get llvm libcall optimizations working

[thestinger]

It doesn't like that our libc signatures don't quite match, and we also currently have to mark all of the wrappers as inline(never) to prevent fixed_stack_segment from being too aggressively bubbled up.

Specifically, I really want dead-store-elimination working on unique pointers like it does with runtimeless Rust:

#[no_std];
#[allow(ctypes, cstack)];

extern {
    fn malloc(size: uint) -> *mut u8;
    fn free(ptr: *mut u8);
    fn abort() -> !;
}

#[lang = "exchange_malloc"]
#[inline(always)]
unsafe fn exchange_malloc(size: uint) -> *mut u8 {
    let ptr = malloc(size);
    if ptr == 0 as *mut u8 {
        abort()
    }
    ptr
}

#[lang = "exchange_free"]
#[inline(always)]
unsafe fn exchange_free(ptr: *mut u8) {
    free(ptr)
}

#[start]
fn main(_: int, _: **u8) -> int {
    let mut _a = ~5;
    0
}

LLVM eliminates the allocation here.

[thestinger]

This is also going to prevent the upcoming escape analysis implementation in LLVM from working.

[thestinger]

This is fixed for malloc and free, at least.