* Clean up some problems with the atomic operation interface discovered

during testing.  In particular, we can't have a cmpset macro that can
  deal with pointers without some really evil voodoo.  So have a
  different macro for pointers
* Add more detailed testing of the atomics using AM's test framework.
  More to come...

This commit was SVN r4191.

configure.ac

@@ -1489,5 +1489,7 @@ AC_CONFIG_FILES([
     src/tools/ompid/Makefile
     src/tools/openmpi/Makefile
     src/tools/wrappers/Makefile
+
+    src/asm/test/Makefile
 ])
 AC_OUTPUT

src/asm/Makefile.am

@@ -14,6 +14,8 @@
 
 include $(top_srcdir)/config/Makefile.options
 
+SUBDIRS = . test
+
 ######################################################################
 #
 # This is a bit complicated.  If there is anything in the library,
@@ -57,13 +59,6 @@ EXTRA_DIST	= \
 
 ######################################################################
 
-TESTS = atomic-test
-check_PROGRAMS = atomic-test
-atomic_test_SOURCES = atomic-test.c
-atomic_test_LDADD = libasm.la
-
-######################################################################
-
 clean-local:
 	rm -f atomic-asm.s
 

src/asm/test/Makefile.am

@@ -0,0 +1,79 @@
+#
+# Copyright (c) 2004-2005 The Trustees of Indiana University.
+#                         All rights reserved.
+# Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
+#                         All rights reserved.
+# Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, 
+#                         University of Stuttgart.  All rights reserved.
+# $COPYRIGHT$
+# 
+# Additional copyrights may follow
+# 
+# $HEADER$
+#
+
+include $(top_srcdir)/config/Makefile.options
+
+noinst_HEADERS = atomic_test.h
+
+TESTS = \
+	atomic_barrier \
+	atomic_spinlock_serial \
+	atomic_spinlock_2 \
+	atomic_spinlock_5 \
+	atomic_spinlock_8 \
+	atomic_cmpset_32_serial \
+	atomic_cmpset_64_serial \
+	atomic-test
+
+check_PROGRAMS = \
+	atomic_barrier \
+	atomic_spinlock_serial \
+	atomic_spinlock_2 \
+	atomic_spinlock_5 \
+	atomic_spinlock_8 \
+	atomic_cmpset_32_serial \
+	atomic_cmpset_64_serial \
+	atomic-test
+
+######################################################################
+
+atomic_barrier_SOURCES = atomic_barrier.c
+atomic_barrier_LDADD = ../libasm.la
+
+######################################################################
+
+atomic_spinlock_serial_SOURCES = \
+	atomic_spinlock_serial.c \
+	atomic_spinlock.c
+atomic_spinlock_serial_LDADD = ../libasm.la
+
+atomic_spinlock_2_SOURCES = \
+	atomic_spinlock_2.c \
+	atomic_spinlock.c
+atomic_spinlock_2_LDADD = ../libasm.la
+
+atomic_spinlock_5_SOURCES = \
+	atomic_spinlock_5.c \
+	atomic_spinlock.c
+atomic_spinlock_5_LDADD = ../libasm.la
+
+atomic_spinlock_8_SOURCES = \
+	atomic_spinlock_8.c \
+	atomic_spinlock.c
+atomic_spinlock_8_LDADD = ../libasm.la
+
+######################################################################
+
+atomic_cmpset_32_serial_SOURCES = atomic_cmpset_32_serial.c
+atomic_cmpset_32_serial_LDADD = ../libasm.la
+
+######################################################################
+
+atomic_cmpset_64_serial_SOURCES = atomic_cmpset_64_serial.c
+atomic_cmpset_64_serial_LDADD = ../libasm.la
+
+######################################################################
+
+atomic_test_SOURCES = atomic-test.c
+atomic_test_LDADD = ../libasm.la

src/asm/test/atomic-test.c

@@ -1,6 +1,20 @@
-#undef OMPI_BUILDING
+/*
+ * Copyright (c) 2004-2005 The Trustees of Indiana University.
+ *                         All rights reserved.
+ * Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
+ *                         All rights reserved.
+ * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, 
+ *                         University of Stuttgart.  All rights reserved.
+ * $COPYRIGHT$
+ * 
+ * Additional copyrights may follow
+ * 
+ * $HEADER$
+ */
 
+#undef OMPI_BUILDING
 #include "ompi_config.h"
+
 #include <assert.h>
 #include <getopt.h>
 #ifdef HAVE_PTHREAD_H
@@ -13,102 +27,6 @@
 
 #include "include/sys/atomic.h"
 
