Path: csiph.com!usenet.pasdenom.info!weretis.net!feeder4.news.weretis.net!news.mixmin.net!eweka.nl!hq-usenetpeers.eweka.nl!xlned.com!feeder1.xlned.com!news2.euro.net!newsgate.cistron.nl!newsgate.news.xs4all.nl!post.news.xs4all.nl!not-for-mail
MIME-Version: 1.0
Sender: joshua.landau.ws@gmail.com
In-Reply-To: <ksejfu$ist$1@ger.gmane.org>
References: <mailman.4865.1374240179.3114.python-list@python.org> <51e967bb$0$29971$c3e8da3$5496439d@news.astraweb.com> <ksbt1a$5q1$1@ger.gmane.org> <CAN1F8qXttLWtMFDED-+gEdOR_5tZmDKtqDsF2kdojCRSMAn7eg@mail.gmail.com> <ksdtu7$ctm$1@ger.gmane.org> <CAN1F8qVv0N1D=JEkvWeQePhZg7fT7ULPPXp+ZTOmu=wQCzXiHQ@mail.gmail.com> <ksejfu$ist$1@ger.gmane.org>
From: Joshua Landau <joshua@landau.ws>
Date: Sat, 20 Jul 2013 19:37:28 +0100
Subject: Re: Find and Replace Simplification
To: Dave Angel <davea@davea.name>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: quoted-printable
Cc: python-list <python-list@python.org>
Precedence: list
Newsgroups: comp.lang.python
Message-ID: <mailman.4928.1374345496.3114.python-list@python.org>
Lines: 63
NNTP-Posting-Host: 2001:888:2000:d::a6
Xref: csiph.com comp.lang.python:50972

On 20 July 2013 19:04, Dave Angel <davea@davea.name> wrote:
> On 07/20/2013 01:03 PM, Joshua Landau wrote:
>>
>> Still, it seems to me that it should be optimizable for sensible
>> builtin types such that .translate is significantly faster, as there's
>> no theoretical extra work that .translate *has* to do that .replace
>> does not, and .replace also has to rebuild the string a lot of times.
>>
>
> translate is going to be faster (than replace) for Unicode if it has a
> "large" table.  For example, to translate from ASCII to EBCDIC, where eve=
ry
> character in the string is replaced by a new one.  I have no idea what th=
e
> cutoff is.  But of course, for a case like ASCII to EBCDIC, it would be v=
ery
> tricky to do it with replaces, probably taking much more than the expecte=
d
> 96 passes.

My timings showed that for ".upper()", doing the full 26 passes "a" ->
"A", it was *way* slower to use .translate than .replace, unless you
used a list or equiv. with much faster lookup. Even then, it was
slower to use .translate.

I agree that for large tables it's obviously going to swing the other
way, but by the time you're running .replace 26 times you wouldn't (at
least I wouldn't) expect it still to be screamingly faster than
.translate.

> translate for byte strings is undoubtedly tons faster.  For byte strings,
> the translation table is 256 bytes, and the inner loop is a simple lookup=
.

For my above test, .translate is about 10x faster than iterated .replace.

> But for Unicode, the table is a dict (or something very like it, I looked=
 at
> the C code, not the Python code).
>
> So for every character in the input string, it does a dict-type lookup,
> before it can even decide if the character is going to change.

The problem can be solved, I'd imagine, for builtin types. Just build
an internal representation upon calling .translate that's faster. It's
especially easy in the list case -- just build a C array=C2=B9 at the start
mapping int -> int and then have really fast C mapping speeds.

For dictionaries, you can do the same thing -- you just have to make
sure you're not breaking any memory barriers.

=C2=B9 I don't do C or other low level languages, so my knowledge in this
area is embarrassingly bad

> Just for reference, the two files I was looking at were:
>
> objects/unicodeobject.c
> objects/bytesobject.c
>
> Extracted from the bz2 downloaded from the page:
>     http://hg.python.org/cpython

I didn't look at bytes first time, I might take a look later.