Mercurial > hg > CbC > CbC_gcc
diff gcc/tree-vect-loop.c @ 131:84e7813d76e9
gcc-8.2
| author | mir3636 |
|---|---|
| date | Thu, 25 Oct 2018 07:37:49 +0900 |
| parents | 04ced10e8804 |
| children | 1830386684a0 |
line wrap: on
line diff
--- a/gcc/tree-vect-loop.c Fri Oct 27 22:46:09 2017 +0900 +++ b/gcc/tree-vect-loop.c Thu Oct 25 07:37:49 2018 +0900 @@ -1,5 +1,5 @@ /* Loop Vectorization - Copyright (C) 2003-2017 Free Software Foundation, Inc. + Copyright (C) 2003-2018 Free Software Foundation, Inc. Contributed by Dorit Naishlos <dorit@il.ibm.com> and Ira Rosen <irar@il.ibm.com> @@ -50,6 +50,10 @@ #include "cgraph.h" #include "tree-cfg.h" #include "tree-if-conv.h" +#include "internal-fn.h" +#include "tree-vector-builder.h" +#include "vec-perm-indices.h" +#include "tree-eh.h" /* Loop Vectorization Pass. @@ -151,6 +155,109 @@ static void vect_estimate_min_profitable_iters (loop_vec_info, int *, int *); +/* Subroutine of vect_determine_vf_for_stmt that handles only one + statement. VECTYPE_MAYBE_SET_P is true if STMT_VINFO_VECTYPE + may already be set for general statements (not just data refs). */ + +static opt_result +vect_determine_vf_for_stmt_1 (stmt_vec_info stmt_info, + bool vectype_maybe_set_p, + poly_uint64 *vf, + vec<stmt_vec_info > *mask_producers) +{ + gimple *stmt = stmt_info->stmt; + + if ((!STMT_VINFO_RELEVANT_P (stmt_info) + && !STMT_VINFO_LIVE_P (stmt_info)) + || gimple_clobber_p (stmt)) + { + if (dump_enabled_p ()) + dump_printf_loc (MSG_NOTE, vect_location, "skip.\n"); + return opt_result::success (); + } + + tree stmt_vectype, nunits_vectype; + opt_result res = vect_get_vector_types_for_stmt (stmt_info, &stmt_vectype, + &nunits_vectype); + if (!res) + return res; + + if (stmt_vectype) + { + if (STMT_VINFO_VECTYPE (stmt_info)) + /* The only case when a vectype had been already set is for stmts + that contain a data ref, or for "pattern-stmts" (stmts generated + by the vectorizer to represent/replace a certain idiom). */ + gcc_assert ((STMT_VINFO_DATA_REF (stmt_info) + || vectype_maybe_set_p) + && STMT_VINFO_VECTYPE (stmt_info) == stmt_vectype); + else if (stmt_vectype == boolean_type_node) + mask_producers->safe_push (stmt_info); + else + STMT_VINFO_VECTYPE (stmt_info) = stmt_vectype; + } + + if (nunits_vectype) + vect_update_max_nunits (vf, nunits_vectype); + + return opt_result::success (); +} + +/* Subroutine of vect_determine_vectorization_factor. Set the vector + types of STMT_INFO and all attached pattern statements and update + the vectorization factor VF accordingly. If some of the statements + produce a mask result whose vector type can only be calculated later, + add them to MASK_PRODUCERS. Return true on success or false if + something prevented vectorization. */ + +static opt_result +vect_determine_vf_for_stmt (stmt_vec_info stmt_info, poly_uint64 *vf, + vec<stmt_vec_info > *mask_producers) +{ + vec_info *vinfo = stmt_info->vinfo; + if (dump_enabled_p ()) + dump_printf_loc (MSG_NOTE, vect_location, "==> examining statement: %G", + stmt_info->stmt); + opt_result res + = vect_determine_vf_for_stmt_1 (stmt_info, false, vf, mask_producers); + if (!res) + return res; + + if (STMT_VINFO_IN_PATTERN_P (stmt_info) + && STMT_VINFO_RELATED_STMT (stmt_info)) + { + gimple *pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info); + stmt_info = STMT_VINFO_RELATED_STMT (stmt_info); + + /* If a pattern statement has def stmts, analyze them too. */ + for (gimple_stmt_iterator si = gsi_start (pattern_def_seq); + !gsi_end_p (si); gsi_next (&si)) + { + stmt_vec_info def_stmt_info = vinfo->lookup_stmt (gsi_stmt (si)); + if (dump_enabled_p ()) + dump_printf_loc (MSG_NOTE, vect_location, + "==> examining pattern def stmt: %G", + def_stmt_info->stmt); + if (!vect_determine_vf_for_stmt_1 (def_stmt_info, true, + vf, mask_producers)) + res = vect_determine_vf_for_stmt_1 (def_stmt_info, true, + vf, mask_producers); + if (!res) + return res; + } + + if (dump_enabled_p ()) + dump_printf_loc (MSG_NOTE, vect_location, + "==> examining pattern statement: %G", + stmt_info->stmt); + res = vect_determine_vf_for_stmt_1 (stmt_info, true, vf, mask_producers); + if (!res) + return res; + } + + return opt_result::success (); +} + /* Function vect_determine_vectorization_factor Determine the vectorization factor (VF). VF is the number of data elements @@ -176,30 +283,21 @@ } */ -static bool +static opt_result vect_determine_vectorization_factor (loop_vec_info loop_vinfo) { struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); unsigned nbbs = loop->num_nodes; - unsigned int vectorization_factor = 0; + poly_uint64 vectorization_factor = 1; tree scalar_type = NULL_TREE; gphi *phi; tree vectype; - unsigned int nunits; stmt_vec_info stmt_info; unsigned i; - HOST_WIDE_INT dummy; - gimple *stmt, *pattern_stmt = NULL; - gimple_seq pattern_def_seq = NULL; - gimple_stmt_iterator pattern_def_si = gsi_none (); - bool analyze_pattern_stmt = false; - bool bool_result; auto_vec<stmt_vec_info> mask_producers; - if (dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, - "=== vect_determine_vectorization_factor ===\n"); + DUMP_VECT_SCOPE ("vect_determine_vectorization_factor"); for (i = 0; i < nbbs; i++) { @@ -209,12 +307,10 @@ gsi_next (&si)) { phi = si.phi (); - stmt_info = vinfo_for_stmt (phi); + stmt_info = loop_vinfo->lookup_stmt (phi); if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, "==> examining phi: "); - dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); - } + dump_printf_loc (MSG_NOTE, vect_location, "==> examining phi: %G", + phi); gcc_assert (stmt_info); @@ -225,477 +321,68 @@ scalar_type = TREE_TYPE (PHI_RESULT (phi)); if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, - "get vectype for scalar type: "); - dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); - dump_printf (MSG_NOTE, "\n"); - } + dump_printf_loc (MSG_NOTE, vect_location, + "get vectype for scalar type: %T\n", + scalar_type); vectype = get_vectype_for_scalar_type (scalar_type); if (!vectype) - { - if (dump_enabled_p ()) - { - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: unsupported " - "data-type "); - dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, - scalar_type); - dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); - } - return false; - } + return opt_result::failure_at (phi, + "not vectorized: unsupported " + "data-type %T\n", + scalar_type); STMT_VINFO_VECTYPE (stmt_info) = vectype; if (dump_enabled_p ()) + dump_printf_loc (MSG_NOTE, vect_location, "vectype: %T\n", + vectype); + + if (dump_enabled_p ()) { - dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); - dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); - dump_printf (MSG_NOTE, "\n"); + dump_printf_loc (MSG_NOTE, vect_location, "nunits = "); + dump_dec (MSG_NOTE, TYPE_VECTOR_SUBPARTS (vectype)); + dump_printf (MSG_NOTE, "\n"); } - nunits = TYPE_VECTOR_SUBPARTS (vectype); - if (dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, "nunits = %d\n", - nunits); - - if (!vectorization_factor - || (nunits > vectorization_factor)) - vectorization_factor = nunits; + vect_update_max_nunits (&vectorization_factor, vectype); } } - for (gimple_stmt_iterator si = gsi_start_bb (bb); - !gsi_end_p (si) || analyze_pattern_stmt;) - { - tree vf_vectype; - - if (analyze_pattern_stmt) - stmt = pattern_stmt; - else - stmt = gsi_stmt (si); - - stmt_info = vinfo_for_stmt (stmt); - - if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, - "==> examining statement: "); - dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); - } - - gcc_assert (stmt_info); - - /* Skip stmts which do not need to be vectorized. */ - if ((!STMT_VINFO_RELEVANT_P (stmt_info) - && !STMT_VINFO_LIVE_P (stmt_info)) - || gimple_clobber_p (stmt)) - { - if (STMT_VINFO_IN_PATTERN_P (stmt_info) - && (pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info)) - && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) - || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) - { - stmt = pattern_stmt; - stmt_info = vinfo_for_stmt (pattern_stmt); - if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, - "==> examining pattern statement: "); - dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); - } - } - else - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, "skip.\n"); - gsi_next (&si); - continue; - } - } - else if (STMT_VINFO_IN_PATTERN_P (stmt_info) - && (pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info)) - && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) - || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) - analyze_pattern_stmt = true; - - /* If a pattern statement has def stmts, analyze them too. */ - if (is_pattern_stmt_p (stmt_info)) - { - if (pattern_def_seq == NULL) - { - pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info); - pattern_def_si = gsi_start (pattern_def_seq); - } - else if (!gsi_end_p (pattern_def_si)) - gsi_next (&pattern_def_si); - if (pattern_def_seq != NULL) - { - gimple *pattern_def_stmt = NULL; - stmt_vec_info pattern_def_stmt_info = NULL; - - while (!gsi_end_p (pattern_def_si)) - { - pattern_def_stmt = gsi_stmt (pattern_def_si); - pattern_def_stmt_info - = vinfo_for_stmt (pattern_def_stmt); - if (STMT_VINFO_RELEVANT_P (pattern_def_stmt_info) - || STMT_VINFO_LIVE_P (pattern_def_stmt_info)) - break; - gsi_next (&pattern_def_si); - } - - if (!gsi_end_p (pattern_def_si)) - { - if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, - "==> examining pattern def stmt: "); - dump_gimple_stmt (MSG_NOTE, TDF_SLIM, - pattern_def_stmt, 0); - } - - stmt = pattern_def_stmt; - stmt_info = pattern_def_stmt_info; - } - else - { - pattern_def_si = gsi_none (); - analyze_pattern_stmt = false; - } - } - else - analyze_pattern_stmt = false; - } - - if (gimple_get_lhs (stmt) == NULL_TREE - /* MASK_STORE has no lhs, but is ok. */ - && (!is_gimple_call (stmt) - || !gimple_call_internal_p (stmt) - || gimple_call_internal_fn (stmt) != IFN_MASK_STORE)) - { - if (is_gimple_call (stmt)) - { - /* Ignore calls with no lhs. These must be calls to - #pragma omp simd functions, and what vectorization factor - it really needs can't be determined until - vectorizable_simd_clone_call. */ - if (!analyze_pattern_stmt && gsi_end_p (pattern_def_si)) - { - pattern_def_seq = NULL; - gsi_next (&si); - } - continue; - } - if (dump_enabled_p ()) - { - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: irregular stmt."); - dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, - 0); - } - return false; - } - - if (VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))) - { - if (dump_enabled_p ()) - { - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: vector stmt in loop:"); - dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); - } - return false; - } - - bool_result = false; - - if (STMT_VINFO_VECTYPE (stmt_info)) - { - /* The only case when a vectype had been already set is for stmts - that contain a dataref, or for "pattern-stmts" (stmts - generated by the vectorizer to represent/replace a certain - idiom). */ - gcc_assert (STMT_VINFO_DATA_REF (stmt_info) - || is_pattern_stmt_p (stmt_info) - || !gsi_end_p (pattern_def_si)); - vectype = STMT_VINFO_VECTYPE (stmt_info); - } - else - { - gcc_assert (!STMT_VINFO_DATA_REF (stmt_info)); - if (gimple_call_internal_p (stmt, IFN_MASK_STORE)) - scalar_type = TREE_TYPE (gimple_call_arg (stmt, 3)); - else - scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); - - /* Bool ops don't participate in vectorization factor - computation. For comparison use compared types to - compute a factor. */ - if (VECT_SCALAR_BOOLEAN_TYPE_P (scalar_type) - && is_gimple_assign (stmt) - && gimple_assign_rhs_code (stmt) != COND_EXPR) - { - if (STMT_VINFO_RELEVANT_P (stmt_info) - || STMT_VINFO_LIVE_P (stmt_info)) - mask_producers.safe_push (stmt_info); - bool_result = true; - - if (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)) - == tcc_comparison - && !VECT_SCALAR_BOOLEAN_TYPE_P - (TREE_TYPE (gimple_assign_rhs1 (stmt)))) - scalar_type = TREE_TYPE (gimple_assign_rhs1 (stmt)); - else - { - if (!analyze_pattern_stmt && gsi_end_p (pattern_def_si)) - { - pattern_def_seq = NULL; - gsi_next (&si); - } - continue; - } - } - - if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, - "get vectype for scalar type: "); - dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); - dump_printf (MSG_NOTE, "\n"); - } - vectype = get_vectype_for_scalar_type (scalar_type); - if (!vectype) - { - if (dump_enabled_p ()) - { - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: unsupported " - "data-type "); - dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, - scalar_type); - dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); - } - return false; - } - - if (!bool_result) - STMT_VINFO_VECTYPE (stmt_info) = vectype; - - if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); - dump_generic_expr (MSG_NOTE, TDF_SLIM, vectype); - dump_printf (MSG_NOTE, "\n"); - } - } - - /* Don't try to compute VF out scalar types if we stmt - produces boolean vector. Use result vectype instead. */ - if (VECTOR_BOOLEAN_TYPE_P (vectype)) - vf_vectype = vectype; - else - { - /* The vectorization factor is according to the smallest - scalar type (or the largest vector size, but we only - support one vector size per loop). */ - if (!bool_result) - scalar_type = vect_get_smallest_scalar_type (stmt, &dummy, - &dummy); - if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, - "get vectype for scalar type: "); - dump_generic_expr (MSG_NOTE, TDF_SLIM, scalar_type); - dump_printf (MSG_NOTE, "\n"); - } - vf_vectype = get_vectype_for_scalar_type (scalar_type); - } - if (!vf_vectype) - { - if (dump_enabled_p ()) - { - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: unsupported data-type "); - dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, - scalar_type); - dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); - } - return false; - } - - if ((GET_MODE_SIZE (TYPE_MODE (vectype)) - != GET_MODE_SIZE (TYPE_MODE (vf_vectype)))) - { - if (dump_enabled_p ()) - { - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: different sized vector " - "types in statement, "); - dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, - vectype); - dump_printf (MSG_MISSED_OPTIMIZATION, " and "); - dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, - vf_vectype); - dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); - } - return false; - } - - if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, "vectype: "); - dump_generic_expr (MSG_NOTE, TDF_SLIM, vf_vectype); - dump_printf (MSG_NOTE, "\n"); - } - - nunits = TYPE_VECTOR_SUBPARTS (vf_vectype); - if (dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, "nunits = %d\n", nunits); - if (!vectorization_factor - || (nunits > vectorization_factor)) - vectorization_factor = nunits; - - if (!analyze_pattern_stmt && gsi_end_p (pattern_def_si)) - { - pattern_def_seq = NULL; - gsi_next (&si); - } + for (gimple_stmt_iterator si = gsi_start_bb (bb); !gsi_end_p (si); + gsi_next (&si)) + { + stmt_info = loop_vinfo->lookup_stmt (gsi_stmt (si)); + opt_result res + = vect_determine_vf_for_stmt (stmt_info, &vectorization_factor, + &mask_producers); + if (!res) + return res; } } /* TODO: Analyze cost. Decide if worth while to vectorize. */ if (dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, "vectorization factor = %d\n", - vectorization_factor); - if (vectorization_factor <= 1) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: unsupported data-type\n"); - return false; - } + { + dump_printf_loc (MSG_NOTE, vect_location, "vectorization factor = "); + dump_dec (MSG_NOTE, vectorization_factor); + dump_printf (MSG_NOTE, "\n"); + } + + if (known_le (vectorization_factor, 1U)) + return opt_result::failure_at (vect_location, + "not vectorized: unsupported data-type\n"); LOOP_VINFO_VECT_FACTOR (loop_vinfo) = vectorization_factor; for (i = 0; i < mask_producers.length (); i++) { - tree mask_type = NULL; - - stmt = STMT_VINFO_STMT (mask_producers[i]); - - if (is_gimple_assign (stmt) - && TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)) == tcc_comparison - && !VECT_SCALAR_BOOLEAN_TYPE_P - (TREE_TYPE (gimple_assign_rhs1 (stmt)))) - { - scalar_type = TREE_TYPE (gimple_assign_rhs1 (stmt)); - mask_type = get_mask_type_for_scalar_type (scalar_type); - - if (!mask_type) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: unsupported mask\n"); - return false; - } - } - else - { - tree rhs; - ssa_op_iter iter; - gimple *def_stmt; - enum vect_def_type dt; - - FOR_EACH_SSA_TREE_OPERAND (rhs, stmt, iter, SSA_OP_USE) - { - if (!vect_is_simple_use (rhs, mask_producers[i]->vinfo, - &def_stmt, &dt, &vectype)) - { - if (dump_enabled_p ()) - { - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: can't compute mask type " - "for statement, "); - dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, - 0); - } - return false; - } - - /* No vectype probably means external definition. - Allow it in case there is another operand which - allows to determine mask type. */ - if (!vectype) - continue; - - if (!mask_type) - mask_type = vectype; - else if (TYPE_VECTOR_SUBPARTS (mask_type) - != TYPE_VECTOR_SUBPARTS (vectype)) - { - if (dump_enabled_p ()) - { - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: different sized masks " - "types in statement, "); - dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, - mask_type); - dump_printf (MSG_MISSED_OPTIMIZATION, " and "); - dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, - vectype); - dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); - } - return false; - } - else if (VECTOR_BOOLEAN_TYPE_P (mask_type) - != VECTOR_BOOLEAN_TYPE_P (vectype)) - { - if (dump_enabled_p ()) - { - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: mixed mask and " - "nonmask vector types in statement, "); - dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, - mask_type); - dump_printf (MSG_MISSED_OPTIMIZATION, " and "); - dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, - vectype); - dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); - } - return false; - } - } - - /* We may compare boolean value loaded as vector of integers. - Fix mask_type in such case. */ - if (mask_type - && !VECTOR_BOOLEAN_TYPE_P (mask_type) - && gimple_code (stmt) == GIMPLE_ASSIGN - && TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)) == tcc_comparison) - mask_type = build_same_sized_truth_vector_type (mask_type); - } - - /* No mask_type should mean loop invariant predicate. - This is probably a subject for optimization in - if-conversion. */ + stmt_info = mask_producers[i]; + opt_tree mask_type = vect_get_mask_type_for_stmt (stmt_info); if (!mask_type) - { - if (dump_enabled_p ()) - { - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: can't compute mask type " - "for statement, "); - dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, - 0); - } - return false; - } - - STMT_VINFO_VECTYPE (mask_producers[i]) = mask_type; - } - - return true; + return opt_result::propagate_failure (mask_type); + STMT_VINFO_VECTYPE (stmt_info) = mask_type; + } + + return opt_result::success (); } @@ -727,13 +414,8 @@ init_expr = unshare_expr (initial_condition_in_loop_num (access_fn, loop_nb)); if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, "step: "); - dump_generic_expr (MSG_NOTE, TDF_SLIM, step_expr); - dump_printf (MSG_NOTE, ", init: "); - dump_generic_expr (MSG_NOTE, TDF_SLIM, init_expr); - dump_printf (MSG_NOTE, "\n"); - } + dump_printf_loc (MSG_NOTE, vect_location, "step: %T, init: %T\n", + step_expr, init_expr); *init = init_expr; *step = step_expr; @@ -757,6 +439,40 @@ return true; } +/* Return true if PHI, described by STMT_INFO, is the inner PHI in + what we are assuming is a double reduction. For example, given + a structure like this: + + outer1: + x_1 = PHI <x_4(outer2), ...>; + ... + + inner: + x_2 = PHI <x_1(outer1), ...>; + ... + x_3 = ...; + ... + + outer2: + x_4 = PHI <x_3(inner)>; + ... + + outer loop analysis would treat x_1 as a double reduction phi and + this function would then return true for x_2. */ + +static bool +vect_inner_phi_in_double_reduction_p (stmt_vec_info stmt_info, gphi *phi) +{ + loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); + use_operand_p use_p; + ssa_op_iter op_iter; + FOR_EACH_PHI_ARG (use_p, phi, op_iter, SSA_OP_USE) + if (stmt_vec_info def_info = loop_vinfo->lookup_def (USE_FROM_PTR (use_p))) + if (STMT_VINFO_DEF_TYPE (def_info) == vect_double_reduction_def) + return true; + return false; +} + /* Function vect_analyze_scalar_cycles_1. Examine the cross iteration def-use cycles of scalar variables @@ -769,13 +485,11 @@ { basic_block bb = loop->header; tree init, step; - auto_vec<gimple *, 64> worklist; + auto_vec<stmt_vec_info, 64> worklist; gphi_iterator gsi; bool double_reduc; - if (dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, - "=== vect_analyze_scalar_cycles ===\n"); + DUMP_VECT_SCOPE ("vect_analyze_scalar_cycles"); /* First - identify all inductions. Reduction detection assumes that all the inductions have been identified, therefore, this order must not be @@ -785,13 +499,10 @@ gphi *phi = gsi.phi (); tree access_fn = NULL; tree def = PHI_RESULT (phi); - stmt_vec_info stmt_vinfo = vinfo_for_stmt (phi); + stmt_vec_info stmt_vinfo = loop_vinfo->lookup_stmt (phi); if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, "Analyze phi: "); - dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); - } + dump_printf_loc (MSG_NOTE, vect_location, "Analyze phi: %G", phi); /* Skip virtual phi's. The data dependences that are associated with virtual defs/uses (i.e., memory accesses) are analyzed elsewhere. */ @@ -806,12 +517,8 @@ { STRIP_NOPS (access_fn); if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, - "Access function of PHI: "); - dump_generic_expr (MSG_NOTE, TDF_SLIM, access_fn); - dump_printf (MSG_NOTE, "\n"); - } + dump_printf_loc (MSG_NOTE, vect_location, + "Access function of PHI: %T\n", access_fn); STMT_VINFO_LOOP_PHI_EVOLUTION_BASE_UNCHANGED (stmt_vinfo) = initial_condition_in_loop_num (access_fn, loop->num); STMT_VINFO_LOOP_PHI_EVOLUTION_PART (stmt_vinfo) @@ -819,11 +526,12 @@ } if (!access_fn + || vect_inner_phi_in_double_reduction_p (stmt_vinfo, phi) || !vect_is_simple_iv_evolution (loop->num, access_fn, &init, &step) || (LOOP_VINFO_LOOP (loop_vinfo) != loop && TREE_CODE (step) != INTEGER_CST)) { - worklist.safe_push (phi); + worklist.safe_push (stmt_vinfo); continue; } @@ -840,23 +548,20 @@ /* Second - identify all reductions and nested cycles. */ while (worklist.length () > 0) { - gimple *phi = worklist.pop (); + stmt_vec_info stmt_vinfo = worklist.pop (); + gphi *phi = as_a <gphi *> (stmt_vinfo->stmt); tree def = PHI_RESULT (phi); - stmt_vec_info stmt_vinfo = vinfo_for_stmt (phi); - gimple *reduc_stmt; if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, "Analyze phi: "); - dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); - } + dump_printf_loc (MSG_NOTE, vect_location, "Analyze phi: %G", phi); gcc_assert (!virtual_operand_p (def) && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_unknown_def_type); - reduc_stmt = vect_force_simple_reduction (loop_vinfo, phi, - &double_reduc, false); - if (reduc_stmt) + stmt_vec_info reduc_stmt_info + = vect_force_simple_reduction (loop_vinfo, stmt_vinfo, + &double_reduc, false); + if (reduc_stmt_info) { if (double_reduc) { @@ -865,8 +570,8 @@ "Detected double reduction.\n"); STMT_VINFO_DEF_TYPE (stmt_vinfo) = vect_double_reduction_def; - STMT_VINFO_DEF_TYPE (vinfo_for_stmt (reduc_stmt)) = - vect_double_reduction_def; + STMT_VINFO_DEF_TYPE (reduc_stmt_info) + = vect_double_reduction_def; } else { @@ -877,8 +582,7 @@ "Detected vectorizable nested cycle.\n"); STMT_VINFO_DEF_TYPE (stmt_vinfo) = vect_nested_cycle; - STMT_VINFO_DEF_TYPE (vinfo_for_stmt (reduc_stmt)) = - vect_nested_cycle; + STMT_VINFO_DEF_TYPE (reduc_stmt_info) = vect_nested_cycle; } else { @@ -887,13 +591,13 @@ "Detected reduction.\n"); STMT_VINFO_DEF_TYPE (stmt_vinfo) = vect_reduction_def; - STMT_VINFO_DEF_TYPE (vinfo_for_stmt (reduc_stmt)) = - vect_reduction_def; + STMT_VINFO_DEF_TYPE (reduc_stmt_info) = vect_reduction_def; /* Store the reduction cycles for possible vectorization in loop-aware SLP if it was not detected as reduction chain. */ - if (! GROUP_FIRST_ELEMENT (vinfo_for_stmt (reduc_stmt))) - LOOP_VINFO_REDUCTIONS (loop_vinfo).safe_push (reduc_stmt); + if (! REDUC_GROUP_FIRST_ELEMENT (reduc_stmt_info)) + LOOP_VINFO_REDUCTIONS (loop_vinfo).safe_push + (reduc_stmt_info); } } } @@ -946,27 +650,28 @@ vect_analyze_scalar_cycles_1 (loop_vinfo, loop->inner); } -/* Transfer group and reduction information from STMT to its pattern stmt. */ +/* Transfer group and reduction information from STMT_INFO to its + pattern stmt. */ static void -vect_fixup_reduc_chain (gimple *stmt) +vect_fixup_reduc_chain (stmt_vec_info stmt_info) { - gimple *firstp = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (stmt)); - gimple *stmtp; - gcc_assert (!GROUP_FIRST_ELEMENT (vinfo_for_stmt (firstp)) - && GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))); - GROUP_SIZE (vinfo_for_stmt (firstp)) = GROUP_SIZE (vinfo_for_stmt (stmt)); + stmt_vec_info firstp = STMT_VINFO_RELATED_STMT (stmt_info); + stmt_vec_info stmtp; + gcc_assert (!REDUC_GROUP_FIRST_ELEMENT (firstp) + && REDUC_GROUP_FIRST_ELEMENT (stmt_info)); + REDUC_GROUP_SIZE (firstp) = REDUC_GROUP_SIZE (stmt_info); do { - stmtp = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (stmt)); - GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmtp)) = firstp; - stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (stmt)); - if (stmt) - GROUP_NEXT_ELEMENT (vinfo_for_stmt (stmtp)) - = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (stmt)); - } - while (stmt); - STMT_VINFO_DEF_TYPE (vinfo_for_stmt (stmtp)) = vect_reduction_def; + stmtp = STMT_VINFO_RELATED_STMT (stmt_info); + REDUC_GROUP_FIRST_ELEMENT (stmtp) = firstp; + stmt_info = REDUC_GROUP_NEXT_ELEMENT (stmt_info); + if (stmt_info) + REDUC_GROUP_NEXT_ELEMENT (stmtp) + = STMT_VINFO_RELATED_STMT (stmt_info); + } + while (stmt_info); + STMT_VINFO_DEF_TYPE (stmtp) = vect_reduction_def; } /* Fixup scalar cycles that now have their stmts detected as patterns. */ @@ -974,18 +679,18 @@ static void vect_fixup_scalar_cycles_with_patterns (loop_vec_info loop_vinfo) { - gimple *first; + stmt_vec_info first; unsigned i; FOR_EACH_VEC_ELT (LOOP_VINFO_REDUCTION_CHAINS (loop_vinfo), i, first) - if (STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (first))) + if (STMT_VINFO_IN_PATTERN_P (first)) { - gimple *next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (first)); + stmt_vec_info next = REDUC_GROUP_NEXT_ELEMENT (first); while (next) { - if (! STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (next))) + if (! STMT_VINFO_IN_PATTERN_P (next)) break; - next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next)); + next = REDUC_GROUP_NEXT_ELEMENT (next); } /* If not all stmt in the chain are patterns try to handle the chain without patterns. */ @@ -993,7 +698,7 @@ { vect_fixup_reduc_chain (first); LOOP_VINFO_REDUCTION_CHAINS (loop_vinfo)[i] - = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (first)); + = STMT_VINFO_RELATED_STMT (first); } } } @@ -1020,9 +725,7 @@ *assumptions = boolean_true_node; *number_of_iterationsm1 = chrec_dont_know; *number_of_iterations = chrec_dont_know; - if (dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, - "=== get_loop_niters ===\n"); + DUMP_VECT_SCOPE ("get_loop_niters"); if (!exit) return cond; @@ -1055,7 +758,8 @@ may_be_zero)); else niter = fold_build3 (COND_EXPR, TREE_TYPE (niter), may_be_zero, - build_int_cst (TREE_TYPE (niter), 0), niter); + build_int_cst (TREE_TYPE (niter), 0), + rewrite_to_non_trapping_overflow (niter)); may_be_zero = NULL_TREE; } @@ -1101,8 +805,8 @@ /* Create and initialize a new loop_vec_info struct for LOOP_IN, as well as stmt_vec_info structs for all the stmts in LOOP_IN. */ -_loop_vec_info::_loop_vec_info (struct loop *loop_in) - : vec_info (vec_info::loop, init_cost (loop_in)), +_loop_vec_info::_loop_vec_info (struct loop *loop_in, vec_info_shared *shared) + : vec_info (vec_info::loop, init_cost (loop_in), shared), loop (loop_in), bbs (XCNEWVEC (basic_block, loop->num_nodes)), num_itersm1 (NULL_TREE), @@ -1110,14 +814,20 @@ num_iters_unchanged (NULL_TREE), num_iters_assumptions (NULL_TREE), th (0), + versioning_threshold (0), vectorization_factor (0), max_vectorization_factor (0), + mask_skip_niters (NULL_TREE), + mask_compare_type (NULL_TREE), unaligned_dr (NULL), peeling_for_alignment (0), ptr_mask (0), + ivexpr_map (NULL), slp_unrolling_factor (1), single_scalar_iteration_cost (0), vectorizable (false), + can_fully_mask_p (true), + fully_masked_p (false), peeling_for_gaps (false), peeling_for_niter (false), operands_swapped (false), @@ -1126,29 +836,6 @@ scalar_loop (NULL), orig_loop_info (NULL) { - /* Create/Update stmt_info for all stmts in the loop. */ - basic_block *body = get_loop_body (loop); - for (unsigned int i = 0; i < loop->num_nodes; i++) - { - basic_block bb = body[i]; - gimple_stmt_iterator si; - - for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) - { - gimple *phi = gsi_stmt (si); - gimple_set_uid (phi, 0); - set_vinfo_for_stmt (phi, new_stmt_vec_info (phi, this)); - } - - for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) - { - gimple *stmt = gsi_stmt (si); - gimple_set_uid (stmt, 0); - set_vinfo_for_stmt (stmt, new_stmt_vec_info (stmt, this)); - } - } - free (body); - /* CHECKME: We want to visit all BBs before their successors (except for latch blocks, for which this assertion wouldn't hold). In the simple case of the loop forms we allow, a dfs order of the BBs would the same @@ -1157,8 +844,39 @@ unsigned int nbbs = dfs_enumerate_from (loop->header, 0, bb_in_loop_p, bbs, loop->num_nodes, loop); gcc_assert (nbbs == loop->num_nodes); + + for (unsigned int i = 0; i < nbbs; i++) + { + basic_block bb = bbs[i]; + gimple_stmt_iterator si; + + for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) + { + gimple *phi = gsi_stmt (si); + gimple_set_uid (phi, 0); + add_stmt (phi); + } + + for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) + { + gimple *stmt = gsi_stmt (si); + gimple_set_uid (stmt, 0); + add_stmt (stmt); + } + } } +/* Free all levels of MASKS. */ + +void +release_vec_loop_masks (vec_loop_masks *masks) +{ + rgroup_masks *rgm; + unsigned int i; + FOR_EACH_VEC_ELT (*masks, i, rgm) + rgm->masks.release (); + masks->release (); +} /* Free all memory used by the _loop_vec_info, as well as all the stmt_vec_info structs of all the stmts in the loop. */ @@ -1173,9 +891,6 @@ for (j = 0; j < nbbs; j++) { basic_block bb = bbs[j]; - for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) - free_stmt_vec_info (gsi_stmt (si)); - for (si = gsi_start_bb (bb); !gsi_end_p (si); ) { gimple *stmt = gsi_stmt (si); @@ -1215,18 +930,138 @@ } } } - - /* Free stmt_vec_info. */ - free_stmt_vec_info (stmt); gsi_next (&si); } } free (bbs); + release_vec_loop_masks (&masks); + delete ivexpr_map; + loop->aux = NULL; } +/* Return an invariant or register for EXPR and emit necessary + computations in the LOOP_VINFO loop preheader. */ + +tree +cse_and_gimplify_to_preheader (loop_vec_info loop_vinfo, tree expr) +{ + if (is_gimple_reg (expr) + || is_gimple_min_invariant (expr)) + return expr; + + if (! loop_vinfo->ivexpr_map) + loop_vinfo->ivexpr_map = new hash_map<tree_operand_hash, tree>; + tree &cached = loop_vinfo->ivexpr_map->get_or_insert (expr); + if (! cached) + { + gimple_seq stmts = NULL; + cached = force_gimple_operand (unshare_expr (expr), + &stmts, true, NULL_TREE); + if (stmts) + { + edge e = loop_preheader_edge (LOOP_VINFO_LOOP (loop_vinfo)); + gsi_insert_seq_on_edge_immediate (e, stmts); + } + } + return cached; +} + +/* Return true if we can use CMP_TYPE as the comparison type to produce + all masks required to mask LOOP_VINFO. */ + +static bool +can_produce_all_loop_masks_p (loop_vec_info loop_vinfo, tree cmp_type) +{ + rgroup_masks *rgm; + unsigned int i; + FOR_EACH_VEC_ELT (LOOP_VINFO_MASKS (loop_vinfo), i, rgm) + if (rgm->mask_type != NULL_TREE + && !direct_internal_fn_supported_p (IFN_WHILE_ULT, + cmp_type, rgm->mask_type, + OPTIMIZE_FOR_SPEED)) + return false; + return true; +} + +/* Calculate the maximum number of scalars per iteration for every + rgroup in LOOP_VINFO. */ + +static unsigned int +vect_get_max_nscalars_per_iter (loop_vec_info loop_vinfo) +{ + unsigned int res = 1; + unsigned int i; + rgroup_masks *rgm; + FOR_EACH_VEC_ELT (LOOP_VINFO_MASKS (loop_vinfo), i, rgm) + res = MAX (res, rgm->max_nscalars_per_iter); + return res; +} + +/* Each statement in LOOP_VINFO can be masked where necessary. Check + whether we can actually generate the masks required. Return true if so, + storing the type of the scalar IV in LOOP_VINFO_MASK_COMPARE_TYPE. */ + +static bool +vect_verify_full_masking (loop_vec_info loop_vinfo) +{ + struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); + unsigned int min_ni_width; + + /* Use a normal loop if there are no statements that need masking. + This only happens in rare degenerate cases: it means that the loop + has no loads, no stores, and no live-out values. */ + if (LOOP_VINFO_MASKS (loop_vinfo).is_empty ()) + return false; + + /* Get the maximum number of iterations that is representable + in the counter type. */ + tree ni_type = TREE_TYPE (LOOP_VINFO_NITERSM1 (loop_vinfo)); + widest_int max_ni = wi::to_widest (TYPE_MAX_VALUE (ni_type)) + 1; + + /* Get a more refined estimate for the number of iterations. */ + widest_int max_back_edges; + if (max_loop_iterations (loop, &max_back_edges)) + max_ni = wi::smin (max_ni, max_back_edges + 1); + + /* Account for rgroup masks, in which each bit is replicated N times. */ + max_ni *= vect_get_max_nscalars_per_iter (loop_vinfo); + + /* Work out how many bits we need to represent the limit. */ + min_ni_width = wi::min_precision (max_ni, UNSIGNED); + + /* Find a scalar mode for which WHILE_ULT is supported. */ + opt_scalar_int_mode cmp_mode_iter; + tree cmp_type = NULL_TREE; + FOR_EACH_MODE_IN_CLASS (cmp_mode_iter, MODE_INT) + { + unsigned int cmp_bits = GET_MODE_BITSIZE (cmp_mode_iter.require ()); + if (cmp_bits >= min_ni_width + && targetm.scalar_mode_supported_p (cmp_mode_iter.require ())) + { + tree this_type = build_nonstandard_integer_type (cmp_bits, true); + if (this_type + && can_produce_all_loop_masks_p (loop_vinfo, this_type)) + { + /* Although we could stop as soon as we find a valid mode, + it's often better to continue until we hit Pmode, since the + operands to the WHILE are more likely to be reusable in + address calculations. */ + cmp_type = this_type; + if (cmp_bits >= GET_MODE_BITSIZE (Pmode)) + break; + } + } + } + + if (!cmp_type) + return false; + + LOOP_VINFO_MASK_COMPARE_TYPE (loop_vinfo) = cmp_type; + return true; +} /* Calculate the cost of one scalar iteration of the loop. */ static void @@ -1234,16 +1069,12 @@ { struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); - int nbbs = loop->num_nodes, factor, scalar_single_iter_cost = 0; + int nbbs = loop->num_nodes, factor; int innerloop_iters, i; - /* Count statements in scalar loop. Using this as scalar cost for a single - iteration for now. - - TODO: Add outer loop support. - - TODO: Consider assigning different costs to different scalar - statements. */ + DUMP_VECT_SCOPE ("vect_compute_single_scalar_iteration_cost"); + + /* Gather costs for statements in the scalar loop. */ /* FORNOW. */ innerloop_iters = 1; @@ -1263,7 +1094,7 @@ for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) { gimple *stmt = gsi_stmt (si); - stmt_vec_info stmt_info = vinfo_for_stmt (stmt); + stmt_vec_info stmt_info = loop_vinfo->lookup_stmt (stmt); if (!is_gimple_assign (stmt) && !is_gimple_call (stmt)) continue; @@ -1287,13 +1118,24 @@ else kind = scalar_stmt; - scalar_single_iter_cost - += record_stmt_cost (&LOOP_VINFO_SCALAR_ITERATION_COST (loop_vinfo), - factor, kind, stmt_info, 0, vect_prologue); + record_stmt_cost (&LOOP_VINFO_SCALAR_ITERATION_COST (loop_vinfo), + factor, kind, stmt_info, 0, vect_prologue); } } - LOOP_VINFO_SINGLE_SCALAR_ITERATION_COST (loop_vinfo) - = scalar_single_iter_cost; + + /* Now accumulate cost. */ + void *target_cost_data = init_cost (loop); + stmt_info_for_cost *si; + int j; + FOR_EACH_VEC_ELT (LOOP_VINFO_SCALAR_ITERATION_COST (loop_vinfo), + j, si) + (void) add_stmt_cost (target_cost_data, si->count, + si->kind, si->stmt_info, si->misalign, + vect_body); + unsigned dummy, body_cost = 0; + finish_cost (target_cost_data, &dummy, &body_cost, &dummy); + destroy_cost_data (target_cost_data); + LOOP_VINFO_SINGLE_SCALAR_ITERATION_COST (loop_vinfo) = body_cost; } @@ -1306,14 +1148,12 @@ - the number of iterations can be analyzed, i.e, a countable loop. The niter could be analyzed under some assumptions. */ -bool +opt_result vect_analyze_loop_form_1 (struct loop *loop, gcond **loop_cond, tree *assumptions, tree *number_of_iterationsm1, tree *number_of_iterations, gcond **inner_loop_cond) { - if (dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, - "=== vect_analyze_loop_form ===\n"); + DUMP_VECT_SCOPE ("vect_analyze_loop_form"); /* Different restrictions apply when we are considering an inner-most loop, vs. an outer (nested) loop. @@ -1334,20 +1174,13 @@ (exit-bb) */ if (loop->num_nodes != 2) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: control flow in loop.