Skip to content

[SYCL] Don't add itt libdevice into linking list for NV,AMD,NativeCPU backend #19603

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 8 commits into
base: sycl
Choose a base branch
from
Open

[SYCL] Don't add itt libdevice into linking list for NV,AMD,NativeCPU backend #19603

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
799c46f
[SYCL] Skip itt libdevice for NV,AMD,NativeCPU backend
jinge90 Jul 28, 2025
17b5572
Merge remote-tracking branch 'upstream/sycl' into simplify_link_nativ…
jinge90 Jul 28, 2025
236c026
fix lit failure
jinge90 Jul 29, 2025
60473a7
Merge remote-tracking branch 'upstream/sycl' into simplify_link_nativ…
jinge90 Jul 29, 2025
97391d0
Merge remote-tracking branch 'upstream/sycl' into simplify_link_nativ…
jinge90 Jul 29, 2025
2216a30
fix lit failure
jinge90 Jul 29, 2025
4f90100
Merge remote-tracking branch 'upstream/sycl' into simplify_link_nativ…
jinge90 Aug 4, 2025
abd7301
Merge remote-tracking branch 'upstream/sycl' into simplify_link_nativ…
jinge90 Aug 8, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 1 addition & 8 deletions clang/lib/Driver/Driver.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5903,13 +5903,6 @@ class OffloadingActionBuilder final {
SmallString<128> LibName(LLCandidate);
llvm::sys::path::append(LibName, DeviceLib);
if (llvm::sys::fs::exists(LibName)) {
// NativeCPU currently only needs libsycl-nativecpu_utils and
// libclc, so temporarily skip other device libs in invocation.
// Todo: remove once NativeCPU tests the other libraries.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If the itt device library should be skipped for NativeCPU, wouldn't keeping this code achieve that, i.e. do you really need to remove this?

Note (as per the original comment) that removing this code will potentially add more device libraries to the NativeCPU link stage, which would slow down linking because NativeCPU currently only needs to link libclc and libsycl-nativecpu_utils. Linking more device libraries in NativeCPU would also need testing for the code linked from these libraries, which seems to go beyond what this PR intends to achieve, which is skipping the itt device lib for the three backends?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hi, @uwedolinsky
Previously, itt device libraries were added to link list together with libsycl-native_utils for nativecpu target, so the logic here is needed to filter out all other libraries except libsycl-native_utils.
Instead of adding unneeded library and then add handling logic to filter out them for nativecpu target, we stops adding itt device libraries into link list when current target is nativecpu, so we don't need to filter out any unneeded libraries.
The PR description seems to be confusing, we just skip "adding" itt device libraries into link list for native cpu/cuda/amd target.
Thanks very much.

if (isNativeCPU &&
!LibName.str().contains("libsycl-nativecpu_utils"))
continue;

++NumOfDeviceLibLinked;
Arg *InputArg = MakeInputArg(Args, C.getDriver().getOpts(),
Args.MakeArgString(LibName));
Expand All @@ -5934,7 +5927,7 @@ class OffloadingActionBuilder final {
}
}

if (!NumOfDeviceLibLinked)
if (!NumOfDeviceLibLinked && !TC->getTriple().isNVPTX())
return false;

// For NVPTX we need to also link libclc at the same stage that we link
Expand Down
22 changes: 15 additions & 7 deletions clang/lib/Driver/ToolChains/SYCL.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -434,7 +434,10 @@ SYCL::getDeviceLibraries(const Compilation &C, const llvm::Triple &TargetTriple,

// For NVPTX and AMDGCN we only use one single bitcode library and ignore
// manually specified SYCL device libraries.
bool IgnoreSingleLibs = TargetTriple.isNVPTX() || TargetTriple.isAMDGCN();
// For NativeCPU, only native_utils devicelib is used.
bool UseSingleLib = TargetTriple.isNVPTX() || TargetTriple.isAMDGCN() ||
TargetTriple.isNativeCPU();
bool IgnoreSingleLib = false;

struct DeviceLibOptInfo {
StringRef DeviceLibName;
Expand Down Expand Up @@ -474,8 +477,10 @@ SYCL::getDeviceLibraries(const Compilation &C, const llvm::Triple &TargetTriple,

