Merge pull request #163 from shintaro-iwasaki/FixMiscBugs

Fix miscellaneous bugs

configure.ac

@@ -415,9 +415,14 @@ fi
 
 # --enable-perf-opt: set CFLAGS for the performance optimization
 if test "x$enable_perf_opt" = "xyes"; then
-    CFLAGS="$CFLAGS -O3 -ftls-model=initial-exec"
-    CXXFLAGS="$CXXFLAGS -O3 -ftls-model=initial-exec"
-    CCASFLAGS="$CCASFLAGS -O3 -ftls-model=initial-exec"
+    CFLAGS="$CFLAGS -O3"
+    CXXFLAGS="$CXXFLAGS -O3"
+    CCASFLAGS="$CCASFLAGS -O3"
+    if test "$CC" != "xlc"; then
+        CFLAGS="$CFLAGS -ftls-model=initial-exec"
+        CXXFLAGS="$CXXFLAGS -ftls-model=initial-exec"
+        CCASFLAGS="$CCASFLAGS -ftls-model=initial-exec"
+    fi
     if test "$CC" = "icc"; then
         CFLAGS="$CFLAGS -ipo"
         CXXFLAGS="$CFLAGS -ipo"
@@ -628,7 +633,9 @@ AC_TRY_COMPILE(
 [have_atomic_builtin=yes],
 [have_atomic_builtin=no]
 )
-if test "x$have_atomic_builtin" = "xyes" ; then
+# __atomic is broken in XLC-16.1.1, so use __sync as a fallback
+# See https://github.com/pmodels/argobots/issues/162 for details.
+if test "$CC" != "xlc" -a "x$have_atomic_builtin" = "xyes" ; then
     AC_DEFINE(ABT_CONFIG_HAVE_ATOMIC_BUILTIN, 1,
               [Define if __atomic builtins are supported])
 fi

src/arch/fcontext/jump_ppc64_sysv_elf_gas.S

@@ -282,7 +282,7 @@ jump_fcontext:
     mtfsf  0xff, %f0  # restore FPSCR
 #ifdef __VSX__
     li    %r10, 144
-    lvx   %v19, %r10, %r1 # load VSCR vector
+    lvx   %v19, %r10, %r1 # load VSCR
     mtvscr %v19           # restore VSCR.  Only the last 32 bits are used
     li    %r14, 160
     lvx  %v20, %r14, %r1  # restore V20

src/arch/fcontext/take_ppc64_sysv_elf_gas.S

@@ -187,7 +187,7 @@ take_fcontext:
     mtfsf  0xff, %f0  # restore FPSCR
 #ifdef __VSX__
     li    %r10, 144
-    lvx   %v19, %r10, %r1 # load VSCR vector
+    lvx   %v19, %r10, %r1 # load VSCR
     mtvscr %v19           # restore VSCR.  Only the last 32 bits are used
     li    %r14, 160
     lvx  %v20, %r14, %r1  # restore V20

src/include/abti_thread.h

@@ -244,6 +244,7 @@ static inline void ABTI_thread_context_switch_sched_to_thread_internal(
                 }
             }
         }
+        ABTI_LOG_SET_SCHED(p_old);
         return;
     }
 #endif

src/thread.c

@@ -1828,7 +1828,7 @@ int ABTI_thread_set_ready(ABTI_local *p_local, ABTI_thread *p_thread)
      * scheduled by another scheduler if it is pushed to a shared pool. */
     while (ABTD_atomic_acquire_load_uint32(&p_thread->request) &
            ABTI_THREAD_REQ_BLOCK)
-        ;
+        ABTD_atomic_pause();
 
     LOG_EVENT("[U%" PRIu64 ":E%d] set ready\n", ABTI_thread_get_id(p_thread),
               p_thread->p_last_xstream->rank);