Re: C++ PATCH to use hash tables for template specialization lookup

[Paolo Carlini <paolo dot carlini at oracle dot com>]

Hi again,
> For now, I'm collecting some data for the attached. When N grows (I
> tried up to ~1000) the difference between patched / unpatched is
> impressive, like O(N) vs O(N^2). Some numbers later...
>   
actually, for that specific benchmark, on the i7 420 I have here at
hand, the results are extremely neat: an almost perfect linear vs
quadratic behavior: probably there is an explanation for it, I never saw
in the lab such clean numbers in my real world experiments ;) For the
record, I just run the front-end under time and plotted the average of
three user times for each point. In both cases release builds, of
course. Other tests in Dave' book, like memoizing.cpp or lookup_cost.cpp
show very similar improvements, other take very little time anyway. I
must also add that, in some preliminary measurements, ICC 11.1 appear to
behave in a way more similar to 4_4-branch than current mainline, thus,
apparently, no hashing, eh, eh ;)

Paolo.

//////////////////////

Attachment: memoizing2.ps
Description: PostScript document