xref: /openbmc/linux/arch/sparc/lib/memset.S (revision 74cad25c)
11da177e4SLinus Torvalds/* linux/arch/sparc/lib/memset.S: Sparc optimized memset, bzero and clear_user code
21da177e4SLinus Torvalds * Copyright (C) 1991,1996 Free Software Foundation
31da177e4SLinus Torvalds * Copyright (C) 1996,1997 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
41da177e4SLinus Torvalds * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu)
51da177e4SLinus Torvalds *
674cad25cSAndreas Larsson * Calls to memset returns initial %o0. Calls to bzero returns 0, if ok, and
774cad25cSAndreas Larsson * number of bytes not yet set if exception occurs and we were called as
874cad25cSAndreas Larsson * clear_user.
91da177e4SLinus Torvalds */
101da177e4SLinus Torvalds
111da177e4SLinus Torvalds#include <asm/ptrace.h>
121da177e4SLinus Torvalds
131da177e4SLinus Torvalds/* Work around cpp -rob */
141da177e4SLinus Torvalds#define ALLOC #alloc
151da177e4SLinus Torvalds#define EXECINSTR #execinstr
161da177e4SLinus Torvalds#define EX(x,y,a,b) 				\
171da177e4SLinus Torvalds98: 	x,y;					\
181da177e4SLinus Torvalds	.section .fixup,ALLOC,EXECINSTR;	\
191da177e4SLinus Torvalds	.align	4;				\
201da177e4SLinus Torvalds99:	ba 30f;					\
211da177e4SLinus Torvalds	 a, b, %o0;				\
221da177e4SLinus Torvalds	.section __ex_table,ALLOC;		\
231da177e4SLinus Torvalds	.align	4;				\
241da177e4SLinus Torvalds	.word	98b, 99b;			\
251da177e4SLinus Torvalds	.text;					\
261da177e4SLinus Torvalds	.align	4
271da177e4SLinus Torvalds
281da177e4SLinus Torvalds#define EXT(start,end,handler) 			\
291da177e4SLinus Torvalds	.section __ex_table,ALLOC;		\
301da177e4SLinus Torvalds	.align	4;				\
311da177e4SLinus Torvalds	.word	start, 0, end, handler;		\
321da177e4SLinus Torvalds	.text;					\
331da177e4SLinus Torvalds	.align	4
341da177e4SLinus Torvalds
351da177e4SLinus Torvalds/* Please don't change these macros, unless you change the logic
361da177e4SLinus Torvalds * in the .fixup section below as well.
371da177e4SLinus Torvalds * Store 64 bytes at (BASE + OFFSET) using value SOURCE. */
381da177e4SLinus Torvalds#define ZERO_BIG_BLOCK(base, offset, source)    \
391da177e4SLinus Torvalds	std	source, [base + offset + 0x00]; \
401da177e4SLinus Torvalds	std	source, [base + offset + 0x08]; \
411da177e4SLinus Torvalds	std	source, [base + offset + 0x10]; \
421da177e4SLinus Torvalds	std	source, [base + offset + 0x18]; \
431da177e4SLinus Torvalds	std	source, [base + offset + 0x20]; \
441da177e4SLinus Torvalds	std	source, [base + offset + 0x28]; \
451da177e4SLinus Torvalds	std	source, [base + offset + 0x30]; \
461da177e4SLinus Torvalds	std	source, [base + offset + 0x38];
471da177e4SLinus Torvalds
481da177e4SLinus Torvalds#define ZERO_LAST_BLOCKS(base, offset, source)	\
491da177e4SLinus Torvalds	std	source, [base - offset - 0x38]; \
501da177e4SLinus Torvalds	std	source, [base - offset - 0x30]; \
511da177e4SLinus Torvalds	std	source, [base - offset - 0x28]; \
521da177e4SLinus Torvalds	std	source, [base - offset - 0x20]; \
531da177e4SLinus Torvalds	std	source, [base - offset - 0x18]; \
541da177e4SLinus Torvalds	std	source, [base - offset - 0x10]; \
