Skip to content

Commit aff37b4

Browse files
author
Jakub Chlanda
committed
Do not merge, this is: llvm/llvm-project#149294
With changes resolved accordingly.
1 parent 27dab6c commit aff37b4

File tree

4 files changed

+113
-75
lines changed

4 files changed

+113
-75
lines changed

‎clang/lib/Driver/ToolChains/AMDGPU.cpp‎

Lines changed: 87 additions & 66 deletions
Original file line numberDiff line numberDiff line change
@@ -31,6 +31,72 @@ using namespace clang::driver::toolchains;
3131
using namespace clang;
3232
using namespace llvm::opt;
3333

34+
RocmInstallationDetector::CommonBitcodeLibsPreferences::
35+
CommonBitcodeLibsPreferences(const Driver &D,
36+
const llvm::opt::ArgList &DriverArgs,
37+
StringRef GPUArch,
38+
const Action::OffloadKind DeviceOffloadingKind,
39+
const bool NeedsASanRT)
40+
: ABIVer(DeviceLibABIVersion::fromCodeObjectVersion(
41+
tools::getAMDGPUCodeObjectVersion(D, DriverArgs))) {
42+
const auto Kind = llvm::AMDGPU::parseArchAMDGCN(GPUArch);
43+
const unsigned ArchAttr = llvm::AMDGPU::getArchAttrAMDGCN(Kind);
44+
45+
IsOpenMP = DeviceOffloadingKind == Action::OFK_OpenMP;
46+
47+
const bool HasWave32 = (ArchAttr & llvm::AMDGPU::FEATURE_WAVE32);
48+
Wave64 =
49+
!HasWave32 || DriverArgs.hasFlag(options::OPT_mwavefrontsize64,
50+
options::OPT_mno_wavefrontsize64, false);
51+
52+
const bool IsKnownOffloading = DeviceOffloadingKind == Action::OFK_OpenMP ||
53+
DeviceOffloadingKind == Action::OFK_HIP;
54+
55+
// Default to enabling f32 denormals on subtargets where fma is fast with
56+
// denormals
57+
const bool DefaultDAZ =
58+
(Kind == llvm::AMDGPU::GK_NONE)
59+
? false
60+
: !((ArchAttr & llvm::AMDGPU::FEATURE_FAST_FMA_F32) &&
61+
(ArchAttr & llvm::AMDGPU::FEATURE_FAST_DENORMAL_F32));
62+
// TODO: There are way too many flags that change this. Do we need to
63+
// check them all?
64+
DAZ = IsKnownOffloading
65+
? DriverArgs.hasFlag(options::OPT_fgpu_flush_denormals_to_zero,
66+
options::OPT_fno_gpu_flush_denormals_to_zero,
67+
DefaultDAZ)
68+
: DriverArgs.hasArg(options::OPT_cl_denorms_are_zero) || DefaultDAZ;
69+
70+
FiniteOnly = DriverArgs.hasArg(options::OPT_cl_finite_math_only) ||
71+
DriverArgs.hasFlag(options::OPT_ffinite_math_only,
72+
options::OPT_fno_finite_math_only, false);
73+
74+
UnsafeMathOpt =
75+
DriverArgs.hasArg(options::OPT_cl_unsafe_math_optimizations) ||
76+
DriverArgs.hasFlag(options::OPT_funsafe_math_optimizations,
77+
options::OPT_fno_unsafe_math_optimizations, false);
78+
79+
FastRelaxedMath = DriverArgs.hasArg(options::OPT_cl_fast_relaxed_math) ||
80+
DriverArgs.hasFlag(options::OPT_ffast_math,
81+
options::OPT_fno_fast_math, false);
82+
83+
const bool DefaultSqrt = IsKnownOffloading ? true : false;
84+
if (DeviceOffloadingKind == Action::OFK_SYCL)
85+
// When using SYCL, sqrt is only correctly rounded if the flag is specified.
86+
CorrectSqrt = DriverArgs.hasArg(options::OPT_foffload_fp32_prec_sqrt);
87+
else
88+
CorrectSqrt =
89+
DriverArgs.hasArg(options::OPT_cl_fp32_correctly_rounded_divide_sqrt) ||
90+
DriverArgs.hasFlag(
91+
options::OPT_fhip_fp32_correctly_rounded_divide_sqrt,
92+
options::OPT_fno_hip_fp32_correctly_rounded_divide_sqrt,
93+
DefaultSqrt);
94+
// GPU Sanitizer currently only supports ASan and is enabled through host
95+
// ASan.
96+
GPUSan = (DriverArgs.hasFlag(options::OPT_fgpu_sanitize,
97+
options::OPT_fno_gpu_sanitize, true) &&
98+
NeedsASanRT);
99+
}
34100
void RocmInstallationDetector::scanLibDevicePath(llvm::StringRef Path) {
35101
assert(!Path.empty());
36102

@@ -883,33 +949,14 @@ void ROCMToolChain::addClangTargetOptions(
883949
ABIVer))
884950
return;
885951

886-
bool Wave64 = isWave64(DriverArgs, Kind);
887-
// TODO: There are way too many flags that change this. Do we need to check
888-
// them all?
889-
bool DAZ = DriverArgs.hasArg(options::OPT_cl_denorms_are_zero) ||
890-
getDefaultDenormsAreZeroForTarget(Kind);
891-
bool FiniteOnly = DriverArgs.hasArg(options::OPT_cl_finite_math_only);
892-
893-
bool UnsafeMathOpt =
894-
DriverArgs.hasArg(options::OPT_cl_unsafe_math_optimizations);
895-
bool FastRelaxedMath = DriverArgs.hasArg(options::OPT_cl_fast_relaxed_math);
896-
bool CorrectSqrt =
897-
DriverArgs.hasArg(options::OPT_cl_fp32_correctly_rounded_divide_sqrt);
898-
899-
// GPU Sanitizer currently only supports ASan and is enabled through host
900-
// ASan.
901-
bool GPUSan = DriverArgs.hasFlag(options::OPT_fgpu_sanitize,
902-
options::OPT_fno_gpu_sanitize, true) &&
903-
getSanitizerArgs(DriverArgs).needsAsanRt();
904-
905952
// Add the OpenCL specific bitcode library.
906953
llvm::SmallVector<BitCodeLibraryInfo, 12> BCLibs;
907954
BCLibs.emplace_back(RocmInstallation->getOpenCLPath().str());
908955

909956
// Add the generic set of libraries.
910957
BCLibs.append(RocmInstallation->getCommonBitcodeLibs(
911-
DriverArgs, LibDeviceFile, Wave64, DAZ, FiniteOnly, UnsafeMathOpt,
912-
FastRelaxedMath, CorrectSqrt, ABIVer, GPUSan, false));
958+
DriverArgs, LibDeviceFile, GpuArch, DeviceOffloadingKind,
959+
getSanitizerArgs(DriverArgs).needsAsanRt()));
913960

914961
for (auto [BCFile, Internalize] : BCLibs) {
915962
if (Internalize)
@@ -946,35 +993,38 @@ bool RocmInstallationDetector::checkCommonBitcodeLibs(
946993

947994
llvm::SmallVector<ToolChain::BitCodeLibraryInfo, 12>
948995
RocmInstallationDetector::getCommonBitcodeLibs(
949-
const llvm::opt::ArgList &DriverArgs, StringRef LibDeviceFile, bool Wave64,
950-
bool DAZ, bool FiniteOnly, bool UnsafeMathOpt, bool FastRelaxedMath,
951-
bool CorrectSqrt, DeviceLibABIVersion ABIVer, bool GPUSan,
952-
bool isOpenMP) const {
996+
const llvm::opt::ArgList &DriverArgs,
997+
StringRef LibDeviceFile, StringRef GPUArch,
998+
const Action::OffloadKind DeviceOffloadingKind,
999+
const bool NeedsASanRT) const {
9531000
llvm::SmallVector<ToolChain::BitCodeLibraryInfo, 12> BCLibs;
9541001

1002+
CommonBitcodeLibsPreferences Pref{D, DriverArgs, GPUArch,
1003+
DeviceOffloadingKind, NeedsASanRT};
1004+
9551005
auto AddBCLib = [&](ToolChain::BitCodeLibraryInfo BCLib,
9561006
bool Internalize = true) {
9571007
BCLib.ShouldInternalize = Internalize;
9581008
BCLibs.emplace_back(BCLib);
9591009
};
9601010
auto AddSanBCLibs = [&]() {
961-
if (GPUSan)
1011+
if (Pref.GPUSan)
9621012
AddBCLib(getAsanRTLPath(), false);
9631013
};
9641014

9651015
AddSanBCLibs();
9661016
AddBCLib(getOCMLPath());
967-
if (!isOpenMP)
1017+
if (!Pref.IsOpenMP)
9681018
AddBCLib(getOCKLPath());
969-
else if (GPUSan && isOpenMP)
1019+
else if (Pref.GPUSan && Pref.IsOpenMP)
9701020
AddBCLib(getOCKLPath(), false);
971-
AddBCLib(getDenormalsAreZeroPath(DAZ));
972-
AddBCLib(getUnsafeMathPath(UnsafeMathOpt || FastRelaxedMath));
973-
AddBCLib(getFiniteOnlyPath(FiniteOnly || FastRelaxedMath));
974-
AddBCLib(getCorrectlyRoundedSqrtPath(CorrectSqrt));
975-
AddBCLib(getWavefrontSize64Path(Wave64));
1021+
AddBCLib(getDenormalsAreZeroPath(Pref.DAZ));
1022+
AddBCLib(getUnsafeMathPath(Pref.UnsafeMathOpt || Pref.FastRelaxedMath));
1023+
AddBCLib(getFiniteOnlyPath(Pref.FiniteOnly || Pref.FastRelaxedMath));
1024+
AddBCLib(getCorrectlyRoundedSqrtPath(Pref.CorrectSqrt));
1025+
AddBCLib(getWavefrontSize64Path(Pref.Wave64));
9761026
AddBCLib(LibDeviceFile);
977-
auto ABIVerPath = getABIVersionPath(ABIVer);
1027+
auto ABIVerPath = getABIVersionPath(Pref.ABIVer);
9781028
if (!ABIVerPath.empty())
9791029
AddBCLib(ABIVerPath);
9801030

@@ -984,7 +1034,7 @@ RocmInstallationDetector::getCommonBitcodeLibs(
9841034
llvm::SmallVector<ToolChain::BitCodeLibraryInfo, 12>
9851035
ROCMToolChain::getCommonDeviceLibNames(
9861036
const llvm::opt::ArgList &DriverArgs, const std::string &GPUArch,
987-
const Action::OffloadKind DeviceOffloadingKind, bool isOpenMP) const {
1037+
Action::OffloadKind DeviceOffloadingKind) const {
9881038
auto Kind = llvm::AMDGPU::parseArchAMDGCN(GPUArch);
9891039
const StringRef CanonArch = llvm::AMDGPU::getArchNameAMDGCN(Kind);
9901040

@@ -995,38 +1045,9 @@ ROCMToolChain::getCommonDeviceLibNames(
9951045
ABIVer))
9961046
return {};
9971047

998-
// If --hip-device-lib is not set, add the default bitcode libraries.
999-
// TODO: There are way too many flags that change this. Do we need to check
1000-
// them all?
1001-
bool DAZ = DriverArgs.hasFlag(options::OPT_fgpu_flush_denormals_to_zero,
1002-
options::OPT_fno_gpu_flush_denormals_to_zero,
1003-
getDefaultDenormsAreZeroForTarget(Kind));
1004-
bool FiniteOnly = DriverArgs.hasFlag(
1005-
options::OPT_ffinite_math_only, options::OPT_fno_finite_math_only, false);
1006-
bool UnsafeMathOpt =
1007-
DriverArgs.hasFlag(options::OPT_funsafe_math_optimizations,
1008-
options::OPT_fno_unsafe_math_optimizations, false);
1009-
bool FastRelaxedMath = DriverArgs.hasFlag(options::OPT_ffast_math,
1010-
options::OPT_fno_fast_math, false);
1011-
bool CorrectSqrt = false;
1012-
if (DeviceOffloadingKind == Action::OFK_SYCL)
1013-
// When using SYCL, sqrt is only correctly rounded if the flag is specified
1014-
CorrectSqrt = DriverArgs.hasArg(options::OPT_foffload_fp32_prec_sqrt);
1015-
else
1016-
CorrectSqrt = DriverArgs.hasFlag(
1017-
options::OPT_fhip_fp32_correctly_rounded_divide_sqrt,
1018-
options::OPT_fno_hip_fp32_correctly_rounded_divide_sqrt, true);
1019-
bool Wave64 = isWave64(DriverArgs, Kind);
1020-
1021-
// GPU Sanitizer currently only supports ASan and is enabled through host
1022-
// ASan.
1023-
bool GPUSan = DriverArgs.hasFlag(options::OPT_fgpu_sanitize,
1024-
options::OPT_fno_gpu_sanitize, true) &&
1025-
getSanitizerArgs(DriverArgs).needsAsanRt();
1026-
10271048
return RocmInstallation->getCommonBitcodeLibs(
1028-
DriverArgs, LibDeviceFile, Wave64, DAZ, FiniteOnly, UnsafeMathOpt,
1029-
FastRelaxedMath, CorrectSqrt, ABIVer, GPUSan, isOpenMP);
1049+
DriverArgs, LibDeviceFile, GPUArch, DeviceOffloadingKind,
1050+
getSanitizerArgs(DriverArgs).needsAsanRt());
10301051
}
10311052

