path: root/cbits/znc.c

#include <stddef.h>
#include <stdbool.h>
#include <stdint.h>
#include <ctype.h>
#include <string.h>
#include <assert.h>
#include <emmintrin.h>

#if 0
#include <stdio.h>
#define DEBUG(...) fprintf(stderr, "[znc2] " __VA_ARGS__)
#else
#define DEBUG(...)
#endif


struct scanner {
  const char *buf;
  size_t len;
  size_t cur;  // cursor
};

enum event_kind {
  PARSEERROR = 0,
  JOIN = 1,
  PART = 2,
  QUIT = 3,
  RENICK = 4,
  TALK = 5,
  NOTICE = 6,
  ACT = 7,
  KICK = 8,
  MODE = 9,
  TOPIC = 10,
};

#pragma pack(1)
struct event {
  uint8_t H, M, S;
  uint8_t kind;  // event_kind
  uint32_t ptr1, ptr2, ptr3;
  uint16_t len1, len2, len3;
};

static inline bool isdigit_char(char c) {
  return c >= '0' && c <= '9';
}

// Exists to avoid UB by strict aliasing rule
static inline uint32_t load_uint32(const char *ptr) {
  uint32_t val;
  memcpy(&val, ptr, 4);
  return val;
}

// Exists to avoid UB by strict aliasing rule
static inline uint64_t load_uint64(const char *ptr) {
  uint64_t val;
  memcpy(&val, ptr, 8);
  return val;
}

// parses "[HH:MM:SS] " including the trailing space
static bool parse_hms(struct scanner *s, struct event *dest) {
  const char *const loc = &s->buf[s->cur];
  const size_t len = s->len - s->cur;
  if (len < 11) return false;
  if (loc[0] != '[' || loc[3] != ':' || loc[6] != ':' || loc[9] != ']' || loc[10] != ' ')  return false;
  if (!isdigit(loc[1]) || !isdigit(loc[2]) ||
      !isdigit(loc[4]) || !isdigit(loc[5]) ||
      !isdigit(loc[7]) || !isdigit(loc[8]))
     return false;

  dest->H = 10 * (loc[1] - '0') + loc[2] - '0';
  dest->M = 10 * (loc[4] - '0') + loc[5] - '0';
  dest->S = 10 * (loc[7] - '0') + loc[8] - '0';
  s->cur += 11;
  return true;
}

static inline bool eat_bytes(struct scanner *s, const char *tok, size_t toklen) {
  const char *const loc = &s->buf[s->cur];
  const size_t len = s->len - s->cur;
  if (len >= toklen && memcmp(loc, tok, toklen) == 0) {
    s->cur += toklen;
    return true;
  }
  return false;
}

#define EAT_BYTES(scanner_, tok_) eat_bytes(scanner_, tok_, strlen(tok_))

static void skip_until(struct scanner *s, char delim) {
  const char *const buf = s->buf;
  const size_t len = s->len;
  const size_t cur = s->cur;
  const char *p = memchr(buf + cur, delim, len - cur);
  if (p == NULL) s->cur = len;
  else s->cur = p - buf;
}

static bool parse_nick(struct scanner *s, uint32_t *destptr, uint16_t *destlen) {
  // Delimiter bytes:
  // \t \n \v \r SP <  >
  // 09 0a 0b 0d 20 3c 3e
  //
  // None of the delimiter bytes have the bit 0x40 set, and the majority of
  // IRC nicks consist entirely of bytes with 0x40 set (which include
  // a-zA-Z[]_`). "Guest12" style nicks do have a numeric suffix, but since the
  // fallback loop is at the end, most of it is done with the fast check anyway.

  const char *const buf = s->buf;
  const size_t len = s->len;
  const size_t startcur = s->cur;
  size_t cur = startcur;

  while (cur < len && (buf[cur] & 0x40) == 0x40) cur++;
  while (cur < len &&
      buf[cur] != '\t' && buf[cur] != '\n' && buf[cur] != '\v' && buf[cur] != '\r' &&
      buf[cur] != ' ' && buf[cur] != '<' && buf[cur] != '>')
    cur++;

  if (cur == startcur) return false;
  if ((uint16_t)(cur - startcur) != cur - startcur) return false;
  *destptr = startcur;
  *destlen = cur - startcur;
  s->cur = cur;
  return true;
}

static bool parse_inside_parens(struct scanner *s, uint32_t *destptr, uint16_t *destlen) {
  const char *const buf = s->buf;
  const size_t len = s->len;
  size_t cur = s->cur;

  if (cur + 1 >= len) return false;
  if (buf[cur] != '(') return false;
  cur++;
  const size_t startcur = cur;

  if (len - cur < 16) goto tail_loop;

  while (cur + 16 <= len) {
    const __m128i vec = _mm_loadu_si128((const __m128i*)(buf + cur));
    if (_mm_movemask_epi8(_mm_cmpeq_epi8(vec, _mm_set1_epi8(')' ))) != 0) break;
    if (_mm_movemask_epi8(_mm_cmpeq_epi8(vec, _mm_set1_epi8('\n'))) != 0) return false;
    cur += 16;
  }

