[pypy-dev] Optimize constant float division by multiplication?

[Toni Mattis]

Hi all,

thanks for the advice. I tried what Armin proposed and like to share my
results with you:

https://bitbucket.org/amintos/pypy/commits/937254cbc554adfb748e3b5eeb44bf765d204b9d?at=default

Keeping in mind what Steve and Maciej pointed out, I restricted the
optimization to floats that are "normal" powers of two. I thought about
checking for infinity, but I could not come up with a scenario where 'x
* (-)0.0' differs from 'x / (-)inf'.

I haven't done exhaustive tests yet, but some of the code where I first
discovered the issue runs a little faster now. Comments and possibly
missed corner cases are welcome (IEEE-754 can be a minefield sometimes).


Thanks,

Toni



Am 06.11.2014 um 10:07 schrieb Armin Rigo:
> Hi Toni,
> 
> On 6 November 2014 10:00, Armin Rigo <arigo at tunes.org> wrote:
>> gcc seems to perform this optimization for divide-by-constant where
>> the constant is exactly a finite power of two that is not a denormal.
>> These are the cases where the result is exactly the same.  We could do
>> it too.
> 
> In short, what is needed is:
> 
> - first check that the optimization you want to do is exact; trying it
> out on "gcc -O2 -S" without any "-ffast-math" flags is a good way to
> know.
> 
> - if it is, then it's a matter of writing some simple code in
> rpython/jit/metainterp/optimizeopt/rewrite.py.  Search for "float_mul"
> here; it will turn for example "f0 * -1.0" into a "float_neg"
> operation, with the comment that it is an exact optimization.
> 
> - don't forget, start by adding a test to test/test_optimizebasic.py
> (search for "float_mul(-1.0, f0)" and add it nearby).
> 
> You might find out that hacking the PyPy JIT at this level is rather easy :-)
> 
> 
> A bientôt,
> 
> Armin.
>