alloca usage
Niels Möller
nisse at lysator.liu.se
Thu Nov 20 21:57:53 CET 2003
It seems that the default configuration of GMP, with WANT_TMP_ALLOCA,
allocates arbitrarily large temporaries on the stack. For example
mpn_tdiv_qr allocates temporary storage proportional to the size of
the input.
Fredrik Hübinette (Pike author) reports that his pike program that
calculates a million decimals of Pi crashes in gmp with a default
configuration of gmp-4.1.2, but works fine if gmp is reconfigured with
--enable-alloca=malloc-reentrant. The problem is made worse by linux
threads (and perhaps threads on other systems as well), which limits
stack space to only 4 MB. (See
http://fredrik.hubbe.net/hacker/viewtopic.php?t=26).
Perhaps it would be a good idea to define TMP_ALLOC as something like
#define TMP_ALLOC(size) \
((size) < PAGESIZE*2) ? alloca((size)) : heap_alloc((size), &tmp_marker)
Then for small inputs, we get only a single comparison of extra
overhead, while for large inputs, the allocation overhead should be
small compared to the cost of the actual computation.
Due to the way the "red zone" stack extension mechanism works, I don't
think it's reliable to allocate more than a few pages of data on the
stack at a time. If you allocate too many pages you may get a
segmentation violation of you access the pages in an unfortunate
order.
Regards,
/Niels
More information about the gmp-devel
mailing list