[PATCH] D85681: [AArch64][SVE] Add patterns for integer mla/mls.

[Paul Walker via Phabricator via llvm-commits]

paulwalker-arm added a comment.

I guess it depends on whether MLA hides the cost of the MUL for all implementations.  We do seem to do this transformation is several places (DAG, ISEL and MachineCombiner) presumably there is a good reason for each placement.

Personally I prefer the DAG route because it forces the creation of a _PRED node, which we'll need sooner or later when the pseudo nodes exist[1].  It will also allow us to convert mla intrinsics for the "all active" case plus fixed length code generation can use it.

That all said this is a journey not a destination so I've no objections to the current patch if that's how far you want to take it today.

[1] I'm pretty sure all the mechanics are in place for MLA like pseudo nodes (at least for _PRED, less so for _MERGE_ZERO).


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D85681/new/

https://reviews.llvm.org/D85681