/* Transformation Utilities for Loop Vectorization. Copyright (C) 2003,2004,2005,2006 Free Software Foundation, Inc. Contributed by Dorit Naishlos 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 2, 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 COPYING. If not, write to the Free Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA. */ #include "config.h" #include "system.h" #include "coretypes.h" #include "tm.h" #include "ggc.h" #include "tree.h" #include "target.h" #include "rtl.h" #include "basic-block.h" #include "diagnostic.h" #include "tree-flow.h" #include "tree-dump.h" #include "timevar.h" #include "cfgloop.h" #include "expr.h" #include "optabs.h" #include "params.h" #include "recog.h" #include "tree-data-ref.h" #include "tree-chrec.h" #include "tree-scalar-evolution.h" #include "tree-vectorizer.h" #include "langhooks.h" #include "tree-pass.h" #include "toplev.h" #include "real.h" /* Utility functions for the code transformation. */ static bool vect_transform_stmt (tree, block_stmt_iterator *, bool *); static tree vect_create_destination_var (tree, tree); static tree vect_create_data_ref_ptr (tree, block_stmt_iterator *, tree, tree *, tree *, bool, tree); static tree vect_create_addr_base_for_vector_ref (tree, tree *, tree); static tree vect_setup_realignment (tree, block_stmt_iterator *, tree *); static tree vect_get_new_vect_var (tree, enum vect_var_kind, const char *); static tree vect_get_vec_def_for_operand (tree, tree, tree *); static tree vect_init_vector (tree, tree, tree); static void vect_finish_stmt_generation (tree stmt, tree vec_stmt, block_stmt_iterator *bsi); static bool vect_is_simple_cond (tree, loop_vec_info); static void update_vuses_to_preheader (tree, struct loop*); static void vect_create_epilog_for_reduction (tree, tree, enum tree_code, tree); static tree get_initial_def_for_reduction (tree, tree, tree *); /* Utility function dealing with loop peeling (not peeling itself). */ static void vect_generate_tmps_on_preheader (loop_vec_info, tree *, tree *, tree *); static tree vect_build_loop_niters (loop_vec_info); static void vect_update_ivs_after_vectorizer (loop_vec_info, tree, edge); static tree vect_gen_niters_for_prolog_loop (loop_vec_info, tree); static void vect_update_init_of_dr (struct data_reference *, tree niters); static void vect_update_inits_of_drs (loop_vec_info, tree); static int vect_min_worthwhile_factor (enum tree_code); /* 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. */ static 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 = "vect_p"; break; default: gcc_unreachable (); } if (name) new_vect_var = create_tmp_var (type, concat (prefix, name, NULL)); else new_vect_var = create_tmp_var (type, prefix); /* Mark vector typed variable as a gimple register variable. */ if (TREE_CODE (type) == VECTOR_TYPE) DECL_GIMPLE_REG_P (new_vect_var) = true; 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. Output: 1. Return an SSA_NAME whose value is the address of the memory location of the first vector of the data reference. 2. If new_stmt_list is not NULL_TREE after return then the caller must insert these statement(s) which define the returned SSA_NAME. FORNOW: We are only handling array accesses with step 1. */ static tree vect_create_addr_base_for_vector_ref (tree stmt, tree *new_stmt_list, tree offset) { stmt_vec_info stmt_info = vinfo_for_stmt (stmt); struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); tree data_ref_base = unshare_expr (DR_BASE_ADDRESS (dr)); tree base_name = build_fold_indirect_ref (data_ref_base); tree vec_stmt; tree addr_base, addr_expr; tree dest, new_stmt; tree base_offset = unshare_expr (DR_OFFSET (dr)); tree init = unshare_expr (DR_INIT (dr)); tree vect_ptr_type, addr_expr2; /* Create base_offset */ base_offset = size_binop (PLUS_EXPR, base_offset, init); dest = create_tmp_var (TREE_TYPE (base_offset), "base_off"); add_referenced_var (dest); base_offset = force_gimple_operand (base_offset, &new_stmt, false, dest); append_to_statement_list_force (new_stmt, new_stmt_list); if (offset) { tree tmp = create_tmp_var (TREE_TYPE (base_offset), "offset"); tree step; /* For interleaved access step we divide STEP by the size of the interleaving group. */ if (DR_GROUP_SIZE (stmt_info)) step = fold_build2 (TRUNC_DIV_EXPR, TREE_TYPE (offset), DR_STEP (dr), build_int_cst (TREE_TYPE (offset), DR_GROUP_SIZE (stmt_info))); else step = DR_STEP (dr); add_referenced_var (tmp); offset = fold_build2 (MULT_EXPR, TREE_TYPE (offset), offset, step); base_offset = fold_build2 (PLUS_EXPR, TREE_TYPE (base_offset), base_offset, offset); base_offset = force_gimple_operand (base_offset, &new_stmt, false, tmp); append_to_statement_list_force (new_stmt, new_stmt_list); } /* base + base_offset */ addr_base = fold_build2 (PLUS_EXPR, TREE_TYPE (data_ref_base), data_ref_base, base_offset); vect_ptr_type = build_pointer_type (STMT_VINFO_VECTYPE (stmt_info)); /* addr_expr = addr_base */ addr_expr = vect_get_new_vect_var (vect_ptr_type, vect_pointer_var, get_name (base_name)); add_referenced_var (addr_expr); vec_stmt = fold_convert (vect_ptr_type, addr_base); addr_expr2 = vect_get_new_vect_var (vect_ptr_type, vect_pointer_var, get_name (base_name)); add_referenced_var (addr_expr2); vec_stmt = force_gimple_operand (vec_stmt, &new_stmt, false, addr_expr2); append_to_statement_list_force (new_stmt, new_stmt_list); if (vect_print_dump_info (REPORT_DETAILS)) { fprintf (vect_dump, "created "); print_generic_expr (vect_dump, vec_stmt, TDF_SLIM); } return vec_stmt; } /* Function vect_create_data_ref_ptr. Create a new pointer to vector type (vp), 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 vector 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_MODIFY_STMT or GIMPLE_MODIFY_STMT . 2. BSI: block_stmt_iterator where new stmts can be added. 3. OFFSET (optional): an offset to be added to the initial address accessed by the data-ref in STMT. 4. ONLY_INIT: indicate if vp is to be updated in the loop, or remain pointing to the initial address. 5. TYPE: if not NULL indicates the required type of the data-ref 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 *vp; vp = (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. Return the pointer. */ static tree vect_create_data_ref_ptr (tree stmt, block_stmt_iterator *bsi ATTRIBUTE_UNUSED, tree offset, tree *initial_address, tree *ptr_incr, bool only_init, tree type) { tree 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 = LOOP_VINFO_LOOP (loop_vinfo); tree vectype = STMT_VINFO_VECTYPE (stmt_info); tree vect_ptr_type; tree vect_ptr; tree tag; tree new_temp; tree vec_stmt; tree new_stmt_list = NULL_TREE; edge pe = loop_preheader_edge (loop); basic_block new_bb; tree vect_ptr_init; struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); base_name = build_fold_indirect_ref (unshare_expr (DR_BASE_ADDRESS (dr))); if (vect_print_dump_info (REPORT_DETAILS)) { tree data_ref_base = base_name; fprintf (vect_dump, "create vector-pointer variable to type: "); print_generic_expr (vect_dump, vectype, TDF_SLIM); if (TREE_CODE (data_ref_base) == VAR_DECL) fprintf (vect_dump, " vectorizing a one dimensional array ref: "); else if (TREE_CODE (data_ref_base) == ARRAY_REF) fprintf (vect_dump, " vectorizing a multidimensional array ref: "); else if (TREE_CODE (data_ref_base) == COMPONENT_REF) fprintf (vect_dump, " vectorizing a record based array ref: "); else if (TREE_CODE (data_ref_base) == SSA_NAME) fprintf (vect_dump, " vectorizing a pointer ref: "); print_generic_expr (vect_dump, base_name, TDF_SLIM); } /** (1) Create the new vector-pointer variable: **/ if (type) vect_ptr_type = build_pointer_type (type); else vect_ptr_type = build_pointer_type (vectype); vect_ptr = vect_get_new_vect_var (vect_ptr_type, vect_pointer_var, get_name (base_name)); add_referenced_var (vect_ptr); /** (2) Add aliasing information to the new vector-pointer: (The points-to info (DR_PTR_INFO) may be defined later.) **/ tag = DR_MEMTAG (dr); gcc_assert (tag); /* If tag is a variable (and NOT_A_TAG) than a new symbol memory tag must be created with tag added to its may alias list. */ if (!MTAG_P (tag)) new_type_alias (vect_ptr, tag, DR_REF (dr)); else set_symbol_mem_tag (vect_ptr, tag); var_ann (vect_ptr)->subvars = DR_SUBVARS (dr); /** (3) Calculate the initial address the vector-pointer, and set the vector-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); pe = loop_preheader_edge (loop); new_bb = bsi_insert_on_edge_immediate (pe, new_stmt_list); gcc_assert (!new_bb); *initial_address = new_temp; /* Create: p = (vectype *) initial_base */ vec_stmt = fold_convert (vect_ptr_type, new_temp); vec_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, vect_ptr, vec_stmt); vect_ptr_init = make_ssa_name (vect_ptr, vec_stmt); GIMPLE_STMT_OPERAND (vec_stmt, 0) = vect_ptr_init; new_bb = bsi_insert_on_edge_immediate (pe, vec_stmt); gcc_assert (!new_bb); /** (4) Handle the updating of the vector-pointer inside the loop: **/ if (only_init) /* No update in loop is required. */ { /* Copy the points-to information if it exists. */ if (DR_PTR_INFO (dr)) duplicate_ssa_name_ptr_info (vect_ptr_init, DR_PTR_INFO (dr)); return vect_ptr_init; } else { block_stmt_iterator incr_bsi; bool insert_after; tree indx_before_incr, indx_after_incr; tree incr; standard_iv_increment_position (loop, &incr_bsi, &insert_after); create_iv (vect_ptr_init, fold_convert (vect_ptr_type, TYPE_SIZE_UNIT (vectype)), NULL_TREE, loop, &incr_bsi, insert_after, &indx_before_incr, &indx_after_incr); incr = bsi_stmt (incr_bsi); set_stmt_info (stmt_ann (incr), new_stmt_vec_info (incr, loop_vinfo)); /* 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)); } merge_alias_info (vect_ptr_init, indx_before_incr); merge_alias_info (vect_ptr_init, indx_after_incr); if (ptr_incr) *ptr_incr = incr; return indx_before_incr; } } /* Function bump_vector_ptr Increment a pointer (to a vector type) by vector-size. Connect the new increment stmt to the existing def-use update-chain of the pointer. 