Two ideas:

1. Nullmove R=3 + depth/6 along with some ShouldWeDoNullmoveSearch
tweaking and calibrating.

2. Probe EGTB everywhere in the search as long as the table/position
is resident in memory.

1 files changed, 236 insertions, 0 deletions

diff --git a/src/fathom/tbprobe.c b/src/fathom/tbprobe.c
index 5eac014..9cce02d 100644
--- a/src/fathom/tbprobe.c
+++ b/src/fathom/tbprobe.c
@@ -1550,6 +1550,24 @@ static bool init_table(struct BaseEntry *be, const char *str, int type)
 
   be->data[type] = data;
 
+#ifndef _WIN32
+  if (type == WDL && be->num <= 5) {
+    // 3-5 man WDL files are small (empirically <=20MB each, <=0.4GB total
+    // across every 3-5 man signature -- see CLAUDE.md's EGTB residency
+    // section), so instead of inferring residency later from whether this
+    // header parse has run (a guess that's only reliable because these
+    // files are small enough to end up fully cached anyway), just wire
+    // the whole mapping into RAM now. Once ready[WDL] is true for one of
+    // these signatures, tb_wdl_is_resident()'s "yes" becomes a guarantee
+    // instead of a proxy -- no mincore() needed for this tier at all.
+    if (mlock(data, (size_t)be->mapping[type]) != 0) {
+      perror("mlock");
+      // Not fatal -- this signature's residency answer just reverts to
+      // being an inference (like the 6-man case) instead of a guarantee.
+    }
+  }
+#endif
+
   bool split = type != DTZ && (data[4] & 0x01);
   if (type == DTM)
     be->dtmLossOnly = data[4] & 0x04;
@@ -1900,6 +1918,224 @@ static int probe_table(const Pos *pos, int s, int *success, const int type)
   return v;
 }
 
