Rollup merge of rust-lang#79039 - thomcc:weakly-relaxing, r=Amanieu

Tighten the bounds on atomic Ordering in std::sys::unix::weak::Weak

This moves reading this from multiple SeqCst reads to Relaxed read + Acquire fence if we are actually going to use the data.

Would love to avoid the Acquire fence, but doing so would need Ordering::Consume, which neither Rust, nor LLVM supports (a shame, since this fence is hardly free on ARM, which is what I was hoping to improve).

r? `@Amanieu` (Sorry for always picking you, but I know a lot of people wouldn't feel comfortable reviewing atomic ordering changes)

library/std/src/sys/unix/weak.rs

@@ -24,7 +24,7 @@
 use crate::ffi::CStr;
 use crate::marker;
 use crate::mem;
-use crate::sync::atomic::{AtomicUsize, Ordering};
+use crate::sync::atomic::{self, AtomicUsize, Ordering};
 
 macro_rules! weak {
     (fn $name:ident($($t:ty),*) -> $ret:ty) => (
@@ -47,15 +47,49 @@ impl<F> Weak<F> {
     pub fn get(&self) -> Option<F> {
         assert_eq!(mem::size_of::<F>(), mem::size_of::<usize>());
         unsafe {
-            if self.addr.load(Ordering::SeqCst) == 1 {
-                self.addr.store(fetch(self.name), Ordering::SeqCst);
-            }
-            match self.addr.load(Ordering::SeqCst) {
+            // Relaxed is fine here because we fence before reading through the
+            // pointer (see the comment below).
+            match self.addr.load(Ordering::Relaxed) {
+                1 => self.initialize(),
                 0 => None,
-                addr => Some(mem::transmute_copy::<usize, F>(&addr)),
+                addr => {
+                    let func = mem::transmute_copy::<usize, F>(&addr);
+                    // The caller is presumably going to read through this value
+                    // (by calling the function we've dlsymed). This means we'd
+                    // need to have loaded it with at least C11's consume
+                    // ordering in order to be guaranteed that the data we read
+                    // from the pointer isn't from before the pointer was
+                    // stored. Rust has no equivalent to memory_order_consume,
+                    // so we use an acquire fence (sorry, ARM).
+                    //
+                    // Now, in practice this likely isn't needed even on CPUs
+                    // where relaxed and consume mean different things. The
+                    // symbols we're loading are probably present (or not) at
+                    // init, and even if they aren't the runtime dynamic loader
+                    // is extremely likely have sufficient barriers internally
+                    // (possibly implicitly, for example the ones provided by
+                    // invoking `mprotect`).
+                    //
+                    // That said, none of that's *guaranteed*, and so we fence.
+                    atomic::fence(Ordering::Acquire);
+                    Some(func)
+                }
             }
         }
     }
+
+    // Cold because it should only happen during first-time initalization.
+    #[cold]
+    unsafe fn initialize(&self) -> Option<F> {
+        let val = fetch(self.name);
+        // This synchronizes with the acquire fence in `get`.
+        self.addr.store(val, Ordering::Release);
+
+        match val {
+            0 => None,
+            addr => Some(mem::transmute_copy::<usize, F>(&addr)),
+        }
+    }
 }
 
 unsafe fn fetch(name: &str) -> usize {