Fix TT write in MultiPV case.
authorAndyGrant <Andrew@grantnet.us>
Mon, 24 Feb 2020 22:32:17 +0000 (23:32 +0100)
committerJoost VandeVondele <Joost.VandeVondele@gmail.com>
Tue, 25 Feb 2020 20:10:10 +0000 (21:10 +0100)
fixes an error reported earlier as https://github.com/official-stockfish/Stockfish/issues/2404 by @AndyGrant.

MultiPV at root shouldn't write to the TT for later lines, as that is neither the eval nor the bestmove for that position.
Fixing this error doesn't matter for playing games (http://tests.stockfishchess.org/tests/view/5dcdbd810ebc590256324a11).

However, it can lead to wrong mate announcements as reported by @uriblass. In particular the following testcase gives
wrong results for the second search, prior to this patch:

```
setoption name MultiPV value 2
position fen 5R2/2kB2p1/p2bR3/8/3p1B2/8/PPP5/2K5 b - - 0 49
go depth 40
position fen 2B2R2/3r2p1/p1kbR3/8/3p1B2/8/PPP5/2K5 b - - 8 48
go depth 40
```

fixes https://github.com/official-stockfish/Stockfish/issues/2561
closes https://github.com/official-stockfish/Stockfish/pull/2562

Only affects MultiPV search.

Bench: 4697493

src/search.cpp

index 3f860ac5e5b9d073e93ab6f1f476a6a49e9c9eff..a32ff4b6b2b209744951d2345a3bf3553b19cb66 100644 (file)
@@ -1339,7 +1339,7 @@ moves_loop: // When in check, search starts from here
     if (PvNode)
         bestValue = std::min(bestValue, maxValue);
 
-    if (!excludedMove)
+    if (!excludedMove && !(rootNode && thisThread->pvIdx))
         tte->save(posKey, value_to_tt(bestValue, ss->ply), ttPv,
                   bestValue >= beta ? BOUND_LOWER :
                   PvNode && bestMove ? BOUND_EXACT : BOUND_UPPER,