[Mlir-commits] [mlir] [mlir] Fix arith tensor constants (PR #206249)

[llvmlistbot at llvm.org]

mygitljf wrote:

@IgWod , @joker-eph Thanks for your review and time.

I updated this to keep the tensor-specific materialization contained in `arith-expand`.

The previous version was too broad: it made ArithToLLVM and the SPIR-V test pass load/depend on the tensor dialect, and the added conversion tests were mostly showing that those pipelines could preserve dynamic tensor arith ops rather than actually lowering them. That made the boundary confusing.

In the new version, `populateCeilFloorDivExpandOpsPatterns` keeps the old default behavior. Only `populateArithExpandOpsPatterns` opts into dynamic tensor constant materialization, where dynamic ranked tensor constants are created with `tensor.splat` using the source tensor dimensions. 

I also removed the ArithToLLVM tensor dependency, the SPIR-V tensor dependency, and the off-contract tests.

https://github.com/llvm/llvm-project/pull/206249