Use unsigned arithmetic for demoted vector plus/minus/mult (PR 88064)

Use unsigned arithmetic for demoted vector plus/minus/mult (PR 88064)

As Jakub pointed out, if we narrow a plus, minus or mult operation based
on the number of bits that consumers need, we have to convert a signed
operation to an unsigned one in order to avoid new undefined behaviour.
This patch does that and generalises vect_convert_input and
vect_recog_over_widening_pattern to cope with the extra casts.
(The changes to both functions are covered by existing tests.)

2018-12-03  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
	PR tree-optimization/88064
	* tree-vect-patterns.c (vect_convert_input): Convert the result of
	an existing cast if it has the right width but the wrong sign.
	Do not test the signedness of the required result when
	considering whether to split an existing cast; instead split to
	a type with the same signedness as the source of the cast, then
	convert it to the opposite signedness where necessary.
	(vect_recog_over_widening_pattern): Handle sign changes between
	the final PLUS_EXPR and the RSHIFT_EXPR.
	(vect_recog_average_pattern): Use an unsigned operation when
	truncating an addition, subtraction or multiplication.  Cast the
	result back to the "real" signedness before promoting.

gcc/testsuite/
	PR tree-optimization/88064
	* gcc.dg/vect/vect-over-widen-23.c: New test.

git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@266829 138bc75d-0d04-0410-961f-82ee72b054a4