Implement black_box using intrinsic

The new implementation allows some `memcpy`s to be optimized away,
so the uninit value in ui/sanitize/memory.rs is constructed directly
onto the return place. Therefore the sanitizer now says that the
value is allocated by `main` rather than `random`.

library/core/src/hint.rs

@@ -152,23 +152,19 @@ pub fn spin_loop() {
 /// backend used. Programs cannot rely on `black_box` for *correctness* in any way.
 ///
 /// [`std::convert::identity`]: crate::convert::identity
-#[cfg_attr(not(miri), inline)]
-#[cfg_attr(miri, inline(never))]
+#[inline]
 #[unstable(feature = "bench_black_box", issue = "64102")]
-#[cfg_attr(miri, allow(unused_mut))]
+#[cfg_attr(not(bootstrap), allow(unused_mut))]
 pub fn black_box<T>(mut dummy: T) -> T {
-    // We need to "use" the argument in some way LLVM can't introspect, and on
-    // targets that support it we can typically leverage inline assembly to do
-    // this. LLVM's interpretation of inline assembly is that it's, well, a black
-    // box. This isn't the greatest implementation since it probably deoptimizes
-    // more than we want, but it's so far good enough.
-
-    #[cfg(not(miri))] // This is just a hint, so it is fine to skip in Miri.
+    #[cfg(bootstrap)]
     // SAFETY: the inline assembly is a no-op.
     unsafe {
-        // FIXME: Cannot use `asm!` because it doesn't support MIPS and other architectures.
         llvm_asm!("" : : "r"(&mut dummy) : "memory" : "volatile");
+        dummy
     }
 
-    dummy
+    #[cfg(not(bootstrap))]
+    {
+        crate::intrinsics::black_box(dummy)
+    }
 }