diff options
| author | Scott Gasch <[email protected]> | 2026-09-16 08:02:21 -0700 |
|---|---|---|
| committer | Scott Gasch <[email protected]> | 2026-09-16 08:02:21 -0700 |
| commit | 966a0ef4da9cf3911eb3da0533179199808717b4 (patch) | |
| tree | e09363b574215ef0d8996155a5ab24314db2dd69 /src/GNUmakefile | |
| parent | b91c8453f3fe21c8ccd3568923d8e39141c9c945 (diff) | |
Retire cPawns/cNonPawns piece-location lists in favor of bbPawns/bbPieces/cKingSq, and hoist two redundant TotalNonPawnCount POPCNT calls out of per-node hot loops.
POSITION no longer stores cPawns[2][8], cNonPawns[2][16], uPawnCount[2],
uNonPawnCount[2][8], or uWhiteSqBishopCount[2] -- every reader now goes
through PawnMaterial/TotalNonPawnCount/WhiteSqBishopCount/KingPosition
(chess.h), computed from the already-incrementally-maintained bbPawns/
bbPieces/cKingSq bitboards via a single -mpopcnt POPCNT instead of a
stored counter. move.c's make/unmake loses the list-swap-on-removal
bookkeeping entirely. GNUmakefile now passes -mpopcnt so the popcount
builtins actually compile to one instruction instead of clang's ~15-op
SWAR fallback.
Measured a small but real nps regression from the naive swap (search.c's
QSearch move-filter and eval.c's _EvalKing were each recomputing the same
TotalNonPawnCount sum on every candidate move / twice per node); hoisted
both to a single call per node instead. Net result across a same-code
self-play match: ~1541.5 knps candidate vs ~1505.7 knps head_reference.
A third hoist (FindEnprisePiece out of SelectBestWithHistory) was tried
and reverted -- its cache isn't actually loop-invariant within a node
(populated from inside Eval(), which doesn't run before every selection
call), caught by a DEBUG-build ASSERT before it landed.
Co-Authored-By: Claude Sonnet 5 <[email protected]>
Diffstat (limited to 'src/GNUmakefile')
| -rw-r--r-- | src/GNUmakefile | 8 |
1 files changed, 7 insertions, 1 deletions
diff --git a/src/GNUmakefile b/src/GNUmakefile index 3613e40..e6d24d8 100644 --- a/src/GNUmakefile +++ b/src/GNUmakefile @@ -97,7 +97,13 @@ ifdef USE_READLINE endif ifdef SIXTYFOUR - PROFILE += -m64 + # -mpopcnt: every x86-64 chip built since ~2008 (and this box's Ryzen + # 9 3900X) has the POPCNT instruction, but clang won't emit it for + # __builtin_popcountll without this -- without it, the builtin lowers + # to a ~15-instruction SWAR bit-trick sequence instead (still faster + # than the old CDECL-call CountBits, but not the ~1-3 cycle single + # instruction the "fast" eval.c wrappers were assumed to get). + PROFILE += -m64 -mpopcnt NASMFLAGS += -d_X64_ else PROFILE += -m32 |