How to organize the project tree for a C++ project using nmake?

Question

There seems to be two major conventions for organizing project files and then many variations.

Convention 1: High-level type directories, project sub-directories

For example, the wxWidgets project uses this style

/solution
   /bin
      /prj1
      /prj2
   /include
      /prj1
      /prj2
   /lib
      /prj1
      /prj2
   /src
      /prj1
      /prj2
   /test
      /prj1
      /prj2

Pros:

• If there are project dependencies, they can be managed from a single file
• Flat build file structure

Cons:

• Since test has it's own header and cpp files, when you generate the unit test applications for EXEs rather than libraries, they need to include the obj files from the application you are testing. This requires you to create inference rules and expand out relative paths for all the source files.
• Reusing any of the projects in another solution requires you to extract the proper files out of the tree structure and modify any build scripts

Convention 2: High-level project dirs, type sub-directories

For example, the Wireshark project uses this style

/solution
   /prj1
      /bin
      /include
      /lib
      /src
      /test
   /prj2
      /bin
      /include
      /lib
      /src
      /test

Pros:

• Projects themselves are self-contained within their folders, making them easier to move and reuse
• Allows for shorter inference rules in the build tools
• Facilitates hierarchical build scripts

Cons:

• If there are dependencies between projects, you need an additional layer of build scripts above the project directories to manage the build order

We currently are using convention 1 on our project and so far it has worked fairly well. Now, I am in the process of adding unit testing (via CxxTest) and facilitating the migration to continuous integration using nmake, Convention 1 is causing some serious headaches in the creation of the proper nmake files.

My primary requirements/goals are:

• Reduce the level of effort to maintain the build scripts of the entire solution.

• De-couple projects and their build steps within a solution from other projects.

• Facilitate continuous integration via the use of build scripts for check-out to release media generation for each commit (obviously leveraging other tools such as CruiseControl as well).

• Make adding or removing additional projects or source files as easy and least error-prone as possible for the developers.

So I ask:

• Are there other pros and cons of either of these methods?
• Is there a clear agrument that favors only one of these conventions?

Answer 1

[A partial answer.]

In "Convention 2: High-level project dirs, type sub-directories," your single con is

If there are dependencies between projects, you need an additional layer of build scripts above the project directories to manage the build order

That can also be viewed as a pro, in many projects.

If you have a lot of repetitive general definitions, one would probably want an include file for the build scripts, where solution-wide constants & parameters could be defined. So the "additional layer of build scripts" will frequently happen anyway, even if there are no (direct) dependencies.

It's a pro in that there's still room for a more modular approach in building. On the other hand, if you want to reuse a project in another, unrelated solution, you would need to compose a different definitions file. (On the other other hand, if there were a single build file for the whole solution, as in Convention 1, you would need a different build script.) As for your maintenance requirement, that's (IMO) very project-dependent.

My feeling leans towards Convention 2, but it's far from a clear win. In fact, your experience with Convention 1, which was working well until recently, may be the biggest pro of all: a team of people with experience with a certain organization is a valuable asset.

Answer 2

Consider using NTFS junction points so you can have both organizations at once. Quick definition: "a junction point is Microsoft's implementation of symbolic links but it only works for directories."

Use Convention 2 for the "real" layout, because it makes the projects easy to move around. Then make a Convention 1 view:

mkdir /solution/test
linkd /solution/test/prj1 /solution/prj1/test
linkd /solution/test/prj2 /solution/prj2/test

Now you have ...

/solution
  /test
    /prj1
    /prj2

... which was the desired result.

You could do the same thing for /src or the other directories if you find it beneficial. Test scripts that benefit from a Convention 1 view live in /solution/test.