Re: [Boost-users] [ptr_map] Warning about returning address of local variable

Subject: Re: [Boost-users] [ptr_map] Warning about returning address of local variable
From: Thorsten Ottosen (tottosen_at_[hidden])
Date: 2018-04-05 15:49:41

Den 02-03-2018 kl. 23:09 skrev Bill Buklis via Boost-users:
> I'm not sure if this a potential bug/problem or just a benign warning, 
> but when using the map_values range adaptor with a ptr_map container I 
> get a warning about returning an address of a local variable. It seems 
> benign as the expected processing happens, but this type of warning 
> generally doesn't sound like a good thing.

It's slightly complicatated, but it is from the range library the 
warning is generated. ptr_map<> defines its iterator reference as 
something which is not a reference:

   ptr_container_detail::ref_pair<F,S>

where S is a pointer. The range lib is then parsing such a ref_pair<> by 
value and returning a reference to the second member.

I think this is a bug, sort of, not in the range library, but in ptr_map 
that creates a fake reference type that is not a real reference type.

The reason that ref_pair is used is to avoid exposing the value type as 
pair<Key,void*>, so it's an artifact of the code-generation 
optimization. Perhaps this is the final nail in the coffin for this 
approach and we need to store T* instead of void*.

I'm curious if you can provoke a crash under heavy optimization? Or if 
you're lucky that because your lambda takes a pointer by copy that you 
escape problems.

> As an aside, I was a little surprised when trying out this little sample 
> that the inclusion of the above headers is enough to require linking 
> with the boost regex library. I figure it's probably a reference in the 
> adaptors somewhere. It's not a big deal. These lib files are normally 
> available anyway.

Yes, the tokenized adaptor is based on regexes:

https://www.boost.org/doc/libs/1_66_0/libs/range/doc/html/range/reference/adaptors/reference/tokenized.html

-Thorsten