Re: Random crashes with ABC enabled
- From: Mike Pall <mike-1403@xxxxxxxxxx>
- To: luajit@xxxxxxxxxxxxx
- Date: Thu, 27 Mar 2014 12:05:13 +0100
Denis Golovan wrote:
> jit.opt.start("maxmcode=10240", "maxtrace=3000", "maxside=400",
> "maxsnap=3000", "maxirconst=2000",
> "hotloop=5", "hotexit=20", "instunroll=25",
> "loopunroll=100", "callunroll=25")
That looks a bit excessive. Though it should still work.
It's hard to give advice without seeing the actual code in
question. But the above looks more like the result of a raging
battle against the JIT compiler, not like a reasonable choice of
parameters.
I wonder whether your problems couldn't be solved better by
working around these issues in your code. E.g. using branch-free
idioms, manually unrolling loops with a low iteration count or
on-the-fly code generation from templates.
> jit.opt.start("-abc") -- without this one if happens really often
Well, FFI accesses are not bounds-checked, anyway. So disabling
ABC elimination does nothing for those loops. The problem must be
somewhere else.
> Could you suggest something?
Just from your description alone it's not possible to suggest
anything you could do to analyze this further. I guess you really
need to distill it into a short, complete and isolated test case.
That said: out-of-bounds FFI write accesses and unanchored FFI
cdata objects are very common causes for intermittent crashes and
various strange symptoms. You should double-check the code for
that. If the problems cease when you use collectgarbage("stop")
then you most certainly have an unanchored FFI cdata object.
--Mike
Other related posts:
