Bug#398316: gcc-4.1: Apparent inline function expansion bug on ARM
Theodore Tso <tytso@mit.edu> writes:
> On Mon, Nov 13, 2006 at 09:49:21AM +0100, Falk Hueffner wrote:
>> Hi,
>>
>> this looks like an aliasing violation to me. bb->list, which is of
>> type __u32*, is accessed via an lvalue of type void*, which is not
>> compatible. Does the problem go away with -fno-strict-aliasing?
>
> Good point, thanks. Yes, it does go away when I use
> -fno-strict-aliasing.
>
> Interestingly, I didn't get nary a peep when I enabled tried running
> gcc "-O2 -Wstrict-aliasing=2". I would have expected gcc to have
> complained about this, but apparently not.
>
> Do you think the problem is with the call to ext2fs_get_mem:
>
> retval = ext2fs_get_mem(bb->size * sizeof(blk_t), &bb->list);
>
> where I pass a __u32* to ext2fs_get_mem as a void*?
>
>
> Or the implementation where I play games with void** and void*:
>
> _INLINE_ errcode_t ext2fs_get_mem(unsigned long size, void *ptr)
> {
> void **pp = (void **)ptr;
>
> *pp = malloc(size);
> ...
Well, neither. All the intermediary types of pointers don't matter;
the actual problem is at the access of bb->list via an lvalue of a bad
type.
> I tried changing the function signature of ext2fs_get_mem() to be:
>
> _INLINE_ errcode_t ext2fs_get_mem(unsigned long size, char *ptr)
>
> ... since according to the spec storing through a char* is supposed to
> be OK, but that didn't seem to fix the problem.
How did you do it exactly? For example memcpy() should work.
> I'd really hate to be forced to change the function signature to
> something like this:
>
> _INLINE_ void * ext2fs_get_mem2(unsigned long size, errcode_t *retval)
>
> just to get around these !@#! type-punning rules, but I can't think of
> anything else that might work. Is there anything you might suggest?
Well, if you really want to retain the signature, I cannot think of
anything better than using memcpy.
--
Falk
Reply to: