memcpy.S (6974B)
1/* SPDX-License-Identifier: GPL-2.0 */ 2/* 3 * 4 * Optimized version of the standard memcpy() function 5 * 6 * Inputs: 7 * in0: destination address 8 * in1: source address 9 * in2: number of bytes to copy 10 * Output: 11 * no return value 12 * 13 * Copyright (C) 2000-2001 Hewlett-Packard Co 14 * Stephane Eranian <eranian@hpl.hp.com> 15 * David Mosberger-Tang <davidm@hpl.hp.com> 16 */ 17#include <asm/asmmacro.h> 18#include <asm/export.h> 19 20GLOBAL_ENTRY(memcpy) 21 22# define MEM_LAT 21 /* latency to memory */ 23 24# define dst r2 25# define src r3 26# define retval r8 27# define saved_pfs r9 28# define saved_lc r10 29# define saved_pr r11 30# define cnt r16 31# define src2 r17 32# define t0 r18 33# define t1 r19 34# define t2 r20 35# define t3 r21 36# define t4 r22 37# define src_end r23 38 39# define N (MEM_LAT + 4) 40# define Nrot ((N + 7) & ~7) 41 42 /* 43 * First, check if everything (src, dst, len) is a multiple of eight. If 44 * so, we handle everything with no taken branches (other than the loop 45 * itself) and a small icache footprint. Otherwise, we jump off to 46 * the more general copy routine handling arbitrary 47 * sizes/alignment etc. 48 */ 49 .prologue 50 .save ar.pfs, saved_pfs 51 alloc saved_pfs=ar.pfs,3,Nrot,0,Nrot 52 .save ar.lc, saved_lc 53 mov saved_lc=ar.lc 54 or t0=in0,in1 55 ;; 56 57 or t0=t0,in2 58 .save pr, saved_pr 59 mov saved_pr=pr 60 61 .body 62 63 cmp.eq p6,p0=in2,r0 // zero length? 64 mov retval=in0 // return dst 65(p6) br.ret.spnt.many rp // zero length, return immediately 66 ;; 67 68 mov dst=in0 // copy because of rotation 69 shr.u cnt=in2,3 // number of 8-byte words to copy 70 mov pr.rot=1<<16 71 ;; 72 73 adds cnt=-1,cnt // br.ctop is repeat/until 74 cmp.gtu p7,p0=16,in2 // copying less than 16 bytes? 75 mov ar.ec=N 76 ;; 77 78 and t0=0x7,t0 79 mov ar.lc=cnt 80 ;; 81 cmp.ne p6,p0=t0,r0 82 83 mov src=in1 // copy because of rotation 84(p7) br.cond.spnt.few .memcpy_short 85(p6) br.cond.spnt.few .memcpy_long 86 ;; 87 nop.m 0 88 ;; 89 nop.m 0 90 nop.i 0 91 ;; 92 nop.m 0 93 ;; 94 .rotr val[N] 95 .rotp p[N] 96 .align 32 971: { .mib 98(p[0]) ld8 val[0]=[src],8 99 nop.i 0 100 brp.loop.imp 1b, 2f 101} 1022: { .mfb 103(p[N-1])st8 [dst]=val[N-1],8 104 nop.f 0 105 br.ctop.dptk.few 1b 106} 107 ;; 108 mov ar.lc=saved_lc 109 mov pr=saved_pr,-1 110 mov ar.pfs=saved_pfs 111 br.ret.sptk.many rp 112 113 /* 114 * Small (<16 bytes) unaligned copying is done via a simple byte-at-the-time 115 * copy loop. This performs relatively poorly on Itanium, but it doesn't 116 * get used very often (gcc inlines small copies) and due to atomicity 117 * issues, we want to avoid read-modify-write of entire words. 118 */ 119 .align 32 120.memcpy_short: 121 adds cnt=-1,in2 // br.ctop is repeat/until 122 mov ar.ec=MEM_LAT 123 brp.loop.imp 1f, 2f 124 ;; 125 mov ar.lc=cnt 126 ;; 127 nop.m 0 128 ;; 129 nop.m 0 130 nop.i 0 131 ;; 132 nop.m 0 133 ;; 134 nop.m 0 135 ;; 136 /* 137 * It is faster to put a stop bit in the loop here because it makes 138 * the pipeline shorter (and latency is what matters on short copies). 139 */ 140 .align 32 1411: { .mib 142(p[0]) ld1 val[0]=[src],1 143 nop.i 0 144 brp.loop.imp 1b, 2f 145} ;; 1462: { .mfb 147(p[MEM_LAT-1])st1 [dst]=val[MEM_LAT-1],1 148 nop.f 0 149 br.ctop.dptk.few 1b 150} ;; 151 mov ar.lc=saved_lc 152 mov pr=saved_pr,-1 153 mov ar.pfs=saved_pfs 154 br.ret.sptk.many rp 155 156 /* 157 * Large (>= 16 bytes) copying is done in a fancy way. Latency isn't 158 * an overriding concern here, but throughput is. We first do 159 * sub-word copying until the destination is aligned, then we check 160 * if the source is also aligned. If so, we do a simple load/store-loop 161 * until there are less than 8 bytes left over and then we do the tail, 162 * by storing the last few bytes using sub-word copying. If the source 163 * is not aligned, we branch off to the non-congruent loop. 164 * 165 * stage: op: 166 * 0 ld 167 * : 168 * MEM_LAT+3 shrp 169 * MEM_LAT+4 st 170 * 171 * On Itanium, the pipeline itself runs without stalls. However, br.ctop 172 * seems to introduce an unavoidable bubble in the pipeline so the overall 173 * latency is 2 cycles/iteration. This gives us a _copy_ throughput 174 * of 4 byte/cycle. Still not bad. 175 */ 176# undef N 177# undef Nrot 178# define N (MEM_LAT + 5) /* number of stages */ 179# define Nrot ((N+1 + 2 + 7) & ~7) /* number of rotating regs */ 180 181#define LOG_LOOP_SIZE 6 182 183.memcpy_long: 184 alloc t3=ar.pfs,3,Nrot,0,Nrot // resize register frame 185 and t0=-8,src // t0 = src & ~7 186 and t2=7,src // t2 = src & 7 187 ;; 188 ld8 t0=[t0] // t0 = 1st source word 189 adds src2=7,src // src2 = (src + 7) 190 sub t4=r0,dst // t4 = -dst 191 ;; 192 and src2=-8,src2 // src2 = (src + 7) & ~7 193 shl t2=t2,3 // t2 = 8*(src & 7) 194 shl t4=t4,3 // t4 = 8*(dst & 7) 195 ;; 196 ld8 t1=[src2] // t1 = 1st source word if src is 8-byte aligned, 2nd otherwise 197 sub t3=64,t2 // t3 = 64-8*(src & 7) 198 shr.u t0=t0,t2 199 ;; 200 add src_end=src,in2 201 shl t1=t1,t3 202 mov pr=t4,0x38 // (p5,p4,p3)=(dst & 7) 203 ;; 204 or t0=t0,t1 205 mov cnt=r0 206 adds src_end=-1,src_end 207 ;; 208(p3) st1 [dst]=t0,1 209(p3) shr.u t0=t0,8 210(p3) adds cnt=1,cnt 211 ;; 212(p4) st2 [dst]=t0,2 213(p4) shr.u t0=t0,16 214(p4) adds cnt=2,cnt 215 ;; 216(p5) st4 [dst]=t0,4 217(p5) adds cnt=4,cnt 218 and src_end=-8,src_end // src_end = last word of source buffer 219 ;; 220 221 // At this point, dst is aligned to 8 bytes and there at least 16-7=9 bytes left to copy: 222 2231:{ add src=cnt,src // make src point to remainder of source buffer 224 sub cnt=in2,cnt // cnt = number of bytes left to copy 225 mov t4=ip 226 } ;; 227 and src2=-8,src // align source pointer 228 adds t4=.memcpy_loops-1b,t4 229 mov ar.ec=N 230 231 and t0=7,src // t0 = src & 7 232 shr.u t2=cnt,3 // t2 = number of 8-byte words left to copy 233 shl cnt=cnt,3 // move bits 0-2 to 3-5 234 ;; 235 236 .rotr val[N+1], w[2] 237 .rotp p[N] 238 239 cmp.ne p6,p0=t0,r0 // is src aligned, too? 240 shl t0=t0,LOG_LOOP_SIZE // t0 = 8*(src & 7) 241 adds t2=-1,t2 // br.ctop is repeat/until 242 ;; 243 add t4=t0,t4 244 mov pr=cnt,0x38 // set (p5,p4,p3) to # of bytes last-word bytes to copy 245 mov ar.lc=t2 246 ;; 247 nop.m 0 248 ;; 249 nop.m 0 250 nop.i 0 251 ;; 252 nop.m 0 253 ;; 254(p6) ld8 val[1]=[src2],8 // prime the pump... 255 mov b6=t4 256 br.sptk.few b6 257 ;; 258 259.memcpy_tail: 260 // At this point, (p5,p4,p3) are set to the number of bytes left to copy (which is 261 // less than 8) and t0 contains the last few bytes of the src buffer: 262(p5) st4 [dst]=t0,4 263(p5) shr.u t0=t0,32 264 mov ar.lc=saved_lc 265 ;; 266(p4) st2 [dst]=t0,2 267(p4) shr.u t0=t0,16 268 mov ar.pfs=saved_pfs 269 ;; 270(p3) st1 [dst]=t0 271 mov pr=saved_pr,-1 272 br.ret.sptk.many rp 273 274/////////////////////////////////////////////////////// 275 .align 64 276 277#define COPY(shift,index) \ 278 1: { .mib \ 279 (p[0]) ld8 val[0]=[src2],8; \ 280 (p[MEM_LAT+3]) shrp w[0]=val[MEM_LAT+3],val[MEM_LAT+4-index],shift; \ 281 brp.loop.imp 1b, 2f \ 282 }; \ 283 2: { .mfb \ 284 (p[MEM_LAT+4]) st8 [dst]=w[1],8; \ 285 nop.f 0; \ 286 br.ctop.dptk.few 1b; \ 287 }; \ 288 ;; \ 289 ld8 val[N-1]=[src_end]; /* load last word (may be same as val[N]) */ \ 290 ;; \ 291 shrp t0=val[N-1],val[N-index],shift; \ 292 br .memcpy_tail 293.memcpy_loops: 294 COPY(0, 1) /* no point special casing this---it doesn't go any faster without shrp */ 295 COPY(8, 0) 296 COPY(16, 0) 297 COPY(24, 0) 298 COPY(32, 0) 299 COPY(40, 0) 300 COPY(48, 0) 301 COPY(56, 0) 302 303END(memcpy) 304EXPORT_SYMBOL(memcpy)