PyPy now "just works" with @

[Pierre-Yves David]

On 12/05/2015 06:22 PM, Gregory Szorc wrote:
> On Sat, Dec 5, 2015 at 5:14 PM, Pierre-Yves David
> <pierre-yves.david at ens-lyon.org <mailto:pierre-yves.david at ens-lyon.org>>
> wrote:
>
>     On 12/05/2015 01:30 PM, Bryan O'Sullivan wrote:
>
>         Very interesting, and thank you for putting in the work to get
>         this going.
>
>         What's the effect of startup time on fast-running commands? A
>         majority
>         of Mercurial invocations take a fraction of a second even on a huge
>         repo, so this is of significant interest.
>
>
>     Basic testing around status and diff show that pypy is still
>     significantly slower on basic operation (even pretty bad hit for diff).
>
>     My mercurial repository
>     ------------------------
>     status
>     python-2.7: 0.040
>     pypy-4.0.1: 0.260
>
>     diff:
>     python-2.7: 0.460
>     pypy-4.0.1: 0.980
>
>     A Mozilla central clone
>     ------------------------
>     status:
>     python-2.7: 0.710
>     pypy-4.0.1: 1.770
>
>     diff:
>     python-2.7: 1.130
>     pypy-4.0.1: 8.880
>
>     log: (252524 changesets)
>     python-2.7: 60.710
>     pypy-4.0.1: 13.200
>
>     Some revset:
>     log --rev 'author(greg) - merge()' (2748 match)
>     python-2.7: 17.740
>     pypy-4.0.1: 8.360
>
>      From here we can explore multiple path:
>
>     - routing different commands to different implementation,
>     - getting chg production ready (mostly a life cycle question at that
>     point),
>     - grabbing the low hanging fruit to have pypy happier,
>     - Getting the pypy devs to look at the less trivial work,
>
>
> It's worth noting that both status and diff also measure differences
> between Mercurial's C and Python implementations, not merely startup
> overhead and PyPy's JIT "warmup" impact. For example, if you profile `hg
> status` for mozilla-central, you'll find that PyPy spends a lot more
> time in calling osutil.listdir(). The difference walking the filesystem
> account for ~75% of the overhead of PyPy on my machine.

Yep, plugging pypy to our C-code will be pretty important to move forward.

-- 
Pierre-Yves David