riscv-gnu-toolchain/gcc.git - Unnamed repository; edit this file 'description' to name the repository.

diff options

Diffstat (limited to 'libjava/java/io/Reader.java')

0 files changed, 0 insertions, 0 deletions

/* Loop Vectorization Copyright (C) 2003-2025 Free Software Foundation, Inc. Contributed by Dorit Naishlos <dorit@il.ibm.com> and Ira Rosen <irar@il.ibm.com> This file is part of GCC. GCC is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 3, or (at your option) any later version. GCC is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with GCC; see the file COPYING3. If not see <http://www.gnu.org/licenses/>. */ #define INCLUDE_ALGORITHM #include "config.h" #include "system.h" #include "coretypes.h" #include "backend.h" #include "target.h" #include "rtl.h" #include "tree.h" #include "gimple.h" #include "cfghooks.h" #include "tree-pass.h" #include "ssa.h" #include "optabs-tree.h" #include "memmodel.h" #include "optabs.h" #include "diagnostic-core.h" #include "fold-const.h" #include "stor-layout.h" #include "cfganal.h" #include "gimplify.h" #include "gimple-iterator.h" #include "gimplify-me.h" #include "tree-ssa-loop-ivopts.h" #include "tree-ssa-loop-manip.h" #include "tree-ssa-loop-niter.h" #include "tree-ssa-loop.h" #include "cfgloop.h" #include "tree-scalar-evolution.h" #include "tree-vectorizer.h" #include "gimple-fold.h" #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" #include "case-cfn-macros.h" #include "langhooks.h" #include "opts.h" /* Loop Vectorization Pass. This pass tries to vectorize loops. For example, the vectorizer transforms the following simple loop: short a[N]; short b[N]; short c[N]; int i; for (i=0; i<N; i++){ a[i] = b[i] + c[i]; } as if it was manually vectorized by rewriting the source code into: typedef int __attribute__((mode(V8HI))) v8hi; short a[N]; short b[N]; short c[N]; int i; v8hi *pa = (v8hi*)a, *pb = (v8hi*)b, *pc = (v8hi*)c; v8hi va, vb, vc; for (i=0; i<N/8; i++){ vb = pb[i]; vc = pc[i]; va = vb + vc; pa[i] = va; } The main entry to this pass is vectorize_loops(), in which the vectorizer applies a set of analyses on a given set of loops, followed by the actual vectorization transformation for the loops that had successfully passed the analysis phase. Throughout this pass we make a distinction between two types of data: scalars (which are represented by SSA_NAMES), and memory references ("data-refs"). These two types of data require different handling both during analysis and transformation. The types of data-refs that the vectorizer currently supports are ARRAY_REFS which base is an array DECL (not a pointer), and INDIRECT_REFS through pointers; both array and pointer accesses are required to have a simple (consecutive) access pattern. Analysis phase: =============== The driver for the analysis phase is vect_analyze_loop(). It applies a set of analyses, some of which rely on the scalar evolution analyzer (scev) developed by Sebastian Pop. During the analysis phase the vectorizer records some information per stmt in a "stmt_vec_info" struct which is attached to each stmt in the loop, as well as general information about the loop as a whole, which is recorded in a "loop_vec_info" struct attached to each loop. Transformation phase: ===================== The loop transformation phase scans all the stmts in the loop, and creates a vector stmt (or a sequence of stmts) for each scalar stmt S in the loop that needs to be vectorized. It inserts the vector code sequence just before the scalar stmt S, and records a pointer to the vector code in STMT_VINFO_VEC_STMT (stmt_info) (stmt_info is the stmt_vec_info struct attached to S). This pointer will be used for the vectorization of following stmts which use the def of stmt S. Stmt S is removed if it writes to memory; otherwise, we rely on dead code elimination for removing it. For example, say stmt S1 was vectorized into stmt VS1: VS1: vb = px[i]; S1: b = x[i]; STMT_VINFO_VEC_STMT (stmt_info (S1)) = VS1 S2: a = b; To vectorize stmt S2, the vectorizer first finds the stmt that defines the operand 'b' (S1), and gets the relevant vector def 'vb' from the vector stmt VS1 pointed to by STMT_VINFO_VEC_STMT (stmt_info (S1)). The resulting sequence would be: VS1: vb = px[i]; S1: b = x[i]; STMT_VINFO_VEC_STMT (stmt_info (S1)) = VS1 VS2: va = vb; S2: a = b; STMT_VINFO_VEC_STMT (stmt_info (S2)) = VS2 Operands that are not SSA_NAMEs, are data-refs that appear in load/store operations (like 'x[i]' in S1), and are handled differently. Target modeling: ================= Currently the only target specific information that is used is the size of the vector (in bytes) - "TARGET_VECTORIZE_UNITS_PER_SIMD_WORD". Targets that can support different sizes of vectors, for now will need to specify one value for "TARGET_VECTORIZE_UNITS_PER_SIMD_WORD". More flexibility will be added in the future. Since we only vectorize operations which vector form can be expressed using existing tree codes, to verify that an operation is supported, the vectorizer checks the relevant optab at the relevant machine_mode (e.g, optab_handler (add_optab, V8HImode)). If the value found is CODE_FOR_nothing, then there's no target support, and we can't vectorize the stmt. For additional information on this project see: http://gcc.gnu.org/projects/tree-ssa/vectorization.html */ static void vect_estimate_min_profitable_iters (loop_vec_info, int *, int *, unsigned *); static stmt_vec_info vect_is_simple_reduction (loop_vec_info, stmt_vec_info, bool *, bool *, bool); /* Function vect_is_simple_iv_evolution. FORNOW: A simple evolution of an induction variables in the loop is considered a polynomial evolution. */ static bool vect_is_simple_iv_evolution (unsigned loop_nb, tree access_fn, tree * init, tree * step) { tree init_expr; tree step_expr; tree evolution_part = evolution_part_in_loop_num (access_fn, loop_nb); basic_block bb; /* When there is no evolution in this loop, the evolution function is not "simple". */ if (evolution_part == NULL_TREE) return false; /* When the evolution is a polynomial of degree >= 2 the evolution function is not "simple". */ if (tree_is_chrec (evolution_part)) return false; step_expr = evolution_part; 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: %T, init: %T\n", step_expr, init_expr); *init = init_expr; *step = step_expr; if (TREE_CODE (step_expr) != INTEGER_CST && (TREE_CODE (step_expr) != SSA_NAME || ((bb = gimple_bb (SSA_NAME_DEF_STMT (step_expr))) && flow_bb_inside_loop_p (get_loop (cfun, loop_nb), bb)) || (!INTEGRAL_TYPE_P (TREE_TYPE (step_expr)) && (!SCALAR_FLOAT_TYPE_P (TREE_TYPE (step_expr)) || !flag_associative_math))) && (TREE_CODE (step_expr) != REAL_CST || !flag_associative_math)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "step unknown.\n"); return false; } return true; } /* Function vect_is_nonlinear_iv_evolution Only support nonlinear induction for integer type 1. neg 2. mul by constant 3. lshift/rshift by constant. For neg induction, return a fake step as integer -1. */ static bool vect_is_nonlinear_iv_evolution (class loop* loop, stmt_vec_info stmt_info, gphi* loop_phi_node, tree *init, tree *step) { tree init_expr, ev_expr, result, op1, op2; gimple* def; if (gimple_phi_num_args (loop_phi_node) != 2) return false; init_expr = PHI_ARG_DEF_FROM_EDGE (loop_phi_node, loop_preheader_edge (loop)); ev_expr = PHI_ARG_DEF_FROM_EDGE (loop_phi_node, loop_latch_edge (loop)); /* Support nonlinear induction only for integer type. */ if (!INTEGRAL_TYPE_P (TREE_TYPE (init_expr))) return false; *init = init_expr; result = PHI_RESULT (loop_phi_node); if (TREE_CODE (ev_expr) != SSA_NAME || ((def = SSA_NAME_DEF_STMT (ev_expr)), false) || !is_gimple_assign (def)) return false; enum tree_code t_code = gimple_assign_rhs_code (def); switch (t_code) { case NEGATE_EXPR: if (gimple_assign_rhs1 (def) != result) return false; *step = build_int_cst (TREE_TYPE (init_expr), -1); STMT_VINFO_LOOP_PHI_EVOLUTION_TYPE (stmt_info) = vect_step_op_neg; break; case RSHIFT_EXPR: case LSHIFT_EXPR: case MULT_EXPR: op1 = gimple_assign_rhs1 (def); op2 = gimple_assign_rhs2 (def); if (TREE_CODE (op2) != INTEGER_CST || op1 != result) return false; *step = op2; if (t_code == LSHIFT_EXPR) STMT_VINFO_LOOP_PHI_EVOLUTION_TYPE (stmt_info) = vect_step_op_shl; else if (t_code == RSHIFT_EXPR) STMT_VINFO_LOOP_PHI_EVOLUTION_TYPE (stmt_info) = vect_step_op_shr; /* NEGATE_EXPR and MULT_EXPR are both vect_step_op_mul. */ else STMT_VINFO_LOOP_PHI_EVOLUTION_TYPE (stmt_info) = vect_step_op_mul; break; default: return false; } STMT_VINFO_LOOP_PHI_EVOLUTION_BASE_UNCHANGED (stmt_info) = *init; STMT_VINFO_LOOP_PHI_EVOLUTION_PART (stmt_info) = *step; 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 (loop_vec_info loop_vinfo, gphi *phi) { 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; } /* Returns true if Phi is a first-order recurrence. A first-order recurrence is a non-reduction recurrence relation in which the value of the recurrence in the current loop iteration equals a value defined in the previous iteration. */ static bool vect_phi_first_order_recurrence_p (loop_vec_info loop_vinfo, class loop *loop, gphi *phi) { /* A nested cycle isn't vectorizable as first order recurrence. */ if (LOOP_VINFO_LOOP (loop_vinfo) != loop) return false; /* Ensure the loop latch definition is from within the loop. */ edge latch = loop_latch_edge (loop); tree ldef = PHI_ARG_DEF_FROM_EDGE (phi, latch); if (TREE_CODE (ldef) != SSA_NAME || SSA_NAME_IS_DEFAULT_DEF (ldef) || is_a <gphi *> (SSA_NAME_DEF_STMT (ldef)) || !flow_bb_inside_loop_p (loop, gimple_bb (SSA_NAME_DEF_STMT (ldef)))) return false; tree def = gimple_phi_result (phi); /* Ensure every use_stmt of the phi node is dominated by the latch definition. */ imm_use_iterator imm_iter; use_operand_p use_p; FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def) if (!is_gimple_debug (USE_STMT (use_p)) && (SSA_NAME_DEF_STMT (ldef) == USE_STMT (use_p) || !vect_stmt_dominates_stmt_p (SSA_NAME_DEF_STMT (ldef), USE_STMT (use_p)))) return false; /* First-order recurrence autovectorization needs shuffle vector. */ tree scalar_type = TREE_TYPE (def); tree vectype = get_vectype_for_scalar_type (loop_vinfo, scalar_type); if (!vectype) return false; return true; } /* Function vect_analyze_scalar_cycles_1. Examine the cross iteration def-use cycles of scalar variables in LOOP. LOOP_VINFO represents the loop that is now being considered for vectorization (can be LOOP, or an outer-loop enclosing LOOP). SLP indicates there will be some subsequent slp analyses or not. */ static void vect_analyze_scalar_cycles_1 (loop_vec_info loop_vinfo, class loop *loop, bool slp) { basic_block bb = loop->header; tree init, step; auto_vec<stmt_vec_info, 64> worklist; gphi_iterator gsi; bool double_reduc, reduc_chain; 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 changed. */ for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) { gphi *phi = gsi.phi (); tree access_fn = NULL; tree def = PHI_RESULT (phi); stmt_vec_info stmt_vinfo = loop_vinfo->lookup_stmt (phi); if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "Analyze phi: %G", (gimple *) phi); /* Skip virtual phi's. The data dependences that are associated with virtual defs/uses (i.e., memory accesses) are analyzed elsewhere. */ if (virtual_operand_p (def)) continue; STMT_VINFO_DEF_TYPE (stmt_vinfo) = vect_unknown_def_type; /* Analyze the evolution function. */ access_fn = analyze_scalar_evolution (loop, def); if (access_fn) { STRIP_NOPS (access_fn); if (dump_enabled_p ()) 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) = evolution_part_in_loop_num (access_fn, loop->num); } if ((!access_fn || vect_inner_phi_in_double_reduction_p (loop_vinfo, phi) || !vect_is_simple_iv_evolution (loop->num, access_fn, &init, &step) || (LOOP_VINFO_LOOP (loop_vinfo) != loop && TREE_CODE (step) != INTEGER_CST)) /* Only handle nonlinear iv for same loop. */ && (LOOP_VINFO_LOOP (loop_vinfo) != loop || !vect_is_nonlinear_iv_evolution (loop, stmt_vinfo, phi, &init, &step))) { worklist.safe_push (stmt_vinfo); continue; } gcc_assert (STMT_VINFO_LOOP_PHI_EVOLUTION_BASE_UNCHANGED (stmt_vinfo) != NULL_TREE); gcc_assert (STMT_VINFO_LOOP_PHI_EVOLUTION_PART (stmt_vinfo) != NULL_TREE); if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "Detected induction.\n"); STMT_VINFO_DEF_TYPE (stmt_vinfo) = vect_induction_def; /* Mark if we have a non-linear IV. */ LOOP_VINFO_NON_LINEAR_IV (loop_vinfo) = STMT_VINFO_LOOP_PHI_EVOLUTION_TYPE (stmt_vinfo) != vect_step_op_add; } /* Second - identify all reductions and nested cycles. */ while (worklist.length () > 0) { stmt_vec_info stmt_vinfo = worklist.pop (); gphi *phi = as_a <gphi *> (stmt_vinfo->stmt); tree def = PHI_RESULT (phi); if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "Analyze phi: %G", (gimple *) phi); gcc_assert (!virtual_operand_p (def) && STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_unknown_def_type); stmt_vec_info reduc_stmt_info = vect_is_simple_reduction (loop_vinfo, stmt_vinfo, &double_reduc, &reduc_chain, slp); if (reduc_stmt_info) { STMT_VINFO_REDUC_DEF (stmt_vinfo) = reduc_stmt_info; STMT_VINFO_REDUC_DEF (reduc_stmt_info) = stmt_vinfo; if (double_reduc) { if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "Detected double reduction.\n"); STMT_VINFO_DEF_TYPE (stmt_vinfo) = vect_double_reduction_def; STMT_VINFO_DEF_TYPE (reduc_stmt_info) = vect_double_reduction_def; /* Make it accessible for SLP vectorization. */ LOOP_VINFO_REDUCTIONS (loop_vinfo).safe_push (reduc_stmt_info); } else { if (loop != LOOP_VINFO_LOOP (loop_vinfo)) { if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "Detected vectorizable nested cycle.\n"); STMT_VINFO_DEF_TYPE (stmt_vinfo) = vect_nested_cycle; } else { if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "Detected reduction.\n"); STMT_VINFO_DEF_TYPE (stmt_vinfo) = 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 (! reduc_chain) LOOP_VINFO_REDUCTIONS (loop_vinfo).safe_push (reduc_stmt_info); } } } else if (vect_phi_first_order_recurrence_p (loop_vinfo, loop, phi)) STMT_VINFO_DEF_TYPE (stmt_vinfo) = vect_first_order_recurrence; else if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "Unknown def-use cycle pattern.\n"); } } /* Function vect_analyze_scalar_cycles. Examine the cross iteration def-use cycles of scalar variables, by analyzing the loop-header PHIs of scalar variables. Classify each cycle as one of the following: invariant, induction, reduction, unknown. We do that for the loop represented by LOOP_VINFO, and also to its inner-loop, if exists. Examples for scalar cycles: Example1: reduction: loop1: for (i=0; i<N; i++) sum += a[i]; Example2: induction: loop2: for (i=0; i<N; i++) a[i] = i; */ static void vect_analyze_scalar_cycles (loop_vec_info loop_vinfo, bool slp) { class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); vect_analyze_scalar_cycles_1 (loop_vinfo, loop, slp); /* When vectorizing an outer-loop, the inner-loop is executed sequentially. Reductions in such inner-loop therefore have different properties than the reductions in the nest that gets vectorized: 1. When vectorized, they are executed in the same order as in the original scalar loop, so we can't change the order of computation when vectorizing them. 2. FIXME: Inner-loop reductions can be used in the inner-loop, so the current checks are too strict. */ if (loop->inner) vect_analyze_scalar_cycles_1 (loop_vinfo, loop->inner, slp); } /* Transfer group and reduction information from STMT_INFO to its pattern stmt. */ static void vect_fixup_reduc_chain (stmt_vec_info stmt_info) { 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 (stmt_info); gcc_checking_assert (STMT_VINFO_DEF_TYPE (stmtp) == STMT_VINFO_DEF_TYPE (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); } /* Fixup scalar cycles that now have their stmts detected as patterns. */ static void vect_fixup_scalar_cycles_with_patterns (loop_vec_info loop_vinfo) { stmt_vec_info first; unsigned i; FOR_EACH_VEC_ELT (LOOP_VINFO_REDUCTION_CHAINS (loop_vinfo), i, first) { stmt_vec_info next = REDUC_GROUP_NEXT_ELEMENT (first); while (next) { if ((STMT_VINFO_IN_PATTERN_P (next) != STMT_VINFO_IN_PATTERN_P (first)) || STMT_VINFO_REDUC_IDX (vect_stmt_to_vectorize (next)) == -1) break; next = REDUC_GROUP_NEXT_ELEMENT (next); } /* If all reduction chain members are well-formed patterns adjust the group to group the pattern stmts instead. */ if (! next && STMT_VINFO_REDUC_IDX (vect_stmt_to_vectorize (first)) != -1) { if (STMT_VINFO_IN_PATTERN_P (first)) { vect_fixup_reduc_chain (first); LOOP_VINFO_REDUCTION_CHAINS (loop_vinfo)[i] = STMT_VINFO_RELATED_STMT (first); } } /* If not all stmt in the chain are patterns or if we failed to update STMT_VINFO_REDUC_IDX dissolve the chain and handle it as regular reduction instead. */ else { stmt_vec_info vinfo = first; stmt_vec_info last = NULL; while (vinfo) { next = REDUC_GROUP_NEXT_ELEMENT (vinfo); REDUC_GROUP_FIRST_ELEMENT (vinfo) = NULL; REDUC_GROUP_NEXT_ELEMENT (vinfo) = NULL; last = vinfo; vinfo = next; } STMT_VINFO_DEF_TYPE (vect_stmt_to_vectorize (first)) = vect_internal_def; loop_vinfo->reductions.safe_push (vect_stmt_to_vectorize (last)); LOOP_VINFO_REDUCTION_CHAINS (loop_vinfo).unordered_remove (i); --i; } } } /* Function vect_get_loop_niters. Determine how many iterations the loop is executed and place it in NUMBER_OF_ITERATIONS. Place the number of latch iterations in NUMBER_OF_ITERATIONSM1. Place the condition under which the niter information holds in ASSUMPTIONS. Return the loop exit conditions. */ static vec<gcond *> vect_get_loop_niters (class loop *loop, const_edge main_exit, tree *assumptions, tree *number_of_iterations, tree *number_of_iterationsm1) { auto_vec<edge> exits = get_loop_exit_edges (loop); vec<gcond *> conds; conds.create (exits.length ()); class tree_niter_desc niter_desc; tree niter_assumptions, niter, may_be_zero; *assumptions = boolean_true_node; *number_of_iterationsm1 = chrec_dont_know; *number_of_iterations = chrec_dont_know; DUMP_VECT_SCOPE ("get_loop_niters"); if (exits.is_empty ()) return conds; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "Loop has %d exits.\n", exits.length ()); edge exit; unsigned int i; FOR_EACH_VEC_ELT (exits, i, exit) { gcond *cond = get_loop_exit_condition (exit); if (cond) conds.safe_push (cond); if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "Analyzing exit %d...\n", i); if (exit != main_exit) continue; may_be_zero = NULL_TREE; if (!number_of_iterations_exit_assumptions (loop, exit, &niter_desc, NULL) || chrec_contains_undetermined (niter_desc.niter)) continue; niter_assumptions = niter_desc.assumptions; may_be_zero = niter_desc.may_be_zero; niter = niter_desc.niter; if (may_be_zero && integer_zerop (may_be_zero)) may_be_zero = NULL_TREE; if (may_be_zero) { if (COMPARISON_CLASS_P (may_be_zero)) { /* Try to combine may_be_zero with assumptions, this can simplify computation of niter expression. */ if (niter_assumptions && !integer_nonzerop (niter_assumptions)) niter_assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, niter_assumptions, fold_build1 (TRUTH_NOT_EXPR, boolean_type_node, may_be_zero)); else niter = fold_build3 (COND_EXPR, TREE_TYPE (niter), may_be_zero, build_int_cst (TREE_TYPE (niter), 0), rewrite_to_non_trapping_overflow (niter)); may_be_zero = NULL_TREE; } else if (integer_nonzerop (may_be_zero)) { *number_of_iterationsm1 = build_int_cst (TREE_TYPE (niter), 0); *number_of_iterations = build_int_cst (TREE_TYPE (niter), 1); continue; } else continue; } /* Loop assumptions are based off the normal exit. */ *assumptions = niter_assumptions; *number_of_iterationsm1 = niter; /* We want the number of loop header executions which is the number of latch executions plus one. ??? For UINT_MAX latch executions this number overflows to zero for loops like do { n++; } while (n != 0); */ if (niter && !chrec_contains_undetermined (niter)) { niter = fold_build2 (PLUS_EXPR, TREE_TYPE (niter), unshare_expr (niter), build_int_cst (TREE_TYPE (niter), 1)); if (TREE_CODE (niter) == INTEGER_CST && TREE_CODE (*number_of_iterationsm1) != INTEGER_CST) { /* If we manage to fold niter + 1 into INTEGER_CST even when niter is some complex expression, ensure back *number_of_iterationsm1 is an INTEGER_CST as well. See PR113210. */ *number_of_iterationsm1 = fold_build2 (PLUS_EXPR, TREE_TYPE (niter), niter, build_minus_one_cst (TREE_TYPE (niter))); } } *number_of_iterations = niter; } if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "All loop exits successfully analyzed.\n"); return conds; } /* Determine the main loop exit for the vectorizer. */ edge vec_init_loop_exit_info (class loop *loop) { /* Before we begin we must first determine which exit is the main one and which are auxilary exits. */ auto_vec<edge> exits = get_loop_exit_edges (loop); if (exits.length () == 1) return exits[0]; /* If we have multiple exits we only support counting IV at the moment. Analyze all exits and return the last one we can analyze. */ class tree_niter_desc niter_desc; edge candidate = NULL; for (edge exit : exits) { if (!get_loop_exit_condition (exit)) continue; if (number_of_iterations_exit_assumptions (loop, exit, &niter_desc, NULL) && !chrec_contains_undetermined (niter_desc.niter)) { tree may_be_zero = niter_desc.may_be_zero; if ((integer_zerop (may_be_zero) /* As we are handling may_be_zero that's not false by rewriting niter to may_be_zero ? 0 : niter we require an empty latch. */ || (single_pred_p (loop->latch) && exit->src == single_pred (loop->latch) && (integer_nonzerop (may_be_zero) || COMPARISON_CLASS_P (may_be_zero)))) && (!candidate || dominated_by_p (CDI_DOMINATORS, exit->src, candidate->src))) candidate = exit; } } return candidate; } /* Function bb_in_loop_p Used as predicate for dfs order traversal of the loop bbs. */ static bool bb_in_loop_p (const_basic_block bb, const void *data) { const class loop *const loop = (const class loop *)data; if (flow_bb_inside_loop_p (loop, bb)) return true; return false; } /* 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 (class loop *loop_in, vec_info_shared *shared) : vec_info (vec_info::loop, shared), loop (loop_in), num_itersm1 (NULL_TREE), num_iters (NULL_TREE), num_iters_unchanged (NULL_TREE), num_iters_assumptions (NULL_TREE), vector_costs (nullptr), scalar_costs (nullptr), th (0), versioning_threshold (0), vectorization_factor (0), main_loop_edge (nullptr), skip_main_loop_edge (nullptr), skip_this_loop_edge (nullptr), reusable_accumulators (), suggested_unroll_factor (1), max_vectorization_factor (0), mask_skip_niters (NULL_TREE), mask_skip_niters_pfa_offset (NULL_TREE), rgroup_compare_type (NULL_TREE), simd_if_cond (NULL_TREE), partial_vector_style (vect_partial_vectors_none), unaligned_dr (NULL), peeling_for_alignment (0), ptr_mask (0), max_spec_read_amount (0), nonlinear_iv (false), ivexpr_map (NULL), scan_map (NULL), slp_unrolling_factor (1), inner_loop_cost_factor (param_vect_inner_loop_cost_factor), vectorizable (false), can_use_partial_vectors_p (param_vect_partial_vector_usage != 0), must_use_partial_vectors_p (false), using_partial_vectors_p (false), using_decrementing_iv_p (false), using_select_vl_p (false), epil_using_partial_vectors_p (false), allow_mutual_alignment (false), partial_load_store_bias (0), peeling_for_gaps (false), peeling_for_niter (false), early_breaks (false), user_unroll (false), no_data_dependencies (false), has_mask_store (false), scalar_loop_scaling (profile_probability::uninitialized ()), scalar_loop (NULL), main_loop_info (NULL), orig_loop_info (NULL), epilogue_vinfo (NULL), drs_advanced_by (NULL_TREE), vec_loop_iv_exit (NULL), vec_epilogue_loop_iv_exit (NULL), scalar_loop_iv_exit (NULL) { /* 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 as reversed postorder traversal, so we are safe. */ bbs = XCNEWVEC (basic_block, loop->num_nodes); 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); if (is_gimple_debug (stmt)) continue; add_stmt (stmt); /* If .GOMP_SIMD_LANE call for the current loop has 3 arguments, the third argument is the #pragma omp simd if (x) condition, when 0, loop shouldn't be vectorized, when non-zero constant, it should be vectorized normally, otherwise versioned with vectorized loop done if the condition is non-zero at runtime. */ if (loop_in->simduid && is_gimple_call (stmt) && gimple_call_internal_p (stmt) && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE && gimple_call_num_args (stmt) >= 3 && TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME && (loop_in->simduid == SSA_NAME_VAR (gimple_call_arg (stmt, 0)))) { tree arg = gimple_call_arg (stmt, 2); if (integer_zerop (arg) || TREE_CODE (arg) == SSA_NAME) simd_if_cond = arg; else gcc_assert (integer_nonzerop (arg)); } } } } /* Free all levels of rgroup CONTROLS. */ void release_vec_loop_controls (vec<rgroup_controls> *controls) { rgroup_controls *rgc; unsigned int i; FOR_EACH_VEC_ELT (*controls, i, rgc) rgc->controls.release (); controls->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. */ _loop_vec_info::~_loop_vec_info () { free (bbs); release_vec_loop_controls (&masks.rgc_vec); release_vec_loop_controls (&lens); delete ivexpr_map; delete scan_map; delete scalar_costs; delete vector_costs; /* When we release an epiloge vinfo that we do not intend to use avoid clearing AUX of the main loop which should continue to point to the main loop vinfo since otherwise we'll leak that. */ if (loop->aux == this) 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_controls *rgm; unsigned int i; FOR_EACH_VEC_ELT (LOOP_VINFO_MASKS (loop_vinfo).rgc_vec, i, rgm) if (rgm->type != NULL_TREE && !direct_internal_fn_supported_p (IFN_WHILE_ULT, cmp_type, rgm->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_controls *rgm; FOR_EACH_VEC_ELT (LOOP_VINFO_MASKS (loop_vinfo).rgc_vec, i, rgm) res = MAX (res, rgm->max_nscalars_per_iter); return res; } /* Calculate the minimum precision necessary to represent: MAX_NITERS * FACTOR as an unsigned integer, where MAX_NITERS is the maximum number of loop header iterations for the original scalar form of LOOP_VINFO. */ static unsigned vect_min_prec_for_max_niters (loop_vec_info loop_vinfo, unsigned int factor) { class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); /* 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); /* Work out how many bits we need to represent the limit. */ return wi::min_precision (max_ni * factor, UNSIGNED); } /* True if the loop needs peeling or partial vectors when vectorized. */ static bool vect_need_peeling_or_partial_vectors_p (loop_vec_info loop_vinfo) { unsigned HOST_WIDE_INT const_vf; HOST_WIDE_INT max_niter = likely_max_stmt_executions_int (LOOP_VINFO_LOOP (loop_vinfo)); unsigned th = LOOP_VINFO_COST_MODEL_THRESHOLD (loop_vinfo); if (!th && LOOP_VINFO_ORIG_LOOP_INFO (loop_vinfo)) th = LOOP_VINFO_COST_MODEL_THRESHOLD (LOOP_VINFO_ORIG_LOOP_INFO (loop_vinfo)); 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))) return true; } else if (LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) /* ??? 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 /* We'd like to use LOOP_VINFO_VERSIONING_THRESHOLD but that's only computed later based on our result. The following is the most conservative approximation. */ > (std::max ((unsigned HOST_WIDE_INT) th, const_vf) / const_vf) * const_vf)))) return true; return false; } /* 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_RGROUP_COMPARE_TYPE. */ static bool vect_verify_full_masking (loop_vec_info 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; /* Produce the rgroup controls. */ for (auto mask : LOOP_VINFO_MASKS (loop_vinfo).mask_set) { vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); tree vectype = mask.first; unsigned nvectors = mask.second; if (masks->rgc_vec.length () < nvectors) masks->rgc_vec.safe_grow_cleared (nvectors, true); rgroup_controls *rgm = &(*masks).rgc_vec[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->type = truth_type_for (vectype); rgm->factor = 1; } } unsigned int max_nscalars_per_iter = vect_get_max_nscalars_per_iter (loop_vinfo); /* Work out how many bits we need to represent the limit. */ min_ni_width = vect_min_prec_for_max_niters (loop_vinfo, max_nscalars_per_iter); /* Find a scalar mode for which WHILE_ULT is supported. */ opt_scalar_int_mode cmp_mode_iter; tree cmp_type = NULL_TREE; tree iv_type = NULL_TREE; widest_int iv_limit = vect_iv_limit_for_partial_vectors (loop_vinfo); unsigned int iv_precision = UINT_MAX; if (iv_limit != -1) iv_precision = wi::min_precision (iv_limit * max_nscalars_per_iter, UNSIGNED); 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, there are at least two reasons why that's not always the best choice: - An IV that's Pmode or wider is more likely to be reusable in address calculations than an IV that's narrower than Pmode. - Doing the comparison in IV_PRECISION or wider allows a natural 0-based IV, whereas using a narrower comparison type requires mitigations against wrap-around. Conversely, if the IV limit is variable, doing the comparison in a wider type than the original type can introduce unnecessary extensions, so picking the widest valid mode is not always a good choice either. Here we prefer the first IV type that's Pmode or wider, and the first comparison type that's IV_PRECISION or wider. (The comparison type must be no wider than the IV type, to avoid extensions in the vector loop.) ??? We might want to try continuing beyond Pmode for ILP32 targets if CMP_BITS < IV_PRECISION. */ iv_type = this_type; if (!cmp_type || iv_precision > TYPE_PRECISION (cmp_type)) cmp_type = this_type; if (cmp_bits >= GET_MODE_BITSIZE (Pmode)) break; } } } if (!cmp_type) { LOOP_VINFO_MASKS (loop_vinfo).rgc_vec.release (); return false; } LOOP_VINFO_RGROUP_COMPARE_TYPE (loop_vinfo) = cmp_type; LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo) = iv_type; LOOP_VINFO_PARTIAL_VECTORS_STYLE (loop_vinfo) = vect_partial_vectors_while_ult; return true; } /* Each statement in LOOP_VINFO can be masked where necessary. Check whether we can actually generate AVX512 style masks. Return true if so, storing the type of the scalar IV in LOOP_VINFO_RGROUP_IV_TYPE. */ static bool vect_verify_full_masking_avx512 (loop_vec_info loop_vinfo) { /* Produce differently organized rgc_vec and differently check we can produce masks. */ /* 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; /* For the decrementing IV we need to represent all values in [0, niter + niter_skip] where niter_skip is the elements we skip in the first iteration for prologue peeling. */ tree iv_type = NULL_TREE; widest_int iv_limit = vect_iv_limit_for_partial_vectors (loop_vinfo); unsigned int iv_precision = UINT_MAX; if (iv_limit != -1) iv_precision = wi::min_precision (iv_limit, UNSIGNED); /* First compute the type for the IV we use to track the remaining scalar iterations. */ opt_scalar_int_mode cmp_mode_iter; FOR_EACH_MODE_IN_CLASS (cmp_mode_iter, MODE_INT) { unsigned int cmp_bits = GET_MODE_BITSIZE (cmp_mode_iter.require ()); if (cmp_bits >= iv_precision && targetm.scalar_mode_supported_p (cmp_mode_iter.require ())) { iv_type = build_nonstandard_integer_type (cmp_bits, true); if (iv_type) break; } } if (!iv_type) return false; /* Produce the rgroup controls. */ for (auto const &mask : LOOP_VINFO_MASKS (loop_vinfo).mask_set) { vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); tree vectype = mask.first; unsigned nvectors = mask.second; /* 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 (); /* We index the rgroup_controls vector with nscalars_per_iter which we keep constant and instead have a varying nvectors, remembering the vector mask with the fewest nV. */ if (masks->rgc_vec.length () < nscalars_per_iter) masks->rgc_vec.safe_grow_cleared (nscalars_per_iter, true); rgroup_controls *rgm = &(*masks).rgc_vec[nscalars_per_iter - 1]; if (!rgm->type || rgm->factor > nvectors) { rgm->type = truth_type_for (vectype); rgm->compare_type = NULL_TREE; rgm->max_nscalars_per_iter = nscalars_per_iter; rgm->factor = nvectors; rgm->bias_adjusted_ctrl = NULL_TREE; } } /* There is no fixed compare type we are going to use but we have to be able to get at one for each mask group. */ unsigned int min_ni_width = wi::min_precision (vect_max_vf (loop_vinfo), UNSIGNED); bool ok = true; for (auto &rgc : LOOP_VINFO_MASKS (loop_vinfo).rgc_vec) { tree mask_type = rgc.type; if (!mask_type) continue; /* For now vect_get_loop_mask only supports integer mode masks when we need to split it. */ if (GET_MODE_CLASS (TYPE_MODE (mask_type)) != MODE_INT || TYPE_PRECISION (TREE_TYPE (mask_type)) != 1) { ok = false; break; } /* If iv_type is usable as compare type use that - we can elide the saturation in that case. */ if (TYPE_PRECISION (iv_type) >= min_ni_width) { tree cmp_vectype = build_vector_type (iv_type, TYPE_VECTOR_SUBPARTS (mask_type)); if (expand_vec_cmp_expr_p (cmp_vectype, mask_type, LT_EXPR)) rgc.compare_type = cmp_vectype; } if (!rgc.compare_type) 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 cmp_type = build_nonstandard_integer_type (cmp_bits, true); if (!cmp_type) continue; /* Check whether we can produce the mask with cmp_type. */ tree cmp_vectype = build_vector_type (cmp_type, TYPE_VECTOR_SUBPARTS (mask_type)); if (expand_vec_cmp_expr_p (cmp_vectype, mask_type, LT_EXPR)) { rgc.compare_type = cmp_vectype; break; } } } if (!rgc.compare_type) { ok = false; break; } } if (!ok) { release_vec_loop_controls (&LOOP_VINFO_MASKS (loop_vinfo).rgc_vec); return false; } LOOP_VINFO_RGROUP_COMPARE_TYPE (loop_vinfo) = error_mark_node; LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo) = iv_type; LOOP_VINFO_PARTIAL_VECTORS_STYLE (loop_vinfo) = vect_partial_vectors_avx512; return true; } /* Check whether we can use vector access with length based on precison comparison. So far, to keep it simple, we only allow the case that the precision of the target supported length is larger than the precision required by loop niters. */ static bool vect_verify_loop_lens (loop_vec_info loop_vinfo) { if (LOOP_VINFO_LENS (loop_vinfo).is_empty ()) return false; if (!VECTOR_MODE_P (loop_vinfo->vector_mode)) return false; machine_mode len_load_mode, len_store_mode; if (!get_len_load_store_mode (loop_vinfo->vector_mode, true) .exists (&len_load_mode)) return false; if (!get_len_load_store_mode (loop_vinfo->vector_mode, false) .exists (&len_store_mode)) return false; signed char partial_load_bias = internal_len_load_store_bias (IFN_LEN_LOAD, len_load_mode); signed char partial_store_bias = internal_len_load_store_bias (IFN_LEN_STORE, len_store_mode); gcc_assert (partial_load_bias == partial_store_bias); if (partial_load_bias == VECT_PARTIAL_BIAS_UNSUPPORTED) return false; /* If the backend requires a bias of -1 for LEN_LOAD, we must not emit len_loads with a length of zero. In order to avoid that we prohibit more than one loop length here. */ if (partial_load_bias == -1 && LOOP_VINFO_LENS (loop_vinfo).length () > 1) return false; LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo) = partial_load_bias; unsigned int max_nitems_per_iter = 1; unsigned int i; rgroup_controls *rgl; /* Find the maximum number of items per iteration for every rgroup. */ FOR_EACH_VEC_ELT (LOOP_VINFO_LENS (loop_vinfo), i, rgl) { unsigned nitems_per_iter = rgl->max_nscalars_per_iter * rgl->factor; max_nitems_per_iter = MAX (max_nitems_per_iter, nitems_per_iter); } /* Work out how many bits we need to represent the length limit. */ unsigned int min_ni_prec = vect_min_prec_for_max_niters (loop_vinfo, max_nitems_per_iter); /* Now use the maximum of below precisions for one suitable IV type: - the IV's natural precision - the precision needed to hold: the maximum number of scalar iterations multiplied by the scale factor (min_ni_prec above) - the Pmode precision If min_ni_prec is less than the precision of the current niters, we perfer to still use the niters type. Prefer to use Pmode and wider IV to avoid narrow conversions. */ unsigned int ni_prec = TYPE_PRECISION (TREE_TYPE (LOOP_VINFO_NITERS (loop_vinfo))); min_ni_prec = MAX (min_ni_prec, ni_prec); min_ni_prec = MAX (min_ni_prec, GET_MODE_BITSIZE (Pmode)); tree iv_type = NULL_TREE; opt_scalar_int_mode tmode_iter; FOR_EACH_MODE_IN_CLASS (tmode_iter, MODE_INT) { scalar_mode tmode = tmode_iter.require (); unsigned int tbits = GET_MODE_BITSIZE (tmode); /* ??? Do we really want to construct one IV whose precision exceeds BITS_PER_WORD? */ if (tbits > BITS_PER_WORD) break; /* Find the first available standard integral type. */ if (tbits >= min_ni_prec && targetm.scalar_mode_supported_p (tmode)) { iv_type = build_nonstandard_integer_type (tbits, true); break; } } if (!iv_type) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "can't vectorize with length-based partial vectors" " because there is no suitable iv type.\n"); return false; } LOOP_VINFO_RGROUP_COMPARE_TYPE (loop_vinfo) = iv_type; LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo) = iv_type; LOOP_VINFO_PARTIAL_VECTORS_STYLE (loop_vinfo) = vect_partial_vectors_len; return true; } /* Calculate the cost of one scalar iteration of the loop. */ static void vect_compute_single_scalar_iteration_cost (loop_vec_info loop_vinfo) { class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); int nbbs = loop->num_nodes, factor; int innerloop_iters, i; DUMP_VECT_SCOPE ("vect_compute_single_scalar_iteration_cost"); /* Gather costs for statements in the scalar loop. */ /* FORNOW. */ innerloop_iters = 1; if (loop->inner) innerloop_iters = LOOP_VINFO_INNER_LOOP_COST_FACTOR (loop_vinfo); for (i = 0; i < nbbs; i++) { gimple_stmt_iterator si; basic_block bb = bbs[i]; if (bb->loop_father == loop->inner) factor = innerloop_iters; else factor = 1; for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) { gimple *stmt = gsi_stmt (si); stmt_vec_info stmt_info = loop_vinfo->lookup_stmt (stmt); if (!is_gimple_assign (stmt) && !is_gimple_call (stmt) && !is_a<gcond *> (stmt)) continue; /* Skip stmts that are not vectorized inside the loop. */ stmt_vec_info vstmt_info = vect_stmt_to_vectorize (stmt_info); if (!STMT_VINFO_RELEVANT_P (vstmt_info) && (!STMT_VINFO_LIVE_P (vstmt_info) || !VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (vstmt_info)))) continue; vect_cost_for_stmt kind; if (STMT_VINFO_DATA_REF (stmt_info)) { if (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info))) kind = scalar_load; else kind = scalar_store; } else if (vect_nop_conversion_p (stmt_info)) continue; else kind = scalar_stmt; /* We are using vect_prologue here to avoid scaling twice by the inner loop factor. */ record_stmt_cost (&LOOP_VINFO_SCALAR_ITERATION_COST (loop_vinfo), factor, kind, stmt_info, 0, vect_prologue); } } /* Now accumulate cost. */ loop_vinfo->scalar_costs = init_cost (loop_vinfo, true); add_stmt_costs (loop_vinfo->scalar_costs, &LOOP_VINFO_SCALAR_ITERATION_COST (loop_vinfo)); loop_vinfo->scalar_costs->finish_cost (nullptr); } /* Function vect_analyze_loop_form. Verify that certain CFG restrictions hold, including: - the loop has a pre-header - the loop has a single entry - nested loops can have only a single exit. - the loop exit condition is simple enough - the number of iterations can be analyzed, i.e, a countable loop. The niter could be analyzed under some assumptions. */ opt_result vect_analyze_loop_form (class loop *loop, gimple *loop_vectorized_call, vect_loop_form_info *info) { DUMP_VECT_SCOPE ("vect_analyze_loop_form"); edge exit_e = vec_init_loop_exit_info (loop); if (!exit_e) return opt_result::failure_at (vect_location, "not vectorized:" " could not determine main exit from" " loop with multiple exits.