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 + vector_size .... 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 - the stmt that updates the pointer in each iteration of the loop. The increment amount across iterations is also 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. Output: Return NEW_DATAREF_PTR as illustrated above. */ static tree bump_vector_ptr (tree dataref_ptr, tree ptr_incr, block_stmt_iterator *bsi, tree stmt) { 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 vptr_type = TREE_TYPE (dataref_ptr); tree ptr_var = SSA_NAME_VAR (dataref_ptr); tree update = fold_convert (vptr_type, TYPE_SIZE_UNIT (vectype)); tree incr_stmt; ssa_op_iter iter; use_operand_p use_p; tree new_dataref_ptr; incr_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, ptr_var, build2 (PLUS_EXPR, vptr_type, dataref_ptr, update)); new_dataref_ptr = make_ssa_name (ptr_var, incr_stmt); GIMPLE_STMT_OPERAND (incr_stmt, 0) = new_dataref_ptr; vect_finish_stmt_generation (stmt, incr_stmt, bsi); /* 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); } /* 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)); merge_alias_info (new_dataref_ptr, dataref_ptr); return new_dataref_ptr; } /* Function vect_create_destination_var. Create a new temporary of type VECTYPE. */ static tree vect_create_destination_var (tree scalar_dest, tree vectype) { tree vec_dest; const 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); new_name = get_name (scalar_dest); if (!new_name) new_name = "var_"; vec_dest = vect_get_new_vect_var (type, vect_simple_var, new_name); add_referenced_var (vec_dest); return vec_dest; } /* Function vect_init_vector. Insert a new stmt (INIT_STMT) that initializes a new vector variable with the vector elements of VECTOR_VAR. Return the DEF of INIT_STMT. It will be used in the vectorization of STMT. */ static tree vect_init_vector (tree stmt, tree vector_var, tree vector_type) { stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); tree new_var; tree init_stmt; tree vec_oprnd; edge pe; tree new_temp; basic_block new_bb; new_var = vect_get_new_vect_var (vector_type, vect_simple_var, "cst_"); add_referenced_var (new_var); init_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, new_var, vector_var); new_temp = make_ssa_name (new_var, init_stmt); GIMPLE_STMT_OPERAND (init_stmt, 0) = new_temp; pe = loop_preheader_edge (loop); new_bb = bsi_insert_on_edge_immediate (pe, init_stmt); gcc_assert (!new_bb); if (vect_print_dump_info (REPORT_DETAILS)) { fprintf (vect_dump, "created new init_stmt: "); print_generic_expr (vect_dump, init_stmt, TDF_SLIM); } vec_oprnd = GIMPLE_STMT_OPERAND (init_stmt, 0); return vec_oprnd; } /* Function vect_get_vec_def_for_operand. OP is an operand in STMT. This function returns a (vector) def that will be used in the vectorized stmt for STMT. In the case that OP is an SSA_NAME which is defined in the loop, then STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def. In case OP is an invariant or constant, a new stmt that creates a vector def needs to be introduced. */ static tree vect_get_vec_def_for_operand (tree op, tree stmt, tree *scalar_def) { tree vec_oprnd; tree vec_stmt; tree def_stmt; stmt_vec_info def_stmt_info = NULL; stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); tree vectype = STMT_VINFO_VECTYPE (stmt_vinfo); int nunits = TYPE_VECTOR_SUBPARTS (vectype); loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); tree vec_inv; tree vec_cst; tree t = NULL_TREE; tree def; int i; enum vect_def_type dt; bool is_simple_use; tree vector_type; if (vect_print_dump_info (REPORT_DETAILS)) { fprintf (vect_dump, "vect_get_vec_def_for_operand: "); print_generic_expr (vect_dump, op, TDF_SLIM); } is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt); gcc_assert (is_simple_use); if (vect_print_dump_info (REPORT_DETAILS)) { if (def) { fprintf (vect_dump, "def = "); print_generic_expr (vect_dump, def, TDF_SLIM); } if (def_stmt) { fprintf (vect_dump, " def_stmt = "); print_generic_expr (vect_dump, def_stmt, TDF_SLIM); } } switch (dt) { /* Case 1: operand is a constant. */ case vect_constant_def: { if (scalar_def) *scalar_def = op; /* Create 'vect_cst_ = {cst,cst,...,cst}' */ if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "Create vector_cst. nunits = %d", nunits); for (i = nunits - 1; i >= 0; --i) { t = tree_cons (NULL_TREE, op, t); } vector_type = get_vectype_for_scalar_type (TREE_TYPE (op)); vec_cst = build_vector (vector_type, t); return vect_init_vector (stmt, vec_cst, vector_type); } /* Case 2: operand is defined outside the loop - loop invariant. */ case vect_invariant_def: { if (scalar_def) *scalar_def = def; /* Create 'vec_inv = {inv,inv,..,inv}' */ if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "Create vector_inv."); for (i = nunits - 1; i >= 0; --i) { t = tree_cons (NULL_TREE, def, t); } /* FIXME: use build_constructor directly. */ vector_type = get_vectype_for_scalar_type (TREE_TYPE (def)); vec_inv = build_constructor_from_list (vector_type, t); return vect_init_vector (stmt, vec_inv, vector_type); } /* Case 3: operand is defined inside the loop. */ case vect_loop_def: { if (scalar_def) *scalar_def = def_stmt; /* Get the def from the vectorized stmt. */ def_stmt_info = vinfo_for_stmt (def_stmt); vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info); gcc_assert (vec_stmt); vec_oprnd = GIMPLE_STMT_OPERAND (vec_stmt, 0); return vec_oprnd; } /* Case 4: operand is defined by a loop header phi - reduction */ case vect_reduction_def: { gcc_assert (TREE_CODE (def_stmt) == PHI_NODE); /* Get the def before the loop */ op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop)); return get_initial_def_for_reduction (stmt, op, scalar_def); } /* Case 5: operand is defined by loop-header phi - induction. */ case vect_induction_def: { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "induction - unsupported."); internal_error ("no support for induction"); /* FORNOW */ } default: gcc_unreachable (); } } /* Function vect_get_vec_def_for_stmt_copy Return a vector-def for an operand. This function is used when the vectorized stmt to be created (by the caller to this function) is a "copy" created in case the vectorized result cannot fit in one vector, and several copies of the vector-stmt are required. In this case the vector-def is retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. DT is the type of the vector def VEC_OPRND. Context: In case the vectorization factor (VF) is bigger than the number of elements that can fit in a vectype (nunits), we have to generate more than one vector stmt to vectorize the scalar stmt. This situation arises when there are multiple data-types operated upon in the loop; the smallest data-type determines the VF, and as a result, when vectorizing stmts operating on wider types we need to create 'VF/nunits' "copies" of the vector stmt (each computing a vector of 'nunits' results, and together computing 'VF' results in each iteration). This function is called when vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in which VF=16 and nunits=4, so the number of copies required is 4): scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT S1: x = load VS1.0: vx.0 = memref0 VS1.1 VS1.1: vx.1 = memref1 VS1.2 VS1.2: vx.2 = memref2 VS1.3 VS1.3: vx.3 = memref3 S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1 VSnew.1: vz1 = vx.1 + ... VSnew.2 VSnew.2: vz2 = vx.2 + ... VSnew.3 VSnew.3: vz3 = vx.3 + ... The vectorization of S1 is explained in vectorizable_load. The vectorization of S2: To create the first vector-stmt out of the 4 copies - VSnew.0 - the function 'vect_get_vec_def_for_operand' is called to get the relevant vector-def for each operand of S2. For operand x it returns the vector-def 'vx.0'. To create the remaining copies of the vector-stmt (VSnew.j), this function is called to get the relevant vector-def for each operand. It is obtained from the respective VS1.j stmt, which is recorded in the STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND. For example, to obtain the vector-def 'vx.1' in order to create the vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'. Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1', and return its def ('vx.1'). Overall, to create the above sequence this function will be called 3 times: vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0); vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1); vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */ static tree vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd) { tree vec_stmt_for_operand; stmt_vec_info def_stmt_info; if (dt == vect_invariant_def || dt == vect_constant_def) { /* Do nothing; can reuse same def. */ ; return vec_oprnd; } vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd); def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand); gcc_assert (def_stmt_info); vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info); gcc_assert (vec_stmt_for_operand); vec_oprnd = GIMPLE_STMT_OPERAND (vec_stmt_for_operand, 0); return vec_oprnd; } /* Function vect_finish_stmt_generation. Insert a new stmt. */ static void vect_finish_stmt_generation (tree stmt, tree vec_stmt, block_stmt_iterator *bsi) { stmt_vec_info stmt_info = vinfo_for_stmt (stmt); loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); bsi_insert_before (bsi, vec_stmt, BSI_SAME_STMT); set_stmt_info (get_stmt_ann (vec_stmt), new_stmt_vec_info (vec_stmt, loop_vinfo)); if (vect_print_dump_info (REPORT_DETAILS)) { fprintf (vect_dump, "add new stmt: "); print_generic_expr (vect_dump, vec_stmt, TDF_SLIM); } /* Make sure bsi points to the stmt that is being vectorized. */ gcc_assert (stmt == bsi_stmt (*bsi)); #ifdef USE_MAPPED_LOCATION SET_EXPR_LOCATION (vec_stmt, EXPR_LOCATION (stmt)); #else SET_EXPR_LOCUS (vec_stmt, EXPR_LOCUS (stmt)); #endif } #define ADJUST_IN_EPILOG 1 /* Function get_initial_def_for_reduction Input: STMT - a stmt that performs a reduction operation in the loop. INIT_VAL - the initial value of the reduction variable Output: SCALAR_DEF - a tree that holds a value to be added to the final result of the reduction (used for "ADJUST_IN_EPILOG" - see below). Return a vector variable, initialized according to the operation that STMT performs. This vector will be used as the initial value of the vector of partial results. Option1 ("ADJUST_IN_EPILOG"): Initialize the vector as follows: add: [0,0,...,0,0] mult: [1,1,...,1,1] min/max: [init_val,init_val,..,init_val,init_val] bit and/or: [init_val,init_val,..,init_val,init_val] and when necessary (e.g. add/mult case) let the caller know that it needs to adjust the result by init_val. Option2: Initialize the vector as follows: add: [0,0,...,0,init_val] mult: [1,1,...,1,init_val] min/max: [init_val,init_val,...,init_val] bit and/or: [init_val,init_val,...,init_val] and no adjustments are needed. For example, for the following code: s = init_val; for (i=0;i= 0; --i) t = tree_cons (NULL_TREE, def, t); if (nelements == nunits - 1) { /* Set the last element of the vector. */ t = tree_cons (NULL_TREE, init_val, t); nelements += 1; } gcc_assert (nelements == nunits); vector_type = get_vectype_for_scalar_type (TREE_TYPE (def)); if (TREE_CODE (init_val) == INTEGER_CST || TREE_CODE (init_val) == REAL_CST) vec = build_vector (vector_type, t); else vec = build_constructor_from_list (vector_type, t); if (!need_epilog_adjust) *scalar_def = NULL_TREE; else *scalar_def = init_val; return vect_init_vector (stmt, vec, vector_type); } /* Function vect_create_epilog_for_reduction Create code at the loop-epilog to finalize the result of a reduction computation. VECT_DEF is a vector of partial results. REDUC_CODE is the tree-code for the epilog reduction. STMT is the scalar reduction stmt that is being vectorized. REDUCTION_PHI is the phi-node that carries the reduction computation. This function: 1. Creates the reduction def-use cycle: sets the the arguments for REDUCTION_PHI: The loop-entry argument is the vectorized initial-value of the reduction. The loop-latch argument is VECT_DEF - the vector of partial sums. 2. "Reduces" the vector of partial results VECT_DEF into a single result, by applying the operation specified by REDUC_CODE if available, or by other means (whole-vector shifts or a scalar loop). The function also creates a new phi node at the loop exit to preserve loop-closed form, as illustrated below. The flow at the entry to this function: loop: vec_def = phi # REDUCTION_PHI VECT_DEF = vector_stmt # vectorized form of STMT s_loop = scalar_stmt # (scalar) STMT loop_exit: s_out0 = phi # (scalar) EXIT_PHI use use The above is transformed by this function into: loop: vec_def = phi # REDUCTION_PHI VECT_DEF = vector_stmt # vectorized form of STMT s_loop = scalar_stmt # (scalar) STMT loop_exit: s_out0 = phi # (scalar) EXIT_PHI v_out1 = phi # NEW_EXIT_PHI v_out2 = reduce s_out3 = extract_field s_out4 = adjust_result use use */ static void vect_create_epilog_for_reduction (tree vect_def, tree stmt, enum tree_code reduc_code, tree reduction_phi) { stmt_vec_info stmt_info = vinfo_for_stmt (stmt); tree vectype; enum machine_mode mode; loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); basic_block exit_bb; tree scalar_dest; tree scalar_type; tree new_phi; block_stmt_iterator exit_bsi; tree vec_dest; tree new_temp; tree new_name; tree epilog_stmt; tree new_scalar_dest, exit_phi; tree bitsize, bitpos, bytesize; enum tree_code code = TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 1)); tree scalar_initial_def; tree vec_initial_def; tree orig_name; imm_use_iterator imm_iter; use_operand_p use_p; bool extract_scalar_result; tree reduction_op; tree orig_stmt; tree use_stmt; tree operation = GIMPLE_STMT_OPERAND (stmt, 1); int op_type; op_type = TREE_CODE_LENGTH (TREE_CODE (operation)); reduction_op = TREE_OPERAND (operation, op_type-1); vectype = get_vectype_for_scalar_type (TREE_TYPE (reduction_op)); mode = TYPE_MODE (vectype); /*** 1. Create the reduction def-use cycle ***/ /* 1.1 set the loop-entry arg of the reduction-phi: */ /* For the case of reduction, vect_get_vec_def_for_operand returns the scalar def before the loop, that defines the initial value of the reduction variable. */ vec_initial_def = vect_get_vec_def_for_operand (reduction_op, stmt, &scalar_initial_def); add_phi_arg (reduction_phi, vec_initial_def, loop_preheader_edge (loop)); /* 1.2 set the loop-latch arg for the reduction-phi: */ add_phi_arg (reduction_phi, vect_def, loop_latch_edge (loop)); if (vect_print_dump_info (REPORT_DETAILS)) { fprintf (vect_dump, "transform reduction: created def-use cycle:"); print_generic_expr (vect_dump, reduction_phi, TDF_SLIM); fprintf (vect_dump, "\n"); print_generic_expr (vect_dump, SSA_NAME_DEF_STMT (vect_def), TDF_SLIM); } /*** 2. Create epilog code The reduction epilog code operates across the elements of the vector of partial results computed by the vectorized loop. The reduction epilog code consists of: step 1: compute the scalar result in a vector (v_out2) step 2: extract the scalar result (s_out3) from the vector (v_out2) step 3: adjust the scalar result (s_out3) if needed. Step 1 can be accomplished using one the following three schemes: (scheme 1) using reduc_code, if available. (scheme 2) using whole-vector shifts, if available. (scheme 3) using a scalar loop. In this case steps 1+2 above are combined. The overall epilog code looks like this: s_out0 = phi # original EXIT_PHI v_out1 = phi # NEW_EXIT_PHI v_out2 = reduce # step 1 s_out3 = extract_field # step 2 s_out4 = adjust_result # step 3 (step 3 is optional, and step2 1 and 2 may be combined). Lastly, the uses of s_out0 are replaced by s_out4. ***/ /* 2.1 Create new loop-exit-phi to preserve loop-closed form: v_out1 = phi */ exit_bb = single_exit (loop)->dest; new_phi = create_phi_node (SSA_NAME_VAR (vect_def), exit_bb); SET_PHI_ARG_DEF (new_phi, single_exit (loop)->dest_idx, vect_def); exit_bsi = bsi_start (exit_bb); /* 2.2 Get the relevant tree-code to use in the epilog for schemes 2,3 (i.e. when reduc_code is not available) and in the final adjustment code (if needed). Also get the original scalar reduction variable as defined in the loop. In case STMT is a "pattern-stmt" (i.e. - it represents a reduction pattern), the tree-code and scalar-def are taken from the original stmt that the pattern-stmt (STMT) replaces. Otherwise (it is a regular reduction) - the tree-code and scalar-def are taken from STMT. */ orig_stmt = STMT_VINFO_RELATED_STMT (stmt_info); if (!orig_stmt) { /* Regular reduction */ orig_stmt = stmt; } else { /* Reduction pattern */ stmt_vec_info stmt_vinfo = vinfo_for_stmt (orig_stmt); gcc_assert (STMT_VINFO_IN_PATTERN_P (stmt_vinfo)); gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo) == stmt); } code = TREE_CODE (GIMPLE_STMT_OPERAND (orig_stmt, 1)); scalar_dest = GIMPLE_STMT_OPERAND (orig_stmt, 0); scalar_type = TREE_TYPE (scalar_dest); new_scalar_dest = vect_create_destination_var (scalar_dest, NULL); bitsize = TYPE_SIZE (scalar_type); bytesize = TYPE_SIZE_UNIT (scalar_type); /* 2.3 Create the reduction code, using one of the three schemes described above. */ if (reduc_code < NUM_TREE_CODES) { /*** Case 1: Create: v_out2 = reduc_expr */ if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "Reduce using direct vector reduction."); vec_dest = vect_create_destination_var (scalar_dest, vectype); epilog_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, vec_dest, build1 (reduc_code, vectype, PHI_RESULT (new_phi))); new_temp = make_ssa_name (vec_dest, epilog_stmt); GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_temp; bsi_insert_after (&exit_bsi, epilog_stmt, BSI_NEW_STMT); extract_scalar_result = true; } else { enum tree_code shift_code = 0; bool have_whole_vector_shift = true; int bit_offset; int element_bitsize = tree_low_cst (bitsize, 1); int vec_size_in_bits = tree_low_cst (TYPE_SIZE (vectype), 1); tree vec_temp; if (vec_shr_optab->handlers[mode].insn_code != CODE_FOR_nothing) shift_code = VEC_RSHIFT_EXPR; else have_whole_vector_shift = false; /* Regardless of whether we have a whole vector shift, if we're