On Friday, 4 April 2014 at 07:43:18 UTC, Bienlein wrote:
On Thursday, 3 April 2014 at 01:55:48 UTC, Andrei Alexandrescu
wrote:
A lot of them could apply to us as well.
https://www.youtube.com/watch?v=TS1lpKBMkgg
Andrei
He's got a point in mentioning things like "def equals(x: Any):
Boolean" and "def compare(x: T, y: T): Int" (although the
latter is not the worst problem I can think of). But the real
message is to me what is said starting from 24:20:
"There remain those periodic stepwise jumps in performance
taking place in the compiler. ... There is a factor of 10 lying
around. It's that bad.
It's so hart to pinpoint what is doing what and why that ain't
nothing possible to modify. You can't make it fast if you can't
change it."
So build time performance problems in Scala is not simply
because the language has so many more features than Java. There
are real problems in the compiler. What was done in D was to
"stabilize" D and call it D1 and then start on D2. I think this
was a wise thing to do. Maybe for the Scala compiler guys it's
time to
stabilize Scala and call it Scala1 and start with Scala2.
Yeah, generally the message I was getting was that you should be
fighting against piling on new features, fighting against
inelegant hacks for performance, and working on improving the
things that you have.
It's like with his argument for compare. Suppose you had a
typesafe enum which was tri-state. Like a type class. Less,
Equal, More. (Basically 'Ordering' from Haskell, though without
dumb unreadable abbreviations.) You get programs which are more
obviously correct, and there's an obvious efficiency gain to be
had there. If you can prove that your values are only Less,
Equal, or More, you could represent that with exactly -1, 0, 1
internally and then it would be obviously better than the
apparently faster C-like thing.
I think this is a really interesting argument. Don't write ugly
things to get performance. Instead write obviously correct things
and then make obvious optimisations. This argument makes me think
a lot about component programming and then optimising that after
you can prove interesting things about it, like inlining lambdas,
etc.