\n"); if (loop_vectorized_call) { tree arg = gimple_call_arg (loop_vectorized_call, 1); class loop *scalar_loop = get_loop (cfun, tree_to_shwi (arg)); edge scalar_exit_e = vec_init_loop_exit_info (scalar_loop); if (!scalar_exit_e) return opt_result::failure_at (vect_location, "not vectorized:" " could not determine main exit from" " loop with multiple exits.\n"); } info->loop_exit = exit_e; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "using as main loop exit: %d -> %d [AUX: %p]\n", exit_e->src->index, exit_e->dest->index, exit_e->aux); /* Check if we have any control flow that doesn't leave the loop. */ basic_block *bbs = get_loop_body (loop); for (unsigned i = 0; i < loop->num_nodes; i++) if (EDGE_COUNT (bbs[i]->succs) != 1 && (EDGE_COUNT (bbs[i]->succs) != 2 || !loop_exits_from_bb_p (bbs[i]->loop_father, bbs[i]))) { free (bbs); return opt_result::failure_at (vect_location, "not vectorized:" " unsupported control flow in loop.\n"); } free (bbs); /* Different restrictions apply when we are considering an inner-most loop, vs. an outer (nested) loop. (FORNOW. May want to relax some of these restrictions in the future). */ info->inner_loop_cond = NULL; if (!loop->inner) { /* Inner-most loop. */ if (empty_block_p (loop->header)) return opt_result::failure_at (vect_location, "not vectorized: empty loop.\n"); } else { class loop *innerloop = loop->inner; edge entryedge; /* Nested loop. We currently require that the loop is doubly-nested, contains a single inner loop with a single exit to the block with the single exit condition in the outer loop. Vectorizable outer-loops look like this: (pre-header) | header <---+ | | inner-loop | | | tail ------+ | (exit-bb) The inner-loop also has the properties expected of inner-most loops as described above. */ if ((loop->inner)->inner || (loop->inner)->next) return opt_result::failure_at (vect_location, "not vectorized:" " multiple nested loops.\n"); entryedge = loop_preheader_edge (innerloop); if (entryedge->src != loop->header || !single_exit (innerloop) || single_exit (innerloop)->dest != EDGE_PRED (loop->latch, 0)->src) return opt_result::failure_at (vect_location, "not vectorized:" " unsupported outerloop form.\n"); /* Analyze the inner-loop. */ vect_loop_form_info inner; opt_result res = vect_analyze_loop_form (loop->inner, NULL, &inner); if (!res) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "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.number_of_iterations)) 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"); info->inner_loop_cond = inner.conds[0]; } 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 latch is empty. */ basic_block latch = loop->latch; do { if (!empty_block_p (latch) || !gimple_seq_empty_p (phi_nodes (latch))) return opt_result::failure_at (vect_location, "not vectorized: latch block not " "empty.\n"); latch = single_pred (latch); } while (single_succ_p (latch)); /* Make sure there is no abnormal exit. */ auto_vec<edge> exits = get_loop_exit_edges (loop); for (edge e : exits) { if (e->flags & EDGE_ABNORMAL) return opt_result::failure_at (vect_location, "not vectorized:" " abnormal loop exit edge.\n"); } info->conds = vect_get_loop_niters (loop, exit_e, &info->assumptions, &info->number_of_iterations, &info->number_of_iterationsm1); if (info->conds.is_empty ()) return opt_result::failure_at (vect_location, "not vectorized: complicated exit condition.\n"); /* Determine what the primary and alternate exit conds are. */ for (unsigned i = 0; i < info->conds.length (); i++) { gcond *cond = info->conds[i]; if (exit_e->src == gimple_bb (cond)) std::swap (info->conds[0], info->conds[i]); } if (integer_zerop (info->assumptions) || !info->number_of_iterations || chrec_contains_undetermined (info->number_of_iterations)) return opt_result::failure_at (info->conds[0], "not vectorized: number of iterations cannot be computed.\n"); if (integer_zerop (info->number_of_iterations)) return opt_result::failure_at (info->conds[0], "not vectorized: number of iterations = 0.\n"); if (!(tree_fits_shwi_p (info->number_of_iterations) && tree_to_shwi (info->number_of_iterations) > 0)) { if (dump_enabled_p ()) { dump_printf_loc (MSG_NOTE, vect_location, "Symbolic number of iterations is "); dump_generic_expr (MSG_NOTE, TDF_DETAILS, info->number_of_iterations); dump_printf (MSG_NOTE, "\n"); } } return opt_result::success (); } /* Create a loop_vec_info for LOOP with SHARED and the vect_analyze_loop_form result. */ loop_vec_info vect_create_loop_vinfo (class loop *loop, vec_info_shared *shared, const vect_loop_form_info *info, loop_vec_info orig_loop_info) { loop_vec_info loop_vinfo = new _loop_vec_info (loop, shared); LOOP_VINFO_NITERSM1 (loop_vinfo) = info->number_of_iterationsm1; LOOP_VINFO_NITERS (loop_vinfo) = info->number_of_iterations; LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo) = info->number_of_iterations; LOOP_VINFO_ORIG_LOOP_INFO (loop_vinfo) = orig_loop_info; if (orig_loop_info && LOOP_VINFO_EPILOGUE_P (orig_loop_info)) LOOP_VINFO_MAIN_LOOP_INFO (loop_vinfo) = LOOP_VINFO_MAIN_LOOP_INFO (orig_loop_info); else LOOP_VINFO_MAIN_LOOP_INFO (loop_vinfo) = orig_loop_info; /* Also record the assumptions for versioning. */ if (!integer_onep (info->assumptions) && !orig_loop_info) LOOP_VINFO_NITERS_ASSUMPTIONS (loop_vinfo) = info->assumptions; for (gcond *cond : info->conds) { stmt_vec_info loop_cond_info = loop_vinfo->lookup_stmt (cond); /* Mark the statement as a condition. */ STMT_VINFO_DEF_TYPE (loop_cond_info) = vect_condition_def; } for (unsigned i = 1; i < info->conds.length (); i ++) LOOP_VINFO_LOOP_CONDS (loop_vinfo).safe_push (info->conds[i]); LOOP_VINFO_LOOP_IV_COND (loop_vinfo) = info->conds[0]; LOOP_VINFO_IV_EXIT (loop_vinfo) = info->loop_exit; /* Check to see if we're vectorizing multiple exits. */ LOOP_VINFO_EARLY_BREAKS (loop_vinfo) = !LOOP_VINFO_LOOP_CONDS (loop_vinfo).is_empty (); if (info->inner_loop_cond) { /* If we have an estimate on the number of iterations of the inner loop use that to limit the scale for costing, otherwise use --param vect-inner-loop-cost-factor literally. */ widest_int nit; if (estimated_stmt_executions (loop->inner, &nit)) LOOP_VINFO_INNER_LOOP_COST_FACTOR (loop_vinfo) = wi::smin (nit, param_vect_inner_loop_cost_factor).to_uhwi (); } return loop_vinfo; } /* Return true if we know that the iteration count is smaller than the vectorization factor. Return false if it isn't, or if we can't be sure either way. */ static bool vect_known_niters_smaller_than_vf (loop_vec_info loop_vinfo) { unsigned int assumed_vf = vect_vf_for_cost (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_VINFO_LOOP (loop_vinfo)); if (max_niter != -1 && (unsigned HOST_WIDE_INT) max_niter < assumed_vf) return true; return false; } /* 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, unsigned *suggested_unroll_factor) { class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); unsigned int assumed_vf = vect_vf_for_cost (loop_vinfo); /* Only loops that can handle partially-populated vectors can have iteration counts less than the vectorization factor. */ if (!LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo) && vect_known_niters_smaller_than_vf (loop_vinfo)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "not vectorized: iteration count smaller than " "vectorization factor.\n"); return 0; } /* If we know the number of iterations we can do better, for the epilogue we can also decide whether the main loop leaves us with enough iterations, prefering a smaller vector epilog then also possibly used for the case we skip the vector loop. */ if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)) { widest_int scalar_niters = wi::to_widest (LOOP_VINFO_NITERSM1 (loop_vinfo)) + 1; if (LOOP_VINFO_EPILOGUE_P (loop_vinfo)) { loop_vec_info orig_loop_vinfo = LOOP_VINFO_ORIG_LOOP_INFO (loop_vinfo); loop_vec_info main_loop_vinfo = LOOP_VINFO_MAIN_LOOP_INFO (loop_vinfo); unsigned lowest_vf = constant_lower_bound (LOOP_VINFO_VECT_FACTOR (orig_loop_vinfo)); int prolog_peeling = 0; if (!vect_use_loop_mask_for_alignment_p (main_loop_vinfo)) prolog_peeling = LOOP_VINFO_PEELING_FOR_ALIGNMENT (main_loop_vinfo); if (prolog_peeling >= 0 && known_eq (LOOP_VINFO_VECT_FACTOR (orig_loop_vinfo), lowest_vf)) { unsigned gap = LOOP_VINFO_PEELING_FOR_GAPS (main_loop_vinfo) ? 1 : 0; scalar_niters = ((scalar_niters - gap - prolog_peeling) % lowest_vf + gap); } } /* Reject vectorizing for a single scalar iteration, even if we could in principle implement that using partial vectors. */ unsigned peeling_gap = LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo); if (scalar_niters <= peeling_gap + 1) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "not vectorized: loop only has a single " "scalar iteration.\n"); return 0; } if (!LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo)) { /* Check that the loop processes at least one full vector. */ poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); if (known_lt (scalar_niters, vf)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "loop does not have enough iterations " "to support vectorization.\n"); return 0; } /* If we need to peel an extra epilogue iteration to handle data accesses with gaps, check that there are enough scalar iterations available. The check above is redundant with this one when peeling for gaps, but the distinction is useful for diagnostics. */ if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) && known_le (scalar_niters, vf)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "loop does not have enough iterations " "to support peeling for gaps.\n"); return 0; } } } /* If using the "very cheap" model. reject cases in which we'd keep a copy of the scalar code (even if we might be able to vectorize it). */ if (loop_cost_model (loop) == VECT_COST_MODEL_VERY_CHEAP && (LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) || LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo))) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "some scalar iterations would need to be peeled\n"); return 0; } int min_profitable_iters, min_profitable_estimate; vect_estimate_min_profitable_iters (loop_vinfo, &min_profitable_iters, &min_profitable_estimate, suggested_unroll_factor); 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"); return -1; } int min_scalar_loop_bound = (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; } /* The static profitablity threshold min_profitable_estimate includes the cost of having to check at runtime whether the scalar loop should be used instead. If it turns out that we don't need or want such a check, the threshold we should use for the static estimate is simply the point at which the vector loop becomes more profitable than the scalar loop. */ if (min_profitable_estimate > min_profitable_iters && !LOOP_REQUIRES_VERSIONING (loop_vinfo) && !LOOP_VINFO_PEELING_FOR_NITER (loop_vinfo) && !LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) && !vect_apply_runtime_profitability_check_p (loop_vinfo)) { if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "no need for a runtime" " choice between the scalar and vector loops\n"); min_profitable_estimate = min_profitable_iters; } /* If the vector loop needs multiple iterations to be beneficial then things are probably too close to call, and the conservative thing would be to stick with the scalar code. */ if (loop_cost_model (loop) == VECT_COST_MODEL_VERY_CHEAP && min_profitable_estimate > (int) vect_vf_for_cost (loop_vinfo)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "one iteration of the vector loop would be" " more expensive than the equivalent number of" " iterations of the scalar loop\n"); return 0; } HOST_WIDE_INT estimated_niter; /* If we are vectorizing an epilogue then we know the maximum number of scalar iterations it will cover is at least one lower than the vectorization factor of the main loop. */ if (LOOP_VINFO_EPILOGUE_P (loop_vinfo)) estimated_niter = vect_vf_for_cost (LOOP_VINFO_ORIG_LOOP_INFO (loop_vinfo)) - 1; else { 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) { 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)) { gimple *stmt = gsi_stmt (gsi); if (is_gimple_debug (stmt)) continue; opt_result res = vect_find_stmt_data_reference (loop, stmt, datarefs, NULL, 0); if (!res) { if (is_gimple_call (stmt) && loop->safelen) { tree fndecl = gimple_call_fndecl (stmt), op; if (fndecl == NULL_TREE && gimple_call_internal_p (stmt, IFN_MASK_CALL)) { fndecl = gimple_call_arg (stmt, 0); gcc_checking_assert (TREE_CODE (fndecl) == ADDR_EXPR); fndecl = TREE_OPERAND (fndecl, 0); gcc_checking_assert (TREE_CODE (fndecl) == FUNCTION_DECL); } if (fndecl != NULL_TREE) { cgraph_node *node = cgraph_node::get (fndecl); if (node != NULL && node->simd_clones != NULL) { unsigned int j, n = gimple_call_num_args (stmt); for (j = 0; j < n; j++) { op = gimple_call_arg (stmt, j); if (DECL_P (op) || (REFERENCE_CLASS_P (op) && get_base_address (op))) break; } op = gimple_call_lhs (stmt); /* Ignore #pragma omp declare simd functions if they don't have data references in the call stmt itself. */ if (j == n && !(op && (DECL_P (op) || (REFERENCE_CLASS_P (op) && get_base_address (op))))) continue; } } } 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_loop_max_datarefs_for_datadeps) return opt_result::failure_at (stmt, "exceeded param " "loop-max-datarefs-for-datadeps\n"); } return opt_result::success (); } /* Determine if operating on full vectors for LOOP_VINFO might leave some scalar iterations still to do. If so, decide how we should handle those scalar iterations. The possibilities are: (1) Make LOOP_VINFO operate on partial vectors instead of full vectors. In this case: LOOP_VINFO_USING_PARTIAL_VECTORS_P == true LOOP_VINFO_EPIL_USING_PARTIAL_VECTORS_P == false LOOP_VINFO_PEELING_FOR_NITER == false (2) Make LOOP_VINFO operate on full vectors and use an epilogue loop to handle the remaining scalar iterations. In this case: LOOP_VINFO_USING_PARTIAL_VECTORS_P == false LOOP_VINFO_PEELING_FOR_NITER == true There are two choices: (2a) Consider vectorizing the epilogue loop at the same VF as the main loop, but using partial vectors instead of full vectors. In this case: LOOP_VINFO_EPIL_USING_PARTIAL_VECTORS_P == true (2b) Consider vectorizing the epilogue loop at lower VFs only. In this case: LOOP_VINFO_EPIL_USING_PARTIAL_VECTORS_P == false */ opt_result vect_determine_partial_vectors_and_peeling (loop_vec_info loop_vinfo) { /* Determine whether there would be any scalar iterations left over. */ bool need_peeling_or_partial_vectors_p = vect_need_peeling_or_partial_vectors_p (loop_vinfo); /* Decide whether to vectorize the loop with partial vectors. */ LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo) = false; LOOP_VINFO_EPIL_USING_PARTIAL_VECTORS_P (loop_vinfo) = false; if (LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) && LOOP_VINFO_MUST_USE_PARTIAL_VECTORS_P (loop_vinfo)) LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo) = true; else if (LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) && need_peeling_or_partial_vectors_p) { /* For partial-vector-usage=1, try to push the handling of partial vectors to the epilogue, with the main loop continuing to operate on full vectors. If we are unrolling we also do not want to use partial vectors. This is to avoid the overhead of generating multiple masks and also to avoid having to execute entire iterations of FALSE masked instructions when dealing with one or less full iterations. ??? We could then end up failing to use partial vectors if we decide to peel iterations into a prologue, and if the main loop then ends up processing fewer than VF iterations. */ if ((param_vect_partial_vector_usage == 1 || loop_vinfo->suggested_unroll_factor > 1) && !LOOP_VINFO_EPILOGUE_P (loop_vinfo) && !vect_known_niters_smaller_than_vf (loop_vinfo)) LOOP_VINFO_EPIL_USING_PARTIAL_VECTORS_P (loop_vinfo) = true; else LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo) = true; } if (LOOP_VINFO_MUST_USE_PARTIAL_VECTORS_P (loop_vinfo) && !LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo)) return opt_result::failure_at (vect_location, "not vectorized: loop needs but cannot " "use partial vectors\n"); if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "operating on %s vectors%s.\n", LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo) ? "partial" : "full", LOOP_VINFO_EPILOGUE_P (loop_vinfo) ? " for epilogue loop" : ""); LOOP_VINFO_PEELING_FOR_NITER (loop_vinfo) = (!LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo) && need_peeling_or_partial_vectors_p); /* We set LOOP_VINFO_USING_SELECT_VL_P as true before loop vectorization analysis that we don't know whether the loop is vectorized by partial vectors (More details see tree-vect-loop-manip.cc). However, SELECT_VL vectorizaton style should only applied on partial vectorization since SELECT_VL is the GIMPLE IR that calculates the number of elements to be process for each iteration. After loop vectorization analysis, Clear LOOP_VINFO_USING_SELECT_VL_P if it is not partial vectorized loop. */ if (!LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo)) LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo) = false; return opt_result::success (); } /* Function vect_analyze_loop_2. Apply a set of analyses on LOOP specified by LOOP_VINFO, the different analyses will record information in some members of LOOP_VINFO. FATAL indicates if some analysis meets fatal error. If one non-NULL pointer SUGGESTED_UNROLL_FACTOR is provided, it's intent to be filled with one worked out suggested unroll factor, while one NULL pointer shows it's going to apply the suggested unroll factor. SINGLE_LANE_SLP_DONE_FOR_SUGGESTED_UF is to hold whether single-lane slp was forced when the suggested unroll factor was worked out. */ static opt_result vect_analyze_loop_2 (loop_vec_info loop_vinfo, bool &fatal, unsigned *suggested_unroll_factor, bool& single_lane_slp_done_for_suggested_uf) { opt_result ok = opt_result::success (); int res; unsigned int max_vf = MAX_VECTORIZATION_FACTOR; loop_vec_info orig_loop_vinfo = NULL; /* If we are dealing with an epilogue then orig_loop_vinfo points to the loop_vec_info of the first vectorized loop. */ if (LOOP_VINFO_EPILOGUE_P (loop_vinfo)) orig_loop_vinfo = LOOP_VINFO_ORIG_LOOP_INFO (loop_vinfo); else orig_loop_vinfo = loop_vinfo; gcc_assert (orig_loop_vinfo); /* The first group of checks is independent of the vector size. */ fatal = true; if (LOOP_VINFO_SIMD_IF_COND (loop_vinfo) && integer_zerop (LOOP_VINFO_SIMD_IF_COND (loop_vinfo))) return opt_result::failure_at (vect_location, "not vectorized: simd if(0)\n"); /* 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)); 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. */ ok = vect_analyze_data_refs (loop_vinfo, &fatal); if (!ok) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "bad data references.\n"); return ok; } /* Check if we are applying unroll factor now. */ bool applying_suggested_uf = loop_vinfo->suggested_unroll_factor > 1; gcc_assert (!applying_suggested_uf || !suggested_unroll_factor); /* When single-lane SLP was forced and we are applying suggested unroll factor, keep that decision here. */ bool force_single_lane = (applying_suggested_uf && single_lane_slp_done_for_suggested_uf); /* Classify all cross-iteration scalar data-flow cycles. Cross-iteration cycles caused by virtual phis are analyzed separately. */ vect_analyze_scalar_cycles (loop_vinfo, !force_single_lane); vect_pattern_recog (loop_vinfo); vect_fixup_scalar_cycles_with_patterns (loop_vinfo); /* Analyze the access patterns of the data-refs in the loop (consecutive, complex, etc.). FORNOW: Only handle consecutive access pattern. */ ok = vect_analyze_data_ref_accesses (loop_vinfo, NULL); if (!ok) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "bad data access.\n"); return ok; } /* Data-flow analysis to detect stmts that do not need to be vectorized. */ ok = vect_mark_stmts_to_be_vectorized (loop_vinfo, &fatal); if (!ok) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "unexpected pattern.\n"); return ok; } /* While the rest of the analysis below depends on it in some way. */ fatal = false; /* Analyze data dependences between the data-refs in the loop and adjust the maximum vectorization factor according to the dependences. FORNOW: fail at the first data dependence that we encounter. */ ok = vect_analyze_data_ref_dependences (loop_vinfo, &max_vf); if (!ok) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "bad data dependence.\n"); return ok; } LOOP_VINFO_MAX_VECT_FACTOR (loop_vinfo) = max_vf; /* Compute the scalar iteration cost. */ vect_compute_single_scalar_iteration_cost (loop_vinfo); bool saved_can_use_partial_vectors_p = LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo); /* This is the point where we can re-start analysis with single-lane SLP forced. */ start_over: /* Check the SLP opportunities in the loop, analyze and build SLP trees. */ ok = vect_analyze_slp (loop_vinfo, loop_vinfo->stmt_vec_infos.length (), force_single_lane); if (!ok) return ok; /* If there are any SLP instances mark them as pure_slp. */ if (!vect_make_slp_decision (loop_vinfo)) return opt_result::failure_at (vect_location, "no stmts to vectorize.\n"); if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "Loop contains only SLP stmts\n"); /* Determine the vectorization factor from the SLP decision. */ LOOP_VINFO_VECT_FACTOR (loop_vinfo) = LOOP_VINFO_SLP_UNROLLING_FACTOR (loop_vinfo); if (dump_enabled_p ()) { dump_printf_loc (MSG_NOTE, vect_location, "vectorization factor = "); dump_dec (MSG_NOTE, LOOP_VINFO_VECT_FACTOR (loop_vinfo)); dump_printf (MSG_NOTE, "\n"); } /* Optimize the SLP graph with the vectorization factor fixed. */ vect_optimize_slp (loop_vinfo); /* Gather the loads reachable from the SLP graph entries. */ vect_gather_slp_loads (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 ()); /* Apply the suggested unrolling factor, this was determined by the backend during finish_cost the first time we ran the analyzis for this vector mode. */ if (applying_suggested_uf) LOOP_VINFO_VECT_FACTOR (loop_vinfo) *= loop_vinfo->suggested_unroll_factor; /* Now the vectorization factor is final. */ 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 = "); dump_dec (MSG_NOTE, vectorization_factor); dump_printf (MSG_NOTE, ", niters = %wd\n", LOOP_VINFO_INT_NITERS (loop_vinfo)); } 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"); loop_vinfo->vector_costs = init_cost (loop_vinfo, false); /* Analyze the alignment of the data-refs in the loop. Fail if a data reference is found that cannot be vectorized. */ ok = vect_analyze_data_refs_alignment (loop_vinfo); if (!ok) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "bad data alignment.\n"); return ok; } /* Prune the list of ddrs to be tested at run-time by versioning for alias. It is important to call pruning after vect_analyze_data_ref_accesses, since we use grouping information gathered by interleaving analysis. */ ok = vect_prune_runtime_alias_test_list (loop_vinfo); if (!ok) 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) return ok; /* Analyze operations in the SLP instances. We can't simply remove unsupported SLP instances as this makes the above SLP kind detection invalid and might also affect the VF. */ if (! vect_slp_analyze_operations (loop_vinfo)) { ok = opt_result::failure_at (vect_location, "unsupported SLP instances\n"); goto again; } /* For now, we don't expect to mix both masking and length approaches for one loop, disable it if both are recorded. */ if (LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) && !LOOP_VINFO_MASKS (loop_vinfo).is_empty () && !LOOP_VINFO_LENS (loop_vinfo).is_empty ()) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "can't vectorize a loop with partial vectors" " because we don't expect to mix different" " approaches with partial vectors for the" " same loop.\n"); LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; } /* If we still have the option of using partial vectors, check whether we can generate the necessary loop controls. */ if (LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)) { if (!LOOP_VINFO_MASKS (loop_vinfo).is_empty ()) { if (!vect_verify_full_masking (loop_vinfo) && !vect_verify_full_masking_avx512 (loop_vinfo)) LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; } else /* !LOOP_VINFO_LENS (loop_vinfo).is_empty () */ if (!vect_verify_loop_lens (loop_vinfo)) LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; } /* If we're vectorizing a loop that uses length "controls" and can iterate more than once, we apply decrementing IV approach in loop control. */ if (LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) && LOOP_VINFO_PARTIAL_VECTORS_STYLE (loop_vinfo) == vect_partial_vectors_len && LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo) == 0 && !(LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) && known_le (LOOP_VINFO_INT_NITERS (loop_vinfo), LOOP_VINFO_VECT_FACTOR (loop_vinfo)))) LOOP_VINFO_USING_DECREMENTING_IV_P (loop_vinfo) = true; /* If a loop uses length controls and has a decrementing loop control IV, we will normally pass that IV through a MIN_EXPR to calcaluate the basis for the length controls. E.g. in a loop that processes one element per scalar iteration, the number of elements would be MIN_EXPR <N, VF>, where N is the number of scalar iterations left. This MIN_EXPR approach allows us to use pointer IVs with an invariant step, since only the final iteration of the vector loop can have inactive lanes. However, some targets have a dedicated instruction for calculating the preferred length, given the total number of elements that still need to be processed. This is encapsulated in the SELECT_VL internal function. If the target supports SELECT_VL, we can use it instead of MIN_EXPR to determine the basis for the length controls. However, unlike the MIN_EXPR calculation, the SELECT_VL calculation can decide to make lanes inactive in any iteration of the vector loop, not just the last iteration. This SELECT_VL approach therefore requires us to use pointer IVs with variable steps. Once we've decided how many elements should be processed by one iteration of the vector loop, we need to populate the rgroup controls. If a loop has multiple rgroups, we need to make sure that those rgroups "line up" (that is, they must be consistent about which elements are active and which aren't). This is done by vect_adjust_loop_lens_control. In principle, it would be possible to use vect_adjust_loop_lens_control on either the result of a MIN_EXPR or the result of a SELECT_VL. However: (1) In practice, it only makes sense to use SELECT_VL when a vector operation will be controlled directly by the result. It is not worth using SELECT_VL if it would only be the input to other calculations. (2) If we use SELECT_VL for an rgroup that has N controls, each associated pointer IV will need N updates by a variable amount (N-1 updates within the iteration and 1 update to move to the next iteration). Because of this, we prefer to use the MIN_EXPR approach whenever there is more than one length control. In addition, SELECT_VL always operates to a granularity of 1 unit. If we wanted to use it to control an SLP operation on N consecutive elements, we would need to make the SELECT_VL inputs measure scalar iterations (rather than elements) and then multiply the SELECT_VL result by N. But using SELECT_VL this way is inefficient because of (1) above. 2. We don't apply SELECT_VL on single-rgroup when both (1) and (2) are satisfied: (1). LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) is true. (2). LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant () is true. Since SELECT_VL (variable step) will make SCEV analysis failed and then we will fail to gain benefits of following unroll optimizations. We prefer using the MIN_EXPR approach in this situation. */ if (LOOP_VINFO_USING_DECREMENTING_IV_P (loop_vinfo)) { tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo); if (direct_internal_fn_supported_p (IFN_SELECT_VL, iv_type, OPTIMIZE_FOR_SPEED) && LOOP_VINFO_LENS (loop_vinfo).length () == 1 && LOOP_VINFO_LENS (loop_vinfo)[0].factor == 1 && (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) || !LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant ())) LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo) = true; /* If any of the SLP instances cover more than a single lane we cannot use .SELECT_VL at the moment, even if the number of lanes is uniform throughout the SLP graph. */ if (LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo)) for (slp_instance inst : LOOP_VINFO_SLP_INSTANCES (loop_vinfo)) if (SLP_TREE_LANES (SLP_INSTANCE_TREE (inst)) != 1 && !(SLP_INSTANCE_KIND (inst) == slp_inst_kind_store && SLP_INSTANCE_TREE (inst)->ldst_lanes)) { LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo) = false; break; } } /* Decide whether this loop_vinfo should use partial vectors or peeling, assuming that the loop will be used as a main loop. We will redo this analysis later if we instead decide to use the loop as an epilogue loop. */ ok = vect_determine_partial_vectors_and_peeling (loop_vinfo); if (!ok) return ok; /* If we're vectorizing an epilogue loop, the vectorized loop either needs to be able to handle fewer than VF scalars, or needs to have a lower VF than the main loop. */ if (LOOP_VINFO_EPILOGUE_P (loop_vinfo) && !LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo)) { poly_uint64 unscaled_vf = exact_div (LOOP_VINFO_VECT_FACTOR (orig_loop_vinfo), orig_loop_vinfo->suggested_unroll_factor); if (maybe_ge (LOOP_VINFO_VECT_FACTOR (loop_vinfo), unscaled_vf)) return opt_result::failure_at (vect_location, "Vectorization factor too high for" " epilogue loop.\n"); } /* If the epilogue needs peeling for gaps but the main loop doesn't give up on the epilogue. */ if (LOOP_VINFO_EPILOGUE_P (loop_vinfo) && LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) && (LOOP_VINFO_PEELING_FOR_GAPS (orig_loop_vinfo) != LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo))) return opt_result::failure_at (vect_location, "Epilogue loop requires peeling for gaps " "but main loop does not.\n"); /* If an epilogue loop is required make sure we can create one. */ if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) || LOOP_VINFO_PEELING_FOR_NITER (loop_vinfo) || LOOP_VINFO_EARLY_BREAKS (loop_vinfo)) { if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "epilog loop required\n"); if (!vect_can_advance_ivs_p (loop_vinfo) || !slpeel_can_duplicate_loop_p (loop, LOOP_VINFO_IV_EXIT (loop_vinfo), LOOP_VINFO_IV_EXIT (loop_vinfo))) { ok = opt_result::failure_at (vect_location, "not vectorized: can't create required " "epilog loop\n"); goto again; } } /* Check the costings of the loop make vectorizing worthwhile. */ res = vect_analyze_loop_costing (loop_vinfo, suggested_unroll_factor); if (res < 0) { ok = opt_result::failure_at (vect_location, "Loop costings may not be worthwhile.\n"); goto again; } if (!res) return opt_result::failure_at (vect_location, "Loop costings not worthwhile.\n"); /* During peeling, we need to check if number of loop iterations is 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 we are analyzing an epilogue we still want to check what its versioning threshold would be. If we decide to vectorize the epilogues we will want to use the lowest versioning threshold of all epilogues and main loop. This will enable us to enter a vectorized epilogue even when versioning the loop. We can't simply check whether the epilogue requires versioning though since we may have skipped some versioning checks when analyzing the epilogue. For instance, checks for alias versioning will be skipped when dealing with epilogues as we assume we already checked them for the main loop. So instead we always check the 'orig_loop_vinfo'. */ if (LOOP_REQUIRES_VERSIONING (orig_loop_vinfo)) { poly_uint64 niters_th = 0; unsigned int th = LOOP_VINFO_COST_MODEL_THRESHOLD (loop_vinfo); 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); } /* Niters for at least one iteration of vectorized loop. */ if (!LOOP_VINFO_USING_PARTIAL_VECTORS_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 += 1; /* Use the same condition as vect_transform_loop to decide when to use the cost to determine a versioning threshold. */ if (vect_apply_runtime_profitability_check_p (loop_vinfo) && ordered_p (th, niters_th)) niters_th = ordered_max (poly_uint64 (th), niters_th); LOOP_VINFO_VERSIONING_THRESHOLD (loop_vinfo) = niters_th; } gcc_assert (known_eq (vectorization_factor, LOOP_VINFO_VECT_FACTOR (loop_vinfo))); single_lane_slp_done_for_suggested_uf = force_single_lane; /* Ok to vectorize! */ LOOP_VINFO_VECTORIZABLE_P (loop_vinfo) = 1; return opt_result::success (); again: /* Ensure that "ok" is false (with an opt_problem if dumping is enabled). */ gcc_assert (!ok); /* Try again with single-lane SLP. */ if (force_single_lane) return ok; /* If we are applying suggested unroll factor, we don't need to re-try any more as we want to keep the SLP mode fixed. */ if (applying_suggested_uf) return ok; /* If there are reduction chains re-trying will fail anyway. */ if (! LOOP_VINFO_REDUCTION_CHAINS (loop_vinfo).is_empty ()) return ok; /* Likewise if the grouped loads or stores in the SLP cannot be handled via interleaving or lane instructions. */ slp_instance instance; slp_tree node; unsigned i, j; FOR_EACH_VEC_ELT (LOOP_VINFO_SLP_INSTANCES (loop_vinfo), i, instance) { if (SLP_TREE_DEF_TYPE (SLP_INSTANCE_TREE (instance)) != vect_internal_def) continue; stmt_vec_info vinfo; vinfo = SLP_TREE_SCALAR_STMTS (SLP_INSTANCE_TREE (instance))[0]; if (! STMT_VINFO_GROUPED_ACCESS (vinfo)) continue; vinfo = DR_GROUP_FIRST_ELEMENT (vinfo); unsigned int size = DR_GROUP_SIZE (vinfo); tree vectype = SLP_TREE_VECTYPE (SLP_INSTANCE_TREE (instance)); if (vect_store_lanes_supported (vectype, size, false) == IFN_LAST && ! 