[Mlir-commits] [mlir] [mlir][tensor] remove tensor.insert constant folding out of canonicalization (PR #142671)

[Mehdi Amini]

https://github.com/joker-eph approved this pull request.

Thanks!

Copying my comment from https://github.com/llvm/llvm-project/pull/142671#discussion_r2130442667

It's probably safer to remove the pattern for now, and maybe think about introducing a "tensor-constant-folding" pass separately. I suspect such pass would want options about a maximum size for folding a constant or other option for the strategy to adopt.
We can also think about having these patterns in canonicalization, but limited to "small" constants (the difficulty being how to define how much small is small enough to be universally safe).

https://github.com/llvm/llvm-project/pull/142671