-/**
- * A testing support library to provide uniform reporting output
- */
-
-static int ompi_n_tests;
-static int ompi_n_success;
-static int ompi_n_failures;
-static char *ompi_description;
-
-static void test_init(char *a)
-{
-    /* local variables */
-    size_t len;
-
-    /* save the descriptive string */
-    len = strlen(a);
-    ompi_description = (char *) malloc(len + 1);
-    assert(ompi_description);
-
-    strcpy(ompi_description, a);
-
-    /* initialize counters */
-    ompi_n_tests = 0;
-    ompi_n_success = 0;
-    ompi_n_failures = 0;
-
-    return;
-
-}
-
-
-static void test_success(void)
-{
-    ompi_n_tests++;
-    ompi_n_success++;
-}
-
-
-static void test_failure(char *a)
-{
-    ompi_n_tests++;
-    ompi_n_failures++;
-
-    fprintf(stderr, " Failure : ");
-    fprintf(stderr, a);
-    fprintf(stderr, "\n");
-    fflush(stderr);
-}
-
-
-static int test_verify_int(int expected_result, int test_result)
-{
-    int return_value;
-
-    return_value = 1;
-    if (expected_result != test_result) {
-        test_failure("Comparison failure");
-        fprintf(stderr, " Expected result: %d\n", expected_result);
-        fprintf(stderr, " Test result: %d\n", test_result);
-        fflush(stderr);
-        return_value = 0;
-    } else {
-        test_success();
-    }
-
-    return return_value;
-}
-
-
-static int test_finalize(void)
-{
-    int return_value;
-
-    return_value = 1;
-
-    if (ompi_n_tests == ompi_n_success) {
-        fprintf(stderr, "SUPPORT: OMPI Test Passed: %s: (%d tests)\n",
-                ompi_description, ompi_n_tests);
-        fflush(stderr);
-    } else {
-        fprintf(stderr,
-                "SUPPORT: OMPI Test failed: %s (%d of %d failed)\n",
-                ompi_description, ompi_n_failures, ompi_n_tests);
-        fflush(stderr);
-        return_value = 0;
-    }
-
-    return return_value;
-}
-
-
-/* note this is for additional output that does NOT go to STDERR but STDOUT */
-static void test_comment (char* userstr)
-{
-	fprintf(stdout, "%s:%s\n", ompi_description, userstr);
-}
 
 /* default options */
 
@@ -122,7 +40,7 @@ int32_t val32;
 int32_t old32;
 int32_t new32;
 
-#ifdef ENABLE_64_BIT
+#if OMPI_HAVE_ATOMIC_MATH_64
 volatile int64_t vol64;
 int64_t val64;
 int64_t old64;
@@ -144,7 +62,7 @@ static void help(void)
     printf("Usage: threadtest [flags]\n"
            "\n"
            "   Flags may be any of\n"
-#ifdef ENABLE_64_BIT
+#if OMPI_HAVE_ATOMIC_MATH_64
            "      -l                do 64-bit tests\n"
 #endif
            "      -r NREPS          number of repetitions\n"
@@ -153,7 +71,7 @@ static void help(void)
            "      -h                print this info\n" "\n"
            "   Numbers may be postfixed with 'k' or 'm'\n\n");
 
-#ifndef ENABLE_64_BIT
+#ifndef OMPI_HAVE_ATOMIC_MATH_64
     printf("   64-bit tests are not enabled in this build of the tests\n\n");
 #endif
 
@@ -216,7 +134,7 @@ static void *thread_main(void *arg)
 
     for (i = 0; i < nreps; i++) {
         ompi_atomic_add_32(&val32, 5);
-#ifdef ENABLE_64_BIT
+#if OMPI_HAVE_ATOMIC_MATH_64
         if (enable_64_bit_tests) {
             ompi_atomic_add_64(&val64, 5);
         }
@@ -231,20 +149,20 @@ static void *thread_main(void *arg)
 int main(int argc, char *argv[])
 {
     int c;
+#if HAVE_PTHREAD_H
     int tid;
     pthread_t *th;
+#endif
 
     /* option processing */
 
-    test_init("atomic operations");
-
     while ((c = getopt(argc, argv, "hlr:t:v")) != -1) {
         switch (c) {
         case 'h':
             help();
             break;
         case 'l':
-#ifdef ENABLE_64_BIT
+#if OMPI_HAVE_ATOMIC_MATH_64
             enable_64_bit_tests = 1;
 #else
 	    usage();
@@ -280,144 +198,144 @@ int main(int argc, char *argv[])
     /* -- cmpset 32-bit tests -- */
 
     vol32 = 42, old32 = 42, new32 = 50;
-    test_verify_int(ompi_atomic_cmpset_32(&vol32, old32, new32), 1);
-    test_verify_int(vol32, new32);
+    assert(ompi_atomic_cmpset_32(&vol32, old32, new32) == 1);
+    assert(vol32 == new32);
 
     vol32 = 42, old32 = 420, new32 = 50;
-    test_verify_int(ompi_atomic_cmpset_32(&vol32, old32, new32), 0);
-    test_verify_int(vol32, 42);
+    assert(ompi_atomic_cmpset_32(&vol32, old32, new32) ==  0);
+    assert(vol32 == 42);
 
     vol32 = 42, old32 = 42, new32 = 50;
-    test_verify_int(ompi_atomic_cmpset_acq_32(&vol32, old32, new32), 1);
-    test_verify_int(vol32, new32);
+    assert(ompi_atomic_cmpset_acq_32(&vol32, old32, new32) == 1);
+    assert(vol32 == new32);
 
     vol32 = 42, old32 = 420, new32 = 50;
-    test_verify_int(ompi_atomic_cmpset_acq_32(&vol32, old32, new32), 0);
-    test_verify_int(vol32, 42);
+    assert(ompi_atomic_cmpset_acq_32(&vol32, old32, new32) == 0);
+    assert(vol32 == 42);
 
     vol32 = 42, old32 = 42, new32 = 50;
-    test_verify_int(ompi_atomic_cmpset_rel_32(&vol32, old32, new32), 1);
-    test_verify_int(vol32, new32);
+    assert(ompi_atomic_cmpset_rel_32(&vol32, old32, new32) ==  1);
+    assert(vol32 == new32);
 
     vol32 = 42, old32 = 420, new32 = 50;
-    test_verify_int(ompi_atomic_cmpset_rel_32(&vol32, old32, new32), 0);
-    test_verify_int(vol32, 42);
+    assert(ompi_atomic_cmpset_rel_32(&vol32, old32, new32) == 0);
+    assert(vol32 == 42);
 
     /* -- cmpset 64-bit tests -- */
 
-#ifdef ENABLE_64_BIT
+#if OMPI_HAVE_ATOMIC_MATH_64
     if (enable_64_bit_tests) {
 	verbose("64 bit serial tests\n");
         vol64 = 42, old64 = 42, new64 = 50;
-        test_verify_int(1, ompi_atomic_cmpset_64(&vol64, old64, new64));
-        test_verify_int(new64, vol64);
+        assert(1 == ompi_atomic_cmpset_64(&vol64, old64, new64));
+        assert(new64 == vol64);
 
 	verbose("64 bit serial test 2\n");
         vol64 = 42, old64 = 420, new64 = 50;
-        test_verify_int(ompi_atomic_cmpset_64(&vol64, old64, new64), 0);
-        test_verify_int(vol64, 42);
+        assert(ompi_atomic_cmpset_64(&vol64, old64, new64) == 0);
+        assert(vol64 == 42);
 
         vol64 = 42, old64 = 42, new64 = 50;
-        test_verify_int(ompi_atomic_cmpset_acq_64(&vol64, old64, new64), 1);
-        test_verify_int(vol64, new64);
+        assert(ompi_atomic_cmpset_acq_64(&vol64, old64, new64) == 1);
+        assert(vol64 == new64);
 
         vol64 = 42, old64 = 420, new64 = 50;
-        test_verify_int(ompi_atomic_cmpset_acq_64(&vol64, old64, new64), 0);
-        test_verify_int(vol64, 42);
+        assert(ompi_atomic_cmpset_acq_64(&vol64, old64, new64) == 0);
+        assert(vol64 == 42);
 
         vol64 = 42, old64 = 42, new64 = 50;
-        test_verify_int(ompi_atomic_cmpset_rel_64(&vol64, old64, new64), 1);
-        test_verify_int(vol64, new64);
+        assert(ompi_atomic_cmpset_rel_64(&vol64, old64, new64) == 1);
+        assert(vol64 == new64);
 
         vol64 = 42, old64 = 420, new64 = 50;
-        test_verify_int(ompi_atomic_cmpset_rel_64(&vol64, old64, new64), 0);
-        test_verify_int(vol64, 42);
+        assert(ompi_atomic_cmpset_rel_64(&vol64, old64, new64) == 0);
+        assert(vol64 == 42);
     }
 #endif
     /* -- cmpset int tests -- */
 
     volint = 42, oldint = 42, newint = 50;
-    test_verify_int(ompi_atomic_cmpset(&volint, oldint, newint), 1);
-    test_verify_int(volint, newint);
+    assert(ompi_atomic_cmpset(&volint, oldint, newint) == 1);
+    assert(volint ==newint);
 
     volint = 42, oldint = 420, newint = 50;
-    test_verify_int(ompi_atomic_cmpset(&volint, oldint, newint), 0);
-    test_verify_int(volint, 42);
+    assert(ompi_atomic_cmpset(&volint, oldint, newint) == 0);
+    assert(volint == 42);
 
     volint = 42, oldint = 42, newint = 50;
-    test_verify_int(ompi_atomic_cmpset_acq(&volint, oldint, newint), 1);
-    test_verify_int(volint, newint);
+    assert(ompi_atomic_cmpset_acq(&volint, oldint, newint) == 1);
+    assert(volint == newint);
 
     volint = 42, oldint = 420, newint = 50;
-    test_verify_int(ompi_atomic_cmpset_acq(&volint, oldint, newint), 0);
-    test_verify_int(volint, 42);
+    assert(ompi_atomic_cmpset_acq(&volint, oldint, newint) == 0);
+    assert(volint == 42);
 
     volint = 42, oldint = 42, newint = 50;
-    test_verify_int(ompi_atomic_cmpset_rel(&volint, oldint, newint), 1);
-    test_verify_int(volint, newint);
+    assert(ompi_atomic_cmpset_rel(&volint, oldint, newint) == 1);
+    assert(volint == newint);
 
     volint = 42, oldint = 420, newint = 50;
-    test_verify_int(ompi_atomic_cmpset_rel(&volint, oldint, newint), 0);
-    test_verify_int(volint, 42);
+    assert(ompi_atomic_cmpset_rel(&volint, oldint, newint) == 0);
+    assert(volint == 42);
 
 
     /* -- cmpset ptr tests -- */
 
     volptr = (void *) 42, oldptr = (void *) 42, newptr = (void *) 50;
-    test_verify_int(ompi_atomic_cmpset(&volptr, oldptr, newptr), 1);
-    test_verify_int(volptr, newptr);
+    assert(ompi_atomic_cmpset_ptr(&volptr, oldptr, newptr) == 1);
+    assert(volptr == newptr);
 
     volptr = (void *) 42, oldptr = (void *) 420, newptr = (void *) 50;
-    test_verify_int(ompi_atomic_cmpset(&volptr, oldptr, newptr), 0);
-    test_verify_int(volptr, (void *) 42);
+    assert(ompi_atomic_cmpset_ptr(&volptr, oldptr, newptr) == 0);
+    assert(volptr == (void *) 42);
 
     volptr = (void *) 42, oldptr = (void *) 42, newptr = (void *) 50;
-    test_verify_int(ompi_atomic_cmpset_acq(&volptr, oldptr, newptr), 1);
-    test_verify_int(volptr, newptr);
+    assert(ompi_atomic_cmpset_acq_ptr(&volptr, oldptr, newptr) == 1);
+    assert(volptr == newptr);
 
     volptr = (void *) 42, oldptr = (void *) 420, newptr = (void *) 50;
-    test_verify_int(ompi_atomic_cmpset_acq(&volptr, oldptr, newptr), 0);
-    test_verify_int(volptr, (void *) 42);
+    assert(ompi_atomic_cmpset_acq_ptr(&volptr, oldptr, newptr) == 0);
+    assert(volptr == (void *) 42);
 
