[clang] [Offload] Do not pass `-fcf-protection=` for offloading (PR #88402)
@@ -0,0 +1,39 @@ +// Check that -fcf-protection does not get passed to the device-side MaskRay wrote: We also have unsupported-option-gpu.c to test various ignored options for GPU. https://github.com/llvm/llvm-project/pull/88402 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [Offload] Do not pass `-fcf-protection=` for offloading (PR #88402)
jhuber6 wrote: I'm wondering if it would be better to make this something that the toolchains handle in `TranslateArgs`. That way we can prevent it from being added at all. Although I don't know how that would allow us to override stuff. There's not really a way to "override" flags like this if we ignore it at all. We could "assume" that if someone does `-Xarch_device -foo` they want the device to use it, but that's not really what the flag means. https://github.com/llvm/llvm-project/pull/88402 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [Offload] Do not pass `-fcf-protection=` for offloading (PR #88402)
@@ -6867,8 +6867,14 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA, CmdArgs.push_back("-nogpulib"); if (Arg *A = Args.getLastArg(options::OPT_fcf_protection_EQ)) { -CmdArgs.push_back( -Args.MakeArgString(Twine("-fcf-protection=") + A->getValue())); +// Do not pass this argument to the offloading device if the target does not +// support it. +// TODO: We need a better way to detect incompatible options for offloading. +if (JA.getOffloadingDeviceKind() == Action::OFK_None || +(!TC.getTriple().isAMDGPU() && !TC.getTriple().isNVPTX() && + !TC.getTriple().isSPIRV())) Artem-B wrote: +1. We have grown too many offloading cases all over the place over time. It was fine when there was only CUDA/NVPTX, was sort of OK when AMDGPU got added, now it gets to be a bit too much. https://github.com/llvm/llvm-project/pull/88402 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [Offload] Do not pass `-fcf-protection=` for offloading (PR #88402)
https://github.com/Artem-B commented: LGTM in principle. https://github.com/llvm/llvm-project/pull/88402 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [Offload] Do not pass `-fcf-protection=` for offloading (PR #88402)
https://github.com/Artem-B edited https://github.com/llvm/llvm-project/pull/88402 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [Offload] Do not pass `-fcf-protection=` for offloading (PR #88402)
@@ -6867,8 +6867,14 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA, CmdArgs.push_back("-nogpulib"); if (Arg *A = Args.getLastArg(options::OPT_fcf_protection_EQ)) { -CmdArgs.push_back( -Args.MakeArgString(Twine("-fcf-protection=") + A->getValue())); +// Do not pass this argument to the offloading device if the target does not +// support it. +// TODO: We need a better way to detect incompatible options for offloading. +if (JA.getOffloadingDeviceKind() == Action::OFK_None || +(!TC.getTriple().isAMDGPU() && !TC.getTriple().isNVPTX() && Artem-B wrote: Nit: I'd collapse negations into one: ``` !(TC.getTriple().isAMDGPU() || TC.getTriple().isNVPTX() || TC.getTriple().isSPIRV()) ``` https://github.com/llvm/llvm-project/pull/88402 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [Offload] Do not pass `-fcf-protection=` for offloading (PR #88402)
