1/* 2 * 3 * Optmized version of the standard do_csum() function 4 * 5 * Return: a 64bit quantity containing the 16bit Internet checksum 6 * 7 * Inputs: 8 * in0: address of buffer to checksum (char *) 9 * in1: length of the buffer (int) 10 * 11 * Copyright (C) 1999, 2001-2002 Hewlett-Packard Co 12 * Stephane Eranian <eranian@hpl.hp.com> 13 * 14 * 02/04/22 Ken Chen <kenneth.w.chen@intel.com> 15 * Data locality study on the checksum buffer. 16 * More optimization cleanup - remove excessive stop bits. 17 * 02/04/08 David Mosberger <davidm@hpl.hp.com> 18 * More cleanup and tuning. 19 * 01/04/18 Jun Nakajima <jun.nakajima@intel.com> 20 * Clean up and optimize and the software pipeline, loading two 21 * back-to-back 8-byte words per loop. Clean up the initialization 22 * for the loop. Support the cases where load latency = 1 or 2. 23 * Set CONFIG_IA64_LOAD_LATENCY to 1 or 2 (default). 24 */ 25 26#include <asm/asmmacro.h> 27 28// 29// Theory of operations: 30// The goal is to go as quickly as possible to the point where 31// we can checksum 16 bytes/loop. Before reaching that point we must 32// take care of incorrect alignment of first byte. 33// 34// The code hereafter also takes care of the "tail" part of the buffer 35// before entering the core loop, if any. The checksum is a sum so it 36// allows us to commute operations. So we do the "head" and "tail" 37// first to finish at full speed in the body. Once we get the head and 38// tail values, we feed them into the pipeline, very handy initialization. 39// 40// Of course we deal with the special case where the whole buffer fits 41// into one 8 byte word. In this case we have only one entry in the pipeline. 42// 43// We use a (LOAD_LATENCY+2)-stage pipeline in the loop to account for 44// possible load latency and also to accommodate for head and tail. 45// 46// The end of the function deals with folding the checksum from 64bits 47// down to 16bits taking care of the carry. 48// 49// This version avoids synchronization in the core loop by also using a 50// pipeline for the accumulation of the checksum in resultx[] (x=1,2). 51// 52// wordx[] (x=1,2) 53// |---| 54// | | 0 : new value loaded in pipeline 55// |---| 56// | | - : in transit data 57// |---| 58// | | LOAD_LATENCY : current value to add to checksum 59// |---| 60// | | LOAD_LATENCY+1 : previous value added to checksum 61// |---| (previous iteration) 62// 63// resultx[] (x=1,2) 64// |---| 65// | | 0 : initial value 66// |---| 67// | | LOAD_LATENCY-1 : new checksum 68// |---| 69// | | LOAD_LATENCY : previous value of checksum 70// |---| 71// | | LOAD_LATENCY+1 : final checksum when out of the loop 72// |---| 73// 74// 75// See RFC1071 "Computing the Internet Checksum" for various techniques for 76// calculating the Internet checksum. 77// 78// NOT YET DONE: 79// - Maybe another algorithm which would take care of the folding at the 80// end in a different manner 81// - Work with people more knowledgeable than me on the network stack 82// to figure out if we could not split the function depending on the 83// type of packet or alignment we get. Like the ip_fast_csum() routine 84// where we know we have at least 20bytes worth of data to checksum. 85// - Do a better job of handling small packets. 86// - Note on prefetching: it was found that under various load, i.e. ftp read/write, 87// nfs read/write, the L1 cache hit rate is at 60% and L2 cache hit rate is at 99.8% 88// on the data that buffer points to (partly because the checksum is often preceded by 89// a copy_from_user()). This finding indiate that lfetch will not be beneficial since 90// the data is already in the cache. 