[deliverable/linux.git] / arch / tile / lib / memcpy_64.c

/*
 * Copyright 2011 Tilera Corporation. All Rights Reserved.
 *
 *   This program is free software; you can redistribute it and/or
 *   modify it under the terms of the GNU General Public License
 *   as published by the Free Software Foundation, version 2.
 *
 *   This program 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, GOOD TITLE or
 *   NON INFRINGEMENT.  See the GNU General Public License for
 *   more details.
 */

#include <linux/types.h>
#include <linux/string.h>
#include <linux/module.h>
/* EXPORT_SYMBOL() is in arch/tile/lib/exports.c since this should be asm. */

/* Must be 8 bytes in size. */
#define op_t uint64_t

/* Threshold value for when to enter the unrolled loops. */
#define	OP_T_THRES	16

#if CHIP_L2_LINE_SIZE() != 64
#error "Assumes 64 byte line size"
#endif

/* How many cache lines ahead should we prefetch? */
#define PREFETCH_LINES_AHEAD 4

/*
 * Provide "base versions" of load and store for the normal code path.
 * The kernel provides other versions for userspace copies.
 */
#define ST(p, v) (*(p) = (v))
#define LD(p) (*(p))

#ifndef USERCOPY_FUNC
#define ST1 ST
#define ST2 ST
#define ST4 ST
#define ST8 ST
#define LD1 LD
#define LD2 LD
#define LD4 LD
#define LD8 LD
#define RETVAL dstv
void *memcpy(void *__restrict dstv, const void *__restrict srcv, size_t n)
#else
/*
 * Special kernel version will provide implementation of the LDn/STn
 * macros to return a count of uncopied bytes due to mm fault.
 */
#define RETVAL 0
int USERCOPY_FUNC(void *__restrict dstv, const void *__restrict srcv, size_t n)
#endif
{
	char *__restrict dst1 = (char *)dstv;
	const char *__restrict src1 = (const char *)srcv;
	const char *__restrict src1_end;
	const char *__restrict prefetch;
	op_t *__restrict dst8;    /* 8-byte pointer to destination memory. */
	op_t final; /* Final bytes to write to trailing word, if any */
	long i;

	if (n < 16) {
		for (; n; n--)
			ST1(dst1++, LD1(src1++));
		return RETVAL;
	}

	/*
	 * Locate the end of source memory we will copy.  Don't
	 * prefetch past this.
	 */
	src1_end = src1 + n - 1;

	/* Prefetch ahead a few cache lines, but not past the end. */
	prefetch = src1;
	for (i = 0; i < PREFETCH_LINES_AHEAD; i++) {
		__insn_prefetch(prefetch);
		prefetch += CHIP_L2_LINE_SIZE();
		prefetch = (prefetch < src1_end) ? prefetch : src1;
	}

	/* Copy bytes until dst is word-aligned. */
	for (; (uintptr_t)dst1 & (sizeof(op_t) - 1); n--)
		ST1(dst1++, LD1(src1++));

	/* 8-byte pointer to destination memory. */
	dst8 = (op_t *)dst1;

	if (__builtin_expect((uintptr_t)src1 & (sizeof(op_t) - 1), 0)) {
		/* Unaligned copy. */

		op_t  tmp0 = 0, tmp1 = 0, tmp2, tmp3;
		const op_t *src8 = (const op_t *) ((uintptr_t)src1 &
						   -sizeof(op_t));
		const void *srci = (void *)src1;
		int m;

		m = (CHIP_L2_LINE_SIZE() << 2) -
			(((uintptr_t)dst8) & ((CHIP_L2_LINE_SIZE() << 2) - 1));
		m = (n < m) ? n : m;
		m /= sizeof(op_t);

		/* Copy until 'dst' is cache-line-aligned. */
		n -= (sizeof(op_t) * m);

		switch (m % 4) {
		case 0:
			if (__builtin_expect(!m, 0))
				goto _M0;
			tmp1 = LD8(src8++);
			tmp2 = LD8(src8++);
			goto _8B3;
		case 2:
			m += 2;
			tmp3 = LD8(src8++);
			tmp0 = LD8(src8++);
			goto _8B1;
		case 3:
			m += 1;
			tmp2 = LD8(src8++);
			tmp3 = LD8(src8++);
			goto _8B2;
		case 1:
			m--;
			tmp0 = LD8(src8++);
			tmp1 = LD8(src8++);
			if (__builtin_expect(!m, 0))
				goto _8B0;
		}

