From 1da177e4c3f41524e886b7f1b8a0c1fc7321cac2 Mon Sep 17 00:00:00 2001 From: Linus Torvalds Date: Sat, 16 Apr 2005 15:20:36 -0700 Subject: Linux-2.6.12-rc2 Initial git repository build. I'm not bothering with the full history, even though we have it. We can create a separate "historical" git archive of that later if we want to, and in the meantime it's about 3.2GB when imported into git - space that would just make the early git days unnecessarily complicated, when we don't have a lot of good infrastructure for it. Let it rip! --- arch/ia64/lib/do_csum.S | 323 ++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 323 insertions(+) create mode 100644 arch/ia64/lib/do_csum.S (limited to 'arch/ia64/lib/do_csum.S') diff --git a/arch/ia64/lib/do_csum.S b/arch/ia64/lib/do_csum.S new file mode 100644 index 00000000000..6bec2fc9f5b --- /dev/null +++ b/arch/ia64/lib/do_csum.S @@ -0,0 +1,323 @@ +/* + * + * Optmized version of the standard do_csum() function + * + * Return: a 64bit quantity containing the 16bit Internet checksum + * + * Inputs: + * in0: address of buffer to checksum (char *) + * in1: length of the buffer (int) + * + * Copyright (C) 1999, 2001-2002 Hewlett-Packard Co + * Stephane Eranian + * + * 02/04/22 Ken Chen + * Data locality study on the checksum buffer. + * More optimization cleanup - remove excessive stop bits. + * 02/04/08 David Mosberger + * More cleanup and tuning. + * 01/04/18 Jun Nakajima + * Clean up and optimize and the software pipeline, loading two + * back-to-back 8-byte words per loop. Clean up the initialization + * for the loop. Support the cases where load latency = 1 or 2. + * Set CONFIG_IA64_LOAD_LATENCY to 1 or 2 (default). + */ + +#include + +// +// Theory of operations: +// The goal is to go as quickly as possible to the point where +// we can checksum 16 bytes/loop. Before reaching that point we must +// take care of incorrect alignment of first byte. +// +// The code hereafter also takes care of the "tail" part of the buffer +// before entering the core loop, if any. The checksum is a sum so it +// allows us to commute operations. So we do the "head" and "tail" +// first to finish at full speed in the body. Once we get the head and +// tail values, we feed them into the pipeline, very handy initialization. +// +// Of course we deal with the special case where the whole buffer fits +// into one 8 byte word. In this case we have only one entry in the pipeline. +// +// We use a (LOAD_LATENCY+2)-stage pipeline in the loop to account for +// possible load latency and also to accommodate for head and tail. +// +// The end of the function deals with folding the checksum from 64bits +// down to 16bits taking care of the carry. +// +// This version avoids synchronization in the core loop by also using a +// pipeline for the accumulation of the checksum in resultx[] (x=1,2). +// +// wordx[] (x=1,2) +// |---| +// | | 0 : new value loaded in pipeline +// |---| +// | | - : in transit data +// |---| +// | | LOAD_LATENCY : current value to add to checksum +// |---| +// | | LOAD_LATENCY+1 : previous value added to checksum +// |---| (previous iteration) +// +// resultx[] (x=1,2) +// |---| +// | | 0 : initial value +// |---| +// | | LOAD_LATENCY-1 : new checksum +// |---| +// | | LOAD_LATENCY : previous value of checksum +// |---| +// | | LOAD_LATENCY+1 : final checksum when out of the loop +// |---| +// +// +// See RFC1071 "Computing the Internet Checksum" for various techniques for +// calculating the Internet checksum. +// +// NOT YET DONE: +// - Maybe another algorithm which would take care of the folding at the +// end in a different manner +// - Work with people more knowledgeable than me on the network stack +// to figure out if we could not split the function