/* $Id: blockops.S,v 1.7 1997/05/20 07:58:28 jj Exp $
 * blockops.S: Common block zero optimized routines.
 *
 * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu)
 */

#include <asm/cprefix.h>

	/* Zero out 64 bytes of memory at (buf + offset).
	 * Assumes %g1 contains zero.
	 */
#define BLAST_BLOCK(buf, offset) \
	std	%g0, [buf + offset + 0x38]; \
	std	%g0, [buf + offset + 0x30]; \
	std	%g0, [buf + offset + 0x28]; \
	std	%g0, [buf + offset + 0x20]; \
	std	%g0, [buf + offset + 0x18]; \
	std	%g0, [buf + offset + 0x10]; \
	std	%g0, [buf + offset + 0x08]; \
	std	%g0, [buf + offset + 0x00];

	/* Copy 32 bytes of memory at (src + offset) to
	 * (dst + offset).
	 */
#define MIRROR_BLOCK(dst, src, offset, t0, t1, t2, t3, t4, t5, t6, t7) \
	ldd	[src + offset + 0x18], t0; \
	ldd	[src + offset + 0x10], t2; \
	ldd	[src + offset + 0x08], t4; \
	ldd	[src + offset + 0x00], t6; \
	std	t0, [dst + offset + 0x18]; \
	std	t2, [dst + offset + 0x10]; \
	std	t4, [dst + offset + 0x08]; \
	std	t6, [dst + offset + 0x00];

	/* Profiling evidence indicates that memset() is
	 * commonly called for blocks of size PAGE_SIZE,
	 * and (2 * PAGE_SIZE) (for kernel stacks)
	 * and with a second arg of zero.  We assume in
	 * all of these cases that the buffer is aligned
	 * on at least an 8 byte boundry.
	 *
	 * Therefore we special case them to make them
	 * as fast as possible.
	 */

	.text
	.align	4
	.globl	C_LABEL(bzero_1page), C_LABEL(__copy_1page)

C_LABEL(bzero_1page):
/* NOTE: If you change the number of insns of this routine, please check
 * arch/sparc/mm/hypersparc.S */
	/* %o0 = buf */
	or	%g0, %g0, %g1
	or	%o0, %g0, %o1
	or	%g0, 0x10, %g2
1:
	BLAST_BLOCK(%o0, 0x00)
	BLAST_BLOCK(%o0, 0x40)
	BLAST_BLOCK(%o0, 0x80)
	BLAST_BLOCK(%o0, 0xc0)
	subcc	%g2, 1, %g2
	bne	1b
	 add	%o0, 0x100, %o0

	retl
	 nop

C_LABEL(__copy_1page):
/* NOTE: If you change the number of insns of this routine, please check
 * arch/sparc/mm/hypersparc.S */
	/* %o0 = dst, %o1 = src */
	or	%g0, 0x10, %g1
1:
	MIRROR_BLOCK(%o0, %o1, 0x00, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5)
	MIRROR_BLOCK(%o0, %o1, 0x20, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5)
	MIRROR_BLOCK(%o0, %o1, 0x40, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5)
	MIRROR_BLOCK(%o0, %o1, 0x60, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5)
	MIRROR_BLOCK(%o0, %o1, 0x80, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5)
	MIRROR_BLOCK(%o0, %o1, 0xa0, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5)
	MIRROR_BLOCK(%o0, %o1, 0xc0, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5)
	MIRROR_BLOCK(%o0, %o1, 0xe0, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5)
	subcc	%g1, 1, %g1
	add	%o0, 0x100, %o0
	bne	1b
	 add	%o1, 0x100, %o1

	retl
	 nop