[PATCH] fold-const.c use of BRANCH_COST

[David Edelsohn]

>>>>> Richard Henderson writes:

>> Yes, BRANCH_COST=1 really does degrade 200.sixtrack that much -- it is
>> completely repeatable.

Richard> But why would you want to set BRANCH_COST that low?

	Because one of the design points for the POWER/PowerPC
architecture was "branches are free" and good branch prediction makes the
cost very low.  I do not want to set BRANCH_COST that low, but it is hard
to argue when setting BRANCH_COST=1 does improve performance of some
testcases.

	The issue is what effect of BRANCH_COST=1 is actually improving
performance.  My investigation lead me to fold_range_test.

	I would like to change the heuristic used by fold_range_test for
PowerPC.  Instead of arguing that the threshold should be some arbitrary
value greater than the setting of BRANCH_COST on PowerPC, it seems more
useful to add port-specific finer granularity of control.

Thanks, David