Blame sysdeps/sh/memcpy.S

Packit 6c4009
/* Copyright (C) 1999-2018 Free Software Foundation, Inc.
Packit 6c4009
   This file is part of the GNU C Library.
Packit 6c4009
   Contributed by Kazumoto Kojima <kkojima@rr.iij4u.or.jp>
Packit 6c4009
   Optimized by Toshiyasu Morita <toshiyasu.morita@hsa.hitachi.com>
Packit 6c4009
Packit 6c4009
   The GNU C Library is free software; you can redistribute it and/or
Packit 6c4009
   modify it under the terms of the GNU Lesser General Public
Packit 6c4009
   License as published by the Free Software Foundation; either
Packit 6c4009
   version 2.1 of the License, or (at your option) any later version.
Packit 6c4009
Packit 6c4009
   The GNU C Library is distributed in the hope that it will be useful,
Packit 6c4009
   but WITHOUT ANY WARRANTY; without even the implied warranty of
Packit 6c4009
   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
Packit 6c4009
   Lesser General Public License for more details.
Packit 6c4009
Packit 6c4009
   You should have received a copy of the GNU Lesser General Public
Packit 6c4009
   License along with the GNU C Library; if not, see
Packit 6c4009
   <http://www.gnu.org/licenses/>.  */
Packit 6c4009
Packit 6c4009
#include <sysdep.h>
Packit 6c4009
Packit 6c4009
/* void *memcpy(void *dst, const void *src, size_t n);
Packit 6c4009
    No overlap between the memory of DST and of SRC are assumed.  */
Packit 6c4009
Packit 6c4009
ENTRY(memcpy)
Packit 6c4009
	mov	r4,r3		/* Save destination.  */
Packit 6c4009
Packit 6c4009
	/* If less than 11 bytes, just do a byte copy.  */
Packit 6c4009
	mov	#11,r0
Packit 6c4009
	cmp/gt	r6,r0
Packit 6c4009
	bt	L_byteloop_init
Packit 6c4009
Packit 6c4009
	/* Check if we need to word-align source.  */
Packit 6c4009
	mov	r5,r0
Packit 6c4009
	tst	#1,r0
Packit 6c4009
	bt	L_wordalign
Packit 6c4009
Packit 6c4009
	mov.b	@r0+,r1		/* Copy one byte.  */
Packit 6c4009
	add	#-1,r6
Packit 6c4009
	mov.b	r1,@r4
Packit 6c4009
	add	#1,r4
Packit 6c4009
Packit 6c4009
	.balignw 4,0x0009
Packit 6c4009
L_wordalign:
Packit 6c4009
	/* Check if we need to longword-align source.  */
Packit 6c4009
	tst	#2,r0
Packit 6c4009
	bt	L_copy
Packit 6c4009
Packit 6c4009
	mov.w	@r0+,r1		/* Copy one word.  */
Packit 6c4009
	add	#-2,r6
Packit 6c4009
#ifdef __BIG_ENDIAN__
Packit 6c4009
	add	#1,r4
Packit 6c4009
	mov.b	r1,@r4
Packit 6c4009
	shlr8	r1
Packit 6c4009
	mov.b	r1,@-r4
Packit 6c4009
	add	#2,r4
Packit 6c4009
#else
Packit 6c4009
	mov.b	r1,@r4
Packit 6c4009
	add	#1,r4
Packit 6c4009
	shlr8	r1
Packit 6c4009
	mov.b	r1,@r4
Packit 6c4009
	add	#1,r4
Packit 6c4009
#endif
Packit 6c4009
L_copy:
Packit 6c4009
	mov	r0,r5
Packit 6c4009
Packit 6c4009
	/* Calculate the correct routine to handle the destination
Packit 6c4009
	   alignment and simultaneously calculate the loop counts for
Packit 6c4009
	   both the 2 word copy loop and byte copy loop.  */
Packit 6c4009
	mova	L_jumptable,r0
Packit 6c4009
	mov	r0,r1
Packit 6c4009
	mov	r4,r0
Packit 6c4009
	mov	r6,r7
Packit 6c4009
	and	#3,r0
Packit 6c4009
	shlr2	r7
Packit 6c4009
	shll	r0
Packit 6c4009
	shlr	r7
Packit 6c4009
	mov.w	@(r0,r1),r2
Packit 6c4009
	mov	#7,r0
Packit 6c4009
	braf	r2
Packit 6c4009
	and	r0,r6
Packit 6c4009
L_base:
Packit 6c4009
Packit 6c4009
	.balign	4
Packit 6c4009
L_jumptable:
Packit 6c4009
	.word	L_copydest0 - L_base
Packit 6c4009
	.word	L_copydest1_or_3 - L_base
Packit 6c4009
	.word	L_copydest2 - L_base
