[llvm] [SLP] Account for GEP pointer-chain cost when root scalars feed load/store indices (PR #192726)

[Ryan Buchner via llvm-commits]

================
@@ -19073,6 +19093,41 @@ InstructionCost BoUpSLP::getTreeCost(InstructionCost TreeCost,
 
     ExtractCost += ExtraCost;
   }
+  // Charge the pointer-chain cost difference once for the root entry when
+  // every external use of its scalars is a GEP feeding a single load/store
+  // (see the detection loop above). Vectorizing the root in this pattern
+  // forces lane extracts (or a vector GEP with unknown stride) to drive the
+  // address computation, which is typically more expensive than keeping the
+  // indices scalar in a unit-stride address chain. Add the delta once rather
+  // than per external use.
+  if (AllUsersGEPSWithStoresLoads && !Pointers.empty()) {
+    const TreeEntry &RootEntry = *VectorizableTree.front();
+    const bool AnyRootKeptAsScalar = any_of(RootEntry.Scalars, [&](Value *V) {
+      return ExternalUsesAsOriginalScalar.contains(V);
+    });
+    const Value *CommonBase = nullptr;
+    bool HaveCommonBase = true;
+    for (const Value *P : Pointers) {
+      const Value *Op = cast<GetElementPtrInst>(P)->getPointerOperand();
----------------
bababuck wrote:

Could we use `getUnderlyingObject` instead?

https://github.com/llvm/llvm-project/pull/192726