Re: [RFC] O(1) PHI arg lookup - finally bootstrapped & regtested

[Kazu Hirata <kazu at cs dot umass dot edu>]

Hi Diego,

> > Any comments?  Jeff and Diego, I would like a general "go" so that I
> > can start submitting split-up patches.
> > 
> Yes.  This is a step in the right direction for compile time performance
> and it also simplifies the relation between edges and PHI arguments.
> Have you quantified the added memory usage?  Also, testing on PR8361,
> tramp3d and the recent goto-intensive PRs would be interesting.

Thanks!  I'll start generating patches.

Here are more numbers:

Common options: -quiet -O2 -o /dev/null

./cc1 and ./cc1plus are compiled with gcc-3.4.2-6.fc3.  They are *not*
bootstrapped compilers.  --disable-checking was used.

Test machine: Pentium M 1.6GHz with 512MB of RAM

All in seconds.

I obtained peak memory usage using top, so the results are not as
accurate as compilation time.

3 runs of ir.ii (from PR 13776)

      original patched
real:  140.528 140.209 (0.227% down)
user:  139.021 139.073 (0.037% up)
peak:    231MB   232MB

3 runs of generate-3.4.ii (from PR 8361)

      original patched
real:   78.904  78.354 (0.697% down)
user:   78.046  77.761 (0.365% down)
peak:    145MB  140MB

3 runs of _num.i (from PR 18419)

      original patched
real:  137.354 129.151 (5.972% down)
user:  134.470 125.646 (6.562% down)
peak:    244MB  245MB

3 runs of tramp3d-v3.ii

      original patched
real:  189.842 187.262 (1.359% down)
user:  188.123 185.704 (1.302% down)
peak:    237MB   235MB

1 run through cc1-i files

      original patched
real:  253.487 249.985 (1.381% down)
user:  245.793 244.160 (0.664% down)

Kazu Hirata