[microsoft] sema for __noop is fairly wrong

[nico]

Bugzilla Link	14081
Version	unspecified
OS	All
Blocks	#14079
Attachments	as UnaryExprOrTypeTrait, As builtin with special handling
CC	@DougGregor,@zmodem,@tritao,@rnk

Extended Description

Eli pointed out in an email that

template<typename T> static void f() { T x = 10; }
void g() { __noop(f<void>()); }

is an interesting case for __noop that compiles with cl but not with clang.

__noop has a few other interesting cases too:

• int __noop = 4 is a syntax error in cl; __noop is a keyword

• int a = __noop(42, 56); assigns 0 to a, but so does int a = __noop;. All of the following are valid, all but the first assign 0 to a:

  __noop;
  a = __noop;
  a = __noop();
  a = (__noop)();
  a = (__noop);
  a = (__noop());
  a = ((__noop)());
  a = ((__noop))();

• __noop is converted to pointers like a 0 literal:

  int (*f2)() = __noop;  // works, but just assings 0 to f2

• ...but it's not a constant expression:

  int arr[__noop() + 1];  // errors out

[nico]

More fun:

#include <typeinfo>
#include <iostream>

#include <stdio.h>
using namespace std;

template<typename T> static void f() { T x = 10 asdf asdf asd asdf }
void g() { __noop(f<void>()); }

class A {
public:
 void f();
};

int main() {
  //&__noop;
  int a[__noop(4 asdfasdf^*^&$^)(4, 5) + 1];
  decltype(__noop) ac;
  ac = 4;

  void (A::*foo)() = __noop;
 
  cout << typeid(__noop).name() << endl;
}

(Note that it's a constexpr in msvs2013)

[rnk]

Hm, unfortunately I didn't notice this bug, so I rediscovered all these wonderful features a few months ago. I added support for paren-less __noop in r212860, but we should probably promote this to a first class Expr node if we want to get it 100% right.

[llvmbot]

I’m working on this issue to make __noop a first class Expr.

--
Ilia Filippov
Software Engineer
Intel Compiler Team
Intel Corp.

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: Nico Weber (nico)

| | | | --- | --- | | Bugzilla Link | [14081](https://llvm.org/bz14081) | | Version | unspecified | | OS | All | | Blocks | llvm/llvm-project#14079 | | Attachments | [as UnaryExprOrTypeTrait](https://user-images.githubusercontent.com/3487/143746161-5d61078b-04e5-4314-ae07-230774ebc69c.gz), [As builtin with special handling](https://user-images.githubusercontent.com/3487/143746162-fd25f418-05d7-4b31-bae3-b29011473dc6.gz) | | CC | @DougGregor,@zmodem,@tritao,@rnk |

Extended Description

Eli pointed out in an email that

template<typename T> static void f() { T x = 10; }
void g() { __noop(f<void>()); }

is an interesting case for __noop that compiles with cl but not with clang.

__noop has a few other interesting cases too:

• int __noop = 4 is a syntax error in cl; __noop is a keyword

• int a = __noop(42, 56); assigns 0 to a, but so does int a = __noop;. All of the following are valid, all but the first assign 0 to a:

  __noop;
  a = __noop;
  a = __noop();
  a = (__noop)();
  a = (__noop);
  a = (__noop());
  a = ((__noop)());
  a = ((__noop))();

• __noop is converted to pointers like a 0 literal:

  int (*f2)() = __noop;  // works, but just assings 0 to f2

• ...but it's not a constant expression:

  int arr[__noop() + 1];  // errors out