[PATCH] D93370: [PowerPC] Add new infrastructure to select load/store instructions, update P8/P9 load/store patterns.

[Amy Kwan via Phabricator via llvm-commits]

amyk added a comment.

In D93370#2459137 <https://reviews.llvm.org/D93370#2459137>, @shchenz wrote:

> I think it is better to explicitly give some reasons why we need this big refactoring, in other words, what's the disadvantage/limitation of legacy implementation? Thank you for the big effort.

That is a good point. Thank you for bringing this up. The primary reason behind this refactoring is that it would allow us to easily exploit prefixed load and stores when the offset is large (such as for immediates that fit within 34-bits).

The current implementation of selecting load/stores is very dependent on the ordering of patterns within TableGen, and we aren't able to easily and reliably generated the prefixed load/stores (also since those patterns would also be written in a different file). 
Essentially, the refactoring of how the loads/stores are handled is meant to provide us with more control over the patterns/different forms to exploit. Additionally, we want to eliminate the dependency of pattern declaration ordering as much as possible so regardless of the order of the patterns that the compiler tries to match, we can generate the best possible code.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D93370/new/

https://reviews.llvm.org/D93370