Update libgit2 submodule.

[nossralf]

This updates the libgit2 submodule to latest master.

As several public struct members as well as some functions' parameters/return values have changed fundamental types (e.g from git_off_t which is an int64_t to git_object_size_t which is an uint64_t) I bumped the minor versions of both libgit2-sys and git2-rs (meaning it's equivalent to a major version bump for 0.x versions). As I'm not familiar with the version increment policy, let me know if that's wrong.

[nossralf]

I guess I'll have to get a VM with Windows up and running tomorrow to figure out why the Windows build fails.

[nossralf]

The change from enum to uint32_t for flags in some structs is what tripped me up. On Windows with msvc, where git_enum! uses i32, the raw enum values must be cast to u32 for rustc to be satisfied.

The libgit2 pull request which introduced this change mentions the ergonomic downsides for C++ that results from using an uint32_t.

As the usage of raw flag values is hidden behind methods in the Rust API, the numeric casts don't affect users of git2-rs unless they are using raw values directly.

[alexcrichton]

Everything looks great to me here thanks!

For the enum changes, could the type definitions for each type be switched to u32 all platforms too? That should remove the need ideally for the as cast when using the named constant values.

[nossralf]

I tried to do that (by using pub enum git_merge_flag_t: u32 in the git_enum! definition for all the affected enums) but it led to errors on Windows for the systest suite related to the enums.

[alexcrichton]

Sorry for the delayed response, but could you gist what the errors look like in that case?

[nossralf]

This is the output I get when I set the git_enum! type to u32 explicitly for these three enums.

RUNNING ALL TESTS
bad git_merge_flag_t signed: rust: 0 (0x0) != c 1 (0x1)
bad git_merge_file_flag_t signed: rust: 0 (0x0) != c 1 (0x1)
bad git_stash_apply_flags signed: rust: 0 (0x0) != c 1 (0x1)

The tests that fail all seem to be generated as such, with just the enum name changed:

#[inline(never)]
#[allow(non_snake_case)]
fn sign_git_stash_apply_flags() {
  extern {
    #[allow(non_snake_case)]
    fn __test_signed_git_stash_apply_flags() -> u32;
  }
  unsafe {
    same(((!(0 as git_stash_apply_flags)) < (0 as git_stash_apply_flags)) as u32,
      __test_signed_git_stash_apply_flags(), "git_stash_apply_flags signed");
  }
}

I don't understand what information the systests are generated from, but that a test that covers signedness fails when the type is set to unsigned at least makes some sense based on the name alone. (If it makes sense to someone who understands the tests more deeply is unclear to me.)

[alexcrichton]

Ah ok no worries! I think that the definitions may be a bit odd at the C layer, but that's ok. Let's go ahead with this and we can improve it later if necessary!