When a large data chunk is to be generated by a function, it can be either acquired from the function’s return value or passed by reference or pointer as an in/out variable. According to my development experience, I have summarized the following principles for choosing the appropriate solution.
-
Even though, there may be compiler optimization when a large object is used as the function’s return value, it is still not recommended. Instead, the large object should be passed to the function as a mutable reference, which can hold the returned data.
-
For the sake of improving the overall efficiency by limiting the number of times of memory allocation and releasing, the memory for the returned object should be allocated outside the working function, if its size can be predetermined. In this way, if the memory of this object can be reused, only one time of memory allocation and releasing is needed.
-
For the sake of user convenience, like matrix and vector operations in Octave, it is obvious that the memory allocation is performed inside the working function.
-
A practical design of the numerical library depends on whether it aims to achieve low level operations with high performance or realize a high level user friendly interface. The former prefers memory allocation outside the working function, while the latter inside.
-
LAPACK is a typical example of low level numerical library, where the memory for all data and work spaces is predefined and allocated.