I've been thinking about CFLAGS such as -mmmx... Does gcc actually have the
ability to actually produce mmx instructions? 


for example: 


mmx.c:

#include <stdio.h>


main()
{
	short x[100000];
	int i;
	
 	for (i=0; i < 100000; i++)
	{
		x[i] += 5;
 
 	}   
	
	
}

gcc -mmmx -O3 --save-temps -funroll-all-loops mmx.c
gives


	.file	"mmx.c"
	.def	___main;	.scl	2;	.type	32;	.endef
	.text
	.align 2
	.p2align 4,,15
.globl _main
	.def	_main;	.scl	2;	.type	32;	.endef
_main:
	pushl	%ebp
	movl	$200008, %eax
	movl	%esp, %ebp
	call	__alloca
	xorl	%eax, %eax
	andl	$-16, %esp
	call	__alloca
	call	___main
	xorl	%eax, %eax
	.p2align 4,,7
L6:
	addw	$5, -200008(%ebp,%eax,2)
	addw	$5, -200006(%ebp,%eax,2)
	addw	$5, -200004(%ebp,%eax,2)
	addw	$5, -200002(%ebp,%eax,2)
	addw	$5, -200000(%ebp,%eax,2)
	addw	$5, -199998(%ebp,%eax,2)
	addw	$5, -199996(%ebp,%eax,2)
	addw	$5, -199994(%ebp,%eax,2)
	addw	$5, -199992(%ebp,%eax,2)
	addw	$5, -199990(%ebp,%eax,2)
	addl	$10, %eax
	cmpl	$99999, %eax
	jle	L6
	leave
	ret




Shouldn't take too much imagination to see how that could be improved with mmx...

Likewise:



fu(int *x)
{
	if (*x == 5)
	  *x = 10;
}



with gcc -march=athlon-xp -O3 --save-temps -funroll-all-loops mmx.c

gives 



_fu:
	pushl	%ebp
	movl	%esp, %ebp
	movl	8(%ebp), %eax
	cmpl	$5, (%eax)
	je	L33
L32:
	leave
	ret
	.p2align 6,,7
L33:
	movl	$10, (%eax)
	jmp	L32


the -march=athlon-xp parameter should imply cmov instruction...(?)


(it doesn't even use muliple leave/ret to eliminate an extra jmp...)



Do the gcc developers include these fancy parameters just to give us ricers a
false sense of speed?


any thoughts?



--
gentoo-performance@g.o mailing list