CIR operation for dynamic_cast

[Lancern]

PR #426 introduces CIRGen support for the dynamic_cast operator and enables a fast path to LLVM lowering. As a follow up, we plan to invent a dedicated operation for dynamic_cast. During LLVM lowering prepare, this new operation will be replaced by what is generated currently for dynamic_cast.

There are multiple possible approaches to add this new operation, however:

The first approach is to reuse the current cir.cast operation and add new variants to represent dynamic_cast. Code below shows a possible CIRGen:

struct Base { /* ... */ };
struct Derived : Base { /* ... */ };

Derived *test(Base *b) {
  // %0 = cir.cast(dynamic, %b : !cir.ptr<!struct_Base>) : !cir.ptr<!struct_Derived>
  return dynamic_cast<Derived *>(b);
}

Derived &test(Base &b) {
  // %0 = cir.cast(dynamic, %b : !cir.ptr<!struct_Base>) : !cir.ptr<!struct_Derived>
  // %is_not_null = cir.cast(ptr_to_bool, %0)
  // %is_null = cir.unary(not, %is_null)
  // cir.if %is_null {
  //   cir.throw ...
  // }
  return dynamic_cast<Derived &>(b);
}

The names of these new variants can be decided later.

The second approach is to invent a new cir.dynamic_cast operation for it:

Derived *test(Base *b) {
  // %0 = cir.dynamic_cast(%b : !cir.ptr<!struct_Base>) : !cir.ptr<!struct_Derived>
  return dynamic_cast<Derived *>(b);
}

Derived &test(Base &b) {
  // %0 = cir.dynamic_cast(%b : !cir.ptr<!struct_Base>) : !cir.ptr<!struct_Derived>
  // %is_not_null = cir.cast(ptr_to_bool, %0)
  // %is_null = cir.unary(not, %is_null)
  // cir.if %is_null {
  //   cir.throw ...
  // }
  return dynamic_cast<Derived &>(b);
}

There may be other approaches as well.

Which one do you prefer?

[bcardosolopes]

I'd say second approach so we're able to add a constraint directly in tablegen for requiring pointer input and output

[Lancern]

Sound reasonable. Let's follow this approach.

[bcardosolopes]

Already implemented, closing