[SYCL][NewOffloadModel] Support -fno-sycl-rdc at compile step (-c) by srividya-sundaram · Pull Request #22833 · intel/llvm

srividya-sundaram · 2026-07-30T21:44:15Z

Summary

This PR adds support for -fno-sycl-rdc -c in the new offload model by finalizing each TU's device code at compile time via clang-linker-wrapper --sycl-device-link --no-sycl-rdc, then embedding the result into the host object via a new -fsycl-include-target-binary CC1 option. The approach mirrors how CUDA/HIP handle no-RDC with -fcuda-include-gpubinary.

Depends on #22832.

…floadModel.

Fix --no-offload-old-driver typo to --no-offload-new-driver in the old offload model usage example. Add a note that compile-step support for -fno-sycl-rdc (matching the old model's -c usage pattern) will be implemented in a follow-up patch via clang-linker-wrapper --sycl-device-link --no-sycl-rdc per translation unit.

Adds support for specifying -fno-sycl-rdc at the compile step in the new offload model, matching the old offload model's usage pattern. Previously, -fno-sycl-rdc was only meaningful at the link step in the new model. Design (mirrors CUDA/HIP -fno-gpu-rdc per-TU finalize model): - BuildOffloadingActions: when -fno-sycl-rdc is passed at compile time, insert a per-TU LinkerWrapperJobAction (OffloadPackager + clang-linker-wrapper --sycl-device-link --no-sycl-rdc) to finalize the TU's device code immediately, producing a self-contained device image. - Clang::ConstructJob: route the finalized image to the new -fsycl-include-target-binary CC1 option instead of -fembed-offload-object, so the host CodeGen embeds and registers the image at compile time rather than deferring it to link time. - CodeGenModule::Release: when -fsycl-include-target-binary is set on the host cc1, read the device image file and call wrapSYCLBinaries to embed the binary and emit __sycl_register_lib/__sycl_unregister_lib ctors into the host module. The final clang-linker-wrapper invocation then finds no SYCL device input and performs a plain host link. - Downstream uses clang-linker-wrapper --sycl-device-link instead of the upstream's clang-sycl-linker, as all Intel SYCL pipeline logic resides in clang-linker-wrapper. New files/options: - -fsycl-include-target-binary <file>: new CC1 option (analog of -fcuda-include-gpubinary) marshalled into CodeGenOpts.SYCLTargetBinaryFileName - clang/test/Driver/sycl-no-rdc-compile-step.cpp: driver tests verifying per-TU pipeline shape, RDC default regression, and link-step regression. Implements compile-step support tracked in CMPLRLLVM-51875. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Copilot

Pull request overview

Adds compile-step (-c) support for -fno-sycl-rdc in the new SYCL offload model by finalizing each TU’s device image during compilation (via a per-TU clang-linker-wrapper --sycl-device-link --no-sycl-rdc) and embedding/registering the finalized image into the host object using a new cc1 option.

Changes:

Teach the driver to insert a per-TU LinkerWrapperJobAction for SYCL NoRDC compilation and route its output to the host compile.
Introduce -fsycl-include-target-binary and plumb it through driver → CodeGen to embed/register the finalized SYCL image at compile time.
Add/update driver + e2e tests and update design documentation around NoRDC usage patterns.

Reviewed changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
sycl/test-e2e/AOT/early_aot.cpp	Adjust RUN lines for old/new model flag placement; adds explanatory note.
sycl/doc/design/OffloadDesign.md	Document NoRDC behavior and usage scenarios for old/new offload models.
sycl/doc/design/NonRelocatableDeviceCode.md	Clarify document scope (old offload model NoRDC design).
clang/test/Driver/sycl-no-rdc-new-driver.cpp	New driver test for `-fno-sycl-rdc` propagation and pipeline shape.
clang/test/Driver/sycl-no-rdc-compile-step.cpp	New driver test covering compile-step NoRDC finalization + cc1 plumbing.
clang/lib/Driver/ToolChains/Clang.cpp	Pass finalized SYCL image to host cc1; propagate NoRDC to linker wrapper and add per-TU device-link mode logic.
clang/lib/Driver/Driver.cpp	Insert per-TU linker-wrapper finalizer action for SYCL NoRDC during compilation.
clang/lib/CodeGen/CodeGenModule.cpp	Read finalized SYCL image and call `wrapSYCLBinaries` to embed/register at compile time.
clang/include/clang/Options/Options.td	Add new `-fsycl-include-target-binary` option (marshalled to CodeGenOpts).
clang/include/clang/Basic/CodeGenOptions.h	Add `SYCLTargetBinaryFileName` storage for the new cc1 option.

