[llvm] [LV] Teach LoopVectorizationLegality about struct vector calls (PR #119221)

[David Sherwood via llvm-commits]

================
@@ -942,11 +954,24 @@ bool LoopVectorizationLegality::canVectorizeInstrs() {
       if (CI && !VFDatabase::getMappings(*CI).empty())
         VecCallVariantsFound = true;
 
+      auto CanWidenInstruction = [this](Instruction const &Inst) {
+        Type *InstTy = Inst.getType();
+        if (isa<CallInst>(Inst) && isa<StructType>(InstTy) &&
+            canWidenCallReturnType(InstTy)) {
+          // TODO: Remove the `StructVecCallFound` flag once vectorizing calls
+          // with struct returns is supported.
+          StructVecCallFound = true;
+          // For now, we only recognize struct values returned from calls where
+          // all users are extractvalue as vectorizable.
+          return all_of(Inst.users(), IsaPred<ExtractValueInst>);
+        }
+        return VectorType::isValidElementType(InstTy) || InstTy->isVoidTy();
----------------
david-arm wrote:

It's a shame this can't just call `canVectorizeTy` here, since it's duplicating the logic in that function. I'm not sure what you think, but you could just write this as:

```
  return !isa<StructType>(InstTy) && canVectorizeTy(StructTy);
```

Alternatively, you could restructure the wider logic a bit to something like this:

```
        Type *InstTy = Inst.getType();
        if (isa<StructType>(InstTy)) {
          if (!isaCallInst>(Inst) || !canWidenCallReturnType(InstTy))
            return false;

          // TODO: Remove the `StructVecCallFound` flag once vectorizing calls
          // with struct returns is supported.
          StructVecCallFound = true;

          // For now, we only recognize struct values returned from calls where
          // all users are extractvalue as vectorizable.
          return all_of(Inst.users(), IsaPred<ExtractValueInst>);
        }
        return canVectorizeTy(InstTy);
```

Any thoughts?


https://github.com/llvm/llvm-project/pull/119221