[PATCH] D48404: Don't modify LibFuncs in DeadArgumentElimination or ArgumentPromotion

[Eli Friedman via Phabricator via llvm-commits]

efriedma added a comment.

Maybe instead of trying to turn off optimizations on internal builtin functions, we should instead try to make optimizations not treat functions with local linkage as builtin?  Not sure.  This came up before on llvmdev, but we didn't really decide one way or the other.

If you're going to go with this approach, we probably also need to fix a bunch of other places that call hasLocalLinkage: llvm::canTrackArgumentsInterprocedurally is used by IPSCCP, GlobalOpt will change the calling convention of function definitions, etc.

Regardless of which approach we choose, DSE should be fixed to use getLibFunc() so it doesn't crash on invalid input.


Repository:
  rL LLVM

https://reviews.llvm.org/D48404