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 = SLP_TREE_REPRESENTATIVE (node); if (STMT_VINFO_GROUPED_ACCESS (vinfo)) { vinfo = DR_GROUP_FIRST_ELEMENT (vinfo); bool single_element_p = !DR_GROUP_NEXT_ELEMENT (vinfo); size = DR_GROUP_SIZE (vinfo); vectype = SLP_TREE_VECTYPE (node); if (vect_load_lanes_supported (vectype, size, false) == IFN_LAST && ! vect_grouped_load_supported (vectype, single_element_p, size)) return opt_result::failure_at (vinfo->stmt, "unsupported grouped load\n"); } } } /* Roll back state appropriately. Force single-lane SLP this time. */ force_single_lane = true; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "re-trying with single-lane SLP\n"); /* Reset the vectorization factor. */ LOOP_VINFO_VECT_FACTOR (loop_vinfo) = 0; /* Free the SLP instances. */ FOR_EACH_VEC_ELT (LOOP_VINFO_SLP_INSTANCES (loop_vinfo), j, instance) vect_free_slp_instance (instance); 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) { basic_block bb = LOOP_VINFO_BBS (loop_vinfo)[i]; for (gimple_stmt_iterator si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) { stmt_vec_info stmt_info = loop_vinfo->lookup_stmt (gsi_stmt (si)); STMT_SLP_TYPE (stmt_info) = not_vect; if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def || STMT_VINFO_DEF_TYPE (stmt_info) == vect_double_reduction_def) { /* vectorizable_reduction adjusts reduction stmt def-types, restore them to that of the PHI. */ STMT_VINFO_DEF_TYPE (STMT_VINFO_REDUC_DEF (stmt_info)) = STMT_VINFO_DEF_TYPE (stmt_info); STMT_VINFO_DEF_TYPE (vect_stmt_to_vectorize (STMT_VINFO_REDUC_DEF (stmt_info))) = STMT_VINFO_DEF_TYPE (stmt_info); } } for (gimple_stmt_iterator si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) { if (is_gimple_debug (gsi_stmt (si))) continue; stmt_vec_info stmt_info = loop_vinfo->lookup_stmt (gsi_stmt (si)); STMT_SLP_TYPE (stmt_info) = not_vect; if (STMT_VINFO_IN_PATTERN_P (stmt_info)) { stmt_vec_info pattern_stmt_info = STMT_VINFO_RELATED_STMT (stmt_info); if (STMT_VINFO_SLP_VECT_ONLY_PATTERN (pattern_stmt_info)) STMT_VINFO_IN_PATTERN_P (stmt_info) = false; gimple *pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info); STMT_SLP_TYPE (pattern_stmt_info) = not_vect; for (gimple_stmt_iterator pi = gsi_start (pattern_def_seq); !gsi_end_p (pi); gsi_next (&pi)) STMT_SLP_TYPE (loop_vinfo->lookup_stmt (gsi_stmt (pi))) = not_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. */ delete loop_vinfo->vector_costs; loop_vinfo->vector_costs = nullptr; /* Reset accumulated rgroup information. */ LOOP_VINFO_MASKS (loop_vinfo).mask_set.empty (); release_vec_loop_controls (&LOOP_VINFO_MASKS (loop_vinfo).rgc_vec); release_vec_loop_controls (&LOOP_VINFO_LENS (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_USE_PARTIAL_VECTORS_P (loop_vinfo) = saved_can_use_partial_vectors_p; LOOP_VINFO_MUST_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo) = false; if (loop_vinfo->scan_map) loop_vinfo->scan_map->empty (); goto start_over; } /* Return true if vectorizing a loop using NEW_LOOP_VINFO appears to be better than vectorizing it using OLD_LOOP_VINFO. Assume that OLD_LOOP_VINFO is better unless something specifically indicates otherwise. Note that this deliberately isn't a partial order. */ static bool vect_better_loop_vinfo_p (loop_vec_info new_loop_vinfo, loop_vec_info old_loop_vinfo) { struct loop *loop = LOOP_VINFO_LOOP (new_loop_vinfo); gcc_assert (LOOP_VINFO_LOOP (old_loop_vinfo) == loop); poly_int64 new_vf = LOOP_VINFO_VECT_FACTOR (new_loop_vinfo); poly_int64 old_vf = LOOP_VINFO_VECT_FACTOR (old_loop_vinfo); /* Always prefer a VF of loop->simdlen over any other VF. */ if (loop->simdlen) { bool new_simdlen_p = known_eq (new_vf, loop->simdlen); bool old_simdlen_p = known_eq (old_vf, loop->simdlen); if (new_simdlen_p != old_simdlen_p) return new_simdlen_p; } const auto *old_costs = old_loop_vinfo->vector_costs; const auto *new_costs = new_loop_vinfo->vector_costs; if (loop_vec_info main_loop = LOOP_VINFO_ORIG_LOOP_INFO (old_loop_vinfo)) return new_costs->better_epilogue_loop_than_p (old_costs, main_loop); return new_costs->better_main_loop_than_p (old_costs); } /* Decide whether to replace OLD_LOOP_VINFO with NEW_LOOP_VINFO. Return true if we should. */ static bool vect_joust_loop_vinfos (loop_vec_info new_loop_vinfo, loop_vec_info old_loop_vinfo) { if (!vect_better_loop_vinfo_p (new_loop_vinfo, old_loop_vinfo)) return false; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "***** Preferring vector mode %s to vector mode %s\n", GET_MODE_NAME (new_loop_vinfo->vector_mode), GET_MODE_NAME (old_loop_vinfo->vector_mode)); return true; } /* Analyze LOOP with VECTOR_MODES[MODE_I] and as epilogue if ORIG_LOOP_VINFO is not NULL. When MASKED_P is not -1 override the default LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P with it. Set AUTODETECTED_VECTOR_MODE if VOIDmode and advance MODE_I to the next mode useful to analyze. Return the loop_vinfo on success and wrapped null on failure. */ static opt_loop_vec_info vect_analyze_loop_1 (class loop *loop, vec_info_shared *shared, const vect_loop_form_info *loop_form_info, loop_vec_info orig_loop_vinfo, const vector_modes &vector_modes, unsigned &mode_i, int masked_p, machine_mode &autodetected_vector_mode, bool &fatal) { loop_vec_info loop_vinfo = vect_create_loop_vinfo (loop, shared, loop_form_info, orig_loop_vinfo); machine_mode vector_mode = vector_modes[mode_i]; loop_vinfo->vector_mode = vector_mode; if (masked_p != -1) loop_vinfo->can_use_partial_vectors_p = masked_p; unsigned int suggested_unroll_factor = 1; bool single_lane_slp_done_for_suggested_uf = false; /* Run the main analysis. */ opt_result res = vect_analyze_loop_2 (loop_vinfo, fatal, &suggested_unroll_factor, single_lane_slp_done_for_suggested_uf); if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "***** Analysis %s with vector mode %s\n", res ? "succeeded" : "failed", GET_MODE_NAME (loop_vinfo->vector_mode)); auto user_unroll = LOOP_VINFO_LOOP (loop_vinfo)->unroll; if (res && !LOOP_VINFO_EPILOGUE_P (loop_vinfo) /* Check to see if the user wants to unroll or if the target wants to. */ && (suggested_unroll_factor > 1 || user_unroll > 1)) { if (suggested_unroll_factor == 1) { int assumed_vf = vect_vf_for_cost (loop_vinfo); suggested_unroll_factor = user_unroll / assumed_vf; if (suggested_unroll_factor > 1) { if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "setting unroll factor to %d based on user requested " "unroll factor %d and suggested vectorization " "factor: %d\n", suggested_unroll_factor, user_unroll, assumed_vf); } } if (suggested_unroll_factor > 1) { if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "***** Re-trying analysis for unrolling" " with unroll factor %d and %s slp.\n", suggested_unroll_factor, single_lane_slp_done_for_suggested_uf ? "single-lane" : ""); loop_vec_info unroll_vinfo = vect_create_loop_vinfo (loop, shared, loop_form_info, NULL); unroll_vinfo->vector_mode = vector_mode; unroll_vinfo->suggested_unroll_factor = suggested_unroll_factor; opt_result new_res = vect_analyze_loop_2 (unroll_vinfo, fatal, NULL, single_lane_slp_done_for_suggested_uf); if (new_res) { delete loop_vinfo; loop_vinfo = unroll_vinfo; LOOP_VINFO_USER_UNROLL (loop_vinfo) = user_unroll > 1; } else delete unroll_vinfo; } } /* Remember the autodetected vector mode. */ if (vector_mode == VOIDmode) autodetected_vector_mode = loop_vinfo->vector_mode; /* Advance mode_i, first skipping modes that would result in the same analysis result. */ while (mode_i + 1 < vector_modes.length () && vect_chooses_same_modes_p (loop_vinfo, vector_modes[mode_i + 1])) { if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "***** The result for vector mode %s would" " be the same\n", GET_MODE_NAME (vector_modes[mode_i + 1])); mode_i += 1; } if (mode_i + 1 < vector_modes.length () && vect_chooses_same_modes_p (autodetected_vector_mode, vector_modes[mode_i + 1])) { if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "***** Skipping vector mode %s, which would" " repeat the analysis for %s\n", GET_MODE_NAME (vector_modes[mode_i + 1]), GET_MODE_NAME (autodetected_vector_mode)); mode_i += 1; } mode_i++; if (!res) { delete loop_vinfo; if (fatal) gcc_checking_assert (orig_loop_vinfo == NULL); return opt_loop_vec_info::propagate_failure (res); } return opt_loop_vec_info::success (loop_vinfo); } /* Function vect_analyze_loop. 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. */ opt_loop_vec_info vect_analyze_loop (class loop *loop, gimple *loop_vectorized_call, vec_info_shared *shared) { 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)))) 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"); /* Analyze the loop form. */ vect_loop_form_info loop_form_info; opt_result res = vect_analyze_loop_form (loop, loop_vectorized_call, &loop_form_info); if (!res) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "bad loop form.\n"); return opt_loop_vec_info::propagate_failure (res); } if (!integer_onep (loop_form_info.assumptions)) { /* We consider to vectorize this loop by versioning it under some assumptions. In order to do this, we need to clear existing information computed by scev and niter analyzer. */ scev_reset_htab (); free_numbers_of_iterations_estimates (loop); /* Also set flag for this loop so that following scev and niter analysis are done under the assumptions. */ loop_constraint_set (loop, LOOP_C_FINITE); } else /* Clear the existing niter information to make sure the nonwrapping flag will be calculated and set propriately. */ free_numbers_of_iterations_estimates (loop); auto_vector_modes vector_modes; /* Autodetect first vector size we try. */ vector_modes.safe_push (VOIDmode); unsigned int autovec_flags = targetm.vectorize.autovectorize_vector_modes (&vector_modes, loop->simdlen != 0); bool pick_lowest_cost_p = ((autovec_flags & VECT_COMPARE_COSTS) && !unlimited_cost_model (loop)); machine_mode autodetected_vector_mode = VOIDmode; opt_loop_vec_info first_loop_vinfo = opt_loop_vec_info::success (NULL); unsigned int mode_i = 0; unsigned HOST_WIDE_INT simdlen = loop->simdlen; /* Keep track of the VF for each mode. Initialize all to 0 which indicates a mode has not been analyzed. */ auto_vec<poly_uint64, 8> cached_vf_per_mode; for (unsigned i = 0; i < vector_modes.length (); ++i) cached_vf_per_mode.safe_push (0); /* First determine the main loop vectorization mode, either the first one that works, starting with auto-detecting the vector mode and then following the targets order of preference, or the one with the lowest cost if pick_lowest_cost_p. */ while (1) { bool fatal; unsigned int last_mode_i = mode_i; /* Set cached VF to -1 prior to analysis, which indicates a mode has failed. */ cached_vf_per_mode[last_mode_i] = -1; opt_loop_vec_info loop_vinfo = vect_analyze_loop_1 (loop, shared, &loop_form_info, NULL, vector_modes, mode_i, -1, autodetected_vector_mode, fatal); if (fatal) break; if (loop_vinfo) { /* Analyzis has been successful so update the VF value. The VF should always be a multiple of unroll_factor and we want to capture the original VF here. */ cached_vf_per_mode[last_mode_i] = exact_div (LOOP_VINFO_VECT_FACTOR (loop_vinfo), loop_vinfo->suggested_unroll_factor); /* Once we hit the desired simdlen for the first time, discard any previous attempts. */ if (simdlen && known_eq (LOOP_VINFO_VECT_FACTOR (loop_vinfo), simdlen)) { delete first_loop_vinfo; first_loop_vinfo = opt_loop_vec_info::success (NULL); simdlen = 0; } else if (pick_lowest_cost_p && first_loop_vinfo && vect_joust_loop_vinfos (loop_vinfo, first_loop_vinfo)) { /* Pick loop_vinfo over first_loop_vinfo. */ delete first_loop_vinfo; first_loop_vinfo = opt_loop_vec_info::success (NULL); } if (first_loop_vinfo == NULL) first_loop_vinfo = loop_vinfo; else { delete loop_vinfo; loop_vinfo = opt_loop_vec_info::success (NULL); } /* Commit to first_loop_vinfo if we have no reason to try alternatives. */ if (!simdlen && !pick_lowest_cost_p) break; } if (mode_i == vector_modes.length () || autodetected_vector_mode == VOIDmode) break; /* Try the next biggest vector size. */ if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "***** Re-trying analysis with vector mode %s\n", GET_MODE_NAME (vector_modes[mode_i])); } if (!first_loop_vinfo) return opt_loop_vec_info::propagate_failure (res); if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "***** Choosing vector mode %s\n", GET_MODE_NAME (first_loop_vinfo->vector_mode)); /* Only vectorize epilogues if PARAM_VECT_EPILOGUES_NOMASK is enabled, SIMDUID is not set, it is the innermost loop and we have either already found the loop's SIMDLEN or there was no SIMDLEN to begin with. TODO: Enable epilogue vectorization for loops with SIMDUID set. */ bool vect_epilogues = (!simdlen && loop->inner == NULL && param_vect_epilogues_nomask && LOOP_VINFO_PEELING_FOR_NITER (first_loop_vinfo) /* No code motion support for multiple epilogues so for now not supported when multiple exits. */ && !LOOP_VINFO_EARLY_BREAKS (first_loop_vinfo) && !loop->simduid && loop_cost_model (loop) > VECT_COST_MODEL_VERY_CHEAP); if (!vect_epilogues) return first_loop_vinfo; /* Now analyze first_loop_vinfo for epilogue vectorization. */ /* For epilogues start the analysis from the first mode. The motivation behind starting from the beginning comes from cases where the VECTOR_MODES array may contain length-agnostic and length-specific modes. Their ordering is not guaranteed, so we could end up picking a mode for the main loop that is after the epilogue's optimal mode. */ int masked_p = -1; if (!unlimited_cost_model (loop) && (first_loop_vinfo->vector_costs->suggested_epilogue_mode (masked_p) != VOIDmode)) { vector_modes[0] = first_loop_vinfo->vector_costs->suggested_epilogue_mode (masked_p); cached_vf_per_mode[0] = 0; } else vector_modes[0] = autodetected_vector_mode; mode_i = 0; bool supports_partial_vectors = (param_vect_partial_vector_usage != 0 || masked_p == 1); machine_mode mask_mode; if (supports_partial_vectors && !partial_vectors_supported_p () && !(VECTOR_MODE_P (first_loop_vinfo->vector_mode) && targetm.vectorize.get_mask_mode (first_loop_vinfo->vector_mode).exists (&mask_mode) && SCALAR_INT_MODE_P (mask_mode))) supports_partial_vectors = false; poly_uint64 first_vinfo_vf = LOOP_VINFO_VECT_FACTOR (first_loop_vinfo); loop_vec_info orig_loop_vinfo = first_loop_vinfo; do { /* Let the user override what the target suggests. */ if (OPTION_SET_P (param_vect_partial_vector_usage)) masked_p = -1; while (1) { /* If the target does not support partial vectors we can shorten the number of modes to analyze for the epilogue as we know we can't pick a mode that would lead to a VF at least as big as the FIRST_VINFO_VF. */ if (!supports_partial_vectors && maybe_ge (cached_vf_per_mode[mode_i], first_vinfo_vf)) { mode_i++; if (mode_i == vector_modes.length ()) break; continue; } /* We would need an exhaustive search to find all modes we skipped but that would lead to the same result as the analysis it was skipped for and where we'd could check cached_vf_per_mode against. Check for the autodetected mode, which is the common situation on x86 which does not perform cost comparison. */ if (!supports_partial_vectors && maybe_ge (cached_vf_per_mode[0], first_vinfo_vf) && vect_chooses_same_modes_p (autodetected_vector_mode, vector_modes[mode_i])) { mode_i++; if (mode_i == vector_modes.length ()) break; continue; } if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "***** Re-trying epilogue analysis with vector " "mode %s\n", GET_MODE_NAME (vector_modes[mode_i])); bool fatal; opt_loop_vec_info loop_vinfo = vect_analyze_loop_1 (loop, shared, &loop_form_info, orig_loop_vinfo, vector_modes, mode_i, masked_p, autodetected_vector_mode, fatal); if (fatal) break; if (loop_vinfo) { if (pick_lowest_cost_p && orig_loop_vinfo->epilogue_vinfo && vect_joust_loop_vinfos (loop_vinfo, orig_loop_vinfo->epilogue_vinfo)) { gcc_assert (vect_epilogues); delete orig_loop_vinfo->epilogue_vinfo; orig_loop_vinfo->epilogue_vinfo = nullptr; } if (!orig_loop_vinfo->epilogue_vinfo) orig_loop_vinfo->epilogue_vinfo = loop_vinfo; else { delete loop_vinfo; loop_vinfo = opt_loop_vec_info::success (NULL); } /* For now only allow one epilogue loop, but allow pick_lowest_cost_p to replace it, so commit to the first epilogue if we have no reason to try alternatives. */ if (!pick_lowest_cost_p) break; } /* Revert back to the default from the suggested prefered epilogue vectorization mode. */ masked_p = -1; if (mode_i == vector_modes.length ()) break; } orig_loop_vinfo = orig_loop_vinfo->epilogue_vinfo; if (!orig_loop_vinfo) break; /* When we selected a first vectorized epilogue, see if the target suggests to have another one. */ masked_p = -1; if (!unlimited_cost_model (loop) && !LOOP_VINFO_USING_PARTIAL_VECTORS_P (orig_loop_vinfo) && (orig_loop_vinfo->vector_costs->suggested_epilogue_mode (masked_p) != VOIDmode)) { vector_modes[0] = orig_loop_vinfo->vector_costs->suggested_epilogue_mode (masked_p); cached_vf_per_mode[0] = 0; mode_i = 0; } else break; } while (1); if (first_loop_vinfo->epilogue_vinfo) { poly_uint64 lowest_th = LOOP_VINFO_VERSIONING_THRESHOLD (first_loop_vinfo); loop_vec_info epilog_vinfo = first_loop_vinfo->epilogue_vinfo; do { poly_uint64 th = LOOP_VINFO_VERSIONING_THRESHOLD (epilog_vinfo); gcc_assert (!LOOP_REQUIRES_VERSIONING (epilog_vinfo) || maybe_ne (lowest_th, 0U)); /* Keep track of the known smallest versioning threshold. */ if (ordered_p (lowest_th, th)) lowest_th = ordered_min (lowest_th, th); epilog_vinfo = epilog_vinfo->epilogue_vinfo; } while (epilog_vinfo); LOOP_VINFO_VERSIONING_THRESHOLD (first_loop_vinfo) = lowest_th; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "***** Choosing epilogue vector mode %s\n", GET_MODE_NAME (first_loop_vinfo->epilogue_vinfo->vector_mode)); } return first_loop_vinfo; } /* 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 (code_helper code, internal_fn *reduc_fn) { /* We support MINUS_EXPR by negating the operand. This also preserves an initial -0.0 since -0.0 - 0.0 (neutral op for MINUS_EXPR) == -0.0 + (-0.0) = -0.0. */ if (code == PLUS_EXPR || code == MINUS_EXPR) { *reduc_fn = IFN_FOLD_LEFT_PLUS; return true; } return false; } /* Function reduction_fn_for_scalar_code Input: CODE - tree_code of a reduction operations. Output: 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 an internal function. Return FALSE if CODE currently cannot be vectorized as reduction. */ bool reduction_fn_for_scalar_code (code_helper code, internal_fn *reduc_fn) { if (code.is_tree_code ()) switch (tree_code (code)) { case MAX_EXPR: *reduc_fn = IFN_REDUC_MAX; return true; case MIN_EXPR: *reduc_fn = IFN_REDUC_MIN; return true; case 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: *reduc_fn = IFN_LAST; return true; default: return false; } else switch (combined_fn (code)) { CASE_CFN_FMAX: *reduc_fn = IFN_REDUC_FMAX; return true; CASE_CFN_FMIN: *reduc_fn = IFN_REDUC_FMIN; return true; default: return false; } } /* If there is a neutral value X such that a reduction would not be affected by the introduction of additional X elements, return that X, otherwise return null. CODE is the code of the reduction and SCALAR_TYPE is type of the scalar elements. If the reduction has just a single initial value then INITIAL_VALUE is that value, otherwise it is null. If AS_INITIAL is TRUE the value is supposed to be used as initial value. In that case no signed zero is returned. */ tree neutral_op_for_reduction (tree scalar_type, code_helper code, tree initial_value, bool as_initial) { if (code.is_tree_code ()) switch (tree_code (code)) { case DOT_PROD_EXPR: case SAD_EXPR: case MINUS_EXPR: case BIT_IOR_EXPR: case BIT_XOR_EXPR: return build_zero_cst (scalar_type); case WIDEN_SUM_EXPR: case PLUS_EXPR: if (!as_initial && HONOR_SIGNED_ZEROS (scalar_type)) return build_real (scalar_type, dconstm0); else 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: return initial_value; default: return NULL_TREE; } else switch (combined_fn (code)) { CASE_CFN_FMIN: CASE_CFN_FMAX: return initial_value; default: return NULL_TREE; } } /* Error reporting helper for vect_is_simple_reduction below. GIMPLE statement STMT is printed with a message MSG. */ static void report_vect_op (dump_flags_t msg_type, gimple *stmt, const char *msg) { dump_printf_loc (msg_type, vect_location, "%s%G", msg, stmt); } /* 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. */ bool needs_fold_left_reduction_p (tree type, code_helper code) { /* CHECKME: check for !flag_finite_math_only too? */ if (SCALAR_FLOAT_TYPE_P (type)) { if (code.is_tree_code ()) switch (tree_code (code)) { case MIN_EXPR: case MAX_EXPR: return false; default: return !flag_associative_math; } else switch (combined_fn (code)) { CASE_CFN_FMIN: CASE_CFN_FMAX: return false; default: return !flag_associative_math; } } if (INTEGRAL_TYPE_P (type)) return (!code.is_tree_code () || !operation_no_trapping_overflow (type, tree_code (code))); 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 has a handled computation expression. Store the main reduction operation in *CODE. */ static bool check_reduction_path (dump_user_location_t loc, loop_p loop, gphi *phi, tree loop_arg, code_helper *code, vec<std::pair<ssa_op_iter, use_operand_p> > &path, bool inner_loop_of_double_reduc) { 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; int sign = -1; *code = ERROR_MARK; for (unsigned i = 1; i < path.length (); ++i) { gimple *use_stmt = USE_STMT (path[i].second); gimple_match_op op; if (!gimple_extract_op (use_stmt, &op)) { fail = true; break; } unsigned int opi = op.num_ops; if (gassign *assign = dyn_cast<gassign *> (use_stmt)) { /* The following make sure we can compute the operand index easily plus it mostly disallows chaining via COND_EXPR condition operands. */ for (opi = 0; opi < op.num_ops; ++opi) if (gimple_assign_rhs1_ptr (assign) + opi == path[i].second->use) break; } else if (gcall *call = dyn_cast<gcall *> (use_stmt)) { for (opi = 0; opi < op.num_ops; ++opi) if (gimple_call_arg_ptr (call, opi) == path[i].second->use) break; } if (opi == op.num_ops) { fail = true; break; } op.code = canonicalize_code (op.code, op.type); if (op.code == MINUS_EXPR) { op.code = PLUS_EXPR; /* Track whether we negate the reduction value each iteration. */ if (op.ops[1] == op.ops[opi]) neg = ! neg; } else if (op.code == IFN_COND_SUB) { op.code = IFN_COND_ADD; /* Track whether we negate the reduction value each iteration. */ if (op.ops[2] == op.ops[opi]) neg = ! neg; } /* For an FMA the reduction code is the PLUS if the addition chain is the reduction. */ else if (op.code == IFN_FMA && opi == 2) op.code = PLUS_EXPR; if (CONVERT_EXPR_CODE_P (op.code) && tree_nop_conversion_p (op.type, TREE_TYPE (op.ops[0]))) ; else if (*code == ERROR_MARK) { *code = op.code; sign = TYPE_SIGN (op.type); } else if (op.code != *code) { fail = true; break; } else if ((op.code == MIN_EXPR || op.code == MAX_EXPR) && sign != TYPE_SIGN (op.type)) { fail = true; break; } /* Check there's only a single stmt the op is used on. For the not value-changing tail and the last stmt allow out-of-loop uses, but not when this is the inner loop of a double reduction. ??? We could relax this and handle arbitrary live stmts by forcing a scalar epilogue for example. */ imm_use_iterator imm_iter; use_operand_p use_p; gimple *op_use_stmt; unsigned cnt = 0; bool cond_fn_p = op.code.is_internal_fn () && (conditional_internal_fn_code (internal_fn (op.code)) != ERROR_MARK); FOR_EACH_IMM_USE_STMT (op_use_stmt, imm_iter, op.ops[opi]) { /* In case of a COND_OP (mask, op1, op2, op1) reduction we should have op1 twice (once as definition, once as else) in the same operation. Enforce this. */ if (cond_fn_p && op_use_stmt == use_stmt) { gcall *call = as_a<gcall *> (use_stmt); unsigned else_pos = internal_fn_else_index (internal_fn (op.code)); if (gimple_call_arg (call, else_pos) != op.ops[opi]) { fail = true; break; } for (unsigned int j = 0; j < gimple_call_num_args (call); ++j) { if (j == else_pos) continue; if (gimple_call_arg (call, j) == op.ops[opi]) cnt++; } } else if (!is_gimple_debug (op_use_stmt) && ((*code != ERROR_MARK || inner_loop_of_double_reduc) || flow_bb_inside_loop_p (loop, gimple_bb (op_use_stmt)))) FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) cnt++; } if (cnt != 1) { fail = true; break; } } return ! fail && ! neg && *code != ERROR_MARK; } 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; code_helper code_; return (check_reduction_path (loc, loop, phi, loop_arg, &code_, path, false) && code_ == code); } /* Function vect_is_simple_reduction (1) Detect a cross-iteration def-use cycle that represents a simple reduction computation. We look for the following pattern: loop_header: a1 = phi < a0, a2 > a3 = ... a2 = operation (a3, a1) or a3 = ... loop_header: a1 = phi < a0, a2 > a2 = operation (a3, a1) such that: 1. operation is commutative and associative and it is safe to change the order of the computation 2. no uses for a2 in the loop (a2 is used out of the loop) 3. no uses of a1 in the loop besides the reduction operation 4. no uses of a1 outside the loop. Conditions 1,4 are tested here. Conditions 2,3 are tested in vect_mark_stmts_to_be_vectorized. (2) Detect a cross-iteration def-use cycle in nested loops, i.e., nested cycles. (3) Detect cycles of phi nodes in outer-loop vectorization, i.e., double reductions: a1 = phi < a0, a2 > inner loop (def of a3) a2 = phi < a3 > (4) Detect condition expressions, ie: for (int i = 0; i < N; i++) if (a[i] < val) ret_val = a[i]; */ static stmt_vec_info vect_is_simple_reduction (loop_vec_info loop_info, stmt_vec_info phi_info, bool *double_reduc, bool *reduc_chain_p, bool slp) { gphi *phi = as_a <gphi *> (phi_info->stmt); gimple *phi_use_stmt = NULL; imm_use_iterator imm_iter; use_operand_p use_p; *double_reduc = false; *reduc_chain_p = false; STMT_VINFO_REDUC_TYPE (phi_info) = TREE_CODE_REDUCTION; tree phi_name = PHI_RESULT (phi); /* ??? If there are no uses of the PHI result the inner loop reduction won't be detected as possibly double-reduction by vectorizable_reduction because that tries to walk the PHI arg from the preheader edge which can be constant. See PR60382. */ if (has_zero_uses (phi_name)) return NULL; class loop *loop = (gimple_bb (phi))->loop_father; unsigned nphi_def_loop_uses = 0; FOR_EACH_IMM_USE_FAST (use_p, imm_iter, phi_name) { gimple *use_stmt = USE_STMT (use_p); if (is_gimple_debug (use_stmt)) continue; if (!flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "intermediate value used outside loop.\n"); return NULL; } /* In case of a COND_OP (mask, op1, op2, op1) reduction we might have op1 twice (once as definition, once as else) in the same operation. Only count it as one. */ if (use_stmt != phi_use_stmt) { nphi_def_loop_uses++; phi_use_stmt = use_stmt; } } tree latch_def = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop)); if (TREE_CODE (latch_def) != SSA_NAME) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "reduction: not ssa_name: %T\n", latch_def); return NULL; } stmt_vec_info def_stmt_info = loop_info->lookup_def (latch_def); if (!def_stmt_info || !flow_bb_inside_loop_p (loop, gimple_bb (def_stmt_info->stmt))) return NULL; bool nested_in_vect_loop = flow_loop_nested_p (LOOP_VINFO_LOOP (loop_info), loop); unsigned nlatch_def_loop_uses = 0; auto_vec<gphi *, 3> lcphis; bool inner_loop_of_double_reduc = false; FOR_EACH_IMM_USE_FAST (use_p, imm_iter, latch_def) { gimple *use_stmt = USE_STMT (use_p); if (is_gimple_debug (use_stmt)) continue; if (flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))) nlatch_def_loop_uses++; else { /* We can have more than one loop-closed PHI. */ lcphis.safe_push (as_a <gphi *> (use_stmt)); if (nested_in_vect_loop && (STMT_VINFO_DEF_TYPE (loop_info->lookup_stmt (use_stmt)) == vect_double_reduction_def)) inner_loop_of_double_reduc = true; } } /* If we are vectorizing an inner reduction we are executing that in the original order only in case we are not dealing with a double reduction. */ if (nested_in_vect_loop && !inner_loop_of_double_reduc) { if (dump_enabled_p ()) report_vect_op (MSG_NOTE, def_stmt_info->stmt, "detected nested cycle: "); return def_stmt_info; } /* When the inner loop of a double reduction ends up with more than one loop-closed PHI we have failed to classify alternate such PHIs as double reduction, leading to wrong code. See PR103237. */ if (inner_loop_of_double_reduc && lcphis.length () != 1) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "unhandle double reduction\n"); return NULL; } /* If this isn't a nested cycle or if the nested cycle reduction value is used ouside of the inner loop we cannot handle uses of the reduction value. */ if (nlatch_def_loop_uses > 1 || nphi_def_loop_uses > 1) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "reduction used in loop.\n"); return NULL; } /* If DEF_STMT is a phi node itself, we expect it to have a single argument defined in the inner loop. */ if (gphi *def_stmt = dyn_cast <gphi *> (def_stmt_info->stmt)) { tree op1 = PHI_ARG_DEF (def_stmt, 0); if (gimple_phi_num_args (def_stmt) != 1 || TREE_CODE (op1) != SSA_NAME) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "unsupported phi node definition.\n"); return NULL; } /* Verify there is an inner cycle composed of the PHI phi_use_stmt and the latch definition 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_gimple_call (def1)) && is_a <gphi *> (phi_use_stmt) && flow_bb_inside_loop_p (loop->inner, gimple_bb (phi_use_stmt)) && (op1 == PHI_ARG_DEF_FROM_EDGE (phi_use_stmt, loop_latch_edge (loop->inner))) && lcphis.length () == 1) { if (dump_enabled_p ()) report_vect_op (MSG_NOTE, def_stmt, "detected double reduction: "); *double_reduc = true; return def_stmt_info; } return NULL; } /* Look for the expression computing latch_def from then loop PHI result. */ auto_vec<std::pair<ssa_op_iter, use_operand_p> > path; code_helper code; if (check_reduction_path (vect_location, loop, phi, latch_def, &code, path, inner_loop_of_double_reduc)) { STMT_VINFO_REDUC_CODE (phi_info) = code; if (code == COND_EXPR && !nested_in_vect_loop) STMT_VINFO_REDUC_TYPE (phi_info) = COND_REDUCTION; /* Fill in STMT_VINFO_REDUC_IDX and gather stmts for an SLP reduction chain for which the additional restriction is that all operations in the chain are the same. */ auto_vec<stmt_vec_info, 8> reduc_chain; unsigned i; bool is_slp_reduc = !nested_in_vect_loop && code != COND_EXPR; for (i = path.length () - 1; i >= 1; --i) { gimple *stmt = USE_STMT (path[i].second); stmt_vec_info stmt_info = loop_info->lookup_stmt (stmt); gimple_match_op op; if (!gimple_extract_op (stmt, &op)) gcc_unreachable (); if (gassign *assign = dyn_cast<gassign *> (stmt)) STMT_VINFO_REDUC_IDX (stmt_info) = path[i].second->use - gimple_assign_rhs1_ptr (assign); else { gcall *call = as_a<gcall *> (stmt); STMT_VINFO_REDUC_IDX (stmt_info) = path[i].second->use - gimple_call_arg_ptr (call, 0); } bool leading_conversion = (CONVERT_EXPR_CODE_P (op.code) && (i == 1 || i == path.length () - 1)); if ((op.code != code && !leading_conversion) /* We can only handle the final value in epilogue generation for reduction chains. */ || (i != 1 && !has_single_use (gimple_get_lhs (stmt)))) is_slp_reduc = false; /* For reduction chains we support a trailing/leading conversions. We do not store those in the actual chain. */ if (leading_conversion) continue; reduc_chain.safe_push (stmt_info); } if (slp && is_slp_reduc && reduc_chain.length () > 1) { for (unsigned i = 0; i < reduc_chain.length () - 1; ++i) { REDUC_GROUP_FIRST_ELEMENT (reduc_chain[i]) = reduc_chain[0]; REDUC_GROUP_NEXT_ELEMENT (reduc_chain[i]) = reduc_chain[i+1]; } REDUC_GROUP_FIRST_ELEMENT (reduc_chain.last ()) = reduc_chain[0]; REDUC_GROUP_NEXT_ELEMENT (reduc_chain.last ()) = NULL; /* Save the chain for further analysis in SLP detection. */ LOOP_VINFO_REDUCTION_CHAINS (loop_info).safe_push (reduc_chain[0]); REDUC_GROUP_SIZE (reduc_chain[0]) = reduc_chain.length (); *reduc_chain_p = true; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "reduction: detected reduction chain\n"); } else if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "reduction: detected reduction\n"); return def_stmt_info; } if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "reduction: unknown pattern\n"); return NULL; } /* Estimate the number of peeled epilogue iterations for LOOP_VINFO. PEEL_ITERS_PROLOGUE is the number of peeled prologue iterations, or -1 if not known. */ static int vect_get_peel_iters_epilogue (loop_vec_info loop_vinfo, int peel_iters_prologue) { int assumed_vf = vect_vf_for_cost (loop_vinfo); if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) || peel_iters_prologue == -1) { if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "cost model: epilogue peel iters set to vf/2 " "because loop iterations are unknown .\n"); return assumed_vf / 2; } else { int niters = LOOP_VINFO_INT_NITERS (loop_vinfo); peel_iters_prologue = MIN (niters, peel_iters_prologue); int 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 = assumed_vf; return peel_iters_epilogue; } } /* Calculate cost of peeling the loop PEEL_ITERS_PROLOGUE times. */ int vect_get_known_peeling_cost (loop_vec_info loop_vinfo, int peel_iters_prologue, int *peel_iters_epilogue, stmt_vector_for_cost *scalar_cost_vec, stmt_vector_for_cost *prologue_cost_vec, stmt_vector_for_cost *epilogue_cost_vec) { int retval = 0; *peel_iters_epilogue = vect_get_peel_iters_epilogue (loop_vinfo, peel_iters_prologue); if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)) { /* If peeled iterations are known but number of scalar loop iterations are unknown, count a taken branch per peeled loop. */ if (peel_iters_prologue > 0) retval = record_stmt_cost (prologue_cost_vec, 1, cond_branch_taken, vect_prologue); if (*peel_iters_epilogue > 0) retval += record_stmt_cost (epilogue_cost_vec, 1, cond_branch_taken, vect_epilogue); } stmt_info_for_cost *si; int j; if (peel_iters_prologue) FOR_EACH_VEC_ELT (*scalar_cost_vec, j, si) 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) retval += record_stmt_cost (epilogue_cost_vec, si->count * *peel_iters_epilogue, si->kind, si->stmt_info, si->misalign, vect_epilogue); return retval; } /* Function vect_estimate_min_profitable_iters Return the number of iterations required for the vector version of the loop to be profitable relative to the cost of the scalar version of the loop. *RET_MIN_PROFITABLE_NITERS is a cost model profitability threshold of iterations for vectorization. -1 value means loop vectorization is not profitable. This returned value may be used for dynamic profitability check. *RET_MIN_PROFITABLE_ESTIMATE is a profitability threshold to be used for static check against estimated number of iterations. */ static void vect_estimate_min_profitable_iters (loop_vec_info loop_vinfo, int *ret_min_profitable_niters, int *ret_min_profitable_estimate, unsigned *suggested_unroll_factor) { int min_profitable_iters; int min_profitable_estimate; int peel_iters_prologue; int peel_iters_epilogue; unsigned vec_inside_cost = 0; int vec_outside_cost = 0; unsigned vec_prologue_cost = 0; unsigned vec_epilogue_cost = 0; int scalar_single_iter_cost = 0; int scalar_outside_cost = 0; int assumed_vf = vect_vf_for_cost (loop_vinfo); int npeel = LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo); vector_costs *target_cost_data = loop_vinfo->vector_costs; /* Cost model disabled. */ if (unlimited_cost_model (LOOP_VINFO_LOOP (loop_vinfo))) { if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "cost model disabled.\n"); *ret_min_profitable_niters = 0; *ret_min_profitable_estimate = 0; return; } /* Requires loop versioning tests to handle misalignment. */ if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo)) { /* FIXME: Make cost depend on complexity of individual check. */ unsigned len = LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo).length (); (void) add_stmt_cost (target_cost_data, len, scalar_stmt, vect_prologue); if (dump_enabled_p ()) dump_printf (MSG_NOTE, "cost model: Adding cost of checks for loop " "versioning to treat misalignment.\n"); } /* Requires loop versioning with alias checks. */ if (LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo)) { /* FIXME: Make cost depend on complexity of individual check. */ unsigned len = LOOP_VINFO_COMP_ALIAS_DDRS (loop_vinfo).length (); (void) add_stmt_cost (target_cost_data, len, scalar_stmt, vect_prologue); len = LOOP_VINFO_CHECK_UNEQUAL_ADDRS (loop_vinfo).length (); if (len) /* Count LEN - 1 ANDs and LEN comparisons. */ (void) add_stmt_cost (target_cost_data, len * 2 - 1, scalar_stmt, 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, vect_prologue); } if (dump_enabled_p ()) dump_printf (MSG_NOTE, "cost model: Adding cost of checks for loop " "versioning aliasing.\n"); } /* Requires loop versioning with niter checks. */ if (LOOP_REQUIRES_VERSIONING_FOR_NITERS (loop_vinfo)) { /* FIXME: Make cost depend on complexity of individual check. */ (void) add_stmt_cost (target_cost_data, 1, vector_stmt, NULL, NULL, NULL_TREE, 0, vect_prologue); if (dump_enabled_p ()) dump_printf (MSG_NOTE, "cost model: Adding cost of checks for loop " "versioning niters.\n"); } if (LOOP_REQUIRES_VERSIONING (loop_vinfo)) (void) add_stmt_cost (target_cost_data, 1, cond_branch_taken, vect_prologue); /* 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. */ scalar_single_iter_cost = (loop_vinfo->scalar_costs->total_cost () * param_vect_scalar_cost_multiplier) / 100; /* Add additional cost for the peeled instructions in prologue and epilogue 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). TODO: Build an expression that represents peel_iters for prologue and epilogue to be used in a run-time test. */ bool prologue_need_br_taken_cost = false; bool prologue_need_br_not_taken_cost = false; /* Calculate peel_iters_prologue. */ if (vect_use_loop_mask_for_alignment_p (loop_vinfo)) peel_iters_prologue = 0; else if (npeel < 0) { peel_iters_prologue = assumed_vf / 2; if (dump_enabled_p ()) dump_printf (MSG_NOTE, "cost model: " "prologue peel iters set to vf/2.\n"); /* If peeled iterations are unknown, count a taken branch and a not taken branch per peeled loop. Even if scalar loop iterations are known, vector iterations are not known since peeled prologue iterations are not known. Hence guards remain the same. */ prologue_need_br_taken_cost = true; prologue_need_br_not_taken_cost = true; } else { peel_iters_prologue = npeel; if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) && peel_iters_prologue > 0) /* If peeled iterations are known but number of scalar loop iterations are unknown, count a taken branch per peeled loop. */ prologue_need_br_taken_cost = true; } bool epilogue_need_br_taken_cost = false; bool epilogue_need_br_not_taken_cost = false; /* Calculate peel_iters_epilogue. */ if (LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo)) /* We need to peel exactly one iteration for gaps. */ peel_iters_epilogue = LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) ? 1 : 0; else if (npeel < 0) { /* If peeling for alignment is unknown, loop bound of main loop becomes unknown. */ peel_iters_epilogue = assumed_vf / 2; if (dump_enabled_p ()) dump_printf (MSG_NOTE, "cost model: " "epilogue peel iters set to vf/2 because " "peeling for alignment is unknown.\n"); /* See the same reason above in peel_iters_prologue calculation. */ epilogue_need_br_taken_cost = true; epilogue_need_br_not_taken_cost = true; } else { peel_iters_epilogue = vect_get_peel_iters_epilogue (loop_vinfo, npeel); if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) && peel_iters_epilogue > 0) /* If peeled iterations are known but number of scalar loop iterations are unknown, count a taken branch per peeled loop. */ epilogue_need_br_taken_cost = true; } stmt_info_for_cost *si; int j; /* Add costs associated with peel_iters_prologue. */ if (peel_iters_prologue) FOR_EACH_VEC_ELT (LOOP_VINFO_SCALAR_ITERATION_COST (loop_vinfo), j, si) { (void) add_stmt_cost (target_cost_data, si->count * peel_iters_prologue, si->kind, si->stmt_info, si->node, si->vectype, si->misalign, vect_prologue); } /* Add costs associated with peel_iters_epilogue. */ if (peel_iters_epilogue) FOR_EACH_VEC_ELT (LOOP_VINFO_SCALAR_ITERATION_COST (loop_vinfo), j, si) { (void) add_stmt_cost (target_cost_data, si->count * peel_iters_epilogue, si->kind, si->stmt_info, si->node, si->vectype, si->misalign, vect_epilogue); } /* Add possible cond_branch_taken/cond_branch_not_taken cost. */ if (prologue_need_br_taken_cost) (void) add_stmt_cost (target_cost_data, 1, cond_branch_taken, vect_prologue); if (prologue_need_br_not_taken_cost) (void) add_stmt_cost (target_cost_data, 1, cond_branch_not_taken, vect_prologue); if (epilogue_need_br_taken_cost) (void) add_stmt_cost (target_cost_data, 1, cond_branch_taken, vect_epilogue); if (epilogue_need_br_not_taken_cost) (void) add_stmt_cost (target_cost_data, 1, cond_branch_not_taken, vect_epilogue); /* Take care of special costs for rgroup controls of partial vectors. */ if (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) && (LOOP_VINFO_PARTIAL_VECTORS_STYLE (loop_vinfo) == vect_partial_vectors_avx512)) { /* Calculate how many masks we need to generate. */ unsigned int num_masks = 0; bool need_saturation = false; for (auto rgm : LOOP_VINFO_MASKS (loop_vinfo).rgc_vec) if (rgm.type) { unsigned nvectors = rgm.factor; num_masks += nvectors; if (TYPE_PRECISION (TREE_TYPE (rgm.compare_type)) < TYPE_PRECISION (LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo))) need_saturation = true; } /* ??? The target isn't able to identify the costs below as producing masks so it cannot penaltize cases where we'd run out of mask registers for example. */ /* ??? We are also failing to account for smaller vector masks we generate by splitting larger masks in vect_get_loop_mask. */ /* In the worst case, we need to generate each mask in the prologue and in the loop body. We need one splat per group and one compare per mask. Sometimes the prologue mask will fold to a constant, so the actual prologue cost might be smaller. However, it's simpler and safer to use the worst-case cost; if this ends up being the tie-breaker between vectorizing or not, then it's probably better not to vectorize. */ (void) add_stmt_cost (target_cost_data, num_masks + LOOP_VINFO_MASKS (loop_vinfo).rgc_vec.length (), vector_stmt, NULL, NULL, NULL_TREE, 0, vect_prologue); (void) add_stmt_cost (target_cost_data, num_masks + LOOP_VINFO_MASKS (loop_vinfo).rgc_vec.length (), vector_stmt, NULL, NULL, NULL_TREE, 0, vect_body); /* When we need saturation we need it both in the prologue and the epilogue. */ if (need_saturation) { (void) add_stmt_cost (target_cost_data, 1, scalar_stmt, NULL, NULL, NULL_TREE, 0, vect_prologue); (void) add_stmt_cost (target_cost_data, 1, scalar_stmt, NULL, NULL, NULL_TREE, 0, vect_body); } } else if (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) && (LOOP_VINFO_PARTIAL_VECTORS_STYLE (loop_vinfo) == vect_partial_vectors_while_ult)) { /* Calculate how many masks we need to generate. */ unsigned int num_masks = 0; rgroup_controls *rgm; unsigned int num_vectors_m1; FOR_EACH_VEC_ELT (LOOP_VINFO_MASKS (loop_vinfo).rgc_vec, num_vectors_m1, rgm) if (rgm->type) num_masks += num_vectors_m1 + 1; gcc_assert (num_masks > 0); /* In the worst case, we need to generate each mask in the prologue and in the loop body. One of the loop body mask instructions replaces the comparison in the scalar loop, and since we don't count the scalar comparison against the scalar body, we shouldn't count that vector instruction against the vector body either. Sometimes we can use unpacks instead of generating prologue masks and sometimes the prologue mask will fold to a constant, so the actual prologue cost might be smaller. However, it's simpler and safer to use the worst-case cost; if this ends up being the tie-breaker between vectorizing or not, then it's probably better not to vectorize. */ (void) add_stmt_cost (target_cost_data, num_masks, vector_stmt, NULL, NULL, NULL_TREE, 0, vect_prologue); (void) add_stmt_cost (target_cost_data, num_masks - 1, vector_stmt, NULL, NULL, NULL_TREE, 0, vect_body); } else if (LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo)) { /* Referring to the functions vect_set_loop_condition_partial_vectors and vect_set_loop_controls_directly, we need to generate each length in the prologue and in the loop body if required. Although there are some possible optimizations, we consider the worst case here. */ bool niters_known_p = LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo); signed char partial_load_store_bias = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); bool need_iterate_p = (!LOOP_VINFO_EPILOGUE_P (loop_vinfo) && !vect_known_niters_smaller_than_vf (loop_vinfo)); /* Calculate how many statements to be added. */ unsigned int prologue_stmts = 0; unsigned int body_stmts = 0; rgroup_controls *rgc; unsigned int num_vectors_m1; FOR_EACH_VEC_ELT (LOOP_VINFO_LENS (loop_vinfo), num_vectors_m1, rgc) if (rgc->type) { /* May need one SHIFT for nitems_total computation. */ unsigned nitems = rgc->max_nscalars_per_iter * rgc->factor; if (nitems != 1 && !niters_known_p) prologue_stmts += 1; /* May need one MAX and one MINUS for wrap around. */ if (vect_rgroup_iv_might_wrap_p (loop_vinfo, rgc)) prologue_stmts += 2; /* Need one MAX and one MINUS for each batch limit excepting for the 1st one. */ prologue_stmts += num_vectors_m1 * 2; unsigned int num_vectors = num_vectors_m1 + 1; /* Need to set up lengths in prologue, only one MIN required for each since start index is zero. */ prologue_stmts += num_vectors; /* If we have a non-zero partial load bias, we need one PLUS to adjust the load length. */ if (partial_load_store_bias != 0) body_stmts += 1; unsigned int length_update_cost = 0; if (LOOP_VINFO_USING_DECREMENTING_IV_P (loop_vinfo)) /* For decrement IV style, Each only need a single SELECT_VL or MIN since beginning to calculate the number of elements need to be processed in current iteration. */ length_update_cost = 1; else /* For increment IV stype, Each may need two MINs and one MINUS to update lengths in body for next iteration. */ length_update_cost = 3; if (need_iterate_p) body_stmts += length_update_cost * num_vectors; } (void) add_stmt_cost (target_cost_data, prologue_stmts, scalar_stmt, vect_prologue); (void) add_stmt_cost (target_cost_data, body_stmts, scalar_stmt, vect_body); } /* FORNOW: The scalar outside cost is incremented in one of the following ways: 1. The vectorizer checks for alignment and aliasing and generates a condition that allows dynamic vectorization. A cost model check is ANDED with the versioning condition. Hence scalar code path now has the added cost of the versioning check. if (cost > th & versioning_check) jmp to vector code Hence run-time scalar is incremented by not-taken branch cost. 2. The vectorizer then checks if a prologue is required. If the cost model check was not done before during versioning, it has to be done before the prologue check. if (cost <= th) prologue = scalar_iters if (prologue == 0) jmp to vector code else execute prologue if (prologue == num_iters) go to exit Hence the run-time scalar cost is incremented by a taken branch, plus a not-taken branch, plus a taken branch cost. 3. The vectorizer then checks if an epilogue is required. If the cost model check was not done before during prologue check, it has to be done with the epilogue check. if (prologue == 0) jmp to vector code else execute prologue if (prologue == num_iters) go to exit vector code: if ((cost <= th) | (scalar_iters-prologue-epilogue == 0)) jmp to epilogue Hence the run-time scalar cost should be incremented by 2 taken branches. TODO: The back end may reorder the BBS's differently and reverse conditions/branch directions. Change the estimates below to something more reasonable. */ /* If the number of iterations is known and we do not do versioning, we can decide whether to vectorize at compile time. Hence the scalar version do not carry cost model guard costs. */ if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) || LOOP_REQUIRES_VERSIONING (loop_vinfo)) { /* Cost model check occurs at versioning. */ if (LOOP_REQUIRES_VERSIONING (loop_vinfo)) scalar_outside_cost += vect_get_stmt_cost (cond_branch_not_taken); else { /* Cost model check occurs at prologue generation. */ if (LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) < 0) scalar_outside_cost += 2 * vect_get_stmt_cost (cond_branch_taken) + vect_get_stmt_cost (cond_branch_not_taken); /* Cost model check occurs at epilogue generation. */ else scalar_outside_cost += 2 * vect_get_stmt_cost (cond_branch_taken); } } /* Complete the target-specific cost calculations. */ loop_vinfo->vector_costs->finish_cost (loop_vinfo->scalar_costs); vec_prologue_cost = loop_vinfo->vector_costs->prologue_cost (); vec_inside_cost = loop_vinfo->vector_costs->body_cost (); vec_epilogue_cost = loop_vinfo->vector_costs->epilogue_cost (); if (suggested_unroll_factor) *suggested_unroll_factor = loop_vinfo->vector_costs->suggested_unroll_factor (); if (suggested_unroll_factor && *suggested_unroll_factor > 1 && LOOP_VINFO_MAX_VECT_FACTOR (loop_vinfo) != MAX_VECTORIZATION_FACTOR && !known_le (LOOP_VINFO_VECT_FACTOR (loop_vinfo) * *suggested_unroll_factor, LOOP_VINFO_MAX_VECT_FACTOR (loop_vinfo))) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "can't unroll as unrolled vectorization factor larger" " than maximum vectorization factor: " HOST_WIDE_INT_PRINT_UNSIGNED "\n", LOOP_VINFO_MAX_VECT_FACTOR (loop_vinfo)); *suggested_unroll_factor = 1; } vec_outside_cost = (int)(vec_prologue_cost + vec_epilogue_cost); if (dump_enabled_p ()) { dump_printf_loc (MSG_NOTE, vect_location, "Cost model analysis: \n"); dump_printf (MSG_NOTE, " Vector inside of loop cost: %d\n", vec_inside_cost); dump_printf (MSG_NOTE, " Vector prologue cost: %d\n", vec_prologue_cost); dump_printf (MSG_NOTE, " Vector epilogue cost: %d\n", vec_epilogue_cost); dump_printf (MSG_NOTE, " Scalar iteration cost: %d\n", scalar_single_iter_cost); dump_printf (MSG_NOTE, " Scalar outside cost: %d\n", scalar_outside_cost); dump_printf (MSG_NOTE, " Vector outside cost: %d\n", vec_outside_cost); dump_printf (MSG_NOTE, " prologue iterations: %d\n", peel_iters_prologue); dump_printf (MSG_NOTE, " epilogue iterations: %d\n", peel_iters_epilogue); } /* Calculate number of iterations required to make the vector version profitable, relative to the loop bodies only. The following condition must hold true: SIC * niters + SOC > VIC * ((niters - NPEEL) / VF) + VOC where SIC = scalar iteration cost, VIC = vector iteration cost, VOC = vector outside cost, VF = vectorization factor, NPEEL = prologue iterations + epilogue iterations, SOC = scalar outside cost for run time cost model check. */ int saving_per_viter = (scalar_single_iter_cost * assumed_vf - vec_inside_cost); if (saving_per_viter <= 0) { if (LOOP_VINFO_LOOP (loop_vinfo)->force_vectorize) 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, "cost model: the vector iteration cost = %d " "divided by the scalar iteration cost = %d " "is greater or equal to the vectorization factor = %d" ".\n", vec_inside_cost, scalar_single_iter_cost, assumed_vf); *ret_min_profitable_niters = -1; *ret_min_profitable_estimate = -1; return; } /* ??? The "if" arm is written to handle all cases; see below for what we would do for !LOOP_VINFO_USING_PARTIAL_VECTORS_P. */ if (LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo)) { /* Rewriting the condition above in terms of the number of vector iterations (vniters) rather than the number of scalar iterations (niters) gives: SIC * (vniters * VF + NPEEL) + SOC > VIC * vniters + VOC <==> vniters * (SIC * VF - VIC) > VOC - SIC * NPEEL - SOC For integer N, X and Y when X > 0: N * X > Y <==> N >= (Y /[floor] X) + 1. */ int outside_overhead = (vec_outside_cost - scalar_single_iter_cost * peel_iters_prologue - scalar_single_iter_cost * peel_iters_epilogue - scalar_outside_cost); /* We're only interested in cases that require at least one vector iteration. */ int min_vec_niters = 1; if (outside_overhead > 0) min_vec_niters = outside_overhead / saving_per_viter + 1; if (dump_enabled_p ()) dump_printf (MSG_NOTE, " Minimum number of vector iterations: %d\n", min_vec_niters); if (LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo)) { /* Now that we know the minimum number of vector iterations, find the minimum niters for which the scalar cost is larger: SIC * niters > VIC * vniters + VOC - SOC We know that the minimum niters is no more than vniters * VF + NPEEL, but it might be (and often is) less than that if a partial vector iteration is cheaper than the equivalent scalar code. */ int threshold = (vec_inside_cost * min_vec_niters + vec_outside_cost - scalar_outside_cost); if (threshold <= 0) min_profitable_iters = 1; else min_profitable_iters = threshold / scalar_single_iter_cost + 1; } else /* Convert the number of vector iterations into a number of scalar iterations. */ min_profitable_iters = (min_vec_niters * assumed_vf + peel_iters_prologue + peel_iters_epilogue); } else { 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 /= saving_per_viter; 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++; } } if (dump_enabled_p ()) dump_printf (MSG_NOTE, " Calculated minimum iters for profitability: %d\n", min_profitable_iters); if (!LOOP_VINFO_USING_PARTIAL_VECTORS_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; else if (min_profitable_iters < peel_iters_prologue) /* For LOOP_VINFO_USING_PARTIAL_VECTORS_P, we need to ensure the vectorized loop executes at least once. */ min_profitable_iters = peel_iters_prologue; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, " Runtime profitability threshold = %d\n", min_profitable_iters); *ret_min_profitable_niters = min_profitable_iters; /* Calculate number of iterations required to make the vector version profitable, relative to the loop bodies only. Non-vectorized variant is SIC * niters and it must win over vector variant on the expected loop trip count. The following condition must hold true: SIC * niters > VIC * ((niters - NPEEL) / VF) + VOC + SOC */ if (vec_outside_cost <= 0) min_profitable_estimate = 0; /* ??? This "else if" arm is written to handle all cases; see below for what we would do for !LOOP_VINFO_USING_PARTIAL_VECTORS_P. */ else if (LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo)) { /* This is a repeat of the code above, but with + SOC rather than - SOC. */ int outside_overhead = (vec_outside_cost - scalar_single_iter_cost * peel_iters_prologue - scalar_single_iter_cost * peel_iters_epilogue + scalar_outside_cost); int min_vec_niters = 1; if (outside_overhead > 0) min_vec_niters = outside_overhead / saving_per_viter + 1; if (LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo)) { int threshold = (vec_inside_cost * min_vec_niters + vec_outside_cost + scalar_outside_cost); min_profitable_estimate = threshold / scalar_single_iter_cost + 1; } else min_profitable_estimate = (min_vec_niters * assumed_vf + peel_iters_prologue + peel_iters_epilogue); } else { 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 * assumed_vf) - vec_inside_cost); } min_profitable_estimate = MAX (min_profitable_estimate, min_profitable_iters); if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, " Static estimate profitability threshold = %d\n", min_profitable_estimate); *ret_min_profitable_estimate = min_profitable_estimate; } /* Writes into SEL a mask for a vec_perm, equivalent to a vec_shr by OFFSET 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_builder *sel) { /* 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 MODE. This is the case if _either_ the platform handles vec_shr_optab, _or_ it supports vec_perm_const with masks for all necessary shift amounts. */ static bool have_whole_vector_shift (machine_mode mode) { if (can_implement_p (vec_shr_optab, mode)) return true; /* Variable-length vectors should be handled via the optab. */ unsigned int nelt; if (!GET_MODE_NUNITS (mode).is_constant (&nelt)) return false; 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); indices.new_vector (sel, 2, nelt); if (!can_vec_perm_const_p (mode, mode, indices, false)) return false; } return true; } /* Return true if (a) STMT_INFO is a DOT_PROD_EXPR reduction whose multiplication operands have differing signs and (b) we intend to emulate the operation using a series of signed DOT_PROD_EXPRs. See vect_emulate_mixed_dot_prod for the actual sequence used. */ static bool vect_is_emulated_mixed_dot_prod (slp_tree slp_node) { stmt_vec_info stmt_info = SLP_TREE_REPRESENTATIVE (slp_node); gassign *assign = dyn_cast<gassign *> (stmt_info->stmt); if (!assign || gimple_assign_rhs_code (assign) != DOT_PROD_EXPR) return false; tree rhs1 = gimple_assign_rhs1 (assign); tree rhs2 = gimple_assign_rhs2 (assign); if (TYPE_SIGN (TREE_TYPE (rhs1)) == TYPE_SIGN (TREE_TYPE (rhs2))) return false; return !directly_supported_p (DOT_PROD_EXPR, SLP_TREE_VECTYPE (slp_node), SLP_TREE_VECTYPE (SLP_TREE_CHILDREN (slp_node)[0]), optab_vector_mixed_sign); } /* TODO: Close dependency between vect_model_*_cost and vectorizable_* functions. Design better to avoid maintenance issues. */ /* Function vect_model_reduction_cost. Models cost for a reduction operation, including the vector ops generated within the strip-mine loop in some cases, the initial definition before the loop, and the epilogue code that must be generated. */ static void vect_model_reduction_cost (loop_vec_info loop_vinfo, slp_tree node, internal_fn reduc_fn, vect_reduction_type reduction_type, int ncopies, stmt_vector_for_cost *cost_vec) { int prologue_cost = 0, epilogue_cost = 0, inside_cost = 0; tree vectype; machine_mode mode; class loop *loop = NULL; if (loop_vinfo) loop = LOOP_VINFO_LOOP (loop_vinfo); /* Condition reductions generate two reductions in the loop. */ if (reduction_type == COND_REDUCTION) ncopies *= 2; vectype = SLP_TREE_VECTYPE (node); mode = TYPE_MODE (vectype); stmt_vec_info orig_stmt_info = vect_orig_stmt (SLP_TREE_REPRESENTATIVE (node)); gimple_match_op op; if (!gimple_extract_op (orig_stmt_info->stmt, &op)) gcc_unreachable (); if (reduction_type == EXTRACT_LAST_REDUCTION) /* No extra instructions are needed in the prologue. The loop body operations are costed in vectorizable_condition. */ inside_cost = 0; else if (reduction_type == FOLD_LEFT_REDUCTION) { /* No extra instructions needed in the prologue. */ prologue_cost = 0; if (reduc_fn != IFN_LAST) /* Count one reduction-like operation per vector. */ inside_cost = record_stmt_cost (cost_vec, ncopies, vec_to_scalar, node, 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, node, 0, vect_body); inside_cost += record_stmt_cost (cost_vec, nelements, scalar_stmt, node, 0, vect_body); } } else { /* Add in the cost of the initial definitions. */ int prologue_stmts; if (reduction_type == COND_REDUCTION) /* For cond reductions we have four vectors: initial index, step, initial result of the data reduction, initial value of the index reduction. */ prologue_stmts = 4; else /* We need the initial reduction value. */ prologue_stmts = 1; prologue_cost += record_stmt_cost (cost_vec, prologue_stmts, scalar_to_vec, node, 0, vect_prologue); } /* 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_info)) { if (reduc_fn != IFN_LAST) { if (reduction_type == COND_REDUCTION) { /* An EQ stmt and an COND_EXPR stmt. */ epilogue_cost += record_stmt_cost (cost_vec, 2, vector_stmt, node, 0, vect_epilogue); /* Reduction of the max index and a reduction of the found values. */ epilogue_cost += record_stmt_cost (cost_vec, 2, vec_to_scalar, node, 0, vect_epilogue); /* A broadcast of the max value. */ epilogue_cost += record_stmt_cost (cost_vec, 1, scalar_to_vec, node, 0, vect_epilogue); } else { epilogue_cost += record_stmt_cost (cost_vec, 1, vector_stmt, node, 0, vect_epilogue); epilogue_cost += record_stmt_cost (cost_vec, 1, vec_to_scalar, node, 0, vect_epilogue); } } else if (reduction_type == COND_REDUCTION) { unsigned estimated_nunits = vect_nunits_for_cost (vectype); /* Extraction of scalar elements. */ epilogue_cost += record_stmt_cost (cost_vec, 2 * estimated_nunits, vec_to_scalar, node, 0, vect_epilogue); /* Scalar max reductions via COND_EXPR / MAX_EXPR. */ epilogue_cost += record_stmt_cost (cost_vec, 2 * estimated_nunits - 3, scalar_stmt, node, 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 (op.type); int element_bitsize = tree_to_uhwi (bitsize); int nelements = vec_size_in_bits / element_bitsize; if (op.code == COND_EXPR) op.code = MAX_EXPR; /* We have a whole vector shift available. */ if (VECTOR_MODE_P (mode) && directly_supported_p (op.code, vectype) && have_whole_vector_shift (mode)) { /* Final reduction via vector shifts and the reduction operator. Also requires scalar extract. */ epilogue_cost += record_stmt_cost (cost_vec, exact_log2 (nelements) * 2, vector_stmt, node, 0, vect_epilogue); epilogue_cost += record_stmt_cost (cost_vec, 1, vec_to_scalar, node, 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 += record_stmt_cost (cost_vec, nelements + nelements - 1, vector_stmt, node, 0, vect_epilogue); } } if (dump_enabled_p ()) dump_printf (MSG_NOTE, "vect_model_reduction_cost: inside_cost = %d, " "prologue_cost = %d, epilogue_cost = %d .\n", inside_cost, prologue_cost, epilogue_cost); } /* SEQ is a sequence of instructions that initialize the reduction described by REDUC_INFO. Emit them in the appropriate place. */ static void vect_emit_reduction_init_stmts (loop_vec_info loop_vinfo, stmt_vec_info reduc_info, gimple *seq) { if (reduc_info->reused_accumulator) { /* When reusing an accumulator from the main loop, we only need initialization instructions if the main loop can be skipped. In that case, emit the initialization instructions at the end of the guard block that does the skip. */ edge skip_edge = loop_vinfo->skip_main_loop_edge; gcc_assert (skip_edge); gimple_stmt_iterator gsi = gsi_last_bb (skip_edge->src); gsi_insert_seq_before (&gsi, seq, GSI_SAME_STMT); } else { /* The normal case: emit the initialization instructions on the preheader edge. */ class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), seq); } } /* Get at the initial defs for the reduction PHIs for REDUC_INFO, which performs a reduction involving GROUP_SIZE scalar statements. 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 (loop_vec_info loop_vinfo, stmt_vec_info reduc_info, tree vector_type, vec<tree> *vec_oprnds, unsigned int number_of_vectors, unsigned int group_size, tree neutral_op) { vec<tree> &initial_values = reduc_info->reduc_initial_values; unsigned HOST_WIDE_INT nunits; unsigned j, number_of_places_left_in_vector; unsigned int i; gcc_assert (group_size == initial_values.length () || 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. For example, we have two scalar operands, s1 and s2 (e.g., group of strided accesses of size two), while NUNITS is four (i.e., four scalars of this type can be packed in a vector). The output vector will contain 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. 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_places_left_in_vector = nunits; bool constant_p = true; tree_vector_builder elts (vector_type, nunits, 1); elts.quick_grow (nunits); gimple_seq ctor_seq = NULL; if (neutral_op && !useless_type_conversion_p (TREE_TYPE (vector_type), TREE_TYPE (neutral_op))) neutral_op = gimple_convert (&ctor_seq, TREE_TYPE (vector_type), neutral_op); for (j = 0; j < nunits * number_of_vectors; ++j) { tree op; i = j % group_size; /* Get the def before the loop. In reduction chain we have only one initial value. Else we have as many as PHIs in the group. */ if (i >= initial_values.length () || (j > i && neutral_op)) op = neutral_op; else { if (!useless_type_conversion_p (TREE_TYPE (vector_type), TREE_TYPE (initial_values[i]))) initial_values[i] = gimple_convert (&ctor_seq, TREE_TYPE (vector_type), initial_values[i]); op = initial_values[i]; } /* Create 'vect_ = {op0,op1,...,opn}'. */ number_of_places_left_in_vector--; elts[nunits - number_of_places_left_in_vector - 1] = op; if (!CONSTANT_CLASS_P (op)) constant_p = false; if (number_of_places_left_in_vector == 0) { 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 && operand_equal_p (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. */ duplicate_and_interleave (loop_vinfo, &ctor_seq, vector_type, elts, number_of_vectors, *vec_oprnds); break; } vec_oprnds->quick_push (init); number_of_places_left_in_vector = nunits; elts.new_vector (vector_type, nunits, 1); elts.quick_grow (nunits); constant_p = true; } } if (ctor_seq != NULL) vect_emit_reduction_init_stmts (loop_vinfo, reduc_info, ctor_seq); } /* For a statement STMT_INFO taking part in a reduction operation return the stmt_vec_info the meta information is stored on. */ stmt_vec_info info_for_reduction (vec_info *vinfo, stmt_vec_info stmt_info) { stmt_info = vect_orig_stmt (stmt_info); gcc_assert (STMT_VINFO_REDUC_DEF (stmt_info)); if (!is_a <gphi *> (stmt_info->stmt) || !VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (stmt_info))) stmt_info = STMT_VINFO_REDUC_DEF (stmt_info); gphi *phi = as_a <gphi *> (stmt_info->stmt); if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_double_reduction_def) { if (gimple_phi_num_args (phi) == 1) stmt_info = STMT_VINFO_REDUC_DEF (stmt_info); } else if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle) { stmt_vec_info info = vinfo->lookup_def (vect_phi_initial_value (phi)); if (info && STMT_VINFO_DEF_TYPE (info) == vect_double_reduction_def) stmt_info = info; } return stmt_info; } /* See if LOOP_VINFO is an epilogue loop whose main loop had a reduction that REDUC_INFO can build on. Adjust REDUC_INFO and return true if so, otherwise return false. */ static bool vect_find_reusable_accumulator (loop_vec_info loop_vinfo, stmt_vec_info reduc_info, tree vectype) { loop_vec_info main_loop_vinfo = LOOP_VINFO_ORIG_LOOP_INFO (loop_vinfo); if (!main_loop_vinfo) return false; if (STMT_VINFO_REDUC_TYPE (reduc_info) != TREE_CODE_REDUCTION) return false; unsigned int num_phis = reduc_info->reduc_initial_values.length (); auto_vec<tree, 16> main_loop_results (num_phis); auto_vec<tree, 16> initial_values (num_phis); if (edge main_loop_edge = loop_vinfo->main_loop_edge) { /* The epilogue loop can be entered either from the main loop or from an earlier guard block. */ edge skip_edge = loop_vinfo->skip_main_loop_edge; for (tree incoming_value : reduc_info->reduc_initial_values) { /* Look for: INCOMING_VALUE = phi<MAIN_LOOP_RESULT(main loop), INITIAL_VALUE(guard block)>. */ gcc_assert (TREE_CODE (incoming_value) == SSA_NAME); gphi *phi = as_a <gphi *> (SSA_NAME_DEF_STMT (incoming_value)); gcc_assert (gimple_bb (phi) == main_loop_edge->dest); tree from_main_loop = PHI_ARG_DEF_FROM_EDGE (phi, main_loop_edge); tree from_skip = PHI_ARG_DEF_FROM_EDGE (phi, skip_edge); main_loop_results.quick_push (from_main_loop); initial_values.quick_push (from_skip); } } else /* The main loop dominates the epilogue loop. */ main_loop_results.splice (reduc_info->reduc_initial_values); /* See if the main loop has the kind of accumulator we need. */ vect_reusable_accumulator *accumulator = main_loop_vinfo->reusable_accumulators.get (main_loop_results[0]); if (!accumulator || num_phis != accumulator->reduc_info->reduc_scalar_results.length () || !std::equal (main_loop_results.begin (), main_loop_results.end (), accumulator->reduc_info->reduc_scalar_results.begin ())) return false; /* Handle the case where we can reduce wider vectors to narrower ones. */ tree old_vectype = TREE_TYPE (accumulator->reduc_input); unsigned HOST_WIDE_INT m; if (!constant_multiple_p (TYPE_VECTOR_SUBPARTS (old_vectype), TYPE_VECTOR_SUBPARTS (vectype), &m)) return false; /* Check the intermediate vector types and operations are available. */ tree prev_vectype = old_vectype; poly_uint64 intermediate_nunits = TYPE_VECTOR_SUBPARTS (old_vectype); while (known_gt (intermediate_nunits, TYPE_VECTOR_SUBPARTS (vectype))) { intermediate_nunits = exact_div (intermediate_nunits, 2); tree intermediate_vectype = get_related_vectype_for_scalar_type (TYPE_MODE (vectype), TREE_TYPE (vectype), intermediate_nunits); if (!intermediate_vectype || !directly_supported_p (STMT_VINFO_REDUC_CODE (reduc_info), intermediate_vectype) || !can_vec_extract (TYPE_MODE (prev_vectype), TYPE_MODE (intermediate_vectype))) return false; prev_vectype = intermediate_vectype; } /* Non-SLP reductions might apply an adjustment after the reduction operation, in order to simplify the initialization of the accumulator. If the epilogue loop carries on from where the main loop left off, it should apply the same adjustment to the final reduction result. If the epilogue loop can also be entered directly (rather than via the main loop), we need to be able to handle that case in the same way, with the same adjustment. (In principle we could add a PHI node to select the correct adjustment, but in practice that shouldn't be necessary.) */ tree main_adjustment = STMT_VINFO_REDUC_EPILOGUE_ADJUSTMENT (accumulator->reduc_info); if (loop_vinfo->main_loop_edge && main_adjustment) { gcc_assert (num_phis == 1); tree initial_value = initial_values[0]; /* Check that we can use INITIAL_VALUE as the adjustment and initialize the accumulator with a neutral value instead. */ if (!operand_equal_p (initial_value, main_adjustment)) return false; code_helper code = STMT_VINFO_REDUC_CODE (reduc_info); initial_values[0] = neutral_op_for_reduction (TREE_TYPE (initial_value), code, initial_value); } STMT_VINFO_REDUC_EPILOGUE_ADJUSTMENT (reduc_info) = main_adjustment; reduc_info->reduc_initial_values.truncate (0); reduc_info->reduc_initial_values.splice (initial_values); reduc_info->reused_accumulator = accumulator; return true; } /* Reduce the vector VEC_DEF down to VECTYPE with reduction operation CODE emitting stmts before GSI. Returns a vector def of VECTYPE. */ static tree vect_create_partial_epilog (tree vec_def, tree vectype, code_helper code, gimple_seq *seq) { unsigned nunits = TYPE_VECTOR_SUBPARTS (TREE_TYPE (vec_def)).to_constant (); unsigned nunits1 = TYPE_VECTOR_SUBPARTS (vectype).to_constant (); tree stype = TREE_TYPE (vectype); tree new_temp = vec_def; while (nunits > nunits1) { nunits /= 2; tree vectype1 = get_related_vectype_for_scalar_type (TYPE_MODE (vectype), stype, nunits); unsigned int bitsize = tree_to_uhwi (TYPE_SIZE (vectype1)); /* 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; gimple *epilog_stmt; 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))); gimple_seq_add_stmt_without_update (seq, epilog_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 (bitsize))); gimple_seq_add_stmt_without_update (seq, epilog_stmt); } else { /* Extract via punning to appropriately sized integer mode vector. */ tree eltype = build_nonstandard_integer_type (bitsize, 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)); gimple_seq_add_stmt_without_update (seq, epilog_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))); gimple_seq_add_stmt_without_update (seq, epilog_stmt); dst1 = make_ssa_name (vectype1); epilog_stmt = gimple_build_assign (dst1, VIEW_CONVERT_EXPR, build1 (VIEW_CONVERT_EXPR, vectype1, tem)); gimple_seq_add_stmt_without_update (seq, epilog_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 (bitsize))); gimple_seq_add_stmt_without_update (seq, epilog_stmt); dst2 = make_ssa_name (vectype1); epilog_stmt = gimple_build_assign (dst2, VIEW_CONVERT_EXPR, build1 (VIEW_CONVERT_EXPR, vectype1, tem)); gimple_seq_add_stmt_without_update (seq, epilog_stmt); } new_temp = gimple_build (seq, code, vectype1, dst1, dst2); } return new_temp; } /* Function vect_create_epilog_for_reduction Create code at the loop-epilog to finalize the result of a reduction computation. STMT_INFO is the scalar reduction stmt that is being vectorized. SLP_NODE is an SLP node containing a group of reduction statements. The first one in this group is STMT_INFO. SLP_NODE_INSTANCE is the SLP node instance containing SLP_NODE REDUC_INDEX says which rhs operand of the STMT_INFO is the reduction phi (counting from 0) LOOP_EXIT is the edge to update in the merge block. In the case of a single exit this edge is always the main loop exit. This function: 1. Completes the reduction def-use cycles. 2. "Reduces" each vector of partial results VECT_DEFS into a single result, 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. The flow at the entry to this function: loop: vec_def = phi <vec_init, null> # REDUCTION_PHI 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> use <s_out0> The above is transformed by this function into: loop: vec_def = phi <vec_init, VECT_DEF> # REDUCTION_PHI 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 v_out2 = reduce <v_out1> s_out3 = extract_field <v_out2, 0> s_out4 = adjust_result <s_out3> use <s_out4> use <s_out4> */ static void vect_create_epilog_for_reduction (loop_vec_info loop_vinfo, stmt_vec_info stmt_info, slp_tree slp_node, slp_instance slp_node_instance, edge loop_exit) { stmt_vec_info reduc_info = info_for_reduction (loop_vinfo, stmt_info); gcc_assert (reduc_info->is_reduc_info); /* For double reductions we need to get at the inner loop reduction stmt which has the meta info attached. Our stmt_info is that of the loop-closed PHI of the inner loop which we remember as def for the reduction PHI generation. */ bool double_reduc = false; if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_double_reduction_def) { double_reduc = true; stmt_info = loop_vinfo->lookup_def (gimple_phi_arg_def (stmt_info->stmt, 0)); stmt_info = vect_stmt_to_vectorize (stmt_info); } code_helper code = STMT_VINFO_REDUC_CODE (reduc_info); internal_fn reduc_fn = STMT_VINFO_REDUC_FN (reduc_info); tree vectype; machine_mode mode; class loop *loop = LOOP_VINFO_LOOP (loop_vinfo), *outer_loop = NULL; basic_block exit_bb; tree scalar_dest; tree scalar_type; gimple *new_phi = NULL, *phi = NULL; gimple_stmt_iterator exit_gsi; tree new_temp = NULL_TREE, new_name, new_scalar_dest; gimple *epilog_stmt = NULL; gimple *exit_phi; tree bitsize; tree def; tree orig_name, scalar_result; imm_use_iterator imm_iter, phi_imm_iter; use_operand_p use_p, phi_use_p; gimple *use_stmt; auto_vec<tree> reduc_inputs; int j, i; vec<tree> &scalar_results = reduc_info->reduc_scalar_results; unsigned int k; /* SLP reduction without reduction chain, e.g., # a1 = phi <a2, a0> # b1 = phi <b2, b0> a2 = operation (a1) b2 = operation (b1) */ const bool slp_reduc = SLP_INSTANCE_KIND (slp_node_instance) != slp_inst_kind_reduc_chain; tree induction_index = NULL_TREE; unsigned int group_size = SLP_TREE_LANES (slp_node); if (nested_in_vect_loop_p (loop, stmt_info)) { outer_loop = loop; loop = loop->inner; gcc_assert (double_reduc); } vectype = STMT_VINFO_REDUC_VECTYPE (reduc_info); gcc_assert (vectype); mode = TYPE_MODE (vectype); tree induc_val = NULL_TREE; tree adjustment_def = NULL; /* Optimize: for induction condition reduction, if we can't use zero for induc_val, use initial_def. */ if (STMT_VINFO_REDUC_TYPE (reduc_info) == INTEGER_INDUC_COND_REDUCTION) induc_val = STMT_VINFO_VEC_INDUC_COND_INITIAL_VAL (reduc_info); else if (double_reduc) ; else adjustment_def = STMT_VINFO_REDUC_EPILOGUE_ADJUSTMENT (reduc_info); stmt_vec_info single_live_out_stmt[] = { stmt_info }; array_slice<const stmt_vec_info> live_out_stmts = single_live_out_stmt; if (slp_reduc) /* All statements produce live-out values. */ live_out_stmts = SLP_TREE_SCALAR_STMTS (slp_node); unsigned vec_num = SLP_TREE_VEC_DEFS (slp_node_instance->reduc_phis).length (); /* For cond reductions we want to create a new vector (INDEX_COND_EXPR) which is updated with the current index of the loop for every match of the original loop's cond_expr (VEC_STMT). This results in a vector containing the last time the condition passed for that vector lane. The first match will be a 1 to allow 0 to be used for non-matching indexes. If there are no matches at all then the vector will be all zeroes. PR92772: This algorithm is broken for architectures that support masked vectors, but do not provide fold_extract_last. */ if (STMT_VINFO_REDUC_TYPE (reduc_info) == COND_REDUCTION) { auto_vec<std::pair<tree, bool>, 2> ccompares; slp_tree cond_node = slp_node_instance->root; while (cond_node != slp_node_instance->reduc_phis) { stmt_vec_info cond_info = SLP_TREE_REPRESENTATIVE (cond_node); int slp_reduc_idx; if (gimple_assign_rhs_code (cond_info->stmt) == COND_EXPR) { gimple *vec_stmt = SSA_NAME_DEF_STMT (SLP_TREE_VEC_DEFS (cond_node)[0]); gcc_assert (gimple_assign_rhs_code (vec_stmt) == VEC_COND_EXPR); ccompares.safe_push (std::make_pair (gimple_assign_rhs1 (vec_stmt), STMT_VINFO_REDUC_IDX (cond_info) == 2)); /* ??? We probably want to have REDUC_IDX on the SLP node? We have both three and four children COND_EXPR nodes dependent on whether the comparison is still embedded as GENERIC. So work backwards. */ slp_reduc_idx = (SLP_TREE_CHILDREN (cond_node).length () - 3 + STMT_VINFO_REDUC_IDX (cond_info)); } else slp_reduc_idx = STMT_VINFO_REDUC_IDX (cond_info); cond_node = SLP_TREE_CHILDREN (cond_node)[slp_reduc_idx]; } gcc_assert (ccompares.length () != 0); tree indx_before_incr, indx_after_incr; poly_uint64 nunits_out = TYPE_VECTOR_SUBPARTS (vectype); int scalar_precision = GET_MODE_PRECISION (SCALAR_TYPE_MODE (TREE_TYPE (vectype))); tree cr_index_scalar_type = make_unsigned_type (scalar_precision); tree cr_index_vector_type = get_related_vectype_for_scalar_type (TYPE_MODE (vectype), cr_index_scalar_type, TYPE_VECTOR_SUBPARTS (vectype)); /* First we create a simple vector induction variable which starts with the values {1,2,3,...} (SERIES_VECT) and increments by the vector size (STEP). */ /* Create a {1,2,3,...