Auto merge of #84876 - alexcrichton:inline-thread-locals-cross-crate, r=Mark-Simulacrum

std: Attempt again to inline thread-local-init across crates

Issue #25088 has been part of `thread_local!` for quite some time now.
Historical attempts have been made to add `#[inline]` to `__getit`
in #43931, #50252, and #59720, but these attempts ended up not landing
at the time due to segfaults on Windows.

In the interim though with `const`-initialized thread locals AFAIK this
is the only remaining bug which is why you might want to use
`#[thread_local]` over `thread_local!`. As a result I figured it was
time to resubmit this and see how it fares on CI and if I can help
debugging any issues that crop up.

Closes #25088

library/std/src/thread/local.rs

@@ -162,6 +162,7 @@ macro_rules! thread_local {
 macro_rules! __thread_local_inner {
     // used to generate the `LocalKey` value for const-initialized thread locals
     (@key $t:ty, const $init:expr) => {{
+        #[cfg_attr(not(target_env = "msvc"), inline)] // see comments below
         unsafe fn __getit() -> $crate::option::Option<&'static $t> {
             const _REQUIRE_UNSTABLE: () = $crate::thread::require_unstable_const_init_thread_local();
 
@@ -260,6 +261,29 @@ macro_rules! __thread_local_inner {
             #[inline]
             fn __init() -> $t { $init }
 
+            // When reading this function you might ask "why is this inlined
+            // everywhere other than MSVC?", and that's a very reasonable
+            // question to ask. The short story is that it segfaults rustc if
+            // this function is inlined. The longer story is that MSVC looks to
+            // not support `extern` references to thread locals across DLL
+            // boundaries. This appears to at least not be supported in the ABI
+            // that LLVM implements.
+            //
+            // Because of this we never inline on MVSC, but we do inline on
+            // other platforms (where external references to thread locals
+            // across DLLs are supported). A better fix for this would be to
+            // inline this function on MSVC, but only for "statically linked"
+            // components. For example if two separately compiled rlibs end up
+            // getting linked into a DLL then it's fine to inline this function
+            // across that boundary. It's only not fine to inline this function
+            // across a DLL boundary. Unfortunately rustc doesn't currently have
+            // this sort of logic available in an attribute, and it's not clear
+            // that rustc is even equipped to answer this (it's more of a Cargo
+            // question kinda). This means that, unfortunately, MSVC gets the
+            // pessimistic path for now where it's never inlined.
+            //
+            // The issue of "should enable on MSVC sometimes" is #84933
+            #[cfg_attr(not(target_env = "msvc"), inline)]
             unsafe fn __getit() -> $crate::option::Option<&'static $t> {
                 #[cfg(all(target_arch = "wasm32", not(target_feature = "atomics")))]
                 static __KEY: $crate::thread::__StaticLocalKeyInner<$t> =

src/test/codegen/auxiliary/thread_local_aux.rs

@@ -0,0 +1,6 @@
+#![crate_type = "lib"]
+#![feature(thread_local_const_init)]
+
+use std::cell::Cell;
+
+thread_local!(pub static A: Cell<u64> = const { Cell::new(0) });

src/test/codegen/thread-local.rs

@@ -0,0 +1,50 @@
+// compile-flags: -O
+// aux-build:thread_local_aux.rs
+// ignore-windows FIXME(#84933)
+// ignore-wasm globals are used instead of thread locals
+// ignore-emscripten globals are used instead of thread locals
+// ignore-android does not use #[thread_local]
+
+#![crate_type = "lib"]
+#![feature(thread_local_const_init)]
+
+extern crate thread_local_aux as aux;
+
+use std::cell::Cell;
+
+thread_local!(static A: Cell<u32> = const { Cell::new(1) });
+
+// CHECK: [[TLS_AUX:@.+]] = external thread_local local_unnamed_addr global i64
+// CHECK: [[TLS:@.+]] = internal thread_local unnamed_addr global
+
+// CHECK-LABEL: @get
+#[no_mangle]
+fn get() -> u32 {
+    // CHECK: %0 = load i32, i32* bitcast ({{.*}} [[TLS]] to i32*)
+    // CHECK-NEXT: ret i32 %0
+    A.with(|a| a.get())
+}
+
+// CHECK-LABEL: @set
+#[no_mangle]
+fn set(v: u32) {
+    // CHECK: store i32 %0, i32* bitcast ({{.*}} [[TLS]] to i32*)
+    // CHECK-NEXT: ret void
+    A.with(|a| a.set(v))
+}
+
+// CHECK-LABEL: @get_aux
+#[no_mangle]
+fn get_aux() -> u64 {
+    // CHECK: %0 = load i64, i64* [[TLS_AUX]]
+    // CHECK-NEXT: ret i64 %0
+    aux::A.with(|a| a.get())
+}
+
+// CHECK-LABEL: @set_aux
+#[no_mangle]
+fn set_aux(v: u64) {
+    // CHECK: store i64 %0, i64* [[TLS_AUX]]
+    // CHECK-NEXT: ret void
+    aux::A.with(|a| a.set(v))
+}