Allow unsafe_code in __thread_local_inner.

[Ms2ger]

This fixes a build error when using thread_local!() in a deny(unsafe_code)
scope in Servo for Android.

[rust-highfive]

r? @nikomatsakis

(rust_highfive has picked a reviewer for you, use r? to override)

[eddyb]

This is one of the weird things I might have cleaned up in my post-const-fn branch - thread_local definitely took the longest out of all of libstd.

Would using ::std::thread_local::__impl::destroy_value::<$t> instead of __destroy work at all?
Is it perhaps an ABI mismatch? If it is, can the destroy_value definition be changed, instead?

Also, take great care in keeping consistency across the supported combinations (IMO they should go in separate files to avoid #[cfg] confusion).

[alexcrichton]

Would using ::std::thread_local::__impl::destroy_value::<$t> instead of __destroy work at all?

Yeah that worked locally for me. I think that I may have just been playing around with various strategies and never revisited this.

@Ms2ger could you update the PR with something like:

diff --git a/src/libstd/thread_local/mod.rs b/src/libstd/thread_local/mod.rs
index 0878029..15cc469 100644
--- a/src/libstd/thread_local/mod.rs
+++ b/src/libstd/thread_local/mod.rs
@@ -205,15 +205,14 @@ macro_rules! __thread_local_inner {

         #[cfg(any(not(any(target_os = "macos", target_os = "linux")), target_arch = "aarch64"))]
         const _INIT: ::std::thread_local::__impl::KeyInner<$t> = {
-            unsafe extern fn __destroy(ptr: *mut u8) {
-                ::std::thread_local::__impl::destroy_value::<$t>(ptr);
-            }
-
             ::std::thread_local::__impl::KeyInner {
                 inner: ::std::cell::UnsafeCell { value: $init },
                 os: ::std::thread_local::__impl::OsStaticKey {
                     inner: ::std::thread_local::__impl::OS_INIT_INNER,
-                    dtor: ::std::option::Option::Some(__destroy as unsafe extern fn(*mut u8)),
+                    dtor: ::std::option::Option::Some(
+                        ::std::thread_local::__impl::destroy_value::<$t>
+                            as unsafe extern fn(*mut u8)
+                    ),
                 },
             }
         };

[eddyb]

Nice, but do you still need the cast? I'm mainly asking because if you do, that looks like a bug/limitation in the "coerce to expected type" logic in typeck.

[alexcrichton]

Oh nice, it is no longer necessary as well!

[Ms2ger]

r? @alexcrichton

[alexcrichton]

@bors: r+ 29aca83

Thanks!

[bors]

⌛ Testing commit 29aca83 with merge 8692e1c...

[bors]

💔 Test failed - auto-win-64-opt

[Manishearth]

@bors: retry