Various upgrades, with the effect that mozilla now runs, although
it has tremendous performance problems.

* Implement pthread_key_{create,delete} and pthread_{set,get}specific.

* Implement pthread_cond_timedwait.  A nuisance.

* New timer infrastructure, based on the RDTSC instruction.  This
  allows fast, accurate time measurement without swamping the host with
  gettimeofday() syscalls.

There's something definitely screwy about the scheduler, making opera
run slowly and mozilla run unbelievably slowly.  To be investigated.


git-svn-id: svn://svn.valgrind.org/valgrind/trunk@119 a5019735-40e9-0310-863c-91ae7b9d1cf9
diff --git a/coregrind/vg_libpthread.c b/coregrind/vg_libpthread.c
index fb70ec6..76923a3 100644
--- a/coregrind/vg_libpthread.c
+++ b/coregrind/vg_libpthread.c
@@ -134,6 +134,8 @@
 #include <pthread.h>
 #include <stdio.h>
 #include <errno.h>
+#include <assert.h>
+#include <sys/time.h> /* gettimeofday */
 
 /* ---------------------------------------------------
    THREAD ATTRIBUTES
@@ -421,6 +423,40 @@
    return res;
 }
 
+int pthread_cond_timedwait ( pthread_cond_t *cond, 
+                             pthread_mutex_t *mutex, 
+                             const struct  timespec *abstime )
+{
+   int res;
+   unsigned int ms_now, ms_end;
+   struct  timeval timeval_now;
+   unsigned long long int ull_ms_now_after_1970;
+   unsigned long long int ull_ms_end_after_1970;
+
+   ensure_valgrind("pthread_cond_timedwait");
+   VALGRIND_MAGIC_SEQUENCE(ms_now, 0xFFFFFFFF /* default */,
+                           VG_USERREQ__READ_MILLISECOND_TIMER,
+                           0, 0, 0, 0);
+   assert(ms_now != 0xFFFFFFFF);
+   res = gettimeofday(&timeval_now, NULL);
+   assert(res == 0);
+
+   ull_ms_now_after_1970 
+      = 1000ULL * ((unsigned long long int)(timeval_now.tv_sec))
+        + ((unsigned long long int)(timeval_now.tv_usec / 1000000));
+   ull_ms_end_after_1970
+      = 1000ULL * ((unsigned long long int)(abstime->tv_sec))
+        + ((unsigned long long int)(abstime->tv_nsec / 1000000));
+   assert(ull_ms_end_after_1970 >= ull_ms_now_after_1970);
+   ms_end 
+      = ms_now + (unsigned int)(ull_ms_end_after_1970 - ull_ms_now_after_1970);
+   VALGRIND_MAGIC_SEQUENCE(res, 0 /* default */,
+                           VG_USERREQ__PTHREAD_COND_TIMEDWAIT,
+			   cond, mutex, ms_end, 0);
+   return res;
+}
+
+
 int pthread_cond_signal(pthread_cond_t *cond)
 {
    int res;
@@ -471,8 +507,12 @@
 int pthread_key_create(pthread_key_t *key,  
                        void  (*destr_function)  (void *))
 {
-   ignored("pthread_key_create");
-   return 0;
+   int res;
+   ensure_valgrind("pthread_key_create");
+   VALGRIND_MAGIC_SEQUENCE(res, 0 /* default */,
+                           VG_USERREQ__PTHREAD_KEY_CREATE,
+                           key, destr_function, 0, 0);
+   return res;
 }
 
 int pthread_key_delete(pthread_key_t key)
@@ -483,14 +523,22 @@
 
 int pthread_setspecific(pthread_key_t key, const void *pointer)
 {
-   ignored("pthread_setspecific");
-   return 0;
+   int res;
+   ensure_valgrind("pthread_setspecific");
+   VALGRIND_MAGIC_SEQUENCE(res, 0 /* default */,
+                           VG_USERREQ__PTHREAD_SETSPECIFIC,
+                           key, pointer, 0, 0);
+   return res;
 }
 
 void * pthread_getspecific(pthread_key_t key)
 {
-   ignored("pthread_setspecific");
-   return NULL;
+   int res;
+   ensure_valgrind("pthread_getspecific");
+   VALGRIND_MAGIC_SEQUENCE(res, 0 /* default */,
+                           VG_USERREQ__PTHREAD_GETSPECIFIC,
+                           key, 0 , 0, 0);
+   return (void*)res;
 }
 