depending on the +// type of packet or alignment we get. Like the ip_fast_csum() routine +// where we know we have at least 20bytes worth of data to checksum. +// - Do a better job of handling small packets. +// - Note on prefetching: it was found that under various load, i.e. ftp read/write, +// nfs read/write, the L1 cache hit rate is at 60% and L2 cache hit rate is at 99.8% +// on the data that buffer points to (partly because the checksum is often preceded by +// a copy_from_user()). This finding indiate that lfetch will not be beneficial since +// the data is already in the cache. +// + +#define saved_pfs r11 +#define hmask r16 +#define tmask r17 +#define first1 r18 +#define firstval r19 +#define firstoff r20 +#define last r21 +#define lastval r22 +#define lastoff r23 +#define saved_lc r24 +#define saved_pr r25 +#define tmp1 r26 +#define tmp2 r27 +#define tmp3 r28 +#define carry1 r29 +#define carry2 r30 +#define first2 r31 + +#define buf in0 +#define len in1 + +#define LOAD_LATENCY 2 // XXX fix me + +#if (LOAD_LATENCY != 1) && (LOAD_LATENCY != 2) +# error "Only 1 or 2 is supported/tested for LOAD_LATENCY." +#endif + +#define PIPE_DEPTH (LOAD_LATENCY+2) +#define ELD p[LOAD_LATENCY] // end of load +#define ELD_1 p[LOAD_LATENCY+1] // and next stage + +// unsigned long do_csum(unsigned char *buf,long len) + +GLOBAL_ENTRY(do_csum) + .prologue + .save ar.pfs, saved_pfs + alloc saved_pfs=ar.pfs,2,16,0,16 + .rotr word1[4], word2[4],result1[LOAD_LATENCY+2],result2[LOAD_LATENCY+2] + .rotp p[PIPE_DEPTH], pC1[2], pC2[2] + mov ret0=r0 // in case we have zero length + cmp.lt p0,p6=r0,len // check for zero length or negative (32bit len) + ;; + add tmp1=buf,len // last byte's address + .save pr, saved_pr + mov saved_pr=pr // preserve predicates (rotation) +(p6) br.ret.spnt.many rp // return if zero or negative length + + mov hmask=-1 // initialize head mask + tbit.nz p15,p0=buf,0 // is buf an odd address? + and first1=-8,buf // 8-byte align down address of first1 element + + and firstoff=7,buf // how many bytes off for first1 element + mov tmask=-1 // initialize tail mask + + ;; + adds tmp2=-1,tmp1 // last-1 + and lastoff=7,tmp1 // how many bytes off for last element + ;; + sub tmp1=8,lastoff // complement to lastoff + and last=-8,tmp2 // address of word containing last byte + ;; + sub tmp3=last,first1 // tmp3=distance from first1 to last + .save ar.lc, saved_lc + mov saved_lc=ar.lc // save lc + cmp.eq p8,p9=last,first1 // everything fits in one word ? + + ld8 firstval=[first1],8 // load, ahead of time, "first1" word + and tmp1=7, tmp1 // make sure that if tmp1==8 -> tmp1=0 + shl tmp2=firstoff,3 // number of bits + ;; +(p9) ld8 lastval=[last] // load, ahead of time, "last" word, if needed + shl tmp1=tmp1,3 // number of bits +(p9) adds tmp3=-8,tmp3 // effectively loaded + ;; +(p8) mov lastval=r0 // we don't need lastval if first1==last + shl hmask=hmask,tmp2 // build head mask, mask off [0,first1off[ + shr.u tmask=tmask,tmp1 // build tail mask, mask off ]8,lastoff] + ;; + .body +#define count tmp3 + +(p8) and hmask=hmask,tmask // apply tail mask to head mask if 1 word only +(p9) and word2[0]=lastval,tmask // mask last it as appropriate + shr.u count=count,3 // how many 8-byte? + ;; + // If count is odd, finish this 8-byte word so that we can + // load two back-to-back 8-byte words per loop thereafter. + and word1[0]=firstval,hmask // and mask it as appropriate + tbit.nz p10,p11=count,0 // if (count is odd) + ;; +(p8) mov result1[0]=word1[0] +(p9) add result1[0]=word1[0],word2[0] + ;; + cmp.ltu p6,p0=result1[0],word1[0] // check the carry + cmp.eq.or.andcm p8,p0=0,count // exit if zero 8-byte + ;; +(p6) adds result1[0]=1,result1[0] +(p8) br.cond.dptk .do_csum_exit // if (within an 8-byte