[Mlir-commits] [mlir] [MLGO][EmitC] Scalarize single-element tensor returns (PR #199686)

[Mehdi Amini]

joker-eph wrote:

Not sure where I should commented, so cross-posting from the original implementation: https://github.com/llvm/llvm-project/pull/194404#issuecomment-4568681377

>> Since the limitations come from EmitC, it felt natural to move the logic there.
> If this is a tensor->tensor dialect pass, I'm not sure I agree.
> This is very speculative as a claim here. I could just as easily see this positioned as a "normalization" of the calling convention (returning a scalar seems cheaper than a tensor).


(Also: why a new PR when this is basically the same pass? Why not keep iterating on the existing PR?)

https://github.com/llvm/llvm-project/pull/199686