Use of anonymous namespaces requires further work/discussion

[tautschnig]

See the post-merge discussion in #929, in particular @peterschrammel's statement.

[pkesseli]

Do we need to make classes (or other kinds of declarations) local to a translation unit?

Yes, I do that frequently. Simple example is if I'm using a small expr_visitort implementation which is only relevant in the current .cpp file, I put it in an anonymous namespace.

Any pointers to that overuse? It's placement at the beginning of a function definition tells you exactly one thing: this function is local to the translation unit.

I'm referring to its differing semantics if it's in a namespace scope vs. when applied to class members. Also, I just find it repetitive:

static int x = 0;
static double y = 1;
static const char DATA[] = "asdf";

Instead, I prefer:

namespace
{
int x = 0;
double y = 1;
const char DATA[] = "asdf";
}

I'm ok with that, as long as a code review on github will show the beginning of the anonymous namespace together with the function definition. Which effectively limites the number of lines to 5.

Limiting their size to 5 lines seems excessive. If we require the semantics of every piece of code to be contained within 5 lines of context, we need to stop using a lot of C++ features (e.g. class access specifiers, virtual functions, ...) :-) .

[pkesseli]

For what it's worth, I don't think it makes sense to refactor out all uses of static for TU-local content. The above comment shows some examples where anonymous namespaces are intuitive or even necessary, but I wouldn't ban the use of static for this purpose either. There are cases, such as single variables or functions, where static is the simpler choice.

I suggest we form a basic guideline along the following:

• Use static for single functions or variables.
• Anonymous namespaces are permitted for:
  • Classes
  • Multiple variables / functions

Basically saying if you only have a single variable or function then use static, otherwise you're allowed to use anonymous namespaces.

CC: @peterschrammel

[tautschnig]

I believe this should be closed via #1031.

[pkesseli]

That's correct. In the CEGIS project, I currently apply the following policy:

• Use anonymous namespaces for local classes
• Use static for local functions and variables.

The current CBMC coding standard permits more liberal use of anonymous namespaces, but I've come around to prefer static for functions and variables after trying it for a while.