     volptr = (void *) 42, oldptr = (void *) 42, newptr = (void *) 50;
-    test_verify_int(ompi_atomic_cmpset_rel(&volptr, oldptr, newptr), 1);
-    test_verify_int(volptr, newptr);
+    assert(ompi_atomic_cmpset_rel_ptr(&volptr, oldptr, newptr) == 1);
+    assert(volptr == newptr);
 
     volptr = (void *) 42, oldptr = (void *) 420, newptr = (void *) 50;
-    test_verify_int(ompi_atomic_cmpset_rel(&volptr, oldptr, newptr), 0);
-    test_verify_int(volptr, (void *) 42);
+    assert(ompi_atomic_cmpset_rel_ptr(&volptr, oldptr, newptr) == 0);
+    assert(volptr == (void *) 42);
 
     /* -- add_32 tests -- */
 
     val32 = 42;
-    test_verify_int(ompi_atomic_add_32(&val32, 5), (42 + 5));
-    test_verify_int((42 + 5), val32);
+    assert(ompi_atomic_add_32(&val32, 5) == (42 + 5));
+    assert((42 + 5) == val32);
 
     /* -- add_64 tests -- */
-#ifdef ENABLE_64_BIT
+#if OMPI_HAVE_ATOMIC_MATH_64
     if (enable_64_bit_tests) {
         val64 = 42;
-        test_verify_int(ompi_atomic_add_64(&val64, 5), (42 + 5));
-        test_verify_int((42 + 5), val64);
+        assert(ompi_atomic_add_64(&val64, 5) == (42 + 5));
+        assert((42 + 5) == val64);
     }
 #endif
     /* -- add_int tests -- */
 
     valint = 42;
     ompi_atomic_add(&valint, 5);
-    test_verify_int((42 + 5), valint);
+    assert((42 + 5) == valint);
 
 
     /* threaded tests */
 
     val32 = 0;
-#ifdef ENABLE_64_BIT
+#if OMPI_HAVE_ATOMIC_MATH_64
     val64 = 0ul;
 #endif
     valint = 0;
 
     /* -- create the thread set -- */
-
+#if HAVE_PTHREAD_H
     th = (pthread_t *) malloc(nthreads * sizeof(pthread_t));
     if (!th) {
         perror("malloc");
@@ -443,15 +361,14 @@ int main(int argc, char *argv[])
     }
     free(th);
 
-    test_verify_int((5 * nthreads * nreps), val32);
-#ifdef ENABLE_64_BIT
+    assert((5 * nthreads * nreps) == val32);
+#if OMPI_HAVE_ATOMIC_MATH_64
     if (enable_64_bit_tests) {
-        test_verify_int((5 * nthreads * nreps), val64);
+        assert((5 * nthreads * nreps) ==  val64);
     }
 #endif
-    test_verify_int((5 * nthreads * nreps), valint);
-
-    test_finalize();
+    assert((5 * nthreads * nreps) == valint);
+#endif
 
     return 0;
 }

src/asm/test/atomic_barrier.c

@@ -0,0 +1,36 @@
+/*
+ * Copyright (c) 2004-2005 The Trustees of Indiana University.
+ *                         All rights reserved.
+ * Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
+ *                         All rights reserved.
+ * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, 
+ *                         University of Stuttgart.  All rights reserved.
+ * $COPYRIGHT$
+ * 
+ * Additional copyrights may follow
+ * 
+ * $HEADER$
+ */
+
+#undef OMPI_BUILDING
+#include "ompi_config.h"
+
+#include <assert.h>
+
+#include "include/sys/atomic.h"
+#include "atomic_test.h"
+
+int
+main(int argc, char *argv[])
+{
+    /* there really isn't a great way to test that the barriers
+       actually barrier, but at least make sure they don't kill the
+       machine.*/
+
+    ompi_atomic_mb();
+    ompi_atomic_rmb();
+    ompi_atomic_wmb();
+
+    return 0;
+}
+

src/asm/test/atomic_cmpset_32_serial.c

@@ -0,0 +1,56 @@
+/*
+ * Copyright (c) 2004-2005 The Trustees of Indiana University.
+ *                         All rights reserved.
+ * Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
+ *                         All rights reserved.
+ * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, 
+ *                         University of Stuttgart.  All rights reserved.
+ * $COPYRIGHT$
+ * 
+ * Additional copyrights may follow
+ * 
+ * $HEADER$
+ */
+
+#undef OMPI_BUILDING
+#include "ompi_config.h"
+
+#include <assert.h>
+
+#include "include/sys/atomic.h"
+
+volatile int32_t vol32;
+int32_t val32;
+int32_t old32;
+int32_t new32;
+
+int
+main(int argc, char *argv[])
+{
+
+    vol32 = 42, old32 = 42, new32 = 50;
+    assert(ompi_atomic_cmpset_32(&vol32, old32, new32) == 1);
+    assert(vol32 == new32);
+
+    vol32 = 42, old32 = 420, new32 = 50;
+    assert(ompi_atomic_cmpset_32(&vol32, old32, new32) ==  0);
+    assert(vol32 == 42);
+
+    vol32 = 42, old32 = 42, new32 = 50;
+    assert(ompi_atomic_cmpset_acq_32(&vol32, old32, new32) == 1);
+    assert(vol32 == new32);
+
+    vol32 = 42, old32 = 420, new32 = 50;
+    assert(ompi_atomic_cmpset_acq_32(&vol32, old32, new32) == 0);
+    assert(vol32 == 42);
+
+    vol32 = 42, old32 = 42, new32 = 50;
+    assert(ompi_atomic_cmpset_rel_32(&vol32, old32, new32) ==  1);
+    assert(vol32 == new32);
+
+    vol32 = 42, old32 = 420, new32 = 50;
+    assert(ompi_atomic_cmpset_rel_32(&vol32, old32, new32) == 0);
+    assert(vol32 == 42);
+
+    return 0;
+}

