https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83201
--- Comment #24 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Sean from comment #23)
> Sorry for digging up the past, but recently ran into a related
> implementation issue in OpenBSD's qsort implementation and came across this
> discussion. While I understand the proposed patch was rejected, it looks
> like it also degraded the implementation to single byte swaps.
>
> Admittedly untested, but a possible solution that I think would make LTO
> happy (avoids aliasing the pointer) and doesn't change the implementation
> profile would be to cast it differently since it's always dealing with a
> simple type alignment check on long/int/char. Something like this:
>
> #define SWAPINIT(TYPE, a, es) swaptype_ ## TYPE = (uintptr_t)a %
> sizeof(TYPE) || es % sizeof(TYPE) ? 2 : es == sizeof(TYPE) ? 0 : 1;
>
> Maybe a more acceptable alternative compromise since it would have a nearly
> identical performance profile with it still swap-optimizing int and long.
Note the issue is not the way the type is computed but the memory
accesses done via
#define swap(a, b) \
if (swaptype_long == 0) { \
long t = *(long *)(a); \
*(long *)(a) = *(long *)(b); \
*(long *)(b) = t; \
} else if (swaptype_int == 0) { \
int t = *(int *)(a); \
*(int *)(a) = *(int *)(b); \
*(int *)(b) = t; \
there exists non-standard ways to preserve the access pattern like
typedef int notbaa_int __attribute__((may_alias));
typedef long notbaa_long __attribute__((may_alias));
#define swap(a, b) \
if (swaptype_long == 0) { \
long t = *(notbaa_long *)(a); \
*(notbaa_long *)(a) = *(notbaa_long *)(b); \
*(notbaa_long *)(b) = t; \
} else if (swaptype_int == 0) { \
int t = *(notbaa_int *)(a); \
*(notbaa_int *)(a) = *(notbaa_int *)(b); \
*(notbaa_int *)(b) = t; \
or portable ones like doing
long t;
memcpy (&t, a, sizeof (long));
memcpy (a, b, sizeof (long));
memcpy (b, &t, sizeof (long));
and hoping for the compiler to optimize this well (GCC will do).
Note at time I proposed the change to the originating project (freebsd?) and
it was accepted there.