Copilot

Pull request overview

Copilot reviewed 10 out of 10 changed files in this pull request and generated no new comments.

srividya-sundaram · 2026-07-31T17:54:54Z

SYCL Pre commit failure is unrelated to the changes in this PR.

srividya-sundaram · 2026-07-31T18:59:01Z

Please note that this PR is a follow-up to #22832 and currently includes its changes as a base. Once #22832 lands, this branch will be rebased on top of it so the diff only shows the compile-step additions.

@maksimsab

…Offload Model (#22832) ## Summary This patch adds `-fno-sycl-rdc` support to the new SYCL offload model in the Clang driver. It is based on the work by @maksimsab in #21973, with review fixes applied. ### What this patch does By default (`-fsycl-rdc`), all device code across translation units is linked together into one module at link time before post-link processing. With `-fno-sycl-rdc`, each object file's device code is processed independently — skipping the `llvm-link` step — which can significantly reduce peak memory and compile time. **Changes:** - `clang/lib/Driver/ToolChains/Clang.cpp`: In `LinkerWrapper::ConstructJob`, propagate `--no-sycl-rdc` to `clang-linker-wrapper` when `-fno-sycl-rdc` is passed at link time. - `sycl/doc/design/OffloadDesign.md`: Document NoRDC mode for the new offload model. - `sycl/doc/design/NonRelocatableDeviceCode.md`: Add a note clarifying that document covers the old offload model only. - `sycl/test-e2e/AOT/early_aot.cpp`: Un-XFAIL the test for the new offload model; adjust RUN lines to pass `-fno-sycl-rdc` at the correct step per model. - `clang/test/Driver/sycl-no-rdc-new-driver.cpp`: New driver test verifying `--no-sycl-rdc` is propagated to `clang-linker-wrapper` correctly. ### Key behavioral difference from the old offload model | Model | Where to pass `-fno-sycl-rdc` | |---|---| | Old (`--no-offload-new-driver`) | At each **compile** step (`-c`) | | New (`--offload-new-driver`) | At the **link** step | This inversion exists because in the new model all SYCL offload processing (post-link, AOT, wrapping) lives in `clang-linker-wrapper` at link time. ### Follow-up work Supporting `-fno-sycl-rdc -c` in the new model (matching the old model's compile-step ergonomics) will be addressed in a follow-up patch. The plan is to invoke `clang-linker-wrapper --sycl-device-link --no-sycl-rdc` per translation unit at compile time, embedding the finalized device image directly into the host object. Supporting -fno-sycl-rdc -c in the new model (matching old model compile-step ergonomics) is addressed in #22833. Fixes: CMPLRLLVM-51875 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Sabianin, Maksim <maksim.sabianin@intel.com> Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

Move embedSYCLNoRDCBinary out of CodeGenModule::Release() into a free function in CGSYCLRuntime.cpp to keep SYCL-specific logic out of common code. CodeGenModule::Release() retains a single call site. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…r wrapper

…p test

KseniyaTikhomirova

SYCL RT changes LGTM

…ile, remove fsycl-include-target-binary alias, use proper DK_Linker forwarding handler

tahonermann

I'm not yet convinced that the diagnostic workaround is needed. Please see the inline comment.

…fix embedSYCLNoRDCBinary ordering, update OffloadDesign.md wording

…rupted module after no-RDC bitcode link failure

Copilot

Pull request overview

Copilot reviewed 22 out of 22 changed files in this pull request and generated 1 comment.