@@ -6867,8 +6867,14 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA, CmdArgs.push_back("-nogpulib"); if (Arg *A = Args.getLastArg(options::OPT_fcf_protection_EQ)) { -CmdArgs.push_back( -Args.MakeArgString(Twine("-fcf-protection=") + A->getValue())); +// Do not pass this argument to the offloading device if the target does not +// support it. +// TODO: We need a better way to detect incompatible options for offloading. +if (JA.getOffloadingDeviceKind() == Action::OFK_None || +(!TC.getTriple().isAMDGPU() && !TC.getTriple().isNVPTX() && + !TC.getTriple().isSPIRV())) arsenm wrote: It is 100% better https://github.com/llvm/llvm-project/pull/88402 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [Offload] Do not pass `-fcf-protection=` for offloading (PR #88402)
@@ -0,0 +1,39 @@ +// Check that -fcf-protection does not get passed to the device-side +// compilation. + +// RUN: %clang -### -x cuda --target=x86_64-unknown-linux-gnu -nogpulib \ +// RUN: -nogpuinc --offload-arch=sm_52 -fcf-protection=full -c %s 2>&1 \ +// RUN: | FileCheck %s --check-prefix=CUDA + +// CUDA: "-cc1" "-triple" "nvptx64-nvidia-cuda" +// CUDA-NOT: "-fcf-protection=full" +// CUDA: "-cc1" "-triple" "x86_64-unknown-linux-gnu" +// CUDA: "-fcf-protection=full" + +// RUN: %clang -### -x hip --target=x86_64-unknown-linux-gnu -nogpulib \ +// RUN: -nogpuinc --offload-arch=gfx90a -fcf-protection=full -c %s 2>&1 \ +// RUN: | FileCheck %s --check-prefix=HIP + +// HIP: "-cc1" "-triple" "amdgcn-amd-amdhsa" +// HIP-NOT: "-fcf-protection=full" +// HIP: "-cc1" "-triple" "x86_64-unknown-linux-gnu" +// HIP: "-fcf-protection=full" + jhuber6 wrote: I don't know how that would be possible. We have options like `-Xarch_device`, but those are stripped by the time we create the synthetic arguments that make it to the tool builder. Potentially we could do this logic where we build the args and just prevent the offloading toolchain from seeing it at all? https://github.com/llvm/llvm-project/pull/88402 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [Offload] Do not pass `-fcf-protection=` for offloading (PR #88402)
@@ -6867,8 +6867,14 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA, CmdArgs.push_back("-nogpulib"); if (Arg *A = Args.getLastArg(options::OPT_fcf_protection_EQ)) { -CmdArgs.push_back( -Args.MakeArgString(Twine("-fcf-protection=") + A->getValue())); +// Do not pass this argument to the offloading device if the target does not +// support it. +// TODO: We need a better way to detect incompatible options for offloading. +if (JA.getOffloadingDeviceKind() == Action::OFK_None || +(!TC.getTriple().isAMDGPU() && !TC.getTriple().isNVPTX() && + !TC.getTriple().isSPIRV())) jhuber6 wrote: I mean we could potentially just have some toolchain check like `isToolchainGPU`, but I don't know if that's exactly any better. https://github.com/llvm/llvm-project/pull/88402 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [Offload] Do not pass `-fcf-protection=` for offloading (PR #88402)