+// Cheap, non-blocking check for whether the WDL table covering a given
+// material signature is already mapped into memory (i.e. init_table()'s
+// one-time mmap()+header-parse has already run for this piece
+// combination, via some earlier probe_table() call this process).
+//
+// This deliberately duplicates probe_table()'s key computation and
+// tbHash[] walk instead of calling it, so that callers can ask "would a
+// probe of this position be fast?" from deep in the search tree without
+// ever taking tbMutex or touching a byte of the underlying file -- it's
+// a handful of popcounts plus one atomic_bool load.
+//
+// Returns:
+//    1  the material signature has a WDL table and it's already resident
+//       (mapped + header parsed) -- a probe here should be near-free.
+//    0  the material signature has a WDL table but it hasn't been
+//       touched by any probe yet this run -- a probe here will pay the
+//       one-time mmap()+parse cost (and is still lazily-mapped, so may
+//       also fault in fresh pages from disk for the data itself).
+//   -1  no WDL table covers this material signature at all (either it's
+//       outside the largest cardinality Fathom found on disk, or it's
+//       KvK / otherwise not tabulated).
+//
+// NOTE: a return of 1 only guarantees the table's *header and index
+// metadata* are parsed, not that every compressed data page this
+// specific position's lookup will touch is already in the OS page
+// cache. For the small 3-5 man files that's essentially the same thing
+// in practice (the whole file ends up cached after the first touch);
+// for large 6-man files it's a weaker guarantee. See CLAUDE.md.
+int tb_wdl_is_resident(uint64_t white, uint64_t black, uint64_t queens,
+                        uint64_t rooks, uint64_t bishops, uint64_t knights,
+                        uint64_t pawns)
+{
+  uint64_t key = popcount(white & queens)  * PRIME_WHITE_QUEEN +
+                 popcount(white & rooks)   * PRIME_WHITE_ROOK +
+                 popcount(white & bishops) * PRIME_WHITE_BISHOP +
+                 popcount(white & knights) * PRIME_WHITE_KNIGHT +
+                 popcount(white & pawns)   * PRIME_WHITE_PAWN +
+                 popcount(black & queens)  * PRIME_BLACK_QUEEN +
+                 popcount(black & rooks)   * PRIME_BLACK_ROOK +
+                 popcount(black & bishops) * PRIME_BLACK_BISHOP +
+                 popcount(black & knights) * PRIME_BLACK_KNIGHT +
+                 popcount(black & pawns)   * PRIME_BLACK_PAWN;
+  if (key == 0ULL) return -1; // KvK: not a tabulated material signature
+
+  int hashIdx = (int)(key >> (64 - TB_HASHBITS));
+  while (tbHash[hashIdx].key && tbHash[hashIdx].key != key)
+    hashIdx = (hashIdx + 1) & ((1 << TB_HASHBITS) - 1);
+  if (!tbHash[hashIdx].ptr ||
+      atomic_load_explicit(&tbHash[hashIdx].error, memory_order_relaxed))
+    return -1;
+
+  struct BaseEntry *be = tbHash[hashIdx].ptr;
+  return atomic_load_explicit(&be->ready[WDL], memory_order_relaxed) ? 1 : 0;
+}
+
+#ifndef _WIN32
+// Page-residency predictor for the 6-man tier, where init_table() can't
+// just mlock() the whole file (some signatures run into the hundreds of
+// MB to low GB -- see CLAUDE.md). Every dereference below is gated by a
+// mincore() check first, so a call to this function never itself blocks
+// on disk I/O: it either confirms every page a real probe_table() lookup
+// would touch is already resident (and, having just read them, returns a
+// verified "yes"), or it stops at the first non-resident page and
+// returns "no" without reading past it.
+//
+// A small direct-mapped cache remembers pages already confirmed resident
+// so repeated probes into the same hot region of a table don't re-pay a
+// mincore() syscall every time. It deliberately only caches "yes"
+// answers -- a page that mincore() reports as not-yet-resident is always
+// re-checked for real next time, so a page that warms up between calls
+// is picked up on its next probe rather than being stuck on a stale "no".
+#define TB_RESIDENCY_PAGE_SIZE   4096u
+#define TB_RESIDENCY_CACHE_BITS  12
+#define TB_RESIDENCY_CACHE_SIZE  (1u << TB_RESIDENCY_CACHE_BITS)
+static atomic_uintptr_t g_ResidencyCache[TB_RESIDENCY_CACHE_SIZE];
+
+static size_t _ResidencyCacheSlot(uintptr_t pageAddr)
+{
+  uintptr_t h = (pageAddr >> 12) * 2654435761u;
+  return (size_t)(h & (TB_RESIDENCY_CACHE_SIZE - 1));
+}
+
+static int _PageIsResident(uintptr_t pageAddr)
+{
+  size_t slot = _ResidencyCacheSlot(pageAddr);
+  if (atomic_load_explicit(&g_ResidencyCache[slot], memory_order_relaxed) == pageAddr)
+    return 1;