\n"); - return false; - } + return opt_result::failure_at (vect_location, + "not vectorized:" + " control flow in loop.\n"); if (empty_block_p (loop->header)) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: empty loop.\n"); - return false; - } + return opt_result::failure_at (vect_location, + "not vectorized: empty loop.\n"); } else { @@ -1372,75 +1205,60 @@ as described above. */ if ((loop->inner)->inner || (loop->inner)->next) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: multiple nested loops.\n"); - return false; - } + return opt_result::failure_at (vect_location, + "not vectorized:" + " multiple nested loops.\n"); if (loop->num_nodes != 5) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: control flow in loop.\n"); - return false; - } + return opt_result::failure_at (vect_location, + "not vectorized:" + " control flow in loop.\n"); entryedge = loop_preheader_edge (innerloop); if (entryedge->src != loop->header || !single_exit (innerloop) || single_exit (innerloop)->dest != EDGE_PRED (loop->latch, 0)->src) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: unsupported outerloop form.\n"); - return false; - } + return opt_result::failure_at (vect_location, + "not vectorized:" + " unsupported outerloop form.\n"); /* Analyze the inner-loop. */ tree inner_niterm1, inner_niter, inner_assumptions; - if (! vect_analyze_loop_form_1 (loop->inner, inner_loop_cond, - &inner_assumptions, &inner_niterm1, - &inner_niter, NULL) - /* Don't support analyzing niter under assumptions for inner - loop. */ - || !integer_onep (inner_assumptions)) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: Bad inner loop.\n"); - return false; - } - - if (!expr_invariant_in_loop_p (loop, inner_niter)) + opt_result res + = vect_analyze_loop_form_1 (loop->inner, inner_loop_cond, + &inner_assumptions, &inner_niterm1, + &inner_niter, NULL); + if (!res) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: inner-loop count not" - " invariant.\n"); - return false; + "not vectorized: Bad inner loop.\n"); + return res; } + /* Don't support analyzing niter under assumptions for inner + loop. */ + if (!integer_onep (inner_assumptions)) + return opt_result::failure_at (vect_location, + "not vectorized: Bad inner loop.\n"); + + if (!expr_invariant_in_loop_p (loop, inner_niter)) + return opt_result::failure_at (vect_location, + "not vectorized: inner-loop count not" + " invariant.\n"); + if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "Considering outer-loop vectorization.\n"); } - if (!single_exit (loop) - || EDGE_COUNT (loop->header->preds) != 2) - { - if (dump_enabled_p ()) - { - if (!single_exit (loop)) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: multiple exits.\n"); - else if (EDGE_COUNT (loop->header->preds) != 2) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: too many incoming edges.\n"); - } - return false; - } + if (!single_exit (loop)) + return opt_result::failure_at (vect_location, + "not vectorized: multiple exits.\n"); + if (EDGE_COUNT (loop->header->preds) != 2) + return opt_result::failure_at (vect_location, + "not vectorized:" + " too many incoming edges.\n"); /* We assume that the loop exit condition is at the end of the loop. i.e, that the loop is represented as a do-while (with a proper if-guard @@ -1448,69 +1266,54 @@ executable statements, and the latch is empty. */ if (!empty_block_p (loop->latch) || !gimple_seq_empty_p (phi_nodes (loop->latch))) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: latch block not empty.\n"); - return false; - } + return opt_result::failure_at (vect_location, + "not vectorized: latch block not empty.\n"); /* Make sure the exit is not abnormal. */ edge e = single_exit (loop); if (e->flags & EDGE_ABNORMAL) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: abnormal loop exit edge.\n"); - return false; - } + return opt_result::failure_at (vect_location, + "not vectorized:" + " abnormal loop exit edge.\n"); *loop_cond = vect_get_loop_niters (loop, assumptions, number_of_iterations, number_of_iterationsm1); if (!*loop_cond) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: complicated exit condition.\n"); - return false; - } + return opt_result::failure_at + (vect_location, + "not vectorized: complicated exit condition.\n"); if (integer_zerop (*assumptions) || !*number_of_iterations || chrec_contains_undetermined (*number_of_iterations)) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: number of iterations cannot be " - "computed.\n"); - return false; - } + return opt_result::failure_at + (*loop_cond, + "not vectorized: number of iterations cannot be computed.\n"); if (integer_zerop (*number_of_iterations)) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: number of iterations = 0.\n"); - return false; - } - - return true; + return opt_result::failure_at + (*loop_cond, + "not vectorized: number of iterations = 0.\n"); + + return opt_result::success (); } /* Analyze LOOP form and return a loop_vec_info if it is of suitable form. */ -loop_vec_info -vect_analyze_loop_form (struct loop *loop) +opt_loop_vec_info +vect_analyze_loop_form (struct loop *loop, vec_info_shared *shared) { tree assumptions, number_of_iterations, number_of_iterationsm1; gcond *loop_cond, *inner_loop_cond = NULL; - if (! vect_analyze_loop_form_1 (loop, &loop_cond, - &assumptions, &number_of_iterationsm1, - &number_of_iterations, &inner_loop_cond)) - return NULL; - - loop_vec_info loop_vinfo = new _loop_vec_info (loop); + opt_result res + = vect_analyze_loop_form_1 (loop, &loop_cond, + &assumptions, &number_of_iterationsm1, + &number_of_iterations, &inner_loop_cond); + if (!res) + return opt_loop_vec_info::propagate_failure (res); + + loop_vec_info loop_vinfo = new _loop_vec_info (loop, shared); LOOP_VINFO_NITERSM1 (loop_vinfo) = number_of_iterationsm1; LOOP_VINFO_NITERS (loop_vinfo) = number_of_iterations; LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo) = number_of_iterations; @@ -1539,14 +1342,18 @@ } } - STMT_VINFO_TYPE (vinfo_for_stmt (loop_cond)) = loop_exit_ctrl_vec_info_type; + stmt_vec_info loop_cond_info = loop_vinfo->lookup_stmt (loop_cond); + STMT_VINFO_TYPE (loop_cond_info) = loop_exit_ctrl_vec_info_type; if (inner_loop_cond) - STMT_VINFO_TYPE (vinfo_for_stmt (inner_loop_cond)) - = loop_exit_ctrl_vec_info_type; + { + stmt_vec_info inner_loop_cond_info + = loop_vinfo->lookup_stmt (inner_loop_cond); + STMT_VINFO_TYPE (inner_loop_cond_info) = loop_exit_ctrl_vec_info_type; + } gcc_assert (!loop->aux); loop->aux = loop_vinfo; - return loop_vinfo; + return opt_loop_vec_info::success (loop_vinfo); } @@ -1560,15 +1367,13 @@ struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); int nbbs = loop->num_nodes; - unsigned int vectorization_factor; + poly_uint64 vectorization_factor; int i; - if (dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, - "=== vect_update_vf_for_slp ===\n"); + DUMP_VECT_SCOPE ("vect_update_vf_for_slp"); vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo); - gcc_assert (vectorization_factor != 0); + gcc_assert (known_ne (vectorization_factor, 0U)); /* If all the stmts in the loop can be SLPed, we perform only SLP, and vectorization factor of the loop is the unrolling factor required by @@ -1582,14 +1387,8 @@ for (gimple_stmt_iterator si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) { - gimple *stmt = gsi_stmt (si); - stmt_vec_info stmt_info = vinfo_for_stmt (stmt); - if (STMT_VINFO_IN_PATTERN_P (stmt_info) - && STMT_VINFO_RELATED_STMT (stmt_info)) - { - stmt = STMT_VINFO_RELATED_STMT (stmt_info); - stmt_info = vinfo_for_stmt (stmt); - } + stmt_vec_info stmt_info = loop_vinfo->lookup_stmt (gsi_stmt (si)); + stmt_info = vect_stmt_to_vectorize (stmt_info); if ((STMT_VINFO_RELEVANT_P (stmt_info) || VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (stmt_info))) && !PURE_SLP_STMT (stmt_info)) @@ -1608,23 +1407,55 @@ { dump_printf_loc (MSG_NOTE, vect_location, "Loop contains SLP and non-SLP stmts\n"); + /* Both the vectorization factor and unroll factor have the form + current_vector_size * X for some rational X, so they must have + a common multiple. */ vectorization_factor - = least_common_multiple (vectorization_factor, + = force_common_multiple (vectorization_factor, LOOP_VINFO_SLP_UNROLLING_FACTOR (loop_vinfo)); } LOOP_VINFO_VECT_FACTOR (loop_vinfo) = vectorization_factor; if (dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, - "Updating vectorization factor to %d\n", - vectorization_factor); + { + dump_printf_loc (MSG_NOTE, vect_location, + "Updating vectorization factor to "); + dump_dec (MSG_NOTE, vectorization_factor); + dump_printf (MSG_NOTE, ".\n"); + } +} + +/* Return true if STMT_INFO describes a double reduction phi and if + the other phi in the reduction is also relevant for vectorization. + This rejects cases such as: + + outer1: + x_1 = PHI <x_3(outer2), ...>; + ... + + inner: + x_2 = ...; + ... + + outer2: + x_3 = PHI <x_2(inner)>; + + if nothing in x_2 or elsewhere makes x_1 relevant. */ + +static bool +vect_active_double_reduction_p (stmt_vec_info stmt_info) +{ + if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_double_reduction_def) + return false; + + return STMT_VINFO_RELEVANT_P (STMT_VINFO_REDUC_DEF (stmt_info)); } /* Function vect_analyze_loop_operations. Scan the loop stmts and make sure they are all vectorizable. */ -static bool +static opt_result vect_analyze_loop_operations (loop_vec_info loop_vinfo) { struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); @@ -1635,9 +1466,10 @@ bool need_to_vectorize = false; bool ok; - if (dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, - "=== vect_analyze_loop_operations ===\n"); + DUMP_VECT_SCOPE ("vect_analyze_loop_operations"); + + stmt_vector_for_cost cost_vec; + cost_vec.create (2); for (i = 0; i < nbbs; i++) { @@ -1649,12 +1481,9 @@ gphi *phi = si.phi (); ok = true; - stmt_info = vinfo_for_stmt (phi); + stmt_info = loop_vinfo->lookup_stmt (phi); if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, "examining phi: "); - dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); - } + dump_printf_loc (MSG_NOTE, vect_location, "examining phi: %G", phi); if (virtual_operand_p (gimple_phi_result (phi))) continue; @@ -1667,41 +1496,29 @@ i.e., this phi is vect_reduction_def), cause this case requires to actually do something here. */ if (STMT_VINFO_LIVE_P (stmt_info) - && STMT_VINFO_DEF_TYPE (stmt_info) - != vect_double_reduction_def) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "Unsupported loop-closed phi in " - "outer-loop.\n"); - return false; - } + && !vect_active_double_reduction_p (stmt_info)) + return opt_result::failure_at (phi, + "Unsupported loop-closed phi" + " in outer-loop.\n"); /* If PHI is used in the outer loop, we check that its operand is defined in the inner loop. */ if (STMT_VINFO_RELEVANT_P (stmt_info)) { tree phi_op; - gimple *op_def_stmt; if (gimple_phi_num_args (phi) != 1) - return false; + return opt_result::failure_at (phi, "unsupported phi"); phi_op = PHI_ARG_DEF (phi, 0); - if (TREE_CODE (phi_op) != SSA_NAME) - return false; - - op_def_stmt = SSA_NAME_DEF_STMT (phi_op); - if (gimple_nop_p (op_def_stmt) - || !flow_bb_inside_loop_p (loop, gimple_bb (op_def_stmt)) - || !vinfo_for_stmt (op_def_stmt)) - return false; - - if (STMT_VINFO_RELEVANT (vinfo_for_stmt (op_def_stmt)) - != vect_used_in_outer - && STMT_VINFO_RELEVANT (vinfo_for_stmt (op_def_stmt)) - != vect_used_in_outer_by_reduction) - return false; + stmt_vec_info op_def_info = loop_vinfo->lookup_def (phi_op); + if (!op_def_info) + return opt_result::failure_at (phi, "unsupported phi"); + + if (STMT_VINFO_RELEVANT (op_def_info) != vect_used_in_outer + && (STMT_VINFO_RELEVANT (op_def_info) + != vect_used_in_outer_by_reduction)) + return opt_result::failure_at (phi, "unsupported phi"); } continue; @@ -1712,52 +1529,58 @@ if ((STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_scope || STMT_VINFO_LIVE_P (stmt_info)) && STMT_VINFO_DEF_TYPE (stmt_info) != vect_induction_def) - { - /* A scalar-dependence cycle that we don't support. */ - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: scalar dependence cycle.\n"); - return false; - } + /* A scalar-dependence cycle that we don't support. */ + return opt_result::failure_at (phi, + "not vectorized:" + " scalar dependence cycle.\n"); if (STMT_VINFO_RELEVANT_P (stmt_info)) { need_to_vectorize = true; if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_induction_def && ! PURE_SLP_STMT (stmt_info)) - ok = vectorizable_induction (phi, NULL, NULL, NULL); + ok = vectorizable_induction (stmt_info, NULL, NULL, NULL, + &cost_vec); else if ((STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def || STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle) && ! PURE_SLP_STMT (stmt_info)) - ok = vectorizable_reduction (phi, NULL, NULL, NULL, NULL); + ok = vectorizable_reduction (stmt_info, NULL, NULL, NULL, NULL, + &cost_vec); } - if (ok && STMT_VINFO_LIVE_P (stmt_info)) - ok = vectorizable_live_operation (phi, NULL, NULL, -1, NULL); + /* SLP PHIs are tested by vect_slp_analyze_node_operations. */ + if (ok + && STMT_VINFO_LIVE_P (stmt_info) + && !PURE_SLP_STMT (stmt_info)) + ok = vectorizable_live_operation (stmt_info, NULL, NULL, -1, NULL, + &cost_vec); if (!ok) - { - if (dump_enabled_p ()) - { - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: relevant phi not " - "supported: "); - dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, phi, 0); - } - return false; - } + return opt_result::failure_at (phi, + "not vectorized: relevant phi not " + "supported: %G", + static_cast <gimple *> (phi)); } for (gimple_stmt_iterator si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) { gimple *stmt = gsi_stmt (si); - if (!gimple_clobber_p (stmt) - && !vect_analyze_stmt (stmt, &need_to_vectorize, NULL, NULL)) - return false; + if (!gimple_clobber_p (stmt)) + { + opt_result res + = vect_analyze_stmt (loop_vinfo->lookup_stmt (stmt), + &need_to_vectorize, + NULL, NULL, &cost_vec); + if (!res) + return res; + } } } /* bbs */ + add_stmt_costs (loop_vinfo->target_cost_data, &cost_vec); + cost_vec.release (); + /* All operations in the loop are either irrelevant (deal with loop control, or dead), or only used outside the loop and can be moved out of the loop (e.g. invariants, inductions). The loop can be @@ -1768,49 +1591,115 @@ if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "All the computation can be taken out of the loop.\n"); - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: redundant loop. no profit to " - "vectorize.\n"); - return false; - } - - return true; + return opt_result::failure_at + (vect_location, + "not vectorized: redundant loop. no profit to vectorize.\n"); + } + + return opt_result::success (); } - -/* Function vect_analyze_loop_2. - - Apply a set of analyses on LOOP, and create a loop_vec_info struct - for it. The different analyses will record information in the - loop_vec_info struct. */ -static bool -vect_analyze_loop_2 (loop_vec_info loop_vinfo, bool &fatal) +/* Analyze the cost of the loop described by LOOP_VINFO. Decide if it + is worthwhile to vectorize. Return 1 if definitely yes, 0 if + definitely no, or -1 if it's worth retrying. */ + +static int +vect_analyze_loop_costing (loop_vec_info loop_vinfo) { - bool ok; - int max_vf = MAX_VECTORIZATION_FACTOR; - int min_vf = 2; - unsigned int n_stmts = 0; - - /* The first group of checks is independent of the vector size. */ - fatal = true; - - /* Find all data references in the loop (which correspond to vdefs/vuses) - and analyze their evolution in the loop. */ - - basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); - - loop_p loop = LOOP_VINFO_LOOP (loop_vinfo); - if (!find_loop_nest (loop, &LOOP_VINFO_LOOP_NEST (loop_vinfo))) + struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); + unsigned int assumed_vf = vect_vf_for_cost (loop_vinfo); + + /* Only fully-masked loops can have iteration counts less than the + vectorization factor. */ + if (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) + { + HOST_WIDE_INT max_niter; + + if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)) + max_niter = LOOP_VINFO_INT_NITERS (loop_vinfo); + else + max_niter = max_stmt_executions_int (loop); + + if (max_niter != -1 + && (unsigned HOST_WIDE_INT) max_niter < assumed_vf) + { + if (dump_enabled_p ()) + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "not vectorized: iteration count smaller than " + "vectorization factor.\n"); + return 0; + } + } + + int min_profitable_iters, min_profitable_estimate; + vect_estimate_min_profitable_iters (loop_vinfo, &min_profitable_iters, + &min_profitable_estimate); + + if (min_profitable_iters < 0) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: loop nest containing two " - "or more consecutive inner loops cannot be " - "vectorized\n"); - return false; - } - + "not vectorized: vectorization not profitable.\n"); + if (dump_enabled_p ()) + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "not vectorized: vector version will never be " + "profitable.\n"); + return -1; + } + + int min_scalar_loop_bound = (PARAM_VALUE (PARAM_MIN_VECT_LOOP_BOUND) + * assumed_vf); + + /* Use the cost model only if it is more conservative than user specified + threshold. */ + unsigned int th = (unsigned) MAX (min_scalar_loop_bound, + min_profitable_iters); + + LOOP_VINFO_COST_MODEL_THRESHOLD (loop_vinfo) = th; + + if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) + && LOOP_VINFO_INT_NITERS (loop_vinfo) < th) + { + if (dump_enabled_p ()) + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "not vectorized: vectorization not profitable.\n"); + if (dump_enabled_p ()) + dump_printf_loc (MSG_NOTE, vect_location, + "not vectorized: iteration count smaller than user " + "specified loop bound parameter or minimum profitable " + "iterations (whichever is more conservative).\n"); + return 0; + } + + HOST_WIDE_INT estimated_niter = estimated_stmt_executions_int (loop); + if (estimated_niter == -1) + estimated_niter = likely_max_stmt_executions_int (loop); + if (estimated_niter != -1 + && ((unsigned HOST_WIDE_INT) estimated_niter + < MAX (th, (unsigned) min_profitable_estimate))) + { + if (dump_enabled_p ()) + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "not vectorized: estimated iteration count too " + "small.\n"); + if (dump_enabled_p ()) + dump_printf_loc (MSG_NOTE, vect_location, + "not vectorized: estimated iteration count smaller " + "than specified loop bound parameter or minimum " + "profitable iterations (whichever is more " + "conservative).\n"); + return -1; + } + + return 1; +} + +static opt_result +vect_get_datarefs_in_loop (loop_p loop, basic_block *bbs, + vec<data_reference_p> *datarefs, + unsigned int *n_stmts) +{ + *n_stmts = 0; for (unsigned i = 0; i < loop->num_nodes; i++) for (gimple_stmt_iterator gsi = gsi_start_bb (bbs[i]); !gsi_end_p (gsi); gsi_next (&gsi)) @@ -1818,9 +1707,9 @@ gimple *stmt = gsi_stmt (gsi); if (is_gimple_debug (stmt)) continue; - ++n_stmts; - if (!find_data_references_in_stmt (loop, stmt, - &LOOP_VINFO_DATAREFS (loop_vinfo))) + ++(*n_stmts); + opt_result res = vect_find_stmt_data_reference (loop, stmt, datarefs); + if (!res) { if (is_gimple_call (stmt) && loop->safelen) { @@ -1852,14 +1741,59 @@ } } } - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: loop contains function " - "calls or data references that cannot " - "be analyzed\n"); - return false; + return res; } + /* If dependence analysis will give up due to the limit on the + number of datarefs stop here and fail fatally. */ + if (datarefs->length () + > (unsigned)PARAM_VALUE (PARAM_LOOP_MAX_DATAREFS_FOR_DATADEPS)) + return opt_result::failure_at (stmt, "exceeded param " + "loop-max-datarefs-for-datadeps\n"); } + return opt_result::success (); +} + +/* Function vect_analyze_loop_2. + + Apply a set of analyses on LOOP, and create a loop_vec_info struct + for it. The different analyses will record information in the + loop_vec_info struct. */ +static opt_result +vect_analyze_loop_2 (loop_vec_info loop_vinfo, bool &fatal, unsigned *n_stmts) +{ + opt_result ok = opt_result::success (); + int res; + unsigned int max_vf = MAX_VECTORIZATION_FACTOR; + poly_uint64 min_vf = 2; + + /* The first group of checks is independent of the vector size. */ + fatal = true; + + /* Find all data references in the loop (which correspond to vdefs/vuses) + and analyze their evolution in the loop. */ + + loop_p loop = LOOP_VINFO_LOOP (loop_vinfo); + + /* Gather the data references and count stmts in the loop. */ + if (!LOOP_VINFO_DATAREFS (loop_vinfo).exists ()) + { + opt_result res + = vect_get_datarefs_in_loop (loop, LOOP_VINFO_BBS (loop_vinfo), + &LOOP_VINFO_DATAREFS (loop_vinfo), + n_stmts); + if (!res) + { + if (dump_enabled_p ()) + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "not vectorized: loop contains function " + "calls or data references that cannot " + "be analyzed\n"); + return res; + } + loop_vinfo->shared->save_datarefs (); + } + else + loop_vinfo->shared->check_datarefs (); /* Analyze the data references and also adjust the minimal vectorization factor according to the loads and stores. */ @@ -1870,7 +1804,7 @@ if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "bad data references.\n"); - return false; + return ok; } /* Classify all cross-iteration scalar data-flow cycles. @@ -1890,7 +1824,7 @@ if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "bad data access.\n"); - return false; + return ok; } /* Data-flow analysis to detect stmts that do not need to be vectorized. */ @@ -1901,7 +1835,7 @@ if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "unexpected pattern.\n"); - return false; + return ok; } /* While the rest of the analysis below depends on it in some way. */ @@ -1913,14 +1847,16 @@ FORNOW: fail at the first data dependence that we encounter. */ ok = vect_analyze_data_ref_dependences (loop_vinfo, &max_vf); - if (!ok - || max_vf < min_vf) + if (!ok) { if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "bad data dependence.\n"); - return false; - } + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "bad data dependence.\n"); + return ok; + } + if (max_vf != MAX_VECTORIZATION_FACTOR + && maybe_lt (max_vf, min_vf)) + return opt_result::failure_at (vect_location, "bad data dependence.\n"); LOOP_VINFO_MAX_VECT_FACTOR (loop_vinfo) = max_vf; ok = vect_determine_vectorization_factor (loop_vinfo); @@ -1929,28 +1865,22 @@ if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "can't determine vectorization factor.\n"); - return false; - } - if (max_vf < LOOP_VINFO_VECT_FACTOR (loop_vinfo)) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "bad data dependence.\n"); - return false; - } + return ok; + } + if (max_vf != MAX_VECTORIZATION_FACTOR + && maybe_lt (max_vf, LOOP_VINFO_VECT_FACTOR (loop_vinfo))) + return opt_result::failure_at (vect_location, "bad data dependence.\n"); /* Compute the scalar iteration cost. */ vect_compute_single_scalar_iteration_cost (loop_vinfo); - int saved_vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo); - HOST_WIDE_INT estimated_niter; + poly_uint64 saved_vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo); unsigned th; - int min_scalar_loop_bound; /* Check the SLP opportunities in the loop, analyze and build SLP trees. */ - ok = vect_analyze_slp (loop_vinfo, n_stmts); + ok = vect_analyze_slp (loop_vinfo, *n_stmts); if (!ok) - return false; + return ok; /* If there are any SLP instances mark them as pure_slp. */ bool slp = vect_make_slp_decision (loop_vinfo); @@ -1963,32 +1893,30 @@ vect_update_vf_for_slp (loop_vinfo); } + bool saved_can_fully_mask_p = LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo); + + /* We don't expect to have to roll back to anything other than an empty + set of rgroups. */ + gcc_assert (LOOP_VINFO_MASKS (loop_vinfo).is_empty ()); + /* This is the point where we can re-start analysis with SLP forced off. */ start_over: /* Now the vectorization factor is final. */ - unsigned vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo); - gcc_assert (vectorization_factor != 0); + poly_uint64 vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo); + gcc_assert (known_ne (vectorization_factor, 0U)); if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) && dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, - "vectorization_factor = %d, niters = " - HOST_WIDE_INT_PRINT_DEC "\n", vectorization_factor, - LOOP_VINFO_INT_NITERS (loop_vinfo)); + { + dump_printf_loc (MSG_NOTE, vect_location, + "vectorization_factor = "); + dump_dec (MSG_NOTE, vectorization_factor); + dump_printf (MSG_NOTE, ", niters = %wd\n", + LOOP_VINFO_INT_NITERS (loop_vinfo)); + } HOST_WIDE_INT max_niter = likely_max_stmt_executions_int (LOOP_VINFO_LOOP (loop_vinfo)); - if ((LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) - && (LOOP_VINFO_INT_NITERS (loop_vinfo) < vectorization_factor)) - || (max_niter != -1 - && (unsigned HOST_WIDE_INT) max_niter < vectorization_factor)) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: iteration count smaller than " - "vectorization factor.\n"); - return false; - } /* Analyze the alignment of the data-refs in the loop. Fail if a data reference is found that cannot be vectorized. */ @@ -1999,7 +1927,7 @@ if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "bad data alignment.\n"); - return false; + return ok; } /* Prune the list of ddrs to be tested at run-time by versioning for alias. @@ -2007,23 +1935,19 @@ since we use grouping information gathered by interleaving analysis. */ ok = vect_prune_runtime_alias_test_list (loop_vinfo); if (!ok) - return false; - - /* Do not invoke vect_enhance_data_refs_alignment for eplilogue - vectorization. */ + return ok; + + /* Do not invoke vect_enhance_data_refs_alignment for epilogue + vectorization, since we do not want to add extra peeling or + add versioning for alignment. */ if (!LOOP_VINFO_EPILOGUE_P (loop_vinfo)) - { /* This pass will decide on using loop versioning and/or loop peeling in order to enhance the alignment of data references in the loop. */ ok = vect_enhance_data_refs_alignment (loop_vinfo); - if (!ok) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "bad data alignment.\n"); - return false; - } - } + else + ok = vect_verify_datarefs_alignment (loop_vinfo); + if (!ok) + return ok; if (slp) { @@ -2033,7 +1957,11 @@ unsigned old_size = LOOP_VINFO_SLP_INSTANCES (loop_vinfo).length (); vect_slp_analyze_operations (loop_vinfo); if (LOOP_VINFO_SLP_INSTANCES (loop_vinfo).length () != old_size) - goto again; + { + ok = opt_result::failure_at (vect_location, + "unsupported SLP instances\n"); + goto again; + } } /* Scan all the remaining operations in the loop that are not subject @@ -2044,111 +1972,86 @@ if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "bad operation or unsupported loop bound.\n"); - return false; + return ok; + } + + /* Decide whether to use a fully-masked loop for this vectorization + factor. */ + LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) + = (LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) + && vect_verify_full_masking (loop_vinfo)); + if (dump_enabled_p ()) + { + if (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) + dump_printf_loc (MSG_NOTE, vect_location, + "using a fully-masked loop.\n"); + else + dump_printf_loc (MSG_NOTE, vect_location, + "not using a fully-masked loop.\n"); } /* If epilog loop is required because of data accesses with gaps, one additional iteration needs to be peeled. Check if there is enough iterations for vectorization. */ if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) - && LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)) - { - int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); + && LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) + && !LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) + { + poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); tree scalar_niters = LOOP_VINFO_NITERSM1 (loop_vinfo); - if (wi::to_widest (scalar_niters) < vf) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, - "loop has no enough iterations to support" - " peeling for gaps.\n"); - return false; - } - } - - /* Analyze cost. Decide if worth while to vectorize. */ - int min_profitable_estimate, min_profitable_iters; - vect_estimate_min_profitable_iters (loop_vinfo, &min_profitable_iters, - &min_profitable_estimate); - - if (min_profitable_iters < 0) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: vectorization not profitable.