@@ -0,0 +1,39 @@ +// Check that -fcf-protection does not get passed to the device-side +// compilation. + +// RUN: %clang -### -x cuda --target=x86_64-unknown-linux-gnu -nogpulib \ +// RUN: -nogpuinc --offload-arch=sm_52 -fcf-protection=full -c %s 2>&1 \ +// RUN: | FileCheck %s --check-prefix=CUDA + +// CUDA: "-cc1" "-triple" "nvptx64-nvidia-cuda" +// CUDA-NOT: "-fcf-protection=full" +// CUDA: "-cc1" "-triple" "x86_64-unknown-linux-gnu" +// CUDA: "-fcf-protection=full" + +// RUN: %clang -### -x hip --target=x86_64-unknown-linux-gnu -nogpulib \ +// RUN: -nogpuinc --offload-arch=gfx90a -fcf-protection=full -c %s 2>&1 \ +// RUN: | FileCheck %s --check-prefix=HIP + +// HIP: "-cc1" "-triple" "amdgcn-amd-amdhsa" +// HIP-NOT: "-fcf-protection=full" +// HIP: "-cc1" "-triple" "x86_64-unknown-linux-gnu" +// HIP: "-fcf-protection=full" + +// RUN: %clang -### -x c --target=x86_64-unknown-linux-gnu -nogpulib -fopenmp=libomp \ +// RUN: -nogpuinc --offload-arch=gfx90a -fcf-protection=full -c %s 2>&1 \ +// RUN: | FileCheck %s --check-prefix=OMP + +// OMP: "-cc1" "-triple" "x86_64-unknown-linux-gnu" +// OMP: "-fcf-protection=full" +// OMP: "-cc1" "-triple" "amdgcn-amd-amdhsa" +// OMP-NOT: "-fcf-protection=full" +// OMP: "-cc1" "-triple" "x86_64-unknown-linux-gnu" +// OMP: "-fcf-protection=full" + +// RUN: %clang -### -x c --target=nvptx64-nvidia-cuda -nogpulib -nogpuinc \ +// RUN: -march=sm_52 -fcf-protection=full -c %s 2>&1 \ +// RUN: | FileCheck %s --check-prefix=DIRECT +// RUN: %clang -### -x c --target=amdgcn-amd-amdhsa -nogpulib -nogpuinc \ +// RUN: -mcpu=gfx90a -fcf-protection=full -c %s 2>&1 \ +// RUN: | FileCheck %s --check-prefix=DIRECT +// DIRECT: "-fcf-protection=full" arsenm wrote: Missing spirv run lines https://github.com/llvm/llvm-project/pull/88402 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [Offload] Do not pass `-fcf-protection=` for offloading (PR #88402)
@@ -0,0 +1,39 @@ +// Check that -fcf-protection does not get passed to the device-side +// compilation. + +// RUN: %clang -### -x cuda --target=x86_64-unknown-linux-gnu -nogpulib \ +// RUN: -nogpuinc --offload-arch=sm_52 -fcf-protection=full -c %s 2>&1 \ +// RUN: | FileCheck %s --check-prefix=CUDA + +// CUDA: "-cc1" "-triple" "nvptx64-nvidia-cuda" +// CUDA-NOT: "-fcf-protection=full" +// CUDA: "-cc1" "-triple" "x86_64-unknown-linux-gnu" +// CUDA: "-fcf-protection=full" + +// RUN: %clang -### -x hip --target=x86_64-unknown-linux-gnu -nogpulib \ +// RUN: -nogpuinc --offload-arch=gfx90a -fcf-protection=full -c %s 2>&1 \ +// RUN: | FileCheck %s --check-prefix=HIP + +// HIP: "-cc1" "-triple" "amdgcn-amd-amdhsa" +// HIP-NOT: "-fcf-protection=full" +// HIP: "-cc1" "-triple" "x86_64-unknown-linux-gnu" +// HIP: "-fcf-protection=full" + arsenm wrote: Can you check explicitly passing this through to the device? https://github.com/llvm/llvm-project/pull/88402 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [Offload] Do not pass `-fcf-protection=` for offloading (PR #88402)
@@ -6867,8 +6867,14 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA, CmdArgs.push_back("-nogpulib"); if (Arg *A = Args.getLastArg(options::OPT_fcf_protection_EQ)) { -CmdArgs.push_back( -Args.MakeArgString(Twine("-fcf-protection=") + A->getValue())); +// Do not pass this argument to the offloading device if the target does not +// support it. +// TODO: We need a better way to detect incompatible options for offloading. +if (JA.getOffloadingDeviceKind() == Action::OFK_None || +(!TC.getTriple().isAMDGPU() && !TC.getTriple().isNVPTX() && + !TC.getTriple().isSPIRV())) arsenm wrote: Can we avoid adding yet another one of these arbitrary architecture list cases https://github.com/llvm/llvm-project/pull/88402 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [Offload] Do not pass `-fcf-protection=` for offloading (PR #88402)
