| 1 |
//===- AMDGPUOpenMP.cpp - AMDGPUOpenMP ToolChain Implementation -*- C++ -*-===// |
| 2 |
// |
| 3 |
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 |
// See https://llvm.org/LICENSE.txt for license information. |
| 5 |
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 |
// |
| 7 |
//===----------------------------------------------------------------------===// |
| 8 |
|
| 9 |
#include "AMDGPUOpenMP.h" |
| 10 |
#include "AMDGPU.h" |
| 11 |
#include "CommonArgs.h" |
| 12 |
#include "ToolChains/ROCm.h" |
| 13 |
#include "clang/Basic/DiagnosticDriver.h" |
| 14 |
#include "clang/Driver/Compilation.h" |
| 15 |
#include "clang/Driver/Driver.h" |
| 16 |
#include "clang/Driver/DriverDiagnostic.h" |
| 17 |
#include "clang/Driver/InputInfo.h" |
| 18 |
#include "clang/Driver/Options.h" |
| 19 |
#include "clang/Driver/Tool.h" |
| 20 |
#include "llvm/ADT/STLExtras.h" |
| 21 |
#include "llvm/Support/FileSystem.h" |
| 22 |
#include "llvm/Support/FormatAdapters.h" |
| 23 |
#include "llvm/Support/FormatVariadic.h" |
| 24 |
#include "llvm/Support/Path.h" |
| 25 |
|
| 26 |
using namespace clang::driver; |
| 27 |
using namespace clang::driver::toolchains; |
| 28 |
using namespace clang::driver::tools; |
| 29 |
using namespace clang; |
| 30 |
using namespace llvm::opt; |
| 31 |
|
| 32 |
AMDGPUOpenMPToolChain::AMDGPUOpenMPToolChain(const Driver &D, |
| 33 |
const llvm::Triple &Triple, |
| 34 |
const ToolChain &HostTC, |
| 35 |
const ArgList &Args) |
| 36 |
: ROCMToolChain(D, Triple, Args), HostTC(HostTC) { |
| 37 |
// Lookup binaries into the driver directory, this is used to |
| 38 |
// discover the clang-offload-bundler executable. |
| 39 |
getProgramPaths().push_back(getDriver().Dir); |
| 40 |
} |
| 41 |
|
| 42 |
void AMDGPUOpenMPToolChain::addClangTargetOptions( |
| 43 |
const llvm::opt::ArgList &DriverArgs, llvm::opt::ArgStringList &CC1Args, |
| 44 |
Action::OffloadKind DeviceOffloadingKind) const { |
| 45 |
HostTC.addClangTargetOptions(DriverArgs, CC1Args, DeviceOffloadingKind); |
| 46 |
|
| 47 |
StringRef GPUArch = DriverArgs.getLastArgValue(options::OPT_march_EQ); |
| 48 |
assert(!GPUArch.empty() && "Must have an explicit GPU arch."); |
| 49 |
|
| 50 |
assert(DeviceOffloadingKind == Action::OFK_OpenMP && |
| 51 |
"Only OpenMP offloading kinds are supported."); |
| 52 |
|
| 53 |
CC1Args.push_back("-target-cpu"); |
| 54 |
CC1Args.push_back(DriverArgs.MakeArgStringRef(GPUArch)); |
| 55 |
CC1Args.push_back("-fcuda-is-device"); |
| 56 |
|
| 57 |
if (DriverArgs.hasArg(options::OPT_nogpulib)) |
| 58 |
return; |
| 59 |
|
| 60 |
for (auto BCFile : getDeviceLibs(DriverArgs)) { |
| 61 |
CC1Args.push_back(BCFile.ShouldInternalize ? "-mlink-builtin-bitcode" |
| 62 |
: "-mlink-bitcode-file"); |
| 63 |
CC1Args.push_back(DriverArgs.MakeArgString(BCFile.Path)); |
| 64 |
} |
| 65 |
|
| 66 |
// Link the bitcode library late if we're using device LTO. |
| 67 |
if (getDriver().isUsingLTO(/* IsOffload */ true)) |
| 68 |
return; |
| 69 |
} |
| 70 |
|
| 71 |
llvm::opt::DerivedArgList *AMDGPUOpenMPToolChain::TranslateArgs( |
| 72 |
const llvm::opt::DerivedArgList &Args, StringRef BoundArch, |
| 73 |
Action::OffloadKind DeviceOffloadKind) const { |
| 74 |
DerivedArgList *DAL = |
| 75 |
HostTC.TranslateArgs(Args, BoundArch, DeviceOffloadKind); |
| 76 |
if (!DAL) |
| 77 |
DAL = new DerivedArgList(Args.getBaseArgs()); |
| 78 |
|
| 79 |
const OptTable &Opts = getDriver().getOpts(); |
| 80 |
|
| 81 |
if (DeviceOffloadKind == Action::OFK_OpenMP) { |
| 82 |
for (Arg *A : Args) |
| 83 |
if (!llvm::is_contained(*DAL, A)) |
| 84 |
DAL->append(A); |
| 85 |
|
| 86 |
