/* Optimized memcpy implementation for CELL BE PowerPC.
   Copyright (C) 2010-2014 Free Software Foundation, Inc.
   This file is part of the GNU C Library.

   The GNU C Library is free software; you can redistribute it and/or
   modify it under the terms of the GNU Lesser General Public
   License as published by the Free Software Foundation; either
   version 2.1 of the License, or (at your option) any later version.

   The GNU C Library is distributed in the hope that it will be useful,
   but WITHOUT ANY WARRANTY; without even the implied warranty of
   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
   Lesser General Public License for more details.

   You should have received a copy of the GNU Lesser General Public
   License along with the GNU C Library; if not, see
   <http://www.gnu.org/licenses/>.  */

#include <sysdep.h>

#define PREFETCH_AHEAD 6	/* no cache lines SRC prefetching ahead  */
#define ZERO_AHEAD 4		/* no cache lines DST zeroing ahead  */

/* memcpy routine optimized for CELL-BE-PPC	v2.0
 *
 * The CELL PPC core has 1 integer unit and 1 load/store unit
 * CELL:
 * 1st level data cache = 32K
 * 2nd level data cache = 512K
 * 3rd level data cache = 0K
 * With 3.2 GHz clockrate the latency to 2nd level cache is >36 clocks,
 * latency to memory is >400 clocks
 * To improve copy performance we need to prefetch source data
 * far ahead to hide this latency
 * For best performance instruction forms ending in "." like "andi."
 * should be avoided as the are implemented in microcode on CELL.
 * The below code is loop unrolled for the CELL cache line of 128 bytes
 */

.align  7

EALIGN (memcpy, 5, 0)
	CALL_MCOUNT

	dcbt	0,r4		/* Prefetch ONE SRC cacheline  */
	cmplwi	cr1,r5,16	/* is size < 16 ?  */
	mr	r6,r3
	blt+	cr1,.Lshortcopy

.Lbigcopy:
	neg	r8,r3		/* LS 3 bits = # bytes to 8-byte dest bdry  */
	clrlwi  r8,r8,32-4	/* align to 16byte boundary  */
	sub     r7,r4,r3
	cmplwi	cr0,r8,0
	beq+	.Ldst_aligned

.Ldst_unaligned:
	mtcrf	0x01,r8		/* put #bytes to boundary into cr7  */
	subf	r5,r8,r5

	bf	cr7*4+3,1f
	lbzx	r0,r7,r6	/* copy 1 byte  */
	stb	r0,0(r6)
	addi	r6,r6,1
1:	bf	cr7*4+2,2f
	lhzx	r0,r7,r6	/* copy 2 byte  */
	sth	r0,0(r6)
	addi	r6,r6,2
2:	bf	cr7*4+1,4f
	lwzx	r0,r7,r6	/* copy 4 byte  */
	stw	r0,0(r6)
	addi	r6,r6,4
4:	bf	cr7*4+0,8f
	lfdx	fp9,r7,r6	/* copy 8 byte  */
	stfd	fp9,0(r6)
	addi	r6,r6,8
8:
	add	r4,r7,r6

.Ldst_aligned:

	cmpwi	cr5,r5,128-1

	neg	r7,r6
	addi	r6,r6,-8	/* prepare for stfdu  */
	addi	r4,r4,-8	/* prepare for lfdu  */

	clrlwi  r7,r7,32-7	/* align to cacheline boundary  */
	ble+	cr5,.Llessthancacheline

	cmplwi	cr6,r7,0
	subf	r5,r7,r5
	srwi	r7,r7,4		/* divide size by 16  */
	srwi	r10,r5,7	/* number of cache lines to copy  */

	cmplwi	r10,0
	li	r11,0		/* number cachelines to copy with prefetch  */
	beq	.Lnocacheprefetch

	cmplwi	r10,PREFETCH_AHEAD
	li	r12,128+8	/* prefetch distance  */
	ble	.Llessthanmaxprefetch

	subi	r11,r10,PREFETCH_AHEAD
	li	r10,PREFETCH_AHEAD

.Llessthanmaxprefetch:
	mtctr	r10

.LprefetchSRC:
	dcbt    r12,r4
	addi    r12,r12,128
	bdnz    .LprefetchSRC

.Lnocacheprefetch:
	mtctr	r7
	cmplwi	cr1,r5,128
	clrlwi  r5,r5,32-7
	beq	cr6,.Lcachelinealigned

.Laligntocacheline:
	lfd	fp9,0x08(r4)
	lfdu	fp10,0x10(r4)
	stfd	fp9,0x08(r6)
	stfdu	fp10,0x10(r6)
	bdnz	.Laligntocacheline


