Gaviota TB crash when probing from multiple threads

[petero2]

I recently discovered that there is a bug in the Gaviota tablebase probing code that can cause an engine using the probing code to crash if it probes from multiple threads. If you compile using a C11 capable compiler, the fix is simple:

Code: Select all

diff --git a/gtb-probe.c b/gtb-probe.c
index dbb612c..b4aab11 100755
--- a/gtb-probe.c
+++ b/gtb-probe.c
@@ -2033,7 +2033,7 @@ egtb_get_id (SQ_CONTENT *w, SQ_CONTENT *b, tbkey_t *id)
        char *t;
        bool_t found;
        tbkey_t i;
-       static tbkey_t cache_i = 0;
+       static _Thread_local tbkey_t cache_i = 0;
 
        assert (PAWN == 1 && KNIGHT == 2 && BISHOP == 3 && ROOK == 4 && QUEEN == 5 && KING == 6);
 

Maybe no engine still in development, other than Texel, is using Gaviota TBs though.

[mar]

thanks Peter. I believe Myrddin uses GTB too

[towforce]

Ideally, requests to the TB would be via a service architecture for several reasons - including this one. However, I do understand that this would be bad for performance, which is not acceptable in a chess engine.

[Daniel Shawul]

I recently had a similar bug in Scorpio where I had the random number seed as static in the game generation code

Code: Select all

thread_local std::mt19937 mtgen(std::random_device{}());

Not sure how much it effect it had though.

[JVMerlino]

thanks Peter. I believe Myrddin uses GTB too

It does, but since I use processes instead of threads, it's not an issue. But I've modified my code anyway just to be sure.

[tmokonen]

Apparently _Thread_local is deprecated in C23. thread_local is now a keyword, rather than a convenience macro.

https://en.cppreference.com/w/c/keyword/_Thread_local