path: root/arch/sh/lib/memcpy-sh4.S

/*
 * "memcpy" implementation of SuperH
 *
 * Copyright (C) 1999  Niibe Yutaka
 * Copyright (c) 2002  STMicroelectronics Ltd
 *   Modified from memcpy.S and micro-optimised for SH4
 *   Stuart Menefy (stuart.menefy@st.com)
 *
 */
#include <linux/linkage.h>

/*
 * void *memcpy(void *dst, const void *src, size_t n);
 *
 * It is assumed that there is no overlap between src and dst.
 * If there is an overlap, then the results are undefined.
 */

	!
	!	GHIJ KLMN OPQR -->  ...G HIJK LMNO PQR.
	!

	! Size is 16 or greater, and may have trailing bytes

	.balign	32
.Lcase1:
	! Read a long word and write a long word at once
	! At the start of each iteration, r7 contains last long load
	add	#-1,r5		!  79 EX
	mov	r4,r2		!   5 MT (0 cycles latency)

	mov.l	@(r0,r5),r7	!  21 LS (2 cycles latency)
	add	#-4,r5		!  50 EX

	add	#7,r2		!  79 EX
	!
#ifdef CONFIG_CPU_LITTLE_ENDIAN
	! 6 cycles, 4 bytes per iteration
3:	mov.l	@(r0,r5),r1	!  21 LS (latency=2)	! NMLK
	mov	r7, r3		!   5 MT (latency=0)	! RQPO

	cmp/hi	r2,r0		!  57 MT
	shll16	r3		! 103 EX

	mov	r1,r6		!   5 MT (latency=0)
	shll8	r3		! 102 EX		! Oxxx

	shlr8	r6		! 106 EX		! xNML
	mov	r1, r7		!   5 MT (latency=0)

	or	r6,r3		!  82 EX		! ONML
	bt/s	3b		! 109 BR

	 mov.l	r3,@-r0		!  30 LS
#else
3:	mov.l	@(r0,r5),r1	!  21 LS (latency=2)	! KLMN
	mov	r7,r3		!   5 MT (latency=0)	! OPQR

	cmp/hi	r2,r0		!  57 MT
	shlr16	r3		! 107 EX

	shlr8	r3		! 106 EX		! xxxO
	mov	r1,r6		!   5 MT (latency=0)

	shll8	r6		! 102 EX		! LMNx
	mov	r1,r7		!   5 MT (latency=0)

	or	r6,r3		!  82 EX		! LMNO
	bt/s	3b		! 109 BR

	 mov.l	r3,@-r0		!  30 LS
#endif
	! Finally, copy a byte at once, if necessary

	add	#4,r5		!  50 EX
	cmp/eq	r4,r0		!  54 MT

	add	#-6,r2		!  50 EX
	bt	9f		! 109 BR

8:	cmp/hi	r2,r0		!  57 MT
	mov.b	@(r0,r5),r1	!  20 LS (latency=2)

	bt/s	8b		! 109 BR

	 mov.b	r1,@-r0		!  29 LS

9:	rts
	 nop


	!
	!	GHIJ KLMN OPQR -->  .GHI JKLM NOPQ R...
	!

	! Size is 16 or greater, and may have trailing bytes

	.balign	32
.Lcase3:
	! Read a long word and write a long word at once
	! At the start of each iteration, r7 contains last long load
	add	#-3,r5		! 79 EX
	mov	r4,r2		!  5 MT (0 cycles latency)

	mov.l	@(r0,r5),r7	! 21 LS (2 cycles latency)
	add	#-4,r5		! 50 EX

	add	#7,r2		!  79 EX
	!
#ifdef CONFIG_CPU_LITTLE_ENDIAN
	! 6 cycles, 4 bytes per iteration
3:	mov.l	@(r0,r5),r1	!  21 LS (latency=2)	! NMLK
	mov	r7, r3		!   5 MT (latency=0)	! RQPO

	cmp/hi	r2,r0		!  57 MT
	shll8	r3		! 102 EX		! QPOx

	mov	r1,r6		!   5 MT (latency=0)
	shlr16	r6		! 107 EX

	shlr8	r6		! 106 EX		! xxxN
	mov	r1, r7		!   5 MT (latency=0)

	or	r6,r3		!  82 EX		! QPON
	bt/s	3b		! 109 BR

	 mov.l	r3,@-r0		!  30 LS
#else
3:	mov	r7,r3		! OPQR
	shlr8	r3		! xOPQ
	mov.l	@(r0,r5),r7	! KLMN
	mov	r7,r6
	shll16	r6
	shll8	r6		! Nxxx
	or	r6,r3		! NOP