xref: /openbmc/linux/arch/sparc/lib/memset.S (revision e51ea544)
1b2441318SGreg Kroah-Hartman/* SPDX-License-Identifier: GPL-2.0 */
21da177e4SLinus Torvalds/* linux/arch/sparc/lib/memset.S: Sparc optimized memset, bzero and clear_user code
31da177e4SLinus Torvalds * Copyright (C) 1991,1996 Free Software Foundation
41da177e4SLinus Torvalds * Copyright (C) 1996,1997 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
51da177e4SLinus Torvalds * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu)
61da177e4SLinus Torvalds *
774cad25cSAndreas Larsson * Calls to memset returns initial %o0. Calls to bzero returns 0, if ok, and
874cad25cSAndreas Larsson * number of bytes not yet set if exception occurs and we were called as
974cad25cSAndreas Larsson * clear_user.
101da177e4SLinus Torvalds */
111da177e4SLinus Torvalds
121da177e4SLinus Torvalds#include <asm/ptrace.h>
13d3867f04SAl Viro#include <asm/export.h>
141da177e4SLinus Torvalds
151da177e4SLinus Torvalds/* Work around cpp -rob */
161da177e4SLinus Torvalds#define ALLOC #alloc
171da177e4SLinus Torvalds#define EXECINSTR #execinstr
181da177e4SLinus Torvalds#define EX(x,y,a,b) 				\
191da177e4SLinus Torvalds98: 	x,y;					\
201da177e4SLinus Torvalds	.section .fixup,ALLOC,EXECINSTR;	\
211da177e4SLinus Torvalds	.align	4;				\
22*e51ea544SAl Viro99:	retl;					\
231da177e4SLinus Torvalds	 a, b, %o0;				\
241da177e4SLinus Torvalds	.section __ex_table,ALLOC;		\
251da177e4SLinus Torvalds	.align	4;				\
261da177e4SLinus Torvalds	.word	98b, 99b;			\
271da177e4SLinus Torvalds	.text;					\
281da177e4SLinus Torvalds	.align	4
291da177e4SLinus Torvalds
301da177e4SLinus Torvalds#define EXT(start,end,handler) 			\
311da177e4SLinus Torvalds	.section __ex_table,ALLOC;		\
321da177e4SLinus Torvalds	.align	4;				\
331da177e4SLinus Torvalds	.word	start, 0, end, handler;		\
341da177e4SLinus Torvalds	.text;					\
351da177e4SLinus Torvalds	.align	4
361da177e4SLinus Torvalds
371da177e4SLinus Torvalds/* Please don't change these macros, unless you change the logic
381da177e4SLinus Torvalds * in the .fixup section below as well.
391da177e4SLinus Torvalds * Store 64 bytes at (BASE + OFFSET) using value SOURCE. */
401da177e4SLinus Torvalds#define ZERO_BIG_BLOCK(base, offset, source)    \
411da177e4SLinus Torvalds	std	source, [base + offset + 0x00]; \
421da177e4SLinus Torvalds	std	source, [base + offset + 0x08]; \
431da177e4SLinus Torvalds	std	source, [base + offset + 0x10]; \
441da177e4SLinus Torvalds	std	source, [base + offset + 0x18]; \
451da177e4SLinus Torvalds	std	source, [base + offset + 0x20]; \
461da177e4SLinus Torvalds	std	source, [base + offset + 0x28]; \
471da177e4SLinus Torvalds	std	source, [base + offset + 0x30]; \
481da177e4SLinus Torvalds	std	source, [base + offset + 0x38];
491da177e4SLinus Torvalds
501da177e4SLinus Torvalds#define ZERO_LAST_BLOCKS(base, offset, source)	\
511da177e4SLinus Torvalds	std	source, [base - offset - 0x38]; \
521da177e4SLinus Torvalds	std	source, [base - offset - 0x30]; \
531da177e4SLinus Torvalds	std	source, [base - offset - 0x28]; \
541da177e4SLinus Torvalds	std	source, [base - offset - 0x20]; \
