/* Data References Analysis and Manipulation Utilities for Vectorization. Copyright (C) 2003-2014 Free Software Foundation, Inc. Contributed by Dorit Naishlos and Ira Rosen 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 . */ #include "config.h" #include "system.h" #include "coretypes.h" #include "dumpfile.h" #include "tm.h" #include "tree.h" #include "stor-layout.h" #include "tm_p.h" #include "target.h" #include "basic-block.h" #include "gimple-pretty-print.h" #include "tree-ssa-alias.h" #include "internal-fn.h" #include "tree-eh.h" #include "gimple-expr.h" #include "is-a.h" #include "gimple.h" #include "gimplify.h" #include "gimple-iterator.h" #include "gimplify-me.h" #include "gimple-ssa.h" #include "tree-phinodes.h" #include "ssa-iterators.h" #include "stringpool.h" #include "tree-ssanames.h" #include "tree-ssa-loop-ivopts.h" #include "tree-ssa-loop-manip.h" #include "tree-ssa-loop.h" #include "dumpfile.h" #include "cfgloop.h" #include "tree-chrec.h" #include "tree-scalar-evolution.h" #include "tree-vectorizer.h" #include "diagnostic-core.h" #include "cgraph.h" /* Need to include rtl.h, expr.h, etc. for optabs. */ #include "expr.h" #include "optabs.h" /* Return true if load- or store-lanes optab OPTAB is implemented for COUNT vectors of type VECTYPE. NAME is the name of OPTAB. */ static bool vect_lanes_optab_supported_p (const char *name, convert_optab optab, tree vectype, unsigned HOST_WIDE_INT count) { enum machine_mode mode, array_mode; bool limit_p; mode = TYPE_MODE (vectype); limit_p = !targetm.array_mode_supported_p (mode, count); array_mode = mode_for_size (count * GET_MODE_BITSIZE (mode), MODE_INT, limit_p); if (array_mode == BLKmode) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "no array mode for %s[" HOST_WIDE_INT_PRINT_DEC "]\n", GET_MODE_NAME (mode), count); return false; } if (convert_optab_handler (optab, array_mode, mode) == CODE_FOR_nothing) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "cannot use %s<%s><%s>\n", name, GET_MODE_NAME (array_mode), GET_MODE_NAME (mode)); return false; } if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "can use %s<%s><%s>\n", name, GET_MODE_NAME (array_mode), GET_MODE_NAME (mode)); return true; } /* Return the smallest scalar part of STMT. This is used to determine the vectype of the stmt. We generally set the vectype according to the type of the result (lhs). For stmts whose result-type is different than the type of the arguments (e.g., demotion, promotion), vectype will be reset appropriately (later). Note that we have to visit the smallest datatype in this function, because that determines the VF. If the smallest datatype in the loop is present only as the rhs of a promotion operation - we'd miss it. Such a case, where a variable of this datatype does not appear in the lhs anywhere in the loop, can only occur if it's an invariant: e.g.: 'int_x = (int) short_inv', which we'd expect to have been optimized away by invariant motion. However, we cannot rely on invariant motion to always take invariants out of the loop, and so in the case of promotion we also have to check the rhs. LHS_SIZE_UNIT and RHS_SIZE_UNIT contain the sizes of the corresponding types. */ tree vect_get_smallest_scalar_type (gimple stmt, HOST_WIDE_INT *lhs_size_unit, HOST_WIDE_INT *rhs_size_unit) { tree scalar_type = gimple_expr_type (stmt); HOST_WIDE_INT lhs, rhs; lhs = rhs = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (scalar_type)); if (is_gimple_assign (stmt) && (gimple_assign_cast_p (stmt) || gimple_assign_rhs_code (stmt) == WIDEN_MULT_EXPR || gimple_assign_rhs_code (stmt) == WIDEN_LSHIFT_EXPR || gimple_assign_rhs_code (stmt) == FLOAT_EXPR)) { tree rhs_type = TREE_TYPE (gimple_assign_rhs1 (stmt)); rhs = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (rhs_type)); if (rhs < lhs) scalar_type = rhs_type; } *lhs_size_unit = lhs; *rhs_size_unit = rhs; return scalar_type; } /* Insert DDR into LOOP_VINFO list of ddrs that may alias and need to be tested at run-time. Return TRUE if DDR was successfully inserted. Return false if versioning is not supported. */ static bool vect_mark_for_runtime_alias_test (ddr_p ddr, loop_vec_info loop_vinfo) { struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); if ((unsigned) PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIAS_CHECKS) == 0) return false; if (dump_enabled_p ()) { dump_printf_loc (MSG_NOTE, vect_location, "mark for run-time aliasing test between "); dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (DDR_A (ddr))); dump_printf (MSG_NOTE, " and "); dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (DDR_B (ddr))); dump_printf (MSG_NOTE, "\n"); } if (optimize_loop_nest_for_size_p (loop)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "versioning not supported when optimizing" " for size.\n"); return false; } /* FORNOW: We don't support versioning with outer-loop vectorization. */ if (loop->inner) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "versioning not yet supported for outer-loops.\n"); return false; } /* FORNOW: We don't support creating runtime alias tests for non-constant step. */ if (TREE_CODE (DR_STEP (DDR_A (ddr))) != INTEGER_CST || TREE_CODE (DR_STEP (DDR_B (ddr))) != INTEGER_CST) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "versioning not yet supported for non-constant " "step\n"); return false; } LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo).safe_push (ddr); return true; } /* Function vect_analyze_data_ref_dependence. Return TRUE if there (might) exist a dependence between a memory-reference DRA and a memory-reference DRB. When versioning for alias may check a dependence at run-time, return FALSE. Adjust *MAX_VF according to the data dependence. */ static bool vect_analyze_data_ref_dependence (struct data_dependence_relation *ddr, loop_vec_info loop_vinfo, int *max_vf) { unsigned int i; struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); struct data_reference *dra = DDR_A (ddr); struct data_reference *drb = DDR_B (ddr); stmt_vec_info stmtinfo_a = vinfo_for_stmt (DR_STMT (dra)); stmt_vec_info stmtinfo_b = vinfo_for_stmt (DR_STMT (drb)); lambda_vector dist_v; unsigned int loop_depth; /* In loop analysis all data references should be vectorizable. */ if (!STMT_VINFO_VECTORIZABLE (stmtinfo_a) || !STMT_VINFO_VECTORIZABLE (stmtinfo_b)) gcc_unreachable (); /* Independent data accesses. */ if (DDR_ARE_DEPENDENT (ddr) == chrec_known) return false; if (dra == drb || (DR_IS_READ (dra) && DR_IS_READ (drb))) return false; /* Unknown data dependence. */ if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) { /* If user asserted safelen consecutive iterations can be executed concurrently, assume independence. */ if (loop->safelen >= 2) { if (loop->safelen < *max_vf) *max_vf = loop->safelen; return false; } if (STMT_VINFO_GATHER_P (stmtinfo_a) || STMT_VINFO_GATHER_P (stmtinfo_b)) { if (dump_enabled_p ()) { dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "versioning for alias not supported for: " "can't determine dependence between "); dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, DR_REF (dra)); dump_printf (MSG_MISSED_OPTIMIZATION, " and "); dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, DR_REF (drb)); dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); } return true; } if (dump_enabled_p ()) { dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "versioning for alias required: " "can't determine dependence between "); dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, DR_REF (dra)); dump_printf (MSG_MISSED_OPTIMIZATION, " and "); dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, DR_REF (drb)); dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); } /* Add to list of ddrs that need to be tested at run-time. */ return !vect_mark_for_runtime_alias_test (ddr, loop_vinfo); } /* Known data dependence. */ if (DDR_NUM_DIST_VECTS (ddr) == 0) { /* If user asserted safelen consecutive iterations can be executed concurrently, assume independence. */ if (loop->safelen >= 2) { if (loop->safelen < *max_vf) *max_vf = loop->safelen; return false; } if (STMT_VINFO_GATHER_P (stmtinfo_a) || STMT_VINFO_GATHER_P (stmtinfo_b)) { if (dump_enabled_p ()) { dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "versioning for alias not supported for: " "bad dist vector for "); dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, DR_REF (dra)); dump_printf (MSG_MISSED_OPTIMIZATION, " and "); dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, DR_REF (drb)); dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); } return true; } if (dump_enabled_p ()) { dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "versioning for alias required: " "bad dist vector for "); dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, DR_REF (dra)); dump_printf (MSG_MISSED_OPTIMIZATION, " and "); dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, DR_REF (drb)); dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); } /* Add to list of ddrs that need to be tested at run-time. */ return !vect_mark_for_runtime_alias_test (ddr, loop_vinfo); } loop_depth = index_in_loop_nest (loop->num, DDR_LOOP_NEST (ddr)); FOR_EACH_VEC_ELT (DDR_DIST_VECTS (ddr), i, dist_v) { int dist = dist_v[loop_depth]; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "dependence distance = %d.\n", dist); if (dist == 0) { if (dump_enabled_p ()) { dump_printf_loc (MSG_NOTE, vect_location, "dependence distance == 0 between "); dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dra)); dump_printf (MSG_NOTE, " and "); dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (drb)); dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); } /* When we perform grouped accesses and perform implicit CSE by detecting equal accesses and doing disambiguation with runtime alias tests like for .. = a[i]; .. = a[i+1]; a[i] = ..; a[i+1] = ..; *p = ..; .. = a[i]; .. = a[i+1]; where we will end up loading { a[i], a[i+1] } once, make sure that inserting group loads before the first load and stores after the last store will do the right thing. */ if ((STMT_VINFO_GROUPED_ACCESS (stmtinfo_a) && GROUP_SAME_DR_STMT (stmtinfo_a)) || (STMT_VINFO_GROUPED_ACCESS (stmtinfo_b) && GROUP_SAME_DR_STMT (stmtinfo_b))) { gimple earlier_stmt; earlier_stmt = get_earlier_stmt (DR_STMT (dra), DR_STMT (drb)); if (DR_IS_WRITE (STMT_VINFO_DATA_REF (vinfo_for_stmt (earlier_stmt)))) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "READ_WRITE dependence in interleaving." "\n"); return true; } } continue; } if (dist > 0 && DDR_REVERSED_P (ddr)) { /* If DDR_REVERSED_P the order of the data-refs in DDR was reversed (to make distance vector positive), and the actual distance is negative. */ if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "dependence distance negative.\n"); continue; } if (abs (dist) >= 2 && abs (dist) < *max_vf) { /* The dependence distance requires reduction of the maximal vectorization factor. */ *max_vf = abs (dist); if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "adjusting maximal vectorization factor to %i\n", *max_vf); } if (abs (dist) >= *max_vf) { /* Dependence distance does not create dependence, as far as vectorization is concerned, in this case. */ if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "dependence distance >= VF.\n"); continue; } if (dump_enabled_p ()) { dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "not vectorized, possible dependence " "between data-refs "); dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dra)); dump_printf (MSG_NOTE, " and "); dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (drb)); dump_printf (MSG_NOTE, "\n"); } return true; } return false; } /* Function vect_analyze_data_ref_dependences. Examine all the data references in the loop, and make sure there do not exist any data dependences between them. Set *MAX_VF according to the maximum vectorization factor the data dependences allow. */ bool vect_analyze_data_ref_dependences (loop_vec_info loop_vinfo, int *max_vf) { unsigned int i; struct data_dependence_relation *ddr; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "=== vect_analyze_data_ref_dependences ===\n"); if (!compute_all_dependences (LOOP_VINFO_DATAREFS (loop_vinfo), &LOOP_VINFO_DDRS (loop_vinfo), LOOP_VINFO_LOOP_NEST (loop_vinfo), true)) return false; FOR_EACH_VEC_ELT (LOOP_VINFO_DDRS (loop_vinfo), i, ddr) if (vect_analyze_data_ref_dependence (ddr, loop_vinfo, max_vf)) return false; return true; } /* Function vect_slp_analyze_data_ref_dependence. Return TRUE if there (might) exist a dependence between a memory-reference DRA and a memory-reference DRB. When versioning for alias may check a dependence at run-time, return FALSE. Adjust *MAX_VF according to the data dependence. */ static bool vect_slp_analyze_data_ref_dependence (struct data_dependence_relation *ddr) { struct data_reference *dra = DDR_A (ddr); struct data_reference *drb = DDR_B (ddr); /* We need to check dependences of statements marked as unvectorizable as well, they still can prohibit vectorization. */ /* Independent data accesses. */ if (DDR_ARE_DEPENDENT (ddr) == chrec_known) return false; if (dra == drb) return false; /* Read-read is OK. */ if (DR_IS_READ (dra) && DR_IS_READ (drb)) return false; /* If dra and drb are part of the same interleaving chain consider them independent. */ if (STMT_VINFO_GROUPED_ACCESS (vinfo_for_stmt (DR_STMT (dra))) && (GROUP_FIRST_ELEMENT (vinfo_for_stmt (DR_STMT (dra))) == GROUP_FIRST_ELEMENT (vinfo_for_stmt (DR_STMT (drb))))) return false; /* Unknown data dependence. */ if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) { if (dump_enabled_p ()) { dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "can't determine dependence between "); dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, DR_REF (dra)); dump_printf (MSG_MISSED_OPTIMIZATION, " and "); dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, DR_REF (drb)); dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); } } else if (dump_enabled_p ()) { dump_printf_loc (MSG_NOTE, vect_location, "determined dependence between "); dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dra)); dump_printf (MSG_NOTE, " and "); dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (drb)); dump_printf (MSG_NOTE, "\n"); } /* We do not vectorize basic blocks with write-write dependencies. */ if (DR_IS_WRITE (dra) && DR_IS_WRITE (drb)) return true; /* If we have a read-write dependence check that the load is before the store. When we vectorize basic blocks, vector load can be only before corresponding scalar load, and vector store can be only after its corresponding scalar store. So the order of the acceses is preserved in case the load is before the store. */ gimple earlier_stmt = get_earlier_stmt (DR_STMT (dra), DR_STMT (drb)); if (DR_IS_READ (STMT_VINFO_DATA_REF (vinfo_for_stmt (earlier_stmt)))) { /* That only holds for load-store pairs taking part in vectorization. */ if (STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (DR_STMT (dra))) && STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (DR_STMT (drb)))) return false; } return true; } /* Function vect_analyze_data_ref_dependences. Examine all the data references in the basic-block, and make sure there do not exist any data dependences between them. Set *MAX_VF according to the maximum vectorization factor the data dependences allow. */ bool vect_slp_analyze_data_ref_dependences (bb_vec_info bb_vinfo) { struct data_dependence_relation *ddr; unsigned int i; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "=== vect_slp_analyze_data_ref_dependences ===\n"); if (!compute_all_dependences (BB_VINFO_DATAREFS (bb_vinfo), &BB_VINFO_DDRS (bb_vinfo), vNULL, true)) return false; FOR_EACH_VEC_ELT (BB_VINFO_DDRS (bb_vinfo), i, ddr) if (vect_slp_analyze_data_ref_dependence (ddr)) return false; return true; } /* Function vect_compute_data_ref_alignment Compute the misalignment of the data reference DR. Output: 1. If during the misalignment computation it is found that the data reference cannot be vectorized then false is returned. 2. DR_MISALIGNMENT (DR) is defined. FOR NOW: No analysis is actually performed. Misalignment is calculated only for trivial cases. TODO. */ static bool vect_compute_data_ref_alignment (struct data_reference *dr) { gimple stmt = DR_STMT (dr); stmt_vec_info stmt_info = vinfo_for_stmt (stmt); loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); struct loop *loop = NULL; tree ref = DR_REF (dr); tree vectype; tree base, base_addr; bool base_aligned; tree misalign; tree aligned_to, alignment; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "vect_compute_data_ref_alignment:\n"); if (loop_vinfo) loop = LOOP_VINFO_LOOP (loop_vinfo); /* Initialize misalignment to unknown. */ SET_DR_MISALIGNMENT (dr, -1); /* Strided loads perform only component accesses, misalignment information is irrelevant for them. */ if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) return true; misalign = DR_INIT (dr); aligned_to = DR_ALIGNED_TO (dr); base_addr = DR_BASE_ADDRESS (dr); vectype = STMT_VINFO_VECTYPE (stmt_info); /* In case the dataref is in an inner-loop of the loop that is being vectorized (LOOP), we use the base and misalignment information relative to the outer-loop (LOOP). This is ok only if the misalignment stays the same throughout the execution of the inner-loop, which is why we have to check that the stride of the dataref in the inner-loop evenly divides by the vector size. */ if (loop && nested_in_vect_loop_p (loop, stmt)) { tree step = DR_STEP (dr); HOST_WIDE_INT dr_step = TREE_INT_CST_LOW (step); if (dr_step % GET_MODE_SIZE (TYPE_MODE (vectype)) == 0) { if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "inner step divides the vector-size.