551da177e4SLinus Torvalds	std	source, [base - offset - 0x08]; \
561da177e4SLinus Torvalds	std	source, [base - offset - 0x00];
571da177e4SLinus Torvalds
581da177e4SLinus Torvalds	.text
591da177e4SLinus Torvalds	.align 4
601da177e4SLinus Torvalds
611da177e4SLinus Torvalds        .globl  __bzero_begin
621da177e4SLinus Torvalds__bzero_begin:
631da177e4SLinus Torvalds
644d14a459SDavid S. Miller	.globl	__bzero
651da177e4SLinus Torvalds	.globl	memset
661da177e4SLinus Torvalds	.globl	__memset_start, __memset_end
671da177e4SLinus Torvalds__memset_start:
681da177e4SLinus Torvaldsmemset:
6974cad25cSAndreas Larsson	mov	%o0, %g1
7074cad25cSAndreas Larsson	mov	1, %g4
711da177e4SLinus Torvalds	and	%o1, 0xff, %g3
721da177e4SLinus Torvalds	sll	%g3, 8, %g2
731da177e4SLinus Torvalds	or	%g3, %g2, %g3
741da177e4SLinus Torvalds	sll	%g3, 16, %g2
751da177e4SLinus Torvalds	or	%g3, %g2, %g3
761da177e4SLinus Torvalds	b	1f
771da177e4SLinus Torvalds	 mov	%o2, %o1
781da177e4SLinus Torvalds3:
791da177e4SLinus Torvalds	cmp	%o2, 3
801da177e4SLinus Torvalds	be	2f
811da177e4SLinus Torvalds	 EX(stb	%g3, [%o0], sub %o1, 0)
821da177e4SLinus Torvalds
831da177e4SLinus Torvalds	cmp	%o2, 2
841da177e4SLinus Torvalds	be	2f
851da177e4SLinus Torvalds	 EX(stb	%g3, [%o0 + 0x01], sub %o1, 1)
861da177e4SLinus Torvalds
871da177e4SLinus Torvalds	EX(stb	%g3, [%o0 + 0x02], sub %o1, 2)
881da177e4SLinus Torvalds2:
891da177e4SLinus Torvalds	sub	%o2, 4, %o2
901da177e4SLinus Torvalds	add	%o1, %o2, %o1
911da177e4SLinus Torvalds	b	4f
921da177e4SLinus Torvalds	 sub	%o0, %o2, %o0
931da177e4SLinus Torvalds
941da177e4SLinus Torvalds__bzero:
9574cad25cSAndreas Larsson	clr	%g4
961da177e4SLinus Torvalds	mov	%g0, %g3
971da177e4SLinus Torvalds1:
981da177e4SLinus Torvalds	cmp	%o1, 7
991da177e4SLinus Torvalds	bleu	7f
1001da177e4SLinus Torvalds	 andcc	%o0, 3, %o2
1011da177e4SLinus Torvalds
1021da177e4SLinus Torvalds	bne	3b
1031da177e4SLinus Torvalds4:
1041da177e4SLinus Torvalds	 andcc	%o0, 4, %g0
1051da177e4SLinus Torvalds
1061da177e4SLinus Torvalds	be	2f
1071da177e4SLinus Torvalds	 mov	%g3, %g2
1081da177e4SLinus Torvalds
1091da177e4SLinus Torvalds	EX(st	%g3, [%o0], sub %o1, 0)
1101da177e4SLinus Torvalds	sub	%o1, 4, %o1
1111da177e4SLinus Torvalds	add	%o0, 4, %o0
1121da177e4SLinus Torvalds2:
1131da177e4SLinus Torvalds	andcc	%o1, 0xffffff80, %o3	! Now everything is 8 aligned and o1 is len to run
1141da177e4SLinus Torvalds	be	9f
1151da177e4SLinus Torvalds	 andcc	%o1, 0x78, %o2
1161da177e4SLinus Torvalds10:
1171da177e4SLinus Torvalds	ZERO_BIG_BLOCK(%o0, 0x00, %g2)
1181da177e4SLinus Torvalds	subcc	%o3, 128, %o3
1191da177e4SLinus Torvalds	ZERO_BIG_BLOCK(%o0, 0x40, %g2)
1201da177e4SLinus Torvalds11:
1211da177e4SLinus Torvalds	EXT(10b, 11b, 20f)
1221da177e4SLinus Torvalds	bne	10b
1231da177e4SLinus Torvalds	 add	%o0, 128, %o0
1241da177e4SLinus Torvalds
1251da177e4SLinus Torvalds	orcc	%o2, %g0, %g0
1261da177e4SLinus Torvalds9:
1271da177e4SLinus Torvalds	be	13f
1281da177e4SLinus Torvalds	 andcc	%o1, 7, %o1
1291da177e4SLinus Torvalds
1301da177e4SLinus Torvalds	srl	%o2, 1, %o3
1311da177e4SLinus Torvalds	set	13f, %o4
1321da177e4SLinus Torvalds	sub	%o4, %o3, %o4
