Add ARM MUSL targets.

The targets are:
- `arm-unknown-linux-musleabi`
- `arm-unknown-linux-musleabihf`
- `armv7-unknown-linux-musleabihf`

These mirror the existing `gnueabi` targets.

All of these targets produce fully static binaries, similar to the
x86 MUSL targets.

For now these targets can only be used with `--rustbuild` builds, as
https://github.com/rust-lang/compiler-rt/pull/22 only made the
necessary compiler-rt changes in the CMake configs, not the plain
GNU Make configs.

I've tested these targets GCC 5.3.0 compiled again musl-1.1.12
(downloaded from http://musl.codu.org/). An example `./configure`
invocation is:

```
./configure \
    --enable-rustbuild
    --target=arm-unknown-linux-musleabi \
    --musl-root="$MUSL_ROOT"
```

where `MUSL_ROOT` points to the `arm-linux-musleabi` prefix.
Usually that path will be of the form
`/foobar/arm-linux-musleabi/arm-linux-musleabi`.

Usually the cross-compile toolchain will live under
`/foobar/arm-linux-musleabi/bin`. That path should either by added
to your `PATH` variable, or you should add a section to your
`config.toml` as follows:

```
[target.arm-unknown-linux-musleabi]
cc = "/foobar/arm-linux-musleabi/bin/arm-linux-musleabi-gcc"
cxx = "/foobar/arm-linux-musleabi/bin/arm-linux-musleabi-g++"
```

As a prerequisite you'll also have to put a cross-compiled static
`libunwind.a` library in `$MUSL_ROOT/lib`. This is similar to [how
the x86_64 MUSL targets are built]
(https://doc.rust-lang.org/book/advanced-linking.html).

2 files changed, 13 insertions, 2 deletions

diff --git a/src/liballoc_jemalloc/build.rs b/src/liballoc_jemalloc/build.rs
index d1b3583d256..dc1b8d6ea98 100644
--- a/src/liballoc_jemalloc/build.rs
+++ b/src/liballoc_jemalloc/build.rs
@@ -73,7 +73,16 @@ fn main() {
                    .replace("\\", "/"))
        .current_dir(&build_dir)
        .env("CC", compiler.path())
-       .env("EXTRA_CFLAGS", cflags)
+       .env("EXTRA_CFLAGS", cflags.clone())
+       // jemalloc generates Makefile deps using GCC's "-MM" flag. This means
+       // that GCC will run the preprocessor, and only the preprocessor, over
+       // jemalloc's source files. If we don't specify CPPFLAGS, then at least
+       // on ARM that step fails with a "Missing implementation for 32-bit
+       // atomic operations" error. This is because no "-march" flag will be
+       // passed to GCC, and then GCC won't define the
+       // "__GCC_HAVE_SYNC_COMPARE_AND_SWAP_4" macro that jemalloc needs to
+       // select an atomic operation implementation.
+       .env("CPPFLAGS", cflags.clone())
        .env("AR", &ar)
        .env("RANLIB", format!("{} s", ar.display()));
 
diff --git a/src/liballoc_jemalloc/lib.rs b/src/liballoc_jemalloc/lib.rs
index 347e97e6ffc..ccf3d978fe4 100644
--- a/src/liballoc_jemalloc/lib.rs
+++ b/src/liballoc_jemalloc/lib.rs
@@ -36,7 +36,9 @@ use libc::{c_int, c_void, size_t};
 #[cfg_attr(target_os = "android", link(name = "gcc"))]
 #[cfg_attr(all(not(windows),
                not(target_os = "android"),
-               not(target_env = "musl")),
+               not(target_env = "musl"),
+               not(target_env = "musleabi"),
+               not(target_env = "musleabihf")),
            link(name = "pthread"))]
 #[cfg(not(cargobuild))]
 extern "C" {}