src/asm/test/atomic_cmpset_64_serial.c

@@ -0,0 +1,61 @@
+/*
+ * Copyright (c) 2004-2005 The Trustees of Indiana University.
+ *                         All rights reserved.
+ * Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
+ *                         All rights reserved.
+ * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, 
+ *                         University of Stuttgart.  All rights reserved.
+ * $COPYRIGHT$
+ * 
+ * Additional copyrights may follow
+ * 
+ * $HEADER$
+ */
+
+#undef OMPI_BUILDING
+#include "ompi_config.h"
+
+#include <assert.h>
+
+#include "include/sys/atomic.h"
+
+#if OMPI_HAVE_ATOMIC_MATH_64
+volatile int64_t vol64;
+int64_t val64;
+int64_t old64;
+int64_t new64;
+#endif
+
+int
+main(int argc, char *argv[])
+{
+#if OMPI_HAVE_ATOMIC_MATH_64
+    vol64 = 42, old64 = 42, new64 = 50;
+    assert(1 == ompi_atomic_cmpset_64(&vol64, old64, new64));
+    assert(new64 == vol64);
+
+    vol64 = 42, old64 = 420, new64 = 50;
+    assert(ompi_atomic_cmpset_64(&vol64, old64, new64) == 0);
+    assert(vol64 == 42);
+
+    vol64 = 42, old64 = 42, new64 = 50;
+    assert(ompi_atomic_cmpset_acq_64(&vol64, old64, new64) == 1);
+    assert(vol64 == new64);
+
+    vol64 = 42, old64 = 420, new64 = 50;
+    assert(ompi_atomic_cmpset_acq_64(&vol64, old64, new64) == 0);
+    assert(vol64 == 42);
+
+    vol64 = 42, old64 = 42, new64 = 50;
+    assert(ompi_atomic_cmpset_rel_64(&vol64, old64, new64) == 1);
+    assert(vol64 == new64);
+
+    vol64 = 42, old64 = 420, new64 = 50;
+    assert(ompi_atomic_cmpset_rel_64(&vol64, old64, new64) == 0);
+    assert(vol64 == 42);
+
+    return 0;
+#else
+    return 77;
+#endif
+}

src/asm/test/atomic_spinlock.c

@@ -0,0 +1,106 @@
+/*
+ * Copyright (c) 2004-2005 The Trustees of Indiana University.
+ *                         All rights reserved.
+ * Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
+ *                         All rights reserved.
+ * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, 
+ *                         University of Stuttgart.  All rights reserved.
+ * $COPYRIGHT$
+ * 
+ * Additional copyrights may follow
+ * 
+ * $HEADER$
+ */
+
+#undef OMPI_BUILDING
+#include "ompi_config.h"
+
+#include <assert.h>
+#ifdef HAVE_PTHREAD_H
+#include <pthread.h>
+#endif
+#include <stdlib.h>
+#include <stdio.h>
+
+#include "include/sys/atomic.h"
+#include "atomic_test.h"
+
+int atomic_verbose = 0;
+
+struct start_info {
+    int tid;
+    int count;
+    ompi_lock_t *lock;
+};
+
+static void* atomic_spinlock_start(void* arg)
+{
+    struct start_info *data = (struct start_info*) arg;
+
+    return (void*) atomic_spinlock_test(data->lock, data->count,
+                                        data->tid);
+}
+
+
+int
+atomic_spinlock_test_th(ompi_lock_t *lock, int count, int id, int thr_count)
+{
+#if HAVE_PTHREAD_H
+    pthread_t *th;
+    int tid, ret = 0;
+    struct start_info *data;
+
+    th = (pthread_t *) malloc(thr_count * sizeof(pthread_t));
+    if (!th) { perror("malloc"); exit(EXIT_FAILURE); }
+    data = (struct start_info *) malloc(thr_count * sizeof(struct start_info));
+    if (!th) { perror("malloc"); exit(EXIT_FAILURE); }
+
+    for (tid = 0; tid < thr_count; tid++) {
+        data[tid].tid = tid;
+        data[tid].count = count;
+        data[tid].lock = lock;
+
+        if (pthread_create(&th[tid], NULL, atomic_spinlock_start, (void *) &(data[tid])) != 0) {
+            perror("pthread_create");
+            exit(EXIT_FAILURE);
+        }
+    }
+
+    /* -- wait for the thread set to finish -- */
+    for (tid = 0; tid < thr_count; tid++) {
+        void *thread_return;
+
+        if (pthread_join(th[tid], &thread_return) != 0) {
+            perror("pthread_join");
+            exit(EXIT_FAILURE);
+        }
+
+        ret += (int) thread_return;
+    }
+    free(data);
+    free(th);
+
+    return ret;
+#else
+    return 77;
+#endif
+}
+
+
+int
+atomic_spinlock_test(ompi_lock_t *lock, int count, int id)
+{
+    int i;
+
+    for (i = 0 ; i < count ; ++i) {
+        ompi_atomic_lock(lock);
+        if (atomic_verbose) { printf("id %03d has the lock (lock)\n", id); }
+        ompi_atomic_unlock(lock);
+
+        while (ompi_atomic_trylock(lock)) { ; }
+        if (atomic_verbose) { printf("id %03d has the lock (trylock)\n", id); }
+        ompi_atomic_unlock(lock);
+    }
+
+    return 0;
+}