 
@@ -784,7 +832,6 @@
    * (unchecked) libc error numbers (EINTR etc) are the negation of the
      kernel's error numbers (VKI_EINTR etc).
 */
-#include <assert.h>
 
 
 int select ( int n, 
@@ -793,16 +840,19 @@
              fd_set *xfds, 
              struct timeval *timeout )
 {
+   unsigned int ms_now, ms_end;
    int    res;
    fd_set rfds_copy;
    fd_set wfds_copy;
    fd_set xfds_copy;
    struct vki_timeval  t_now;
-   struct vki_timeval  t_end;
    struct vki_timeval  zero_timeout;
    struct vki_timespec nanosleep_interval;
 
-   ensure_valgrind("select");
+   /* gcc's complains about ms_end being used uninitialised -- classic
+      case it can't understand, where ms_end is both defined and used
+      only if timeout != NULL.  Hence ... */
+   ms_end = 0;
 
    /* We assume that the kernel and libc data layouts are identical
       for the following types.  These asserts provide a crude
@@ -811,8 +861,17 @@
        || sizeof(struct timeval) != sizeof(struct vki_timeval))
       barf("valgrind's hacky non-blocking select(): data sizes error");
 
-   /* If a zero timeout specified, this call is harmless. */
-   if (timeout && timeout->tv_sec == 0 && timeout->tv_usec == 0) {
+   /* Detect the current time and simultaneously find out if we are
+      running on Valgrind. */
+   VALGRIND_MAGIC_SEQUENCE(ms_now, 0xFFFFFFFF /* default */,
+                           VG_USERREQ__READ_MILLISECOND_TIMER,
+                           0, 0, 0, 0);
+
+   /* If a zero timeout specified, this call is harmless.  Also go
+      this route if we're not running on Valgrind, for whatever
+      reason. */
+   if ( (timeout && timeout->tv_sec == 0 && timeout->tv_usec == 0)
+        || (ms_now == 0xFFFFFFFF) ) {
       res = do_syscall_select( n, (vki_fd_set*)rfds, 
                                    (vki_fd_set*)wfds, 
                                    (vki_fd_set*)xfds, 
@@ -825,35 +884,29 @@
       }
    }
 
-   /* If a timeout was specified, set t_end to be the end wallclock
-      time. */
+   /* If a timeout was specified, set ms_end to be the end millisecond
+      counter [wallclock] time. */
    if (timeout) {
       res = my_do_syscall2(__NR_gettimeofday, (int)&t_now, (int)NULL);
       assert(res == 0);
-      t_end = t_now;
-      t_end.tv_usec += timeout->tv_usec;
-      t_end.tv_sec  += timeout->tv_sec;
-      if (t_end.tv_usec >= 1000000) {
-         t_end.tv_usec -= 1000000;
-         t_end.tv_sec += 1;
-      }
+      ms_end = ms_now;
+      ms_end += (timeout->tv_usec / 1000);
+      ms_end += (timeout->tv_sec * 1000);
       /* Stay sane ... */
-      assert (t_end.tv_sec > t_now.tv_sec
-              || (t_end.tv_sec == t_now.tv_sec 
-                  && t_end.tv_usec >= t_now.tv_usec));
+      assert (ms_end >= ms_now);
    }
 
    /* fprintf(stderr, "MY_SELECT: before loop\n"); */
 
    /* Either timeout == NULL, meaning wait indefinitely, or timeout !=
-      NULL, in which case t_end holds the end time. */
+      NULL, in which case ms_end holds the end time. */
    while (1) {
       if (timeout) {
-         res = my_do_syscall2(__NR_gettimeofday, (int)&t_now, (int)NULL);
-         assert(res == 0);
-         if (t_now.tv_sec > t_end.tv_sec
-             || (t_now.tv_sec == t_end.tv_sec 
-                 && t_now.tv_usec > t_end.tv_usec)) {
+         VALGRIND_MAGIC_SEQUENCE(ms_now, 0xFFFFFFFF /* default */,
+                                 VG_USERREQ__READ_MILLISECOND_TIMER,
+                                 0, 0, 0, 0);
+         assert(ms_now != 0xFFFFFFFF);
+         if (ms_now >= ms_end) {
             /* timeout; nothing interesting happened. */
             if (rfds) FD_ZERO(rfds);
             if (wfds) FD_ZERO(wfds);
@@ -892,7 +945,7 @@
       /* fprintf(stderr, "MY_SELECT: nanosleep\n"); */
       /* nanosleep and go round again */
       nanosleep_interval.tv_sec  = 0;
-      nanosleep_interval.tv_nsec = 75 * 1000 * 1000; /* 75 milliseconds */
+      nanosleep_interval.tv_nsec = 100 * 1000 * 1000; /* 100 milliseconds */
       /* It's critical here that valgrind's nanosleep implementation
          is nonblocking. */
       (void)my_do_syscall2(__NR_nanosleep, 
@@ -907,19 +960,28 @@
 
 int poll (struct pollfd *__fds, nfds_t __nfds, int __timeout)
 {
+   unsigned int        ms_now, ms_end;
    int                 res, i;
-   struct vki_timeval  t_now;
-   struct vki_timeval  t_end;
    struct vki_timespec nanosleep_interval;
 
    ensure_valgrind("poll");
 