if (!DAL->hasArg(options::OPT_march_EQ)) { |
| 87 |
StringRef Arch = BoundArch; |
| 88 |
if (Arch.empty()) { |
| 89 |
auto ArchsOrErr = getSystemGPUArchs(Args); |
| 90 |
if (!ArchsOrErr) { |
| 91 |
std::string ErrMsg = |
| 92 |
llvm::formatv("{0}", llvm::fmt_consume(ArchsOrErr.takeError())); |
| 93 |
getDriver().Diag(diag::err_drv_undetermined_gpu_arch) |
| 94 |
<< llvm::Triple::getArchTypeName(getArch()) << ErrMsg << "-march"; |
| 95 |
Arch = CudaArchToString(CudaArch::HIPDefault); |
| 96 |
} else { |
| 97 |
Arch = Args.MakeArgString(ArchsOrErr->front()); |
| 98 |
} |
| 99 |
} |
| 100 |
DAL->AddJoinedArg(nullptr, Opts.getOption(options::OPT_march_EQ), Arch); |
| 101 |
} |
| 102 |
|
| 103 |
return DAL; |
| 104 |
} |
| 105 |
|
| 106 |
for (Arg *A : Args) { |
| 107 |
DAL->append(A); |
| 108 |
} |
| 109 |
|
| 110 |
if (!BoundArch.empty()) { |
| 111 |
DAL->eraseArg(options::OPT_march_EQ); |
| 112 |
DAL->AddJoinedArg(nullptr, Opts.getOption(options::OPT_march_EQ), |
| 113 |
BoundArch); |
| 114 |
} |
| 115 |
|
| 116 |
return DAL; |
| 117 |
} |
| 118 |
|
| 119 |
void AMDGPUOpenMPToolChain::addClangWarningOptions( |
| 120 |
ArgStringList &CC1Args) const { |
| 121 |
HostTC.addClangWarningOptions(CC1Args); |
| 122 |
} |
| 123 |
|
| 124 |
ToolChain::CXXStdlibType |
| 125 |
AMDGPUOpenMPToolChain::GetCXXStdlibType(const ArgList &Args) const { |
| 126 |
return HostTC.GetCXXStdlibType(Args); |
| 127 |
} |
| 128 |
|
| 129 |
void AMDGPUOpenMPToolChain::AddClangSystemIncludeArgs( |
| 130 |
const ArgList &DriverArgs, ArgStringList &CC1Args) const { |
| 131 |
HostTC.AddClangSystemIncludeArgs(DriverArgs, CC1Args); |
| 132 |
} |
| 133 |
|
| 134 |
void AMDGPUOpenMPToolChain::AddIAMCUIncludeArgs(const ArgList &Args, |
| 135 |
ArgStringList &CC1Args) const { |
| 136 |
HostTC.AddIAMCUIncludeArgs(Args, CC1Args); |
| 137 |
} |
| 138 |
|
| 139 |
SanitizerMask AMDGPUOpenMPToolChain::getSupportedSanitizers() const { |
| 140 |
// The AMDGPUOpenMPToolChain only supports sanitizers in the sense that it |
| 141 |
// allows sanitizer arguments on the command line if they are supported by the |
| 142 |
// host toolchain. The AMDGPUOpenMPToolChain will actually ignore any command |
| 143 |
// line arguments for any of these "supported" sanitizers. That means that no |
| 144 |
// sanitization of device code is actually supported at this time. |
| 145 |
// |
| 146 |
// This behavior is necessary because the host and device toolchains |
| 147 |
// invocations often share the command line, so the device toolchain must |
| 148 |
// tolerate flags meant only for the host toolchain. |
| 149 |
return HostTC.getSupportedSanitizers(); |
| 150 |
} |
| 151 |
|
| 152 |
VersionTuple |
| 153 |
AMDGPUOpenMPToolChain::computeMSVCVersion(const Driver *D, |
| 154 |
const ArgList &Args) const { |
| 155 |
return HostTC.computeMSVCVersion(D, Args); |
| 156 |
} |
| 157 |
|
| 158 |
llvm::SmallVector<ToolChain::BitCodeLibraryInfo, 12> |
| 159 |
AMDGPUOpenMPToolChain::getDeviceLibs(const llvm::opt::ArgList &Args) const { |
| 160 |
if (Args.hasArg(options::OPT_nogpulib)) |
| 161 |
return {}; |
| 162 |
|
| 163 |
if (!RocmInstallation.hasDeviceLibrary()) { |
| 164 |
getDriver().Diag(diag::err_drv_no_rocm_device_lib) << 0; |
| 165 |
return {}; |
| 166 |
} |
| 167 |
|
| 168 |
StringRef GpuArch = getProcessorFromTargetID( |
| 169 |
getTriple(), Args.getLastArgValue(options::OPT_march_EQ)); |
| 170 |
|
| 171 |
SmallVector<BitCodeLibraryInfo, 12> BCLibs; |
| 172 |
for (auto BCLib : getCommonDeviceLibNames(Args, GpuArch.str(), |
| 173 |
/*IsOpenMP=*/true)) |
| 174 |
BCLibs.emplace_back(BCLib); |
| 175 |
|
| 176 |
return BCLibs; |
| 177 |
} |