https://github.com/jhuber6 updated https://github.com/llvm/llvm-project/pull/88402 >From 9b9cbaa09425a706eaf3bb8e85a824ef89b61a9f Mon Sep 17 00:00:00 2001 From: Joseph Huber Date: Thu, 11 Apr 2024 10:36:15 -0500 Subject: [PATCH] [Offload] Do not pass `-fcf-protection=` for offloading Summary: This patch prevents the `-fcf-protection=` flag from being passed to the device compilation during offloading. This is not supported on CUDA and AMD devices, but if the user is compiling with fcf protection for the host it will fail to compile. We have a lot of these cases with various hacked together solutions, it would be nice to have a single solution to detect from the driver if a feature like this can be used for offloading, but for now this should resolve the issue. Fixe: https://github.com/llvm/llvm-project/issues/86450 --- clang/lib/Driver/ToolChains/Clang.cpp | 10 - clang/test/Driver/hip-options.hip | 8 ++-- clang/test/Driver/offload-no-fcf-protection.c | 39 +++ 3 files changed, 51 insertions(+), 6 deletions(-) create mode 100644 clang/test/Driver/offload-no-fcf-protection.c diff --git a/clang/lib/Driver/ToolChains/Clang.cpp b/clang/lib/Driver/ToolChains/Clang.cpp index 766a9b91e3c0ad..1264ffa1ef7c8e 100644 --- a/clang/lib/Driver/ToolChains/Clang.cpp +++ b/clang/lib/Driver/ToolChains/Clang.cpp @@ -6867,8 +6867,14 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA, CmdArgs.push_back("-nogpulib"); if (Arg *A = Args.getLastArg(options::OPT_fcf_protection_EQ)) { -CmdArgs.push_back( -Args.MakeArgString(Twine("-fcf-protection=") + A->getValue())); +// Do not pass this argument to the offloading device if the target does not +// support it. +// TODO: We need a better way to detect incompatible options for offloading. +if (JA.getOffloadingDeviceKind() == Action::OFK_None || +(!TC.getTriple().isAMDGPU() && !TC.getTriple().isNVPTX() && + !TC.getTriple().isSPIRV())) + CmdArgs.push_back( + Args.MakeArgString(Twine("-fcf-protection=") + A->getValue())); } if (Arg *A = Args.getLastArg(options::OPT_mfunction_return_EQ)) diff --git a/clang/test/Driver/hip-options.hip b/clang/test/Driver/hip-options.hip index 2ba9032f16946b..679f3084ace14e 100644 --- a/clang/test/Driver/hip-options.hip +++ b/clang/test/Driver/hip-options.hip @@ -43,10 +43,10 @@ // MLLVM-NOT: "-mllvm"{{.*}}"-amdgpu-early-inline-all=true"{{.*}}"-mllvm"{{.*}}"-amdgpu-early-inline-all=true" // RUN: %clang -### -Xarch_device -g -nogpulib -nogpuinc --cuda-gpu-arch=gfx900 \ -// RUN: -Xarch_device -fcf-protection=branch -Xarch_device -mllvm=--inline-threshold=100 \ +// RUN: -Xarch_device -mllvm=--inline-threshold=100 \ // RUN: --cuda-gpu-arch=gfx906 %s 2>&1 | FileCheck -check-prefix=DEV %s -// DEV: "-cc1"{{.*}} "-fcuda-is-device" {{.*}} "-debug-info-kind={{.*}}" {{.*}} "-fcf-protection=branch" {{.*}}"-mllvm" "--inline-threshold=100" -// DEV: "-cc1"{{.*}} "-fcuda-is-device" {{.*}} "-debug-info-kind={{.*}}" {{.*}} "-fcf-protection=branch" {{.*}}"-mllvm" "--inline-threshold=100" +// DEV: "-cc1"{{.*}} "-fcuda-is-device" {{.*}} "-debug-info-kind={{.*}}"{{.*}}"-mllvm" "--inline-threshold=100" +// DEV: "-cc1"{{.*}} "-fcuda-is-device" {{.*}} "-debug-info-kind={{.*}}"{{.*}}"-mllvm" "--inline-threshold=100" // DEV-NOT: clang{{.*}} {{.*}} "-debug-info-kind={{.*}}" // RUN: %clang -### -Xarch_host -g -nogpulib -nogpuinc --cuda-gpu-arch=gfx900 \ @@ -244,4 +244,4 @@ // RUN: 2>&1 | FileCheck -check-prefix=NO-WARN-ATOMIC %s // NO-WARN-ATOMIC: clang{{.