emulating the operation via tree-vect-generic, we don't want to use it. Only the first round of the reduction is likely to still be profitable via emulation. */ /* ??? It might be better to emit a reduction tree code here, so that tree-vect-generic can expand the first round via bit tricks. */ if (!VECTOR_MODE_P (mode)) have_whole_vector_shift = false; else { optab optab = optab_for_tree_code (code, vectype); if (optab->handlers[mode].insn_code == CODE_FOR_nothing) have_whole_vector_shift = false; } if (have_whole_vector_shift) { /*** Case 2: Create: for (offset = VS/2; offset >= element_size; offset/=2) { Create: va' = vec_shift Create: va = vop } */ if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "Reduce using vector shifts"); vec_dest = vect_create_destination_var (scalar_dest, vectype); new_temp = PHI_RESULT (new_phi); for (bit_offset = vec_size_in_bits/2; bit_offset >= element_bitsize; bit_offset /= 2) { tree bitpos = size_int (bit_offset); epilog_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, vec_dest, build2 (shift_code, vectype, new_temp, bitpos)); new_name = make_ssa_name (vec_dest, epilog_stmt); GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_name; bsi_insert_after (&exit_bsi, epilog_stmt, BSI_NEW_STMT); epilog_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, vec_dest, build2 (code, vectype, new_name, new_temp)); new_temp = make_ssa_name (vec_dest, epilog_stmt); GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_temp; bsi_insert_after (&exit_bsi, epilog_stmt, BSI_NEW_STMT); } extract_scalar_result = true; } else { tree rhs; /*** Case 3: Create: s = extract_field for (offset = element_size; offset < vector_size; offset += element_size;) { Create: s' = extract_field Create: s = op } */ if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "Reduce using scalar code. "); vec_temp = PHI_RESULT (new_phi); vec_size_in_bits = tree_low_cst (TYPE_SIZE (vectype), 1); rhs = build3 (BIT_FIELD_REF, scalar_type, vec_temp, bitsize, bitsize_zero_node); BIT_FIELD_REF_UNSIGNED (rhs) = TYPE_UNSIGNED (scalar_type); epilog_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, new_scalar_dest, rhs); new_temp = make_ssa_name (new_scalar_dest, epilog_stmt); GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_temp; bsi_insert_after (&exit_bsi, epilog_stmt, BSI_NEW_STMT); for (bit_offset = element_bitsize; bit_offset < vec_size_in_bits; bit_offset += element_bitsize) { tree bitpos = bitsize_int (bit_offset); tree rhs = build3 (BIT_FIELD_REF, scalar_type, vec_temp, bitsize, bitpos); BIT_FIELD_REF_UNSIGNED (rhs) = TYPE_UNSIGNED (scalar_type); epilog_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, new_scalar_dest, rhs); new_name = make_ssa_name (new_scalar_dest, epilog_stmt); GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_name; bsi_insert_after (&exit_bsi, epilog_stmt, BSI_NEW_STMT); epilog_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, new_scalar_dest, build2 (code, scalar_type, new_name, new_temp)); new_temp = make_ssa_name (new_scalar_dest, epilog_stmt); GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_temp; bsi_insert_after (&exit_bsi, epilog_stmt, BSI_NEW_STMT); } extract_scalar_result = false; } } /* 2.4 Extract the final scalar result. Create: s_out3 = extract_field */ if (extract_scalar_result) { tree rhs; if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "extract scalar result"); if (BYTES_BIG_ENDIAN) bitpos = size_binop (MULT_EXPR, bitsize_int (TYPE_VECTOR_SUBPARTS (vectype) - 1), TYPE_SIZE (scalar_type)); else bitpos = bitsize_zero_node; rhs = build3 (BIT_FIELD_REF, scalar_type, new_temp, bitsize, bitpos); BIT_FIELD_REF_UNSIGNED (rhs) = TYPE_UNSIGNED (scalar_type); epilog_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, new_scalar_dest, rhs); new_temp = make_ssa_name (new_scalar_dest, epilog_stmt); GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_temp; bsi_insert_after (&exit_bsi, epilog_stmt, BSI_NEW_STMT); } /* 2.4 Adjust the final result by the initial value of the reduction variable. (When such adjustment is not needed, then 'scalar_initial_def' is zero). Create: s_out4 = scalar_expr */ if (scalar_initial_def) { epilog_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, new_scalar_dest, build2 (code, scalar_type, new_temp, scalar_initial_def)); new_temp = make_ssa_name (new_scalar_dest, epilog_stmt); GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_temp; bsi_insert_after (&exit_bsi, epilog_stmt, BSI_NEW_STMT); } /* 2.6 Replace uses of s_out0 with uses of s_out3 */ /* Find the loop-closed-use at the loop exit of the original scalar result. (The reduction result is expected to have two immediate uses - one at the latch block, and one at the loop exit). */ exit_phi = NULL; FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) { if (!flow_bb_inside_loop_p (loop, bb_for_stmt (USE_STMT (use_p)))) { exit_phi = USE_STMT (use_p); break; } } /* We expect to have found an exit_phi because of loop-closed-ssa form. */ gcc_assert (exit_phi); /* Replace the uses: */ orig_name = PHI_RESULT (exit_phi); FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, orig_name) FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) SET_USE (use_p, new_temp); } /* Function vectorizable_reduction. Check if STMT performs a reduction operation that can be vectorized. If VEC_STMT is also passed, vectorize the STMT: create a vectorized stmt to replace it, put it in VEC_STMT, and insert it at BSI. Return FALSE if not a vectorizable STMT, TRUE otherwise. This function also handles reduction idioms (patterns) that have been recognized in advance during vect_pattern_recog. In this case, STMT may be of this form: X = pattern_expr (arg0, arg1, ..., X) and it's STMT_VINFO_RELATED_STMT points to the last stmt in the original sequence that had been detected and replaced by the pattern-stmt (STMT). In some cases of reduction patterns, the type of the reduction variable X is different than the type of the other arguments of STMT. In such cases, the vectype that is used when transforming STMT into a vector stmt is different than the vectype that is used to determine the vectorization factor, because it consists of a different number of elements than the actual number of elements that are being operated upon in parallel. For example, consider an accumulation of shorts into an int accumulator. On some targets it's possible to vectorize this pattern operating on 8 shorts at a time (hence, the vectype for purposes of determining the vectorization factor should be V8HI); on the other hand, the vectype that is used to create the vector form is actually V4SI (the type of the result). Upon entry to this function, STMT_VINFO_VECTYPE records the vectype that indicates what is the actual level of parallelism (V8HI in the example), so that the right vectorization factor would be derived. This vectype corresponds to the type of arguments to the reduction stmt, and should *NOT* be used to create the vectorized stmt. The right vectype for the vectorized stmt is obtained from the type of the result X: get_vectype_for_scalar_type (TREE_TYPE (X)) This means that, contrary to "regular" reductions (or "regular" stmts in general), the following equation: STMT_VINFO_VECTYPE == get_vectype_for_scalar_type (TREE_TYPE (X)) does *NOT* necessarily hold for reduction patterns. */ bool vectorizable_reduction (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt) { tree vec_dest; tree scalar_dest; tree op; tree loop_vec_def0 = NULL_TREE, loop_vec_def1 = NULL_TREE; 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 loop *loop = LOOP_VINFO_LOOP (loop_vinfo); tree operation; enum tree_code code, orig_code, epilog_reduc_code = 0; enum machine_mode vec_mode; int op_type; optab optab, reduc_optab; tree new_temp = NULL_TREE; tree def, def_stmt; enum vect_def_type dt; tree new_phi; tree scalar_type; bool is_simple_use; tree orig_stmt; stmt_vec_info orig_stmt_info; tree expr = NULL_TREE; int i; int nunits = TYPE_VECTOR_SUBPARTS (vectype); int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; stmt_vec_info prev_stmt_info; tree reduc_def; tree new_stmt = NULL_TREE; int j; gcc_assert (ncopies >= 1); /* 1. Is vectorizable reduction? */ /* Not supportable if the reduction variable is used in the loop. */ if (STMT_VINFO_RELEVANT_P (stmt_info)) return false; if (!STMT_VINFO_LIVE_P (stmt_info)) return false; /* Make sure it was already recognized as a reduction computation. */ if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_reduction_def) return false; /* 2. Has this been recognized as a reduction pattern? Check if STMT represents a pattern that has been recognized in earlier analysis stages. For stmts that represent a pattern, the STMT_VINFO_RELATED_STMT field records the last stmt in the original sequence that constitutes the pattern. */ orig_stmt = STMT_VINFO_RELATED_STMT (stmt_info); if (orig_stmt) { orig_stmt_info = vinfo_for_stmt (orig_stmt); gcc_assert (STMT_VINFO_RELATED_STMT (orig_stmt_info) == stmt); gcc_assert (STMT_VINFO_IN_PATTERN_P (orig_stmt_info)); gcc_assert (!STMT_VINFO_IN_PATTERN_P (stmt_info)); } /* 3. Check the operands of the operation. The first operands are defined inside the loop body. The last operand is the reduction variable, which is defined by the loop-header-phi. */ gcc_assert (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT); operation = GIMPLE_STMT_OPERAND (stmt, 1); code = TREE_CODE (operation); op_type = TREE_CODE_LENGTH (code); if (op_type != binary_op && op_type != ternary_op) return false; scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0); scalar_type = TREE_TYPE (scalar_dest); /* All uses but the last are expected to be defined in the loop. The last use is the reduction variable. */ for (i = 0; i < op_type-1; i++) { op = TREE_OPERAND (operation, i); is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt); gcc_assert (is_simple_use); gcc_assert (dt == vect_loop_def || dt == vect_invariant_def || dt == vect_constant_def); } op = TREE_OPERAND (operation, i); is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt); gcc_assert (is_simple_use); gcc_assert (dt == vect_reduction_def); gcc_assert (TREE_CODE (def_stmt) == PHI_NODE); if (orig_stmt) gcc_assert (orig_stmt == vect_is_simple_reduction (loop, def_stmt)); else gcc_assert (stmt == vect_is_simple_reduction (loop, def_stmt)); if (STMT_VINFO_LIVE_P (vinfo_for_stmt (def_stmt))) return false; /* 4. Supportable by target? */ /* 4.1. check support for the operation in the loop */ optab = optab_for_tree_code (code, vectype); if (!optab) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "no optab."); return false; } vec_mode = TYPE_MODE (vectype); if (optab->handlers[(int) vec_mode].insn_code == CODE_FOR_nothing) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "op not supported by target."); if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD || LOOP_VINFO_VECT_FACTOR (loop_vinfo) < vect_min_worthwhile_factor (code)) return false; if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "proceeding using word mode."); } /* Worthwhile without SIMD support? */ if (!VECTOR_MODE_P (TYPE_MODE (vectype)) && LOOP_VINFO_VECT_FACTOR (loop_vinfo) < vect_min_worthwhile_factor (code)) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "not worthwhile without SIMD support."); return false; } /* 4.2. Check support for the epilog operation. If STMT represents a reduction pattern, then the type of the reduction variable may be different than the type of the rest of the arguments. For example, consider the case of accumulation of shorts into an int accumulator; The original code: S1: int_a = (int) short_a; orig_stmt-> S2: int_acc = plus ; was replaced with: STMT: int_acc = widen_sum This means that: 1. The tree-code that is used to create the vector operation in the epilog code (that reduces the partial results) is not the tree-code of STMT, but is rather the tree-code of the original stmt from the pattern that STMT is replacing. I.e, in the example above we want to use 'widen_sum' in the loop, but 'plus' in the epilog. 2. The type (mode) we use to check available target support for the vector operation to be created in the *epilog*, is determined by the type of the reduction variable (in the example above we'd check this: plus_optab[vect_int_mode]). However the type (mode) we use to check available target support for the vector operation to be created *inside the loop*, is determined by the type of the other arguments to STMT (in the example we'd check this: widen_sum_optab[vect_short_mode]). This is contrary to "regular" reductions, in which the types of all the arguments are the same as the type of the reduction variable. For "regular" reductions we can therefore use the same vector type (and also the same tree-code) when generating the epilog code and when generating the code inside the loop. */ if (orig_stmt) { /* This is a reduction pattern: get the vectype from the type of the reduction variable, and get the tree-code from orig_stmt. */ orig_code = TREE_CODE (GIMPLE_STMT_OPERAND (orig_stmt, 1)); vectype = get_vectype_for_scalar_type (TREE_TYPE (def)); vec_mode = TYPE_MODE (vectype); } else { /* Regular reduction: use the same vectype and tree-code as used for the vector code inside the loop can be used for the epilog code. */ orig_code = code; } if (!reduction_code_for_scalar_code (orig_code, &epilog_reduc_code)) return false; reduc_optab = optab_for_tree_code (epilog_reduc_code, vectype); if (!reduc_optab) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "no optab for reduction."); epilog_reduc_code = NUM_TREE_CODES; } if (reduc_optab->handlers[(int) vec_mode].insn_code == CODE_FOR_nothing) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "reduc op not supported by target."); epilog_reduc_code = NUM_TREE_CODES; } if (!vec_stmt) /* transformation not required. */ { STMT_VINFO_TYPE (stmt_info) = reduc_vec_info_type; return true; } /** Transform. **/ if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "transform reduction."); /* Create the destination vector */ vec_dest = vect_create_destination_var (scalar_dest, vectype); /* Create the reduction-phi that defines the reduction-operand. */ new_phi = create_phi_node (vec_dest, loop->header); /* In case the vectorization factor (VF) is bigger than the number of elements that we can fit in a vectype (nunits), we have to generate more than one vector stmt - i.e - we need to "unroll" the vector stmt by a factor VF/nunits. For more details see documentation in vectorizable_operation. */ prev_stmt_info = NULL; for (j = 0; j < ncopies; j++) { /* Handle uses. */ if (j == 0) { op = TREE_OPERAND (operation, 0); loop_vec_def0 = vect_get_vec_def_for_operand (op, stmt, NULL); if (op_type == ternary_op) { op = TREE_OPERAND (operation, 1); loop_vec_def1 = vect_get_vec_def_for_operand (op, stmt, NULL); } /* Get the vector def for the reduction variable from the phi node */ reduc_def = PHI_RESULT (new_phi); } else { enum vect_def_type dt = vect_unknown_def_type; /* Dummy */ loop_vec_def0 = vect_get_vec_def_for_stmt_copy (dt, loop_vec_def0); if (op_type == ternary_op) loop_vec_def1 = vect_get_vec_def_for_stmt_copy (dt, loop_vec_def1); /* Get the vector def for the reduction variable from the vectorized reduction operation generated in the previous iteration (j-1) */ reduc_def = GIMPLE_STMT_OPERAND (new_stmt ,0); } /* Arguments are ready. create the new vector stmt. */ if (op_type == binary_op) expr = build2 (code, vectype, loop_vec_def0, reduc_def); else expr = build3 (code, vectype, loop_vec_def0, loop_vec_def1, reduc_def); new_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, vec_dest, expr); new_temp = make_ssa_name (vec_dest, new_stmt); GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp; vect_finish_stmt_generation (stmt, new_stmt, bsi); if (j == 0) STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; else STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; prev_stmt_info = vinfo_for_stmt (new_stmt); } /* Finalize the reduction-phi (set it's arguments) and create the epilog reduction code. */ vect_create_epilog_for_reduction (new_temp, stmt, epilog_reduc_code, new_phi); return true; } /* Checks if CALL can be vectorized in type VECTYPE. Returns true if the target has a vectorized version of the function, or false if the function cannot be vectorized. */ bool vectorizable_function (tree call, tree vectype) { tree fndecl = get_callee_fndecl (call); /* We only handle functions that do not read or clobber memory -- i.e. const or novops ones. */ if (!(call_expr_flags (call) & (ECF_CONST | ECF_NOVOPS))) return false; if (!fndecl || TREE_CODE (fndecl) != FUNCTION_DECL || !DECL_BUILT_IN (fndecl)) return false; if (targetm.vectorize.builtin_vectorized_function (DECL_FUNCTION_CODE (fndecl), vectype)) return true; return false; } /* Returns an expression that performs a call to vectorized version of FNDECL in type VECTYPE, with the arguments given by ARGS. If extra statements need to be generated, they are inserted before BSI. */ static tree build_vectorized_function_call (tree fndecl, tree vectype, tree args) { tree vfndecl; enum built_in_function code = DECL_FUNCTION_CODE (fndecl); /* The target specific builtin should be available. */ vfndecl = targetm.vectorize.builtin_vectorized_function (code, vectype); gcc_assert (vfndecl != NULL_TREE); return build_function_call_expr (vfndecl, args); } /* Function vectorizable_call. Check if STMT performs a function call that can be vectorized. If VEC_STMT is also passed, vectorize the STMT: create a vectorized stmt to replace it, put it in VEC_STMT, and insert it at BSI. Return FALSE if not a vectorizable STMT, TRUE otherwise. */ bool vectorizable_call (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt) { tree vec_dest; tree scalar_dest; tree operation; tree op, args, type; tree vec_oprnd, vargs, *pvargs_end; 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); tree fndecl, rhs, new_temp, def, def_stmt; enum vect_def_type dt; /* Is STMT a vectorizable call? */ if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT) return false; if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) != SSA_NAME) return false; operation = GIMPLE_STMT_OPERAND (stmt, 1); if (TREE_CODE (operation) != CALL_EXPR) return false; /* For now, we only vectorize functions if a target specific builtin is available. TODO -- in some cases, it might be profitable to insert the calls for pieces of the vector, in order to be able to vectorize other operations in the loop. */ if (!vectorizable_function (operation, vectype)) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "function is not vectorizable."); return false; } gcc_assert (ZERO_SSA_OPERANDS (stmt, SSA_OP_ALL_VIRTUALS)); for (args = TREE_OPERAND (operation, 1); args; args = TREE_CHAIN (args)) { op = TREE_VALUE (args); if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt)) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "use not simple."); return false; } } if (!vec_stmt) /* transformation not required. */ { STMT_VINFO_TYPE (stmt_info) = call_vec_info_type; return true; } /** Transform. **/ if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "transform operation."); /* Handle def. */ scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0); vec_dest = vect_create_destination_var (scalar_dest, vectype); /* Handle uses. */ vargs = NULL_TREE; pvargs_end = &vargs; for (args = TREE_OPERAND (operation, 1); args; args = TREE_CHAIN (args)) { op = TREE_VALUE (args); vec_oprnd = vect_get_vec_def_for_operand (op, stmt, NULL); *pvargs_end = tree_cons (NULL_TREE, vec_oprnd, NULL_TREE); pvargs_end = &TREE_CHAIN (*pvargs_end); } fndecl = get_callee_fndecl (operation); rhs = build_vectorized_function_call (fndecl, vectype, vargs); *vec_stmt = build2 (GIMPLE_MODIFY_STMT, vectype, vec_dest, rhs); new_temp = make_ssa_name (vec_dest, *vec_stmt); GIMPLE_STMT_OPERAND (*vec_stmt, 0) = new_temp; vect_finish_stmt_generation (stmt, *vec_stmt, bsi); /* The call in STMT might prevent it from being removed in dce. We however cannot remove it here, due to the way the ssa name it defines is mapped to the new definition. So just replace rhs of the statement with something harmless. */ type = TREE_TYPE (scalar_dest); GIMPLE_STMT_OPERAND (stmt, 1) = fold_convert (type, integer_zero_node); return true; } /* Function vectorizable_assignment. Check if STMT performs an assignment (copy) that can be vectorized. If VEC_STMT is also passed, vectorize the STMT: create a vectorized stmt to replace it, put it in VEC_STMT, and insert it at BSI. Return FALSE if not a vectorizable STMT, TRUE otherwise. */ bool vectorizable_assignment (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt) { tree vec_dest; tree scalar_dest; tree op; tree vec_oprnd; 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); tree new_temp; tree def, def_stmt; enum vect_def_type dt; int nunits = TYPE_VECTOR_SUBPARTS (vectype); int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; gcc_assert (ncopies >= 1); if (ncopies > 1) return false; /* FORNOW */ /* Is vectorizable assignment? */ if (!STMT_VINFO_RELEVANT_P (stmt_info)) return false; gcc_assert (STMT_VINFO_DEF_TYPE (stmt_info) == vect_loop_def); if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT) return false; scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0); if (TREE_CODE (scalar_dest) != SSA_NAME) return false; op = GIMPLE_STMT_OPERAND (stmt, 1); if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt)) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "use not simple."); return false; } if (!vec_stmt) /* transformation not required. */ { STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type; return true; } /** Transform. **/ if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "transform assignment."); /* Handle def. */ vec_dest = vect_create_destination_var (scalar_dest, vectype); /* Handle use. */ op = GIMPLE_STMT_OPERAND (stmt, 1); vec_oprnd = vect_get_vec_def_for_operand (op, stmt, NULL); /* Arguments are ready. create the new vector stmt. */ *vec_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, vec_dest, vec_oprnd); new_temp = make_ssa_name (vec_dest, *vec_stmt); GIMPLE_STMT_OPERAND (*vec_stmt, 0) = new_temp; vect_finish_stmt_generation (stmt, *vec_stmt, bsi); return true; } /* Function vect_min_worthwhile_factor. For a loop where we could vectorize the operation indicated by CODE, return the minimum vectorization factor that makes it worthwhile to use generic vectors. */ static int vect_min_worthwhile_factor (enum tree_code code) { switch (code) { case PLUS_EXPR: case MINUS_EXPR: case NEGATE_EXPR: return 4; case BIT_AND_EXPR: case BIT_IOR_EXPR: case BIT_XOR_EXPR: case BIT_NOT_EXPR: return 2; default: return INT_MAX; } } /* Function vectorizable_operation. Check if STMT performs a binary or unary operation that can be vectorized. If VEC_STMT is also passed, vectorize the STMT: create a vectorized stmt to replace it, put it in VEC_STMT, and insert it at BSI. Return FALSE if not a vectorizable STMT, TRUE otherwise. */ bool vectorizable_operation (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt) { tree vec_dest; tree scalar_dest; tree operation; tree op0, op1 = NULL; tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE; 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); enum tree_code code; enum machine_mode vec_mode; tree new_temp; int op_type; optab optab; int icode; enum machine_mode optab_op2_mode; tree def, def_stmt; enum vect_def_type dt0, dt1; tree new_stmt; stmt_vec_info prev_stmt_info; int nunits_in = TYPE_VECTOR_SUBPARTS (vectype); int nunits_out; tree vectype_out; int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; int j; gcc_assert (ncopies >= 1); /* Is STMT a vectorizable binary/unary operation? */ if (!STMT_VINFO_RELEVANT_P (stmt_info)) return false; gcc_assert (STMT_VINFO_DEF_TYPE (stmt_info) == vect_loop_def); if (STMT_VINFO_LIVE_P (stmt_info)) { /* FORNOW: not yet supported. */ if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "value used after loop."); return false; } if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT) return false; if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) != SSA_NAME) return false; scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0); vectype_out = get_vectype_for_scalar_type (TREE_TYPE (scalar_dest)); nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); if (nunits_out != nunits_in) return false; operation = GIMPLE_STMT_OPERAND (stmt, 1); code = TREE_CODE (operation); optab = optab_for_tree_code (code, vectype); /* Support only unary or binary operations. */ op_type = TREE_CODE_LENGTH (code); if (op_type != unary_op && op_type != binary_op) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "num. args = %d (not unary/binary op).", op_type); return false; } op0 = TREE_OPERAND (operation, 0); if (!vect_is_simple_use (op0, loop_vinfo, &def_stmt, &def, &dt0)) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "use not simple."); return false; } if (op_type == binary_op) { op1 = TREE_OPERAND (operation, 1); if (!vect_is_simple_use (op1, loop_vinfo, &def_stmt, &def, &dt1)) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "use not simple."); return false; } } /* Supportable by target? */ if (!optab) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "no optab."); return false; } vec_mode = TYPE_MODE (vectype); icode = (int) optab->handlers[(int) vec_mode].insn_code; if (icode == CODE_FOR_nothing) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "op not supported by target."); if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD || LOOP_VINFO_VECT_FACTOR (loop_vinfo) < vect_min_worthwhile_factor (code)) return false; if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "proceeding using word mode."); } /* Worthwhile without SIMD support? */ if (!VECTOR_MODE_P (TYPE_MODE (vectype)) && LOOP_VINFO_VECT_FACTOR (loop_vinfo) < vect_min_worthwhile_factor (code)) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "not worthwhile without SIMD support."); return false; } if (code == LSHIFT_EXPR || code == RSHIFT_EXPR) { /* FORNOW: not yet supported. */ if (!VECTOR_MODE_P (vec_mode)) return false; /* Invariant argument is needed for a vector shift by a scalar shift operand. */ optab_op2_mode = insn_data[icode].operand[2].mode; if (! (VECTOR_MODE_P (optab_op2_mode) || dt1 == vect_constant_def || dt1 == vect_invariant_def)) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "operand mode requires invariant argument."); return false; } } if (!vec_stmt) /* transformation not required. */ { STMT_VINFO_TYPE (stmt_info) = op_vec_info_type; return true; } /** Transform. **/ if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "transform binary/unary operation."); /* Handle def. */ vec_dest = vect_create_destination_var (scalar_dest, vectype); /* In case the vectorization factor (VF) is bigger than the number of elements that we can fit in a vectype (nunits), we have to generate more than one vector stmt - i.e - we need to "unroll" the vector stmt by a factor VF/nunits. In doing so, we record a pointer from one copy of the vector stmt to the next, in the field STMT_VINFO_RELATED_STMT. This is necessary in order to allow following stages to find the correct vector defs to be used when vectorizing stmts that use the defs of the current stmt. The example below illustrates the vectorization process when VF=16 and nunits=4 (i.e - we need to create 4 vectorized stmts): before vectorization: RELATED_STMT VEC_STMT S1: x = memref - - S2: z = x + 1 - - step 1: vectorize stmt S1 (done in vectorizable_load. See more details there): RELATED_STMT VEC_STMT VS1_0: vx0 = memref0 VS1_1 - VS1_1: vx1 = memref1 VS1_2 - VS1_2: vx2 = memref2 VS1_3 - VS1_3: vx3 = memref3 - - S1: x = load - VS1_0 S2: z = x + 1 - - step2: vectorize stmt S2 (done here): To vectorize stmt S2 we first need to find the relevant vector def for the first operand 'x'. This is, as usual, obtained from the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt that defines 'x' (S1). This way we find the stmt VS1_0, and the relevant vector def 