jhuber6 updated this revision to Diff 456441.
jhuber6 added a comment.
Updating. I realized all of the math-related ones are already covered by driver
options for AMDGPU passing the appropriate fp contract to the frontend. This
patch gets rid of most of that handling and just uses those directly. Also
makes it easier to test.
We also check if the `+wavefront64` feature was explicitly turned on as part of
@yaxunl's suggestion.
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/lib/CodeGen/CodeGenModule.cpp
clang/lib/CodeGen/TargetInfo.cpp
clang/lib/CodeGen/TargetInfo.h
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 +1,49 @@
+// Check that we generate all the expected default features for the target.
+// RUN: %clang_cc1 -x hip -triple amdgcn-amd-amdhsa -target-cpu gfx90a -S -emit-llvm -o - %s | FileCheck %s --check-prefix=GFX90A
+// RUN: %clang_cc1 -x hip -triple amdgcn-amd-amdhsa -target-cpu gfx1030 -S -emit-llvm -o - %s | FileCheck %s --check-prefix=GFX1030
+
+// GFX90A: @__oclc_wavefrontsize64 = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1
+// GFX90A: @__oclc_daz_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0
+// GFX90A: @__oclc_finite_only_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0
+// GFX90A: @__oclc_unsafe_math_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0
+// GFX90A: @__oclc_correctly_rounded_sqrt32 = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1
+// GFX90A: @__oclc_ISA_version = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i32 9010
+// GFX90A: @__oclc_ABI_version = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i32 400
+
+// GFX1030: @__oclc_wavefrontsize64 = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0
+// GFX1030: @__oclc_daz_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0
+// GFX1030: @__oclc_finite_only_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0
+// GFX1030: @__oclc_unsafe_math_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0
+// GFX1030: @__oclc_correctly_rounded_sqrt32 = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1
+// GFX1030: @__oclc_ISA_version = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i32 10048
+// GFX1030: @__oclc_ABI_version = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i32 400
+
+// Check that we can override the wavefront features.
+// RUN: %clang_cc1 -x hip -triple amdgcn-amd-amdhsa -target-cpu gfx1030 -target-feature +wavefrontsize64 \
+// RUN: -S -emit-llvm -o - %s | FileCheck %s --check-prefix=WAVEFRONT
+// WAVEFRONT: @__oclc_wavefrontsize64 = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1
+
+// Check that we can enable denormalization at zero.
+// RUN: %clang_cc1 -x hip -triple amdgcn-amd-amdhsa -target-cpu gfx90a -fdenormal-fp-math-f32=preserve-sign,preserve-sign \
+// RUN: -S -emit-llvm -o - %s | FileCheck %s --check-prefix=DENORM-AT-ZERO
+// DENORM-AT-ZERO: @__oclc_daz_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1
+
+// Check that we can enable finite math.
+// RUN: %clang_cc1 -x hip -triple amdgcn-amd-amdhsa -target-cpu gfx90a -ffinite-math-only \
+// RUN: -S -emit-llvm -o - %s | FileCheck %s --check-prefix=FINITE-MATH
+// FINITE-MATH: @__oclc_finite_only_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1
+// FINITE-MATH: @__oclc_unsafe_math_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0
+
+// Check that we can enable unsafe math.
+// RUN: %clang_cc1 -x hip -triple amdgcn-amd-amdhsa -target-cpu gfx90a -menable-unsafe-fp-math \
+// RUN: -S -emit-llvm -o - %s | FileCheck %s --check-prefix=UNSAFE-MATH
+// UNSAFE-MATH: @__oclc_finite_only_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0
+// UNSAFE-MATH: @__oclc_unsafe_math_opt = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1
+
+// Check that we can disable correctly rounded square roots.
+// RUN: %clang_cc1 -x hip -triple amdgcn-amd-amdhsa -target-cpu gfx90a -fno-hip-fp32-correctly-rounded-divide-sqrt \
+// RUN: -S -emit-llvm -o - %s | FileCheck %s --check-prefix=CORRECT-SQRT
+// CORRECT-SQRT: @__oclc_correctly_rounded_sqrt32 = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 0
+// RUN: %clang_cc1 -x cl -triple amdgcn-amd-amdhsa -target-cpu gfx90a -cl-fp32-correctly-rounded-divide-sqrt \
+// RUN: -disable-llvm-optzns -S -emit-llvm -o - %s | FileCheck %s --check-prefix=CL-CORRECT-SQRT
+// CL-CORRECT-SQRT: @__oclc_correctly_rounded_sqrt32 = linkonce_odr hidden local_unnamed_addr addrspace(4) constant i8 1
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 +9406,75 @@
}
}
+/// 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));
+ };
+
+ // The wavefront size is 64 if defined by the target or explicitly specified
+ // by the user.
+ bool Wavefront64 =
+ !(Features & llvm::AMDGPU::FEATURE_WAVE32) ||
+ llvm::is_contained(CGM.getTarget().getTargetOpts().FeaturesAsWritten,
+ "+wavefrontsize64");
+
+ // Different math flags set by the current floating point contract.
+ bool RelaxedMath = CGM.getLangOpts().FastMath;
+ bool UnsafeMath = CGM.getLangOpts().UnsafeFPMath;
+ bool DenormAtZero = CGM.getCodeGenOpts().FP32DenormalMode ==
+ llvm::DenormalMode::getPreserveSign();
+ bool FiniteOnly =
+ CGM.getLangOpts().NoHonorInfs || CGM.getLangOpts().NoHonorNaNs;
+
+ // Set correct square root rounding depending on the target lanauge.
+ bool CorrectSqrt = CGM.getLangOpts().OpenCL
+ ? CGM.getCodeGenOpts().OpenCLCorrectlyRoundedDivSqrt
+ : CGM.getCodeGenOpts().HIPCorrectlyRoundedDivSqrt;
+
+ // Control constants for math operations.
+ AddGlobal("__oclc_wavefrontsize64", Wavefront64, /*Size=*/8);
+ AddGlobal("__oclc_daz_opt", DenormAtZero, /*Size=*/8);
+ AddGlobal("__oclc_finite_only_opt", FiniteOnly || RelaxedMath, /*Size=*/8);
+ AddGlobal("__oclc_unsafe_math_opt", UnsafeMath || RelaxedMath, /*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
@@ -951,6 +951,7 @@
if (getCodeGenOpts().SkipRaxSetup)
getModule().addModuleFlag(llvm::Module::Override, "SkipRaxSetup", 1);
+ getTargetCodeGenInfo().emitTargetGlobals(*this);
getTargetCodeGenInfo().emitTargetMetadata(*this, MangledDeclNames);
EmitBackendOptionsMetadata(getCodeGenOpts());
Index: clang/include/clang/Basic/CodeGenOptions.def
===================================================================
--- clang/include/clang/Basic/CodeGenOptions.def
+++ clang/include/clang/Basic/CodeGenOptions.def
@@ -192,6 +192,7 @@
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
