John Carmack applauds D's pure attribute

[Andrew Wiley]

On Sun, Feb 26, 2012 at 11:05 AM, Paulo Pinto <pjmlp at progtools.org> wrote:
> Am 26.02.2012 17:34, schrieb so:
>
>> On Sunday, 26 February 2012 at 15:58:41 UTC, H. S. Teoh wrote:
>>
>>> Would this even be an issue on multicore systems where the GC can run
>>> concurrently? As long as the stop-the-world parts are below some given
>>> threshold.
>>
>>
>> If it is possible to guarantee that i don't think anyone would bother
>> with manual MM.
>
>
> Well, some game studios seem to be quite happy with XNA, which implies using
> a GC:
>
> http://infinite-flight.com/if/index.html
>

I don't really see why you keep bringing up these examples. This is a
performance issue, which means you can certainly ignore it and things
will still work, just not as well. I've seen 3d games in Java, but
they always suffer from an awkward pause at fairly regular intervals.
This is why the AAA shops are still writing most of the engines in
C++.
You will always be able to find examples of developers that simply
chose to ignore the issue for one reason or another.

To make it clear, I'm not trying to antagonize you here. I agree that
GC is in general a superior technical solution to manual memory
management, and given the research going into GC technology, I'm sure
that long term it's probably a good idea.

However, I disagree with your statement that "the main issue is that
the GC needs to be optimized, not that manual memory management is
required."
Making a GC that can run fast enough to make this sort of thing a
non-issue is currently so hard that it can only be used in certain
niche situations. That will probably change, but it will probably
change over the course of several years. Manual memory management,
however, is here now and dead simple to use so long as the programmer
understands the semantics. Programming in that model is harder, but
not nearly as bad as, say, thread-based concurrency with race
conditions and deadlock. Manual memory management is much simpler to
deal with than many other things programmers already take on
voluntarily.
When you want your realtime application to behave in a certain way,
would you rather spend months or years working on the GC and program
in a completely difficult style to deal with the issue, or use manual
memory management *now* and deal with the slightly more difficult
programming model? Cost/benefit wise, GC just doesn't make a lot of
sense in this sort of scenario unless you have a lot of resources to
burn or a specific reason to choose a GC-mandatory platform.

Again, I'm not saying GC is bad, I'm saying that in this area, the
cost/benefit ratio doesn't say you should spend your time improving
the GC to make things work. For everyone else, GC is great, and I
applaud David Simcha's efforts to improve D's GC performance.