*}} "-triple" "amdgcn-amd-amdhsa" {{.*}} "-Werror=atomic-alignment" {{.*}} "-Wno-error=atomic-alignment" // NO-WARN-ATOMIC-NOT: clang{{.*}} "-triple" "x86_64-unknown-linux-gnu" {{.*}} "-Werror=atomic-alignment" -// NO-WARN-ATOMIC-NOT: clang{{.*}} "-triple" "x86_64-unknown-linux-gnu" {{.*}} "-Wno-error=atomic-alignment" \ No newline at end of file +// NO-WARN-ATOMIC-NOT: clang{{.*}} "-triple" "x86_64-unknown-linux-gnu" {{.*}} "-Wno-error=atomic-alignment" diff --git a/clang/test/Driver/offload-no-fcf-protection.c b/clang/test/Driver/offload-no-fcf-protection.c new file mode 100644 index 00..ef275881edbaf1 --- /dev/null +++ b/clang/test/Driver/offload-no-fcf-protection.c @@ -0,0 +1,39 @@ +// Check that -fcf-protection does not get passed to the device-side +// compilation. + +// RUN: %clang -### -x cuda --target=x86_64-unknown-linux-gnu -nogpulib \ +// RUN: -nogpuinc --offload-arch=sm_52 -fcf-protection=full -c %s 2>&1 \ +// RUN: | FileCheck %s --check-prefix=CUDA + +// CUDA: "-cc1" "-triple" "nvptx64-nvidia-cuda" +// CUDA-NOT: "-fcf-protection=full" +// CUDA: "-cc1" "-triple" "x86_64-unknown-linux-gnu" +// CUDA: "-fcf-protection=full" + +// RUN: %clang -### -x hip --target=x86_64-unknown-linux-gnu -nogpulib \ +// RUN: -nogpuinc --offload-arch=gfx90a -fcf-protection=full -c %s 2>&1 \ +// RUN: | FileCheck %s --check-prefix=HIP + +// HIP: "-cc1" "-
[clang] [Offload] Do not pass `-fcf-protection=` for offloading (PR #88402)
llvmbot wrote: @llvm/pr-subscribers-clang @llvm/pr-subscribers-clang-driver Author: Joseph Huber (jhuber6) Changes Summary: This patch prevents the `-fcf-protection=` flag from being passed to the device compilation during offloading. This is not supported on CUDA and AMD devices, but if the user is compiling with fcf protection for the host it will fail to compile. We have a lot of these cases with various hacked together solutions, it would be nice to have a single solution to detect from the driver if a feature like this can be used for offloading, but for now this should resolve the issue. Fixes: https://github.com/llvm/llvm-project/issues/86450 --- Full diff: https://github.com/llvm/llvm-project/pull/88402.diff 2 Files Affected: - (modified) clang/lib/Driver/ToolChains/Clang.cpp (+8-2) - (added) clang/test/Driver/offload-no-fcf-protection.c (+39) ``diff diff --git a/clang/lib/Driver/ToolChains/Clang.cpp b/clang/lib/Driver/ToolChains/Clang.cpp index 766a9b91e3c0ad..1264ffa1ef7c8e 100644 --- a/clang/lib/Driver/ToolChains/Clang.cpp +++ b/clang/lib/Driver/ToolChains/Clang.cpp @@ -6867,8 +6867,14 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA, CmdArgs.push_back("-nogpulib"); if (Arg *A = Args.getLastArg(options::OPT_fcf_protection_EQ)) { -CmdArgs.push_back( -Args.MakeArgString(Twine("-fcf-protection=") + A->getValue())); +// Do not pass this argument to the offloading device if the target does not +// support it. +// TODO: We need a better way to detect incompatible options for offloading. +if (JA.getOffloadingDeviceKind() == Action::OFK_None || +(!TC.getTriple().isAMDGPU() && !TC.getTriple().isNVPTX() && + !TC.getTriple().isSPIRV())) + CmdArgs.push_back( + Args.MakeArgString(Twine("-fcf-protection=") + A->getValue())); } if (Arg *A = Args.getLastArg(options::OPT_mfunction_return_EQ)) diff --git a/clang/test/Driver/offload-no-fcf-protection.c