unsafe::reinterpret_cast can cause valgrind errors

[jdm]

Sorry for the vague summary, but I've reduced this testcase about as far as I can.

import std::{str, option, unsafe};
use std;

type metadata<T> = {node: int, data: @T};
type compile_unit_md = {path: str};
type metadata_cache = [@debug_metadata];

tag debug_metadata {
    compile_unit_metadata(@metadata<compile_unit_md>);
}

fn get_compile_unit_metadata(cache: metadata_cache, full_path: str)
    -> @metadata<compile_unit_md> unsafe {
    ret alt *cache[0] {
      compile_unit_metadata(md) { unsafe::reinterpret_cast(md) }
    };
}

fn main() {
    let cache = [@compile_unit_metadata(@{node: 1, data: @{path: "test/a"}})];
    get_compile_unit_metadata(cache, "test/a/b");
}

When running the above, I get output like:

godot:build jdm$ ~/src/rustbug
rustbug(24826,0xb0305000) malloc: *** error for object 0x11001a0: incorrect checksum for freed object - object was probably modified after being freed.
*** set a breakpoint in malloc_error_break to debug

and valgrind has lots of entries for glue_drop15:

==24813== Invalid read of size 4
==24813==    at 0x30A6: glue_drop15 (in /Users/jdm/src/rustbug)
==24813==    by 0x232A: main::_e820af82886e4814 (in /Users/jdm/src/rustbug)
==24813==    by 0x259A: _rust_main (in /Users/jdm/src/rustbug)
==24813==    by 0x34E717: task_start_wrapper (in /Users/jdm/src/rust/build/stage2/lib/rustc/i686-apple-darwin/lib/librustrt.dylib)
==24813==    by 0xDEADBEEE: ???
==24813==  Address 0x1c08d60 is 0 bytes inside a block of size 12 free'd
==24813==    at 0x1243A: free (vg_replace_malloc.c:430)
==24813==    by 0x351B79: upcall_free (in /Users/jdm/src/rust/build/stage2/lib/rustc/i686-apple-darwin/lib/librustrt.dylib)
==24813==    by 0x3165: glue_free16 (in /Users/jdm/src/rustbug)
==24813==    by 0x30E3: glue_drop15 (in /Users/jdm/src/rustbug)
==24813==    by 0x2201: main::_e820af82886e4814 (in /Users/jdm/src/rustbug)
==24813==    by 0x259A: _rust_main (in /Users/jdm/src/rustbug)
==24813==    by 0x34E717: task_start_wrapper (in /Users/jdm/src/rust/build/stage2/lib/rustc/i686-apple-darwin/lib/librustrt.dylib)
==24813==    by 0xDEADBEEE: ???

If I assign the result of get_compile_metadata to a local var valgrind still complains, but the malloc error doesn't seem to show up any more.

[jdm]

fn get_compile_unit_metadata(cache: metadata_cache, full_path: str)
    -> @metadata<compile_unit_md> unsafe {
    ret alt *cache[0] {
      compile_unit_metadata(md) { unsafe::reinterpret_cast(md) }
    };
}

Looks like the unsafe::reinterpret_cast here is the catalyst. This version of get_compile_unit_metadata displays the valgrind errors, and the one that just returns md without the cast does not.

[jdm]

import std::unsafe;
use std;

tag test {
    a(@int);
}

fn main() unsafe {
    let val = a(@1);
    log_err alt val {
      a(i) { unsafe::reinterpret_cast::<@int, @int>(i) }
    };
}

16:08 13:06 < brson> jdm: the problem is thaht when you reinterpret cast that box, now you have two typed pointers to a box, but you didn't ever bump the refcount, so you ending up dropping the ref twice
16:08 13:06 < brson> the solution here is that after you do the reinterpret cast, run unsafe::leak on the old pointer

[jdm]

Brian's suggestion resolved the valgrind messages.