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

[Mehdi Amini]

joker-eph wrote:

I think that is reasonable, with one caveat: making an operation more static (turning an SSA value into an attribute) can generate invalid IR (the verifier would check the attribute where it didn't check the SSA value, I consider operation accepting either attributes or SSA value to be an anti-pattern by the way). 
The patterns are responsible to preserve correctness (that is checking the folding is valid), but with encoding you can't really know I think.

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