PATCH: tune ARM's rtx_costs function

[James Lemke]

> Why are these times so much higher than when optimizing (nearly double)?
> 
> Can you tell me what the following gives at -O0?

Some other changes went in between when I ran with -O1 and -O0.  I was
focused on differences between patches, but you're right the times grew.

I have re-run all the tests with exactly the same version of compiler
source, except for the subject lines in arm.md.

=============
For "time make CXXFLAGS_FOR_TARGET=-O0 all-target-libstdc++-v3":

Ref time with (no_new_pseudos ? 0 : preserve_subexpressions_p ())..
real    11m31.243s
user    7m40.630s
sys     3m34.570s

Time with (!no_new_pseudos)..
User time is about the same (-0.03%).
real    11m41.039s
user    7m40.470s
sys     3m38.640s

Time with (optimize && !no_new_pseudos)..
User time is about the same (+0.4%).
real    11m39.165s
user    7m42.840s
sys     3m36.360s

=============
For "time make CXXFLAGS_FOR_TARGET=-O1 all-target-libstdc++-v3":
User time is....

Ref time with (no_new_pseudos ? 0 : preserve_subexpressions_p ())..
real    12m30.848s
user    8m31.350s
sys     3m33.970s

Time with (!no_new_pseudos)..
User time is about the same (-0.2%).
real    12m19.366s
user    8m30.320s
sys     3m36.060s

-- 
James Lemke   jim@wasabisystems.com   Orillia, Ontario
http://www.wasabisystems.com