551da177e4SLinus Torvalds	std	source, [base - offset - 0x18]; \
561da177e4SLinus Torvalds	std	source, [base - offset - 0x10]; \
571da177e4SLinus Torvalds	std	source, [base - offset - 0x08]; \
581da177e4SLinus Torvalds	std	source, [base - offset - 0x00];
591da177e4SLinus Torvalds
601da177e4SLinus Torvalds	.text
611da177e4SLinus Torvalds	.align 4
621da177e4SLinus Torvalds
631da177e4SLinus Torvalds        .globl  __bzero_begin
641da177e4SLinus Torvalds__bzero_begin:
651da177e4SLinus Torvalds
664d14a459SDavid S. Miller	.globl	__bzero
67f5a651f1SNagarathnam Muthusamy	.type	__bzero,#function
681da177e4SLinus Torvalds	.globl	memset
69d3867f04SAl Viro	EXPORT_SYMBOL(__bzero)
70d3867f04SAl Viro	EXPORT_SYMBOL(memset)
711da177e4SLinus Torvalds	.globl	__memset_start, __memset_end
721da177e4SLinus Torvalds__memset_start:
731da177e4SLinus Torvaldsmemset:
7474cad25cSAndreas Larsson	mov	%o0, %g1
7574cad25cSAndreas Larsson	mov	1, %g4
761da177e4SLinus Torvalds	and	%o1, 0xff, %g3
771da177e4SLinus Torvalds	sll	%g3, 8, %g2
781da177e4SLinus Torvalds	or	%g3, %g2, %g3
791da177e4SLinus Torvalds	sll	%g3, 16, %g2
801da177e4SLinus Torvalds	or	%g3, %g2, %g3
811da177e4SLinus Torvalds	b	1f
821da177e4SLinus Torvalds	 mov	%o2, %o1
831da177e4SLinus Torvalds3:
841da177e4SLinus Torvalds	cmp	%o2, 3
851da177e4SLinus Torvalds	be	2f
861da177e4SLinus Torvalds	 EX(stb	%g3, [%o0], sub %o1, 0)
871da177e4SLinus Torvalds
881da177e4SLinus Torvalds	cmp	%o2, 2
891da177e4SLinus Torvalds	be	2f
901da177e4SLinus Torvalds	 EX(stb	%g3, [%o0 + 0x01], sub %o1, 1)
911da177e4SLinus Torvalds
921da177e4SLinus Torvalds	EX(stb	%g3, [%o0 + 0x02], sub %o1, 2)
931da177e4SLinus Torvalds2:
941da177e4SLinus Torvalds	sub	%o2, 4, %o2
951da177e4SLinus Torvalds	add	%o1, %o2, %o1
961da177e4SLinus Torvalds	b	4f
971da177e4SLinus Torvalds	 sub	%o0, %o2, %o0
981da177e4SLinus Torvalds
991da177e4SLinus Torvalds__bzero:
10074cad25cSAndreas Larsson	clr	%g4
1011da177e4SLinus Torvalds	mov	%g0, %g3
1021da177e4SLinus Torvalds1:
1031da177e4SLinus Torvalds	cmp	%o1, 7
1041da177e4SLinus Torvalds	bleu	7f
1051da177e4SLinus Torvalds	 andcc	%o0, 3, %o2
1061da177e4SLinus Torvalds
1071da177e4SLinus Torvalds	bne	3b
1081da177e4SLinus Torvalds4:
1091da177e4SLinus Torvalds	 andcc	%o0, 4, %g0
1101da177e4SLinus Torvalds
1111da177e4SLinus Torvalds	be	2f
1121da177e4SLinus Torvalds	 mov	%g3, %g2
1131da177e4SLinus Torvalds
1141da177e4SLinus Torvalds	EX(st	%g3, [%o0], sub %o1, 0)
1151da177e4SLinus Torvalds	sub	%o1, 4, %o1
1161da177e4SLinus Torvalds	add	%o0, 4, %o0
1171da177e4SLinus Torvalds2:
1181da177e4SLinus Torvalds	andcc	%o1, 0xffffff80, %o3	! Now everything is 8 aligned and o1 is len to run
1191da177e4SLinus Torvalds	be	9f
1201da177e4SLinus Torvalds	 andcc	%o1, 0x78, %o2
1211da177e4SLinus Torvalds10:
1221da177e4SLinus Torvalds	ZERO_BIG_BLOCK(%o0, 0x00, %g2)
1231da177e4SLinus Torvalds	subcc	%o3, 128, %o3
1241da177e4SLinus Torvalds	ZERO_BIG_BLOCK(%o0, 0x40, %g2)
1251da177e4SLinus Torvalds11:
1261da177e4SLinus Torvalds	EXT(10b, 11b, 20f)
1271da177e4SLinus Torvalds	bne	10b
1281da177e4SLinus Torvalds	 add	%o0, 128, %o0