// Make sure that internal libraries are still linked against
// when -fno-sycl-device-lib contains "all" and single libraries
// should be ignored.
IgnoreSingleLibs = IgnoreSingleLibs && !ExcludeDeviceLibs;
// should be ignored. For NativeCPU, the native_cpu utils library
// is always linked without '-only-needed' flag.
IgnoreSingleLib =
UseSingleLib && ExcludeDeviceLibs && !TargetTriple.isNativeCPU();

for (const auto &K : DeviceLibLinkInfo.keys())
DeviceLibLinkInfo[K] = (K == "internal") || !ExcludeDeviceLibs;
Expand All @@ -490,21 +495,24 @@ SYCL::getDeviceLibraries(const Compilation &C, const llvm::Triple &TargetTriple,
<< A->getSpelling() << Val;
}
DeviceLibLinkInfo[Val] = !ExcludeDeviceLibs;
PrintUnusedExcludeWarning = IgnoreSingleLibs && ExcludeDeviceLibs;
PrintUnusedExcludeWarning = UseSingleLib && ExcludeDeviceLibs;
}
if (PrintUnusedExcludeWarning)
C.getDriver().Diag(diag::warn_drv_unused_argument) << A->getSpelling();
}
}

if (TargetTriple.isNVPTX() && IgnoreSingleLibs)
if (TargetTriple.isNVPTX() && !IgnoreSingleLib)
LibraryList.push_back(
Args.MakeArgString("devicelib-nvptx64-nvidia-cuda.bc"));

if (TargetTriple.isAMDGCN() && IgnoreSingleLibs)
if (TargetTriple.isAMDGCN() && !IgnoreSingleLib)
LibraryList.push_back(Args.MakeArgString("devicelib-amdgcn-amd-amdhsa.bc"));

if (IgnoreSingleLibs)
if (TargetTriple.isNativeCPU() && !IgnoreSingleLib)
LibraryList.push_back(Args.MakeArgString("libsycl-nativecpu_utils.bc"));

if (UseSingleLib)
return LibraryList;

using SYCLDeviceLibsList = SmallVector<DeviceLibOptInfo, 5>;
Expand Down
15 changes: 0 additions & 15 deletions clang/test/Driver/sycl-device-lib-amdgcn.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,21 +4,6 @@

// UNSUPPORTED: system-windows

// Check if internal libraries are still linked against when linkage of all
// device libs is manually excluded.
// RUN: %clangxx -ccc-print-phases -std=c++11 -fsycl -fno-sycl-device-lib=all --sysroot=%S/Inputs/SYCL \
// RUN: -fsycl-targets=amdgcn-amd-amdhsa -Xsycl-target-backend --offload-arch=gfx906 -fsycl-instrument-device-code %s 2>&1 \
// RUN: | FileCheck -check-prefix=CHK-NO-DEVLIB %s