Suppressed comments (1)

sycl/test-e2e/AOT/early_aot.cpp:8

The "Old offload model" RUN lines don’t explicitly force the old model (e.g. via --no-offload-new-driver). In configurations where the new offload model is enabled by default, these commands would exercise the new model instead, reducing coverage and making the section header misleading.

// Old offload model: -fno-sycl-rdc at compile step only.
// RUN: %clangxx -fsycl -fsycl-targets=spir64_gen -Xsycl-target-backend=spir64_gen %gpu_aot_target_opts -fno-sycl-rdc -c -DADD_CPP %s -o %t_add.o
// RUN: %clangxx -fsycl -fsycl-targets=spir64_gen -Xsycl-target-backend=spir64_gen %gpu_aot_target_opts -fno-sycl-rdc -c -DSUB_CPP %s -o %t_sub.o
// RUN: %clangxx -fsycl -DMAIN_CPP %s %t_add.o %t_sub.o -o %t.out

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…mbedSYCLNoRDCBinary

Copilot

Pull request overview

Copilot reviewed 21 out of 21 changed files in this pull request and generated 3 comments.

Suppressed comments (3)

clang/include/clang/Options/Options.td:1931

The PR description specifies a new -fsycl-include-target-binary CC1 option, but this change defines only -foffload-include-binary (and aliases the CUDA spelling to it). As written, the promised SYCL option is not available; either implement the described spelling or update the PR contract and all documentation/tests consistently.

def foffload_include_binary : Separate<["-"], "foffload-include-binary">,
  Visibility<[CC1Option]>,
  HelpText<"Incorporate offloading device-side binary into host object file.">,
  MarshallingInfoString<CodeGenOpts<"OffloadBinaryToEmbedFile">>;

clang/lib/Driver/Driver.cpp:8390

The !isa<LinkJobAction> condition also selects preprocessing and syntax-only host actions, even though this finalizer is intended for compile steps that produce a host object/IR. For -fsycl -fno-sycl-rdc -E without -o (the special preprocessing branch above only handles an explicit output), this adds an image linker-wrapper to a preprocessing-only pipeline whose device actions have no linkable image. Restrict this branch to host actions with a real compilable output and leave preprocessing/syntax-only handling on their existing paths.

  } else if (C.isOffloadingHostKind(Action::OFK_SYCL) &&
             tools::SYCL::shouldDoPerObjectFileLinking(C) &&
             !isa<LinkJobAction>(HostAction)) {

clang/lib/Driver/ToolChains/Clang.cpp:12165

getFinalPhase describes the whole driver invocation, not this nested per-TU action. For a compile-and-link command such as clang++ --offload-new-driver -fsycl -fno-sycl-rdc a.cpp -o a.out, the action graph still creates this OFK_SYCL finalizer (as the new pipeline check demonstrates), but the final phase is Link, so both flags are suppressed. The wrapper then takes the normal runWrapperAndCompile path and returns an object, while the host cc1 is given -foffload-include-binary and tries to parse that file as wrapper bitcode. Forward the flags based on the per-TU action rather than the global final phase, and update the link-step test to cover this distinction.

      if (C.getDriver().getFinalPhase(C.getArgs()) != phases::Link &&
          JA.isDeviceOffloading(Action::OFK_SYCL)) {

…-step embedding

…adLinkModules

tahonermann

Thank you for working with me through the link module control flow changes this week @srividya-sundaram! This looks much cleaner to me now.

I added a couple of comments to note minor fixes.

Just something I want to note to be aware of. With these changes, the no-rdc SYCL device image (if specified) will be linked in everywhere loadLinkModules() is called (which currently includes CodeGenAction::CreateASTConsumer(), CodeGenAction::loadModule(), and CIRGenAction::BeginSourceFileAction()). I think that is a good thing, but it isn't how the modules returned from getBitcodeModuleList() (a list populated by, I think, -mlink-bitcode-file) are handled, so there might be some subtle reason to do things differently. I don't recommend making a change unless/until there is a demonstrable reason to do so.

