,
+ const Fortran::common::MathOptionsBase &mathOpts) {
OptLevel = level;
StackArrays = opts.StackArrays;
Underscoring = opts.Underscoring;
@@ -36,6 +38,15 @@ struct MLIRToLLVMPassPipelineConfig {
DebugInfo = opts.getDebugInfo();
AliasAnalysis = opts.AliasAnalysis;
FramePointerKind = opts.getFramePointer();
+ // The logic for setting these attributes is intended to match the logic
+ // used in Clang.
+ NoInfsFPMath = mathOpts.getNoHonorInfs();
+ NoNaNsFPMath = mathOpts.getNoHonorNaNs();
+ ApproxFuncFPMath = mathOpts.getApproxFunc();
+ NoSignedZerosFPMath = mathOpts.getNoSignedZeros();
+ UnsafeFPMath = mathOpts.getAssociativeMath() &&
+ mathOpts.getReciprocalMath() && NoSignedZerosFPMath &&
+ ApproxFuncFPMath && mathOpts.getFPContractEnabled();
}
llvm::OptimizationLevel OptLevel; ///< optimisation level
@@ -49,6 +60,13 @@ struct MLIRToLLVMPassPipelineConfig {
llvm::FramePointerKind::None; ///< Add frame pointer to functions.
unsigned VScaleMin = 0; ///< SVE vector range minimum.
unsigned VScaleMax = 0; ///< SVE vector range maximum.
+ bool NoInfsFPMath = false; ///< Set no-infs-fp-math attribute for functions.
+ bool NoNaNsFPMath = false; ///< Set no-nans-fp-math attribute for functions.
+ bool ApproxFuncFPMath =
+ false; ///< Set approx-func-fp-math attribute for functions.
+ bool NoSignedZerosFPMath =
+ false; ///< Set no-signed-zeros-fp-math attribute for functions.
+ bool UnsafeFPMath = false; ///< Set unsafe-fp-math attribute for functions.
};
struct OffloadModuleOpts {
diff --git a/flang/lib/Frontend/FrontendActions.cpp b/flang/lib/Frontend/FrontendActions.cpp
index 4738f7ba57042..25e0fe847876d 100644
--- a/flang/lib/Frontend/FrontendActions.cpp
+++ b/flang/lib/Frontend/FrontendActions.cpp
@@ -785,6 +785,7 @@ void CodeGenAction::generateLLVMIR() {
CompilerInstance &ci = this->getInstance();
auto opts = ci.getInvocation().getCodeGenOpts();
+ auto mathOpts = ci.getInvocation().getLoweringOpts().getMathOptions();
llvm::OptimizationLevel level = mapToLevel(opts);
fir::support::loadDialects(*mlirCtx);
@@ -797,7 +798,7 @@ void CodeGenAction::generateLLVMIR() {
pm.addPass(std::make_unique<Fortran::lower::VerifierPass>());
pm.enableVerifier(/*verifyPasses=*/true);
- MLIRToLLVMPassPipelineConfig config(level, opts);
+ MLIRToLLVMPassPipelineConfig config(level, opts, mathOpts);
if (auto vsr = getVScaleRange(ci)) {
config.VScaleMin = vsr->first;
diff --git a/flang/lib/Optimizer/Transforms/FunctionAttr.cpp b/flang/lib/Optimizer/Transforms/FunctionAttr.cpp
index 55b908ba5d861..55ffc4ed43933 100644
--- a/flang/lib/Optimizer/Transforms/FunctionAttr.cpp
+++ b/flang/lib/Optimizer/Transforms/FunctionAttr.cpp
@@ -27,6 +27,11 @@ class FunctionAttrPass : public fir::impl::FunctionAttrBase<FunctionAttrPass> {
public:
FunctionAttrPass(const fir::FunctionAttrOptions &options) {
framePointerKind = options.framePointerKind;
+ noInfsFPMath = options.noInfsFPMath;
+ noNaNsFPMath = options.noNaNsFPMath;
+ approxFuncFPMath = options.approxFuncFPMath;
+ noSignedZerosFPMath = options.noSignedZerosFPMath;
+ unsafeFPMath = options.unsafeFPMath;
}
FunctionAttrPass() {}
void runOnOperation() override;
@@ -45,14 +50,32 @@ void FunctionAttrPass::runOnOperation() {
func->setAttr("frame_pointer", mlir::LLVM::FramePointerKindAttr::get(
context, framePointerKind));
+ if (noInfsFPMath)
+ func->setAttr("no_infs_fp_math", mlir::BoolAttr::get(context, true));
+ if (noNaNsFPMath)
