patch applied (packages/integer-gmp): Patch GMP to always use the GHC allocation functions

[Simon Marlow]

On 19/09/2009 21:51, Duncan Coutts wrote:
> On Sat, 2009-09-19 at 22:08 +0200, Matthias Kilian wrote:
>> On Mon, Jun 22, 2009 at 01:35:28PM -0700, Ian Lynagh wrote:
>>> Sun Jun 21 11:29:57 PDT 2009  Ian Lynagh<igloo at earth.li>
>>>    * Patch GMP to always use the GHC allocation functions
>>>    This works around a crash (only on OS X for some reason?) where the GHCi
>>>    GMP uses the systems memory allocator. We should fix this properly, by
>>>    making ghci run constructor functions.
>>>
>>>      M ./cbits/alloc.c -6 +6
>>>      M ./gmp/ghc.mk -1 +1
>>>      A ./gmp/tarball/gmp-4.2.4-nodoc-patched.tar.bz2
>>>      R ./gmp/tarball/gmp-4.2.4-nodoc.tar.bz2
>>>      A ./gmp/tarball/patch
>>
>> Can't this been done using mp_set_memory_functions() instead of
>> patching memory.c?
>
> Yes, that's what the constructor does, but GHCi is not running the
> constructor, because its linker does not know about them.
>
> However, it still does not make any sense because GHCi itself is linked
> against GMP (because GHCi depends on base which depends on integer-gmp)
> so it already calls mp_set_memory_functions(), so it should work even if
> GHCi does not call the constructor when loading its copy of the
> integer-gmp. The only way I can imagine it going wrong is if somehow
> they are using two different instances of GMP, eg one statically linked
> and one dynamically.

I expect that's what happens, but I can't see *how* it happens.  Looking 
at configure.ac and integer-gmp.buildinfo.in, it appears that 
extra-libraries is empty when GMP is not found.  So in that case, how do 
we end up linking libgmp.a?  And how does it get loaded by GHCi?

Cheers,
	Simon