Core dump using HiPE (possibly my fault ;-))

[Mikael Pettersson]

Mikael Pettersson writes:
 > Thierry Mallard writes:
 >  > Greetings,
 >  > 
 >  > Vlad Dumitrescu helped me optimize a quick bench about computing some
 >  > numbers. You'll find the corresponding source attached to this mail.
 >  > 
 >  > I got a core dump when trying to run this test using HiPE :
 >  > 
 >  > -=-=-=-
 >  > [shaman@REDACTED test1]$ erl
 >  > Erlang (BEAM) emulator version 5.2 [source] [hipe]
 >  > 
 >  > Eshell V5.2  (abort with ^G)
 >  > 1> c(test1, [native, {hipe, [o3]}]).
 >  > {ok,test1}
 >  > 2> test1:run(10000, 10000).
 >  > Generating list of 10000 floats ... <0.36.0>
 >  > done
 >  > original Test time : 107.411 seconds 
 >  > 3> no next heap size found: 197920092, offset 0
 >  >                                                Aborted (core dumped)
 >  > -=-=--
 >  > 
 >  > The core dump goes away when i change the min_heap_size to 100 000
 >  > (instead of 1 000 000), so i'm not sure whereas this is a bug or simply
 >  > a bad use of min_heap_size.. 
 > 
 > Thanks for the bug report. I'll look into this problem.
 > So far I've been able to reproduce it on x86 with both R9B-0 (whoa, that's
 > ancient :->) and our current code base. I'm testing SPARC currently.

Ok the bug also occurs on SPARC. I think I know what it is (a procedure
in the garbage collector which isn't aware of native code stacks; it
didn't need to be before last summer when we added more serious compilation
of floating-point code). I'll try to come up with a patch "soonish".

/Mikael