[llvm] [SLP]Charge extract latency for extracts feeding memory addresses (PR #224634)

[Alexey Bataev via llvm-commits]

================
@@ -19757,6 +19761,27 @@ InstructionCost BoUpSLP::getTreeCost(InstructionCost TreeCost,
       It->second = DT->findNearestCommonDominator(It->second, UseBlock);
   }
 
+  // An extract feeds a memory address if any external user of the scalar is
+  // a GEP index or the pointer operand of a load/store. The extract is priced
+  // once per scalar, so all its uses are checked for order independence.
+  auto IsExtractOnAddressPath = [&](Value *Scalar) {
+    return any_of(ExternalUses, [&](const ExternalUser &EU) {
+      if (EU.Scalar != Scalar || EphValues.count(EU.User))
+        return false;
+      Value *Usr = LookThroughIndexCast(EU.User);
+      if (auto *GEP = dyn_cast_if_present<GetElementPtrInst>(Usr))
+        // Conservatively assume an address use if there are too many to check.
+        return GEP->hasNUsesOrMore(UsesLimit) ||
+               any_of(GEP->users(), [GEP](User *U) {
+                 return match(U,
+                              m_CombineOr(m_Load(m_Specific(GEP)),
+                                          m_Store(m_Value(), m_Specific(GEP))));
----------------
alexey-bataev wrote:

It is not a store-related, it is address-related, which feeds into mem operations

https://github.com/llvm/llvm-project/pull/224634