[boost] Re: N1456 Hash table implementation

From: David Abrahams (dave_at_[hidden])
Date: 2004-01-04 08:26:18

"Wyss, Felix" <FelixW_at_[hidden]> writes:

>> > I think it would be better to mandate that resize()
>> >   1) guarantees iterators remain valid
> [Wyss, Felix] I would argue that resize() is semantically an invariant
> operation on a hash table; it doesn't modify the content.  
> It's an operation to adjust the memory consumption and lookup
> performance tradeoff at runtime.  

Just like std::vector<T>::reserve?  Should that guarantee that
iterators remain valid and not throw exceptions?

> As a previous poster pointed out,
> the SGI implementation guarantees that iterators remain valid.
>
>> >   2) provides the strong exception guarantee
> [Wyss, Felix] Same here, resize() is a semantically invariant operation.

I don't see why that should have any bearing.

> It's easy to implement a hash table that provides the strong exception
> guarantee for resize() by caching the hash code.  
> This even increases performance for types with expensive compares.
> I thus dare asking: Why not?

It incurs a space cost that some people may not want to pay and
decreases performance for some types.

-- 
Dave Abrahams
Boost Consulting
www.boost-consulting.com