[NewOffloadModel] Remove compiler backend option and linker option to be passed as argument for ClangLinkerWrapper #20691

YixingZhang007 · 2025-11-19T18:59:45Z

In another patch #19579, compilation and linking options used in ClangLinkerWrapper are changed to be consumed from the SYCLImage, instead of sycl_backend_compile_options_EQ and sycl_target_link_options_EQ that are passed as argument to the ClangLinkerWrapper. Since the compiler option and linker option passed into ClangLinkerWrapper are no longer useful, this PR remove them to be passed into ClangLinkerWrapper. The specific changes are listed below.

The code for constructing the BackendOptString and LinkOptString and append to the argument for calling ClangLinkerWrapper are removed in Clang.cpp.
Remove gpu_tool_arg_EQ and cpu_tool_arg_EQ options specification in LinkerWrapperOpts.td. We are now passing the link time backend compile option for GPU and CPU through OPT_device_compiler_args_EQ to ClangLinkerWrapper, as coded in Clang.cpp.
In ClangLinkerWrapper.cpp, the option in OPT_device_compiler_args_EQ are added to SYCLBackendOptions and handled together with other compiler option passed through SYCLImage.
Modify the tests that are associated with these changes, such as those that are checking the compiler backend option and linker option passed into ClangLinkerWrapper.

With this patch, the following SYCL E2E tests pass with the new offloading model:

KernelAndProgram/kernel-bundle-merge-options-env.cpp
KernelAndProgram/level-zero-link-flags.cpp
ESIMD/private_memory/pm_access_1.cpp
ESIMD/private_memory/pm_access_2.cpp
ESIMD/private_memory/pm_access_3.cpp

Copilot

Pull request overview

This PR removes the passing of backend compiler and linker options via --gpu-tool-arg and --cpu-tool-arg arguments to ClangLinkerWrapper. Instead, these options are now consumed directly from SYCLImage metadata using the --device-compiler argument format. The change simplifies the option passing mechanism by consolidating all backend compile options through a single path.

Remove obsolete gpu_tool_arg_EQ and cpu_tool_arg_EQ option definitions
Refactor option handling in ClangLinkerWrapper to use --device-compiler format instead
Update runSYCLPostLinkTool to accept HasGPUTool boolean parameter instead of checking removed option

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
clang/tools/clang-linker-wrapper/LinkerWrapperOpts.td	Removes `gpu_tool_arg_EQ` and `cpu_tool_arg_EQ` option definitions
clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp	Updates option processing to use `HasGPUTool` parameter and adds handling for `OPT_device_compiler_args_EQ`
clang/lib/Driver/ToolChains/Clang.cpp	Refactors argument construction to pass backend options via `--device-compiler` instead of removed wrapper options
clang/lib/Driver/Driver.cpp	Adds `GetUseNewOffloadDriverForSYCLOffload` utility function to determine when to use new offload driver
clang/include/clang/Driver/Driver.h	Declares `GetUseNewOffloadDriverForSYCLOffload` function
clang/test/Driver/*.cpp	Updates test expectations to check for new `--device-compiler` argument format

clang/lib/Driver/Driver.cpp

clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp

YixingZhang007 · 2025-12-02T16:26:51Z

5 SYCL E2E tests passed (test names are listed in the PR description, and 3 pm_access_n test were only failing in Self Build CI before), and no SYCL E2E test or driver test regression occurred when the new offloading model was enabled as default together with my changes (the CI before removing new offloading changes from this PR can be found at https://github.com/intel/llvm/actions/runs/19864471834/job/56923211306?pr=20691)

YuriPlyakhin · 2025-12-02T19:49:26Z

clang/lib/Driver/ToolChains/Clang.cpp

-            "--device-linker=" + TC->getTripleString() + "=" + Arg));
+      // SYCL offload toolchains handle device compiler and linker argument forwarding using SYCLImage.
+      // Therefore, we skip forwarding these arguments here when a SYCL offload toolchain is present.
+      if (!C.hasOffloadToolChain<Action::OFK_SYCL>()) {


Is it possible to avoid this SYCL-specific customization? What happens if we keep this for SYCL as well?
If I understand correctly @mdtoguchi : "The expected behavior is for any device options to be handled at both compile and link time. ", we need to keep this code and pass user-specified options to clang-linker-wrapper same as for any other programming model.

Where does CompilerArgs and LinkerArgs come from? I mean, is it from some clang++ arguments?

YuriPlyakhin · 2025-12-02T20:12:24Z

clang/test/Driver/sycl-offload-new-driver.cpp


-/// Test option passing behavior for clang-offload-wrapper options.
-// RUN: %clangxx --target=x86_64-unknown-linux-gnu -fsycl --offload-new-driver \
-// RUN:          -Xsycl-target-backend -backend-opt -### %s 2>&1 \


Per my understanding, we still need to pass -backend-opt to clang-linker-wrapper through device-compiler option.
so this test should be just updated to use proper clang-linker-wrapper option.
Similar for linker options.

