2010-02-22 Pedro Alves <pedro@codesourcery.com>
[deliverable/binutils-gdb.git] / gdb / breakpoint.c
index 8c97949d7dccdeef1533a3be3376361371c2d508..4224c76cbc5a340cf77758b9827d36fbca7d6727 100644 (file)
@@ -127,6 +127,9 @@ static int breakpoint_address_match (struct address_space *aspace1,
                                     struct address_space *aspace2,
                                     CORE_ADDR addr2);
 
+static int watchpoint_locations_match (struct bp_location *loc1,
+                                      struct bp_location *loc2);
+
 static void breakpoints_info (char *, int);
 
 static void breakpoint_1 (int, int);
@@ -1485,10 +1488,43 @@ Note: automatically using hardware breakpoints for read-only addresses.\n"));
              watchpoints.  It's not clear that it's necessary... */
           && bpt->owner->disposition != disp_del_at_next_stop)
     {
-      val = target_insert_watchpoint (bpt->address, 
+      val = target_insert_watchpoint (bpt->address,
                                      bpt->length,
                                      bpt->watchpoint_type);
-      bpt->inserted = (val != -1);
+
+      /* If trying to set a read-watchpoint, and it turns out it's not
+        supported, try emulating one with an access watchpoint.  */
+      if (val == 1 && bpt->watchpoint_type == hw_read)
+       {
+         struct bp_location *loc, **loc_temp;
+
+         /* But don't try to insert it, if there's already another
+            hw_access location that would be considered a duplicate
+            of this one.  */
+         ALL_BP_LOCATIONS (loc, loc_temp)
+           if (loc != bpt
+               && loc->watchpoint_type == hw_access
+               && watchpoint_locations_match (bpt, loc))
+             {
+               bpt->duplicate = 1;
+               bpt->inserted = 1;
+               bpt->target_info = loc->target_info;
+               bpt->watchpoint_type = hw_access;
+               val = 0;
+               break;
+             }
+
+         if (val == 1)
+           {
+             val = target_insert_watchpoint (bpt->address,
+                                             bpt->length,
+                                             hw_access);
+             if (val == 0)
+               bpt->watchpoint_type = hw_access;
+           }
+       }
+
+      bpt->inserted = (val == 0);
     }
 
   else if (bpt->owner->type == bp_catchpoint)
@@ -3434,11 +3470,67 @@ bpstat_check_watchpoint (bpstat bs)
            case WP_VALUE_CHANGED:
              if (b->type == bp_read_watchpoint)
                {
-                 /* Don't stop: read watchpoints shouldn't fire if
-                    the value has changed.  This is for targets
-                    which cannot set read-only watchpoints.  */
-                 bs->print_it = print_it_noop;
-                 bs->stop = 0;
+                 /* There are two cases to consider here:
+
+                    1. we're watching the triggered memory for reads.
+                    In that case, trust the target, and always report
+                    the watchpoint hit to the user.  Even though
+                    reads don't cause value changes, the value may
+                    have changed since the last time it was read, and
+                    since we're not trapping writes, we will not see
+                    those, and as such we should ignore our notion of
+                    old value.
+
+                    2. we're watching the triggered memory for both
+                    reads and writes.  There are two ways this may
+                    happen:
+
+                    2.1. this is a target that can't break on data
+                    reads only, but can break on accesses (reads or
+                    writes), such as e.g., x86.  We detect this case
+                    at the time we try to insert read watchpoints.
+
+                    2.2. otherwise, the target supports read
+                    watchpoints, but, the user set an access or write
+                    watchpoint watching the same memory as this read
+                    watchpoint.
+
+                    If we're watching memory writes as well as reads,
+                    ignore watchpoint hits when we find that the
+                    value hasn't changed, as reads don't cause
+                    changes.  This still gives false positives when
+                    the program writes the same value to memory as
+                    what there was already in memory (we will confuse
+                    it for a read), but it's much better than
+                    nothing.  */
+
+                 int other_write_watchpoint = 0;
+
+                 if (bl->watchpoint_type == hw_read)
+                   {
+                     struct breakpoint *other_b;
+
+                     ALL_BREAKPOINTS (other_b)
+                       if ((other_b->type == bp_hardware_watchpoint
+                            || other_b->type == bp_access_watchpoint)
+                           && (other_b->watchpoint_triggered
+                               == watch_triggered_yes))
+                         {
+                           other_write_watchpoint = 1;
+                           break;
+                         }
+                   }
+
+                 if (other_write_watchpoint
+                     || bl->watchpoint_type == hw_access)
+                   {
+                     /* We're watching the same memory for writes,
+                        and the value changed since the last time we
+                        updated it, so this trap must be for a write.
+                        Ignore it.  */
+                     bs->print_it = print_it_noop;
+                     bs->stop = 0;
+                   }
                }
              break;
            case WP_VALUE_NOT_CHANGED:
@@ -4697,6 +4789,12 @@ breakpoint_address_is_meaningful (struct breakpoint *bpt)
 static int
 watchpoint_locations_match (struct bp_location *loc1, struct bp_location *loc2)
 {
+  /* Note that this checks the owner's type, not the location's.  In
+     case the target does not support read watchpoints, but does
+     support access watchpoints, we'll have bp_read_watchpoint
+     watchpoints with hw_access locations.  Those should be considered
+     duplicates of hw_read locations.  The hw_read locations will
+     become hw_access locations later.  */
   return (loc1->owner->type == loc2->owner->type
          && loc1->pspace->aspace == loc2->pspace->aspace
          && loc1->address == loc2->address
@@ -8459,6 +8557,16 @@ update_global_location_list (int should_insert)
                          /* For the sake of should_be_inserted.
                             Duplicates check below will fix up this later.  */
                          loc2->duplicate = 0;
+
+                         /* Read watchpoint locations are switched to
+                            access watchpoints, if the former are not
+                            supported, but the latter are.  */
+                         if (is_hardware_watchpoint (old_loc->owner))
+                           {
+                             gdb_assert (is_hardware_watchpoint (loc2->owner));
+                             loc2->watchpoint_type = old_loc->watchpoint_type;
+                           }
+
                          if (loc2 != old_loc && should_be_inserted (loc2))
                            {
                              loc2->inserted = 1;
This page took 0.027098 seconds and 4 git commands to generate.