Re: Use of Variants

[Akim Demaille]

Well, I have now better results without Boost Variants.  More
work is still needed, but it is now faster with (Bison) variants
than with the union (which is expected, since the later makes
heavy use of malloc to store pointers to objects instead of objects).

This is in addition of a +10% speed up I had in lalr1.cc by fusing
the three stacks into a single one.

> $ _build/i386-apple-darwin9.4.0/etc/bench.pl --cflags=-O2 -i  
> 400 
>                                                                                         Subroutine 
>  Benchmark::mytime redefined at /System/Library/Perl/5.8.8/ 
> Benchmark.pm line 459.
> Benchmark: timing 400 iterations of f-union, f-union-debug, f- 
> variant, f-variant-debug...
>    f-union: 1.92616 wallclock secs ( 1.14 cusr +  0.48 csys =  1.62  
> CPU) @ 246.91/s (n=400)
> f-union-debug: 1.94717 wallclock secs ( 1.15 cusr +  0.48 csys =   
> 1.63 CPU) @ 245.40/s (n=400)
>  f-variant: 1.86092 wallclock secs ( 1.04 cusr +  0.49 csys =  1.53  
> CPU) @ 261.44/s (n=400)
> f-variant-debug: 1.84779 wallclock secs ( 1.05 cusr +  0.48 csys =   
> 1.53 CPU) @ 261.44/s (n=400)
>                  Rate f-union-debug        f-union f-variant- 
> debug     f-variant
> f-union-debug   245/s            --            -1%              
> -6%           -6%
> f-union         247/s            1%             --              
> -6%           -6%
> f-variant-debug 261/s            7%             6%               
> --           -0%
> f-variant       261/s            7%             6%               
> 0%            --
> Sizes:
>         f-union:    37kB
>       f-variant:    37kB
> f-variant-debug:    42kB
>   f-union-debug:    42kB

There is not even (currently) any significant penalty on the
size.

This is much better than what I have yesterday, with stock Boost
Variants:

>        union: 38.9151 wallclock secs (22.03 cusr +  9.36 csys =  
> 31.39 CPU) @ 233.51/s (n=7330)
>  union-debug: 36.3611 wallclock secs (20.65 cusr +  8.93 csys =  
> 29.58 CPU) @ 231.71/s (n=6854)
>      variant: 47.37   wallclock secs (32.54 cusr +  8.88 csys =  
> 41.42 CPU) @ 169.07/s (n=7003)
> variant-debug: 48.9951 wallclock secs (33.82 cusr +  9.30 csys =  
> 43.12 CPU) @ 169.78/s (n=7321)
>
>               Rate       variant variant-debug   union-debug          
> union
> variant       169/s            --           -0%           
> -27%          -28%
> variant-debug 170/s            0%            --           
> -27%          -27%
> union-debug   232/s           37%           36%             
> --           -1%
> union         234/s           38%           38%             
> 1%            --
>
> Sizes:
>      variant:     108kB
> variant-debug:     120kB
>        union:      65kB
>  union-debug:      69kB