Re: [PATCH] Overhaul __builtin_constant_p processing (take 2)

[Neil Booth <neil at daikokuya dot co dot uk>]

Roger Sayle wrote:-

> You'll have to trust me that we'll probably eventually recover this
> loss by combining it with GCSE's local_cprop, but that won't happen
> over night.  I'd be more concerned about the slow downs that we
> haven't predicted, and don't have any plans to remedy.

OK.  However, the slowdown you showed is approximately another 30%
of a CPP pass.  It's really not cheap in wall-time terms; just that
GCC is so slow that 1% or so doesn't appear to be much.  But it's
1% of a bloated number.
 
> Alternatively, my still unreviewed patch "More aggressive jump
> bypassing", http://gcc.gnu.org/ml/gcc-patches/2003-01/msg01164.html,
> knocked over 12 seconds of the bootstrap time of just the C, C++
> and FORTRAN compilers as measured by Andreas' SPEC2000 benchmarking
> scripts when run last June.  Of course, Moore's law also predicts
> that computers are 25% faster now than they were then... :>

Well Moore is wrong; my 350 MHz K6-2 is the exact same speed it was
3 years ago.  I could easily develop GCC on that machine with a complete
bootstrap at that time, when I started, in about 2 - 3 hours.

Now, it takes about 3 or 4 times longer.  That machine has become
useless for routine GCC development.  If I added Java and Ada in,
I think a full bootstrap would be approaching 24 hours.  It's
ridiculous.  Why should I have to keep upgrading hardware just to
make a GCC bootstrap in a reasonable amount of time, when compilers
around 5 years ago could optimize almost as well as GCC in less time,
but only required a 100MHz Pentium?

We need to fix the data structures, the algorithms and (I hope) get
rid of GC, because the current ones clearly aren't working.

I don't want to pick on you; at least you were honest about what effect
your patch had.  Most others seem to go in invisibly.

Neil.