+ func->setAttr("no_nans_fp_math", mlir::BoolAttr::get(context, true));
+ if (approxFuncFPMath)
+ func->setAttr("approx_func_fp_math", mlir::BoolAttr::get(context, true));
+ if (noSignedZerosFPMath)
+ func->setAttr("no_signed_zeros_fp_math",
+ mlir::BoolAttr::get(context, true));
+ if (unsafeFPMath)
+ func->setAttr("unsafe_fp_math", mlir::BoolAttr::get(context, true));
+
LLVM_DEBUG(llvm::dbgs() << "=== End " DEBUG_TYPE " ===\n");
}
-std::unique_ptr<mlir::Pass>
-fir::createFunctionAttrPass(fir::FunctionAttrTypes &functionAttr) {
+std::unique_ptr<mlir::Pass> fir::createFunctionAttrPass(
+ fir::FunctionAttrTypes &functionAttr, bool noInfsFPMath, bool noNaNsFPMath,
+ bool approxFuncFPMath, bool noSignedZerosFPMath, bool unsafeFPMath) {
FunctionAttrOptions opts;
// Frame pointer
opts.framePointerKind = functionAttr.framePointerKind;
+ opts.noInfsFPMath = noInfsFPMath;
+ opts.noNaNsFPMath = noNaNsFPMath;
+ opts.approxFuncFPMath = approxFuncFPMath;
+ opts.noSignedZerosFPMath = noSignedZerosFPMath;
+ opts.unsafeFPMath = unsafeFPMath;
return std::make_unique<FunctionAttrPass>(opts);
}
diff --git a/flang/test/Driver/func-attr-fast-math.f90 b/flang/test/Driver/func-attr-fast-math.f90
new file mode 100644
index 0000000000000..2f078dd1b2183
--- /dev/null
+++ b/flang/test/Driver/func-attr-fast-math.f90
@@ -0,0 +1,20 @@
+! Test that -mframe-pointer can accept only specific values and when given an invalid value, check it raises an error.
+
+! RUN: %flang -O1 -emit-llvm -S -o - %s 2>&1| FileCheck %s --check-prefix=CHECK-NOFASTMATH
+! RUN: %flang -Ofast -emit-llvm -S -o - %s 2>&1| FileCheck %s --check-prefix=CHECK-OFAST
+! RUN: %flang -O1 -ffast-math -emit-llvm -S -o - %s 2>&1| FileCheck %s --check-prefix=CHECK-FFAST-MATH
+
+subroutine func
+end subroutine func
+
+! CHECK-NOFASTMATH-LABEL: define void @func_() local_unnamed_addr
+! CHECK-NOFASTMATH-SAME: #[[ATTRS:[0-9]+]]
+! CHECK-NOFASTMATH: attributes #[[ATTRS]] = { mustprogress nofree norecurse nosync nounwind willreturn memory(none) }
+!
+! CHECK-OFAST-LABEL: define void @func_() local_unnamed_addr
+! CHECK-OFAST-SAME: #[[ATTRS:[0-9]+]]
+! CHECK-OFAST: attributes #[[ATTRS]] = { mustprogress nofree norecurse nosync nounwind willreturn memory(none) "approx-func-fp-math"="true" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "unsafe-fp-math"="true" }
+
+! CHECK-FFAST-MATH-LABEL: define void @func_() local_unnamed_addr
+! CHECK-FFAST-MATH-SAME: #[[ATTRS:[0-9]+]]
+! CHECK-FFAST-MATH: attributes #[[ATTRS]] = { mustprogress nofree norecurse nosync nounwind willreturn memory(none) "approx-func-fp-math"="true" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "unsafe-fp-math"="true" }
>From a8c899b8add4bc40b732ba72d74a7b602cb70058 Mon Sep 17 00:00:00 2001
From: Alex Bradbury <asb at igalia.com>
Date: Tue, 30 Jan 2024 15:48:00 +0000
Subject: [PATCH 2/3] Modify func-attr-fast-math.f90 test case to be resilient
to other attributes being produced (e.g. target-cpu, target-features)
---
flang/test/Driver/func-attr-fast-math.f90 | 16 ++++++++--------
1 file changed, 8 insertions(+), 8 deletions(-)
diff --git a/flang/test/Driver/func-attr-fast-math.f90 b/flang/test/Driver/func-attr-fast-math.f90
index 2f078dd1b2183..a801d4e3cb1ed 100644
--- a/flang/test/Driver/func-attr-fast-math.f90
+++ b/flang/test/Driver/func-attr-fast-math.f90
@@ -1,20 +1,20 @@
-! Test that -mframe-pointer can accept only specific values and when given an invalid value, check it raises an error.
