On Mon, Sep 16, 2019 at 2:30 PM Linus Torvalds
<torva...@linux-foundation.org> wrote:
>
> On Mon, Sep 16, 2019 at 10:41 AM Andy Lutomirski <l...@kernel.org> wrote:
> >
> > After some experimentation, I think y'all are just doing it wrong.
> > GCC is very clever about this as long as it's given the chance. This
> > test, for example, generates excellent code:
> >
> > #include <string.h>
> >
> > __THROW __nonnull ((1)) __attribute__((always_inline)) void
> > *memset(void *s, int c, size_t n)
> > {
> > asm volatile ("nop");
> > return s;
> > }
> >
> > /* generates 'nop' */
> > void zero(void *dest, size_t size)
> > {
> > __builtin_memset(dest, 0, size);
> > }
>
> I think the point was that we'd like to get the default memset (for
> when __builtin_memset() doesn't generate inline code) also inlined
> into just "rep stosb", instead of that tail-call "jmp memset".
Well, when I wrote this email, I *thought* it was inlining the
'memset' function, but maybe I just can't read gcc's output today.
It seems like gcc is maybe smart enough to occasionally optimize
memset just because it's called 'memset'. This generates good code:
#include <stddef.h>
inline void *memset(void *dest, int c, size_t n)
{
/* Boris' implementation */
void *ret, *dummy;
asm volatile("push %%rdi\n\t"
"mov %%rax, %%rsi\n\t"
"mov %%rcx, %%rdx\n\t"
"andl $7,%%edx\n\t"
"shrq $3,%%rcx\n\t"
"movzbl %%sil,%%esi\n\t"
"movabs $0x0101010101010101,%%rax\n\t"
"imulq %%rsi,%%rax\n\t"
"rep stosq\n\t"
"movl %%edx,%%ecx\n\t"
"rep stosb\n\t"
"pop %%rax\n\t"
: "=&D" (ret), "=c" (dummy)
: "0" (dest), "a" (c), "c" (n)
: "rsi", "rdx", "memory");
return ret;
}
int one_word(void)
{
int x;
memset(&x, 0, sizeof(x));
return x;
}
So maybe Boris' patch is good after all.