word) +(p11) br.cond.dptk .do_csum16 // if (count is even) + + // Here count is odd. + ld8 word1[1]=[first1],8 // load an 8-byte word + cmp.eq p9,p10=1,count // if (count == 1) + adds count=-1,count // loaded an 8-byte word + ;; + add result1[0]=result1[0],word1[1] + ;; + cmp.ltu p6,p0=result1[0],word1[1] + ;; +(p6) adds result1[0]=1,result1[0] +(p9) br.cond.sptk .do_csum_exit // if (count == 1) exit + // Fall through to caluculate the checksum, feeding result1[0] as + // the initial value in result1[0]. + // + // Calculate the checksum loading two 8-byte words per loop. + // +.do_csum16: + add first2=8,first1 + shr.u count=count,1 // we do 16 bytes per loop + ;; + adds count=-1,count + mov carry1=r0 + mov carry2=r0 + brp.loop.imp 1f,2f + ;; + mov ar.ec=PIPE_DEPTH + mov ar.lc=count // set lc + mov pr.rot=1<<16 + // result1[0] must be initialized in advance. + mov result2[0]=r0 + ;; + .align 32 +1: +(ELD_1) cmp.ltu pC1[0],p0=result1[LOAD_LATENCY],word1[LOAD_LATENCY+1] +(pC1[1])adds carry1=1,carry1 +(ELD_1) cmp.ltu pC2[0],p0=result2[LOAD_LATENCY],word2[LOAD_LATENCY+1] +(pC2[1])adds carry2=1,carry2 +(ELD) add result1[LOAD_LATENCY-1]=result1[LOAD_LATENCY],word1[LOAD_LATENCY] +(ELD) add result2[LOAD_LATENCY-1]=result2[LOAD_LATENCY],word2[LOAD_LATENCY] +2: +(p[0]) ld8 word1[0]=[first1],16 +(p[0]) ld8 word2[0]=[first2],16 + br.ctop.sptk 1b + ;; + // Since len is a 32-bit value, carry cannot be larger than a 64-bit value. +(pC1[1])adds carry1=1,carry1 // since we miss the last one +(pC2[1])adds carry2=1,carry2 + ;; + add result1[LOAD_LATENCY+1]=result1[LOAD_LATENCY+1],carry1 + add result2[LOAD_LATENCY+1]=result2[LOAD_LATENCY+1],carry2 + ;; + cmp.ltu p6,p0=result1[LOAD_LATENCY+1],carry1 + cmp.ltu p7,p0=result2[LOAD_LATENCY+1],carry2 + ;; +(p6) adds result1[LOAD_LATENCY+1]=1,result1[LOAD_LATENCY+1] +(p7) adds result2[LOAD_LATENCY+1]=1,result2[LOAD_LATENCY+1] + ;; + add result1[0]=result1[LOAD_LATENCY+1],result2[LOAD_LATENCY+1] + ;; + cmp.ltu p6,p0=result1[0],result2[LOAD_LATENCY+1] + ;; +(p6) adds result1[0]=1,result1[0] + ;; +.do_csum_exit: + // + // now fold 64 into 16 bits taking care of carry + // that's not very good because it has lots of sequentiality + // + mov tmp3=0xffff + zxt4 tmp1=result1[0] + shr.u tmp2=result1[0],32 + ;; + add result1[0]=tmp1,tmp2 + ;; + and tmp1=result1[0],tmp3 + shr.u tmp2=result1[0],16 + ;; + add result1[0]=tmp1,tmp2 + ;; + and tmp1=result1[0],tmp3 + shr.u tmp2=result1[0],16 + ;; + add result1[0]=tmp1,tmp2 + ;; + and tmp1=result1[0],tmp3 + shr.u tmp2=result1[0],16 + ;; + add ret0=tmp1,tmp2 + mov pr=saved_pr,0xffffffffffff0000 + ;; + // if buf was odd then swap bytes + mov ar.pfs=saved_pfs // restore ar.ec +(p15) mux1 ret0=ret0,@rev // reverse word + ;; + mov ar.lc=saved_lc +(p15) shr.u ret0=ret0,64-16 // + shift back to position = swap bytes + br.ret.sptk.many rp + +// I (Jun Nakajima) wrote an equivalent code (see below), but it was +// not much better than the original. So keep the original there so that +// someone else can challenge. +// +// shr.u word1[0]=result1[0],32 +// zxt4 result1[0]=result1[0] +// ;; +// add result1[0]=result1[0],word1[0] +// ;; +// zxt2 result2[0]=result1[0] +// extr.u word1[0]=result1[0],16,16 +// shr.u carry1=result1[0],32 +// ;; +// add result2[0]=result2[0],word1[0] +// ;; +// add result2[0]=result2[0],carry1 +// ;; +// extr.u ret0=result2[0],16,16 +// ;; +// add ret0=ret0,result2[0] +// ;; +// zxt2 ret0=ret0 +// mov ar.pfs=saved_pfs // restore ar.ec +// mov pr=saved_pr,0xffffffffffff0000 +// ;; +// // if buf was odd then swap bytes +// mov ar.lc=saved_lc +//(p15) mux1 ret0=ret0,@rev // reverse word +// ;; +//(p15) shr.u ret0=ret0,64-16 // + shift back to position = swap bytes +// br.ret.sptk.many rp + +END(do_csum) -- cgit v1.2.3