It looks like there might be some interactions with -flto=thin that might warrant some additional tests. In CodeGenAction::loadModule(), if CI.getCodeGenOpts().ThinLTOIndexFile.empty() is false, the loadLinkModules() call won't be reached. I don't know if that is a good thing or otherwise.

srividya-sundaram · 2026-08-21T21:45:50Z

Thank you for the helpful offline discussions, @tahonermann!

Just something I want to note to be aware of. With these changes, the no-rdc SYCL device image (if specified) will be linked in everywhere loadLinkModules() is called (which currently includes CodeGenAction::CreateASTConsumer(), CodeGenAction::loadModule(), and CIRGenAction::BeginSourceFileAction()). I think that is a good thing, but it isn't how the modules returned from getBitcodeModuleList() (a list populated by, I think, -mlink-bitcode-file) are handled, so there might be some subtle reason to do things differently. I don't recommend making a change unless/until there is a demonstrable reason to do so.

It appears that loadModule is called only when the input is IR
In the normal SYCL no-RDC compile flow the input is C++ , so loadModule never gets called.
So this is probably not relevant/an issue with our SYCL no-rdc-flow

tahonermann

Thanks, @srividya-sundaram, the updates look great!

maksimsab and others added 6 commits July 30, 2026 13:44

[SYCL][NewOffloadModel] Enable -no-sycl-rdc in Clang Driver for NewOf…

90fe7b6

…floadModel.

fix test for cross-platform execution

998d20f

[SYCL][Driver] Fix tests for -fno-sycl-rdc compile-step support

07a16e7

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

[SYCL] Fix clang-format issues

2a41c6f

srividya-sundaram requested a review from Copilot July 30, 2026 22:34

Copilot started reviewing on behalf of srividya-sundaram July 30, 2026 22:35 View session

Copilot AI reviewed Jul 30, 2026

View reviewed changes

Comment thread clang/lib/CodeGen/CodeGenModule.cpp Outdated

Comment thread clang/lib/Driver/ToolChains/Clang.cpp Outdated

Comment thread clang/lib/CodeGen/CodeGenModule.cpp Outdated

Comment thread sycl/doc/design/OffloadDesign.md Outdated

Comment thread sycl/test-e2e/AOT/early_aot.cpp Outdated

[SYCL] Address Copilot review comments on PR intel#22833

d8ece47

srividya-sundaram requested a review from Copilot July 31, 2026 01:46

Copilot started reviewing on behalf of srividya-sundaram July 31, 2026 01:47 View session

Copilot AI reviewed Jul 31, 2026

View reviewed changes

srividya-sundaram mentioned this pull request Jul 31, 2026

[SYCL][NewOffloadModel] Enable -fno-sycl-rdc in Clang Driver for New Offload Model #22832

Merged

srividya-sundaram marked this pull request as ready for review July 31, 2026 18:59

srividya-sundaram requested review from a team as code owners July 31, 2026 18:59

srividya-sundaram requested a review from KseniyaTikhomirova July 31, 2026 18:59

sarnex reviewed Jul 31, 2026

View reviewed changes

mdtoguchi reviewed Jul 31, 2026

View reviewed changes

Comment thread clang/test/Driver/sycl-no-rdc-compile-step.cpp Outdated

srividya-sundaram and others added 5 commits July 31, 2026 16:36

[SYCL][Driver] Use getFinalPhase to guard --sycl-device-link in linke…

7a9c782

…r wrapper

[SYCL][Driver] Trim verbose comment in SYCL no-RDC host offload block

a45367c

[SYCL][Driver] Add -fno-sycl-rdc -flto tests to sycl-no-rdc-new-driver

a2bc665

[SYCL][Driver] Trim verbose comment blocks in sycl-no-rdc-compile-ste…

0e6ce2e

…p test

tahonermann reviewed Aug 10, 2026

View reviewed changes

Comment thread clang/lib/CodeGen/CGSYCLRuntime.cpp Outdated

KseniyaTikhomirova approved these changes Aug 14, 2026

View reviewed changes

YuriPlyakhin reviewed Aug 14, 2026

View reviewed changes

Comment thread clang/lib/CodeGen/CodeGenModule.cpp Outdated

srividya-sundaram added 4 commits August 16, 2026 16:27

Merge branch 'sycl' into sycl-no-rdc-compile-step

d244f46

[SYCL][CodeGen] Address review feedback: rename OffloadBinaryToEmbedF…

1cefcf3

…ile, remove fsycl-include-target-binary alias, use proper DK_Linker forwarding handler

Merge branch 'sycl' into sycl-no-rdc-compile-step

f5843b2

Fix llvm::make_scope_exit error.