src/asm/test/atomic_spinlock_2.c

@@ -0,0 +1,33 @@
+/*
+ * Copyright (c) 2004-2005 The Trustees of Indiana University.
+ *                         All rights reserved.
+ * Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
+ *                         All rights reserved.
+ * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, 
+ *                         University of Stuttgart.  All rights reserved.
+ * $COPYRIGHT$
+ * 
+ * Additional copyrights may follow
+ * 
+ * $HEADER$
+ */
+
+#undef OMPI_BUILDING
+#include "ompi_config.h"
+
+#include <assert.h>
+
+#include "include/sys/atomic.h"
+#include "atomic_test.h"
+
+int
+main(int argc, char *argv[])
+{
+    int ret = 77;
+    ompi_lock_t lock;
+
+    ompi_atomic_init(&lock, OMPI_ATOMIC_UNLOCKED);
+    ret = atomic_spinlock_test_th(&lock, TEST_REPS, 0, 2);
+
+    return ret;
+}

src/asm/test/atomic_spinlock_5.c

@@ -0,0 +1,33 @@
+/*
+ * Copyright (c) 2004-2005 The Trustees of Indiana University.
+ *                         All rights reserved.
+ * Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
+ *                         All rights reserved.
+ * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, 
+ *                         University of Stuttgart.  All rights reserved.
+ * $COPYRIGHT$
+ * 
+ * Additional copyrights may follow
+ * 
+ * $HEADER$
+ */
+
+#undef OMPI_BUILDING
+#include "ompi_config.h"
+
+#include <assert.h>
+
+#include "include/sys/atomic.h"
+#include "atomic_test.h"
+
+int
+main(int argc, char *argv[])
+{
+    int ret = 77;
+    ompi_lock_t lock;
+
+    ompi_atomic_init(&lock, OMPI_ATOMIC_UNLOCKED);
+    ret = atomic_spinlock_test_th(&lock, TEST_REPS, 0, 5);
+
+    return ret;
+}

src/asm/test/atomic_spinlock_8.c

@@ -0,0 +1,33 @@
+/*
+ * Copyright (c) 2004-2005 The Trustees of Indiana University.
+ *                         All rights reserved.
+ * Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
+ *                         All rights reserved.
+ * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, 
+ *                         University of Stuttgart.  All rights reserved.
+ * $COPYRIGHT$
+ * 
+ * Additional copyrights may follow
+ * 
+ * $HEADER$
+ */
+
+#undef OMPI_BUILDING
+#include "ompi_config.h"
+
+#include <assert.h>
+
+#include "include/sys/atomic.h"
+#include "atomic_test.h"
+
+int
+main(int argc, char *argv[])
+{
+    int ret = 77;
+    ompi_lock_t lock;
+
+    ompi_atomic_init(&lock, OMPI_ATOMIC_UNLOCKED);
+    ret = atomic_spinlock_test_th(&lock, TEST_REPS, 0, 8);
+
+    return ret;
+}

src/asm/test/atomic_spinlock_serial.c

@@ -0,0 +1,33 @@
+/*
+ * Copyright (c) 2004-2005 The Trustees of Indiana University.
+ *                         All rights reserved.
+ * Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
+ *                         All rights reserved.
+ * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, 
+ *                         University of Stuttgart.  All rights reserved.
+ * $COPYRIGHT$
+ * 
+ * Additional copyrights may follow
+ * 
+ * $HEADER$
+ */
+
+#undef OMPI_BUILDING
+#include "ompi_config.h"
+
+#include <assert.h>
+
+#include "include/sys/atomic.h"
+#include "atomic_test.h"
+
+int
+main(int argc, char *argv[])
+{
+    int ret;
+    ompi_lock_t lock;
+
+    ompi_atomic_init(&lock, OMPI_ATOMIC_UNLOCKED);
+    ret = atomic_spinlock_test(&lock, TEST_REPS, 0);
+
+    return ret;
+}

src/include/sys/atomic.h

@@ -27,10 +27,10 @@
  * The following #defines will be true / false based on
  * assembly support:
  *
- *  \c OMPI_HAVE_MEM_BARRIER atomic memory barriers
- *  \c OMPI_HAVE_ATOMIC_SPINLOCKS atomic spinlocks
- *  \c OMPI_HAVE_ATOMIC_MATH_32 if 32 bit add/sub/cmpset can be done "atomicly"
- *  \c OMPI_HAVE_ATOMIC_MATH_64 if 32 bit add/sub/cmpset can be done "atomicly"
+ *  - \c OMPI_HAVE_MEM_BARRIER atomic memory barriers
+ *  - \c OMPI_HAVE_ATOMIC_SPINLOCKS atomic spinlocks
+ *  - \c OMPI_HAVE_ATOMIC_MATH_32 if 32 bit add/sub/cmpset can be done "atomicly"
+ *  - \c OMPI_HAVE_ATOMIC_MATH_64 if 32 bit add/sub/cmpset can be done "atomicly"
  *
  * Note that for the Atomic math, atomic add/sub may be implemented as
  * C code using ompi_atomic_cmpset.  The appearance of atomic
@@ -82,6 +82,7 @@ extern "C" {
 #include "include/sys/sparc64/atomic.h"
 #endif
 
+#ifndef DOXYGEN
 /* compare and set operations can't really be emulated from software,
    so if these defines aren't already set, they should be set to 0
    now */
@@ -91,6 +92,7 @@ extern "C" {
 #ifndef OMPI_HAVE_ATOMIC_CMPSET_64
 #define OMPI_HAVE_ATOMIC_CMPSET_64 0
 #endif
+#endif /* DOXYGEN */
 
