This is what I finally came up with:
I can check before compilation which toolboxes are needed and call license() accordingly. Or I can implement a built-in check into the executable itself. (Called with a special parameter after compilation triggers a self-check for available toolboxes.) In either case I need the names of the required toolboxes.
I tried several ways to generate a list of toolboxes. In short: Running the program in Matlab and then entering license('inuse') is not very reliable. depfun() descends way to much. mydepfun() and fdep() do not descend enough.
I think the problem with mydepfun() and fdep() is that they do not descend into the \toolbox\shared folder. So I took mydepfun() from Tobias Kienzler (link to the original sources) and modified it:
function [list,callers,tboxes_found] = i_scan(f)
func = i_function_name(f);
[list,~,~,~,~,~,callers,~] = depfun(func,'-toponly','-quiet');
toolboxroot = fullfile(matlabroot,'toolbox');
sharedroot = strcat(toolboxroot, filesep, 'shared');
intoolbox = strncmpi(list,toolboxroot,numel(toolboxroot));
inshared = strncmpi(list,sharedroot, numel(sharedroot));
tboxes_found = list(intoolbox & ~inshared);
tboxes_found = regexpi(tboxes_found, '[\\/]toolbox[\\/](.+?)[\\/]', 'tokens');
tboxes_found = cellfun(@(cfun) cfun{1}, tboxes_found);
list = list(~intoolbox | inshared);
callers = callers(~intoolbox | inshared);
for jj = 1:numel(list)
c = callers{jj};
cs = cell(numel(c),1);
for kk = 1:numel(c)
cs{kk} = list{c(kk)};
end;
callers{jj} = cs;
end;
This way i_scan(f) is returning the toolboxes and also descends into \toolbox\shared. The main function of mydepfun() just collects the toolboxes:
function [filelist,callers,toolboxes] = mydepfun(fn,recursive)
.
.
toolboxes = {};
[filelist,callers,tboxes_found] = i_scan(foundfile);
toolboxes = [toolboxes; tboxes_found];
.
.
[newlist,newcallers,tboxes_found] = i_scan(toscan{1});
toolboxes = [toolboxes; tboxes_found];
.
.
toolboxes = unique(toolboxes);
The toolboxes listed are the ones our source code uses. The modified mydepfun() seems to work fine. (Apart from the typical problems caused by elements only resolved during run time like eval(), function handles, callbacks etc.)
And: The dependeny walkers I have seen - like mydepfun() - are using depfun() inside. depfun() is not reliable since it silently ignores all source code not on the path (also its return prob_files is empty in this case). So care has to be taken that the Matlab path is set correctly. (Also any additional pathes are problematic since Matlab may take unexpectedly functions with the same name from other locations.)
After all, I think, this is a good way to make my build process more reliable.
/Zweikeks