X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=rts%2Fposix%2FGetTime.c;h=34fcd5c2a7b462b99baf43d6b4d85f2ab9514948;hb=4c6a3f787abcaed009a574196d82237d9ae64fc8;hp=89d83a31cb35102a701f3519260daaa7fa4c990b;hpb=ab1e183aa7801f9260a9cbb6edbd553cf8249b25;p=ghc-hetmet.git

diff --git a/rts/posix/GetTime.c b/rts/posix/GetTime.c
index 89d83a3..34fcd5c 100644
--- a/rts/posix/GetTime.c
+++ b/rts/posix/GetTime.c
@@ -32,6 +32,10 @@
 # include <sys/times.h>
 #endif
 
+#ifdef USE_PAPI
+# include <papi.h>
+#endif
+
 #if ! ((defined(HAVE_GETRUSAGE) && !irix_HOST_OS) || defined(HAVE_TIMES))
 #error No implementation for getProcessCPUTime() available.
 #endif
@@ -44,7 +48,7 @@ Ticks getProcessCPUTime(void)
 {
     struct rusage t;
     getrusage(RUSAGE_SELF, &t);
-    return (t.ru_utime.tv_sec * TICKS_PER_SECOND + 
+    return ((Ticks)t.ru_utime.tv_sec * TICKS_PER_SECOND +
 	    ((Ticks)t.ru_utime.tv_usec * TICKS_PER_SECOND)/1000000);
 }
 
@@ -52,7 +56,7 @@ Ticks getProcessElapsedTime(void)
 {
     struct timeval tv;
     gettimeofday(&tv, (struct timezone *) NULL);
-    return (tv.tv_sec * TICKS_PER_SECOND +
+    return ((Ticks)tv.tv_sec * TICKS_PER_SECOND +
 	    ((Ticks)tv.tv_usec * TICKS_PER_SECOND)/1000000);
 }
 
@@ -68,9 +72,17 @@ void getProcessTimes(Ticks *user, Ticks *elapsed)
 
 Ticks getProcessCPUTime(void)
 {
+#if !defined(THREADED_RTS) && USE_PAPI
+    long long usec;
+    if ((usec = PAPI_get_virt_usec()) < 0) {
+	barf("PAPI_get_virt_usec: %lld", usec);
+    }
+    return ((usec * TICKS_PER_SECOND) / 1000000);
+#else
     Ticks user, elapsed;
     getProcessTimes(&user,&elapsed);
     return user;
+#endif
 }
 
 Ticks getProcessElapsedTime(void)
@@ -115,16 +127,27 @@ void getProcessTimes(Ticks *user, Ticks *elapsed)
 
 Ticks getThreadCPUTime(void)
 {
-#if defined(HAVE_CLOCK_GETTIME) && defined(CLOCK_THREAD_CPUTIME_ID)
-    // clock_gettime() gives us per-thread CPU time.  It isn't
-    // reliable on Linux, but it's the best we have.
-    struct timespec ts;
-    clock_gettime(CLOCK_THREAD_CPUTIME_ID, &ts);
-    return (ts.tv_sec * TICKS_PER_SECOND + 
-	    ((Ticks)ts.tv_nsec * TICKS_PER_SECOND) / 1000000000);
-#else
-    return getProcessCPUTime();
+#if USE_PAPI
+    long long usec;
+    if ((usec = PAPI_get_virt_usec()) < 0) {
+	barf("PAPI_get_virt_usec: %lld", usec);
+    }
+    return ((usec * TICKS_PER_SECOND) / 1000000);
+
+#elif !defined(BE_CONSERVATIVE) && defined(HAVE_CLOCK_GETTIME) && defined (_POSIX_THREAD_CPUTIME) && defined(CLOCK_THREAD_CPUTIME_ID) && defined(HAVE_SYSCONF)
+    if (sysconf(_POSIX_THREAD_CPUTIME) != -1) {
+        // clock_gettime() gives us per-thread CPU time.  It isn't
+        // reliable on Linux, but it's the best we have.
+        struct timespec ts;
+        int res;
+        res = clock_gettime(CLOCK_THREAD_CPUTIME_ID, &ts);
+        if (res == 0) {
+            return ((Ticks)ts.tv_sec * TICKS_PER_SECOND +
+                    ((Ticks)ts.tv_nsec * TICKS_PER_SECOND) / 1000000000);
+        }
+    }
 #endif
+    return getProcessCPUTime();
 }
 
 nat