[llvm] [LV] Don't simplify wide binops to constants if non-uniform (PR #121898)

[via llvm-commits]

ayalz wrote:

> ... I've been thinking about some potential alternatives, but it seems like none of the options are really great :( 

An add Reduction with a constant addend is an Induction, by definition, but a zero addend leads SCEV (rightfully) to consider the Phi an invariant rather than an AddRec?

Such a redundant invariant Phi chain best be RAUW'd with its start value? Better apply such a scalar preliminary cleanup before reaching LV, rather than within it?


https://github.com/llvm/llvm-project/pull/121898