On Friday, 3 April 2015 at 17:55:00 UTC, Dicebot wrote:
On Friday, 3 April 2015 at 17:25:51 UTC, Ben Boeckel wrote:
On Fri, Apr 03, 2015 at 17:10:31 +0000, Dicebot via Digitalmars-d-announce wrote:
On Friday, 3 April 2015 at 17:03:35 UTC, Atila Neves wrote:
> . Separate compilation. One file changes, only one file > gets rebuilt

This immediately has caught my eye as huge "no" in the description. We must ban C style separate compilation, there is simply no way to move forward otherwise. At the very least not endorse it in any way.

Why? Other than the -fversion=... stuff, what is really blocking this? I personally find unity builds to not be worth it, but I don't see anything blocking separate compilation for D if dependencies are set up


There are 2 big problems with C-style separate compilation:


Complicates whole-program optimization possibilities. Old school object files are simply not good enough to preserve information necessary to produce optimized builds and we are not in position to create own metadata + linker combo to circumvent that. This also applies to attribute inference which has become a really important development direction to handle growing attribute hell.

During last D Berlin Meetup we had an interesting conversation on attribute inference topic with Martin Nowak and dropping legacy C-style separate compilation seemed to be recognized as unavoidable to implement anything decent in that domain.


Ironically, it is just very slow. Those who come from C world got used to using separate compilation to speed up rebuilds but it doesn't work that way in D. It may look better if you change only 1 or 2 module but as amount of modified modules grows, incremental rebuild quickly becomes _slower_ than full program build with all files processed in one go. It can sometimes result in order of magnitude slowdown (personal experience).

Difference from C is that repeated imports are very cheap in D (you don't copy-paste module content again and again like with headers) but at the same time semantic analysis of imported module is more expensive (because D semantics are more complicated). When you do separate compilation you discard already processed imports and repeat it again and again from the very beginning for each new compiled file, accumulating huge slowdown for application in total.

To get best compilation speed in D you want to process as many modules with shared imports at one time as possible. At the same time for really big projects it becomes not feasible at some point, especially if CTFE is heavily used and memory consumption explodes. In that case best approach is partial separate compilation - decoupling parts of a program as static libraries and doing parallel compilation of each separate library - but still compiling each library in one go. That allows to get parallelization without doing the same costly work again and again.


It's true that it's not always faster to compile each module separately, I already knew that. It seems to me, however, that when that's actually the case, the practical difference is negligible. Even if 10x slower, the linker will take longer anyway. Because it'll all still be under a second. That's been my experience anyway. i.e. It's either faster or it doesn't make much of a difference.

All I know is I've seen a definite improvement in my edit-compile-unittest cycle by compiling modules separately.

How would the decoupling happen? Is the user supposed to partition the binary into suitable static libraries? Or is the system supposed to be smart enough to figure that out?


Reply via email to