tail_loop:
  while (cur < len) {
    if (buf[cur] == ')') goto close_found;
    if (buf[cur] == '\n') return false;
    cur++;
  }

close_found:
  if ((uint16_t)(cur - startcur) != cur - startcur) return false;
  *destptr = startcur;
  *destlen = cur - startcur;
  s->cur = cur + 1;  // skip the closing paren
  return true;
}

static bool parse_enclose_tail(
    struct scanner *s, uint32_t *destptr, uint16_t *destlen, char leftdelim, char rightdelim
) {
  const char *const buf = s->buf;
  const size_t len = s->len;
  size_t cur = s->cur;

  if (cur + 1 >= len) return false;
  if (buf[cur] != leftdelim) return false;
  cur++;
  const size_t startcur = cur;

  const char *p = memchr(buf + cur, '\n', len - cur);
  if (p == NULL) cur = len;
  else cur = p - buf;

  if (cur == startcur) return false;
  const size_t rightdelim_cur = cur - 1;
  if (buf[rightdelim_cur] != rightdelim) return false;

  if ((uint16_t)(rightdelim_cur - startcur) != rightdelim_cur - startcur) return false;
  *destptr = startcur;
  *destlen = rightdelim_cur - startcur;
  s->cur = cur;  // skip the rightdelim, point to the newline
  return true;
}

static bool parse_remaining(struct scanner *s, uint32_t *destptr, uint16_t *destlen) {
  const char *const buf = s->buf;
  const size_t len = s->len;
  size_t cur = s->cur;

  if (cur >= len) return false;

  const size_t startcur = cur;
  const char *p = memchr(buf + cur, '\n', len - cur);
  if (p == NULL) return false;
  cur = p - buf;

  if ((uint16_t)(cur - startcur) != cur - startcur) return false;
  *destptr = startcur;
  *destlen = cur - startcur;
  s->cur = cur;
  return true;
}

// Assumes the "Joins: " etc. prefix has already been consumed
static bool parse_useract(struct scanner *s, struct event *dest, bool withtail) {
  if (!parse_nick(s, &dest->ptr1, &dest->len1)) return false;
  DEBUG(" parse_useract nick\n");
  if (!EAT_BYTES(s, " ")) return false;
  DEBUG(" parse_useract space\n");
  if (!parse_inside_parens(s, &dest->ptr2, &dest->len2)) return false;
  DEBUG(" parse_useract parens\n");
  if (withtail) {
    if (!EAT_BYTES(s, " ")) return false;
    DEBUG(" parse_useract space2\n");
    if (!parse_enclose_tail(s, &dest->ptr3, &dest->len3, '(', ')')) return false;
    DEBUG(" parse_useract tail\n");
  }
  return true;
}

static bool parse_eventdata(struct scanner *s, struct event *dest) {
  DEBUG(" parse_eventdata cur=%zu\n", s->cur);
  if (EAT_BYTES(s, "*** ")) {
    if (EAT_BYTES(s, "Joins: ")) {
      dest->kind = JOIN;
      if (!parse_useract(s, dest, false)) return false;
      // znc prints nickserv account name after parentheses since znc 1.9.0; skip it
      skip_until(s, '\n');
      return true;
    }
    if (EAT_BYTES(s, "Parts: ")) {
      dest->kind = PART;
      return parse_useract(s, dest, true);
    }
    if (EAT_BYTES(s, "Quits: ")) {
      dest->kind = QUIT;
      return parse_useract(s, dest, true);
    }

    if (!parse_nick(s, &dest->ptr1, &dest->len1)) return false;
    if (EAT_BYTES(s, " is now known as ")) {
      dest->kind = RENICK;
      return parse_nick(s, &dest->ptr2, &dest->len2);
    }
    if (EAT_BYTES(s, " sets mode: ")) {
      dest->kind = MODE;
      return parse_remaining(s, &dest->ptr2, &dest->len2);
    }
    if (EAT_BYTES(s, " was kicked by ")) {
      dest->kind = KICK;
      if (!parse_nick(s, &dest->ptr2, &dest->len2)) return false;
      if (!EAT_BYTES(s, " ")) return false;
      return parse_enclose_tail(s, &dest->ptr3, &dest->len3, '(', ')');
    }
    if (EAT_BYTES(s, " changes topic to ")) {
      dest->kind = TOPIC;
      return parse_enclose_tail(s, &dest->ptr2, &dest->len2, '\'', '\'');
    }

    return false;
  }

  if (EAT_BYTES(s, "<")) {
    // DEBUG(" Ate '<'\n");
    dest->kind = TALK;
    if (!parse_nick(s, &dest->ptr1, &dest->len1)) return false;
    // DEBUG(" Nick %u %hu\n", dest->ptr1, dest->len1);
    if (!EAT_BYTES(s, "> ")) return false;
    // DEBUG(" Ate '> '\n");
    return parse_remaining(s, &dest->ptr2, &dest->len2);
  }

