[LV] Only skip scalarization overhead for members used as address.

Refine logic to scalarize interleave group member: only skip
scalarization overhead for member being used as addresses. For others,
use the regular scalar memory op cost.

This currently doesn't trigger in practice as far as I could find, but
fixes a potential divergence between VPlan- and legacy cost models.

It fixes a concrete divergence with a follow-up patch,
llvm#161276.

Author: fhahn

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -5750,13 +5750,18 @@ void LoopVectorizationCostModel::setCostBasedWideningDecision(ElementCount VF) {
              getMemoryInstructionCost(I, ElementCount::getFixed(1))));
         UpdateMemOpUserCost(cast<LoadInst>(I));
       } else if (const auto *Group = getInterleavedAccessGroup(I)) {
-        // Scalarize an interleave group of address loads.
-        for (unsigned I = 0; I < Group->getFactor(); ++I) {
-          if (Instruction *Member = Group->getMember(I)) {
-            setWideningDecision(
-                Member, VF, CM_Scalarize,
-                (VF.getKnownMinValue() *
-                 getMemoryInstructionCost(Member, ElementCount::getFixed(1))));
+        // Scalarize all members of this interleaved group when any member
+        // is used as an address. The address-used load skips scalarization
+        // overhead, other members include it.
+        for (unsigned Idx = 0; Idx < Group->getFactor(); ++Idx) {
+          if (Instruction *Member = Group->getMember(Idx)) {
+            InstructionCost Cost =
+                AddrDefs.contains(Member)
+                    ? (VF.getKnownMinValue() *
+                       getMemoryInstructionCost(Member,
+                                                ElementCount::getFixed(1)))
+                    : getMemInstScalarizationCost(Member, VF);
+            setWideningDecision(Member, VF, CM_Scalarize, Cost);
             UpdateMemOpUserCost(cast<LoadInst>(Member));
           }
         }