Thanks for pointing this out!

I think for backend option and linker option that are passed at compile time, they are stored in the SYCLImage like --image=file=... compile-opts=... -backend-opt. The option that are stored in the SYCLImage got directly extracted in clang-linker-wrapper (this is done is another PR #19579 and extractSYCLCompileLinkOptions function found at https://github.com/YixingZhang007/llvm/blob/fe390ac353c65ec85ccd2b3bf7883a1efca55a5e/clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp#L2218).

Only the -backend-opt that are passed at link time for AOT, which should be through format clang++ ... -fsycl-targets=spir64_gen -Xsycl-target-backend=spir64_gen -backend-gen-opt will added to device-compiler and passed to the clang-linker-wrapper.

Therefore, for this test and the one got deleted below, -backend-opt and -link-opt are passed through the SYCLImage but not directly added as argument for clang-linker-wrapper :)

Yes, I understand how it works with your implementation. And I challenge that. Despite we extracted and put options to SYCLImage on compile stage, we still want to pass these options to link stage as well.
Example of when this is needed is if we provide both source and object files as an input to clang++.

YuriPlyakhin · 2025-12-02T20:17:02Z

clang/test/Driver/sycl-offload.c

-
-// RUN:   %clang -### -target x86_64-unknown-linux-gnu -fsycl --offload-new-driver -fsycl-targets=spir64-unknown-unknown -Xsycl-target-backend "-DFOO1 -DFOO2" %s 2>&1 \
-// RUN:   | FileCheck -check-prefix=CHK-TOOLS-OPTS %s
-// CHK-TOOLS-OPTS: clang-linker-wrapper{{.*}} "--sycl-backend-compile-options=-DFOO1 -DFOO2"


similar here, per my understanding, need to pass options using device-compiler

YuriPlyakhin · 2025-12-02T20:17:37Z

clang/test/Driver/sycl-offload.c


-// RUN:   %clang -### -target x86_64-unknown-linux-gnu -fsycl --offload-new-driver -fsycl-targets=spir64-unknown-unknown -Xsycl-target-linker "-DFOO1 -DFOO2" %s 2>&1 \
-// RUN:   | FileCheck -check-prefix=CHK-TOOLS-OPTS2 %s
-// CHK-TOOLS-OPTS2: clang-linker-wrapper{{.*}} "--sycl-target-link-options=-DFOO1 -DFOO2"


same here, but for device-linker

YuriPlyakhin · 2025-12-02T20:19:20Z

clang/test/Driver/sycl-offload.cpp

-// RUN:    | FileCheck -check-prefixes=SYCL_TARGET_OPT %s
-// RUN:  %clangxx -### -target x86_64-unknown-linux-gnu -fsycl --offload-new-driver -Xsycl-target-backend=spir64 -DFOO -Xsycl-target-linker=spir64 -DFOO2 %S/Inputs/SYCL/objlin64.o 2>&1 \
-// RUN:    | FileCheck -check-prefixes=SYCL_TARGET_OPT %s
-// SYCL_TARGET_OPT: clang-linker-wrapper{{.*}} "--sycl-backend-compile-options=-DFOO" "--sycl-target-link-options=-DFOO2"


same here...

YuriPlyakhin · 2025-12-02T20:44:57Z

clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp

+
+      SmallVector<StringRef, 16> Args;
+      StringRef(CompileLinkOptionsOrErr->first).split(Args, ' ');
+      bool HasGPUTool =


The name of the variable is a bit confusing to me. Would something like this be better?
IsDevicePassedWithSyclTargetBackend. Maybe a little bit verbose, but maybe it is OK.

YuriPlyakhin · 2025-12-02T20:47:31Z

clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp

 static Expected<std::vector<module_split::SplitModule>>
-runSYCLPostLinkTool(ArrayRef<StringRef> InputFiles, const ArgList &Args) {
+runSYCLPostLinkTool(ArrayRef<StringRef> InputFiles, const ArgList &Args,
+                    bool HasGPUTool) {


new parameter needs to be documented in comments above

YuriPlyakhin · 2025-12-02T20:48:25Z

clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp

-  StringRef OptTool = (IsCPU) ? Args.getLastArgValue(OPT_cpu_tool_arg_EQ)
-                              : Args.getLastArgValue(OPT_gpu_tool_arg_EQ);
-  OptTool.split(CmdArgs, " ", /*MaxSplit=*/-1, /*KeepEmpty=*/false);
  return;


this return is also unnecessary, right?

YuriPlyakhin · 2025-12-02T20:52:54Z

clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp

    }
  }
-
-  StringRef OptTool = (IsCPU) ? Args.getLastArgValue(OPT_cpu_tool_arg_EQ)


Args parameter is not used anymore in this function and can be removed.

Or maybe we need to handle OPT_device_compiler_args_EQ here.
I think it depends on how such options are handled for other programming models. We need to try to be as close as possible to upstream...

Thanks for the suggestion!

