Syntax error when compiling C example with 32-bit address space

[ararmine]

Clang issues a syntax error for the following C example example:
Tried with x86-64 clang 19.1.0 version.

1. cat example.c

typedef int (*__ptr32 fptr1)();
typedef int (* __ptr32 fptr5)(fptr1);
int f5 (int (int()));
int f5 (fptr5 fp)
{
  return 5;
}

2. clang example.c -fms-extensions

example.c:4:5: error: conflicting types for 'f5'
int f5 (fptr5 fp)
    ^
example.c:3:5: note: previous declaration is here
int f5 (int (int()));
    ^
1 error generated.

NOTE: When compiling the example as a C++ source, there is no error.

x64 MSVC v.19 gives a warning for the C example:

example.c
<source>(4): warning C4028: formal parameter 1 different from declaration

I am not sure if this a bug.

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: None (ararmine)

Clang issues a syntax error for the following C example example: Tried with **x86-64 clang 19.1.0** version.

1. cat example.c

typedef int (*__ptr32 fptr1)();
typedef int (* __ptr32 fptr5)(fptr1);
int f5 (int (int()));
int f5 (fptr5 fp)
{
  return 5;
}

2. clang example.c -fms-extensions

example.c:4:5: error: conflicting types for 'f5'
int f5 (fptr5 fp)
    ^
example.c:3:5: note: previous declaration is here
int f5 (int (int()));
    ^
1 error generated.

NOTE: When compiling the example as a C++ source, there is no error.

x64 MSVC v.19 gives a warning for the C example:

example.c
<source>(4): warning C4028: formal parameter 1 different from declaration

I am not sure if this a bug.

[AaronBallman]

I think Clang's behavior is correct because the types are different; MSVC is being slightly more permissive and perhaps we want to follow suit for compatibility?

CC @compnerd @rnk @majnemer @zmodem for additional opinions

[zygoloid]

More data: in C++ mode, MSVC accepts:

typedef int (*__ptr32 fptr1)();
typedef int (* __ptr32 fptr5)(fptr1);
int f5 (int (int())) { return 4; }
int f5 (fptr5 fp) { return 5; }

and treats these declarations as two different functions. In C mode, MSVC accepts:

int f5 (int p);
int f5 (void *p) { return 5; }

with only a warning. So this has nothing to do with __ptr32 -- MSVC just allows prototypes to differ between declaration and definition in C (with a warning).

[ararmine]

Clang accepts a little simplified version of the initial example, in C mode:

typedef int (* __ptr32 fptr5)();
int f5 (int());
int f5 (fptr5 fp)
{
  return 0;
}

Shouldn't Clang issue a similar error for this case?

[AaronBallman]

Clang accepts a little simplified version of the initial example, in C mode:

typedef int (* __ptr32 fptr5)();
int f5 (int());
int f5 (fptr5 fp)
{
  return 0;
}

Shouldn't Clang issue a similar error for this case?

Yes, I believe it should because the first f5 has a type of int (*)(int (*)()) and the second f5 as a type of int (*)(int (* __ptr32)()) and those are different types (at least, on a non-64-bit build). MSDN is silent about just how this qualifier works, so I could be wrong.

[compnerd]

With the goal of MSVC compatibility, I think that it makes sense to actually mimic the MSVC behaviour and simply warn on the differing prototype/definition in C mode and treat the functions as overloads in C++ mode. This type of subtle difference is something which could prevent a smooth migration from MSVC to clang even though it is meant to be a drop-in replacement.

[rnk]

From a code health and correctness perspective, I think Clang's behavior is good. Mismatched declaration and definition prototypes can create all sorts of hard to diagnose UB, and fixing those declarations feels like worthwhile effort during any MSVC->Clang C codebase migration.

Our guiding policy has historically been to be MSVC compatible enough to compile system headers, and this feels like it would go beyond that.

[AaronBallman]

Hmm, this extension is confusing to me. MSVC does not appear to make the qualifier part of the type identity; otherwise I would expect this to be a valid overload set: https://godbolt.org/z/zjWohWaq4

So if this qualifier is part of the type identity, then I think Clang's behavior is correct and more useful than MSVC's behavior. However, because this qualifier seems to be... something else... it may be more defensible to follow MSVC's behavior.

[zygoloid]

Hmm, this extension is confusing to me. MSVC does not appear to make the qualifier part of the type identity; otherwise I would expect this to be a valid overload set: https://godbolt.org/z/zjWohWaq4

I guess MSVC is ignoring all top-level qualifiers when determining whether an overload is the same function (not just cv-qualifiers).