\n"); - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: vector version will never be " - "profitable.\n"); + if (known_lt (wi::to_widest (scalar_niters), vf)) + return opt_result::failure_at (vect_location, + "loop has no enough iterations to" + " support peeling for gaps.\n"); + } + + /* Check the costings of the loop make vectorizing worthwhile. */ + res = vect_analyze_loop_costing (loop_vinfo); + if (res < 0) + { + ok = opt_result::failure_at (vect_location, + "Loop costings may not be worthwhile.\n"); goto again; } - - min_scalar_loop_bound = (PARAM_VALUE (PARAM_MIN_VECT_LOOP_BOUND) - * vectorization_factor); - - /* Use the cost model only if it is more conservative than user specified - threshold. */ - th = (unsigned) MAX (min_scalar_loop_bound, min_profitable_iters); - - LOOP_VINFO_COST_MODEL_THRESHOLD (loop_vinfo) = th; - - if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) - && LOOP_VINFO_INT_NITERS (loop_vinfo) < th) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: vectorization not profitable.\n"); - if (dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, - "not vectorized: iteration count smaller than user " - "specified loop bound parameter or minimum profitable " - "iterations (whichever is more conservative).\n"); - goto again; - } - - estimated_niter - = estimated_stmt_executions_int (LOOP_VINFO_LOOP (loop_vinfo)); - if (estimated_niter == -1) - estimated_niter = max_niter; - if (estimated_niter != -1 - && ((unsigned HOST_WIDE_INT) estimated_niter - < MAX (th, (unsigned) min_profitable_estimate))) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: estimated iteration count too " - "small.\n"); - if (dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, - "not vectorized: estimated iteration count smaller " - "than specified loop bound parameter or minimum " - "profitable iterations (whichever is more " - "conservative).\n"); - goto again; - } + if (!res) + return opt_result::failure_at (vect_location, + "Loop costings not worthwhile.\n"); /* Decide whether we need to create an epilogue loop to handle remaining scalar iterations. */ - th = ((LOOP_VINFO_COST_MODEL_THRESHOLD (loop_vinfo) - / LOOP_VINFO_VECT_FACTOR (loop_vinfo)) - * LOOP_VINFO_VECT_FACTOR (loop_vinfo)); - - if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) - && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) > 0) - { - if (ctz_hwi (LOOP_VINFO_INT_NITERS (loop_vinfo) - - LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo)) - < exact_log2 (LOOP_VINFO_VECT_FACTOR (loop_vinfo))) + th = LOOP_VINFO_COST_MODEL_THRESHOLD (loop_vinfo); + + unsigned HOST_WIDE_INT const_vf; + if (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) + /* The main loop handles all iterations. */ + LOOP_VINFO_PEELING_FOR_NITER (loop_vinfo) = false; + else if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) + && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) >= 0) + { + /* Work out the (constant) number of iterations that need to be + peeled for reasons other than niters. */ + unsigned int peel_niter = LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo); + if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo)) + peel_niter += 1; + if (!multiple_p (LOOP_VINFO_INT_NITERS (loop_vinfo) - peel_niter, + LOOP_VINFO_VECT_FACTOR (loop_vinfo))) LOOP_VINFO_PEELING_FOR_NITER (loop_vinfo) = true; } else if (LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) - || (tree_ctz (LOOP_VINFO_NITERS (loop_vinfo)) - < (unsigned)exact_log2 (LOOP_VINFO_VECT_FACTOR (loop_vinfo)) - /* In case of versioning, check if the maximum number of - iterations is greater than th. If they are identical, - the epilogue is unnecessary. */ + /* ??? When peeling for gaps but not alignment, we could + try to check whether the (variable) niters is known to be + VF * N + 1. That's something of a niche case though. */ + || LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) + || !LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (&const_vf) + || ((tree_ctz (LOOP_VINFO_NITERS (loop_vinfo)) + < (unsigned) exact_log2 (const_vf)) + /* In case of versioning, check if the maximum number of + iterations is greater than th. If they are identical, + the epilogue is unnecessary. */ && (!LOOP_REQUIRES_VERSIONING (loop_vinfo) - || (unsigned HOST_WIDE_INT) max_niter > th))) + || ((unsigned HOST_WIDE_INT) max_niter + > (th / const_vf) * const_vf)))) LOOP_VINFO_PEELING_FOR_NITER (loop_vinfo) = true; /* If an epilogue loop is required make sure we can create one. */ @@ -2162,10 +2065,9 @@ single_exit (LOOP_VINFO_LOOP (loop_vinfo)))) { - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "not vectorized: can't create required " - "epilog loop\n"); + ok = opt_result::failure_at (vect_location, + "not vectorized: can't create required " + "epilog loop\n"); goto again; } } @@ -2174,47 +2076,50 @@ enough for both peeled prolog loop and vector loop. This check can be merged along with threshold check of loop versioning, so increase threshold for this case if necessary. */ - if (LOOP_REQUIRES_VERSIONING (loop_vinfo) - && (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) - || LOOP_VINFO_PEELING_FOR_NITER (loop_vinfo))) - { - unsigned niters_th; - - /* Niters for peeled prolog loop. */ - if (LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) < 0) - { - struct data_reference *dr = LOOP_VINFO_UNALIGNED_DR (loop_vinfo); - tree vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (DR_STMT (dr))); - - niters_th = TYPE_VECTOR_SUBPARTS (vectype) - 1; + if (LOOP_REQUIRES_VERSIONING (loop_vinfo)) + { + poly_uint64 niters_th = 0; + + if (!vect_use_loop_mask_for_alignment_p (loop_vinfo)) + { + /* Niters for peeled prolog loop. */ + if (LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) < 0) + { + dr_vec_info *dr_info = LOOP_VINFO_UNALIGNED_DR (loop_vinfo); + tree vectype = STMT_VINFO_VECTYPE (dr_info->stmt); + niters_th += TYPE_VECTOR_SUBPARTS (vectype) - 1; + } + else + niters_th += LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo); } - else - niters_th = LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo); /* Niters for at least one iteration of vectorized loop. */ - niters_th += LOOP_VINFO_VECT_FACTOR (loop_vinfo); + if (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) + niters_th += LOOP_VINFO_VECT_FACTOR (loop_vinfo); /* One additional iteration because of peeling for gap. */ if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo)) - niters_th++; - if (LOOP_VINFO_COST_MODEL_THRESHOLD (loop_vinfo) < niters_th) - LOOP_VINFO_COST_MODEL_THRESHOLD (loop_vinfo) = niters_th; - } - - gcc_assert (vectorization_factor - == (unsigned)LOOP_VINFO_VECT_FACTOR (loop_vinfo)); + niters_th += 1; + LOOP_VINFO_VERSIONING_THRESHOLD (loop_vinfo) = niters_th; + } + + gcc_assert (known_eq (vectorization_factor, + LOOP_VINFO_VECT_FACTOR (loop_vinfo))); /* Ok to vectorize! */ - return true; + return opt_result::success (); again: + /* Ensure that "ok" is false (with an opt_problem if dumping is enabled). */ + gcc_assert (!ok); + /* Try again with SLP forced off but if we didn't do any SLP there is no point in re-trying. */ if (!slp) - return false; + return ok; /* If there are reduction chains re-trying will fail anyway. */ if (! LOOP_VINFO_REDUCTION_CHAINS (loop_vinfo).is_empty ()) - return false; + return ok; /* Likewise if the grouped loads or stores in the SLP cannot be handled via interleaving or lane instructions. */ @@ -2224,27 +2129,29 @@ FOR_EACH_VEC_ELT (LOOP_VINFO_SLP_INSTANCES (loop_vinfo), i, instance) { stmt_vec_info vinfo; - vinfo = vinfo_for_stmt - (SLP_TREE_SCALAR_STMTS (SLP_INSTANCE_TREE (instance))[0]); + vinfo = SLP_TREE_SCALAR_STMTS (SLP_INSTANCE_TREE (instance))[0]; if (! STMT_VINFO_GROUPED_ACCESS (vinfo)) continue; - vinfo = vinfo_for_stmt (STMT_VINFO_GROUP_FIRST_ELEMENT (vinfo)); - unsigned int size = STMT_VINFO_GROUP_SIZE (vinfo); + vinfo = DR_GROUP_FIRST_ELEMENT (vinfo); + unsigned int size = DR_GROUP_SIZE (vinfo); tree vectype = STMT_VINFO_VECTYPE (vinfo); - if (! vect_store_lanes_supported (vectype, size) - && ! vect_grouped_store_supported (vectype, size)) - return false; + if (! vect_store_lanes_supported (vectype, size, false) + && ! known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U) + && ! vect_grouped_store_supported (vectype, size)) + return opt_result::failure_at (vinfo->stmt, + "unsupported grouped store\n"); FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (instance), j, node) { - vinfo = vinfo_for_stmt (SLP_TREE_SCALAR_STMTS (node)[0]); - vinfo = vinfo_for_stmt (STMT_VINFO_GROUP_FIRST_ELEMENT (vinfo)); - bool single_element_p = !STMT_VINFO_GROUP_NEXT_ELEMENT (vinfo); - size = STMT_VINFO_GROUP_SIZE (vinfo); + vinfo = SLP_TREE_SCALAR_STMTS (node)[0]; + vinfo = DR_GROUP_FIRST_ELEMENT (vinfo); + bool single_element_p = !DR_GROUP_NEXT_ELEMENT (vinfo); + size = DR_GROUP_SIZE (vinfo); vectype = STMT_VINFO_VECTYPE (vinfo); - if (! vect_load_lanes_supported (vectype, size) + if (! vect_load_lanes_supported (vectype, size, false) && ! vect_grouped_load_supported (vectype, single_element_p, size)) - return false; + return opt_result::failure_at (vinfo->stmt, + "unsupported grouped load\n"); } } @@ -2258,7 +2165,7 @@ LOOP_VINFO_VECT_FACTOR (loop_vinfo) = saved_vectorization_factor; /* Free the SLP instances. */ FOR_EACH_VEC_ELT (LOOP_VINFO_SLP_INSTANCES (loop_vinfo), j, instance) - vect_free_slp_instance (instance); + vect_free_slp_instance (instance, false); LOOP_VINFO_SLP_INSTANCES (loop_vinfo).release (); /* Reset SLP type to loop_vect on all stmts. */ for (i = 0; i < LOOP_VINFO_LOOP (loop_vinfo)->num_nodes; ++i) @@ -2267,39 +2174,42 @@ for (gimple_stmt_iterator si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) { - stmt_vec_info stmt_info = vinfo_for_stmt (gsi_stmt (si)); + stmt_vec_info stmt_info = loop_vinfo->lookup_stmt (gsi_stmt (si)); STMT_SLP_TYPE (stmt_info) = loop_vect; } for (gimple_stmt_iterator si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) { - stmt_vec_info stmt_info = vinfo_for_stmt (gsi_stmt (si)); + stmt_vec_info stmt_info = loop_vinfo->lookup_stmt (gsi_stmt (si)); STMT_SLP_TYPE (stmt_info) = loop_vect; if (STMT_VINFO_IN_PATTERN_P (stmt_info)) { - stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)); + gimple *pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info); + stmt_info = STMT_VINFO_RELATED_STMT (stmt_info); STMT_SLP_TYPE (stmt_info) = loop_vect; - for (gimple_stmt_iterator pi - = gsi_start (STMT_VINFO_PATTERN_DEF_SEQ (stmt_info)); + for (gimple_stmt_iterator pi = gsi_start (pattern_def_seq); !gsi_end_p (pi); gsi_next (&pi)) - { - gimple *pstmt = gsi_stmt (pi); - STMT_SLP_TYPE (vinfo_for_stmt (pstmt)) = loop_vect; - } + STMT_SLP_TYPE (loop_vinfo->lookup_stmt (gsi_stmt (pi))) + = loop_vect; } } } /* Free optimized alias test DDRS. */ + LOOP_VINFO_LOWER_BOUNDS (loop_vinfo).truncate (0); LOOP_VINFO_COMP_ALIAS_DDRS (loop_vinfo).release (); LOOP_VINFO_CHECK_UNEQUAL_ADDRS (loop_vinfo).release (); /* Reset target cost data. */ destroy_cost_data (LOOP_VINFO_TARGET_COST_DATA (loop_vinfo)); LOOP_VINFO_TARGET_COST_DATA (loop_vinfo) = init_cost (LOOP_VINFO_LOOP (loop_vinfo)); + /* Reset accumulated rgroup information. */ + release_vec_loop_masks (&LOOP_VINFO_MASKS (loop_vinfo)); /* Reset assorted flags. */ LOOP_VINFO_PEELING_FOR_NITER (loop_vinfo) = false; LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = false; LOOP_VINFO_COST_MODEL_THRESHOLD (loop_vinfo) = 0; + LOOP_VINFO_VERSIONING_THRESHOLD (loop_vinfo) = 0; + LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = saved_can_fully_mask_p; goto start_over; } @@ -2310,40 +2220,44 @@ for it. The different analyses will record information in the loop_vec_info struct. If ORIG_LOOP_VINFO is not NULL epilogue must be vectorized. */ -loop_vec_info -vect_analyze_loop (struct loop *loop, loop_vec_info orig_loop_vinfo) +opt_loop_vec_info +vect_analyze_loop (struct loop *loop, loop_vec_info orig_loop_vinfo, + vec_info_shared *shared) { - loop_vec_info loop_vinfo; - unsigned int vector_sizes; + auto_vector_sizes vector_sizes; /* Autodetect first vector size we try. */ current_vector_size = 0; - vector_sizes = targetm.vectorize.autovectorize_vector_sizes (); - - if (dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, - "===== analyze_loop_nest =====\n"); + targetm.vectorize.autovectorize_vector_sizes (&vector_sizes); + unsigned int next_size = 0; + + DUMP_VECT_SCOPE ("analyze_loop_nest"); if (loop_outer (loop) && loop_vec_info_for_loop (loop_outer (loop)) && LOOP_VINFO_VECTORIZABLE_P (loop_vec_info_for_loop (loop_outer (loop)))) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, - "outer-loop already vectorized.\n"); - return NULL; - } - + return opt_loop_vec_info::failure_at (vect_location, + "outer-loop already vectorized.\n"); + + if (!find_loop_nest (loop, &shared->loop_nest)) + return opt_loop_vec_info::failure_at + (vect_location, + "not vectorized: loop nest containing two or more consecutive inner" + " loops cannot be vectorized\n"); + + unsigned n_stmts = 0; + poly_uint64 autodetected_vector_size = 0; while (1) { /* Check the CFG characteristics of the loop (nesting, entry/exit). */ - loop_vinfo = vect_analyze_loop_form (loop); + opt_loop_vec_info loop_vinfo + = vect_analyze_loop_form (loop, shared); if (!loop_vinfo) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "bad loop form.\n"); - return NULL; + return loop_vinfo; } bool fatal = false; @@ -2351,7 +2265,8 @@ if (orig_loop_vinfo) LOOP_VINFO_ORIG_LOOP_INFO (loop_vinfo) = orig_loop_vinfo; - if (vect_analyze_loop_2 (loop_vinfo, fatal)) + opt_result res = vect_analyze_loop_2 (loop_vinfo, fatal, &n_stmts); + if (res) { LOOP_VINFO_VECTORIZABLE_P (loop_vinfo) = 1; @@ -2360,59 +2275,93 @@ delete loop_vinfo; - vector_sizes &= ~current_vector_size; + if (next_size == 0) + autodetected_vector_size = current_vector_size; + + if (next_size < vector_sizes.length () + && known_eq (vector_sizes[next_size], autodetected_vector_size)) + next_size += 1; + if (fatal - || vector_sizes == 0 - || current_vector_size == 0) - return NULL; + || next_size == vector_sizes.length () + || known_eq (current_vector_size, 0U)) + return opt_loop_vec_info::propagate_failure (res); /* Try the next biggest vector size. */ - current_vector_size = 1 << floor_log2 (vector_sizes); + current_vector_size = vector_sizes[next_size++]; if (dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, - "***** Re-trying analysis with " - "vector size %d\n", current_vector_size); + { + dump_printf_loc (MSG_NOTE, vect_location, + "***** Re-trying analysis with " + "vector size "); + dump_dec (MSG_NOTE, current_vector_size); + dump_printf (MSG_NOTE, "\n"); + } } } - -/* Function reduction_code_for_scalar_code +/* Return true if there is an in-order reduction function for CODE, storing + it in *REDUC_FN if so. */ + +static bool +fold_left_reduction_fn (tree_code code, internal_fn *reduc_fn) +{ + switch (code) + { + case PLUS_EXPR: + *reduc_fn = IFN_FOLD_LEFT_PLUS; + return true; + + default: + return false; + } +} + +/* Function reduction_fn_for_scalar_code Input: CODE - tree_code of a reduction operations. Output: - REDUC_CODE - the corresponding tree-code to be used to reduce the - vector of partial results into a single scalar result, or ERROR_MARK + REDUC_FN - the corresponding internal function to be used to reduce the + vector of partial results into a single scalar result, or IFN_LAST if the operation is a supported reduction operation, but does not have - such a tree-code. + such an internal function. Return FALSE if CODE currently cannot be vectorized as reduction. */ static bool -reduction_code_for_scalar_code (enum tree_code code, - enum tree_code *reduc_code) +reduction_fn_for_scalar_code (enum tree_code code, internal_fn *reduc_fn) { switch (code) { case MAX_EXPR: - *reduc_code = REDUC_MAX_EXPR; + *reduc_fn = IFN_REDUC_MAX; return true; case MIN_EXPR: - *reduc_code = REDUC_MIN_EXPR; + *reduc_fn = IFN_REDUC_MIN; return true; case PLUS_EXPR: - *reduc_code = REDUC_PLUS_EXPR; + *reduc_fn = IFN_REDUC_PLUS; return true; + case BIT_AND_EXPR: + *reduc_fn = IFN_REDUC_AND; + return true; + + case BIT_IOR_EXPR: + *reduc_fn = IFN_REDUC_IOR; + return true; + + case BIT_XOR_EXPR: + *reduc_fn = IFN_REDUC_XOR; + return true; + case MULT_EXPR: case MINUS_EXPR: - case BIT_IOR_EXPR: - case BIT_XOR_EXPR: - case BIT_AND_EXPR: - *reduc_code = ERROR_MARK; + *reduc_fn = IFN_LAST; return true; default: @@ -2420,6 +2369,54 @@ } } +/* If there is a neutral value X such that SLP reduction NODE would not + be affected by the introduction of additional X elements, return that X, + otherwise return null. CODE is the code of the reduction. REDUC_CHAIN + is true if the SLP statements perform a single reduction, false if each + statement performs an independent reduction. */ + +static tree +neutral_op_for_slp_reduction (slp_tree slp_node, tree_code code, + bool reduc_chain) +{ + vec<stmt_vec_info> stmts = SLP_TREE_SCALAR_STMTS (slp_node); + stmt_vec_info stmt_vinfo = stmts[0]; + tree vector_type = STMT_VINFO_VECTYPE (stmt_vinfo); + tree scalar_type = TREE_TYPE (vector_type); + struct loop *loop = gimple_bb (stmt_vinfo->stmt)->loop_father; + gcc_assert (loop); + + switch (code) + { + case WIDEN_SUM_EXPR: + case DOT_PROD_EXPR: + case SAD_EXPR: + case PLUS_EXPR: + case MINUS_EXPR: + case BIT_IOR_EXPR: + case BIT_XOR_EXPR: + return build_zero_cst (scalar_type); + + case MULT_EXPR: + return build_one_cst (scalar_type); + + case BIT_AND_EXPR: + return build_all_ones_cst (scalar_type); + + case MAX_EXPR: + case MIN_EXPR: + /* For MIN/MAX the initial values are neutral. A reduction chain + has only a single initial value, so that value is neutral for + all statements. */ + if (reduc_chain) + return PHI_ARG_DEF_FROM_EDGE (stmt_vinfo->stmt, + loop_preheader_edge (loop)); + return NULL_TREE; + + default: + return NULL_TREE; + } +} /* Error reporting helper for vect_is_simple_reduction below. GIMPLE statement STMT is printed with a message MSG. */ @@ -2427,10 +2424,23 @@ static void report_vect_op (dump_flags_t msg_type, gimple *stmt, const char *msg) { - dump_printf_loc (msg_type, vect_location, "%s", msg); - dump_gimple_stmt (msg_type, TDF_SLIM, stmt, 0); + dump_printf_loc (msg_type, vect_location, "%s%G", msg, stmt); } +/* DEF_STMT_INFO occurs in a loop that contains a potential reduction + operation. Return true if the results of DEF_STMT_INFO are something + that can be accumulated by such a reduction. */ + +static bool +vect_valid_reduction_input_p (stmt_vec_info def_stmt_info) +{ + return (is_gimple_assign (def_stmt_info->stmt) + || is_gimple_call (def_stmt_info->stmt) + || STMT_VINFO_DEF_TYPE (def_stmt_info) == vect_induction_def + || (gimple_code (def_stmt_info->stmt) == GIMPLE_PHI + && STMT_VINFO_DEF_TYPE (def_stmt_info) == vect_internal_def + && !is_loop_header_bb_p (gimple_bb (def_stmt_info->stmt)))); +} /* Detect SLP reduction of the form: @@ -2455,8 +2465,8 @@ struct loop *loop = (gimple_bb (phi))->loop_father; struct loop *vect_loop = LOOP_VINFO_LOOP (loop_info); enum tree_code code; - gimple *current_stmt = NULL, *loop_use_stmt = NULL, *first, *next_stmt; - stmt_vec_info use_stmt_info, current_stmt_info; + gimple *loop_use_stmt = NULL; + stmt_vec_info use_stmt_info, current_stmt_info = NULL; tree lhs; imm_use_iterator imm_iter; use_operand_p use_p; @@ -2517,19 +2527,18 @@ return false; /* Insert USE_STMT into reduction chain. */ - use_stmt_info = vinfo_for_stmt (loop_use_stmt); - if (current_stmt) + use_stmt_info = loop_info->lookup_stmt (loop_use_stmt); + if (current_stmt_info) { - current_stmt_info = vinfo_for_stmt (current_stmt); - GROUP_NEXT_ELEMENT (current_stmt_info) = loop_use_stmt; - GROUP_FIRST_ELEMENT (use_stmt_info) - = GROUP_FIRST_ELEMENT (current_stmt_info); + REDUC_GROUP_NEXT_ELEMENT (current_stmt_info) = use_stmt_info; + REDUC_GROUP_FIRST_ELEMENT (use_stmt_info) + = REDUC_GROUP_FIRST_ELEMENT (current_stmt_info); } else - GROUP_FIRST_ELEMENT (use_stmt_info) = loop_use_stmt; + REDUC_GROUP_FIRST_ELEMENT (use_stmt_info) = use_stmt_info; lhs = gimple_assign_lhs (loop_use_stmt); - current_stmt = loop_use_stmt; + current_stmt_info = use_stmt_info; size++; } @@ -2539,34 +2548,24 @@ /* Swap the operands, if needed, to make the reduction operand be the second operand. */ lhs = PHI_RESULT (phi); - next_stmt = GROUP_FIRST_ELEMENT (vinfo_for_stmt (current_stmt)); - while (next_stmt) - { + stmt_vec_info next_stmt_info = REDUC_GROUP_FIRST_ELEMENT (current_stmt_info); + while (next_stmt_info) + { + gassign *next_stmt = as_a <gassign *> (next_stmt_info->stmt); if (gimple_assign_rhs2 (next_stmt) == lhs) { tree op = gimple_assign_rhs1 (next_stmt); - gimple *def_stmt = NULL; - - if (TREE_CODE (op) == SSA_NAME) - def_stmt = SSA_NAME_DEF_STMT (op); + stmt_vec_info def_stmt_info = loop_info->lookup_def (op); /* Check that the other def is either defined in the loop ("vect_internal_def"), or it's an induction (defined by a loop-header phi-node). */ - if (def_stmt - && gimple_bb (def_stmt) - && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt)) - && (is_gimple_assign (def_stmt) - || is_gimple_call (def_stmt) - || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt)) - == vect_induction_def - || (gimple_code (def_stmt) == GIMPLE_PHI - && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt)) - == vect_internal_def - && !is_loop_header_bb_p (gimple_bb (def_stmt))))) + if (def_stmt_info + && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt_info->stmt)) + && vect_valid_reduction_input_p (def_stmt_info)) { lhs = gimple_assign_lhs (next_stmt); - next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); + next_stmt_info = REDUC_GROUP_NEXT_ELEMENT (next_stmt_info); continue; } @@ -2575,31 +2574,18 @@ else { tree op = gimple_assign_rhs2 (next_stmt); - gimple *def_stmt = NULL; - - if (TREE_CODE (op) == SSA_NAME) - def_stmt = SSA_NAME_DEF_STMT (op); + stmt_vec_info def_stmt_info = loop_info->lookup_def (op); /* Check that the other def is either defined in the loop ("vect_internal_def"), or it's an induction (defined by a loop-header phi-node). */ - if (def_stmt - && gimple_bb (def_stmt) - && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt)) - && (is_gimple_assign (def_stmt) - || is_gimple_call (def_stmt) - || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt)) - == vect_induction_def - || (gimple_code (def_stmt) == GIMPLE_PHI - && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt)) - == vect_internal_def - && !is_loop_header_bb_p (gimple_bb (def_stmt))))) + if (def_stmt_info + && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt_info->stmt)) + && vect_valid_reduction_input_p (def_stmt_info)) { if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, "swapping oprnds: "); - dump_gimple_stmt (MSG_NOTE, TDF_SLIM, next_stmt, 0); - } + dump_printf_loc (MSG_NOTE, vect_location, "swapping oprnds: %G", + next_stmt); swap_ssa_operands (next_stmt, gimple_assign_rhs1_ptr (next_stmt), @@ -2614,17 +2600,159 @@ } lhs = gimple_assign_lhs (next_stmt); - next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); + next_stmt_info = REDUC_GROUP_NEXT_ELEMENT (next_stmt_info); } /* Save the chain for further analysis in SLP detection. */ - first = GROUP_FIRST_ELEMENT (vinfo_for_stmt (current_stmt)); - LOOP_VINFO_REDUCTION_CHAINS (loop_info).safe_push (first); - GROUP_SIZE (vinfo_for_stmt (first)) = size; + stmt_vec_info first_stmt_info + = REDUC_GROUP_FIRST_ELEMENT (current_stmt_info); + LOOP_VINFO_REDUCTION_CHAINS (loop_info).safe_push (first_stmt_info); + REDUC_GROUP_SIZE (first_stmt_info) = size; return true; } +/* Return true if we need an in-order reduction for operation CODE + on type TYPE. NEED_WRAPPING_INTEGRAL_OVERFLOW is true if integer + overflow must wrap. */ + +static bool +needs_fold_left_reduction_p (tree type, tree_code code, + bool need_wrapping_integral_overflow) +{ + /* CHECKME: check for !flag_finite_math_only too? */ + if (SCALAR_FLOAT_TYPE_P (type)) + switch (code) + { + case MIN_EXPR: + case MAX_EXPR: + return false; + + default: + return !flag_associative_math; + } + + if (INTEGRAL_TYPE_P (type)) + { + if (!operation_no_trapping_overflow (type, code)) + return true; + if (need_wrapping_integral_overflow + && !TYPE_OVERFLOW_WRAPS (type) + && operation_can_overflow (code)) + return true; + return false; + } + + if (SAT_FIXED_POINT_TYPE_P (type)) + return true; + + return false; +} + +/* Return true if the reduction PHI in LOOP with latch arg LOOP_ARG and + reduction operation CODE has a handled computation expression. */ + +bool +check_reduction_path (dump_user_location_t loc, loop_p loop, gphi *phi, + tree loop_arg, enum tree_code code) +{ + auto_vec<std::pair<ssa_op_iter, use_operand_p> > path; + auto_bitmap visited; + tree lookfor = PHI_RESULT (phi); + ssa_op_iter curri; + use_operand_p curr = op_iter_init_phiuse (&curri, phi, SSA_OP_USE); + while (USE_FROM_PTR (curr) != loop_arg) + curr = op_iter_next_use (&curri); + curri.i = curri.numops; + do + { + path.safe_push (std::make_pair (curri, curr)); + tree use = USE_FROM_PTR (curr); + if (use == lookfor) + break; + gimple *def = SSA_NAME_DEF_STMT (use); + if (gimple_nop_p (def) + || ! flow_bb_inside_loop_p (loop, gimple_bb (def))) + { +pop: + do + { + std::pair<ssa_op_iter, use_operand_p> x = path.pop (); + curri = x.first; + curr = x.second; + do + curr = op_iter_next_use (&curri); + /* Skip already visited or non-SSA operands (from iterating + over PHI args). */ + while (curr != NULL_USE_OPERAND_P + && (TREE_CODE (USE_FROM_PTR (curr)) != SSA_NAME + || ! bitmap_set_bit (visited, + SSA_NAME_VERSION + (USE_FROM_PTR (curr))))); + } + while (curr == NULL_USE_OPERAND_P && ! path.is_empty ()); + if (curr == NULL_USE_OPERAND_P) + break; + } + else + { + if (gimple_code (def) == GIMPLE_PHI) + curr = op_iter_init_phiuse (&curri, as_a <gphi *>(def), SSA_OP_USE); + else + curr = op_iter_init_use (&curri, def, SSA_OP_USE); + while (curr != NULL_USE_OPERAND_P + && (TREE_CODE (USE_FROM_PTR (curr)) != SSA_NAME + || ! bitmap_set_bit (visited, + SSA_NAME_VERSION + (USE_FROM_PTR (curr))))) + curr = op_iter_next_use (&curri); + if (curr == NULL_USE_OPERAND_P) + goto pop; + } + } + while (1); + if (dump_file && (dump_flags & TDF_DETAILS)) + { + dump_printf_loc (MSG_NOTE, loc, "reduction path: "); + unsigned i; + std::pair<ssa_op_iter, use_operand_p> *x; + FOR_EACH_VEC_ELT (path, i, x) + dump_printf (MSG_NOTE, "%T ", USE_FROM_PTR (x->second)); + dump_printf (MSG_NOTE, "\n"); + } + + /* Check whether the reduction path detected is valid. */ + bool fail = path.length () == 0; + bool neg = false; + for (unsigned i = 1; i < path.length (); ++i) + { + gimple *use_stmt = USE_STMT (path[i].second); + tree op = USE_FROM_PTR (path[i].second); + if (! has_single_use (op) + || ! is_gimple_assign (use_stmt)) + { + fail = true; + break; + } + if (gimple_assign_rhs_code (use_stmt) != code) + { + if (code == PLUS_EXPR + && gimple_assign_rhs_code (use_stmt) == MINUS_EXPR) + { + /* Track whether we negate the reduction value each iteration. */ + if (gimple_assign_rhs2 (use_stmt) == op) + neg = ! neg; + } + else + { + fail = true; + break; + } + } + } + return ! fail && ! neg; +} + /* Function vect_is_simple_reduction @@ -2670,15 +2798,16 @@ */ -static gimple * -vect_is_simple_reduction (loop_vec_info loop_info, gimple *phi, +static stmt_vec_info +vect_is_simple_reduction (loop_vec_info loop_info, stmt_vec_info phi_info, bool *double_reduc, bool need_wrapping_integral_overflow, enum vect_reduction_type *v_reduc_type) { + gphi *phi = as_a <gphi *> (phi_info->stmt); struct loop *loop = (gimple_bb (phi))->loop_father; struct loop *vect_loop = LOOP_VINFO_LOOP (loop_info); - gimple *def_stmt, *def1 = NULL, *def2 = NULL, *phi_use_stmt = NULL; + gimple *phi_use_stmt = NULL; enum tree_code orig_code, code; tree op1, op2, op3 = NULL_TREE, op4 = NULL_TREE; tree type; @@ -2731,22 +2860,22 @@ if (TREE_CODE (loop_arg) != SSA_NAME) { if (dump_enabled_p ()) - { - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "reduction: not ssa_name: "); - dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, loop_arg); - dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); - } + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "reduction: not ssa_name: %T\n", loop_arg); return NULL; } - def_stmt = SSA_NAME_DEF_STMT (loop_arg); - if (is_gimple_assign (def_stmt)) + stmt_vec_info def_stmt_info = loop_info->lookup_def (loop_arg); + if (!def_stmt_info + || !flow_bb_inside_loop_p (loop, gimple_bb (def_stmt_info->stmt))) + return NULL; + + if (gassign *def_stmt = dyn_cast <gassign *> (def_stmt_info->stmt)) { name = gimple_assign_lhs (def_stmt); phi_def = false; } - else if (gimple_code (def_stmt) == GIMPLE_PHI) + else if (gphi *def_stmt = dyn_cast <gphi *> (def_stmt_info->stmt)) { name = PHI_RESULT (def_stmt); phi_def = true; @@ -2754,17 +2883,12 @@ else { if (dump_enabled_p ()) - { - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "reduction: unhandled reduction operation: "); - dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, def_stmt, 0); - } + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "reduction: unhandled reduction operation: %G", + def_stmt_info->stmt); return NULL; } - if (! flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))) - return NULL; - nloop_uses = 0; auto_vec<gphi *, 3> lcphis; FOR_EACH_IMM_USE_FAST (use_p, imm_iter, name) @@ -2790,6 +2914,7 @@ defined in the inner loop. */ if (phi_def) { + gphi *def_stmt = as_a <gphi *> (def_stmt_info->stmt); op1 = PHI_ARG_DEF (def_stmt, 0); if (gimple_phi_num_args (def_stmt) != 1 @@ -2802,12 +2927,13 @@ return NULL; } - def1 = SSA_NAME_DEF_STMT (op1); + gimple *def1 = SSA_NAME_DEF_STMT (op1); if (gimple_bb (def1) && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt)) && loop->inner && flow_bb_inside_loop_p (loop->inner, gimple_bb (def1)) && is_gimple_assign (def1) + && is_a <gphi *> (phi_use_stmt) && flow_bb_inside_loop_p (loop->inner, gimple_bb (phi_use_stmt))) { if (dump_enabled_p ()) @@ -2815,7 +2941,7 @@ "detected double reduction: "); *double_reduc = true; - return def_stmt; + return def_stmt_info; } return NULL; @@ -2841,6 +2967,7 @@ } } + gassign *def_stmt = as_a <gassign *> (def_stmt_info->stmt); bool nested_in_vect_loop = flow_loop_nested_p (vect_loop, loop); code = orig_code = gimple_assign_rhs_code (def_stmt); @@ -2916,98 +3043,40 @@ if (dump_enabled_p ()) { dump_printf_loc (MSG_NOTE, vect_location, - "reduction: multiple types: operation type: "); - dump_generic_expr (MSG_NOTE, TDF_SLIM, type); - dump_printf (MSG_NOTE, ", operands types: "); - dump_generic_expr (MSG_NOTE, TDF_SLIM, - TREE_TYPE (op1)); - dump_printf (MSG_NOTE, ","); - dump_generic_expr (MSG_NOTE, TDF_SLIM, - TREE_TYPE (op2)); + "reduction: multiple types: operation type: " + "%T, operands types: %T,%T", + type, TREE_TYPE (op1), TREE_TYPE (op2)); if (op3) - { - dump_printf (MSG_NOTE, ","); - dump_generic_expr (MSG_NOTE, TDF_SLIM, - TREE_TYPE (op3)); - } + dump_printf (MSG_NOTE, ",%T", TREE_TYPE (op3)); if (op4) - { - dump_printf (MSG_NOTE, ","); - dump_generic_expr (MSG_NOTE, TDF_SLIM, - TREE_TYPE (op4)); - } + dump_printf (MSG_NOTE, ",%T", TREE_TYPE (op4)); dump_printf (MSG_NOTE, "\n"); } return NULL; } - /* Check that it's ok to change the order of the computation. + /* Check whether it's ok to change the order of the computation. Generally, when vectorizing a reduction we change the order of the computation. This may change the behavior of the program in some cases, so we need to check that this is ok. One exception is when vectorizing an outer-loop: the inner-loop is executed sequentially, and therefore vectorizing reductions in the inner-loop during outer-loop vectorization is safe. */ - - if (*v_reduc_type != COND_REDUCTION - && check_reduction) - { - /* CHECKME: check for !flag_finite_math_only too? */ - if (SCALAR_FLOAT_TYPE_P (type) && !flag_associative_math) - { - /* Changing the order of operations changes the semantics. */ - if (dump_enabled_p ()) - report_vect_op (MSG_MISSED_OPTIMIZATION, def_stmt, - "reduction: unsafe fp math optimization: "); - return NULL; - } - else if (INTEGRAL_TYPE_P (type)) - { - if (!operation_no_trapping_overflow (type, code)) - { - /* Changing the order of operations changes the semantics. */ - if (dump_enabled_p ()) - report_vect_op (MSG_MISSED_OPTIMIZATION, def_stmt, - "reduction: unsafe int math optimization" - " (overflow traps): "); - return NULL; - } - if (need_wrapping_integral_overflow - && !TYPE_OVERFLOW_WRAPS (type) - && operation_can_overflow (code)) - { - /* Changing the order of operations changes the semantics. */ - if (dump_enabled_p ()) - report_vect_op (MSG_MISSED_OPTIMIZATION, def_stmt, - "reduction: unsafe int math optimization" - " (overflow doesn't wrap): "); - return NULL; - } - } - else if (SAT_FIXED_POINT_TYPE_P (type)) - { - /* Changing the order of operations changes the semantics. */ - if (dump_enabled_p ()) - report_vect_op (MSG_MISSED_OPTIMIZATION, def_stmt, - "reduction: unsafe fixed-point math optimization: "); - return