[Mlir-commits] [mlir] [mlir][tensor] Preserve encoding in more canonicalizers (concat, reshape, pad) (PR #207241)

[Mehdi Amini]

joker-eph wrote:

> ConcatOp::inferResultType built the result type with no encoding. When all inputs share an encoding, propagate it; when they differ, drop it (concat semantics are undefined across differing encodings, so we do not pick one).

It's not clear to me that this is correct: the encoding could relate to the inputs dimensions and be totally inconsistent with the concatenated ones.

Overall there are similar issues with most of the "fixes" at hand with encoding: these "opaque metadata" poses fundamental problems for transformations.

https://github.com/llvm/llvm-project/pull/207241