-
Notifications
You must be signed in to change notification settings - Fork 798
[SYCL] Fix linkage adjustment of kernels #19771
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.
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
[SYCL] Fix linkage adjustment of kernels #19771
Conversation
This patch resolves intel#19409 Two main changes were done: 1. Corrected method of detection that kernel is defined as an inline function. The original code detecting this was introduced in intel#338, but by some reason we looked into a headers-provided wrapper that calls a kernel instead of a kernel itself. Alongside with fixing this, a dedicated clang-level test was added to check this behavior 2. Fixed a bug introduced by incorrect conflict resolution (commit fcd95a9) with llvm/llvm-project#137882. Specifically, when deciding if a function (that is expected to be a SYCL kernel) linkage should be promoted to non-discardable we used to check for OpenCL kernel attribute. That attribute is not present in SYCL headers and therefore was only generated for actual SYCL kernels emitted by front-end. However, with kernel attributes unified, we now promote linkage of non-kernel wrapper function that has `sycl_kernel` attribute attached to it explicitly by SYCL headers. The fix here is to only react to the kernel attribute if it was added implicitly, to preserve original behavior.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
A couple of NITs, otherwise LGTM. What I don't really get is how that worked before. IMO linkage of sycl_kernel attributed function should not really matter.
|
…kernel-linkage-adjustment
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
ESIMD + FE LGTM but will leave actual FE review to Mariya :)
@@ -12824,7 +12824,8 @@ static GVALinkage adjustGVALinkageForAttributes(const ASTContext &Context, | |||
if (Context.shouldExternalize(D)) | |||
return GVA_StrongExternal; | |||
} else if (Context.getLangOpts().SYCLIsDevice && | |||
D->hasAttr<DeviceKernelAttr>()) { | |||
(D->hasAttr<DeviceKernelAttr>() && |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Lol sorry about all the problems caused from unifying the attributes, I originally just wanted to add a new attribute for SPIR kernels but upstream suggested I unify them all, thanks for fixing this
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
To be fair, this whole thing with fixing-up the linkage is also a little bit weird, so no worries :)
…kernel-linkage-adjustment
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Joint Matrix changes LGTM
This patch resolves #19409
Two main changes were done:
Corrected method of detection that kernel is defined as an inline function. The original code detecting this was introduced in [SYCL] Allow an inlined kernel to be called from multiple TUs. #338, but by some reason we looked into a headers-provided wrapper that calls a kernel instead of a kernel itself. Alongside with fixing this, a dedicated clang-level test was added to check this behavior
Fixed a bug introduced by incorrect conflict resolution (commit fcd95a9) with [clang] Simplify device kernel attributes llvm/llvm-project#137882. Specifically, when deciding if a function (that is expected to be a SYCL kernel) linkage should be promoted to non-discardable we used to check for OpenCL kernel attribute. That attribute is not present in SYCL headers and therefore was only generated for actual SYCL kernels emitted by front-end. However, with kernel attributes unified, we now promote linkage of non-kernel wrapper function that has
sycl_kernel
attribute attached to it explicitly by SYCL headers. The fix here is to only react to the kernel attribute if it was added implicitly, to preserve original behavior.