MDEV-26720: rw_lock: Prefer fetch_sub() to fetch_and()

rw_lock::write_unlock(): Revert part of
commit d46b424 (MDEV-24142)
to make the IA-32 and AMD64 implementation faster.

storage/innobase/include/rw_lock.h

@@ -196,8 +196,12 @@ class rw_lock
   /** Release an exclusive lock */
   void write_unlock()
   {
-    IF_DBUG_ASSERT(auto l=,)
-    lock.fetch_and(~WRITER, std::memory_order_release);
+    /* Below, we use fetch_sub(WRITER) instead of fetch_and(~WRITER).
+    The reason is that on IA-32 and AMD64 it translates into the 80486
+    instruction LOCK XADD, while fetch_and() translates into a loop
+    around LOCK CMPXCHG. For other ISA either form should be fine. */
+    static_assert(WRITER == 1U << 31, "compatibility");
+    IF_DBUG_ASSERT(auto l=,) lock.fetch_sub(WRITER, std::memory_order_release);
     /* the write lock must have existed */
 #ifdef SUX_LOCK_GENERIC
     DBUG_ASSERT((l & (WRITER | UPDATER)) == WRITER);