On 2025-03-10 21:55, Aurelien Jarno wrote: > Hi Bill, > > On 2025-03-10 18:48, Bill Allombert wrote: > > On Mon, Mar 10, 2025 at 12:24:43PM +0100, Bill Allombert wrote: > > > On Fri, Mar 07, 2025 at 11:45:10AM +0100, Bill Allombert wrote: > > > > I am still unsure wether there is a race condition in PARI or not, > > > > however if > > > > there is, it is clear that 2.41 makes the issue much worse. > > > > > > Hello Aurélien, > > > > > > I have made some progress. I have made a test program that does not use > > > PARI/GP. > > > You need to compile with -O0 to hit the race condition. > > > > > > gcc pthread1.c -Wall -O0 -g -pthread -o pthread1 > > > > I join a simplified version. > > > > gcc pthread3.c -Wall -O0 -g -pthread -o pthread3 > > > > for i in `seq 1 100`; do echo -n "$i " && sid ./pthread3 || break; done > > > > This crashes or hangs with 2.41 while it works fine with 2.36. > > I have just tested that before and with upstream commit 89b53077d2a5, > and I confirm that your testcase passes without that commit and fails > with it. Great job!
I also noticed it failed the same way on arm64. > > Please tell me if you would like me to do more experiments. > > If you have time, yes, that's very useful. I'll also try to have a > closer look at your testcase, but it already seems relatively small. I am not sure I understand everything yet, but it seems that building pari with -fexceptions is a good workaround for the issue. Aurelien -- Aurelien Jarno GPG: 4096R/1DDD8C9B aurel...@aurel32.net http://aurel32.net