91// 92 93#define saved_pfs r11 94#define hmask r16 95#define tmask r17 96#define first1 r18 97#define firstval r19 98#define firstoff r20 99#define last r21 100#define lastval r22 101#define lastoff r23 102#define saved_lc r24 103#define saved_pr r25 104#define tmp1 r26 105#define tmp2 r27 106#define tmp3 r28 107#define carry1 r29 108#define carry2 r30 109#define first2 r31 110 111#define buf in0 112#define len in1 113 114#define LOAD_LATENCY 2 // XXX fix me 115 116#if (LOAD_LATENCY != 1) && (LOAD_LATENCY != 2) 117# error "Only 1 or 2 is supported/tested for LOAD_LATENCY." 118#endif 119 120#define PIPE_DEPTH (LOAD_LATENCY+2) 121#define ELD p[LOAD_LATENCY] // end of load 122#define ELD_1 p[LOAD_LATENCY+1] // and next stage 123 124// unsigned long do_csum(unsigned char *buf,long len) 125 126GLOBAL_ENTRY(do_csum) 127 .prologue 128 .save ar.pfs, saved_pfs 129 alloc saved_pfs=ar.pfs,2,16,0,16 130 .rotr word1[4], word2[4],result1[LOAD_LATENCY+2],result2[LOAD_LATENCY+2] 131 .rotp p[PIPE_DEPTH], pC1[2], pC2[2] 132 mov ret0=r0 // in case we have zero length 133 cmp.lt p0,p6=r0,len // check for zero length or negative (32bit len) 134 ;; 135 add tmp1=buf,len // last byte's address 136 .save pr, saved_pr 137 mov saved_pr=pr // preserve predicates (rotation) 138(p6) br.ret.spnt.many rp // return if zero or negative length 139 140 mov hmask=-1 // initialize head mask 141 tbit.nz p15,p0=buf,0 // is buf an odd address? 142 and first1=-8,buf // 8-byte align down address of first1 element 143 144 and firstoff=7,buf // how many bytes off for first1 element 145 mov tmask=-1 // initialize tail mask 146 147 ;; 148 adds tmp2=-1,tmp1 // last-1 149 and lastoff=7,tmp1 // how many bytes off for last element 150 ;; 151 sub tmp1=8,lastoff // complement to lastoff 152 and last=-8,tmp2 // address of word containing last byte 153 ;; 154 sub tmp3=last,first1 // tmp3=distance from first1 to last 155 .save ar.lc, saved_lc 156 mov saved_lc=ar.lc // save lc 157 cmp.eq p8,p9=last,first1 // everything fits in one word ? 158 159 ld8 firstval=[first1],8 // load, ahead of time, "first1" word 160 and tmp1=7, tmp1 // make sure that if tmp1==8 -> tmp1=0 161 shl tmp2=firstoff,3 // number of bits 162 ;; 163(p9) ld8 lastval=[last] // load, ahead of time, "last" word, if needed 164 shl tmp1=tmp1,3 // number of bits 165(p9) adds tmp3=-8,tmp3 // effectively loaded 166 ;; 167(p8) mov lastval=r0 // we don't need lastval if first1==last 168 shl hmask=hmask,tmp2 // build head mask, mask off [0,first1off[ 169 shr.u tmask=tmask,tmp1 // build tail mask, mask off ]8,lastoff] 170 ;; 171 .body 172#define count tmp3 173 174(p8) and hmask=hmask,tmask // apply tail mask to head mask if 1 word only 175(p9) and word2[0]=lastval,tmask // mask last it as appropriate 176 shr.u count=count,3 // how many 8-byte? 177 ;; 178 // If count is odd, finish this 8-byte word so that we can 179 // load two back-to-back 8-byte words per loop thereafter. 180 and word1[0]=firstval,hmask // and mask it as appropriate 181 tbit.nz p10,p11=count,0 // if (count is odd) 182 ;; 183(p8) mov result1[0]=word1[0] 184(p9) add result1[0]=word1[0],word2[0] 185 ;; 186 cmp.ltu p6,p0=result1[0],word1[0] // check the carry 187 cmp.eq.or.andcm p8,p0=0,count // exit if zero 8-byte 188 ;; 189(p6) adds result1[0]=1,result1[0] 190(p8) br.cond.dptk .do_csum_exit // if (within an 8-byte word) 191(p11) br.cond.dptk .do_csum16 // if (count is even) 192 193 // Here count is odd. 194 ld8 word1[1]=[first1],8 // load an 8-byte word 195 cmp.eq p9,p10=1,count // if (count == 1) 196 adds count=-1,count // loaded an 8-byte word 197 ;; 198 add result1[0]=result1[0],word1[1] 199 ;; 200 cmp.ltu p6,p0=result1[0],word1[1] 201 ;; 202(p6) adds result1[0]=1,result1[0] 203(p9) br.cond.sptk .do_csum_exit // if (count == 1) exit 204 // Fall through to caluculate the checksum, feeding result1[0] as 205 // the initial value in result1[0]. 