[ast-optimizer-branch] Simplification is pretty expensive

[Jan Hubicka]

> It should not be a surprise, but re-writing the trees is a very
> slow proposition.  I've started tracking the ast branch with the
> SPEC95 tester and the build times are embarrassing (bootstrap
> times are 65% higher).

Note that Andreas is already tracking the branch, so he has some history
about the development, but since the simplification has been broken for
a while, there is just huge gap in the results.
> 
> At least the SPEC scores don't seem to have been affected.  It's

For Andrea's testing they definitly are.  The scores seems to be
consistently worse with -ftree-ssa

> time to start profiling.  Although much of the slowdown could be
> due to silly implementation blunders, we may find opportunities
> in transformations that happen after simplification:
> 
> - The current SSA algorithms are naive.  There are several
>   alternatives that could be used.
> 
> - Maybe we can lobotomize the tree->rtl pass now that it doesn't
>   have to deal with complex trees (granted this is only true for
>   the C front end now).

Is this sutable to speedup (not slowdown) the non-optimizing compilation
where we may want to skip simplification?

Honza
> 
> 
> Diego.