[Mlir-commits] [mlir] [mlir][vector] Don't fold in_bounds for negative constant indices (PR #219681)

[Alessandro Potenza]

alepot55 wrote:

Both of those are better reasons than the suggestions they answer.

The merge-order point I had not weighed at all. I was looking at the duplication and not at what a shared helper costs two patches that are deliberately independent, and forcing an order on them to save four lines is a bad trade. Same for `maxStart < 0`: a bail that skips a `ValueBoundsConstraintSet` query and a bail that skips two integer comparisons are not the same kind of guard, so "match the sibling" was the wrong instinct.

Agreed on the cross-references going in with whichever lands second. The `kDynamic` comment is the one I would most want in the tree anyway, since the `isDynamicDim` return that makes the subtraction safe is fifteen lines up and nothing connects the two today.


https://github.com/llvm/llvm-project/pull/219681