myths about python 3

Terry Reedy tjreedy at udel.edu
Thu Jan 28 07:39:14 CET 2010


On 1/27/2010 8:36 PM, alex23 wrote:
> Terry Reedy<tjre... at udel.edu>  wrote:
>> Actually, Unladen Swallow is now targeted at 3.1; its developers have
>> conservatively proposed its integration in CPython 3.3.

This statement was to counter the 'myth' that US was only targeted at 
2.x when the current situation is quite the opposite.

>> I would not be completely shocked if it happens in 3.2.

I was initially rather dubious about the idea. I based the above
on the team's acceptance of and response to reasonable requirements.
In particular, several people said that the speed/space traceoff
should be optional, and that compilation 'without llvm' should really
be without, not just with llvm present but disabled. Instead of arguing,
Colin went ahead and patched the build process to make it be this way.

> Why do I feel like there's less of an onus on Unladen Swallow to
> _actually prove itself in substantial real world usage_ before
> integration into CPython than there is on even the smallest of modules
> for inclusion in the standard library?

I have no idea. It will have to improve its speedup more before 
adoption. I will not be surprised if that happens.

> Are we really expected to just ditch everything we know about
> CPython's performance characteristics just for some questionable and
> possibly uneven gains?
>
> I've been a big supporter of Py3 from the beginning, but this repeated
> claim of US becoming the mainline interpreter for 3.x

US is not a new or separate interpreter. It will be an optional jit 
replacement for one component of CPython, the eval loop. All the code 
for builting functions, types, and modules will be untouched, as will 
their big O performance characteristics.

> pretty much kills dead a lot of my interest.

If you can still have a binary free of the traceoff, why would you care?

 > What am I not seeing amidst the high
> memory usage and variable performance results of US's _custom-made_
> benchmarks? Doesn't that fact alone worry anyone else? Or that LLVM is
> listed as only having "partial support" with non-Cygwin x86 Windows?

They claim they have pretty well fixed that. They know that complete 
Windows support, including 64 bit versions, is a necessity.

Terry Jan Reedy




More information about the Python-list mailing list