} vector. */ 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); tree vec_step = build_vector_from_val (cr_index_vector_type, step); /* Create an induction variable. */ gimple_stmt_iterator incr_gsi; bool insert_after; vect_iv_increment_position (LOOP_VINFO_IV_EXIT (loop_vinfo), &incr_gsi, &insert_after); create_iv (series_vect, PLUS_EXPR, vec_step, NULL_TREE, loop, &incr_gsi, insert_after, &indx_before_incr, &indx_after_incr); /* Next create a new phi node vector (NEW_PHI_TREE) which starts filled with zeros (VEC_ZERO). */ /* Create a vector of 0s. */ tree zero = build_zero_cst (cr_index_scalar_type); tree vec_zero = build_vector_from_val (cr_index_vector_type, zero); /* 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); add_phi_arg (as_a <gphi *> (new_phi), vec_zero, loop_preheader_edge (loop), UNKNOWN_LOCATION); /* Now take the condition from the loops original cond_exprs and produce a new cond_exprs (INDEX_COND_EXPR) which for every match uses values from the induction variable (INDEX_BEFORE_INCR) otherwise uses values from the phi node (NEW_PHI_TREE). Finally, we update the phi (NEW_PHI_TREE) to take the value of the new cond_expr (INDEX_COND_EXPR). */ gimple_seq stmts = NULL; for (int i = ccompares.length () - 1; i != -1; --i) { tree ccompare = ccompares[i].first; if (ccompares[i].second) new_phi_tree = gimple_build (&stmts, VEC_COND_EXPR, cr_index_vector_type, ccompare, indx_before_incr, new_phi_tree); else new_phi_tree = gimple_build (&stmts, VEC_COND_EXPR, cr_index_vector_type, ccompare, new_phi_tree, indx_before_incr); } gsi_insert_seq_before (&incr_gsi, stmts, GSI_SAME_STMT); /* Update the phi with the vec cond. */ induction_index = new_phi_tree; add_phi_arg (as_a <gphi *> (new_phi), induction_index, loop_latch_edge (loop), UNKNOWN_LOCATION); } /* 2. Create epilog code. The reduction epilog code operates across the elements of the vector of partial results computed by the vectorized loop. The reduction epilog code consists of: step 1: compute the scalar result in a vector (v_out2) step 2: extract the scalar result (s_out3) from the vector (v_out2) 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_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. The overall epilog code looks like this: s_out0 = phi <s_loop> # original EXIT_PHI v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI v_out2 = reduce <v_out1> # step 1 s_out3 = extract_field <v_out2, 0> # step 2 s_out4 = adjust_result <s_out3> # step 3 (step 3 is optional, and steps 1 and 2 may be combined). Lastly, the uses of s_out0 are replaced by s_out4. */ /* 2.1 Create new loop-exit-phis to preserve loop-closed form: v_out1 = phi <VECT_DEF> Store them in NEW_PHIS. */ if (double_reduc) loop = outer_loop; /* We need to reduce values in all exits. */ exit_bb = loop_exit->dest; exit_gsi = gsi_after_labels (exit_bb); reduc_inputs.create (vec_num); for (unsigned i = 0; i < vec_num; i++) { gimple_seq stmts = NULL; def = vect_get_slp_vect_def (slp_node, i); tree new_def = copy_ssa_name (def); phi = create_phi_node (new_def, exit_bb); if (LOOP_VINFO_IV_EXIT (loop_vinfo) == loop_exit) SET_PHI_ARG_DEF (phi, loop_exit->dest_idx, def); else { for (unsigned k = 0; k < gimple_phi_num_args (phi); k++) SET_PHI_ARG_DEF (phi, k, def); } new_def = gimple_convert (&stmts, vectype, new_def); reduc_inputs.quick_push (new_def); gsi_insert_seq_before (&exit_gsi, stmts, GSI_SAME_STMT); } /* 2.2 Get the relevant tree-code to use in the epilog for schemes 2,3 (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 taken from the original stmt that the pattern-stmt (STMT) replaces. Otherwise (it is a regular reduction) - the tree-code and scalar-def are taken from STMT. */ stmt_vec_info orig_stmt_info = vect_orig_stmt (stmt_info); if (orig_stmt_info != stmt_info) { /* Reduction pattern */ gcc_assert (STMT_VINFO_IN_PATTERN_P (orig_stmt_info)); gcc_assert (STMT_VINFO_RELATED_STMT (orig_stmt_info) == stmt_info); } scalar_dest = gimple_get_lhs (orig_stmt_info->stmt); scalar_type = TREE_TYPE (scalar_dest); scalar_results.truncate (0); scalar_results.reserve_exact (group_size); new_scalar_dest = vect_create_destination_var (scalar_dest, NULL); bitsize = TYPE_SIZE (scalar_type); /* True if we should implement SLP_REDUC using native reduction operations instead of scalar operations. */ const bool 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> a2 = operation (a1) a3 = operation (a2), we may end up with more than one vector result. Here we reduce them to one vector. The same is true for a SLP reduction, e.g., # a1 = phi <a2, a0> # b1 = phi <b2, b0> a2 = operation (a1) b2 = operation (a2), where we can end up with more than one vector as well. We can easily accumulate vectors when the number of vector elements is a multiple of the SLP group size. The same is true if we couldn't use a single defuse cycle. */ if (!slp_reduc || direct_slp_reduc || (slp_reduc && constant_multiple_p (TYPE_VECTOR_SUBPARTS (vectype), group_size))) { gimple_seq stmts = NULL; tree single_input = reduc_inputs[0]; for (k = 1; k < reduc_inputs.length (); k++) single_input = gimple_build (&stmts, code, vectype, single_input, reduc_inputs[k]); gsi_insert_seq_before (&exit_gsi, stmts, GSI_SAME_STMT); reduc_inputs.truncate (0); reduc_inputs.safe_push (single_input); } tree orig_reduc_input = reduc_inputs[0]; /* If this loop is an epilogue loop that can be skipped after the main loop, we can only share a reduction operation between the main loop and the epilogue if we put it at the target of the skip edge. We can still reuse accumulators if this check fails. Doing so has the minor(?) benefit of making the epilogue loop's scalar result independent of the main loop's scalar result. */ bool unify_with_main_loop_p = false; if (reduc_info->reused_accumulator && loop_vinfo->skip_this_loop_edge && single_succ_p (exit_bb) && single_succ (exit_bb) == loop_vinfo->skip_this_loop_edge->dest) { unify_with_main_loop_p = true; basic_block reduc_block = loop_vinfo->skip_this_loop_edge->dest; reduc_inputs[0] = make_ssa_name (vectype); gphi *new_phi = create_phi_node (reduc_inputs[0], reduc_block); add_phi_arg (new_phi, orig_reduc_input, single_succ_edge (exit_bb), UNKNOWN_LOCATION); add_phi_arg (new_phi, reduc_info->reused_accumulator->reduc_input, loop_vinfo->skip_this_loop_edge, UNKNOWN_LOCATION); exit_gsi = gsi_after_labels (reduc_block); } /* Shouldn't be used beyond this point. */ exit_bb = nullptr; if (STMT_VINFO_REDUC_TYPE (reduc_info) == COND_REDUCTION && reduc_fn != IFN_LAST) { /* For condition reductions, we have a vector (REDUC_INPUTS 0) containing various data values where the condition matched and another vector (INDUCTION_INDEX) containing all the indexes of those matches. We need to extract the last matching index (which will be the index with highest value) and use this to index into the data vector. For the case where there were no matches, the data vector will contain all default values and the index vector will be all zeros. */ /* Get various versions of the type of the vector of indexes. */ tree index_vec_type = TREE_TYPE (induction_index); gcc_checking_assert (TYPE_UNSIGNED (index_vec_type)); tree index_scalar_type = TREE_TYPE (index_vec_type); tree index_vec_cmp_type = truth_type_for (index_vec_type); /* Get an unsigned integer version of the type of the data vector. */ int scalar_precision = GET_MODE_PRECISION (SCALAR_TYPE_MODE (scalar_type)); tree scalar_type_unsigned = make_unsigned_type (scalar_precision); tree vectype_unsigned = get_same_sized_vectype (scalar_type_unsigned, vectype); /* First we need to create a vector (ZERO_VEC) of zeros and another vector (MAX_INDEX_VEC) filled with the last matching index, which we can create using a MAX reduction and then expanding. In the case where the loop never made any matches, the max index will be zero. */ /* Vector of {0, 0, 0,...}. */ tree zero_vec = build_zero_cst (vectype); /* Find maximum value from the vector of found indexes. */ tree max_index = make_ssa_name (index_scalar_type); 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,...}. */ tree max_index_vec = make_ssa_name (index_vec_type); tree max_index_vec_rhs = build_vector_from_val (index_vec_type, max_index); gimple *max_index_vec_stmt = gimple_build_assign (max_index_vec, max_index_vec_rhs); gsi_insert_before (&exit_gsi, max_index_vec_stmt, GSI_SAME_STMT); /* Next we compare the new vector (MAX_INDEX_VEC) full of max indexes with the vector (INDUCTION_INDEX) of found indexes, choosing values from the data vector (REDUC_INPUTS 0) for matches, 0 (ZERO_VEC) otherwise. Only one value should match, resulting in a vector (VEC_COND) with one data value and the rest zeros. In the case where the loop never made any matches, every index will match, resulting in a vector with all data values (which will all be the default value). */ /* Compare the max index vector to the vector of found indexes to find the position of the max value. */ tree vec_compare = make_ssa_name (index_vec_cmp_type); gimple *vec_compare_stmt = gimple_build_assign (vec_compare, EQ_EXPR, induction_index, max_index_vec); gsi_insert_before (&exit_gsi, vec_compare_stmt, GSI_SAME_STMT); /* Use the compare to choose either values from the data vector or zero. */ tree vec_cond = make_ssa_name (vectype); gimple *vec_cond_stmt = gimple_build_assign (vec_cond, VEC_COND_EXPR, vec_compare, reduc_inputs[0], zero_vec); gsi_insert_before (&exit_gsi, vec_cond_stmt, GSI_SAME_STMT); /* Finally we need to extract the data value from the vector (VEC_COND) into a scalar (MATCHED_DATA_REDUC). Logically we want to do a OR reduction, but because this doesn't exist, we can use a MAX reduction instead. The data value might be signed or a float so we need to cast it first. In the case where the loop never made any matches, the data values are all identical, and so will reduce down correctly. */ /* Make the matched data values unsigned. */ tree vec_cond_cast = make_ssa_name (vectype_unsigned); tree vec_cond_cast_rhs = build1 (VIEW_CONVERT_EXPR, vectype_unsigned, vec_cond); gimple *vec_cond_cast_stmt = gimple_build_assign (vec_cond_cast, VIEW_CONVERT_EXPR, vec_cond_cast_rhs); gsi_insert_before (&exit_gsi, vec_cond_cast_stmt, GSI_SAME_STMT); /* Reduce down to a scalar value. */ tree data_reduc = make_ssa_name (scalar_type_unsigned); 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 result. */ gimple_seq stmts = NULL; new_temp = gimple_build (&stmts, VIEW_CONVERT_EXPR, scalar_type, data_reduc); gsi_insert_seq_before (&exit_gsi, stmts, GSI_SAME_STMT); scalar_results.safe_push (new_temp); } else if (STMT_VINFO_REDUC_TYPE (reduc_info) == COND_REDUCTION && reduc_fn == IFN_LAST) { /* Condition reduction without supported IFN_REDUC_MAX. Generate idx = 0; idx_val = induction_index[0]; val = data_reduc[0]; for (idx = 0, val = init, i = 0; i < nelts; ++i) if (induction_index[i] > idx_val) val = data_reduc[i], idx_val = induction_index[i]; return val; */ tree data_eltype = TREE_TYPE (vectype); tree idx_eltype = TREE_TYPE (TREE_TYPE (induction_index)); unsigned HOST_WIDE_INT el_size = tree_to_uhwi (TYPE_SIZE (idx_eltype)); 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) { tree old_idx_val = idx_val; tree old_val = val; idx_val = make_ssa_name (idx_eltype); epilog_stmt = gimple_build_assign (idx_val, BIT_FIELD_REF, build3 (BIT_FIELD_REF, idx_eltype, induction_index, bitsize_int (el_size), bitsize_int (off))); gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); val = make_ssa_name (data_eltype); epilog_stmt = gimple_build_assign (val, BIT_FIELD_REF, build3 (BIT_FIELD_REF, data_eltype, reduc_inputs[0], bitsize_int (el_size), bitsize_int (off))); gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); if (off != 0) { tree new_idx_val = idx_val; if (off != v_size - el_size) { new_idx_val = make_ssa_name (idx_eltype); epilog_stmt = gimple_build_assign (new_idx_val, MAX_EXPR, idx_val, old_idx_val); gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); } tree cond = make_ssa_name (boolean_type_node); epilog_stmt = gimple_build_assign (cond, GT_EXPR, idx_val, old_idx_val); gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); tree new_val = make_ssa_name (data_eltype); epilog_stmt = gimple_build_assign (new_val, COND_EXPR, cond, val, old_val); gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); idx_val = new_idx_val; val = new_val; } } /* Convert the reduced value back to the result type and set as the result. */ gimple_seq stmts = NULL; val = gimple_convert (&stmts, scalar_type, val); gsi_insert_seq_before (&exit_gsi, stmts, GSI_SAME_STMT); scalar_results.safe_push (val); } /* 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_fn != IFN_LAST && (!slp_reduc || group_size == 1)) { tree tmp; tree vec_elem_type; /* Case 1: Create: v_out2 = reduc_expr <v_out1> */ if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "Reduce using direct vector reduction.\n"); gimple_seq stmts = NULL; vec_elem_type = TREE_TYPE (vectype); new_temp = gimple_build (&stmts, as_combined_fn (reduc_fn), vec_elem_type, reduc_inputs[0]); new_temp = gimple_convert (&stmts, scalar_type, new_temp); gsi_insert_seq_before (&exit_gsi, stmts, GSI_SAME_STMT); if ((STMT_VINFO_REDUC_TYPE (reduc_info) == INTEGER_INDUC_COND_REDUCTION) && induc_val) { /* 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 = make_ssa_name (boolean_type_node); epilog_stmt = gimple_build_assign (zcompare, EQ_EXPR, new_temp, induc_val); gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); tree initial_def = reduc_info->reduc_initial_values[0]; tmp = make_ssa_name (new_scalar_dest); epilog_stmt = gimple_build_assign (tmp, COND_EXPR, zcompare, initial_def, new_temp); gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); new_temp = tmp; } scalar_results.safe_push (new_temp); } else if (direct_slp_reduc) { /* Here we create one vector for each of the 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 (reduc_inputs.length () == 1); gcc_assert (pow2p_hwi (group_size)); 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 = truth_type_for (index_type); /* Create a vector that, for each element, identifies which of the 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; tree neutral_op = neutral_op_for_reduction (TREE_TYPE (vectype), code, NULL_TREE, false); 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 = reduc_info->reduc_initial_values[i]; scalar_value = gimple_convert (&seq, TREE_TYPE (vectype), scalar_value); 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 REDUC_INPUTS[0] 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 ? reduc_inputs[0] : vector_identity; VEC is now suitable for a full vector reduction. */ tree vec = gimple_build (&seq, VEC_COND_EXPR, vectype, sel, reduc_inputs[0], 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; tree vec_temp; gcc_assert (slp_reduc || reduc_inputs.length () == 1); /* 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 stype = TREE_TYPE (vectype); unsigned nunits = TYPE_VECTOR_SUBPARTS (vectype).to_constant (); unsigned nunits1 = nunits; if ((mode1 = targetm.vectorize.split_reduction (mode)) != mode && reduc_inputs.length () == 1) { nunits1 = GET_MODE_NUNITS (mode1).to_constant (); /* For SLP reductions we have to make sure lanes match up, but since we're doing individual element final reduction reducing vector width here is even more important. ??? We can also separate lanes with permutes, for the common case of power-of-two group-size odd/even extracts would work. */ if (slp_reduc && nunits != nunits1) { nunits1 = least_common_multiple (nunits1, group_size); gcc_assert (exact_log2 (nunits1) != -1 && nunits1 <= nunits); } } if (!slp_reduc && (mode1 = targetm.vectorize.split_reduction (mode)) != mode) nunits1 = GET_MODE_NUNITS (mode1).to_constant (); tree vectype1 = get_related_vectype_for_scalar_type (TYPE_MODE (vectype), stype, nunits1); reduce_with_shift = have_whole_vector_shift (mode1); if (!VECTOR_MODE_P (mode1) || !directly_supported_p (code, vectype1)) 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. */ gimple_seq stmts = NULL; new_temp = vect_create_partial_epilog (reduc_inputs[0], vectype1, code, &stmts); gsi_insert_seq_before (&exit_gsi, stmts, GSI_SAME_STMT); reduc_inputs[0] = new_temp; if (reduce_with_shift && (!slp_reduc || group_size == 1)) { 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 (vectype1); /* Case 2: Create: for (offset = nelements/2; offset >= 1; offset/=2) { Create: va' = vec_shift <va, offset> Create: va = vop <va, va'> } */ tree rhs; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "Reduce using vector shifts\n"); gimple_seq stmts = NULL; new_temp = gimple_convert (&stmts, vectype1, new_temp); for (elt_offset = nelements / 2; elt_offset >= 1; elt_offset /= 2) { calc_vec_perm_mask_for_shift (elt_offset, nelements, &sel); indices.new_vector (sel, 2, nelements); tree mask = vect_gen_perm_mask_any (vectype1, indices); new_name = gimple_build (&stmts, VEC_PERM_EXPR, vectype1, new_temp, zero_vec, mask); new_temp = gimple_build (&stmts, code, vectype1, new_name, new_temp); } gsi_insert_seq_before (&exit_gsi, stmts, GSI_SAME_STMT); /* 2.4 Extract the final scalar result. Create: s_out3 = extract_field <v_out2, bitpos> */ if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "extract scalar result\n"); rhs = build3 (BIT_FIELD_REF, scalar_type, new_temp, bitsize, 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); gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); scalar_results.safe_push (new_temp); } else { /* Case 3: Create: s = extract_field <v_out2, 0> for (offset = element_size; offset < vector_size; offset += element_size;) { Create: s' = extract_field <v_out2, offset> Create: s = op <s, s'> // For non SLP cases } */ if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "Reduce using scalar code.\n"); int vec_size_in_bits = tree_to_uhwi (TYPE_SIZE (vectype1)); int element_bitsize = tree_to_uhwi (bitsize); tree compute_type = TREE_TYPE (vectype); gimple_seq stmts = NULL; FOR_EACH_VEC_ELT (reduc_inputs, i, vec_temp) { int bit_offset; new_temp = gimple_build (&stmts, BIT_FIELD_REF, compute_type, vec_temp, bitsize, bitsize_zero_node); /* In SLP we don't need to apply reduction operation, so we just collect s' values in SCALAR_RESULTS. */ if (slp_reduc) scalar_results.safe_push (new_temp); for (bit_offset = element_bitsize; bit_offset < vec_size_in_bits; bit_offset += element_bitsize) { tree bitpos = bitsize_int (bit_offset); new_name = gimple_build (&stmts, BIT_FIELD_REF, compute_type, vec_temp, bitsize, bitpos); if (slp_reduc) { /* In SLP we don't need to apply reduction operation, so we just collect s' values in SCALAR_RESULTS. */ new_temp = new_name; scalar_results.safe_push (new_name); } else new_temp = gimple_build (&stmts, code, compute_type, new_name, new_temp); } } /* The only case where we need to reduce scalar results in a SLP reduction, is unrolling. If the size of SCALAR_RESULTS is greater than GROUP_SIZE, we reduce them combining elements modulo GROUP_SIZE. */ if (slp_reduc) { tree res, first_res, new_res; /* Reduce multiple scalar results in case of SLP unrolling. */ for (j = group_size; scalar_results.iterate (j, &res); j++) { first_res = scalar_results[j % group_size]; new_res = gimple_build (&stmts, code, compute_type, first_res, res); scalar_results[j % group_size] = new_res; } scalar_results.truncate (group_size); for (k = 0; k < group_size; k++) scalar_results[k] = gimple_convert (&stmts, scalar_type, scalar_results[k]); } else { /* Reduction chain - we have one scalar to keep in SCALAR_RESULTS. */ new_temp = gimple_convert (&stmts, scalar_type, new_temp); scalar_results.safe_push (new_temp); } gsi_insert_seq_before (&exit_gsi, stmts, GSI_SAME_STMT); } if ((STMT_VINFO_REDUC_TYPE (reduc_info) == INTEGER_INDUC_COND_REDUCTION) && induc_val) { /* 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 = make_ssa_name (boolean_type_node); epilog_stmt = gimple_build_assign (zcompare, EQ_EXPR, scalar_results[0], induc_val); gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); tree initial_def = reduc_info->reduc_initial_values[0]; tree tmp = make_ssa_name (new_scalar_dest); epilog_stmt = gimple_build_assign (tmp, COND_EXPR, zcompare, initial_def, scalar_results[0]); gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); scalar_results[0] = tmp; } } /* 2.5 Adjust the final result by the initial value of the reduction variable. (When such adjustment is not needed, then 'adjustment_def' is zero). For example, if code is PLUS we create: new_temp = loop_exit_def + adjustment_def */ if (adjustment_def) { gcc_assert (!slp_reduc || group_size == 1); gimple_seq stmts = NULL; if (double_reduc) { gcc_assert (VECTOR_TYPE_P (TREE_TYPE (adjustment_def))); adjustment_def = gimple_convert (&stmts, vectype, adjustment_def); new_temp = gimple_build (&stmts, code, vectype, reduc_inputs[0], adjustment_def); } else { new_temp = scalar_results[0]; gcc_assert (TREE_CODE (TREE_TYPE (adjustment_def)) != VECTOR_TYPE); adjustment_def = gimple_convert (&stmts, TREE_TYPE (vectype), adjustment_def); new_temp = gimple_convert (&stmts, TREE_TYPE (vectype), new_temp); new_temp = gimple_build (&stmts, code, TREE_TYPE (vectype), new_temp, adjustment_def); new_temp = gimple_convert (&stmts, scalar_type, new_temp); } epilog_stmt = gimple_seq_last_stmt (stmts); gsi_insert_seq_before (&exit_gsi, stmts, GSI_SAME_STMT); scalar_results[0] = new_temp; } /* Record this operation if it could be reused by the epilogue loop. */ if (STMT_VINFO_REDUC_TYPE (reduc_info) == TREE_CODE_REDUCTION && reduc_inputs.length () == 1) loop_vinfo->reusable_accumulators.put (scalar_results[0], { orig_reduc_input, reduc_info }); if (double_reduc) loop = outer_loop; /* 2.6 Handle the loop-exit phis. Replace the uses of scalar loop-exit phis with new adjusted scalar results, i.e., replace use <s_out0> with use <s_out4>. Transform: loop_exit: s_out0 = phi <s_loop> # (scalar) EXIT_PHI v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI v_out2 = reduce <v_out1> s_out3 = extract_field <v_out2, 0> s_out4 = adjust_result <s_out3> use <s_out0> use <s_out0> into: loop_exit: s_out0 = phi <s_loop> # (scalar) EXIT_PHI v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI v_out2 = reduce <v_out1> s_out3 = extract_field <v_out2, 0> s_out4 = adjust_result <s_out3> use <s_out4> use <s_out4> */ gcc_assert (live_out_stmts.size () == scalar_results.length ()); auto_vec<gimple *> phis; for (k = 0; k < live_out_stmts.size (); k++) { stmt_vec_info scalar_stmt_info = vect_orig_stmt (live_out_stmts[k]); scalar_dest = gimple_get_lhs (scalar_stmt_info->stmt); /* Find the loop-closed-use at the loop exit of the original scalar result. (The reduction result is expected to have two immediate uses, one at the latch block, and one at the loop exit). For double reductions we are looking for exit phis of the outer loop. */ FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) { if (!flow_bb_inside_loop_p (loop, gimple_bb (USE_STMT (use_p)))) { if (!is_gimple_debug (USE_STMT (use_p)) && gimple_bb (USE_STMT (use_p)) == loop_exit->dest) phis.safe_push (USE_STMT (use_p)); } else { if (double_reduc && gimple_code (USE_STMT (use_p)) == GIMPLE_PHI) { tree phi_res = PHI_RESULT (USE_STMT (use_p)); FOR_EACH_IMM_USE_FAST (phi_use_p, phi_imm_iter, phi_res) { if (!flow_bb_inside_loop_p (loop, gimple_bb (USE_STMT (phi_use_p))) && !is_gimple_debug (USE_STMT (phi_use_p))) phis.safe_push (USE_STMT (phi_use_p)); } } } } FOR_EACH_VEC_ELT (phis, i, exit_phi) { /* Replace the uses: */ orig_name = PHI_RESULT (exit_phi); /* Look for a single use at the target of the skip edge. */ if (unify_with_main_loop_p) { use_operand_p use_p; gimple *user; if (!single_imm_use (orig_name, &use_p, &user)) gcc_unreachable (); orig_name = gimple_get_lhs (user); } scalar_result = scalar_results[k]; FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, orig_name) { FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) SET_USE (use_p, scalar_result); update_stmt (use_stmt); } } phis.truncate (0); } } /* 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. If MASK is nonzero mask the input and then operate on it unconditionally. 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 mask) { 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); /* Re-create a VEC_COND_EXPR to mask the input here in order to be able to perform an unconditional element-wise reduction of it. */ if (mask) { tree masked_vector_rhs = make_temp_ssa_name (vectype, NULL, "masked_vector_rhs"); tree neutral_op = neutral_op_for_reduction (scalar_type, code, NULL_TREE, false); tree vector_identity = build_vector_from_val (vectype, neutral_op); gassign *select = gimple_build_assign (masked_vector_rhs, VEC_COND_EXPR, mask, vector_rhs, vector_identity); gsi_insert_before (gsi, select, GSI_SAME_STMT); vector_rhs = masked_vector_rhs; } 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); /* Fold the vector extract, combining it with a previous reversal like seen in PR90579. */ auto gsi2 = gsi_for_stmt (stmt); if (fold_stmt (&gsi2, follow_all_ssa_edges)) update_stmt (gsi_stmt (gsi2)); 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; } /* Get a masked internal function equivalent to REDUC_FN. VECTYPE_IN is the type of the vector input. */ static internal_fn get_masked_reduction_fn (internal_fn reduc_fn, tree vectype_in) { internal_fn mask_reduc_fn; internal_fn mask_len_reduc_fn; switch (reduc_fn) { case IFN_FOLD_LEFT_PLUS: mask_reduc_fn = IFN_MASK_FOLD_LEFT_PLUS; mask_len_reduc_fn = IFN_MASK_LEN_FOLD_LEFT_PLUS; break; default: return IFN_LAST; } if (direct_internal_fn_supported_p (mask_reduc_fn, vectype_in, OPTIMIZE_FOR_SPEED)) return mask_reduc_fn; if (direct_internal_fn_supported_p (mask_len_reduc_fn, vectype_in, OPTIMIZE_FOR_SPEED)) return mask_len_reduc_fn; return IFN_LAST; } /* 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 (loop_vec_info loop_vinfo, stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, slp_tree slp_node, code_helper code, internal_fn reduc_fn, int num_ops, tree vectype_in, int reduc_index, vec_loop_masks *masks, vec_loop_lens *lens) { class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); tree vectype_out = SLP_TREE_VECTYPE (slp_node); internal_fn mask_reduc_fn = get_masked_reduction_fn (reduc_fn, vectype_in); gcc_assert (!nested_in_vect_loop_p (loop, stmt_info)); bool is_cond_op = false; if (!code.is_tree_code ()) { code = conditional_internal_fn_code (internal_fn (code)); gcc_assert (code != ERROR_MARK); is_cond_op = true; } gcc_assert (TREE_CODE_LENGTH (tree_code (code)) == binary_op); gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (vectype_out), TYPE_VECTOR_SUBPARTS (vectype_in))); /* ??? We should, when transforming the cycle PHI, record the existing scalar def as vector def so looking up the vector def works. This would also allow generalizing this for reduction paths of length > 1 and/or SLP reductions. */ slp_tree reduc_node = SLP_TREE_CHILDREN (slp_node)[reduc_index]; tree reduc_var = vect_get_slp_scalar_def (reduc_node, 0); /* The operands either come from a binary operation or an IFN_COND operation. The former is a gimple assign with binary rhs and the latter is a gimple call with four arguments. */ gcc_assert (num_ops == 2 || num_ops == 4); int group_size = 1; stmt_vec_info scalar_dest_def_info; auto_vec<tree> vec_oprnds0, vec_opmask; vect_get_slp_defs (SLP_TREE_CHILDREN (slp_node)[(is_cond_op ? 2 : 0) + (1 - reduc_index)], &vec_oprnds0); group_size = SLP_TREE_SCALAR_STMTS (slp_node).length (); scalar_dest_def_info = SLP_TREE_SCALAR_STMTS (slp_node)[group_size - 1]; /* For an IFN_COND_OP we also need the vector mask operand. */ if (is_cond_op) vect_get_slp_defs (SLP_TREE_CHILDREN (slp_node)[0], &vec_opmask); gimple *sdef = vect_orig_stmt (scalar_dest_def_info)->stmt; tree scalar_dest = gimple_get_lhs (sdef); tree scalar_type = TREE_TYPE (scalar_dest); int vec_num = vec_oprnds0.length (); 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); if (!HONOR_SIGNED_ZEROS (vectype_out)) ; else { gcc_assert (!HONOR_SIGN_DEPENDENT_ROUNDING (vectype_out)); vector_identity = const_unop (NEGATE_EXPR, vectype_out, vector_identity); } } 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; tree len = NULL_TREE; tree bias = NULL_TREE; if (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) { tree loop_mask = vect_get_loop_mask (loop_vinfo, gsi, masks, vec_num, vectype_in, i); if (is_cond_op) mask = prepare_vec_mask (loop_vinfo, TREE_TYPE (loop_mask), loop_mask, vec_opmask[i], gsi); else mask = loop_mask; } else if (is_cond_op) mask = vec_opmask[i]; if (LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo)) { len = vect_get_loop_len (loop_vinfo, gsi, lens, vec_num, vectype_in, i, 1); signed char biasval = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); bias = build_int_cst (intQI_type_node, biasval); if (!is_cond_op) mask = build_minus_one_cst (truth_type_for (vectype_in)); } /* 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 (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) && mask && mask_reduc_fn == IFN_LAST) 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 || (mask && mask_reduc_fn != IFN_LAST)) { if (mask && len && mask_reduc_fn == IFN_MASK_LEN_FOLD_LEFT_PLUS) new_stmt = gimple_build_call_internal (mask_reduc_fn, 5, reduc_var, def0, mask, len, bias); else if (mask && mask_reduc_fn == IFN_MASK_FOLD_LEFT_PLUS) new_stmt = gimple_build_call_internal (mask_reduc_fn, 3, reduc_var, def0, mask); else 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, tree_code (code), reduc_var, def0, mask); 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, true); } if (i == vec_num - 1) { gimple_set_lhs (new_stmt, scalar_dest); vect_finish_replace_stmt (loop_vinfo, scalar_dest_def_info, new_stmt); } else vect_finish_stmt_generation (loop_vinfo, scalar_dest_def_info, new_stmt, gsi); slp_node->push_vec_def (new_stmt); } return true; } /* Function is_nonwrapping_integer_induction. Check if STMT_VINO (which is part of loop LOOP) both increments and does not cause overflow. */ static bool is_nonwrapping_integer_induction (stmt_vec_info stmt_vinfo, class loop *loop) { 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 (phi)); widest_int ni, max_loop_value, lhs_max; 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 max size of the loop will not wrap. */ if (TYPE_OVERFLOW_UNDEFINED (lhs_type)) return true; if (! max_stmt_executions (loop, &ni)) return false; max_loop_value = wi::mul (wi::to_widest (step), ni, TYPE_SIGN (lhs_type), &overflow); if (overflow) return false; max_loop_value = wi::add (wi::to_widest (base), max_loop_value, TYPE_SIGN (lhs_type), &overflow); if (overflow) return false; return (wi::min_precision (max_loop_value, TYPE_SIGN (lhs_type)) <= TYPE_PRECISION (lhs_type)); } /* Check if masking can be supported by inserting a conditional expression. CODE is the code for the operation. COND_FN is the conditional internal function, if it exists. VECTYPE_IN is the type of the vector input. */ static bool use_mask_by_cond_expr_p (code_helper code, internal_fn cond_fn, tree vectype_in) { if (cond_fn != IFN_LAST && direct_internal_fn_supported_p (cond_fn, vectype_in, OPTIMIZE_FOR_SPEED)) return false; if (code.is_tree_code ()) switch (tree_code (code)) { case DOT_PROD_EXPR: case SAD_EXPR: return true; default: break; } return false; } /* Insert a conditional expression to enable masked vectorization. CODE is the code for the operation. VOP is the array of operands. MASK is the loop mask. GSI is a statement iterator used to place the new conditional expression. */ static void build_vect_cond_expr (code_helper code, tree vop[3], tree mask, gimple_stmt_iterator *gsi) { switch (tree_code (code)) { case DOT_PROD_EXPR: { tree vectype = TREE_TYPE (vop[1]); tree zero = build_zero_cst (vectype); tree masked_op1 = make_temp_ssa_name (vectype, NULL, "masked_op1"); gassign *select = gimple_build_assign (masked_op1, VEC_COND_EXPR, mask, vop[1], zero); gsi_insert_before (gsi, select, GSI_SAME_STMT); vop[1] = masked_op1; break; } case SAD_EXPR: { tree vectype = TREE_TYPE (vop[1]); tree masked_op1 = make_temp_ssa_name (vectype, NULL, "masked_op1"); gassign *select = gimple_build_assign (masked_op1, VEC_COND_EXPR, mask, vop[1], vop[0]); gsi_insert_before (gsi, select, GSI_SAME_STMT); vop[1] = masked_op1; break; } default: gcc_unreachable (); } } /* Given an operation with CODE in loop reduction path whose reduction PHI is specified by REDUC_INFO, the operation has TYPE of scalar result, and its input vectype is represented by VECTYPE_IN. The vectype of vectorized result may be different from VECTYPE_IN, either in base type or vectype lanes, lane-reducing operation is the case. This function check if it is possible, and how to perform partial vectorization on the operation in the context of LOOP_VINFO. */ static void vect_reduction_update_partial_vector_usage (loop_vec_info loop_vinfo, stmt_vec_info reduc_info, slp_tree slp_node, code_helper code, tree type, tree vectype_in) { enum vect_reduction_type reduc_type = STMT_VINFO_REDUC_TYPE (reduc_info); internal_fn reduc_fn = STMT_VINFO_REDUC_FN (reduc_info); internal_fn cond_fn = get_conditional_internal_fn (code, type); if (reduc_type != FOLD_LEFT_REDUCTION && !use_mask_by_cond_expr_p (code, cond_fn, vectype_in) && (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 operate on partial vectors because" " no conditional operation is available.\n"); LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; } else if (reduc_type == FOLD_LEFT_REDUCTION && reduc_fn == IFN_LAST && !expand_vec_cond_expr_p (vectype_in, truth_type_for (vectype_in))) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "can't operate on partial vectors because" " no conditional operation is available.\n"); LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; } else if (reduc_type == FOLD_LEFT_REDUCTION && internal_fn_mask_index (reduc_fn) == -1 && FLOAT_TYPE_P (vectype_in) && HONOR_SIGN_DEPENDENT_ROUNDING (vectype_in)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "can't operate on partial vectors because" " signed zeros cannot be preserved.\n"); LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; } else { internal_fn mask_reduc_fn = get_masked_reduction_fn (reduc_fn, vectype_in); vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); vec_loop_lens *lens = &LOOP_VINFO_LENS (loop_vinfo); unsigned nvectors = vect_get_num_copies (loop_vinfo, slp_node, vectype_in); if (mask_reduc_fn == IFN_MASK_LEN_FOLD_LEFT_PLUS) vect_record_loop_len (loop_vinfo, lens, nvectors, vectype_in, 1); else vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype_in, NULL); } } /* Check if STMT_INFO is a lane-reducing operation that can be vectorized in the context of LOOP_VINFO, and vector cost will be recorded in COST_VEC, and the analysis is for slp if SLP_NODE is not NULL. For a lane-reducing operation, the loop reduction path that it lies in, may contain normal operation, or other lane-reducing operation of different input type size, an example as: int sum = 0; for (i) { ... sum += d0[i] * d1[i]; // dot-prod <vector(16) char> sum += w[i]; // widen-sum <vector(16) char> sum += abs(s0[i] - s1[i]); // sad <vector(8) short> sum += n[i]; // normal <vector(4) int> ... } Vectorization factor is essentially determined by operation whose input vectype has the most lanes ("vector(16) char" in the example), while we need to choose input vectype with the least lanes ("vector(4) int" in the example) to determine effective number of vector reduction PHIs. */ bool vectorizable_lane_reducing (loop_vec_info loop_vinfo, stmt_vec_info stmt_info, slp_tree slp_node, stmt_vector_for_cost *cost_vec) { gimple *stmt = stmt_info->stmt; if (!lane_reducing_stmt_p (stmt)) return false; tree type = TREE_TYPE (gimple_assign_lhs (stmt)); if (!INTEGRAL_TYPE_P (type)) return false; /* Do not try to vectorize bit-precision reductions. */ if (!type_has_mode_precision_p (type)) return false; /* TODO: Support lane-reducing operation that does not directly participate in loop reduction. */ if (!STMT_VINFO_REDUC_DEF (vect_orig_stmt (stmt_info)) || STMT_VINFO_REDUC_IDX (stmt_info) < 0) return false; stmt_vec_info reduc_info = info_for_reduction (loop_vinfo, stmt_info); /* Lane-reducing pattern inside any inner loop of LOOP_VINFO is not recoginized. */ gcc_assert (!nested_in_vect_loop_p (LOOP_VINFO_LOOP (loop_vinfo), stmt_info)); gcc_assert (STMT_VINFO_REDUC_TYPE (reduc_info) == TREE_CODE_REDUCTION); for (int i = 0; i < (int) gimple_num_ops (stmt) - 1; i++) { slp_tree slp_op; tree op; tree vectype; enum vect_def_type dt; if (!vect_is_simple_use (loop_vinfo, slp_node, i, &op, &slp_op, &dt, &vectype)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "use not simple.\n"); return false; } if (!vectype) { vectype = get_vectype_for_scalar_type (loop_vinfo, TREE_TYPE (op), slp_op); if (!vectype) return false; } if (!vect_maybe_update_slp_op_vectype (slp_op, vectype)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "incompatible vector types for invariants\n"); return false; } if (i == STMT_VINFO_REDUC_IDX (stmt_info)) continue; /* There should be at most one cycle def in the stmt. */ if (VECTORIZABLE_CYCLE_DEF (dt)) return false; } tree vectype_in = SLP_TREE_VECTYPE (SLP_TREE_CHILDREN (slp_node)[0]); gcc_assert (vectype_in); /* Compute number of effective vector statements for costing. */ unsigned int ncopies_for_cost = vect_get_num_copies (loop_vinfo, slp_node, vectype_in); gcc_assert (ncopies_for_cost >= 1); if (vect_is_emulated_mixed_dot_prod (slp_node)) { /* We need extra two invariants: one that contains the minimum signed value and one that contains half of its negative. */ int prologue_stmts = 2; unsigned cost = record_stmt_cost (cost_vec, prologue_stmts, scalar_to_vec, slp_node, 0, vect_prologue); if (dump_enabled_p ()) dump_printf (MSG_NOTE, "vectorizable_lane_reducing: " "extra prologue_cost = %d .\n", cost); /* Three dot-products and a subtraction. */ ncopies_for_cost *= 4; } record_stmt_cost (cost_vec, (int) ncopies_for_cost, vector_stmt, slp_node, 0, vect_body); if (LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)) { enum tree_code code = gimple_assign_rhs_code (stmt); vect_reduction_update_partial_vector_usage (loop_vinfo, reduc_info, slp_node, code, type, vectype_in); } /* Transform via vect_transform_reduction. */ SLP_TREE_TYPE (slp_node) = reduc_vec_info_type; return true; } /* Function vectorizable_reduction. 