// CHK-NO-DEVLIB-NOT: {{[0-9]+}}: input, "{{.*}}devicelib-amdgcn-amd-amdhsa.bc", ir, (device-sycl, gfx906)
// CHK-NO-DEVLIB: [[LIB1:[0-9]+]]: input, "{{.*}}libsycl-itt-user-wrappers.bc", ir, (device-sycl, gfx906)
// CHK-NO-DEVLIB-NOT: {{[0-9]+}}: input, "{{.*}}devicelib-amdgcn-amd-amdhsa.bc", ir, (device-sycl, gfx906)
// CHK-NO-DEVLIB: [[LIB2:[0-9]+]]: input, "{{.*}}libsycl-itt-compiler-wrappers.bc", ir, (device-sycl, gfx906)
// CHK-NO-DEVLIB-NOT: {{[0-9]+}}: input, "{{.*}}devicelib-amdgcn-amd-amdhsa.bc", ir, (device-sycl, gfx906)
// CHK-NO-DEVLIB: [[LIB3:[0-9]+]]: input, "{{.*}}libsycl-itt-stubs.bc", ir, (device-sycl, gfx906)
// CHK-NO-DEVLIB-NOT: {{[0-9]+}}: input, "{{.*}}devicelib-amdgcn-amd-amdhsa.bc", ir, (device-sycl, gfx906)
// CHK-NO-DEVLIB: {{[0-9]+}}: linker, {{{.*}}[[LIB1]], [[LIB2]], [[LIB3]]{{.*}}}, ir, (device-sycl, gfx906)

// Check that the -fsycl-device-lib flag has no effect when "all" is specified.
// RUN: %clangxx -ccc-print-phases -std=c++11 -fsycl -fsycl-device-lib=all --sysroot=%S/Inputs/SYCL \
// RUN: -fsycl-targets=amdgcn-amd-amdhsa -Xsycl-target-backend --offload-arch=gfx906 %s 2>&1 \
Expand Down
15 changes: 0 additions & 15 deletions clang/test/Driver/sycl-device-lib-nvptx.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,21 +4,6 @@

// UNSUPPORTED: system-windows

// Check if internal libraries are still linked against when linkage of all
// device libs is manually excluded.
// RUN: %clangxx -ccc-print-phases -std=c++11 -fsycl -fno-sycl-device-lib=all --sysroot=%S/Inputs/SYCL \
// RUN: -fsycl-targets=nvptx64-nvidia-cuda -fsycl-instrument-device-code %s 2>&1 \
// RUN: | FileCheck -check-prefix=CHK-NO-DEVLIB %s

// CHK-NO-DEVLIB-NOT: {{[0-9]+}}: input, "{{.*}}devicelib-nvptx64-nvidia-cuda.bc", ir, (device-sycl, sm_50)
// CHK-NO-DEVLIB: [[LIB1:[0-9]+]]: input, "{{.*}}libsycl-itt-user-wrappers.bc", ir, (device-sycl, sm_50)
// CHK-NO-DEVLIB-NOT: {{[0-9]+}}: input, "{{.*}}devicelib-nvptx64-nvidia-cuda.bc", ir, (device-sycl, sm_50)
// CHK-NO-DEVLIB: [[LIB2:[0-9]+]]: input, "{{.*}}libsycl-itt-compiler-wrappers.bc", ir, (device-sycl, sm_50)
// CHK-NO-DEVLIB-NOT: {{[0-9]+}}: input, "{{.*}}devicelib-nvptx64-nvidia-cuda.bc", ir, (device-sycl, sm_50)
// CHK-NO-DEVLIB: [[LIB3:[0-9]+]]: input, "{{.*}}libsycl-itt-stubs.bc", ir, (device-sycl, sm_50)
// CHK-NO-DEVLIB-NOT: {{[0-9]+}}: input, "{{.*}}devicelib-nvptx64-nvidia-cuda.bc", ir, (device-sycl, sm_50)
// CHK-NO-DEVLIB: {{[0-9]+}}: linker, {{{.*}}[[LIB1]], [[LIB2]], [[LIB3]]{{.*}}}, ir, (device-sycl, sm_50)

