[Mlir-commits] [mlir] [mlir][vector] Add vector.transpose with unit-dim to vector.shape_cast pattern (PR #72105)

[Nicolas Vasilache]

nicolasvasilache wrote:

@joker-eph side note here, "the better guarantees because no data movement" is not accurate anymore.
While it was the intent for the op to fold away into other ops all the way to memory, we now also have lowerings that end up generating vector.shuffle operations.

I am personally not clear which one is a "more canonical" form atm, I have personally not considered canonicalizations from one form to the other.


https://github.com/llvm/llvm-project/pull/72105