 /**********************************************************************
  *
@@ -98,7 +100,7 @@ extern "C" {
  *                   but can't inline
  *
  *********************************************************************/
-#ifndef OMPI_HAVE_ATOMIC_MEM_BARRIER
+#if !defined(OMPI_HAVE_ATOMIC_MEM_BARRIER) && !defined(DOXYGEN)
 /* no way to emulate in C code */
 #define OMPI_HAVE_ATOMIC_MEM_BARRIER 0
 #endif
@@ -106,16 +108,38 @@ extern "C" {
 #if defined(DOXYGEN) || OMPI_HAVE_ATOMIC_MEM_BARRIER
 /**
  * Memory barrier
+ *
+ * Will use system-specific features to instruct the processor and
+ * memory controller that all writes and reads that have been posted
+ * before the call to \c ompi_atomic_mb() must appear to have
+ * completed before the next read or write.
+ *
+ * \note This can have some expensive side effects, including flushing
+ * the pipeline, preventing the cpu from reordering instructions, and
+ * generally grinding the memory controller's performance.  Use only
+ * if you need *both* read and write barriers.
  */
 void ompi_atomic_mb(void);
 
 /**
  * Read memory barrier
+ *
+ * Use system-specific features to instruct the processor and memory
+ * conrtoller that all reads that have been posted before the call to
+ * \c ompi_atomic_rmb() must appear to have been completed before the
+ * next read.  Nothing is said about the ordering of writes when using
+ * \c ompi_atomic_rmb().
  */
 void ompi_atomic_rmb(void);
 
 /**
  * Write memory barrier.
+ *
+ * Use system-specific features to instruct the processor and memory
+ * conrtoller that all writes that have been posted before the call to
+ * \c ompi_atomic_rmb() must appear to have been completed before the
+ * next write.  Nothing is said about the ordering of reads when using
+ * \c ompi_atomic_rmb().
  */
 void ompi_atomic_wmb(void);
 
@@ -138,13 +162,12 @@ struct ompi_lock_t {
 };
 typedef struct ompi_lock_t ompi_lock_t;
 
-#ifndef OMPI_HAVE_ATOMIC_SPINLOCKS
+#if !defined(OMPI_HAVE_ATOMIC_SPINLOCKS) && !defined(DOXYGEN)
 #define OMPI_HAVE_ATOMIC_SPINLOCKS (OMPI_HAVE_ATOMIC_CMPSET_32 || OMPI_HAVE_ATOMIC_CMPSET_64)
 #endif
 
 #if defined(DOXYGEN) || OMPI_HAVE_ATOMIC_SPINLOCKS
 
-
 /**
  * Enumeration of lock states
  */
@@ -195,7 +218,7 @@ static inline void ompi_atomic_unlock(ompi_lock_t *lock);
  * Atomic math operations
  *
  *********************************************************************/
-#ifndef OMPI_HAVE_ATOMIC_CMPSET_32
+#if !defined(OMPI_HAVE_ATOMIC_CMPSET_32) && !defined(DOXYGEN)
 #define OMPI_HAVE_ATOMIC_CMPSET_32 0
 #endif
 #if defined(DOXYGEN) || OMPI_HAVE_ATOMIC_CMPSET_32
@@ -208,7 +231,7 @@ int ompi_atomic_cmpset_rel_32(volatile int32_t *addr, int32_t oldval,
 #endif
 
 
-#ifndef OMPI_HAVE_ATOMIC_CMPSET_64
+#if !defined(OMPI_HAVE_ATOMIC_CMPSET_64) && !defined(DOXYGEN)
 #define OMPI_HAVE_ATOMIC_CMPSET_64 0
 #endif
 #if defined(DOXYGEN) || OMPI_HAVE_ATOMIC_CMPSET_64
@@ -220,7 +243,7 @@ int ompi_atomic_cmpset_rel_64(volatile int64_t *addr, int64_t oldval,
                               int64_t newval);
 #endif
 
-#ifndef OMPI_HAVE_ATOMIC_MATH_32
+#if !defined(OMPI_HAVE_ATOMIC_MATH_32) && !defined(DOXYGEN)
 /* define to 0 for these tests.  WIll fix up later. */
 #define OMPI_HAVE_ATOMIC_MATH_32 0
 #endif
@@ -277,11 +300,26 @@ static inline void ompi_atomic_add_xx(volatile void* addr,
 static inline void ompi_atomic_sub_xx(volatile void* addr, 
                                       int32_t value, size_t length);
 
+static inline int ompi_atomic_cmpset_ptr(volatile void* addr, 
+                                         void* oldval, 
+                                         void* newval);
+static inline int ompi_atomic_cmpset_acq_ptr(volatile void* addr, 
+                                             void* oldval, 
+                                             void* newval);
+static inline int ompi_atomic_cmpset_rel_ptr(volatile void* addr, 
+                                             void* oldval, 
+                                             void* newval);
+static inline int ompi_atomic_add_pt(volatile void* addr, 
+                                            void* delta);
+static inline int ompi_atomic_sub_ptr(volatile void* addr, 
+                                             void* delta);
 
 /**
  * Atomic compare and set of pointer with relaxed semantics. This
- * macro detect at compile time the type of the first argument 
- * and choose the correct function to be called.
+ * macro detect at compile time the type of the first argument and
+ * choose the correct function to be called.  
+ *
+ * \note This macro should only be used for integer types.
  *
  * @param addr          Address of <TYPE>.
  * @param oldval        Comparison value <TYPE>.
@@ -293,12 +331,13 @@ static inline void ompi_atomic_sub_xx(volatile void* addr,
    ompi_atomic_cmpset_xx( (volatile void*)(ADDR), (int64_t)(OLDVAL), \
                           (int64_t)(NEWVAL), sizeof(*(ADDR)) )
 