		do {
			tmp2 = LD8(src8++);
			tmp0 =  __insn_dblalign(tmp0, tmp1, srci);
			ST8(dst8++, tmp0);
_8B3:
			tmp3 = LD8(src8++);
			tmp1 = __insn_dblalign(tmp1, tmp2, srci);
			ST8(dst8++, tmp1);
_8B2:
			tmp0 = LD8(src8++);
			tmp2 = __insn_dblalign(tmp2, tmp3, srci);
			ST8(dst8++, tmp2);
_8B1:
			tmp1 = LD8(src8++);
			tmp3 = __insn_dblalign(tmp3, tmp0, srci);
			ST8(dst8++, tmp3);
			m -= 4;
		} while (m);

_8B0:
		tmp0 = __insn_dblalign(tmp0, tmp1, srci);
		ST8(dst8++, tmp0);
		src8--;

_M0:
		if (__builtin_expect(n >= CHIP_L2_LINE_SIZE(), 0)) {
			op_t tmp4, tmp5, tmp6, tmp7, tmp8;

			prefetch = ((const char *)src8) +
				CHIP_L2_LINE_SIZE() * PREFETCH_LINES_AHEAD;

			for (tmp0 = LD8(src8++); n >= CHIP_L2_LINE_SIZE();
			     n -= CHIP_L2_LINE_SIZE()) {
				/* Prefetch and advance to next line to
				   prefetch, but don't go past the end.  */
				__insn_prefetch(prefetch);

				/* Make sure prefetch got scheduled
				   earlier.  */
				__asm__ ("" : : : "memory");

				prefetch += CHIP_L2_LINE_SIZE();
				prefetch = (prefetch < src1_end) ? prefetch :
					(const char *) src8;

				tmp1 = LD8(src8++);
				tmp2 = LD8(src8++);
				tmp3 = LD8(src8++);
				tmp4 = LD8(src8++);
				tmp5 = LD8(src8++);
				tmp6 = LD8(src8++);
				tmp7 = LD8(src8++);
				tmp8 = LD8(src8++);

				tmp0 = __insn_dblalign(tmp0, tmp1, srci);
				tmp1 = __insn_dblalign(tmp1, tmp2, srci);
				tmp2 = __insn_dblalign(tmp2, tmp3, srci);
				tmp3 = __insn_dblalign(tmp3, tmp4, srci);
				tmp4 = __insn_dblalign(tmp4, tmp5, srci);
				tmp5 = __insn_dblalign(tmp5, tmp6, srci);
				tmp6 = __insn_dblalign(tmp6, tmp7, srci);
				tmp7 = __insn_dblalign(tmp7, tmp8, srci);

				__insn_wh64(dst8);

				ST8(dst8++, tmp0);
				ST8(dst8++, tmp1);
				ST8(dst8++, tmp2);
				ST8(dst8++, tmp3);
				ST8(dst8++, tmp4);
				ST8(dst8++, tmp5);
				ST8(dst8++, tmp6);
				ST8(dst8++, tmp7);

				tmp0 = tmp8;
			}
			src8--;
		}

		/* Copy the rest 8-byte chunks. */
		if (n >= sizeof(op_t)) {
			tmp0 = LD8(src8++);
			for (; n >= sizeof(op_t); n -= sizeof(op_t)) {
				tmp1 = LD8(src8++);
				tmp0 = __insn_dblalign(tmp0, tmp1, srci);
				ST8(dst8++, tmp0);
				tmp0 = tmp1;
			}
			src8--;
		}

		if (n == 0)
			return RETVAL;

		tmp0 = LD8(src8++);
		tmp1 = ((const char *)src8 <= src1_end)
			? LD8((op_t *)src8) : 0;
		final = __insn_dblalign(tmp0, tmp1, srci);

	} else {
		/* Aligned copy. */

		const op_t *__restrict src8 = (const op_t *)src1;

		/* src8 and dst8 are both word-aligned. */
		if (n >= CHIP_L2_LINE_SIZE()) {
			/* Copy until 'dst' is cache-line-aligned. */
			for (; (uintptr_t)dst8 & (CHIP_L2_LINE_SIZE() - 1);
			     n -= sizeof(op_t))
				ST8(dst8++, LD8(src8++));

			for (; n >= CHIP_L2_LINE_SIZE(); ) {
				op_t tmp0, tmp1, tmp2, tmp3;
				op_t tmp4, tmp5, tmp6, tmp7;

				/*
				 * Prefetch and advance to next line
				 * to prefetch, but don't go past the
				 * end.
				 */
				__insn_prefetch(prefetch);