\n"); misalign = STMT_VINFO_DR_INIT (stmt_info); aligned_to = STMT_VINFO_DR_ALIGNED_TO (stmt_info); base_addr = STMT_VINFO_DR_BASE_ADDRESS (stmt_info); } else { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "inner step doesn't divide the vector-size.\n"); misalign = NULL_TREE; } } /* Similarly, if we're doing basic-block vectorization, we can only use base and misalignment information relative to an innermost loop if the misalignment stays the same throughout the execution of the loop. As above, this is the case if the stride of the dataref evenly divides by the vector size. */ if (!loop) { tree step = DR_STEP (dr); HOST_WIDE_INT dr_step = TREE_INT_CST_LOW (step); if (dr_step % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "SLP: step doesn't divide the vector-size.\n"); misalign = NULL_TREE; } } base = build_fold_indirect_ref (base_addr); alignment = ssize_int (TYPE_ALIGN (vectype)/BITS_PER_UNIT); if ((aligned_to && tree_int_cst_compare (aligned_to, alignment) < 0) || !misalign) { if (dump_enabled_p ()) { dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "Unknown alignment for access: "); dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, base); dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); } return true; } if ((DECL_P (base) && tree_int_cst_compare (ssize_int (DECL_ALIGN_UNIT (base)), alignment) >= 0) || (TREE_CODE (base_addr) == SSA_NAME && tree_int_cst_compare (ssize_int (TYPE_ALIGN_UNIT (TREE_TYPE ( TREE_TYPE (base_addr)))), alignment) >= 0) || (get_pointer_alignment (base_addr) >= TYPE_ALIGN (vectype))) base_aligned = true; else base_aligned = false; if (!base_aligned) { /* Do not change the alignment of global variables here if flag_section_anchors is enabled as we already generated RTL for other functions. Most global variables should have been aligned during the IPA increase_alignment pass. */ if (!vect_can_force_dr_alignment_p (base, TYPE_ALIGN (vectype)) || (TREE_STATIC (base) && flag_section_anchors)) { if (dump_enabled_p ()) { dump_printf_loc (MSG_NOTE, vect_location, "can't force alignment of ref: "); dump_generic_expr (MSG_NOTE, TDF_SLIM, ref); dump_printf (MSG_NOTE, "\n"); } return true; } /* Force the alignment of the decl. NOTE: This is the only change to the code we make during the analysis phase, before deciding to vectorize the loop. */ if (dump_enabled_p ()) { dump_printf_loc (MSG_NOTE, vect_location, "force alignment of "); dump_generic_expr (MSG_NOTE, TDF_SLIM, ref); dump_printf (MSG_NOTE, "\n"); } ((dataref_aux *)dr->aux)->base_decl = base; ((dataref_aux *)dr->aux)->base_misaligned = true; } /* If this is a backward running DR then first access in the larger vectype actually is N-1 elements before the address in the DR. Adjust misalign accordingly. */ if (tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0) { tree offset = ssize_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); /* DR_STEP(dr) is the same as -TYPE_SIZE of the scalar type, otherwise we wouldn't be here. */ offset = fold_build2 (MULT_EXPR, ssizetype, offset, DR_STEP (dr)); /* PLUS because DR_STEP was negative. */ misalign = size_binop (PLUS_EXPR, misalign, offset); } /* Modulo alignment. */ misalign = size_binop (FLOOR_MOD_EXPR, misalign, alignment); if (!tree_fits_uhwi_p (misalign)) { /* Negative or overflowed misalignment value. */ if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "unexpected misalign value\n"); return false; } SET_DR_MISALIGNMENT (dr, tree_to_uhwi (misalign)); if (dump_enabled_p ()) { dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "misalign = %d bytes of ref ", DR_MISALIGNMENT (dr)); dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, ref); dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); } return true; } /* Function vect_compute_data_refs_alignment Compute the misalignment of data references in the loop. Return FALSE if a data reference is found that cannot be vectorized. */ static bool vect_compute_data_refs_alignment (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo) { vec datarefs; struct data_reference *dr; unsigned int i; if (loop_vinfo) datarefs = LOOP_VINFO_DATAREFS (loop_vinfo); else datarefs = BB_VINFO_DATAREFS (bb_vinfo); FOR_EACH_VEC_ELT (datarefs, i, dr) if (STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (DR_STMT (dr))) && !vect_compute_data_ref_alignment (dr)) { if (bb_vinfo) { /* Mark unsupported statement as unvectorizable. */ STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (DR_STMT (dr))) = false; continue; } else return false; } return true; } /* Function vect_update_misalignment_for_peel DR - the data reference whose misalignment is to be adjusted. DR_PEEL - the data reference whose misalignment is being made zero in the vector loop by the peel. NPEEL - the number of iterations in the peel loop if the misalignment of DR_PEEL is known at compile time. */ static void vect_update_misalignment_for_peel (struct data_reference *dr, struct data_reference *dr_peel, int npeel) { unsigned int i; vec same_align_drs; struct data_reference *current_dr; int dr_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dr)))); int dr_peel_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dr_peel)))); stmt_vec_info stmt_info = vinfo_for_stmt (DR_STMT (dr)); stmt_vec_info peel_stmt_info = vinfo_for_stmt (DR_STMT (dr_peel)); /* For interleaved data accesses the step in the loop must be multiplied by the size of the interleaving group. */ if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) dr_size *= GROUP_SIZE (vinfo_for_stmt (GROUP_FIRST_ELEMENT (stmt_info))); if (STMT_VINFO_GROUPED_ACCESS (peel_stmt_info)) dr_peel_size *= GROUP_SIZE (peel_stmt_info); /* It can be assumed that the data refs with the same alignment as dr_peel are aligned in the vector loop. */ same_align_drs = STMT_VINFO_SAME_ALIGN_REFS (vinfo_for_stmt (DR_STMT (dr_peel))); FOR_EACH_VEC_ELT (same_align_drs, i, current_dr) { if (current_dr != dr) continue; gcc_assert (DR_MISALIGNMENT (dr) / dr_size == DR_MISALIGNMENT (dr_peel) / dr_peel_size); SET_DR_MISALIGNMENT (dr, 0); return; } if (known_alignment_for_access_p (dr) && known_alignment_for_access_p (dr_peel)) { bool negative = tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0; int misal = DR_MISALIGNMENT (dr); tree vectype = STMT_VINFO_VECTYPE (stmt_info); misal += negative ? -npeel * dr_size : npeel * dr_size; misal &= (TYPE_ALIGN (vectype) / BITS_PER_UNIT) - 1; SET_DR_MISALIGNMENT (dr, misal); return; } if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "Setting misalignment to -1.\n"); SET_DR_MISALIGNMENT (dr, -1); } /* Function vect_verify_datarefs_alignment Return TRUE if all data references in the loop can be handled with respect to alignment. */ bool vect_verify_datarefs_alignment (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo) { vec datarefs; struct data_reference *dr; enum dr_alignment_support supportable_dr_alignment; unsigned int i; if (loop_vinfo) datarefs = LOOP_VINFO_DATAREFS (loop_vinfo); else datarefs = BB_VINFO_DATAREFS (bb_vinfo); FOR_EACH_VEC_ELT (datarefs, i, dr) { gimple stmt = DR_STMT (dr); stmt_vec_info stmt_info = vinfo_for_stmt (stmt); if (!STMT_VINFO_RELEVANT_P (stmt_info)) continue; /* For interleaving, only the alignment of the first access matters. Skip statements marked as not vectorizable. */ if ((STMT_VINFO_GROUPED_ACCESS (stmt_info) && GROUP_FIRST_ELEMENT (stmt_info) != stmt) || !STMT_VINFO_VECTORIZABLE (stmt_info)) continue; /* Strided loads perform only component accesses, alignment is irrelevant for them. */ if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) continue; supportable_dr_alignment = vect_supportable_dr_alignment (dr, false); if (!supportable_dr_alignment) { if (dump_enabled_p ()) { if (DR_IS_READ (dr)) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "not vectorized: unsupported unaligned load."); else dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "not vectorized: unsupported unaligned " "store."); dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, DR_REF (dr)); dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); } return false; } if (supportable_dr_alignment != dr_aligned && dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "Vectorizing an unaligned access.\n"); } return true; } /* Given an memory reference EXP return whether its alignment is less than its size. */ static bool not_size_aligned (tree exp) { if (!tree_fits_uhwi_p (TYPE_SIZE (TREE_TYPE (exp)))) return true; return (tree_to_uhwi (TYPE_SIZE (TREE_TYPE (exp))) > get_object_alignment (exp)); } /* Function vector_alignment_reachable_p Return true if vector alignment for DR is reachable by peeling a few loop iterations. Return false otherwise. */ static bool vector_alignment_reachable_p (struct data_reference *dr) { gimple stmt = DR_STMT (dr); stmt_vec_info stmt_info = vinfo_for_stmt (stmt); tree vectype = STMT_VINFO_VECTYPE (stmt_info); if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) { /* For interleaved access we peel only if number of iterations in the prolog loop ({VF - misalignment}), is a multiple of the number of the interleaved accesses. */ int elem_size, mis_in_elements; int nelements = TYPE_VECTOR_SUBPARTS (vectype); /* FORNOW: handle only known alignment. */ if (!known_alignment_for_access_p (dr)) return false; elem_size = GET_MODE_SIZE (TYPE_MODE (vectype)) / nelements; mis_in_elements = DR_MISALIGNMENT (dr) / elem_size; if ((nelements - mis_in_elements) % GROUP_SIZE (stmt_info)) return false; } /* If misalignment is known at the compile time then allow peeling only if natural alignment is reachable through peeling. */ if (known_alignment_for_access_p (dr) && !aligned_access_p (dr)) { HOST_WIDE_INT elmsize = int_cst_value (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); if (dump_enabled_p ()) { dump_printf_loc (MSG_NOTE, vect_location, "data size =" HOST_WIDE_INT_PRINT_DEC, elmsize); dump_printf (MSG_NOTE, ". misalignment = %d.\n", DR_MISALIGNMENT (dr)); } if (DR_MISALIGNMENT (dr) % elmsize) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "data size does not divide the misalignment.\n"); return false; } } if (!known_alignment_for_access_p (dr)) { tree type = TREE_TYPE (DR_REF (dr)); bool is_packed = not_size_aligned (DR_REF (dr)); if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "Unknown misalignment, is_packed = %d\n",is_packed); if ((TYPE_USER_ALIGN (type) && !is_packed) || targetm.vectorize.vector_alignment_reachable (type, is_packed)) return true; else return false; } return true; } /* Calculate the cost of the memory access represented by DR. */ static void vect_get_data_access_cost (struct data_reference *dr, unsigned int *inside_cost, unsigned int *outside_cost, stmt_vector_for_cost *body_cost_vec) { gimple stmt = DR_STMT (dr); stmt_vec_info stmt_info = vinfo_for_stmt (stmt); int nunits = TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info)); loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); int ncopies = vf / nunits; if (DR_IS_READ (dr)) vect_get_load_cost (dr, ncopies, true, inside_cost, outside_cost, NULL, body_cost_vec, false); else vect_get_store_cost (dr, ncopies, inside_cost, body_cost_vec); if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "vect_get_data_access_cost: inside_cost = %d, " "outside_cost = %d.\n", *inside_cost, *outside_cost); } /* Insert DR into peeling hash table with NPEEL as key. */ static void vect_peeling_hash_insert (loop_vec_info loop_vinfo, struct data_reference *dr, int npeel) { struct _vect_peel_info elem, *slot; _vect_peel_info **new_slot; bool supportable_dr_alignment = vect_supportable_dr_alignment (dr, true); elem.npeel = npeel; slot = LOOP_VINFO_PEELING_HTAB (loop_vinfo).find (&elem); if (slot) slot->count++; else { slot = XNEW (struct _vect_peel_info); slot->npeel = npeel; slot->dr = dr; slot->count = 1; new_slot = LOOP_VINFO_PEELING_HTAB (loop_vinfo).find_slot (slot, INSERT); *new_slot = slot; } if (!supportable_dr_alignment && unlimited_cost_model (LOOP_VINFO_LOOP (loop_vinfo))) slot->count += VECT_MAX_COST; } /* Traverse peeling hash table to find peeling option that aligns maximum number of data accesses. */ int vect_peeling_hash_get_most_frequent (_vect_peel_info **slot, _vect_peel_extended_info *max) { vect_peel_info elem = *slot; if (elem->count > max->peel_info.count || (elem->count == max->peel_info.count && max->peel_info.npeel > elem->npeel)) { max->peel_info.npeel = elem->npeel; max->peel_info.count = elem->count; max->peel_info.dr = elem->dr; } return 1; } /* Traverse peeling hash table and calculate cost for each peeling option. Find the one with the lowest cost. */ int vect_peeling_hash_get_lowest_cost (_vect_peel_info **slot, _vect_peel_extended_info *min) { vect_peel_info elem = *slot; int save_misalignment, dummy; unsigned int inside_cost = 0, outside_cost = 0, i; gimple stmt = DR_STMT (elem->dr); stmt_vec_info stmt_info = vinfo_for_stmt (stmt); loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); vec datarefs = LOOP_VINFO_DATAREFS (loop_vinfo); struct data_reference *dr; stmt_vector_for_cost prologue_cost_vec, body_cost_vec, epilogue_cost_vec; int single_iter_cost; prologue_cost_vec.create (2); body_cost_vec.create (2); epilogue_cost_vec.create (2); FOR_EACH_VEC_ELT (datarefs, i, dr) { stmt = DR_STMT (dr); stmt_info = vinfo_for_stmt (stmt); /* For interleaving, only the alignment of the first access matters. */ if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && GROUP_FIRST_ELEMENT (stmt_info) != stmt) continue; save_misalignment = DR_MISALIGNMENT (dr); vect_update_misalignment_for_peel (dr, elem->dr, elem->npeel); vect_get_data_access_cost (dr, &inside_cost, &outside_cost, &body_cost_vec); SET_DR_MISALIGNMENT (dr, save_misalignment); } single_iter_cost = vect_get_single_scalar_iteration_cost (loop_vinfo); outside_cost += vect_get_known_peeling_cost (loop_vinfo, elem->npeel, &dummy, single_iter_cost, &prologue_cost_vec, &epilogue_cost_vec); /* Prologue and epilogue costs are added to the target model later. These costs depend only on the scalar iteration cost, the number of peeling iterations finally chosen, and the number of misaligned statements. So discard the information found here. */ prologue_cost_vec.release (); epilogue_cost_vec.release (); if (inside_cost < min->inside_cost || (inside_cost == min->inside_cost && outside_cost < min->outside_cost)) { min->inside_cost = inside_cost; min->outside_cost = outside_cost; min->body_cost_vec.release (); min->body_cost_vec = body_cost_vec; min->peel_info.dr = elem->dr; min->peel_info.npeel = elem->npeel; } else body_cost_vec.release (); return 1; } /* Choose best peeling option by traversing peeling hash table and either choosing an option with the lowest cost (if cost model is enabled) or the option that aligns as many accesses as possible. */ static struct data_reference * vect_peeling_hash_choose_best_peeling (loop_vec_info loop_vinfo, unsigned int *npeel, stmt_vector_for_cost *body_cost_vec) { struct _vect_peel_extended_info res; res.peel_info.dr = NULL; res.body_cost_vec = stmt_vector_for_cost (); if (!unlimited_cost_model (LOOP_VINFO_LOOP (loop_vinfo))) { res.inside_cost = INT_MAX; res.outside_cost = INT_MAX; LOOP_VINFO_PEELING_HTAB (loop_vinfo) .traverse <_vect_peel_extended_info *, vect_peeling_hash_get_lowest_cost> (&res); } else { res.peel_info.count = 0; LOOP_VINFO_PEELING_HTAB (loop_vinfo) .traverse <_vect_peel_extended_info *, vect_peeling_hash_get_most_frequent> (&res); } *npeel = res.peel_info.npeel; *body_cost_vec = res.body_cost_vec; return res.peel_info.dr; } /* Function vect_enhance_data_refs_alignment This pass will use loop versioning and loop peeling in order to enhance the alignment of data references in the loop. FOR NOW: we assume that whatever versioning/peeling takes place, only the original loop is to be vectorized. Any other loops that are created by the transformations performed in this pass - are not supposed to be vectorized. This restriction will be relaxed. This pass will require a cost model to guide it whether to apply peeling or versioning or a combination of the two. For example, the scheme that intel uses when given a loop with several memory accesses, is as follows: choose one memory access ('p') which alignment you want to force by doing peeling. Then, either (1) generate a loop in which 'p' is aligned and all other accesses are not necessarily aligned, or (2) use loop versioning to generate one loop in which all accesses are aligned, and another loop in which only 'p' is necessarily aligned. ("Automatic Intra-Register Vectorization for the Intel Architecture", Aart J.C. Bik, Milind Girkar, Paul M. Grey and Ximmin Tian, International Journal of Parallel Programming, Vol. 30, No. 2, April 2002.) Devising a cost model is the most critical aspect of this work. It will guide us on which access to peel for, whether to use loop versioning, how many versions to create, etc. The cost model will probably consist of generic considerations as well as target specific considerations (on powerpc for example, misaligned stores are more painful than misaligned loads). Here are the general steps involved in alignment enhancements: -- original loop, before alignment analysis: for (i=0; i datarefs = LOOP_VINFO_DATAREFS (loop_vinfo); struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); enum dr_alignment_support supportable_dr_alignment; struct data_reference *dr0 = NULL, *first_store = NULL; struct data_reference *dr; unsigned int i, j; bool do_peeling = false; bool do_versioning = false; bool stat; gimple stmt; stmt_vec_info stmt_info; unsigned int npeel = 0; bool all_misalignments_unknown = true; unsigned int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); unsigned possible_npeel_number = 