lp:~mwhudson/pypy/imported-guard-value-counting

Created by Michael Hudson-Doyle on 2010-03-15 and last modified on 2010-03-15
Get this branch:
bzr branch lp:~mwhudson/pypy/imported-guard-value-counting
Only Michael Hudson-Doyle can upload to this branch. If you are Michael Hudson-Doyle please log in for upload directions.

Branch merges

Related bugs

Related blueprints

Branch information

Owner:
Michael Hudson-Doyle
Project:
PyPy
Status:
Development

Recent revisions

23862. By arigo on 2010-03-05

Check in my working copy. That branch will probably be abandonned
and maybe restarted some day.

23861. By arigo on 2010-02-26

Start implementing the idea that there are really two kinds of Descrs
for guards: one before optimization, and the other after optimization.
They need slightly different attributes, and the goal is to let
optimization decide exactly which subclass of ResumeGuardDescr is
needed for each guard.

23860. By arigo on 2010-02-25

A branch in which to try the following idea: on failing GUARD_VALUE,
put not just one counter but one counter per value. This should help
because a GUARD_VALUE that keeps seeing new values is not useful to
start tracing from.

23859. By arigo on 2010-02-25

Merge branch/debug-vref3:

Revert r71374, as it appears that there are cases in which
the performance matters -- i.e. the fact that some fields
may remain virtual across a call to a function that can
force virtualrefs or virtualizables. (According to fijal
he sees a 30% difference on one benchmark.)

Fix it differently in optimizeopt.py, by saying that calls
to such functions always force at least the 'forced' field
of virtualrefs (but not necessarily other fields).

23858. By arigo on 2010-02-25

Kill the unused GOND_CALL_GC_MALLOC. In the x86 backend,
the same result is achieved by _fastpath_malloc.

23857. By arigo on 2010-02-25

Attempt to be more Windows-friendly.

23856. By fijal on 2010-02-24

Ooops, it was not supposed to go to this branch

23855. By fijal on 2010-02-24

An approach to having read/write not copy stuff. Let's see how well it performs
before actually trying hard to make lltype happy

23854. By agaynor on 2010-02-24

Make a bunch of the itertools classes into new-style classes.

23853. By arigo on 2010-02-24

Merge the gc-better-hash branch.

It improves the identityhash of young objects -- they are now not just
the address, because the address of young objects is always within the
small nursery; they are the address plus a number that changes at every
nursery collection.

Also, it contains a bit of work to let the identityhash logic be tested
by test_direct.py, at least on SemiSpace and Generation GC -- there is
still an issue with testing it on the Hybrid GC. (Of course, it's still
fully tested by test_newgc.py.)

Branch metadata

Branch format:
Branch format 7
Repository format:
Bazaar repository format 2a (needs bzr 1.16 or later)
Stacked on:
lp:pypy
This branch contains Public information 
Everyone can see this information.