Comment by ponce | Hacker News Reader

Comment by ponce | Hacker News Reader

Parent Full thread

There is an awful lot of problems with C++ though.

- The tool are certaintly not rock solid. Parsing requires arbitrary look-ahead and semantic pass, order of headers inclusion matters, #if and dirty tricks are everywhere

- compilation is slow

- non standard extensions makes painful portable C++ across compilers and platform, because you need them and they can have incompatible syntax

- C++ const is useless when it comes to optimization or concurrency.

- any real world project involving C++ is going to be challenging if you're not working with top C++ library implementers or emasculating rules (like "no exceptions" or "nothing goes in the constructor" at Google)

Anyway, even if C++1x was already here and completely supported (which C++03 is not), D would still have an edge.

Despite its current problems D greatly enhance productivity over C++. Refactoring is _way_ faster when you have import and no headers to maintain. Properties enhance expressivity a lot. You have access to the machine, value semantics, and reference returns, so there is nothing conceptually wrong with it.

Also, in D there is much less skill diffference between library writers and typical programmers as compared to C++.

Finally if you really want C++1x features, I sincerely advises to try D because there is a lot of overlapping.

> - compilation is slow

What I've learned is that people don't really care about full builds when doing large projects. The separation into header and source files means that when you change a source file, only this one has to be recompiled. You only do massive rebuilds when modifying headers.

In case of D, when you modify a module, the change can possibly propagate indefinitely. Even when you just change the body of a function and think the modification should be isolated, consider that the build tool must assume this function may potentially be used at compile-time to generate code for other modules. The changes propagate as far as they would with .h files (unless the build tool can do semantic analysis or make unsafe assumptions).

To this fact, add how DMD only outputs template instances into one object file when compiling multiple modules at the same time. If you're lucky, at the next incremental build it will output them into the same modules. If you're not, you end up with unresolved references. This template emission method is an optimization - the alternative is a lot of bloat that object files generated by C++ compilers suffer from. I agree with Walter's position that what DMD does should be the default. And it is possible to make a good build tool that deals with this issue. I've tried doing that for DMD-Win, but its case, such a build tool is currently infeasible due to toolchain issues.

You can do "header" files with D, too. They're called .di files and the compiler will even generate them automatically, or they can be built by hand.

The compile speed advantage D has over C/C++ is that in the latter the header files must be reparsed for every source file. In D, the header files only need to be reparsed once, and then it is looked at symbolically for the rest of the source modules.

D is also faster at compiling because the lexical grammar is designed to require little processing.

That latter advantage is true, but D's "header" files are currently impractical. Creating them by hand is a major source of pain, greater than with C/C++. The automatic generation on the other hand is troublesome for at least two reasons:

1) It's very sensitive to the contents of .d files. For instance, It leaves the bodies of inline-able functions in, thus if you modify such a function in the .d file, it will again mean a change spill to the .di and modules importing it (directly and indirectly). Similarly if you just reorder the functions in a module or change a private constant. This could partially be mitigated by a sufficiently smart build tool that tracks symbol importing, aliasing and usage in order to infer the modules that need to be recompiled. Sadly, this is quite complex and no current tool does it at the moment.

2) It assumes none of the functions will be used at compile-time (by stripping away function bodies depending on whether they may be inlined), which forces you to stuff them all into compile-time-only modules.

Reply on news.ycombinator.com