+
+  unsigned char vec = 0;
+  if (mincore((void *)pageAddr, TB_RESIDENCY_PAGE_SIZE, (char *)&vec) != 0)
+    return 0; // shouldn't happen for a page inside a live mapping; be conservative
+
+  if (vec & 1) {
+    atomic_store_explicit(&g_ResidencyCache[slot], pageAddr, memory_order_relaxed);
+    return 1;
+  }
+  return 0;
+}
+
+static int _RangeIsResident(const void *addr, size_t len)
+{
+  uintptr_t start = (uintptr_t)addr & ~(uintptr_t)(TB_RESIDENCY_PAGE_SIZE - 1);
+  uintptr_t end = ((uintptr_t)addr + len + TB_RESIDENCY_PAGE_SIZE - 1) &
+                   ~(uintptr_t)(TB_RESIDENCY_PAGE_SIZE - 1);
+  for (uintptr_t p = start; p < end; p += TB_RESIDENCY_PAGE_SIZE)
+    if (!_PageIsResident(p))
+      return 0;
+  return 1;
+}
+
+// Precise, non-blocking "would a real WDL probe of this exact position
+// hit resident pages only?" check. Unlike tb_wdl_is_resident() (which
+// only knows whether a material signature's header has been parsed at
+// all), this computes the same indexTable/sizeTable/data-block addresses
+// decompress_pairs() would touch for this specific position and confirms
+// each one is actually in the OS page cache before trusting the next.
+//
+// RETURN: 1 if a subsequent real probe is expected to be fully resident
+// (no page fault / disk read), 0 otherwise (table not loaded, material
+// untabulated, or some page on the lookup path isn't resident yet).
+int tb_wdl_probe_is_resident(uint64_t white, uint64_t black, uint64_t kings,
+                              uint64_t queens, uint64_t rooks,
+                              uint64_t bishops, uint64_t knights,
+                              uint64_t pawns, bool turn)
+{
+  Pos pos = { white, black, kings, queens, rooks, bishops, knights, pawns,
+              0, 0, turn };
+
+  uint64_t key = calc_key(&pos, false);
+  if (key == 0ULL)
+    return 0;
+
+  int hashIdx = (int)(key >> (64 - TB_HASHBITS));
+  while (tbHash[hashIdx].key && tbHash[hashIdx].key != key)
+    hashIdx = (hashIdx + 1) & ((1 << TB_HASHBITS) - 1);
+  if (!tbHash[hashIdx].ptr ||
+      atomic_load_explicit(&tbHash[hashIdx].error, memory_order_relaxed))
+    return 0;
+
+  struct BaseEntry *be = tbHash[hashIdx].ptr;
+  if (!atomic_load_explicit(&be->ready[WDL], memory_order_relaxed))
+    return 0;
+
+  // 3-5 man WDL files are mlock()'d whole at load time (init_table()) --
+  // ready[WDL] already IS the guarantee for this tier, no page check
+  // needed.
+  if (be->num <= 5)
+    return 1;
+
+  bool bside, flip;
+  if (!be->symmetric) {
+    flip = key != be->key;
+    bside = (pos.turn == WHITE) == flip;
+  } else {
+    flip = pos.turn != WHITE;
+    bside = false;
+  }
+
+  struct EncInfo *ei = first_ei(be, WDL);
+  int p[TB_PIECES];
+  size_t idx;
+
+  if (!be->hasPawns) {
+    ei = &ei[bside];
+    for (int i = 0; i < be->num;)
+      i = fill_squares(&pos, ei->pieces, flip, 0, p, i);
+    idx = encode_piece(p, ei, be);
+  } else {
+    int i = fill_squares(&pos, ei->pieces, flip, flip ? 0x38 : 0, p, 0);
+    int t = leading_pawn(p, be, FILE_ENC);
+    ei = &ei[t + 4 * bside];
+    while (i < be->num)
+      i = fill_squares(&pos, ei->pieces, flip, flip ? 0x38 : 0, p, i);
+    idx = encode_pawn_f(p, ei, be);
+  }
+
+  struct PairsData *d = ei->precomp;
+  if (!d->idxBits)
+    return 1; // constant-value table -- no data page to check at all
+
+  uint32_t mainIdx = (uint32_t)(idx >> d->idxBits);
+  uint8_t *idxEntry = d->indexTable + 6 * (size_t)mainIdx;
+  if (!_RangeIsResident(idxEntry, 6))
+    return 0;
+
+  uint32_t block;
+  memcpy(&block, idxEntry, sizeof(block));
+  block = from_le_u32(block);
+  uint16_t idxOffset = *(uint16_t *)(idxEntry + 4);
+
+  int litIdx = (int)((idx & (((size_t)1 << d->idxBits) - 1)) -
+                      ((size_t)1 << (d->idxBits - 1)));
+  litIdx += from_le_u16(idxOffset);
+
+  if (litIdx < 0) {
+    while (litIdx < 0) {
+      uint32_t prev = block - 1;
+      if (!_RangeIsResident(&d->sizeTable[prev], sizeof(uint16_t)))
+        return 0;
+      block = prev;
+      litIdx += d->sizeTable[block] + 1;
+    }
+  } else {
+    if (!_RangeIsResident(&d->sizeTable[block], sizeof(uint16_t)))
+      return 0;
+    while (litIdx > d->sizeTable[block]) {
+      litIdx -= d->sizeTable[block] + 1;
+      block++;
+      if (!_RangeIsResident(&d->sizeTable[block], sizeof(uint16_t)))
+        return 0;
+    }
+  }
+
+  uint8_t *dataBlock = d->data + ((size_t)block << d->blockSize);
+  return _RangeIsResident(dataBlock, sizeof(uint32_t) * 2);
+}
+#endif // !_WIN32
+
 static int probe_wdl_table(const Pos *pos, int *success)
 {
   return probe_table(pos, 0, success, WDL);