Although the first pattern match wins in the sense that no later
function can match the *old* gimple statement, it still seems worth
letting them match the *new* gimple statements, just like we would if
the original IR had included that sequence from the outset.

This is mostly true after the later patch for PR85694, where e.g. we
could recognise:

   signed char a;
   int ap = (int) a;
   int res = ap * 3;

as the pattern:

   short ap' = (short) a;
   short res = ap' * 3;     // S1: definition statement
   int res = (int) res;     // S2: pattern statement

and then apply the mult pattern to "ap' * 3".  The patch needs to
come first (without its own test cases) so that the main over-widening
patch doesn't regress anything.

2018-06-30  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
	* gimple-iterator.c (gsi_for_stmt): Add a new overload that takes
	the containing gimple_seq *.
	* gimple-iterator.h (gsi_for_stmt): Declare it.
	* tree-vect-patterns.c (vect_recog_dot_prod_pattern)
	(vect_recog_sad_pattern, vect_recog_widen_sum_pattern)
	(vect_recog_widen_shift_pattern, vect_recog_rotate_pattern)
	(vect_recog_vector_vector_shift_pattern, vect_recog_divmod_pattern)
	(vect_recog_mask_conversion_pattern): Remove STMT_VINFO_IN_PATTERN_P
	checks.
	(vect_init_pattern_stmt, vect_set_pattern_stmt): New functions,
	split out from...
	(vect_mark_pattern_stmts): ...here.  Handle cases in which the
	statement being replaced is part of an existing pattern
	definition sequence, inserting the new pattern statements before
	the original one.
	(vect_pattern_recog_1): Don't return a bool.  If the statement
	is already part of a pattern, instead apply pattern matching
	to the pattern definition statements.  Don't clear the
	STMT_VINFO_RELATED_STMT if is_pattern_stmt_p.
	(vect_pattern_recog): Don't break after the first match;
	continue processing the pattern definition statements instead.
	Don't bail out for STMT_VINFO_IN_PATTERN_P here.

From-SVN: r262275