A little (fairly random) list of coding habit changes I tried in my weekend projects (mainly Oryol) over the past few years and how they worked out, in no particular order. This is all strictly IMHO, don’t take this as ghospel / the-one-and-only-truth ;)

What worked well:

• no setters/getters: I was used to make class members private and write public or protected setter/getter methods. One day I removed all setters/getters from Oryol to see ‘how it feels’, and it just felt right (still does). Class declarations in headers are much easier to grok, and code is more readable. Const-correctness still works, the only thing you lose is the ability to set a breakpoint into the setter/getter methods, but I can count on one (maybe two) hands when I ever needed that.

• no more ‘big ideas’: I stopped long ago believing in ‘big architecture ideas’ that dominate an entire project (Oryol is one result of this, although the ‘old way’ is still visible in places like the Messaging module). Stuff like ‘everything is an object’, ‘everything communicates by message passing’, ‘all objects have names’, ‘all objects live in a big tree’, everything this, everything that… ESPECIALLY if these big ideas are cast into centralized code. At one point you inevitably start working around the big idea because new things just don’t quite fit in anymore. It is better for a project to have many ‘small, local ideas’ and very little shared code. Big ideas can still have their place, but only isolated in small leaf-modules where they cannot infect an entire project.

• no inheritance: I hardly use class inheritance anymore, this is more an unintended side effect of other habit changes (mostly because I stopped caring about ‘designing for code-reuse and extensibility’, both are well-intended, but dangerous lies).

• fixed upper bounds, object pools and pre-allocation: This is hard to put into an existing code base but works very well in Oryol. Whenever there need to be several things of the same type, Oryol has an upper limit of how many of those things can be alive at the same time, either as compile-time option or set at application startup. If a limit is reached, the application is terminated with an error. Usually those upper bounds exist as a ‘rough estimate’ in the mind of the programmer anyway (most programming solutions I’ve seen work well for some ‘intended’ number of items, but completely fall apart if the number of items is a few order-of-magnitudes higher). Defining this upper limit instead of letting everything grow infinitely has (at least) 2 advantages:
  • the application can pre-allocate big chunks of memory at startup which can simplify memory management greatly
  • more obvious leak detection and debugging, if objects of a specific type leak, it is detected early when its pool runs out of free items, instead of much later when the process runs out of memory, it’s also obvious what type of object is leaking

• no pointers: This was not really intended but happened as an automatic side-effect of keeping objects in fixed-size arrays. There are hardly any pointers (of the raw or smart flavour) in Oryol, there are also hardly any calls to new/delete/make_shared/whatever. Instead an ‘object handle’ is usually a simple array index, sometimes with an unique value as ‘dangling reference protection’, or it can be any other type of ‘id’ that can be mapped to one (or more!) memory locations.

• no tiny heap objects, no pointer-chasing: …another side-effect of grouping (and processing) objects in flat arrays. There’s a lot of C++ code out there which creates thousands of tiny objects on the heap, held together by smart pointers and long chains of pointer-chasing call like get_this_obj()->get_that_obj()->do_this(). This is the typical ‘death-by-a-thousand-cuts’ code which is slow without any obvious hotspots and thus nearly impossible to optimize.

• C++11 range-based for loops: I started using this when possible, but still fallback to the old method quite often for any ‘non-trivial’ loops

What kinda worked:

• upper/lower case to indicate public/private: I thought Go’s ‘everything that starts with lower case is private’ is a neat idea, especially since C++’s public/private/friend declaration is quite useless (IMHO). What I want from a public/private mechanism is a hint for an API user that a class is not for ‘public consumption’. I don’t even need to enforce this through a compile error, just a hint is enough. So I started to use upper case for public classes, members and methods, and lower case for ‘private’ stuff. I think this was only partly successful:
  • once ‘the dam was broken’ I started to use all types of naming schemes, camel case, snake case, etc… especially in small sample projects, this looks completely terrible and I need to find a more consistent style again
  • even in Oryol I still mix conventions, in some places, a ‘hard’ private: block makes sense, and in other places all class members are public since they need to be accessible from other places in the same modul, and are only ‘tagged’ as private through the lower case name
• C++11 auto: I’ve started to use ‘auto’ sparingly if the underlying type is either clear from looking at the code, or not at all important. Using auto or not can be a subtle hint to someone reading the code. If I want to make a point that a type is important in a specific place, I don’t use an auto, and the other way around.
• C++11 lambdas (when used wisely): When I absolutely need to pass a function object with captured arguments I usually do this through a lambda now instead of using std::function/std::bind directly (that’s just too ugly). I wouldn’t use this as a ‘big idea’ for building an entire system around it though since std::function objects are quite fat and complex under the hood.

What I don’t want to give up just yet:

• this->: I’m still a fan of using ‘this’ for member access inside class methods, otherwise I find it hard to differentiate between local variables and member variables (and don’t get me started on the m_ notation)

• const correctness: I still think const-correctness is a very useful concept, even though I have to admit that I can’t think of a case ‘where it actually prevented a bug’ (may be that’s exactly the point of using const, I don’t know). I even started to use const for local variables that should not be mutable, but I think that might go a bit too far, especially since I’m not always consistent with this. I would actually love if everything in C++ was const by default and would have to be explicetly marked as mutable, will never happen of course.

What didn’t work:

Only one thing comes to mind, probably because I forgot about stuff that obviously didn’t work pretty fast:

• inline code inside class declaration: I tried for a while to move all inline code up into the class declaration instead of having it isolated below the declaration in ‘standalone methods’. I gave up on this except for very tiny and unimportant classes because it looks just terrible. I consider the header file the most important highlevel documentation of a class, and intermixing the only important part with implementation details just looks wrong. That’s also a gripe I have with other languages that interleave interface declaration with implementation, but I guess I’m in the minority there :)