[llvm] [SLP]Better cost estimation for masked gather or "clustered" loads. (PR #105858)

[Alexey Bataev via llvm-commits]

================
@@ -4790,105 +4790,174 @@ BoUpSLP::LoadsState BoUpSLP::canVectorizeLoads(
       }
     }
   }
-  auto CheckForShuffledLoads = [&, &TTI = *TTI](Align CommonAlignment) {
+  // Correctly identify compare the cost of loads + shuffles rather than
+  // strided/masked gather loads. Returns true if vectorized + shuffles
+  // representation is better than just gather.
+  auto CheckForShuffledLoads = [&, &TTI = *TTI](Align CommonAlignment,
+                                                bool ProfitableGatherPointers) {
+    // Compare masked gather cost and loads + insert subvector costs.
+    TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
+    auto [ScalarGEPCost, VectorGEPCost] =
+        getGEPCosts(TTI, PointerOps, PointerOps.front(),
+                    Instruction::GetElementPtr, CostKind, ScalarTy, VecTy);
+    // Estimate the cost of masked gather GEP. If not a splat, roughly
+    // estimate as a buildvector, otherwise estimate as splat.
+    if (static_cast<unsigned>(count_if(
+            PointerOps, IsaPred<GetElementPtrInst>)) < PointerOps.size() - 1 ||
+        any_of(PointerOps, [&](Value *V) {
+          return getUnderlyingObject(V) !=
+                 getUnderlyingObject(PointerOps.front());
+        }))
+      VectorGEPCost += TTI.getScalarizationOverhead(
+          VecTy, APInt::getAllOnes(VecTy->getElementCount().getKnownMinValue()),
----------------
alexey-bataev wrote:

Yep, will fix

https://github.com/llvm/llvm-project/pull/105858