Speed up eval/move-gen/SEE; remove EvalLookForDanger, overhaul _EvalKing

SPRT-confirmed improvement (1814 games, st 4.0): candidate nps=1498118
vs head=1378448, avg depth 14.30 vs 13.87, ELO=+24.2+/-16.0,
LOWER95=0.5118, LLR=+3.126 crossing the H1 bound.

Most of the diff is non-semantic speed work across eval.c, generate.c,
movesup.c, see.c, probe.c and friends. Two real semantic changes:
removed EvalLookForDanger and overhauled _EvalKing's king-safety scoring.

Co-Authored-By: Claude Sonnet 5 <[email protected]>

1 files changed, 122 insertions, 13 deletions

diff --git a/src/chess.h b/src/chess.h
index d637905..fc2ef85 100755
--- a/src/chess.h
+++ b/src/chess.h
@@ -778,7 +778,6 @@ typedef struct _POSITION
     BITBOARD bbTrappedEscapes[2][6];
     ULONG uNumTrapped[2];
     ULONG uArmyScaler[2];
-    ULONG uClosedScaler;
     SCORE iScore[2];
     SCORE iReducedMaterialDownScaler[2];
     ULONG uMinorsAtHome[2];
@@ -833,7 +832,7 @@ typedef struct _KEY_POINTER
 }
 KEY_POINTER;
 
-typedef union _GENERATOR_FLAGS 
+typedef union _GENERATOR_FLAGS
 {
     struct {
         UCHAR uMoveCount;
@@ -868,6 +867,31 @@ typedef struct _MOVE_STACK
     //
     ULONG uUnblockedKeyValue[MAX_PLY_PER_SEARCH];
     KEY_POINTER sUnblocked[MAX_PLY_PER_SEARCH][128];
+    // _FindUnblockedSquaresBB's output: one bit per square unblocked
+    // for the opposing king this ply (membership only -- the
+    // direction-back-to-king half of the old sUnblocked table is
+    // recomputed on demand by DIRECTION_BETWEEN_SQUARES, a table
+    // lookup in release builds, so it needs no storage of its own).
+    // Kept unconditional (not #ifdef UNBLOCKED_BITBOARD-gated) so
+    // testgenerate.c's harness can exercise _FindUnblockedSquaresBB
+    // and compare it against the mailbox table regardless of which
+    // toggle this build was compiled with.
+    BITBOARD bbUnblocked[MAX_PLY_PER_SEARCH];
+    // Which generation (uUnblockedKeyValue[uPly] value) bbUnblocked[uPly]
+    // was last actually built for -- WouldGiveCheck (generate.c) compares
+    // this against the live uUnblockedKeyValue[uPly] to decide whether
+    // the cached bitboard is still valid for the node now occupying this
+    // ply, or needs a lazy rebuild. GenerateMoves no longer builds
+    // bbUnblocked eagerly; it just bumps uUnblockedKeyValue[uPly], same
+    // as root.c's init comment there describes -- the actual build is
+    // deferred to WouldGiveCheck's first call this generation, so a node
+    // whose every candidate move gets rejected before ever needing
+    // WouldGiveCheck (e.g. every capture futility-pruned in QSearch)
+    // never pays for it at all. Starts zeroed; every real ticket in
+    // uUnblockedKeyValue starts at 1 or higher (root.c) and only
+    // increases, so 0 is a permanently safe "never built yet" sentinel
+    // -- no special seeding needed here, unlike that field.
+    ULONG uUnblockedBuiltFor[MAX_PLY_PER_SEARCH];
 
     //
     // The main move list, a long series of moves, their values and some
@@ -922,6 +946,20 @@ MOVE_STACK;
 #define GENERATE_CAPTURES_PROMS_CHECKS (3)
 #define GENERATE_CAPTURES_PROMS        (4)
 #define GENERATE_DONT_SCORE            (5)
+
+// generate.c's _GenerateCapturesAndPromotions (GENERATE_CAPTURES_PROMS's
+// direct, non-filtered generator) exists only when every piece type is
+// on its bitboard generator -- see that function's block comment in
+// generate.c. Declared here, not in generate.c, so every translation
+// unit that needs to know whether it exists (main.c's test-registration
+// list, testgenerate.c's own tests) sees the same answer generate.c's
+// GenerateMoves dispatch does.
+#if defined(GENERATE_KNIGHT_BITBOARD) && defined(GENERATE_BISHOP_BITBOARD) && \
+    defined(GENERATE_ROOK_BITBOARD) && defined(GENERATE_QUEEN_BITBOARD) && \
+    defined(GENERATE_KING_BITBOARD) && defined(GENERATE_PAWN_BITBOARD)
+#define CAPTURES_PROMS_BITBOARD_PATH 1
+#endif
+
 #ifdef TEST
 #define GENERATE_ALL_MOVES_CHECK_OK    (6)
 #endif
@@ -1138,6 +1176,18 @@ typedef struct _COUNTERS
     }
     kingsafetyhash;
 