1291da177e4SLinus Torvalds
1301da177e4SLinus Torvalds	orcc	%o2, %g0, %g0
1311da177e4SLinus Torvalds9:
1321da177e4SLinus Torvalds	be	13f
1331da177e4SLinus Torvalds	 andcc	%o1, 7, %o1
1341da177e4SLinus Torvalds
1351da177e4SLinus Torvalds	srl	%o2, 1, %o3
1361da177e4SLinus Torvalds	set	13f, %o4
1371da177e4SLinus Torvalds	sub	%o4, %o3, %o4
1381da177e4SLinus Torvalds	jmp	%o4
1391da177e4SLinus Torvalds	 add	%o0, %o2, %o0
1401da177e4SLinus Torvalds
1411da177e4SLinus Torvalds12:
1421da177e4SLinus Torvalds	ZERO_LAST_BLOCKS(%o0, 0x48, %g2)
1431da177e4SLinus Torvalds	ZERO_LAST_BLOCKS(%o0, 0x08, %g2)
1441da177e4SLinus Torvalds13:
1451da177e4SLinus Torvalds	be	8f
1461da177e4SLinus Torvalds	 andcc	%o1, 4, %g0
1471da177e4SLinus Torvalds
1481da177e4SLinus Torvalds	be	1f
1491da177e4SLinus Torvalds	 andcc	%o1, 2, %g0
1501da177e4SLinus Torvalds
1511da177e4SLinus Torvalds	EX(st	%g3, [%o0], and %o1, 7)
1521da177e4SLinus Torvalds	add	%o0, 4, %o0
1531da177e4SLinus Torvalds1:
1541da177e4SLinus Torvalds	be	1f
1551da177e4SLinus Torvalds	 andcc	%o1, 1, %g0
1561da177e4SLinus Torvalds
1571da177e4SLinus Torvalds	EX(sth	%g3, [%o0], and %o1, 3)
1581da177e4SLinus Torvalds	add	%o0, 2, %o0
1591da177e4SLinus Torvalds1:
1601da177e4SLinus Torvalds	bne,a	8f
1611da177e4SLinus Torvalds	 EX(stb	%g3, [%o0], and %o1, 1)
1621da177e4SLinus Torvalds8:
16374cad25cSAndreas Larsson	b	0f
16474cad25cSAndreas Larsson	 nop
1651da177e4SLinus Torvalds7:
1661da177e4SLinus Torvalds	be	13b
1671da177e4SLinus Torvalds	 orcc	%o1, 0, %g0
1681da177e4SLinus Torvalds
1691da177e4SLinus Torvalds	be	0f
1701da177e4SLinus Torvalds8:
1711da177e4SLinus Torvalds	 add	%o0, 1, %o0
1721da177e4SLinus Torvalds	subcc	%o1, 1, %o1
173f61698e6SAlexander Shmelev	bne	8b
1741da177e4SLinus Torvalds	 EX(stb	%g3, [%o0 - 1], add %o1, 1)
1751da177e4SLinus Torvalds0:
17674cad25cSAndreas Larsson	andcc	%g4, 1, %g0
17774cad25cSAndreas Larsson	be	5f
17874cad25cSAndreas Larsson	 nop
17974cad25cSAndreas Larsson	retl
18074cad25cSAndreas Larsson	 mov	%g1, %o0
18174cad25cSAndreas Larsson5:
1821da177e4SLinus Torvalds	retl
1831da177e4SLinus Torvalds	 clr	%o0
1841da177e4SLinus Torvalds__memset_end:
1851da177e4SLinus Torvalds
1861da177e4SLinus Torvalds	.section .fixup,#alloc,#execinstr
1871da177e4SLinus Torvalds	.align	4
1881da177e4SLinus Torvalds20:
1891da177e4SLinus Torvalds	cmp	%g2, 8
1901da177e4SLinus Torvalds	bleu	1f
1911da177e4SLinus Torvalds	 and	%o1, 0x7f, %o1
1921da177e4SLinus Torvalds	sub	%g2, 9, %g2
1931da177e4SLinus Torvalds	add	%o3, 64, %o3
1941da177e4SLinus Torvalds1:
1951da177e4SLinus Torvalds	sll	%g2, 3, %g2
1961da177e4SLinus Torvalds	add	%o3, %o1, %o0
197*e51ea544SAl Viro	retl
1981da177e4SLinus Torvalds	 sub	%o0, %g2, %o0
1991da177e4SLinus Torvalds21:
2001da177e4SLinus Torvalds	mov	8, %o0
2011da177e4SLinus Torvalds	and	%o1, 7, %o1
2021da177e4SLinus Torvalds	sub	%o0, %g2, %o0
2031da177e4SLinus Torvalds	sll	%o0, 3, %o0
204*e51ea544SAl Viro	retl
2051da177e4SLinus Torvalds	 add	%o0, %o1, %o0
2061da177e4SLinus Torvalds
2071da177e4SLinus Torvalds	.globl __bzero_end
2081da177e4SLinus Torvalds__bzero_end:
209