+! REQUIRES: x86-registered-target
-! RUN: %flang -O1 -emit-llvm -S -o - %s 2>&1| FileCheck %s --check-prefix=CHECK-NOFASTMATH
-! RUN: %flang -Ofast -emit-llvm -S -o - %s 2>&1| FileCheck %s --check-prefix=CHECK-OFAST
-! RUN: %flang -O1 -ffast-math -emit-llvm -S -o - %s 2>&1| FileCheck %s --check-prefix=CHECK-FFAST-MATH
+! RUN: %flang --target=x86_64 -O1 -emit-llvm -S -o - %s 2>&1| FileCheck %s --check-prefix=CHECK-NOFASTMATH
+! RUN: %flang --target=x86_64 -Ofast -emit-llvm -S -o - %s 2>&1| FileCheck %s --check-prefix=CHECK-OFAST
+! RUN: %flang --target=x86_64 -O1 -ffast-math -emit-llvm -S -o - %s 2>&1| FileCheck %s --check-prefix=CHECK-FFAST-MATH
subroutine func
end subroutine func
! CHECK-NOFASTMATH-LABEL: define void @func_() local_unnamed_addr
! CHECK-NOFASTMATH-SAME: #[[ATTRS:[0-9]+]]
-! CHECK-NOFASTMATH: attributes #[[ATTRS]] = { mustprogress nofree norecurse nosync nounwind willreturn memory(none) }
-!
+! CHECK-NOT fp-math"=
+
! CHECK-OFAST-LABEL: define void @func_() local_unnamed_addr
! CHECK-OFAST-SAME: #[[ATTRS:[0-9]+]]
-! CHECK-OFAST: attributes #[[ATTRS]] = { mustprogress nofree norecurse nosync nounwind willreturn memory(none) "approx-func-fp-math"="true" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "unsafe-fp-math"="true" }
+! CHECK-OFAST: attributes #[[ATTRS]] = { {{.*}}"approx-func-fp-math"="true" {{.*}}"no-infs-fp-math"="true" {{.*}}"no-nans-fp-math"="true" {{.*}}"no-signed-zeros-fp-math"="true" {{.*}}"unsafe-fp-math"="true"{{.*}} }
! CHECK-FFAST-MATH-LABEL: define void @func_() local_unnamed_addr
! CHECK-FFAST-MATH-SAME: #[[ATTRS:[0-9]+]]
-! CHECK-FFAST-MATH: attributes #[[ATTRS]] = { mustprogress nofree norecurse nosync nounwind willreturn memory(none) "approx-func-fp-math"="true" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" "unsafe-fp-math"="true" }
+! CHECK-FFAST-MATH: attributes #[[ATTRS]] = { {{.*}}"approx-func-fp-math"="true" {{.*}}"no-infs-fp-math"="true" {{.*}}"no-nans-fp-math"="true" {{.*}}"no-signed-zeros-fp-math"="true" {{.*}}"unsafe-fp-math"="true"{{.*}} }
>From 02f0070c69331296d5b0b02b7e0d24bf730b3b91 Mon Sep 17 00:00:00 2001
From: Alex Bradbury <asb at igalia.com>
Date: Tue, 30 Jan 2024 17:15:09 +0000
Subject: [PATCH 3/3] func-attr-fast-math.f90 doesn't need to be x86-specific
---
flang/test/Driver/func-attr-fast-math.f90 | 8 +++-----
1 file changed, 3 insertions(+), 5 deletions(-)
diff --git a/flang/test/Driver/func-attr-fast-math.f90 b/flang/test/Driver/func-attr-fast-math.f90
index a801d4e3cb1ed..05824a6078a09 100644
--- a/flang/test/Driver/func-attr-fast-math.f90
+++ b/flang/test/Driver/func-attr-fast-math.f90
@@ -1,8 +1,6 @@
-! REQUIRES: x86-registered-target
-
-! RUN: %flang --target=x86_64 -O1 -emit-llvm -S -o - %s 2>&1| FileCheck %s --check-prefix=CHECK-NOFASTMATH
-! RUN: %flang --target=x86_64 -Ofast -emit-llvm -S -o - %s 2>&1| FileCheck %s --check-prefix=CHECK-OFAST
-! RUN: %flang --target=x86_64 -O1 -ffast-math -emit-llvm -S -o - %s 2>&1| FileCheck %s --check-prefix=CHECK-FFAST-MATH
+! RUN: %flang -O1 -emit-llvm -S -o - %s 2>&1| FileCheck %s --check-prefix=CHECK-NOFASTMATH
+! RUN: %flang -Ofast -emit-llvm -S -o - %s 2>&1| FileCheck %s --check-prefix=CHECK-OFAST
+! RUN: %flang -O1 -ffast-math -emit-llvm -S -o - %s 2>&1| FileCheck %s --check-prefix=CHECK-FFAST-MATH
subroutine func
end subroutine func
More information about the flang-commits
mailing list