// Check that the -fsycl-device-lib flag has no effect when "all" is specified.
// RUN: %clangxx -ccc-print-phases -std=c++11 -fsycl -fsycl-device-lib=all --sysroot=%S/Inputs/SYCL \
// RUN: -fsycl-targets=nvptx64-nvidia-cuda %s 2>&1 \
Expand Down
62 changes: 28 additions & 34 deletions clang/test/Driver/sycl-offload-nvptx.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,7 +37,7 @@
/// Check phases w/out specifying a compute capability.
// RUN: %clangxx -ccc-print-phases --sysroot=%S/Inputs/SYCL -std=c++11 \
// RUN: -target x86_64-unknown-linux-gnu -fsycl -fno-sycl-device-lib=all \
// RUN: -fsycl-instrument-device-code -fsycl-targets=nvptx64-nvidia-cuda %s 2>&1 \
// RUN: -fsycl-targets=nvptx64-nvidia-cuda %s 2>&1 \
// RUN: -fsycl-libspirv-path=%S/Inputs/SYCL/share/clc/remangled-l32-signed_char.libspirv-nvptx64-nvidia-cuda.bc \
// RUN: --cuda-path=%S/Inputs/CUDA_111/usr/local/cuda \
// RUN: | FileCheck -check-prefix=CHK-PHASES-NO-CC %s
Expand All @@ -54,27 +54,24 @@
// CHK-PHASES-NO-CC: 7: backend, {6}, assembler, (host-sycl)
// CHK-PHASES-NO-CC: 8: assembler, {7}, object, (host-sycl)
// CHK-PHASES-NO-CC: 9: linker, {4}, ir, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 10: input, "{{.*}}libsycl-itt-user-wrappers.bc", ir, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 11: input, "{{.*}}libsycl-itt-compiler-wrappers.bc", ir, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 12: input, "{{.*}}libsycl-itt-stubs.bc", ir, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 13: input, "{{.*}}libspirv-nvptx64{{.*}}", ir, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 14: input, "{{.*}}libdevice{{.*}}", ir, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 15: linker, {9, 10, 11, 12, 13, 14}, ir, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 16: sycl-post-link, {15}, ir, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 17: file-table-tform, {16}, ir, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 18: backend, {17}, assembler, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 19: assembler, {18}, object, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 20: linker, {18, 19}, cuda-fatbin, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 21: foreach, {17, 20}, cuda-fatbin, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 22: file-table-tform, {16, 21}, tempfiletable, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 23: clang-offload-wrapper, {22}, object, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 24: offload, "device-sycl (nvptx64-nvidia-cuda:sm_50)" {23}, object
// CHK-PHASES-NO-CC: 25: linker, {8, 24}, image, (host-sycl)
// CHK-PHASES-NO-CC: 10: input, "{{.*}}libspirv-nvptx64{{.*}}", ir, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 11: input, "{{.*}}libdevice{{.*}}", ir, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 12: linker, {9, 10, 11}, ir, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 13: sycl-post-link, {12}, ir, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 14: file-table-tform, {13}, ir, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 15: backend, {14}, assembler, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 16: assembler, {15}, object, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 17: linker, {15, 16}, cuda-fatbin, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 18: foreach, {14, 17}, cuda-fatbin, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 19: file-table-tform, {13, 18}, tempfiletable, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 20: clang-offload-wrapper, {19}, object, (device-sycl, sm_50)
