[llvm] Fix scalar overload name constructed by ReplaceWithVeclib.cpp (PR #111095)

[Tex Riddell via llvm-commits]

================
@@ -100,27 +100,38 @@ static void replaceWithTLIFunction(IntrinsicInst *II, VFInfo &Info,
 static bool replaceWithCallToVeclib(const TargetLibraryInfo &TLI,
                                     IntrinsicInst *II) {
   assert(II != nullptr && "Intrinsic cannot be null");
+  Intrinsic::ID IID = II->getIntrinsicID();
+  if (IID == Intrinsic::not_intrinsic)
+    return false;
   // At the moment VFABI assumes the return type is always widened unless it is
   // a void type.
   auto *VTy = dyn_cast<VectorType>(II->getType());
   ElementCount EC(VTy ? VTy->getElementCount() : ElementCount::getFixed(0));
+  Type *ScalarRetTy = II->getType()->getScalarType();
   // Compute the argument types of the corresponding scalar call and check that
   // all vector operands match the previously found EC.
   SmallVector<Type *, 8> ScalarArgTypes;
-  Intrinsic::ID IID = II->getIntrinsicID();
+
+  // OloadTys collects types used in scalar intrinsic overload name.
+  SmallVector<Type *, 3> OloadTys;
+  if (!ScalarRetTy->isVoidTy() &&
+      isVectorIntrinsicWithOverloadTypeAtArg(IID, -1))
+    OloadTys.push_back(ScalarRetTy);
+
   for (auto Arg : enumerate(II->args())) {
     auto *ArgTy = Arg.value()->getType();
-    if (isVectorIntrinsicWithScalarOpAtArg(IID, Arg.index())) {
-      ScalarArgTypes.push_back(ArgTy);
-    } else if (auto *VectorArgTy = dyn_cast<VectorType>(ArgTy)) {
-      ScalarArgTypes.push_back(VectorArgTy->getElementType());
+    auto *ScalarArgTy = ArgTy->getScalarType();
+    ScalarArgTypes.push_back(ScalarArgTy);
+    if (isVectorIntrinsicWithOverloadTypeAtArg(IID, Arg.index()))
+        OloadTys.push_back(ScalarArgTy);
+    if (auto *VectorArgTy = dyn_cast<VectorType>(ArgTy)) {
----------------
tex3d wrote:

In any case, we definitely should not do this up front:
```
if (!isVectorIntrinsicWithScalarOpAtArg(IID, Arg.index()))
   return false;
```

Otherwise, any intrinsic with a vector operand will be skipped.  That's not what this code does.  It expects a vector or scalar operand.  If vector, the scalar type is used and it checks that the size matches, otherwise, it's a scalar operand, this scalar type is used, and `isVectorIntrinsicWithScalarOpAtArg` should return true (it would return false for vector operands).

https://github.com/llvm/llvm-project/pull/111095