[PATCH] D88823: [CodeGen][TailDuplicator] Don't duplicate blocks with INLINEASM_BR

[James Y Knight via Phabricator via llvm-commits]

jyknight added a comment.

> Okay. Abandoning this change.

Sorry, maybe I buried the lede too far into the message? "I think the fix you have here is OK" was at the end. :)

> One thing, I believe we're trying to turn "asm goto" into something it was never meant to be. If you look at gcc's output for this code.

Certainly, as discussed on the email thread with Segher, Clang/LLVM has interpreted a restriction on the semantics for asm-goto which GCC does _not_ have. LLVM treats asm-goto similar to the block-label and indirect goto features -- and thus requires that a critical edge to the indirect target cannot be split by making a new target block. GCC does not have such a restriction. We can discuss whether we should modify the semantics of asm-goto in Clang, but not part of this bugfix.

> This side steps the issue of resolving PHI nodes.

Sure. But, in LLVM, PHIs also appear in a landingpad, which has the exact same behavior. "findPHICopyInsertPoint" being able to place a COPY earlier in a block is not something brand-new for asm-goto. The PHI code is intended to be able to handle this placement, and does so correctly, in this example at least.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D88823/new/

https://reviews.llvm.org/D88823