b/clang/test/Driver/offload-no-fcf-protection.c new file mode 100644 index 00..ef275881edbaf1 --- /dev/null +++ b/clang/test/Driver/offload-no-fcf-protection.c @@ -0,0 +1,39 @@ +// Check that -fcf-protection does not get passed to the device-side +// compilation. + +// RUN: %clang -### -x cuda --target=x86_64-unknown-linux-gnu -nogpulib \ +// RUN: -nogpuinc --offload-arch=sm_52 -fcf-protection=full -c %s 2>&1 \ +// RUN: | FileCheck %s --check-prefix=CUDA + +// CUDA: "-cc1" "-triple" "nvptx64-nvidia-cuda" +// CUDA-NOT: "-fcf-protection=full" +// CUDA: "-cc1" "-triple" "x86_64-unknown-linux-gnu" +// CUDA: "-fcf-protection=full" + +// RUN: %clang -### -x hip --target=x86_64-unknown-linux-gnu -nogpulib \ +// RUN: -nogpuinc --offload-arch=gfx90a -fcf-protection=full -c %s 2>&1 \ +// RUN: | FileCheck %s --check-prefix=HIP + +// HIP: "-cc1" "-triple" "amdgcn-amd-amdhsa" +// HIP-NOT: "-fcf-protection=full" +// HIP: "-cc1" "-triple" "x86_64-unknown-linux-gnu" +// HIP: "-fcf-protection=full" + +// RUN: %clang -### -x c --target=x86_64-unknown-linux-gnu -nogpulib -fopenmp=libomp \ +// RUN: -nogpuinc --offload-arch=gfx90a -fcf-protection=full -c %s 2>&1 \ +// RUN: | FileCheck %s --check-prefix=OMP + +// OMP: "-cc1" "-triple" "x86_64-unknown-linux-gnu" +// OMP: "-fcf-protection=full" +// OMP: "-cc1" "-triple" "amdgcn-amd-amdhsa" +// OMP-NOT: "-fcf-protection=full" +// OMP: "-cc1" "-triple" "x86_64-unknown-linux-gnu" +// OMP: "-fcf-protection=full" + +// RUN: %clang -### -x c --target=nvptx64-nvidia-cuda -nogpulib -nogpuinc \ +// RUN: -march=sm_52 -fcf-protection=full -c %s 2>&1 \ +// RUN: | FileCheck %s --check-prefix=DIRECT +// RUN: %clang -### -x c --target=amdgcn-amd-amdhsa -nogpulib -nogpuinc \ +// RUN: -mcpu=gfx90a -fcf-protection=full -c %s 2>&1 \ +// RUN: | FileCheck %s --check-prefix=DIRECT +// DIRECT: "-fcf-protection=full" `` https://github.com/llvm/llvm-project/pull/88402 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [Offload] Do not pass `-fcf-protection=` for offloading (PR #88402)
https://github.com/jhuber6 edited https://github.com/llvm/llvm-project/pull/88402 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] [Offload] Do not pass `-fcf-protection=` for offloading (PR #88402)
https://github.com/jhuber6 created https://github.com/llvm/llvm-project/pull/88402 Summary: This patch prevents the `-fcf-protection=` flag from being passed to the device compilation during offloading. This is not supported on CUDA and AMD devices, but if the user is compiling with fcf protection for the host it will fail to compile. We have a lot of these cases with various hacked together solutions, it would be nice to have a single solution to detect from the driver if a feature like this can be used for offloading, but for now this should resolve the issue. Fixe: https://github.com/llvm/llvm-project/issues/86450 >From e228f8bb81bd44bc3c76d0d61836f068612a9d1e Mon Sep 17 00:00:00 2001 From: Joseph Huber Date: Thu, 11 Apr 2024 10:36:15 -0500 Subject: [PATCH] [Offload] Do not pass `-fcf-protection=` for offloading Summary: This patch prevents the `-fcf-protection=` flag from being passed to the device compilation during offloading. This is not supported on CUDA and AMD devices, but if the user is compiling with fcf protection for the host it will fail to compile. We have a lot of these cases with various hacked together