NULL; - } - } + if (check_reduction + && *v_reduc_type == TREE_CODE_REDUCTION + && needs_fold_left_reduction_p (type, code, + need_wrapping_integral_overflow)) + *v_reduc_type = FOLD_LEFT_REDUCTION; /* Reduction is safe. We're dealing with one of the following: 1) integer arithmetic and no trapv 2) floating point arithmetic, and special flags permit this optimization 3) nested cycle (i.e., outer loop vectorization). */ - if (TREE_CODE (op1) == SSA_NAME) - def1 = SSA_NAME_DEF_STMT (op1); - - if (TREE_CODE (op2) == SSA_NAME) - def2 = SSA_NAME_DEF_STMT (op2); - - if (code != COND_EXPR - && ((!def1 || gimple_nop_p (def1)) && (!def2 || gimple_nop_p (def2)))) + stmt_vec_info def1_info = loop_info->lookup_def (op1); + stmt_vec_info def2_info = loop_info->lookup_def (op2); + if (code != COND_EXPR && !def1_info && !def2_info) { if (dump_enabled_p ()) report_vect_op (MSG_NOTE, def_stmt, "reduction: no defs for operands: "); @@ -3018,38 +3087,24 @@ the other def is either defined in the loop ("vect_internal_def"), or it's an induction (defined by a loop-header phi-node). */ - if (def2 && def2 == phi + if (def2_info + && def2_info->stmt == phi && (code == COND_EXPR - || !def1 || gimple_nop_p (def1) - || !flow_bb_inside_loop_p (loop, gimple_bb (def1)) - || (def1 && flow_bb_inside_loop_p (loop, gimple_bb (def1)) - && (is_gimple_assign (def1) - || is_gimple_call (def1) - || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def1)) - == vect_induction_def - || (gimple_code (def1) == GIMPLE_PHI - && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def1)) - == vect_internal_def - && !is_loop_header_bb_p (gimple_bb (def1))))))) + || !def1_info + || !flow_bb_inside_loop_p (loop, gimple_bb (def1_info->stmt)) + || vect_valid_reduction_input_p (def1_info))) { if (dump_enabled_p ()) report_vect_op (MSG_NOTE, def_stmt, "detected reduction: "); - return def_stmt; - } - - if (def1 && def1 == phi + return def_stmt_info; + } + + if (def1_info + && def1_info->stmt == phi && (code == COND_EXPR - || !def2 || gimple_nop_p (def2) - || !flow_bb_inside_loop_p (loop, gimple_bb (def2)) - || (def2 && flow_bb_inside_loop_p (loop, gimple_bb (def2)) - && (is_gimple_assign (def2) - || is_gimple_call (def2) - || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def2)) - == vect_induction_def - || (gimple_code (def2) == GIMPLE_PHI - && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def2)) - == vect_internal_def - && !is_loop_header_bb_p (gimple_bb (def2))))))) + || !def2_info + || !flow_bb_inside_loop_p (loop, gimple_bb (def2_info->stmt)) + || vect_valid_reduction_input_p (def2_info))) { if (! nested_in_vect_loop && orig_code != MINUS_EXPR) { @@ -3101,7 +3156,7 @@ report_vect_op (MSG_NOTE, def_stmt, "detected reduction: "); } - return def_stmt; + return def_stmt_info; } /* Try to find SLP reduction chain. */ @@ -3114,121 +3169,22 @@ report_vect_op (MSG_NOTE, def_stmt, "reduction: detected reduction chain: "); - return def_stmt; + return def_stmt_info; } /* Dissolve group eventually half-built by vect_is_slp_reduction. */ - gimple *first = GROUP_FIRST_ELEMENT (vinfo_for_stmt (def_stmt)); + stmt_vec_info first = REDUC_GROUP_FIRST_ELEMENT (def_stmt_info); while (first) { - gimple *next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (first)); - GROUP_FIRST_ELEMENT (vinfo_for_stmt (first)) = NULL; - GROUP_NEXT_ELEMENT (vinfo_for_stmt (first)) = NULL; + stmt_vec_info next = REDUC_GROUP_NEXT_ELEMENT (first); + REDUC_GROUP_FIRST_ELEMENT (first) = NULL; + REDUC_GROUP_NEXT_ELEMENT (first) = NULL; first = next; } /* Look for the expression computing loop_arg from loop PHI result. */ - auto_vec<std::pair<ssa_op_iter, use_operand_p> > path; - auto_bitmap visited; - tree lookfor = PHI_RESULT (phi); - ssa_op_iter curri; - use_operand_p curr = op_iter_init_phiuse (&curri, as_a <gphi *>(phi), - SSA_OP_USE); - while (USE_FROM_PTR (curr) != loop_arg) - curr = op_iter_next_use (&curri); - curri.i = curri.numops; - do - { - path.safe_push (std::make_pair (curri, curr)); - tree use = USE_FROM_PTR (curr); - if (use == lookfor) - break; - gimple *def = SSA_NAME_DEF_STMT (use); - if (gimple_nop_p (def) - || ! flow_bb_inside_loop_p (loop, gimple_bb (def))) - { -pop: - do - { - std::pair<ssa_op_iter, use_operand_p> x = path.pop (); - curri = x.first; - curr = x.second; - do - curr = op_iter_next_use (&curri); - /* Skip already visited or non-SSA operands (from iterating - over PHI args). */ - while (curr != NULL_USE_OPERAND_P - && (TREE_CODE (USE_FROM_PTR (curr)) != SSA_NAME - || ! bitmap_set_bit (visited, - SSA_NAME_VERSION - (USE_FROM_PTR (curr))))); - } - while (curr == NULL_USE_OPERAND_P && ! path.is_empty ()); - if (curr == NULL_USE_OPERAND_P) - break; - } - else - { - if (gimple_code (def) == GIMPLE_PHI) - curr = op_iter_init_phiuse (&curri, as_a <gphi *>(def), SSA_OP_USE); - else - curr = op_iter_init_use (&curri, def, SSA_OP_USE); - while (curr != NULL_USE_OPERAND_P - && (TREE_CODE (USE_FROM_PTR (curr)) != SSA_NAME - || ! bitmap_set_bit (visited, - SSA_NAME_VERSION - (USE_FROM_PTR (curr))))) - curr = op_iter_next_use (&curri); - if (curr == NULL_USE_OPERAND_P) - goto pop; - } - } - while (1); - if (dump_file && (dump_flags & TDF_DETAILS)) - { - dump_printf_loc (MSG_NOTE, vect_location, - "reduction path: "); - unsigned i; - std::pair<ssa_op_iter, use_operand_p> *x; - FOR_EACH_VEC_ELT (path, i, x) - { - dump_generic_expr (MSG_NOTE, TDF_SLIM, USE_FROM_PTR (x->second)); - dump_printf (MSG_NOTE, " "); - } - dump_printf (MSG_NOTE, "\n"); - } - - /* Check whether the reduction path detected is valid. */ - bool fail = path.length () == 0; - bool neg = false; - for (unsigned i = 1; i < path.length (); ++i) - { - gimple *use_stmt = USE_STMT (path[i].second); - tree op = USE_FROM_PTR (path[i].second); - if (! has_single_use (op) - || ! is_gimple_assign (use_stmt)) - { - fail = true; - break; - } - if (gimple_assign_rhs_code (use_stmt) != code) - { - if (code == PLUS_EXPR - && gimple_assign_rhs_code (use_stmt) == MINUS_EXPR) - { - /* Track whether we negate the reduction value each iteration. */ - if (gimple_assign_rhs2 (use_stmt) == op) - neg = ! neg; - } - else - { - fail = true; - break; - } - } - } - if (! fail && ! neg) - return def_stmt; + if (check_reduction_path (vect_location, loop, phi, loop_arg, code)) + return def_stmt_info; if (dump_enabled_p ()) { @@ -3243,24 +3199,24 @@ in-place if it enables detection of more reductions. Arguments as there. */ -gimple * -vect_force_simple_reduction (loop_vec_info loop_info, gimple *phi, +stmt_vec_info +vect_force_simple_reduction (loop_vec_info loop_info, stmt_vec_info phi_info, bool *double_reduc, bool need_wrapping_integral_overflow) { enum vect_reduction_type v_reduc_type; - gimple *def = vect_is_simple_reduction (loop_info, phi, double_reduc, - need_wrapping_integral_overflow, - &v_reduc_type); - if (def) - { - stmt_vec_info reduc_def_info = vinfo_for_stmt (phi); - STMT_VINFO_REDUC_TYPE (reduc_def_info) = v_reduc_type; - STMT_VINFO_REDUC_DEF (reduc_def_info) = def; - reduc_def_info = vinfo_for_stmt (def); - STMT_VINFO_REDUC_DEF (reduc_def_info) = phi; - } - return def; + stmt_vec_info def_info + = vect_is_simple_reduction (loop_info, phi_info, double_reduc, + need_wrapping_integral_overflow, + &v_reduc_type); + if (def_info) + { + STMT_VINFO_REDUC_TYPE (phi_info) = v_reduc_type; + STMT_VINFO_REDUC_DEF (phi_info) = def_info; + STMT_VINFO_REDUC_TYPE (def_info) = v_reduc_type; + STMT_VINFO_REDUC_DEF (def_info) = phi_info; + } + return def_info; } /* Calculate cost of peeling the loop PEEL_ITERS_PROLOGUE times. */ @@ -3272,11 +3228,11 @@ stmt_vector_for_cost *epilogue_cost_vec) { int retval = 0; - int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); + int assumed_vf = vect_vf_for_cost (loop_vinfo); if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)) { - *peel_iters_epilogue = vf/2; + *peel_iters_epilogue = assumed_vf / 2; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "cost model: epilogue peel iters set to vf/2 " @@ -3294,35 +3250,27 @@ int niters = LOOP_VINFO_INT_NITERS (loop_vinfo); peel_iters_prologue = niters < peel_iters_prologue ? niters : peel_iters_prologue; - *peel_iters_epilogue = (niters - peel_iters_prologue) % vf; + *peel_iters_epilogue = (niters - peel_iters_prologue) % assumed_vf; /* If we need to peel for gaps, but no peeling is required, we have to peel VF iterations. */ if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) && !*peel_iters_epilogue) - *peel_iters_epilogue = vf; + *peel_iters_epilogue = assumed_vf; } stmt_info_for_cost *si; int j; if (peel_iters_prologue) FOR_EACH_VEC_ELT (*scalar_cost_vec, j, si) - { - stmt_vec_info stmt_info - = si->stmt ? vinfo_for_stmt (si->stmt) : NULL; - retval += record_stmt_cost (prologue_cost_vec, - si->count * peel_iters_prologue, - si->kind, stmt_info, si->misalign, - vect_prologue); - } + retval += record_stmt_cost (prologue_cost_vec, + si->count * peel_iters_prologue, + si->kind, si->stmt_info, si->misalign, + vect_prologue); if (*peel_iters_epilogue) FOR_EACH_VEC_ELT (*scalar_cost_vec, j, si) - { - stmt_vec_info stmt_info - = si->stmt ? vinfo_for_stmt (si->stmt) : NULL; - retval += record_stmt_cost (epilogue_cost_vec, - si->count * *peel_iters_epilogue, - si->kind, stmt_info, si->misalign, - vect_epilogue); - } + retval += record_stmt_cost (epilogue_cost_vec, + si->count * *peel_iters_epilogue, + si->kind, si->stmt_info, si->misalign, + vect_epilogue); return retval; } @@ -3356,7 +3304,7 @@ unsigned vec_epilogue_cost = 0; int scalar_single_iter_cost = 0; int scalar_outside_cost = 0; - int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); + int assumed_vf = vect_vf_for_cost (loop_vinfo); int npeel = LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo); void *target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); @@ -3393,6 +3341,18 @@ /* Count LEN - 1 ANDs and LEN comparisons. */ (void) add_stmt_cost (target_cost_data, len * 2 - 1, scalar_stmt, NULL, 0, vect_prologue); + len = LOOP_VINFO_LOWER_BOUNDS (loop_vinfo).length (); + if (len) + { + /* Count LEN - 1 ANDs and LEN comparisons. */ + unsigned int nstmts = len * 2 - 1; + /* +1 for each bias that needs adding. */ + for (unsigned int i = 0; i < len; ++i) + if (!LOOP_VINFO_LOWER_BOUNDS (loop_vinfo)[i].unsigned_p) + nstmts += 1; + (void) add_stmt_cost (target_cost_data, nstmts, scalar_stmt, + NULL, 0, vect_prologue); + } dump_printf (MSG_NOTE, "cost model: Adding cost of checks for loop " "versioning aliasing.\n"); @@ -3425,7 +3385,7 @@ = LOOP_VINFO_SINGLE_SCALAR_ITERATION_COST (loop_vinfo); /* Add additional cost for the peeled instructions in prologue and epilogue - loop. + loop. (For fully-masked loops there will be no peeling.) FORNOW: If we don't know the value of peel_iters for prologue or epilogue at compile-time - we assume it's vf/2 (the worst would be vf-1). @@ -3433,15 +3393,33 @@ TODO: Build an expression that represents peel_iters for prologue and epilogue to be used in a run-time test. */ - if (npeel < 0) - { - peel_iters_prologue = vf/2; + if (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) + { + peel_iters_prologue = 0; + peel_iters_epilogue = 0; + + if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo)) + { + /* We need to peel exactly one iteration. */ + peel_iters_epilogue += 1; + stmt_info_for_cost *si; + int j; + FOR_EACH_VEC_ELT (LOOP_VINFO_SCALAR_ITERATION_COST (loop_vinfo), + j, si) + (void) add_stmt_cost (target_cost_data, si->count, + si->kind, si->stmt_info, si->misalign, + vect_epilogue); + } + } + else if (npeel < 0) + { + peel_iters_prologue = assumed_vf / 2; dump_printf (MSG_NOTE, "cost model: " "prologue peel iters set to vf/2.\n"); /* If peeling for alignment is unknown, loop bound of main loop becomes unknown. */ - peel_iters_epilogue = vf/2; + peel_iters_epilogue = assumed_vf / 2; dump_printf (MSG_NOTE, "cost model: " "epilogue peel iters set to vf/2 because " "peeling for alignment is unknown.\n"); @@ -3462,15 +3440,13 @@ int j; FOR_EACH_VEC_ELT (LOOP_VINFO_SCALAR_ITERATION_COST (loop_vinfo), j, si) { - struct _stmt_vec_info *stmt_info - = si->stmt ? vinfo_for_stmt (si->stmt) : NULL; (void) add_stmt_cost (target_cost_data, si->count * peel_iters_prologue, - si->kind, stmt_info, si->misalign, + si->kind, si->stmt_info, si->misalign, vect_prologue); (void) add_stmt_cost (target_cost_data, si->count * peel_iters_epilogue, - si->kind, stmt_info, si->misalign, + si->kind, si->stmt_info, si->misalign, vect_epilogue); } } @@ -3493,20 +3469,12 @@ &epilogue_cost_vec); FOR_EACH_VEC_ELT (prologue_cost_vec, j, si) - { - struct _stmt_vec_info *stmt_info - = si->stmt ? vinfo_for_stmt (si->stmt) : NULL; - (void) add_stmt_cost (data, si->count, si->kind, stmt_info, - si->misalign, vect_prologue); - } + (void) add_stmt_cost (data, si->count, si->kind, si->stmt_info, + si->misalign, vect_prologue); FOR_EACH_VEC_ELT (epilogue_cost_vec, j, si) - { - struct _stmt_vec_info *stmt_info - = si->stmt ? vinfo_for_stmt (si->stmt) : NULL; - (void) add_stmt_cost (data, si->count, si->kind, stmt_info, - si->misalign, vect_epilogue); - } + (void) add_stmt_cost (data, si->count, si->kind, si->stmt_info, + si->misalign, vect_epilogue); prologue_cost_vec.release (); epilogue_cost_vec.release (); @@ -3620,30 +3588,32 @@ PL_ITERS = prologue iterations, EP_ITERS= epilogue iterations SOC = scalar outside cost for run time cost model check. */ - if ((scalar_single_iter_cost * vf) > (int) vec_inside_cost) - { - if (vec_outside_cost <= 0) + if ((scalar_single_iter_cost * assumed_vf) > (int) vec_inside_cost) + { + min_profitable_iters = ((vec_outside_cost - scalar_outside_cost) + * assumed_vf + - vec_inside_cost * peel_iters_prologue + - vec_inside_cost * peel_iters_epilogue); + if (min_profitable_iters <= 0) min_profitable_iters = 0; else - { - min_profitable_iters = ((vec_outside_cost - scalar_outside_cost) * vf - - vec_inside_cost * peel_iters_prologue - - vec_inside_cost * peel_iters_epilogue) - / ((scalar_single_iter_cost * vf) - - vec_inside_cost); - - if ((scalar_single_iter_cost * vf * min_profitable_iters) - <= (((int) vec_inside_cost * min_profitable_iters) - + (((int) vec_outside_cost - scalar_outside_cost) * vf))) - min_profitable_iters++; - } + { + min_profitable_iters /= ((scalar_single_iter_cost * assumed_vf) + - vec_inside_cost); + + if ((scalar_single_iter_cost * assumed_vf * min_profitable_iters) + <= (((int) vec_inside_cost * min_profitable_iters) + + (((int) vec_outside_cost - scalar_outside_cost) + * assumed_vf))) + min_profitable_iters++; + } } /* vector version will never be profitable. */ else { if (LOOP_VINFO_LOOP (loop_vinfo)->force_vectorize) - warning_at (vect_location, OPT_Wopenmp_simd, "vectorization " - "did not happen for a simd loop"); + warning_at (vect_location.get_location_t (), OPT_Wopenmp_simd, + "vectorization did not happen for a simd loop"); if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, @@ -3651,7 +3621,7 @@ "divided by the scalar iteration cost = %d " "is greater or equal to the vectorization factor = %d" ".\n", - vec_inside_cost, scalar_single_iter_cost, vf); + vec_inside_cost, scalar_single_iter_cost, assumed_vf); *ret_min_profitable_niters = -1; *ret_min_profitable_estimate = -1; return; @@ -3661,9 +3631,10 @@ " Calculated minimum iters for profitability: %d\n", min_profitable_iters); - /* We want the vectorized loop to execute at least once. */ - if (min_profitable_iters < (vf + peel_iters_prologue)) - min_profitable_iters = vf + peel_iters_prologue; + if (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) + && min_profitable_iters < (assumed_vf + peel_iters_prologue)) + /* We want the vectorized loop to execute at least once. */ + min_profitable_iters = assumed_vf + peel_iters_prologue; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, @@ -3683,10 +3654,11 @@ min_profitable_estimate = 0; else { - min_profitable_estimate = ((vec_outside_cost + scalar_outside_cost) * vf + min_profitable_estimate = ((vec_outside_cost + scalar_outside_cost) + * assumed_vf - vec_inside_cost * peel_iters_prologue - vec_inside_cost * peel_iters_epilogue) - / ((scalar_single_iter_cost * vf) + / ((scalar_single_iter_cost * assumed_vf) - vec_inside_cost); } min_profitable_estimate = MAX (min_profitable_estimate, min_profitable_iters); @@ -3702,12 +3674,13 @@ vector elements (not bits) for a vector with NELT elements. */ static void calc_vec_perm_mask_for_shift (unsigned int offset, unsigned int nelt, - vec_perm_indices *sel) + vec_perm_builder *sel) { - unsigned int i; - - for (i = 0; i < nelt; i++) - sel->quick_push ((i + offset) & (2 * nelt - 1)); + /* The encoding is a single stepped pattern. Any wrap-around is handled + by vec_perm_indices. */ + sel->new_vector (nelt, 1, 3); + for (unsigned int i = 0; i < 3; i++) + sel->quick_push (i + offset); } /* Checks whether the target supports whole-vector shifts for vectors of mode @@ -3719,17 +3692,18 @@ if (optab_handler (vec_shr_optab, mode) != CODE_FOR_nothing) return true; - if (direct_optab_handler (vec_perm_const_optab, mode) == CODE_FOR_nothing) + /* Variable-length vectors should be handled via the optab. */ + unsigned int nelt; + if (!GET_MODE_NUNITS (mode).is_constant (&nelt)) return false; - unsigned int i, nelt = GET_MODE_NUNITS (mode); - auto_vec_perm_indices sel (nelt); - - for (i = nelt/2; i >= 1; i/=2) - { - sel.truncate (0); + vec_perm_builder sel; + vec_perm_indices indices; + for (unsigned int i = nelt / 2; i >= 1; i /= 2) + { calc_vec_perm_mask_for_shift (i, nelt, &sel); - if (!can_vec_perm_p (mode, false, &sel)) + indices.new_vector (sel, 2, nelt); + if (!can_vec_perm_const_p (mode, indices, false)) return false; } return true; @@ -3745,104 +3719,127 @@ the loop, and the epilogue code that must be generated. */ static void -vect_model_reduction_cost (stmt_vec_info stmt_info, enum tree_code reduc_code, - int ncopies) +vect_model_reduction_cost (stmt_vec_info stmt_info, internal_fn reduc_fn, + int ncopies, stmt_vector_for_cost *cost_vec) { - int prologue_cost = 0, epilogue_cost = 0; + int prologue_cost = 0, epilogue_cost = 0, inside_cost; enum tree_code code; optab optab; tree vectype; - gimple *orig_stmt; machine_mode mode; loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); struct loop *loop = NULL; - void *target_cost_data; if (loop_vinfo) - { - loop = LOOP_VINFO_LOOP (loop_vinfo); - target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); - } - else - target_cost_data = BB_VINFO_TARGET_COST_DATA (STMT_VINFO_BB_VINFO (stmt_info)); + loop = LOOP_VINFO_LOOP (loop_vinfo); /* Condition reductions generate two reductions in the loop. */ - if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) == COND_REDUCTION) + vect_reduction_type reduction_type + = STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info); + if (reduction_type == COND_REDUCTION) ncopies *= 2; - /* Cost of reduction op inside loop. */ - unsigned inside_cost = add_stmt_cost (target_cost_data, ncopies, vector_stmt, - stmt_info, 0, vect_body); - vectype = STMT_VINFO_VECTYPE (stmt_info); mode = TYPE_MODE (vectype); - orig_stmt = STMT_VINFO_RELATED_STMT (stmt_info); - - if (!orig_stmt) - orig_stmt = STMT_VINFO_STMT (stmt_info); - - code = gimple_assign_rhs_code (orig_stmt); - - /* Add in cost for initial definition. - For cond reduction we have four vectors: initial index, step, initial - result of the data reduction, initial value of the index reduction. */ - int prologue_stmts = STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) - == COND_REDUCTION ? 4 : 1; - prologue_cost += add_stmt_cost (target_cost_data, prologue_stmts, - scalar_to_vec, stmt_info, 0, - vect_prologue); + stmt_vec_info orig_stmt_info = vect_orig_stmt (stmt_info); + + code = gimple_assign_rhs_code (orig_stmt_info->stmt); + + if (reduction_type == EXTRACT_LAST_REDUCTION + || reduction_type == FOLD_LEFT_REDUCTION) + { + /* No extra instructions needed in the prologue. */ + prologue_cost = 0; + + if (reduction_type == EXTRACT_LAST_REDUCTION || reduc_fn != IFN_LAST) + /* Count one reduction-like operation per vector. */ + inside_cost = record_stmt_cost (cost_vec, ncopies, vec_to_scalar, + stmt_info, 0, vect_body); + else + { + /* Use NELEMENTS extracts and NELEMENTS scalar ops. */ + unsigned int nelements = ncopies * vect_nunits_for_cost (vectype); + inside_cost = record_stmt_cost (cost_vec, nelements, + vec_to_scalar, stmt_info, 0, + vect_body); + inside_cost += record_stmt_cost (cost_vec, nelements, + scalar_stmt, stmt_info, 0, + vect_body); + } + } + else + { + /* Add in cost for initial definition. + For cond reduction we have four vectors: initial index, step, + initial result of the data reduction, initial value of the index + reduction. */ + int prologue_stmts = reduction_type == COND_REDUCTION ? 4 : 1; + prologue_cost += record_stmt_cost (cost_vec, prologue_stmts, + scalar_to_vec, stmt_info, 0, + vect_prologue); + + /* Cost of reduction op inside loop. */ + inside_cost = record_stmt_cost (cost_vec, ncopies, vector_stmt, + stmt_info, 0, vect_body); + } /* Determine cost of epilogue code. We have a reduction operator that will reduce the vector in one statement. Also requires scalar extract. */ - if (!loop || !nested_in_vect_loop_p (loop, orig_stmt)) - { - if (reduc_code != ERROR_MARK) - { - if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) == COND_REDUCTION) + if (!loop || !nested_in_vect_loop_p (loop, orig_stmt_info)) + { + if (reduc_fn != IFN_LAST) + { + if (reduction_type == COND_REDUCTION) { /* An EQ stmt and an COND_EXPR stmt. */ - epilogue_cost += add_stmt_cost (target_cost_data, 2, - vector_stmt, stmt_info, 0, - vect_epilogue); + epilogue_cost += record_stmt_cost (cost_vec, 2, + vector_stmt, stmt_info, 0, + vect_epilogue); /* Reduction of the max index and a reduction of the found values. */ - epilogue_cost += add_stmt_cost (target_cost_data, 2, - vec_to_scalar, stmt_info, 0, - vect_epilogue); + epilogue_cost += record_stmt_cost (cost_vec, 2, + vec_to_scalar, stmt_info, 0, + vect_epilogue); /* A broadcast of the max value. */ - epilogue_cost += add_stmt_cost (target_cost_data, 1, - scalar_to_vec, stmt_info, 0, - vect_epilogue); + epilogue_cost += record_stmt_cost (cost_vec, 1, + scalar_to_vec, stmt_info, 0, + vect_epilogue); } else { - epilogue_cost += add_stmt_cost (target_cost_data, 1, vector_stmt, - stmt_info, 0, vect_epilogue); - epilogue_cost += add_stmt_cost (target_cost_data, 1, - vec_to_scalar, stmt_info, 0, - vect_epilogue); + epilogue_cost += record_stmt_cost (cost_vec, 1, vector_stmt, + stmt_info, 0, vect_epilogue); + epilogue_cost += record_stmt_cost (cost_vec, 1, + vec_to_scalar, stmt_info, 0, + vect_epilogue); } } - else if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) == COND_REDUCTION) - { - unsigned nunits = TYPE_VECTOR_SUBPARTS (vectype); + else if (reduction_type == COND_REDUCTION) + { + unsigned estimated_nunits = vect_nunits_for_cost (vectype); /* Extraction of scalar elements. */ - epilogue_cost += add_stmt_cost (target_cost_data, 2 * nunits, - vec_to_scalar, stmt_info, 0, - vect_epilogue); + epilogue_cost += record_stmt_cost (cost_vec, + 2 * estimated_nunits, + vec_to_scalar, stmt_info, 0, + vect_epilogue); /* Scalar max reductions via COND_EXPR / MAX_EXPR. */ - epilogue_cost += add_stmt_cost (target_cost_data, 2 * nunits - 3, - scalar_stmt, stmt_info, 0, - vect_epilogue); + epilogue_cost += record_stmt_cost (cost_vec, + 2 * estimated_nunits - 3, + scalar_stmt, stmt_info, 0, + vect_epilogue); } + else if (reduction_type == EXTRACT_LAST_REDUCTION + || reduction_type == FOLD_LEFT_REDUCTION) + /* No extra instructions need in the epilogue. */ + ; else { int vec_size_in_bits = tree_to_uhwi (TYPE_SIZE (vectype)); tree bitsize = - TYPE_SIZE (TREE_TYPE (gimple_assign_lhs (orig_stmt))); + TYPE_SIZE (TREE_TYPE (gimple_assign_lhs (orig_stmt_info->stmt))); int element_bitsize = tree_to_uhwi (bitsize); int nelements = vec_size_in_bits / element_bitsize; @@ -3859,21 +3856,21 @@ { /* Final reduction via vector shifts and the reduction operator. Also requires scalar extract. */ - epilogue_cost += add_stmt_cost (target_cost_data, - exact_log2 (nelements) * 2, - vector_stmt, stmt_info, 0, - vect_epilogue); - epilogue_cost += add_stmt_cost (target_cost_data, 1, - vec_to_scalar, stmt_info, 0, - vect_epilogue); + epilogue_cost += record_stmt_cost (cost_vec, + exact_log2 (nelements) * 2, + vector_stmt, stmt_info, 0, + vect_epilogue); + epilogue_cost += record_stmt_cost (cost_vec, 1, + vec_to_scalar, stmt_info, 0, + vect_epilogue); } else /* Use extracts and reduction op for final reduction. For N elements, we have N extracts and N-1 reduction ops. */ - epilogue_cost += add_stmt_cost (target_cost_data, - nelements + nelements - 1, - vector_stmt, stmt_info, 0, - vect_epilogue); + epilogue_cost += record_stmt_cost (cost_vec, + nelements + nelements - 1, + vector_stmt, stmt_info, 0, + vect_epilogue); } } @@ -3890,22 +3887,21 @@ Models cost for induction operations. */ static void -vect_model_induction_cost (stmt_vec_info stmt_info, int ncopies) +vect_model_induction_cost (stmt_vec_info stmt_info, int ncopies, + stmt_vector_for_cost *cost_vec) { - loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); - void *target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); unsigned inside_cost, prologue_cost; if (PURE_SLP_STMT (stmt_info)) return; /* loop cost for vec_loop. */ - inside_cost = add_stmt_cost (target_cost_data, ncopies, vector_stmt, - stmt_info, 0, vect_body); + inside_cost = record_stmt_cost (cost_vec, ncopies, vector_stmt, + stmt_info, 0, vect_body); /* prologue cost for vec_init and vec_step. */ - prologue_cost = add_stmt_cost (target_cost_data, 2, scalar_to_vec, - stmt_info, 0, vect_prologue); + prologue_cost = record_stmt_cost (cost_vec, 2, scalar_to_vec, + stmt_info, 0, vect_prologue); if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, @@ -3918,15 +3914,15 @@ /* Function get_initial_def_for_reduction Input: - STMT - a stmt that performs a reduction operation in the loop. + STMT_VINFO - a stmt that performs a reduction operation in the loop. INIT_VAL - the initial value of the reduction variable Output: ADJUSTMENT_DEF - a tree that holds a value to be added to the final result of the reduction (used for adjusting the epilog - see below). - Return a vector variable, initialized according to the operation that STMT - performs. This vector will be used as the initial value of the - vector of partial results. + Return a vector variable, initialized according to the operation that + STMT_VINFO performs. This vector will be used as the initial value + of the vector of partial results. Option1 (adjust in epilog): Initialize the vector as follows: add/bit or/xor: [0,0,...,0,0] @@ -3947,7 +3943,7 @@ for (i=0;i<n;i++) s = s + a[i]; - STMT is 's = s + a[i]', and the reduction variable is 's'. + STMT_VINFO is 's = s + a[i]', and the reduction variable is 's'. For a vector of 4 units, we want to return either [0,0,0,init_val], or [0,0,0,0] and let the caller know that it needs to adjust the result at the end by 'init_val'. @@ -3959,60 +3955,30 @@ A cost model should help decide between these two schemes. */ tree -get_initial_def_for_reduction (gimple *stmt, tree init_val, +get_initial_def_for_reduction (stmt_vec_info stmt_vinfo, tree init_val, tree *adjustment_def) { - stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); tree scalar_type = TREE_TYPE (init_val); tree vectype = get_vectype_for_scalar_type (scalar_type); - int nunits; - enum tree_code code = gimple_assign_rhs_code (stmt); + enum tree_code code = gimple_assign_rhs_code (stmt_vinfo->stmt); tree def_for_init; tree init_def; - int i; - bool nested_in_vect_loop = false; REAL_VALUE_TYPE real_init_val = dconst0; int int_init_val = 0; - gimple *def_stmt = NULL; gimple_seq stmts = NULL; gcc_assert (vectype); - nunits = TYPE_VECTOR_SUBPARTS (vectype); gcc_assert (POINTER_TYPE_P (scalar_type) || INTEGRAL_TYPE_P (scalar_type) || SCALAR_FLOAT_TYPE_P (scalar_type)); - if (nested_in_vect_loop_p (loop, stmt)) - nested_in_vect_loop = true; - else - gcc_assert (loop == (gimple_bb (stmt))->loop_father); - - /* In case of double reduction we only create a vector variable to be put - in the reduction phi node. The actual statement creation is done in - vect_create_epilog_for_reduction. */ - if (adjustment_def && nested_in_vect_loop - && TREE_CODE (init_val) == SSA_NAME - && (def_stmt = SSA_NAME_DEF_STMT (init_val)) - && gimple_code (def_stmt) == GIMPLE_PHI - && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt)) - && vinfo_for_stmt (def_stmt) - && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt)) - == vect_double_reduction_def) - { - *adjustment_def = NULL; - return vect_create_destination_var (init_val, vectype); - } - - /* In case of a nested reduction do not use an adjustment def as - that case is not supported by the epilogue generation correctly - if ncopies is not one. */ - if (adjustment_def && nested_in_vect_loop) - { - *adjustment_def = NULL; - return vect_get_vec_def_for_operand (init_val, stmt); - } + gcc_assert (nested_in_vect_loop_p (loop, stmt_vinfo) + || loop == (gimple_bb (stmt_vinfo->stmt))->loop_father); + + vect_reduction_type reduction_type + = STMT_VINFO_VEC_REDUCTION_TYPE (stmt_vinfo); switch (code) { @@ -4026,7 +3992,7 @@ case MULT_EXPR: case BIT_AND_EXPR: { - /* ADJUSMENT_DEF is NULL when called from + /* ADJUSTMENT_DEF is NULL when called from vect_create_epilog_for_reduction to vectorize double reduction. */ if (adjustment_def) *adjustment_def = init_val; @@ -4049,14 +4015,21 @@ /* Option1: the first element is '0' or '1' as well. */ init_def = gimple_build_vector_from_val (&stmts, vectype, def_for_init); + else if (!TYPE_VECTOR_SUBPARTS (vectype).is_constant ()) + { + /* Option2 (variable length): the first element is INIT_VAL. */ + init_def = gimple_build_vector_from_val (&stmts, vectype, + def_for_init); + init_def = gimple_build (&stmts, CFN_VEC_SHL_INSERT, + vectype, init_def, init_val); + } else { /* Option2: the first element is INIT_VAL. */ - auto_vec<tree, 32> elts (nunits); + tree_vector_builder elts (vectype, 1, 2); elts.quick_push (init_val); - for (i = 1; i < nunits; ++i) - elts.quick_push (def_for_init); - init_def = gimple_build_vector (&stmts, vectype, elts); + elts.quick_push (def_for_init); + init_def = gimple_build_vector (&stmts, &elts); } } break; @@ -4068,9 +4041,10 @@ if (adjustment_def) { *adjustment_def = NULL_TREE; - if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_vinfo) != COND_REDUCTION) + if (reduction_type != COND_REDUCTION + && reduction_type != EXTRACT_LAST_REDUCTION) { - init_def = vect_get_vec_def_for_operand (init_val, stmt); + init_def = vect_get_vec_def_for_operand (init_val, stmt_vinfo); break; } } @@ -4089,78 +4063,39 @@ } /* Get at the initial defs for the reduction PHIs in SLP_NODE. - NUMBER_OF_VECTORS is the number of vector defs to create. */ + NUMBER_OF_VECTORS is the number of vector defs to create. + If NEUTRAL_OP is nonnull, introducing extra elements of that + value will not change the result. */ static void get_initial_defs_for_reduction (slp_tree slp_node, vec<tree> *vec_oprnds, unsigned int number_of_vectors, - enum tree_code code, bool reduc_chain) + bool reduc_chain, tree neutral_op) { - vec<gimple *> stmts = SLP_TREE_SCALAR_STMTS (slp_node); - gimple *stmt = stmts[0]; - stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); - unsigned nunits; + vec<stmt_vec_info> stmts = SLP_TREE_SCALAR_STMTS (slp_node); + stmt_vec_info stmt_vinfo = stmts[0]; + unsigned HOST_WIDE_INT nunits; unsigned j, number_of_places_left_in_vector; - tree vector_type, scalar_type; + tree vector_type; tree vop; int group_size = stmts.length (); unsigned int vec_num, i; unsigned number_of_copies = 1; vec<tree> voprnds; voprnds.create (number_of_vectors); - tree neutral_op = NULL; struct loop *loop; + auto_vec<tree, 16> permute_results; vector_type = STMT_VINFO_VECTYPE (stmt_vinfo); - scalar_type = TREE_TYPE (vector_type); - nunits = TYPE_VECTOR_SUBPARTS (vector_type); gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def); - loop = (gimple_bb (stmt))->loop_father; + loop = (gimple_bb (stmt_vinfo->stmt))->loop_father; gcc_assert (loop); edge pe = loop_preheader_edge (loop); - /* op is the reduction operand of the first stmt already. */ - /* For additional copies (see the explanation of NUMBER_OF_COPIES below) - we need either neutral operands or the original operands. See - get_initial_def_for_reduction() for details. */ - switch (code) - { - case WIDEN_SUM_EXPR: - case DOT_PROD_EXPR: - case SAD_EXPR: - case PLUS_EXPR: - case MINUS_EXPR: - case BIT_IOR_EXPR: - case BIT_XOR_EXPR: - neutral_op = build_zero_cst (scalar_type); - break; - - case MULT_EXPR: - neutral_op = build_one_cst (scalar_type); - break; - - case BIT_AND_EXPR: - neutral_op = build_all_ones_cst (scalar_type); - break; - - /* For MIN/MAX we don't have an easy neutral operand but - the initial values can be used fine here. Only for - a reduction chain we have to force a neutral element. */ - case MAX_EXPR: - case MIN_EXPR: - if (! reduc_chain) - neutral_op = NULL; - else - neutral_op = PHI_ARG_DEF_FROM_EDGE (stmt, pe); - break; - - default: - gcc_assert (! reduc_chain); - neutral_op = NULL; - } + gcc_assert (!reduc_chain || neutral_op); /* NUMBER_OF_COPIES is the number of times we need to use the same values in created vectors. It is greater than 1 if unrolling is performed. @@ -4171,21 +4106,25 @@ two copies of each scalar operand: {s1, s2, s1, s2}. (NUMBER_OF_COPIES will be 2). - If GROUP_SIZE > NUNITS, the scalars will be split into several vectors - containing the operands. + If REDUC_GROUP_SIZE > NUNITS, the scalars will be split into several + vectors containing the operands. For example, NUNITS is four as before, and the group size is 8 (s1, s2, ..., s8). We will create two vectors {s1, s2, s3, s4} and {s5, s6, s7, s8}. */ + if (!TYPE_VECTOR_SUBPARTS (vector_type).is_constant (&nunits)) + nunits = group_size; + number_of_copies = nunits * number_of_vectors / group_size; number_of_places_left_in_vector = nunits; - auto_vec<tree, 32> elts (nunits); + bool constant_p = true; + tree_vector_builder elts (vector_type, nunits, 1); elts.quick_grow (nunits); for (j = 0; j < number_of_copies; j++) { - for (i = group_size - 1; stmts.iterate (i, &stmt); i--) + for (i = group_size - 1; stmts.iterate (i, &stmt_vinfo); i--) { tree op; /* Get the def before the loop. In reduction chain we have only @@ -4195,21 +4134,58 @@ && neutral_op) op = neutral_op; else - op = PHI_ARG_DEF_FROM_EDGE (stmt, pe); + op = PHI_ARG_DEF_FROM_EDGE (stmt_vinfo->stmt, pe); /* Create 'vect_ = {op0,op1,...,opn}'. */ number_of_places_left_in_vector--; elts[number_of_places_left_in_vector] = op; + if (!CONSTANT_CLASS_P (op)) + constant_p = false; if (number_of_places_left_in_vector == 0) { gimple_seq ctor_seq = NULL; - tree init = gimple_build_vector (&ctor_seq, vector_type, elts); + tree init; + if (constant_p && !neutral_op + ? multiple_p (TYPE_VECTOR_SUBPARTS (vector_type), nunits) + : known_eq (TYPE_VECTOR_SUBPARTS (vector_type), nunits)) + /* Build the vector directly from ELTS. */ + init = gimple_build_vector (&ctor_seq, &elts); + else if (neutral_op) + { + /* Build a vector of the neutral value and shift the + other elements into place. */ + init = gimple_build_vector_from_val (&ctor_seq, vector_type, + neutral_op); + int k = nunits; + while (k > 0 && elts[k - 1] == neutral_op) + k -= 1; + while (k > 0) + { + k -= 1; + init = gimple_build (&ctor_seq, CFN_VEC_SHL_INSERT, + vector_type, init, elts[k]); + } + } + else + { + /* First time round, duplicate ELTS to fill the + required number of vectors, then cherry pick the + appropriate result for each iteration. */ + if (vec_oprnds->is_empty ()) + duplicate_and_interleave (&ctor_seq, vector_type, elts, + number_of_vectors, + permute_results); + init = permute_results[number_of_vectors - j - 1]; + } if (ctor_seq != NULL) gsi_insert_seq_on_edge_immediate (pe, ctor_seq); voprnds.quick_push (init); number_of_places_left_in_vector = nunits; + elts.new_vector (vector_type, nunits, 1); + elts.quick_grow (nunits); + constant_p = true; } } } @@ -4260,20 +4236,27 @@ VECT_DEFS is list of vector of partial results, i.e., the lhs's of vector reduction statements. - STMT is the scalar reduction stmt that is being vectorized. + STMT_INFO is the scalar reduction stmt that is being vectorized. NCOPIES is > 1 in case the vectorization factor (VF) is bigger than the number of elements that we can fit in a vectype (nunits). In this case we have to generate more than one vector stmt - i.e - we need to "unroll" the vector stmt by a factor VF/nunits. For more details see documentation in vectorizable_operation. - REDUC_CODE is the tree-code for the epilog reduction. + REDUC_FN is the internal function for the epilog reduction. REDUCTION_PHIS is a list of the phi-nodes that carry the reduction computation. REDUC_INDEX is the index of the operand in the right hand side of the statement that is defined by REDUCTION_PHI. DOUBLE_REDUC is TRUE if double reduction phi nodes should be handled. SLP_NODE is an SLP node containing a group of reduction statements. The - first one in this group is STMT. + first one in this group is STMT_INFO. + INDUC_VAL is for INTEGER_INDUC_COND_REDUCTION the value to use for the case + when the COND_EXPR is never true in the loop. For MAX_EXPR, it needs to + be smaller than any value of the IV in the loop, for MIN_EXPR larger than + any value of the IV in the loop. + INDUC_CODE is the code for epilog reduction if INTEGER_INDUC_COND_REDUCTION. + NEUTRAL_OP is the value given by neutral_op_for_slp_reduction; it is + null if this is not an SLP reduction This function: 1. Creates the reduction def-use cycles: sets the arguments for @@ -4282,7 +4265,7 @@ The loop-latch argument is taken from VECT_DEFS - the vector of partial sums. 2. "Reduces" each vector of partial results VECT_DEFS into a single result, - by applying the operation specified by REDUC_CODE if available, or by + by calling the function specified by REDUC_FN if available, or by other means (whole-vector shifts or a scalar loop). The function also creates a new phi node at the loop exit to preserve loop-closed form, as illustrated below. @@ -4291,8 +4274,8 @@ loop: vec_def = phi <null, null> # REDUCTION_PHI - VECT_DEF = vector_stmt # vectorized form of STMT - s_loop = scalar_stmt # (scalar) STMT + VECT_DEF = vector_stmt # vectorized form of STMT_INFO + s_loop = scalar_stmt # (scalar) STMT_INFO loop_exit: s_out0 = phi <s_loop> # (scalar) EXIT_PHI use <s_out0> @@ -4302,8 +4285,8 @@ loop: vec_def = phi <vec_init, VECT_DEF> # REDUCTION_PHI - VECT_DEF = vector_stmt # vectorized form of STMT - s_loop = scalar_stmt # (scalar) STMT + VECT_DEF = vector_stmt # vectorized form of STMT_INFO + s_loop = scalar_stmt # (scalar) STMT_INFO loop_exit: s_out0 = phi <s_loop> # (scalar) EXIT_PHI v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI @@ -4315,15 +4298,17 @@ */ static void -vect_create_epilog_for_reduction (vec<tree> vect_defs, gimple *stmt, +vect_create_epilog_for_reduction (vec<tree> vect_defs, + stmt_vec_info stmt_info, gimple *reduc_def_stmt, - int ncopies, enum tree_code reduc_code, - vec<gimple *> reduction_phis, + int ncopies, internal_fn reduc_fn, + vec<stmt_vec_info> reduction_phis, bool double_reduc, slp_tree slp_node, - slp_instance slp_node_instance) + slp_instance slp_node_instance, + tree induc_val, enum tree_code induc_code, + tree neutral_op) { - stmt_vec_info stmt_info = vinfo_for_stmt (stmt); stmt_vec_info prev_phi_info; tree vectype; machine_mode mode; @@ -4333,11 +4318,12 @@ tree scalar_dest; tree scalar_type; gimple *new_phi = NULL, *phi; + stmt_vec_info phi_info; gimple_stmt_iterator exit_gsi; tree vec_dest; tree new_temp = NULL_TREE, new_dest, new_name, new_scalar_dest; gimple *epilog_stmt = NULL; - enum tree_code code = gimple_assign_rhs_code (stmt); + enum tree_code code = gimple_assign_rhs_code (stmt_info->stmt); gimple *exit_phi; tree bitsize; tree adjustment_def = NULL; @@ -4346,25 +4332,26 @@ tree orig_name, scalar_result; imm_use_iterator imm_iter, phi_imm_iter; use_operand_p use_p, phi_use_p; - gimple *use_stmt, *orig_stmt, *reduction_phi = NULL; + gimple *use_stmt; + stmt_vec_info reduction_phi_info = NULL; bool nested_in_vect_loop = false; auto_vec<gimple *> new_phis; - auto_vec<gimple *> inner_phis; - enum vect_def_type dt = vect_unknown_def_type; + auto_vec<stmt_vec_info> inner_phis; int j, i; auto_vec<tree> scalar_results; unsigned int group_size = 1, k, ratio; auto_vec<tree> vec_initial_defs; auto_vec<gimple *> phis; bool slp_reduc = false; + bool direct_slp_reduc; tree new_phi_result; - gimple *inner_phi = NULL; + stmt_vec_info inner_phi = NULL; tree induction_index = NULL_TREE; if (slp_node) group_size = SLP_TREE_SCALAR_STMTS (slp_node).length (); - if (nested_in_vect_loop_p (loop, stmt)) + if (nested_in_vect_loop_p (loop, stmt_info)) { outer_loop = loop; loop = loop->inner; @@ -4400,25 +4387,52 @@ unsigned vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); vec_initial_defs.reserve (vec_num); get_initial_defs_for_reduction (slp_node_instance->reduc_phis, - &vec_initial_defs, vec_num, code, - GROUP_FIRST_ELEMENT (stmt_info)); + &vec_initial_defs, vec_num, + REDUC_GROUP_FIRST_ELEMENT (stmt_info), + neutral_op); } else { /* Get at the scalar def before the loop, that defines the initial value of the reduction variable. */ - gimple *def_stmt; initial_def = PHI_ARG_DEF_FROM_EDGE (reduc_def_stmt, loop_preheader_edge (loop)); - vect_is_simple_use (initial_def, loop_vinfo, &def_stmt, &initial_def_dt); - vec_initial_def = get_initial_def_for_reduction (stmt, initial_def, - &adjustment_def); + /* Optimize: if initial_def is for REDUC_MAX smaller than the base + and we can't use zero for induc_val, use initial_def. Similarly + for REDUC_MIN and initial_def larger than the base. */ + if (TREE_CODE (initial_def) == INTEGER_CST + && (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) + == INTEGER_INDUC_COND_REDUCTION) + && !integer_zerop (induc_val) + && ((induc_code == MAX_EXPR + && tree_int_cst_lt (initial_def, induc_val)) + || (induc_code == MIN_EXPR + && tree_int_cst_lt (induc_val, initial_def)))) + induc_val = initial_def; + + if (double_reduc) + /* In case of double reduction we only create a vector variable + to be put in the reduction phi node. The actual statement + creation is done later in this function. */ + vec_initial_def = vect_create_destination_var (initial_def, vectype); + else if (nested_in_vect_loop) + { + /* Do not use an adjustment def as that case is not supported + correctly if ncopies is not one. */ + vect_is_simple_use (initial_def, loop_vinfo, &initial_def_dt); + vec_initial_def = vect_get_vec_def_for_operand (initial_def, + stmt_info); + } + else + vec_initial_def + = get_initial_def_for_reduction (stmt_info, initial_def, + &adjustment_def); vec_initial_defs.create (1); vec_initial_defs.quick_push (vec_initial_def); } /* Set phi nodes arguments. */ - FOR_EACH_VEC_ELT (reduction_phis, i, phi) + FOR_EACH_VEC_ELT (reduction_phis, i, phi_info) { tree vec_init_def = vec_initial_defs[i]; tree def = vect_defs[i]; @@ -4426,15 +4440,15 @@ { if (j != 0) { - phi = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (phi)); + phi_info = STMT_VINFO_RELATED_STMT (phi_info); if (nested_in_vect_loop) vec_init_def - = vect_get_vec_def_for_stmt_copy (initial_def_dt, - vec_init_def); + = vect_get_vec_def_for_stmt_copy (loop_vinfo, vec_init_def); } /* Set the loop-entry arg of the reduction-phi. */ + gphi *phi = as_a <gphi *> (phi_info->stmt); if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) == INTEGER_INDUC_COND_REDUCTION) { @@ -4444,29 +4458,26 @@ gcc_assert (i == 0); tree vec_init_def_type = TREE_TYPE (vec_init_def); - tree zero_vec = build_zero_cst (vec_init_def_type); - - add_phi_arg (as_a <gphi *> (phi), zero_vec, - loop_preheader_edge (loop), UNKNOWN_LOCATION); + tree induc_val_vec + = build_vector_from_val (vec_init_def_type, induc_val); + + add_phi_arg (phi, induc_val_vec, loop_preheader_edge (loop), + UNKNOWN_LOCATION); } else - add_phi_arg (as_a <gphi *> (phi), vec_init_def, - loop_preheader_edge (loop), UNKNOWN_LOCATION); + add_phi_arg (phi, vec_init_def, loop_preheader_edge (loop), + UNKNOWN_LOCATION); /* Set the loop-latch arg for the reduction-phi. */ if (j > 0) - def = vect_get_vec_def_for_stmt_copy (vect_unknown_def_type, def); - - add_phi_arg (as_a <gphi *> (phi), def, loop_latch_edge (loop), - UNKNOWN_LOCATION); + def = vect_get_vec_def_for_stmt_copy (loop_vinfo, def); + + add_phi_arg (phi, def, loop_latch_edge (loop), UNKNOWN_LOCATION); if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, - "transform reduction: created def-use cycle: "); - dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); - dump_gimple_stmt (MSG_NOTE, TDF_SLIM, SSA_NAME_DEF_STMT (def), 0); - } + dump_printf_loc (MSG_NOTE, vect_location, + "transform reduction: created def-use cycle: %G%G", + phi, SSA_NAME_DEF_STMT (def)); } } @@ -4480,10 +4491,9 @@ if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) == COND_REDUCTION) { tree indx_before_incr, indx_after_incr; - int nunits_out = TYPE_VECTOR_SUBPARTS (vectype); - int k; - - gimple *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); + poly_uint64 nunits_out = TYPE_VECTOR_SUBPARTS (vectype); + + gimple *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info)->stmt; gcc_assert (gimple_assign_rhs_code (vec_stmt) == VEC_COND_EXPR); int scalar_precision @@ -4497,10 +4507,7 @@ vector size (STEP). */ /* Create a {1,2,3,...} vector. */ - auto_vec<tree, 32> vtemp (nunits_out); - for (k = 0; k < nunits_out; ++k) - vtemp.quick_push (build_int_cst (cr_index_scalar_type, k + 1)); - tree series_vect = build_vector (cr_index_vector_type, vtemp); + tree series_vect = build_index_vector (cr_index_vector_type, 1, 1); /* Create a vector of the step value. */ tree step = build_int_cst (cr_index_scalar_type, nunits_out); @@ -4523,8 +4530,7 @@ /* Create a vector phi node. */ tree new_phi_tree = make_ssa_name (cr_index_vector_type); new_phi = create_phi_node (new_phi_tree, loop->header); - set_vinfo_for_stmt (new_phi, - new_stmt_vec_info (new_phi, loop_vinfo)); + loop_vinfo->add_stmt (new_phi); add_phi_arg (as_a <gphi *> (new_phi), vec_zero, loop_preheader_edge (loop), UNKNOWN_LOCATION); @@ -4549,10 +4555,8 @@ gimple *index_condition = gimple_build_assign (induction_index, index_cond_expr); gsi_insert_before (&incr_gsi, index_condition, GSI_SAME_STMT); - stmt_vec_info index_vec_info = new_stmt_vec_info (index_condition, - loop_vinfo); + stmt_vec_info index_vec_info = loop_vinfo->add_stmt (index_condition); STMT_VINFO_VECTYPE (index_vec_info) = cr_index_vector_type; - set_vinfo_for_stmt (index_condition, index_vec_info); /* Update the phi with the vec cond. */ add_phi_arg (as_a <gphi *> (new_phi), induction_index, @@ -4569,7 +4573,7 @@ step 3: adjust the scalar result (s_out3) if needed. Step 1 can be accomplished using one the following three schemes: - (scheme 1) using reduc_code, if available. + (scheme 1) using reduc_fn, if available. (scheme 2) using whole-vector shifts, if available. (scheme 3) using a scalar loop. In this case steps 1+2 above are combined. @@ -4599,17 +4603,17 @@ { tree new_def = copy_ssa_name (def); phi = create_phi_node (new_def, exit_bb); - set_vinfo_for_stmt (phi, new_stmt_vec_info (phi, loop_vinfo)); + stmt_vec_info phi_info = loop_vinfo->add_stmt (phi); if (j == 0) new_phis.quick_push (phi); else { - def = vect_get_vec_def_for_stmt_copy (dt, def); - STMT_VINFO_RELATED_STMT (prev_phi_info) = phi; + def = vect_get_vec_def_for_stmt_copy (loop_vinfo, def); + STMT_VINFO_RELATED_STMT (prev_phi_info) = phi_info; } SET_PHI_ARG_DEF (phi, single_exit (loop)->dest_idx, def); - prev_phi_info = vinfo_for_stmt (phi); + prev_phi_info = phi_info; } } @@ -4622,26 +4626,24 @@ inner_phis.create (vect_defs.length ()); FOR_EACH_VEC_ELT (new_phis, i, phi) { + stmt_vec_info phi_info = loop_vinfo->lookup_stmt (phi); tree new_result = copy_ssa_name (PHI_RESULT (phi)); gphi *outer_phi = create_phi_node (new_result, exit_bb); SET_PHI_ARG_DEF (outer_phi, single_exit (loop)->dest_idx, PHI_RESULT (phi)); - set_vinfo_for_stmt (outer_phi, new_stmt_vec_info (outer_phi, - loop_vinfo)); - inner_phis.quick_push (phi); + prev_phi_info = loop_vinfo->add_stmt (outer_phi); + inner_phis.quick_push (phi_info); new_phis[i] = outer_phi; - prev_phi_info = vinfo_for_stmt (outer_phi); - while (STMT_VINFO_RELATED_STMT (vinfo_for_stmt (phi))) + while (STMT_VINFO_RELATED_STMT (phi_info)) { - phi = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (phi)); - new_result = copy_ssa_name (PHI_RESULT (phi)); + phi_info = STMT_VINFO_RELATED_STMT (phi_info); + new_result = copy_ssa_name (PHI_RESULT (phi_info->stmt)); outer_phi = create_phi_node (new_result, exit_bb); SET_PHI_ARG_DEF (outer_phi, single_exit (loop)->dest_idx, - PHI_RESULT (phi)); - set_vinfo_for_stmt (outer_phi, new_stmt_vec_info (outer_phi, - loop_vinfo)); - STMT_VINFO_RELATED_STMT (prev_phi_info) = outer_phi; - prev_phi_info = vinfo_for_stmt (outer_phi); + PHI_RESULT (phi_info->stmt)); + stmt_vec_info outer_phi_info = loop_vinfo->add_stmt (outer_phi); + STMT_VINFO_RELATED_STMT (prev_phi_info) = outer_phi_info; + prev_phi_info = outer_phi_info; } } } @@ -4649,7 +4651,7 @@ exit_gsi = gsi_after_labels (exit_bb); /* 2.2 Get the relevant tree-code to use in the epilog for schemes 2,3 - (i.e. when reduc_code is not available) and in the final adjustment + (i.e. when reduc_fn is not available) and in the final adjustment code (if needed). Also get the original scalar reduction variable as defined in the loop. In case STMT is a "pattern-stmt" (i.e. - it represents a reduction pattern), the tree-code and scalar-def are @@ -4657,27 +4659,21 @@ Otherwise (it is a regular reduction) - the tree-code and scalar-def are taken from STMT. */ - orig_stmt = STMT_VINFO_RELATED_STMT (stmt_info); - if (!orig_stmt) - { - /* Regular reduction */ - orig_stmt = stmt; - } - else + stmt_vec_info orig_stmt_info = vect_orig_stmt (stmt_info); + if (orig_stmt_info != stmt_info) { /* Reduction pattern */ - stmt_vec_info stmt_vinfo = vinfo_for_stmt (orig_stmt); - gcc_assert (STMT_VINFO_IN_PATTERN_P (stmt_vinfo)); - gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo) == stmt); - } - - code = gimple_assign_rhs_code (orig_stmt); + gcc_assert (STMT_VINFO_IN_PATTERN_P (orig_stmt_info)); + gcc_assert (STMT_VINFO_RELATED_STMT (orig_stmt_info) == stmt_info); + } + + code = gimple_assign_rhs_code (orig_stmt_info->stmt); /* For MINUS_EXPR the initial vector is [init_val,0,...,0], therefore, partial results are added and not subtracted. */ if (code == MINUS_EXPR) code = PLUS_EXPR; - scalar_dest = gimple_assign_lhs (orig_stmt); + scalar_dest = gimple_assign_lhs (orig_stmt_info->stmt); scalar_type = TREE_TYPE (scalar_dest); scalar_results.create (group_size); new_scalar_dest = vect_create_destination_var (scalar_dest, NULL); @@ -4697,7 +4693,13 @@ # b1 = phi <b2, b0> a2 = operation (a1) b2 = operation (b1) */ - slp_reduc = (slp_node && !GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))); + slp_reduc = (slp_node && !REDUC_GROUP_FIRST_ELEMENT (stmt_info)); + + /* True if we should implement SLP_REDUC using native reduction operations + instead of scalar operations. */ + direct_slp_reduc = (reduc_fn != IFN_LAST + && slp_reduc + && !TYPE_VECTOR_SUBPARTS (vectype).is_constant ()); /* In case of reduction chain, e.g., # a1 = phi <a3, a0> @@ -4706,7 +4708,7 @@ we may end up with more than one vector result. Here we reduce them to one vector. */ - if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))) + if (REDUC_GROUP_FIRST_ELEMENT (stmt_info) || direct_slp_reduc) { tree first_vect = PHI_RESULT (new_phis[0]); gassign *new_vec_stmt = NULL; @@ -4736,11 +4738,11 @@ tree first_vect = PHI_RESULT (new_phis[0]); gassign *new_vec_stmt = NULL; vec_dest = vect_create_destination_var (scalar_dest, vectype); - gimple *next_phi = new_phis[0]; + stmt_vec_info next_phi_info = loop_vinfo->lookup_stmt (new_phis[0]); for (int k = 1; k < ncopies; ++k) { - next_phi = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (next_phi)); - tree second_vect = PHI_RESULT (next_phi); + next_phi_info = STMT_VINFO_RELATED_STMT (next_phi_info); + tree second_vect = PHI_RESULT (next_phi_info->stmt); tree tem = make_ssa_name (vec_dest, new_vec_stmt); new_vec_stmt = gimple_build_assign (tem, code, first_vect, second_vect); @@ -4755,7 +4757,7 @@ new_phi_result = PHI_RESULT (new_phis[0]); if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) == COND_REDUCTION - && reduc_code != ERROR_MARK) + && reduc_fn != IFN_LAST) { /* For condition reductions, we have a vector (NEW_PHI_RESULT) containing various data values where the condition matched and another vector @@ -4793,8 +4795,9 @@ /* Find maximum value from the vector of found indexes. */ tree max_index = make_ssa_name (index_scalar_type); - gimple *max_index_stmt = gimple_build_assign (max_index, REDUC_MAX_EXPR, - induction_index); + gcall *max_index_stmt = gimple_build_call_internal (IFN_REDUC_MAX, + 1, induction_index); + gimple_call_set_lhs (max_index_stmt, max_index); gsi_insert_before (&exit_gsi, max_index_stmt, GSI_SAME_STMT); /* Vector of {max_index, max_index, max_index,...}. */ @@ -4849,13 +4852,9 @@ /* Reduce down to a scalar value. */ tree data_reduc = make_ssa_name (scalar_type_unsigned); - optab ot = optab_for_tree_code (REDUC_MAX_EXPR, vectype_unsigned, - optab_default); - gcc_assert (optab_handler (ot, TYPE_MODE (vectype_unsigned)) - != CODE_FOR_nothing); - gimple *data_reduc_stmt = gimple_build_assign (data_reduc, - REDUC_MAX_EXPR, - vec_cond_cast); + gcall *data_reduc_stmt = gimple_build_call_internal (IFN_REDUC_MAX, + 1, vec_cond_cast); + gimple_call_set_lhs (data_reduc_stmt, data_reduc); gsi_insert_before (&exit_gsi, data_reduc_stmt, GSI_SAME_STMT); /* Convert the reduced value back to the result type and set as the @@ -4867,9 +4866,9 @@ scalar_results.safe_push (new_temp); } else if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) == COND_REDUCTION - && reduc_code == ERROR_MARK) - { - /* Condition redution without supported REDUC_MAX_EXPR. Generate + && reduc_fn == IFN_LAST) + { + /* Condition reduction without supported IFN_REDUC_MAX. Generate idx = 0; idx_val = induction_index[0]; val = data_reduc[0]; @@ -4881,8 +4880,11 @@ tree data_eltype = TREE_TYPE (TREE_TYPE (new_phi_result)); tree idx_eltype = TREE_TYPE (TREE_TYPE (induction_index)); unsigned HOST_WIDE_INT el_size = tree_to_uhwi (TYPE_SIZE (idx_eltype)); - unsigned HOST_WIDE_INT v_size - = el_size * TYPE_VECTOR_SUBPARTS (TREE_TYPE (induction_index)); + poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (TREE_TYPE (induction_index)); + /* Enforced by vectorizable_reduction, which ensures we have target + support before allowing a conditional reduction on variable-length + vectors. */ + unsigned HOST_WIDE_INT v_size = el_size * nunits.to_constant (); tree idx_val = NULL_TREE, val = NULL_TREE; for (unsigned HOST_WIDE_INT off = 0; off < v_size; off += el_size) { @@ -4939,7 +4941,7 @@ /* 2.3 Create the reduction code, using one of the three schemes described above. In SLP we simply need to extract all the elements from the vector (without reducing them), so we use scalar shifts. */ - else if (reduc_code != ERROR_MARK && !slp_reduc) + else if (reduc_fn != IFN_LAST && !slp_reduc) { tree tmp; tree vec_elem_type; @@ -4954,32 +4956,39 @@ vec_elem_type = TREE_TYPE (TREE_TYPE (new_phi_result)); if (!useless_type_conversion_p (scalar_type, vec_elem_type)) { - tree tmp_dest = - vect_create_destination_var (scalar_dest, vec_elem_type); - tmp = build1 (reduc_code, vec_elem_type, new_phi_result); - epilog_stmt = gimple_build_assign (tmp_dest, tmp); + tree tmp_dest + = vect_create_destination_var (scalar_dest, vec_elem_type); + epilog_stmt = gimple_build_call_internal (reduc_fn, 1, + new_phi_result); + gimple_set_lhs (epilog_stmt, tmp_dest); new_temp = make_ssa_name (tmp_dest, epilog_stmt); - gimple_assign_set_lhs (epilog_stmt, new_temp); + gimple_set_lhs (epilog_stmt, new_temp); gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); - tmp = build1 (NOP_EXPR, scalar_type, new_temp); + epilog_stmt = gimple_build_assign (new_scalar_dest, NOP_EXPR, + new_temp); } else - tmp = build1 (reduc_code, scalar_type, new_phi_result); - - epilog_stmt = gimple_build_assign (new_scalar_dest, tmp); + { + epilog_stmt = gimple_build_call_internal (reduc_fn, 1, + new_phi_result); + gimple_set_lhs (epilog_stmt, new_scalar_dest); + } + new_temp = make_ssa_name (new_scalar_dest, epilog_stmt); - gimple_assign_set_lhs (epilog_stmt, new_temp); + gimple_set_lhs (epilog_stmt, new_temp); gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); - if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) - == INTEGER_INDUC_COND_REDUCTION) - { - /* Earlier we set the initial value to be zero. Check the result - and if it is zero then replace with the original initial - value. */ - tree zero = build_zero_cst (scalar_type); - tree zcompare = build2 (EQ_EXPR, boolean_type_node, new_temp, zero); + if ((STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) + == INTEGER_INDUC_COND_REDUCTION) + && !operand_equal_p (initial_def, induc_val, 0)) + { + /* Earlier we set the initial value to be a vector if induc_val + values. Check the result and if it is induc_val then replace + with the original initial value, unless induc_val is + the same as initial_def already. */ + tree zcompare = build2 (EQ_EXPR, boolean_type_node, new_temp, + induc_val); tmp = make_ssa_name (new_scalar_dest); epilog_stmt = gimple_build_assign (tmp, COND_EXPR, zcompare, @@ -4990,40 +4999,220 @@ scalar_results.safe_push (new_temp); } + else if (direct_slp_reduc) + { + /* Here we create one vector for each of the REDUC_GROUP_SIZE results, + with the elements for other SLP statements replaced with the + neutral value. We can then do a normal reduction on each vector. */ + + /* Enforced by vectorizable_reduction. */ + gcc_assert (new_phis.length () == 1); + gcc_assert (pow2p_hwi (group_size)); + + slp_tree orig_phis_slp_node = slp_node_instance->reduc_phis; + vec<stmt_vec_info> orig_phis + = SLP_TREE_SCALAR_STMTS (orig_phis_slp_node); + gimple_seq seq = NULL; + + /* Build a vector {0, 1, 2, ...}, with the same number of elements + and the same element size as VECTYPE. */ + tree index = build_index_vector (vectype, 0, 1); + tree index_type = TREE_TYPE (index); + tree index_elt_type = TREE_TYPE (index_type); + tree mask_type = build_same_sized_truth_vector_type (index_type); + + /* Create a vector that, for each element, identifies which of + the REDUC_GROUP_SIZE results should use it. */ + tree index_mask = build_int_cst (index_elt_type, group_size - 1); + index = gimple_build (&seq, BIT_AND_EXPR, index_type, index, + build_vector_from_val (index_type, index_mask)); + + /* Get a neutral vector value. This is simply a splat of the neutral + scalar value if we have one, otherwise the initial scalar value + is itself a neutral value. */ + tree vector_identity = NULL_TREE; + if (neutral_op) + vector_identity = gimple_build_vector_from_val (&seq, vectype, + neutral_op); + for (unsigned int i = 0; i < group_size; ++i) + { + /* If there's no univeral neutral value, we can use the + initial scalar value from the original PHI. This is used + for MIN and MAX reduction, for example. */ + if (!neutral_op) + { + tree scalar_value + = PHI_ARG_DEF_FROM_EDGE (orig_phis[i]->stmt, + loop_preheader_edge (loop)); + vector_identity = gimple_build_vector_from_val (&seq, vectype, + scalar_value); + } + + /* Calculate the equivalent of: + + sel[j] = (index[j] == i); + + which selects the elements of NEW_PHI_RESULT that should + be included in the result. */ + tree compare_val = build_int_cst (index_elt_type, i); + compare_val = build_vector_from_val (index_type, compare_val); + tree sel = gimple_build (&seq, EQ_EXPR, mask_type, + index, compare_val); + + /* Calculate the equivalent of: + + vec = seq ? new_phi_result : vector_identity; + + VEC is now suitable for a full vector reduction. */ + tree vec = gimple_build (&seq, VEC_COND_EXPR, vectype, + sel, new_phi_result, vector_identity); + + /* Do the reduction and convert it to the appropriate type. */ + tree scalar = gimple_build (&seq, as_combined_fn (reduc_fn), + TREE_TYPE (vectype), vec); + scalar = gimple_convert (&seq, scalar_type, scalar); + scalar_results.safe_push (scalar); + } + gsi_insert_seq_before (&exit_gsi, seq, GSI_SAME_STMT); + } else { - bool reduce_with_shift = have_whole_vector_shift (mode); - int element_bitsize = tree_to_uhwi (bitsize); - int vec_size_in_bits = tree_to_uhwi (TYPE_SIZE (vectype)); + bool reduce_with_shift; tree vec_temp; - /* COND reductions all do the final reduction with MAX_EXPR. */ + /* COND reductions all do the final reduction with MAX_EXPR + or MIN_EXPR. */ if (code == COND_EXPR) - code = MAX_EXPR; - - /* Regardless of whether we have a whole vector shift, if we're - emulating the operation via tree-vect-generic, we don't want - to use it. Only the first round of the reduction is likely - to still be profitable via emulation. */ - /* ??? It might be better to emit a reduction tree code here, so that - tree-vect-generic can expand the first round via bit tricks. */ - if (!VECTOR_MODE_P (mode)) - reduce_with_shift = false; + { + if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) + == INTEGER_INDUC_COND_REDUCTION) + code = induc_code; + else if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) + == CONST_COND_REDUCTION) + code = STMT_VINFO_VEC_CONST_COND_REDUC_CODE (stmt_info); + else + code = MAX_EXPR; + } + + /* See if the target wants to do the final (shift) reduction + in a vector mode of smaller size and first reduce upper/lower + halves against each other. */ + enum machine_mode mode1 = mode; + tree vectype1 = vectype; + unsigned sz = tree_to_uhwi (TYPE_SIZE_UNIT (vectype)); + unsigned sz1 = sz; + if (!slp_reduc + && (mode1 = targetm.vectorize.split_reduction (mode)) != mode) + sz1 = GET_MODE_SIZE (mode1).to_constant (); + + vectype1 = get_vectype_for_scalar_type_and_size (scalar_type, sz1); + reduce_with_shift = have_whole_vector_shift (mode1); + if (!VECTOR_MODE_P (mode1)) + reduce_with_shift = false; else - { - optab optab = optab_for_tree_code (code, vectype, optab_default); - if (optab_handler (optab, mode) == CODE_FOR_nothing) - reduce_with_shift = false; - } + { + optab optab = optab_for_tree_code (code, vectype1, optab_default); + if (optab_handler (optab, mode1) == CODE_FOR_nothing) + reduce_with_shift = false; + } + + /* First reduce the vector to the desired vector size we should + do shift reduction on by combining upper and lower halves. */ + new_temp = new_phi_result; + while (sz > sz1) + { + gcc_assert (!slp_reduc); + sz /= 2; + vectype1 = get_vectype_for_scalar_type_and_size (scalar_type, sz); + + /* The target has to make sure we support lowpart/highpart + extraction, either via direct vector extract or through + an integer mode punning. */ + tree dst1, dst2; + if (convert_optab_handler (vec_extract_optab, + TYPE_MODE (TREE_TYPE (new_temp)), + TYPE_MODE (vectype1)) + != CODE_FOR_nothing) + { + /* Extract sub-vectors directly once vec_extract becomes + a conversion optab. */ + dst1 = make_ssa_name (vectype1); + epilog_stmt + = gimple_build_assign (dst1, BIT_FIELD_REF, + build3 (BIT_FIELD_REF, vectype1, + new_temp, TYPE_SIZE (vectype1), + bitsize_int (0))); + gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); + dst2 = make_ssa_name (vectype1); + epilog_stmt + = gimple_build_assign (dst2, BIT_FIELD_REF, + build3 (BIT_FIELD_REF, vectype1, + new_temp, TYPE_SIZE (vectype1), + bitsize_int (sz * BITS_PER_UNIT))); + gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); + } + else + { + /* Extract via punning to appropriately sized integer mode + vector. */ + tree eltype = build_nonstandard_integer_type (sz * BITS_PER_UNIT, + 1); + tree etype = build_vector_type (eltype, 2); + gcc_assert (convert_optab_handler (vec_extract_optab, + TYPE_MODE (etype), + TYPE_MODE (eltype)) + != CODE_FOR_nothing); + tree tem = make_ssa_name (etype); + epilog_stmt = gimple_build_assign (tem, VIEW_CONVERT_EXPR, + build1 (VIEW_CONVERT_EXPR, + etype, new_temp)); + gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); + new_temp = tem; + tem = make_ssa_name (eltype); + epilog_stmt + = gimple_build_assign (tem, BIT_FIELD_REF, + build3 (BIT_FIELD_REF, eltype, + new_temp, TYPE_SIZE (eltype), + bitsize_int (0))); + gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); + dst1 = make_ssa_name (vectype1); + epilog_stmt = gimple_build_assign (dst1, VIEW_CONVERT_EXPR, + build1 (VIEW_CONVERT_EXPR, + vectype1, tem)); + gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); + tem = make_ssa_name (eltype); + epilog_stmt + = gimple_build_assign (tem, BIT_FIELD_REF, + build3 (BIT_FIELD_REF, eltype, + new_temp, TYPE_SIZE (eltype), + bitsize_int (sz * BITS_PER_UNIT))); + gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); + dst2 = make_ssa_name (vectype1); + epilog_stmt = gimple_build_assign (dst2, VIEW_CONVERT_EXPR, + build1 (VIEW_CONVERT_EXPR, + vectype1, tem)); + gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); + } + + new_temp = make_ssa_name (vectype1); + epilog_stmt = gimple_build_assign (new_temp, code, dst1, dst2); + gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); + } if (reduce_with_shift && !slp_reduc) - { - int nelements = vec_size_in_bits / element_bitsize; - auto_vec_perm_indices sel (nelements); + { + int element_bitsize = tree_to_uhwi (bitsize); + /* Enforced by vectorizable_reduction, which disallows SLP reductions + for variable-length vectors and also requires direct target support + for loop reductions. */ + int vec_size_in_bits = tree_to_uhwi (TYPE_SIZE (vectype1)); + int nelements = vec_size_in_bits / element_bitsize; + vec_perm_builder sel; + vec_perm_indices indices; int elt_offset; - tree zero_vec = build_zero_cst (vectype); + tree zero_vec = build_zero_cst (vectype1); /* Case 2: Create: for (offset = nelements/2; offset >= 1; offset/=2) { @@ -5037,15 +5226,15 @@ dump_printf_loc (MSG_NOTE, vect_location, "Reduce using vector shifts\n"); - vec_dest = vect_create_destination_var (scalar_dest, vectype); - new_temp = new_phi_result; + mode1 = TYPE_MODE (vectype1); + vec_dest = vect_create_destination_var (scalar_dest, vectype1); for (elt_offset = nelements / 2; elt_offset >= 1; elt_offset /= 2) { - sel.truncate (0); calc_vec_perm_mask_for_shift (elt_offset, nelements, &sel); - tree mask = vect_gen_perm_mask_any (vectype, sel); + indices.new_vector (sel, 2, nelements); + tree mask = vect_gen_perm_mask_any (vectype1, indices); epilog_stmt = gimple_build_assign (vec_dest, VEC_PERM_EXPR, new_temp, zero_vec, mask); new_name = make_ssa_name (vec_dest, epilog_stmt); @@ -5090,7 +5279,8 @@ dump_printf_loc (MSG_NOTE, vect_location, "Reduce using scalar code.