				/* Make sure prefetch got scheduled
				   earlier.  */
				__asm__ ("" : : : "memory");

				prefetch += CHIP_L2_LINE_SIZE();
				prefetch = (prefetch < src1_end) ? prefetch :
					(const char *)src8;

				/*
				 * Do all the loads before wh64.  This
				 * is necessary if [src8, src8+7] and
				 * [dst8, dst8+7] share the same cache
				 * line and dst8 <= src8, as can be
				 * the case when called from memmove,
				 * or with code tested on x86 whose
				 * memcpy always works with forward
				 * copies.
				 */
				tmp0 = LD8(src8++);
				tmp1 = LD8(src8++);
				tmp2 = LD8(src8++);
				tmp3 = LD8(src8++);
				tmp4 = LD8(src8++);
				tmp5 = LD8(src8++);
				tmp6 = LD8(src8++);
				tmp7 = LD8(src8++);

				/* wh64 and wait for tmp7 load completion. */
				__asm__ ("move %0, %0; wh64 %1\n"
					 : : "r"(tmp7), "r"(dst8));

				ST8(dst8++, tmp0);
				ST8(dst8++, tmp1);
				ST8(dst8++, tmp2);
				ST8(dst8++, tmp3);
				ST8(dst8++, tmp4);
				ST8(dst8++, tmp5);
				ST8(dst8++, tmp6);
				ST8(dst8++, tmp7);

				n -= CHIP_L2_LINE_SIZE();
			}
#if CHIP_L2_LINE_SIZE() != 64
# error "Fix code that assumes particular L2 cache line size."
#endif
		}

		for (; n >= sizeof(op_t); n -= sizeof(op_t))
			ST8(dst8++, LD8(src8++));

		if (__builtin_expect(n == 0, 1))
			return RETVAL;

		final = LD8(src8);
	}

	/* n != 0 if we get here.  Write out any trailing bytes. */
	dst1 = (char *)dst8;
#ifndef __BIG_ENDIAN__
	if (n & 4) {
		ST4((uint32_t *)dst1, final);
		dst1 += 4;
		final >>= 32;
		n &= 3;
	}
	if (n & 2) {
		ST2((uint16_t *)dst1, final);
		dst1 += 2;
		final >>= 16;
		n &= 1;
	}
	if (n)
		ST1((uint8_t *)dst1, final);
#else
	if (n & 4) {
		ST4((uint32_t *)dst1, final >> 32);
		dst1 += 4;
        }
        else
        {
		final >>= 32;
        }
	if (n & 2) {
		ST2((uint16_t *)dst1, final >> 16);
		dst1 += 2;
        }
        else
        {
		final >>= 16;
        }
	if (n & 1)
		ST1((uint8_t *)dst1, final >> 8);
#endif

