jhuber6 updated this revision to Diff 453021.
jhuber6 added a comment.
Adjusting, adding code generation options for the other constants and changing
to use linkonce ODR linkage.
I attempted to follow Jon's suggestion and group it with the existing code. but
all the existing handling for this occurs in the driver. So I don't think
there's a convenient way to drop in this functionality without adding a new
function as in this patch.
Repository:
rG LLVM Github Monorepo
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D130096/new/
https://reviews.llvm.org/D130096
Files:
clang/include/clang/Basic/CodeGenOptions.def
clang/include/clang/Basic/CodeGenOptions.h
clang/include/clang/Driver/Options.td
clang/lib/CodeGen/CodeGenModule.cpp
clang/lib/CodeGen/TargetInfo.cpp
clang/lib/CodeGen/TargetInfo.h
clang/lib/Frontend/CompilerInvocation.cpp
clang/test/CodeGen/amdgcn-control-constants.c
Index: clang/test/CodeGen/amdgcn-control-constants.c
===================================================================
--- /dev/null
+++ clang/test/CodeGen/amdgcn-control-constants.c
@@ -0,0 +1,54 @@
+// RUN: %clang_cc1 -x c -triple amdgcn-amd-amdhsa -target-cpu gfx90a -S -emit-llvm -o - %s | FileCheck %s --check-prefix=GFX90A
+// RUN: %clang_cc1 -x c -triple amdgcn-amd-amdhsa -target-cpu gfx1030 -S -emit-llvm -o - %s | FileCheck %s --check-prefix=GFX1030
+// RUN: %clang_cc1 -x c -triple amdgcn-amd-amdhsa -target-cpu gfx908 -ffast-math -S -emit-llvm -o - %s | FileCheck %s --check-prefix=FAST
+// RUN: %clang_cc1 -x c -triple amdgcn-amd-amdhsa -target-cpu gfx908 -ffinite-math-only -S -emit-llvm -o - %s | FileCheck %s --check-prefix=FINITE
+// RUN: %clang_cc1 -x c -triple amdgcn-amd-amdhsa -target-cpu gfx703 -fgpu-flush-denormals-to-zero -S -emit-llvm -o - %s | FileCheck %s --check-prefix=DAZ
+// RUN: %clang_cc1 -x c -triple amdgcn-amd-amdhsa -target-cpu gfx908 -funsafe-math-optimizations -S -emit-llvm -o - %s | FileCheck %s --check-prefix=UNSAFE-MATH
+
+// GFX90A: @__oclc_daz_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0, align 1
+// GFX90A: @__oclc_wavefrontsize64 = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1, align 1
+// GFX90A: @__oclc_finite_only_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0, align 1
+// GFX90A: @__oclc_unsafe_math_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0, align 1
+// GFX90A: @__oclc_correctly_rounded_sqrt32 = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1, align 1
+// GFX90A: @__oclc_ISA_version = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i32 9010, align 4
+// GFX90A: @__oclc_ABI_version = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i32 400, align 4
+
+// GFX1030: @__oclc_daz_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0, align 1
+// GFX1030: @__oclc_wavefrontsize64 = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0, align 1
+// GFX1030: @__oclc_finite_only_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0, align 1
+// GFX1030: @__oclc_unsafe_math_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0, align 1
+// GFX1030: @__oclc_correctly_rounded_sqrt32 = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1, align 1
+// GFX1030: @__oclc_ISA_version = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i32 10048, align 4
+// GFX1030: @__oclc_ABI_version = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i32 400, align 4
+
+// FAST: @__oclc_daz_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0, align 1
+// FAST: @__oclc_wavefrontsize64 = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1, align 1
+// FAST: @__oclc_finite_only_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1, align 1
+// FAST: @__oclc_unsafe_math_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1, align 1
+// FAST: @__oclc_correctly_rounded_sqrt32 = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1, align 1
+// FAST: @__oclc_ISA_version = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i32 9008, align 4
+// FAST: @__oclc_ABI_version = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i32 400, align 4
+
+// FINITE: @__oclc_daz_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0, align 1
+// FINITE: @__oclc_wavefrontsize64 = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1, align 1
+// FINITE: @__oclc_finite_only_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1, align 1
