[llvm] [AArch64][CostModel] Consider the cost of const vector (PR #117539)

[Alexey Bataev via llvm-commits]

================
@@ -11019,8 +11036,54 @@ BoUpSLP::getEntryCost(const TreeEntry *E, ArrayRef<Value *> VectorizedVals,
 
   bool NeedToShuffleReuses = !E->ReuseShuffleIndices.empty();
   if (E->isGather()) {
-    if (allConstant(VL))
+    if (allConstant(VL)) {
+      auto IsAllowedScalarTy = [&](const Type *T) {
+        return T->isFloatTy() || T->isDoubleTy() || T->isIntegerTy();
+      };
+      if (IsAllowedScalarTy(E->Scalars.front()->getType())) {
+        InstructionCost ScalarCost, VectorCost;
+
+        auto IsDuplicateEntry = [&](const TreeEntry *E) {
+          auto *TE = getTreeEntry(E->Scalars[0]);
+          if (TE != E) {
+            auto It = MultiNodeScalars.find(E->Scalars[0]);
+            if (It != MultiNodeScalars.end()) {
+              auto *TEIt = find_if(It->getSecond(), [&](TreeEntry *ME) {
+                return ME->isSame(VL);
+              });
+              if (TEIt != It->getSecond().end())
+                return true;
+            }
+          }
+          return false;
+        };
+
+        // FIXME: If there is more than 1 SLP tree realizing the same const
+        // vector, codegen will realize it only once. Hence, no need to consider
+        // the cost of const vector twice. But, currently we can't check if the
+        // tree entry is present in other SLP tree.
+        if (!isSplat(E->Scalars) && !all_of(E->Scalars, IsaPred<UndefValue>) &&
+            !IsDuplicateEntry(E)) {
+          // Get unique scalars
+          SmallDenseSet<Value *> UniqScalars;
+          for (auto *V : E->Scalars)
+            UniqScalars.insert(V);
+
+          // Constant is realized by having a mov/fmov into GPR. So,
+          // ScalarCost = #UniqScalars
+          ScalarCost = (UniqScalars.size());
+
+          // FIXME: Ideally, getGatherCost API should be used for this but
+          // currently, this API returns zero cost with all constants.
+          VectorCost =
+              TTI->getConstVectCost(Instruction::Load, FinalVecTy, Align(), 0,
+                                    CostKind, TTI::OperandValueInfo(), nullptr,
+                                    /*AllConstants=*/true, ScalarCost);
+        }
+        return VectorCost - ScalarCost;
+      }
----------------
alexey-bataev wrote:

This code should be in TTI completely, not here. Here we correctly consider the cost of constant vector as 0, it must be handled in other cost function (check getArithmeticInstCost in X86 for reference).

https://github.com/llvm/llvm-project/pull/117539