	return RETVAL;
}

#ifdef USERCOPY_FUNC
#undef ST1
#undef ST2
#undef ST4
#undef ST8
#undef LD1
#undef LD2
#undef LD4
#undef LD8
#undef USERCOPY_FUNC
#endif
Commit	Line	Data
18aecc2b CM	1	/*
	2	* Copyright 2011 Tilera Corporation. All Rights Reserved.
	3	*
	4	* This program is free software; you can redistribute it and/or
	5	* modify it under the terms of the GNU General Public License
	6	* as published by the Free Software Foundation, version 2.
	7	*
	8	* This program is distributed in the hope that it will be useful, but
	9	* WITHOUT ANY WARRANTY; without even the implied warranty of
	10	* MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
	11	* NON INFRINGEMENT. See the GNU General Public License for
	12	* more details.
	13	*/
	14
	15	#include <linux/types.h>
	16	#include <linux/string.h>
	17	#include <linux/module.h>
18aecc2b CM	18	/* EXPORT_SYMBOL() is in arch/tile/lib/exports.c since this should be asm. */
	19
	20	/* Must be 8 bytes in size. */
c53c70a9	21	#define op_t uint64_t
18aecc2b	22
c53c70a9 CM	23	/* Threshold value for when to enter the unrolled loops. */
	24	#define OP_T_THRES 16
	25
	26	#if CHIP_L2_LINE_SIZE() != 64
	27	#error "Assumes 64 byte line size"
18aecc2b CM	28	#endif
	29
	30	/* How many cache lines ahead should we prefetch? */
c53c70a9	31	#define PREFETCH_LINES_AHEAD 4
18aecc2b CM	32
	33	/*
	34	* Provide "base versions" of load and store for the normal code path.
	35	* The kernel provides other versions for userspace copies.
	36	*/
	37	#define ST(p, v) (*(p) = (v))
	38	#define LD(p) (*(p))
	39
	40	#ifndef USERCOPY_FUNC
	41	#define ST1 ST
	42	#define ST2 ST
	43	#define ST4 ST
	44	#define ST8 ST
	45	#define LD1 LD
	46	#define LD2 LD
	47	#define LD4 LD
	48	#define LD8 LD
	49	#define RETVAL dstv
	50	void memcpy(void __restrict dstv, const void *__restrict srcv, size_t n)
	51	#else
	52	/*
	53	* Special kernel version will provide implementation of the LDn/STn
	54	* macros to return a count of uncopied bytes due to mm fault.
	55	*/
	56	#define RETVAL 0
	57	int USERCOPY_FUNC(void __restrict dstv, const void __restrict srcv, size_t n)
	58	#endif
	59	{
	60	char __restrict dst1 = (char )dstv;
	61	const char __restrict src1 = (const char )srcv;
	62	const char *__restrict src1_end;
	63	const char *__restrict prefetch;
c53c70a9 CM	64	op_t __restrict dst8; / 8-byte pointer to destination memory. */
c53c70a9 CM	65	op_t final; /* Final bytes to write to trailing word, if any */
18aecc2b CM	66	long i;
	67
	68	if (n < 16) {
	69	for (; n; n--)
	70	ST1(dst1++, LD1(src1++));
	71	return RETVAL;
	72	}
	73
	74	/*
	75	* Locate the end of source memory we will copy. Don't
	76	* prefetch past this.
	77	*/
	78	src1_end = src1 + n - 1;
	79
	80	/* Prefetch ahead a few cache lines, but not past the end. */
	81	prefetch = src1;
	82	for (i = 0; i < PREFETCH_LINES_AHEAD; i++) {
	83	__insn_prefetch(prefetch);
	84	prefetch += CHIP_L2_LINE_SIZE();
c53c70a9	85	prefetch = (prefetch < src1_end) ? prefetch : src1;
18aecc2b CM	86	}
	87
	88	/* Copy bytes until dst is word-aligned. */
c53c70a9	89	for (; (uintptr_t)dst1 & (sizeof(op_t) - 1); n--)
18aecc2b CM	90	ST1(dst1++, LD1(src1++));
	91
	92	/* 8-byte pointer to destination memory. */
c53c70a9 CM	93	dst8 = (op_t *)dst1;
	94
	95	if (__builtin_expect((uintptr_t)src1 & (sizeof(op_t) - 1), 0)) {
	96	/* Unaligned copy. */
	97
