[v3] DR 282 and libstdc++/4402

[Benjamin Kosnik]

> My only complaint is all those (repeated!) calls to numpunct virtuals 
> in each format operation.  Those details should be cached on the first 
> format operation, and re-used thereafter until the next imbue call.  
> There used to be infrastructure for such caching; did it get stripped
> out?

There is no performance analysis to support your assertion. Please provide 
it. Which (repeated!) calls are you referencing?

The old caching code was removed. It was incomplete and, as you yourself 
commented to me last we talked about it, largely experimental. I thought 
it best to not constrain the library ABI to this design.

Let's get complete char/wchar_t functionality and then do caching where 
necessary, based on real numbers not hand waving... 

-benjamin