solutions, it would be nice to have a single solution to detect from the driver if a feature like this can be used for offloading, but for now this should resolve the issue. Fixe: https://github.com/llvm/llvm-project/issues/86450 --- clang/lib/Driver/ToolChains/Clang.cpp | 10 - clang/test/Driver/offload-no-fcf-protection.c | 39 +++ 2 files changed, 47 insertions(+), 2 deletions(-) create mode 100644 clang/test/Driver/offload-no-fcf-protection.c diff --git a/clang/lib/Driver/ToolChains/Clang.cpp b/clang/lib/Driver/ToolChains/Clang.cpp index 766a9b91e3c0ad..1264ffa1ef7c8e 100644 --- a/clang/lib/Driver/ToolChains/Clang.cpp +++ b/clang/lib/Driver/ToolChains/Clang.cpp @@ -6867,8 +6867,14 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA, CmdArgs.push_back("-nogpulib"); if (Arg *A = Args.getLastArg(options::OPT_fcf_protection_EQ)) { -CmdArgs.push_back( -Args.MakeArgString(Twine("-fcf-protection=") + A->getValue())); +// Do not pass this argument to the offloading device if the target does not +// support it. +// TODO: We need a better way to detect incompatible options for offloading. +if (JA.getOffloadingDeviceKind() == Action::OFK_None || +(!TC.getTriple().isAMDGPU() && !TC.getTriple().isNVPTX() && + !TC.getTriple().isSPIRV())) + CmdArgs.push_back( + Args.MakeArgString(Twine("-fcf-protection=") + A->getValue())); } if (Arg *A = Args.getLastArg(options::OPT_mfunction_return_EQ)) diff --git a/clang/test/Driver/offload-no-fcf-protection.c b/clang/test/Driver/offload-no-fcf-protection.c new file mode 100644 index 00..ef275881edbaf1 --- /dev/null +++ b/clang/test/Driver/offload-no-fcf-protection.c @@ -0,0 +1,39 @@ +// Check that -fcf-protection does not get passed to the device-side +// compilation. + +// RUN: %clang -### -x cuda --target=x86_64-unknown-linux-gnu -nogpulib \ +// RUN: -nogpuinc --offload-arch=sm_52 -fcf-protection=full -c %s 2>&1 \ +// RUN: | FileCheck %s --check-prefix=CUDA + +// CUDA: "-cc1" "-triple" "nvptx64-nvidia-cuda" +// CUDA-NOT: "-fcf-protection=full" +// CUDA: "-cc1" "-triple" "x86_64-unknown-linux-gnu" +// CUDA: "-fcf-protection=full" + +// RUN: %clang -### -x hip --target=x86_64-unknown-linux-gnu -nogpulib \ +// RUN: -nogpuinc --offload-arch=gfx90a -fcf-protection=full -c %s 2>&1 \ +// RUN: | FileCheck %s --check-prefix=HIP + +// HIP: "-cc1" "-triple" "amdgcn-amd-amdhsa" +// HIP-NOT: "-fcf-protection=full" +// HIP: "-cc1" "-triple" "x86_64-unknown-linux-gnu" +// HIP: "-fcf-protection=full" + +// RUN: %clang -### -x c --target=x86_64-unknown-linux-gnu -nogpulib -fopenmp=libomp \ +// RUN: -nogpuinc --offload-arch=gfx90a -fcf-protection=full -c %s 2>&1 \ +// RUN: | FileCheck %s --check-prefix=OMP + +// OMP: "-cc1" "-triple" "x86_64-unknown-linux-gnu" +// OMP: "-fcf-protection=full" +// OMP: "-cc1" "-triple" "amdgcn-amd-amdhsa" +// OMP-NOT: "-fcf-protection=full" +// OMP: "-cc1" "-triple" "x86_64-unknown-linux-gnu" +// OMP: "-fcf-protection=full" + +// RUN: %clang -### -x c --target=nvptx64-nvidia-cuda -nogpulib -nogpuinc \ +// RUN: -march=sm_52 -fcf-protection=full -c %s 2>&1 \ +// RUN: | FileCheck %s --check-prefix=DIRECT +// RUN: %clang -### -x c --target=amdgcn-amd-amdhsa -nogpulib -nogpuinc \ +// RUN: -mcpu=gfx90a -fcf-protection=full -c %s 2>&1 \ +// RUN: | FileCheck %s --check-prefix=DIRECT +// DIRECT: "-fcf-protection=full" ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits