emit SLP vectorized loads earlier

emit SLP vectorized loads earlier

This makes sure to emit SLP vectorized loads where the first scalar
load is.  This makes SLP dependence checking more powerful because
hoisting loads can use TBAA and it increases the freedom for
vector placement when there are constraints from live lanes.

Vectorized shifts block inserting vectorized stmts always after
vectorized defs because it ends up using the original scalar
operand even when the SLP graph indicates the shift operand
is vectorized (and we actually emit and cost those stmts).

vect_slp_analyze_and_verify_node_alignment shows we need alignment
for too many places, this is a temporary solution and my plan
is to have a single meta-info for a dataref group instead
(also getting rid of DR_GROUP_FIRST/NEXT_ELEMENT).

2020-06-24  Richard Biener  <rguenther@suse.de>

	* tree-vectorizer.h (vect_find_first_scalar_stmt_in_slp):
	Declare.
	* tree-vect-data-refs.c (vect_preserves_scalar_order_p):
	Simplify for new position of vectorized SLP loads.
	(vect_slp_analyze_node_dependences): Adjust for it.
	(vect_slp_analyze_and_verify_node_alignment): Compute alignment
	for the first stmts dataref.
	* tree-vect-slp.c (vect_find_first_scalar_stmt_in_slp): New.
	(vect_schedule_slp_instance): Emit loads before the
	first scalar stmt.
	* tree-vect-stmts.c (vectorizable_load): Do what the comment
	says and use vect_find_first_scalar_stmt_in_slp.