1*eea9507aSLey Foon Tan /* Extracted from GLIBC memcpy.c and memcopy.h, which is:
2*eea9507aSLey Foon Tan Copyright (C) 1991, 1992, 1993, 1997, 2004 Free Software Foundation, Inc.
3*eea9507aSLey Foon Tan This file is part of the GNU C Library.
4*eea9507aSLey Foon Tan Contributed by Torbjorn Granlund (tege@sics.se).
5*eea9507aSLey Foon Tan
6*eea9507aSLey Foon Tan The GNU C Library is free software; you can redistribute it and/or
7*eea9507aSLey Foon Tan modify it under the terms of the GNU Lesser General Public
8*eea9507aSLey Foon Tan License as published by the Free Software Foundation; either
9*eea9507aSLey Foon Tan version 2.1 of the License, or (at your option) any later version.
10*eea9507aSLey Foon Tan
11*eea9507aSLey Foon Tan The GNU C Library is distributed in the hope that it will be useful,
12*eea9507aSLey Foon Tan but WITHOUT ANY WARRANTY; without even the implied warranty of
13*eea9507aSLey Foon Tan MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14*eea9507aSLey Foon Tan Lesser General Public License for more details.
15*eea9507aSLey Foon Tan
16*eea9507aSLey Foon Tan You should have received a copy of the GNU Lesser General Public
17*eea9507aSLey Foon Tan License along with the GNU C Library; if not, see
18*eea9507aSLey Foon Tan <http://www.gnu.org/licenses/>. */
19*eea9507aSLey Foon Tan
20*eea9507aSLey Foon Tan #include <linux/types.h>
21*eea9507aSLey Foon Tan
22*eea9507aSLey Foon Tan /* Type to use for aligned memory operations.
23*eea9507aSLey Foon Tan This should normally be the biggest type supported by a single load
24*eea9507aSLey Foon Tan and store. */
25*eea9507aSLey Foon Tan #define op_t unsigned long int
26*eea9507aSLey Foon Tan #define OPSIZ (sizeof(op_t))
27*eea9507aSLey Foon Tan
28*eea9507aSLey Foon Tan /* Optimal type for storing bytes in registers. */
29*eea9507aSLey Foon Tan #define reg_char char
30*eea9507aSLey Foon Tan
31*eea9507aSLey Foon Tan #define MERGE(w0, sh_1, w1, sh_2) (((w0) >> (sh_1)) | ((w1) << (sh_2)))
32*eea9507aSLey Foon Tan
33*eea9507aSLey Foon Tan /* Copy exactly NBYTES bytes from SRC_BP to DST_BP,
34*eea9507aSLey Foon Tan without any assumptions about alignment of the pointers. */
35*eea9507aSLey Foon Tan #define BYTE_COPY_FWD(dst_bp, src_bp, nbytes) \
36*eea9507aSLey Foon Tan do { \
37*eea9507aSLey Foon Tan size_t __nbytes = (nbytes); \
38*eea9507aSLey Foon Tan while (__nbytes > 0) { \
39*eea9507aSLey Foon Tan unsigned char __x = ((unsigned char *) src_bp)[0]; \
40*eea9507aSLey Foon Tan src_bp += 1; \
41*eea9507aSLey Foon Tan __nbytes -= 1; \
42*eea9507aSLey Foon Tan ((unsigned char *) dst_bp)[0] = __x; \
43*eea9507aSLey Foon Tan dst_bp += 1; \
44*eea9507aSLey Foon Tan } \
45*eea9507aSLey Foon Tan } while (0)
46*eea9507aSLey Foon Tan
47*eea9507aSLey Foon Tan /* Copy *up to* NBYTES bytes from SRC_BP to DST_BP, with
48*eea9507aSLey Foon Tan the assumption that DST_BP is aligned on an OPSIZ multiple. If
49*eea9507aSLey Foon Tan not all bytes could be easily copied, store remaining number of bytes
50*eea9507aSLey Foon Tan in NBYTES_LEFT, otherwise store 0. */
