[Mlir-commits] [mlir] [MLIR][XeGPU] Rematerialize layout conversion sources after conflict resolution (PR #230669)

[Jianhui Li]

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

----------------
Jianhui-Li wrote:

//To fix the operand/result mismatch issue similar to the issue Frank filed for ...
Can you explain what is the issue here? 
I don't think it is a clean solution mixing rematerializeWithLayout() and sinkElementwiseConversions() together. They play different roles: one is about sink a chain of elementwise op through conversion (single user but has larger inst_data/lane_data layout), the other is to copy these elementwise ops (two users has different layout requirement).  

// Sink would still have to check that those operands are cheap before deciding to sink
Is this the reason that this function call isTriviallyRematerializable(definingOp)? Maybe worth another name. 


https://github.com/llvm/llvm-project/pull/230669