[pypy-svn] r68987 - pypy/branch/gc-jit-hack/pypy/jit/metainterp

[arigo at codespeak.net]

Author: arigo
Date: Thu Nov  5 12:46:09 2009
New Revision: 68987

Modified:
   pypy/branch/gc-jit-hack/pypy/jit/metainterp/pyjitpl.py
Log:
This is the best result so far.  At least in this case I understand the
results: they seem to clearly show that a lot of the objects allocated
during jit-tracing actually survive.  The results are not good enough --
they say that jit-tracing takes only half the time, but the truth is
that it's always soon followed by a lengthy minor collection, so that
the total time (as well as the total GC time) end up not really
improved.  Closing this branch for now.



Modified: pypy/branch/gc-jit-hack/pypy/jit/metainterp/pyjitpl.py
==============================================================================
--- pypy/branch/gc-jit-hack/pypy/jit/metainterp/pyjitpl.py	(original)
+++ pypy/branch/gc-jit-hack/pypy/jit/metainterp/pyjitpl.py	Thu Nov  5 12:46:09 2009
@@ -1086,14 +1086,13 @@
             from pypy.rpython.lltypesystem import lltype, lloperation
             from pypy.rlib import rgc
             if enlarge:
+                rgc.collect(0)
                 newsize = 10*1024*1024         # XXX a rather custom number
                 if sys.maxint > 2147483647:
                     newsize *= 2
             else:
                 newsize = 0
             lloperation.llop.gc_resize_nursery(lltype.Void, newsize)
-            if enlarge:
-                rgc.collect(0)
 
 # ____________________________________________________________