51*eea9507aSLey Foon Tan /* extern void _wordcopy_fwd_aligned __P ((long int, long int, size_t)); */
52*eea9507aSLey Foon Tan /* extern void _wordcopy_fwd_dest_aligned __P ((long int, long int, size_t)); */
53*eea9507aSLey Foon Tan #define WORD_COPY_FWD(dst_bp, src_bp, nbytes_left, nbytes) \
54*eea9507aSLey Foon Tan do { \
55*eea9507aSLey Foon Tan if (src_bp % OPSIZ == 0) \
56*eea9507aSLey Foon Tan _wordcopy_fwd_aligned(dst_bp, src_bp, (nbytes) / OPSIZ);\
57*eea9507aSLey Foon Tan else \
58*eea9507aSLey Foon Tan _wordcopy_fwd_dest_aligned(dst_bp, src_bp, (nbytes) / OPSIZ);\
59*eea9507aSLey Foon Tan src_bp += (nbytes) & -OPSIZ; \
60*eea9507aSLey Foon Tan dst_bp += (nbytes) & -OPSIZ; \
61*eea9507aSLey Foon Tan (nbytes_left) = (nbytes) % OPSIZ; \
62*eea9507aSLey Foon Tan } while (0)
63*eea9507aSLey Foon Tan
64*eea9507aSLey Foon Tan
65*eea9507aSLey Foon Tan /* Threshold value for when to enter the unrolled loops. */
66*eea9507aSLey Foon Tan #define OP_T_THRES 16
67*eea9507aSLey Foon Tan
68*eea9507aSLey Foon Tan /* _wordcopy_fwd_aligned -- Copy block beginning at SRCP to
69*eea9507aSLey Foon Tan block beginning at DSTP with LEN `op_t' words (not LEN bytes!).
70*eea9507aSLey Foon Tan Both SRCP and DSTP should be aligned for memory operations on `op_t's. */
71*eea9507aSLey Foon Tan /* stream-lined (read x8 + write x8) */
_wordcopy_fwd_aligned(long int dstp,long int srcp,size_t len)72*eea9507aSLey Foon Tan static void _wordcopy_fwd_aligned(long int dstp, long int srcp, size_t len)
73*eea9507aSLey Foon Tan {
74*eea9507aSLey Foon Tan while (len > 7) {
75*eea9507aSLey Foon Tan register op_t a0, a1, a2, a3, a4, a5, a6, a7;
76*eea9507aSLey Foon Tan
77*eea9507aSLey Foon Tan a0 = ((op_t *) srcp)[0];
78*eea9507aSLey Foon Tan a1 = ((op_t *) srcp)[1];
79*eea9507aSLey Foon Tan a2 = ((op_t *) srcp)[2];
80*eea9507aSLey Foon Tan a3 = ((op_t *) srcp)[3];
81*eea9507aSLey Foon Tan a4 = ((op_t *) srcp)[4];
82*eea9507aSLey Foon Tan a5 = ((op_t *) srcp)[5];
83*eea9507aSLey Foon Tan a6 = ((op_t *) srcp)[6];
84*eea9507aSLey Foon Tan a7 = ((op_t *) srcp)[7];
85*eea9507aSLey Foon Tan ((op_t *) dstp)[0] = a0;
86*eea9507aSLey Foon Tan ((op_t *) dstp)[1] = a1;
87*eea9507aSLey Foon Tan ((op_t *) dstp)[2] = a2;
88*eea9507aSLey Foon Tan ((op_t *) dstp)[3] = a3;
89*eea9507aSLey Foon Tan ((op_t *) dstp)[4] = a4;
90*eea9507aSLey Foon Tan ((op_t *) dstp)[5] = a5;
91*eea9507aSLey Foon Tan ((op_t *) dstp)[6] = a6;
92*eea9507aSLey Foon Tan ((op_t *) dstp)[7] = a7;
93*eea9507aSLey Foon Tan
94*eea9507aSLey Foon Tan srcp += 8 * OPSIZ;
95*eea9507aSLey Foon Tan dstp += 8 * OPSIZ;
96*eea9507aSLey Foon Tan len -= 8;
97*eea9507aSLey Foon Tan }
98*eea9507aSLey Foon Tan while (len > 0) {
99*eea9507aSLey Foon Tan *(op_t *)dstp = *(op_t *)srcp;
100*eea9507aSLey Foon Tan
101*eea9507aSLey Foon Tan srcp += OPSIZ;
102*eea9507aSLey Foon Tan dstp += OPSIZ;
103*eea9507aSLey Foon Tan len -= 1;
104*eea9507aSLey Foon Tan }
105*eea9507aSLey Foon Tan }
106*eea9507aSLey Foon Tan
107*eea9507aSLey Foon Tan /* _wordcopy_fwd_dest_aligned -- Copy block beginning at SRCP to
108*eea9507aSLey Foon Tan block beginning at DSTP with LEN `op_t' words (not LEN bytes!).
