[PATCH] D78059: [llvm][STLExtras] Add various type_trait utilities currently present in MLIR

[David Blaikie via Phabricator via llvm-commits]

dblaikie added a comment.

Not sure if the other LLVM type traits have tests - but could you check & probably add some static_assert tests to demonstrate/validate the functionality being added here?



================
Comment at: llvm/include/llvm/ADT/STLExtras.h:109-145
+/// This class provides various trait information about a callable object.
+///   * To access the number of arguments: Traits::num_args
+///   * To access the type of an argument: Traits::arg_t<i>
+///   * To access the type of the result:  Traits::result_t
+template <typename T, bool isClass = std::is_class<T>::value>
+struct function_traits : public function_traits<decltype(&T::operator())> {};
+
----------------
Use of this device makes me a bit uncomfortable - such a thing would be incompatible with overloading, default arguments, etc. 

A general idea that's gone around inside Google but I can't find a good reference for it externally, is to only depend on the "call only interface" (ie: never take the address of a function you don't own - only call it with whatever types it was intended to have and treat it as though it returns the type it's documented to return). This ensures maximum flexibility for the function vendor to refactor it - add default arguments or overload the function as part of migrating/improving/refactoring the API, etc.

Is there any chance the existing uses of the mlir-equivalent of this trait could reasonably avoid it & instead rely on the call-only contract when interacting with functions/functors?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D78059/new/

https://reviews.llvm.org/D78059