[Patch] libstdc++/16612

[Paolo Carlini]

Benjamin Kosnik wrote:

>I guess I'm still not sure why this is conditional. Can you explain your
>reasoning here? I don't see the _S_empty_rep checks as being necessary
>for correctness or performance. What am I missing?
>  
>
Indeed, _S_empty_rep is not necessary for correctness, is an optimization.

However, when general applications are concerned (not using shared 
memory, that is), is definitely a *good* optimization. In order to 
appreciate that just construct a testcase involving lots of 
default-constructed strings... Thanks to _S_empty_rep, slow, unnecessary 
allocation of dynamic memory is delayed 'til really necessary.

The proper fix, in my opinion, involves storing short strings in the 
string object itself (that is, both the empty string, consisting only of 
charT() and also larger strings up to, say, 8-16 chars), without using 
dynamic memory (and its slowness), neither static memory (and its 
problems vs 16612). I'm working on that, seems feasible in a reasonable 
amount of time.

That said, perhaps we can add tricks to 3.4/4.0 helping the performance 
when lots of default constructed strings are involved, we'll see soon 
what we can do in this area too... In that case, 
--enable-fully-dynamic-string will go, and we'll be all happier! But 
frankly I doubt that we can really achieve that within the 3.4/4.0 ABI, 
would too good to be true...

Paolo.