Andrew Morrow <andrew.c.morrow@gmail.com> writes:
> I am no language laywer, so I don't know for certain what is going on
> here, but it does seem to be that one of the following must be true:
> 1) There is something illegal/undefined about the address argument to
> memset in the original version, due to the int32 * and char * casts.
I think you are absolutely right --- the cast to int32* allows the
compiler to assume that the pointer is word-aligned, and even casting
it back to char * or void * at the memset call wouldn't really undo
the damage. So really, correct coding of the macro should be along
the lines of
{ void *_vstart = (void *)(start); \
int _val = (val); \
Size _len = (len); \
\
if ((((long) _vstart) & INT_ALIGN_MASK) == 0 && \
(_len & INT_ALIGN_MASK) == 0 && \
_val == 0 && \
_len <= MEMSET_LOOP_LIMIT) \
{ \
int32 * _start = (int32 *) (_vstart); \
int32 * _stop = (int32 *) ((char *) _start + _len); \
while (_start < _stop) \
*_start++ = 0; \
} \
else \
memset(_vstart, _val, _len); \
} while (0)
Interesting that we have not seen this before.
regards, tom lane
