From: Andreas Ericsson <ae@op5.se>
Date: Sun, 09 Sep 2007 01:36:25 GMT
Subject: Re: [RFC] Convert builin-mailinfo.c to use The Better String Library.
Message-ID: <46E34E19.8050402@op5.se>
In-Reply-To: <20070909003718.GE13385@artemis.corp>

Pierre Habouzit wrote:
> On Sat, Sep 08, 2007 at 11:50:34PM +0000, Andreas Ericsson wrote:
> 
>>>> You can tell C compilers to
>>>> check all array accesses, but that is a performance issue.
>>> Runtime checking of arrays in D is a performance issue too, so it is 
>>> selectable via a command line switch.
>> Same as in C then.
> 
>   HAHAHAHAHAHA. Please, who do you try to convince here ? Except in the
> local scope, there is few differences between a foo* and a foo[] in C.
> 

"Runtime checking of arrays is a performance issue." It's true whether it's
done manually by the coder or by the compiler. The difference is that in C,
you get to choose where it should be done.


>>> But more importantly,
>>> 2) For dynamically sized arrays, the dimension of the array is carried
>>> with the array, so loops automatically loop the correct number of times.
>>> No runtime check is necessary, and it's easier for the code reviewer to
>>> visually check the code for correctness.
>> But this introduces handy but, strictly speaking, unnecessary overhead
>> as well, meaning, in short; 'D is slower than C, but easier to write
>> code in'.
> 
>   That's BS. See the strbuf API I've been pushing recently ? It has
> simplified git's code a lot, because each time git had to deal with a
> growing string, it had to deal with at least three variables: the buffer
> pointer, the current occupied length, and its allocated size. That was
> three thing to have variable names for, and to pass to functions.
> 

Yup. I applaud your efforts, but it does come with a slight overhead,
except where it replaces faulty code. In practice, it's probably better
to use the api for all the string-handling, as none of it is performance-
critical.


>   Now instead, it's just one struct. D gives that gratis. There is no
> performance loss because you _need_ to do the same. How do you deal with
> dynamic arrays if you dont't store their lenght and size somewhere ? Or
> are you the kind of programmer that write:
> 
>   /* 640kb should be enough for everyone… */
>   some_type *array = malloc(640 << 10);
> 

No, but it would depend on what I am to do with it.

> 
>> So in essence, it's a bit like Python, but a teensy bit faster and a
>> lot easier to shoot yourself in the foot with.
> 
>> What was the niche you were going for when you thought up D? It can't
>> have been systems programming, because *any* extra baggage is baggage
>> one would like to get rid of. If it was application programming I fail
>> to see how one more language would help, as there will be portability
>> problems galore and it's still considerably slower to develop in than
>> fe Python, while at the same time being considerably easier to mess up
>> in.
> 
>   Right now I'm just laughing. There is for sure overheads in some
> places of D, but the example you take, and what you try to attack in D
> is definitely not where you lose any kind of performance. You could have
> attacked the GC instead (which is after all an easy classical target).
> 

I was asking what role D was designed to fill. I didn't mean it as an
attack, but re-reading what I wrote earlier I see it came off a bit harsh.


>   Just to evaluate the silliness of your arguments:
>   * http://www.digitalmars.com/d/comparison.html so that you can tell
>     what the D features really are,

You may notice that the feature-list is being provided by the creators
and marketeers of the D language. Walter Bright certainly seems like a
nice enough person, but it's possible it's a tad biased.


>   * http://shootout.alioth.debian.org/gp4/benchmark.php?test=all&lang=all
>     so that you can know what the D performance really is about. Of
>     course those are only micro benchmarks, but well, python is "just"
>     15 times slower than D, and D seems to be 10% slower.


I get it to 7.7xC and 1.2xC, respectively, but whatever. It still means
performance-critical apps will be written in C, while
insert-script-language-of-choice will still be used for prototyping and
not-so performance-critical apps.


> Well then I'm
>     okay with D, I'm ready to buy 10% faster CPUs and avoid a lot of
>     painful debugging time. In my world, 10% faster hardware is cheaper
>     by many orders of magnitude than skilled programmers, but YMMV.
> 

I'm curious as to how many fewer bugs D developers write compared to C
programmers. I guess it's hard to do a fair test given the comparatively
shallow pool of D gurus around, but it'd still be interesting to see a
practical test. 20% increase in runtime is certainly acceptable for
never having to see a bug again, but is it acceptable for 10% fewer bugs?
Or 20% fewer?

-- 
Andreas Ericsson                   andreas.ericsson@op5.se
OP5 AB                             www.op5.se
Tel: +46 8-230225                  Fax: +46 8-230231