/**
* Take in a collection of exclusive sub engines and produces a tamarama, also
* returns via out_top_remap, a mapping indicating how tops in the subengines in
* relate to the tamarama's tops.
*/
aligned_unique_ptr<NFA> buildTamarama(const TamaInfo &tamaInfo, const u32 queue,
map<pair<const NFA *, u32>, u32> &out_top_remap) {
vector<u32> top_base;
remapTops(tamaInfo, top_base, out_top_remap);
size_t subSize = tamaInfo.subengines.size();
DEBUG_PRINTF("subSize:%lu\n", subSize);
size_t total_size =
sizeof(NFA) + // initial NFA structure
sizeof(Tamarama) + // Tamarama structure
sizeof(u32) * subSize + // base top event value for subengines,
// used for top remapping at runtime
sizeof(u32) * subSize + 64; // offsets to subengines in bytecode and
// padding for subengines
for (const auto &sub : tamaInfo.subengines) {
total_size += ROUNDUP_CL(sub->length);
}
// use subSize as a sentinel value for no active subengines,
// so add one to subSize here
u32 activeIdxSize = calcPackedBytes(subSize + 1);
aligned_unique_ptr<NFA> nfa = aligned_zmalloc_unique<NFA>(total_size);
nfa->type = verify_u8(TAMARAMA_NFA_0);
nfa->length = verify_u32(total_size);
nfa->queueIndex = queue;
char *ptr = (char *)nfa.get() + sizeof(NFA);
char *base_offset = ptr;
Tamarama *t = (Tamarama *)ptr;
t->numSubEngines = verify_u32(subSize);
t->activeIdxSize = verify_u8(activeIdxSize);
ptr += sizeof(Tamarama);
copy_bytes(ptr, top_base);
ptr += byte_length(top_base);
u32 *offsets = (u32*)ptr;
char *sub_nfa_offset = ptr + sizeof(u32) * subSize;
copyInSubnfas(base_offset, *nfa, tamaInfo, offsets, sub_nfa_offset,
activeIdxSize);
assert((size_t)(sub_nfa_offset - (char *)nfa.get()) <= total_size);
return nfa;
}
template <class LbrStruct> static
void fillNfa(NFA *nfa, lbr_common *c, ReportID report, const depth &repeatMin,
const depth &repeatMax, u32 minPeriod, enum RepeatType rtype) {
assert(nfa);
RepeatStateInfo rsi(rtype, repeatMin, repeatMax, minPeriod);
DEBUG_PRINTF("selected %s model for {%s,%s} repeat\n",
repeatTypeName(rtype), repeatMin.str().c_str(),
repeatMax.str().c_str());
// Fill the lbr_common structure first. Note that the RepeatInfo structure
// directly follows the LbrStruct.
const u32 info_offset = sizeof(LbrStruct);
c->repeatInfoOffset = info_offset;
c->report = report;
RepeatInfo *info = (RepeatInfo *)((char *)c + info_offset);
info->type = verify_u8(rtype);
info->repeatMin = depth_to_u32(repeatMin);
info->repeatMax = depth_to_u32(repeatMax);
info->stateSize = rsi.stateSize;
info->packedCtrlSize = rsi.packedCtrlSize;
info->horizon = rsi.horizon;
info->minPeriod = minPeriod;
copy_bytes(&info->packedFieldSizes, rsi.packedFieldSizes);
info->patchCount = rsi.patchCount;
info->patchSize = rsi.patchSize;
info->encodingSize = rsi.encodingSize;
info->patchesOffset = rsi.patchesOffset;
// Fill the NFA structure.
nfa->nPositions = repeatMin;
nfa->streamStateSize = verify_u32(rsi.packedCtrlSize + rsi.stateSize);
nfa->scratchStateSize = (u32)sizeof(lbr_state);
nfa->minWidth = verify_u32(repeatMin);
nfa->maxWidth = repeatMax.is_finite() ? verify_u32(repeatMax) : 0;
// Fill the lbr table for sparse lbr model.