1; tree vectype; unsigned int nelements, mis, same_align_drs_max = 0; stmt_vector_for_cost body_cost_vec = stmt_vector_for_cost (); if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "=== vect_enhance_data_refs_alignment ===\n"); /* While cost model enhancements are expected in the future, the high level view of the code at this time is as follows: A) If there is a misaligned access then see if peeling to align this access can make all data references satisfy vect_supportable_dr_alignment. If so, update data structures as needed and return true. B) If peeling wasn't possible and there is a data reference with an unknown misalignment that does not satisfy vect_supportable_dr_alignment then see if loop versioning checks can be used to make all data references satisfy vect_supportable_dr_alignment. If so, update data structures as needed and return true. C) If neither peeling nor versioning were successful then return false if any data reference does not satisfy vect_supportable_dr_alignment. D) Return true (all data references satisfy vect_supportable_dr_alignment). Note, Possibility 3 above (which is peeling and versioning together) is not being done at this time. */ /* (1) Peeling to force alignment. */ /* (1.1) Decide whether to perform peeling, and how many iterations to peel: Considerations: + How many accesses will become aligned due to the peeling - How many accesses will become unaligned due to the peeling, and the cost of misaligned accesses. - The cost of peeling (the extra runtime checks, the increase in code size). */ FOR_EACH_VEC_ELT (datarefs, i, dr) { stmt = DR_STMT (dr); stmt_info = vinfo_for_stmt (stmt); if (!STMT_VINFO_RELEVANT_P (stmt_info)) continue; /* For interleaving, only the alignment of the first access matters. */ if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && GROUP_FIRST_ELEMENT (stmt_info) != stmt) continue; /* For invariant accesses there is nothing to enhance. */ if (integer_zerop (DR_STEP (dr))) continue; /* Strided loads perform only component accesses, alignment is irrelevant for them. */ if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) continue; supportable_dr_alignment = vect_supportable_dr_alignment (dr, true); do_peeling = vector_alignment_reachable_p (dr); if (do_peeling) { if (known_alignment_for_access_p (dr)) { unsigned int npeel_tmp; bool negative = tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0; /* Save info about DR in the hash table. */ if (!LOOP_VINFO_PEELING_HTAB (loop_vinfo).is_created ()) LOOP_VINFO_PEELING_HTAB (loop_vinfo).create (1); vectype = STMT_VINFO_VECTYPE (stmt_info); nelements = TYPE_VECTOR_SUBPARTS (vectype); mis = DR_MISALIGNMENT (dr) / GET_MODE_SIZE (TYPE_MODE ( TREE_TYPE (DR_REF (dr)))); npeel_tmp = (negative ? (mis - nelements) : (nelements - mis)) & (nelements - 1); /* For multiple types, it is possible that the bigger type access will have more than one peeling option. E.g., a loop with two types: one of size (vector size / 4), and the other one of size (vector size / 8). Vectorization factor will 8. If both access are misaligned by 3, the first one needs one scalar iteration to be aligned, and the second one needs 5. But the the first one will be aligned also by peeling 5 scalar iterations, and in that case both accesses will be aligned. Hence, except for the immediate peeling amount, we also want to try to add full vector size, while we don't exceed vectorization factor. We do this automtically for cost model, since we calculate cost for every peeling option. */ if (unlimited_cost_model (LOOP_VINFO_LOOP (loop_vinfo))) possible_npeel_number = vf /nelements; /* Handle the aligned case. We may decide to align some other access, making DR unaligned. */ if (DR_MISALIGNMENT (dr) == 0) { npeel_tmp = 0; if (unlimited_cost_model (LOOP_VINFO_LOOP (loop_vinfo))) possible_npeel_number++; } for (j = 0; j < possible_npeel_number; j++) { gcc_assert (npeel_tmp <= vf); vect_peeling_hash_insert (loop_vinfo, dr, npeel_tmp); npeel_tmp += nelements; } all_misalignments_unknown = false; /* Data-ref that was chosen for the case that all the misalignments are unknown is not relevant anymore, since we have a data-ref with known alignment. */ dr0 = NULL; } else { /* If we don't know any misalignment values, we prefer peeling for data-ref that has the maximum number of data-refs with the same alignment, unless the target prefers to align stores over load. */ if (all_misalignments_unknown) { unsigned same_align_drs = STMT_VINFO_SAME_ALIGN_REFS (stmt_info).length (); if (!dr0 || same_align_drs_max < same_align_drs) { same_align_drs_max = same_align_drs; dr0 = dr; } /* For data-refs with the same number of related accesses prefer the one where the misalign computation will be invariant in the outermost loop. */ else if (same_align_drs_max == same_align_drs) { struct loop *ivloop0, *ivloop; ivloop0 = outermost_invariant_loop_for_expr (loop, DR_BASE_ADDRESS (dr0)); ivloop = outermost_invariant_loop_for_expr (loop, DR_BASE_ADDRESS (dr)); if ((ivloop && !ivloop0) || (ivloop && ivloop0 && flow_loop_nested_p (ivloop, ivloop0))) dr0 = dr; } if (!first_store && DR_IS_WRITE (dr)) first_store = dr; } /* If there are both known and unknown misaligned accesses in the loop, we choose peeling amount according to the known accesses. */ if (!supportable_dr_alignment) { dr0 = dr; if (!first_store && DR_IS_WRITE (dr)) first_store = dr; } } } else { if (!aligned_access_p (dr)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "vector alignment may not be reachable\n"); break; } } } /* Check if we can possibly peel the loop. */ if (!vect_can_advance_ivs_p (loop_vinfo) || !slpeel_can_duplicate_loop_p (loop, single_exit (loop))) do_peeling = false; if (do_peeling && all_misalignments_unknown && vect_supportable_dr_alignment (dr0, false)) { /* Check if the target requires to prefer stores over loads, i.e., if misaligned stores are more expensive than misaligned loads (taking drs with same alignment into account). */ if (first_store && DR_IS_READ (dr0)) { unsigned int load_inside_cost = 0, load_outside_cost = 0; unsigned int store_inside_cost = 0, store_outside_cost = 0; unsigned int load_inside_penalty = 0, load_outside_penalty = 0; unsigned int store_inside_penalty = 0, store_outside_penalty = 0; stmt_vector_for_cost dummy; dummy.create (2); vect_get_data_access_cost (dr0, &load_inside_cost, &load_outside_cost, &dummy); vect_get_data_access_cost (first_store, &store_inside_cost, &store_outside_cost, &dummy); dummy.release (); /* Calculate the penalty for leaving FIRST_STORE unaligned (by aligning the load DR0). */ load_inside_penalty = store_inside_cost; load_outside_penalty = store_outside_cost; for (i = 0; STMT_VINFO_SAME_ALIGN_REFS (vinfo_for_stmt ( DR_STMT (first_store))).iterate (i, &dr); i++) if (DR_IS_READ (dr)) { load_inside_penalty += load_inside_cost; load_outside_penalty += load_outside_cost; } else { load_inside_penalty += store_inside_cost; load_outside_penalty += store_outside_cost; } /* Calculate the penalty for leaving DR0 unaligned (by aligning the FIRST_STORE). */ store_inside_penalty = load_inside_cost; store_outside_penalty = load_outside_cost; for (i = 0; STMT_VINFO_SAME_ALIGN_REFS (vinfo_for_stmt ( DR_STMT (dr0))).iterate (i, &dr); i++) if (DR_IS_READ (dr)) { store_inside_penalty += load_inside_cost; store_outside_penalty += load_outside_cost; } else { store_inside_penalty += store_inside_cost; store_outside_penalty += store_outside_cost; } if (load_inside_penalty > store_inside_penalty || (load_inside_penalty == store_inside_penalty && load_outside_penalty > store_outside_penalty)) dr0 = first_store; } /* In case there are only loads with different unknown misalignments, use peeling only if it may help to align other accesses in the loop. */ if (!first_store && !STMT_VINFO_SAME_ALIGN_REFS ( vinfo_for_stmt (DR_STMT (dr0))).length () && vect_supportable_dr_alignment (dr0, false) != dr_unaligned_supported) do_peeling = false; } if (do_peeling && !dr0) { /* Peeling is possible, but there is no data access that is not supported unless aligned. So we try to choose the best possible peeling. */ /* We should get here only if there are drs with known misalignment. */ gcc_assert (!all_misalignments_unknown); /* Choose the best peeling from the hash table. */ dr0 = vect_peeling_hash_choose_best_peeling (loop_vinfo, &npeel, &body_cost_vec); if (!dr0 || !npeel) do_peeling = false; } if (do_peeling) { stmt = DR_STMT (dr0); stmt_info = vinfo_for_stmt (stmt); vectype = STMT_VINFO_VECTYPE (stmt_info); nelements = TYPE_VECTOR_SUBPARTS (vectype); if (known_alignment_for_access_p (dr0)) { bool negative = tree_int_cst_compare (DR_STEP (dr0), size_zero_node) < 0; if (!npeel) { /* Since it's known at compile time, compute the number of iterations in the peeled loop (the peeling factor) for use in updating DR_MISALIGNMENT values. The peeling factor is the vectorization factor minus the misalignment as an element count. */ mis = DR_MISALIGNMENT (dr0); mis /= GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dr0)))); npeel = ((negative ? mis - nelements : nelements - mis) & (nelements - 1)); } /* For interleaved data access every iteration accesses all the members of the group, therefore we divide the number of iterations by the group size. */ stmt_info = vinfo_for_stmt (DR_STMT (dr0)); if (STMT_VINFO_GROUPED_ACCESS (stmt_info)) npeel /= GROUP_SIZE (stmt_info); if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "Try peeling by %d\n", npeel); } /* Ensure that all data refs can be vectorized after the peel. */ FOR_EACH_VEC_ELT (datarefs, i, dr) { int save_misalignment; if (dr == dr0) continue; stmt = DR_STMT (dr); stmt_info = vinfo_for_stmt (stmt); /* For interleaving, only the alignment of the first access matters. */ if (STMT_VINFO_GROUPED_ACCESS (stmt_info) && GROUP_FIRST_ELEMENT (stmt_info) != stmt) continue; /* Strided loads perform only component accesses, alignment is irrelevant for them. */ if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) continue; save_misalignment = DR_MISALIGNMENT (dr); vect_update_misalignment_for_peel (dr, dr0, npeel); supportable_dr_alignment = vect_supportable_dr_alignment (dr, false); SET_DR_MISALIGNMENT (dr, save_misalignment); if (!supportable_dr_alignment) { do_peeling = false; break; } } if (do_peeling && known_alignment_for_access_p (dr0) && npeel == 0) { stat = vect_verify_datarefs_alignment (loop_vinfo, NULL); if (!stat) do_peeling = false; else { body_cost_vec.release (); return stat; } } if (do_peeling) { unsigned max_allowed_peel = PARAM_VALUE (PARAM_VECT_MAX_PEELING_FOR_ALIGNMENT); if (max_allowed_peel != (unsigned)-1) { unsigned max_peel = npeel; if (max_peel == 0) { gimple dr_stmt = DR_STMT (dr0); stmt_vec_info vinfo = vinfo_for_stmt (dr_stmt); tree vtype = STMT_VINFO_VECTYPE (vinfo); max_peel = TYPE_VECTOR_SUBPARTS (vtype) - 1; } if (max_peel > max_allowed_peel) { do_peeling = false; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "Disable peeling, max peels reached: %d\n", max_peel); } } } if (do_peeling) { stmt_info_for_cost *si; void *data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo); /* (1.2) Update the DR_MISALIGNMENT of each data reference DR_i. If the misalignment of DR_i is identical to that of dr0 then set DR_MISALIGNMENT (DR_i) to zero. If the misalignment of DR_i and dr0 are known at compile time then increment DR_MISALIGNMENT (DR_i) by the peeling factor times the element size of DR_i (MOD the vectorization factor times the size). Otherwise, the misalignment of DR_i must be set to unknown. */ FOR_EACH_VEC_ELT (datarefs, i, dr) if (dr != dr0) vect_update_misalignment_for_peel (dr, dr0, npeel); LOOP_VINFO_UNALIGNED_DR (loop_vinfo) = dr0; if (npeel) LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) = npeel; else LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) = DR_MISALIGNMENT (dr0); SET_DR_MISALIGNMENT (dr0, 0); if (dump_enabled_p ()) { dump_printf_loc (MSG_NOTE, vect_location, "Alignment of access forced using peeling.\n"); dump_printf_loc (MSG_NOTE, vect_location, "Peeling for alignment will be applied.\n"); } /* We've delayed passing the inside-loop peeling costs to the target cost model until we were sure peeling would happen. Do so now. */ if (body_cost_vec.exists ()) { FOR_EACH_VEC_ELT (body_cost_vec, i, si) { struct _stmt_vec_info *stmt_info = si->stmt ? vinfo_for_stmt (si->stmt) : NULL; (void) add_stmt_cost (data, si->count, si->kind, stmt_info, si->misalign, vect_body); } body_cost_vec.release (); } stat = vect_verify_datarefs_alignment (loop_vinfo, NULL); gcc_assert (stat); return stat; } } body_cost_vec.release (); /* (2) Versioning to force alignment. */ /* Try versioning if: 1) optimize loop for speed 2) there is at least one unsupported misaligned data ref with an unknown misalignment, and 3) all misaligned data refs with a known misalignment are supported, and 4) the number of runtime alignment checks is within reason. */ do_versioning = optimize_loop_nest_for_speed_p (loop) && (!loop->inner); /* FORNOW */ if (do_versioning) { FOR_EACH_VEC_ELT (datarefs, i, dr) { stmt = DR_STMT (dr); stmt_info = vinfo_for_stmt (stmt); /* For interleaving, only the alignment of the first access matters. */ if (aligned_access_p (dr) || (STMT_VINFO_GROUPED_ACCESS (stmt_info) && GROUP_FIRST_ELEMENT (stmt_info) != stmt)) continue; /* Strided loads perform only component accesses, alignment is irrelevant for them. */ if (STMT_VINFO_STRIDE_LOAD_P (stmt_info)) continue; supportable_dr_alignment = vect_supportable_dr_alignment (dr, false); if (!supportable_dr_alignment) { gimple stmt; int mask; tree vectype; if (known_alignment_for_access_p (dr) || LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo).length () >= (unsigned) PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIGNMENT_CHECKS)) { do_versioning = false; break; } stmt = DR_STMT (dr); vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt)); gcc_assert (vectype); /* The rightmost bits of an aligned address must be zeros. Construct the mask needed for this test. For example, GET_MODE_SIZE for the vector mode V4SI is 16 bytes so the mask must be 15 = 0xf. */ mask = GET_MODE_SIZE (TYPE_MODE (vectype)) - 1; /* FORNOW: use the same mask to test all potentially unaligned references in the loop. The vectorizer currently supports a single vector size, see the reference to GET_MODE_NUNITS (TYPE_MODE (vectype)) where the vectorization factor is computed. */ gcc_assert (!LOOP_VINFO_PTR_MASK (loop_vinfo) || LOOP_VINFO_PTR_MASK (loop_vinfo) == mask); LOOP_VINFO_PTR_MASK (loop_vinfo) = mask; LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo).safe_push ( DR_STMT (dr)); } } /* Versioning requires at least one misaligned data reference. */ if (!LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo)) do_versioning = false; else if (!do_versioning) LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo).truncate (0); } if (do_versioning) { vec may_misalign_stmts = LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo); gimple stmt; /* It can now be assumed that the data references in the statements in LOOP_VINFO_MAY_MISALIGN_STMTS will be aligned in the version of the loop being vectorized. */ FOR_EACH_VEC_ELT (may_misalign_stmts, i, stmt) { stmt_vec_info stmt_info = vinfo_for_stmt (stmt); dr = STMT_VINFO_DATA_REF (stmt_info); SET_DR_MISALIGNMENT (dr, 0); if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "Alignment of access forced using versioning.\n"); } if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "Versioning for alignment will be applied.\n"); /* Peeling and versioning can't be done together at this time. */ gcc_assert (! (do_peeling && do_versioning)); stat = vect_verify_datarefs_alignment (loop_vinfo, NULL); gcc_assert (stat); return stat; } /* This point is reached if neither peeling nor versioning is being done. */ gcc_assert (! (do_peeling || do_versioning)); stat = vect_verify_datarefs_alignment (loop_vinfo, NULL); return stat; } /* Function vect_find_same_alignment_drs. Update group and alignment relations according to the chosen vectorization factor. */ static void vect_find_same_alignment_drs (struct data_dependence_relation *ddr, loop_vec_info loop_vinfo) { unsigned int i; struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); int vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo); struct data_reference *dra = DDR_A (ddr); struct data_reference *drb = DDR_B (ddr); stmt_vec_info stmtinfo_a = vinfo_for_stmt (DR_STMT (dra)); stmt_vec_info stmtinfo_b = vinfo_for_stmt (DR_STMT (drb)); int dra_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dra)))); int drb_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (drb)))); lambda_vector dist_v; unsigned int loop_depth; if (DDR_ARE_DEPENDENT (ddr) == chrec_known) return; if (dra == drb) return; if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) return; /* Loop-based vectorization and known data dependence. */ if (DDR_NUM_DIST_VECTS (ddr) == 0) return; /* Data-dependence analysis reports a distance vector of zero for data-references that overlap only in the first iteration but have different sign step (see PR45764). So as a sanity check require equal DR_STEP. */ if (!operand_equal_p (DR_STEP (dra), DR_STEP (drb), 0)) return; loop_depth = index_in_loop_nest (loop->num, DDR_LOOP_NEST (ddr)); FOR_EACH_VEC_ELT (DDR_DIST_VECTS (ddr), i, dist_v) { int dist = dist_v[loop_depth]; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "dependence distance = %d.\n", dist); /* Same loop iteration. */ if (dist == 0 || (dist % vectorization_factor == 0 && dra_size == drb_size)) { /* Two references with distance zero have the same alignment. */ STMT_VINFO_SAME_ALIGN_REFS (stmtinfo_a).safe_push (drb); STMT_VINFO_SAME_ALIGN_REFS (stmtinfo_b).safe_push (dra); if (dump_enabled_p ()) { dump_printf_loc (MSG_NOTE, vect_location, "accesses have the same alignment.