\n"); - vec_size_in_bits = tree_to_uhwi (TYPE_SIZE (vectype)); + int vec_size_in_bits = tree_to_uhwi (TYPE_SIZE (vectype1)); + int element_bitsize = tree_to_uhwi (bitsize); FOR_EACH_VEC_ELT (new_phis, i, new_phi) { int bit_offset; @@ -5099,7 +5289,7 @@ else vec_temp = gimple_assign_lhs (new_phi); tree rhs = build3 (BIT_FIELD_REF, scalar_type, vec_temp, bitsize, - bitsize_zero_node); + bitsize_zero_node); epilog_stmt = gimple_build_assign (new_scalar_dest, rhs); new_temp = make_ssa_name (new_scalar_dest, epilog_stmt); gimple_assign_set_lhs (epilog_stmt, new_temp); @@ -5143,8 +5333,8 @@ /* The only case where we need to reduce scalar results in SLP, is unrolling. If the size of SCALAR_RESULTS is greater than - GROUP_SIZE, we reduce them combining elements modulo - GROUP_SIZE. */ + REDUC_GROUP_SIZE, we reduce them combining elements modulo + REDUC_GROUP_SIZE. */ if (slp_reduc) { tree res, first_res, new_res; @@ -5168,14 +5358,16 @@ scalar_results.safe_push (new_temp); } - if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) - == INTEGER_INDUC_COND_REDUCTION) - { - /* Earlier we set the initial value to be zero. Check the result - and if it is zero then replace with the original initial - value. */ - tree zero = build_zero_cst (scalar_type); - tree zcompare = build2 (EQ_EXPR, boolean_type_node, new_temp, zero); + if ((STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) + == INTEGER_INDUC_COND_REDUCTION) + && !operand_equal_p (initial_def, induc_val, 0)) + { + /* Earlier we set the initial value to be a vector if induc_val + values. Check the result and if it is induc_val then replace + with the original initial value, unless induc_val is + the same as initial_def already. */ + tree zcompare = build2 (EQ_EXPR, boolean_type_node, new_temp, + induc_val); tree tmp = make_ssa_name (new_scalar_dest); epilog_stmt = gimple_build_assign (tmp, COND_EXPR, zcompare, @@ -5219,10 +5411,9 @@ gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); if (nested_in_vect_loop) { - set_vinfo_for_stmt (epilog_stmt, - new_stmt_vec_info (epilog_stmt, loop_vinfo)); - STMT_VINFO_RELATED_STMT (vinfo_for_stmt (epilog_stmt)) = - STMT_VINFO_RELATED_STMT (vinfo_for_stmt (new_phi)); + stmt_vec_info epilog_stmt_info = loop_vinfo->add_stmt (epilog_stmt); + STMT_VINFO_RELATED_STMT (epilog_stmt_info) + = STMT_VINFO_RELATED_STMT (loop_vinfo->lookup_stmt (new_phi)); if (!double_reduc) scalar_results.quick_push (new_temp); @@ -5262,26 +5453,23 @@ /* In SLP reduction chain we reduce vector results into one vector if - necessary, hence we set here GROUP_SIZE to 1. SCALAR_DEST is the LHS of - the last stmt in the reduction chain, since we are looking for the loop - exit phi node. */ - if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))) - { - gimple *dest_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[group_size - 1]; - /* Handle reduction patterns. */ - if (STMT_VINFO_RELATED_STMT (vinfo_for_stmt (dest_stmt))) - dest_stmt = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (dest_stmt)); - - scalar_dest = gimple_assign_lhs (dest_stmt); + necessary, hence we set here REDUC_GROUP_SIZE to 1. SCALAR_DEST is the + LHS of the last stmt in the reduction chain, since we are looking for + the loop exit phi node. */ + if (REDUC_GROUP_FIRST_ELEMENT (stmt_info)) + { + stmt_vec_info dest_stmt_info + = vect_orig_stmt (SLP_TREE_SCALAR_STMTS (slp_node)[group_size - 1]); + scalar_dest = gimple_assign_lhs (dest_stmt_info->stmt); group_size = 1; } - /* In SLP we may have several statements in NEW_PHIS and REDUCTION_PHIS (in - case that GROUP_SIZE is greater than vectorization factor). Therefore, we - need to match SCALAR_RESULTS with corresponding statements. The first - (GROUP_SIZE / number of new vector stmts) scalar results correspond to - the first vector stmt, etc. - (RATIO is equal to (GROUP_SIZE / number of new vector stmts)). */ + /* In SLP we may have several statements in NEW_PHIS and REDUCTION_PHIS (in + case that REDUC_GROUP_SIZE is greater than vectorization factor). + Therefore, we need to match SCALAR_RESULTS with corresponding statements. + The first (REDUC_GROUP_SIZE / number of new vector stmts) scalar results + correspond to the first vector stmt, etc. + (RATIO is equal to (REDUC_GROUP_SIZE / number of new vector stmts)). */ if (group_size > new_phis.length ()) { ratio = group_size / new_phis.length (); @@ -5290,24 +5478,25 @@ else ratio = 1; + stmt_vec_info epilog_stmt_info = NULL; for (k = 0; k < group_size; k++) { if (k % ratio == 0) { - epilog_stmt = new_phis[k / ratio]; - reduction_phi = reduction_phis[k / ratio]; + epilog_stmt_info = loop_vinfo->lookup_stmt (new_phis[k / ratio]); + reduction_phi_info = reduction_phis[k / ratio]; if (double_reduc) inner_phi = inner_phis[k / ratio]; } if (slp_reduc) { - gimple *current_stmt = SLP_TREE_SCALAR_STMTS (slp_node)[k]; - - orig_stmt = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (current_stmt)); - /* SLP statements can't participate in patterns. */ - gcc_assert (!orig_stmt); - scalar_dest = gimple_assign_lhs (current_stmt); + stmt_vec_info scalar_stmt_info = SLP_TREE_SCALAR_STMTS (slp_node)[k]; + + orig_stmt_info = STMT_VINFO_RELATED_STMT (scalar_stmt_info); + /* SLP statements can't participate in patterns. */ + gcc_assert (!orig_stmt_info); + scalar_dest = gimple_assign_lhs (scalar_stmt_info->stmt); } phis.create (3); @@ -5326,7 +5515,8 @@ { if (outer_loop) { - stmt_vec_info exit_phi_vinfo = vinfo_for_stmt (exit_phi); + stmt_vec_info exit_phi_vinfo + = loop_vinfo->lookup_stmt (exit_phi); gphi *vect_phi; /* FORNOW. Currently not supporting the case that an inner-loop @@ -5339,7 +5529,7 @@ if (double_reduc) STMT_VINFO_VEC_STMT (exit_phi_vinfo) = inner_phi; else - STMT_VINFO_VEC_STMT (exit_phi_vinfo) = epilog_stmt; + STMT_VINFO_VEC_STMT (exit_phi_vinfo) = epilog_stmt_info; if (!double_reduc || STMT_VINFO_DEF_TYPE (exit_phi_vinfo) != vect_double_reduction_def) @@ -5363,10 +5553,8 @@ FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, orig_name) { stmt_vec_info use_stmt_vinfo; - stmt_vec_info new_phi_vinfo; tree vect_phi_init, preheader_arg, vect_phi_res; basic_block bb = gimple_bb (use_stmt); - gimple *use; /* Check that USE_STMT is really double reduction phi node. */ @@ -5374,7 +5562,7 @@ || gimple_phi_num_args (use_stmt) != 2 || bb->loop_father != outer_loop) continue; - use_stmt_vinfo = vinfo_for_stmt (use_stmt); + use_stmt_vinfo = loop_vinfo->lookup_stmt (use_stmt); if (!use_stmt_vinfo || STMT_VINFO_DEF_TYPE (use_stmt_vinfo) != vect_double_reduction_def) @@ -5390,41 +5578,38 @@ /* Create vector phi node. */ vect_phi = create_phi_node (vec_initial_def, bb); - new_phi_vinfo = new_stmt_vec_info (vect_phi, - loop_vec_info_for_loop (outer_loop)); - set_vinfo_for_stmt (vect_phi, new_phi_vinfo); + loop_vec_info_for_loop (outer_loop)->add_stmt (vect_phi); /* Create vs0 - initial def of the double reduction phi. */ preheader_arg = PHI_ARG_DEF_FROM_EDGE (use_stmt, loop_preheader_edge (outer_loop)); vect_phi_init = get_initial_def_for_reduction - (stmt, preheader_arg, NULL); + (stmt_info, preheader_arg, NULL); /* Update phi node arguments with vs0 and vs2. */ add_phi_arg (vect_phi, vect_phi_init, loop_preheader_edge (outer_loop), UNKNOWN_LOCATION); - add_phi_arg (vect_phi, PHI_RESULT (inner_phi), - loop_latch_edge (outer_loop), UNKNOWN_LOCATION); + add_phi_arg (vect_phi, PHI_RESULT (inner_phi->stmt), + loop_latch_edge (outer_loop), UNKNOWN_LOCATION); if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, - "created double reduction phi node: "); - dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vect_phi, 0); - } + dump_printf_loc (MSG_NOTE, vect_location, + "created double reduction phi node: %G", + vect_phi); vect_phi_res = PHI_RESULT (vect_phi); /* Replace the use, i.e., set the correct vs1 in the regular reduction phi node. FORNOW, NCOPIES is always 1, so the loop is redundant. */ - use = reduction_phi; - for (j = 0; j < ncopies; j++) - { - edge pr_edge = loop_preheader_edge (loop); - SET_PHI_ARG_DEF (use, pr_edge->dest_idx, vect_phi_res); - use = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (use)); - } + stmt_vec_info use_info = reduction_phi_info; + for (j = 0; j < ncopies; j++) + { + edge pr_edge = loop_preheader_edge (loop); + SET_PHI_ARG_DEF (as_a <gphi *> (use_info->stmt), + pr_edge->dest_idx, vect_phi_res); + use_info = STMT_VINFO_RELATED_STMT (use_info); + } } } } @@ -5481,31 +5666,222 @@ } } +/* Return a vector of type VECTYPE that is equal to the vector select + operation "MASK ? VEC : IDENTITY". Insert the select statements + before GSI. */ + +static tree +merge_with_identity (gimple_stmt_iterator *gsi, tree mask, tree vectype, + tree vec, tree identity) +{ + tree cond = make_temp_ssa_name (vectype, NULL, "cond"); + gimple *new_stmt = gimple_build_assign (cond, VEC_COND_EXPR, + mask, vec, identity); + gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT); + return cond; +} + +/* Successively apply CODE to each element of VECTOR_RHS, in left-to-right + order, starting with LHS. Insert the extraction statements before GSI and + associate the new scalar SSA names with variable SCALAR_DEST. + Return the SSA name for the result. */ + +static tree +vect_expand_fold_left (gimple_stmt_iterator *gsi, tree scalar_dest, + tree_code code, tree lhs, tree vector_rhs) +{ + tree vectype = TREE_TYPE (vector_rhs); + tree scalar_type = TREE_TYPE (vectype); + tree bitsize = TYPE_SIZE (scalar_type); + unsigned HOST_WIDE_INT vec_size_in_bits = tree_to_uhwi (TYPE_SIZE (vectype)); + unsigned HOST_WIDE_INT element_bitsize = tree_to_uhwi (bitsize); + + for (unsigned HOST_WIDE_INT bit_offset = 0; + bit_offset < vec_size_in_bits; + bit_offset += element_bitsize) + { + tree bitpos = bitsize_int (bit_offset); + tree rhs = build3 (BIT_FIELD_REF, scalar_type, vector_rhs, + bitsize, bitpos); + + gassign *stmt = gimple_build_assign (scalar_dest, rhs); + rhs = make_ssa_name (scalar_dest, stmt); + gimple_assign_set_lhs (stmt, rhs); + gsi_insert_before (gsi, stmt, GSI_SAME_STMT); + + stmt = gimple_build_assign (scalar_dest, code, lhs, rhs); + tree new_name = make_ssa_name (scalar_dest, stmt); + gimple_assign_set_lhs (stmt, new_name); + gsi_insert_before (gsi, stmt, GSI_SAME_STMT); + lhs = new_name; + } + return lhs; +} + +/* Perform an in-order reduction (FOLD_LEFT_REDUCTION). STMT_INFO is the + statement that sets the live-out value. REDUC_DEF_STMT is the phi + statement. CODE is the operation performed by STMT_INFO and OPS are + its scalar operands. REDUC_INDEX is the index of the operand in + OPS that is set by REDUC_DEF_STMT. REDUC_FN is the function that + implements in-order reduction, or IFN_LAST if we should open-code it. + VECTYPE_IN is the type of the vector input. MASKS specifies the masks + that should be used to control the operation in a fully-masked loop. */ + +static bool +vectorize_fold_left_reduction (stmt_vec_info stmt_info, + gimple_stmt_iterator *gsi, + stmt_vec_info *vec_stmt, slp_tree slp_node, + gimple *reduc_def_stmt, + tree_code code, internal_fn reduc_fn, + tree ops[3], tree vectype_in, + int reduc_index, vec_loop_masks *masks) +{ + loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); + struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); + tree vectype_out = STMT_VINFO_VECTYPE (stmt_info); + stmt_vec_info new_stmt_info = NULL; + + int ncopies; + if (slp_node) + ncopies = 1; + else + ncopies = vect_get_num_copies (loop_vinfo, vectype_in); + + gcc_assert (!nested_in_vect_loop_p (loop, stmt_info)); + gcc_assert (ncopies == 1); + gcc_assert (TREE_CODE_LENGTH (code) == binary_op); + gcc_assert (reduc_index == (code == MINUS_EXPR ? 0 : 1)); + gcc_assert (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) + == FOLD_LEFT_REDUCTION); + + if (slp_node) + gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (vectype_out), + TYPE_VECTOR_SUBPARTS (vectype_in))); + + tree op0 = ops[1 - reduc_index]; + + int group_size = 1; + stmt_vec_info scalar_dest_def_info; + auto_vec<tree> vec_oprnds0; + if (slp_node) + { + vect_get_vec_defs (op0, NULL_TREE, stmt_info, &vec_oprnds0, NULL, + slp_node); + group_size = SLP_TREE_SCALAR_STMTS (slp_node).length (); + scalar_dest_def_info = SLP_TREE_SCALAR_STMTS (slp_node)[group_size - 1]; + } + else + { + tree loop_vec_def0 = vect_get_vec_def_for_operand (op0, stmt_info); + vec_oprnds0.create (1); + vec_oprnds0.quick_push (loop_vec_def0); + scalar_dest_def_info = stmt_info; + } + + tree scalar_dest = gimple_assign_lhs (scalar_dest_def_info->stmt); + tree scalar_type = TREE_TYPE (scalar_dest); + tree reduc_var = gimple_phi_result (reduc_def_stmt); + + int vec_num = vec_oprnds0.length (); + gcc_assert (vec_num == 1 || slp_node); + tree vec_elem_type = TREE_TYPE (vectype_out); + gcc_checking_assert (useless_type_conversion_p (scalar_type, vec_elem_type)); + + tree vector_identity = NULL_TREE; + if (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) + vector_identity = build_zero_cst (vectype_out); + + tree scalar_dest_var = vect_create_destination_var (scalar_dest, NULL); + int i; + tree def0; + FOR_EACH_VEC_ELT (vec_oprnds0, i, def0) + { + gimple *new_stmt; + tree mask = NULL_TREE; + if (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) + mask = vect_get_loop_mask (gsi, masks, vec_num, vectype_in, i); + + /* Handle MINUS by adding the negative. */ + if (reduc_fn != IFN_LAST && code == MINUS_EXPR) + { + tree negated = make_ssa_name (vectype_out); + new_stmt = gimple_build_assign (negated, NEGATE_EXPR, def0); + gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT); + def0 = negated; + } + + if (mask) + def0 = merge_with_identity (gsi, mask, vectype_out, def0, + vector_identity); + + /* On the first iteration the input is simply the scalar phi + result, and for subsequent iterations it is the output of + the preceding operation. */ + if (reduc_fn != IFN_LAST) + { + new_stmt = gimple_build_call_internal (reduc_fn, 2, reduc_var, def0); + /* For chained SLP reductions the output of the previous reduction + operation serves as the input of the next. For the final statement + the output cannot be a temporary - we reuse the original + scalar destination of the last statement. */ + if (i != vec_num - 1) + { + gimple_set_lhs (new_stmt, scalar_dest_var); + reduc_var = make_ssa_name (scalar_dest_var, new_stmt); + gimple_set_lhs (new_stmt, reduc_var); + } + } + else + { + reduc_var = vect_expand_fold_left (gsi, scalar_dest_var, code, + reduc_var, def0); + new_stmt = SSA_NAME_DEF_STMT (reduc_var); + /* Remove the statement, so that we can use the same code paths + as for statements that we've just created. */ + gimple_stmt_iterator tmp_gsi = gsi_for_stmt (new_stmt); + gsi_remove (&tmp_gsi, false); + } + + if (i == vec_num - 1) + { + gimple_set_lhs (new_stmt, scalar_dest); + new_stmt_info = vect_finish_replace_stmt (scalar_dest_def_info, + new_stmt); + } + else + new_stmt_info = vect_finish_stmt_generation (scalar_dest_def_info, + new_stmt, gsi); + + if (slp_node) + SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); + } + + if (!slp_node) + STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; + + return true; +} /* Function is_nonwrapping_integer_induction. - Check if STMT (which is part of loop LOOP) both increments and + Check if STMT_VINO (which is part of loop LOOP) both increments and does not cause overflow. */ static bool -is_nonwrapping_integer_induction (gimple *stmt, struct loop *loop) +is_nonwrapping_integer_induction (stmt_vec_info stmt_vinfo, struct loop *loop) { - stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); + gphi *phi = as_a <gphi *> (stmt_vinfo->stmt); tree base = STMT_VINFO_LOOP_PHI_EVOLUTION_BASE_UNCHANGED (stmt_vinfo); tree step = STMT_VINFO_LOOP_PHI_EVOLUTION_PART (stmt_vinfo); - tree lhs_type = TREE_TYPE (gimple_phi_result (stmt)); + tree lhs_type = TREE_TYPE (gimple_phi_result (phi)); widest_int ni, max_loop_value, lhs_max; - bool overflow = false; + wi::overflow_type overflow = wi::OVF_NONE; /* Make sure the loop is integer based. */ if (TREE_CODE (base) != INTEGER_CST || TREE_CODE (step) != INTEGER_CST) return false; - /* Check that the induction increments. */ - if (tree_int_cst_sgn (step) == -1) - return false; - /* Check that the max size of the loop will not wrap. */ if (TYPE_OVERFLOW_UNDEFINED (lhs_type)) @@ -5530,17 +5906,18 @@ /* Function vectorizable_reduction. - Check if STMT performs a reduction operation that can be vectorized. - If VEC_STMT is also passed, vectorize the STMT: create a vectorized + Check if STMT_INFO performs a reduction operation that can be vectorized. + If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized stmt to replace it, put it in VEC_STMT, and insert it at GSI. - Return FALSE if not a vectorizable STMT, TRUE otherwise. + Return true if STMT_INFO is vectorizable in this way. This function also handles reduction idioms (patterns) that have been - recognized in advance during vect_pattern_recog. In this case, STMT may be - of this form: + recognized in advance during vect_pattern_recog. In this case, STMT_INFO + may be of this form: X = pattern_expr (arg0, arg1, ..., X) - and it's STMT_VINFO_RELATED_STMT points to the last stmt in the original - sequence that had been detected and replaced by the pattern-stmt (STMT). + and its STMT_VINFO_RELATED_STMT points to the last stmt in the original + sequence that had been detected and replaced by the pattern-stmt + (STMT_INFO). This function also handles reduction of condition expressions, for example: for (int i = 0; i < N; i++) @@ -5552,9 +5929,9 @@ index into the vector of results. In some cases of reduction patterns, the type of the reduction variable X is - different than the type of the other arguments of STMT. - In such cases, the vectype that is used when transforming STMT into a vector - stmt is different than the vectype that is used to determine the + different than the type of the other arguments of STMT_INFO. + In such cases, the vectype that is used when transforming STMT_INFO into + a vector stmt is different than the vectype that is used to determine the vectorization factor, because it consists of a different number of elements than the actual number of elements that are being operated upon in parallel. @@ -5578,78 +5955,70 @@ does *NOT* necessarily hold for reduction patterns. */ bool -vectorizable_reduction (gimple *stmt, gimple_stmt_iterator *gsi, - gimple **vec_stmt, slp_tree slp_node, - slp_instance slp_node_instance) +vectorizable_reduction (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, + stmt_vec_info *vec_stmt, slp_tree slp_node, + slp_instance slp_node_instance, + stmt_vector_for_cost *cost_vec) { tree vec_dest; tree scalar_dest; - stmt_vec_info stmt_info = vinfo_for_stmt (stmt); tree vectype_out = STMT_VINFO_VECTYPE (stmt_info); tree vectype_in = NULL_TREE; loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); - enum tree_code code, orig_code, epilog_reduc_code; + enum tree_code code, orig_code; + internal_fn reduc_fn; machine_mode vec_mode; int op_type; - optab optab, reduc_optab; + optab optab; tree new_temp = NULL_TREE; - gimple *def_stmt; enum vect_def_type dt, cond_reduc_dt = vect_unknown_def_type; + stmt_vec_info cond_stmt_vinfo = NULL; + enum tree_code cond_reduc_op_code = ERROR_MARK; tree scalar_type; bool is_simple_use; - gimple *orig_stmt; - stmt_vec_info orig_stmt_info = NULL; int i; int ncopies; int epilog_copies; stmt_vec_info prev_stmt_info, prev_phi_info; bool single_defuse_cycle = false; - gimple *new_stmt = NULL; + stmt_vec_info new_stmt_info = NULL; int j; tree ops[3]; enum vect_def_type dts[3]; bool nested_cycle = false, found_nested_cycle_def = false; bool double_reduc = false; basic_block def_bb; - struct loop * def_stmt_loop, *outer_loop = NULL; + struct loop * def_stmt_loop; tree def_arg; - gimple *def_arg_stmt; auto_vec<tree> vec_oprnds0; auto_vec<tree> vec_oprnds1; auto_vec<tree> vec_oprnds2; auto_vec<tree> vect_defs; - auto_vec<gimple *> phis; + auto_vec<stmt_vec_info> phis; int vec_num; tree def0, tem; - bool first_p = true; tree cr_index_scalar_type = NULL_TREE, cr_index_vector_type = NULL_TREE; tree cond_reduc_val = NULL_TREE; /* Make sure it was already recognized as a reduction computation. */ - if (STMT_VINFO_DEF_TYPE (vinfo_for_stmt (stmt)) != vect_reduction_def - && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (stmt)) != vect_nested_cycle) + if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_reduction_def + && STMT_VINFO_DEF_TYPE (stmt_info) != vect_nested_cycle) return false; - if (nested_in_vect_loop_p (loop, stmt)) - { - outer_loop = loop; + if (nested_in_vect_loop_p (loop, stmt_info)) + { loop = loop->inner; nested_cycle = true; } - /* In case of reduction chain we switch to the first stmt in the chain, but - we don't update STMT_INFO, since only the last stmt is marked as reduction - and has reduction properties. */ - if (GROUP_FIRST_ELEMENT (stmt_info) - && GROUP_FIRST_ELEMENT (stmt_info) != stmt) - { - stmt = GROUP_FIRST_ELEMENT (stmt_info); - first_p = false; - } - - if (gimple_code (stmt) == GIMPLE_PHI) - { + if (REDUC_GROUP_FIRST_ELEMENT (stmt_info)) + gcc_assert (slp_node + && REDUC_GROUP_FIRST_ELEMENT (stmt_info) == stmt_info); + + if (gphi *phi = dyn_cast <gphi *> (stmt_info->stmt)) + { + tree phi_result = gimple_phi_result (phi); /* Analysis is fully done on the reduction stmt invocation. */ if (! vec_stmt) { @@ -5660,23 +6029,33 @@ return true; } - gimple *reduc_stmt = STMT_VINFO_REDUC_DEF (stmt_info); - if (STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (reduc_stmt))) - reduc_stmt = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (reduc_stmt)); - - gcc_assert (is_gimple_assign (reduc_stmt)); + if (STMT_VINFO_REDUC_TYPE (stmt_info) == FOLD_LEFT_REDUCTION) + /* Leave the scalar phi in place. Note that checking + STMT_VINFO_VEC_REDUCTION_TYPE (as below) only works + for reductions involving a single statement. */ + return true; + + stmt_vec_info reduc_stmt_info = STMT_VINFO_REDUC_DEF (stmt_info); + reduc_stmt_info = vect_stmt_to_vectorize (reduc_stmt_info); + + if (STMT_VINFO_VEC_REDUCTION_TYPE (reduc_stmt_info) + == EXTRACT_LAST_REDUCTION) + /* Leave the scalar phi in place. */ + return true; + + gassign *reduc_stmt = as_a <gassign *> (reduc_stmt_info->stmt); for (unsigned k = 1; k < gimple_num_ops (reduc_stmt); ++k) { tree op = gimple_op (reduc_stmt, k); - if (op == gimple_phi_result (stmt)) + if (op == phi_result) continue; if (k == 1 && gimple_assign_rhs_code (reduc_stmt) == COND_EXPR) continue; - tem = get_vectype_for_scalar_type (TREE_TYPE (op)); - if (! vectype_in - || TYPE_VECTOR_SUBPARTS (tem) < TYPE_VECTOR_SUBPARTS (vectype_in)) - vectype_in = tem; + if (!vectype_in + || (GET_MODE_SIZE (SCALAR_TYPE_MODE (TREE_TYPE (vectype_in))) + < GET_MODE_SIZE (SCALAR_TYPE_MODE (TREE_TYPE (op))))) + vectype_in = get_vectype_for_scalar_type (TREE_TYPE (op)); break; } gcc_assert (vectype_in); @@ -5686,15 +6065,11 @@ else ncopies = vect_get_num_copies (loop_vinfo, vectype_in); - use_operand_p use_p; - gimple *use_stmt; + stmt_vec_info use_stmt_info; if (ncopies > 1 - && (STMT_VINFO_RELEVANT (vinfo_for_stmt (reduc_stmt)) - <= vect_used_only_live) - && single_imm_use (gimple_phi_result (stmt), &use_p, &use_stmt) - && (use_stmt == reduc_stmt - || (STMT_VINFO_RELATED_STMT (vinfo_for_stmt (use_stmt)) - == reduc_stmt))) + && STMT_VINFO_RELEVANT (reduc_stmt_info) <= vect_used_only_live + && (use_stmt_info = loop_vinfo->lookup_single_use (phi_result)) + && vect_stmt_to_vectorize (use_stmt_info) == reduc_stmt_info) single_defuse_cycle = true; /* Create the destination vector */ @@ -5703,9 +6078,10 @@ if (slp_node) /* The size vect_schedule_slp_instance computes is off for us. */ - vec_num = ((LOOP_VINFO_VECT_FACTOR (loop_vinfo) - * SLP_TREE_SCALAR_STMTS (slp_node).length ()) - / TYPE_VECTOR_SUBPARTS (vectype_in)); + vec_num = vect_get_num_vectors + (LOOP_VINFO_VECT_FACTOR (loop_vinfo) + * SLP_TREE_SCALAR_STMTS (slp_node).length (), + vectype_in); else vec_num = 1; @@ -5720,18 +6096,18 @@ /* Create the reduction-phi that defines the reduction operand. */ gimple *new_phi = create_phi_node (vec_dest, loop->header); - set_vinfo_for_stmt (new_phi, - new_stmt_vec_info (new_phi, loop_vinfo)); + stmt_vec_info new_phi_info = loop_vinfo->add_stmt (new_phi); if (slp_node) - SLP_TREE_VEC_STMTS (slp_node).quick_push (new_phi); + SLP_TREE_VEC_STMTS (slp_node).quick_push (new_phi_info); else { if (j == 0) - STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_phi; + STMT_VINFO_VEC_STMT (stmt_info) + = *vec_stmt = new_phi_info; else - STMT_VINFO_RELATED_STMT (prev_phi_info) = new_phi; - prev_phi_info = vinfo_for_stmt (new_phi); + STMT_VINFO_RELATED_STMT (prev_phi_info) = new_phi_info; + prev_phi_info = new_phi_info; } } } @@ -5744,7 +6120,7 @@ /* Not supportable if the reduction variable is used in the loop, unless it's a reduction chain. */ if (STMT_VINFO_RELEVANT (stmt_info) > vect_used_in_outer - && !GROUP_FIRST_ELEMENT (stmt_info)) + && !REDUC_GROUP_FIRST_ELEMENT (stmt_info)) return false; /* Reductions that are not used even in an enclosing outer-loop, @@ -5760,10 +6136,9 @@ the STMT_VINFO_RELATED_STMT field records the last stmt in the original sequence that constitutes the pattern. */ - orig_stmt = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (stmt)); - if (orig_stmt) - { - orig_stmt_info = vinfo_for_stmt (orig_stmt); + stmt_vec_info orig_stmt_info = STMT_VINFO_RELATED_STMT (stmt_info); + if (orig_stmt_info) + { gcc_assert (STMT_VINFO_IN_PATTERN_P (orig_stmt_info)); gcc_assert (!STMT_VINFO_IN_PATTERN_P (stmt_info)); } @@ -5772,7 +6147,7 @@ inside the loop body. The last operand is the reduction variable, which is defined by the loop-header-phi. */ - gcc_assert (is_gimple_assign (stmt)); + gassign *stmt = as_a <gassign *> (stmt_info->stmt); /* Flatten RHS. */ switch (get_gimple_rhs_class (gimple_assign_rhs_code (stmt))) @@ -5818,7 +6193,7 @@ The last use is the reduction variable. In case of nested cycle this assumption is not true: we use reduc_index to record the index of the reduction variable. */ - gimple *reduc_def_stmt = NULL; + stmt_vec_info reduc_def_info = NULL; int reduc_index = -1; for (i = 0; i < op_type; i++) { @@ -5826,13 +6201,14 @@ if (i == 0 && code == COND_EXPR) continue; - is_simple_use = vect_is_simple_use (ops[i], loop_vinfo, - &def_stmt, &dts[i], &tem); + stmt_vec_info def_stmt_info; + is_simple_use = vect_is_simple_use (ops[i], loop_vinfo, &dts[i], &tem, + &def_stmt_info); dt = dts[i]; gcc_assert (is_simple_use); if (dt == vect_reduction_def) { - reduc_def_stmt = def_stmt; + reduc_def_info = def_stmt_info; reduc_index = i; continue; } @@ -5841,7 +6217,8 @@ /* To properly compute ncopies we are interested in the widest input type in case we're looking at a widening accumulation. */ if (!vectype_in - || TYPE_VECTOR_SUBPARTS (vectype_in) > TYPE_VECTOR_SUBPARTS (tem)) + || (GET_MODE_SIZE (SCALAR_TYPE_MODE (TREE_TYPE (vectype_in))) + < GET_MODE_SIZE (SCALAR_TYPE_MODE (TREE_TYPE (tem))))) vectype_in = tem; } @@ -5853,11 +6230,11 @@ return false; if (dt == vect_nested_cycle) - { - found_nested_cycle_def = true; - reduc_def_stmt = def_stmt; - reduc_index = i; - } + { + found_nested_cycle_def = true; + reduc_def_info = def_stmt_info; + reduc_index = i; + } if (i == 1 && code == COND_EXPR) { @@ -5867,9 +6244,13 @@ cond_reduc_dt = dt; cond_reduc_val = ops[i]; } - if (dt == vect_induction_def && def_stmt != NULL - && is_nonwrapping_integer_induction (def_stmt, loop)) - cond_reduc_dt = dt; + if (dt == vect_induction_def + && def_stmt_info + && is_nonwrapping_integer_induction (def_stmt_info, loop)) + { + cond_reduc_dt = dt; + cond_stmt_vinfo = def_stmt_info; + } } } @@ -5880,13 +6261,25 @@ directy used in stmt. */ if (reduc_index == -1) { - if (orig_stmt) - reduc_def_stmt = STMT_VINFO_REDUC_DEF (orig_stmt_info); + if (STMT_VINFO_REDUC_TYPE (stmt_info) == FOLD_LEFT_REDUCTION) + { + if (dump_enabled_p ()) + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "in-order reduction chain without SLP.\n"); + return false; + } + + if (orig_stmt_info) + reduc_def_info = STMT_VINFO_REDUC_DEF (orig_stmt_info); else - reduc_def_stmt = STMT_VINFO_REDUC_DEF (stmt_info); - } - - if (! reduc_def_stmt || gimple_code (reduc_def_stmt) != GIMPLE_PHI) + reduc_def_info = STMT_VINFO_REDUC_DEF (stmt_info); + } + + if (! reduc_def_info) + return false; + + gphi *reduc_def_phi = dyn_cast <gphi *> (reduc_def_info->stmt); + if (!reduc_def_phi) return false; if (!(reduc_index == -1 @@ -5901,29 +6294,44 @@ /* For pattern recognized stmts, orig_stmt might be a reduction, but some helper statements for the pattern might not, or might be COND_EXPRs with reduction uses in the condition. */ - gcc_assert (orig_stmt); + gcc_assert (orig_stmt_info); return false; } - stmt_vec_info reduc_def_info = vinfo_for_stmt (reduc_def_stmt); + /* PHIs should not participate in patterns. */ + gcc_assert (!STMT_VINFO_RELATED_STMT (reduc_def_info)); enum vect_reduction_type v_reduc_type = STMT_VINFO_REDUC_TYPE (reduc_def_info); - gimple *tmp = STMT_VINFO_REDUC_DEF (reduc_def_info); + stmt_vec_info tmp = STMT_VINFO_REDUC_DEF (reduc_def_info); STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) = v_reduc_type; /* If we have a condition reduction, see if we can simplify it further. */ if (v_reduc_type == COND_REDUCTION) { - if (cond_reduc_dt == vect_induction_def) + /* TODO: We can't yet handle reduction chains, since we need to treat + each COND_EXPR in the chain specially, not just the last one. + E.g. for: + + x_1 = PHI <x_3, ...> + x_2 = a_2 ? ... : x_1; + x_3 = a_3 ? ... : x_2; + + we're interested in the last element in x_3 for which a_2 || a_3 + is true, whereas the current reduction chain handling would + vectorize x_2 as a normal VEC_COND_EXPR and only treat x_3 + as a reduction operation. */ + if (reduc_index == -1) { if (dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, - "condition expression based on " - "integer induction.\n"); - STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) - = INTEGER_INDUC_COND_REDUCTION; + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "conditional reduction chains not supported\n"); + return false; } + /* vect_is_simple_reduction ensured that operand 2 is the + loop-carried operand. */ + gcc_assert (reduc_index == 2); + /* Loop peeling modifies initial value of reduction PHI, which makes the reduction stmt to be transformed different to the original stmt analyzed. We need to record reduction code for @@ -5936,6 +6344,57 @@ gcc_assert (cond_reduc_dt == vect_constant_def); STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) = CONST_COND_REDUCTION; } + else if (direct_internal_fn_supported_p (IFN_FOLD_EXTRACT_LAST, + vectype_in, OPTIMIZE_FOR_SPEED)) + { + if (dump_enabled_p ()) + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "optimizing condition reduction with" + " FOLD_EXTRACT_LAST.\n"); + STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) = EXTRACT_LAST_REDUCTION; + } + else if (cond_reduc_dt == vect_induction_def) + { + tree base + = STMT_VINFO_LOOP_PHI_EVOLUTION_BASE_UNCHANGED (cond_stmt_vinfo); + tree step = STMT_VINFO_LOOP_PHI_EVOLUTION_PART (cond_stmt_vinfo); + + gcc_assert (TREE_CODE (base) == INTEGER_CST + && TREE_CODE (step) == INTEGER_CST); + cond_reduc_val = NULL_TREE; + /* Find a suitable value, for MAX_EXPR below base, for MIN_EXPR + above base; punt if base is the minimum value of the type for + MAX_EXPR or maximum value of the type for MIN_EXPR for now. */ + if (tree_int_cst_sgn (step) == -1) + { + cond_reduc_op_code = MIN_EXPR; + if (tree_int_cst_sgn (base) == -1) + cond_reduc_val = build_int_cst (TREE_TYPE (base), 0); + else if (tree_int_cst_lt (base, + TYPE_MAX_VALUE (TREE_TYPE (base)))) + cond_reduc_val + = int_const_binop (PLUS_EXPR, base, integer_one_node); + } + else + { + cond_reduc_op_code = MAX_EXPR; + if (tree_int_cst_sgn (base) == 1) + cond_reduc_val = build_int_cst (TREE_TYPE (base), 0); + else if (tree_int_cst_lt (TYPE_MIN_VALUE (TREE_TYPE (base)), + base)) + cond_reduc_val + = int_const_binop (MINUS_EXPR, base, integer_one_node); + } + if (cond_reduc_val) + { + if (dump_enabled_p ()) + dump_printf_loc (MSG_NOTE, vect_location, + "condition expression based on " + "integer induction.