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 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_INFO may be of this form: X = pattern_expr (arg0, arg1, ..., X) 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++) if (a[i] < value) last = a[i]; This is handled by vectorising the loop and creating an additional vector containing the loop indexes for which "a[i] < value" was true. In the function epilogue this is reduced to a single max value and then used to 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_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. For example, consider an accumulation of shorts into an int accumulator. On some targets it's possible to vectorize this pattern operating on 8 shorts at a time (hence, the vectype for purposes of determining the vectorization factor should be V8HI); on the other hand, the vectype that is used to create the vector form is actually V4SI (the type of the result). Upon entry to this function, STMT_VINFO_VECTYPE records the vectype that indicates what is the actual level of parallelism (V8HI in the example), so that the right vectorization factor would be derived. This vectype corresponds to the type of arguments to the reduction stmt, and should *NOT* be used to create the vectorized stmt. The right vectype for the vectorized stmt is obtained from the type of the result X: get_vectype_for_scalar_type (vinfo, TREE_TYPE (X)) This means that, contrary to "regular" reductions (or "regular" stmts in general), the following equation: STMT_VINFO_VECTYPE == get_vectype_for_scalar_type (vinfo, TREE_TYPE (X)) does *NOT* necessarily hold for reduction patterns. */ bool vectorizable_reduction (loop_vec_info loop_vinfo, stmt_vec_info stmt_info, slp_tree slp_node, slp_instance slp_node_instance, stmt_vector_for_cost *cost_vec) { tree vectype_in = NULL_TREE; class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); enum vect_def_type cond_reduc_dt = vect_unknown_def_type; stmt_vec_info cond_stmt_vinfo = NULL; int i; int ncopies; bool single_defuse_cycle = false; bool nested_cycle = false; bool double_reduc = false; tree cr_index_scalar_type = NULL_TREE, cr_index_vector_type = NULL_TREE; tree cond_reduc_val = NULL_TREE; const bool reduc_chain = SLP_INSTANCE_KIND (slp_node_instance) == slp_inst_kind_reduc_chain; /* Make sure it was already recognized as a reduction computation. */ if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_reduction_def && STMT_VINFO_DEF_TYPE (stmt_info) != vect_double_reduction_def && STMT_VINFO_DEF_TYPE (stmt_info) != vect_nested_cycle) return false; /* The stmt we store reduction analysis meta on. */ stmt_vec_info reduc_info = info_for_reduction (loop_vinfo, stmt_info); reduc_info->is_reduc_info = true; if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_nested_cycle) { if (is_a <gphi *> (stmt_info->stmt)) { /* We eventually need to set a vector type on invariant arguments. */ unsigned j; slp_tree child; FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (slp_node), j, child) if (!vect_maybe_update_slp_op_vectype (child, SLP_TREE_VECTYPE (slp_node))) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "incompatible vector types for " "invariants\n"); return false; } /* Analysis for double-reduction is done on the outer loop PHI, nested cycles have no further restrictions. */ SLP_TREE_TYPE (slp_node) = cycle_phi_info_type; } else SLP_TREE_TYPE (slp_node) = reduc_vec_info_type; return true; } stmt_vec_info phi_info = stmt_info; if (!is_a <gphi *> (stmt_info->stmt)) { SLP_TREE_TYPE (slp_node) = reduc_vec_info_type; return true; } if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_double_reduction_def) { if (gimple_bb (stmt_info->stmt) != loop->header) { /* For SLP we arrive here for both the inner loop LC PHI and the outer loop PHI. The latter is what we want to analyze the reduction with. The LC PHI is handled by vectorizable_lc_phi. */ return gimple_phi_num_args (as_a <gphi *> (stmt_info->stmt)) == 2; } use_operand_p use_p; gimple *use_stmt; bool res = single_imm_use (gimple_phi_result (stmt_info->stmt), &use_p, &use_stmt); gcc_assert (res); phi_info = loop_vinfo->lookup_stmt (use_stmt); } slp_node_instance->reduc_phis = slp_node; /* ??? We're leaving slp_node to point to the PHIs, we only need it to get at the number of vector stmts which wasn't yet initialized for the instance root. */ /* PHIs should not participate in patterns. */ gcc_assert (!STMT_VINFO_RELATED_STMT (phi_info)); gphi *reduc_def_phi = as_a <gphi *> (phi_info->stmt); /* Verify following REDUC_IDX from the latch def leads us back to the PHI and compute the reduction chain length. Discover the real reduction operation stmt on the way (stmt_info and slp_for_stmt_info). */ tree reduc_def = PHI_ARG_DEF_FROM_EDGE (reduc_def_phi, loop_latch_edge (gimple_bb (reduc_def_phi)->loop_father)); unsigned reduc_chain_length = 0; bool only_slp_reduc_chain = true; stmt_info = NULL; slp_tree slp_for_stmt_info = NULL; slp_tree vdef_slp = slp_node_instance->root; /* For double-reductions we start SLP analysis at the inner loop LC PHI which is the def of the outer loop live stmt. */ if (STMT_VINFO_DEF_TYPE (reduc_info) == vect_double_reduction_def) vdef_slp = SLP_TREE_CHILDREN (vdef_slp)[0]; while (reduc_def != PHI_RESULT (reduc_def_phi)) { stmt_vec_info def = loop_vinfo->lookup_def (reduc_def); stmt_vec_info vdef = vect_stmt_to_vectorize (def); int reduc_idx = STMT_VINFO_REDUC_IDX (vdef); if (reduc_idx == -1) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "reduction chain broken by patterns.\n"); return false; } if (!REDUC_GROUP_FIRST_ELEMENT (vdef)) only_slp_reduc_chain = false; /* For epilogue generation live members of the chain need to point back to the PHI via their original stmt for info_for_reduction to work. For SLP we need to look at all lanes here - even though we only will vectorize from the SLP node with live lane zero the other live lanes also need to be identified as part of a reduction to be able to skip code generation for them. */ for (auto s : SLP_TREE_SCALAR_STMTS (vdef_slp)) if (STMT_VINFO_LIVE_P (s)) STMT_VINFO_REDUC_DEF (vect_orig_stmt (s)) = phi_info; gimple_match_op op; if (!gimple_extract_op (vdef->stmt, &op)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "reduction chain includes unsupported" " statement type.\n"); return false; } if (CONVERT_EXPR_CODE_P (op.code)) { if (!tree_nop_conversion_p (op.type, TREE_TYPE (op.ops[0]))) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "conversion in the reduction chain.\n"); return false; } vdef_slp = SLP_TREE_CHILDREN (vdef_slp)[0]; } else { /* First non-conversion stmt. */ if (!stmt_info) { stmt_info = vdef; slp_for_stmt_info = vdef_slp; } if (lane_reducing_op_p (op.code)) { /* The last operand of lane-reducing operation is for reduction. */ gcc_assert (reduc_idx > 0 && reduc_idx == (int) op.num_ops - 1); slp_tree op_node = SLP_TREE_CHILDREN (vdef_slp)[0]; tree vectype_op = SLP_TREE_VECTYPE (op_node); tree type_op = TREE_TYPE (op.ops[0]); if (!vectype_op) { vectype_op = get_vectype_for_scalar_type (loop_vinfo, type_op); if (!vectype_op || !vect_maybe_update_slp_op_vectype (op_node, vectype_op)) return false; } /* For lane-reducing operation vectorizable analysis needs the reduction PHI information. */ STMT_VINFO_REDUC_DEF (def) = phi_info; /* To accommodate lane-reducing operations of mixed input vectypes, choose input vectype with the least lanes for the reduction PHI statement, which would result in the most ncopies for vectorized reduction results. */ if (!vectype_in || (GET_MODE_SIZE (SCALAR_TYPE_MODE (TREE_TYPE (vectype_in))) < GET_MODE_SIZE (SCALAR_TYPE_MODE (type_op)))) vectype_in = vectype_op; } else if (!vectype_in) vectype_in = SLP_TREE_VECTYPE (slp_node); if (!REDUC_GROUP_FIRST_ELEMENT (vdef)) vdef_slp = SLP_TREE_CHILDREN (vdef_slp)[reduc_idx]; } reduc_def = op.ops[reduc_idx]; reduc_chain_length++; } /* PHIs should not participate in patterns. */ gcc_assert (!STMT_VINFO_RELATED_STMT (phi_info)); if (nested_in_vect_loop_p (loop, stmt_info)) { loop = loop->inner; nested_cycle = true; } /* STMT_VINFO_REDUC_DEF doesn't point to the first but the last element. */ if (REDUC_GROUP_FIRST_ELEMENT (stmt_info)) { gcc_assert (!REDUC_GROUP_NEXT_ELEMENT (stmt_info)); stmt_info = REDUC_GROUP_FIRST_ELEMENT (stmt_info); } if (REDUC_GROUP_FIRST_ELEMENT (stmt_info)) gcc_assert (REDUC_GROUP_FIRST_ELEMENT (stmt_info) == stmt_info); /* 1. Is vectorizable reduction? */ /* 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 && !reduc_chain) return false; /* Reductions that are not used even in an enclosing outer-loop, are expected to be "live" (used out of the loop). */ if (STMT_VINFO_RELEVANT (stmt_info) == vect_unused_in_scope && !STMT_VINFO_LIVE_P (stmt_info)) return false; /* 2. Has this been recognized as a reduction pattern? Check if STMT represents a pattern that has been recognized in earlier analysis stages. For stmts that represent a pattern, the STMT_VINFO_RELATED_STMT field records the last stmt in the original sequence that constitutes the pattern. */ 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)); } /* 3. Check the operands of the operation. The first operands are defined inside the loop body. The last operand is the reduction variable, which is defined by the loop-header-phi. */ tree vectype_out = SLP_TREE_VECTYPE (slp_for_stmt_info); STMT_VINFO_REDUC_VECTYPE (reduc_info) = vectype_out; gimple_match_op op; if (!gimple_extract_op (stmt_info->stmt, &op)) gcc_unreachable (); bool lane_reducing = lane_reducing_op_p (op.code); if (!POINTER_TYPE_P (op.type) && !INTEGRAL_TYPE_P (op.type) && !SCALAR_FLOAT_TYPE_P (op.type)) return false; /* Do not try to vectorize bit-precision reductions. */ if (!type_has_mode_precision_p (op.type)) return false; /* Lane-reducing ops also never can be used in a SLP reduction group since we'll mix lanes belonging to different reductions. But it's OK to use them in a reduction chain or when the reduction group has just one element. */ if (lane_reducing && !reduc_chain && SLP_TREE_LANES (slp_node) > 1) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "lane-reducing reduction in reduction group.\n"); return false; } /* All uses but the last are expected to be defined in the loop. 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. */ slp_tree *slp_op = XALLOCAVEC (slp_tree, op.num_ops); tree *vectype_op = XALLOCAVEC (tree, op.num_ops); /* We need to skip an extra operand for COND_EXPRs with embedded comparison. */ unsigned opno_adjust = 0; if (op.code == COND_EXPR && COMPARISON_CLASS_P (op.ops[0])) opno_adjust = 1; for (i = 0; i < (int) op.num_ops; i++) { /* The condition of COND_EXPR is checked in vectorizable_condition(). */ if (i == 0 && op.code == COND_EXPR) continue; stmt_vec_info def_stmt_info; enum vect_def_type dt; if (!vect_is_simple_use (loop_vinfo, slp_for_stmt_info, i + opno_adjust, &op.ops[i], &slp_op[i], &dt, &vectype_op[i], &def_stmt_info)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "use not simple.\n"); return false; } /* Skip reduction operands, and for an IFN_COND_OP we might hit the reduction operand twice (once as definition, once as else). */ if (op.ops[i] == op.ops[STMT_VINFO_REDUC_IDX (stmt_info)]) continue; /* There should be only one cycle def in the stmt, the one leading to reduc_def. */ if (VECTORIZABLE_CYCLE_DEF (dt)) return false; if (!vectype_op[i]) vectype_op[i] = get_vectype_for_scalar_type (loop_vinfo, TREE_TYPE (op.ops[i]), slp_op[i]); /* Record how the non-reduction-def value of COND_EXPR is defined. ??? For a chain of multiple CONDs we'd have to match them up all. */ if (op.code == COND_EXPR && reduc_chain_length == 1) { if (dt == vect_constant_def) { cond_reduc_dt = dt; cond_reduc_val = op.ops[i]; } else 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; } } } enum vect_reduction_type reduction_type = STMT_VINFO_REDUC_TYPE (phi_info); STMT_VINFO_REDUC_TYPE (reduc_info) = reduction_type; /* If we have a condition reduction, see if we can simplify it further. */ if (reduction_type == COND_REDUCTION) { if (SLP_TREE_LANES (slp_node) != 1) return false; /* When the condition uses the reduction value in the condition, fail. */ if (STMT_VINFO_REDUC_IDX (stmt_info) == 0) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "condition depends on previous iteration\n"); return false; } if (reduc_chain_length == 1 && (direct_internal_fn_supported_p (IFN_FOLD_EXTRACT_LAST, vectype_in, OPTIMIZE_FOR_SPEED) || direct_internal_fn_supported_p (IFN_LEN_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_REDUC_TYPE (reduc_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; enum tree_code cond_reduc_op_code = ERROR_MARK; tree res = PHI_RESULT (STMT_VINFO_STMT (cond_stmt_vinfo)); if (!types_compatible_p (TREE_TYPE (res), TREE_TYPE (base))) ; /* 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. */ else 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_REDUC_CODE (reduc_info) = cond_reduc_op_code; STMT_VINFO_VEC_INDUC_COND_INITIAL_VAL (reduc_info) = cond_reduc_val; STMT_VINFO_REDUC_TYPE (reduc_info) = INTEGER_INDUC_COND_REDUCTION; } } else if (cond_reduc_dt == vect_constant_def) { enum vect_def_type cond_initial_dt; tree cond_initial_val = vect_phi_initial_value (reduc_def_phi); 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))) { tree e = fold_binary (LE_EXPR, boolean_type_node, cond_initial_val, cond_reduc_val); if (e && (integer_onep (e) || integer_zerop (e))) { if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "condition expression based on " "compile time constant.\n"); /* Record reduction code at analysis stage. */ STMT_VINFO_REDUC_CODE (reduc_info) = integer_onep (e) ? MAX_EXPR : MIN_EXPR; STMT_VINFO_REDUC_TYPE (reduc_info) = CONST_COND_REDUCTION; } } } } if (STMT_VINFO_LIVE_P (phi_info)) return false; ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); gcc_assert (ncopies >= 1); poly_uint64 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); if (nested_cycle) { gcc_assert (STMT_VINFO_DEF_TYPE (reduc_info) == vect_double_reduction_def); double_reduc = true; } /* 4.2. Check support for the epilog operation. If STMT represents a reduction pattern, then the type of the reduction variable may be different than the type of the rest of the arguments. For example, consider the case of accumulation of shorts into an int accumulator; The original code: S1: int_a = (int) short_a; orig_stmt-> S2: int_acc = plus <int_a ,int_acc>; was replaced with: STMT: int_acc = widen_sum <short_a, int_acc> This means that: 1. The tree-code that is used to create the vector operation in the epilog code (that reduces the partial results) is not the tree-code of STMT, but is rather the tree-code of the original stmt from the pattern that STMT is replacing. I.e, in the example above we want to use 'widen_sum' in the loop, but 'plus' in the epilog. 2. The type (mode) we use to check available target support for the vector operation to be created in the *epilog*, is determined by the type of the reduction variable (in the example above we'd check this: optab_handler (plus_optab, vect_int_mode])). However the type (mode) we use to check available target support for the vector operation to be created *inside the loop*, is determined by the type of the other arguments to STMT (in the example we'd check this: optab_handler (widen_sum_optab, vect_short_mode)). This is contrary to "regular" reductions, in which the types of all the arguments are the same as the type of the reduction variable. For "regular" reductions we can therefore use the same vector type (and also the same tree-code) when generating the epilog code and when generating the code inside the loop. */ code_helper orig_code = STMT_VINFO_REDUC_CODE (phi_info); /* If conversion might have created a conditional operation like IFN_COND_ADD already. Use the internal code for the following checks. */ if (orig_code.is_internal_fn ()) { tree_code new_code = conditional_internal_fn_code (internal_fn (orig_code)); orig_code = new_code != ERROR_MARK ? new_code : orig_code; } STMT_VINFO_REDUC_CODE (reduc_info) = orig_code; reduction_type = STMT_VINFO_REDUC_TYPE (reduc_info); if (reduction_type == TREE_CODE_REDUCTION) { /* 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. Likewise when we are vectorizing a series of reductions using SLP and the VF is one the reductions are performed in scalar order. */ if (!reduc_chain && known_eq (LOOP_VINFO_VECT_FACTOR (loop_vinfo), 1u)) ; else if (needs_fold_left_reduction_p (op.type, orig_code)) { /* When vectorizing a reduction chain w/o SLP the reduction PHI is not directy used in stmt. */ if (!only_slp_reduc_chain && reduc_chain_length != 1) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "in-order reduction chain without SLP.\n"); return false; } /* Code generation doesn't support function calls other than .COND_*. */ if (!op.code.is_tree_code () && !(op.code.is_internal_fn () && conditional_internal_fn_code (internal_fn (op.code)) != ERROR_MARK)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "in-order reduction chain operation not " "supported.\n"); return false; } STMT_VINFO_REDUC_TYPE (reduc_info) = reduction_type = FOLD_LEFT_REDUCTION; } else if (!commutative_binary_op_p (orig_code, op.type) || !associative_binary_op_p (orig_code, op.type)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "reduction: not commutative/associative\n"); return false; } } if ((reduction_type == COND_REDUCTION || reduction_type == INTEGER_INDUC_COND_REDUCTION || reduction_type == CONST_COND_REDUCTION || reduction_type == EXTRACT_LAST_REDUCTION) && 1 && ncopies > 1) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "multiple types in condition reduction.\n"); return false; } internal_fn 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"); reduc_fn = IFN_LAST; } } else { if (!nested_cycle || double_reduc) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "no reduc code for scalar code.\n"); return false; } } } else if (reduction_type == COND_REDUCTION) { int scalar_precision = GET_MODE_PRECISION (SCALAR_TYPE_MODE (op.type)); cr_index_scalar_type = make_unsigned_type (scalar_precision); cr_index_vector_type = get_same_sized_vectype (cr_index_scalar_type, vectype_out); if (direct_internal_fn_supported_p (IFN_REDUC_MAX, cr_index_vector_type, OPTIMIZE_FOR_SPEED)) reduc_fn = IFN_REDUC_MAX; } STMT_VINFO_REDUC_FN (reduc_info) = reduc_fn; 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; } /* For SLP reductions, see if there is a neutral value we can use. */ tree neutral_op = NULL_TREE; tree initial_value = NULL_TREE; if (reduc_chain) initial_value = vect_phi_initial_value (reduc_def_phi); neutral_op = neutral_op_for_reduction (TREE_TYPE (vectype_out), orig_code, initial_value); 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_TREE_LANES (slp_node) > 1 && !reduc_chain) { /* 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 (!reduc_chain && !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_LANES (slp_node); if (!neutral_op && !can_duplicate_and_interleave_p (loop_vinfo, group_size, TREE_TYPE (vectype_out))) { 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 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; } } if (reduction_type == COND_REDUCTION) { widest_int ni; if (! max_loop_iterations (loop, &ni)) { if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "loop count not known, cannot create cond " "reduction.\n"); return false; } /* Convert backedges to iterations. */ ni += 1; /* The additional index will be the same type as the condition. Check that the loop can fit into this less one (because we'll use up the zero slot for when there are no matches). */ tree max_index = TYPE_MAX_VALUE (cr_index_scalar_type); if (wi::geu_p (ni, wi::to_widest (max_index))) { if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "loop size is greater than data size.\n"); return false; } } /* In case the vectorization factor (VF) is bigger than the number of elements that we can fit in a vectype (nunits), 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. */ /* If the reduction is used in an outer loop we need to generate VF intermediate results, like so (e.g. for ncopies=2): r0 = phi (init, r0) r1 = phi (init, r1) r0 = x0 + r0; r1 = x1 + r1; (i.e. we generate VF results in 2 registers). In this case we have a separate def-use cycle for each copy, and therefore for each copy we get the vector def for the reduction variable from the respective phi node created for this copy. Otherwise (the reduction is unused in the loop nest), we can combine together intermediate results, like so (e.g. for ncopies=2): r = phi (init, r) r = x0 + r; r = x1 + r; (i.e. we generate VF/2 results in a single register). In this case for each copy we get the vector def for the reduction variable from the vectorized reduction operation generated in the previous iteration. This only works when we see both the reduction PHI and its only consumer in vectorizable_reduction and there are no intermediate stmts participating. When unrolling we want each unrolled iteration to have its own reduction accumulator since one of the main goals of unrolling a reduction is to reduce the aggregate loop-carried latency. */ if (ncopies > 1 && !reduc_chain && SLP_TREE_LANES (slp_node) == 1 && (STMT_VINFO_RELEVANT (stmt_info) <= vect_used_only_live) && reduc_chain_length == 1 && loop_vinfo->suggested_unroll_factor == 1) single_defuse_cycle = true; if (single_defuse_cycle && !lane_reducing) { gcc_assert (op.code != COND_EXPR); /* 4. check support for the operation in the loop This isn't necessary for the lane reduction codes, since they can only be produced by pattern matching, and it's up to the pattern matcher to test for support. The main reason for specifically skipping this step is to avoid rechecking whether mixed-sign dot-products can be implemented using signed dot-products. */ machine_mode vec_mode = TYPE_MODE (vectype_in); if (!directly_supported_p (op.code, vectype_in, optab_vector)) { if (dump_enabled_p ()) dump_printf (MSG_NOTE, "op not supported by target.\n"); if (maybe_ne (GET_MODE_SIZE (vec_mode), UNITS_PER_WORD) || !vect_can_vectorize_without_simd_p (op.code)) single_defuse_cycle = false; else if (dump_enabled_p ()) dump_printf (MSG_NOTE, "proceeding using word mode.\n"); } if (vect_emulated_vector_p (vectype_in) && !vect_can_vectorize_without_simd_p (op.code)) { if (dump_enabled_p ()) dump_printf (MSG_NOTE, "using word mode not possible.\n"); return false; } } if (dump_enabled_p () && single_defuse_cycle) dump_printf_loc (MSG_NOTE, vect_location, "using single def-use cycle for reduction by reducing " "multiple vectors to one in the loop body\n"); STMT_VINFO_FORCE_SINGLE_CYCLE (reduc_info) = single_defuse_cycle; /* For lane-reducing operation, the below processing related to single defuse-cycle will be done in its own vectorizable function. One more thing to note is that the operation must not be involved in fold-left reduction. */ single_defuse_cycle &= !lane_reducing; if (single_defuse_cycle || reduction_type == FOLD_LEFT_REDUCTION) for (i = 0; i < (int) op.num_ops; i++) if (!vect_maybe_update_slp_op_vectype (slp_op[i], vectype_op[i])) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "incompatible vector types for invariants\n"); return false; } vect_model_reduction_cost (loop_vinfo, slp_for_stmt_info, reduc_fn, reduction_type, ncopies, cost_vec); /* Cost the reduction op inside the loop if transformed via vect_transform_reduction for non-lane-reducing operation. Otherwise this is costed by the separate vectorizable_* routines. */ if (single_defuse_cycle) record_stmt_cost (cost_vec, ncopies, vector_stmt, slp_for_stmt_info, 0, vect_body); if (dump_enabled_p () && reduction_type == FOLD_LEFT_REDUCTION) dump_printf_loc (MSG_NOTE, vect_location, "using an in-order (fold-left) reduction.\n"); SLP_TREE_TYPE (slp_node) = cycle_phi_info_type; /* All but single defuse-cycle optimized and fold-left reductions go through their own vectorizable_* routines. */ if (!single_defuse_cycle && reduction_type != FOLD_LEFT_REDUCTION) { stmt_vec_info tem = vect_stmt_to_vectorize (STMT_VINFO_REDUC_DEF (phi_info)); if (REDUC_GROUP_FIRST_ELEMENT (tem)) { gcc_assert (!REDUC_GROUP_NEXT_ELEMENT (tem)); tem = REDUC_GROUP_FIRST_ELEMENT (tem); } STMT_VINFO_DEF_TYPE (vect_orig_stmt (tem)) = vect_internal_def; STMT_VINFO_DEF_TYPE (tem) = vect_internal_def; } else if (LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)) vect_reduction_update_partial_vector_usage (loop_vinfo, reduc_info, slp_node, op.code, op.type, vectype_in); return true; } /* STMT_INFO is a dot-product reduction whose multiplication operands have different signs. Emit a sequence to emulate the operation using a series of signed DOT_PROD_EXPRs and return the last statement generated. VEC_DEST is the result of the vector operation and VOP lists its inputs. */ static gassign * vect_emulate_mixed_dot_prod (loop_vec_info loop_vinfo, stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, tree vec_dest, tree vop[3]) { tree wide_vectype = signed_type_for (TREE_TYPE (vec_dest)); tree narrow_vectype = signed_type_for (TREE_TYPE (vop[0])); tree narrow_elttype = TREE_TYPE (narrow_vectype); gimple *new_stmt; /* Make VOP[0] the unsigned operand VOP[1] the signed operand. */ if (!TYPE_UNSIGNED (TREE_TYPE (vop[0]))) std::swap (vop[0], vop[1]); /* Convert all inputs to signed types. */ for (int i = 0; i < 3; ++i) if (TYPE_UNSIGNED (TREE_TYPE (vop[i]))) { tree tmp = make_ssa_name (signed_type_for (TREE_TYPE (vop[i]))); new_stmt = gimple_build_assign (tmp, NOP_EXPR, vop[i]); vect_finish_stmt_generation (loop_vinfo, stmt_info, new_stmt, gsi); vop[i] = tmp; } /* In the comments below we assume 8-bit inputs for simplicity, but the approach works for any full integer type. */ /* Create a vector of -128. */ tree min_narrow_elttype = TYPE_MIN_VALUE (narrow_elttype); tree min_narrow = build_vector_from_val (narrow_vectype, min_narrow_elttype); /* Create a vector of 64. */ auto half_wi = wi::lrshift (wi::to_wide (min_narrow_elttype), 1); tree half_narrow = wide_int_to_tree (narrow_elttype, half_wi); half_narrow = build_vector_from_val (narrow_vectype, half_narrow); /* Emit: SUB_RES = VOP[0] - 128. */ tree sub_res = make_ssa_name (narrow_vectype); new_stmt = gimple_build_assign (sub_res, PLUS_EXPR, vop[0], min_narrow); vect_finish_stmt_generation (loop_vinfo, stmt_info, new_stmt, gsi); /* Emit: STAGE1 = DOT_PROD_EXPR <VOP[1], 64, VOP[2]>; STAGE2 = DOT_PROD_EXPR <VOP[1], 64, STAGE1>; STAGE3 = DOT_PROD_EXPR <SUB_RES, -128, STAGE2>; on the basis that x * y == (x - 128) * y + 64 * y + 64 * y Doing the two 64 * y steps first allows more time to compute x. */ tree stage1 = make_ssa_name (wide_vectype); new_stmt = gimple_build_assign (stage1, DOT_PROD_EXPR, vop[1], half_narrow, vop[2]); vect_finish_stmt_generation (loop_vinfo, stmt_info, new_stmt, gsi); tree stage2 = make_ssa_name (wide_vectype); new_stmt = gimple_build_assign (stage2, DOT_PROD_EXPR, vop[1], half_narrow, stage1); vect_finish_stmt_generation (loop_vinfo, stmt_info, new_stmt, gsi); tree stage3 = make_ssa_name (wide_vectype); new_stmt = gimple_build_assign (stage3, DOT_PROD_EXPR, sub_res, vop[1], stage2); vect_finish_stmt_generation (loop_vinfo, stmt_info, new_stmt, gsi); /* Convert STAGE3 to the reduction type. */ return gimple_build_assign (vec_dest, CONVERT_EXPR, stage3); } /* Transform the definition stmt STMT_INFO of a reduction PHI backedge value. */ bool vect_transform_reduction (loop_vec_info loop_vinfo, stmt_vec_info stmt_info, gimple_stmt_iterator *gsi, slp_tree slp_node) { tree vectype_out = SLP_TREE_VECTYPE (slp_node); class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); unsigned vec_num; stmt_vec_info reduc_info = info_for_reduction (loop_vinfo, stmt_info); gcc_assert (reduc_info->is_reduc_info); if (nested_in_vect_loop_p (loop, stmt_info)) { loop = loop->inner; gcc_assert (STMT_VINFO_DEF_TYPE (reduc_info) == vect_double_reduction_def); } gimple_match_op op; if (!gimple_extract_op (stmt_info->stmt, &op)) gcc_unreachable (); /* All uses but the last are expected to be defined in the loop. 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. */ int reduc_index = STMT_VINFO_REDUC_IDX (stmt_info); tree vectype_in = SLP_TREE_VECTYPE (SLP_TREE_CHILDREN (slp_node)[0]); vec_num = vect_get_num_copies (loop_vinfo, slp_node, vectype_in); code_helper code = canonicalize_code (op.code, op.type); internal_fn cond_fn = get_conditional_internal_fn (code, op.type); vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); vec_loop_lens *lens = &LOOP_VINFO_LENS (loop_vinfo); bool mask_by_cond_expr = use_mask_by_cond_expr_p (code, cond_fn, vectype_in); /* Transform. */ tree new_temp = NULL_TREE; auto_vec<tree> vec_oprnds[3]; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "transform reduction.\n"); /* A binary COND_OP reduction must have the same definition and else value. */ bool cond_fn_p = code.is_internal_fn () && conditional_internal_fn_code (internal_fn (code)) != ERROR_MARK; if (cond_fn_p) { gcc_assert (code == IFN_COND_ADD || code == IFN_COND_SUB || code == IFN_COND_MUL || code == IFN_COND_AND || code == IFN_COND_IOR || code == IFN_COND_XOR || code == IFN_COND_MIN || code == IFN_COND_MAX); gcc_assert (op.num_ops == 4 && (op.ops[reduc_index] == op.ops[internal_fn_else_index ((internal_fn) code)])); } bool masked_loop_p = LOOP_VINFO_FULLY_MASKED_P (loop_vinfo); vect_reduction_type reduction_type = STMT_VINFO_REDUC_TYPE (reduc_info); if (reduction_type == FOLD_LEFT_REDUCTION) { internal_fn reduc_fn = STMT_VINFO_REDUC_FN (reduc_info); gcc_assert (code.is_tree_code () || cond_fn_p); return vectorize_fold_left_reduction (loop_vinfo, stmt_info, gsi, slp_node, code, reduc_fn, op.num_ops, vectype_in, reduc_index, masks, lens); } bool single_defuse_cycle = STMT_VINFO_FORCE_SINGLE_CYCLE (reduc_info); bool lane_reducing = lane_reducing_op_p (code); gcc_assert (single_defuse_cycle || lane_reducing); if (lane_reducing) { /* The last operand of lane-reducing op is for reduction. */ gcc_assert (reduc_index == (int) op.num_ops - 1); } /* Create the destination vector */ tree scalar_dest = gimple_get_lhs (stmt_info->stmt); tree vec_dest = vect_create_destination_var (scalar_dest, vectype_out); /* Get NCOPIES vector definitions for all operands except the reduction definition. */ if (!cond_fn_p) { gcc_assert (reduc_index >= 0 && reduc_index <= 2); vect_get_vec_defs (loop_vinfo, slp_node, single_defuse_cycle && reduc_index == 0 ? NULL_TREE : op.ops[0], &vec_oprnds[0], single_defuse_cycle && reduc_index == 1 ? NULL_TREE : op.ops[1], &vec_oprnds[1], op.num_ops == 3 && !(single_defuse_cycle && reduc_index == 2) ? op.ops[2] : NULL_TREE, &vec_oprnds[2]); } else { /* For a conditional operation pass the truth type as mask vectype. */ gcc_assert (single_defuse_cycle && (reduc_index == 1 || reduc_index == 2)); vect_get_vec_defs (loop_vinfo, slp_node, op.ops[0], &vec_oprnds[0], reduc_index == 1 ? NULL_TREE : op.ops[1], &vec_oprnds[1], reduc_index == 2 ? NULL_TREE : op.ops[2], &vec_oprnds[2]); } /* For single def-use cycles get one copy of the vectorized reduction definition. */ if (single_defuse_cycle) { vect_get_vec_defs (loop_vinfo, slp_node, reduc_index == 0 ? op.ops[0] : NULL_TREE, &vec_oprnds[0], reduc_index == 1 ? op.ops[1] : NULL_TREE, &vec_oprnds[1], reduc_index == 2 ? op.ops[2] : NULL_TREE, &vec_oprnds[2]); } else if (lane_reducing) { /* For normal reduction, consistency between vectorized def/use is naturally ensured when mapping from scalar statement. But if lane- reducing op is involved in reduction, thing would become somewhat complicated in that the op's result and operand for accumulation are limited to less lanes than other operands, which certainly causes def/use mismatch on adjacent statements around the op if do not have any kind of specific adjustment. One approach is to refit lane- reducing op in the way of introducing new trivial pass-through copies to fix possible def/use gap, so as to make it behave like a normal op. And vector reduction PHIs are always generated to the full extent, no matter lane-reducing op exists or not. If some copies or PHIs are actually superfluous, they would be cleaned up by passes after vectorization. An example for single-lane slp, lane-reducing ops with mixed input vectypes in a reduction chain, is given as below. Similarly, this handling is applicable for multiple-lane slp as well. int sum = 1; for (i) { sum += d0[i] * d1[i]; // dot-prod <vector(16) char> sum += w[i]; // widen-sum <vector(16) char> sum += abs(s0[i] - s1[i]); // sad <vector(8) short> sum += n[i]; // normal <vector(4) int> } The vector size is 128-bit，vectorization factor is 16. Reduction statements would be transformed as: vector<4> int sum_v0 = { 0, 0, 0, 1 }; vector<4> int sum_v1 = { 0, 0, 0, 0 }; vector<4> int sum_v2 = { 0, 0, 0, 0 }; vector<4> int sum_v3 = { 0, 0, 0, 0 }; for (i / 16) { sum_v0 = DOT_PROD (d0_v0[i: 0 ~ 15], d1_v0[i: 0 ~ 15], sum_v0); sum_v1 = sum_v1; // copy sum_v2 = sum_v2; // copy sum_v3 = sum_v3; // copy sum_v0 = sum_v0; // copy sum_v1 = WIDEN_SUM (w_v1[i: 0 ~ 15], sum_v1); sum_v2 = sum_v2; // copy sum_v3 = sum_v3; // copy sum_v0 = sum_v0; // copy sum_v1 = SAD (s0_v1[i: 0 ~ 7 ], s1_v1[i: 0 ~ 7 ], sum_v1); sum_v2 = SAD (s0_v2[i: 8 ~ 15], s1_v2[i: 8 ~ 15], sum_v2); sum_v3 = sum_v3; // copy sum_v0 += n_v0[i: 0 ~ 3 ]; sum_v1 += n_v1[i: 4 ~ 7 ]; sum_v2 += n_v2[i: 8 ~ 11]; sum_v3 += n_v3[i: 12 ~ 15]; } Moreover, for a higher instruction parallelism in final vectorized loop, it is considered to make those effective vector lane-reducing ops be distributed evenly among all def-use cycles. In the above example, DOT_PROD, WIDEN_SUM and SADs are generated into disparate cycles, instruction dependency among them could be eliminated. */ unsigned effec_ncopies = vec_oprnds[0].length (); unsigned total_ncopies = vec_oprnds[reduc_index].length (); gcc_assert (effec_ncopies <= total_ncopies); if (effec_ncopies < total_ncopies) { for (unsigned i = 0; i < op.num_ops - 1; i++) { gcc_assert (vec_oprnds[i].length () == effec_ncopies); vec_oprnds[i].safe_grow_cleared (total_ncopies); } } tree reduc_vectype_in = vectype_in; gcc_assert (reduc_vectype_in); unsigned effec_reduc_ncopies = vect_get_num_copies (loop_vinfo, slp_node, reduc_vectype_in); gcc_assert (effec_ncopies <= effec_reduc_ncopies); if (effec_ncopies < effec_reduc_ncopies) { /* Find suitable def-use cycles to generate vectorized statements into, and reorder operands based on the selection. */ unsigned curr_pos = reduc_info->reduc_result_pos; unsigned next_pos = (curr_pos + effec_ncopies) % effec_reduc_ncopies; gcc_assert (curr_pos < effec_reduc_ncopies); reduc_info->reduc_result_pos = next_pos; if (curr_pos) { unsigned count = effec_reduc_ncopies - effec_ncopies; unsigned start = curr_pos - count; if ((int) start < 0) { count = curr_pos; start = 0; } for (unsigned i = 0; i < op.num_ops - 1; i++) { for (unsigned j = effec_ncopies; j > start; j--) { unsigned k = j - 1; std::swap (vec_oprnds[i][k], vec_oprnds[i][k + count]); gcc_assert (!vec_oprnds[i][k]); } } } } } bool emulated_mixed_dot_prod = vect_is_emulated_mixed_dot_prod (slp_node); unsigned num = vec_oprnds[reduc_index == 0 ? 