09-Dec-2014 20:54, H. S. Teoh via Digitalmars-d пишет:
On Tue, Dec 09, 2014 at 07:16:56PM +0300, Dmitry Olshansky via Digitalmars-d
wrote:
08-Dec-2014 18:18, H. S. Teoh via Digitalmars-d пишет:
On Mon, Dec 08, 2014 at 08:33:16AM +0000, Russel Winder via Digitalmars-d wrote:
[...]
As with any of these situation the convoluted hardcoded for a
specific processor code, especially assembly language will always
win. I don't care about that, I care about the fastest
comprehensible code that is portable simply by compilation or
execution. Based on this, Java does well, so does some Groovy
perhaps surprisingly, also Scala. C++ does well especially with TBB
(though as an API it leaves a lot to be desired). D is OK but only
using ldc2 or gdc, dmd sucks.
[...]
Yeah, I find in my own experience that gdc -O3 tends to produce code
that's consistently ~20% faster than dmd -O, especially in
compute-intensive code.
And that's not nearly enough. Also both LDC & GDC often can't inline
many functions from phobos due to separate compilation.
[...]
Really? Most of the Phobos function I use are templates, so inlining
shouldn't be a problem, should it? Besides, gdc is far better at
inlining that dmd ever was, though of course there are some constructs
that the front-end doesn't inline, and the backend doesn't have enough
info to do so. This is an area that should be improved.
std.ascii.isWhite ... and there are plenty of things our templates
inevitably unfold to. I mean come on phobos library is big pile of
object code, it can't be all templates.
Last time I checked if you copy-paste isWhite it to your source code it
gets much faster then std one because of inlining.
--
Dmitry Olshansky