The reason I didn't move the handling of OPT_device_compiler_args_EQ to here is that -device xxx may be passed as a link-time backend argument, and we need to check for that in linkAndWrapDeviceFiles (found at https://github.com/YixingZhang007/llvm/blob/fe390ac353c65ec85ccd2b3bf7883a1efca55a5e/clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp#L2250) so that we can determine whether we need to add the device name in front of the output table.

Therefore, to avoid processing OPT_device_compiler_args_EQ twice, I add OPT_device_compiler_args_EQ to CompileLinkOptionsOrErr in linkAndWrapDeviceFiles, and CompileLinkOptionsOrErr is later passed as BackendOptions to this function.

ok, now I see that order of options did not change. So, I guess, just remove the Args param as not used.

clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp

YixingZhang007 requested review from a team as code owners November 19, 2025 18:59

YixingZhang007 marked this pull request as draft November 19, 2025 18:59

YixingZhang007 had a problem deploying to WindowsCILock November 19, 2025 19:02 — with GitHub Actions Failure

YixingZhang007 had a problem deploying to WindowsCILock November 19, 2025 19:38 — with GitHub Actions Failure

YixingZhang007 had a problem deploying to WindowsCILock November 20, 2025 03:49 — with GitHub Actions Failure

YixingZhang007 had a problem deploying to WindowsCILock November 20, 2025 04:10 — with GitHub Actions Failure

YixingZhang007 had a problem deploying to WindowsCILock November 20, 2025 04:50 — with GitHub Actions Failure

YixingZhang007 had a problem deploying to WindowsCILock November 25, 2025 15:05 — with GitHub Actions Failure

YixingZhang007 had a problem deploying to WindowsCILock November 25, 2025 15:40 — with GitHub Actions Failure

YixingZhang007 had a problem deploying to WindowsCILock November 25, 2025 19:23 — with GitHub Actions Error

YixingZhang007 had a problem deploying to WindowsCILock November 25, 2025 19:38 — with GitHub Actions Failure

YixingZhang007 had a problem deploying to WindowsCILock November 25, 2025 20:08 — with GitHub Actions Failure

YixingZhang007 had a problem deploying to WindowsCILock November 25, 2025 21:27 — with GitHub Actions Failure

YixingZhang007 had a problem deploying to WindowsCILock November 25, 2025 22:08 — with GitHub Actions Failure

YixingZhang007 force-pushed the fix_linker_unrecognized_option_issue branch from 43ec9a6 to 2c3b7bc Compare November 25, 2025 22:39

YixingZhang007 had a problem deploying to WindowsCILock November 25, 2025 22:39 — with GitHub Actions Failure

YixingZhang007 had a problem deploying to WindowsCILock November 25, 2025 23:37 — with GitHub Actions Failure

YixingZhang007 had a problem deploying to WindowsCILock December 2, 2025 16:13 — with GitHub Actions Failure

YixingZhang007 requested a review from Copilot December 2, 2025 16:21

Copilot AI reviewed Dec 2, 2025

View reviewed changes

clang/lib/Driver/Driver.cpp Outdated Show resolved Hide resolved

clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp Outdated Show resolved Hide resolved

code clean up

f4730f1

YixingZhang007 had a problem deploying to WindowsCILock December 2, 2025 16:39 — with GitHub Actions Error

intel deleted a comment from Copilot AI Dec 2, 2025

code clean up

2f11978

YixingZhang007 had a problem deploying to WindowsCILock December 2, 2025 17:12 — with GitHub Actions Error

code clean up

fe390ac

YixingZhang007 requested a review from maksimsab December 2, 2025 17:18

YixingZhang007 marked this pull request as ready for review December 2, 2025 17:18

YixingZhang007 temporarily deployed to WindowsCILock December 2, 2025 17:18 — with GitHub Actions Inactive

YixingZhang007 requested review from YuriPlyakhin and sarnex December 2, 2025 17:19

YixingZhang007 had a problem deploying to WindowsCILock December 2, 2025 17:52 — with GitHub Actions Failure

YixingZhang007 temporarily deployed to WindowsCILock December 2, 2025 17:52 — with GitHub Actions Inactive

YuriPlyakhin reviewed Dec 2, 2025

View reviewed changes

clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp Show resolved Hide resolved

[NewOffloadModel] Remove compiler backend option and linker option to be passed as argument for ClangLinkerWrapper #20691

Are you sure you want to change the base?

[NewOffloadModel] Remove compiler backend option and linker option to be passed as argument for ClangLinkerWrapper #20691

Conversation

YixingZhang007 commented Nov 19, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

YixingZhang007 commented Dec 2, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

YixingZhang007 Dec 2, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

YixingZhang007 Dec 2, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

YixingZhang007 commented Nov 19, 2025 •

edited

Loading

YixingZhang007 commented Dec 2, 2025 •

edited

Loading

YixingZhang007 Dec 2, 2025 •

edited

Loading

YixingZhang007 Dec 2, 2025 •

edited

Loading