10321053
bool AMDGPUToolChain::shouldSkipSanitizeOption(

‎clang/lib/Driver/ToolChains/AMDGPU.h‎

Lines changed: 1 addition & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -147,8 +147,7 @@ class LLVM_LIBRARY_VISIBILITY ROCMToolChain : public AMDGPUToolChain {
147147
llvm::SmallVector<BitCodeLibraryInfo, 12>
148148
getCommonDeviceLibNames(const llvm::opt::ArgList &DriverArgs,
149149
const std::string &GPUArch,
150-
const Action::OffloadKind DeviceOffloadingKind,
151-
bool isOpenMP = false) const;
150+
const Action::OffloadKind DeviceOffloadingKind) const;
152151

153152
SanitizerMask getSupportedSanitizers() const override {
154153
return SanitizerKind::Address;

‎clang/lib/Driver/ToolChains/AMDGPUOpenMP.cpp‎

Lines changed: 1 addition & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -360,8 +360,7 @@ AMDGPUOpenMPToolChain::getDeviceLibs(
360360

361361
SmallVector<BitCodeLibraryInfo, 12> BCLibs;
362362
for (auto BCLib :
363-
getCommonDeviceLibNames(Args, GpuArch.str(), DeviceOffloadingKind,
364-
/*IsOpenMP=*/true))
363+
getCommonDeviceLibNames(Args, GpuArch.str(), DeviceOffloadingKind))
365364
BCLibs.emplace_back(BCLib);
366365

367366
return BCLibs;

‎clang/lib/Driver/ToolChains/ROCm.h‎

Lines changed: 24 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -18,6 +18,7 @@
1818
#include "llvm/ADT/StringMap.h"
1919
#include "llvm/Option/ArgList.h"
2020
#include "llvm/Support/VersionTuple.h"
21+
#include "llvm/TargetParser/TargetParser.h"
2122
#include "llvm/TargetParser/Triple.h"
2223

2324
namespace clang {
@@ -77,6 +78,24 @@ class RocmInstallationDetector {
7778
SPACKReleaseStr(SPACKReleaseStr.str()) {}
7879
};
7980

81+
struct CommonBitcodeLibsPreferences {
82+
CommonBitcodeLibsPreferences(const Driver &D,
83+
const llvm::opt::ArgList &DriverArgs,
84+
StringRef GPUArch,
85+
const Action::OffloadKind DeviceOffloadingKind,
86+
const bool NeedsASanRT);
87+
88+
DeviceLibABIVersion ABIVer;
89+
bool IsOpenMP;
90+
bool Wave64;
91+
bool DAZ;
92+
bool FiniteOnly;
93+
bool UnsafeMathOpt;
94+
bool FastRelaxedMath;
95+
bool CorrectSqrt;
96+
bool GPUSan;
97+
};
98+
8099
const Driver &D;
81100
bool HasHIPRuntime = false;
82101
bool HasDeviceLibrary = false;
@@ -175,11 +194,11 @@ class RocmInstallationDetector {
175194

176195
/// Get file paths of default bitcode libraries common to AMDGPU based
177196
/// toolchains.
178-
llvm::SmallVector<ToolChain::BitCodeLibraryInfo, 12> getCommonBitcodeLibs(
179-
const llvm::opt::ArgList &DriverArgs, StringRef LibDeviceFile,
180-
bool Wave64, bool DAZ, bool FiniteOnly, bool UnsafeMathOpt,
181-
bool FastRelaxedMath, bool CorrectSqrt, DeviceLibABIVersion ABIVer,
182-
bool GPUSan, bool isOpenMP) const;
197+
llvm::SmallVector<ToolChain::BitCodeLibraryInfo, 12>
198+
getCommonBitcodeLibs(const llvm::opt::ArgList &DriverArgs,
199+
StringRef LibDeviceFile, StringRef GPUArch,
200+
const Action::OffloadKind DeviceOffloadingKind,
201+
const bool NeedsASanRT) const;
183202
/// Check file paths of default bitcode libraries common to AMDGPU based
184203
/// toolchains. \returns false if there are invalid or missing files.
185204
bool checkCommonBitcodeLibs(StringRef GPUArch, StringRef LibDeviceFile,

0 commit comments

Comments
 (0)