[clang-linker-wrapper] Pass on --cuda-path to child clang processes #149107
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
When using clang-linker-wrapper with --cuda-path, it does not get passed on to the child device linking processes. This causes it to fail when cuda linking is involved and nvlink is not in $PATH. This patch lets the child linking process find nvlink through --cuda-path.
Member
|
@llvm/pr-subscribers-clang Author: Ivan R. Ivanov (ivanradanov) ChangesWhen using clang-linker-wrapper with --cuda-path, it does not get passed on to the child device linking processes. This causes it to fail when cuda linking is involved and nvlink is not in $PATH. This patch lets the child linking process find nvlink through --cuda-path. Full diff: https://github.com/llvm/llvm-project/pull/149107.diff 1 Files Affected:
diff --git a/clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp b/clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp
index 9d34b62da20f5..b9e20a6534bf6 100644
--- a/clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp
+++ b/clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp
@@ -567,6 +567,8 @@ Expected<StringRef> clang(ArrayRef<StringRef> InputFiles, const ArgList &Args,
CmdArgs.append({"-Xlinker", Args.MakeArgString(Arg)});
for (StringRef Arg : Args.getAllArgValues(OPT_compiler_arg_EQ))
CmdArgs.push_back(Args.MakeArgString(Arg));
+ for (StringRef Arg : Args.getAllArgValues(OPT_cuda_path_EQ))
+ CmdArgs.push_back(Args.MakeArgString("--cuda-path=" + Arg));
if (Error Err = executeCommands(*ClangPath, CmdArgs))
return std::move(Err);
|
jhuber6
reviewed
Jul 16, 2025
Contributor
jhuber6
left a comment
jhuber6
left a comment
There was a problem hiding this comment.
This is supposed to be handled by the forwarding interface in the Clang toolchain. Doing something like this you should see it.
$ clang input.c -fopenmp --offload-arch=sm_89 --cuda-path=/opt/cuda -###
"clang-linker-wrapper" "--device-compiler=nvptx64-nvidia-cuda=--cuda-path=/opt/cuda"Are you not getting this? That's probably a bug so it'd help to have some examples with how you're invoking this.
Contributor
Author
|
I see. I was doing this which only passes I suppose under the current infra, all the offloading compilation arguments need to be present when linking as well, so as to invoke the appropriate toolchains, and the above usage is not recommended? |
Contributor
|
It's passed to the linker wrapper because it needs |
Contributor
Author
Gives me so the child of |
Contributor
That's odd, can you give me the full output of |
Contributor
Author
|
Contributor
I see, this is because there's no CUDA toolchain that's created in this case. Normally the |
Contributor
Author
|
passing gives us the appropriate flags. That means the cuda toolchain was created, correct? I wonder if we need a step in clang that looks at all the .o files for sections that need device linking and concats the archs, and reinvokes itself with --offload-arch=<all_collected_arches> (although it is clang-linker-wrapper's job to do the parsing of the .o files for that so kind of weird to have clang do it) But then in theory the appropriate toolchains should be created. Perhaps it can only kick in when -foffload-via-llvm is on, but no --offload-archs are specified, i.e. we are asking clang to figure the appropriate offload archs. That step could actually be handled by clang-offload-wrapper - you would get Pretty convoluted so I don't know if it's appropriate. Anyways, @jhuber6 thank you for taking a look, I think I will close this for now. |
Contributor
|
Yeah that's definitely not ideal, I'll need to think of a way to handle that more gracefully. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
When using clang-linker-wrapper with --cuda-path, it does not get passed on to the child device linking processes. This causes it to fail when cuda linking is involved and nvlink is not in $PATH. This patch lets the child linking process find nvlink through --cuda-path.