[PATCH] D49347: [WIP][DebugInfo][LV] DebugLoc in induction PHI

[Vedant Kumar via Phabricator via llvm-commits]

vsk added a comment.

In https://reviews.llvm.org/D49347#1168565, @gramanas wrote:

> > Can there be multiple induction phis in the original loop's header? If so, why is picking the location from the first phi correct?
>
> I am not sure. I don't think there can be many induction phis in a loop that hasn't been vectorized yet.


Why not? The impression I get from skimming IndVarSimplify is that any phi in the loop header may be an induction variable.

> While creating the vectorized loop skeleton, `createInductuonVariable()` is called once, so I thought the first phi
>  found will probably be the only one as well.

It doesn't hurt to try and verify this. Why not add an assertion that there are no more phis -- since if they are the DebugLoc selection would be suspect -- and test it in a stage2 build?


Repository:
  rL LLVM

https://reviews.llvm.org/D49347