[llvm] [LV] Only use legacy scalarization costs with replicate regions. (PR #212738)

[David Sherwood via llvm-commits]

================
@@ -147,8 +147,8 @@ define void @test(ptr noalias nocapture %points, ptr noalias nocapture readonly
 ; ENABLED_MASKED_STRIDED:  LV: Found an estimated cost of 1 for VF 1 For instruction: store i16 %0, ptr %arrayidx6, align 2
 ; ENABLED_MASKED_STRIDED:  Cost of 2 for VF 2: profitable to scalarize store i16 %0, ptr %arrayidx6, align 2
 ; ENABLED_MASKED_STRIDED:  Cost of 4 for VF 4: profitable to scalarize store i16 %0, ptr %arrayidx6, align 2
-; ENABLED_MASKED_STRIDED:  Cost of 8 for VF 8: profitable to scalarize store i16 %0, ptr %arrayidx6, align 2
-; ENABLED_MASKED_STRIDED:  Cost of 16.5 for VF 16: profitable to scalarize store i16 %0, ptr %arrayidx6, align 2
+; ENABLED_MASKED_STRIDED:  Cost of 12 for VF 8: INTERLEAVE-GROUP with factor 3, ir<%arrayidx6>, ir<%cmp1>
----------------
david-arm wrote:

Are we now making a different widening decision that's worse than before? From the CHECK lines it's hard to see if we were already interleaving.

https://github.com/llvm/llvm-project/pull/212738