CbC/CbC_gcc: gcc/tree-vect-loop-manip.c comparison

comparison gcc/tree-vect-loop-manip.c @ 145:1830386684a0

gcc-9.2.0

author	anatofuz
date	Thu, 13 Feb 2020 11:34:05 +0900
parents	84e7813d76e9
children

comparison

equal deleted inserted replaced

-:84e7813d76e9
+:1830386684a0
 /* Vectorizer Specific Loop Manipulations
-Copyright (C) 2003-2018 Free Software Foundation, Inc.
+Copyright (C) 2003-2020 Free Software Foundation, Inc.
 Contributed by Dorit Naishlos <dorit@il.ibm.com>
 and Ira Rosen <irar@il.ibm.com>
 This file is part of GCC.
 #include "tree-ssa-loop-niter.h"
 #include "internal-fn.h"
 #include "stor-layout.h"
 #include "optabs-query.h"
 #include "vec-perm-indices.h"
+#include "insn-config.h"
+#include "rtl.h"
+#include "recog.h"
 /*************************************************************************
 Simple Loop Peeling Utilities
 Utilities to support loop peeling for vectorization purposes.
 gimple *stmt;
 use_operand_p use_p;
 ssa_op_iter iter;
 edge e;
 edge_iterator ei;
-struct loop *loop = bb->loop_father;
+class loop *loop = bb->loop_father;
-struct loop *outer_loop = NULL;
+class loop *outer_loop = NULL;
 if (rename_from_outer_loop)
 {
 gcc_assert (loop);
 outer_loop = loop_outer (loop);
 /* Define one loop mask MASK from loop LOOP.  INIT_MASK is the value that
 the mask should have during the first iteration and NEXT_MASK is the
 value that it should have on subsequent iterations.  */
 static void
-vect_set_loop_mask (struct loop *loop, tree mask, tree init_mask,
+vect_set_loop_mask (class loop *loop, tree mask, tree init_mask,
 		    tree next_mask)
 {
 gphi *phi = create_phi_node (mask, loop->header);
 add_phi_arg (phi, init_mask, loop_preheader_edge (loop), UNKNOWN_LOCATION);
 add_phi_arg (phi, next_mask, loop_latch_edge (loop), UNKNOWN_LOCATION);
 }
 /* Add SEQ to the end of LOOP's preheader block.  */
 static void
-add_preheader_seq (struct loop *loop, gimple_seq seq)
+add_preheader_seq (class loop *loop, gimple_seq seq)
 {
 if (seq)
 {
 edge pe = loop_preheader_edge (loop);
 basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq);
 }
 /* Add SEQ to the beginning of LOOP's header block.  */
 static void
-add_header_seq (struct loop *loop, gimple_seq seq)
+add_header_seq (class loop *loop, gimple_seq seq)
 {
 if (seq)
 {
 gimple_stmt_iterator gsi = gsi_after_labels (loop->header);
 gsi_insert_seq_before (&gsi, seq, GSI_SAME_STMT);
 			       rgroup_masks *src_rgm)
 {
 tree src_masktype = src_rgm->mask_type;
 tree dest_masktype = dest_rgm->mask_type;
 machine_mode src_mode = TYPE_MODE (src_masktype);
+insn_code icode1, icode2;
 if (dest_rgm->max_nscalars_per_iter <= src_rgm->max_nscalars_per_iter
-&& optab_handler (vec_unpacku_hi_optab, src_mode) != CODE_FOR_nothing
+&& (icode1 = optab_handler (vec_unpacku_hi_optab,
-&& optab_handler (vec_unpacku_lo_optab, src_mode) != CODE_FOR_nothing)
+				  src_mode)) != CODE_FOR_nothing
+&& (icode2 = optab_handler (vec_unpacku_lo_optab,
+				  src_mode)) != CODE_FOR_nothing)
 {
 /* Unpacking the source masks gives at least as many mask bits as
 	 we need.  We can then VIEW_CONVERT any excess bits away.  */
-tree unpack_masktype = vect_halve_mask_nunits (src_masktype);
+machine_mode dest_mode = insn_data[icode1].operand[0].mode;
+gcc_assert (dest_mode == insn_data[icode2].operand[0].mode);
+tree unpack_masktype = vect_halve_mask_nunits (src_masktype, dest_mode);
 for (unsigned int i = 0; i < dest_rgm->masks.length (); ++i)
 	{
 	  tree src = src_rgm->masks[i / 2];
 	  tree dest = dest_rgm->masks[i];
 	  tree_code code = ((i & 1) == (BYTES_BIG_ENDIAN ? 0 : 1)
 all the masks in RGM and return a mask that is nonzero when the loop
 needs to iterate.  Add any new preheader statements to PREHEADER_SEQ.
 Use LOOP_COND_GSI to insert code before the exit gcond.
 RGM belongs to loop LOOP.  The loop originally iterated NITERS
-times and has been vectorized according to LOOP_VINFO.  Each iteration
+times and has been vectorized according to LOOP_VINFO.
-of the vectorized loop handles VF iterations of the scalar loop.
 If NITERS_SKIP is nonnull, the first iteration of the vectorized loop
 starts with NITERS_SKIP dummy iterations of the scalar loop before
 the real work starts.  The mask elements for these dummy iterations
 must be 0, to ensure that the extra iterations do not have an effect.
 This means that we cannot guarantee that such an induction variable
 would ever hit a value that produces a set of all-false masks for RGM.  */
 static tree
-vect_set_loop_masks_directly (struct loop *loop, loop_vec_info loop_vinfo,
+vect_set_loop_masks_directly (class loop *loop, loop_vec_info loop_vinfo,
 			      gimple_seq *preheader_seq,
 			      gimple_stmt_iterator loop_cond_gsi,
-			      rgroup_masks *rgm, tree vf,
+			      rgroup_masks *rgm, tree niters, tree niters_skip,
-			      tree niters, tree niters_skip,
 			      bool might_wrap_p)
 {
 tree compare_type = LOOP_VINFO_MASK_COMPARE_TYPE (loop_vinfo);
+tree iv_type = LOOP_VINFO_MASK_IV_TYPE (loop_vinfo);
 tree mask_type = rgm->mask_type;
 unsigned int nscalars_per_iter = rgm->max_nscalars_per_iter;
 poly_uint64 nscalars_per_mask = TYPE_VECTOR_SUBPARTS (mask_type);
+poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
 /* Calculate the maximum number of scalar values that the rgroup
 handles in total, the number that it handles for each iteration
 of the vector loop, and the number that it should skip during the
 first iteration of the vector loop.  */
 tree nscalars_total = niters;
-tree nscalars_step = vf;
+tree nscalars_step = build_int_cst (iv_type, vf);
 tree nscalars_skip = niters_skip;
 if (nscalars_per_iter != 1)
 {
 /* We checked before choosing to use a fully-masked loop that these
 	 multiplications don't overflow.  */
-tree factor = build_int_cst (compare_type, nscalars_per_iter);
+tree compare_factor = build_int_cst (compare_type, nscalars_per_iter);
+tree iv_factor = build_int_cst (iv_type, nscalars_per_iter);
 nscalars_total = gimple_build (preheader_seq, MULT_EXPR, compare_type,
-				     nscalars_total, factor);
+				     nscalars_total, compare_factor);
-nscalars_step = gimple_build (preheader_seq, MULT_EXPR, compare_type,
+nscalars_step = gimple_build (preheader_seq, MULT_EXPR, iv_type,
-				    nscalars_step, factor);
+				    nscalars_step, iv_factor);
 if (nscalars_skip)
 	nscalars_skip = gimple_build (preheader_seq, MULT_EXPR, compare_type,
-				      nscalars_skip, factor);
+				      nscalars_skip, compare_factor);
 }
 /* Create an induction variable that counts the number of scalars
 processed.  */
 tree index_before_incr, index_after_incr;
 gimple_stmt_iterator incr_gsi;
 bool insert_after;
+standard_iv_increment_position (loop, &incr_gsi, &insert_after);
+create_iv (build_int_cst (iv_type, 0), nscalars_step, NULL_TREE, loop,
+	     &incr_gsi, insert_after, &index_before_incr, &index_after_incr);
 tree zero_index = build_int_cst (compare_type, 0);
-standard_iv_increment_position (loop, &incr_gsi, &insert_after);
-create_iv (zero_index, nscalars_step, NULL_TREE, loop, &incr_gsi,
-	     insert_after, &index_before_incr, &index_after_incr);