69230b9

tahonermann reviewed Aug 17, 2026

View reviewed changes

Comment thread clang/lib/CodeGen/CGSYCLRuntime.cpp Outdated

YuriPlyakhin reviewed Aug 18, 2026

View reviewed changes

Comment thread clang/lib/CodeGen/CodeGenModule.cpp Outdated

YuriPlyakhin reviewed Aug 18, 2026

View reviewed changes

Comment thread sycl/doc/design/OffloadDesign.md Outdated

YuriPlyakhin reviewed Aug 18, 2026

View reviewed changes

Comment thread sycl/doc/design/OffloadDesign.md Outdated

srividya-sundaram added 2 commits August 18, 2026 19:27

[SYCL][CodeGen] Address review feedback: align Options.td placement, …

ccfe16a

…fix embedSYCLNoRDCBinary ordering, update OffloadDesign.md wording

[SYCL] Ignore -fno-sycl-rdc at link step; guard Release() against cor…

251023d

…rupted module after no-RDC bitcode link failure

srividya-sundaram requested a lite review from Copilot August 19, 2026 17:31

Copilot started reviewing on behalf of srividya-sundaram August 19, 2026 17:32 View session

Copilot AI reviewed Aug 19, 2026

View reviewed changes

Comment thread clang/lib/Driver/ToolChains/Clang.cpp

srividya-sundaram and others added 2 commits August 20, 2026 11:26

[SYCL] Fix misleading comment in embedSYCLNoRDCBinary

f9aa9e1

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

[SYCL] Move no-RDC wrapper .bc linking into loadLinkModules; remove e…

5f2697e

…mbedSYCLNoRDCBinary

srividya-sundaram requested a lite review from Copilot August 20, 2026 22:56

Copilot started reviewing on behalf of srividya-sundaram August 20, 2026 22:56 View session

Copilot AI reviewed Aug 20, 2026

View reviewed changes

Comment thread clang/lib/CodeGen/ModuleLinker.cpp Outdated

Comment thread clang/lib/Driver/ToolChains/Clang.cpp

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

srividya-sundaram added 3 commits August 20, 2026 16:17

[SYCL] Add loadLinkModules error path tests for -fno-sycl-rdc compile…

a3ffeb7

…-step embedding

[SYCL] Exclude SYCL-CUDA host mode

06dc495

[SYCL] Fix data layout mismatch when linking no-RDC wrapper .bc in lo…

b51d41c

…adLinkModules

tahonermann requested changes Aug 21, 2026

View reviewed changes

Comment thread clang/lib/CodeGen/ModuleLinker.cpp

Comment thread clang/lib/CodeGen/ModuleLinker.cpp

Comment thread clang/lib/CodeGen/ModuleLinker.cpp Outdated

srividya-sundaram added 2 commits August 21, 2026 11:56

Address review comments.

b74bd4a

Fix clang format.

515d6d9

srividya-sundaram requested a review from YuriPlyakhin August 21, 2026 21:47

tahonermann approved these changes Aug 21, 2026

View reviewed changes

Uh oh!

Conversation

srividya-sundaram commented Jul 30, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

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!

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

srividya-sundaram commented Jul 31, 2026

Uh oh!

srividya-sundaram commented Jul 31, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

KseniyaTikhomirova left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

tahonermann left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Uh oh!

Uh oh!

tahonermann left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

srividya-sundaram commented Aug 21, 2026

Uh oh!

tahonermann left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

8 participants

srividya-sundaram commented Jul 30, 2026 •

edited

Loading