	98	op_t tmp0 = 0, tmp1 = 0, tmp2, tmp3;
	99	const op_t src8 = (const op_t ) ((uintptr_t)src1 &
	100	-sizeof(op_t));
	101	const void srci = (void )src1;
	102	int m;
	103
	104	m = (CHIP_L2_LINE_SIZE() << 2) -
	105	(((uintptr_t)dst8) & ((CHIP_L2_LINE_SIZE() << 2) - 1));
	106	m = (n < m) ? n : m;
	107	m /= sizeof(op_t);
	108
	109	/* Copy until 'dst' is cache-line-aligned. */
	110	n -= (sizeof(op_t) * m);
	111
	112	switch (m % 4) {
	113	case 0:
	114	if (__builtin_expect(!m, 0))
	115	goto _M0;
	116	tmp1 = LD8(src8++);
	117	tmp2 = LD8(src8++);
	118	goto _8B3;
	119	case 2:
	120	m += 2;
	121	tmp3 = LD8(src8++);
	122	tmp0 = LD8(src8++);
	123	goto _8B1;
	124	case 3:
	125	m += 1;
	126	tmp2 = LD8(src8++);
	127	tmp3 = LD8(src8++);
	128	goto _8B2;
	129	case 1:
	130	m--;
	131	tmp0 = LD8(src8++);
	132	tmp1 = LD8(src8++);
	133	if (__builtin_expect(!m, 0))
	134	goto _8B0;
	135	}
	136
	137	do {
	138	tmp2 = LD8(src8++);
	139	tmp0 = __insn_dblalign(tmp0, tmp1, srci);
	140	ST8(dst8++, tmp0);
	141	_8B3:
	142	tmp3 = LD8(src8++);
	143	tmp1 = __insn_dblalign(tmp1, tmp2, srci);
	144	ST8(dst8++, tmp1);
	145	_8B2:
	146	tmp0 = LD8(src8++);
	147	tmp2 = __insn_dblalign(tmp2, tmp3, srci);
	148	ST8(dst8++, tmp2);
	149	_8B1:
	150	tmp1 = LD8(src8++);
	151	tmp3 = __insn_dblalign(tmp3, tmp0, srci);
	152	ST8(dst8++, tmp3);
	153	m -= 4;
	154	} while (m);
	155
	156	_8B0:
157	tmp0 = __insn_dblalign(tmp0, tmp1, srci);
158	ST8(dst8++, tmp0);
159	src8--;
160
161	_M0:
162	if (__builtin_expect(n >= CHIP_L2_LINE_SIZE(), 0)) {
163	op_t tmp4, tmp5, tmp6, tmp7, tmp8;
164
165	prefetch = ((const char *)src8) +
166	CHIP_L2_LINE_SIZE() * PREFETCH_LINES_AHEAD;
167
168	for (tmp0 = LD8(src8++); n >= CHIP_L2_LINE_SIZE();
169	n -= CHIP_L2_LINE_SIZE()) {
170	/* Prefetch and advance to next line to
171	prefetch, but don't go past the end. */
172	__insn_prefetch(prefetch);
173
174	/* Make sure prefetch got scheduled
175	earlier. */
176	__asm__ ("" : : : "memory");
177
178	prefetch += CHIP_L2_LINE_SIZE();
179	prefetch = (prefetch < src1_end) ? prefetch :
180	(const char *) src8;
181
182	tmp1 = LD8(src8++);
183	tmp2 = LD8(src8++);
184	tmp3 = LD8(src8++);
185	tmp4 = LD8(src8++);
186	tmp5 = LD8(src8++);
187	tmp6 = LD8(src8++);
188	tmp7 = LD8(src8++);
189	tmp8 = LD8(src8++);
190
191	tmp0 = __insn_dblalign(tmp0, tmp1, srci);
192	tmp1 = __insn_dblalign(tmp1, tmp2, srci);
193	tmp2 = __insn_dblalign(tmp2, tmp3, srci);
194	tmp3 = __insn_dblalign(tmp3, tmp4, srci);
195	tmp4 = __insn_dblalign(tmp4, tmp5, srci);
196	tmp5 = __insn_dblalign(tmp5, tmp6, srci);
197	tmp6 = __insn_dblalign(tmp6, tmp7, srci);
198	tmp7 = __insn_dblalign(tmp7, tmp8, srci);
199
200	__insn_wh64(dst8);
201
202	ST8(dst8++, tmp0);
203	ST8(dst8++, tmp1);
204	ST8(dst8++, tmp2);
205	ST8(dst8++, tmp3);
206	ST8(dst8++, tmp4);
207	ST8(dst8++, tmp5);
208	ST8(dst8++, tmp6);
209	ST8(dst8++, tmp7);
210
211	tmp0 = tmp8;
212	}
213	src8--;
214	}
215
216	/* Copy the rest 8-byte chunks. */
217	if (n >= sizeof(op_t)) {
218	tmp0 = LD8(src8++);
219	for (; n >= sizeof(op_t); n -= sizeof(op_t)) {
220	tmp1 = LD8(src8++);
221	tmp0 = __insn_dblalign(tmp0, tmp1, srci);
222	ST8(dst8++, tmp0);
223	tmp0 = tmp1;
224	}
225	src8--;
18aecc2b CM	226	}
	227
	228	if (n == 0)
	229	return RETVAL;
	230
c53c70a9 CM	231	tmp0 = LD8(src8++);
	232	tmp1 = ((const char *)src8 <= src1_end)
	233	? LD8((op_t *)src8) : 0;
