Rollup merge of #41143 - stjepang:optimize-bool-fetch-nand, r=nagisa

Optimize AtomicBool::fetch_nand

This is an attempt to push the PR #40563 to completion.

Benchmark: [source](https://gist.github.com/stjepang/023f5025623f5474184f9f4dfd6379ae)
Improvement:

```
 name  old_ ns/iter  new_ce_ ns/iter  diff ns/iter   diff %
 1t    146,440       89,904                -56,536  -38.61%
 2t    561,456       316,835              -244,621  -43.57%
 4t    2,822,821     1,005,424          -1,817,397  -64.38%
```

r? @eddyb
cc @alexcrichton @nagisa

src/libcore/sync/atomic.rs

@@ -539,17 +539,16 @@ pub fn fetch_nand(&self, val: bool, order: Ordering) -> bool {
         // We can't use atomic_nand here because it can result in a bool with
         // an invalid value. This happens because the atomic operation is done
         // with an 8-bit integer internally, which would set the upper 7 bits.
-        // So we just use a compare-exchange loop instead, which is what the
-        // intrinsic actually expands to anyways on many platforms.
-        let mut old = self.load(Relaxed);
-        loop {
-            let new = !(old && val);
-            match self.compare_exchange_weak(old, new, order, Relaxed) {
-                Ok(_) => break,
-                Err(x) => old = x,
-            }
+        // So we just use fetch_xor or swap instead.
+        if val {
+            // !(x & true) == !x
+            // We must invert the bool.
+            self.fetch_xor(true, order)
+        } else {
+            // !(x & false) == true
+            // We must set the bool to true.
+            self.swap(true, order)
         }
-        old
     }
 
     /// Logical "or" with a boolean value.

src/libcore/tests/atomic.rs

@@ -24,10 +24,23 @@ fn bool_() {
 #[test]
 fn bool_and() {
     let a = AtomicBool::new(true);
-    assert_eq!(a.fetch_and(false, SeqCst),true);
+    assert_eq!(a.fetch_and(false, SeqCst), true);
     assert_eq!(a.load(SeqCst),false);
 }
 
+#[test]
+fn bool_nand() {
+    let a = AtomicBool::new(false);
+    assert_eq!(a.fetch_nand(false, SeqCst), false);
+    assert_eq!(a.load(SeqCst), true);
+    assert_eq!(a.fetch_nand(false, SeqCst), true);
+    assert_eq!(a.load(SeqCst), true);
+    assert_eq!(a.fetch_nand(true, SeqCst), true);
+    assert_eq!(a.load(SeqCst), false);
+    assert_eq!(a.fetch_nand(true, SeqCst), false);
+    assert_eq!(a.load(SeqCst), true);
+}
+
 #[test]
 fn uint_and() {
     let x = AtomicUsize::new(0xf731);