Author: sewardj
Date: 2007-10-01 11:21:07 +0100 (Mon, 01 Oct 2007)
New Revision: 6927
Log:
Improve performance for OpenOffice-sized applications by messing with
N_SECMAP_BITS, and update some associated comments.
Modified:
branches/THRCHECK/thrcheck/tc_main.c
Modified: branches/THRCHECK/thrcheck/tc_main.c
===================================================================
--- branches/THRCHECK/thrcheck/tc_main.c 2007-09-30 10:48:32 UTC (rev
6926)
+++ branches/THRCHECK/thrcheck/tc_main.c 2007-10-01 10:21:07 UTC (rev
6927)
@@ -571,7 +571,14 @@
/* Primary Map is the usual FiniteMap Addr SecMap* w/ cache */
-#define N_SECMAP_BITS 11
+/* See comments below on shadow_mem_make_NoAccess re performance
+ effects of N_SECMAP_BITS settings. On a 2.4GHz Core2,
+ starting/quitting OOo (32-bit), I have these rough numbers:
+ N_SECMAP_BITS = 11 2m23
+ N_SECMAP_BITS = 12 1m58
+ N_SECMAP_BITS = 13 1m53
+*/
+#define N_SECMAP_BITS 13
#define N_SECMAP_ARANGE (1 << N_SECMAP_BITS)
#define N_SECMAP_MASK (~(N_SECMAP_ARANGE - 1))
#define N_SECMAP_W32S (N_SECMAP_ARANGE / 4)
@@ -2499,35 +2506,35 @@
Otherwise, the range contains some locks. Then we have to do all
the complexity below and only then set the range to NoAccess.
- // make up a set containing the which are deleted
- ToDelete = NULL
- scan the range in map_locks,
- for each lock lk found {
- add lk to ToDelete
-
- case lk of
- UnlockedNew. simply free up the Lock
- UnlockedBy. makes a reference to Segment (via
- map_segments) but just leave that.
- Maybe we will later GC the segment graph.
+ 1. Make up a set containing the locks which are deleted:
- LockedBy. Bad case. Report error.
- remove from from the currlocks set of the
- owning thread
- }
+ ToDelete = NULL
- modify all shadow words, by removing ToDelete from the lockset
- of all ShM and ShR states.
+ for each lk in map_locks {
+ if lk's guest addr falls in the range to memory be deleted
+ add lk to ToDelete
- Finally set the range to NoAccess.
+ if lk is held, issue an error message - freeing memory
+ containing a held lock
+ }
+
+ 2. Modify all shadow words, by removing ToDelete from the lockset
+ of all ShM and ShR states. Note this involves a complete scan
+ over map_shmem, which is very expensive according to OProfile.
+ Hence it depends critically on the size of each entry in
+ map_shmem. See comments on definition of N_SECMAP_BITS above.
- Optimisation 1: skip SecMaps which do not have .hasShared set
+ 3. Set the range to NoAccess. Clear the .hasShared and .anyLocks
+ hint bits for any completely vacated SecMaps.
- Optimisation 2: for each SecMap, have a summary lock set which is
- the union of all locks mentioned in locksets on this page.
- Then skip the page if the summary lockset does not intersect with
- ToDelete.
+ Optimisation 1 (implemented): skip SecMaps which do not have
+ .hasShared set
+ Optimisation 2 (not implemented, needs rethink): for each SecMap,
+ have a summary lock set which is the union of all locks mentioned
+ in locksets on this page. Then skip the page if the summary
+ lockset does not intersect with ToDelete.
+
that's potentially cheap, since the usual lockset refinement
only shrinks locksets; hence there is no point in updating the
summary lockset for ShM/R -> ShM/R transitions
@@ -3229,7 +3236,7 @@
}
// FIXME: error-if: exiting thread holds any locks
- // or shouw evh__pre_thread_ll_exit do that?
+ // or should evh__pre_thread_ll_exit do that?
/* Delete thread from ShM/ShR thread sets and restore Excl states
where appropriate */
-------------------------------------------------------------------------
This SF.net email is sponsored by: Microsoft
Defy all challenges. Microsoft(R) Visual Studio 2005.
http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/
_______________________________________________
Valgrind-developers mailing list
Valgrind-developers@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/valgrind-developers
