remove SLP_TREE_TWO_OPERATORS, add SLP permutation node

This removes the SLP_TREE_TWO_OPERATORS hack in favor of having
explicit SLP nodes for both computations and the blend operation.
For this introduce a generic merge + select + permute SLP node
(with implementation limits).

Building upon earlier patches it adds vect_stmt_dominates_stmt_p
and the ability to compute a vector insertion place from
vectorized stmts (which now have UID zero) as needed for
the permute node.

2020-06-17  Richard Biener  <rguenther@suse.de>

	* tree-vectorizer.h (_slp_tree::two_operators): Remove.
	(_slp_tree::lane_permutation): New member.
	(_slp_tree::code): Likewise.
	(SLP_TREE_TWO_OPERATORS): Remove.
	(SLP_TREE_LANE_PERMUTATION): New.
	(SLP_TREE_CODE): Likewise.
	(vect_stmt_dominates_stmt_p): Declare.
	* tree-vectorizer.c (vect_stmt_dominates_stmt_p): New function.
	* tree-vect-stmts.c (vect_model_simple_cost): Remove
	SLP_TREE_TWO_OPERATORS handling.
	* tree-vect-slp.c (_slp_tree::_slp_tree): Amend.
	(_slp_tree::~_slp_tree): Likewise.
	(vect_two_operations_perm_ok_p): Remove.
	(vect_build_slp_tree_1): Remove verification of two-operator
	permutation here.
	(vect_build_slp_tree_2): When we have two different operators
	build two computation SLP nodes and a blend.
	(vect_print_slp_tree): Print the lane permutation if it exists.
	(slp_copy_subtree): Copy it.
	(vect_slp_rearrange_stmts): Re-arrange it.
	(vect_slp_analyze_node_operations_1): Handle SLP_TREE_CODE
	VEC_PERM_EXPR explicitely.
	(vect_schedule_slp_instance): Likewise.  Remove old
	SLP_TREE_TWO_OPERATORS code.
	(vectorizable_slp_permutation): New function.

1 files changed, 9 insertions, 4 deletions

diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
index 6c830ad..a200991 100644
--- a/gcc/tree-vectorizer.h
+++ b/gcc/tree-vectorizer.h
@@ -135,6 +135,10 @@ struct _slp_tree {
   /* Load permutation relative to the stores, NULL if there is no
      permutation.  */
   vec<unsigned> load_permutation;
+  /* Lane permutation of the operands scalar lanes encoded as pairs
+     of { operand number, lane number }.  The number of elements
+     denotes the number of output lanes.  */
+  vec<std::pair<unsigned, unsigned> > lane_permutation;
 
   tree vectype;
   /* Vectorized stmt/s.  */
@@ -151,12 +155,12 @@ struct _slp_tree {
   /* The maximum number of vector elements for the subtree rooted
      at this node.  */
   poly_uint64 max_nunits;
-  /* Whether the scalar computations use two different operators.  */
-  bool two_operators;
   /* The DEF type of this node.  */
   enum vect_def_type def_type;
   /* The number of scalar lanes produced by this node.  */
   unsigned int lanes;
+  /* The operation of this node.  */
+  enum tree_code code;
 };
 
 
@@ -195,11 +199,12 @@ public:
 #define SLP_TREE_VEC_DEFS(S)                     (S)->vec_defs
 #define SLP_TREE_NUMBER_OF_VEC_STMTS(S)          (S)->vec_stmts_size
 #define SLP_TREE_LOAD_PERMUTATION(S)             (S)->load_permutation
-#define SLP_TREE_TWO_OPERATORS(S)		 (S)->two_operators
+#define SLP_TREE_LANE_PERMUTATION(S)             (S)->lane_permutation
 #define SLP_TREE_DEF_TYPE(S)			 (S)->def_type
 #define SLP_TREE_VECTYPE(S)			 (S)->vectype
 #define SLP_TREE_REPRESENTATIVE(S)		 (S)->representative
 #define SLP_TREE_LANES(S)			 (S)->lanes
+#define SLP_TREE_CODE(S)			 (S)->code
 
 /* Key for map that records association between
    scalar conditions and corresponding loop mask, and
@@ -1932,6 +1937,6 @@ void vect_pattern_recog (vec_info *);
 unsigned vectorize_loops (void);
 void vect_free_loop_info_assumptions (class loop *);
 gimple *vect_loop_vectorized_call (class loop *, gcond **cond = NULL);
-
+bool vect_stmt_dominates_stmt_p (gimple *, gimple *);
 
 #endif  /* GCC_TREE_VECTORIZER_H  */