[PATCH] add support for -Wmismatched-dealloc

[Martin Sebor]

On 12/8/20 5:07 PM, Joseph Myers wrote:
> I don't see any definition of __attr_dealloc (presumably should be a macro
> in misc/sys/cdefs.h) in this patch (or in the glibc source tree).

Whoops!

> 
> Given all the functions in stdio.h with the same list of deallocation
> functions, there should probably be another macro there to expand to
> __attribute_malloc__ __attr_dealloc (fclose, 1) __attr_dealloc (freopen,
> 3) __attr_dealloc_freopen64.  (That would also apply to open_wmemstream in
> wchar.h, but I suppose you have the issue there with functions not being
> declared in wchar.h, only in stdio.h?)

You're right that adding the attribute to open_wmemstream runs into
the same problem as declaring the ctermid argument with L_ctermid
elements in <unistd.h>: fclose isn't declared in <wchar.h>.  I did
some more work on adding the attribute to other functions and found
out that the same problem also happens between tempnam() in <stdio.h>
and reallocarray() in <stdlib.h>.

I spent some time working around this but in the end it turned out
to be too convoluted so I decided to make the attribute a little
smarter.  Instead of associating all allocation functions with all
deallocation functions (such as fdopen, fopen, fopen64, etc. with
fclose, freopen, and freopen64) I changed it so that an allocator
only needs to be associated with a single deallocator (a reallocator
also needs to be associated with itself).  That makes things quite
a bit simpler.

The attached patch implements this for <stdio.h>, <stdlib.h>, and
<wchar.h>.  To get around the <wchar.h> dependency on <stdio.h> it
uses __REDIRECT to introduce a reserved alias for fclose.

Besides running the test suite I tested it with my own test and also
by adding the same declarations to the GCC test suite and verifying
it triggers warnings as expected.

The GCC patches needed to make this simpler scheme work haven't been
reviewed yet so this work has a dependency on them getting approved.

I grepped for __attribute_malloc__ in Glibc headers to see if there
are other APIs that would benefit from the same annotation but found
none.  At the same time, I don't have the impression that malloc is
used on all the APIs it could be.  Are there any that you or anyone
else can think of that might be worth looking at?

Martin
-------------- next part --------------
A non-text attachment was scrubbed...
Name: glibc-attr-malloc.diff
Type: text/x-patch
Size: 10910 bytes
Desc: not available
URL: <https://sourceware.org/pipermail/libc-alpha/attachments/20201211/210fcf09/attachment-0001.bin>