[PATCH] D28489: [CodeGen] Generalize MacroFusion for any instructions pair

[Andrew Trick via Phabricator via llvm-commits]

atrick added a comment.

I think you should find a way to do this without calling shouldScheduleAdjacent on every DAG node. It's fine to say you've tested compilation time but what really matters here are the pathological cases with very large blocks. It's rare for instructions in the middle of blocks to have fusion opportunities, so it's wrong to introduce this potential cost for all blocks.


Repository:
  rL LLVM

https://reviews.llvm.org/D28489