Re: [Boost-users] OT: std::map<>::operator[] and default constructor

From: Howard Hinnant (hinnant_at_[hidden])
Date: 2005-02-15 11:56:05

On Feb 15, 2005, at 11:09 AM, Noel Yap wrote:

> Please direct me to a more appropriate list if this is too off topic:
>
> It's come to my attention that our vendor's std::map<>::operator[] 
> /always/ calls the default constructor of the key even when the key 
> already exists in the map.
>
> This is what they say:
>
>     It turned out to be a documentation bug. Our implementation is in 
> sync
>     with the standard.
>
>     From 23.3.1.2 [lib.map.access]:
>
>     T& operator[](const key_type& x);
>
>     -1- Returns: (*((insert(make_pair(x, T()))).first)).second.
>
>     Notice the T() in the call to make_pair() above: that's an 
> invocation of
>     the default ctor of the mapped type (i.e., Aoeu in the user's test
>     case). So the program behaves as expected. If you do not like this
>     behavior, replacing the expression map[16] in the program with 
> map.find
>     (16)->second will avoid the construction of the unused temporary. I
>     modified the test case to incorporate this change.
>
>
> My inclination is to say the standard is a spec, not an 
> implementation.  Allowing std::map<>::operator[] not to call the 
> default constructor requires no more and promises no less than the 
> spec.  All this is true unless, of course, there's some order of 
> complexity issue I'm missing.  But someone else had said, "an 
> implementation has to have the same 'observable behaviour' as the 
> abstract machine defined in the standard.  Calling the default ctor 
> definitely counts as observable behaviour, so I don't think RW have 
> any choice but to implement it this way".

I believe RW is technically correct.  However there is a DR on this 
issue with WP status:

http://www.open-std.org/jtc1/sc22/wg21/docs/lwg-defects.html#334

The proposed resolution no longer requires calling the default 
constructor (nor does it forbid it).  WP status means that the 
committee has officially voted in favor of this resolution, but it is 
not yet official.  It has been applied to the draft working paper for 
C++0X.

The Metrowerks implementation has purposefully gone non-standard on 
this issue (until C++0X becomes official) as we believe it results in a 
higher quality library for our customers.  I.e. we call the default 
constructor only if it doesn't already exist in the map.

> What's the best place to get an official answer to the real intent of 
> std::map<>::operator[]?  Must it /always/ call the default 
> constructor?

The newsgroup comp.std.c++ .

-Howard