* Optimal parser
*********************************/
-static U32 ZSTD_totalLen(ZSTD_optimal_t sol)
-{
- return sol.litlen + sol.mlen;
-}
-
#if 0 /* debug */
static void
ZSTD_optimal_t* const opt = optStatePtr->priceTable;
ZSTD_match_t* const matches = optStatePtr->matchTable;
- ZSTD_optimal_t lastSequence;
+ ZSTD_optimal_t lastStretch;
ZSTD_optLdm_t optLdm;
- ZSTD_memset(&lastSequence, 0, sizeof(ZSTD_optimal_t));
+ ZSTD_memset(&lastStretch, 0, sizeof(ZSTD_optimal_t));
optLdm.seqStore = ms->ldmSeqStore ? *ms->ldmSeqStore : kNullRawSeqStore;
optLdm.endPosInBlock = optLdm.startPosInBlock = optLdm.offset = 0;
U32 nbMatches = getAllMatches(matches, ms, &nextToUpdate3, ip, iend, rep, ll0, minMatch);
ZSTD_optLdm_processMatchCandidate(&optLdm, matches, &nbMatches,
(U32)(ip-istart), (U32)(iend-ip));
- if (!nbMatches) { ip++; continue; }
+ if (!nbMatches) {
+ DEBUGLOG(8, "no match found at cPos %u", (unsigned)(ip-istart));
+ ip++;
+ continue;
+ }
/* initialize opt[0] */
- { U32 i ; for (i=0; i<ZSTD_REP_NUM; i++) opt[0].rep[i] = rep[i]; }
- opt[0].mlen = 0; /* means is_a_literal */
+ opt[0].mlen = 0; /* there are only literals so far */
opt[0].litlen = litlen;
- /* We don't need to include the actual price of the literals because
- * it is static for the duration of the forward pass, and is included
- * in every price. We include the literal length to avoid negative
- * prices when we subtract the previous literal length.
+ /* No need to include the actual price of the literals before the segment
+ * because it is static for the duration of the forward pass, and is included
+ * in every subsequent price. We include the literal length as the cost variation
+ * of litlen depends on the value of litlen.
*/
opt[0].price = LL_PRICE(litlen);
+ ZSTD_STATIC_ASSERT(sizeof(opt[0].rep[0]) == sizeof(rep[0]));
+ memcpy(&opt[0].rep, rep, sizeof(opt[0].rep));
/* large match -> immediate encoding */
{ U32 const maxML = matches[nbMatches-1].len;
nbMatches, maxML, maxOffBase, (U32)(ip-prefixStart));
if (maxML > sufficient_len) {
- lastSequence.litlen = litlen;
- lastSequence.mlen = maxML;
- lastSequence.off = maxOffBase;
- DEBUGLOG(6, "large match (%u>%u), immediate encoding",
+ lastStretch.litlen = 0;
+ lastStretch.mlen = maxML;
+ lastStretch.off = maxOffBase;
+ DEBUGLOG(6, "large match (%u>%u) => immediate encoding",
maxML, sufficient_len);
cur = 0;
- last_pos = ZSTD_totalLen(lastSequence);
+ last_pos = maxML;
goto _shortestPath;
} }
/* set prices for first matches starting position == 0 */
assert(opt[0].price >= 0);
- { U32 const literalsPrice = (U32)opt[0].price + (U32)LL_PRICE(0);
- U32 pos;
+ { U32 pos;
U32 matchNb;
for (pos = 1; pos < minMatch; pos++) {
- opt[pos].price = ZSTD_MAX_PRICE; /* mlen, litlen and price will be fixed during forward scanning */
- opt[pos].mlen = 0;
+ opt[pos].price = ZSTD_MAX_PRICE;
+ /* will be updated later on at match check */
}
for (matchNb = 0; matchNb < nbMatches; matchNb++) {
U32 const offBase = matches[matchNb].off;
U32 const end = matches[matchNb].len;
for ( ; pos <= end ; pos++ ) {
- U32 const matchPrice = ZSTD_getMatchPrice(offBase, pos, optStatePtr, optLevel);
- U32 const sequencePrice = literalsPrice + matchPrice;
