Move code documentation around.

Fix volatile warning in MSVC.

Include/cpython/pyatomic.h

@@ -1,6 +1,3 @@
-#ifndef Py_ATOMIC_H
-#define Py_ATOMIC_H
-
 // This header provides cross-platform low-level atomic operations
 // similar to C11 atomics.
 //
@@ -10,17 +7,22 @@
 // The "_relaxed" suffix for load and store operations indicates the "relaxed"
 // memory order. They don't provide synchronization, but (roughly speaking)
 // guarantee somewhat sane behavior for races instead of undefined behavior.
-// In practice, they correspond to "normal" hardware load and store instructions,
-// so they are almost as inexpensive as plain loads and stores in C.
+// In practice, they correspond to "normal" hardware load and store
+// instructions, so they are almost as inexpensive as plain loads and stores
+// in C.
 //
 // Note that atomic read-modify-write operations like _Py_atomic_add_* return
 // the previous value of the atomic variable, not the new value.
 //
-// See https://en.cppreference.com/w/c/atomic for more information on C11 atomics.
+// See https://en.cppreference.com/w/c/atomic for more information on C11
+// atomics.
 // See https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2020/p2055r0.pdf
 // "A Relaxed Guide to memory_order_relaxed" for discussion of and common usage
 // or relaxed atomics.
 
+#ifndef Py_ATOMIC_H
+#define Py_ATOMIC_H
+
 // Atomically adds `value` to `address` and returns the previous value
 static inline int
 _Py_atomic_add_int(int *address, int value);

Include/cpython/pyatomic_gcc.h

@@ -1,12 +1,12 @@
-#ifndef Py_ATOMIC_GCC_H
-#  error "this header file must not be included directly"
-#endif
-
 // This is the implementation of Python atomic operations using GCC's built-in
 // functions that match the C+11 memory model. This implementation is preferred
 // for GCC compatible compilers, such as Clang. These functions are available in
 // GCC 4.8+ without needing to compile with --std=c11 or --std=gnu11.
 
+#ifndef Py_ATOMIC_GCC_H
+#  error "this header file must not be included directly"
+#endif
+
 static inline int
 _Py_atomic_add_int(int *address, int value)
 {

Include/cpython/pyatomic_msc.h

@@ -1,10 +1,10 @@
+// This is the implementation of Python atomic operations for MSVC if the
+// compiler does not support C11 or C++11 atomics.
+
 #ifndef Py_ATOMIC_MSC_H
 #  error "this header file must not be included directly"
 #endif
 
-// This is the implementation of Python atomic operations for MSVC if the
-// compiler does not support C11 or C++11 atomics.
-
 #include <intrin.h>
 
 
@@ -85,7 +85,7 @@ _Py_atomic_add_uint32(uint32_t *address, uint32_t value)
 static inline uint64_t
 _Py_atomic_add_uint64(uint64_t *address, uint64_t value)
 {
-    return (uint64_t)_Py_atomic_add_int64((volatile int64_t*)address, (int64_t)value);
+    return (uint64_t)_Py_atomic_add_int64((int64_t*)address, (int64_t)value);
 }
 
 static inline uintptr_t
@@ -270,7 +270,7 @@ _Py_atomic_exchange_uint32(uint32_t *address, uint32_t value)
 static inline uint64_t
 _Py_atomic_exchange_uint64(uint64_t *address, uint64_t value)
 {
-    return (uint64_t)_Py_atomic_exchange_int64((volatile __int64*)address, (__int64)value);
+    return (uint64_t)_Py_atomic_exchange_int64((int64_t *)address, (int64_t)value);
 }
 
 static inline uintptr_t
@@ -693,7 +693,7 @@ _Py_atomic_store_uint32(uint32_t *address, uint32_t value)
 static inline void
 _Py_atomic_store_uint64(uint64_t *address, uint64_t value)
 {
-    _Py_atomic_exchange_int64((volatile __int64*)address, (__int64)value);
+    _Py_atomic_exchange_int64((int64_t *)address, (int64_t)value);
 }
 
 static inline void

Include/cpython/pyatomic_std.h

@@ -1,11 +1,11 @@
-#ifndef Py_ATOMIC_STD_H
-#  error "this header file must not be included directly"
-#endif
-
 // This is the implementation of Python atomic operations using C++11 or C11
 // atomics. Note that the pyatomic_gcc.h implementation is preferred for GCC
 // compatible compilers, even if they support C++11 atomics.
 
+#ifndef Py_ATOMIC_STD_H
+#  error "this header file must not be included directly"
+#endif
+
 #ifdef __cplusplus
 extern "C++" {
 #include <atomic>