+    // Separate from kingsafetyhash above -- _GetKingActionBonus's
+    // cache (eval.c) is keyed and populated completely independently
+    // (see KING_SAFETY_HASH_ENTRY's u64ActionKey comment), so its own
+    // hit rate needs its own counters rather than conflating with
+    // _GetFileStormDefects'.
+    struct
+    {
+        UINT64 u64Probes;
+        UINT64 u64Hits;
+    }
+    kingactionhash;
+
     struct
     {
         UINT64 u64TotalNodeCount;
@@ -1746,6 +1796,21 @@ typedef struct _PAWN_HASH_ENTRY
     UCHAR uCountPerFile[2][10];
     UCHAR uNumRammedPawns;
     UCHAR uNumUnmovedPawns[2];
+    // Total pawns on the board, both colors -- _EvalRook indexes
+    // ROOK_VALUE_AS_PAWNS_COME_OFF with this once per rook (up to 4x
+    // per Eval() call); also feeds uClosedScaler's own formula below,
+    // in both cases replacing a fresh FastCountBits(bbPawns[WHITE] |
+    // bbPawns[BLACK]) with a value this entry already has cached.
+    UCHAR uTotalPawnCount;
+    // "How closed/busy is the board" scaler _EvalKnight indexes
+    // KNIGHT_IN_CLOSED_POSITION with (eval.c's EvalPawns computes the
+    // formula) -- used to live as a POSITION field recomputed from
+    // scratch on every Eval() call even though every input (total
+    // pawn count, uNumUnmovedPawns, uNumRammedPawns) is already pure
+    // pawn-structure data this same entry is keyed on. Same "finish
+    // moving a pawn-only term into the pawn hash" pattern as
+    // iBishopColorScore below and 5fbd935's original move.
+    UCHAR uClosedScaler;
     SHORT iBishopColorScore[2][2];  // [color][IS_SQUARE_WHITE(bishop's
                                      // square)] -- good/bad-bishop
                                      // pawn-structure score, pure
@@ -1771,6 +1836,17 @@ typedef struct _KING_SAFETY_HASH_ENTRY
     // reports old counter units into KING_DEFECT_DUMP_INFO, unrelated
     // to this cache.
     SCORE iFileStormDefects[2];
+    // Independent validity tag for iActionBonus below, decoupled from
+    // u64Key above on purpose: _GetKingActionBonus (eval.c) is called
+    // from the opposite half of _EvalKing's safety/activity split from
+    // _GetFileStormDefects (only when the ENEMY lacks mating material,
+    // where file-storm's own populate-on-miss never runs), so a hit on
+    // u64Key would say nothing about whether iActionBonus was ever
+    // actually computed for this key. Per-king (keyed on pawn sig +
+    // just THIS king's own square, not both kings'), unlike
+    // iFileStormDefects -- see _GetKingActionBonus's header comment.
+    UINT64 u64ActionKey;
+    SCORE iActionBonus;
 }
 KING_SAFETY_HASH_ENTRY;
 
@@ -2799,7 +2875,7 @@ InCheck(POSITION *pos, ULONG uSide);
 FLAG
 IsAttackedBB(COOR cTest, POSITION *pos, ULONG uSide);
 
-FLAG
+FLAG FORCEINLINE
 InCheckBB(POSITION *pos, ULONG uSide);
 
 // ISATTACKED_BITBOARD defined -> the BB versions above resolve plain
@@ -2828,10 +2904,10 @@ StripMove(CHAR *szMove);
 ULONG
 LooksLikeMove(CHAR *szData);
 
-void FASTCALL
+void FORCEINLINE
 SelectBestWithHistory(SEARCHER_THREAD_CONTEXT *ctx, ULONG u);
 
-void FASTCALL
+void FORCEINLINE
 SelectBestNoHistory(SEARCHER_THREAD_CONTEXT *ctx, ULONG u);
 
 void FASTCALL
@@ -2942,6 +3018,24 @@ _RookAttacksBB(COOR c, BITBOARD bbOccupied);
 BITBOARD
 _BishopAttacksBB(COOR c, BITBOARD bbOccupied);
 
