Re: [Boost-users] [container.string] why isn't memcpy used in the internals?

Subject: Re: [Boost-users] [container.string] why isn't memcpy used in the internals?
From: Ion Gaztañaga (igaztanaga_at_[hidden])
Date: 2014-02-10 03:54:36

El 10/02/2014 9:18, Adam Romanek escribió:
> Hi!
>
> Recently I've been trying to improve the performance of a code that
> makes heavy use of std::string (lots of constructions and destructions).
> My first thought was to use boost::container::string as it does not use
> reference-counting and uses small-string optimization, so it should be a
> perfect replacement for std::string when it comes to performance.
>
> Right... But the problem is that switching to boost::container::string
> only made things worse in the general case. I wrote "in the general
> case" because when strings to be held are small and SSO may be applied
> then boost::container::string outperforms std::string considerably.
> However, in the general case constructing boost::container::string from
> a C string is significantly slower than constructing std::string.
>
> My investigation showed that std::string uses memcpy to copy the C
> string into its buffer while boost::container::string does not. The
> memcpy comes from std::char_traits::copy so maybe
> boost::container::string could use it too? I guess the rest of
> operations from std::char_traits (find, move, assign etc.) could be
> utilized too as std::char_traits has been carefully crafted with
> performance in mind.
>
> Should I create a ticket for this issue?

Yes, please. I'm slowly trying to benchmark and improve Boost.Container 
performance, I haven't started optimizing basic_string so your ticket 
could be a good start point.

In a quick inspection I see that the constructor is dispatched to 
basic_string::assign(It, It) and that assignment is not optimized at 
all. I should definitely write a assignment function that dispatches to 
std::char_traits::copy, as it could be used to optimize several member 
functions.

Best,

Ion