[PATCH] D117140: [LV] Always create VPWidenCanonicalIVRecipe, optimize away later.

[Ayal Zaks via Phabricator via llvm-commits]

Ayal accepted this revision.
Ayal added a comment.
This revision is now accepted and ready to land.

In D117140#3253999 <https://reviews.llvm.org/D117140#3253999>, @fhahn wrote:

> Remove unnecessary early exit.
>
> In D117140#3251403 <https://reviews.llvm.org/D117140#3251403>, @Ayal wrote:
>
>> In D117140#3251026 <https://reviews.llvm.org/D117140#3251026>, @fhahn wrote:
>>
>>> In D117140#3250984 <https://reviews.llvm.org/D117140#3250984>, @Ayal wrote:
>>>
>>>> This is fine, thanks for accommodating, one last thought: does Canonical mean starting at zero, or possibly at EPResumeVal?
>>>
>>> I think checking for zero a the moment should suffice, as EPResumeVal will only be set as start value just before epilogue VPlan execution. This should also match the current behavior before the change (and  D117551 <https://reviews.llvm.org/D117551> which now contains the addition if `isCanonical`). Once a single plan contains both vector loops we might need some changes in that direction though.
>>
>> Resetting start of canonicalIV to EPResumeVal just before epiloque VPlan execution might invalidate passes that relied on this value being zero, including removeRedundantCanonicalIVs()? A test would be good, either confirming or reassuring :-)
>
> Added an assert in 165e36bf180e <https://reviews.llvm.org/rG165e36bf180eecb32fa5442fd5c4d43388a136fb>.

Thanks!
For epilog VPlan execution, the start of all IV's should be bumped according to EPResumeVal; perhaps better to model CanonicalIVStartValue as an abstract VPValue of VPlan similar to TripCount and BackedgeTakenCount, w/o claiming it is zero prematurely.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D117140/new/

https://reviews.llvm.org/D117140