\n"); + STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) + = INTEGER_INDUC_COND_REDUCTION; + } + } else if (cond_reduc_dt == vect_constant_def) { enum vect_def_type cond_initial_dt; @@ -5944,8 +6403,7 @@ = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop)); gcc_assert (cond_reduc_val != NULL_TREE); - vect_is_simple_use (cond_initial_val, loop_vinfo, - &def_stmt, &cond_initial_dt); + vect_is_simple_use (cond_initial_val, loop_vinfo, &cond_initial_dt); if (cond_initial_dt == vect_constant_def && types_compatible_p (TREE_TYPE (cond_initial_val), TREE_TYPE (cond_reduc_val))) @@ -5968,16 +6426,16 @@ } } - if (orig_stmt) - gcc_assert (tmp == orig_stmt - || GROUP_FIRST_ELEMENT (vinfo_for_stmt (tmp)) == orig_stmt); + if (orig_stmt_info) + gcc_assert (tmp == orig_stmt_info + || REDUC_GROUP_FIRST_ELEMENT (tmp) == orig_stmt_info); else /* We changed STMT to be the first stmt in reduction chain, hence we check that in this case the first element in the chain is STMT. */ - gcc_assert (stmt == tmp - || GROUP_FIRST_ELEMENT (vinfo_for_stmt (tmp)) == stmt); - - if (STMT_VINFO_LIVE_P (vinfo_for_stmt (reduc_def_stmt))) + gcc_assert (tmp == stmt_info + || REDUC_GROUP_FIRST_ELEMENT (tmp) == stmt_info); + + if (STMT_VINFO_LIVE_P (reduc_def_info)) return false; if (slp_node) @@ -5988,13 +6446,15 @@ gcc_assert (ncopies >= 1); vec_mode = TYPE_MODE (vectype_in); + poly_uint64 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); if (code == COND_EXPR) { /* Only call during the analysis stage, otherwise we'll lose STMT_VINFO_TYPE. */ - if (!vec_stmt && !vectorizable_condition (stmt, gsi, NULL, - ops[reduc_index], 0, NULL)) + if (!vec_stmt && !vectorizable_condition (stmt_info, gsi, NULL, + ops[reduc_index], 0, NULL, + cost_vec)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, @@ -6033,7 +6493,7 @@ if (dump_enabled_p ()) dump_printf (MSG_NOTE, "op not supported by target.\n"); - if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD + if (maybe_ne (GET_MODE_SIZE (vec_mode), UNITS_PER_WORD) || !vect_worthwhile_without_simd_p (loop_vinfo, code)) return false; @@ -6088,13 +6548,15 @@ (and also the same tree-code) when generating the epilog code and when generating the code inside the loop. */ - if (orig_stmt) + vect_reduction_type reduction_type + = STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info); + if (orig_stmt_info + && (reduction_type == TREE_CODE_REDUCTION + || reduction_type == FOLD_LEFT_REDUCTION)) { /* This is a reduction pattern: get the vectype from the type of the reduction variable, and get the tree-code from orig_stmt. */ - gcc_assert (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) - == TREE_CODE_REDUCTION); - orig_code = gimple_assign_rhs_code (orig_stmt); + orig_code = gimple_assign_rhs_code (orig_stmt_info->stmt); gcc_assert (vectype_out); vec_mode = TYPE_MODE (vectype_out); } @@ -6109,55 +6571,48 @@ /* For simple condition reductions, replace with the actual expression we want to base our reduction around. */ - if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) == CONST_COND_REDUCTION) + if (reduction_type == CONST_COND_REDUCTION) { orig_code = STMT_VINFO_VEC_CONST_COND_REDUC_CODE (stmt_info); gcc_assert (orig_code == MAX_EXPR || orig_code == MIN_EXPR); } - else if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) - == INTEGER_INDUC_COND_REDUCTION) - orig_code = MAX_EXPR; + else if (reduction_type == INTEGER_INDUC_COND_REDUCTION) + orig_code = cond_reduc_op_code; } if (nested_cycle) { - def_bb = gimple_bb (reduc_def_stmt); + def_bb = gimple_bb (reduc_def_phi); def_stmt_loop = def_bb->loop_father; - def_arg = PHI_ARG_DEF_FROM_EDGE (reduc_def_stmt, + def_arg = PHI_ARG_DEF_FROM_EDGE (reduc_def_phi, loop_preheader_edge (def_stmt_loop)); - if (TREE_CODE (def_arg) == SSA_NAME - && (def_arg_stmt = SSA_NAME_DEF_STMT (def_arg)) - && gimple_code (def_arg_stmt) == GIMPLE_PHI - && flow_bb_inside_loop_p (outer_loop, gimple_bb (def_arg_stmt)) - && vinfo_for_stmt (def_arg_stmt) - && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_arg_stmt)) - == vect_double_reduction_def) + stmt_vec_info def_arg_stmt_info = loop_vinfo->lookup_def (def_arg); + if (def_arg_stmt_info + && (STMT_VINFO_DEF_TYPE (def_arg_stmt_info) + == vect_double_reduction_def)) double_reduc = true; } - epilog_reduc_code = ERROR_MARK; - - if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) != COND_REDUCTION) - { - if (reduction_code_for_scalar_code (orig_code, &epilog_reduc_code)) - { - reduc_optab = optab_for_tree_code (epilog_reduc_code, vectype_out, - optab_default); - if (!reduc_optab) - { - if (dump_enabled_p ()) - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "no optab for reduction.\n"); - - epilog_reduc_code = ERROR_MARK; - } - else if (optab_handler (reduc_optab, vec_mode) == CODE_FOR_nothing) + reduc_fn = IFN_LAST; + + if (reduction_type == TREE_CODE_REDUCTION + || reduction_type == FOLD_LEFT_REDUCTION + || reduction_type == INTEGER_INDUC_COND_REDUCTION + || reduction_type == CONST_COND_REDUCTION) + { + if (reduction_type == FOLD_LEFT_REDUCTION + ? fold_left_reduction_fn (orig_code, &reduc_fn) + : reduction_fn_for_scalar_code (orig_code, &reduc_fn)) + { + if (reduc_fn != IFN_LAST + && !direct_internal_fn_supported_p (reduc_fn, vectype_out, + OPTIMIZE_FOR_SPEED)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "reduc op not supported by target.\n"); - epilog_reduc_code = ERROR_MARK; + reduc_fn = IFN_LAST; } } else @@ -6172,23 +6627,32 @@ } } } - else + else if (reduction_type == COND_REDUCTION) { int scalar_precision = GET_MODE_PRECISION (SCALAR_TYPE_MODE (scalar_type)); cr_index_scalar_type = make_unsigned_type (scalar_precision); - cr_index_vector_type = build_vector_type - (cr_index_scalar_type, TYPE_VECTOR_SUBPARTS (vectype_out)); - - optab = optab_for_tree_code (REDUC_MAX_EXPR, cr_index_vector_type, - optab_default); - if (optab_handler (optab, TYPE_MODE (cr_index_vector_type)) - != CODE_FOR_nothing) - epilog_reduc_code = REDUC_MAX_EXPR; - } - - if ((double_reduc - || STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) != TREE_CODE_REDUCTION) + cr_index_vector_type = build_vector_type (cr_index_scalar_type, + nunits_out); + + if (direct_internal_fn_supported_p (IFN_REDUC_MAX, cr_index_vector_type, + OPTIMIZE_FOR_SPEED)) + reduc_fn = IFN_REDUC_MAX; + } + + if (reduction_type != EXTRACT_LAST_REDUCTION + && (!nested_cycle || double_reduc) + && reduc_fn == IFN_LAST + && !nunits_out.is_constant ()) + { + if (dump_enabled_p ()) + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "missing target support for reduction on" + " variable-length vectors.\n"); + return false; + } + + if ((double_reduc || reduction_type != TREE_CODE_REDUCTION) && ncopies > 1) { if (dump_enabled_p ()) @@ -6198,6 +6662,101 @@ return false; } + /* For SLP reductions, see if there is a neutral value we can use. */ + tree neutral_op = NULL_TREE; + if (slp_node) + neutral_op = neutral_op_for_slp_reduction + (slp_node_instance->reduc_phis, code, + REDUC_GROUP_FIRST_ELEMENT (stmt_info) != NULL); + + if (double_reduc && reduction_type == FOLD_LEFT_REDUCTION) + { + /* We can't support in-order reductions of code such as this: + + for (int i = 0; i < n1; ++i) + for (int j = 0; j < n2; ++j) + l += a[j]; + + since GCC effectively transforms the loop when vectorizing: + + for (int i = 0; i < n1 / VF; ++i) + for (int j = 0; j < n2; ++j) + for (int k = 0; k < VF; ++k) + l += a[j]; + + which is a reassociation of the original operation. */ + if (dump_enabled_p ()) + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "in-order double reduction not supported.\n"); + + return false; + } + + if (reduction_type == FOLD_LEFT_REDUCTION + && slp_node + && !REDUC_GROUP_FIRST_ELEMENT (stmt_info)) + { + /* We cannot use in-order reductions in this case because there is + an implicit reassociation of the operations involved. */ + if (dump_enabled_p ()) + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "in-order unchained SLP reductions not supported.\n"); + return false; + } + + /* For double reductions, and for SLP reductions with a neutral value, + we construct a variable-length initial vector by loading a vector + full of the neutral value and then shift-and-inserting the start + values into the low-numbered elements. */ + if ((double_reduc || neutral_op) + && !nunits_out.is_constant () + && !direct_internal_fn_supported_p (IFN_VEC_SHL_INSERT, + vectype_out, OPTIMIZE_FOR_SPEED)) + { + if (dump_enabled_p ()) + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "reduction on variable-length vectors requires" + " target support for a vector-shift-and-insert" + " operation.\n"); + return false; + } + + /* Check extra constraints for variable-length unchained SLP reductions. */ + if (STMT_SLP_TYPE (stmt_info) + && !REDUC_GROUP_FIRST_ELEMENT (stmt_info) + && !nunits_out.is_constant ()) + { + /* We checked above that we could build the initial vector when + there's a neutral element value. Check here for the case in + which each SLP statement has its own initial value and in which + that value needs to be repeated for every instance of the + statement within the initial vector. */ + unsigned int group_size = SLP_TREE_SCALAR_STMTS (slp_node).length (); + scalar_mode elt_mode = SCALAR_TYPE_MODE (TREE_TYPE (vectype_out)); + if (!neutral_op + && !can_duplicate_and_interleave_p (group_size, elt_mode)) + { + if (dump_enabled_p ()) + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "unsupported form of SLP reduction for" + " variable-length vectors: cannot build" + " initial vector.\n"); + return false; + } + /* The epilogue code relies on the number of elements being a multiple + of the group size. The duplicate-and-interleave approach to setting + up the the initial vector does too. */ + if (!multiple_p (nunits_out, group_size)) + { + if (dump_enabled_p ()) + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "unsupported form of SLP reduction for" + " variable-length vectors: the vector size" + " is not a multiple of the number of results.\n"); + return false; + } + } + /* In case of widenning multiplication by a constant, we update the type of the constant to be the type of the other operand. We check that the constant fits the type in the pattern recognition pass. */ @@ -6218,7 +6777,7 @@ } } - if (STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info) == COND_REDUCTION) + if (reduction_type == COND_REDUCTION) { widest_int ni; @@ -6275,13 +6834,12 @@ This only works when we see both the reduction PHI and its only consumer in vectorizable_reduction and there are no intermediate stmts participating. */ - use_operand_p use_p; - gimple *use_stmt; + stmt_vec_info use_stmt_info; + tree reduc_phi_result = gimple_phi_result (reduc_def_phi); if (ncopies > 1 && (STMT_VINFO_RELEVANT (stmt_info) <= vect_used_only_live) - && single_imm_use (gimple_phi_result (reduc_def_stmt), &use_p, &use_stmt) - && (use_stmt == stmt - || STMT_VINFO_RELATED_STMT (vinfo_for_stmt (use_stmt)) == stmt)) + && (use_stmt_info = loop_vinfo->lookup_single_use (reduc_phi_result)) + && vect_stmt_to_vectorize (use_stmt_info) == stmt_info) { single_defuse_cycle = true; epilog_copies = 1; @@ -6304,10 +6862,46 @@ return false; } + if (slp_node) + vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); + else + vec_num = 1; + + internal_fn cond_fn = get_conditional_internal_fn (code); + vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); + if (!vec_stmt) /* transformation not required. */ { - if (first_p) - vect_model_reduction_cost (stmt_info, epilog_reduc_code, ncopies); + vect_model_reduction_cost (stmt_info, reduc_fn, ncopies, cost_vec); + if (loop_vinfo && LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo)) + { + if (reduction_type != FOLD_LEFT_REDUCTION + && (cond_fn == IFN_LAST + || !direct_internal_fn_supported_p (cond_fn, vectype_in, + OPTIMIZE_FOR_SPEED))) + { + if (dump_enabled_p ()) + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "can't use a fully-masked loop because no" + " conditional operation is available.\n"); + LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; + } + else if (reduc_index == -1) + { + if (dump_enabled_p ()) + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "can't use a fully-masked loop for chained" + " reductions.\n"); + LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; + } + else + vect_record_loop_mask (loop_vinfo, masks, ncopies * vec_num, + vectype_in); + } + if (dump_enabled_p () + && reduction_type == FOLD_LEFT_REDUCTION) + dump_printf_loc (MSG_NOTE, vect_location, + "using an in-order (fold-left) reduction.\n"); STMT_VINFO_TYPE (stmt_info) = reduc_vec_info_type; return true; } @@ -6321,16 +6915,27 @@ if (code == COND_EXPR) gcc_assert (ncopies == 1); + bool masked_loop_p = LOOP_VINFO_FULLY_MASKED_P (loop_vinfo); + + if (reduction_type == FOLD_LEFT_REDUCTION) + return vectorize_fold_left_reduction + (stmt_info, gsi, vec_stmt, slp_node, reduc_def_phi, code, + reduc_fn, ops, vectype_in, reduc_index, masks); + + if (reduction_type == EXTRACT_LAST_REDUCTION) + { + gcc_assert (!slp_node); + return vectorizable_condition (stmt_info, gsi, vec_stmt, + NULL, reduc_index, NULL, NULL); + } + /* Create the destination vector */ vec_dest = vect_create_destination_var (scalar_dest, vectype_out); prev_stmt_info = NULL; prev_phi_info = NULL; - if (slp_node) - vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); - else - { - vec_num = 1; + if (!slp_node) + { vec_oprnds0.create (1); vec_oprnds1.create (1); if (op_type == ternary_op) @@ -6345,16 +6950,16 @@ if (slp_node) phis.splice (SLP_TREE_VEC_STMTS (slp_node_instance->reduc_phis)); else - phis.quick_push (STMT_VINFO_VEC_STMT (vinfo_for_stmt (reduc_def_stmt))); + phis.quick_push (STMT_VINFO_VEC_STMT (reduc_def_info)); for (j = 0; j < ncopies; j++) { if (code == COND_EXPR) { gcc_assert (!slp_node); - vectorizable_condition (stmt, gsi, vec_stmt, - PHI_RESULT (phis[0]), - reduc_index, NULL); + vectorizable_condition (stmt_info, gsi, vec_stmt, + PHI_RESULT (phis[0]->stmt), + reduc_index, NULL, NULL); /* Multiple types are not supported for condition. */ break; } @@ -6389,12 +6994,12 @@ else { vec_oprnds0.quick_push - (vect_get_vec_def_for_operand (ops[0], stmt)); + (vect_get_vec_def_for_operand (ops[0], stmt_info)); vec_oprnds1.quick_push - (vect_get_vec_def_for_operand (ops[1], stmt)); + (vect_get_vec_def_for_operand (ops[1], stmt_info)); if (op_type == ternary_op) vec_oprnds2.quick_push - (vect_get_vec_def_for_operand (ops[2], stmt)); + (vect_get_vec_def_for_operand (ops[2], stmt_info)); } } else @@ -6404,22 +7009,25 @@ gcc_assert (reduc_index != -1 || ! single_defuse_cycle); if (single_defuse_cycle && reduc_index == 0) - vec_oprnds0[0] = gimple_assign_lhs (new_stmt); + vec_oprnds0[0] = gimple_get_lhs (new_stmt_info->stmt); else vec_oprnds0[0] - = vect_get_vec_def_for_stmt_copy (dts[0], vec_oprnds0[0]); + = vect_get_vec_def_for_stmt_copy (loop_vinfo, + vec_oprnds0[0]); if (single_defuse_cycle && reduc_index == 1) - vec_oprnds1[0] = gimple_assign_lhs (new_stmt); + vec_oprnds1[0] = gimple_get_lhs (new_stmt_info->stmt); else vec_oprnds1[0] - = vect_get_vec_def_for_stmt_copy (dts[1], vec_oprnds1[0]); + = vect_get_vec_def_for_stmt_copy (loop_vinfo, + vec_oprnds1[0]); if (op_type == ternary_op) { if (single_defuse_cycle && reduc_index == 2) - vec_oprnds2[0] = gimple_assign_lhs (new_stmt); + vec_oprnds2[0] = gimple_get_lhs (new_stmt_info->stmt); else vec_oprnds2[0] - = vect_get_vec_def_for_stmt_copy (dts[2], vec_oprnds2[0]); + = vect_get_vec_def_for_stmt_copy (loop_vinfo, + vec_oprnds2[0]); } } } @@ -6427,17 +7035,41 @@ FOR_EACH_VEC_ELT (vec_oprnds0, i, def0) { tree vop[3] = { def0, vec_oprnds1[i], NULL_TREE }; - if (op_type == ternary_op) - vop[2] = vec_oprnds2[i]; - - new_temp = make_ssa_name (vec_dest, new_stmt); - new_stmt = gimple_build_assign (new_temp, code, - vop[0], vop[1], vop[2]); - vect_finish_stmt_generation (stmt, new_stmt, gsi); + if (masked_loop_p) + { + /* Make sure that the reduction accumulator is vop[0]. */ + if (reduc_index == 1) + { + gcc_assert (commutative_tree_code (code)); + std::swap (vop[0], vop[1]); + } + tree mask = vect_get_loop_mask (gsi, masks, vec_num * ncopies, + vectype_in, i * ncopies + j); + gcall *call = gimple_build_call_internal (cond_fn, 4, mask, + vop[0], vop[1], + vop[0]); + new_temp = make_ssa_name (vec_dest, call); + gimple_call_set_lhs (call, new_temp); + gimple_call_set_nothrow (call, true); + new_stmt_info + = vect_finish_stmt_generation (stmt_info, call, gsi); + } + else + { + if (op_type == ternary_op) + vop[2] = vec_oprnds2[i]; + + gassign *new_stmt = gimple_build_assign (vec_dest, code, + vop[0], vop[1], vop[2]); + new_temp = make_ssa_name (vec_dest, new_stmt); + gimple_assign_set_lhs (new_stmt, new_temp); + new_stmt_info + = vect_finish_stmt_generation (stmt_info, new_stmt, gsi); + } if (slp_node) { - SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); + SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt_info); vect_defs.quick_push (new_temp); } else @@ -6448,22 +7080,23 @@ continue; if (j == 0) - STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; + STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt_info; else - STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; - - prev_stmt_info = vinfo_for_stmt (new_stmt); + STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt_info; + + prev_stmt_info = new_stmt_info; } /* Finalize the reduction-phi (set its arguments) and create the epilog reduction code. */ if ((!single_defuse_cycle || code == COND_EXPR) && !slp_node) - vect_defs[0] = gimple_assign_lhs (*vec_stmt); - - vect_create_epilog_for_reduction (vect_defs, stmt, reduc_def_stmt, - epilog_copies, - epilog_reduc_code, phis, - double_reduc, slp_node, slp_node_instance); + vect_defs[0] = gimple_get_lhs ((*vec_stmt)->stmt); + + vect_create_epilog_for_reduction (vect_defs, stmt_info, reduc_def_phi, + epilog_copies, reduc_fn, phis, + double_reduc, slp_node, slp_node_instance, + cond_reduc_val, cond_reduc_op_code, + neutral_op); return true; } @@ -6473,7 +7106,7 @@ For a loop where we could vectorize the operation indicated by CODE, return the minimum vectorization factor that makes it worthwhile to use generic vectors. */ -int +static unsigned int vect_min_worthwhile_factor (enum tree_code code) { switch (code) @@ -6502,24 +7135,25 @@ vect_worthwhile_without_simd_p (vec_info *vinfo, tree_code code) { loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo); + unsigned HOST_WIDE_INT value; return (loop_vinfo - && (LOOP_VINFO_VECT_FACTOR (loop_vinfo) - >= vect_min_worthwhile_factor (code))); + && LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (&value) + && value >= vect_min_worthwhile_factor (code)); } /* Function vectorizable_induction - Check if PHI performs an induction computation that can be vectorized. + Check if STMT_INFO performs an induction computation that can be vectorized. If VEC_STMT is also passed, vectorize the induction PHI: create a vectorized phi to replace it, put it in VEC_STMT, and add it to the same basic block. - Return FALSE if not a vectorizable STMT, TRUE otherwise. */ + Return true if STMT_INFO is vectorizable in this way. */ bool -vectorizable_induction (gimple *phi, +vectorizable_induction (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi ATTRIBUTE_UNUSED, - gimple **vec_stmt, slp_tree slp_node) + stmt_vec_info *vec_stmt, slp_tree slp_node, + stmt_vector_for_cost *cost_vec) { - stmt_vec_info stmt_info = vinfo_for_stmt (phi); loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); unsigned ncopies; @@ -6534,7 +7168,7 @@ gphi *induction_phi; tree induc_def, vec_dest; tree init_expr, step_expr; - int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); + poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); unsigned i; tree expr; gimple_seq stmts; @@ -6544,9 +7178,9 @@ edge latch_e; tree loop_arg; gimple_stmt_iterator si; - basic_block bb = gimple_bb (phi); - - if (gimple_code (phi) != GIMPLE_PHI) + + gphi *phi = dyn_cast <gphi *> (stmt_info->stmt); + if (!phi) return false; if (!STMT_VINFO_RELEVANT_P (stmt_info)) @@ -6557,7 +7191,7 @@ return false; tree vectype = STMT_VINFO_VECTYPE (stmt_info); - unsigned nunits = TYPE_VECTOR_SUBPARTS (vectype); + poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); if (slp_node) ncopies = 1; @@ -6566,7 +7200,7 @@ gcc_assert (ncopies >= 1); /* FORNOW. These restrictions should be relaxed. */ - if (nested_in_vect_loop_p (loop, phi)) + if (nested_in_vect_loop_p (loop, stmt_info)) { imm_use_iterator imm_iter; use_operand_p use_p; @@ -6603,7 +7237,7 @@ } if (exit_phi) { - stmt_vec_info exit_phi_vinfo = vinfo_for_stmt (exit_phi); + stmt_vec_info exit_phi_vinfo = loop_vinfo->lookup_stmt (exit_phi); if (!(STMT_VINFO_RELEVANT_P (exit_phi_vinfo) && !STMT_VINFO_LIVE_P (exit_phi_vinfo))) { @@ -6622,13 +7256,21 @@ iv_loop = loop; gcc_assert (iv_loop == (gimple_bb (phi))->loop_father); + if (slp_node && !nunits.is_constant ()) + { + /* The current SLP code creates the initial value element-by-element. */ + if (dump_enabled_p ()) + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "SLP induction not supported for variable-length" + " vectors.\n"); + return false; + } + if (!vec_stmt) /* transformation not required. */ { STMT_VINFO_TYPE (stmt_info) = induc_vec_info_type; - if (dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, - "=== vectorizable_induction ===\n"); - vect_model_induction_cost (stmt_info, ncopies); + DUMP_VECT_SCOPE ("vectorizable_induction"); + vect_model_induction_cost (stmt_info, ncopies, cost_vec); return true; } @@ -6652,9 +7294,33 @@ init_expr = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (iv_loop)); + stmts = NULL; + if (!nested_in_vect_loop) + { + /* Convert the initial value to the desired type. */ + tree new_type = TREE_TYPE (vectype); + init_expr = gimple_convert (&stmts, new_type, init_expr); + + /* If we are using the loop mask to "peel" for alignment then we need + to adjust the start value here. */ + tree skip_niters = LOOP_VINFO_MASK_SKIP_NITERS (loop_vinfo); + if (skip_niters != NULL_TREE) + { + if (FLOAT_TYPE_P (vectype)) + skip_niters = gimple_build (&stmts, FLOAT_EXPR, new_type, + skip_niters); + else + skip_niters = gimple_convert (&stmts, new_type, skip_niters); + tree skip_step = gimple_build (&stmts, MULT_EXPR, new_type, + skip_niters, step_expr); + init_expr = gimple_build (&stmts, MINUS_EXPR, new_type, + init_expr, skip_step); + } + } + /* Convert the step to the desired type. */ - stmts = NULL; step_expr = gimple_convert (&stmts, TREE_TYPE (vectype), step_expr); + if (stmts) { new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); @@ -6662,6 +7328,7 @@ } /* Find the first insertion point in the BB. */ + basic_block bb = gimple_bb (phi); si = gsi_after_labels (bb); /* For SLP induction we have to generate several IVs as for example @@ -6670,14 +7337,8 @@ [VF*S, VF*S, VF*S, VF*S] for all. */ if (slp_node) { - /* Convert the init to the desired type. */ - stmts = NULL; - init_expr = gimple_convert (&stmts, TREE_TYPE (vectype), init_expr); - if (stmts) - { - new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); - gcc_assert (!new_bb); - } + /* Enforced above. */ + unsigned int const_nunits = nunits.to_constant (); /* Generate [VF*S, VF*S, ... ]. */ if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (step_expr))) @@ -6690,32 +7351,33 @@ new_name = fold_build2 (MULT_EXPR, TREE_TYPE (step_expr), expr, step_expr); if (! CONSTANT_CLASS_P (new_name)) - new_name = vect_init_vector (phi, new_name, + new_name = vect_init_vector (stmt_info, new_name, TREE_TYPE (step_expr), NULL); new_vec = build_vector_from_val (vectype, new_name); - vec_step = vect_init_vector (phi, new_vec, vectype, NULL); + vec_step = vect_init_vector (stmt_info, new_vec, vectype, NULL); /* Now generate the IVs. */ unsigned group_size = SLP_TREE_SCALAR_STMTS (slp_node).length (); unsigned nvects = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); - unsigned elts = nunits * nvects; - unsigned nivs = least_common_multiple (group_size, nunits) / nunits; + unsigned elts = const_nunits * nvects; + unsigned nivs = least_common_multiple (group_size, + const_nunits) / const_nunits; gcc_assert (elts % group_size == 0); tree elt = init_expr; unsigned ivn; for (ivn = 0; ivn < nivs; ++ivn) { - auto_vec<tree, 32> elts (nunits); + tree_vector_builder elts (vectype, const_nunits, 1); stmts = NULL; - for (unsigned eltn = 0; eltn < nunits; ++eltn) + for (unsigned eltn = 0; eltn < const_nunits; ++eltn) { - if (ivn*nunits + eltn >= group_size - && (ivn*nunits + eltn) % group_size == 0) + if (ivn*const_nunits + eltn >= group_size + && (ivn * const_nunits + eltn) % group_size == 0) elt = gimple_build (&stmts, PLUS_EXPR, TREE_TYPE (elt), elt, step_expr); elts.quick_push (elt); } - vec_init = gimple_build_vector (&stmts, vectype, elts); + vec_init = gimple_build_vector (&stmts, &elts); if (stmts) { new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); @@ -6725,29 +7387,29 @@ /* Create the induction-phi that defines the induction-operand. */ vec_dest = vect_get_new_vect_var (vectype, vect_simple_var, "vec_iv_"); induction_phi = create_phi_node (vec_dest, iv_loop->header); - set_vinfo_for_stmt (induction_phi, - new_stmt_vec_info (induction_phi, loop_vinfo)); + stmt_vec_info induction_phi_info + = loop_vinfo->add_stmt (induction_phi); induc_def = PHI_RESULT (induction_phi); /* Create the iv update inside the loop */ vec_def = make_ssa_name (vec_dest); new_stmt = gimple_build_assign (vec_def, PLUS_EXPR, induc_def, vec_step); gsi_insert_before (&si, new_stmt, GSI_SAME_STMT); - set_vinfo_for_stmt (new_stmt, new_stmt_vec_info (new_stmt, loop_vinfo)); + loop_vinfo->add_stmt (new_stmt); /* Set the arguments of the phi node: */ add_phi_arg (induction_phi, vec_init, pe, UNKNOWN_LOCATION); add_phi_arg (induction_phi, vec_def, loop_latch_edge (iv_loop), UNKNOWN_LOCATION); - SLP_TREE_VEC_STMTS (slp_node).quick_push (induction_phi); + SLP_TREE_VEC_STMTS (slp_node).quick_push (induction_phi_info); } /* Re-use IVs when we can. */ if (ivn < nvects) { unsigned vfp - = least_common_multiple (group_size, nunits) / group_size; + = least_common_multiple (group_size, const_nunits) / group_size; /* Generate [VF'*S, VF'*S, ... ]. */ if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (step_expr))) { @@ -6759,13 +7421,13 @@ new_name = fold_build2 (MULT_EXPR, TREE_TYPE (step_expr), expr, step_expr); if (! CONSTANT_CLASS_P (new_name)) - new_name = vect_init_vector (phi, new_name, + new_name = vect_init_vector (stmt_info, new_name, TREE_TYPE (step_expr), NULL); new_vec = build_vector_from_val (vectype, new_name); - vec_step = vect_init_vector (phi, new_vec, vectype, NULL); + vec_step = vect_init_vector (stmt_info, new_vec, vectype, NULL); for (; ivn < nvects; ++ivn) { - gimple *iv = SLP_TREE_VEC_STMTS (slp_node)[ivn - nivs]; + gimple *iv = SLP_TREE_VEC_STMTS (slp_node)[ivn - nivs]->stmt; tree def; if (gimple_code (iv) == GIMPLE_PHI) def = gimple_phi_result (iv); @@ -6781,9 +7443,8 @@ gimple_stmt_iterator tgsi = gsi_for_stmt (iv); gsi_insert_after (&tgsi, new_stmt, GSI_CONTINUE_LINKING); } - set_vinfo_for_stmt (new_stmt, - new_stmt_vec_info (new_stmt, loop_vinfo)); - SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt); + SLP_TREE_VEC_STMTS (slp_node).quick_push + (loop_vinfo->add_stmt (new_stmt)); } } @@ -6796,7 +7457,7 @@ /* iv_loop is nested in the loop to be vectorized. init_expr had already been created during vectorization of previous stmts. We obtain it from the STMT_VINFO_VEC_STMT of the defining stmt. */ - vec_init = vect_get_vec_def_for_operand (init_expr, phi); + vec_init = vect_get_vec_def_for_operand (init_expr, stmt_info); /* If the initial value is not of proper type, convert it. */ if (!useless_type_conversion_p (vectype, TREE_TYPE (vec_init))) { @@ -6811,8 +7472,7 @@ new_bb = gsi_insert_on_edge_immediate (loop_preheader_edge (iv_loop), new_stmt); gcc_assert (!new_bb); - set_vinfo_for_stmt (new_stmt, - new_stmt_vec_info (new_stmt, loop_vinfo)); + loop_vinfo->add_stmt (new_stmt); } } else @@ -6822,18 +7482,45 @@ stmts = NULL; new_name = gimple_convert (&stmts, TREE_TYPE (vectype), init_expr); - auto_vec<tree, 32> elts (nunits); - elts.quick_push (new_name); - for (i = 1; i < nunits; i++) - { - /* Create: new_name_i = new_name + step_expr */ - new_name = gimple_build (&stmts, PLUS_EXPR, TREE_TYPE (new_name), - new_name, step_expr); + unsigned HOST_WIDE_INT const_nunits; + if (nunits.is_constant (&const_nunits)) + { + tree_vector_builder elts (vectype, const_nunits, 1); elts.quick_push (new_name); + for (i = 1; i < const_nunits; i++) + { + /* Create: new_name_i = new_name + step_expr */ + new_name = gimple_build (&stmts, PLUS_EXPR, TREE_TYPE (new_name), + new_name, step_expr); + elts.quick_push (new_name); + } + /* Create a vector from [new_name_0, new_name_1, ..., + new_name_nunits-1] */ + vec_init = gimple_build_vector (&stmts, &elts); } - /* Create a vector from [new_name_0, new_name_1, ..., - new_name_nunits-1] */ - vec_init = gimple_build_vector (&stmts, vectype, elts); + else if (INTEGRAL_TYPE_P (TREE_TYPE (step_expr))) + /* Build the initial value directly from a VEC_SERIES_EXPR. */ + vec_init = gimple_build (&stmts, VEC_SERIES_EXPR, vectype, + new_name, step_expr); + else + { + /* Build: + [base, base, base, ...] + + (vectype) [0, 1, 2, ...] * [step, step, step, ...]. */ + gcc_assert (SCALAR_FLOAT_TYPE_P (TREE_TYPE (step_expr))); + gcc_assert (flag_associative_math); + tree index = build_index_vector (vectype, 0, 1); + tree base_vec = gimple_build_vector_from_val (&stmts, vectype, + new_name); + tree step_vec = gimple_build_vector_from_val (&stmts, vectype, + step_expr); + vec_init = gimple_build (&stmts, FLOAT_EXPR, vectype, index); + vec_init = gimple_build (&stmts, MULT_EXPR, vectype, + vec_init, step_vec); + vec_init = gimple_build (&stmts, PLUS_EXPR, vectype, + vec_init, base_vec); + } + if (stmts) { new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); @@ -6872,7 +7559,7 @@ gcc_assert (CONSTANT_CLASS_P (new_name) || TREE_CODE (new_name) == SSA_NAME); new_vec = build_vector_from_val (vectype, t); - vec_step = vect_init_vector (phi, new_vec, vectype, NULL); + vec_step = vect_init_vector (stmt_info, new_vec, vectype, NULL); /* Create the following def-use cycle: @@ -6889,22 +7576,21 @@ /* Create the induction-phi that defines the induction-operand. */ vec_dest = vect_get_new_vect_var (vectype, vect_simple_var, "vec_iv_"); induction_phi = create_phi_node (vec_dest, iv_loop->header); - set_vinfo_for_stmt (induction_phi, - new_stmt_vec_info (induction_phi, loop_vinfo)); + stmt_vec_info induction_phi_info = loop_vinfo->add_stmt (induction_phi); induc_def = PHI_RESULT (induction_phi); /* Create the iv update inside the loop */ vec_def = make_ssa_name (vec_dest); new_stmt = gimple_build_assign (vec_def, PLUS_EXPR, induc_def, vec_step); gsi_insert_before (&si, new_stmt, GSI_SAME_STMT); - set_vinfo_for_stmt (new_stmt, new_stmt_vec_info (new_stmt, loop_vinfo)); + stmt_vec_info new_stmt_info = loop_vinfo->add_stmt (new_stmt); /* Set the arguments of the phi node: */ add_phi_arg (induction_phi, vec_init, pe, UNKNOWN_LOCATION); add_phi_arg (induction_phi, vec_def, loop_latch_edge (iv_loop), UNKNOWN_LOCATION); - STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = induction_phi; + STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = induction_phi_info; /* In case that vectorization factor (VF) is bigger than the number of elements that we can fit in a vectype (nunits), we have to generate @@ -6939,10 +7625,10 @@ gcc_assert (CONSTANT_CLASS_P (new_name) || TREE_CODE (new_name) == SSA_NAME); new_vec = build_vector_from_val (vectype, t); - vec_step = vect_init_vector (phi, new_vec, vectype, NULL); + vec_step = vect_init_vector (stmt_info, new_vec, vectype, NULL); vec_def = induc_def; - prev_stmt_vinfo = vinfo_for_stmt (induction_phi); + prev_stmt_vinfo = induction_phi_info; for (i = 1; i < ncopies; i++) { /* vec_i = vec_prev + vec_step */ @@ -6952,10 +7638,9 @@ gimple_assign_set_lhs (new_stmt, vec_def); gsi_insert_before (&si, new_stmt, GSI_SAME_STMT); - set_vinfo_for_stmt (new_stmt, - new_stmt_vec_info (new_stmt, loop_vinfo)); - STMT_VINFO_RELATED_STMT (prev_stmt_vinfo) = new_stmt; - prev_stmt_vinfo = vinfo_for_stmt (new_stmt); + new_stmt_info = loop_vinfo->add_stmt (new_stmt); + STMT_VINFO_RELATED_STMT (prev_stmt_vinfo) = new_stmt_info; + prev_stmt_vinfo = new_stmt_info; } } @@ -6978,56 +7663,52 @@ } if (exit_phi) { - stmt_vec_info stmt_vinfo = vinfo_for_stmt (exit_phi); + stmt_vec_info stmt_vinfo = loop_vinfo->lookup_stmt (exit_phi); /* FORNOW. Currently not supporting the case that an inner-loop induction is not used in the outer-loop (i.e. only outside the outer-loop). */ gcc_assert (STMT_VINFO_RELEVANT_P (stmt_vinfo) && !STMT_VINFO_LIVE_P (stmt_vinfo)); - STMT_VINFO_VEC_STMT (stmt_vinfo) = new_stmt; + STMT_VINFO_VEC_STMT (stmt_vinfo) = new_stmt_info; if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, - "vector of inductions after inner-loop:"); - dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0); - } + dump_printf_loc (MSG_NOTE, vect_location, + "vector of inductions after inner-loop:%G", + new_stmt); } } if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, - "transform induction: created def-use cycle: "); - dump_gimple_stmt (MSG_NOTE, TDF_SLIM, induction_phi, 0); - dump_gimple_stmt (MSG_NOTE, TDF_SLIM, - SSA_NAME_DEF_STMT (vec_def), 0); - } + dump_printf_loc (MSG_NOTE, vect_location, + "transform induction: created def-use cycle: %G%G", + induction_phi, SSA_NAME_DEF_STMT (vec_def)); return true; } /* Function vectorizable_live_operation. - STMT computes a value that is used outside the loop. Check if + STMT_INFO computes a value that is used outside the loop. Check if it can be supported. */ bool -vectorizable_live_operation (gimple *stmt, +vectorizable_live_operation (stmt_vec_info stmt_info, gimple_stmt_iterator *gsi ATTRIBUTE_UNUSED, slp_tree slp_node, int slp_index, - gimple **vec_stmt) + stmt_vec_info *vec_stmt, + stmt_vector_for_cost *) { - stmt_vec_info stmt_info = vinfo_for_stmt (stmt); loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); imm_use_iterator imm_iter; tree lhs, lhs_type, bitsize, vec_bitsize; tree vectype = STMT_VINFO_VECTYPE (stmt_info); - int nunits = TYPE_VECTOR_SUBPARTS (vectype); + poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); int ncopies; gimple *use_stmt; auto_vec<tree> vec_oprnds; + int vec_entry = 0; + poly_uint64 vec_index = 0; gcc_assert (STMT_VINFO_LIVE_P (stmt_info)); @@ -7035,7 +7716,7 @@ return false; /* FORNOW. CHECKME. */ - if (nested_in_vect_loop_p (loop, stmt)) + if (nested_in_vect_loop_p (loop, stmt_info)) return false; /* If STMT is not relevant and it is a simple assignment and its inputs are @@ -7043,7 +7724,7 @@ scalar value that it computes will be used. */ if (!STMT_VINFO_RELEVANT_P (stmt_info)) { - gcc_assert (is_simple_and_all_uses_invariant (stmt, loop_vinfo)); + gcc_assert (is_simple_and_all_uses_invariant (stmt_info, loop_vinfo)); if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "statement is simple and uses invariant. Leaving in " @@ -7056,13 +7737,74 @@ else ncopies = vect_get_num_copies (loop_vinfo, vectype); + if (slp_node) + { + gcc_assert (slp_index >= 0); + + int num_scalar = SLP_TREE_SCALAR_STMTS (slp_node).length (); + int num_vec = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); + + /* Get the last occurrence of the scalar index from the concatenation of + all the slp vectors. Calculate which slp vector it is and the index + within. */ + poly_uint64 pos = (num_vec * nunits) - num_scalar + slp_index; + + /* Calculate which vector contains the result, and which lane of + that vector we need. */ + if (!can_div_trunc_p (pos, nunits, &vec_entry, &vec_index)) + { + if (dump_enabled_p ()) + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "Cannot determine which vector holds the" + " final result.