WIP: Work on magic nop instructions

objc2/benches/autorelease.rs

@@ -134,6 +134,7 @@ macro_rules! main_with_warmup {
             )+
         }
 
+        // Needed to get DYLD to resolve the stubs
         fn warmup() {
             $(
                 warmup_fns::$f();

objc2/src/rc/id.rs

@@ -274,12 +274,82 @@ impl<T: Message, O: Ownership> Id<T, O> {
 
     /// TODO
     #[doc(alias = "objc_retainAutoreleasedReturnValue")]
+    // This relies heavily on being inlined right after `objc_msgSend`.
     #[inline(always)]
     pub unsafe fn retain_autoreleased(ptr: NonNull<T>) -> Id<T, O> {
-        // SAFETY: Same as `retain`, `objc_retainAutoreleasedReturnValue` is
-        // just an optimization.
+        // Not supported on TARGET_OS_WIN32
+        #[cfg(all(apple, not(target_os = "windows")))]
+        {
+            // Add magic nop instruction to participate in the fast
+            // autorelease scheme.
+            //
+            // We will unconditionally emit these instructions, even if they
+            // end up being unused (for example because we're unlucky with
+            // inlining, some other work is done between the objc_msgSend and
+            // this, or the runtime version is too old to support it).
+            //
+            // See `callerAcceptsOptimizedReturn` in `objc-object.h`:
+            // https://github.com/apple-oss-distributions/objc4/blob/objc4-838/runtime/objc-object.h#L1209-L1377
+            // and this StackOverflow answer for some background on why the
+            // design is like it is: https://stackoverflow.com/a/23765612.
+            //
+            // It may seem like there should be a better way to do this, but
+            // emitting raw assembly is exactly what Clang and Swift does:
+            // swiftc: https://github.com/apple/swift/blob/swift-5.5.3-RELEASE/lib/IRGen/GenObjC.cpp#L148-L173
+            // Clang: https://github.com/llvm/llvm-project/blob/889317d47b7f046cf0e68746da8f7f264582fb5b/clang/lib/CodeGen/CGObjC.cpp#L2339-L2373
+            //
+            // SAFETY:
+            // Based on https://doc.rust-lang.org/stable/reference/inline-assembly.html#rules-for-inline-assembly
+            //
+            // We don't care about the value of the register (so it's okay to
+            // be undefined), and its value is preserved.
+            //
+            // nomem: No reads or writes to memory are performed (this `mov`
+            //   operates entirely on registers).
+            // preserves_flags: `mov` doesn't modify any flags.
+            // nostack: We don't touch the stack.
+
+            // Supported since macOS 10.7.
+            #[cfg(target_arch = "x86_64")]
+            {} // x86_64 looks at the next call instruction
+
+            // Supported since macOS 10.8.
+            #[cfg(target_arch = "arm")]
+            unsafe {
+                core::arch::asm!("mov r7, r7", options(nomem, preserves_flags, nostack))
+            };
+
+            // Supported since macOS 10.10.
+            #[cfg(target_arch = "aarch64")]
+            unsafe {
+                core::arch::asm!("mov fp, fp", options(nomem, preserves_flags, nostack))
+            };
+
+            // Supported since macOS 10.12.
+            #[cfg(target_arch = "x86")]
+            unsafe {
+                core::arch::asm!("mov ebp, ebp", options(nomem, preserves_flags, nostack))
+            };
+        }
 
         let ptr = ptr.as_ptr() as *mut objc_sys::objc_object;
+
+        // objc_autoreleaseReturnValue / objc_retainAutoreleasedReturnValue:
+
+        // #![feature(asm_sym)]
+        // #[cfg(target_arch = "x86_64")]
+        // unsafe {
+        //     core::arch::asm!(
+        //         "mov rdi, rax",
+        //         "call {}",
+        //         sym objc2::ffi::objc_retainAutoreleasedReturnValue,
+        //         inout("rax") obj,
+        //         clobber_abi("C"),
+        //     );
+        // }
+
+        // SAFETY: Same as `retain`, `objc_retainAutoreleasedReturnValue` is
+        // just an optimization.
         let res = unsafe { objc_sys::objc_retainAutoreleasedReturnValue(ptr) };
         debug_assert_eq!(
             res, ptr,