A lot of ICU implementation code writes strings and other contents to buffers that require a fair bit of boilerplate memory management. Usually, a fixed-length stack buffer is used to cover the most common cases. An initial guess for the needed capacity may exceed the stack buffer size, requiring immediate heap allocation. An overflow may occur during processing, requiring heap allocation or reallocation, with or without copying the current contents. Some functions need a destination buffer with the same capacity as one provided on input, such as string case mapping functions when source and destination are the same buffer. When the function exits, a heap-allocated buffer must be freed, and there might be several exit points from the function.
I want to encapsulate all of this into a templatized class which frees the heap memory in the destructor.
This would be ICU-internal, only for arrays of Plain Old Datatypes, and using uprv_malloc() and uprv_free().
Implemented in the trunk after ICU 4.4 Milestone 3 (4.3.3), see MaybeStackArray in cmemory.h. There is also a MaybeStackHeaderAndArray class in ICU 4.6 Milestone 1 (4.5.1), combining a header struct with a contiguously adjacent array.
Not always the best solution
In some cases where MaybeStackArray would be a good fit as a replacement of the existing C or C-style code, it would actually be much simpler to just use UnicodeString, UVector or similar existing C++ classes. (ICU 4.6 Milestone 1 (4.5.1) adds a CharString class for char * strings.) We should try to use those if possible.