\n"); dump_printf (MSG_NOTE, "dependence distance modulo vf == 0 between "); dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dra)); dump_printf (MSG_NOTE, " and "); dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (drb)); dump_printf (MSG_NOTE, "\n"); } } } } /* Function vect_analyze_data_refs_alignment Analyze the alignment of the data-references in the loop. Return FALSE if a data reference is found that cannot be vectorized. */ bool vect_analyze_data_refs_alignment (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo) { if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "=== vect_analyze_data_refs_alignment ===\n"); /* Mark groups of data references with same alignment using data dependence information. */ if (loop_vinfo) { vec ddrs = LOOP_VINFO_DDRS (loop_vinfo); struct data_dependence_relation *ddr; unsigned int i; FOR_EACH_VEC_ELT (ddrs, i, ddr) vect_find_same_alignment_drs (ddr, loop_vinfo); } if (!vect_compute_data_refs_alignment (loop_vinfo, bb_vinfo)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "not vectorized: can't calculate alignment " "for data ref.\n"); return false; } return true; } /* Analyze groups of accesses: check that DR belongs to a group of accesses of legal size, step, etc. Detect gaps, single element interleaving, and other special cases. Set grouped access info. Collect groups of strided stores for further use in SLP analysis. */ static bool vect_analyze_group_access (struct data_reference *dr) { tree step = DR_STEP (dr); tree scalar_type = TREE_TYPE (DR_REF (dr)); HOST_WIDE_INT type_size = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (scalar_type)); gimple stmt = DR_STMT (dr); stmt_vec_info stmt_info = vinfo_for_stmt (stmt); loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); HOST_WIDE_INT dr_step = TREE_INT_CST_LOW (step); HOST_WIDE_INT groupsize, last_accessed_element = 1; bool slp_impossible = false; struct loop *loop = NULL; if (loop_vinfo) loop = LOOP_VINFO_LOOP (loop_vinfo); /* For interleaving, GROUPSIZE is STEP counted in elements, i.e., the size of the interleaving group (including gaps). */ groupsize = absu_hwi (dr_step) / type_size; /* Not consecutive access is possible only if it is a part of interleaving. */ if (!GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))) { /* Check if it this DR is a part of interleaving, and is a single element of the group that is accessed in the loop. */ /* Gaps are supported only for loads. STEP must be a multiple of the type size. The size of the group must be a power of 2. */ if (DR_IS_READ (dr) && (dr_step % type_size) == 0 && groupsize > 0 && exact_log2 (groupsize) != -1) { GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) = stmt; GROUP_SIZE (vinfo_for_stmt (stmt)) = groupsize; if (dump_enabled_p ()) { dump_printf_loc (MSG_NOTE, vect_location, "Detected single element interleaving "); dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dr)); dump_printf (MSG_NOTE, " step "); dump_generic_expr (MSG_NOTE, TDF_SLIM, step); dump_printf (MSG_NOTE, "\n"); } if (loop_vinfo) { if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "Data access with gaps requires scalar " "epilogue loop\n"); if (loop->inner) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "Peeling for outer loop is not" " supported\n"); return false; } LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true; } return true; } if (dump_enabled_p ()) { dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "not consecutive access "); dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); } if (bb_vinfo) { /* Mark the statement as unvectorizable. */ STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (DR_STMT (dr))) = false; return true; } return false; } if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) == stmt) { /* First stmt in the interleaving chain. Check the chain. */ gimple next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (stmt)); struct data_reference *data_ref = dr; unsigned int count = 1; tree prev_init = DR_INIT (data_ref); gimple prev = stmt; HOST_WIDE_INT diff, gaps = 0; unsigned HOST_WIDE_INT count_in_bytes; while (next) { /* Skip same data-refs. In case that two or more stmts share data-ref (supported only for loads), we vectorize only the first stmt, and the rest get their vectorized loads from the first one. */ if (!tree_int_cst_compare (DR_INIT (data_ref), DR_INIT (STMT_VINFO_DATA_REF ( vinfo_for_stmt (next))))) { if (DR_IS_WRITE (data_ref)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "Two store stmts share the same dr.\n"); return false; } /* For load use the same data-ref load. */ GROUP_SAME_DR_STMT (vinfo_for_stmt (next)) = prev; prev = next; next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next)); continue; } prev = next; data_ref = STMT_VINFO_DATA_REF (vinfo_for_stmt (next)); /* All group members have the same STEP by construction. */ gcc_checking_assert (operand_equal_p (DR_STEP (data_ref), step, 0)); /* Check that the distance between two accesses is equal to the type size. Otherwise, we have gaps. */ diff = (TREE_INT_CST_LOW (DR_INIT (data_ref)) - TREE_INT_CST_LOW (prev_init)) / type_size; if (diff != 1) { /* FORNOW: SLP of accesses with gaps is not supported. */ slp_impossible = true; if (DR_IS_WRITE (data_ref)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "interleaved store with gaps\n"); return false; } gaps += diff - 1; } last_accessed_element += diff; /* Store the gap from the previous member of the group. If there is no gap in the access, GROUP_GAP is always 1. */ GROUP_GAP (vinfo_for_stmt (next)) = diff; prev_init = DR_INIT (data_ref); next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next)); /* Count the number of data-refs in the chain. */ count++; } /* COUNT is the number of accesses found, we multiply it by the size of the type to get COUNT_IN_BYTES. */ count_in_bytes = type_size * count; /* Check that the size of the interleaving (including gaps) is not greater than STEP. */ if (dr_step != 0 && absu_hwi (dr_step) < count_in_bytes + gaps * type_size) { if (dump_enabled_p ()) { dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "interleaving size is greater than step for "); dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, DR_REF (dr)); dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); } return false; } /* Check that the size of the interleaving is equal to STEP for stores, i.e., that there are no gaps. */ if (dr_step != 0 && absu_hwi (dr_step) != count_in_bytes) { if (DR_IS_READ (dr)) { slp_impossible = true; /* There is a gap after the last load in the group. This gap is a difference between the groupsize and the number of elements. When there is no gap, this difference should be 0. */ GROUP_GAP (vinfo_for_stmt (stmt)) = groupsize - count; } else { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "interleaved store with gaps\n"); return false; } } /* Check that STEP is a multiple of type size. */ if (dr_step != 0 && (dr_step % type_size) != 0) { if (dump_enabled_p ()) { dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "step is not a multiple of type size: step "); dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, step); dump_printf (MSG_MISSED_OPTIMIZATION, " size "); dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, TYPE_SIZE_UNIT (scalar_type)); dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); } return false; } if (groupsize == 0) groupsize = count; GROUP_SIZE (vinfo_for_stmt (stmt)) = groupsize; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "Detected interleaving of size %d\n", (int)groupsize); /* SLP: create an SLP data structure for every interleaving group of stores for further analysis in vect_analyse_slp. */ if (DR_IS_WRITE (dr) && !slp_impossible) { if (loop_vinfo) LOOP_VINFO_GROUPED_STORES (loop_vinfo).safe_push (stmt); if (bb_vinfo) BB_VINFO_GROUPED_STORES (bb_vinfo).safe_push (stmt); } /* There is a gap in the end of the group. */ if (groupsize - last_accessed_element > 0 && loop_vinfo) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "Data access with gaps requires scalar " "epilogue loop\n"); if (loop->inner) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "Peeling for outer loop is not supported\n"); return false; } LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true; } } return true; } /* Analyze the access pattern of the data-reference DR. In case of non-consecutive accesses call vect_analyze_group_access() to analyze groups of accesses. */ static bool vect_analyze_data_ref_access (struct data_reference *dr) { tree step = DR_STEP (dr); tree scalar_type = TREE_TYPE (DR_REF (dr)); gimple stmt = DR_STMT (dr); stmt_vec_info stmt_info = vinfo_for_stmt (stmt); loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); struct loop *loop = NULL; if (loop_vinfo) loop = LOOP_VINFO_LOOP (loop_vinfo); if (loop_vinfo && !step) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "bad data-ref access in loop\n"); return false; } /* Allow invariant loads in not nested loops. */ if (loop_vinfo && integer_zerop (step)) { GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) = NULL; if (nested_in_vect_loop_p (loop, stmt)) { if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "zero step in inner loop of nest\n"); return false; } return DR_IS_READ (dr); } if (loop && nested_in_vect_loop_p (loop, stmt)) { /* Interleaved accesses are not yet supported within outer-loop vectorization for references in the inner-loop. */ GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) = NULL; /* For the rest of the analysis we use the outer-loop step. */ step = STMT_VINFO_DR_STEP (stmt_info); if (integer_zerop (step)) { if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "zero step in outer loop.\n"); if (DR_IS_READ (dr)) return true; else return false; } } /* Consecutive? */ if (TREE_CODE (step) == INTEGER_CST) { HOST_WIDE_INT dr_step = TREE_INT_CST_LOW (step); if (!tree_int_cst_compare (step, TYPE_SIZE_UNIT (scalar_type)) || (dr_step < 0 && !compare_tree_int (TYPE_SIZE_UNIT (scalar_type), -dr_step))) { /* Mark that it is not interleaving. */ GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) = NULL; return true; } } if (loop && nested_in_vect_loop_p (loop, stmt)) { if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "grouped access in outer loop.\n"); return false; } /* Assume this is a DR handled by non-constant strided load case. */ if (TREE_CODE (step) != INTEGER_CST) return STMT_VINFO_STRIDE_LOAD_P (stmt_info); /* Not consecutive access - check if it's a part of interleaving group. */ return vect_analyze_group_access (dr); } /* A helper function used in the comparator function to sort data references. T1 and T2 are two data references to be compared. The function returns -1, 0, or 1. */ static int compare_tree (tree t1, tree t2) { int i, cmp; enum tree_code code; char tclass; if (t1 == t2) return 0; if (t1 == NULL) return -1; if (t2 == NULL) return 1; if (TREE_CODE (t1) != TREE_CODE (t2)) return TREE_CODE (t1) < TREE_CODE (t2) ? -1 : 1; code = TREE_CODE (t1); switch (code) { /* For const values, we can just use hash values for comparisons. */ case INTEGER_CST: case REAL_CST: case FIXED_CST: case STRING_CST: case COMPLEX_CST: case VECTOR_CST: { hashval_t h1 = iterative_hash_expr (t1, 0); hashval_t h2 = iterative_hash_expr (t2, 0); if (h1 != h2) return h1 < h2 ? -1 : 1; break; } case SSA_NAME: cmp = compare_tree (SSA_NAME_VAR (t1), SSA_NAME_VAR (t2)); if (cmp != 0) return cmp; if (SSA_NAME_VERSION (t1) != SSA_NAME_VERSION (t2)) return SSA_NAME_VERSION (t1) < SSA_NAME_VERSION (t2) ? -1 : 1; break; default: tclass = TREE_CODE_CLASS (code); /* For var-decl, we could compare their UIDs. */ if (tclass == tcc_declaration) { if (DECL_UID (t1) != DECL_UID (t2)) return DECL_UID (t1) < DECL_UID (t2) ? -1 : 1; break; } /* For expressions with operands, compare their operands recursively. */ for (i = TREE_OPERAND_LENGTH (t1) - 1; i >= 0; --i) { cmp = compare_tree (TREE_OPERAND (t1, i), TREE_OPERAND (t2, i)); if (cmp != 0) return cmp; } } return 0; } /* Compare two data-references DRA and DRB to group them into chunks suitable for grouping. */ static int dr_group_sort_cmp (const void *dra_, const void *drb_) { data_reference_p dra = *(data_reference_p *)const_cast(dra_); data_reference_p drb = *(data_reference_p *)const_cast(drb_); int cmp; /* Stabilize sort. */ if (dra == drb) return 0; /* Ordering of DRs according to base. */ if (!operand_equal_p (DR_BASE_ADDRESS (dra), DR_BASE_ADDRESS (drb), 0)) { cmp = compare_tree (DR_BASE_ADDRESS (dra), DR_BASE_ADDRESS (drb)); if (cmp != 0) return cmp; } /* And according to DR_OFFSET. */ if (!dr_equal_offsets_p (dra, drb)) { cmp = compare_tree (DR_OFFSET (dra), DR_OFFSET (drb)); if (cmp != 0) return cmp; } /* Put reads before writes. */ if (DR_IS_READ (dra) != DR_IS_READ (drb)) return DR_IS_READ (dra) ? -1 : 1; /* Then sort after access size. */ if (!operand_equal_p (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dra))), TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (drb))), 0)) { cmp = compare_tree (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dra))), TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (drb)))); if (cmp != 0) return cmp; } /* And after step. */ if (!operand_equal_p (DR_STEP (dra), DR_STEP (drb), 0)) { cmp = compare_tree (DR_STEP (dra), DR_STEP (drb)); if (cmp != 0) return cmp; } /* Then sort after DR_INIT. In case of identical DRs sort after stmt UID. */ cmp = tree_int_cst_compare (DR_INIT (dra), DR_INIT (drb)); if (cmp == 0) return gimple_uid (DR_STMT (dra)) < gimple_uid (DR_STMT (drb)) ? -1 : 1; return cmp; } /* Function vect_analyze_data_ref_accesses. Analyze the access pattern of all the data references in the loop. FORNOW: the only access pattern that is considered vectorizable is a simple step 1 (consecutive) access. FORNOW: handle only arrays and pointer accesses. */ bool vect_analyze_data_ref_accesses (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo) { unsigned int i; vec datarefs; struct data_reference *dr; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "=== vect_analyze_data_ref_accesses ===\n"); if (loop_vinfo) datarefs = LOOP_VINFO_DATAREFS (loop_vinfo); else datarefs = BB_VINFO_DATAREFS (bb_vinfo); if (datarefs.is_empty ()) return true; /* Sort the array of datarefs to make building the interleaving chains linear. */ qsort (datarefs.address (), datarefs.length (), sizeof (data_reference_p), dr_group_sort_cmp); /* Build the interleaving chains. */ for (i = 0; i < datarefs.length () - 1;) { data_reference_p dra = datarefs[i]; stmt_vec_info stmtinfo_a = vinfo_for_stmt (DR_STMT (dra)); stmt_vec_info lastinfo = NULL; for (i = i + 1; i < datarefs.length (); ++i) { data_reference_p drb = datarefs[i]; stmt_vec_info stmtinfo_b = vinfo_for_stmt (DR_STMT (drb)); /* ??? Imperfect sorting (non-compatible types, non-modulo accesses, same accesses) can lead to a group to be artificially split here as we don't just skip over those. If it really matters we can push those to a worklist and re-iterate over them. The we can just skip ahead to the next DR here. */ /* Check that the data-refs have same first location (except init) and they are both either store or load (not load and store). */ if (DR_IS_READ (dra) != DR_IS_READ (drb) || !operand_equal_p (DR_BASE_ADDRESS (dra), DR_BASE_ADDRESS (drb), 0) || !dr_equal_offsets_p (dra, drb)) break; /* Check that the data-refs have the same constant size and step. */ tree sza = TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dra))); tree szb = TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (drb))); if (!tree_fits_uhwi_p (sza) || !tree_fits_uhwi_p (szb) || !tree_int_cst_equal (sza, szb) || !tree_fits_shwi_p (DR_STEP (dra)) || !tree_fits_shwi_p (DR_STEP (drb)) || !tree_int_cst_equal (DR_STEP (dra), DR_STEP (drb))) break; /* Do not place the same access in the interleaving chain twice. */ if (tree_int_cst_compare (DR_INIT (dra), DR_INIT (drb)) == 0) break; /* Check the types are compatible. ??? We don't distinguish this during sorting. */ if (!types_compatible_p (TREE_TYPE (DR_REF (dra)), TREE_TYPE (DR_REF (drb)))) break; /* Sorting has ensured that DR_INIT (dra) <= DR_INIT (drb). */ HOST_WIDE_INT init_a = TREE_INT_CST_LOW (DR_INIT (dra)); HOST_WIDE_INT init_b = TREE_INT_CST_LOW (DR_INIT (drb)); gcc_assert (init_a < init_b); /* If init_b == init_a + the size of the type * k, we have an interleaving, and DRA is accessed before DRB. */ HOST_WIDE_INT type_size_a = tree_to_uhwi (sza); if ((init_b - init_a) % type_size_a != 0) break; /* The step (if not zero) is greater than the difference between data-refs' inits. This splits groups into suitable sizes. */ HOST_WIDE_INT step = tree_to_shwi (DR_STEP (dra)); if (step != 0 && step <= (init_b - init_a)) break; if (dump_enabled_p ()) { dump_printf_loc (MSG_NOTE, vect_location, "Detected interleaving "); dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dra)); dump_printf (MSG_NOTE, " and "); dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (drb)); dump_printf (MSG_NOTE, "\n"); } /* Link the found element into the group list. */ if (!GROUP_FIRST_ELEMENT (stmtinfo_a)) { GROUP_FIRST_ELEMENT (stmtinfo_a) = DR_STMT (dra); lastinfo = stmtinfo_a; } GROUP_FIRST_ELEMENT (stmtinfo_b) = DR_STMT (dra); GROUP_NEXT_ELEMENT (lastinfo) = DR_STMT (drb); lastinfo = stmtinfo_b; } } FOR_EACH_VEC_ELT (datarefs, i, dr) if (STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (DR_STMT (dr))) && !vect_analyze_data_ref_access (dr)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "not vectorized: complicated access pattern.