[PATCH] D74035: [AMDGPU] Cluster MIMG instructions

[Jay Foad via Phabricator via llvm-commits]

foad added a comment.

> This absolutely needs a test with sampled images.

I've added one.

> More generally, I'm worried about whether the generic infrastructure will do something useful with the extra address arguments, so this needs quite a bit of test coverage about what happens with different texture coordinate patterns.

The only code that knows about the address operands is SIInstrInfo::getMemOperandsWithOffset (which produces a list of address operands) and SIInstrInfo::shouldClusterMemOps (which consumes it). The logic is currently very simple: cluster if the first address operand (the image descriptor) is the same.

I'm sure the logic could be much cleverer, but I'd like to leave that for later patches with their own test cases. Until that happens, I'm not sure that it's worth producing tests for different texture coordinate patterns.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D74035/new/

https://reviews.llvm.org/D74035