[PATCH] D150526: [X86] Add X86FixupVectorConstantsPass to re-fold AVX512 vector load folds as broadcast folds

[Simon Pilgrim via Phabricator via llvm-commits]

RKSimon added a comment.

> Will it increase compile time much?

Iterating the the basic blocks is very cheap, the key is to avoid extracting the constant raw bits data too frequently - I think I've addressed this by pulling out the getConstantFromPool call, but only trying to find a splat (and generate the new constant) if we have a suitable opcode.

The cost of the lookupBroadcastFoldTable calls are reduced by sorting the tables on first use (as we already do for the other folding tables) and using a lower_bound search.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D150526/new/

https://reviews.llvm.org/D150526