1331da177e4SLinus Torvalds	jmp	%o4
1341da177e4SLinus Torvalds	 add	%o0, %o2, %o0
1351da177e4SLinus Torvalds
1361da177e4SLinus Torvalds12:
1371da177e4SLinus Torvalds	ZERO_LAST_BLOCKS(%o0, 0x48, %g2)
1381da177e4SLinus Torvalds	ZERO_LAST_BLOCKS(%o0, 0x08, %g2)
1391da177e4SLinus Torvalds13:
1401da177e4SLinus Torvalds	be	8f
1411da177e4SLinus Torvalds	 andcc	%o1, 4, %g0
1421da177e4SLinus Torvalds
1431da177e4SLinus Torvalds	be	1f
1441da177e4SLinus Torvalds	 andcc	%o1, 2, %g0
1451da177e4SLinus Torvalds
1461da177e4SLinus Torvalds	EX(st	%g3, [%o0], and %o1, 7)
1471da177e4SLinus Torvalds	add	%o0, 4, %o0
1481da177e4SLinus Torvalds1:
1491da177e4SLinus Torvalds	be	1f
1501da177e4SLinus Torvalds	 andcc	%o1, 1, %g0
1511da177e4SLinus Torvalds
1521da177e4SLinus Torvalds	EX(sth	%g3, [%o0], and %o1, 3)
1531da177e4SLinus Torvalds	add	%o0, 2, %o0
1541da177e4SLinus Torvalds1:
1551da177e4SLinus Torvalds	bne,a	8f
1561da177e4SLinus Torvalds	 EX(stb	%g3, [%o0], and %o1, 1)
1571da177e4SLinus Torvalds8:
15874cad25cSAndreas Larsson	b	0f
15974cad25cSAndreas Larsson	 nop
1601da177e4SLinus Torvalds7:
1611da177e4SLinus Torvalds	be	13b
1621da177e4SLinus Torvalds	 orcc	%o1, 0, %g0
1631da177e4SLinus Torvalds
1641da177e4SLinus Torvalds	be	0f
1651da177e4SLinus Torvalds8:
1661da177e4SLinus Torvalds	 add	%o0, 1, %o0
1671da177e4SLinus Torvalds	subcc	%o1, 1, %o1
168f61698e6SAlexander Shmelev	bne	8b
1691da177e4SLinus Torvalds	 EX(stb	%g3, [%o0 - 1], add %o1, 1)
1701da177e4SLinus Torvalds0:
17174cad25cSAndreas Larsson	andcc	%g4, 1, %g0
17274cad25cSAndreas Larsson	be	5f
17374cad25cSAndreas Larsson	 nop
17474cad25cSAndreas Larsson	retl
17574cad25cSAndreas Larsson	 mov	%g1, %o0
17674cad25cSAndreas Larsson5:
1771da177e4SLinus Torvalds	retl
1781da177e4SLinus Torvalds	 clr	%o0
1791da177e4SLinus Torvalds__memset_end:
1801da177e4SLinus Torvalds
1811da177e4SLinus Torvalds	.section .fixup,#alloc,#execinstr
1821da177e4SLinus Torvalds	.align	4
1831da177e4SLinus Torvalds20:
1841da177e4SLinus Torvalds	cmp	%g2, 8
1851da177e4SLinus Torvalds	bleu	1f
1861da177e4SLinus Torvalds	 and	%o1, 0x7f, %o1
1871da177e4SLinus Torvalds	sub	%g2, 9, %g2
1881da177e4SLinus Torvalds	add	%o3, 64, %o3
1891da177e4SLinus Torvalds1:
1901da177e4SLinus Torvalds	sll	%g2, 3, %g2
1911da177e4SLinus Torvalds	add	%o3, %o1, %o0
1921da177e4SLinus Torvalds	b 30f
1931da177e4SLinus Torvalds	 sub	%o0, %g2, %o0
1941da177e4SLinus Torvalds21:
1951da177e4SLinus Torvalds	mov	8, %o0
1961da177e4SLinus Torvalds	and	%o1, 7, %o1
1971da177e4SLinus Torvalds	sub	%o0, %g2, %o0
1981da177e4SLinus Torvalds	sll	%o0, 3, %o0
1991da177e4SLinus Torvalds	b 30f
2001da177e4SLinus Torvalds	 add	%o0, %o1, %o0
2011da177e4SLinus Torvalds30:
2021da177e4SLinus Torvalds/* %o4 is faulting address, %o5 is %pc where fault occurred */
2031da177e4SLinus Torvalds	save	%sp, -104, %sp
2041da177e4SLinus Torvalds	mov	%i5, %o0
2051da177e4SLinus Torvalds	mov	%i7, %o1
2061da177e4SLinus Torvalds	call	lookup_fault
2071da177e4SLinus Torvalds	 mov	%i4, %o2
2081da177e4SLinus Torvalds	ret
2091da177e4SLinus Torvalds	 restore
2101da177e4SLinus Torvalds
2111da177e4SLinus Torvalds	.globl __bzero_end
2121da177e4SLinus Torvalds__bzero_end:
213