Minor GMP-ECM bug

[jasonp]

In the course of plugging calls to GMP-ECM into msieve, I've noticed a minor bug in GMP 6.1.3; the block of debug code starting at stage2.c:761 calls ecm_findmatch, which assumes an ECM stage 2 computation. However, P+-1 both go through this code path, and I had to do some unnecessary digging when I got crashes trying P-1 with lots of verbosity.

Also: regarding all the threads with people compiling machine-specific versions of GMP, does everyone know that GMP 4.2.1 and onward has a --enable-fat configure option, that builds multiple machine-specific versions of all the low-level GMP calls and chooses the best version at runtime? It looks like you only need one binary to achieve high performance on all x86 machines...

Finally: Alex, does your inria address work? I got two bounces trying to contact you.

[akruppa]

Thanks for the bug report. This was already fixed in the SVN but there was no patch for 6.1.3 yet. I'm attaching it. Sorry for the wasted time digging through the code. The patch also fixes a bug when deleting treefiles if stage 2 was interrupted during building F from roots.

I'll send you my correct Loria email address via PM.

Alex

[jasonp]

Originally Posted by akruppa

Thanks for the bug report. This was already fixed in the SVN but there was no patch for 6.1.3 yet. I'm attaching it. Sorry for the wasted time digging through the code. The patch also fixes a bug when deleting treefiles if stage 2 was interrupted during building F from roots.

No problem with the time spent, I've been looking through papers and this was an excuse to dig into some very nice code.