Packit 6c4009
	.word	L_copydest1_or_3 - L_base
Packit 6c4009
Packit 6c4009
	.balign	4
Packit 6c4009
	/* Copy routine for (dest mod 4) == 1 or == 3.  */
Packit 6c4009
L_copydest1_or_3:
Packit 6c4009
	add	#-1,r4
Packit 6c4009
	.balignw 4,0x0009
Packit 6c4009
L_copydest1_or_3_loop:
Packit 6c4009
	mov.l	@r5+,r0		/* Read first longword.  */
Packit 6c4009
	dt	r7
Packit 6c4009
	mov.l	@r5+,r1		/* Read second longword.  */
Packit 6c4009
#ifdef __BIG_ENDIAN__
Packit 6c4009
	/* Write first longword as byte, word, byte.  */
Packit 6c4009
	mov.b	r0,@(4,r4)
Packit 6c4009
	shlr8	r0
Packit 6c4009
	mov.w	r0,@(2,r4)
Packit 6c4009
	shlr16	r0
Packit 6c4009
	mov.b	r0,@(1,r4)
Packit 6c4009
	mov	r1,r0
Packit 6c4009
	/* Write second longword as byte, word, byte.  */
Packit 6c4009
	mov.b	r0,@(8,r4)
Packit 6c4009
	shlr8	r0
Packit 6c4009
	mov.w	r0,@(6,r4)
Packit 6c4009
	shlr16	r0
Packit 6c4009
	mov.b	r0,@(5,r4)
Packit 6c4009
#else
Packit 6c4009
	/* Write first longword as byte, word, byte.  */
Packit 6c4009
	mov.b	r0,@(1,r4)
Packit 6c4009
	shlr8	r0
Packit 6c4009
	mov.w	r0,@(2,r4)
Packit 6c4009
	shlr16	r0
Packit 6c4009
	mov.b	r0,@(4,r4)
Packit 6c4009
	mov	r1,r0
Packit 6c4009
	/* Write second longword as byte, word, byte.  */
Packit 6c4009
	mov.b	r0,@(5,r4)
Packit 6c4009
	shlr8	r0
Packit 6c4009
	mov.w	r0,@(6,r4)
Packit 6c4009
	shlr16	r0
Packit 6c4009
	mov.b	r0,@(8,r4)
Packit 6c4009
#endif
Packit 6c4009
	bf/s	L_copydest1_or_3_loop
Packit 6c4009
	add	#8,r4
Packit 6c4009
Packit 6c4009
	bra	L_byteloop_init
Packit 6c4009
	add	#1,r4
Packit 6c4009
Packit 6c4009
	.balign 4
Packit 6c4009
	/* Copy routine for (dest mod 4) == 2.  */
Packit 6c4009
L_copydest2:
Packit 6c4009
L_copydest2_loop:
Packit 6c4009
	mov.l	@r5+,r0
Packit 6c4009
	dt	r7
Packit 6c4009
	mov.l	@r5+,r1
Packit 6c4009
#ifdef __BIG_ENDIAN__
Packit 6c4009
	mov.w	r0,@(2,r4)
Packit 6c4009
	shlr16	r0
Packit 6c4009
	mov.w	r0,@r4
Packit 6c4009
	mov	r1,r0
Packit 6c4009
	mov.w	r0,@(6,r4)
Packit 6c4009
	shlr16	r0
Packit 6c4009
	mov.w	r0,@(4,r4)
Packit 6c4009
#else
Packit 6c4009
	mov.w	r0,@r4
Packit 6c4009
	shlr16	r0
Packit 6c4009
	mov.w	r0,@(2,r4)
Packit 6c4009
	mov	r1,r0
Packit 6c4009
	mov.w	r0,@(4,r4)
Packit 6c4009
	shlr16	r0
Packit 6c4009
	mov.w	r0,@(6,r4)
Packit 6c4009
#endif
Packit 6c4009
	bf/s	L_copydest2_loop
Packit 6c4009
	add	#8,r4
Packit 6c4009
Packit 6c4009
	bra	L_byteloop_init
Packit 6c4009
	nop
Packit 6c4009
Packit 6c4009
	.balign 4
Packit 6c4009
	/* Copy routine for (dest mod 4) == 0.  */
Packit 6c4009
L_copydest0:
Packit 6c4009
	add	#-8,r4
Packit 6c4009
	.balignw 4,0x0009
Packit 6c4009
L_copydest0_loop:
Packit 6c4009
	mov.l	@r5+,r0
Packit 6c4009
	dt	r7
Packit 6c4009
	mov.l	@r5+,r1
Packit 6c4009
	add	#8,r4
Packit 6c4009
	mov.l	r0,@r4
Packit 6c4009
	bf/s	L_copydest0_loop
Packit 6c4009
	mov.l	r1,@(4,r4)
Packit 6c4009
Packit 6c4009
	add	#8,r4		/* Fall through.  */
Packit 6c4009
Packit 6c4009
L_byteloop_init:
Packit 6c4009
	tst	r6,r6
Packit 6c4009
	bt	L_exit
Packit 6c4009
Packit 6c4009
	.balignw 4,0x0009
Packit 6c4009
	/* Copy remaining bytes.  */
Packit 6c4009
L_byteloop:
Packit 6c4009
	mov.b	@r5+,r0
Packit 6c4009
	dt	r6
Packit 6c4009
	mov.b	r0,@r4
Packit 6c4009
	bf/s	L_byteloop
Packit 6c4009
	add	#1,r4
Packit 6c4009
Packit 6c4009
L_exit:
Packit 6c4009
	rts
Packit 6c4009
	mov	r3,r0		/* Return destination.  */
Packit 6c4009
END(memcpy)
Packit 6c4009
libc_hidden_builtin_def (memcpy)