[PATCH] D73480: [VectorCombine] new IR transform pass for partial vector ops

[Sanjay Patel via Phabricator via llvm-commits]

spatel added a comment.

In D73480#1850467 <https://reviews.llvm.org/D73480#1850467>, @efriedma wrote:

> I think the general idea of a pass that uses vector cost models to do peephole optimizations on vectors makes sense.  A couple general concerns:
>
> 1. If we're running it in the middle of the pipeline, we need to be careful the transforms don't conflict with instcombine.


AFAIK, the closest we've come to overlapping/conflicting with vector<->scalar transforms is D50992 <https://reviews.llvm.org/D50992>. I'd be happy to abandon that in favor of a cost-aware transform here. Similarly, we could move some questionable shuffle transforms and/or vector demanded elements analysis out of InstCombine to live here. That would likely have a side benefit of making the optimizer slightly faster overall since we don't need to run this pass as often as InstCombine.

> 2. The vectorizer's cost model is optimized for throughput; I'm a little worried we'll run into issues with latency if we depend too much on the cost model for scalar code.

We do have: TargetTransformInfo::getInstructionLatency(const Instruction *I)
...but I don't see it used/overridden anywhere in tree currently.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D73480/new/

https://reviews.llvm.org/D73480