How does copy-on-modify operate on structs and classes?

I understand that Matlab uses "copy-on-modify" so matrices, structs and class objects are only copied inside a function when the item passed in is modified.

Question: When matlab modifies a member of a struct or class in a function, is the entire struct or class object copied, or only the member which was modified?

Why is this important? Because if a struct or class has some very large matrices embedded, then a function which modifies a small part of the entire object would either copy/modify only the small part it works with (good!), or it might duplicate the entire object, even if it only works with a small piece (bad!).

A somewhat trivial example:

s.owner="charlie";
s.data=rand(10000,10000);
s=change_owner(s,"charlene");
. . .
function t = change_owner(t,new_owner);
    t.owner = new_owner;
end

If change_owner(...) sees the entire object as modified, then it has to duplicate the nearly 1 GB data matrix, in order to change the owner field, which makes this function very inefficient. But if matlab recognizes that only the "owner" member is modified, then it can avoid duplicating the data matrix, and thus operate efficiently on the struct.

The same question obviously would apply to a class object as to a struct...

Answer 1

[James Tursa]

Question: When matlab modifies a member of a struct or class in a function, is the entire struct or class object copied, or only the member which was modified?

Only the member being modified is deep copied. The other members are typically reference copied only. E.g., in the following you can see that the data pointer value pr did not change ... i.e. the memory of the data field was not deep copied.

>> s.owner = 'charlie';
>> s.data = rand(1,5);
>> format debug
>> s.data
ans =

Structure address = 54747f90
m = 1
n = 5
pr = 11c661dc0
pi = 0
    0.8147    0.9058    0.1270    0.9134    0.6324
>> s = change_owner(s,'charlene')
s = 
    owner: 'charlene'
     data: [0.8147 0.9058 0.1270 0.9134 0.6324]
>> s.data
ans =

Structure address = 54747f90
m = 1
n = 5
pr = 11c661dc0
pi = 0
    0.8147    0.9058    0.1270    0.9134    0.6324

[Ian]

great. thx.

"format debug" is a new one on me. Thanks for the example.

For anyone else who reads this thread: A little googling turned up the following interesting links on memory management, parameter passing, copy-on-write and "format debug" (which appears to be an undocumented Matlab feature AFAICT).

First, Loren's (2006) post regarding Matlab's memory management has some helpful info both in her explanation and in the comments

https://blogs.mathworks.com/loren/2006/05/10/memory-management-for-functions-and-variables/

Second, some very insightful examples demonstrated by the use of "format debug" are in the links found at

https://groups.google.com/forum/#!searchin/comp.soft-sys.matlab/subject$3A%22Question$20about$20%22format$20debug%22%22

[Ian]

Curious...when I turn on format debug, I don't get info on the imaginary pointer, in either R2018b or R2019a, even when I explicity make data complex or purely imaginary:

>> version

ans =

    '9.6.0.1072779 (R2019a)'

>> format debug
>> s.data='charlie'; s.data=rand(1,3);
>> s.data

ans =


Structure address = 13feab970
m = 1
n = 3
pr = 600004a7b3e0

    0.4854    0.8003    0.1419

>> s.data=s.data + i*rand(1,3)

s = 

  struct with fields:

    owner: 'charlie'
     data: [0.4854 + 0.4218i 0.8003 + 0.9157i 0.1419 + 0.7922i]

>> s.data

ans =


Structure address = 13fea2e10
m = 1
n = 3
pr = 6040020849e0

   0.4854 + 0.4218i   0.8003 + 0.9157i   0.1419 + 0.7922i

>> s.data = i*rand(1,3);
>> s.data

ans =


Structure address = 13fe9bf90
m = 1
n = 3
pr = 60000309ad80

   0.0000 + 0.9595i   0.0000 + 0.6557i   0.0000 + 0.0357i 

[Bruno Luong]

MATLAB stores interleave complex numbers starting from R2018b. So there is no longer Pi pointer.

[Ian]

Ah. Thnx.

I learn something new every day...