Filling drop blocks LLVM from optimizing away allocations.

[eddyb]

Testcase by @tsion (run on playpen):

#![crate_type = "lib"]
#![feature(filling_drop)]

extern "C" {
    fn __rust_allocate(_: usize, _: usize) -> &'static mut i8;
    fn __rust_deallocate(_: *mut i8, _: usize, _: usize);
}

pub fn foo(x: i32) -> i32 {
    unsafe {
        let b = __rust_allocate(4, 4) as *mut _ as *mut i32;
        *b = x;
        let r = *b;

        // Comment this check and watch the IR shrink.
        if b != std::mem::POST_DROP_USIZE as *mut i32 {
            __rust_deallocate(b as *mut i8, 4, 4);
        }

        r
    }
}

A complete implementation of #5016 should fix this, although I'm surprised there is a check at all here (run on playpen):

#![crate_type = "lib"]

pub fn foo(x: i32) -> i32 {
    let b = Box::new(x);
    *b
}

cc @alexcrichton @pnkfelix

[jonas-schievink]

Filling drop is no more

[eddyb]

My second example is now:

define i32 @_ZN8rust_out3foo17hcacd841310a2f504E(i32) unnamed_addr #0 personality i32 (i32, i32, i64, %"8.unwind::libunwind::_Unwind_Exception"*, %"8.unwind::libunwind::_Unwind_Context"*)* @rust_eh_personality {
entry-block:
  ret i32 %0
}