+ int const matchPrice = (int)ZSTD_getMatchPrice(offBase, pos, optStatePtr, optLevel);
+ int const sequencePrice = opt[0].price + matchPrice;
DEBUGLOG(7, "rPos:%u => set initial price : %.2f",
- pos, ZSTD_fCost((int)sequencePrice));
+ pos, ZSTD_fCost(sequencePrice));
opt[pos].mlen = pos;
opt[pos].off = offBase;
- opt[pos].litlen = litlen;
- opt[pos].price = (int)sequencePrice;
+ opt[pos].litlen = 0; /* end of match */
+ opt[pos].price = sequencePrice + LL_PRICE(0);
}
opt[pos].price = ZSTD_MAX_PRICE;
}
DEBUGLOG(7, "cPos:%zi==rPos:%u", inr-istart, cur);
/* Fix current position with one literal if cheaper */
- { U32 const litlen = (opt[cur-1].mlen == 0) ? opt[cur-1].litlen + 1 : 1;
+ { U32 const litlen = opt[cur-1].litlen + 1;
int const price = opt[cur-1].price
+ LIT_PRICE(ip+cur-1)
+ LL_INCPRICE(litlen);
DEBUGLOG(7, "cPos:%zi==rPos:%u : better price (%.2f<=%.2f) using literal (ll==%u) (hist:%u,%u,%u)",
inr-istart, cur, ZSTD_fCost(price), ZSTD_fCost(opt[cur].price), litlen,
opt[cur-1].rep[0], opt[cur-1].rep[1], opt[cur-1].rep[2]);
- if ( (optLevel == 2) /* additional check only for high modes */
+ if ( 0 && (optLevel == 2) /* additional check only for high modes */
&& (opt[cur].mlen > 0) /* interrupt a match */
&& (LL_INCPRICE(1) < 0) ) /* ll1 is cheaper than ll0 */
{
DEBUGLOG(7, "==> match+1lit is cheaper (%.2f < %.2f) !!!", ZSTD_fCost(with1literal), ZSTD_fCost(withMoreLiterals));
/* do not take this literal */
} else {
- opt[cur].mlen = 0;
- opt[cur].off = 0;
+ opt[cur].mlen = opt[cur-1].mlen;
+ opt[cur].off = opt[cur-1].off;
opt[cur].litlen = litlen;
opt[cur].price = price;
}
} else {
- /* normal case: take the literal, it's expected to be cheaper */
- opt[cur].mlen = 0;
- opt[cur].off = 0;
+ /* normal case: take the literal, it's expected to be cheaper at position @cur */
+ opt[cur].mlen = opt[cur-1].mlen;
+ opt[cur].off = opt[cur-1].off;
opt[cur].litlen = litlen;
opt[cur].price = price;
}
*/
ZSTD_STATIC_ASSERT(sizeof(opt[cur].rep) == sizeof(repcodes_t));
assert(cur >= opt[cur].mlen);
- if (opt[cur].mlen != 0) {
+ if (opt[cur].litlen == 0) {
+ /* just finished a match => alter offset history */
U32 const prev = cur - opt[cur].mlen;
- repcodes_t const newReps = ZSTD_newRep(opt[prev].rep, opt[cur].off, opt[cur].litlen==0);
+ repcodes_t const newReps = ZSTD_newRep(opt[prev].rep, opt[cur].off, opt[prev].litlen==0);
ZSTD_memcpy(opt[cur].rep, &newReps, sizeof(repcodes_t));
} else {
ZSTD_memcpy(opt[cur].rep, opt[cur - 1].rep, sizeof(repcodes_t));
if ( (optLevel==0) /*static_test*/
&& (opt[cur+1].price <= opt[cur].price + (BITCOST_MULTIPLIER/2)) ) {
- DEBUGLOG(7, "move to next rPos:%u : price is <=", cur+1);
+ DEBUGLOG(7, "skip current position : next rPos(%u) price is cheaper", cur+1);
continue; /* skip unpromising positions; about ~+6% speed, -0.01 ratio */
}
assert(opt[cur].price >= 0);
- { U32 const ll0 = (opt[cur].mlen != 0);
- U32 const litlen = (opt[cur].mlen == 0) ? opt[cur].litlen : 0;
- U32 const previousPrice = (U32)opt[cur].price;
- U32 const basePrice = previousPrice + (U32)LL_PRICE(0);
+ { U32 const ll0 = (opt[cur].litlen == 0);
+ int const previousPrice = opt[cur].price;