+// FINITE: @__oclc_unsafe_math_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0, align 1
+// FINITE: @__oclc_correctly_rounded_sqrt32 = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1, align 1
+// FINITE: @__oclc_ISA_version = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i32 9008, align 4
+// FINITE: @__oclc_ABI_version = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i32 400, align 4
+
+// DAZ: @__oclc_daz_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1, align 1
+// DAZ: @__oclc_wavefrontsize64 = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1, align 1
+// DAZ: @__oclc_finite_only_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0, align 1
+// DAZ: @__oclc_unsafe_math_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0, align 1
+// DAZ: @__oclc_correctly_rounded_sqrt32 = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1, align 1
+// DAZ: @__oclc_ISA_version = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i32 7003, align 4
+// DAZ: @__oclc_ABI_version = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i32 400, align 4
+
+// UNSAFE-MATH: @__oclc_daz_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0, align 1
+// UNSAFE-MATH: @__oclc_wavefrontsize64 = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1, align 1
+// UNSAFE-MATH: @__oclc_finite_only_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0, align 1
+// UNSAFE-MATH: @__oclc_unsafe_math_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0, align 1
+// UNSAFE-MATH: @__oclc_correctly_rounded_sqrt32 = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1, align 1
+// UNSAFE-MATH: @__oclc_ISA_version = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i32 9008, align 4
+// UNSAFE-MATH: @__oclc_ABI_version = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i32 400, align 4
Index: clang/lib/Frontend/CompilerInvocation.cpp
===================================================================
--- clang/lib/Frontend/CompilerInvocation.cpp
+++ clang/lib/Frontend/CompilerInvocation.cpp
@@ -1562,6 +1562,12 @@
if (!Opts.EmitVersionIdentMetadata)
GenerateArg(Args, OPT_Qn, SA);
+ if (Opts.AMDGPUDenormAtZero)
+ GenerateArg(Args,
+ *Opts.AMDGPUDenormAtZero ? OPT_fgpu_flush_denormals_to_zero
+ : OPT_fno_gpu_flush_denormals_to_zero,
+ SA);
+
switch (Opts.FiniteLoops) {
case CodeGenOptions::FiniteLoopsKind::Language:
break;
@@ -1668,6 +1674,13 @@
Opts.setDebugInfo(codegenoptions::LimitedDebugInfo);
}
+ // Forward the flag value to override the default target behaviour on AMDGPU
+ // targets.
+ if (Args.hasArg(OPT_fgpu_flush_denormals_to_zero))
+ Opts.AMDGPUDenormAtZero = true;
+ else if (Args.hasArg(OPT_fno_gpu_flush_denormals_to_zero))
+ Opts.AMDGPUDenormAtZero = false;
+
for (const auto &Arg : Args.getAllArgValues(OPT_fdebug_prefix_map_EQ)) {
auto Split = StringRef(Arg).split('=');
Opts.DebugPrefixMap.insert(
Index: clang/lib/CodeGen/TargetInfo.h
===================================================================
--- clang/lib/CodeGen/TargetInfo.h
+++ clang/lib/CodeGen/TargetInfo.h
@@ -76,6 +76,9 @@
CodeGen::CodeGenModule &CGM,
const llvm::MapVector<GlobalDecl, StringRef> &MangledDeclNames) const {}
+ /// Provides a convenient hook to handle extra target-specific globals.
+ virtual void emitTargetGlobals(CodeGen::CodeGenModule &CGM) const {}
+
/// Any further codegen related checks that need to be done on a function call
/// in a target specific manner.
virtual void checkFunctionCallABI(CodeGenModule &CGM, SourceLocation CallLoc,
Index: clang/lib/CodeGen/TargetInfo.cpp
===================================================================
--- clang/lib/CodeGen/TargetInfo.cpp
+++ clang/lib/CodeGen/TargetInfo.cpp
@@ -33,6 +33,7 @@
#include "llvm/IR/IntrinsicsS390.h"
#include "llvm/IR/Type.h"
#include "llvm/Support/MathExtras.h"
+#include "llvm/Support/TargetParser.h"
#include "llvm/Support/raw_ostream.h"
#include <algorithm>
@@ -9288,6 +9289,8 @@
void setFunctionDeclAttributes(const FunctionDecl *FD, llvm::Function *F,
CodeGenModule &CGM) const;
+ void emitTargetGlobals(CodeGen::CodeGenModule &CGM) const override;
+
void setTargetAttributes(const Decl *D, llvm::GlobalValue *GV,
CodeGen::CodeGenModule &M) const override;
unsigned getOpenCLKernelCallingConv() const override;
@@ -9403,6 +9406,69 @@
}
}
+/// Emits control constants used to change per-architecture behaviour in the
+/// AMDGPU ROCm device libraries.
