[RFC] Merge criteria for tree-ssa

[Jan Hubicka]

> On Friday 06 February 2004 00:24, Jan Hubicka wrote:
> > > Compile Time Performance
> > > ------------------------
> > >
> > >   - Bootstrap times within 1-5% of mainline.
> > >
> > >   - POOMA, DLV and MICO compile times within 1-5% of mainline.
> > >
> > >   - Memory consumption within 1-5% of mainline (POOMA, DLV and
> > >     MICO seem good tests).
> >
> > All these tests hits inline limits.  For instance the tree-ssa estimate
> > gerald's testcase to 160000 instruction, while mainline to 140000
> > instructions.  THat means that tree-ssa happens to inline fewer
> > functions so you get kind of apples-to-orranges comparsion.
> 
> So we need to calibrate the estimate_insn function for GIMPLE trees?

I already did that, but stil the comparsions are simply not comparable.
You can't teach estimate_insns to ignore garbage produced by lowering
and with current organization you can't elliminate the garbage, so you
have to count it and you get greater counts than before.

In general this is not a bug - you get different estimates from trees,
from gimple and from insns and it is dificult to compare them in some
sensible way.

Honza
> 
> Gr.
> Steven