.Lcachelinealigned:		/* copy while cache lines  */

	blt-	cr1,.Llessthancacheline	/* size <128  */

.Louterloop:
	cmpwi   r11,0
	mtctr	r11
	beq-	.Lendloop

	li	r11,128*ZERO_AHEAD +8	/* DCBZ dist  */

.align	4
	/* Copy whole cachelines, optimized by prefetching SRC cacheline  */
.Lloop:				/* Copy aligned body  */
	dcbt	r12,r4		/* PREFETCH SOURCE some cache lines ahead  */
	lfd	fp9, 0x08(r4)
	dcbz	r11,r6
	lfd	fp10, 0x10(r4)	/* 4 register stride copy is optimal  */
	lfd	fp11, 0x18(r4)	/* to hide 1st level cache latency.  */
	lfd	fp12, 0x20(r4)
	stfd	fp9, 0x08(r6)
	stfd	fp10, 0x10(r6)
	stfd	fp11, 0x18(r6)
	stfd	fp12, 0x20(r6)
	lfd	fp9, 0x28(r4)
	lfd	fp10, 0x30(r4)
	lfd	fp11, 0x38(r4)
	lfd	fp12, 0x40(r4)
	stfd	fp9, 0x28(r6)
	stfd	fp10, 0x30(r6)
	stfd	fp11, 0x38(r6)
	stfd	fp12, 0x40(r6)
	lfd	fp9, 0x48(r4)
	lfd	fp10, 0x50(r4)
	lfd	fp11, 0x58(r4)
	lfd	fp12, 0x60(r4)
	stfd	fp9, 0x48(r6)
	stfd	fp10, 0x50(r6)
	stfd	fp11, 0x58(r6)
	stfd	fp12, 0x60(r6)
	lfd	fp9, 0x68(r4)
	lfd	fp10, 0x70(r4)
	lfd	fp11, 0x78(r4)
	lfdu	fp12, 0x80(r4)
	stfd	fp9, 0x68(r6)
	stfd	fp10, 0x70(r6)
	stfd	fp11, 0x78(r6)
	stfdu	fp12, 0x80(r6)

	bdnz	.Lloop

.Lendloop:
	cmpwi	r10,0
	slwi	r10,r10,2	/* adjust from 128 to 32 byte stride  */
	beq-	.Lendloop2
	mtctr	r10

.Lloop2:			/* Copy aligned body  */
	lfd	fp9, 0x08(r4)
	lfd	fp10, 0x10(r4)
	lfd	fp11, 0x18(r4)
	lfdu	fp12, 0x20(r4)
	stfd	fp9, 0x08(r6)
	stfd	fp10, 0x10(r6)
	stfd	fp11, 0x18(r6)
	stfdu	fp12, 0x20(r6)

	bdnz	.Lloop2
.Lendloop2:

.Llessthancacheline:		/* less than cache to do ?  */
	cmplwi	cr0,r5,16
	srwi	r7,r5,4		/* divide size by 16  */
	blt-	.Ldo_lt16
	mtctr	r7

.Lcopy_remaining:
	lfd	fp9,0x08(r4)
	lfdu	fp10,0x10(r4)
	stfd	fp9,0x08(r6)
	stfdu	fp10,0x10(r6)
	bdnz	.Lcopy_remaining

.Ldo_lt16:			/* less than 16 ?  */
	cmplwi	cr0,r5,0	/* copy remaining bytes (0-15)  */
	beqlr+			/* no rest to copy  */
	addi	r4,r4,8
	addi	r6,r6,8

.Lshortcopy:			/* SIMPLE COPY to handle size =< 15 bytes  */
	mtcrf	0x01,r5
	sub	r7,r4,r6
	bf-	cr7*4+0,8f
	lfdx	fp9,r7,r6	/* copy 8 byte  */
	stfd	fp9,0(r6)
	addi	r6,r6,8
8:
	bf	cr7*4+1,4f
	lwzx	r0,r7,r6	/* copy 4 byte  */
	stw	r0,0(r6)
	addi	r6,r6,4
4:
	bf	cr7*4+2,2f
	lhzx	r0,r7,r6	/* copy 2 byte  */
	sth	r0,0(r6)
	addi	r6,r6,2
2:
	bf	cr7*4+3,1f
	lbzx	r0,r7,r6	/* copy 1 byte  */
	stb	r0,0(r6)
1:	blr

END (memcpy)
libc_hidden_builtin_def (memcpy)