Preparse old-style octals as strings

[Jeroen Demeyer]

To solve #17807, we preparse 0100 as Integer('0100') instead of Integer(0100): thanks to #17413, this will give a deprecation warning so that users should know something funny is going on:

sage: 0100
/usr/local/src/sage-git/local/lib/python2.7/site-packages/IPython/core/interactiveshell.py:2883: DeprecationWarning: use 0o as octal prefix instead of 0
If you do not want this number to be interpreted as octal, remove the leading zeros.
See http://trac.sagemath.org/17413 for details.
  exec(code_obj, self.user_global_ns, self.user_ns)
64

We also preparse new-style binary/octal/hex strings as integers instead of strings, which should be faster, see 5.

[Jeroen Demeyer]

New commits:

​533af4c

Preparse integers as strings

[Nils Bruin]

I would think the slow-down for small integers makes the proposed changes very unattractive. The whole idea of preparsing 100 as Integer(100) is that in the AST and the generated byte code, the constant 100 is already stored as a numerical one, and hence conversion is much faster.

For instance, in code like:

sage: %timeit L=[Integer('1000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000') for i in range(10000r)]
100 loops, best of 3: 11.4 ms per loop
sage: %timeit L=[Integer(1000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000) for i in range(10000r)]
100 loops, best of 3: 4.66 ms per loop

you really start to notice this.

preparse_file mitigates issues like this a little bit by factoring out numeric constants, so that their conversion doesn't happen in inner loops.

If you want to store your numerical constants as strings in the byte code and have fast conversion, you should probably store them in hex. It's a little more work in the preparser, but the result is more compact and conversion to a numerical constant is truly linear in number of bits.

[Jeroen Demeyer]

Replying to nbruin:

The whole idea of preparsing 100 as Integer(100) is that in the AST and the generated byte code, the constant 100 is already stored as a numerical one

Is that really important? The "timeit" example is of course a bit artificial.

[Nils Bruin]

Replying to jdemeyer:

Replying to nbruin:

The whole idea of preparsing 100 as Integer(100) is that in the AST and the generated byte code, the constant 100 is already stored as a numerical one

Is that really important? The "timeit" example is of course a bit artificial.

I think so. Storing the numerical constant as a string forces the decimal-to-binary conversion to happen at runtime. If you just write it as a python integer, the conversion happens at parsing, so by the time it's an ast and the compiler looks at it, you're already dealing with an integer.

Unfortunately, python doesn't permit compile-time macros (perhaps if we're ever going to rewrite our preparser to properly parse, we can combine it with ​MacroPy and then we can have Integer object creation at compile time)

I am fairly certain that nearly all code that gets run with sage has TONS of small integer literals in it. Probably for many applications, people would get better performance writing 100r, but of course nobody will. This ticket would deteriorate the much more common situation in favour of a small gain in the extremely rare situation that someone wants to include an incredibly large integer literal in their code (talk about artificial--I don't think having a (small) numerical literal in an inner loop is artificial at all!). And the extremely rare situation can be solved by writing quotes already (and by the time you're writing such a long string you might as well convert it to hex anyway, in which case Python and GMP are (at least asymptotically) comparable.

If you use preparse_file instead of preparse, you'll see that numerical constants get pushed outside (which has its own set of problems, see #11542), which mitigates the string-conversion-in-inner-loop.

[Jeroen Demeyer]

Fine, you convinced me.

[git]

Branch pushed to git repo; I updated commit sha1. This was a forced push. New commits:

​f72c449

Preparse old-style octal numbers as strings

[Frédéric Chapoton]

New commits:

​293c474	Merge branch 'u/jdemeyer/ticket/17808' into 6.7.b1
​b469c4c	trac #17808 taking care of pt tutorial

[git]

Branch pushed to git repo; I updated commit sha1. New commits:

​b243aa3	Merge branch 'public/ticket/17808' into 7.2.b6
​52fd0ce	trac #17808 handling es and ja tutorial

[Frédéric Chapoton]

is the title still acurately describing the content of the ticket ?

[Jeroen Demeyer]

Replying to chapoton:

is the title still acurately describing the content of the ticket ?

Yes.

[Jeroen Demeyer]

Although the ticket does change a little bit more than that, I added a sentence to the description.

[Frédéric Chapoton]

ok, looks good to me