if (rtype == REPEAT_SPARSE_OPTIMAL_P) {
u64a *table = getTable<LbrStruct>(nfa);
// Adjust table length according to the optimal patch length.
size_t len = nfa->length;
assert((u32)repeatMax >= rsi.patchSize);
len -= sizeof(u64a) * ((u32)repeatMax - rsi.patchSize);
nfa->length = verify_u32(len);
info->length = verify_u32(sizeof(RepeatInfo)
+ sizeof(u64a) * (rsi.patchSize + 1));
copy_bytes(table, rsi.table);
}
}
template <class LbrStruct> static
aligned_unique_ptr<NFA> makeLbrNfa(NFAEngineType nfa_type,
enum RepeatType rtype,
const depth &repeatMax) {
size_t tableLen = 0;
if (rtype == REPEAT_SPARSE_OPTIMAL_P) {
tableLen = sizeof(u64a) * (repeatMax + 1);
}
size_t len = sizeof(NFA) + sizeof(LbrStruct) + sizeof(RepeatInfo) +
tableLen + sizeof(u64a);
aligned_unique_ptr<NFA> nfa = aligned_zmalloc_unique<NFA>(len);
nfa->type = verify_u8(nfa_type);
nfa->length = verify_u32(len);
return nfa;
}
vector<u8> findLeftOffsetStopAlphabet(const CastleProto &castle,
UNUSED som_type som) {
const depth max_width = findMaxWidth(castle);
DEBUG_PRINTF("castle has reach %s and max width %s\n",
describeClass(castle.reach()).c_str(),
max_width.str().c_str());
const CharReach escape = ~castle.reach(); // invert reach for stop chars.
u32 d = min(max_width, depth(MAX_STOP_DEPTH));
const u8 mask = verify_u8((1U << d) - 1);
vector<u8> stop(N_CHARS, 0);
for (size_t c = escape.find_first(); c != escape.npos;
c = escape.find_next(c)) {
stop[c] |= mask;
}
return stop;
}
static
bool findDVerm(const vector<const hwlmLiteral *> &lits, AccelAux *aux) {
const hwlmLiteral &first = *lits.front();
struct candidate {
candidate(void)
: c1(0), c2(0), max_offset(0), b5insens(false), valid(false) {}
candidate(const hwlmLiteral &base, u32 offset)
: c1(base.s[offset]), c2(base.s[offset + 1]), max_offset(0),
b5insens(false), valid(true) {}
char c1;
char c2;
u32 max_offset;
bool b5insens;
bool valid;
bool operator>(const candidate &other) const {
if (!valid) {
return false;
}
if (!other.valid) {
return true;
}
if (other.cdiffers() && !cdiffers()) {
return false;
}
if (!other.cdiffers() && cdiffers()) {
return true;
}
if (!other.b5insens && b5insens) {
return false;
}
if (other.b5insens && !b5insens) {
return true;
}
if (max_offset > other.max_offset) {
return false;
}
return true;
}
bool cdiffers(void) const {
if (!b5insens) {
return c1 != c2;
}
return (c1 & CASE_CLEAR) != (c2 & CASE_CLEAR);
}
};
candidate best;
for (u32 i = 0; i < MIN(MAX_ACCEL_OFFSET, first.s.length()) - 1; i++) {
candidate curr(first, i);
/* check to see if this pair appears in each string */
for (const auto &lit_ptr : lits) {
const hwlmLiteral &lit = *lit_ptr;
if (lit.nocase && (ourisalpha(curr.c1) || ourisalpha(curr.c2))) {
curr.b5insens = true; /* no choice but to be case insensitive */
}
bool found = false;
bool found_nc = false;
for (u32 j = 0;
!found && j < MIN(MAX_ACCEL_OFFSET, lit.s.length()) - 1; j++) {
found |= curr.c1 == lit.s[j] && curr.c2 == lit.s[j + 1];
found_nc |= (curr.c1 & CASE_CLEAR) == (lit.s[j] & CASE_CLEAR)
