package GMP, get rid of gmp_wrapper.c

[StefanKarpinski]

BigInt and BigFloat should be provided as a GMP package rather than living in extras. Also, looking through gmp_wrapper.c, it seems like we can do all of this stuff without the wrapper entirely if we generated a libgmp dynamic library instead.

[ViralBShah]

I see that on OS X, there is a $JULIAHOME/usr/lib/libgmp.dylib - so it does seem that a dynamic library is installed for GMP. Looking through gmp_wrapper.c, it does look like we can get rid of it.

[aviks]

For the record, there are two things that need done for removing gmp_wrapper.c ... it may be doable

• GMP uses structs as arguments (mpz_t, mpf_t). Within julia, they are stored as void pointers. The wrapper de-rerences them before passing them on to GMP.
• the exported function names are mangled in the shared library, in a system dependent way.

[StefanKarpinski]

Ah! Yes, I knew there was a reason for this, but I failed to notice the dereferencing when I was looking through the file. What's up with the exported function name mangling? The gmp_wrapper.c file doesn't seem to have mangled names when calling – are those macros or something?

[Keno]

For reasons that are somewhat obscure, the deref in there isn't actually a deref (a pointer cast to void* would work just as well), so that's actually no problem (it has to do with the interaction of passing an array to a function that expects a pointer and the typedefs used by GMP). They work great for usual C usage in that they behave just as if they were usual structs, but they are actually not.

[StefanKarpinski]

The source is so convoluted I can't even figure out what the definition of gmp_t is. Can you expand a bit?

[Keno]

Sure, Jeff, Jameson and I had quite a bit of fun with this one Yesterday.

Essentially there is a

struct mzp_struct;

which is a BigInt. The implementation is not important. Then there are

typedef mpz_struct[1] mpz_t;
typedef mpz_struct* mpz_ptr;

Now, what we are doing is

mpz_t *integer = malloc(sizeof(mpz_t))
mpz_init(*integer)

(Note these are all somewhat simplyfied and don't show the actual definition, but should be enough to understand why).
Now sice

sizeof(type[N]) = N*sizeof(type)

we have

sizeof(mpz_t) = sizeof(mpz_struct[1]) = sizeof(mpz_struct)

So at least we are allocating the right amount of memory. But notice also that what we are returning is essentially a mpz_ptr (since it's a pointer to an area of memory the size of mpz_struct).

Now, why does this still work? Well, consider the prototype of mpz_init

void mpz_init(mpz_ptr integer);

From our above analysis, we see that this should clearly work:

mpz_t *integer = malloc(sizeof(mpz_t))
mpz_init((mpz_ptr)integer)

but here is the tricky part that involves passing an array to a function that expects a pointer (think of it as passing a char[] to a function that expect a char*). Do do so, the function essentially takes the address of whatever array we are passing to it and passes that as a pointer (but still gives the impression of passing the array by value). So essentially what it's doing is, if we have

mpz_init((mpz_ptr)integer)
mpz_struct[1] st;
mpz_ptr ptr = st; //note this already does the conversion I talked about
//These three are equivalent
mpz_init(st);
mpz_init(&(st[0])); //note: st[0]==*ptr
mpz_init(ptr)

So since mpz_t* already points to the memory we need, we need to deref it to essentially cancel the taking of address done by the function call. Hope that helps.

[StefanKarpinski]

Jeez. This feels like C++. Thanks for explaining.

[ViralBShah]

I guess we cannot really do away with the wrapper then. Close this?

[Keno]

Yeah, we can, the GMP functions all just take a pointer.

[aviks]

On the name mangling,

in deps/gmp-5.0.5/gmp.h, generated at compile time from gmp-h.in

#define mpz_set_str __gmpz_set_str

aviks ~/dev/julia $ nm usr/lib/libgmp.dylib  | grep mpz_set_str
0000000000015610 T ___gmpz_set_str

I am not sure how these differ from one system to another.

[Keno]

I don't think that varies from system to system at all actually.

[andreasnoack]

Closed by a42f263