Gcc builtin review: memset

[OndÅej BÃlka <neleai at seznam dot cz>]

First quick statement that memccpy is ok as gcc doesn't do anything.

Then memset suffers for same flaws as memcpy. If modify previous memcpy
benchmark to following then for size 120 and obsolete glibc-2.13 you
have an around 25% performance regression.

a.c:
#include <string.h>
int memcpy2(char *c, char *s)
{
  return memset(c, 0, 120);
}

An size bloat is lesser.

Then there is optimization that on some archs you should call
__bzero (s, n) instead memset (s, 0, n). A effective __bzero consist of
small prologue that jumps to memset after initialization. It saves
cycles of creating mask from c and passing additional argument.

A main diadvantage would be renaming as bzero doesn't return anything
but x64 does return same value as memset.

Also as I mentioned ufunc a memset is target to get better performance
without troubles that you couldn't select cpu. Note that on haswell
using avx2 instincs could improve performance. How are you going exploit
that with gcc?