&& (curr.c2 & CASE_CLEAR) == (lit.s[j + 1] & CASE_CLEAR);
if (curr.b5insens) {
found = found_nc;
}
}
if (!curr.b5insens && !found && found_nc) {
curr.b5insens = true;
found = true;
}
if (!found) {
goto next_candidate;
}
}
/* check to find the max offset where this appears */
for (const auto &lit_ptr : lits) {
const hwlmLiteral &lit = *lit_ptr;
for (u32 j = 0; j < MIN(MAX_ACCEL_OFFSET, lit.s.length()) - 1;
j++) {
bool found = false;
if (curr.b5insens) {
found = (curr.c1 & CASE_CLEAR) == (lit.s[j] & CASE_CLEAR)
&& (curr.c2 & CASE_CLEAR) == (lit.s[j + 1] & CASE_CLEAR);
} else {
found = curr.c1 == lit.s[j] && curr.c2 == lit.s[j + 1];
}
if (found) {
curr.max_offset = MAX(curr.max_offset, j);
break;
}
}
}
if (curr > best) {
best = curr;
}
next_candidate:;
}
if (!best.valid) {
return false;
}
aux->dverm.offset = verify_u8(best.max_offset);
if (!best.b5insens) {
aux->dverm.accel_type = ACCEL_DVERM;
aux->dverm.c1 = best.c1;
aux->dverm.c2 = best.c2;
DEBUG_PRINTF("built dverm for %02hhx%02hhx\n",
aux->dverm.c1, aux->dverm.c2);
} else {
aux->dverm.accel_type = ACCEL_DVERM_NOCASE;
aux->dverm.c1 = best.c1 & CASE_CLEAR;
aux->dverm.c2 = best.c2 & CASE_CLEAR;
DEBUG_PRINTF("built dverm nc for %02hhx%02hhx\n",
aux->dverm.c1, aux->dverm.c2);
}
return true;
}
static
void findForwardAccelScheme(const vector<hwlmLiteral> &lits,
hwlm_group_t expected_groups, AccelAux *aux) {
DEBUG_PRINTF("building accel expected=%016llx\n", expected_groups);
u32 min_len = MAX_ACCEL_OFFSET;
vector<const hwlmLiteral *> filtered_lits;
filterLits(lits, expected_groups, &filtered_lits, &min_len);
if (filtered_lits.empty()) {
return;
}
if (findDVerm(filtered_lits, aux)
|| findSVerm(filtered_lits, aux)) {
return;
}
vector<CharReach> reach(MAX_ACCEL_OFFSET, CharReach());
for (const auto &lit : lits) {
if (!(lit.groups & expected_groups)) {
continue;
}
for (u32 i = 0; i < MAX_ACCEL_OFFSET && i < lit.s.length(); i++) {
unsigned char c = lit.s[i];
if (lit.nocase) {
DEBUG_PRINTF("adding %02hhx to %u\n", mytoupper(c), i);
DEBUG_PRINTF("adding %02hhx to %u\n", mytolower(c), i);
reach[i].set(mytoupper(c));
reach[i].set(mytolower(c));
} else {
DEBUG_PRINTF("adding %02hhx to %u\n", c, i);
reach[i].set(c);
}
}
}
u32 min_count = ~0U;
u32 min_offset = ~0U;
for (u32 i = 0; i < min_len; i++) {
size_t count = reach[i].count();
DEBUG_PRINTF("offset %u is %s (reach %zu)\n", i,
describeClass(reach[i]).c_str(), count);
if (count < min_count) {
min_count = (u32)count;
min_offset = i;
}
}
assert(min_offset <= min_len);
if (min_count > MAX_SHUFTI_WIDTH) {
DEBUG_PRINTF("min shufti with %u chars is too wide\n", min_count);
return;
}
const CharReach &cr = reach[min_offset];
if (shuftiBuildMasks(cr, &aux->shufti.lo, &aux->shufti.hi) != -1) {
DEBUG_PRINTF("built shufti for %s (%zu chars, offset %u)\n",
describeClass(cr).c_str(), cr.count(), min_offset);
aux->shufti.accel_type = ACCEL_SHUFTI;
aux->shufti.offset = verify_u8(min_offset);
return;
}
DEBUG_PRINTF("fail\n");
}