\n"); if (bb_vinfo) { /* Mark the statement as not vectorizable. */ STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (DR_STMT (dr))) = false; continue; } else return false; } return true; } /* Operator == between two dr_with_seg_len objects. This equality operator is used to make sure two data refs are the same one so that we will consider to combine the aliasing checks of those two pairs of data dependent data refs. */ static bool operator == (const dr_with_seg_len& d1, const dr_with_seg_len& d2) { return operand_equal_p (DR_BASE_ADDRESS (d1.dr), DR_BASE_ADDRESS (d2.dr), 0) && compare_tree (d1.offset, d2.offset) == 0 && compare_tree (d1.seg_len, d2.seg_len) == 0; } /* Function comp_dr_with_seg_len_pair. Comparison function for sorting objects of dr_with_seg_len_pair_t so that we can combine aliasing checks in one scan. */ static int comp_dr_with_seg_len_pair (const void *p1_, const void *p2_) { const dr_with_seg_len_pair_t* p1 = (const dr_with_seg_len_pair_t *) p1_; const dr_with_seg_len_pair_t* p2 = (const dr_with_seg_len_pair_t *) p2_; const dr_with_seg_len &p11 = p1->first, &p12 = p1->second, &p21 = p2->first, &p22 = p2->second; /* For DR pairs (a, b) and (c, d), we only consider to merge the alias checks if a and c have the same basic address snd step, and b and d have the same address and step. Therefore, if any a&c or b&d don't have the same address and step, we don't care the order of those two pairs after sorting. */ int comp_res; if ((comp_res = compare_tree (DR_BASE_ADDRESS (p11.dr), DR_BASE_ADDRESS (p21.dr))) != 0) return comp_res; if ((comp_res = compare_tree (DR_BASE_ADDRESS (p12.dr), DR_BASE_ADDRESS (p22.dr))) != 0) return comp_res; if ((comp_res = compare_tree (DR_STEP (p11.dr), DR_STEP (p21.dr))) != 0) return comp_res; if ((comp_res = compare_tree (DR_STEP (p12.dr), DR_STEP (p22.dr))) != 0) return comp_res; if ((comp_res = compare_tree (p11.offset, p21.offset)) != 0) return comp_res; if ((comp_res = compare_tree (p12.offset, p22.offset)) != 0) return comp_res; return 0; } template static void swap (T& a, T& b) { T c (a); a = b; b = c; } /* Function vect_vfa_segment_size. Create an expression that computes the size of segment that will be accessed for a data reference. The functions takes into account that realignment loads may access one more vector. Input: DR: The data reference. LENGTH_FACTOR: segment length to consider. Return an expression whose value is the size of segment which will be accessed by DR. */ static tree vect_vfa_segment_size (struct data_reference *dr, tree length_factor) { tree segment_length; if (integer_zerop (DR_STEP (dr))) segment_length = TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr))); else segment_length = size_binop (MULT_EXPR, fold_convert (sizetype, DR_STEP (dr)), fold_convert (sizetype, length_factor)); if (vect_supportable_dr_alignment (dr, false) == dr_explicit_realign_optimized) { tree vector_size = TYPE_SIZE_UNIT (STMT_VINFO_VECTYPE (vinfo_for_stmt (DR_STMT (dr)))); segment_length = size_binop (PLUS_EXPR, segment_length, vector_size); } return segment_length; } /* Function vect_prune_runtime_alias_test_list. Prune a list of ddrs to be tested at run-time by versioning for alias. Merge several alias checks into one if possible. Return FALSE if resulting list of ddrs is longer then allowed by PARAM_VECT_MAX_VERSION_FOR_ALIAS_CHECKS, otherwise return TRUE. */ bool vect_prune_runtime_alias_test_list (loop_vec_info loop_vinfo) { vec may_alias_ddrs = LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo); vec& comp_alias_ddrs = LOOP_VINFO_COMP_ALIAS_DDRS (loop_vinfo); int vect_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo); tree scalar_loop_iters = LOOP_VINFO_NITERS (loop_vinfo); ddr_p ddr; unsigned int i; tree length_factor; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "=== vect_prune_runtime_alias_test_list ===\n"); if (may_alias_ddrs.is_empty ()) return true; /* Basically, for each pair of dependent data refs store_ptr_0 and load_ptr_0, we create an expression: ((store_ptr_0 + store_segment_length_0) <= load_ptr_0) || (load_ptr_0 + load_segment_length_0) <= store_ptr_0)) for aliasing checks. However, in some cases we can decrease the number of checks by combining two checks into one. For example, suppose we have another pair of data refs store_ptr_0 and load_ptr_1, and if the following condition is satisfied: load_ptr_0 < load_ptr_1 && load_ptr_1 - load_ptr_0 - load_segment_length_0 < store_segment_length_0 (this condition means, in each iteration of vectorized loop, the accessed memory of store_ptr_0 cannot be between the memory of load_ptr_0 and load_ptr_1.) we then can use only the following expression to finish the alising checks between store_ptr_0 & load_ptr_0 and store_ptr_0 & load_ptr_1: ((store_ptr_0 + store_segment_length_0) <= load_ptr_0) || (load_ptr_1 + load_segment_length_1 <= store_ptr_0)) Note that we only consider that load_ptr_0 and load_ptr_1 have the same basic address. */ comp_alias_ddrs.create (may_alias_ddrs.length ()); /* First, we collect all data ref pairs for aliasing checks. */ FOR_EACH_VEC_ELT (may_alias_ddrs, i, ddr) { struct data_reference *dr_a, *dr_b; gimple dr_group_first_a, dr_group_first_b; tree segment_length_a, segment_length_b; gimple stmt_a, stmt_b; dr_a = DDR_A (ddr); stmt_a = DR_STMT (DDR_A (ddr)); dr_group_first_a = GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt_a)); if (dr_group_first_a) { stmt_a = dr_group_first_a; dr_a = STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt_a)); } dr_b = DDR_B (ddr); stmt_b = DR_STMT (DDR_B (ddr)); dr_group_first_b = GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt_b)); if (dr_group_first_b) { stmt_b = dr_group_first_b; dr_b = STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt_b)); } if (!operand_equal_p (DR_STEP (dr_a), DR_STEP (dr_b), 0)) length_factor = scalar_loop_iters; else length_factor = size_int (vect_factor); segment_length_a = vect_vfa_segment_size (dr_a, length_factor); segment_length_b = vect_vfa_segment_size (dr_b, length_factor); dr_with_seg_len_pair_t dr_with_seg_len_pair (dr_with_seg_len (dr_a, segment_length_a), dr_with_seg_len (dr_b, segment_length_b)); if (compare_tree (DR_BASE_ADDRESS (dr_a), DR_BASE_ADDRESS (dr_b)) > 0) swap (dr_with_seg_len_pair.first, dr_with_seg_len_pair.second); comp_alias_ddrs.safe_push (dr_with_seg_len_pair); } /* Second, we sort the collected data ref pairs so that we can scan them once to combine all possible aliasing checks. */ comp_alias_ddrs.qsort (comp_dr_with_seg_len_pair); /* Third, we scan the sorted dr pairs and check if we can combine alias checks of two neighbouring dr pairs. */ for (size_t i = 1; i < comp_alias_ddrs.length (); ++i) { /* Deal with two ddrs (dr_a1, dr_b1) and (dr_a2, dr_b2). */ dr_with_seg_len *dr_a1 = &comp_alias_ddrs[i-1].first, *dr_b1 = &comp_alias_ddrs[i-1].second, *dr_a2 = &comp_alias_ddrs[i].first, *dr_b2 = &comp_alias_ddrs[i].second; /* Remove duplicate data ref pairs. */ if (*dr_a1 == *dr_a2 && *dr_b1 == *dr_b2) { if (dump_enabled_p ()) { dump_printf_loc (MSG_NOTE, vect_location, "found equal ranges "); dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dr_a1->dr)); dump_printf (MSG_NOTE, ", "); dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dr_b1->dr)); dump_printf (MSG_NOTE, " and "); dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dr_a2->dr)); dump_printf (MSG_NOTE, ", "); dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dr_b2->dr)); dump_printf (MSG_NOTE, "\n"); } comp_alias_ddrs.ordered_remove (i--); continue; } if (*dr_a1 == *dr_a2 || *dr_b1 == *dr_b2) { /* We consider the case that DR_B1 and DR_B2 are same memrefs, and DR_A1 and DR_A2 are two consecutive memrefs. */ if (*dr_a1 == *dr_a2) { swap (dr_a1, dr_b1); swap (dr_a2, dr_b2); } if (!operand_equal_p (DR_BASE_ADDRESS (dr_a1->dr), DR_BASE_ADDRESS (dr_a2->dr), 0) || !tree_fits_shwi_p (dr_a1->offset) || !tree_fits_shwi_p (dr_a2->offset)) continue; HOST_WIDE_INT diff = (tree_to_shwi (dr_a2->offset) - tree_to_shwi (dr_a1->offset)); /* Now we check if the following condition is satisfied: DIFF - SEGMENT_LENGTH_A < SEGMENT_LENGTH_B where DIFF = DR_A2->OFFSET - DR_A1->OFFSET. However, SEGMENT_LENGTH_A or SEGMENT_LENGTH_B may not be constant so we have to make a best estimation. We can get the minimum value of SEGMENT_LENGTH_B as a constant, represented by MIN_SEG_LEN_B, then either of the following two conditions can guarantee the one above: 1: DIFF <= MIN_SEG_LEN_B 2: DIFF - SEGMENT_LENGTH_A < MIN_SEG_LEN_B */ HOST_WIDE_INT min_seg_len_b = (TREE_CODE (dr_b1->seg_len) == INTEGER_CST) ? TREE_INT_CST_LOW (dr_b1->seg_len) : vect_factor; if (diff <= min_seg_len_b || (TREE_CODE (dr_a1->seg_len) == INTEGER_CST && diff - (HOST_WIDE_INT) TREE_INT_CST_LOW (dr_a1->seg_len) < min_seg_len_b)) { dr_a1->seg_len = size_binop (PLUS_EXPR, dr_a2->seg_len, size_int (diff)); comp_alias_ddrs.ordered_remove (i--); } } } if ((int) comp_alias_ddrs.length () > PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIAS_CHECKS)) { if (dump_enabled_p ()) { dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "disable versioning for alias - max number of " "generated checks exceeded.\n"); } return false; } return true; } /* Check whether a non-affine read in stmt is suitable for gather load and if so, return a builtin decl for that operation. */ tree vect_check_gather (gimple stmt, loop_vec_info loop_vinfo, tree *basep, tree *offp, int *scalep) { HOST_WIDE_INT scale = 1, pbitpos, pbitsize; struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); stmt_vec_info stmt_info = vinfo_for_stmt (stmt); struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); tree offtype = NULL_TREE; tree decl, base, off; enum machine_mode pmode; int punsignedp, pvolatilep; base = DR_REF (dr); /* For masked loads/stores, DR_REF (dr) is an artificial MEM_REF, see if we can use the def stmt of the address. */ if (is_gimple_call (stmt) && gimple_call_internal_p (stmt) && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD || gimple_call_internal_fn (stmt) == IFN_MASK_STORE) && TREE_CODE (base) == MEM_REF && TREE_CODE (TREE_OPERAND (base, 0)) == SSA_NAME && integer_zerop (TREE_OPERAND (base, 1)) && !expr_invariant_in_loop_p (loop, TREE_OPERAND (base, 0))) { gimple def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (base, 0)); if (is_gimple_assign (def_stmt) && gimple_assign_rhs_code (def_stmt) == ADDR_EXPR) base = TREE_OPERAND (gimple_assign_rhs1 (def_stmt), 0); } /* The gather builtins need address of the form loop_invariant + vector * {1, 2, 4, 8} or loop_invariant + sign_extend (vector) * { 1, 2, 4, 8 }. Unfortunately DR_BASE_ADDRESS/DR_OFFSET can be a mixture of loop invariants/SSA_NAMEs defined in the loop, with casts, multiplications and additions in it. To get a vector, we need a single SSA_NAME that will be defined in the loop and will contain everything that is not loop invariant and that can be vectorized. The following code attempts to find such a preexistng SSA_NAME OFF and put the loop invariants into a tree BASE that can be gimplified before the loop. */ base = get_inner_reference (base, &pbitsize, &pbitpos, &off, &pmode, &punsignedp, &pvolatilep, false); gcc_assert (base != NULL_TREE && (pbitpos % BITS_PER_UNIT) == 0); if (TREE_CODE (base) == MEM_REF) { if (!integer_zerop (TREE_OPERAND (base, 1))) { if (off == NULL_TREE) { double_int moff = mem_ref_offset (base); off = double_int_to_tree (sizetype, moff); } else off = size_binop (PLUS_EXPR, off, fold_convert (sizetype, TREE_OPERAND (base, 1))); } base = TREE_OPERAND (base, 0); } else base = build_fold_addr_expr (base); if (off == NULL_TREE) off = size_zero_node; /* If base is not loop invariant, either off is 0, then we start with just the constant offset in the loop invariant BASE and continue with base as OFF, otherwise give up. We could handle that case by gimplifying the addition of base + off into some SSA_NAME and use that as off, but for now punt. */ if (!expr_invariant_in_loop_p (loop, base)) { if (!integer_zerop (off)) return NULL_TREE; off = base; base = size_int (pbitpos / BITS_PER_UNIT); } /* Otherwise put base + constant offset into the loop invariant BASE and continue with OFF. */ else { base = fold_convert (sizetype, base); base = size_binop (PLUS_EXPR, base, size_int (pbitpos / BITS_PER_UNIT)); } /* OFF at this point may be either a SSA_NAME or some tree expression from get_inner_reference. Try to peel off loop invariants from it into BASE as long as possible. */ STRIP_NOPS (off); while (offtype == NULL_TREE) { enum tree_code code; tree op0, op1, add = NULL_TREE; if (TREE_CODE (off) == SSA_NAME) { gimple def_stmt = SSA_NAME_DEF_STMT (off); if (expr_invariant_in_loop_p (loop, off)) return NULL_TREE; if (gimple_code (def_stmt) != GIMPLE_ASSIGN) break; op0 = gimple_assign_rhs1 (def_stmt); code = gimple_assign_rhs_code (def_stmt); op1 = gimple_assign_rhs2 (def_stmt); } else { if (get_gimple_rhs_class (TREE_CODE (off)) == GIMPLE_TERNARY_RHS) return NULL_TREE; code = TREE_CODE (off); extract_ops_from_tree (off, &code, &op0, &op1); } switch (code) { case POINTER_PLUS_EXPR: case PLUS_EXPR: if (expr_invariant_in_loop_p (loop, op0)) { add = op0; off = op1; do_add: add = fold_convert (sizetype, add); if (scale != 1) add = size_binop (MULT_EXPR, add, size_int (scale)); base = size_binop (PLUS_EXPR, base, add); continue; } if (expr_invariant_in_loop_p (loop, op1)) { add = op1; off = op0; goto do_add; } break; case MINUS_EXPR: if (expr_invariant_in_loop_p (loop, op1)) { add = fold_convert (sizetype, op1); add = size_binop (MINUS_EXPR, size_zero_node, add); off = op0; goto do_add; } break; case MULT_EXPR: if (scale == 1 && tree_fits_shwi_p (op1)) { scale = tree_to_shwi (op1); off = op0; continue; } break; case SSA_NAME: off = op0; continue; CASE_CONVERT: if (!POINTER_TYPE_P (TREE_TYPE (op0)) && !INTEGRAL_TYPE_P (TREE_TYPE (op0))) break; if (TYPE_PRECISION (TREE_TYPE (op0)) == TYPE_PRECISION (TREE_TYPE (off))) { off = op0; continue; } if (TYPE_PRECISION (TREE_TYPE (op0)) < TYPE_PRECISION (TREE_TYPE (off))) { off = op0; offtype = TREE_TYPE (off); STRIP_NOPS (off); continue; } break; default: break; } break; } /* If at the end OFF still isn't a SSA_NAME or isn't defined in the loop, punt. */ if (TREE_CODE (off) != SSA_NAME || expr_invariant_in_loop_p (loop, off)) return NULL_TREE; if (offtype == NULL_TREE) offtype = TREE_TYPE (off); decl = targetm.vectorize.builtin_gather (STMT_VINFO_VECTYPE (stmt_info), offtype, scale); if (decl == NULL_TREE) return NULL_TREE; if (basep) *basep = base; if (offp) *offp = off; if (scalep) *scalep = scale; return decl; } /* Function vect_analyze_data_refs. Find all the data references in the loop or basic block. The general structure of the analysis of data refs in the vectorizer is as follows: 1- vect_analyze_data_refs(loop/bb): call compute_data_dependences_for_loop/bb to find and analyze all data-refs in the loop/bb and their dependences. 2- vect_analyze_dependences(): apply dependence testing using ddrs. 3- vect_analyze_drs_alignment(): check that ref_stmt.alignment is ok. 4- vect_analyze_drs_access(): check that ref_stmt.step is ok. */ bool vect_analyze_data_refs (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo, int *min_vf) { struct loop *loop = NULL; basic_block bb = NULL; unsigned int i; vec datarefs; struct data_reference *dr; tree scalar_type; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location, "=== vect_analyze_data_refs ===\n"); if (loop_vinfo) { basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); loop = LOOP_VINFO_LOOP (loop_vinfo); datarefs = LOOP_VINFO_DATAREFS (loop_vinfo); if (!find_loop_nest (loop, &LOOP_VINFO_LOOP_NEST (loop_vinfo))) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "not vectorized: loop contains function calls" " or data references that cannot be analyzed\n"); return false; } for (i = 0; i < loop->num_nodes; i++) { gimple_stmt_iterator gsi; for (gsi = gsi_start_bb (bbs[i]); !gsi_end_p (gsi); gsi_next (&gsi)) { gimple stmt = gsi_stmt (gsi); if (!find_data_references_in_stmt (loop, stmt, &datarefs)) { if (is_gimple_call (stmt) && loop->safelen) { tree fndecl = gimple_call_fndecl (stmt), op; if (fndecl != NULL_TREE) { struct cgraph_node *node = cgraph_get_node (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; } } } LOOP_VINFO_DATAREFS (loop_vinfo) = datarefs; if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "not vectorized: loop contains function " "calls or data references that cannot " "be analyzed\n"); return false; } } } LOOP_VINFO_DATAREFS (loop_vinfo) = datarefs; } else { gimple_stmt_iterator gsi; bb = BB_VINFO_BB (bb_vinfo); for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) { gimple stmt = gsi_stmt (gsi); if (!find_data_references_in_stmt (NULL, stmt, &BB_VINFO_DATAREFS (bb_vinfo))) { /* Mark the rest of the basic-block as unvectorizable. */ for (; !gsi_end_p (gsi); gsi_next (&gsi)) { stmt = gsi_stmt (gsi); STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (stmt)) = false; } break; } } datarefs = BB_VINFO_DATAREFS (bb_vinfo); } /* Go through the data-refs, check that the analysis succeeded. Update pointer from stmt_vec_info struct to DR and vectype. */ FOR_EACH_VEC_ELT (datarefs, i, dr) { gimple stmt; stmt_vec_info stmt_info; tree base, offset, init; bool gather = false; bool simd_lane_access = false; int vf; again: if (!dr || !DR_REF (dr)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "not vectorized: unhandled data-ref\n"); return false; } stmt = DR_STMT (dr); stmt_info = vinfo_for_stmt (stmt); /* Discard clobbers from the dataref vector. We will remove clobber stmts during vectorization. */ if (gimple_clobber_p (stmt)) { if (i == datarefs.length () - 1) { datarefs.pop (); break; } datarefs[i] = datarefs.pop (); goto again; } /* Check that analysis of the data-ref succeeded. */ if (!DR_BASE_ADDRESS (dr) || !DR_OFFSET (dr) || !DR_INIT (dr) || !DR_STEP (dr)) { bool maybe_gather = DR_IS_READ (dr) && !TREE_THIS_VOLATILE (DR_REF (dr)) && targetm.vectorize.builtin_gather != NULL; bool maybe_simd_lane_access = loop_vinfo && loop->simduid; /* If target supports vector gather loads, or if this might be a SIMD lane access, see if they can't be used. */ if (loop_vinfo && (maybe_gather || maybe_simd_lane_access) && !nested_in_vect_loop_p (loop, stmt)) { struct data_reference *newdr = create_data_ref (NULL, loop_containing_stmt (stmt), DR_REF (dr), stmt, true); gcc_assert (newdr != NULL && DR_REF (newdr)); if (DR_BASE_ADDRESS (newdr) && DR_OFFSET (newdr) && DR_INIT (newdr) && DR_STEP (newdr) && integer_zerop (DR_STEP (newdr))) { if (maybe_simd_lane_access) { tree off = DR_OFFSET (newdr); STRIP_NOPS (off); if (TREE_CODE (DR_INIT (newdr)) == INTEGER_CST && TREE_CODE (off) == MULT_EXPR && tree_fits_uhwi_p (TREE_OPERAND (off, 1))) { tree step = TREE_OPERAND (off, 1); off = TREE_OPERAND (off, 0); STRIP_NOPS (off); if (CONVERT_EXPR_P (off) && TYPE_PRECISION (TREE_TYPE (TREE_OPERAND (off, 0))) < TYPE_PRECISION (TREE_TYPE (off))) off = TREE_OPERAND (off, 0); if (TREE_CODE (off) == SSA_NAME) { gimple def = SSA_NAME_DEF_STMT (off); tree reft = TREE_TYPE (DR_REF (newdr)); if (gimple_call_internal_p (def) && gimple_call_internal_fn (def) == IFN_GOMP_SIMD_LANE) { tree arg = gimple_call_arg (def, 0); gcc_assert (TREE_CODE (arg) == SSA_NAME); arg = SSA_NAME_VAR (arg); if (arg == loop->simduid /* For now. */ && tree_int_cst_equal (TYPE_SIZE_UNIT (reft), step)) { DR_OFFSET (newdr) = ssize_int (0); DR_STEP (newdr) = step; DR_ALIGNED_TO (newdr) = size_int (BIGGEST_ALIGNMENT); dr = newdr; simd_lane_access = true; } } } } } if (!simd_lane_access && maybe_gather) { dr = newdr; gather = true; } } if (!gather && !simd_lane_access) free_data_ref (newdr); } if (!gather && !simd_lane_access) { if (dump_enabled_p ()) { dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "not vectorized: data ref analysis " "failed "); dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); } if (bb_vinfo) break; return false; } } if (TREE_CODE (DR_BASE_ADDRESS (dr)) == INTEGER_CST) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "not vectorized: base addr of dr is a " "constant\n"); if (bb_vinfo) break; if (gather || simd_lane_access) free_data_ref (dr); return false; } if (TREE_THIS_VOLATILE (DR_REF (dr))) { if (dump_enabled_p ()) { dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "not vectorized: volatile type "); dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); } if (bb_vinfo) break; return false; } if (stmt_can_throw_internal (stmt)) { if (dump_enabled_p ()) { dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "not vectorized: statement can throw an " "exception "); dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); } if (bb_vinfo) break; if (gather || simd_lane_access) free_data_ref (dr); return false; } if (TREE_CODE (DR_REF (dr)) == COMPONENT_REF && DECL_BIT_FIELD (TREE_OPERAND (DR_REF (dr), 1))) { if (dump_enabled_p ()) { dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "not vectorized: statement is bitfield " "access "); dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); } if (bb_vinfo) break; if (gather || simd_lane_access) free_data_ref (dr); return false; } base = unshare_expr (DR_BASE_ADDRESS (dr)); offset = unshare_expr (DR_OFFSET (dr)); init = unshare_expr (DR_INIT (dr)); if (is_gimple_call (stmt) && (!gimple_call_internal_p (stmt) || (gimple_call_internal_fn (stmt) != IFN_MASK_LOAD && gimple_call_internal_fn (stmt) != IFN_MASK_STORE))) { if (dump_enabled_p ()) { dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "not vectorized: dr in a call "); dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); } if (bb_vinfo) break; if (gather || simd_lane_access) free_data_ref (dr); return false; } /* Update DR field in stmt_vec_info struct. */ /* If the dataref is in an inner-loop of the loop that is considered for for vectorization, we also want to analyze the access relative to the outer-loop (DR contains information only relative to the inner-most enclosing loop). We do that by building a reference to the first location accessed by the inner-loop, and analyze it relative to the outer-loop. */ if (loop && nested_in_vect_loop_p (loop, stmt)) { tree outer_step, outer_base, outer_init; HOST_WIDE_INT pbitsize, pbitpos; tree poffset; enum machine_mode pmode; int punsignedp, pvolatilep; affine_iv base_iv, offset_iv; tree dinit; /* Build a reference to the first location accessed by the inner-loop: *(BASE+INIT). (The first location is actually BASE+INIT+OFFSET, but we add OFFSET separately later). */ tree inner_base = build_fold_indirect_ref (fold_build_pointer_plus (base, init)); if (dump_enabled_p ()) { dump_printf_loc (MSG_NOTE, vect_location, "analyze in outer-loop: "); dump_generic_expr (MSG_NOTE, TDF_SLIM, inner_base); dump_printf (MSG_NOTE, "\n"); } outer_base = get_inner_reference (inner_base, &pbitsize, &pbitpos, &poffset, &pmode, &punsignedp, &pvolatilep, false); gcc_assert (outer_base != NULL_TREE); if (pbitpos % BITS_PER_UNIT != 0) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "failed: bit offset alignment.\n"); return false; } outer_base = build_fold_addr_expr (outer_base); if (!simple_iv (loop, loop_containing_stmt (stmt), outer_base, &base_iv, false)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "failed: evolution of base is not affine.\n"); return false; } if (offset) { if (poffset) poffset = fold_build2 (PLUS_EXPR, TREE_TYPE (offset), offset, poffset); else poffset = offset; } if (!poffset) { offset_iv.base = ssize_int (0); offset_iv.step = ssize_int (0); } else if (!simple_iv (loop, loop_containing_stmt (stmt), poffset, &offset_iv, false)) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "evolution of offset is not affine.\n"); return false; } outer_init = ssize_int (pbitpos / BITS_PER_UNIT); split_constant_offset (base_iv.base, &base_iv.base, &dinit); outer_init = size_binop (PLUS_EXPR, outer_init, dinit); split_constant_offset (offset_iv.base, &offset_iv.base, &dinit); outer_init = size_binop (PLUS_EXPR, outer_init, dinit); outer_step = size_binop (PLUS_EXPR, fold_convert (ssizetype, base_iv.step), fold_convert (ssizetype, offset_iv.step)); STMT_VINFO_DR_STEP (stmt_info) = outer_step; /* FIXME: Use canonicalize_base_object_address (base_iv.base); */ STMT_VINFO_DR_BASE_ADDRESS (stmt_info) = base_iv.base; STMT_VINFO_DR_INIT (stmt_info) = outer_init; STMT_VINFO_DR_OFFSET (stmt_info) = fold_convert (ssizetype, offset_iv.base); STMT_VINFO_DR_ALIGNED_TO (stmt_info) = size_int (highest_pow2_factor (offset_iv.base)); if (dump_enabled_p ()) { dump_printf_loc (MSG_NOTE, vect_location, "\touter base_address: "); dump_generic_expr (MSG_NOTE, TDF_SLIM, STMT_VINFO_DR_BASE_ADDRESS (stmt_info)); dump_printf (MSG_NOTE, "\n\touter offset from base address: "); dump_generic_expr (MSG_NOTE, TDF_SLIM, STMT_VINFO_DR_OFFSET (stmt_info)); dump_printf (MSG_NOTE, "\n\touter constant offset from base address: "); dump_generic_expr (MSG_NOTE, TDF_SLIM, STMT_VINFO_DR_INIT (stmt_info)); dump_printf (MSG_NOTE, "\n\touter step: "); dump_generic_expr (MSG_NOTE, TDF_SLIM, STMT_VINFO_DR_STEP (stmt_info)); dump_printf (MSG_NOTE, "\n\touter aligned to: "); dump_generic_expr (MSG_NOTE, TDF_SLIM, STMT_VINFO_DR_ALIGNED_TO (stmt_info)); dump_printf (MSG_NOTE, "\n"); } } if (STMT_VINFO_DATA_REF (stmt_info)) { if (dump_enabled_p ()) { dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "not vectorized: more than one data ref " "in stmt: "); dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); } if (bb_vinfo) break; if (gather || simd_lane_access) free_data_ref (dr); return false; } STMT_VINFO_DATA_REF (stmt_info) = dr; if (simd_lane_access) { STMT_VINFO_SIMD_LANE_ACCESS_P (stmt_info) = true; datarefs[i] = dr; } /* Set vectype for STMT. */ scalar_type = TREE_TYPE (DR_REF (dr)); STMT_VINFO_VECTYPE (stmt_info) = get_vectype_for_scalar_type (scalar_type); if (!STMT_VINFO_VECTYPE (stmt_info)) { if (dump_enabled_p ()) { dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "not vectorized: no vectype for stmt: "); dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); dump_printf (MSG_MISSED_OPTIMIZATION, " scalar_type: "); dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_DETAILS, scalar_type); dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); } if (bb_vinfo) break; if (gather || simd_lane_access) { STMT_VINFO_DATA_REF (stmt_info) = NULL; free_data_ref (dr); } return false; } else { if (dump_enabled_p ()) { dump_printf_loc (MSG_NOTE, vect_location, "got vectype for stmt: "); dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0); dump_generic_expr (MSG_NOTE, TDF_SLIM, STMT_VINFO_VECTYPE (stmt_info)); dump_printf (MSG_NOTE, "\n"); } } /* Adjust the minimal vectorization factor according to the vector type. */ vf = TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info)); if (vf > *min_vf) *min_vf = vf; if (gather) { tree off; gather = 0 != vect_check_gather (stmt, loop_vinfo, NULL, &off, NULL); if (gather && get_vectype_for_scalar_type (TREE_TYPE (off)) == NULL_TREE) gather = false; if (!gather) { STMT_VINFO_DATA_REF (stmt_info) = NULL; free_data_ref (dr); if (dump_enabled_p ()) { dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "not vectorized: not suitable for gather " "load "); dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); } return false; } datarefs[i] = dr; STMT_VINFO_GATHER_P (stmt_info) = true; } else if (loop_vinfo && TREE_CODE (DR_STEP (dr)) != INTEGER_CST) { if (nested_in_vect_loop_p (loop, stmt) || !DR_IS_READ (dr)) { if (dump_enabled_p ()) { dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "not vectorized: not suitable for strided " "load "); dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0); dump_printf (MSG_MISSED_OPTIMIZATION, "\n"); } return false; } STMT_VINFO_STRIDE_LOAD_P (stmt_info) = true; } } /* If we stopped analysis at the first dataref we could not analyze when trying to vectorize a basic-block mark the rest of the datarefs as not vectorizable and truncate the vector of datarefs. That avoids spending useless time in analyzing their dependence. */ if (i != datarefs.length ()) { gcc_assert (bb_vinfo != NULL); for (unsigned j = i; j < datarefs.length (); ++j) { data_reference_p dr = datarefs[j]; STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (DR_STMT (dr))) = false; free_data_ref (dr); } datarefs.truncate (i); } return true; } /* Function vect_get_new_vect_var. Returns a name for a new variable. The current naming scheme appends the prefix "vect_" or "vect_p" (depending on the value of VAR_KIND) to the name of vectorizer generated variables, and appends that to NAME if provided. */ tree vect_get_new_vect_var (tree type, enum vect_var_kind var_kind, const char *name) { const char *prefix; tree new_vect_var; switch (var_kind) { case vect_simple_var: prefix = "vect"; break; case vect_scalar_var: prefix = "stmp"; break; case vect_pointer_var: prefix = "vectp"; break; default: gcc_unreachable (); } if (name) { char* tmp = concat (prefix, "_", name, NULL); new_vect_var = create_tmp_reg (type, tmp); free (tmp); } else new_vect_var = create_tmp_reg (type, prefix); return new_vect_var; } /* Function vect_create_addr_base_for_vector_ref. Create an expression that computes the address of the first memory location that will be accessed for a data reference. Input: STMT: The statement containing the data reference. NEW_STMT_LIST: Must be initialized to NULL_TREE or a statement list. OFFSET: Optional. If supplied, it is be added to the initial address. LOOP: Specify relative to which loop-nest should the address be computed. For example, when the dataref is in an inner-loop nested in an outer-loop that is now being vectorized, LOOP can be either the outer-loop, or the inner-loop. The first memory location accessed by the following dataref ('in' points to short): for (i=0; iloop_father) { struct loop *outer_loop = LOOP_VINFO_LOOP (loop_vinfo); gcc_assert (nested_in_vect_loop_p (outer_loop, stmt)); data_ref_base = unshare_expr (STMT_VINFO_DR_BASE_ADDRESS (stmt_info)); base_offset = unshare_expr (STMT_VINFO_DR_OFFSET (stmt_info)); init = unshare_expr (STMT_VINFO_DR_INIT (stmt_info)); } else { data_ref_base = unshare_expr (DR_BASE_ADDRESS (dr)); base_offset = unshare_expr (DR_OFFSET (dr)); init = unshare_expr (DR_INIT (dr)); } if (loop_vinfo) base_name = get_name (data_ref_base); else { base_offset = ssize_int (0); init = ssize_int (0); base_name = get_name (DR_REF (dr)); } /* Create base_offset */ base_offset = size_binop (PLUS_EXPR, fold_convert (sizetype, base_offset), fold_convert (sizetype, init)); if (offset) { offset = fold_build2 (MULT_EXPR, sizetype, fold_convert (sizetype, offset), step); base_offset = fold_build2 (PLUS_EXPR, sizetype, base_offset, offset); } /* base + base_offset */ if (loop_vinfo) addr_base = fold_build_pointer_plus (data_ref_base, base_offset); else { addr_base = build1 (ADDR_EXPR, build_pointer_type (TREE_TYPE (DR_REF (dr))), unshare_expr (DR_REF (dr))); } vect_ptr_type = build_pointer_type (STMT_VINFO_VECTYPE (stmt_info)); addr_base = fold_convert (vect_ptr_type, addr_base); dest = vect_get_new_vect_var (vect_ptr_type, vect_pointer_var, base_name); addr_base = force_gimple_operand (addr_base, &seq, false, dest); gimple_seq_add_seq (new_stmt_list, seq); if (DR_PTR_INFO (dr) && TREE_CODE (addr_base) == SSA_NAME) { duplicate_ssa_name_ptr_info (addr_base, DR_PTR_INFO (dr)); if (offset) mark_ptr_info_alignment_unknown (SSA_NAME_PTR_INFO (addr_base)); } if (dump_enabled_p ()) { dump_printf_loc (MSG_NOTE, vect_location, "created "); dump_generic_expr (MSG_NOTE, TDF_SLIM, addr_base); dump_printf (MSG_NOTE, "\n"); } return addr_base; } /* Function vect_create_data_ref_ptr. Create a new pointer-to-AGGR_TYPE variable (ap), that points to the first location accessed in the loop by STMT, along with the def-use update chain to appropriately advance the pointer through the loop iterations. Also set aliasing information for the pointer. This pointer is used by the callers to this function to create a memory reference expression for vector load/store access. Input: 1. STMT: a stmt that references memory. Expected to be of the form GIMPLE_ASSIGN or GIMPLE_ASSIGN . 2. AGGR_TYPE: the type of the reference, which should be either a vector or an array. 3. AT_LOOP: the loop where the vector memref is to be created. 4. OFFSET (optional): an offset to be added to the initial address accessed by the data-ref in STMT. 5. BSI: location where the new stmts are to be placed if there is no loop 6. ONLY_INIT: indicate if ap is to be updated in the loop, or remain pointing to the initial address. Output: 1. Declare a new ptr to vector_type, and have it point to the base of the data reference (initial addressed accessed by the data reference). For example, for vector of type V8HI, the following code is generated: v8hi *ap; ap = (v8hi *)initial_address; if OFFSET is not supplied: initial_address = &a[init]; if OFFSET is supplied: initial_address = &a[init + OFFSET]; Return the initial_address in INITIAL_ADDRESS. 2. If ONLY_INIT is true, just return the initial pointer. Otherwise, also update the pointer in each iteration of the loop. Return the increment stmt that updates the pointer in PTR_INCR. 3. Set INV_P to true if the access pattern of the data reference in the vectorized loop is invariant. Set it to false otherwise. 4. Return the pointer. */ tree vect_create_data_ref_ptr (gimple stmt, tree aggr_type, struct loop *at_loop, tree offset, tree *initial_address, gimple_stmt_iterator *gsi, gimple *ptr_incr, bool only_init, bool *inv_p) { const char *base_name; stmt_vec_info stmt_info = vinfo_for_stmt (stmt); loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); struct loop *loop = NULL; bool nested_in_vect_loop = false; struct loop *containing_loop = NULL; tree aggr_ptr_type; tree aggr_ptr; tree new_temp; gimple vec_stmt; gimple_seq new_stmt_list = NULL; edge pe = NULL; basic_block new_bb; tree aggr_ptr_init; struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); tree aptr; gimple_stmt_iterator incr_gsi; bool insert_after; tree indx_before_incr, indx_after_incr; gimple incr; tree step; bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); gcc_assert (TREE_CODE (aggr_type) == ARRAY_TYPE || TREE_CODE (aggr_type) == VECTOR_TYPE); if (loop_vinfo) { loop = LOOP_VINFO_LOOP (loop_vinfo); nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); containing_loop = (gimple_bb (stmt))->loop_father; pe = loop_preheader_edge (loop); } else { gcc_assert (bb_vinfo); only_init = true; *ptr_incr = NULL; } /* Check the step (evolution) of the load in LOOP, and record whether it's invariant. */ if (nested_in_vect_loop) step = STMT_VINFO_DR_STEP (stmt_info); else step = DR_STEP (STMT_VINFO_DATA_REF (stmt_info)); if (integer_zerop (step)) *inv_p = true; else *inv_p = false; /* Create an expression for the first address accessed by this load in LOOP. */ base_name = get_name (DR_BASE_ADDRESS (dr)); if (dump_enabled_p ()) { tree dr_base_type = TREE_TYPE (DR_BASE_OBJECT (dr)); dump_printf_loc (MSG_NOTE, vect_location, "create %s-pointer variable to type: ", get_tree_code_name (TREE_CODE (aggr_type))); dump_generic_expr (MSG_NOTE, TDF_SLIM, aggr_type); if (TREE_CODE (dr_base_type) == ARRAY_TYPE) dump_printf (MSG_NOTE, " vectorizing an array ref: "); else if (TREE_CODE (dr_base_type) == VECTOR_TYPE) dump_printf (MSG_NOTE, " vectorizing a vector ref: "); else if (TREE_CODE (dr_base_type) == RECORD_TYPE) dump_printf (MSG_NOTE, " vectorizing a record based array ref: "); else dump_printf (MSG_NOTE, " vectorizing a pointer ref: "); dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_BASE_OBJECT (dr)); dump_printf (MSG_NOTE, "\n"); } /* (1) Create the new aggregate-pointer variable. Vector and array types inherit the alias set of their component type by default so we need to use a ref-all pointer if the data reference does not conflict with the created aggregated data reference because it is not addressable. */ bool need_ref_all = false; if (!alias_sets_conflict_p (get_alias_set (aggr_type), get_alias_set (DR_REF (dr)))) need_ref_all = true; /* Likewise for any of the data references in the stmt group. */ else if (STMT_VINFO_GROUP_SIZE (stmt_info) > 1) { gimple orig_stmt = STMT_VINFO_GROUP_FIRST_ELEMENT (stmt_info); do { stmt_vec_info sinfo = vinfo_for_stmt (orig_stmt); struct data_reference *sdr = STMT_VINFO_DATA_REF (sinfo); if (!alias_sets_conflict_p (get_alias_set (aggr_type), get_alias_set (DR_REF (sdr)))) { need_ref_all = true; break; } orig_stmt = STMT_VINFO_GROUP_NEXT_ELEMENT (sinfo); } while (orig_stmt); } aggr_ptr_type = build_pointer_type_for_mode (aggr_type, ptr_mode, need_ref_all); aggr_ptr = vect_get_new_vect_var (aggr_ptr_type, vect_pointer_var, base_name); /* Note: If the dataref is in an inner-loop nested in LOOP, and we are vectorizing LOOP (i.e., outer-loop vectorization), we need to create two def-use update cycles for the pointer: one relative to the outer-loop (LOOP), which is what steps (3) and (4) below do. The other is relative to the inner-loop (which is the inner-most loop containing the dataref), and this is done be step (5) below. When vectorizing inner-most loops, the vectorized loop (LOOP) is also the inner-most loop, and so steps (3),(4) work the same, and step (5) is redundant. Steps (3),(4) create the following: vp0 = &base_addr; LOOP: vp1 = phi(vp0,vp2) ... ... vp2 = vp1 + step goto LOOP If there is an inner-loop nested in loop, then step (5) will also be applied, and an additional update in the inner-loop will be created: vp0 = &base_addr; LOOP: vp1 = phi(vp0,vp2) ... inner: vp3 = phi(vp1,vp4) vp4 = vp3 + inner_step if () goto inner ... vp2 = vp1 + step if () goto LOOP */ /* (2) Calculate the initial address of the aggregate-pointer, and set the aggregate-pointer to point to it before the loop. */ /* Create: (&(base[init_val+offset]) in the loop preheader. */ new_temp = vect_create_addr_base_for_vector_ref (stmt, &new_stmt_list, offset, loop); if (new_stmt_list) { if (pe) { new_bb = gsi_insert_seq_on_edge_immediate (pe, new_stmt_list); gcc_assert (!new_bb); } else gsi_insert_seq_before (gsi, new_stmt_list, GSI_SAME_STMT); } *initial_address = new_temp; /* Create: p = (aggr_type *) initial_base */ if (TREE_CODE (new_temp) != SSA_NAME || !useless_type_conversion_p (aggr_ptr_type, TREE_TYPE (new_temp))) { vec_stmt = gimple_build_assign (aggr_ptr, fold_convert (aggr_ptr_type, new_temp)); aggr_ptr_init = make_ssa_name (aggr_ptr, vec_stmt); /* Copy the points-to information if it exists. */ if (DR_PTR_INFO (dr)) duplicate_ssa_name_ptr_info (aggr_ptr_init, DR_PTR_INFO (dr)); gimple_assign_set_lhs (vec_stmt, aggr_ptr_init); if (pe) { new_bb = gsi_insert_on_edge_immediate (pe, vec_stmt); gcc_assert (!new_bb); } else gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT); } else aggr_ptr_init = new_temp; /* (3) Handle the updating of the aggregate-pointer inside the loop. This is needed when ONLY_INIT is false, and also when AT_LOOP is the inner-loop nested in LOOP (during outer-loop vectorization). */ /* No update in loop is required. */ if (only_init && (!loop_vinfo || at_loop == loop)) aptr = aggr_ptr_init; else { /* The step of the aggregate pointer is the type size. */ tree iv_step = TYPE_SIZE_UNIT (aggr_type); /* One exception to the above is when the scalar step of the load in LOOP is zero. In this case the step here is also zero. */ if (*inv_p) iv_step = size_zero_node; else if (tree_int_cst_sgn (step) == -1) iv_step = fold_build1 (NEGATE_EXPR, TREE_TYPE (iv_step), iv_step); standard_iv_increment_position (loop, &incr_gsi, &insert_after); create_iv (aggr_ptr_init, fold_convert (aggr_ptr_type, iv_step), aggr_ptr, loop, &incr_gsi, insert_after, &indx_before_incr, &indx_after_incr); incr = gsi_stmt (incr_gsi); set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo, NULL)); /* Copy the points-to information if it exists. */ if (DR_PTR_INFO (dr)) { duplicate_ssa_name_ptr_info (indx_before_incr, DR_PTR_INFO (dr)); duplicate_ssa_name_ptr_info (indx_after_incr, DR_PTR_INFO (dr)); } if (ptr_incr) *ptr_incr = incr; aptr = indx_before_incr; } if (!nested_in_vect_loop || only_init) return aptr; /* (4) Handle the updating of the aggregate-pointer inside the inner-loop nested in LOOP, if exists. */ gcc_assert (nested_in_vect_loop); if (!only_init) { standard_iv_increment_position (containing_loop, &incr_gsi, &insert_after); create_iv (aptr, fold_convert (aggr_ptr_type, DR_STEP (dr)), aggr_ptr, containing_loop, &incr_gsi, insert_after, &indx_before_incr, &indx_after_incr); incr = gsi_stmt (incr_gsi); set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo, NULL)); /* Copy the points-to information if it exists. */ if (DR_PTR_INFO (dr)) { duplicate_ssa_name_ptr_info (indx_before_incr, DR_PTR_INFO (dr)); duplicate_ssa_name_ptr_info (indx_after_incr, DR_PTR_INFO (dr)); } if (ptr_incr) *ptr_incr = incr; return indx_before_incr; } else gcc_unreachable (); } /* Function bump_vector_ptr Increment a pointer (to a vector type) by vector-size. If requested, i.e. if PTR-INCR is given, then also connect the new increment stmt to the existing def-use update-chain of the pointer, by modifying the PTR_INCR as illustrated below: The pointer def-use update-chain before this function: DATAREF_PTR = phi (p_0, p_2) .... PTR_INCR: p_2 = DATAREF_PTR + step The pointer def-use update-chain after this function: DATAREF_PTR = phi (p_0, p_2) .... NEW_DATAREF_PTR = DATAREF_PTR + BUMP .... PTR_INCR: p_2 = NEW_DATAREF_PTR + step Input: DATAREF_PTR - ssa_name of a pointer (to vector type) that is being updated in the loop. PTR_INCR - optional. The stmt that updates the pointer in each iteration of the loop. The increment amount across iterations is expected to be vector_size. BSI - location where the new update stmt is to be placed. STMT - the original scalar memory-access stmt that is being vectorized. BUMP - optional. The offset by which to bump the pointer. If not given, the offset is assumed to be vector_size. Output: Return NEW_DATAREF_PTR as illustrated above. */ tree bump_vector_ptr (tree dataref_ptr, gimple ptr_incr, gimple_stmt_iterator *gsi, gimple stmt, tree bump) { stmt_vec_info stmt_info = vinfo_for_stmt (stmt); struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); tree vectype = STMT_VINFO_VECTYPE (stmt_info); tree update = TYPE_SIZE_UNIT (vectype); gimple incr_stmt; ssa_op_iter iter; use_operand_p use_p; tree new_dataref_ptr; if (bump) update = bump; new_dataref_ptr = copy_ssa_name (dataref_ptr, NULL); incr_stmt = gimple_build_assign_with_ops (POINTER_PLUS_EXPR, new_dataref_ptr, dataref_ptr, update); vect_finish_stmt_generation (stmt, incr_stmt, gsi); /* Copy the points-to information if it exists. */ if (DR_PTR_INFO (dr)) { duplicate_ssa_name_ptr_info (new_dataref_ptr, DR_PTR_INFO (dr)); mark_ptr_info_alignment_unknown (SSA_NAME_PTR_INFO (new_dataref_ptr)); } if (!ptr_incr) return new_dataref_ptr; /* Update the vector-pointer's cross-iteration increment. */ FOR_EACH_SSA_USE_OPERAND (use_p, ptr_incr, iter, SSA_OP_USE) { tree use = USE_FROM_PTR (use_p); if (use == dataref_ptr) SET_USE (use_p, new_dataref_ptr); else gcc_assert (tree_int_cst_compare (use, update) == 0); } return new_dataref_ptr; } /* Function vect_create_destination_var. Create a new temporary of type VECTYPE. */ tree vect_create_destination_var (tree scalar_dest, tree vectype) { tree vec_dest; const char *name; char *new_name; tree type; enum vect_var_kind kind; kind = vectype ? vect_simple_var : vect_scalar_var; type = vectype ? vectype : TREE_TYPE (scalar_dest); gcc_assert (TREE_CODE (scalar_dest) == SSA_NAME); name = get_name (scalar_dest); if (name) asprintf (&new_name, "%s_%u", name, SSA_NAME_VERSION (scalar_dest)); else asprintf (&new_name, "_%u", SSA_NAME_VERSION (scalar_dest)); vec_dest = vect_get_new_vect_var (type, kind, new_name); free (new_name); return vec_dest; } /* Function vect_grouped_store_supported. Returns TRUE if interleave high and interleave low permutations are supported, and FALSE otherwise. */ bool vect_grouped_store_supported (tree vectype, unsigned HOST_WIDE_INT count) { enum machine_mode mode = TYPE_MODE (vectype); /* vect_permute_store_chain requires the group size to be a power of two. */ if (exact_log2 (count) == -1) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "the size of the group of accesses" " is not a power of 2\n"); return false; } /* Check that the permutation is supported. */ if (VECTOR_MODE_P (mode)) { unsigned int i, nelt = GET_MODE_NUNITS (mode); unsigned char *sel = XALLOCAVEC (unsigned char, nelt); for (i = 0; i < nelt / 2; i++) { sel[i * 2] = i; sel[i * 2 + 1] = i + nelt; } if (can_vec_perm_p (mode, false, sel)) { for (i = 0; i < nelt; i++) sel[i] += nelt / 2; if (can_vec_perm_p (mode, false, sel)) return true; } } if (dump_enabled_p ()) dump_printf (MSG_MISSED_OPTIMIZATION, "interleave op not supported by target.\n"); return false; } /* Return TRUE if vec_store_lanes is available for COUNT vectors of type VECTYPE. */ bool vect_store_lanes_supported (tree vectype, unsigned HOST_WIDE_INT count) { return vect_lanes_optab_supported_p ("vec_store_lanes", vec_store_lanes_optab, vectype, count); } /* Function vect_permute_store_chain. Given a chain of interleaved stores in DR_CHAIN of LENGTH that must be a power of 2, generate interleave_high/low stmts to reorder the data correctly for the stores. Return the final references for stores in RESULT_CHAIN. E.g., LENGTH is 4 and the scalar type is short, i.e., VF is 8. The input is 4 vectors each containing 8 elements. We assign a number to each element, the input sequence is: 1st vec: 0 1 2 3 4 5 6 7 2nd vec: 8 9 10 11 12 13 14 15 3rd vec: 16 17 18 19 20 21 22 23 4th vec: 24 25 26 27 28 29 30 31 The output sequence should be: 1st vec: 0 8 16 24 1 9 17 25 2nd vec: 2 10 18 26 3 11 19 27 3rd vec: 4 12 20 28 5 13 21 30 4th vec: 6 14 22 30 7 15 23 31 i.e., we interleave the contents of the four vectors in their order. We use interleave_high/low instructions to create such output. The input of each interleave_high/low operation is two vectors: 1st vec 2nd vec 0 1 2 3 4 5 6 7 the even elements of the result vector are obtained left-to-right from the high/low elements of the first vector. The odd elements of the result are obtained left-to-right from the high/low elements of the second vector. The output of interleave_high will be: 0 4 1 5 and of interleave_low: 2 6 3 7 The permutation is done in log LENGTH stages. In each stage interleave_high and interleave_low stmts are created for each pair of vectors in DR_CHAIN, where the first argument is taken from the first half of DR_CHAIN and the second argument from it's second half. In our example, I1: interleave_high (1st vec, 3rd vec) I2: interleave_low (1st vec, 3rd vec) I3: interleave_high (2nd vec, 4th vec) I4: interleave_low (2nd vec, 4th vec) The output for the first stage is: I1: 0 16 1 17 2 18 3 19 I2: 4 20 5 21 6 22 7 23 I3: 8 24 9 25 10 26 11 27 I4: 12 28 13 29 14 30 15 31 The output of the second stage, i.e. the final result is: I1: 0 8 16 24 1 9 17 25 I2: 2 10 18 26 3 11 19 27 I3: 4 12 20 28 5 13 21 30 I4: 6 14 22 30 7 15 23 31. */ void vect_permute_store_chain (vec dr_chain, unsigned int length, gimple stmt, gimple_stmt_iterator *gsi, vec *result_chain) { tree vect1, vect2, high, low; gimple perm_stmt; tree vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt)); tree perm_mask_low, perm_mask_high; unsigned int i, n; unsigned int j, nelt = TYPE_VECTOR_SUBPARTS (vectype); unsigned char *sel = XALLOCAVEC (unsigned char, nelt); result_chain->quick_grow (length); memcpy (result_chain->address (), dr_chain.address (), length * sizeof (tree)); for (i = 0, n = nelt / 2; i < n; i++) { sel[i * 2] = i; sel[i * 2 + 1] = i + nelt; } perm_mask_high = vect_gen_perm_mask (vectype, sel); gcc_assert (perm_mask_high != NULL); for (i = 0; i < nelt; i++) sel[i] += nelt / 2; perm_mask_low = vect_gen_perm_mask (vectype, sel); gcc_assert (perm_mask_low != NULL); for (i = 0, n = exact_log2 (length); i < n; i++) { for (j = 0; j < length/2; j++) { vect1 = dr_chain[j]; vect2 = dr_chain[j+length/2]; /* Create interleaving stmt: high = VEC_PERM_EXPR */ high = make_temp_ssa_name (vectype, NULL, "vect_inter_high"); perm_stmt = gimple_build_assign_with_ops (VEC_PERM_EXPR, high, vect1, vect2, perm_mask_high); vect_finish_stmt_generation (stmt, perm_stmt, gsi); (*result_chain)[2*j] = high; /* Create interleaving stmt: low = VEC_PERM_EXPR */ low = make_temp_ssa_name (vectype, NULL, "vect_inter_low"); perm_stmt = gimple_build_assign_with_ops (VEC_PERM_EXPR, low, vect1, vect2, perm_mask_low); vect_finish_stmt_generation (stmt, perm_stmt, gsi); (*result_chain)[2*j+1] = low; } memcpy (dr_chain.address (), result_chain->address (), length * sizeof (tree)); } } /* Function vect_setup_realignment This function is called when vectorizing an unaligned load using the dr_explicit_realign[_optimized] scheme. This function generates the following code at the loop prolog: p = initial_addr; x msq_init = *(floor(p)); # prolog load realignment_token = call target_builtin; loop: x msq = phi (msq_init, ---) The stmts marked with x are generated only for the case of dr_explicit_realign_optimized. The code above sets up a new (vector) pointer, pointing to the first location accessed by STMT, and a "floor-aligned" load using that pointer. It also generates code to compute the "realignment-token" (if the relevant target hook was defined), and creates a phi-node at the loop-header bb whose arguments are the result of the prolog-load (created by this function) and the result of a load that takes place in the loop (to be created by the caller to this function). For the case of dr_explicit_realign_optimized: The caller to this function uses the phi-result (msq) to create the realignment code inside the loop, and sets up the missing phi argument, as follows: loop: msq = phi (msq_init, lsq) lsq = *(floor(p')); # load in loop result = realign_load (msq, lsq, realignment_token); For the case of dr_explicit_realign: loop: msq = *(floor(p)); # load in loop p' = p + (VS-1); lsq = *(floor(p')); # load in loop result = realign_load (msq, lsq, realignment_token); Input: STMT - (scalar) load stmt to be vectorized. This load accesses a memory location that may be unaligned. BSI - place where new code is to be inserted. ALIGNMENT_SUPPORT_SCHEME - which of the two misalignment handling schemes is used. Output: REALIGNMENT_TOKEN - the result of a call to the builtin_mask_for_load target hook, if defined. Return value - the result of the loop-header phi node. */ tree vect_setup_realignment (gimple stmt, gimple_stmt_iterator *gsi, tree *realignment_token, enum dr_alignment_support alignment_support_scheme, tree init_addr, struct loop **at_loop) { stmt_vec_info stmt_info = vinfo_for_stmt (stmt); tree vectype = STMT_VINFO_VECTYPE (stmt_info); loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); struct loop *loop = NULL; edge pe = NULL; tree scalar_dest = gimple_assign_lhs (stmt); tree vec_dest; gimple inc; tree ptr; tree data_ref; gimple new_stmt; basic_block new_bb; tree msq_init = NULL_TREE; tree new_temp; gimple phi_stmt; tree msq = NULL_TREE; gimple_seq stmts = NULL; bool inv_p; bool compute_in_loop = false; bool nested_in_vect_loop = false; struct loop *containing_loop = (gimple_bb (stmt))->loop_father; struct loop *loop_for_initial_load = NULL; if (loop_vinfo) { loop = LOOP_VINFO_LOOP (loop_vinfo); nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); } gcc_assert (alignment_support_scheme == dr_explicit_realign || alignment_support_scheme == dr_explicit_realign_optimized); /* We need to generate three things: 1. the misalignment computation 2. the extra vector load (for the optimized realignment scheme). 3. the phi node for the two vectors from which the realignment is done (for the optimized realignment scheme). */ /* 1. Determine where to generate the misalignment computation. If INIT_ADDR is NULL_TREE, this indicates that the misalignment calculation will be generated by this function, outside the loop (in the preheader). Otherwise, INIT_ADDR had already been computed for us by the caller, inside the loop. Background: If the misalignment remains fixed throughout the iterations of the loop, then both realignment schemes are applicable, and also the misalignment computation can be done outside LOOP. This is because we are vectorizing LOOP, and so the memory accesses in LOOP advance in steps that are a multiple of VS (the Vector Size), and therefore the misalignment in different vectorized LOOP iterations is always the same. The problem arises only if the memory access is in an inner-loop nested inside LOOP, which is now being vectorized using outer-loop vectorization. This is the only case when the misalignment of the memory access may not remain fixed throughout the iterations of the inner-loop (as explained in detail in vect_supportable_dr_alignment). In this case, not only is the optimized realignment scheme not applicable, but also the misalignment computation (and generation of the realignment token that is passed to REALIGN_LOAD) have to be done inside the loop. In short, INIT_ADDR indicates whether we are in a COMPUTE_IN_LOOP mode or not, which in turn determines if the misalignment is computed inside the inner-loop, or outside LOOP. */ if (init_addr != NULL_TREE || !loop_vinfo) { compute_in_loop = true; gcc_assert (alignment_support_scheme == dr_explicit_realign); } /* 2. Determine where to generate the extra vector load. For the optimized realignment scheme, instead of generating two vector loads in each iteration, we generate a single extra vector load in the preheader of the loop, and in each iteration reuse the result of the vector load from the previous iteration. In case the memory access is in an inner-loop nested inside LOOP, which is now being vectorized using outer-loop vectorization, we need to determine whether this initial vector load should be generated at the preheader of the inner-loop, or can be generated at the preheader of LOOP. If the memory access has no evolution in LOOP, it can be generated in the preheader of LOOP. Otherwise, it has to be generated inside LOOP (in the preheader of the inner-loop). */ if (nested_in_vect_loop) { tree outerloop_step = STMT_VINFO_DR_STEP (stmt_info); bool invariant_in_outerloop = (tree_int_cst_compare (outerloop_step, size_zero_node) == 0); loop_for_initial_load = (invariant_in_outerloop ? loop : loop->inner); } else loop_for_initial_load = loop; if (at_loop) *at_loop = loop_for_initial_load; if (loop_for_initial_load) pe = loop_preheader_edge (loop_for_initial_load); /* 3. For the case of the optimized realignment, create the first vector load at the loop preheader. */ if (alignment_support_scheme == dr_explicit_realign_optimized) { /* Create msq_init = *(floor(p1)) in the loop preheader */ gcc_assert (!compute_in_loop); vec_dest = vect_create_destination_var (scalar_dest, vectype); ptr = vect_create_data_ref_ptr (stmt, vectype, loop_for_initial_load, NULL_TREE, &init_addr, NULL, &inc, true, &inv_p); new_temp = copy_ssa_name (ptr, NULL); new_stmt = gimple_build_assign_with_ops (BIT_AND_EXPR, new_temp, ptr, build_int_cst (TREE_TYPE (ptr), -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); gcc_assert (!new_bb); data_ref = build2 (MEM_REF, TREE_TYPE (vec_dest), new_temp, build_int_cst (reference_alias_ptr_type (DR_REF (dr)), 0)); new_stmt = gimple_build_assign (vec_dest, data_ref); new_temp = make_ssa_name (vec_dest, new_stmt); gimple_assign_set_lhs (new_stmt, new_temp); if (pe) { new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); gcc_assert (!new_bb); } else gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT); msq_init = gimple_assign_lhs (new_stmt); } /* 4. Create realignment token using a target builtin, if available. It is done either inside the containing loop, or before LOOP (as determined above). */ if (targetm.vectorize.builtin_mask_for_load) { tree builtin_decl; /* Compute INIT_ADDR - the initial addressed accessed by this memref. */ if (!init_addr) { /* Generate the INIT_ADDR computation outside LOOP. */ init_addr = vect_create_addr_base_for_vector_ref (stmt, &stmts, NULL_TREE, loop); if (loop) { pe = loop_preheader_edge (loop); new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); gcc_assert (!new_bb); } else gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT); } builtin_decl = targetm.vectorize.builtin_mask_for_load (); new_stmt = gimple_build_call (builtin_decl, 1, init_addr); vec_dest = vect_create_destination_var (scalar_dest, gimple_call_return_type (new_stmt)); new_temp = make_ssa_name (vec_dest, new_stmt); gimple_call_set_lhs (new_stmt, new_temp); if (compute_in_loop) gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT); else { /* Generate the misalignment computation outside LOOP. */ pe = loop_preheader_edge (loop); new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); gcc_assert (!new_bb); } *realignment_token = gimple_call_lhs (new_stmt); /* The result of the CALL_EXPR to this builtin is determined from the value of the parameter and no global variables are touched which makes the builtin a "const" function. Requiring the builtin to have the "const" attribute makes it unnecessary to call mark_call_clobbered. */ gcc_assert (TREE_READONLY (builtin_decl)); } if (alignment_support_scheme == dr_explicit_realign) return msq; gcc_assert (!compute_in_loop); gcc_assert (alignment_support_scheme == dr_explicit_realign_optimized); /* 5. Create msq = phi in loop */ pe = loop_preheader_edge (containing_loop); vec_dest = vect_create_destination_var (scalar_dest, vectype); msq = make_ssa_name (vec_dest, NULL); phi_stmt = create_phi_node (msq, containing_loop->header); add_phi_arg (phi_stmt, msq_init, pe, UNKNOWN_LOCATION); return msq; } /* Function vect_grouped_load_supported. Returns TRUE if even and odd permutations are supported, and FALSE otherwise. */ bool vect_grouped_load_supported (tree vectype, unsigned HOST_WIDE_INT count) { enum machine_mode mode = TYPE_MODE (vectype); /* vect_permute_load_chain requires the group size to be a power of two. */ if (exact_log2 (count) == -1) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "the size of the group of accesses" " is not a power of 2\n"); return false; } /* Check that the permutation is supported. */ if (VECTOR_MODE_P (mode)) { unsigned int i, nelt = GET_MODE_NUNITS (mode); unsigned char *sel = XALLOCAVEC (unsigned char, nelt); for (i = 0; i < nelt; i++) sel[i] = i * 2; if (can_vec_perm_p (mode, false, sel)) { for (i = 0; i < nelt; i++) sel[i] = i * 2 + 1; if (can_vec_perm_p (mode, false, sel)) return true; } } if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, "extract even/odd not supported by target\n"); return false; } /* Return TRUE if vec_load_lanes is available for COUNT vectors of type VECTYPE. */ bool vect_load_lanes_supported (tree vectype, unsigned HOST_WIDE_INT count) { return vect_lanes_optab_supported_p ("vec_load_lanes", vec_load_lanes_optab, vectype, count); } /* Function vect_permute_load_chain. Given a chain of interleaved loads in DR_CHAIN of LENGTH that must be a power of 2, generate extract_even/odd stmts to reorder the input data correctly. Return the final references for loads in RESULT_CHAIN. E.g., LENGTH is 4 and the scalar type is short, i.e., VF is 8. The input is 4 vectors each containing 8 elements. We assign a number to each element, the input sequence is: 1st vec: 0 1 2 3 4 5 6 7 2nd vec: 8 9 10 11 12 13 14 15 3rd vec: 16 17 18 19 20 21 22 23 4th vec: 24 25 26 27 28 29 30 31 The output sequence should be: 1st vec: 0 4 8 12 16 20 24 28 2nd vec: 1 5 9 13 17 21 25 29 3rd vec: 2 6 10 14 18 22 26 30 4th vec: 3 7 11 15 19 23 27 31 i.e., the first output vector should contain the first elements of each interleaving group, etc. We use extract_even/odd instructions to create such output. The input of each extract_even/odd operation is two vectors 1st vec 2nd vec 0 1 2 3 4 5 6 7 and the output is the vector of extracted even/odd elements. The output of extract_even will be: 0 2 4 6 and of extract_odd: 1 3 5 7 The permutation is done in log LENGTH stages. In each stage extract_even and extract_odd stmts are created for each pair of vectors in DR_CHAIN in their order. In our example, E1: extract_even (1st vec, 2nd vec) E2: extract_odd (1st vec, 2nd vec) E3: extract_even (3rd vec, 4th vec) E4: extract_odd (3rd vec, 4th vec) The output for the first stage will be: E1: 0 2 4 6 8 10 12 14 E2: 1 3 5 7 9 11 13 15 E3: 16 18 20 22 24 26 28 30 E4: 17 19 21 23 25 27 29 31 In order to proceed and create the correct sequence for the next stage (or for the correct output, if the second stage is the last one, as in our example), we first put the output of extract_even operation and then the output of extract_odd in RESULT_CHAIN (which is then copied to DR_CHAIN). The input for the second stage is: 1st vec (E1): 0 2 4 6 8 10 12 14 2nd vec (E3): 16 18 20 22 24 26 28 30 3rd vec (E2): 1 3 5 7 9 11 13 15 4th vec (E4): 17 19 21 23 25 27 29 31 The output of the second stage: E1: 0 4 8 12 16 20 24 28 E2: 2 6 10 14 18 22 26 30 E3: 1 5 9 13 17 21 25 29 E4: 3 7 11 15 19 23 27 31 And RESULT_CHAIN after reordering: 1st vec (E1): 0 4 8 12 16 20 24 28 2nd vec (E3): 1 5 9 13 17 21 25 29 3rd vec (E2): 2 6 10 14 18 22 26 30 4th vec (E4): 3 7 11 15 19 23 27 31. */ static void vect_permute_load_chain (vec dr_chain, unsigned int length, gimple stmt, gimple_stmt_iterator *gsi, vec *result_chain) { tree data_ref, first_vect, second_vect; tree perm_mask_even, perm_mask_odd; gimple perm_stmt; tree vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt)); unsigned int i, j, log_length = exact_log2 (length); unsigned nelt = TYPE_VECTOR_SUBPARTS (vectype); unsigned char *sel = XALLOCAVEC (unsigned char, nelt); result_chain->quick_grow (length); memcpy (result_chain->address (), dr_chain.address (), length * sizeof (tree)); for (i = 0; i < nelt; ++i) sel[i] = i * 2; perm_mask_even = vect_gen_perm_mask (vectype, sel); gcc_assert (perm_mask_even != NULL); for (i = 0; i < nelt; ++i) sel[i] = i * 2 + 1; perm_mask_odd = vect_gen_perm_mask (vectype, sel); gcc_assert (perm_mask_odd != NULL); for (i = 0; i < log_length; i++) { for (j = 0; j < length; j += 2) { first_vect = dr_chain[j]; second_vect = dr_chain[j+1]; /* data_ref = permute_even (first_data_ref, second_data_ref); */ data_ref = make_temp_ssa_name (vectype, NULL, "vect_perm_even"); perm_stmt = gimple_build_assign_with_ops (VEC_PERM_EXPR, data_ref, first_vect, second_vect, perm_mask_even); vect_finish_stmt_generation (stmt, perm_stmt, gsi); (*result_chain)[j/2] = data_ref; /* data_ref = permute_odd (first_data_ref, second_data_ref); */ data_ref = make_temp_ssa_name (vectype, NULL, "vect_perm_odd"); perm_stmt = gimple_build_assign_with_ops (VEC_PERM_EXPR, data_ref, first_vect, second_vect, perm_mask_odd); vect_finish_stmt_generation (stmt, perm_stmt, gsi); (*result_chain)[j/2+length/2] = data_ref; } memcpy (dr_chain.address (), result_chain->address (), length * sizeof (tree)); } } /* Function vect_transform_grouped_load. Given a chain of input interleaved data-refs (in DR_CHAIN), build statements to perform their permutation and ascribe the result vectorized statements to the scalar statements. */ void vect_transform_grouped_load (gimple stmt, vec dr_chain, int size, gimple_stmt_iterator *gsi) { vec result_chain = vNULL; /* DR_CHAIN contains input data-refs that are a part of the interleaving. RESULT_CHAIN is the output of vect_permute_load_chain, it contains permuted vectors, that are ready for vector computation. */ result_chain.create (size); vect_permute_load_chain (dr_chain, size, stmt, gsi, &result_chain); vect_record_grouped_load_vectors (stmt, result_chain); result_chain.release (); } /* RESULT_CHAIN contains the output of a group of grouped loads that were generated as part of the vectorization of STMT. Assign the statement for each vector to the associated scalar statement. */ void vect_record_grouped_load_vectors (gimple stmt, vec result_chain) { gimple first_stmt = GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)); gimple next_stmt, new_stmt; unsigned int i, gap_count; tree tmp_data_ref; /* Put a permuted data-ref in the VECTORIZED_STMT field. Since we scan the chain starting from it's first node, their order corresponds the order of data-refs in RESULT_CHAIN. */ next_stmt = first_stmt; gap_count = 1; FOR_EACH_VEC_ELT (result_chain, i, tmp_data_ref) { if (!next_stmt) break; /* Skip the gaps. Loads created for the gaps will be removed by dead code elimination pass later. No need to check for the first stmt in the group, since it always exists. GROUP_GAP is the number of steps in elements from the previous access (if there is no gap GROUP_GAP is 1). We skip loads that correspond to the gaps. */ if (next_stmt != first_stmt && gap_count < GROUP_GAP (vinfo_for_stmt (next_stmt))) { gap_count++; continue; } while (next_stmt) { new_stmt = SSA_NAME_DEF_STMT (tmp_data_ref); /* We assume that if VEC_STMT is not NULL, this is a case of multiple copies, and we put the new vector statement in the first available RELATED_STMT. */ if (!STMT_VINFO_VEC_STMT (vinfo_for_stmt (next_stmt))) STMT_VINFO_VEC_STMT (vinfo_for_stmt (next_stmt)) = new_stmt; else { if (!GROUP_SAME_DR_STMT (vinfo_for_stmt (next_stmt))) { gimple prev_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt (next_stmt)); gimple rel_stmt = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (prev_stmt)); while (rel_stmt) { prev_stmt = rel_stmt; rel_stmt = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (rel_stmt)); } STMT_VINFO_RELATED_STMT (vinfo_for_stmt (prev_stmt)) = new_stmt; } } next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); gap_count = 1; /* If NEXT_STMT accesses the same DR as the previous statement, put the same TMP_DATA_REF as its vectorized statement; otherwise get the next data-ref from RESULT_CHAIN. */ if (!next_stmt || !GROUP_SAME_DR_STMT (vinfo_for_stmt (next_stmt))) break; } } } /* Function vect_force_dr_alignment_p. Returns whether the alignment of a DECL can be forced to be aligned on ALIGNMENT bit boundary. */ bool vect_can_force_dr_alignment_p (const_tree decl, unsigned int alignment) { if (TREE_CODE (decl) != VAR_DECL) return false; /* We cannot change alignment of common or external symbols as another translation unit may contain a definition with lower alignment. The rules of common symbol linking mean that the definition will override the common symbol. The same is true for constant pool entries which may be shared and are not properly merged by LTO. */ if (DECL_EXTERNAL (decl) || DECL_COMMON (decl) || DECL_IN_CONSTANT_POOL (decl)) return false; if (TREE_ASM_WRITTEN (decl)) return false; /* Do not override the alignment as specified by the ABI when the used attribute is set. */ if (DECL_PRESERVE_P (decl)) return false; /* Do not override explicit alignment set by the user when an explicit section name is also used. This is a common idiom used by many software projects. */ if (DECL_SECTION_NAME (decl) != NULL_TREE && !DECL_HAS_IMPLICIT_SECTION_NAME_P (decl)) return false; if (TREE_STATIC (decl)) return (alignment <= MAX_OFILE_ALIGNMENT); else return (alignment <= MAX_STACK_ALIGNMENT); } /* Return whether the data reference DR is supported with respect to its alignment. If CHECK_ALIGNED_ACCESSES is TRUE, check if the access is supported even it is aligned, i.e., check if it is possible to vectorize it with different alignment. */ enum dr_alignment_support vect_supportable_dr_alignment (struct data_reference *dr, bool check_aligned_accesses) { gimple stmt = DR_STMT (dr); stmt_vec_info stmt_info = vinfo_for_stmt (stmt); tree vectype = STMT_VINFO_VECTYPE (stmt_info); enum machine_mode mode = TYPE_MODE (vectype); loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); struct loop *vect_loop = NULL; bool nested_in_vect_loop = false; if (aligned_access_p (dr) && !check_aligned_accesses) return dr_aligned; /* For now assume all conditional loads/stores support unaligned access without any special code. */ if (is_gimple_call (stmt) && gimple_call_internal_p (stmt) && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD || gimple_call_internal_fn (stmt) == IFN_MASK_STORE)) return dr_unaligned_supported; if (loop_vinfo) { vect_loop = LOOP_VINFO_LOOP (loop_vinfo); nested_in_vect_loop = nested_in_vect_loop_p (vect_loop, stmt); } /* Possibly unaligned access. */ /* We can choose between using the implicit realignment scheme (generating a misaligned_move stmt) and the explicit realignment scheme (generating aligned loads with a REALIGN_LOAD). There are two variants to the explicit realignment scheme: optimized, and unoptimized. We can optimize the realignment only if the step between consecutive vector loads is equal to the vector size. Since the vector memory accesses advance in steps of VS (Vector Size) in the vectorized loop, it is guaranteed that the misalignment amount remains the same throughout the execution of the vectorized loop. Therefore, we can create the "realignment token" (the permutation mask that is passed to REALIGN_LOAD) at the loop preheader. However, in the case of outer-loop vectorization, when vectorizing a memory access in the inner-loop nested within the LOOP that is now being vectorized, while it is guaranteed that the misalignment of the vectorized memory access will remain the same in different outer-loop iterations, it is *not* guaranteed that is will remain the same throughout the execution of the inner-loop. This is because the inner-loop advances with the original scalar step (and not in steps of VS). If the inner-loop step happens to be a multiple of VS, then the misalignment remains fixed and we can use the optimized realignment scheme. For example: for (i=0; i; vs += va; v1 = v2; } } } */ if (DR_IS_READ (dr)) { bool is_packed = false; tree type = (TREE_TYPE (DR_REF (dr))); if (optab_handler (vec_realign_load_optab, mode) != CODE_FOR_nothing && (!targetm.vectorize.builtin_mask_for_load || targetm.vectorize.builtin_mask_for_load ())) { tree vectype = STMT_VINFO_VECTYPE (stmt_info); if ((nested_in_vect_loop && (TREE_INT_CST_LOW (DR_STEP (dr)) != GET_MODE_SIZE (TYPE_MODE (vectype)))) || !loop_vinfo) return dr_explicit_realign; else return dr_explicit_realign_optimized; } if (!known_alignment_for_access_p (dr)) is_packed = not_size_aligned (DR_REF (dr)); if ((TYPE_USER_ALIGN (type) && !is_packed) || targetm.vectorize. support_vector_misalignment (mode, type, DR_MISALIGNMENT (dr), is_packed)) /* Can't software pipeline the loads, but can at least do them. */ return dr_unaligned_supported; } else { bool is_packed = false; tree type = (TREE_TYPE (DR_REF (dr))); if (!known_alignment_for_access_p (dr)) is_packed = not_size_aligned (DR_REF (dr)); if ((TYPE_USER_ALIGN (type) && !is_packed) || targetm.vectorize. support_vector_misalignment (mode, type, DR_MISALIGNMENT (dr), is_packed)) return dr_unaligned_supported; } /* Unsupported. */ return dr_unaligned_unsupported; }