'vx0'. Having found 'vx0' we can generate the vector stmt VS2_0, and as usual, record it in the STMT_VINFO_VEC_STMT of stmt S2. When creating the second copy (VS2_1), we obtain the relevant vector def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of stmt VS1_0. This way we find the stmt VS1_1 and the relevant vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0. Similarly when creating stmts VS2_2 and VS2_3. This is the resulting chain of stmts and pointers: RELATED_STMT VEC_STMT VS1_0: vx0 = memref0 VS1_1 - VS1_1: vx1 = memref1 VS1_2 - VS1_2: vx2 = memref2 VS1_3 - VS1_3: vx3 = memref3 - - S1: x = load - VS1_0 VS2_0: vz0 = vx0 + v1 VS2_1 - VS2_1: vz1 = vx1 + v1 VS2_2 - VS2_2: vz2 = vx2 + v1 VS2_3 - VS2_3: vz3 = vx3 + v1 - - S2: z = x + 1 - VS2_0 */ prev_stmt_info = NULL; for (j = 0; j < ncopies; j++) { /* Handle uses. */ if (j == 0) { vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); if (op_type == binary_op) { if (code == LSHIFT_EXPR || code == RSHIFT_EXPR) { /* Vector shl and shr insn patterns can be defined with scalar operand 2 (shift operand). In this case, use constant or loop invariant op1 directly, without extending it to vector mode first. */ optab_op2_mode = insn_data[icode].operand[2].mode; if (!VECTOR_MODE_P (optab_op2_mode)) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "operand 1 using scalar mode."); vec_oprnd1 = op1; } } if (!vec_oprnd1) vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, NULL); } } else { vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt0, vec_oprnd0); if (op_type == binary_op) vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt1, vec_oprnd1); } /* Arguments are ready. create the new vector stmt. */ if (op_type == binary_op) new_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, vec_dest, build2 (code, vectype, vec_oprnd0, vec_oprnd1)); else new_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, vec_dest, build1 (code, vectype, vec_oprnd0)); new_temp = make_ssa_name (vec_dest, new_stmt); GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp; vect_finish_stmt_generation (stmt, new_stmt, bsi); if (j == 0) STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; else STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; prev_stmt_info = vinfo_for_stmt (new_stmt); } return true; } /* Function vectorizable_type_demotion Check if STMT performs a binary or unary operation that involves type demotion, and if it can be vectorized. If VEC_STMT is also passed, vectorize the STMT: create a vectorized stmt to replace it, put it in VEC_STMT, and insert it at BSI. Return FALSE if not a vectorizable STMT, TRUE otherwise. */ bool vectorizable_type_demotion (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt) { tree vec_dest; tree scalar_dest; tree operation; tree op0; tree vec_oprnd0=NULL, vec_oprnd1=NULL; stmt_vec_info stmt_info = vinfo_for_stmt (stmt); loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); enum tree_code code; tree new_temp; tree def, def_stmt; enum vect_def_type dt0; tree new_stmt; stmt_vec_info prev_stmt_info; int nunits_in; int nunits_out; tree vectype_out; int ncopies; int j; tree expr; tree vectype_in; tree scalar_type; optab optab; enum machine_mode vec_mode; /* Is STMT a vectorizable type-demotion operation? */ if (!STMT_VINFO_RELEVANT_P (stmt_info)) return false; gcc_assert (STMT_VINFO_DEF_TYPE (stmt_info) == vect_loop_def); if (STMT_VINFO_LIVE_P (stmt_info)) { /* FORNOW: not yet supported. */ if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "value used after loop."); return false; } if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT) return false; if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) != SSA_NAME) return false; operation = GIMPLE_STMT_OPERAND (stmt, 1); code = TREE_CODE (operation); if (code != NOP_EXPR && code != CONVERT_EXPR) return false; op0 = TREE_OPERAND (operation, 0); vectype_in = get_vectype_for_scalar_type (TREE_TYPE (op0)); nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0); scalar_type = TREE_TYPE (scalar_dest); vectype_out = get_vectype_for_scalar_type (scalar_type); nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); if (nunits_in != nunits_out / 2) /* FORNOW */ return false; ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out; gcc_assert (ncopies >= 1); if (! INTEGRAL_TYPE_P (scalar_type) || !INTEGRAL_TYPE_P (TREE_TYPE (op0))) return false; /* Check the operands of the operation. */ if (!vect_is_simple_use (op0, loop_vinfo, &def_stmt, &def, &dt0)) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "use not simple."); return false; } /* Supportable by target? */ code = VEC_PACK_MOD_EXPR; optab = optab_for_tree_code (VEC_PACK_MOD_EXPR, vectype_in); if (!optab) return false; vec_mode = TYPE_MODE (vectype_in); if (optab->handlers[(int) vec_mode].insn_code == CODE_FOR_nothing) return false; STMT_VINFO_VECTYPE (stmt_info) = vectype_in; if (!vec_stmt) /* transformation not required. */ { STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type; return true; } /** Transform. **/ if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "transform type demotion operation. ncopies = %d.", ncopies); /* Handle def. */ vec_dest = vect_create_destination_var (scalar_dest, vectype_out); /* In case the vectorization factor (VF) is bigger than the number of elements that we can fit in a vectype (nunits), we have to generate more than one vector stmt - i.e - we need to "unroll" the vector stmt by a factor VF/nunits. */ prev_stmt_info = NULL; for (j = 0; j < ncopies; j++) { /* Handle uses. */ if (j == 0) { enum vect_def_type dt = vect_unknown_def_type; /* Dummy */ vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd0); } else { vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt0, vec_oprnd1); vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt0, vec_oprnd0); } /* Arguments are ready. Create the new vector stmt. */ expr = build2 (code, vectype_out, vec_oprnd0, vec_oprnd1); new_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, vec_dest, expr); new_temp = make_ssa_name (vec_dest, new_stmt); GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp; vect_finish_stmt_generation (stmt, new_stmt, bsi); if (j == 0) STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; else STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; prev_stmt_info = vinfo_for_stmt (new_stmt); } *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); return true; } /* Function vect_gen_widened_results_half Create a vector stmt whose code, type, number of arguments, and result variable are CODE, VECTYPE, OP_TYPE, and VEC_DEST, and its arguments are VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI. In the case that CODE is a CALL_EXPR, this means that a call to DECL needs to be created (DECL is a function-decl of a target-builtin). STMT is the original scalar stmt that we are vectorizing. */ static tree vect_gen_widened_results_half (enum tree_code code, tree vectype, tree decl, tree vec_oprnd0, tree vec_oprnd1, int op_type, tree vec_dest, block_stmt_iterator *bsi, tree stmt) { tree vec_params; tree expr; tree new_stmt; tree new_temp; tree sym; ssa_op_iter iter; /* Generate half of the widened result: */ if (code == CALL_EXPR) { /* Target specific support */ vec_params = build_tree_list (NULL_TREE, vec_oprnd0); if (op_type == binary_op) vec_params = tree_cons (NULL_TREE, vec_oprnd1, vec_params); expr = build_function_call_expr (decl, vec_params); } else { /* Generic support */ gcc_assert (op_type == TREE_CODE_LENGTH (code)); if (op_type == binary_op) expr = build2 (code, vectype, vec_oprnd0, vec_oprnd1); else expr = build1 (code, vectype, vec_oprnd0); } new_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, vec_dest, expr); new_temp = make_ssa_name (vec_dest, new_stmt); GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp; vect_finish_stmt_generation (stmt, new_stmt, bsi); if (code == CALL_EXPR) { FOR_EACH_SSA_TREE_OPERAND (sym, new_stmt, iter, SSA_OP_ALL_VIRTUALS) { if (TREE_CODE (sym) == SSA_NAME) sym = SSA_NAME_VAR (sym); mark_sym_for_renaming (sym); } } return new_stmt; } /* Function vectorizable_type_promotion Check if STMT performs a binary or unary operation that involves type promotion, and if it can be vectorized. If VEC_STMT is also passed, vectorize the STMT: create a vectorized stmt to replace it, put it in VEC_STMT, and insert it at BSI. Return FALSE if not a vectorizable STMT, TRUE otherwise. */ bool vectorizable_type_promotion (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt) { tree vec_dest; tree scalar_dest; tree operation; tree op0, op1 = NULL; tree vec_oprnd0=NULL, vec_oprnd1=NULL; stmt_vec_info stmt_info = vinfo_for_stmt (stmt); loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); enum tree_code code, code1 = CODE_FOR_nothing, code2 = CODE_FOR_nothing; tree decl1 = NULL_TREE, decl2 = NULL_TREE; int op_type; tree def, def_stmt; enum vect_def_type dt0, dt1; tree new_stmt; stmt_vec_info prev_stmt_info; int nunits_in; int nunits_out; tree vectype_out; int ncopies; int j; tree vectype_in; /* Is STMT a vectorizable type-promotion operation? */ if (!STMT_VINFO_RELEVANT_P (stmt_info)) return false; gcc_assert (STMT_VINFO_DEF_TYPE (stmt_info) == vect_loop_def); if (STMT_VINFO_LIVE_P (stmt_info)) { /* FORNOW: not yet supported. */ if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "value used after loop."); return