[llvm] [DAG] Fold extract_subvector(insert_subvector(x, y, c1), c2) --> extract_subvector(y,c2-c1) (PR #87925)

[Steven Johnson via llvm-commits]

steven-johnson wrote:

I think this is a subtle bug in the changed code here -- I think the code above doesn't consider the possibility of scalable / fixed vector combinations, and Halide's tests (which tend to exercise a lot of corner cases in LLVM vector support) are breaking this. I'd like to request this get rolled back (to keep Halide unbroken) until we can work out exactly why this is breaking in this fashion, and hopefully re-land it after doing testing with Halide.

TL;DR: this needs to be rolled back.



https://github.com/llvm/llvm-project/pull/87925