+// generate.c's bitboard replacement for _FindUnblockedSquares, exposed
+// non-static so testgenerate.c's correctness/speed harness can call it
+// directly regardless of whether UNBLOCKED_BITBOARD is defined for
+// this build.
+void
+_FindUnblockedSquaresBB(MOVE_STACK *pStack,
+                        POSITION *pos);
+
+// The mailbox original, exposed non-static (compiled whenever
+// !UNBLOCKED_BITBOARD, or under DEBUG/TEST for the cross-check and
+// the test harness above) under the same build conditions generate.c
+// itself uses -- see generate.c.
+#if !defined(UNBLOCKED_BITBOARD) || defined(DEBUG) || defined(TEST)
+void
+_FindUnblockedSquares(MOVE_STACK *pStack,
+                      POSITION *pos);
+#endif
+
 void
 _GenerateRookBB(MOVE_STACK *pStack,
                 POSITION *pos,
@@ -3057,7 +3151,7 @@ GenerateWhiteKing(MOVE_STACK *pStack,
                   POSITION *pos,
                   COOR cKing);
 
-FLAG
+FLAG FORCEINLINE
 WouldGiveCheck(IN SEARCHER_THREAD_CONTEXT *ctx,
                IN MOVE mv);
 
@@ -3095,6 +3189,9 @@ TestGenerateQueenSpeed(void);
 void
 TestGenerateAllMovesSpeed(void);
 
+void
+TestFindUnblockedSquaresSpeed(void);
+
 #endif
 
 //
@@ -3148,6 +3245,7 @@ extern MOVE HASHMOVE;
 extern MOVE RECOGNMOVE;
 extern MOVE DRAWMOVE;
 extern MOVE MATEMOVE;
+extern MOVE EGTBMOVE;
 extern FLAG g_fIsWhiteSquare[128];
 extern BITBOARD BBFILE[8];
 extern BITBOARD BBRANK[9];
@@ -3707,7 +3805,7 @@ CanWeReduceOrPruneMove(IN SEARCHER_THREAD_CONTEXT *ctx,
                     IN FLAG fThreat,
                     IN FLAG fIsLeftoverMove);
 
-INT
+INT FORCEINLINE
 GetLMRReduction(IN SEARCHER_THREAD_CONTEXT *ctx,
                 IN ULONG uRemainingDepth,
                 IN ULONG uLegalMoves,
@@ -3789,7 +3887,7 @@ GetMoveSee(IN SEARCHER_THREAD_CONTEXT *ctx,
             IN MOVE mv,
             IN ULONG uMoveNum);
 
-void
+void FORCEINLINE
 ComputeMoveExtension(IN OUT SEARCHER_THREAD_CONTEXT *ctx,
                      IN SCORE iAlpha,
                      IN SCORE iBeta,
@@ -3813,7 +3911,7 @@ MateDistancePruningCutoff(IN ULONG uPly,
                           IN OUT SCORE *piAlpha,
                           IN OUT SCORE *piBeta);
 
-FLAG
+FLAG FORCEINLINE
 CommonSearchInit(IN SEARCHER_THREAD_CONTEXT *ctx,
                  IN OUT SCORE *piAlpha,
                  IN OUT SCORE *piBeta,
@@ -3849,14 +3947,25 @@ TestQSearchSanity(void);
 //
 #define SEE_HEAPS
 
-SCORE
+SCORE FORCEINLINE
 SEE(POSITION *pos,
     MOVE mv);
 
+// The real exchange-walk SEE implementation, with no early exit --
+// see SEE()'s own header comment in see.c for the cheap no-defender
+// check that now sits in front of this. Exposed non-static (unlike
+// its historical merged-into-SEE status) so testsee.c's
+// TestSEENoDefenderFastPath can call it directly by name for an
+// exact-value correctness/speed comparison against SEE(), same
+// convention as _GenerateAllMoves/_GenerateAllMovesBB.
+SCORE
+_SEEFull(POSITION *pos,
+         MOVE mv);
+
 // Sign-only fast path for SEE(pos, mv) >= 0 -- see its own header
 // comment in see.c. Exposed non-static so GetMoveSee (searchsup.c) can
 // call it and testsee.c can diff it against real SEE().
-FLAG
+FLAG FORCEINLINE
 SeeFastSign(POSITION *pos,
             MOVE mv);
 
@@ -4449,7 +4558,7 @@ _GetAttacksBB(SEE_LIST *pList,
 // call it with a king-vacated occupancy bitboard. Deliberately
 // excludes pawns -- see its own header comment in see.c; callers
 // needing pawn attacks must check g_PawnAttackOriginBB separately.
-BITBOARD
+BITBOARD FORCEINLINE
 _WhoAttacksSquareBB(POSITION *pos,
                      COOR cSquare,
                      ULONG uSide,
@@ -4807,7 +4916,7 @@ GenerateRandomLegalSymetricPosition(POSITION *pos);
 void
 InitializeInteriorNodeRecognizers(void);
 
-ULONG
+ULONG FORCEINLINE
 RecognLookup(SEARCHER_THREAD_CONTEXT *ctx,
              SCORE *piScore,
              ULONG uEgtbProbeMode);