  if (EAT_BYTES(s, "-")) {
    dest->kind = NOTICE;
    uint32_t ptr1;
    uint16_t len1;
    if (!parse_nick(s, &ptr1, &len1)) return false;

    if (len1 > 0 && s->buf[ptr1 + len1 - 1] == '-') len1--;
    if (len1 == 0) return false;
    dest->ptr1 = ptr1;
    dest->len1 = len1;

    if (!EAT_BYTES(s, " ")) return false;
    return parse_remaining(s, &dest->ptr2, &dest->len2);
  }

  if (EAT_BYTES(s, "* ")) {
    dest->kind = ACT;
    if (!parse_nick(s, &dest->ptr1, &dest->len1)) return false;
    if (!EAT_BYTES(s, " ")) return false;
    return parse_remaining(s, &dest->ptr2, &dest->len2);
  }

  return false;
}

static size_t tirclogv_parse_znc_loop(
    uint8_t *events_, size_t max_events, const char *buf, size_t len
) {
  // offsets are stored in a uint32_t, so the buffer can't be too large
  if ((uint32_t)len != len) return 0;

  assert(sizeof(struct event) == 22);
  struct event *events = (struct event*)events_;
  struct event dummyevent;

  struct scanner s = (struct scanner){.buf = buf, .len = len, .cur = 0};
  size_t numev = 0;
  while (s.cur < s.len && numev < max_events) {
    DEBUG("cur=%zu len=%zu\n", s.cur, s.len);
    if (!parse_hms(&s, events ? &events[numev] : &dummyevent)) goto parseerror;
    DEBUG(" hms\n");
    if (!parse_eventdata(&s, events ? &events[numev] : &dummyevent)) goto parseerror;
    DEBUG(" evdata cur=%zu len=%zu buf[cur]=%u\n", s.cur, s.len, (unsigned)s.buf[s.cur]);
    if (s.cur != s.len && s.buf[s.cur] != '\n') goto parseerror;
    s.cur++;  // skip the newline
    numev++;
    continue;

  parseerror:
    assert(PARSEERROR == 0);
    if (events) memset(&events[numev], 0, sizeof(struct event));
    numev++;
    const char *p = memchr(buf + s.cur, '\n', len - s.cur);
    if (p == NULL) break;
    s.cur += (p + 1) - (buf + s.cur);
  }

  DEBUG("ret numev=%zu\n", numev);
  return numev;
}

// Assumes there is space for evs_allocated * sizeof(struct event) items in events.
size_t tirclogv_parse_znc(uint8_t *events, size_t evs_allocated, const char *buf, size_t len) {
  return tirclogv_parse_znc_loop(events, evs_allocated, buf, len);
}

size_t tirclogv_count_lines(const char *buf, size_t len) {
  size_t numln = 0;

  if (len == 0) return 0;
  if (buf[len - 1] == '\n') len--;  // we actually count the number of '\n' below

  const int vecwidth = 16;
  const int unrollcount = 2;

  while (len >= vecwidth * unrollcount) {
    const int iter_size = vecwidth * unrollcount;
    size_t niters = len / iter_size;
    if (niters > 256) niters = 256;  // make sure we're not overflowing the bytes

    assert(unrollcount == 2);
    __m128i acc0, acc1;
    acc0 = acc1 = _mm_set1_epi8(0);
    for (size_t i = 0; i < niters; i++) {
      const __m128i vec0 = _mm_loadu_si128((const __m128i*)(buf + vecwidth * (unrollcount * i + 0)));
      const __m128i vec1 = _mm_loadu_si128((const __m128i*)(buf + vecwidth * (unrollcount * i + 1)));
      // since cmpeq returns 0xff/0x00 for true/false, "+= (eq == 0xff)" == "-= eq"
      // (GCC is able to see that "+= (cmpeq(...) == 0xff)" can be optimised to
      // "-= cmpeq(...)", but then proceeds to mess things up and generates
      // bad, quirky code anyway, so we do the optimisation manually)
      acc0 = _mm_sub_epi8(acc0, _mm_cmpeq_epi8(vec0, _mm_set1_epi8('\n')));
      acc1 = _mm_sub_epi8(acc1, _mm_cmpeq_epi8(vec1, _mm_set1_epi8('\n')));
    }
    acc0 = _mm_add_epi8(acc0, acc1);

    uint8_t bytes[vecwidth];
    memcpy(bytes, &acc0, vecwidth);
#pragma GCC novector  // the generated vector code is rather dumb (lots of punpck{l,h})
    for (int i = 0; i < vecwidth; i++) numln += bytes[i];

    buf += niters * iter_size;
    len -= niters * iter_size;
  }

#pragma GCC novector
  for (size_t i = 0; i < len; i++)
    numln += buf[i] == '\n';

  return numln;
}