Re: Python-on-guile

[Mikael Djurfeldt]

Nice!

I guess it would be nice if "continue" *could* be compiled efficiently. And, as you indicate, perhaps that would amount to efficiently compiling let/ec.

Best regards,

Mikael

On Sat, Apr 24, 2021 at 5:19 PM Stefan Israelsson Tampe <stefan.itampe@gmail.com> wrote:

Guile is 3x faster then fastest python-on-guile which is 2x faster then python3 Cpython

attached is a guile corresponding program.

On Sat, Apr 24, 2021 at 4:41 PM Stefan Israelsson Tampe <stefan.itampe@gmail.com> wrote:

To note is that 'continue' is killing performance for python-on-guile programs, so by changing the

code to not use continue lead to python-on-guile running twice the speed of python3. The reason is that 

the while loop is used as

(while (...)

   (let/ec continue

        ...))

And the let/ec is probably not optimally compiled. Python-on-guile will check the loop for continue usage and if not then it will skip the let/ec.

I attached the code not using continue

On Sat, Apr 24, 2021 at 2:59 PM Stefan Israelsson Tampe <stefan.itampe@gmail.com> wrote:

Actually changing in (language python compile),

(define (letec f)

  (let/ec x (f x)))) 

To

(define-syntax-rule (letec f)

  (let/ec x (f x)))) 

Actually lead to similar speeds as python3.

On Sat, Apr 24, 2021 at 1:26 PM Stefan Israelsson Tampe <stefan.itampe@gmail.com> wrote:

Pro tip, when running this on guile the scheme code that it compilse to is located in log.txt.

If you ,opt the resulting code in a guile session you might be able to pinpoint issues that 

delays the code execution.

On Sat, Apr 24, 2021 at 12:04 PM Mikael Djurfeldt <mikael@djurfeldt.com> wrote:

(I should perhaps add that my script doesn't benchmark the object system but rather loops, conditionals and integer arithmetic.)

Den fre 23 apr. 2021 17:00Mikael Djurfeldt <mikael@djurfeldt.com> skrev:

Hi,

Yesterday, Andy committed new code to the compiler, some of which concerned skipping some arity checking.

Also, Stefan meanwhile committed something called "reworked object system" to his python-on-guile.

Sorry for coming with unspecific information (don't have time to track down the details) but I noticed that my benchmark script written in Python, and which computes the 20:th Ramanujan number, now runs 60% faster than before these changes.

This means that python-on-guile running on guile3 master executes python code only 2.6 times slower than the CPython python3 interpreter itself. :-)

Have a nice weekend all,

Mikael