> From: Erik
> Newsgroups: gmane.comp.gcc.help
> Subject: Re: Optimisation puzzle
> Date: Fri, 16 Mar 2007 12:53:16 +0100
[]
> I tried with the following program: void q(const unsigned int); void
> f() {for (unsigned int x = 0; x != 10; x++) q(77);}
>
> and built it with "gcc -std=c99 -Os -Wall -Wextra -Werror -S":
> .L2:
> subl $12, %esp
> incl %ebx
> pushl $77
> call q
> addl $16, %esp
> cmpl $10, %ebx
> jne .L2
>
> As you can see, "pushl $77" is still inside the loop even though q
> promises to not change its argument. This must be a bug.
Maybe this somehow linked to the fact, that parameter by value are always
copied and there's no way for caller to know about what happened to that
*copy*. E.g. "void q(const unsigned int b);" means b is read-only inside q,
but any caller sets up b for q, i.e it create/change it. Anyways this
can be optimized. And i think, gdb will be ok with such code.
That isn't working program, you've tested, is it? I tried working one,
but it was little bit hard to get non-unrolled loop. See even one more
instruction, that seems to be useless, unless it somehow speeds-up
register access (or any other CPU's magic).
Of course things are completely different with q being static.
AMD64 code:
..globl q
.type q, @function
q: _mtune=k8_ q: _mtune=nocona_
..LFB2: .LFB2:
mov %edi, %edi *NOTE* movsd a(%rip), %xmm1
movlpd a(%rip), %xmm1 mov %edi, %edi *NOTE*
cvtsi2sdq %rdi, %xmm0 cvtsi2sdq %rdi, %xmm0
sqrtsd %xmm0, %xmm0 sqrtsd %xmm0, %xmm0
addsd %xmm0, %xmm1 addsd %xmm0, %xmm1
movsd %xmm1, a(%rip) movsd %xmm1, a(%rip)
ret ret
..LFE2: .LFE2:
.size q, .-q
..globl f
.type f, @function
f:
..LFB3:
pushq %rbx
..LCFI0:
movl %edi, %ebx
..L7: movl $77, %edi
movl $77, %edi .L7_OPERFORMANCE:
call q call q
decl %ebx decl %ebx
jne .L7 jne .L7_OPERFORMANCE
popq %rbx
ret
..LFE3:
.size f, .-f
Code is:
#include <math.h>
static double a;
void q(const unsigned int b)
{
a += sqrt(b);
}
void f(int c)
{
do {
q(77);
} while (--c);
}
int main(int argc, char *argv[])
{
f(argc);
return 0;
}
Kind regards.
_____
