Not my goal to bashing or not Stroustrup or to talk too much
about C++ here, but I found this paper that deals a bit with
allocators:
http://www.scs.stanford.edu/~dm/home/papers/c++-new.html
(not sure if already posted in the forum).
It criticizes quite heavily the new operator in C++.
It starts with:
"
These are some notes I have on C++'s operator new. Basically, I
find its syntax downright hateful, and really wish the language
had dealt more cleanly with allocation and construction. I wish
one could pass around function pointers to constructors, give
constructors knowledge of which memory allocators an object was
allocated with, implement a robust debugging malloc, have safer
per-class allocators, or per-class allocators that have access to
constructor arguments. There do exist some slightly gross hacks
for working around some of the problems. In the end, I show how
to avoid using operator new at all. More general constructs in
the language can achieve similar objectives with more
flexibility. You may find the replacement allocator proposed here
fairly disgusting. Just keep in mind that something far worse is
built right into the language.
"
and concludes with:
"
When a programing language doesn't support some necessary
operation, one shouldn't simply add new syntax for that specific
operation. Instead, one should ask, "How can I ammend the
language to make this operation implementable in a standard
library?" The answer may be a much simpler, cleaner, and more
powerful set of mechanisms than the one tailored for a specific
problem.
C++ needed a way to perform type-safe memory allocation. Such a
scheme could have been implemented entirely as a library
function, given the right support. Such support might have
included ways to infer whether classes have destructors at
compile time, support for calling constructors directly and even
taking their addresses, and more. These features might even have
been useful in more contexts than memory allocation.
Instead, Stroustrup introduced operator new, a disgusting piece
of syntax inadequate for the job. After many complaints, new's
functionality was finally extended to include per-class operator
new[] and placement new, still an imperfect solution.
"
Now, why I am interested in the topic: sometimes I feel like it's
OK to let the GC manage the memory, but definitely I am not ready
to give up the deterministic call of destructors. Scoping classes
for that is kinda ugly if not by default (yes, biased opinion).
But, OTOH, maybe it is a confusion in my head that comes from the
fact that "constructing" an object means both allocating and
constructing, while "destructing" means both deallocating and
destructing. I sometimes just feel that construction/destruction
shall be separated form allocation/deallocation.
I am not sure about the impact on optimizations, but this will
simplify delegating memory management to some memory manager of
choice (I think).
