[pypy-dev] sys._getframe(1).f_locals is very slow on PyPy 1.8

[Jacob Hallén]

One of the ways the JIT makes things go faster is by not generating any frames 
at all. When you ask for the contents of a frame, a virtual frame is 
generated, to provide the information you are asking for. This takes a fair 
bit of time, making your code slow.

Jacob Hallén

Sunday 27 May 2012 you wrote:
> Hi Ronny,
> 
> Thank you for your reply.
> You mean that sys._getframe() will remain slow on PyPy.
> I'll try other way (instead of sys._getframe()) to make Tenjin faster on
> PyPy.
> 
> --
> regards,
> makoto kuwata
> 
> On Sun, May 27, 2012 at 6:09 PM, Ronny Pfannschmidt
> 
> <Ronny.Pfannschmidt at gmx.de> wrote:
> > Hi Makoto,
> > 
> > currently sys._getframe() is mostly killing the jit
> > since it currently needs access to all frame objects,
> > 
> > if i remember correct (arigo, fijal or anto will know for sure),
> > the jit currently just bails out if it gets a sys._getframe call
> > 
> > (which easyly explain the ~34 times slower  in your test)
> > 
> > i think there are ways to support it,
> > but i currently don't have the time/knowledge to check which of them work
> > 
> > -- Ronny
> > 
> > On 05/27/2012 06:24 AM, Makoto Kuwata wrote:
> >> Hi,
> >> 
> >> I found that sys._getframe() (or sys._getframe().f_locals) is
> >> very slow in PyPy 1.8.
> >> For example, the following takes only 0.0398 sec on PyPy 1.8:
> >> 
> >>       def f5(x): return f4(x)
> >>       def f4(x): return f3(x)
> >>       def f2(x): return f1(x)
> >>       def f1(x): reutrn None               ## just return None
> >>       for _ in xrange(1000000):
> >>           f5(0)
> >> 
> >> But the next one takes 1.3563 sec.
> >> 
> >>       def f5(x): return f4(x)
> >>       def f4(x): return f3(x)
> >>       def f2(x): return f1(x)
> >>       def f1(x): reutrn sys._getframe(1)    ## calls sys._getframe(1)
> >>       for _ in xrange(1000000):
> >>           f5(0)
> >> 
> >> Is there any plan to make it faster in the future release?
> >> 
> >> 
> >> Background
> >> ----------
> >> 
> >> I'm an author of Tenjin, a high-speed template engine for Python.
> >> Tenjin supports PyPy officially, and I received a report that
> >> there are some cases in which Tenjin is very slow on PyPy.
> >> 
> >> I investigated report and found that it is the sys._getframe()
> >> and sys._getframe().f_locals that makes Tenjin much slow on PyPy.
> >> 
> >> Tenjin uses sys._getframe(1).f_locals in order to access to
> >> caller's local variables. It is necessary for Tenjin.
> >> 
> >> 
> >> Benchmark
> >> ---------
> >> 
> >> Here is a benchmark script to measure cost of sys._getframe(1)
> >> and sys._getframe(1).f_locals.
> >> (You must install Benchmarker library to run it.)
> >> 
> >> https://gist.github.com/2802160
> >> 
> >> Benchmark result shows that:
> >> 
> >> * Cost of sys._getframe(1) and sys._getframe(1).f_locals is
> >>   not so much on CPython but very big on PyPy.
> >> * Nested depth os function call affects to cost of them.
> >>   For example:
> >> 
> >>       def f5(x): return f4(x)
> >>       def f4(x): return f3(x)
> >>       def f2(x): return f1(x)
> >>       def f1(x): reutrn sys._getframe(1)
> >>       f5(0)
> >> 
> >>   is much slower than:
> >> 
> >>       def f2(x): return f1(x)
> >>       def f1(x): reutrn sys._getframe(1)
> >>       f2(0)
> >> 
> >> 
> >> 
> >> By the way, PyPy is awesome.
> >> It is not only faster than CPython but also has high-level
> >> compatibility with CPython. Almost of all test script of Tenjin
> >> works on PyPy. Great.
> >> I hope Tenjin works much faster on PyPy.
> >> 
> >> --
> >> regards,
> >> makoto kuwata
> >> _______________________________________________
> >> pypy-dev mailing list
> >> pypy-dev at python.org
> >> http://mail.python.org/mailman/listinfo/pypy-dev
> 
> _______________________________________________
> pypy-dev mailing list
> pypy-dev at python.org
> http://mail.python.org/mailman/listinfo/pypy-dev