x86_64: Use xchg instead of lock or

Follow-up to #156.

src/imp/atomic128/x86_64.rs

@@ -68,23 +68,6 @@ macro_rules! ptr_modifier {
     };
 }
 
-#[cfg(not(any(portable_atomic_no_outline_atomics, target_env = "sgx")))]
-#[cfg(target_feature = "sse")]
-#[cfg(target_pointer_width = "32")]
-macro_rules! sp {
-    () => {
-        "esp"
-    };
-}
-#[cfg(not(any(portable_atomic_no_outline_atomics, target_env = "sgx")))]
-#[cfg(target_feature = "sse")]
-#[cfg(target_pointer_width = "64")]
-macro_rules! sp {
-    () => {
-        "rsp"
-    };
-}
-
 // Unlike AArch64 and RISC-V, x86's assembler doesn't check instruction
 // requirements for the currently enabled target features. In the first place,
 // there is no option in the x86 assembly for such case, like ARM .arch_extension,
@@ -200,16 +183,21 @@ unsafe fn atomic_store_vmovdqa(dst: *mut u128, val: u128, order: Ordering) {
                 );
             }
             Ordering::SeqCst => {
+                let p = core::cell::UnsafeCell::new(core::mem::MaybeUninit::<u64>::uninit());
                 asm!(
                     concat!("vmovdqa xmmword ptr [{dst", ptr_modifier!(), "}], {val}"),
-                    // Equivalent to mfence, but is 10-35% faster at least in simple cases on Coffee Lake (https://github.com/taiki-e/portable-atomic/pull/156).
-                    // Based on x86_32 64-bit atomic SeqCst store using SSE generated by LLVM.
-                    // https://godbolt.org/z/9sKEr8YWc
-                    concat!("lock or dword ptr [", sp!(), "], 0"),
+                    // Equivalent to mfence, but is up to 3.1x faster on Coffee Lake and up to 2.4x faster on Raptor Lake-H at least in simple cases.
+                    // - https://github.com/taiki-e/portable-atomic/pull/156
+                    // - LLVM uses lock or for x86_32 64-bit atomic SeqCst store using SSE https://godbolt.org/z/9sKEr8YWc
+                    // - Windows uses xchg for x86_32 for MemoryBarrier https://learn.microsoft.com/en-us/windows/win32/api/winnt/nf-winnt-memorybarrier
+                    // - MSVC STL uses lock inc https://github.com/microsoft/STL/pull/740
+                    // - boost uses lock or https://github.com/boostorg/atomic/commit/559eba81af71386cedd99f170dc6101c6ad7bf22
+                    concat!("xchg qword ptr [{p", ptr_modifier!(), "}], {tmp}"),
                     dst = in(reg) dst,
                     val = in(xmm_reg) val,
-                    // Do not use `preserves_flags` because OR modifies the OF, CF, SF, ZF, and PF flags.
-                    options(nostack),
+                    p = inout(reg) p.get() => _,
+                    tmp = lateout(reg) _,
+                    options(nostack, preserves_flags),
                 );
             }
             _ => unreachable!(),