CopyToDma: Fold CastOp and ExpandShapeOp into air::DmaMemcpyOp

[erwei-xilinx]

No description provided.

[fifield]

It seems like condenseMemrefDataReorderingToAIRDma is a canonicalization pattern? Should it be implemented as one so that it runs whenever the canonicalilzer runs and not just during this pass? It might also simplify the code to consider each case as a simple canonicalization pattern, instead of as nested if's trying to do the whole thing at once.

[erwei-xilinx]

It seems like condenseMemrefDataReorderingToAIRDma is a canonicalization pattern? Should it be implemented as one so that it runs whenever the canonicalilzer runs and not just during this pass? It might also simplify the code to consider each case as a simple canonicalization pattern, instead of as nested if's trying to do the whole thing at once.

Totally agreed. The lack of such canonicalizer has been creating issues for air.dma_memcpy_nd and air.channel.put/get, as subsequent passes can no longer handle those memref ops. I updated the issue #707 to keep track of this.