apr_atomic: Generic apr_atomic_read64() needs a mutex on 32bit systems (tearing).

A 64bit load on a 32 bit CPU/system uses two instructions (tearing), so ensure
atomicity with regard to other atomic functions by using the (same) lock.

test_atomics_threaded64() fails because of this on 32bit systems. PR 66457.

* atomic/unix/mutex64.c(apr_atomic_read64):
  Use locking when APR_SIZEOF_VOIDP < 8


atomic64: Generic apr_atomic_read64() to always use a lock.

Don't play games with sizeof(void*) to determine whether a raw load
intruction is atomic or not. Systems that fall back to the generic
implementation are not eligible for the compiler builtins or CPU
native atomic intructions already, and we don't want to reimplement
that here (e.g. alignment, ordering guarantees, ...).

* atomic/unix/mutex64.c(apr_atomic_read64):
  No #ifdefery, always take the lock.

Follow up to r1907541.


atomic: No raw 64bit load/store on 32bit systems or anything but x86_64 or s390x.

Raw 64 bit load and store need two intructions on 32bit systems (tearing) so
they are not atomic, and only x86(_64) and s390(x) have stong mempry ordering
guarantees. Always use builtin functions for the cases where raw load/store
don't work as expected.

* atomic/unix/builtins.c, atomic/unix/builtins64.c:
  Use an accept-list rather than a reject-list to define WEAK_MEMORY_ORDERING.
  Test APR_SIZEOF_VOIDP < 8 to force usage of __sync builtins for _read{32,64}
  and _set{32,64} on 32bit systems when __atomic_{load,store} buitlins are not
  available.


configure: Test apr_uint64_t alignment for 64bit atomic builtins usability.

On some systems the __atomic builtins may be available only through libatomic
or fall back to libatomic when the atomic operations are not issued on a
suitably aligned address (64bit atomics on 8-byte aligned addresses only for
instance).

Modify the tests for HAVE_ATOMIC_BUILTINS64 and HAVE__ATOMIC_BUILTINS64 such
that the address for the atomic operations is not aligned (unless 64bit ints
always have the suitable alignment, i.e. mainly 64bit systems..).
Also, use the __atomic_always_lock_free() builtin to fail the test when the
compiler already knows about the alignment issue (falling back to libatomic,
which we don't require/want).

With this, 64bit builtins should be selected only for platforms that can
natively handle atomics on any apr_uin64_t (since the APR has no dedicated
8-byte aligned 64bit type for now), while the generic/mutex implementation
is used for others.


testatomic: initialize in the test the globals used by it.

Just in case the test is later reordered (e.g. test_atomics_threaded64
and test_atomics_threaded_setread64 use the same atomic_ops64 variable).


atomic: test 4-bytes aligned and/or cross-cacheline atomics (on 32bit systems).


Merges r1907541, r1907678, r1907637, r1907642, r1907677, r1907985 from trunk.
Merges r1907679, r1907998 from 1.8.x.
Submitted by: ylavic


git-svn-id: https://svn.apache.org/repos/asf/apr/apr/branches/1.7.x@1908002 13f79535-47bb-0310-9956-ffa450edef68

3 files changed, 18 insertions, 9 deletions

diff --git a/atomic/unix/builtins.c b/atomic/unix/builtins.c
index 22b828c3c..e7b117c71 100644
--- a/atomic/unix/builtins.c
+++ b/atomic/unix/builtins.c
@@ -18,10 +18,11 @@
 
 #ifdef USE_ATOMICS_BUILTINS
 
-#if defined(__arm__) || defined(__powerpc__) || defined(__powerpc64__)
-#define WEAK_MEMORY_ORDERING 1
-#else
+#if defined(__i386__) || defined(__x86_64__) \
+    || defined(__s390__) || defined(__s390x__)
 #define WEAK_MEMORY_ORDERING 0
+#else
+#define WEAK_MEMORY_ORDERING 1
 #endif
 
 APR_DECLARE(apr_status_t) apr_atomic_init(apr_pool_t *p)
diff --git a/atomic/unix/builtins64.c b/atomic/unix/builtins64.c
index 7d8422548..f0cdb309b 100644
--- a/atomic/unix/builtins64.c
+++ b/atomic/unix/builtins64.c
@@ -18,17 +18,18 @@
 
 #ifdef USE_ATOMICS_BUILTINS64
 
-#if defined(__arm__) || defined(__powerpc__) || defined(__powerpc64__)
-#define WEAK_MEMORY_ORDERING 1
-#else
+#if defined(__i386__) || defined(__x86_64__) \
+    || defined(__s390__) || defined(__s390x__)
 #define WEAK_MEMORY_ORDERING 0
+#else
+#define WEAK_MEMORY_ORDERING 1
 #endif
 
 APR_DECLARE(apr_uint64_t) apr_atomic_read64(volatile apr_uint64_t *mem)
 {
 #if HAVE__ATOMIC_BUILTINS
     return __atomic_load_n(mem, __ATOMIC_SEQ_CST);
-#elif WEAK_MEMORY_ORDERING
+#elif WEAK_MEMORY_ORDERING || APR_SIZEOF_VOIDP < 8
     /* No __sync_load() available => apr_atomic_add64(mem, 0) */
     return __sync_fetch_and_add(mem, 0);
 #else
@@ -40,7 +41,7 @@ APR_DECLARE(void) apr_atomic_set64(volatile apr_uint64_t *mem, apr_uint64_t val)
 {
 #if HAVE__ATOMIC_BUILTINS
     __atomic_store_n(mem, val, __ATOMIC_SEQ_CST);
-#elif WEAK_MEMORY_ORDERING
+#elif WEAK_MEMORY_ORDERING || APR_SIZEOF_VOIDP < 8
     /* No __sync_store() available => apr_atomic_xchg64(mem, val) */
     __sync_synchronize();
     __sync_lock_test_and_set(mem, val);
diff --git a/atomic/unix/mutex64.c b/atomic/unix/mutex64.c
index 9fc44af61..a6a7ceb0b 100644
--- a/atomic/unix/mutex64.c
+++ b/atomic/unix/mutex64.c
@@ -96,7 +96,14 @@ apr_status_t apr__atomic_generic64_init(apr_pool_t *p)
 
 APR_DECLARE(apr_uint64_t) apr_atomic_read64(volatile apr_uint64_t *mem)
 {
-    return *mem;
+    apr_uint64_t cur_value;
+    DECLARE_MUTEX_LOCKED(mutex, mem);
+
+    cur_value = *mem;
+
+    MUTEX_UNLOCK(mutex);
+
+    return cur_value;
 }
 
 APR_DECLARE(void) apr_atomic_set64(volatile apr_uint64_t *mem, apr_uint64_t val)