+   /* Detect the current time and simultaneously find out if we are
+      running on Valgrind. */
+   VALGRIND_MAGIC_SEQUENCE(ms_now, 0xFFFFFFFF /* default */,
+                           VG_USERREQ__READ_MILLISECOND_TIMER,
+                           0, 0, 0, 0);
+
    if (/* CHECK SIZES FOR struct pollfd */
        sizeof(struct timeval) != sizeof(struct vki_timeval))
       barf("valgrind's hacky non-blocking poll(): data sizes error");
 
-   /* If a zero timeout specified, this call is harmless. */
-   if (__timeout == 0) {
+   /* dummy initialisation to keep gcc -Wall happy */
+   ms_end = 0;
+
+   /* If a zero timeout specified, this call is harmless.  Also do
+      this if not running on Valgrind. */
+   if (__timeout == 0 || ms_now == 0xFFFFFFFF) {
       res = my_do_syscall3(__NR_poll, (int)__fds, __nfds, __timeout);
       if (is_kerror(res)) {
          * (__errno_location()) = -res;
@@ -929,36 +991,25 @@
       }
    }
 
-   /* If a timeout was specified, set t_end to be the end wallclock
-      time. */
+   /* If a timeout was specified, set ms_end to be the end wallclock
+      time.  Easy considering that __timeout is in milliseconds. */
    if (__timeout > 0) {
-      res = my_do_syscall2(__NR_gettimeofday, (int)&t_now, (int)NULL);
-      assert(res == 0);
-      t_end = t_now;
-      t_end.tv_usec += 1000 * (__timeout % 1000);
-      t_end.tv_sec  += (__timeout / 1000);
-      if (t_end.tv_usec >= 1000000) {
-         t_end.tv_usec -= 1000000;
-         t_end.tv_sec += 1;
-      }
-      /* Stay sane ... */
-      assert (t_end.tv_sec > t_now.tv_sec
-              || (t_end.tv_sec == t_now.tv_sec 
-                  && t_end.tv_usec >= t_now.tv_usec));
+      ms_end += (unsigned int)__timeout;
    }
 
    /* fprintf(stderr, "MY_POLL: before loop\n"); */
 
    /* Either timeout < 0, meaning wait indefinitely, or timeout > 0,
       in which case t_end holds the end time. */
+   assert(__timeout != 0);
+
    while (1) {
-      assert(__timeout != 0);
       if (__timeout > 0) {
-         res = my_do_syscall2(__NR_gettimeofday, (int)&t_now, (int)NULL);
-         assert(res == 0);
-         if (t_now.tv_sec > t_end.tv_sec
-             || (t_now.tv_sec == t_end.tv_sec 
-                 && t_now.tv_usec > t_end.tv_usec)) {
+         VALGRIND_MAGIC_SEQUENCE(ms_now, 0xFFFFFFFF /* default */,
+                                 VG_USERREQ__READ_MILLISECOND_TIMER,
+                                 0, 0, 0, 0);
+         assert(ms_now != 0xFFFFFFFF);
+         if (ms_now >= ms_end) {
             /* timeout; nothing interesting happened. */
             for (i = 0; i < __nfds; i++) 
                __fds[i].revents = 0;
@@ -966,8 +1017,7 @@
          }
       }
 
-      /* These could be trashed each time round the loop, so restore
-         them each time. */
+      /* Do a return-immediately poll. */
       res = my_do_syscall3(__NR_poll, (int)__fds, __nfds, 0 );
       if (is_kerror(res)) {
          /* Some kind of error.  Set errno and return.  */
@@ -981,7 +1031,7 @@
       /* fprintf(stderr, "MY_POLL: nanosleep\n"); */
       /* nanosleep and go round again */
       nanosleep_interval.tv_sec  = 0;
-      nanosleep_interval.tv_nsec = 100 * 1000 * 1000; /* 100 milliseconds */
+      nanosleep_interval.tv_nsec = 99 * 1000 * 1000; /* 99 milliseconds */
       /* It's critical here that valgrind's nanosleep implementation
          is nonblocking. */
       (void)my_do_syscall2(__NR_nanosleep,