Re: [RFC][BZ #16549, #16410] Remove pthread_(cond)wait assembly implementations?

[OndÅej BÃlka <neleai at seznam dot cz>]

On Tue, Feb 11, 2014 at 09:27:20AM -0800, Roland McGrath wrote:
> These were written in assembly to begin with because C versions did not
> perform adequately.  It's entirely possible the compiler has improved
> enough to change that calculus.  But actual performance analysis is
> required.  

I used following benchmark for timedwait and deleting assembly
implementation actually improves performance by around 5000 cycles.

old one:

cond wait spend time: 0.000067360
cond wait spend time: 0.000064957
cond wait spend time: 0.000055014
cond wait spend time: 0.000063775
cond wait spend time: 0.000064073
cond wait spend time: 0.000063578
cond wait spend time: 0.000063339
cond wait spend time: 0.000064064
cond wait spend time: 0.000063312
cond wait spend time: 0.000063331

new one:

cond wait spend time: 0.000064140
cond wait spend time: 0.000062096
cond wait spend time: 0.000061748
cond wait spend time: 0.000061601
cond wait spend time: 0.000061581
cond wait spend time: 0.000061620
cond wait spend time: 0.000061725
cond wait spend time: 0.000061611
cond wait spend time: 0.000061601
cond wait spend time: 0.000061591



diff --git a/benchtests/Makefile b/benchtests/Makefile
index 8bfb039..2736f58 100644
--- a/benchtests/Makefile
+++ b/benchtests/Makefile
@@ -33,8 +33,12 @@ string-bench-all := $(string-bench)
 
 stdlib-bench := strtod
 
-benchset := $(string-bench-all) $(stdlib-bench)
+nptl-bench := pthread_cond_timedwait
 
+
+benchset := $(nptl-bench) $(string-bench-all) $(stdlib-bench)
+
+LDLIBS-bench-pthread_cond_timedwait = -lpthread
 LDLIBS-bench-acos = -lm
 LDLIBS-bench-acosh = -lm
 LDLIBS-bench-asin = -lm
diff --git a/benchtests/bench-pthread_cond_timedwait.c b/benchtests/bench-pthread_cond_timedwait.c
new file mode 100644
index 0000000..ec21893
--- /dev/null
+++ b/benchtests/bench-pthread_cond_timedwait.c
@@ -0,0 +1,42 @@
+#include <stdio.h>
+#include <pthread.h>
+
+int
+main ()
+{
+  struct timespec t1, t2, t3;
+  pthread_mutex_t mutex;
+  pthread_cond_t cond = PTHREAD_COND_INITIALIZER;
+  pthread_mutex_init (&mutex, NULL);
+
+
+  for (int i = 0; i < 10; i++)
+    {
+      clock_gettime (CLOCK_REALTIME, &t1);
+      t3 = t1;
+      t3.tv_nsec += 10000;
+      pthread_mutex_lock (&mutex);
+      pthread_cond_timedwait (&cond, &mutex, &t3);
+      pthread_mutex_unlock (&mutex);
+
+      clock_gettime (CLOCK_REALTIME, &t2);
+
+      if (t2.tv_nsec >= t1.tv_nsec)
+	{
+	  t2.tv_sec -= t1.tv_sec;
+	  t2.tv_nsec -= t1.tv_nsec;
+	}
+      else
+	{
+	  t2.tv_sec = t2.tv_sec - t1.tv_sec - 1;
+	  t2.tv_nsec = t2.tv_nsec + (1000000000 - t1.tv_nsec);
+	}
+
+      printf ("cond wait spend time: %d.%.9d\n", t2.tv_sec, t2.tv_nsec);
+    }
+  pthread_mutex_destroy (&mutex);
+  pthread_cond_destroy (&cond);
+
+
+  return 0;
+}