109*eea9507aSLey Foon Tan DSTP should be aligned for memory operations on `op_t's, but SRCP must
110*eea9507aSLey Foon Tan *not* be aligned. */
111*eea9507aSLey Foon Tan /* stream-lined (read x4 + write x4) */
_wordcopy_fwd_dest_aligned(long int dstp,long int srcp,size_t len)112*eea9507aSLey Foon Tan static void _wordcopy_fwd_dest_aligned(long int dstp, long int srcp,
113*eea9507aSLey Foon Tan size_t len)
114*eea9507aSLey Foon Tan {
115*eea9507aSLey Foon Tan op_t ap;
116*eea9507aSLey Foon Tan int sh_1, sh_2;
117*eea9507aSLey Foon Tan
118*eea9507aSLey Foon Tan /* Calculate how to shift a word read at the memory operation
119*eea9507aSLey Foon Tan aligned srcp to make it aligned for copy. */
120*eea9507aSLey Foon Tan
121*eea9507aSLey Foon Tan sh_1 = 8 * (srcp % OPSIZ);
122*eea9507aSLey Foon Tan sh_2 = 8 * OPSIZ - sh_1;
123*eea9507aSLey Foon Tan
124*eea9507aSLey Foon Tan /* Make SRCP aligned by rounding it down to the beginning of the `op_t'
125*eea9507aSLey Foon Tan it points in the middle of. */
126*eea9507aSLey Foon Tan srcp &= -OPSIZ;
127*eea9507aSLey Foon Tan ap = ((op_t *) srcp)[0];
128*eea9507aSLey Foon Tan srcp += OPSIZ;
129*eea9507aSLey Foon Tan
130*eea9507aSLey Foon Tan while (len > 3) {
131*eea9507aSLey Foon Tan op_t a0, a1, a2, a3;
132*eea9507aSLey Foon Tan
133*eea9507aSLey Foon Tan a0 = ((op_t *) srcp)[0];
134*eea9507aSLey Foon Tan a1 = ((op_t *) srcp)[1];
135*eea9507aSLey Foon Tan a2 = ((op_t *) srcp)[2];
136*eea9507aSLey Foon Tan a3 = ((op_t *) srcp)[3];
137*eea9507aSLey Foon Tan ((op_t *) dstp)[0] = MERGE(ap, sh_1, a0, sh_2);
138*eea9507aSLey Foon Tan ((op_t *) dstp)[1] = MERGE(a0, sh_1, a1, sh_2);
139*eea9507aSLey Foon Tan ((op_t *) dstp)[2] = MERGE(a1, sh_1, a2, sh_2);
140*eea9507aSLey Foon Tan ((op_t *) dstp)[3] = MERGE(a2, sh_1, a3, sh_2);
141*eea9507aSLey Foon Tan
142*eea9507aSLey Foon Tan ap = a3;
143*eea9507aSLey Foon Tan srcp += 4 * OPSIZ;
144*eea9507aSLey Foon Tan dstp += 4 * OPSIZ;
145*eea9507aSLey Foon Tan len -= 4;
146*eea9507aSLey Foon Tan }
147*eea9507aSLey Foon Tan while (len > 0) {
148*eea9507aSLey Foon Tan register op_t a0;
149*eea9507aSLey Foon Tan
150*eea9507aSLey Foon Tan a0 = ((op_t *) srcp)[0];
151*eea9507aSLey Foon Tan ((op_t *) dstp)[0] = MERGE(ap, sh_1, a0, sh_2);
152*eea9507aSLey Foon Tan
153*eea9507aSLey Foon Tan ap = a0;
154*eea9507aSLey Foon Tan srcp += OPSIZ;
155*eea9507aSLey Foon Tan dstp += OPSIZ;
156*eea9507aSLey Foon Tan len -= 1;
