kind of regression -fprofile-* between r126538 and r126587

[Thomas Veith]

Hi *,

when compiling crafty-20.14 with

CFLAGS='$(CFLAGS) -Wall -pipe -D_REENTRANT -march=i686 -O3 \
                   -fprofile-use -fbranch-probabilities -fomit-frame-pointer \
                   -ftree-vectorize -ftree-vectorizer-verbose=1 -msse2 \
                   -fno-gcse -mpreferred-stack-boundary=2'

with rev. 125687 crafty breaks severly -- with rev. 126538 it compiles and 
runs smoothly.

As you can see, there are a lot of options involved and crafty is a big 
program so I'm not able to provide a reduced testcase and file a PR in 
bugzilla. But maybe someone can take this as a hint.

Best regards,
Thomas

p.s.: when compiling without -fprofile-use, there is an almost 10% speedup 
from r126538 to r126587.. congrats!!!

r126538 -> Raw nodes per second: 1043604
r126587 -> Raw nodes per second: 1132738