[llvm] [LoopVectorizer] Prune VFs based on plan register pressure (PR #132190)

[David Green via llvm-commits]

================
@@ -12,28 +12,28 @@ define i32 @mla_i32(ptr noalias nocapture readonly %A, ptr noalias nocapture rea
 ; CHECK:       for.body.preheader:
 ; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK:       vector.ph:
-; CHECK-NEXT:    [[N_RND_UP:%.*]] = add i32 [[N]], 15
-; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i32 [[N_RND_UP]], 16
+; CHECK-NEXT:    [[N_RND_UP:%.*]] = add i32 [[N]], 7
----------------
davemgreen wrote:

This should not be changing, it was OK as it was before. https://godbolt.org/z/KnT1xaeTf

Is it changing because of the large reduction registers? Those wouldn't be expected to be in the final code. Perhaps https://github.com/llvm/llvm-project/pull/113903 will help to represent the register usage more correctly.

https://github.com/llvm/llvm-project/pull/132190