[llvm] [LoopVectorize] Don't scalarize predicated instruction with optsize (PR #129265)

[David Sherwood via llvm-commits]

================
@@ -67,150 +67,44 @@ define void @example1() optsize {
   ret void
 }
 
-; Can vectorize in 'optsize' mode by masking the needed tail.
+; Can't vectorize in 'optsize' mode as it would require masking the needed tail.
----------------
david-arm wrote:

Isn't the underlying problem here that vectorising leads to significantly increased code due to scalarising the memory operations? The new comment suggests to me that tail-folding and optsize are incompatible, but that's only true if you end up scalarising.

https://github.com/llvm/llvm-project/pull/129265