[Mlir-commits] [mlir] [mlir][vector] Canonicalize transfer_{read, write} vector<1xT> (PR #196598)

[Krzysztof Drewniak]

krzysz00 wrote:

> That would be an argument to turn loads into transfer_read...

If we were raising a program to transfer-normal form, yes it would.

The claim I'm making here is that there's code - consider also things like this PR in IREE https://github.com/iree-org/iree/pull/24449 - that *expects* everything to look like a transfer_read/transfer_write and doesn't want to have to special-case in these allegedly "canonical" vector.load operations

That is, turning your transfer_* operations into load/store/maskedload/maskedstore/[your favorite platform intrinsic]/... is, in general, a big lever you pull at some defined point in your compilation pipeline. This pattern avoids breaking that layering.

https://github.com/llvm/llvm-project/pull/196598