[PATCH]: Fix PR tree-optimization/20601

[Daniel Berlin]

This is a 4.0 and 4.1 regression.

I missed one of the loop early exit conditions that needed to be tested
before saying a value is the same constant along all edges (IE the loop
exited early with the "all_same" value still true, but not valid because
it hadn't finished checking all the edges)

Fixed with this patch
Jakub came up with a nice execute test that breaks, which i've added to
the testsutie.

Bootstrapped and regtested on the 4.0 branch on i686-pc-linux-gnu
Bootstrapped and regtested on the 4.1 branch on powerpc-linux-gnu

Committed to mainline and 4.0 branch
--Dan
-------------- next part --------------
A non-text attachment was scrubbed...
Name: pr20601.diff
Type: text/x-patch
Size: 3043 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20050323/701abcad/attachment.bin>