 tree test_index, test_limit, first_limit;
 gimple_stmt_iterator *test_gsi;
 if (might_wrap_p)
 {
 /* In principle the loop should stop iterating once the incremented
 	   NSCALARS_TOTAL -[sat] (NSCALARS_STEP - NSCALARS_SKIP)
 	 where the rightmost subtraction can be done directly in
 	 COMPARE_TYPE.  */
 test_index = index_before_incr;
-tree adjust = nscalars_step;
+tree adjust = gimple_convert (preheader_seq, compare_type,
+				    nscalars_step);
 if (nscalars_skip)
 	adjust = gimple_build (preheader_seq, MINUS_EXPR, compare_type,
 			       adjust, nscalars_skip);
 test_limit = gimple_build (preheader_seq, MAX_EXPR, compare_type,
 				 nscalars_total, adjust);
 				   test_limit, nscalars_skip);
 test_gsi = &loop_cond_gsi;
 first_limit = test_limit;
 }
+/* Convert the IV value to the comparison type (either a no-op or
+a demotion).  */
+gimple_seq test_seq = NULL;
+test_index = gimple_convert (&test_seq, compare_type, test_index);
+gsi_insert_seq_before (test_gsi, test_seq, GSI_SAME_STMT);
 /* Provide a definition of each mask in the group.  */
 tree next_mask = NULL_TREE;
 tree mask;
 unsigned int i;
 Insert the branch-back condition before LOOP_COND_GSI and return the
 final gcond.  */
 static gcond *
-vect_set_loop_condition_masked (struct loop *loop, loop_vec_info loop_vinfo,
+vect_set_loop_condition_masked (class loop *loop, loop_vec_info loop_vinfo,
 				tree niters, tree final_iv,
 				bool niters_maybe_zero,
 				gimple_stmt_iterator loop_cond_gsi)
 {
 gimple_seq preheader_seq = NULL;
 gimple_seq header_seq = NULL;
 tree compare_type = LOOP_VINFO_MASK_COMPARE_TYPE (loop_vinfo);
 unsigned int compare_precision = TYPE_PRECISION (compare_type);
-unsigned HOST_WIDE_INT max_vf = vect_max_vf (loop_vinfo);
 tree orig_niters = niters;
 /* Type of the initial value of NITERS.  */
 tree ni_actual_type = TREE_TYPE (niters);
 unsigned int ni_actual_precision = TYPE_PRECISION (ni_actual_type);
+tree niters_skip = LOOP_VINFO_MASK_SKIP_NITERS (loop_vinfo);
 /* Convert NITERS to the same size as the compare.  */
 if (compare_precision > ni_actual_precision
 && niters_maybe_zero)
 {
 			     niters, build_one_cst (compare_type));
 }
 else
 niters = gimple_convert (&preheader_seq, compare_type, niters);
-/* Convert skip_niters to the right type.  */
+widest_int iv_limit = vect_iv_limit_for_full_masking (loop_vinfo);
-tree niters_skip = LOOP_VINFO_MASK_SKIP_NITERS (loop_vinfo);
-/* Now calculate the value that the induction variable must be able
-to hit in order to ensure that we end the loop with an all-false mask.
-This involves adding the maximum number of inactive trailing scalar
-iterations.  */
-widest_int iv_limit;
-bool known_max_iters = max_loop_iterations (loop, &iv_limit);
-if (known_max_iters)
-{
-if (niters_skip)
-	{
-	  /* Add the maximum number of skipped iterations to the
-	     maximum iteration count.  */
-	  if (TREE_CODE (niters_skip) == INTEGER_CST)
-	    iv_limit += wi::to_widest (niters_skip);
-	  else
-	    iv_limit += max_vf - 1;
-	}
-/* IV_LIMIT is the maximum number of latch iterations, which is also
-	 the maximum in-range IV value.  Round this value down to the previous
-	 vector alignment boundary and then add an extra full iteration.  */
-poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
-iv_limit = (iv_limit & -(int) known_alignment (vf)) + max_vf;
-}
-/* Get the vectorization factor in tree form.  */
-tree vf = build_int_cst (compare_type,
-			   LOOP_VINFO_VECT_FACTOR (loop_vinfo));
 /* Iterate over all the rgroups and fill in their masks.  We could use
 the first mask from any rgroup for the loop condition; here we
 arbitrarily pick the last.  */
 tree test_mask = NULL_TREE;
 	  }
 	/* See whether zero-based IV would ever generate all-false masks
 	   before wrapping around.  */
 	bool might_wrap_p
-	  = (!known_max_iters
+	  = (iv_limit == -1
 	     || (wi::min_precision (iv_limit * rgm->max_nscalars_per_iter,
 				    UNSIGNED)
 		 > compare_precision));
 	/* Set up all masks for this group.  */
 	test_mask = vect_set_loop_masks_directly (loop, loop_vinfo,
 						  &preheader_seq,
-						  loop_cond_gsi, rgm, vf,
+						  loop_cond_gsi, rgm,
 						  niters, niters_skip,
 						  might_wrap_p);
 }
 /* Emit all accumulated statements.  */
 /* Like vect_set_loop_condition, but handle the case in which there
 are no loop masks.  */
 static gcond *
-vect_set_loop_condition_unmasked (struct loop *loop, tree niters,
+vect_set_loop_condition_unmasked (class loop *loop, tree niters,
 				  tree step, tree final_iv,
 				  bool niters_maybe_zero,
 				  gimple_stmt_iterator loop_cond_gsi)
 {
 tree indx_before_incr, indx_after_incr;
 N * STEP on exit from the loop.
 Assumption: the exit-condition of LOOP is the last stmt in the loop.  */
 void
-vect_set_loop_condition (struct loop *loop, loop_vec_info loop_vinfo,
+vect_set_loop_condition (class loop *loop, loop_vec_info loop_vinfo,
 			 tree niters, tree step, tree final_iv,
 			 bool niters_maybe_zero)
 {
 gcond *cond_stmt;
 gcond *orig_cond = get_loop_exit_condition (loop);
 	  gsi_next (&gsi_to);
 	  continue;
 	}
 if (TREE_CODE (from_arg) != SSA_NAME)
 	gcc_assert (operand_equal_p (from_arg, to_arg, 0));
-else
+else if (TREE_CODE (to_arg) == SSA_NAME
+	       && from_arg != to_arg)
 	{
 	  if (get_current_def (to_arg) == NULL_TREE)
-	    set_current_def (to_arg, get_current_def (from_arg));
+	    {
+	      gcc_assert (types_compatible_p (TREE_TYPE (to_arg),
+					      TREE_TYPE (get_current_def
+							   (from_arg))));
+	      set_current_def (to_arg, get_current_def (from_arg));
+	    }
 	}
 gsi_next (&gsi_from);
 gsi_next (&gsi_to);
 }
 on E which is either the entry or exit of LOOP.  If SCALAR_LOOP is
 non-NULL, assume LOOP and SCALAR_LOOP are equivalent and copy the
 basic blocks from SCALAR_LOOP instead of LOOP, but to either the
 entry or exit of LOOP.  */
-struct loop *
+class loop *
-slpeel_tree_duplicate_loop_to_edge_cfg (struct loop *loop,
+slpeel_tree_duplicate_loop_to_edge_cfg (class loop *loop,
-					struct loop *scalar_loop, edge e)
+					class loop *scalar_loop, edge e)
 {
-struct loop *new_loop;
+class loop *new_loop;
 basic_block *new_bbs, *bbs, *pbbs;
 bool at_exit;
 bool was_imm_dom;
 basic_block exit_dest;
 edge exit, new_exit;
 (3) its exit condition is the last stmt in the header
 (4) E is the entry/exit edge of LOOP.
 */
 bool
-slpeel_can_duplicate_loop_p (const struct loop *loop, const_edge e)
+slpeel_can_duplicate_loop_p (const class loop *loop, const_edge e)
 {
 edge exit_e = single_exit (loop);
 edge entry_e = loop_preheader_edge (loop);
 gcond *orig_cond = get_loop_exit_condition (loop);
 gimple_stmt_iterator loop_exit_gsi = gsi_last_bb (exit_e->src);
 /* If the loop has a virtual PHI, but exit bb doesn't, create a virtual PHI
 in the exit bb and rename all the uses after the loop.  This simplifies
 the *guard[12] routines, which assume loop closed SSA form for all PHIs
 (but normally loop closed SSA form doesn't require virtual PHIs to be
 in the same form).  Doing this early simplifies the checking what
-uses should be renamed.  */
+uses should be renamed.
-static void
+If we create a new phi after the loop, return the definition that
-create_lcssa_for_virtual_phi (struct loop *loop)
+applies on entry to the loop, otherwise return null.  */
+static tree
+create_lcssa_for_virtual_phi (class loop *loop)
 {
 gphi_iterator gsi;
 edge exit_e = single_exit (loop);
 for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi))
 	    FOR_EACH_IMM_USE_STMT (stmt, imm_iter, vop)
 	      if (stmt != new_phi
 		  && !flow_bb_inside_loop_p (loop, gimple_bb (stmt)))
 		FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
 		  SET_USE (use_p, new_vop);
+	    return PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop));
 	  }
 	break;
 }
+return NULL_TREE;
 }
 /* Function vect_get_loop_location.
 Extract the location of the loop in the source code.
 If the loop is not well formed for vectorization, an estimated
 location is calculated.
 Return the loop location if succeed and NULL if not.  */
 dump_user_location_t
-find_loop_location (struct loop *loop)
+find_loop_location (class loop *loop)
 {
 gimple *stmt = NULL;
 basic_block bb;
 gimple_stmt_iterator si;
 These restrictions will be relaxed in the future.  */
 bool
 vect_can_advance_ivs_p (loop_vec_info loop_vinfo)
 {
-struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
+class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
 basic_block bb = loop->header;
 gphi_iterator gsi;
 /* Analyze phi functions of the loop header.  */
 static void
 vect_update_ivs_after_vectorizer (loop_vec_info loop_vinfo,
 				  tree niters, edge update_e)
 {
 gphi_iterator gsi, gsi1;
-struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
+class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
 basic_block update_bb = update_e->dest;
 basic_block exit_bb = single_exit (loop)->dest;
 /* Make sure there exists a single-predecessor exit bb:  */
 gcc_assert (single_pred_p (exit_bb));
 {
 dr_vec_info *dr_info = LOOP_VINFO_UNALIGNED_DR (loop_vinfo);
 stmt_vec_info stmt_info = dr_info->stmt;
 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
-unsigned int target_align = DR_TARGET_ALIGNMENT (dr_info);
+poly_uint64 target_align = DR_TARGET_ALIGNMENT (dr_info);
-gcc_assert (target_align != 0);
+unsigned HOST_WIDE_INT target_align_c;
+tree target_align_minus_1;
 bool negative = tree_int_cst_compare (DR_STEP (dr_info->dr),
 					size_zero_node) < 0;
 tree offset = (negative
 		 ? size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1)
 		 : size_zero_node);
 tree start_addr = vect_create_addr_base_for_vector_ref (stmt_info, seq,
 							  offset);
 tree type = unsigned_type_for (TREE_TYPE (start_addr));
-tree target_align_minus_1 = build_int_cst (type, target_align - 1);
+if (target_align.is_constant (&target_align_c))
+target_align_minus_1 = build_int_cst (type, target_align_c - 1);
+else
+{
+tree vla = build_int_cst (type, target_align);
+tree vla_align = fold_build2 (BIT_AND_EXPR, type, vla,
+				    fold_build2 (MINUS_EXPR, type,
+						 build_int_cst (type, 0), vla));
+target_align_minus_1 = fold_build2 (MINUS_EXPR, type, vla_align,
+					  build_int_cst (type, 1));
+}
 HOST_WIDE_INT elem_size
 = int_cst_value (TYPE_SIZE_UNIT (TREE_TYPE (vectype)));
 tree elem_size_log = build_int_cst (type, exact_log2 (elem_size));
 /* Create:  misalign_in_bytes = addr & (target_align - 1).  */
 tree niters_type = TREE_TYPE (LOOP_VINFO_NITERS (loop_vinfo));
 gimple_seq stmts = NULL, new_stmts = NULL;
 tree iters, iters_name;
 stmt_vec_info stmt_info = dr_info->stmt;
 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
-unsigned int target_align = DR_TARGET_ALIGNMENT (dr_info);
+poly_uint64 target_align = DR_TARGET_ALIGNMENT (dr_info);
 if (LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) > 0)
 {
 int npeel = LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo);
 {
 tree misalign_in_elems = get_misalign_in_elems (&stmts, loop_vinfo);
 tree type = TREE_TYPE (misalign_in_elems);
 HOST_WIDE_INT elem_size
 	= int_cst_value (TYPE_SIZE_UNIT (TREE_TYPE (vectype)));
-HOST_WIDE_INT align_in_elems = target_align / elem_size;
+/* We only do prolog peeling if the target alignment is known at compile
-tree align_in_elems_minus_1 = build_int_cst (type, align_in_elems - 1);
+time.  */
+poly_uint64 align_in_elems =
+	exact_div (target_align, elem_size);
+tree align_in_elems_minus_1 =
+	build_int_cst (type, align_in_elems - 1);
 tree align_in_elems_tree = build_int_cst (type, align_in_elems);
 /* Create:  (niters_type) ((align_in_elems - misalign_in_elems)
 				 & (align_in_elems - 1)).  */
 bool negative = tree_int_cst_compare (DR_STEP (dr_info->dr),
 else
 	iters = fold_build2 (MINUS_EXPR, type, align_in_elems_tree,
 			     misalign_in_elems);
 iters = fold_build2 (BIT_AND_EXPR, type, iters, align_in_elems_minus_1);
 iters = fold_convert (niters_type, iters);
-*bound = align_in_elems - 1;
+unsigned HOST_WIDE_INT align_in_elems_c;
+if (align_in_elems.is_constant (&align_in_elems_c))
+	*bound = align_in_elems_c - 1;
+else
+	*bound = -1;
 }
 if (dump_enabled_p ())
 dump_printf_loc (MSG_NOTE, vect_location,
 		     "niters for prolog loop: %T\n", iters);
 If CODE is PLUS, the vector loop starts NITERS iterations after the
 scalar one, otherwise CODE is MINUS and the vector loop starts NITERS
 iterations before the scalar one (using masking to skip inactive
 elements).  This function updates the information recorded in DR to
 account for the difference.  Specifically, it updates the OFFSET
-field of DR.  */
+field of DR_INFO.  */
 static void
-vect_update_init_of_dr (struct data_reference *dr, tree niters, tree_code code)
+vect_update_init_of_dr (dr_vec_info *dr_info, tree niters, tree_code code)
 {
-tree offset = DR_OFFSET (dr);
+struct data_reference *dr = dr_info->dr;
+tree offset = dr_info->offset;
+if (!offset)
+offset = build_zero_cst (sizetype);
 niters = fold_build2 (MULT_EXPR, sizetype,
 			fold_convert (sizetype, niters),
 			fold_convert (sizetype, DR_STEP (dr)));
 offset = fold_build2 (code, sizetype,
 			fold_convert (sizetype, offset), niters);
-DR_OFFSET (dr) = offset;
+dr_info->offset = offset;
 }
 /* Function vect_update_inits_of_drs
 Apply vect_update_inits_of_dr to all accesses in LOOP_VINFO.
 CODE and NITERS are as for vect_update_inits_of_dr.  */
-static void
+void
 vect_update_inits_of_drs (loop_vec_info loop_vinfo, tree niters,
 			  tree_code code)
 {
 unsigned int i;
 vec<data_reference_p> datarefs = LOOP_VINFO_DATAREFS (loop_vinfo);
 struct data_reference *dr;
 DUMP_VECT_SCOPE ("vect_update_inits_of_dr");
-/* Adjust niters to sizetype and insert stmts on loop preheader edge.  */
+/* Adjust niters to sizetype.  We used to insert the stmts on loop preheader
+here, but since we might use these niters to update the epilogues niters
+and data references we can't insert them here as this definition might not
+always dominate its uses.  */
 if (!types_compatible_p (sizetype, TREE_TYPE (niters)))
-{
+niters = fold_convert (sizetype, niters);
-gimple_seq seq;
-edge pe = loop_preheader_edge (LOOP_VINFO_LOOP (loop_vinfo));
-tree var = create_tmp_var (sizetype, "prolog_loop_adjusted_niters");
-niters = fold_convert (sizetype, niters);
-niters = force_gimple_operand (niters, &seq, false, var);
-if (seq)
-	{
-	  basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq);
-	  gcc_assert (!new_bb);
-	}
-}
 FOR_EACH_VEC_ELT (datarefs, i, dr)
 {
 dr_vec_info *dr_info = loop_vinfo->lookup_dr (dr);
 if (!STMT_VINFO_GATHER_SCATTER_P (dr_info->stmt))
-	vect_update_init_of_dr (dr, niters, code);
+	vect_update_init_of_dr (dr_info, niters, code);
 }
 }
 /* For the information recorded in LOOP_VINFO prepare the loop for peeling
 by masking.  This involves calculating the number of iterations to
 				     tree niters_vector,
 				     tree *niters_vector_mult_vf_ptr)
 {
 /* We should be using a step_vector of VF if VF is variable.  */
 int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo).to_constant ();
-struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
+class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
 tree type = TREE_TYPE (niters_vector);
 tree log_vf = build_int_cst (type, exact_log2 (vf));
 basic_block exit_bb = single_exit (loop)->dest;
 gcc_assert (niters_vector_mult_vf_ptr != NULL);
 						    &stmts, true, var);
 gimple_stmt_iterator gsi = gsi_start_bb (exit_bb);
 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
 }
 *niters_vector_mult_vf_ptr = niters_vector_mult_vf;
+}
+/* LCSSA_PHI is a lcssa phi of EPILOG loop which is copied from LOOP,
+this function searches for the corresponding lcssa phi node in exit
+bb of LOOP.  If it is found, return the phi result; otherwise return
+NULL.  */
+static tree
+find_guard_arg (class loop *loop, class loop *epilog ATTRIBUTE_UNUSED,
+		gphi *lcssa_phi)
+{
+gphi_iterator gsi;
+edge e = single_exit (loop);
+gcc_assert (single_pred_p (e->dest));
+for (gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi); gsi_next (&gsi))
+{
+gphi *phi = gsi.phi ();
+if (operand_equal_p (PHI_ARG_DEF (phi, 0),
+			   PHI_ARG_DEF (lcssa_phi, 0), 0))
+	return PHI_RESULT (phi);
+}
+return NULL_TREE;
 }
 /* Function slpeel_tree_duplicate_loop_to_edge_cfg duplciates FIRST/SECOND
 from SECOND/FIRST and puts it at the original loop's preheader/exit
 edge, the two loops are arranged as below:
 second loop, i.e, between_bb in the example code.  With PHIs updated,
 the second loop will execute rest iterations of the first.  */
 static void
 slpeel_update_phi_nodes_for_loops (loop_vec_info loop_vinfo,
-				   struct loop *first, struct loop *second,
+				   class loop *first, class loop *second,
 				   bool create_lcssa_for_iv_phis)
 {
 gphi_iterator gsi_update, gsi_orig;
-struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
+class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
 edge first_latch_e = EDGE_SUCC (first->latch, 0);
 edge second_preheader_e = loop_preheader_edge (second);
 basic_block between_bb = single_exit (first)->dest;
 	}
 /* Update PHI node in the second loop by replacing arg on the loop's
 	 incoming edge.  */
 adjust_phi_and_debug_stmts (update_phi, second_preheader_e, arg);
+}
+/* For epilogue peeling we have to make sure to copy all LC PHIs
+for correct vectorization of live stmts.  */
+if (loop == first)
+{
+basic_block orig_exit = single_exit (second)->dest;
+for (gsi_orig = gsi_start_phis (orig_exit);
+	   !gsi_end_p (gsi_orig); gsi_next (&gsi_orig))
+	{
+	  gphi *orig_phi = gsi_orig.phi ();
+	  tree orig_arg = PHI_ARG_DEF (orig_phi, 0);
+	  if (TREE_CODE (orig_arg) != SSA_NAME || virtual_operand_p  (orig_arg))
+	    continue;
+	  /* Already created in the above loop.   */
+	  if (find_guard_arg (first, second, orig_phi))
+	    continue;
+	  tree new_res = copy_ssa_name (orig_arg);
+	  gphi *lcphi = create_phi_node (new_res, between_bb);
+	  add_phi_arg (lcphi, orig_arg, single_exit (first), UNKNOWN_LOCATION);
+	}
 }
 }
 /* Function slpeel_add_loop_guard adds guard skipping from the beginning
 of SKIP_LOOP to the beginning of UPDATE_LOOP.  GUARD_EDGE and MERGE_EDGE
 This function creates PHI nodes at merge_bb and replaces the use of i_5
 in the update_loop's PHI node with the result of new PHI result.  */
 static void
-slpeel_update_phi_nodes_for_guard1 (struct loop *skip_loop,
+slpeel_update_phi_nodes_for_guard1 (class loop *skip_loop,
-				    struct loop *update_loop,
+				    class loop *update_loop,
 				    edge guard_edge, edge merge_edge)
 {
-source_location merge_loc, guard_loc;
+location_t merge_loc, guard_loc;
 edge orig_e = loop_preheader_edge (skip_loop);
 edge update_e = loop_preheader_edge (update_loop);
 gphi_iterator gsi_orig, gsi_update;
 for ((gsi_orig = gsi_start_phis (skip_loop->header),
 add_phi_arg (new_phi, guard_arg, guard_edge, guard_loc);
 /* Update phi in UPDATE_PHI.  */
 adjust_phi_and_debug_stmts (update_phi, update_e, new_res);
 }
-}
-/* LCSSA_PHI is a lcssa phi of EPILOG loop which is copied from LOOP,
-this function searches for the corresponding lcssa phi node in exit
-bb of LOOP.  If it is found, return the phi result; otherwise return
-NULL.  */
-static tree
-find_guard_arg (struct loop *loop, struct loop *epilog ATTRIBUTE_UNUSED,
-		gphi *lcssa_phi)
-{
-gphi_iterator gsi;
-edge e = single_exit (loop);
-gcc_assert (single_pred_p (e->dest));
-for (gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi); gsi_next (&gsi))
-{
-gphi *phi = gsi.phi ();
-if (operand_equal_p (PHI_ARG_DEF (phi, 0),
-			   PHI_ARG_DEF (lcssa_phi, 0), 0))
-	return PHI_RESULT (phi);
-}
-return NULL_TREE;
 }
 /* LOOP and EPILOG are two consecutive loops in CFG and EPILOG is copied
 from LOOP.  Function slpeel_add_loop_guard adds guard skipping from a
 point between the two loops to the end of EPILOG.  Edges GUARD_EDGE
 the arg of the original PHI in exit_bb, arg for GUARD_EDGE is defined
 by LOOP and is found in the exit bb of LOOP.  Arg of the original PHI
 in exit_bb will also be updated.  */
 static void
-slpeel_update_phi_nodes_for_guard2 (struct loop *loop, struct loop *epilog,
+slpeel_update_phi_nodes_for_guard2 (class loop *loop, class loop *epilog,
 				    edge guard_edge, edge merge_edge)
 {
 gphi_iterator gsi;
 basic_block merge_bb = guard_edge->dest;
 for (gsi = gsi_start_phis (exit_bb); !gsi_end_p (gsi); gsi_next (&gsi))
 {
 gphi *update_phi = gsi.phi ();
 tree old_arg = PHI_ARG_DEF (update_phi, 0);
-/* This loop-closed-phi actually doesn't represent a use out of the
-	 loop - the phi arg is a constant.  */
+tree merge_arg = NULL_TREE;
-if (TREE_CODE (old_arg) != SSA_NAME)
-	continue;
+/* If the old argument is a SSA_NAME use its current_def.  */
+if (TREE_CODE (old_arg) == SSA_NAME)
-tree merge_arg = get_current_def (old_arg);
+	merge_arg = get_current_def (old_arg);
+/* If it's a constant or doesn't have a current_def, just use the old
+	 argument.  */
 if (!merge_arg)
 	merge_arg = old_arg;
 tree guard_arg = find_guard_arg (loop, epilog, update_phi);
 /* If the var is live after loop but not a reduction, we simply
 /* EPILOG loop is duplicated from the original loop for vectorizing,
 the arg of its loop closed ssa PHI needs to be updated.  */
 static void
-slpeel_update_phi_nodes_for_lcssa (struct loop *epilog)
+slpeel_update_phi_nodes_for_lcssa (class loop *epilog)
 {
 gphi_iterator gsi;
 basic_block exit_bb = single_exit (epilog)->dest;
 gcc_assert (single_pred_p (exit_bb));
 merge_bb_3:
 Note this function peels prolog and epilog only if it's necessary,
 as well as guards.
-Returns created epilogue or NULL.
+This function returns the epilogue loop if a decision was made to vectorize
+it, otherwise NULL.
+The analysis resulting in this epilogue loop's loop_vec_info was performed
+in the same vect_analyze_loop call as the main loop's.  At that time
+vect_analyze_loop constructs a list of accepted loop_vec_info's for lower
+vectorization factors than the main loop.  This list is stored in the main
+loop's loop_vec_info in the 'epilogue_vinfos' member.  Everytime we decide to
+vectorize the epilogue loop for a lower vectorization factor,  the
+loop_vec_info sitting at the top of the epilogue_vinfos list is removed,
+updated and linked to the epilogue loop.  This is later used to vectorize
+the epilogue.  The reason the loop_vec_info needs updating is that it was
+constructed based on the original main loop, and the epilogue loop is a
+copy of this loop, so all links pointing to statements in the original loop
+need updating.  Furthermore, these loop_vec_infos share the
+data_reference's records, which will also need to be updated.
 TODO: Guard for prefer_scalar_loop should be emitted along with
 versioning conditions if loop versioning is needed.  */
-struct loop *
+class loop *
 vect_do_peeling (loop_vec_info loop_vinfo, tree niters, tree nitersm1,
 		 tree *niters_vector, tree *step_vector,
 		 tree *niters_vector_mult_vf_var, int th,
-		 bool check_profitability, bool niters_no_overflow)
+		 bool check_profitability, bool niters_no_overflow,
+		 tree *advance)
 {
 edge e, guard_e;
 tree type = TREE_TYPE (niters), guard_cond;
 basic_block guard_bb, guard_to;
 profile_probability prob_prolog, prob_vector, prob_epilog;
 int estimated_vf;
 int prolog_peeling = 0;
+bool vect_epilogues = loop_vinfo->epilogue_vinfos.length () > 0;
+/* We currently do not support prolog peeling if the target alignment is not
+known at compile time.  'vect_gen_prolog_loop_niters' depends on the
+target alignment being constant.  */
+dr_vec_info *dr_info = LOOP_VINFO_UNALIGNED_DR (loop_vinfo);
+if (dr_info && !DR_TARGET_ALIGNMENT (dr_info).is_constant ())
+return NULL;
 if (!vect_use_loop_mask_for_alignment_p (loop_vinfo))
 prolog_peeling = LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo);
 poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
 poly_uint64 bound_epilog = 0;
 if (estimated_vf == 2)
 estimated_vf = 3;
 prob_prolog = prob_epilog = profile_probability::guessed_always ()
 			.apply_scale (estimated_vf - 1, estimated_vf);
-struct loop *prolog, *epilog = NULL, *loop = LOOP_VINFO_LOOP (loop_vinfo);
+class loop *prolog, *epilog = NULL, *loop = LOOP_VINFO_LOOP (loop_vinfo);
-struct loop *first_loop = loop;
+class loop *first_loop = loop;
 bool irred_flag = loop_preheader_edge (loop)->flags & EDGE_IRREDUCIBLE_LOOP;
+/* We might have a queued need to update virtual SSA form.  As we
+delete the update SSA machinery below after doing a regular
+incremental SSA update during loop copying make sure we don't
+lose that fact.
+???  Needing to update virtual SSA form by renaming is unfortunate
+but not all of the vectorizer code inserting new loads / stores
+properly assigns virtual operands to those statements.  */
+update_ssa (TODO_update_ssa_only_virtuals);
 create_lcssa_for_virtual_phi (loop);
-update_ssa (TODO_update_ssa_only_virtuals);
+/* If we're vectorizing an epilogue loop, the update_ssa above will
+have ensured that the virtual operand is in SSA form throughout the
+vectorized main loop.  Normally it is possible to trace the updated
+vector-stmt vdefs back to scalar-stmt vdefs and vector-stmt vuses
+back to scalar-stmt vuses, meaning that the effect of the SSA update
+remains local to the main loop.  However, there are rare cases in
+which the vectorized loop has vdefs even when the original scalar
+loop didn't.  For example, vectorizing a load with IFN_LOAD_LANES
+introduces clobbers of the temporary vector array, which in turn
+needs new vdefs.  If the scalar loop doesn't write to memory, these
+new vdefs will be the only ones in the vector loop.
+In that case, update_ssa will have added a new virtual phi to the
+main loop, which previously didn't need one.  Ensure that we (locally)
+maintain LCSSA form for the virtual operand, just as we would have
+done if the virtual phi had existed from the outset.  This makes it
+easier to duplicate the scalar epilogue loop below.  */
+tree vop_to_rename = NULL_TREE;
+if (loop_vec_info orig_loop_vinfo = LOOP_VINFO_ORIG_LOOP_INFO (loop_vinfo))
+{
+class loop *orig_loop = LOOP_VINFO_LOOP (orig_loop_vinfo);
+vop_to_rename = create_lcssa_for_virtual_phi (orig_loop);
+}
 if (MAY_HAVE_DEBUG_BIND_STMTS)
 {
 gcc_assert (!adjust_vec.exists ());
 adjust_vec.create (32);
 so that we can also get the upper bound on the number of iterations.  */
 tree niters_prolog;
 int bound_prolog = 0;
 if (prolog_peeling)
 niters_prolog = vect_gen_prolog_loop_niters (loop_vinfo, anchor,
-						 &bound_prolog);
+						  &bound_prolog);
 else
 niters_prolog = build_int_cst (type, 0);
+loop_vec_info epilogue_vinfo = NULL;
+if (vect_epilogues)
+{
+epilogue_vinfo = loop_vinfo->epilogue_vinfos[0];
+loop_vinfo->epilogue_vinfos.ordered_remove (0);
+}
+tree niters_vector_mult_vf = NULL_TREE;
+/* Saving NITERs before the loop, as this may be changed by prologue.  */
+tree before_loop_niters = LOOP_VINFO_NITERS (loop_vinfo);
+edge update_e = NULL, skip_e = NULL;
+unsigned int lowest_vf = constant_lower_bound (vf);
+/* If we know the number of scalar iterations for the main loop we should
+check whether after the main loop there are enough iterations left over
+for the epilogue.  */
+if (vect_epilogues
+&& LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
+&& prolog_peeling >= 0
+&& known_eq (vf, lowest_vf))
+{
+unsigned HOST_WIDE_INT eiters
+	= (LOOP_VINFO_INT_NITERS (loop_vinfo)
+	   - LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo));
+eiters -= prolog_peeling;
+eiters
+	= eiters % lowest_vf + LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo);
+unsigned int ratio;
+unsigned int epilogue_gaps
+	= LOOP_VINFO_PEELING_FOR_GAPS (epilogue_vinfo);
+while (!(constant_multiple_p
+	       (GET_MODE_SIZE (loop_vinfo->vector_mode),
+		GET_MODE_SIZE (epilogue_vinfo->vector_mode), &ratio)
+	       && eiters >= lowest_vf / ratio + epilogue_gaps))
+	{
+	  delete epilogue_vinfo;
+	  epilogue_vinfo = NULL;
+	  if (loop_vinfo->epilogue_vinfos.length () == 0)
+	    {
+	      vect_epilogues = false;
+	      break;
+	    }
+	  epilogue_vinfo = loop_vinfo->epilogue_vinfos[0];
+	  loop_vinfo->epilogue_vinfos.ordered_remove (0);
+	  epilogue_gaps = LOOP_VINFO_PEELING_FOR_GAPS (epilogue_vinfo);
+	}
+}
 /* Prolog loop may be skipped.  */
 bool skip_prolog = (prolog_peeling != 0);
-/* Skip to epilog if scalar loop may be preferred.  It's only needed
+/* Skip this loop to epilog when there are not enough iterations to enter this
-when we peel for epilog loop and when it hasn't been checked with
+vectorized loop.  If true we should perform runtime checks on the NITERS
-loop versioning.  */
+to check whether we should skip the current vectorized loop.  If we know
+the number of scalar iterations we may choose to add a runtime check if
+this number "maybe" smaller than the number of iterations required
+when we know the number of scalar iterations may potentially
+be smaller than the number of iterations required to enter this loop, for
+this we use the upper bounds on the prolog and epilog peeling.  When we
+don't know the number of iterations and don't require versioning it is
+because we have asserted that there are enough scalar iterations to enter
+the main loop, so this skip is not necessary.  When we are versioning then
+we only add such a skip if we have chosen to vectorize the epilogue.  */
 bool skip_vector = (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
 		      ? maybe_lt (LOOP_VINFO_INT_NITERS (loop_vinfo),
 				  bound_prolog + bound_epilog)
-		      : !LOOP_REQUIRES_VERSIONING (loop_vinfo));
+		      : (!LOOP_REQUIRES_VERSIONING (loop_vinfo)
+			 || vect_epilogues));
 /* Epilog loop must be executed if the number of iterations for epilog
 loop is known at compile time, otherwise we need to add a check at
 the end of vector loop and skip to the end of epilog loop.  */
 bool skip_epilog = (prolog_peeling < 0
 		      || !LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
 	  scale_loop_profile (loop, prob_vector, 0);
 	}
 }
 dump_user_location_t loop_loc = find_loop_location (loop);
-struct loop *scalar_loop = LOOP_VINFO_SCALAR_LOOP (loop_vinfo);
+class loop *scalar_loop = LOOP_VINFO_SCALAR_LOOP (loop_vinfo);
+if (vect_epilogues)
+/* Make sure to set the epilogue's epilogue scalar loop, such that we can
+use the original scalar loop as remaining epilogue if necessary.  */
+LOOP_VINFO_SCALAR_LOOP (epilogue_vinfo)
+= LOOP_VINFO_SCALAR_LOOP (loop_vinfo);
 if (prolog_peeling)
 {
 e = loop_preheader_edge (loop);
 if (!slpeel_can_duplicate_loop_p (loop, e))
 	{
 	{
 	  dump_printf_loc (MSG_MISSED_OPTIMIZATION, loop_loc,
 			   "slpeel_tree_duplicate_loop_to_edge_cfg failed.\n");
 	  gcc_unreachable ();
 	}
+prolog->force_vectorize = false;
 slpeel_update_phi_nodes_for_loops (loop_vinfo, prolog, loop, true);
 first_loop = prolog;
 reset_original_copy_tables ();
 /* Update the number of iterations for prolog loop.  */
 	  slpeel_update_phi_nodes_for_guard1 (prolog, loop, guard_e, e);
 	  scale_bbs_frequencies (&bb_after_prolog, 1, prob_prolog);
 	  scale_loop_profile (prolog, prob_prolog, bound_prolog);
 	}
 /* Update init address of DRs.  */
 vect_update_inits_of_drs (loop_vinfo, niters_prolog, PLUS_EXPR);
 /* Update niters for vector loop.  */
 LOOP_VINFO_NITERS (loop_vinfo)
 	= fold_build2 (MINUS_EXPR, type, niters, niters_prolog);
 	{
 	  dump_printf_loc (MSG_MISSED_OPTIMIZATION, loop_loc,
 			   "loop can't be duplicated to exit edge.\n");
 	  gcc_unreachable ();
 	}
-/* Peel epilog and put it on exit edge of loop.  */
+/* Peel epilog and put it on exit edge of loop.  If we are vectorizing
-epilog = slpeel_tree_duplicate_loop_to_edge_cfg (loop, scalar_loop, e);
+	 said epilog then we should use a copy of the main loop as a starting
+	 point.  This loop may have already had some preliminary transformations
+	 to allow for more optimal vectorization, for example if-conversion.
+	 If we are not vectorizing the epilog then we should use the scalar loop
+	 as the transformations mentioned above make less or no sense when not
+	 vectorizing.  */
+epilog = vect_epilogues ? get_loop_copy (loop) : scalar_loop;
+if (vop_to_rename)
+	{
+	  /* Vectorizing the main loop can sometimes introduce a vdef to
+	     a loop that previously didn't have one; see the comment above
+	     the definition of VOP_TO_RENAME for details.  The definition
+	     D that holds on E will then be different from the definition
+	     VOP_TO_RENAME that holds during SCALAR_LOOP, so we need to
+	     rename VOP_TO_RENAME to D when copying the loop.
+	     The virtual operand is in LCSSA form for the main loop,
+	     and no stmt between the main loop and E needs a vdef,
+	     so we know that D is provided by a phi rather than by a
+	     vdef on a normal gimple stmt.  */
+	  basic_block vdef_bb = e->src;
+	  gphi *vphi;
+	  while (!(vphi = get_virtual_phi (vdef_bb)))
+	    vdef_bb = get_immediate_dominator (CDI_DOMINATORS, vdef_bb);
+	  gcc_assert (vop_to_rename != gimple_phi_result (vphi));
+	  set_current_def (vop_to_rename, gimple_phi_result (vphi));
+	}
+epilog = slpeel_tree_duplicate_loop_to_edge_cfg (loop, epilog, e);
 if (!epilog)
 	{
 	  dump_printf_loc (MSG_MISSED_OPTIMIZATION, loop_loc,
 			   "slpeel_tree_duplicate_loop_to_edge_cfg failed.\n");
 	  gcc_unreachable ();
 	}
+epilog->force_vectorize = false;
 slpeel_update_phi_nodes_for_loops (loop_vinfo, loop, epilog, false);
 /* Scalar version loop may be preferred.  In this case, add guard
 	 and skip to epilog.  Note this only happens when the number of
 	 iterations of loop is unknown at compile time, otherwise this
 	  guard_to = split_edge (loop_preheader_edge (epilog));
 	  guard_e = slpeel_add_loop_guard (guard_bb, guard_cond,
 					   guard_to, guard_bb,
 					   prob_vector.invert (),
 					   irred_flag);
+	  skip_e = guard_e;
 	  e = EDGE_PRED (guard_to, 0);
 	  e = (e != guard_e ? e : EDGE_PRED (guard_to, 1));
 	  slpeel_update_phi_nodes_for_guard1 (first_loop, epilog, guard_e, e);
 	  /* Simply propagate profile info from guard_bb to guard_to which is
 				    (prob_vector));
 	  free (bbs);
 	}
 basic_block bb_before_epilog = loop_preheader_edge (epilog)->src;
-tree niters_vector_mult_vf;
 /* If loop is peeled for non-zero constant times, now niters refers to
 	 orig_niters - prolog_peeling, it won't overflow even the orig_niters
 	 overflows.  */
 niters_no_overflow |= (prolog_peeling > 0);
 vect_gen_vector_loop_niters (loop_vinfo, niters,
 	vect_gen_vector_loop_niters_mult_vf (loop_vinfo, *niters_vector,
 					     &niters_vector_mult_vf);
 /* Update IVs of original loop as if they were advanced by
 	 niters_vector_mult_vf steps.  */
 gcc_checking_assert (vect_can_advance_ivs_p (loop_vinfo));
-edge update_e = skip_vector ? e : loop_preheader_edge (epilog);
+update_e = skip_vector ? e : loop_preheader_edge (epilog);
 vect_update_ivs_after_vectorizer (loop_vinfo, niters_vector_mult_vf,
 					update_e);
 if (skip_epilog)
 	{
 delete_update_ssa ();
 adjust_vec_debug_stmts ();
 scev_reset ();
 }
+if (vect_epilogues)
+{
+epilog->aux = epilogue_vinfo;
+LOOP_VINFO_LOOP (epilogue_vinfo) = epilog;
+loop_constraint_clear (epilog, LOOP_C_INFINITE);
+/* We now must calculate the number of NITERS performed by the previous
+	 loop and EPILOGUE_NITERS to be performed by the epilogue.  */
+tree niters = fold_build2 (PLUS_EXPR, TREE_TYPE (niters_vector_mult_vf),
+				 niters_prolog, niters_vector_mult_vf);
+/* If skip_vector we may skip the previous loop, we insert a phi-node to
+	 determine whether we are coming from the previous vectorized loop
+	 using the update_e edge or the skip_vector basic block using the
+	 skip_e edge.  */
+if (skip_vector)
+	{
+	  gcc_assert (update_e != NULL && skip_e != NULL);
+	  gphi *new_phi = create_phi_node (make_ssa_name (TREE_TYPE (niters)),
+					   update_e->dest);
+	  tree new_ssa = make_ssa_name (TREE_TYPE (niters));
+	  gimple *stmt = gimple_build_assign (new_ssa, niters);
+	  gimple_stmt_iterator gsi;
+	  if (TREE_CODE (niters_vector_mult_vf) == SSA_NAME
+	      && SSA_NAME_DEF_STMT (niters_vector_mult_vf)->bb != NULL)
+	    {
+	      gsi = gsi_for_stmt (SSA_NAME_DEF_STMT (niters_vector_mult_vf));
+	      gsi_insert_after (&gsi, stmt, GSI_NEW_STMT);
+	    }
+	  else
+	    {
+	      gsi = gsi_last_bb (update_e->src);
+	      gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
+	    }
+	  niters = new_ssa;
+	  add_phi_arg (new_phi, niters, update_e, UNKNOWN_LOCATION);
+	  add_phi_arg (new_phi, build_zero_cst (TREE_TYPE (niters)), skip_e,
+		       UNKNOWN_LOCATION);
+	  niters = PHI_RESULT (new_phi);
+	}
+/* Subtract the number of iterations performed by the vectorized loop
+	 from the number of total iterations.  */
+tree epilogue_niters = fold_build2 (MINUS_EXPR, TREE_TYPE (niters),
+					  before_loop_niters,
+					  niters);
+LOOP_VINFO_NITERS (epilogue_vinfo) = epilogue_niters;
+LOOP_VINFO_NITERSM1 (epilogue_vinfo)
+	= fold_build2 (MINUS_EXPR, TREE_TYPE (epilogue_niters),
+		       epilogue_niters,
+		       build_one_cst (TREE_TYPE (epilogue_niters)));
+/* Set ADVANCE to the number of iterations performed by the previous
+	 loop and its prologue.  */
+*advance = niters;
+/* Redo the peeling for niter analysis as the NITERs and alignment
+	 may have been updated to take the main loop into account.  */
+determine_peel_for_niter (epilogue_vinfo);
+}
 adjust_vec.release ();
 free_original_copy_tables ();
-return epilog;
+return vect_epilogues ? epilog : NULL;
 }
 /* Function vect_create_cond_for_niters_checks.
 Create a conditional expression that represents the run-time checks for
 cost model threshold TH.
 The versioning precondition(s) are placed in *COND_EXPR and
 *COND_EXPR_STMT_LIST.  */
-void
+class loop *
 vect_loop_versioning (loop_vec_info loop_vinfo,
-		      unsigned int th, bool check_profitability,
+		      gimple *loop_vectorized_call)
-		      poly_uint64 versioning_threshold)
+{
-{
+class loop *loop = LOOP_VINFO_LOOP (loop_vinfo), *nloop;
-struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo), *nloop;
+class loop *scalar_loop = LOOP_VINFO_SCALAR_LOOP (loop_vinfo);
-struct loop *scalar_loop = LOOP_VINFO_SCALAR_LOOP (loop_vinfo);
 basic_block condition_bb;
 gphi_iterator gsi;
 gimple_stmt_iterator cond_exp_gsi;
 basic_block merge_bb;
 basic_block new_exit_bb;
 gimple_seq gimplify_stmt_list = NULL;
 tree scalar_loop_iters = LOOP_VINFO_NITERSM1 (loop_vinfo);
 bool version_align = LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo);
 bool version_alias = LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo);
 bool version_niter = LOOP_REQUIRES_VERSIONING_FOR_NITERS (loop_vinfo);
+poly_uint64 versioning_threshold
-if (check_profitability)
+= LOOP_VINFO_VERSIONING_THRESHOLD (loop_vinfo);
+tree version_simd_if_cond
+= LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND (loop_vinfo);
+unsigned th = LOOP_VINFO_COST_MODEL_THRESHOLD (loop_vinfo);
+if (vect_apply_runtime_profitability_check_p (loop_vinfo)
+&& !ordered_p (th, versioning_threshold))
 cond_expr = fold_build2 (GE_EXPR, boolean_type_node, scalar_loop_iters,
 			     build_int_cst (TREE_TYPE (scalar_loop_iters),
 					    th - 1));
 if (maybe_ne (versioning_threshold, 0U))
 {
 if (version_niter)
 vect_create_cond_for_niters_checks (loop_vinfo, &cond_expr);
 if (cond_expr)
-cond_expr = force_gimple_operand_1 (cond_expr, &cond_expr_stmt_list,
+cond_expr = force_gimple_operand_1 (unshare_expr (cond_expr),
+					&cond_expr_stmt_list,
 					is_gimple_condexpr, NULL_TREE);
 if (version_align)
 vect_create_cond_for_align_checks (loop_vinfo, &cond_expr,
 				       &cond_expr_stmt_list);
 if (version_alias)
 {
 vect_create_cond_for_unequal_addrs (loop_vinfo, &cond_expr);
 vect_create_cond_for_lower_bounds (loop_vinfo, &cond_expr);
 vect_create_cond_for_alias_checks (loop_vinfo, &cond_expr);
+}
+if (version_simd_if_cond)
+{
+gcc_assert (dom_info_available_p (CDI_DOMINATORS));
+if (flag_checking)
+	if (basic_block bb
+	    = gimple_bb (SSA_NAME_DEF_STMT (version_simd_if_cond)))
+	  gcc_assert (bb != loop->header
+		      && dominated_by_p (CDI_DOMINATORS, loop->header, bb)
+		      && (scalar_loop == NULL
+			  || (bb != scalar_loop->header
+			      && dominated_by_p (CDI_DOMINATORS,
+						 scalar_loop->header, bb))));
+tree zero = build_zero_cst (TREE_TYPE (version_simd_if_cond));
+tree c = fold_build2 (NE_EXPR, boolean_type_node,
+			    version_simd_if_cond, zero);
+if (cond_expr)
+cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
+				 c, cond_expr);
+else
+cond_expr = c;
+if (dump_enabled_p ())
+	dump_printf_loc (MSG_NOTE, vect_location,
+			 "created versioning for simd if condition check.\n");
 }
 cond_expr = force_gimple_operand_1 (unshare_expr (cond_expr),
 				      &gimplify_stmt_list,
 				      is_gimple_condexpr, NULL_TREE);
 gimple_seq_add_seq (&cond_expr_stmt_list, gimplify_stmt_list);
-initialize_original_copy_tables ();
+/* Compute the outermost loop cond_expr and cond_expr_stmt_list are
-if (scalar_loop)
+invariant in.  */
-{
+class loop *outermost = outermost_invariant_loop_for_expr (loop, cond_expr);
-edge scalar_e;
+for (gimple_stmt_iterator gsi = gsi_start (cond_expr_stmt_list);
-basic_block preheader, scalar_preheader;
+!gsi_end_p (gsi); gsi_next (&gsi))
+{
-/* We don't want to scale SCALAR_LOOP's frequencies, we need to
+gimple *stmt = gsi_stmt (gsi);
-	 scale LOOP's frequencies instead.  */
+update_stmt (stmt);
-nloop = loop_version (scalar_loop, cond_expr, &condition_bb,
+ssa_op_iter iter;
+use_operand_p use_p;
+basic_block def_bb;
+FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE)
+	if ((def_bb = gimple_bb (SSA_NAME_DEF_STMT (USE_FROM_PTR (use_p))))
+	    && flow_bb_inside_loop_p (outermost, def_bb))
+	  outermost = superloop_at_depth (loop, bb_loop_depth (def_bb) + 1);
+}
+/* Search for the outermost loop we can version.  Avoid versioning of
+non-perfect nests but allow if-conversion versioned loops inside.  */
+class loop *loop_to_version = loop;
+if (flow_loop_nested_p (outermost, loop))
+{
+if (dump_enabled_p ())
+	dump_printf_loc (MSG_NOTE, vect_location,
+			 "trying to apply versioning to outer loop %d\n",
+			 outermost->num);
+if (outermost->num == 0)
+	outermost = superloop_at_depth (loop, 1);
+/* And avoid applying versioning on non-perfect nests.  */
+while (loop_to_version != outermost
+	     && single_exit (loop_outer (loop_to_version))
+	     && (!loop_outer (loop_to_version)->inner->next
+		 || vect_loop_vectorized_call (loop_to_version))
+	     && (!loop_outer (loop_to_version)->inner->next
+		 || !loop_outer (loop_to_version)->inner->next->next))
+	loop_to_version = loop_outer (loop_to_version);
+}
+/* Apply versioning.  If there is already a scalar version created by
+if-conversion re-use that.  Note we cannot re-use the copy of
+an if-converted outer-loop when vectorizing the inner loop only.  */
+gcond *cond;
+if ((!loop_to_version->inner || loop == loop_to_version)
+&& loop_vectorized_call)
+{
+gcc_assert (scalar_loop);
+condition_bb = gimple_bb (loop_vectorized_call);
+cond = as_a <gcond *> (last_stmt (condition_bb));
+gimple_cond_set_condition_from_tree (cond, cond_expr);
+update_stmt (cond);
+if (cond_expr_stmt_list)
+	{
+	  cond_exp_gsi = gsi_for_stmt (loop_vectorized_call);
+	  gsi_insert_seq_before (&cond_exp_gsi, cond_expr_stmt_list,
+				 GSI_SAME_STMT);
+	}
+/* if-conversion uses profile_probability::always () for both paths,
+	 reset the paths probabilities appropriately.  */
+edge te, fe;
+extract_true_false_edges_from_block (condition_bb, &te, &fe);
+te->probability = prob;
+fe->probability = prob.invert ();
+/* We can scale loops counts immediately but have to postpone
+scaling the scalar loop because we re-use it during peeling.  */
+scale_loop_frequencies (loop_to_version, te->probability);
+LOOP_VINFO_SCALAR_LOOP_SCALING (loop_vinfo) = fe->probability;
+nloop = scalar_loop;
+if (dump_enabled_p ())
+	dump_printf_loc (MSG_NOTE, vect_location,
+			 "reusing %sloop version created by if conversion\n",
+			 loop_to_version != loop ? "outer " : "");
+}
+else
+{
+if (loop_to_version != loop
+	  && dump_enabled_p ())
+	dump_printf_loc (MSG_NOTE, vect_location,
+			 "applying loop versioning to outer loop %d\n",
+			 loop_to_version->num);
+initialize_original_copy_tables ();
+nloop = loop_version (loop_to_version, cond_expr, &condition_bb,
 			    prob, prob.invert (), prob, prob.invert (), true);
-scale_loop_frequencies (loop, prob);
+gcc_assert (nloop);
-/* CONDITION_BB was created above SCALAR_LOOP's preheader,
+nloop = get_loop_copy (loop);
-	 while we need to move it above LOOP's preheader.  */
-e = loop_preheader_edge (loop);
+/* Kill off IFN_LOOP_VECTORIZED_CALL in the copy, nobody will
-scalar_e = loop_preheader_edge (scalar_loop);
+reap those otherwise;  they also refer to the original
-/* The vector loop preheader might not be empty, since new
+	 loops.  */
-	 invariants could have been created while analyzing the loop.  */
+class loop *l = loop;
-gcc_assert (single_pred_p (e->src));
+while (gimple *call = vect_loop_vectorized_call (l))
-gcc_assert (empty_block_p (scalar_e->src)
+	{
-		  && single_pred_p (scalar_e->src));
+	  call = SSA_NAME_DEF_STMT (get_current_def (gimple_call_lhs (call)));
-gcc_assert (single_pred_p (condition_bb));
+	  fold_loop_internal_call (call, boolean_false_node);
-preheader = e->src;
+	  l = loop_outer (l);
-scalar_preheader = scalar_e->src;
+	}
-scalar_e = find_edge (condition_bb, scalar_preheader);
+free_original_copy_tables ();
-e = single_pred_edge (preheader);
-redirect_edge_and_branch_force (single_pred_edge (condition_bb),
+if (cond_expr_stmt_list)
-				      scalar_preheader);
+	{
-redirect_edge_and_branch_force (scalar_e, preheader);
+	  cond_exp_gsi = gsi_last_bb (condition_bb);
-redirect_edge_and_branch_force (e, condition_bb);
+	  gsi_insert_seq_before (&cond_exp_gsi, cond_expr_stmt_list,
-set_immediate_dominator (CDI_DOMINATORS, condition_bb,
+				 GSI_SAME_STMT);
-			       single_pred (condition_bb));
+	}
-set_immediate_dominator (CDI_DOMINATORS, scalar_preheader,
-			       single_pred (scalar_preheader));
+/* Loop versioning violates an assumption we try to maintain during
-set_immediate_dominator (CDI_DOMINATORS, preheader,
+	 vectorization - that the loop exit block has a single predecessor.
-			       condition_bb);
+	 After versioning, the exit block of both loop versions is the same
-}
+	 basic block (i.e. it has two predecessors). Just in order to simplify
-else
+	 following transformations in the vectorizer, we fix this situation
-nloop = loop_version (loop, cond_expr, &condition_bb,
+	 here by adding a new (empty) block on the exit-edge of the loop,
-			  prob, prob.invert (), prob, prob.invert (), true);
+	 with the proper loop-exit phis to maintain loop-closed-form.
+	 If loop versioning wasn't done from loop, but scalar_loop instead,
+	 merge_bb will have already just a single successor.  */
+merge_bb = single_exit (loop_to_version)->dest;
+if (EDGE_COUNT (merge_bb->preds) >= 2)
+	{
+	  gcc_assert (EDGE_COUNT (merge_bb->preds) >= 2);
+	  new_exit_bb = split_edge (single_exit (loop_to_version));
+	  new_exit_e = single_exit (loop_to_version);
+	  e = EDGE_SUCC (new_exit_bb, 0);
+	  for (gsi = gsi_start_phis (merge_bb); !gsi_end_p (gsi);
+	       gsi_next (&gsi))
+	    {
+	      tree new_res;
+	      orig_phi = gsi.phi ();
+	      new_res = copy_ssa_name (PHI_RESULT (orig_phi));
+	      new_phi = create_phi_node (new_res, new_exit_bb);
+	      arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, e);
+	      add_phi_arg (new_phi, arg, new_exit_e,
+			   gimple_phi_arg_location_from_edge (orig_phi, e));
+	      adjust_phi_and_debug_stmts (orig_phi, e, PHI_RESULT (new_phi));
+	    }
+	}
+update_ssa (TODO_update_ssa);
+}
 if (version_niter)
 {
 /* The versioned loop could be infinite, we need to clear existing
 	 niter information which is copied from the original loop.  */
 			 vect_location,
 "loop versioned for vectorization to enhance "
 			 "alignment\n");
 }
-free_original_copy_tables ();
+return nloop;
-/* Loop versioning violates an assumption we try to maintain during
+}
-vectorization - that the loop exit block has a single predecessor.
-After versioning, the exit block of both loop versions is the same
-basic block (i.e. it has two predecessors). Just in order to simplify
-following transformations in the vectorizer, we fix this situation
-here by adding a new (empty) block on the exit-edge of the loop,
-with the proper loop-exit phis to maintain loop-closed-form.
-If loop versioning wasn't done from loop, but scalar_loop instead,
-merge_bb will have already just a single successor.  */
-merge_bb = single_exit (loop)->dest;
-if (scalar_loop == NULL || EDGE_COUNT (merge_bb->preds) >= 2)
-{
-gcc_assert (EDGE_COUNT (merge_bb->preds) >= 2);
-new_exit_bb = split_edge (single_exit (loop));
-new_exit_e = single_exit (loop);
-e = EDGE_SUCC (new_exit_bb, 0);
-for (gsi = gsi_start_phis (merge_bb); !gsi_end_p (gsi); gsi_next (&gsi))
-	{
-	  tree new_res;
-	  orig_phi = gsi.phi ();
-	  new_res = copy_ssa_name (PHI_RESULT (orig_phi));
-	  new_phi = create_phi_node (new_res, new_exit_bb);
-	  arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, e);
-	  add_phi_arg (new_phi, arg, new_exit_e,
-		       gimple_phi_arg_location_from_edge (orig_phi, e));
-	  adjust_phi_and_debug_stmts (orig_phi, e, PHI_RESULT (new_phi));
-	}
-}
-/* End loop-exit-fixes after versioning.  */
-if (cond_expr_stmt_list)
-{
-cond_exp_gsi = gsi_last_bb (condition_bb);
-gsi_insert_seq_before (&cond_exp_gsi, cond_expr_stmt_list,
-			     GSI_SAME_STMT);
-}
-update_ssa (TODO_update_ssa);
-}

Mercurial > hg > CbC > CbC_gcc

comparison gcc/tree-vect-loop-manip.c @ 145:1830386684a0