[Mlir-commits] [mlir] [mlir][affine] Handle generic memory ops in loop fusion dependence analysis (PR #224657)

[llvmlistbot at llvm.org]

================

----------------
lipracer wrote:

Minor / non-blocking: mayAccessMemRef uses getEffectedValues, which skips effects without a value. Ops with a value-less effect (e.g. gpu.printf's bare MemoryEffects<[MemWrite]>) are thus treated as accessing nothing. Operand-bound effects (MemWriteAt/MemReadAt) are fine, but a value-less read/write aliasing relevant memory would be under-approximated. Couldn't find an in-tree op that's both value-less and touches a dependent memref, so this is theoretical — flagging for someone familiar with the effect model.

https://github.com/llvm/llvm-project/pull/224657