[autovect] [patch] another cost-model fix

[Dorit Nuzman <DORIT at il dot ibm dot com>]

In computing the cost of the reduction epilogue there was a bug which
caused us to never take the path of the epilogue that uses vector-shifts
(instead we always took the path of the epilogue that uses scalar
operations). This caused to estimate higher costs for reduction epilogues
on powerpc (in one testcase the cost was computed to be 32, instead of 10).
So now with this fix a couple loops in the powerpc costmodel testsuite get
vectorized when they didn't before.

Bootstrapped with vectorization enabled on powerpc-linux. Committed to
autovect-branch.

dorit

        * tree-vect-transform.c (vect_model_reduction_cost): Fix epilogue
        cost computation.

        * gcc.dg/vect/costmodel/ppc/costmodel-vect-reduc-1char.c: Loops now
        get vectorized.

(See attached file: costmodelfix.july3.txt)

Attachment: costmodelfix.july3.txt
Description: Text document