\n"); + return false; + } + } + if (!vec_stmt) - /* No transformation required. */ - return true; - - /* If stmt has a related stmt, then use that for getting the lhs. */ - if (is_pattern_stmt_p (stmt_info)) - stmt = STMT_VINFO_RELATED_STMT (stmt_info); + { + /* No transformation required. */ + if (LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo)) + { + if (!direct_internal_fn_supported_p (IFN_EXTRACT_LAST, vectype, + OPTIMIZE_FOR_SPEED)) + { + if (dump_enabled_p ()) + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "can't use a fully-masked loop because " + "the target doesn't support extract last " + "reduction.\n"); + LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; + } + else if (slp_node) + { + if (dump_enabled_p ()) + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "can't use a fully-masked loop because an " + "SLP statement is live after the loop.\n"); + LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; + } + else if (ncopies > 1) + { + if (dump_enabled_p ()) + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, + "can't use a fully-masked loop because" + " ncopies is greater than 1.\n"); + LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false; + } + else + { + gcc_assert (ncopies == 1 && !slp_node); + vect_record_loop_mask (loop_vinfo, + &LOOP_VINFO_MASKS (loop_vinfo), + 1, vectype); + } + } + return true; + } + + /* Use the lhs of the original scalar statement. */ + gimple *stmt = vect_orig_stmt (stmt_info)->stmt; lhs = (is_a <gphi *> (stmt)) ? gimple_phi_result (stmt) : gimple_get_lhs (stmt); @@ -7077,20 +7819,14 @@ tree vec_lhs, bitstart; if (slp_node) { - gcc_assert (slp_index >= 0); - - int num_scalar = SLP_TREE_SCALAR_STMTS (slp_node).length (); - int num_vec = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); - - /* Get the last occurrence of the scalar index from the concatenation of - all the slp vectors. Calculate which slp vector it is and the index - within. */ - int pos = (num_vec * nunits) - num_scalar + slp_index; - int vec_entry = pos / nunits; - int vec_index = pos % nunits; + gcc_assert (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)); /* Get the correct slp vectorized stmt. */ - vec_lhs = gimple_get_lhs (SLP_TREE_VEC_STMTS (slp_node)[vec_entry]); + gimple *vec_stmt = SLP_TREE_VEC_STMTS (slp_node)[vec_entry]->stmt; + if (gphi *phi = dyn_cast <gphi *> (vec_stmt)) + vec_lhs = gimple_phi_result (phi); + else + vec_lhs = gimple_get_lhs (vec_stmt); /* Get entry to use. */ bitstart = bitsize_int (vec_index); @@ -7099,26 +7835,48 @@ else { enum vect_def_type dt = STMT_VINFO_DEF_TYPE (stmt_info); - vec_lhs = vect_get_vec_def_for_operand_1 (stmt, dt); + vec_lhs = vect_get_vec_def_for_operand_1 (stmt_info, dt); + gcc_checking_assert (ncopies == 1 + || !LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)); /* For multiple copies, get the last copy. */ for (int i = 1; i < ncopies; ++i) - vec_lhs = vect_get_vec_def_for_stmt_copy (vect_unknown_def_type, - vec_lhs); + vec_lhs = vect_get_vec_def_for_stmt_copy (loop_vinfo, vec_lhs); /* Get the last lane in the vector. */ bitstart = int_const_binop (MINUS_EXPR, vec_bitsize, bitsize); } - /* Create a new vectorized stmt for the uses of STMT and insert outside the - loop. */ gimple_seq stmts = NULL; - tree bftype = TREE_TYPE (vectype); - if (VECTOR_BOOLEAN_TYPE_P (vectype)) - bftype = build_nonstandard_integer_type (tree_to_uhwi (bitsize), 1); - tree new_tree = build3 (BIT_FIELD_REF, bftype, vec_lhs, bitsize, bitstart); - new_tree = force_gimple_operand (fold_convert (lhs_type, new_tree), &stmts, - true, NULL_TREE); + tree new_tree; + if (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) + { + /* Emit: + + SCALAR_RES = EXTRACT_LAST <VEC_LHS, MASK> + + where VEC_LHS is the vectorized live-out result and MASK is + the loop mask for the final iteration. */ + gcc_assert (ncopies == 1 && !slp_node); + tree scalar_type = TREE_TYPE (STMT_VINFO_VECTYPE (stmt_info)); + tree mask = vect_get_loop_mask (gsi, &LOOP_VINFO_MASKS (loop_vinfo), + 1, vectype, 0); + tree scalar_res = gimple_build (&stmts, CFN_EXTRACT_LAST, + scalar_type, mask, vec_lhs); + + /* Convert the extracted vector element to the required scalar type. */ + new_tree = gimple_convert (&stmts, lhs_type, scalar_res); + } + else + { + tree bftype = TREE_TYPE (vectype); + if (VECTOR_BOOLEAN_TYPE_P (vectype)) + bftype = build_nonstandard_integer_type (tree_to_uhwi (bitsize), 1); + new_tree = build3 (BIT_FIELD_REF, bftype, vec_lhs, bitsize, bitstart); + new_tree = force_gimple_operand (fold_convert (lhs_type, new_tree), + &stmts, true, NULL_TREE); + } + if (stmts) gsi_insert_seq_on_edge_immediate (single_exit (loop), stmts); @@ -7146,17 +7904,17 @@ return true; } -/* Kill any debug uses outside LOOP of SSA names defined in STMT. */ +/* Kill any debug uses outside LOOP of SSA names defined in STMT_INFO. */ static void -vect_loop_kill_debug_uses (struct loop *loop, gimple *stmt) +vect_loop_kill_debug_uses (struct loop *loop, stmt_vec_info stmt_info) { ssa_op_iter op_iter; imm_use_iterator imm_iter; def_operand_p def_p; gimple *ustmt; - FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) + FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt_info->stmt, op_iter, SSA_OP_DEF) { FOR_EACH_IMM_USE_STMT (ustmt, imm_iter, DEF_FROM_PTR (def_p)) { @@ -7218,6 +7976,97 @@ return false; } +/* Return a mask type with half the number of elements as TYPE. */ + +tree +vect_halve_mask_nunits (tree type) +{ + poly_uint64 nunits = exact_div (TYPE_VECTOR_SUBPARTS (type), 2); + return build_truth_vector_type (nunits, current_vector_size); +} + +/* Return a mask type with twice as many elements as TYPE. */ + +tree +vect_double_mask_nunits (tree type) +{ + poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (type) * 2; + return build_truth_vector_type (nunits, current_vector_size); +} + +/* Record that a fully-masked version of LOOP_VINFO would need MASKS to + contain a sequence of NVECTORS masks that each control a vector of type + VECTYPE. */ + +void +vect_record_loop_mask (loop_vec_info loop_vinfo, vec_loop_masks *masks, + unsigned int nvectors, tree vectype) +{ + gcc_assert (nvectors != 0); + if (masks->length () < nvectors) + masks->safe_grow_cleared (nvectors); + rgroup_masks *rgm = &(*masks)[nvectors - 1]; + /* The number of scalars per iteration and the number of vectors are + both compile-time constants. */ + unsigned int nscalars_per_iter + = exact_div (nvectors * TYPE_VECTOR_SUBPARTS (vectype), + LOOP_VINFO_VECT_FACTOR (loop_vinfo)).to_constant (); + if (rgm->max_nscalars_per_iter < nscalars_per_iter) + { + rgm->max_nscalars_per_iter = nscalars_per_iter; + rgm->mask_type = build_same_sized_truth_vector_type (vectype); + } +} + +/* Given a complete set of masks MASKS, extract mask number INDEX + for an rgroup that operates on NVECTORS vectors of type VECTYPE, + where 0 <= INDEX < NVECTORS. Insert any set-up statements before GSI. + + See the comment above vec_loop_masks for more details about the mask + arrangement. */ + +tree +vect_get_loop_mask (gimple_stmt_iterator *gsi, vec_loop_masks *masks, + unsigned int nvectors, tree vectype, unsigned int index) +{ + rgroup_masks *rgm = &(*masks)[nvectors - 1]; + tree mask_type = rgm->mask_type; + + /* Populate the rgroup's mask array, if this is the first time we've + used it. */ + if (rgm->masks.is_empty ()) + { + rgm->masks.safe_grow_cleared (nvectors); + for (unsigned int i = 0; i < nvectors; ++i) + { + tree mask = make_temp_ssa_name (mask_type, NULL, "loop_mask"); + /* Provide a dummy definition until the real one is available. */ + SSA_NAME_DEF_STMT (mask) = gimple_build_nop (); + rgm->masks[i] = mask; + } + } + + tree mask = rgm->masks[index]; + if (maybe_ne (TYPE_VECTOR_SUBPARTS (mask_type), + TYPE_VECTOR_SUBPARTS (vectype))) + { + /* A loop mask for data type X can be reused for data type Y + if X has N times more elements than Y and if Y's elements + are N times bigger than X's. In this case each sequence + of N elements in the loop mask will be all-zero or all-one. + We can then view-convert the mask so that each sequence of + N elements is replaced by a single element. */ + gcc_assert (multiple_p (TYPE_VECTOR_SUBPARTS (mask_type), + TYPE_VECTOR_SUBPARTS (vectype))); + gimple_seq seq = NULL; + mask_type = build_same_sized_truth_vector_type (vectype); + mask = gimple_build (&seq, VIEW_CONVERT_EXPR, mask_type, mask); + if (seq) + gsi_insert_seq_before (gsi, seq, GSI_SAME_STMT); + } + return mask; +} + /* Scale profiling counters by estimation for LOOP which is vectorized by factor VF. */ @@ -7229,20 +8078,14 @@ gcov_type new_est_niter = niter_for_unrolled_loop (loop, vf); profile_count freq_h = loop->header->count, freq_e = preheader->count (); - /* Use frequency only if counts are zero. */ - if (!(freq_h > 0) && !(freq_e > 0)) - { - freq_h = profile_count::from_gcov_type (loop->header->frequency); - freq_e = profile_count::from_gcov_type (EDGE_FREQUENCY (preheader)); - } - if (freq_h > 0) + if (freq_h.nonzero_p ()) { profile_probability p; /* Avoid dropping loop body profile counter to 0 because of zero count in loop's preheader. */ - if (!(freq_e > profile_count::from_gcov_type (1))) - freq_e = profile_count::from_gcov_type (1); + if (!(freq_e == profile_count::zero ())) + freq_e = freq_e.force_nonzero (); p = freq_e.apply_scale (new_est_niter + 1, 1).probability_in (freq_h); scale_loop_frequencies (loop, p); } @@ -7258,6 +8101,52 @@ scale_bbs_frequencies (&loop->latch, 1, exit_l->probability / prob); } +/* Vectorize STMT_INFO if relevant, inserting any new instructions before GSI. + When vectorizing STMT_INFO as a store, set *SEEN_STORE to its + stmt_vec_info. */ + +static void +vect_transform_loop_stmt (loop_vec_info loop_vinfo, stmt_vec_info stmt_info, + gimple_stmt_iterator *gsi, stmt_vec_info *seen_store) +{ + struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); + poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); + + if (dump_enabled_p ()) + dump_printf_loc (MSG_NOTE, vect_location, + "------>vectorizing statement: %G", stmt_info->stmt); + + if (MAY_HAVE_DEBUG_BIND_STMTS && !STMT_VINFO_LIVE_P (stmt_info)) + vect_loop_kill_debug_uses (loop, stmt_info); + + if (!STMT_VINFO_RELEVANT_P (stmt_info) + && !STMT_VINFO_LIVE_P (stmt_info)) + return; + + if (STMT_VINFO_VECTYPE (stmt_info)) + { + poly_uint64 nunits + = TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info)); + if (!STMT_SLP_TYPE (stmt_info) + && maybe_ne (nunits, vf) + && dump_enabled_p ()) + /* For SLP VF is set according to unrolling factor, and not + to vector size, hence for SLP this print is not valid. */ + dump_printf_loc (MSG_NOTE, vect_location, "multiple-types.\n"); + } + + /* Pure SLP statements have already been vectorized. We still need + to apply loop vectorization to hybrid SLP statements. */ + if (PURE_SLP_STMT (stmt_info)) + return; + + if (dump_enabled_p ()) + dump_printf_loc (MSG_NOTE, vect_location, "transform statement.\n"); + + if (vect_transform_stmt (stmt_info, gsi, NULL, NULL)) + *seen_store = stmt_info; +} + /* Function vect_transform_loop. The analysis phase has determined that the loop is vectorizable. @@ -7273,27 +8162,26 @@ basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); int nbbs = loop->num_nodes; int i; - tree niters_vector = NULL; - int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); - bool grouped_store; - bool slp_scheduled = false; - gimple *stmt, *pattern_stmt; - gimple_seq pattern_def_seq = NULL; - gimple_stmt_iterator pattern_def_si = gsi_none (); - bool transform_pattern_stmt = false; + tree niters_vector = NULL_TREE; + tree step_vector = NULL_TREE; + tree niters_vector_mult_vf = NULL_TREE; + poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); + unsigned int lowest_vf = constant_lower_bound (vf); + gimple *stmt; bool check_profitability = false; - int th; - - if (dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, "=== vec_transform_loop ===\n"); + unsigned int th; + + DUMP_VECT_SCOPE ("vec_transform_loop"); + + loop_vinfo->shared->check_datarefs (); /* Use the more conservative vectorization threshold. If the number of iterations is constant assume the cost check has been performed by our caller. If the threshold makes all loops profitable that - run at least the vectorization factor number of times checking - is pointless, too. */ + run at least the (estimated) vectorization factor number of times + checking is pointless, too. */ th = LOOP_VINFO_COST_MODEL_THRESHOLD (loop_vinfo); - if (th >= LOOP_VINFO_VECT_FACTOR (loop_vinfo) + if (th >= vect_vf_for_cost (loop_vinfo) && !LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)) { if (dump_enabled_p ()) @@ -7318,7 +8206,17 @@ if (LOOP_REQUIRES_VERSIONING (loop_vinfo)) { - vect_loop_versioning (loop_vinfo, th, check_profitability); + poly_uint64 versioning_threshold + = LOOP_VINFO_VERSIONING_THRESHOLD (loop_vinfo); + if (check_profitability + && ordered_p (poly_uint64 (th), versioning_threshold)) + { + versioning_threshold = ordered_max (poly_uint64 (th), + versioning_threshold); + check_profitability = false; + } + vect_loop_versioning (loop_vinfo, th, check_profitability, + versioning_threshold); check_profitability = false; } @@ -7342,17 +8240,24 @@ LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo) = niters; tree nitersm1 = unshare_expr (LOOP_VINFO_NITERSM1 (loop_vinfo)); bool niters_no_overflow = loop_niters_no_overflow (loop_vinfo); - epilogue = vect_do_peeling (loop_vinfo, niters, nitersm1, &niters_vector, th, + epilogue = vect_do_peeling (loop_vinfo, niters, nitersm1, &niters_vector, + &step_vector, &niters_vector_mult_vf, th, check_profitability, niters_no_overflow); + if (niters_vector == NULL_TREE) { - if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)) - niters_vector - = build_int_cst (TREE_TYPE (LOOP_VINFO_NITERS (loop_vinfo)), - LOOP_VINFO_INT_NITERS (loop_vinfo) / vf); + if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) + && !LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) + && known_eq (lowest_vf, vf)) + { + niters_vector + = build_int_cst (TREE_TYPE (LOOP_VINFO_NITERS (loop_vinfo)), + LOOP_VINFO_INT_NITERS (loop_vinfo) / lowest_vf); + step_vector = build_one_cst (TREE_TYPE (niters)); + } else vect_gen_vector_loop_niters (loop_vinfo, niters, &niters_vector, - niters_no_overflow); + &step_vector, niters_no_overflow); } /* 1) Make sure the loop header has exactly two entries @@ -7362,6 +8267,19 @@ split_edge (loop_preheader_edge (loop)); + if (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) + && vect_use_loop_mask_for_alignment_p (loop_vinfo)) + /* This will deal with any possible peeling. */ + vect_prepare_for_masked_peels (loop_vinfo); + + /* Schedule the SLP instances first, then handle loop vectorization + below. */ + if (!loop_vinfo->slp_instances.is_empty ()) + { + DUMP_VECT_SCOPE ("scheduling SLP instances"); + vect_schedule_slp (loop_vinfo); + } + /* FORNOW: the vectorizer supports only loops which body consist of one basic block (header + empty latch). When the vectorizer will support more involved loop forms, the order by which the BBs are @@ -7377,25 +8295,22 @@ { gphi *phi = si.phi (); if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, - "------>vectorizing phi: "); - dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0); - } - stmt_info = vinfo_for_stmt (phi); + dump_printf_loc (MSG_NOTE, vect_location, + "------>vectorizing phi: %G", phi); + stmt_info = loop_vinfo->lookup_stmt (phi); if (!stmt_info) continue; - if (MAY_HAVE_DEBUG_STMTS && !STMT_VINFO_LIVE_P (stmt_info)) - vect_loop_kill_debug_uses (loop, phi); + if (MAY_HAVE_DEBUG_BIND_STMTS && !STMT_VINFO_LIVE_P (stmt_info)) + vect_loop_kill_debug_uses (loop, stmt_info); if (!STMT_VINFO_RELEVANT_P (stmt_info) && !STMT_VINFO_LIVE_P (stmt_info)) continue; if (STMT_VINFO_VECTYPE (stmt_info) - && (TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info)) - != (unsigned HOST_WIDE_INT) vf) + && (maybe_ne + (TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info)), vf)) && dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "multiple-types.\n"); @@ -7406,207 +8321,98 @@ { if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "transform phi.\n"); - vect_transform_stmt (phi, NULL, NULL, NULL, NULL); + vect_transform_stmt (stmt_info, NULL, NULL, NULL); } } - pattern_stmt = NULL; for (gimple_stmt_iterator si = gsi_start_bb (bb); - !gsi_end_p (si) || transform_pattern_stmt;) - { - bool is_store; - - if (transform_pattern_stmt) - stmt = pattern_stmt; - else + !gsi_end_p (si);) + { + stmt = gsi_stmt (si); + /* During vectorization remove existing clobber stmts. */ + if (gimple_clobber_p (stmt)) + { + unlink_stmt_vdef (stmt); + gsi_remove (&si, true); + release_defs (stmt); + } + else { - stmt = gsi_stmt (si); - /* During vectorization remove existing clobber stmts. */ - if (gimple_clobber_p (stmt)) + stmt_info = loop_vinfo->lookup_stmt (stmt); + + /* vector stmts created in the outer-loop during vectorization of + stmts in an inner-loop may not have a stmt_info, and do not + need to be vectorized. */ + stmt_vec_info seen_store = NULL; + if (stmt_info) { - unlink_stmt_vdef (stmt); - gsi_remove (&si, true); - release_defs (stmt); - continue; + if (STMT_VINFO_IN_PATTERN_P (stmt_info)) + { + gimple *def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info); + for (gimple_stmt_iterator subsi = gsi_start (def_seq); + !gsi_end_p (subsi); gsi_next (&subsi)) + { + stmt_vec_info pat_stmt_info + = loop_vinfo->lookup_stmt (gsi_stmt (subsi)); + vect_transform_loop_stmt (loop_vinfo, pat_stmt_info, + &si, &seen_store); + } + stmt_vec_info pat_stmt_info + = STMT_VINFO_RELATED_STMT (stmt_info); + vect_transform_loop_stmt (loop_vinfo, pat_stmt_info, &si, + &seen_store); + } + vect_transform_loop_stmt (loop_vinfo, stmt_info, &si, + &seen_store); + } + gsi_next (&si); + if (seen_store) + { + if (STMT_VINFO_GROUPED_ACCESS (seen_store)) + /* Interleaving. If IS_STORE is TRUE, the + vectorization of the interleaving chain was + completed - free all the stores in the chain. */ + vect_remove_stores (DR_GROUP_FIRST_ELEMENT (seen_store)); + else + /* Free the attached stmt_vec_info and remove the stmt. */ + loop_vinfo->remove_stmt (stmt_info); } } - - if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, - "------>vectorizing statement: "); - dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); - } - - stmt_info = vinfo_for_stmt (stmt); - - /* vector stmts created in the outer-loop during vectorization of - stmts in an inner-loop may not have a stmt_info, and do not - need to be vectorized. */ - if (!stmt_info) + } + + /* Stub out scalar statements that must not survive vectorization. + Doing this here helps with grouped statements, or statements that + are involved in patterns. */ + for (gimple_stmt_iterator gsi = gsi_start_bb (bb); + !gsi_end_p (gsi); gsi_next (&gsi)) + { + gcall *call = dyn_cast <gcall *> (gsi_stmt (gsi)); + if (call && gimple_call_internal_p (call, IFN_MASK_LOAD)) { - gsi_next (&si); - continue; - } - - if (MAY_HAVE_DEBUG_STMTS && !STMT_VINFO_LIVE_P (stmt_info)) - vect_loop_kill_debug_uses (loop, stmt); - - if (!STMT_VINFO_RELEVANT_P (stmt_info) - && !STMT_VINFO_LIVE_P (stmt_info)) - { - if (STMT_VINFO_IN_PATTERN_P (stmt_info) - && (pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info)) - && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) - || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) - { - stmt = pattern_stmt; - stmt_info = vinfo_for_stmt (stmt); - } - else - { - gsi_next (&si); - continue; - } - } - else if (STMT_VINFO_IN_PATTERN_P (stmt_info) - && (pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info)) - && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) - || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) - transform_pattern_stmt = true; - - /* If pattern statement has def stmts, vectorize them too. */ - if (is_pattern_stmt_p (stmt_info)) - { - if (pattern_def_seq == NULL) - { - pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info); - pattern_def_si = gsi_start (pattern_def_seq); - } - else if (!gsi_end_p (pattern_def_si)) - gsi_next (&pattern_def_si); - if (pattern_def_seq != NULL) + tree lhs = gimple_get_lhs (call); + if (!VECTOR_TYPE_P (TREE_TYPE (lhs))) { - gimple *pattern_def_stmt = NULL; - stmt_vec_info pattern_def_stmt_info = NULL; - - while (!gsi_end_p (pattern_def_si)) - { - pattern_def_stmt = gsi_stmt (pattern_def_si); - pattern_def_stmt_info - = vinfo_for_stmt (pattern_def_stmt); - if (STMT_VINFO_RELEVANT_P (pattern_def_stmt_info) - || STMT_VINFO_LIVE_P (pattern_def_stmt_info)) - break; - gsi_next (&pattern_def_si); - } - - if (!gsi_end_p (pattern_def_si)) - { - if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, - "==> vectorizing pattern def " - "stmt: "); - dump_gimple_stmt (MSG_NOTE, TDF_SLIM, - pattern_def_stmt, 0); - } - - stmt = pattern_def_stmt; - stmt_info = pattern_def_stmt_info; - } - else - { - pattern_def_si = gsi_none (); - transform_pattern_stmt = false; - } - } - else - transform_pattern_stmt = false; - } - - if (STMT_VINFO_VECTYPE (stmt_info)) - { - unsigned int nunits - = (unsigned int) - TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info)); - if (!STMT_SLP_TYPE (stmt_info) - && nunits != (unsigned int) vf - && dump_enabled_p ()) - /* For SLP VF is set according to unrolling factor, and not - to vector size, hence for SLP this print is not valid. */ - dump_printf_loc (MSG_NOTE, vect_location, "multiple-types.\n"); - } - - /* SLP. Schedule all the SLP instances when the first SLP stmt is - reached. */ - if (STMT_SLP_TYPE (stmt_info)) - { - if (!slp_scheduled) - { - slp_scheduled = true; - - if (dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, - "=== scheduling SLP instances ===\n"); - - vect_schedule_slp (loop_vinfo); - } - - /* Hybrid SLP stmts must be vectorized in addition to SLP. */ - if (!vinfo_for_stmt (stmt) || PURE_SLP_STMT (stmt_info)) - { - if (!transform_pattern_stmt && gsi_end_p (pattern_def_si)) - { - pattern_def_seq = NULL; - gsi_next (&si); - } - continue; + tree zero = build_zero_cst (TREE_TYPE (lhs)); + gimple *new_stmt = gimple_build_assign (lhs, zero); + gsi_replace (&gsi, new_stmt, true); } } - - /* -------- vectorize statement ------------ */ - if (dump_enabled_p ()) - dump_printf_loc (MSG_NOTE, vect_location, "transform statement.\n"); - - grouped_store = false; - is_store = vect_transform_stmt (stmt, &si, &grouped_store, NULL, NULL); - if (is_store) - { - if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) - { - /* Interleaving. If IS_STORE is TRUE, the vectorization of the - interleaving chain was completed - free all the stores in - the chain. */ - gsi_next (&si); - vect_remove_stores (GROUP_FIRST_ELEMENT (stmt_info)); - } - else - { - /* Free the attached stmt_vec_info and remove the stmt. */ - gimple *store = gsi_stmt (si); - free_stmt_vec_info (store); - unlink_stmt_vdef (store); - gsi_remove (&si, true); - release_defs (store); - } - - /* Stores can only appear at the end of pattern statements. */ - gcc_assert (!transform_pattern_stmt); - pattern_def_seq = NULL; - } - else if (!transform_pattern_stmt && gsi_end_p (pattern_def_si)) - { - pattern_def_seq = NULL; - gsi_next (&si); - } - } /* stmts in BB */ + } } /* BBs in loop */ - slpeel_make_loop_iterate_ntimes (loop, niters_vector); - - scale_profile_for_vect_loop (loop, vf); - + /* The vectorization factor is always > 1, so if we use an IV increment of 1. + a zero NITERS becomes a nonzero NITERS_VECTOR. */ + if (integer_onep (step_vector)) + niters_no_overflow = true; + vect_set_loop_condition (loop, loop_vinfo, niters_vector, step_vector, + niters_vector_mult_vf, !niters_no_overflow); + + unsigned int assumed_vf = vect_vf_for_cost (loop_vinfo); + scale_profile_for_vect_loop (loop, assumed_vf); + + /* True if the final iteration might not handle a full vector's + worth of scalar iterations. */ + bool final_iter_may_be_partial = LOOP_VINFO_FULLY_MASKED_P (loop_vinfo); /* The minimum number of iterations performed by the epilogue. This is 1 when peeling for gaps because we always need a final scalar iteration. */ @@ -7614,18 +8420,41 @@ /* +1 to convert latch counts to loop iteration counts, -min_epilogue_iters to remove iterations that cannot be performed by the vector code. */ - int bias = 1 - min_epilogue_iters; + int bias_for_lowest = 1 - min_epilogue_iters; + int bias_for_assumed = bias_for_lowest; + int alignment_npeels = LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo); + if (alignment_npeels && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) + { + /* When the amount of peeling is known at compile time, the first + iteration will have exactly alignment_npeels active elements. + In the worst case it will have at least one. */ + int min_first_active = (alignment_npeels > 0 ? alignment_npeels : 1); + bias_for_lowest += lowest_vf - min_first_active; + bias_for_assumed += assumed_vf - min_first_active; + } /* In these calculations the "- 1" converts loop iteration counts back to latch counts. */ if (loop->any_upper_bound) loop->nb_iterations_upper_bound - = wi::udiv_floor (loop->nb_iterations_upper_bound + bias, vf) - 1; + = (final_iter_may_be_partial + ? wi::udiv_ceil (loop->nb_iterations_upper_bound + bias_for_lowest, + lowest_vf) - 1 + : wi::udiv_floor (loop->nb_iterations_upper_bound + bias_for_lowest, + lowest_vf) - 1); if (loop->any_likely_upper_bound) loop->nb_iterations_likely_upper_bound - = wi::udiv_floor (loop->nb_iterations_likely_upper_bound + bias, vf) - 1; + = (final_iter_may_be_partial + ? wi::udiv_ceil (loop->nb_iterations_likely_upper_bound + + bias_for_lowest, lowest_vf) - 1 + : wi::udiv_floor (loop->nb_iterations_likely_upper_bound + + bias_for_lowest, lowest_vf) - 1); if (loop->any_estimate) loop->nb_iterations_estimate - = wi::udiv_floor (loop->nb_iterations_estimate + bias, vf) - 1; + = (final_iter_may_be_partial + ? wi::udiv_ceil (loop->nb_iterations_estimate + bias_for_assumed, + assumed_vf) - 1 + : wi::udiv_floor (loop->nb_iterations_estimate + bias_for_assumed, + assumed_vf) - 1); if (dump_enabled_p ()) { @@ -7639,16 +8468,19 @@ dump_printf (MSG_NOTE, "\n"); } else - dump_printf_loc (MSG_NOTE, vect_location, - "LOOP EPILOGUE VECTORIZED (VS=%d)\n", - current_vector_size); + { + dump_printf_loc (MSG_NOTE, vect_location, + "LOOP EPILOGUE VECTORIZED (VS="); + dump_dec (MSG_NOTE, current_vector_size); + dump_printf (MSG_NOTE, ")\n"); + } } /* Free SLP instances here because otherwise stmt reference counting won't work. */ slp_instance instance; FOR_EACH_VEC_ELT (LOOP_VINFO_SLP_INSTANCES (loop_vinfo), i, instance) - vect_free_slp_instance (instance); + vect_free_slp_instance (instance, true); LOOP_VINFO_SLP_INSTANCES (loop_vinfo).release (); /* Clear-up safelen field since its value is invalid after vectorization since vectorized loop can have loop-carried dependencies. */ @@ -7658,30 +8490,39 @@ if (LOOP_VINFO_EPILOGUE_P (loop_vinfo)) epilogue = NULL; + if (!PARAM_VALUE (PARAM_VECT_EPILOGUES_NOMASK)) + epilogue = NULL; + if (epilogue) { - unsigned int vector_sizes - = targetm.vectorize.autovectorize_vector_sizes (); - vector_sizes &= current_vector_size - 1; - - if (!PARAM_VALUE (PARAM_VECT_EPILOGUES_NOMASK)) - epilogue = NULL; - else if (!vector_sizes) - epilogue = NULL; - else if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) - && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) >= 0) - { - int smallest_vec_size = 1 << ctz_hwi (vector_sizes); - int ratio = current_vector_size / smallest_vec_size; - int eiters = LOOP_VINFO_INT_NITERS (loop_vinfo) - - LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo); - eiters = eiters % vf; - - epilogue->nb_iterations_upper_bound = eiters - 1; - - if (eiters < vf / ratio) - epilogue = NULL; - } + auto_vector_sizes vector_sizes; + targetm.vectorize.autovectorize_vector_sizes (&vector_sizes); + unsigned int next_size = 0; + + if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) + && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) >= 0 + && known_eq (vf, lowest_vf)) + { + unsigned int eiters + = (LOOP_VINFO_INT_NITERS (loop_vinfo) + - LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo)); + eiters = eiters % lowest_vf; + epilogue->nb_iterations_upper_bound = eiters - 1; + + unsigned int ratio; + while (next_size < vector_sizes.length () + && !(constant_multiple_p (current_vector_size, + vector_sizes[next_size], &ratio) + && eiters >= lowest_vf / ratio)) + next_size += 1; + } + else + while (next_size < vector_sizes.length () + && maybe_lt (current_vector_size, vector_sizes[next_size])) + next_size += 1; + + if (next_size == vector_sizes.length ()) + epilogue = NULL; } if (epilogue) @@ -7781,7 +8622,7 @@ efalse = make_edge (bb, store_bb, EDGE_FALSE_VALUE); /* Put STORE_BB to likely part. */ efalse->probability = profile_probability::unlikely (); - store_bb->frequency = PROB_ALWAYS - EDGE_FREQUENCY (efalse); + store_bb->count = efalse->count (); make_single_succ_edge (store_bb, join_bb, EDGE_FALLTHRU); if (dom_info_available_p (CDI_DOMINATORS)) set_immediate_dominator (CDI_DOMINATORS, store_bb, bb); @@ -7825,11 +8666,8 @@ /* Setup GSI_TO to the non-empty block start. */ gsi_to = gsi_start_bb (store_bb); if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, - "Move stmt to created bb\n"); - dump_gimple_stmt (MSG_NOTE, TDF_SLIM, last, 0); - } + dump_printf_loc (MSG_NOTE, vect_location, + "Move stmt to created bb\n%G", last); /* Move all stored value producers if possible. */ while (!gsi_end_p (gsi)) { @@ -7893,11 +8731,8 @@ /* Can move STMT1 to STORE_BB. */ if (dump_enabled_p ()) - { - dump_printf_loc (MSG_NOTE, vect_location, - "Move stmt to created bb\n"); - dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt1, 0); - } + dump_printf_loc (MSG_NOTE, vect_location, + "Move stmt to created bb\n%G", stmt1); gsi_move_before (&gsi_from, &gsi_to); /* Shift GSI_TO for further insertion. */ gsi_prev (&gsi_to);