Re: AVR byte swap optimization

[Rask Ingemann Lambertsen <rask at sygehus dot dk>]

On Fri, Nov 17, 2006 at 04:30:53PM -0700, Shaun Jackman wrote:
> Is there anything I can do to help GCC along here? I'm using GCC 4.1.0
> with -O2.
> 
> I won't bother to show bswap_32 here, which produces a real disaster!
> Think 47 instructions, for what should be 6.

   You may get better code if you write it something like this:

#include <string.h>

uint32_t bswap_32 (uint32_t x)
{
	unsigned char c[4], temp;

	memcpy (c, &x, 4);
	temp = c[0];
	c[0] = c[3];
	c[3] = temp;
	temp = c[1];
	c[1] = c[2];
	c[2] = temp;
	memcpy (&x, c, 4);

	return (x);
}

   It isn't only on the AVR that bswap_32() is nontrivial to get right.
These two versions would rule on the i386 if GCC would be just a little bit
smarter:

#include <string.h>
#define BSWAP_16(x) \
	((((x) >> 8) & 0xff) | (((x) & 0xff) << 8))

uint32_t bswap_32_a (uint32_t y)
{
	uint16_t d[2];
	uint32_t x = y;

	memcpy (d, &x, sizeof (d));
	d[0] = BSWAP_16 (d[0]);
	memcpy (&x, d, sizeof (x));
	x = ((x >> 16) & 0xffff) | ((x & 0xffff) << 16);
	memcpy (d, &x, sizeof (d));
	d[0] = BSWAP_16 (d[0]);
	memcpy (&x, d, sizeof (x));

	return (x);
}
/*
bswap_32_a:
	subl	$16, %esp
	movl	20(%esp), %eax
	movl	%eax, 12(%esp)
	rolw	$8, 12(%esp)
	roll	$16, 12(%esp)
	rolw	$8, 12(%esp)
	movl	12(%esp), %eax
	addl	$16, %esp
	ret
*/
uint32_t bswap_32_b (uint32_t y)
{
	union { uint16_t d[2]; uint32_t x; } t;
	t.x = y;

	t.d[0] = BSWAP_16 (t.d[0]);
	t.x = ((t.x >> 16) & 0xffff) | ((t.x & 0xffff) << 16);
	t.d[0] = BSWAP_16 (t.d[0]);

	return (t.x);
}
/*
bswap_32_b:
	movl	4(%esp), %edx
	movl	%edx, %eax
	rolw	$8, %ax
	movw	%ax, %dx
	movl	%edx, %eax
	roll	$16, %eax
	movl	%eax, %edx
	rolw	$8, %ax
	movw	%ax, %dx
	movl	%edx, %eax
	ret
*/

-- 
Rask Ingemann Lambertsen