+ int const basePrice = previousPrice + LL_PRICE(0);
U32 nbMatches = getAllMatches(matches, ms, &nextToUpdate3, inr, iend, opt[cur].rep, ll0, minMatch);
U32 matchNb;
continue;
}
- { U32 const maxML = matches[nbMatches-1].len;
- DEBUGLOG(7, "cPos:%zi==rPos:%u, found %u matches, of maxLength=%u",
- inr-istart, cur, nbMatches, maxML);
-
- if ( (maxML > sufficient_len)
- || (cur + maxML >= ZSTD_OPT_NUM) ) {
- lastSequence.mlen = maxML;
- lastSequence.off = matches[nbMatches-1].off;
- lastSequence.litlen = litlen;
- cur -= (opt[cur].mlen==0) ? opt[cur].litlen : 0; /* last sequence is actually only literals, fix cur to last match - note : may underflow, in which case, it's first sequence, and it's okay */
- last_pos = cur + ZSTD_totalLen(lastSequence);
- if (cur > ZSTD_OPT_NUM) cur = 0; /* underflow => first match */
+ { U32 const longestML = matches[nbMatches-1].len;
+ DEBUGLOG(7, "cPos:%zi==rPos:%u, found %u matches, of longest ML=%u",
+ inr-istart, cur, nbMatches, longestML);
+
+ if ( (longestML > sufficient_len)
+ || (cur + longestML >= ZSTD_OPT_NUM) ) {
+ lastStretch.mlen = longestML;
+ lastStretch.off = matches[nbMatches-1].off;
+ lastStretch.litlen = 0;
+ last_pos = cur + longestML;
goto _shortestPath;
} }
U32 mlen;
DEBUGLOG(7, "testing match %u => offBase=%4u, mlen=%2u, llen=%2u",
- matchNb, matches[matchNb].off, lastML, litlen);
+ matchNb, matches[matchNb].off, lastML, opt[cur].litlen);
for (mlen = lastML; mlen >= startML; mlen--) { /* scan downward */
U32 const pos = cur + mlen;
- int const price = (int)basePrice + (int)ZSTD_getMatchPrice(offset, mlen, optStatePtr, optLevel);
+ int const price = basePrice + (int)ZSTD_getMatchPrice(offset, mlen, optStatePtr, optLevel);
if ((pos > last_pos) || (price < opt[pos].price)) {
DEBUGLOG(7, "rPos:%u (ml=%2u) => new better price (%.2f<%.2f)",
while (last_pos < pos) { opt[last_pos+1].price = ZSTD_MAX_PRICE; last_pos++; } /* fill empty positions */
opt[pos].mlen = mlen;
opt[pos].off = offset;
- opt[pos].litlen = litlen;
+ opt[pos].litlen = 0;
opt[pos].price = price;
} else {
DEBUGLOG(7, "rPos:%u (ml=%2u) => new price is worse (%.2f>=%.2f)",
opt[last_pos+1].price = ZSTD_MAX_PRICE;
} /* for (cur = 1; cur <= last_pos; cur++) */
- lastSequence = opt[last_pos];
- cur = last_pos > ZSTD_totalLen(lastSequence) ? last_pos - ZSTD_totalLen(lastSequence) : 0; /* single sequence, and it starts before `ip` */
- assert(cur < ZSTD_OPT_NUM); /* control overflow*/
+ lastStretch = opt[last_pos];
+ assert(cur >= lastStretch.mlen);
+ cur = last_pos - lastStretch.mlen;
_shortestPath: /* cur, last_pos, best_mlen, best_off have to be set */
assert(opt[0].mlen == 0);
+ assert(last_pos >= lastStretch.mlen);
+ assert(cur == last_pos - lastStretch.mlen);
+ assert(lastStretch.rep[0] != 0);
+
+ if (lastStretch.mlen==0) {
+ /* no solution : all matches have been converted into literals */
+ assert(lastStretch.litlen == (ip - anchor) + last_pos);
+ ip += last_pos;
+ continue;
+ }
+ assert(lastStretch.off > 0);
- /* Set the next chunk's repcodes based on the repcodes of the beginning
- * of the last match, and the last sequence. This avoids us having to
- * update them while traversing the sequences.