+void AMDGPUTargetCodeGenInfo::emitTargetGlobals(
+ CodeGen::CodeGenModule &CGM) const {
+ if (!CGM.getTriple().isAMDGCN())
+ return;
+ StringRef CPU = CGM.getTarget().getTargetOpts().CPU;
+ llvm::AMDGPU::GPUKind Kind = llvm::AMDGPU::parseArchAMDGCN(CPU);
+ unsigned Features = llvm::AMDGPU::getArchAttrAMDGCN(Kind);
+ if (Kind == llvm::AMDGPU::GK_NONE)
+ return;
+
+ unsigned Minor;
+ unsigned Major;
+ StringRef Identifier = CPU.drop_while([](char C) { return !isDigit(C); });
+ if (Identifier.take_back(2).getAsInteger(16, Minor) ||
+ Identifier.drop_back(2).getAsInteger(10, Major))
+ return;
+
+ auto AddGlobal = [&](StringRef Name, unsigned Value, unsigned Size) {
+ if (CGM.getModule().getNamedGlobal(Name))
+ return;
+
+ auto *Type =
+ llvm::IntegerType::getIntNTy(CGM.getModule().getContext(), Size);
+ auto *GV = new llvm::GlobalVariable(
+ CGM.getModule(), Type, true,
+ llvm::GlobalValue::LinkageTypes::LinkOnceODRLinkage,
+ llvm::ConstantInt::get(Type, Value), Name, nullptr,
+ llvm::GlobalValue::ThreadLocalMode::NotThreadLocal,
+ CGM.getContext().getTargetAddressSpace(LangAS::opencl_constant));
+ GV->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Local);
+ GV->setVisibility(llvm::GlobalValue::VisibilityTypes::HiddenVisibility);
+ GV->setAlignment(CGM.getDataLayout().getABITypeAlign(Type));
+ };
+
+ bool DenormAtZero =
+ CGM.getCodeGenOpts().AMDGPUDenormAtZero
+ ? *CGM.getCodeGenOpts().AMDGPUDenormAtZero
+ : !((Features & llvm::AMDGPU::FEATURE_FAST_FMA_F32) &&
+ (Features & llvm::AMDGPU::FEATURE_FAST_DENORMAL_F32));
+ bool Wavefront64 = !(Features & llvm::AMDGPU::FEATURE_WAVE32);
+ bool FastRelaxedMath = CGM.getLangOpts().FastMath;
+ bool FiniteOnly =
+ CGM.getLangOpts().NoHonorInfs || CGM.getLangOpts().NoHonorNaNs;
+ bool UnsafeMath = CGM.getCodeGenOpts().UnsafeMathOptimizations;
+ bool CorrectSqrt = CGM.getCodeGenOpts().HIPCorrectlyRoundedDivSqrt;
+
+ // Control constants for math operations.
+ AddGlobal("__oclc_daz_opt", DenormAtZero, /*Size=*/8);
+ AddGlobal("__oclc_wavefrontsize64", Wavefront64, /*Size=*/8);
+ AddGlobal("__oclc_finite_only_opt", FiniteOnly || FastRelaxedMath,
+ /*Size=*/8);
+ AddGlobal("__oclc_unsafe_math_opt", UnsafeMath || FastRelaxedMath,
+ /*Size=*/8);
+ AddGlobal("__oclc_correctly_rounded_sqrt32", CorrectSqrt, /*Size=*/8);
+
+ // Control constants for the system.
