1 /* Optimized version of the standard memcpy() function.
2 This file is part of the GNU C Library.
3 Copyright (C) 2000, 2001 Free Software Foundation, Inc.
4 Contributed by Dan Pop <Dan.Pop@cern.ch>.
6 The GNU C Library is free software; you can redistribute it and/or
7 modify it under the terms of the GNU Lesser General Public
8 License as published by the Free Software Foundation; either
9 version 2.1 of the License, or (at your option) any later version.
11 The GNU C Library is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 Lesser General Public License for more details.
16 You should have received a copy of the GNU Lesser General Public
17 License along with the GNU C Library; if not, write to the Free
18 Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
28 An assembly implementation of the algorithm used by the generic C
29 version from glibc. The case when all three arguments are multiples
30 of 8 is treated separatedly, for extra performance.
32 In this form, it assumes little endian mode. For big endian mode,
33 sh1 must be computed using an extra instruction: sub sh1 = 64, sh1
34 and the order of r[MEMLAT] and r[MEMLAT+1] must be reverted in the
63 (p[0]) ld8 r[0] = [asrc], 8 ; /* w1 */ \
64 (p[MEMLAT+1]) st8 [dest] = value, 8 ; \
65 (p[MEMLAT]) shrp value = r[MEMLAT], r[MEMLAT+1], shift ; \
68 br.ctop.sptk .loop##shift ; \
69 br.cond.sptk .cpyfew ; /* deal with the remaining bytes */
72 #define Nrot (((2*MEMLAT+3) + 7) & ~7)
76 alloc r2 = ar.pfs, 3, Nrot - 3, 0, Nrot
77 .rotr r[MEMLAT + 2], q[MEMLAT + 1]
79 mov ret0 = in0 // return value = dest
81 mov saved_pr = pr // save the predicate registers
83 mov saved_lc = ar.lc // save the loop counter
85 or tmp3 = in0, in1 ;; // tmp3 = dest | src
86 or tmp3 = tmp3, in2 // tmp3 = dest | src | len
87 mov dest = in0 // dest
90 sub tmp2 = r0, in0 // tmp2 = -dest
91 cmp.eq p6, p0 = in2, r0 // if (len == 0)
92 (p6) br.cond.spnt .restore_and_exit;;// return dest;
93 and tmp4 = 7, tmp3 // tmp4 = (dest | src | len) & 7
94 shr.u loopcnt = len, 4 ;; // loopcnt = len / 16
95 cmp.ne p6, p0 = tmp4, r0 // if ((dest | src | len) & 7 != 0)
96 (p6) br.cond.sptk .next // goto next;
98 // The optimal case, when dest, src and len are all multiples of 8
100 and tmp3 = 0xf, len // tmp3 = len % 16
101 mov pr.rot = 1 << 16 // set rotating predicates
102 mov ar.ec = MEMLAT + 1 ;; // set the epilog counter
103 cmp.ne p6, p0 = tmp3, r0 // do we have to copy an extra word?
104 adds loopcnt = -1, loopcnt;; // --loopcnt
105 (p6) ld8 value = [src], 8;;
106 (p6) st8 [dest] = value, 8 // copy the "extra" word
107 mov ar.lc = loopcnt // set the loop counter
108 cmp.eq p6, p0 = 8, len
109 (p6) br.cond.spnt .restore_and_exit;;// there was only one word to copy
111 adds asrc = 8, src ;;
114 (p[0]) ld8 r[0] = [src], 16
115 (p[0]) ld8 q[0] = [asrc], 16
116 (p[MEMLAT]) st8 [dest] = r[MEMLAT], 16
117 (p[MEMLAT]) st8 [adest] = q[MEMLAT], 16
120 mov pr = saved_pr, -1 // restore the predicate registers
121 mov ar.lc = saved_lc // restore the loop counter
124 cmp.ge p6, p0 = OP_T_THRES, len // is len <= OP_T_THRES
125 and loopcnt = 7, tmp2 // loopcnt = -dest % 8
126 (p6) br.cond.spnt .cpyfew // copy byte by byte
128 cmp.eq p6, p0 = loopcnt, r0
129 (p6) br.cond.sptk .dest_aligned
130 sub len = len, loopcnt // len -= -dest % 8
131 adds loopcnt = -1, loopcnt // --loopcnt
134 .l1: // copy -dest % 8 bytes
135 ld1 value = [src], 1 // value = *src++
137 st1 [dest] = value, 1 // *dest++ = value
140 and sh1 = 7, src // sh1 = src % 8
141 and tmp2 = -8, len // tmp2 = len & -OPSIZ
142 and asrc = -8, src // asrc = src & -OPSIZ -- align src
143 shr.u loopcnt = len, 3 // loopcnt = len / 8
144 and len = 7, len;; // len = len % 8
145 adds loopcnt = -1, loopcnt // --loopcnt
146 addl tmp4 = @ltoff(.table), gp
147 addl tmp3 = @ltoff(.loop56), gp
148 mov ar.ec = MEMLAT + 1 // set EC
149 mov pr.rot = 1 << 16;; // set rotating predicates
150 mov ar.lc = loopcnt // set LC
151 cmp.eq p6, p0 = sh1, r0 // is the src aligned?
152 (p6) br.cond.sptk .src_aligned
153 add src = src, tmp2 // src += len & -OPSIZ
154 shl sh1 = sh1, 3 // sh1 = 8 * (src % 8)
155 ld8 ploop56 = [tmp3] // ploop56 = &loop56
156 ld8 ptable = [tmp4];; // ptable = &table
157 add tmp3 = ptable, sh1;; // tmp3 = &table + sh1
158 mov ar.ec = MEMLAT + 1 + 1 // one more pass needed
159 ld8 tmp4 = [tmp3];; // tmp4 = loop offset
160 sub loopaddr = ploop56,tmp4 // loopadd = &loop56 - loop offset
161 ld8 r[1] = [asrc], 8;; // w0
163 br b6 // jump to the appropriate loop
175 (p[0]) ld8 r[0] = [src], 8
176 (p[MEMLAT]) st8 [dest] = r[MEMLAT], 8
179 cmp.eq p6, p0 = len, r0 // is len == 0 ?
180 adds len = -1, len // --len;
181 (p6) br.cond.spnt .restore_and_exit ;;
186 st1 [dest] = value, 1
189 mov pr = saved_pr, -1 // restore the predicate registers
190 mov ar.lc = saved_lc // restore the loop counter
194 data8 0 // dummy entry
195 data8 .loop56 - .loop8
196 data8 .loop56 - .loop16
197 data8 .loop56 - .loop24
198 data8 .loop56 - .loop32
199 data8 .loop56 - .loop40
200 data8 .loop56 - .loop48
201 data8 .loop56 - .loop56