- */
- if (lastSequence.mlen != 0) {
- repcodes_t const reps = ZSTD_newRep(opt[cur].rep, lastSequence.off, lastSequence.litlen==0);
- ZSTD_memcpy(rep, &reps, sizeof(reps));
+ /* Update offset history */
+ if (lastStretch.litlen == 0) {
+ /* finishing on a match : update offset history */
+ repcodes_t const reps = ZSTD_newRep(opt[cur].rep, lastStretch.off, opt[cur].litlen==0);
+ ZSTD_memcpy(rep, &reps, sizeof(repcodes_t));
} else {
- ZSTD_memcpy(rep, opt[cur].rep, sizeof(repcodes_t));
+ ZSTD_memcpy(rep, lastStretch.rep, sizeof(repcodes_t));
+ assert(cur >= lastStretch.litlen);
+ cur -= lastStretch.litlen;
}
- { U32 const storeEnd = cur + 1;
+ /* let's write the shortest path solution
+ * solution is stored in @opt,
+ * in reverse order,
+ * starting from @storeEnd (==cur+1)
+ * (effectively partially overwriting @opt).
+ * Content is changed too:
+ * - So far, @opt stored stretches, aka a match followed by literals
+ * - Now, it will store sequences, aka literals followed by a match
+ */
+ { U32 const storeEnd = cur + 2;
U32 storeStart = storeEnd;
- U32 seqPos = cur;
+ U32 stretchPos = cur;
+ ZSTD_optimal_t nextStretch;
DEBUGLOG(6, "start reverse traversal (last_pos:%u, cur:%u)",
last_pos, cur); (void)last_pos;
assert(storeEnd < ZSTD_OPT_NUM);
- DEBUGLOG(6, "last sequence copied into pos=%u (llen=%u,mlen=%u,ofc=%u)",
- storeEnd, lastSequence.litlen, lastSequence.mlen, lastSequence.off);
- opt[storeEnd] = lastSequence;
- while (seqPos > 0) {
- U32 const backDist = ZSTD_totalLen(opt[seqPos]);
+ DEBUGLOG(6, "last stretch copied into pos=%u (llen=%u,mlen=%u,ofc=%u)",
+ storeEnd, lastStretch.litlen, lastStretch.mlen, lastStretch.off);
+ if (lastStretch.litlen > 0) {
+ /* last "sequence" is unfinished: just a bunch of literals */
+ opt[storeEnd].litlen = lastStretch.litlen;
+ opt[storeEnd].mlen = 0;
+ storeStart = storeEnd-1;
+ opt[storeStart] = lastStretch;
+ } {
+ opt[storeEnd] = lastStretch; /* note: litlen will be fixed */
+ storeStart = storeEnd;
+ }
+ while (1) {
+ nextStretch = opt[stretchPos];
+ opt[storeStart].litlen = nextStretch.litlen;
+ DEBUGLOG(6, "selected sequence (llen=%u,mlen=%u,ofc=%u)",
+ opt[storeStart].litlen, opt[storeStart].mlen, opt[storeStart].off);
+ if (nextStretch.mlen == 0) {
+ /* reaching beginning of segment */
+ break;
+ }
storeStart--;
- DEBUGLOG(6, "sequence from rPos=%u copied into pos=%u (llen=%u,mlen=%u,ofc=%u)",
- seqPos, storeStart, opt[seqPos].litlen, opt[seqPos].mlen, opt[seqPos].off);
- opt[storeStart] = opt[seqPos];
- seqPos = (seqPos > backDist) ? seqPos - backDist : 0;
+ opt[storeStart] = nextStretch; /* note: litlen will be fixed */
+ assert(nextStretch.litlen + nextStretch.mlen <= stretchPos);
+ stretchPos -= nextStretch.litlen + nextStretch.mlen;
}
/* save sequences */
anchor += advance;
ip = anchor;
} }
+ DEBUGLOG(7, "new offset history : %u, %u, %u", rep[0], rep[1], rep[2]);
+
+ /* update all costs */
ZSTD_setBasePrices(optStatePtr, optLevel);
}
} /* while (ip < ilimit) */
* Consequently, this can only work if no data has been previously loaded in tables,
* aka, no dictionary, no prefix, no ldm preprocessing.
* The compression ratio gain is generally small (~0.5% on first block),
- ** the cost is 2x cpu time on first block. */
+ * the cost is 2x cpu time on first block. */
assert(srcSize <= ZSTD_BLOCKSIZE_MAX);
if ( (ms->opt.litLengthSum==0) /* first block */
&& (seqStore->sequences == seqStore->sequencesStart) /* no ldm */
projects / thirdparty / zstd.git / commitdiff