1 : 0].length (); unsigned mask_index = 0; for (unsigned i = 0; i < num; ++i) { gimple *new_stmt; tree vop[3] = { vec_oprnds[0][i], vec_oprnds[1][i], NULL_TREE }; if (!vop[0] || !vop[1]) { tree reduc_vop = vec_oprnds[reduc_index][i]; /* If could not generate an effective vector statement for current portion of reduction operand, insert a trivial copy to simply handle over the operand to other dependent statements. */ gcc_assert (reduc_vop); if (TREE_CODE (reduc_vop) == SSA_NAME && !SSA_NAME_IS_DEFAULT_DEF (reduc_vop)) new_stmt = SSA_NAME_DEF_STMT (reduc_vop); else { new_temp = make_ssa_name (vec_dest); new_stmt = gimple_build_assign (new_temp, reduc_vop); vect_finish_stmt_generation (loop_vinfo, stmt_info, new_stmt, gsi); } } else if (masked_loop_p && !mask_by_cond_expr) { /* No conditional ifns have been defined for lane-reducing op yet. */ gcc_assert (!lane_reducing); /* Make sure that the reduction accumulator is vop[0]. */ if (reduc_index == 1) { gcc_assert (commutative_binary_op_p (code, op.type)); std::swap (vop[0], vop[1]); } tree mask = vect_get_loop_mask (loop_vinfo, gsi, masks, vec_num, vectype_in, mask_index++); 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); vect_finish_stmt_generation (loop_vinfo, stmt_info, call, gsi); new_stmt = call; } else { if (op.num_ops >= 3) vop[2] = vec_oprnds[2][i]; if (masked_loop_p && mask_by_cond_expr) { tree mask = vect_get_loop_mask (loop_vinfo, gsi, masks, vec_num, vectype_in, mask_index++); build_vect_cond_expr (code, vop, mask, gsi); } if (emulated_mixed_dot_prod) new_stmt = vect_emulate_mixed_dot_prod (loop_vinfo, stmt_info, gsi, vec_dest, vop); else if (code.is_internal_fn () && !cond_fn_p) new_stmt = gimple_build_call_internal (internal_fn (code), op.num_ops, vop[0], vop[1], vop[2]); else if (code.is_internal_fn () && cond_fn_p) new_stmt = gimple_build_call_internal (internal_fn (code), op.num_ops, vop[0], vop[1], vop[2], vop[reduc_index]); else new_stmt = gimple_build_assign (vec_dest, tree_code (op.code), vop[0], vop[1], vop[2]); new_temp = make_ssa_name (vec_dest, new_stmt); gimple_set_lhs (new_stmt, new_temp); vect_finish_stmt_generation (loop_vinfo, stmt_info, new_stmt, gsi); } if (single_defuse_cycle && i < num - 1) vec_oprnds[reduc_index].safe_push (gimple_get_lhs (new_stmt)); else slp_node->push_vec_def (new_stmt); } return true; } /* Transform phase of a cycle PHI. */ bool vect_transform_cycle_phi (loop_vec_info loop_vinfo, stmt_vec_info stmt_info, slp_tree slp_node, slp_instance slp_node_instance) { tree vectype_out = SLP_TREE_VECTYPE (slp_node); class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); int i; bool nested_cycle = false; int vec_num; const bool reduc_chain = SLP_INSTANCE_KIND (slp_node_instance) == slp_inst_kind_reduc_chain; if (nested_in_vect_loop_p (loop, stmt_info)) { loop = loop->inner; nested_cycle = true; } stmt_vec_info reduc_info = info_for_reduction (loop_vinfo, stmt_info); gcc_assert (reduc_info->is_reduc_info); if (STMT_VINFO_REDUC_TYPE (reduc_info) == EXTRACT_LAST_REDUCTION || STMT_VINFO_REDUC_TYPE (reduc_info) == FOLD_LEFT_REDUCTION) /* Leave the scalar phi in place. */ return true; vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); /* Check whether we should use a single PHI node and accumulate vectors to one before the backedge. */ if (STMT_VINFO_FORCE_SINGLE_CYCLE (reduc_info)) vec_num = 1; /* Create the destination vector */ gphi *phi = as_a <gphi *> (stmt_info->stmt); tree vec_dest = vect_create_destination_var (gimple_phi_result (phi), vectype_out); /* Get the loop-entry arguments. */ tree vec_initial_def = NULL_TREE; auto_vec<tree> vec_initial_defs; vec_initial_defs.reserve (vec_num); /* 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 (STMT_VINFO_REDUC_TYPE (reduc_info) == INTEGER_INDUC_COND_REDUCTION) { gcc_assert (SLP_TREE_LANES (slp_node) == 1); tree initial_def = vect_phi_initial_value (phi); reduc_info->reduc_initial_values.safe_push (initial_def); tree induc_val = STMT_VINFO_VEC_INDUC_COND_INITIAL_VAL (reduc_info); if (TREE_CODE (initial_def) == INTEGER_CST && !integer_zerop (induc_val) && ((STMT_VINFO_REDUC_CODE (reduc_info) == MAX_EXPR && tree_int_cst_lt (initial_def, induc_val)) || (STMT_VINFO_REDUC_CODE (reduc_info) == MIN_EXPR && tree_int_cst_lt (induc_val, initial_def)))) { induc_val = initial_def; /* Communicate we used the initial_def to epilouge generation. */ STMT_VINFO_VEC_INDUC_COND_INITIAL_VAL (reduc_info) = NULL_TREE; } vec_initial_defs.quick_push (build_vector_from_val (vectype_out, induc_val)); } else if (nested_cycle) { unsigned phi_idx = loop_preheader_edge (loop)->dest_idx; vect_get_slp_defs (SLP_TREE_CHILDREN (slp_node)[phi_idx], &vec_initial_defs); } else { gcc_assert (slp_node == slp_node_instance->reduc_phis); vec<tree> &initial_values = reduc_info->reduc_initial_values; vec<stmt_vec_info> &stmts = SLP_TREE_SCALAR_STMTS (slp_node); unsigned int num_phis = stmts.length (); if (reduc_chain) num_phis = 1; initial_values.reserve (num_phis); for (unsigned int i = 0; i < num_phis; ++i) { gphi *this_phi = as_a<gphi *> (stmts[i]->stmt); initial_values.quick_push (vect_phi_initial_value (this_phi)); } if (vec_num == 1) vect_find_reusable_accumulator (loop_vinfo, reduc_info, vectype_out); if (!initial_values.is_empty ()) { tree initial_value = (num_phis == 1 ? initial_values[0] : NULL_TREE); code_helper code = STMT_VINFO_REDUC_CODE (reduc_info); tree neutral_op = neutral_op_for_reduction (TREE_TYPE (vectype_out), code, initial_value); /* Try to simplify the vector initialization by applying an adjustment after the reduction has been performed. This can also break a critical path but on the other hand requires to keep the initial value live across the loop. */ if (neutral_op && initial_values.length () == 1 && !reduc_info->reused_accumulator && STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def && !operand_equal_p (neutral_op, initial_values[0])) { STMT_VINFO_REDUC_EPILOGUE_ADJUSTMENT (reduc_info) = initial_values[0]; initial_values[0] = neutral_op; } get_initial_defs_for_reduction (loop_vinfo, reduc_info, vectype_out, &vec_initial_defs, vec_num, stmts.length (), neutral_op); } } if (vec_initial_def) { vec_initial_defs.create (1); vec_initial_defs.quick_push (vec_initial_def); } if (auto *accumulator = reduc_info->reused_accumulator) { tree def = accumulator->reduc_input; if (!useless_type_conversion_p (vectype_out, TREE_TYPE (def))) { unsigned int nreduc; bool res = constant_multiple_p (TYPE_VECTOR_SUBPARTS (TREE_TYPE (def)), TYPE_VECTOR_SUBPARTS (vectype_out), &nreduc); gcc_assert (res); gimple_seq stmts = NULL; /* Reduce the single vector to a smaller one. */ if (nreduc != 1) { /* Perform the reduction in the appropriate type. */ tree rvectype = vectype_out; if (!useless_type_conversion_p (TREE_TYPE (vectype_out), TREE_TYPE (TREE_TYPE (def)))) rvectype = build_vector_type (TREE_TYPE (TREE_TYPE (def)), TYPE_VECTOR_SUBPARTS (vectype_out)); def = vect_create_partial_epilog (def, rvectype, STMT_VINFO_REDUC_CODE (reduc_info), &stmts); } /* The epilogue loop might use a different vector mode, like VNx2DI vs. V2DI. */ if (TYPE_MODE (vectype_out) != TYPE_MODE (TREE_TYPE (def))) { tree reduc_type = build_vector_type_for_mode (TREE_TYPE (TREE_TYPE (def)), TYPE_MODE (vectype_out)); def = gimple_convert (&stmts, reduc_type, def); } /* Adjust the input so we pick up the partially reduced value for the skip edge in vect_create_epilog_for_reduction. */ accumulator->reduc_input = def; /* And the reduction could be carried out using a different sign. */ if (!useless_type_conversion_p (vectype_out, TREE_TYPE (def))) def = gimple_convert (&stmts, vectype_out, def); edge e; if ((e = loop_vinfo->main_loop_edge) || (e = loop_vinfo->skip_this_loop_edge)) { /* While we'd like to insert on the edge this will split blocks and disturb bookkeeping, we also will eventually need this on the skip edge. Rely on sinking to fixup optimal placement and insert in the pred. */ gimple_stmt_iterator gsi = gsi_last_bb (e->src); /* Insert before a cond that eventually skips the epilogue. */ if (!gsi_end_p (gsi) && stmt_ends_bb_p (gsi_stmt (gsi))) gsi_prev (&gsi); gsi_insert_seq_after (&gsi, stmts, GSI_CONTINUE_LINKING); } else gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); } if (loop_vinfo->main_loop_edge) vec_initial_defs[0] = vect_get_main_loop_result (loop_vinfo, def, vec_initial_defs[0]); else vec_initial_defs.safe_push (def); } /* Generate the reduction PHIs upfront. */ for (i = 0; i < vec_num; i++) { tree vec_init_def = vec_initial_defs[i]; /* Create the reduction-phi that defines the reduction operand. */ gphi *new_phi = create_phi_node (vec_dest, loop->header); add_phi_arg (new_phi, vec_init_def, loop_preheader_edge (loop), UNKNOWN_LOCATION); /* The loop-latch arg is set in epilogue processing. */ slp_node->push_vec_def (new_phi); } return true; } /* Vectorizes LC PHIs. */ bool vectorizable_lc_phi (loop_vec_info loop_vinfo, stmt_vec_info stmt_info, slp_tree slp_node) { if (!loop_vinfo || !is_a <gphi *> (stmt_info->stmt) || gimple_phi_num_args (stmt_info->stmt) != 1) return false; if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def && STMT_VINFO_DEF_TYPE (stmt_info) != vect_double_reduction_def) return false; /* Deal with copies from externs or constants that disguise as loop-closed PHI nodes (PR97886). */ if (!vect_maybe_update_slp_op_vectype (SLP_TREE_CHILDREN (slp_node)[0], SLP_TREE_VECTYPE (slp_node))) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "incompatible vector types for invariants\n"); return false; } /* ??? This can happen with data vs. mask uses of boolean. */ if (!useless_type_conversion_p (SLP_TREE_VECTYPE (slp_node), SLP_TREE_VECTYPE (SLP_TREE_CHILDREN (slp_node)[0]))) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "missed mask promotion\n"); return false; } SLP_TREE_TYPE (slp_node) = lc_phi_info_type; return true; } bool vect_transform_lc_phi (loop_vec_info loop_vinfo, stmt_vec_info stmt_info, slp_tree slp_node) { tree vectype = SLP_TREE_VECTYPE (slp_node); tree scalar_dest = gimple_phi_result (stmt_info->stmt); basic_block bb = gimple_bb (stmt_info->stmt); edge e = single_pred_edge (bb); tree vec_dest = vect_create_destination_var (scalar_dest, vectype); auto_vec<tree> vec_oprnds; vect_get_vec_defs (loop_vinfo, slp_node, gimple_phi_arg_def (stmt_info->stmt, 0), &vec_oprnds); for (unsigned i = 0; i < vec_oprnds.length (); i++) { /* Create the vectorized LC PHI node. */ gphi *new_phi = create_phi_node (vec_dest, bb); add_phi_arg (new_phi, vec_oprnds[i], e, UNKNOWN_LOCATION); slp_node->push_vec_def (new_phi); } return true; } /* Vectorizes PHIs. */ bool vectorizable_phi (bb_vec_info, stmt_vec_info stmt_info, slp_tree slp_node, stmt_vector_for_cost *cost_vec) { if (!is_a <gphi *> (stmt_info->stmt) || !slp_node) return false; if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def) return false; tree vectype = SLP_TREE_VECTYPE (slp_node); if (cost_vec) /* transformation not required. */ { slp_tree child; unsigned i; FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (slp_node), i, child) if (!child) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "PHI node with unvectorized backedge def\n"); return false; } else if (!vect_maybe_update_slp_op_vectype (child, vectype)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "incompatible vector types for invariants\n"); return false; } else if (SLP_TREE_DEF_TYPE (child) == vect_internal_def && !useless_type_conversion_p (vectype, SLP_TREE_VECTYPE (child))) { /* With bools we can have mask and non-mask precision vectors or different non-mask precisions. while pattern recog is supposed to guarantee consistency here bugs in it can cause mismatches (PR103489 and PR103800 for example). Deal with them here instead of ICEing later. */ if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "incompatible vector type setup from " "bool pattern detection\n"); return false; } /* For single-argument PHIs assume coalescing which means zero cost for the scalar and the vector PHIs. This avoids artificially favoring the vector path (but may pessimize it in some cases). */ if (gimple_phi_num_args (as_a <gphi *> (stmt_info->stmt)) > 1) record_stmt_cost (cost_vec, SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node), vector_stmt, slp_node, vectype, 0, vect_body); SLP_TREE_TYPE (slp_node) = phi_info_type; return true; } tree scalar_dest = gimple_phi_result (stmt_info->stmt); basic_block bb = gimple_bb (stmt_info->stmt); tree vec_dest = vect_create_destination_var (scalar_dest, vectype); auto_vec<gphi *> new_phis; for (unsigned i = 0; i < gimple_phi_num_args (stmt_info->stmt); ++i) { slp_tree child = SLP_TREE_CHILDREN (slp_node)[i]; /* Skip not yet vectorized defs. */ if (SLP_TREE_DEF_TYPE (child) == vect_internal_def && SLP_TREE_VEC_DEFS (child).is_empty ()) continue; auto_vec<tree> vec_oprnds; vect_get_slp_defs (SLP_TREE_CHILDREN (slp_node)[i], &vec_oprnds); if (!new_phis.exists ()) { new_phis.create (vec_oprnds.length ()); for (unsigned j = 0; j < vec_oprnds.length (); j++) { /* Create the vectorized LC PHI node. */ new_phis.quick_push (create_phi_node (vec_dest, bb)); slp_node->push_vec_def (new_phis[j]); } } edge e = gimple_phi_arg_edge (as_a <gphi *> (stmt_info->stmt), i); for (unsigned j = 0; j < vec_oprnds.length (); j++) add_phi_arg (new_phis[j], vec_oprnds[j], e, UNKNOWN_LOCATION); } /* We should have at least one already vectorized child. */ gcc_assert (new_phis.exists ()); return true; } /* Vectorizes first order recurrences. An overview of the transformation is described below. Suppose we have the following loop. int t = 0; for (int i = 0; i < n; ++i) { b[i] = a[i] - t; t = a[i]; } There is a first-order recurrence on 'a'. For this loop, the scalar IR looks (simplified) like: scalar.preheader: init = 0; scalar.body: i = PHI <0(scalar.preheader), i+1(scalar.body)> _2 = PHI <(init(scalar.preheader), <_1(scalar.body)> _1 = a[i] b[i] = _1 - _2 if (i < n) goto scalar.body In this example, _2 is a recurrence because it's value depends on the previous iteration. We vectorize this as (VF = 4) vector.preheader: vect_init = vect_cst(..., ..., ..., 0) vector.body i = PHI <0(vector.preheader), i+4(vector.body)> vect_1 = PHI <vect_init(vector.preheader), v2(vector.body)> vect_2 = a[i, i+1, i+2, i+3]; vect_3 = vec_perm (vect_1, vect_2, { 3, 4, 5, 6 }) b[i, i+1, i+2, i+3] = vect_2 - vect_3 if (..) goto vector.body In this function, vectorizable_recurr, we code generate both the vector PHI node and the permute since those together compute the vectorized value of the scalar PHI. We do not yet have the backedge value to fill in there nor into the vec_perm. Those are filled in vect_schedule_scc. TODO: Since the scalar loop does not have a use of the recurrence outside of the loop the natural way to implement peeling via vectorizing the live value doesn't work. For now peeling of loops with a recurrence is not implemented. For SLP the supported cases are restricted to those requiring a single vector recurrence PHI. */ bool vectorizable_recurr (loop_vec_info loop_vinfo, stmt_vec_info stmt_info, slp_tree slp_node, stmt_vector_for_cost *cost_vec) { if (!loop_vinfo || !is_a<gphi *> (stmt_info->stmt)) return false; gphi *phi = as_a<gphi *> (stmt_info->stmt); /* So far we only support first-order recurrence auto-vectorization. */ if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_first_order_recurrence) return false; tree vectype = SLP_TREE_VECTYPE (slp_node); unsigned ncopies = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); poly_int64 nunits = TYPE_VECTOR_SUBPARTS (vectype); unsigned dist = SLP_TREE_LANES (slp_node); /* We need to be able to make progress with a single vector. */ if (maybe_gt (dist * 2, nunits)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "first order recurrence exceeds half of " "a vector\n"); return false; } /* We need to be able to build a { ..., a, b } init vector with dist number of distinct trailing values. Always possible when dist == 1 or when nunits is constant or when the initializations are uniform. */ tree uniform_initval = NULL_TREE; edge pe = loop_preheader_edge (LOOP_VINFO_LOOP (loop_vinfo)); for (stmt_vec_info s : SLP_TREE_SCALAR_STMTS (slp_node)) { gphi *phi = as_a <gphi *> (s->stmt); if (! uniform_initval) uniform_initval = PHI_ARG_DEF_FROM_EDGE (phi, pe); else if (! operand_equal_p (uniform_initval, PHI_ARG_DEF_FROM_EDGE (phi, pe))) { uniform_initval = NULL_TREE; break; } } if (!uniform_initval && !nunits.is_constant ()) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "cannot build initialization vector for " "first order recurrence\n"); return false; } /* First-order recurrence autovectorization needs to handle permutation with indices = [nunits-1, nunits, nunits+1, ...]. */ vec_perm_builder sel (nunits, 1, 3); for (int i = 0; i < 3; ++i) sel.quick_push (nunits - dist + i); vec_perm_indices indices (sel, 2, nunits); if (cost_vec) /* transformation not required. */ { if (!can_vec_perm_const_p (TYPE_MODE (vectype), TYPE_MODE (vectype), indices)) return false; /* We eventually need to set a vector type on invariant arguments. */ unsigned j; slp_tree child; FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (slp_node), j, child) if (!vect_maybe_update_slp_op_vectype (child, vectype)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "incompatible vector types for " "invariants\n"); return false; } /* Verify we have set up compatible types. */ edge le = loop_latch_edge (LOOP_VINFO_LOOP (loop_vinfo)); slp_tree latch_def = SLP_TREE_CHILDREN (slp_node)[le->dest_idx]; tree latch_vectype = SLP_TREE_VECTYPE (latch_def); if (!types_compatible_p (latch_vectype, vectype)) return false; /* The recurrence costs the initialization vector and one permute for each copy. With SLP the prologue value is explicitly represented and costed separately. */ unsigned prologue_cost = 0; unsigned inside_cost = record_stmt_cost (cost_vec, ncopies, vector_stmt, slp_node, 0, vect_body); if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "vectorizable_recurr: inside_cost = %d, " "prologue_cost = %d .\n", inside_cost, prologue_cost); SLP_TREE_TYPE (slp_node) = recurr_info_type; return true; } tree vec_init; if (! uniform_initval) { vec<constructor_elt, va_gc> *v = NULL; vec_alloc (v, nunits.to_constant ()); for (unsigned i = 0; i < nunits.to_constant () - dist; ++i) CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, build_zero_cst (TREE_TYPE (vectype))); for (stmt_vec_info s : SLP_TREE_SCALAR_STMTS (slp_node)) { gphi *phi = as_a <gphi *> (s->stmt); tree preheader = PHI_ARG_DEF_FROM_EDGE (phi, pe); if (!useless_type_conversion_p (TREE_TYPE (vectype), TREE_TYPE (preheader))) { gimple_seq stmts = NULL; preheader = gimple_convert (&stmts, TREE_TYPE (vectype), preheader); gsi_insert_seq_on_edge_immediate (pe, stmts); } CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, preheader); } vec_init = build_constructor (vectype, v); } else vec_init = uniform_initval; vec_init = vect_init_vector (loop_vinfo, stmt_info, vec_init, vectype, NULL); /* Create the vectorized first-order PHI node. */ tree vec_dest = vect_get_new_vect_var (vectype, vect_simple_var, "vec_recur_"); basic_block bb = gimple_bb (phi); gphi *new_phi = create_phi_node (vec_dest, bb); add_phi_arg (new_phi, vec_init, pe, UNKNOWN_LOCATION); /* Insert shuffles the first-order recurrence autovectorization. result = VEC_PERM <vec_recur, vect_1, index[nunits-1, nunits, ...]>. */ tree perm = vect_gen_perm_mask_checked (vectype, indices); /* Insert the required permute after the latch definition. The second and later operands are tentative and will be updated when we have vectorized the latch definition. */ edge le = loop_latch_edge (LOOP_VINFO_LOOP (loop_vinfo)); gimple *latch_def = SSA_NAME_DEF_STMT (PHI_ARG_DEF_FROM_EDGE (phi, le)); gimple_stmt_iterator gsi2 = gsi_for_stmt (latch_def); gsi_next (&gsi2); for (unsigned i = 0; i < ncopies; ++i) { vec_dest = make_ssa_name (vectype); gassign *vperm = gimple_build_assign (vec_dest, VEC_PERM_EXPR, i == 0 ? gimple_phi_result (new_phi) : NULL, NULL, perm); vect_finish_stmt_generation (loop_vinfo, stmt_info, vperm, &gsi2); slp_node->push_vec_def (vperm); } return true; } /* Return true if VECTYPE represents a vector that requires lowering by the vector lowering pass. */ bool vect_emulated_vector_p (tree vectype) { return (!VECTOR_MODE_P (TYPE_MODE (vectype)) && (!VECTOR_BOOLEAN_TYPE_P (vectype) || TYPE_PRECISION (TREE_TYPE (vectype)) != 1)); } /* Return true if we can emulate CODE on an integer mode representation of a vector. */ bool vect_can_vectorize_without_simd_p (tree_code code) { switch (code) { case PLUS_EXPR: case MINUS_EXPR: case NEGATE_EXPR: case BIT_AND_EXPR: case BIT_IOR_EXPR: case BIT_XOR_EXPR: case BIT_NOT_EXPR: return true; default: return false; } } /* Likewise, but taking a code_helper. */ bool vect_can_vectorize_without_simd_p (code_helper code) { return (code.is_tree_code () && vect_can_vectorize_without_simd_p (tree_code (code))); } /* Create vector init for vectorized iv. */ static tree vect_create_nonlinear_iv_init (gimple_seq* stmts, tree init_expr, tree step_expr, poly_uint64 nunits, tree vectype, enum vect_induction_op_type induction_type) { unsigned HOST_WIDE_INT const_nunits; tree vec_shift, vec_init, new_name; unsigned i; tree itype = TREE_TYPE (vectype); /* iv_loop is the loop to be vectorized. Create: vec_init = [X, X+S, X+2*S, X+3*S] (S = step_expr, X = init_expr). */ new_name = gimple_convert (stmts, itype, init_expr); switch (induction_type) { case vect_step_op_shr: case vect_step_op_shl: /* Build the Initial value from shift_expr. */ vec_init = gimple_build_vector_from_val (stmts, vectype, new_name); vec_shift = gimple_build (stmts, VEC_SERIES_EXPR, vectype, build_zero_cst (itype), step_expr); vec_init = gimple_build (stmts, (induction_type == vect_step_op_shr ? RSHIFT_EXPR : LSHIFT_EXPR), vectype, vec_init, vec_shift); break; case vect_step_op_neg: { vec_init = gimple_build_vector_from_val (stmts, vectype, new_name); tree vec_neg = gimple_build (stmts, NEGATE_EXPR, vectype, vec_init); /* The encoding has 2 interleaved stepped patterns. */ vec_perm_builder sel (nunits, 2, 3); sel.quick_grow (6); for (i = 0; i < 3; i++) { sel[2 * i] = i; sel[2 * i + 1] = i + nunits; } vec_perm_indices indices (sel, 2, nunits); /* Don't use vect_gen_perm_mask_checked since can_vec_perm_const_p may fail when vec_init is const vector. In that situation vec_perm is not really needed. */ tree perm_mask_even = vect_gen_perm_mask_any (vectype, indices); vec_init = gimple_build (stmts, VEC_PERM_EXPR, vectype, vec_init, vec_neg, perm_mask_even); } break; case vect_step_op_mul: { /* Use unsigned mult to avoid UD integer overflow. */ gcc_assert (nunits.is_constant (&const_nunits)); tree utype = unsigned_type_for (itype); tree uvectype = build_vector_type (utype, TYPE_VECTOR_SUBPARTS (vectype)); new_name = gimple_convert (stmts, utype, new_name); vec_init = gimple_build_vector_from_val (stmts, uvectype, new_name); tree_vector_builder elts (uvectype, const_nunits, 1); tree elt_step = build_one_cst (utype); elts.quick_push (elt_step); for (i = 1; i < const_nunits; i++) { /* Create: new_name_i = new_name + step_expr. */ elt_step = gimple_build (stmts, MULT_EXPR, utype, elt_step, step_expr); elts.quick_push (elt_step); } /* Create a vector from [new_name_0, new_name_1, ..., new_name_nunits-1]. */ tree vec_mul = gimple_build_vector (stmts, &elts); vec_init = gimple_build (stmts, MULT_EXPR, uvectype, vec_init, vec_mul); vec_init = gimple_convert (stmts, vectype, vec_init); } break; default: gcc_unreachable (); } return vec_init; } /* Peel init_expr by skip_niter for induction_type. */ tree vect_peel_nonlinear_iv_init (gimple_seq* stmts, tree init_expr, tree skip_niters, tree step_expr, enum vect_induction_op_type induction_type) { gcc_assert (TREE_CODE (skip_niters) == INTEGER_CST); tree type = TREE_TYPE (init_expr); unsigned prec = TYPE_PRECISION (type); switch (induction_type) { case vect_step_op_neg: if (TREE_INT_CST_LOW (skip_niters) % 2) init_expr = gimple_build (stmts, NEGATE_EXPR, type, init_expr); /* else no change. */ break; case vect_step_op_shr: case vect_step_op_shl: skip_niters = gimple_convert (stmts, type, skip_niters); step_expr = gimple_build (stmts, MULT_EXPR, type, step_expr, skip_niters); /* When shift mount >= precision, need to avoid UD. In the original loop, there's no UD, and according to semantic, init_expr should be 0 for lshr, ashl, and >>= (prec - 1) for ashr. */ if (!tree_fits_uhwi_p (step_expr) || tree_to_uhwi (step_expr) >= prec) { if (induction_type == vect_step_op_shl || TYPE_UNSIGNED (type)) init_expr = build_zero_cst (type); else init_expr = gimple_build (stmts, RSHIFT_EXPR, type, init_expr, wide_int_to_tree (type, prec - 1)); } else init_expr = gimple_build (stmts, (induction_type == vect_step_op_shr ? RSHIFT_EXPR : LSHIFT_EXPR), type, init_expr, step_expr); break; case vect_step_op_mul: { tree utype = unsigned_type_for (type); init_expr = gimple_convert (stmts, utype, init_expr); wide_int skipn = wi::to_wide (skip_niters); wide_int begin = wi::to_wide (step_expr); auto_mpz base, exp, mod, res; wi::to_mpz (begin, base, TYPE_SIGN (type)); wi::to_mpz (skipn, exp, UNSIGNED); mpz_ui_pow_ui (mod, 2, TYPE_PRECISION (type)); mpz_powm (res, base, exp, mod); begin = wi::from_mpz (utype, res, true); tree mult_expr = wide_int_to_tree (utype, begin); init_expr = gimple_build (stmts, MULT_EXPR, utype, init_expr, mult_expr); init_expr = gimple_convert (stmts, type, init_expr); } break; default: gcc_unreachable (); } return init_expr; } /* Create vector step for vectorized iv. */ static tree vect_create_nonlinear_iv_step (gimple_seq* stmts, tree step_expr, poly_uint64 vf, enum vect_induction_op_type induction_type) { tree expr = build_int_cst (TREE_TYPE (step_expr), vf); tree new_name = NULL; /* Step should be pow (step, vf) for mult induction. */ if (induction_type == vect_step_op_mul) { gcc_assert (vf.is_constant ()); wide_int begin = wi::to_wide (step_expr); for (unsigned i = 0; i != vf.to_constant () - 1; i++) begin = wi::mul (begin, wi::to_wide (step_expr)); new_name = wide_int_to_tree (TREE_TYPE (step_expr), begin); } else if (induction_type == vect_step_op_neg) /* Do nothing. */ ; else new_name = gimple_build (stmts, MULT_EXPR, TREE_TYPE (step_expr), expr, step_expr); return new_name; } static tree vect_create_nonlinear_iv_vec_step (loop_vec_info loop_vinfo, stmt_vec_info stmt_info, tree new_name, tree vectype, enum vect_induction_op_type induction_type) { /* No step is needed for neg induction. */ if (induction_type == vect_step_op_neg) return NULL; tree t = unshare_expr (new_name); gcc_assert (CONSTANT_CLASS_P (new_name) || TREE_CODE (new_name) == SSA_NAME); tree new_vec = build_vector_from_val (vectype, t); tree vec_step = vect_init_vector (loop_vinfo, stmt_info, new_vec, vectype, NULL); return vec_step; } /* Update vectorized iv with vect_step, induc_def is init. */ static tree vect_update_nonlinear_iv (gimple_seq* stmts, tree vectype, tree induc_def, tree vec_step, enum vect_induction_op_type induction_type) { tree vec_def = induc_def; switch (induction_type) { case vect_step_op_mul: { /* Use unsigned mult to avoid UD integer overflow. */ tree uvectype = build_vector_type (unsigned_type_for (TREE_TYPE (vectype)), TYPE_VECTOR_SUBPARTS (vectype)); vec_def = gimple_convert (stmts, uvectype, vec_def); vec_step = gimple_convert (stmts, uvectype, vec_step); vec_def = gimple_build (stmts, MULT_EXPR, uvectype, vec_def, vec_step); vec_def = gimple_convert (stmts, vectype, vec_def); } break; case vect_step_op_shr: vec_def = gimple_build (stmts, RSHIFT_EXPR, vectype, vec_def, vec_step); break; case vect_step_op_shl: vec_def = gimple_build (stmts, LSHIFT_EXPR, vectype, vec_def, vec_step); break; case vect_step_op_neg: vec_def = induc_def; /* Do nothing. */ break; default: gcc_unreachable (); } return vec_def; } /* Function vectorizable_nonlinear_induction Check if STMT_INFO performs an nonlinear 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 true if STMT_INFO is vectorizable in this way. */ static bool vectorizable_nonlinear_induction (loop_vec_info loop_vinfo, stmt_vec_info stmt_info, slp_tree slp_node, stmt_vector_for_cost *cost_vec) { class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); unsigned ncopies; bool nested_in_vect_loop = false; class loop *iv_loop; tree vec_def; edge pe = loop_preheader_edge (loop); basic_block new_bb; tree vec_init, vec_step; tree new_name; gimple *new_stmt; gphi *induction_phi; tree induc_def, vec_dest; tree init_expr, step_expr; tree niters_skip; poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); unsigned i; gimple_stmt_iterator si; gphi *phi = dyn_cast <gphi *> (stmt_info->stmt); tree vectype = SLP_TREE_VECTYPE (slp_node); poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); enum vect_induction_op_type induction_type = STMT_VINFO_LOOP_PHI_EVOLUTION_TYPE (stmt_info); gcc_assert (induction_type > vect_step_op_add); ncopies = vect_get_num_copies (loop_vinfo, slp_node, vectype); gcc_assert (ncopies >= 1); /* FORNOW. Only handle nonlinear induction in the same loop. */ if (nested_in_vect_loop_p (loop, stmt_info)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "nonlinear induction in nested loop.\n"); return false; } iv_loop = loop; gcc_assert (iv_loop == (gimple_bb (phi))->loop_father); /* TODO: Support multi-lane SLP for nonlinear iv. There should be separate vector iv update for each iv and a permutation to generate wanted vector iv. */ if (SLP_TREE_LANES (slp_node) > 1) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "SLP induction not supported for nonlinear" " induction.\n"); return false; } if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype))) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "floating point nonlinear induction vectorization" " not supported.\n"); return false; } step_expr = STMT_VINFO_LOOP_PHI_EVOLUTION_PART (stmt_info); init_expr = vect_phi_initial_value (phi); gcc_assert (step_expr != NULL_TREE && init_expr != NULL && TREE_CODE (step_expr) == INTEGER_CST); /* step_expr should be aligned with init_expr, .i.e. uint64 a >> 1, step is int, but vector<uint64> shift is used. */ step_expr = fold_convert (TREE_TYPE (vectype), step_expr); if (TREE_CODE (init_expr) == INTEGER_CST) init_expr = fold_convert (TREE_TYPE (vectype), init_expr); else if (!tree_nop_conversion_p (TREE_TYPE (vectype), TREE_TYPE (init_expr))) { /* INIT_EXPR could be a bit_field, bail out for such case. */ if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "nonlinear induction vectorization failed:" " component type of vectype is not a nop conversion" " from type of init_expr.\n"); return false; } switch (induction_type) { case vect_step_op_neg: if (maybe_eq (TYPE_VECTOR_SUBPARTS (vectype), 1u)) return false; if (TREE_CODE (init_expr) != INTEGER_CST && TREE_CODE (init_expr) != REAL_CST) { /* Check for backend support of NEGATE_EXPR and vec_perm. */ if (!directly_supported_p (NEGATE_EXPR, vectype)) return false; /* The encoding has 2 interleaved stepped patterns. */ vec_perm_builder sel (nunits, 2, 3); machine_mode mode = TYPE_MODE (vectype); sel.quick_grow (6); for (i = 0; i < 3; i++) { sel[i * 2] = i; sel[i * 2 + 1] = i + nunits; } vec_perm_indices indices (sel, 2, nunits); if (!can_vec_perm_const_p (mode, mode, indices)) return false; } break; case vect_step_op_mul: { /* Check for backend support of MULT_EXPR. */ if (!directly_supported_p (MULT_EXPR, vectype)) return false; /* ?? How to construct vector step for variable number vector. [ 1, step, pow (step, 2), pow (step, 4), .. ]. */ if (!vf.is_constant ()) return false; } break; case vect_step_op_shr: /* Check for backend support of RSHIFT_EXPR. */ if (!directly_supported_p (RSHIFT_EXPR, vectype, optab_vector)) return false; /* Don't shift more than type precision to avoid UD. */ if (!tree_fits_uhwi_p (step_expr) || maybe_ge (nunits * tree_to_uhwi (step_expr), TYPE_PRECISION (TREE_TYPE (init_expr)))) return false; break; case vect_step_op_shl: /* Check for backend support of RSHIFT_EXPR. */ if (!directly_supported_p (LSHIFT_EXPR, vectype, optab_vector)) return false; /* Don't shift more than type precision to avoid UD. */ if (!tree_fits_uhwi_p (step_expr) || maybe_ge (nunits * tree_to_uhwi (step_expr), TYPE_PRECISION (TREE_TYPE (init_expr)))) return false; break; default: gcc_unreachable (); } if (cost_vec) /* transformation not required. */ { unsigned inside_cost = 0, prologue_cost = 0; /* loop cost for vec_loop. Neg induction doesn't have any inside_cost. */ inside_cost = record_stmt_cost (cost_vec, ncopies, vector_stmt, slp_node, 0, vect_body); /* loop cost for vec_loop. Neg induction doesn't have any inside_cost. */ if (induction_type == vect_step_op_neg) inside_cost = 0; /* prologue cost for vec_init and vec_step. */ prologue_cost = record_stmt_cost (cost_vec, 2, scalar_to_vec, slp_node, 0, vect_prologue); if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "vect_model_induction_cost: inside_cost = %d, " "prologue_cost = %d. \n", inside_cost, prologue_cost); SLP_TREE_TYPE (slp_node) = induc_vec_info_type; DUMP_VECT_SCOPE ("vectorizable_nonlinear_induction"); return true; } /* Transform. */ /* Compute a vector variable, initialized with the first VF values of the induction variable. E.g., for an iv with IV_PHI='X' and evolution S, for a vector of 4 units, we want to compute: [X, X + S, X + 2*S, X + 3*S]. */ if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "transform induction phi.\n"); pe = loop_preheader_edge (iv_loop); /* Find the first insertion point in the BB. */ basic_block bb = gimple_bb (phi); si = gsi_after_labels (bb); gimple_seq stmts = NULL; niters_skip = LOOP_VINFO_MASK_SKIP_NITERS (loop_vinfo); /* If we are using the loop mask to "peel" for alignment then we need to adjust the start value here. */ if (niters_skip != NULL_TREE) init_expr = vect_peel_nonlinear_iv_init (&stmts, init_expr, niters_skip, step_expr, induction_type); vec_init = vect_create_nonlinear_iv_init (&stmts, init_expr, step_expr, nunits, vectype, induction_type); if (stmts) { new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); gcc_assert (!new_bb); } stmts = NULL; new_name = vect_create_nonlinear_iv_step (&stmts, step_expr, vf, induction_type); if (stmts) { new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); gcc_assert (!new_bb); } vec_step = vect_create_nonlinear_iv_vec_step (loop_vinfo, stmt_info, new_name, vectype, induction_type); /* Create the following def-use cycle: loop prolog: vec_init = ... vec_step = ... loop: vec_iv = PHI <vec_init, vec_loop> ... STMT ... vec_loop = vec_iv + vec_step; */ /* 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); induc_def = PHI_RESULT (induction_phi); /* Create the iv update inside the loop. */ stmts = NULL; vec_def = vect_update_nonlinear_iv (&stmts, vectype, induc_def, vec_step, induction_type); gsi_insert_seq_before (&si, stmts, GSI_SAME_STMT); new_stmt = SSA_NAME_DEF_STMT (vec_def); /* 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_node->push_vec_def (induction_phi); /* 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 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. */ if (ncopies > 1) { stmts = NULL; /* FORNOW. This restriction should be relaxed. */ gcc_assert (!nested_in_vect_loop); new_name = vect_create_nonlinear_iv_step (&stmts, step_expr, nunits, induction_type); vec_step = vect_create_nonlinear_iv_vec_step (loop_vinfo, stmt_info, new_name, vectype, induction_type); vec_def = induc_def; for (i = 1; i < ncopies; i++) { /* vec_i = vec_prev + vec_step. */ stmts = NULL; vec_def = vect_update_nonlinear_iv (&stmts, vectype, vec_def, vec_step, induction_type); gsi_insert_seq_before (&si, stmts, GSI_SAME_STMT); new_stmt = SSA_NAME_DEF_STMT (vec_def); slp_node->push_vec_def (new_stmt); } } if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "transform induction: created def-use cycle: %G%G", (gimple *) induction_phi, SSA_NAME_DEF_STMT (vec_def)); return true; } /* Function vectorizable_induction 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 true if STMT_INFO is vectorizable in this way. */ bool vectorizable_induction (loop_vec_info loop_vinfo, stmt_vec_info stmt_info, slp_tree slp_node, stmt_vector_for_cost *cost_vec) { class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); bool nested_in_vect_loop = false; class loop *iv_loop; tree vec_def; edge pe = loop_preheader_edge (loop); basic_block new_bb; tree vec_init = NULL_TREE, vec_step, t; tree new_name; gphi *induction_phi; tree induc_def, vec_dest; poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); unsigned i; tree expr; tree index_vectype = NULL_TREE; gimple_stmt_iterator si; enum vect_induction_op_type induction_type = STMT_VINFO_LOOP_PHI_EVOLUTION_TYPE (stmt_info); gphi *phi = dyn_cast <gphi *> (stmt_info->stmt); if (!phi) return false; if (!STMT_VINFO_RELEVANT_P (stmt_info)) return false; /* Make sure it was recognized as induction computation. */ if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_induction_def) return false; /* Handle nonlinear induction in a separate place. */ if (induction_type != vect_step_op_add) return vectorizable_nonlinear_induction (loop_vinfo, stmt_info, slp_node, cost_vec); tree vectype = SLP_TREE_VECTYPE (slp_node); poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); /* FORNOW. These restrictions should be relaxed. */ if (nested_in_vect_loop_p (loop, stmt_info)) { imm_use_iterator imm_iter; use_operand_p use_p; gimple *exit_phi; edge latch_e; tree loop_arg; exit_phi = NULL; latch_e = loop_latch_edge (loop->inner); loop_arg = PHI_ARG_DEF_FROM_EDGE (phi, latch_e); FOR_EACH_IMM_USE_FAST (use_p, imm_iter, loop_arg) { gimple *use_stmt = USE_STMT (use_p); if (is_gimple_debug (use_stmt)) continue; if (!flow_bb_inside_loop_p (loop->inner, gimple_bb (use_stmt))) { exit_phi = use_stmt; break; } } if (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))) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "inner-loop induction only used outside " "of the outer vectorized loop.\n"); return false; } } nested_in_vect_loop = true; iv_loop = loop->inner; } else iv_loop = loop; gcc_assert (iv_loop == (gimple_bb (phi))->loop_father); if (!nunits.is_constant () && SLP_TREE_LANES (slp_node) != 1) { /* The current SLP code creates the step 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 (FLOAT_TYPE_P (vectype) && !param_vect_induction_float) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "floating point induction vectorization disabled\n"); return false; } tree step_expr = STMT_VINFO_LOOP_PHI_EVOLUTION_PART (stmt_info); gcc_assert (step_expr != NULL_TREE); if (INTEGRAL_TYPE_P (TREE_TYPE (step_expr)) && !type_has_mode_precision_p (TREE_TYPE (step_expr))) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "bit-precision induction vectorization not " "supported.\n"); return false; } tree stept = TREE_TYPE (step_expr); tree step_vectype = get_same_sized_vectype (stept, vectype); stept = TREE_TYPE (step_vectype); /* Check for target support of the vectorized arithmetic used here. */ if (!target_supports_op_p (step_vectype, PLUS_EXPR, optab_default) || !target_supports_op_p (step_vectype, MINUS_EXPR, optab_default)) return false; if (!nunits.is_constant ()) { if (!target_supports_op_p (step_vectype, MULT_EXPR, optab_default)) return false; /* FLOAT_EXPR when computing VEC_INIT for float inductions. */ if (SCALAR_FLOAT_TYPE_P (stept)) { tree index_type = build_nonstandard_integer_type (GET_MODE_BITSIZE (SCALAR_TYPE_MODE (stept)), 1); index_vectype = build_vector_type (index_type, nunits); if (!can_float_p (TYPE_MODE (step_vectype), TYPE_MODE (index_vectype), 1)) return false; } } if (cost_vec) /* transformation not required. */ { unsigned inside_cost = 0, prologue_cost = 0; /* We eventually need to set a vector type on invariant arguments. */ unsigned j; slp_tree child; FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (slp_node), j, child) if (!vect_maybe_update_slp_op_vectype (child, SLP_TREE_VECTYPE (slp_node))) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "incompatible vector types for " "invariants\n"); return false; } /* loop cost for vec_loop. */ inside_cost = record_stmt_cost (cost_vec, SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node), vector_stmt, slp_node, 0, vect_body); /* prologue cost for vec_init (if not nested) and step. */ prologue_cost = record_stmt_cost (cost_vec, 1 + !nested_in_vect_loop, scalar_to_vec, slp_node, 0, vect_prologue); if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "vect_model_induction_cost: inside_cost = %d, " "prologue_cost = %d .\n", inside_cost, prologue_cost); SLP_TREE_TYPE (slp_node) = induc_vec_info_type; DUMP_VECT_SCOPE ("vectorizable_induction"); return true; } /* Transform. */ /* Compute a vector variable, initialized with the first VF values of the induction variable. E.g., for an iv with IV_PHI='X' and evolution S, for a vector of 4 units, we want to compute: [X, X + S, X + 2*S, X + 3*S]. */ if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "transform induction phi.