157*eea9507aSLey Foon Tan }
158*eea9507aSLey Foon Tan }
159*eea9507aSLey Foon Tan
memcpy(void * dstpp,const void * srcpp,size_t len)160*eea9507aSLey Foon Tan void *memcpy(void *dstpp, const void *srcpp, size_t len)
161*eea9507aSLey Foon Tan {
162*eea9507aSLey Foon Tan unsigned long int dstp = (long int) dstpp;
163*eea9507aSLey Foon Tan unsigned long int srcp = (long int) srcpp;
164*eea9507aSLey Foon Tan
165*eea9507aSLey Foon Tan /* Copy from the beginning to the end. */
166*eea9507aSLey Foon Tan
167*eea9507aSLey Foon Tan /* If there not too few bytes to copy, use word copy. */
168*eea9507aSLey Foon Tan if (len >= OP_T_THRES) {
169*eea9507aSLey Foon Tan /* Copy just a few bytes to make DSTP aligned. */
170*eea9507aSLey Foon Tan len -= (-dstp) % OPSIZ;
171*eea9507aSLey Foon Tan BYTE_COPY_FWD(dstp, srcp, (-dstp) % OPSIZ);
172*eea9507aSLey Foon Tan
173*eea9507aSLey Foon Tan /* Copy whole pages from SRCP to DSTP by virtual address
174*eea9507aSLey Foon Tan manipulation, as much as possible. */
175*eea9507aSLey Foon Tan
176*eea9507aSLey Foon Tan /* PAGE_COPY_FWD_MAYBE (dstp, srcp, len, len); */
177*eea9507aSLey Foon Tan
178*eea9507aSLey Foon Tan /* Copy from SRCP to DSTP taking advantage of the known
179*eea9507aSLey Foon Tan alignment of DSTP. Number of bytes remaining is put in the
180*eea9507aSLey Foon Tan third argument, i.e. in LEN. This number may vary from
181*eea9507aSLey Foon Tan machine to machine. */
182*eea9507aSLey Foon Tan
183*eea9507aSLey Foon Tan WORD_COPY_FWD(dstp, srcp, len, len);
184*eea9507aSLey Foon Tan
185*eea9507aSLey Foon Tan /* Fall out and copy the tail. */
186*eea9507aSLey Foon Tan }
187*eea9507aSLey Foon Tan
188*eea9507aSLey Foon Tan /* There are just a few bytes to copy. Use byte memory operations. */
189*eea9507aSLey Foon Tan BYTE_COPY_FWD(dstp, srcp, len);
190*eea9507aSLey Foon Tan
191*eea9507aSLey Foon Tan return dstpp;
192*eea9507aSLey Foon Tan }
193*eea9507aSLey Foon Tan
memcpyb(void * dstpp,const void * srcpp,unsigned len)194*eea9507aSLey Foon Tan void *memcpyb(void *dstpp, const void *srcpp, unsigned len)
195*eea9507aSLey Foon Tan {
196*eea9507aSLey Foon Tan unsigned long int dstp = (long int) dstpp;
197*eea9507aSLey Foon Tan unsigned long int srcp = (long int) srcpp;
198*eea9507aSLey Foon Tan
199*eea9507aSLey Foon Tan BYTE_COPY_FWD(dstp, srcp, len);
200*eea9507aSLey Foon Tan
201*eea9507aSLey Foon Tan return dstpp;
202*eea9507aSLey Foon Tan }
203