[PATCH] D104870: [SimplifyCFG] Tail-merging all blocks with `unreachable` terminator

[Heejin Ahn via Phabricator via llvm-commits]

aheejin added a comment.

In D104870#2839468 <https://reviews.llvm.org/D104870#2839468>, @rnk wrote:

> This change deserves some discussion. To address the issues with regions, we shouldn't rely on my hazily remembered understandings, we should consult the experts.
>
> - For wasm, let's try @aheejin and @dschuff
> - For GPUs, let's try @nhaehnle and @arsenm
> - WinEH isn't an issue, we future proofed it against this transform
> - For debug info, let's try @aprantl and @jmorse
>
> To those I just added to the review, the question is, will tail merging calls to noreturn functions in IR (BranchFolding already does it in codegen) be an issue for the subsystems you contribute to.

Thanks for letting me know! I guess WinEH is fine because WinEH makes sure to clone <https://github.com/llvm/llvm-project/blob/0646e3625499b08a3ac9efd48396f3b463a19139/llvm/lib/CodeGen/WinEHPrepare.cpp#L744-L947> blocks that belong to multiple funclets, right? I think Wasm should be OK as well as WinEH works, because we use <https://github.com/llvm/llvm-project/blob/79c98279b6cdab6f208043865ab7b4ac4612cec5/llvm/lib/CodeGen/TargetPassConfig.cpp#L925> WinEHPrepare, so we benefit from this transformation as well.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D104870/new/

https://reviews.llvm.org/D104870