\n"); pe = loop_preheader_edge (iv_loop); /* 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 with group size 3 we need [i0, i1, i2, i0 + S0] [i1 + S1, i2 + S2, i0 + 2*S0, i1 + 2*S1] [i2 + 2*S2, i0 + 3*S0, i1 + 3*S1, i2 + 3*S2]. */ gimple_stmt_iterator incr_si; bool insert_after; standard_iv_increment_position (iv_loop, &incr_si, &insert_after); /* The initial values are vectorized, but any lanes > group_size need adjustment. */ slp_tree init_node = SLP_TREE_CHILDREN (slp_node)[pe->dest_idx]; /* Gather steps. Since we do not vectorize inductions as cycles we have to reconstruct the step from SCEV data. */ unsigned group_size = SLP_TREE_LANES (slp_node); tree *steps = XALLOCAVEC (tree, group_size); tree *inits = XALLOCAVEC (tree, group_size); stmt_vec_info phi_info; FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (slp_node), i, phi_info) { steps[i] = STMT_VINFO_LOOP_PHI_EVOLUTION_PART (phi_info); if (!init_node) inits[i] = gimple_phi_arg_def (as_a<gphi *> (phi_info->stmt), pe->dest_idx); } /* Now generate the IVs. */ unsigned nvects = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); gcc_assert (multiple_p (nunits * nvects, group_size)); unsigned nivs; unsigned HOST_WIDE_INT const_nunits; if (nested_in_vect_loop) nivs = nvects; else if (nunits.is_constant (&const_nunits)) { /* Compute the number of distinct IVs we need. First reduce group_size if it is a multiple of const_nunits so we get one IV for a group_size of 4 but const_nunits 2. */ unsigned group_sizep = group_size; if (group_sizep % const_nunits == 0) group_sizep = group_sizep / const_nunits; nivs = least_common_multiple (group_sizep, const_nunits) / const_nunits; } else { gcc_assert (SLP_TREE_LANES (slp_node) == 1); nivs = 1; } gimple_seq init_stmts = NULL; tree lupdate_mul = NULL_TREE; if (!nested_in_vect_loop) { if (nunits.is_constant (&const_nunits)) { /* The number of iterations covered in one vector iteration. */ unsigned lup_mul = (nvects * const_nunits) / group_size; lupdate_mul = build_vector_from_val (step_vectype, SCALAR_FLOAT_TYPE_P (stept) ? build_real_from_wide (stept, lup_mul, UNSIGNED) : build_int_cstu (stept, lup_mul)); } else { if (SCALAR_FLOAT_TYPE_P (stept)) { tree tem = build_int_cst (integer_type_node, vf); lupdate_mul = gimple_build (&init_stmts, FLOAT_EXPR, stept, tem); } else lupdate_mul = build_int_cst (stept, vf); lupdate_mul = gimple_build_vector_from_val (&init_stmts, step_vectype, lupdate_mul); } } tree peel_mul = NULL_TREE; if (LOOP_VINFO_MASK_SKIP_NITERS (loop_vinfo)) { if (SCALAR_FLOAT_TYPE_P (stept)) peel_mul = gimple_build (&init_stmts, FLOAT_EXPR, stept, LOOP_VINFO_MASK_SKIP_NITERS (loop_vinfo)); else peel_mul = gimple_convert (&init_stmts, stept, LOOP_VINFO_MASK_SKIP_NITERS (loop_vinfo)); peel_mul = gimple_build_vector_from_val (&init_stmts, step_vectype, peel_mul); /* If early break then we have to create a new PHI which we can use as an offset to adjust the induction reduction in early exits. This is because when peeling for alignment using masking, the first few elements of the vector can be inactive. As such if we find the entry in the first iteration we have adjust the starting point of the scalar code. We do this by creating a new scalar PHI that keeps track of whether we are the first iteration of the loop (with the additional masking) or whether we have taken a loop iteration already. The generated sequence: pre-header: bb1: i_1 = <number of leading inactive elements> header: bb2: i_2 = PHI <i_1(bb1), 0(latch)> … early-exit: bb3: i_3 = iv_step * i_2 + PHI<vector-iv> The first part of the adjustment to create i_1 and i_2 are done here and the last part creating i_3 is done in vectorizable_live_operations when the induction extraction is materialized. */ if (LOOP_VINFO_EARLY_BREAKS (loop_vinfo) && !LOOP_VINFO_MASK_NITERS_PFA_OFFSET (loop_vinfo)) { auto skip_niters = LOOP_VINFO_MASK_SKIP_NITERS (loop_vinfo); tree ty_skip_niters = TREE_TYPE (skip_niters); tree break_lhs_phi = vect_get_new_vect_var (ty_skip_niters, vect_scalar_var, "pfa_iv_offset"); gphi *nphi = create_phi_node (break_lhs_phi, bb); add_phi_arg (nphi, skip_niters, pe, UNKNOWN_LOCATION); add_phi_arg (nphi, build_zero_cst (ty_skip_niters), loop_latch_edge (iv_loop), UNKNOWN_LOCATION); LOOP_VINFO_MASK_NITERS_PFA_OFFSET (loop_vinfo) = PHI_RESULT (nphi); } } tree step_mul = NULL_TREE; unsigned ivn; auto_vec<tree> vec_steps; for (ivn = 0; ivn < nivs; ++ivn) { gimple_seq stmts = NULL; bool invariant = true; if (nunits.is_constant (&const_nunits)) { tree_vector_builder step_elts (step_vectype, const_nunits, 1); tree_vector_builder init_elts (vectype, const_nunits, 1); tree_vector_builder mul_elts (step_vectype, const_nunits, 1); for (unsigned eltn = 0; eltn < const_nunits; ++eltn) { /* The scalar steps of the IVs. */ tree elt = steps[(ivn*const_nunits + eltn) % group_size]; elt = gimple_convert (&init_stmts, TREE_TYPE (step_vectype), elt); step_elts.quick_push (elt); if (!init_node) { /* The scalar inits of the IVs if not vectorized. */ elt = inits[(ivn*const_nunits + eltn) % group_size]; if (!useless_type_conversion_p (TREE_TYPE (vectype), TREE_TYPE (elt))) elt = gimple_build (&init_stmts, VIEW_CONVERT_EXPR, TREE_TYPE (vectype), elt); init_elts.quick_push (elt); } /* The number of steps to add to the initial values. */ unsigned mul_elt = (ivn*const_nunits + eltn) / group_size; mul_elts.quick_push (SCALAR_FLOAT_TYPE_P (stept) ? build_real_from_wide (stept, mul_elt, UNSIGNED) : build_int_cstu (stept, mul_elt)); } vec_step = gimple_build_vector (&init_stmts, &step_elts); step_mul = gimple_build_vector (&init_stmts, &mul_elts); if (!init_node) vec_init = gimple_build_vector (&init_stmts, &init_elts); } else { if (init_node) ; else if (INTEGRAL_TYPE_P (TREE_TYPE (steps[0]))) { new_name = gimple_convert (&init_stmts, stept, inits[0]); /* Build the initial value directly as a VEC_SERIES_EXPR. */ vec_init = gimple_build (&init_stmts, VEC_SERIES_EXPR, step_vectype, new_name, steps[0]); if (!useless_type_conversion_p (vectype, step_vectype)) vec_init = gimple_build (&init_stmts, VIEW_CONVERT_EXPR, vectype, vec_init); } else { /* Build: [base, base, base, ...] + (vectype) [0, 1, 2, ...] * [step, step, step, ...]. */ gcc_assert (SCALAR_FLOAT_TYPE_P (TREE_TYPE (steps[0]))); gcc_assert (flag_associative_math); gcc_assert (index_vectype != NULL_TREE); tree index = build_index_vector (index_vectype, 0, 1); new_name = gimple_convert (&init_stmts, TREE_TYPE (steps[0]), inits[0]); tree base_vec = gimple_build_vector_from_val (&init_stmts, step_vectype, new_name); tree step_vec = gimple_build_vector_from_val (&init_stmts, step_vectype, steps[0]); vec_init = gimple_build (&init_stmts, FLOAT_EXPR, step_vectype, index); vec_init = gimple_build (&init_stmts, MULT_EXPR, step_vectype, vec_init, step_vec); vec_init = gimple_build (&init_stmts, PLUS_EXPR, step_vectype, vec_init, base_vec); if (!useless_type_conversion_p (vectype, step_vectype)) vec_init = gimple_build (&init_stmts, VIEW_CONVERT_EXPR, vectype, vec_init); } /* iv_loop is nested in the loop to be vectorized. Generate: vec_step = [S, S, S, S] */ t = unshare_expr (steps[0]); gcc_assert (CONSTANT_CLASS_P (t) || TREE_CODE (t) == SSA_NAME); vec_step = gimple_build_vector_from_val (&init_stmts, step_vectype, t); } vec_steps.safe_push (vec_step); if (peel_mul) { if (!step_mul) { gcc_assert (!nunits.is_constant ()); step_mul = gimple_build (&init_stmts, MINUS_EXPR, step_vectype, build_zero_cst (step_vectype), peel_mul); } else step_mul = gimple_build (&init_stmts, MINUS_EXPR, step_vectype, step_mul, peel_mul); } /* 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); induc_def = PHI_RESULT (induction_phi); /* Create the iv update inside the loop */ tree up = vec_step; if (lupdate_mul) { if (LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo)) { /* When we're using loop_len produced by SELEC_VL, the non-final iterations are not always processing VF elements. So vectorize induction variable instead of _21 = vect_vec_iv_.6_22 + { VF, ... }; We should generate: _35 = .SELECT_VL (ivtmp_33, VF); vect_cst__22 = [vec_duplicate_expr] _35; _21 = vect_vec_iv_.6_22 + vect_cst__22; */ vec_loop_lens *lens = &LOOP_VINFO_LENS (loop_vinfo); tree len = vect_get_loop_len (loop_vinfo, NULL, lens, 1, vectype, 0, 0); if (SCALAR_FLOAT_TYPE_P (stept)) expr = gimple_build (&stmts, FLOAT_EXPR, stept, len); else expr = gimple_convert (&stmts, stept, len); lupdate_mul = gimple_build_vector_from_val (&stmts, step_vectype, expr); up = gimple_build (&stmts, MULT_EXPR, step_vectype, vec_step, lupdate_mul); } else up = gimple_build (&init_stmts, MULT_EXPR, step_vectype, vec_step, lupdate_mul); } vec_def = gimple_convert (&stmts, step_vectype, induc_def); vec_def = gimple_build (&stmts, PLUS_EXPR, step_vectype, vec_def, up); vec_def = gimple_convert (&stmts, vectype, vec_def); insert_iv_increment (&incr_si, insert_after, stmts); add_phi_arg (induction_phi, vec_def, loop_latch_edge (iv_loop), UNKNOWN_LOCATION); if (init_node) vec_init = vect_get_slp_vect_def (init_node, ivn); if (!nested_in_vect_loop && step_mul && !integer_zerop (step_mul)) { gcc_assert (invariant); vec_def = gimple_convert (&init_stmts, step_vectype, vec_init); up = gimple_build (&init_stmts, MULT_EXPR, step_vectype, vec_step, step_mul); vec_def = gimple_build (&init_stmts, PLUS_EXPR, step_vectype, vec_def, up); vec_init = gimple_convert (&init_stmts, vectype, vec_def); } /* Set the arguments of the phi node: */ add_phi_arg (induction_phi, vec_init, pe, UNKNOWN_LOCATION); slp_node->push_vec_def (induction_phi); } if (!nested_in_vect_loop) { /* Fill up to the number of vectors we need for the whole group. */ if (nunits.is_constant (&const_nunits)) nivs = least_common_multiple (group_size, const_nunits) / const_nunits; else nivs = 1; vec_steps.reserve (nivs-ivn); for (; ivn < nivs; ++ivn) { slp_node->push_vec_def (SLP_TREE_VEC_DEFS (slp_node)[0]); vec_steps.quick_push (vec_steps[0]); } } /* Re-use IVs when we can. We are generating further vector stmts by adding VF' * stride to the IVs generated above. */ if (ivn < nvects) { if (nunits.is_constant (&const_nunits)) { unsigned vfp = (least_common_multiple (group_size, const_nunits) / group_size); lupdate_mul = build_vector_from_val (step_vectype, SCALAR_FLOAT_TYPE_P (stept) ? build_real_from_wide (stept, vfp, UNSIGNED) : build_int_cstu (stept, vfp)); } else { if (SCALAR_FLOAT_TYPE_P (stept)) { tree tem = build_int_cst (integer_type_node, nunits); lupdate_mul = gimple_build (&init_stmts, FLOAT_EXPR, stept, tem); } else lupdate_mul = build_int_cst (stept, nunits); lupdate_mul = gimple_build_vector_from_val (&init_stmts, step_vectype, lupdate_mul); } for (; ivn < nvects; ++ivn) { gimple *iv = SSA_NAME_DEF_STMT (SLP_TREE_VEC_DEFS (slp_node)[ivn - nivs]); tree def = gimple_get_lhs (iv); if (ivn < 2*nivs) vec_steps[ivn - nivs] = gimple_build (&init_stmts, MULT_EXPR, step_vectype, vec_steps[ivn - nivs], lupdate_mul); gimple_seq stmts = NULL; def = gimple_convert (&stmts, step_vectype, def); def = gimple_build (&stmts, PLUS_EXPR, step_vectype, def, vec_steps[ivn % nivs]); def = gimple_convert (&stmts, vectype, def); if (gimple_code (iv) == GIMPLE_PHI) gsi_insert_seq_before (&si, stmts, GSI_SAME_STMT); else { gimple_stmt_iterator tgsi = gsi_for_stmt (iv); gsi_insert_seq_after (&tgsi, stmts, GSI_CONTINUE_LINKING); } slp_node->push_vec_def (def); } } new_bb = gsi_insert_seq_on_edge_immediate (pe, init_stmts); gcc_assert (!new_bb); return true; } /* Function vectorizable_live_operation_1. helper function for vectorizable_live_operation. */ static tree vectorizable_live_operation_1 (loop_vec_info loop_vinfo, basic_block exit_bb, tree vectype, slp_tree slp_node, tree bitsize, tree bitstart, tree vec_lhs, tree lhs_type, gimple_stmt_iterator *exit_gsi) { gcc_assert (single_pred_p (exit_bb) || LOOP_VINFO_EARLY_BREAKS (loop_vinfo)); tree vec_lhs_phi = copy_ssa_name (vec_lhs); gimple *phi = create_phi_node (vec_lhs_phi, exit_bb); for (unsigned i = 0; i < gimple_phi_num_args (phi); i++) SET_PHI_ARG_DEF (phi, i, vec_lhs); gimple_seq stmts = NULL; tree new_tree; /* If bitstart is 0 then we can use a BIT_FIELD_REF */ if (integer_zerop (bitstart)) { tree scalar_res = gimple_build (&stmts, BIT_FIELD_REF, TREE_TYPE (vectype), vec_lhs_phi, bitsize, bitstart); /* Convert the extracted vector element to the scalar type. */ new_tree = gimple_convert (&stmts, lhs_type, scalar_res); } else if (LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo)) { /* Emit: SCALAR_RES = VEC_EXTRACT <VEC_LHS, LEN + BIAS - 1> where VEC_LHS is the vectorized live-out result and MASK is the loop mask for the final iteration. */ gcc_assert (SLP_TREE_LANES (slp_node) == 1); gimple_seq tem = NULL; gimple_stmt_iterator gsi = gsi_last (tem); tree len = vect_get_loop_len (loop_vinfo, &gsi, &LOOP_VINFO_LENS (loop_vinfo), 1, vectype, 0, 1); gimple_seq_add_seq (&stmts, tem); /* BIAS - 1. */ signed char biasval = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); tree bias_minus_one = int_const_binop (MINUS_EXPR, build_int_cst (TREE_TYPE (len), biasval), build_one_cst (TREE_TYPE (len))); /* LAST_INDEX = LEN + (BIAS - 1). */ tree last_index = gimple_build (&stmts, PLUS_EXPR, TREE_TYPE (len), len, bias_minus_one); /* SCALAR_RES = VEC_EXTRACT <VEC_LHS, LEN + BIAS - 1>. */ tree scalar_res = gimple_build (&stmts, CFN_VEC_EXTRACT, TREE_TYPE (vectype), vec_lhs_phi, last_index); /* Convert the extracted vector element to the scalar type. */ new_tree = gimple_convert (&stmts, lhs_type, scalar_res); } else 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 (SLP_TREE_LANES (slp_node) == 1); tree scalar_type = TREE_TYPE (vectype); gimple_seq tem = NULL; gimple_stmt_iterator gsi = gsi_last (tem); tree mask = vect_get_loop_mask (loop_vinfo, &gsi, &LOOP_VINFO_MASKS (loop_vinfo), 1, vectype, 0); tree scalar_res; gimple_seq_add_seq (&stmts, tem); scalar_res = gimple_build (&stmts, CFN_EXTRACT_LAST, scalar_type, mask, vec_lhs_phi); /* Convert the extracted vector element to the 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_phi, bitsize, bitstart); new_tree = force_gimple_operand (fold_convert (lhs_type, new_tree), &stmts, true, NULL_TREE); } *exit_gsi = gsi_after_labels (exit_bb); if (stmts) gsi_insert_seq_before (exit_gsi, stmts, GSI_SAME_STMT); return new_tree; } /* Function vectorizable_live_operation. STMT_INFO computes a value that is used outside the loop. Check if it can be supported. */ bool vectorizable_live_operation (vec_info *vinfo, stmt_vec_info stmt_info, slp_tree slp_node, slp_instance slp_node_instance, int slp_index, bool vec_stmt_p, stmt_vector_for_cost *cost_vec) { loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo); imm_use_iterator imm_iter; tree lhs, lhs_type, bitsize; tree vectype = SLP_TREE_VECTYPE (slp_node); poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype); gimple *use_stmt; use_operand_p use_p; auto_vec<tree> vec_oprnds; int vec_entry = 0; poly_uint64 vec_index = 0; gcc_assert (STMT_VINFO_LIVE_P (stmt_info) || LOOP_VINFO_EARLY_BREAKS (loop_vinfo)); /* If a stmt of a reduction is live, vectorize it via vect_create_epilog_for_reduction. vectorizable_reduction assessed validity so just trigger the transform here. */ if (STMT_VINFO_REDUC_DEF (vect_orig_stmt (stmt_info))) { if (!vec_stmt_p) return true; /* For SLP reductions we vectorize the epilogue for all involved stmts together. For SLP reduction chains we only get here once. */ if (SLP_INSTANCE_KIND (slp_node_instance) == slp_inst_kind_reduc_group && slp_index != 0) return true; stmt_vec_info reduc_info = info_for_reduction (loop_vinfo, stmt_info); gcc_assert (reduc_info->is_reduc_info); if (STMT_VINFO_REDUC_TYPE (reduc_info) == FOLD_LEFT_REDUCTION || STMT_VINFO_REDUC_TYPE (reduc_info) == EXTRACT_LAST_REDUCTION) return true; if (!LOOP_VINFO_EARLY_BREAKS (loop_vinfo) || !LOOP_VINFO_EARLY_BREAKS_VECT_PEELED (loop_vinfo)) vect_create_epilog_for_reduction (loop_vinfo, stmt_info, slp_node, slp_node_instance, LOOP_VINFO_IV_EXIT (loop_vinfo)); /* If early break we only have to materialize the reduction on the merge block, but we have to find an alternate exit first. */ if (LOOP_VINFO_EARLY_BREAKS (loop_vinfo)) { slp_tree phis_node = slp_node_instance->reduc_phis; stmt_info = SLP_TREE_REPRESENTATIVE (phis_node); for (auto exit : get_loop_exit_edges (LOOP_VINFO_LOOP (loop_vinfo))) if (exit != LOOP_VINFO_IV_EXIT (loop_vinfo)) { vect_create_epilog_for_reduction (loop_vinfo, stmt_info, phis_node, slp_node_instance, exit); break; } if (LOOP_VINFO_EARLY_BREAKS_VECT_PEELED (loop_vinfo)) vect_create_epilog_for_reduction (loop_vinfo, stmt_info, phis_node, slp_node_instance, LOOP_VINFO_IV_EXIT (loop_vinfo)); } return true; } /* If STMT is not relevant and it is a simple assignment and its inputs are invariant then it can remain in place, unvectorized. The original last scalar value that it computes will be used. */ if (!STMT_VINFO_RELEVANT_P (stmt_info)) { 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 " "place.\n"); return true; } gcc_assert (slp_index >= 0); /* 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 num_scalar = SLP_TREE_LANES (slp_node); int num_vec = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); 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_p) { /* No transformation required. */ if (loop_vinfo && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)) { if (SLP_TREE_LANES (slp_node) != 1) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "can't operate on partial vectors " "because an SLP statement is live after " "the loop.\n"); LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; } else if (SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node) > 1) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "can't operate on partial vectors " "because ncopies is greater than 1.\n"); LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; } else { if (direct_internal_fn_supported_p (IFN_EXTRACT_LAST, vectype, OPTIMIZE_FOR_SPEED)) vect_record_loop_mask (loop_vinfo, &LOOP_VINFO_MASKS (loop_vinfo), 1, vectype, NULL); else if (can_vec_extract_var_idx_p ( TYPE_MODE (vectype), TYPE_MODE (TREE_TYPE (vectype)))) vect_record_loop_len (loop_vinfo, &LOOP_VINFO_LENS (loop_vinfo), 1, vectype, 1); else { if (dump_enabled_p ()) dump_printf_loc ( MSG_MISSED_OPTIMIZATION, vect_location, "can't operate on partial vectors " "because the target doesn't support extract " "last reduction.\n"); LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false; } } } /* ??? Enable for loop costing as well. */ if (!loop_vinfo) record_stmt_cost (cost_vec, 1, vec_to_scalar, slp_node, 0, vect_epilogue); return true; } /* Use the lhs of the original scalar statement. */ gimple *stmt = vect_orig_stmt (stmt_info)->stmt; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "extracting lane for live " "stmt %G", stmt); lhs = gimple_get_lhs (stmt); lhs_type = TREE_TYPE (lhs); bitsize = vector_element_bits_tree (vectype); /* Get the vectorized lhs of STMT and the lane to use (counted in bits). */ gcc_assert (!loop_vinfo || ((!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) && !LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo)) || SLP_TREE_LANES (slp_node) == 1)); /* Get the correct slp vectorized stmt. */ tree vec_lhs = SLP_TREE_VEC_DEFS (slp_node)[vec_entry]; gimple *vec_stmt = SSA_NAME_DEF_STMT (vec_lhs); /* In case we need to early break vectorize also get the first stmt. */ tree vec_lhs0 = SLP_TREE_VEC_DEFS (slp_node)[0]; /* Get entry to use. */ tree bitstart = bitsize_int (vec_index); bitstart = int_const_binop (MULT_EXPR, bitsize, bitstart); if (loop_vinfo) { /* Ensure the VEC_LHS for lane extraction stmts satisfy loop-closed PHI requirement, insert one phi node for it. It looks like: loop; BB: # lhs' = PHI <lhs> ==> loop; BB: # vec_lhs' = PHI <vec_lhs> new_tree = lane_extract <vec_lhs', ...>; lhs' = new_tree; */ class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); /* Check if we have a loop where the chosen exit is not the main exit, in these cases for an early break we restart the iteration the vector code did. For the live values we want the value at the start of the iteration rather than at the end. */ edge main_e = LOOP_VINFO_IV_EXIT (loop_vinfo); bool all_exits_as_early_p = LOOP_VINFO_EARLY_BREAKS_VECT_PEELED (loop_vinfo); FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, lhs) if (!is_gimple_debug (use_stmt) && !flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))) FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) { edge e = gimple_phi_arg_edge (as_a <gphi *> (use_stmt), phi_arg_index_from_use (use_p)); gcc_assert (loop_exit_edge_p (loop, e)); bool main_exit_edge = e == main_e; tree tmp_vec_lhs = vec_lhs; tree tmp_bitstart = bitstart; /* For early exit where the exit is not in the BB that leads to the latch then we're restarting the iteration in the scalar loop. So get the first live value. */ bool early_break_first_element_p = (all_exits_as_early_p || !main_exit_edge) && STMT_VINFO_DEF_TYPE (stmt_info) == vect_induction_def; if (early_break_first_element_p) { tmp_vec_lhs = vec_lhs0; tmp_bitstart = build_zero_cst (TREE_TYPE (bitstart)); } gimple_stmt_iterator exit_gsi; tree new_tree = vectorizable_live_operation_1 (loop_vinfo, e->dest, vectype, slp_node, bitsize, tmp_bitstart, tmp_vec_lhs, lhs_type, &exit_gsi); auto gsi = gsi_for_stmt (use_stmt); if (early_break_first_element_p && LOOP_VINFO_MASK_NITERS_PFA_OFFSET (loop_vinfo)) { tree step_expr = STMT_VINFO_LOOP_PHI_EVOLUTION_PART (stmt_info); tree break_lhs_phi = LOOP_VINFO_MASK_NITERS_PFA_OFFSET (loop_vinfo); tree ty_skip_niters = TREE_TYPE (break_lhs_phi); gimple_seq iv_stmts = NULL; /* Now create the PHI for the outside loop usage to retrieve the value for the offset counter. */ tree rphi_step = gimple_convert (&iv_stmts, ty_skip_niters, step_expr); tree tmp2 = gimple_build (&iv_stmts, MULT_EXPR, ty_skip_niters, rphi_step, break_lhs_phi); if (POINTER_TYPE_P (TREE_TYPE (new_tree))) { tmp2 = gimple_convert (&iv_stmts, sizetype, tmp2); tmp2 = gimple_build (&iv_stmts, POINTER_PLUS_EXPR, TREE_TYPE (new_tree), new_tree, tmp2); } else { tmp2 = gimple_convert (&iv_stmts, TREE_TYPE (new_tree), tmp2); tmp2 = gimple_build (&iv_stmts, PLUS_EXPR, TREE_TYPE (new_tree), new_tree, tmp2); } new_tree = tmp2; gsi_insert_seq_before (&exit_gsi, iv_stmts, GSI_SAME_STMT); } tree lhs_phi = gimple_phi_result (use_stmt); remove_phi_node (&gsi, false); gimple *copy = gimple_build_assign (lhs_phi, new_tree); gsi_insert_before (&exit_gsi, copy, GSI_SAME_STMT); break; } /* There a no further out-of-loop uses of lhs by LC-SSA construction. */ FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, lhs) gcc_assert (is_gimple_debug (use_stmt) || flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))); } else { /* For basic-block vectorization simply insert the lane-extraction. */ 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); gimple_seq stmts = NULL; new_tree = force_gimple_operand (fold_convert (lhs_type, new_tree), &stmts, true, NULL_TREE); if (TREE_CODE (new_tree) == SSA_NAME && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs)) SSA_NAME_OCCURS_IN_ABNORMAL_PHI (new_tree) = 1; if (is_a <gphi *> (vec_stmt)) { gimple_stmt_iterator si = gsi_after_labels (gimple_bb (vec_stmt)); gsi_insert_seq_before (&si, stmts, GSI_SAME_STMT); } else { gimple_stmt_iterator si = gsi_for_stmt (vec_stmt); gsi_insert_seq_after (&si, stmts, GSI_SAME_STMT); } /* Replace use of lhs with newly computed result. If the use stmt is a single arg PHI, just replace all uses of PHI result. It's necessary because lcssa PHI defining lhs may be before newly inserted stmt. */ use_operand_p use_p; stmt_vec_info use_stmt_info; FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, lhs) if (!is_gimple_debug (use_stmt) && (!(use_stmt_info = vinfo->lookup_stmt (use_stmt)) || !PURE_SLP_STMT (vect_stmt_to_vectorize (use_stmt_info)))) { /* ??? This can happen when the live lane ends up being rooted in a vector construction code-generated by an external SLP node (and code-generation for that already happened). See gcc.dg/vect/bb-slp-47.c. Doing this is what would happen if that vector CTOR were not code-generated yet so it is not too bad. ??? In fact we'd likely want to avoid this situation in the first place. */ if (TREE_CODE (new_tree) == SSA_NAME && !SSA_NAME_IS_DEFAULT_DEF (new_tree) && gimple_code (use_stmt) != GIMPLE_PHI && !vect_stmt_dominates_stmt_p (SSA_NAME_DEF_STMT (new_tree), use_stmt)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "Using original scalar computation for " "live lane because use preceeds vector " "def\n"); continue; } /* ??? It can also happen that we end up pulling a def into a loop where replacing out-of-loop uses would require a new LC SSA PHI node. Retain the original scalar in those cases as well. PR98064. */ if (TREE_CODE (new_tree) == SSA_NAME && !SSA_NAME_IS_DEFAULT_DEF (new_tree) && (gimple_bb (use_stmt)->loop_father != gimple_bb (vec_stmt)->loop_father) && !flow_loop_nested_p (gimple_bb (vec_stmt)->loop_father, gimple_bb (use_stmt)->loop_father)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "Using original scalar computation for " "live lane because there is an out-of-loop " "definition for it\n"); continue; } FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) SET_USE (use_p, new_tree); update_stmt (use_stmt); } } return true; } /* Given loop represented by LOOP_VINFO, return true if computation of LOOP_VINFO_NITERS (= LOOP_VINFO_NITERSM1 + 1) doesn't overflow, false otherwise. */ static bool loop_niters_no_overflow (loop_vec_info loop_vinfo) { /* Constant case. */ if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)) { tree cst_niters = LOOP_VINFO_NITERS (loop_vinfo); tree cst_nitersm1 = LOOP_VINFO_NITERSM1 (loop_vinfo); gcc_assert (TREE_CODE (cst_niters) == INTEGER_CST); gcc_assert (TREE_CODE (cst_nitersm1) == INTEGER_CST); if (wi::to_widest (cst_nitersm1) < wi::to_widest (cst_niters)) return true; } widest_int max; class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); /* Check the upper bound of loop niters. */ if (get_max_loop_iterations (loop, &max)) { tree type = TREE_TYPE (LOOP_VINFO_NITERS (loop_vinfo)); signop sgn = TYPE_SIGN (type); widest_int type_max = widest_int::from (wi::max_value (type), sgn); if (max < type_max) return true; } return false; } /* Return a mask type with half the number of elements as OLD_TYPE, given that it should have mode NEW_MODE. */ tree vect_halve_mask_nunits (tree old_type, machine_mode new_mode) { poly_uint64 nunits = exact_div (TYPE_VECTOR_SUBPARTS (old_type), 2); return build_truth_vector_type_for_mode (nunits, new_mode); } /* Return a mask type with twice as many elements as OLD_TYPE, given that it should have mode NEW_MODE. */ tree vect_double_mask_nunits (tree old_type, machine_mode new_mode) { poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (old_type) * 2; return build_truth_vector_type_for_mode (nunits, new_mode); } /* 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. If SCALAR_MASK is nonnull, the fully-masked loop would AND these vector masks with the vector version of SCALAR_MASK. */ void vect_record_loop_mask (loop_vec_info loop_vinfo, vec_loop_masks *masks, unsigned int nvectors, tree vectype, tree scalar_mask) { gcc_assert (nvectors != 0); if (scalar_mask) { scalar_cond_masked_key cond (scalar_mask, nvectors); loop_vinfo->scalar_cond_masked_set.add (cond); } masks->mask_set.add (std::make_pair (vectype, nvectors)); } /* 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 (loop_vec_info loop_vinfo, gimple_stmt_iterator *gsi, vec_loop_masks *masks, unsigned int nvectors, tree vectype, unsigned int index) { if (LOOP_VINFO_PARTIAL_VECTORS_STYLE (loop_vinfo) == vect_partial_vectors_while_ult) { rgroup_controls *rgm = &(masks->rgc_vec)[nvectors - 1]; tree mask_type = rgm->type; /* Populate the rgroup's mask array, if this is the first time we've used it. */ if (rgm->controls.is_empty ()) { rgm->controls.safe_grow_cleared (nvectors, true); 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->controls[i] = mask; } } tree mask = rgm->controls[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 = truth_type_for (vectype); mask = gimple_build (&seq, VIEW_CONVERT_EXPR, mask_type, mask); if (seq) gsi_insert_seq_before (gsi, seq, GSI_SAME_STMT); } return mask; } else if (LOOP_VINFO_PARTIAL_VECTORS_STYLE (loop_vinfo) == vect_partial_vectors_avx512) { /* 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 (); rgroup_controls *rgm = &masks->rgc_vec[nscalars_per_iter - 1]; /* The stored nV is dependent on the mask type produced. */ gcc_assert (exact_div (nvectors * TYPE_VECTOR_SUBPARTS (vectype), TYPE_VECTOR_SUBPARTS (rgm->type)).to_constant () == rgm->factor); nvectors = rgm->factor; /* Populate the rgroup's mask array, if this is the first time we've used it. */ if (rgm->controls.is_empty ()) { rgm->controls.safe_grow_cleared (nvectors, true); for (unsigned int i = 0; i < nvectors; ++i) { tree mask = make_temp_ssa_name (rgm->type, NULL, "loop_mask"); /* Provide a dummy definition until the real one is available. */ SSA_NAME_DEF_STMT (mask) = gimple_build_nop (); rgm->controls[i] = mask; } } if (known_eq (TYPE_VECTOR_SUBPARTS (rgm->type), TYPE_VECTOR_SUBPARTS (vectype))) return rgm->controls[index]; /* Split the vector if needed. Since we are dealing with integer mode masks with AVX512 we can operate on the integer representation performing the whole vector shifting. */ unsigned HOST_WIDE_INT factor; bool ok = constant_multiple_p (TYPE_VECTOR_SUBPARTS (rgm->type), TYPE_VECTOR_SUBPARTS (vectype), &factor); gcc_assert (ok); gcc_assert (GET_MODE_CLASS (TYPE_MODE (rgm->type)) == MODE_INT); tree mask_type = truth_type_for (vectype); gcc_assert (GET_MODE_CLASS (TYPE_MODE (mask_type)) == MODE_INT); unsigned vi = index / factor; unsigned vpart = index % factor; tree vec = rgm->controls[vi]; gimple_seq seq = NULL; vec = gimple_build (&seq, VIEW_CONVERT_EXPR, lang_hooks.types.type_for_mode (TYPE_MODE (rgm->type), 1), vec); /* For integer mode masks simply shift the right bits into position. */ if (vpart != 0) vec = gimple_build (&seq, RSHIFT_EXPR, TREE_TYPE (vec), vec, build_int_cst (integer_type_node, (TYPE_VECTOR_SUBPARTS (vectype) * vpart))); vec = gimple_convert (&seq, lang_hooks.types.type_for_mode (TYPE_MODE (mask_type), 1), vec); vec = gimple_build (&seq, VIEW_CONVERT_EXPR, mask_type, vec); if (seq) gsi_insert_seq_before (gsi, seq, GSI_SAME_STMT); return vec; } else gcc_unreachable (); } /* Record that LOOP_VINFO would need LENS to contain a sequence of NVECTORS lengths for controlling an operation on VECTYPE. The operation splits each element of VECTYPE into FACTOR separate subelements, measuring the length as a number of these subelements. */ void vect_record_loop_len (loop_vec_info loop_vinfo, vec_loop_lens *lens, unsigned int nvectors, tree vectype, unsigned int factor) { gcc_assert (nvectors != 0); if (lens->length () < nvectors) lens->safe_grow_cleared (nvectors, true); rgroup_controls *rgl = &(*lens)[nvectors - 1]; /* The number of scalars per iteration, scalar occupied bytes 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 (rgl->max_nscalars_per_iter < nscalars_per_iter) { /* For now, we only support cases in which all loads and stores fall back to VnQI or none do. */ gcc_assert (!rgl->max_nscalars_per_iter || (rgl->factor == 1 && factor == 1) || (rgl->max_nscalars_per_iter * rgl->factor == nscalars_per_iter * factor)); rgl->max_nscalars_per_iter = nscalars_per_iter; rgl->type = vectype; rgl->factor = factor; } } /* Given a complete set of lengths LENS, extract length number INDEX for an rgroup that operates on NVECTORS vectors of type VECTYPE, where 0 <= INDEX < NVECTORS. Return a value that contains FACTOR multipled by the number of elements that should be processed. Insert any set-up statements before GSI. */ tree vect_get_loop_len (loop_vec_info loop_vinfo, gimple_stmt_iterator *gsi, vec_loop_lens *lens, unsigned int nvectors, tree vectype, unsigned int index, unsigned int factor) { rgroup_controls *rgl = &(*lens)[nvectors - 1]; bool use_bias_adjusted_len = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo) != 0; /* Populate the rgroup's len array, if this is the first time we've used it. */ if (rgl->controls.is_empty ()) { rgl->controls.safe_grow_cleared (nvectors, true); for (unsigned int i = 0; i < nvectors; ++i) { tree len_type = LOOP_VINFO_RGROUP_COMPARE_TYPE (loop_vinfo); gcc_assert (len_type != NULL_TREE); tree len = make_temp_ssa_name (len_type, NULL, "loop_len"); /* Provide a dummy definition until the real one is available. */ SSA_NAME_DEF_STMT (len) = gimple_build_nop (); rgl->controls[i] = len; if (use_bias_adjusted_len) { gcc_assert (i == 0); tree adjusted_len = make_temp_ssa_name (len_type, NULL, "adjusted_loop_len"); SSA_NAME_DEF_STMT (adjusted_len) = gimple_build_nop (); rgl->bias_adjusted_ctrl = adjusted_len; } } } if (use_bias_adjusted_len) return rgl->bias_adjusted_ctrl; tree loop_len = rgl->controls[index]; if (rgl->factor == 1 && factor == 1) { poly_int64 nunits1 = TYPE_VECTOR_SUBPARTS (rgl->type); poly_int64 nunits2 = TYPE_VECTOR_SUBPARTS (vectype); if (maybe_ne (nunits1, nunits2)) { /* A loop len 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. */ gcc_assert (multiple_p (nunits1, nunits2)); factor = exact_div (nunits1, nunits2).to_constant (); tree iv_type = LOOP_VINFO_RGROUP_IV_TYPE (loop_vinfo); gimple_seq seq = NULL; loop_len = gimple_build (&seq, EXACT_DIV_EXPR, iv_type, loop_len, build_int_cst (iv_type, factor)); if (seq) gsi_insert_seq_before (gsi, seq, GSI_SAME_STMT); } } return loop_len; } /* Generate the tree for the loop len mask and return it. Given the lens, nvectors, vectype, index and factor to gen the len mask as below. tree len_mask = VCOND_MASK_LEN (compare_mask, ones, zero, len, bias) */ tree vect_gen_loop_len_mask (loop_vec_info loop_vinfo, gimple_stmt_iterator *gsi, gimple_stmt_iterator *cond_gsi, vec_loop_lens *lens, unsigned int nvectors, tree vectype, tree stmt, unsigned int index, unsigned int factor) { tree all_one_mask = build_all_ones_cst (vectype); tree all_zero_mask = build_zero_cst (vectype); tree len = vect_get_loop_len (loop_vinfo, gsi, lens, nvectors, vectype, index, factor); tree bias = build_int_cst (intQI_type_node, LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo)); tree len_mask = make_temp_ssa_name (TREE_TYPE (stmt), NULL, "vec_len_mask"); gcall *call = gimple_build_call_internal (IFN_VCOND_MASK_LEN, 5, stmt, all_one_mask, all_zero_mask, len, bias); gimple_call_set_lhs (call, len_mask); gsi_insert_before (cond_gsi, call, GSI_SAME_STMT); return len_mask; } /* Scale profiling counters by estimation for LOOP which is vectorized by factor VF. If FLAT is true, the loop we started with had unrealistically flat profile. */ static void scale_profile_for_vect_loop (class loop *loop, edge exit_e, unsigned vf, bool flat) { /* For flat profiles do not scale down proportionally by VF and only cap by known iteration count bounds. */ if (flat) { if (dump_file && (dump_flags & TDF_DETAILS)) fprintf (dump_file, "Vectorized loop profile seems flat; not scaling iteration " "count down by the vectorization factor %i\n", vf); scale_loop_profile (loop, profile_probability::always (), get_likely_max_loop_iterations_int (loop)); return; } /* Loop body executes VF fewer times and exit increases VF times. */ profile_count entry_count = loop_preheader_edge (loop)->count (); /* If we have unreliable loop profile avoid dropping entry count below header count. This can happen since loops has unrealistically low trip counts. */ while (vf > 1 && loop->header->count > entry_count && loop->header->count < entry_count * vf) { if (dump_file && (dump_flags & TDF_DETAILS)) fprintf (dump_file, "Vectorization factor %i seems too large for profile " "prevoiusly believed to be consistent; reducing.\n", vf); vf /= 2; } if (entry_count.nonzero_p ()) set_edge_probability_and_rescale_others (exit_e, entry_count.probability_in (loop->header->count / vf)); /* Avoid producing very large exit probability when we do not have sensible profile. */ else if (exit_e->probability < profile_probability::always () / (vf * 2)) set_edge_probability_and_rescale_others (exit_e, exit_e->probability * vf); loop->latch->count = single_pred_edge (loop->latch)->count (); scale_loop_profile (loop, profile_probability::always () / vf, get_likely_max_loop_iterations_int (loop)); } /* Update EPILOGUE's loop_vec_info. EPILOGUE was constructed as a copy of the original loop that has now been vectorized. The inits of the data_references need to be advanced with the number of iterations of the main loop. This has been computed in vect_do_peeling and is stored in parameter ADVANCE. Since the loop_vec_info of this EPILOGUE was constructed for the original loop, its stmt_vec_infos all point to the original statements. These need to be updated to point to their corresponding copies. The data_reference's connections also need to be updated. Their corresponding dr_vec_info need to be reconnected to the EPILOGUE's stmt_vec_infos, their statements need to point to their corresponding copy. */ static void update_epilogue_loop_vinfo (class loop *epilogue, tree advance) { loop_vec_info epilogue_vinfo = loop_vec_info_for_loop (epilogue); hash_map<tree,tree> mapping; gimple *orig_stmt, *new_stmt; gimple_stmt_iterator epilogue_gsi; gphi_iterator epilogue_phi_gsi; stmt_vec_info stmt_vinfo = NULL, related_vinfo; basic_block *epilogue_bbs = get_loop_body (epilogue); unsigned i; free (LOOP_VINFO_BBS (epilogue_vinfo)); LOOP_VINFO_BBS (epilogue_vinfo) = epilogue_bbs; LOOP_VINFO_NBBS (epilogue_vinfo) = epilogue->num_nodes; /* The EPILOGUE loop is a copy of the original loop so they share the same gimple UIDs. In this loop we update the loop_vec_info of the EPILOGUE to point to the copied statements. */ for (unsigned i = 0; i < epilogue->num_nodes; ++i) { for (epilogue_phi_gsi = gsi_start_phis (epilogue_bbs[i]); !gsi_end_p (epilogue_phi_gsi); gsi_next (&epilogue_phi_gsi)) { new_stmt = epilogue_phi_gsi.phi (); gcc_assert (gimple_uid (new_stmt) > 0); stmt_vinfo = epilogue_vinfo->stmt_vec_infos[gimple_uid (new_stmt) - 1]; STMT_VINFO_STMT (stmt_vinfo) = new_stmt; } for (epilogue_gsi = gsi_start_bb (epilogue_bbs[i]); !gsi_end_p (epilogue_gsi); gsi_next (&epilogue_gsi)) { new_stmt = gsi_stmt (epilogue_gsi); if (is_gimple_debug (new_stmt)) continue; gcc_assert (gimple_uid (new_stmt) > 0); stmt_vinfo = epilogue_vinfo->stmt_vec_infos[gimple_uid (new_stmt) - 1]; STMT_VINFO_STMT (stmt_vinfo) = new_stmt; related_vinfo = STMT_VINFO_RELATED_STMT (stmt_vinfo); if (related_vinfo != NULL && related_vinfo != stmt_vinfo) { gimple *stmt = STMT_VINFO_STMT (related_vinfo); /* Set BB such that the assert in 'get_initial_defs_for_reduction' is able to determine that the BB of the related stmt is inside this loop. */ gimple_set_bb (stmt, gimple_bb (new_stmt)); related_vinfo = STMT_VINFO_RELATED_STMT (related_vinfo); gcc_assert (related_vinfo == NULL || related_vinfo == stmt_vinfo); } } } struct data_reference *dr; vec<data_reference_p> datarefs = LOOP_VINFO_DATAREFS (epilogue_vinfo); FOR_EACH_VEC_ELT (datarefs, i, dr) { orig_stmt = DR_STMT (dr); gcc_assert (gimple_uid (orig_stmt) > 0); stmt_vinfo = epilogue_vinfo->stmt_vec_infos[gimple_uid (orig_stmt) - 1]; DR_STMT (dr) = STMT_VINFO_STMT (stmt_vinfo); } /* Advance data_reference's with the number of iterations of the previous loop and its prologue. */ vect_update_inits_of_drs (epilogue_vinfo, advance, PLUS_EXPR); /* Remember the advancement made. */ LOOP_VINFO_DRS_ADVANCED_BY (epilogue_vinfo) = advance; } /* When vectorizing early break statements instructions that happen before the early break in the current BB need to be moved to after the early break. This function deals with that and assumes that any validity checks has already been performed. While moving the instructions if it encounters a VUSE or VDEF it then corrects the VUSES as it moves the statements along. GDEST is the location in which to insert the new statements. */ static void move_early_exit_stmts (loop_vec_info loop_vinfo) { DUMP_VECT_SCOPE ("move_early_exit_stmts"); if (LOOP_VINFO_EARLY_BRK_STORES (loop_vinfo).is_empty ()) return; /* Move all stmts that need moving. */ basic_block dest_bb = LOOP_VINFO_EARLY_BRK_DEST_BB (loop_vinfo); gimple_stmt_iterator dest_gsi = gsi_after_labels (dest_bb);


context:
space:
mode: