[llvm] [NVPTX] Customize getScalarizationOverhead (PR #128077)

[Justin Fargnoli via llvm-commits]

================
@@ -100,6 +101,42 @@ class NVPTXTTIImpl : public BasicTTIImplBase<NVPTXTTIImpl> {
       TTI::OperandValueInfo Op2Info = {TTI::OK_AnyValue, TTI::OP_None},
       ArrayRef<const Value *> Args = {}, const Instruction *CxtI = nullptr);
 
+  InstructionCost getScalarizationOverhead(VectorType *InTy,
+                                           const APInt &DemandedElts,
+                                           bool Insert, bool Extract,
+                                           TTI::TargetCostKind CostKind,
+                                           ArrayRef<Value *> VL = {}) {
+    if (!InTy->getElementCount().isFixed())
+      return InstructionCost::getInvalid();
+
+    auto VT = getTLI()->getValueType(DL, InTy);
+    auto NumElements = InTy->getElementCount().getFixedValue();
+    InstructionCost Cost = 0;
+    if (Insert && !VL.empty()) {
+      bool AllConstant = all_of(seq(NumElements), [&](int Idx) {
+        return !DemandedElts[Idx] || isa<Constant>(VL[Idx]);
+      });
+      if (AllConstant) {
+        Cost += TTI::TCC_Free;
+        Insert = false;
+      }
+    }
+    if (Insert && Isv2x16VT(VT)) {
+      // Can be built in a single mov
+      Cost += 1;
+      Insert = false;
----------------
justinfargnoli wrote:

It doesn't look like this is true: https://godbolt.org/z/TzodGMqf7

https://github.com/llvm/llvm-project/pull/128077