Re: What to do about compile time slowdowns for 4.1?

[Daniel Berlin <dberlin at dberlin dot org>]

On Tue, 2005-08-30 at 15:37 +0200, Sebastian Pop wrote:
> Steven Bosscher wrote:
> > 

> tree level: 7 times
> Breakpoint 3, number_of_iterations_cond (type=0x2aaaaadf84d0, base0=0x2aaaaae03d80, step0=0x0, code=LE_EXPR, base1=0x2aaaaafd5b90, step1=0x2aaaaafc8d20, 
>     niter=0xc174b0) at tree-ssa-loop-niter.c:436
> 
> rtl level: once
> Breakpoint 4, iv_number_of_iterations (loop=0xc324b0, insn=0x2aaaaafdf730, condition=0x2aaaaafe31a0, desc=0x7fffff912ff0) at loop-iv.c:2416
> 
> It seems to me that we have to reuse the computed information instead
> of throwing away all the results and all the estimations at the end of
> each tree level loop optimization...  it probably was a good idea when
> there were a few loop optimizations, but discarding loop informations
> just because we have probably invalidated a part of the ssa_names is
> not the way I want to walk.

I agree completely.
We need to be a lot smarter about invalidating certain types of
expensive to compute information.  Another good example of where we do
this is alias information.

I'm working on this from the alias analysis side of things, but right
now it takes a back seat to improving the actual information we expose,
etc.