Fix phi backedge detection in backprop (PR85989)

Fix phi backedge detection in backprop (PR85989)

This PR is a nasty wrong code bug due to my fluffing a test for a
backedge in gimple-ssa-backprop.c.  Backedges are supposed to be
from definitions in the statement we're visiting to uses in statements
that we haven't visited yet.  However, the check failed to account for
PHIs in the current block that had already been processed, so if two
PHIs in the same block referenced each other, we'd treat both
references as backedges.

In more detail:

The first phase of the pass goes through all statements in an arbitrary
order, making optimistic assumptions about any statements that haven't
been visited yet.  The second phase then calculates a correct
(supposedly maximal) fixed point.

Although the first phase order is arbitrary in principle, we still use
the CFG rpo to cut down on the backedges.  This means that the only
thing that's truly arbitrary is the order that we process the PHIs
in a block.  Any order should be OK and should eventually give the
same results.  But we have to follow whatever order we pick,
and the pass wasn't doing that.

2018-06-01  Richard Sandiford  <richard.sandiford@linaro.org>

gcc/
	PR tree-optimization/85989
	* gimple-ssa-backprop.c (backprop::m_visited_phis): New member
	variable.
	(backprop::intersect_uses): Check it when deciding whether this
	is a backedge reference.
	(backprop::process_block): Add each phi to m_visited_phis
	after visiting it, then clear it at the end.

gcc/testsuite/
	PR tree-optimization/85989
	* gcc.dg/torture/pr85989.c: New test.

git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@261064 138bc75d-0d04-0410-961f-82ee72b054a4