clang redefines size_t silently but then complains

[filcab]

Bugzilla Link	2427
Resolution	FIXED
Resolved on	Nov 07, 2018 00:21
Version	unspecified
OS	All
CC	@efriedma-quic

Extended Description

clang doesn't compile the following code, although gcc happily compiles:

#include <stddef.h>
extern void* f(void** p, size_t s);

#include <stdlib.h>

void* f(void** p, size_t s)
{
}

I think the problem is with the redefinition of size_t in stdlib.h (a system header) to something different from what is defined in stddef.h.

In Mac OS X, clang complains:
[filcab@farnsworth ~] $ clang b.c
b.c:7:7: error: conflicting types for 'f'
void* f(void** p, size_t s)
^
b.c:2:14: error: previous declaration is here
extern void* f(void** p, size_t s);
^
2 diagnostics generated.

[lattner]

This is another "duplicate typedefs in system headers" should be ignored thing.

[filcab]

What's weird is that clang isn't complaining about the redefinition (so it may also be muffled somewhere) but it's actually redefining instead of ignoring everything.

[efriedma-quic]

What's mystifying me is that the old size_t and the new size_t should be compatible. How exactly is your stdlib.h defining size_t? And can you try the following testcase?

#include <stddef.h>
extern size_t s;
#include <stdlib.h>
extern size_t s;

[lattner]

It sounds to me that the redefinition isn't getting canonicalized right, so we end up with two size_t's floating around.

[efriedma-quic]

It sounds to me that the redefinition isn't getting canonicalized right, so we
end up with two size_t's floating around.

We don't "canonicalize" redefinitions of typedefs; we just let them shadow the earlier typedef.

[filcab]

What's mystifying me is that the old size_t and the new size_t should be
compatible. How exactly is your stdlib.h defining size_t?
stdlib.h includes types.h and then:

70 "/usr/include/i386/_types.h"

(snip)

typedef unsigned int __darwin_size_t;

And can you try the following testcase?

#include <stddef.h>
extern size_t s;
#include <stdlib.h>
extern size_t s;

$ clang b.c
b.c:4:15: error: redefinition of 's'
extern size_t s;
^
b.c:2:15: error: previous definition is here
extern size_t s;
^
2 diagnostics generated.

[lattner]

That would be the problem. That means that two types pointing to two different typedef decls are different, even though they represent the same type. That would cause them to be considered incompatible probably

[efriedma-quic]

That would be the problem. That means that two types pointing to two different
typedef decls are different, even though they represent the same type. That
would cause them to be considered incompatible probably

I don't know about C++, but that's not an issue in C. No matter how you typedef an type in C, it's still compatible with itself.

The issue is here:
typedef unsigned int __darwin_size_t;

size_t on an Intel Mac should definitely be unsigned long; I just had someone on IRC check with gcc. Not sure what's going wrong here. Maybe some gcc fixincludes magic that clang can't take advantage of?

Unfortunately, it's extremely difficult to cleanly fix this in the compiler... an extremely hacky fix would be to force any definition of size_t to the appropriate type if it's defined in a system header. It would also be possible to just change the type of size_t in clang to unsigned int (that wouldn't be consistent with gcc on Darwin, but any program depending on that isn't portable anyway).

[llvmbot]

This should be resolved. Please confirm and close.

[filcab]

Yes, it works here too. Marking as closed (if I can).