Build all platforms
cc @alexcrichton this needs a more thorough review. It turns out libc failed to build on a lot of older Rust versions for various reasons and platforms, so it took a while to fix the build on all of them.
This PR fixes the build on all platforms and all Rust version down to the
minimum Rust version supported by libc: Rust 1.13.0.
The `build.rs` is extended with logic to detect the newer Rust features used by
`libc` since Rust 1.13.0:
* Rust 1.19.0: `untagged_unions`. APIs using untagged unions are gated on
`cfg(libc_unions)` and not available on older Rust versions.
* Rust 1.25.0: `repr(align)`. Because `repr(align)` cannot be parsed by older
Rust versions, all uses of `repr(align)` are split into `align.rs` and
`no_align.rs` modules, which are gated on the `cfg(libc_align)` at the top
level. These modules sometimes contain macros that are expanded at the top
level to avoid privacy issues (`pub(crate)` is not available in older Rust
versions). Closes#1242 .
* Rust : `const` `mem::size_of`. These uses are worked around with hardcoded
constants on older Rust versions.
Also, `repr(packed)` structs cannot automatically `derive()` some traits like
`Debug`. These have been moved into `s_no_extra_traits!` and the lint of missing
`Debug` implementations on public items is silenced for these. We can manually
implement the `extra_traits` for these in a follow up PR. This is tracked
in #1243. Also, `extra_traits` does not enable `align` manually anymore.
Since `f64::to_bits` is not available in older Rust versions, its usage
has been replaced with a `transmute` to an `u64` which is what that method
does under the hood.
Closes#1232 .
This commit adds a `ci/build.sh` script that checks that libc builds correctly
for some common configurations (`--no-default-features`, `default`,
`extra_traits`) on most targets supported by Rust since Rust 1.13.0 (the oldest
Rust version that libc supports).
The build matrix is refactored into two stages.
The first stage is called `tools-and-build-and-tier1` and it aims to discover
issues quickly by running the documentation and linter builds, as well as
checking that the library builds correctly on all targets in all supported
channels and "problematic" Rust versions; Rust versions adding major new
features like `repr(align)`, `union`, etc. This first stage also runs
libc-test for the tier-1 targets on linux and osx. These builds finish
quickly because no emulation is necessary.
The second stage is called `tier2` and it runs libc-test for all other targets
for which we are currently able to do so.
Closes#1229 .
Check style using rustfmt and reformat
A couple of recent PRs attempt to reformat the library using rustfmt and run into issues against libc's style checker because rustfmt default settings differ.
This PR uses the rustfmt-preview from nightly to check that the library is appropriately formatted, adding a style as similar to libc's style as possible, so that both style checkers work properly simultaneously.
Fix cmsg(3) bugs for musl and OSX
This PR fixes bugs in the cmsg(3) family of functions for Linux/musl and OSX, introduced by PR #1098 and PR #1212 . It also adds an integration test which hopefully will validate these functions on every platform.
Since these are defined in C as macros, they must be reimplemented in
libc as Rust functions. They're hard to get exactly right, and they
vary from platform to platform. The test builds custom C code that uses
the real macros, and compares its output to the Rust versions' output
for various inputs.
Skip the CMSG_NXTHDR test on sparc64 linux because it hits a Bus Error.
Issue #1239
Skip the entire cmsg test program on s390x because it dumps core
seemingly before the kernel finishes booting.
Issue #1240
RFC 2235 - Implement PartialEq,Eq,Hash,Debug for all types
First pass at implementing [RFC2235](https://github.com/rust-lang/rfcs/blob/master/text/2235-libc-struct-traits.md). I just started with `PartialEq`/`Eq` and tested locally on my x64 Linux system. Definitely going to run into other types for other platforms that need this treatment, but thought I'd start small.
Open question is whether this should also rely on the `align` feature, which should improve which types can be auto-derived versus manually implemented (as it removed a lot of odd-sized padding arrays). My first pass here doesn't do that, but I might test it out to see if it does simplify quite a few structs. I think it might also be nice to have as it makes it easier for contributors to add new structs.
Part of rust-lang/rust#57715
Re-enable android build jobs
This performs a `yes | ./sdk/tools/bin/sdkmanager --licenses` - the sole purpose of the `--licenses` flag is to allow the user to accept all licenses a priori, so that all other commands don't show any licenses.
For some reason, after all licenses have been accepted, the install step says that some licenses are still not accepted... so we need to accept those too... Why? I have no idea, and accepting licenses while installing still appears to make things hang a little, but it only does so for one non-accepted license, instead of for 5-6 like before, so now things do work.
