Coder: issue replacing variables with constant values

I am experiencing an issue with Coder. If I create a structure 'V' with a field 'A' and assign it a value of 3, and later call V.A in the m-code, the Coder output will not call v.a - it will replace it with "3". A basic example of the current code layout is below:

function [x]=helloworld(y)

persistent myconfig

if isempty(myconfig)

   myconfig =loadmyconfig();

end

[x]=addnums(myconfig)

within loadmyconfig.m:

function [Data]= loadmyconfig ()

   Data = struct(‘junk’,double(1),...
        ‘A’,1,...
        ‘B’,41)

within addnums.m:

function [y]= addnums (myconfig)

y = myconfig.A+ myconfig.B;

would become, when compiled as an executable: within addnums.c:

%function [y]= addnums (myconfig) y = 1+ 41;

I was unable to locate anything in the online help that addresses this issue. Does anyone have any suggestions?

Answer 1

[Arnab De]

If an expression (such as myconfig.A) has a constant value in all executions of the program, MATLAB Coder tries to use that constant value in place of the expression. This optimization makes the generated code efficient. Why do you consider this to be a problem?

[Jared]

In the example, the values of A and B need to be changed outside of the initial version that is compiled, and this is not possible without recompiling the function using the current Coder output.

Essentially we need to deploy this code and have the ability to adjust the values contained within the persistent structure ('myconfig') without recompiling the entire function and redeploying it on the system. This is the whole reason we are not just hard-coding constant values and going through the trouble of creating the myconfig structure.

[Arnab De]

Hi Jared,

I understand your problem now. Please see my answer below.

[A.H.T.Eranga De Silva]

Hi Arnab and Jared,

Could you please let me know the solution for this issue?

I also encountered with the same issue in MATLAB

Thank you...

Eranga

[Adam Danz]

@A.H.T.Eranga De Silva The order of the answers change according to votes and which answer is accepted. He's referencing the other answer on this page.

Answer 2

[Arnab De]

MATLAB Coder only respects the interface of the entry point function. Although you did not specify your codegen command, I guess it was something like

codegen helloworld -args {1} ...;

which specifies helloworld as the only entry point function. Therefore Coder optimized other functions thinking that they are not called from outside. If you want to call loadmyconfig and addnums from outside, you must specify them as entry point functions. Your codegen command should be

a = struct('junk', 1, 'A', 1, 'B', 1); 
codegen loadmyconfig addnums -args {a} ...;

The GUI also lets you specify multiple entry points. You do not need to add helloworld unless you need it. Do not add functions that you are not going to call from outside as entry points, because you'll lose out on optimizations.

BTW, you mentioned that you compiled the MATLAB code to executable. However, it appears that for your purpose, it will be better to compile them to a static/dynamic library. Also, it is recommended that you first test your code by converting it to mex.

[Jared]

Thanks for the reply. Unfortunately, specifying two entry point functions does not correct the issue illustrated in the first post. Also, compiling to a library as opposed to an executable returns the same results.

I have tested the compiled executable, and the results obtained using the compiled C do match the MATLAB results. The only issue appears to be this "optimization" that is overwriting variable names with constants.

[Arnab De]

Hi Jared,

I tried the program you gave us and specifying multiple entry points does work for me. Probably there is still some miscommunication. Can you please give us a small program and the exact codegen command you are using? Also, paste the signatures of the generated functions. Which version of MATLAB are you using?

[Jared]

Here is the m-code:

% hello world

function x=helloworld(y)

persistent myconfig

myconfig=loadmyconfig();

x=addnums(myconfig)+y;

end

% load my config

function Data=loadmyconfig()

Data=struct('junk',double(1),...

    'A',1,...
    'B',41);

end

% add nums

function y=addnums(myconfig)

y=myconfig.A+myconfig.B;

Calling the codegen command in the order previously listed generates an error. The code below is what I used instead:

y=double(7);

cfg = coder.config('exe');

cfg.CustomSource = 'main.c';

codegen -config cfg helloworld -args {y} loadmyconfig

This generates (helloworld.c):

real_T helloworld(real_T y) { return 42.0 + y; }

I am using 2012b.

[Arnab De]

As I said, Coder only respects the interface of the entry point functions. You specified helloworld and loadmyconfig as entry points and it did not change the signatures of those functions --- see helloworld.h/.c and loadmyconfig.h/.c. If you added addnums to your entry point list, it would have preserved the interface of addnums too (it would be in addnums.h/.c). Additionally, the structure would be declared in helloworld_types.h. However, the compiler is free to optimize the codes of these functions in any way it sees fit. In your case, you'll be calling the functions loadmyconfig and addnums from outside --- so as long as the interface of these functions are preserved, it should work for you. It does not matter if it has optimized calls to these functions in helloworld.c. Run the codegen command

 a = struct('junk', 1, 'A', 1, 'B', 1); 
 codegen -config:lib helloworld -args {1} loadmyconfig addnums -args {a};

You should find the generated files in codegen/lib/helloworld/.

I also suggest that you go through the tutorials and documentation once. You can also contact MathWorks tech support.