Define gc counted allocator functions with libc API

[andreasnoack]

...and use them in CHOLMOD wrappers. This makes the memory consumption from SuiteSparse visible to our GC. Right now SuiteSparse object looks tiny to the GC even though they are huge and in consequence much memory is left allocated when it could be freed.

In particular, I'd like comments on the gc.c part of this as I haven't touched that part of the code base before.

[vtjnash]

this needs to add 16 bytes, to keep p aligned

[carnaval]

does malloc has any alignment guarantee beyond sizeof(void*) ?

[vtjnash]

yes, on almost all modern platforms, it guarantees 16-byte alignment

edit: if size > 16

[carnaval]

we might as well have this be 16 aligned then. I would think that libraries actually needing such alignment would be defensive about it, but let's not be greedy for 8 bytes (aka the Jameson way :p )

[yuyichao]

Actually I just come across malloc_usable_size in GLibC, Can we use it on systems where it's available?

[andreasnoack]

Thanks for the comments. Here is a new attempt.

[vtjnash]

Actually I just come across malloc_usable_size in GLibC, Can we use it on systems where it's available?

it's not portable, although something similar is available on most systems

[ScottPJones]

There are aligned_alloc and posix_memalign, depending on platform, that can give you 16 byte alignment.

[yuyichao]

it's not portable, although something similar is available on most systems

Depend on what's the allocation pattern of cholmod I'm just wondering how much memory we'll save by not storing the same information (size) twice.

Edit: and whether it's worth writing a non-portable optimization for it.

[carnaval]

Well it's a non breaking change to go back and use platform dependent functions, so let's merge this first. Most C code that cares about perf wont use malloc for small things anyway.

[yuyichao]

Agree. (especially the non-breaking improvement part).

[andreasnoack]

Tests are passing now. Had to update the .travis.yml. @vtjnash would you take a look at the changes to jl_gc.c and see if they seem right? Everything should be 16 byte aligned now.

[vtjnash]

lgtm