[SYCL] Move SYCLLowerWGLocalMemoryPass to OptimizerEarlyEPCallback #16347
Conversation
The pass transforms __sycl_allocateLocalMemory call to access of global variable @WGLocalMem . Move the transform to beginning of the optimizer since the access could enable more optimization than the function call. In addition, intel gpu compiler has a pass to transform global variable in addrspace(3) to alloca, we must run SYCLLowerWGLocalMemoryPass before it.
|
Similar to #16183 but tweak the position of the pass to satisfy the requirements. |
|
@jsji, is this PR ready for review? Thanks for putting the motivation for the change to the description. It sounds reasonable except this part:
DPC++ runs all passes before the translation to SPIR-V and IGC runs its passes after the translation to SPIR-V, so existing location satisfies the requirement already. |
Update comments. Co-authored-by: Victor Lomuller <[email protected]>
There was a problem hiding this comment.
FE change looks ok to me but I do not know the functionality well enough to approve the new pass location. I defer to @bader for approval
| @@ -1043,7 +1043,14 @@ void EmitAssemblyHelper::RunOptimizationPipeline( | |||
| /*ExcludeAspects=*/{"fp64"})); | |||
| MPM.addPass(SYCLPropagateJointMatrixUsagePass()); | |||
| }); | |||
| else if (LangOpts.SYCLIsHost && !LangOpts.SYCLESIMDBuildHostCode) | |||
| PB.registerOptimizerEarlyEPCallback([](ModulePassManager &MPM, | |||
There was a problem hiding this comment.
What is the reason to use registerOptimizerEarlyEPCallback instead of registerPipelineStartEPCallback?
There was a problem hiding this comment.
As mentioned in https://github.com/intel/llvm/pull/16183/files/43526397fa2b4c559666fdfd6f302bfba7409255#r1857546652, we need to make sure this pass run after AlwaysIniner, moving to PipelineStart is what was tried in #16183, it is before alwaysinliner and too early.
There was a problem hiding this comment.
@jsji, could you please extend the comment at lines 1049-1050 with the requirement to run AlwaysInliner pass before SYCLLowerWGLocalMemoryPass?
|
If we are sure that the pass needs to be moved to an earlier stage, I had an idea that we can remove the pass's implicit dependency on alwaysinliner pass and move the pass to pipeline start: wenju-he@e6ce584 |
Thanks @wenju-he ! Yes, looks good idea to handle the inlining of these two functions inside the pass itself. Please post your PR for review instead. |
|
|
Close, use #16356 instead. |
The pass transforms __sycl_allocateLocalMemory call to access of global
variable @WGLocalMem .
Move the transform to beginning of the optimizer since the access
could enable more optimization than the function call.