-
 /**
  * Atomic compare and set of pointer with acquire semantics. This
  * macro detect at compile time the type of the first argument 
  * and choose the correct function to be called.
  *
+ * \note This macro should only be used for integer types.
+ *
  * @param addr          Address of <TYPE>.
  * @param oldval        Comparison value <TYPE>.
  * @param newval        New value to set if comparision is true <TYPE>.
@@ -315,6 +354,8 @@ static inline void ompi_atomic_sub_xx(volatile void* addr,
  * macro detect at compile time the type of the first argument 
  * and choose the correct function to b
  *
+ * \note This macro should only be used for integer types.
+ *
  * @param addr          Address of <TYPE>.
  * @param oldval        Comparison value <TYPE>.
  * @param newval        New value to set if comparision is true <TYPE>.
@@ -331,6 +372,8 @@ static inline void ompi_atomic_sub_xx(volatile void* addr,
  * macro detect at compile time the type of the first argument 
  * and choose the correct function to be called.
  *
+ * \note This macro should only be used for integer types.
+ *
  * @param addr          Address of <TYPE>
  * @param delta         Value to add (converted to <TYPE>).
  */
@@ -343,6 +386,8 @@ static inline void ompi_atomic_sub_xx(volatile void* addr,
  * macro detect at compile time the type of the first argument 
  * and choose the correct function to be called.
  *
+ * \note This macro should only be used for integer types.
+ *
  * @param addr          Address of <TYPE>
  * @param delta         Value to substract (converted to <TYPE>).
  */
@@ -366,4 +411,7 @@ static inline void ompi_atomic_sub_xx(volatile void* addr,
 }
 #endif
 
+
+
+
 #endif /* OMPI_SYS_ATOMIC_H */

src/include/sys/atomic_impl.h

@@ -221,6 +221,87 @@ ompi_atomic_sub_xx(volatile void* addr, int32_t value, size_t length)
 #endif /* (OMPI_HAVE_ATOMIC_CMPSET_32 || OMPI_HAVE_ATOMIC_CMPSET_64) */
 
 
+static inline int
+ompi_atomic_cmpset_ptr(volatile void* addr, 
+                       void* oldval, 
+                       void* newval)
+{
+#if SIZEOF_VOID_P == 4 && OMPI_HAVE_ATOMIC_CMPSET_32
+    return ompi_atomic_cmpset_32((int32_t*) addr, (unsigned long) oldval, 
+                                 (unsigned long) newval);
+#elif SIZEOF_VOID_P == 8 && OMPI_HAVE_ATOMIC_CMPSET_64
+    return ompi_atomic_cmpset_64((int64_t*) addr, (unsigned long) oldval, 
+                                 (unsigned long) newval);
+#else
+    abort();
+    return 0;
+#endif
+}
+
+
+static inline int
+ompi_atomic_cmpset_acq_ptr(volatile void* addr, 
+                           void* oldval, 
+                           void* newval)
+{
+#if SIZEOF_VOID_P == 4 && OMPI_HAVE_ATOMIC_CMPSET_32
+    return ompi_atomic_cmpset_acq_32((int32_t*) addr, (unsigned long) oldval, 
+                                     (unsigned long) newval);
+#elif SIZEOF_VOID_P == 8 && OMPI_HAVE_ATOMIC_CMPSET_64
+    return ompi_atomic_cmpset_acq_64((int64_t*) addr, (unsigned long) oldval, 
+                                     (unsigned long) newval);
+#else
+    abort();
+    return 0;
+#endif
+}
+
+
+static inline int ompi_atomic_cmpset_rel_ptr(volatile void* addr, 
+                                             void* oldval, 
+                                             void* newval)
+{
+#if SIZEOF_VOID_P == 4 && OMPI_HAVE_ATOMIC_CMPSET_32
+    return ompi_atomic_cmpset_rel_32((int32_t*) addr, (unsigned long) oldval, 
+                                     (unsigned long) newval);
+#elif SIZEOF_VOID_P == 8 && OMPI_HAVE_ATOMIC_CMPSET_64
+    return ompi_atomic_cmpset_rel_64((int64_t*) addr, (unsigned long) oldval, 
+                                     (unsigned long) newval);
+#else
+    abort();
+    return 0;
+#endif
+}
+
+
+static inline int ompi_atomic_add_pt(volatile void* addr, 
+                                            void* delta)
+{
+#if SIZEOF_VOID_P == 4 && OMPI_HAVE_ATOMIC_CMPSET_32
+    return ompi_atomic_add_32((int32_t*) addr, (unsigned long) delta);
+#elif SIZEOF_VOID_P == 8 && OMPI_HAVE_ATOMIC_CMPSET_64
+    return ompi_atomic_add_64((int64_t*) addr, (unsigned long) delta);
+#else
+    abort();
+    return 0;
+#endif
+}
+
+
+static inline int ompi_atomic_sub_ptr(volatile void* addr, 
+                                             void* delta)
+{
+#if SIZEOF_VOID_P == 4 && OMPI_HAVE_ATOMIC_CMPSET_32
+    return ompi_atomic_sub_32((int32_t*) addr, (unsigned long) delta);
+#elif SIZEOF_VOID_P == 8 && OMPI_HAVE_ATOMIC_CMPSET_64
+    return ompi_atomic_sub_64((int64_t*) addr, (unsigned long) delta);
+#else
+    abort();
+    return 0;
+#endif
+}
+
+
 /**********************************************************************
  *
  * Atomic spinlocks