	234	final = __insn_dblalign(tmp0, tmp1, srci);
18aecc2b	235
18aecc2b CM	236	} else {
	237	/* Aligned copy. */
	238
c53c70a9	239	const op_t __restrict src8 = (const op_t )src1;
18aecc2b CM	240
	241	/* src8 and dst8 are both word-aligned. */
	242	if (n >= CHIP_L2_LINE_SIZE()) {
	243	/* Copy until 'dst' is cache-line-aligned. */
	244	for (; (uintptr_t)dst8 & (CHIP_L2_LINE_SIZE() - 1);
c53c70a9	245	n -= sizeof(op_t))
18aecc2b CM	246	ST8(dst8++, LD8(src8++));
	247
	248	for (; n >= CHIP_L2_LINE_SIZE(); ) {
c53c70a9 CM	249	op_t tmp0, tmp1, tmp2, tmp3;
c53c70a9 CM	250	op_t tmp4, tmp5, tmp6, tmp7;
18aecc2b CM	251
	252	/*
	253	* Prefetch and advance to next line
c53c70a9 CM	254	* to prefetch, but don't go past the
c53c70a9 CM	255	* end.
18aecc2b CM	256	*/
18aecc2b CM	257	__insn_prefetch(prefetch);
c53c70a9 CM	258
	259	/* Make sure prefetch got scheduled
	260	earlier. */
	261	__asm__ ("" : : : "memory");
	262
18aecc2b	263	prefetch += CHIP_L2_LINE_SIZE();
c53c70a9	264	prefetch = (prefetch < src1_end) ? prefetch :
18aecc2b CM	265	(const char *)src8;
	266
	267	/*
c53c70a9 CM	268	* Do all the loads before wh64. This
	269	* is necessary if [src8, src8+7] and
	270	* [dst8, dst8+7] share the same cache
	271	* line and dst8 <= src8, as can be
	272	* the case when called from memmove,
	273	* or with code tested on x86 whose
	274	* memcpy always works with forward
	275	* copies.
18aecc2b	276	*/
c53c70a9 CM	277	tmp0 = LD8(src8++);
	278	tmp1 = LD8(src8++);
	279	tmp2 = LD8(src8++);
	280	tmp3 = LD8(src8++);
	281	tmp4 = LD8(src8++);
	282	tmp5 = LD8(src8++);
	283	tmp6 = LD8(src8++);
	284	tmp7 = LD8(src8++);
	285
	286	/* wh64 and wait for tmp7 load completion. */
	287	__asm__ ("move %0, %0; wh64 %1\n"
	288	: : "r"(tmp7), "r"(dst8));
18aecc2b	289
c53c70a9 CM	290	ST8(dst8++, tmp0);
	291	ST8(dst8++, tmp1);
	292	ST8(dst8++, tmp2);
	293	ST8(dst8++, tmp3);
	294	ST8(dst8++, tmp4);
	295	ST8(dst8++, tmp5);
	296	ST8(dst8++, tmp6);
	297	ST8(dst8++, tmp7);
	298
	299	n -= CHIP_L2_LINE_SIZE();
18aecc2b	300	}
c53c70a9 CM	301	#if CHIP_L2_LINE_SIZE() != 64
	302	# error "Fix code that assumes particular L2 cache line size."
	303	#endif
18aecc2b CM	304	}
18aecc2b CM	305
c53c70a9	306	for (; n >= sizeof(op_t); n -= sizeof(op_t))
18aecc2b CM	307	ST8(dst8++, LD8(src8++));
	308
	309	if (__builtin_expect(n == 0, 1))
	310	return RETVAL;
	311
	312	final = LD8(src8);
	313	}
	314
	315	/* n != 0 if we get here. Write out any trailing bytes. */
	316	dst1 = (char *)dst8;
1efea40d	317	#ifndef __BIG_ENDIAN__
18aecc2b CM	318	if (n & 4) {
	319	ST4((uint32_t *)dst1, final);
	320	dst1 += 4;
	321	final >>= 32;
	322	n &= 3;
	323	}
	324	if (n & 2) {
	325	ST2((uint16_t *)dst1, final);
	326	dst1 += 2;
	327	final >>= 16;
	328	n &= 1;
	329	}
	330	if (n)
	331	ST1((uint8_t *)dst1, final);
1efea40d CM	332	#else
	333	if (n & 4) {
	334	ST4((uint32_t *)dst1, final >> 32);
	335	dst1 += 4;
	336	}
	337	else
	338	{
	339	final >>= 32;
	340	}
	341	if (n & 2) {
	342	ST2((uint16_t *)dst1, final >> 16);
	343	dst1 += 2;
	344	}
	345	else
	346	{
	347	final >>= 16;
	348	}
	349	if (n & 1)
	350	ST1((uint8_t *)dst1, final >> 8);
	351	#endif
18aecc2b CM	352
	353	return RETVAL;
	354	}
	355
18aecc2b CM	356	#ifdef USERCOPY_FUNC
	357	#undef ST1
	358	#undef ST2
	359	#undef ST4
	360	#undef ST8
	361	#undef LD1
	362	#undef LD2
	363	#undef LD4
	364	#undef LD8
	365	#undef USERCOPY_FUNC
	366	#endif