206 // 207 // Calculate the checksum loading two 8-byte words per loop. 208 // 209.do_csum16: 210 add first2=8,first1 211 shr.u count=count,1 // we do 16 bytes per loop 212 ;; 213 adds count=-1,count 214 mov carry1=r0 215 mov carry2=r0 216 brp.loop.imp 1f,2f 217 ;; 218 mov ar.ec=PIPE_DEPTH 219 mov ar.lc=count // set lc 220 mov pr.rot=1<<16 221 // result1[0] must be initialized in advance. 222 mov result2[0]=r0 223 ;; 224 .align 32 2251: 226(ELD_1) cmp.ltu pC1[0],p0=result1[LOAD_LATENCY],word1[LOAD_LATENCY+1] 227(pC1[1])adds carry1=1,carry1 228(ELD_1) cmp.ltu pC2[0],p0=result2[LOAD_LATENCY],word2[LOAD_LATENCY+1] 229(pC2[1])adds carry2=1,carry2 230(ELD) add result1[LOAD_LATENCY-1]=result1[LOAD_LATENCY],word1[LOAD_LATENCY] 231(ELD) add result2[LOAD_LATENCY-1]=result2[LOAD_LATENCY],word2[LOAD_LATENCY] 2322: 233(p[0]) ld8 word1[0]=[first1],16 234(p[0]) ld8 word2[0]=[first2],16 235 br.ctop.sptk 1b 236 ;; 237 // Since len is a 32-bit value, carry cannot be larger than a 64-bit value. 238(pC1[1])adds carry1=1,carry1 // since we miss the last one 239(pC2[1])adds carry2=1,carry2 240 ;; 241 add result1[LOAD_LATENCY+1]=result1[LOAD_LATENCY+1],carry1 242 add result2[LOAD_LATENCY+1]=result2[LOAD_LATENCY+1],carry2 243 ;; 244 cmp.ltu p6,p0=result1[LOAD_LATENCY+1],carry1 245 cmp.ltu p7,p0=result2[LOAD_LATENCY+1],carry2 246 ;; 247(p6) adds result1[LOAD_LATENCY+1]=1,result1[LOAD_LATENCY+1] 248(p7) adds result2[LOAD_LATENCY+1]=1,result2[LOAD_LATENCY+1] 249 ;; 250 add result1[0]=result1[LOAD_LATENCY+1],result2[LOAD_LATENCY+1] 251 ;; 252 cmp.ltu p6,p0=result1[0],result2[LOAD_LATENCY+1] 253 ;; 254(p6) adds result1[0]=1,result1[0] 255 ;; 256.do_csum_exit: 257 // 258 // now fold 64 into 16 bits taking care of carry 259 // that's not very good because it has lots of sequentiality 260 // 261 mov tmp3=0xffff 262 zxt4 tmp1=result1[0] 263 shr.u tmp2=result1[0],32 264 ;; 265 add result1[0]=tmp1,tmp2 266 ;; 267 and tmp1=result1[0],tmp3 268 shr.u tmp2=result1[0],16 269 ;; 270 add result1[0]=tmp1,tmp2 271 ;; 272 and tmp1=result1[0],tmp3 273 shr.u tmp2=result1[0],16 274 ;; 275 add result1[0]=tmp1,tmp2 276 ;; 277 and tmp1=result1[0],tmp3 278 shr.u tmp2=result1[0],16 279 ;; 280 add ret0=tmp1,tmp2 281 mov pr=saved_pr,0xffffffffffff0000 282 ;; 283 // if buf was odd then swap bytes 284 mov ar.pfs=saved_pfs // restore ar.ec 285(p15) mux1 ret0=ret0,@rev // reverse word 286 ;; 287 mov ar.lc=saved_lc 288(p15) shr.u ret0=ret0,64-16 // + shift back to position = swap bytes 289 br.ret.sptk.many rp 290 291// I (Jun Nakajima) wrote an equivalent code (see below), but it was 292// not much better than the original. So keep the original there so that 293// someone else can challenge. 294// 295// shr.u word1[0]=result1[0],32 296// zxt4 result1[0]=result1[0] 297// ;; 298// add result1[0]=result1[0],word1[0] 299// ;; 300// zxt2 result2[0]=result1[0] 301// extr.u word1[0]=result1[0],16,16 302// shr.u carry1=result1[0],32 303// ;; 304// add result2[0]=result2[0],word1[0] 305// ;; 306// add result2[0]=result2[0],carry1 307// ;; 308// extr.u ret0=result2[0],16,16 309// ;; 310// add ret0=ret0,result2[0] 311// ;; 312// zxt2 ret0=ret0 313// mov ar.pfs=saved_pfs // restore ar.ec 314// mov pr=saved_pr,0xffffffffffff0000 315// ;; 316// // if buf was odd then swap bytes 317// mov ar.lc=saved_lc 318//(p15) mux1 ret0=ret0,@rev // reverse word 319// ;; 320//(p15) shr.u ret0=ret0,64-16 // + shift back to position = swap bytes 321// br.ret.sptk.many rp 322 323END(do_csum) 324