// CHK-PHASES-NO-CC: 21: offload, "device-sycl (nvptx64-nvidia-cuda:sm_50)" {20}, object
// CHK-PHASES-NO-CC: 22: linker, {8, 21}, image, (host-sycl)
//
/// Check phases specifying a compute capability.
// RUN: %clangxx -ccc-print-phases --sysroot=%S/Inputs/SYCL -std=c++11 \
// RUN: -target x86_64-unknown-linux-gnu -fsycl -fno-sycl-device-lib=all \
// RUN: -fsycl-instrument-device-code -fsycl-targets=nvptx64-nvidia-cuda \
// RUN: -fsycl-targets=nvptx64-nvidia-cuda \
// RUN: -fsycl-libspirv-path=%S/Inputs/SYCL/share/clc/remangled-l32-signed_char.libspirv-nvptx64-nvidia-cuda.bc \
// RUN: --cuda-path=%S/Inputs/CUDA_111/usr/local/cuda \
// RUN: -Xsycl-target-backend "--cuda-gpu-arch=sm_35" %s 2>&1 \
Expand All @@ -92,22 +89,19 @@
// CHK-PHASES: 7: backend, {6}, assembler, (host-sycl)
// CHK-PHASES: 8: assembler, {7}, object, (host-sycl)
// CHK-PHASES: 9: linker, {4}, ir, (device-sycl, sm_35)
// CHK-PHASES: 10: input, "{{.*}}libsycl-itt-user-wrappers.bc", ir, (device-sycl, sm_35)
// CHK-PHASES: 11: input, "{{.*}}libsycl-itt-compiler-wrappers.bc", ir, (device-sycl, sm_35)
// CHK-PHASES: 12: input, "{{.*}}libsycl-itt-stubs.bc", ir, (device-sycl, sm_35)
// CHK-PHASES: 13: input, "{{.*}}libspirv-nvptx64{{.*}}", ir, (device-sycl, sm_35)
// CHK-PHASES: 14: input, "{{.*}}libdevice{{.*}}", ir, (device-sycl, sm_35)
// CHK-PHASES: 15: linker, {9, 10, 11, 12, 13, 14}, ir, (device-sycl, sm_35)
// CHK-PHASES: 16: sycl-post-link, {15}, ir, (device-sycl, sm_35)
// CHK-PHASES: 17: file-table-tform, {16}, ir, (device-sycl, sm_35)
// CHK-PHASES: 18: backend, {17}, assembler, (device-sycl, sm_35)
// CHK-PHASES: 19: assembler, {18}, object, (device-sycl, sm_35)
// CHK-PHASES: 20: linker, {18, 19}, cuda-fatbin, (device-sycl, sm_35)
// CHK-PHASES: 21: foreach, {17, 20}, cuda-fatbin, (device-sycl, sm_35)
// CHK-PHASES: 22: file-table-tform, {16, 21}, tempfiletable, (device-sycl, sm_35)
// CHK-PHASES: 23: clang-offload-wrapper, {22}, object, (device-sycl, sm_35)
// CHK-PHASES: 24: offload, "device-sycl (nvptx64-nvidia-cuda:sm_35)" {23}, object
// CHK-PHASES: 25: linker, {8, 24}, image, (host-sycl)
// CHK-PHASES: 10: input, "{{.*}}libspirv-nvptx64{{.*}}", ir, (device-sycl, sm_35)
// CHK-PHASES: 11: input, "{{.*}}libdevice{{.*}}", ir, (device-sycl, sm_35)
// CHK-PHASES: 12: linker, {9, 10, 11}, ir, (device-sycl, sm_35)
// CHK-PHASES: 13: sycl-post-link, {12}, ir, (device-sycl, sm_35)
// CHK-PHASES: 14: file-table-tform, {13}, ir, (device-sycl, sm_35)
// CHK-PHASES: 15: backend, {14}, assembler, (device-sycl, sm_35)
// CHK-PHASES: 16: assembler, {15}, object, (device-sycl, sm_35)
// CHK-PHASES: 17: linker, {15, 16}, cuda-fatbin, (device-sycl, sm_35)
// CHK-PHASES: 18: foreach, {14, 17}, cuda-fatbin, (device-sycl, sm_35)
// CHK-PHASES: 19: file-table-tform, {13, 18}, tempfiletable, (device-sycl, sm_35)
// CHK-PHASES: 20: clang-offload-wrapper, {19}, object, (device-sycl, sm_35)
// CHK-PHASES: 21: offload, "device-sycl (nvptx64-nvidia-cuda:sm_35)" {20}, object
// CHK-PHASES: 22: linker, {8, 21}, image, (host-sycl)

/// Check calling preprocessor only
// RUN: %clangxx -E -fsycl -fsycl-targets=nvptx64-nvidia-cuda -ccc-print-phases %s 2>&1 \
Expand Down
Loading