+ AddGlobal("__oclc_ISA_version", Minor + Major * 1000, /*Size=*/32);
+ AddGlobal("__oclc_ABI_version",
+ CGM.getTarget().getTargetOpts().CodeObjectVersion, /*Size=*/32);
+}
+
void AMDGPUTargetCodeGenInfo::setTargetAttributes(
const Decl *D, llvm::GlobalValue *GV, CodeGen::CodeGenModule &M) const {
if (requiresAMDGPUProtectedVisibility(D, GV)) {
Index: clang/lib/CodeGen/CodeGenModule.cpp
===================================================================
--- clang/lib/CodeGen/CodeGenModule.cpp
+++ clang/lib/CodeGen/CodeGenModule.cpp
@@ -936,6 +936,7 @@
if (getCodeGenOpts().SkipRaxSetup)
getModule().addModuleFlag(llvm::Module::Override, "SkipRaxSetup", 1);
+ getTargetCodeGenInfo().emitTargetGlobals(*this);
getTargetCodeGenInfo().emitTargetMetadata(*this, MangledDeclNames);
EmitBackendOptionsMetadata(getCodeGenOpts());
Index: clang/include/clang/Driver/Options.td
===================================================================
--- clang/include/clang/Driver/Options.td
+++ clang/include/clang/Driver/Options.td
@@ -963,9 +963,9 @@
HelpText<"Ignore environment variables to detect CUDA installation">;
def ptxas_path_EQ : Joined<["--"], "ptxas-path=">, Group<i_Group>,
HelpText<"Path to ptxas (used for compiling CUDA code)">;
-def fgpu_flush_denormals_to_zero : Flag<["-"], "fgpu-flush-denormals-to-zero">,
+def fgpu_flush_denormals_to_zero : Flag<["-"], "fgpu-flush-denormals-to-zero">, Flags<[CC1Option]>,
HelpText<"Flush denormal floating point values to zero in CUDA/HIP device mode.">;
-def fno_gpu_flush_denormals_to_zero : Flag<["-"], "fno-gpu-flush-denormals-to-zero">;
+def fno_gpu_flush_denormals_to_zero : Flag<["-"], "fno-gpu-flush-denormals-to-zero">, Flags<[CC1Option]>;
def fcuda_flush_denormals_to_zero : Flag<["-"], "fcuda-flush-denormals-to-zero">,
Alias<fgpu_flush_denormals_to_zero>;
def fno_cuda_flush_denormals_to_zero : Flag<["-"], "fno-cuda-flush-denormals-to-zero">,
@@ -1864,10 +1864,10 @@
} // end -f[no-]sanitize* flags
-def funsafe_math_optimizations : Flag<["-"], "funsafe-math-optimizations">,
- Group<f_Group>;
-def fno_unsafe_math_optimizations : Flag<["-"], "fno-unsafe-math-optimizations">,
- Group<f_Group>;
+def funsafe_math_optimizations : Flag<["-"], "funsafe-math-optimizations">, Flags<[CC1Option]>,
+ Group<f_Group>, MarshallingInfoFlag<CodeGenOpts<"UnsafeMathOptimizations">>;
+def fno_unsafe_math_optimizations : Flag<["-"], "fno-unsafe-math-optimizations">, Flags<[CC1Option]>,
+ Group<f_Group>, MarshallingInfoFlag<CodeGenOpts<"UnsafeMathOptimizations">>;
def fassociative_math : Flag<["-"], "fassociative-math">, Group<f_Group>;
def fno_associative_math : Flag<["-"], "fno-associative-math">, Group<f_Group>;
defm reciprocal_math : BoolFOption<"reciprocal-math",
Index: clang/include/clang/Basic/CodeGenOptions.h
===================================================================
--- clang/include/clang/Basic/CodeGenOptions.h
+++ clang/include/clang/Basic/CodeGenOptions.h
@@ -407,6 +407,9 @@
const char *Argv0 = nullptr;
std::vector<std::string> CommandLineArgs;
+ /// Override the system default denormalization behavior for AMDGPU.
+ Optional<bool> AMDGPUDenormAtZero;
+
/// The minimum hotness value a diagnostic needs in order to be included in
/// optimization diagnostics.
///
Index: clang/include/clang/Basic/CodeGenOptions.def
===================================================================
--- clang/include/clang/Basic/CodeGenOptions.def
+++ clang/include/clang/Basic/CodeGenOptions.def
@@ -193,6 +193,7 @@
CODEGENOPT(HIPSaveKernelArgName, 1, 0) ///< Set when -fhip-kernel-arg-name is enabled.
CODEGENOPT(UniqueInternalLinkageNames, 1, 0) ///< Internal Linkage symbols get unique names.
CODEGENOPT(SplitMachineFunctions, 1, 0) ///< Split machine functions using profile information.
+CODEGENOPT(UnsafeMathOptimizations, 1, 0) ///< Set when -funsafe-math-optimizations.
/// When false, this attempts to generate code as if the result of an
/// overflowing conversion matches the overflowing behavior of a target's native
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
