[clang] [llvm] InstCombine: Order shufflevector operands by complexity (PR #113212)

[Matthias Braun via cfe-commits]

MatzeB wrote:

> This idea here makes sense, I guess, but I'm a little worried about unexpected side-effects... not sure how reliably the backend handles commuted masks, and I don't think we have any testing infrastructure that would catch that sort of issue.

Was wondering about this too, but at least it should be easy to fix things if some targets somehow depend on the order and produce worse code in one ordering but not the other. Given that these targets could have reached this situation without this change it may be a good thing to flush out this behavior...

https://github.com/llvm/llvm-project/pull/113212