[PATCH] add support for -Wmismatched-dealloc

[Florian Weimer]

* Martin Sebor:

> On 1/4/21 9:07 AM, Florian Weimer wrote:
>> * Martin Sebor via Libc-alpha:
>> 
>>> diff --git a/wcsmbs/wchar.h b/wcsmbs/wchar.h
>>> index 9cf8b05a87..4c1c7f1119 100644
>>> --- a/wcsmbs/wchar.h
>>> +++ b/wcsmbs/wchar.h
>>> @@ -151,7 +151,8 @@ extern size_t wcsxfrm_l (wchar_t *__s1, const wchar_t *__s2,
>>>   			 size_t __n, locale_t __loc) __THROW;
>>>     /* Duplicate S, returning an identical malloc'd string.  */
>>> -extern wchar_t *wcsdup (const wchar_t *__s) __THROW __attribute_malloc__;
>>> +extern wchar_t *wcsdup (const wchar_t *__s) __THROW
>>> +  __attribute_malloc__ __attr_dealloc_free;
>>>   #endif
>>>     /* Find the first occurrence of WC in WCS.  */
>>> @@ -562,9 +563,18 @@ extern wchar_t *wcpncpy (wchar_t *__restrict __dest,
>>>   /* Wide character I/O functions.  */
>>>     #if defined __USE_XOPEN2K8 || __GLIBC_USE (LIB_EXT2)
>>> +#  ifdef __REDIRECT
>>> +/* Declare the __fclose alias and associate it as a deallocator
>>> +   with open_wmemstream below.  */
>>> +extern int __REDIRECT (__fclose, (FILE *), fclose);
>>> +#    define __attr_dealloc_fclose __attr_dealloc (__fclose, 1)
>>> +#  else
>>> +#    define __attr_dealloc_fclose /* empty */
>>> +#  endif
>>>   /* Like OPEN_MEMSTREAM, but the stream is wide oriented and produces
>>>      a wide character string.  */
>>> -extern __FILE *open_wmemstream (wchar_t **__bufloc, size_t *__sizeloc) __THROW;
>>> +extern __FILE *open_wmemstream (wchar_t **__bufloc, size_t *__sizeloc) __THROW
>>> +  __attribute_malloc__ __attr_dealloc_fclose;
>>>   #endif
>>>     #if defined __USE_ISOC95 || defined __USE_UNIX98
>> Why is an alias for fclose needed here, but not for free?
>
> Because fclose is not a built-in so there's no __builtin_fclose
> to associate open_wmemstream with.  free is a built-in and so
> __attr_dealloc_free just references __builtin_free and doesn't
> need an explicit declaration.

Ahh, that explains the discrepancy.

I'm a bit worried that the __fclose alias causes problems.  Would it be
possible to add __builtin_fclose to GCC instead?

Based on how this patch appears to make both __fclose and fclose
acceptable as a deallocator, GCC resolves redirects as part of the
matching check.  I wonder if this constrains the usefulness of the
attribute in some way.  I can imagine situations where at the source
level, different deallocators should be used (say to support debugging
builds), but release builds redirect different deallocators to the same
implementation.

Thanks,
Florian
-- 
Red Hat GmbH, https://de.redhat.com/ , Registered seat: Grasbrunn,
Commercial register: Amtsgericht Muenchen, HRB 153243,
Managing Directors: Charles Cachera, Brian Klemm, Laurie Krebs, Michael O'Neill