Loop unroll fixes

[Zoltan Hidvegi <hzoli at hzoli dot 2y dot net>]

Sorry to bug you all with that again, but I wonder if the fix for PR
3384 could be applied to the 3.0 branch and the mainline (on the
mainline the doloop.c patch conflicts with Bernd Smith's change to fix
a subset of the bugs fixed by my patches that also generates a slighly
less efficient code).  See PR 3384, and threads
http://gcc.gnu.org/ml/gcc-patches/2001-07/msg01080.html
http://gcc.gnu.org/ml/gcc-bugs/2001-07/msg00750.html

I think this is a bad regression from 2.95.2.  -funroll-loops is a
powerful and widely used optimization.  GNATS already have a high
priority assigned to this bug, please, someone look at it and either
check it in or do some alrernative fix.  Bootstrapped and passed the
regressions on powerpc-ibm-aix4.3.3 and on i686-pc-linux-gnu, at least
it did some time back just before 3.0.1 was released, but there has
been no change in the files involved since than.

Thanks very much,

Zoli