false; } if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT) return false; if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) != SSA_NAME) return false; operation = GIMPLE_STMT_OPERAND (stmt, 1); code = TREE_CODE (operation); if (code != NOP_EXPR && code != WIDEN_MULT_EXPR) return false; op0 = TREE_OPERAND (operation, 0); vectype_in = get_vectype_for_scalar_type (TREE_TYPE (op0)); nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in); ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in; gcc_assert (ncopies >= 1); scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0); vectype_out = get_vectype_for_scalar_type (TREE_TYPE (scalar_dest)); nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out); if (nunits_out != nunits_in / 2) /* FORNOW */ return false; if (! INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest)) || !INTEGRAL_TYPE_P (TREE_TYPE (op0))) return false; /* Check the operands of the operation. */ if (!vect_is_simple_use (op0, loop_vinfo, &def_stmt, &def, &dt0)) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "use not simple."); return false; } op_type = TREE_CODE_LENGTH (code); if (op_type == binary_op) { op1 = TREE_OPERAND (operation, 1); if (!vect_is_simple_use (op1, loop_vinfo, &def_stmt, &def, &dt1)) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "use not simple."); return false; } } /* Supportable by target? */ if (!supportable_widening_operation (code, stmt, vectype_in, &decl1, &decl2, &code1, &code2)) return false; STMT_VINFO_VECTYPE (stmt_info) = vectype_in; if (!vec_stmt) /* transformation not required. */ { STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type; return true; } /** Transform. **/ if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "transform type promotion operation. ncopies = %d.", ncopies); /* Handle def. */ vec_dest = vect_create_destination_var (scalar_dest, vectype_out); /* In case the vectorization factor (VF) is bigger than the number of elements that we can fit in a vectype (nunits), we have to generate more than one vector stmt - i.e - we need to "unroll" the vector stmt by a factor VF/nunits. */ prev_stmt_info = NULL; for (j = 0; j < ncopies; j++) { /* Handle uses. */ if (j == 0) { vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL); if (op_type == binary_op) vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, NULL); } else { vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt0, vec_oprnd0); if (op_type == binary_op) vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt1, vec_oprnd1); } /* Arguments are ready. Create the new vector stmt. We are creating two vector defs because the widened result does not fit in one vector. The vectorized stmt can be expressed as a call to a taregt builtin, or a using a tree-code. */ /* Generate first half of the widened result: */ new_stmt = vect_gen_widened_results_half (code1, vectype_out, decl1, vec_oprnd0, vec_oprnd1, op_type, vec_dest, bsi, stmt); if (j == 0) STMT_VINFO_VEC_STMT (stmt_info) = new_stmt; else STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; prev_stmt_info = vinfo_for_stmt (new_stmt); /* Generate second half of the widened result: */ new_stmt = vect_gen_widened_results_half (code2, vectype_out, decl2, vec_oprnd0, vec_oprnd1, op_type, vec_dest, bsi, stmt); STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; prev_stmt_info = vinfo_for_stmt (new_stmt); } *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); return true; } /* Function vect_strided_store_supported. Returns TRUE is INTERLEAVE_HIGH and INTERLEAVE_LOW operations are supported, and FALSE otherwise. */ static bool vect_strided_store_supported (tree vectype) { optab interleave_high_optab, interleave_low_optab; int mode; mode = (int) TYPE_MODE (vectype); /* Check that the operation is supported. */ interleave_high_optab = optab_for_tree_code (VEC_INTERLEAVE_HIGH_EXPR, vectype); interleave_low_optab = optab_for_tree_code (VEC_INTERLEAVE_LOW_EXPR, vectype); if (!interleave_high_optab || !interleave_low_optab) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "no optab for interleave."); return false; } if (interleave_high_optab->handlers[(int) mode].insn_code == CODE_FOR_nothing || interleave_low_optab->handlers[(int) mode].insn_code == CODE_FOR_nothing) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "interleave op not supported by target."); return false; } return true; } /* 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. */ static bool vect_permute_store_chain (VEC(tree,heap) *dr_chain, unsigned int length, tree stmt, block_stmt_iterator *bsi, VEC(tree,heap) **result_chain) { tree perm_dest, perm_stmt, vect1, vect2, high, low; tree vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt)); tree scalar_dest; int i; unsigned int j; VEC(tree,heap) *first, *second; scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0); first = VEC_alloc (tree, heap, length/2); second = VEC_alloc (tree, heap, length/2); /* Check that the operation is supported. */ if (!vect_strided_store_supported (vectype)) return false; *result_chain = VEC_copy (tree, heap, dr_chain); for (i = 0; i < exact_log2 (length); i++) { for (j = 0; j < length/2; j++) { vect1 = VEC_index (tree, dr_chain, j); vect2 = VEC_index (tree, dr_chain, j+length/2); /* Create interleaving stmt: in the case of big endian: high = interleave_high (vect1, vect2) and in the case of little endian: high = interleave_low (vect1, vect2). */ perm_dest = create_tmp_var (vectype, "vect_inter_high"); DECL_GIMPLE_REG_P (perm_dest) = 1; add_referenced_var (perm_dest); if (BYTES_BIG_ENDIAN) perm_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, perm_dest, build2 (VEC_INTERLEAVE_HIGH_EXPR, vectype, vect1, vect2)); else perm_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, perm_dest, build2 (VEC_INTERLEAVE_LOW_EXPR, vectype, vect1, vect2)); high = make_ssa_name (perm_dest, perm_stmt); GIMPLE_STMT_OPERAND (perm_stmt, 0) = high; vect_finish_stmt_generation (stmt, perm_stmt, bsi); VEC_replace (tree, *result_chain, 2*j, high); /* Create interleaving stmt: in the case of big endian: low = interleave_low (vect1, vect2) and in the case of little endian: low = interleave_high (vect1, vect2). */ perm_dest = create_tmp_var (vectype, "vect_inter_low"); DECL_GIMPLE_REG_P (perm_dest) = 1; add_referenced_var (perm_dest); if (BYTES_BIG_ENDIAN) perm_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, perm_dest, build2 (VEC_INTERLEAVE_LOW_EXPR, vectype, vect1, vect2)); else perm_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, perm_dest, build2 (VEC_INTERLEAVE_HIGH_EXPR, vectype, vect1, vect2)); low = make_ssa_name (perm_dest, perm_stmt); GIMPLE_STMT_OPERAND (perm_stmt, 0) = low; vect_finish_stmt_generation (stmt, perm_stmt, bsi); VEC_replace (tree, *result_chain, 2*j+1, low); } dr_chain = VEC_copy (tree, heap, *result_chain); } return true; } /* Function vectorizable_store. Check if STMT defines a non scalar data-ref (array/pointer/structure) that can be vectorized. If VEC_STMT is also passed, vectorize the STMT: create a vectorized stmt to replace it, put it in VEC_STMT, and insert it at BSI. Return FALSE if not a vectorizable STMT, TRUE otherwise. */ bool vectorizable_store (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt) { tree scalar_dest; tree data_ref; tree op; tree vec_oprnd = NULL_TREE; stmt_vec_info stmt_info = vinfo_for_stmt (stmt); struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL; tree vectype = STMT_VINFO_VECTYPE (stmt_info); loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); enum machine_mode vec_mode; tree dummy; enum dr_alignment_support alignment_support_cheme; ssa_op_iter iter; def_operand_p def_p; tree def, def_stmt; enum vect_def_type dt; stmt_vec_info prev_stmt_info = NULL; tree dataref_ptr = NULL_TREE; int nunits = TYPE_VECTOR_SUBPARTS (vectype); int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; int j; tree next_stmt, first_stmt; bool strided_store = false; unsigned int group_size, i; VEC(tree,heap) *dr_chain = NULL, *oprnds = NULL, *result_chain = NULL; gcc_assert (ncopies >= 1); /* Is vectorizable store? */ if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT) return false; scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0); if (TREE_CODE (scalar_dest) != ARRAY_REF && TREE_CODE (scalar_dest) != INDIRECT_REF && !DR_GROUP_FIRST_DR (stmt_info)) return false; op = GIMPLE_STMT_OPERAND (stmt, 1); if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt)) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "use not simple."); return false; } vec_mode = TYPE_MODE (vectype); /* FORNOW. In some cases can vectorize even if data-type not supported (e.g. - array initialization with 0). */ if (mov_optab->handlers[(int)vec_mode].insn_code == CODE_FOR_nothing) return false; if (!STMT_VINFO_DATA_REF (stmt_info)) return false; if (DR_GROUP_FIRST_DR (stmt_info)) { strided_store = true; if (!vect_strided_store_supported (vectype)) return false; } if (!vec_stmt) /* transformation not required. */ { STMT_VINFO_TYPE (stmt_info) = store_vec_info_type; return true; } /** Transform. **/ if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "transform store. ncopies = %d",ncopies); if (strided_store) { first_stmt = DR_GROUP_FIRST_DR (stmt_info); first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++; /* We vectorize all the stmts of the interleaving group when we reach the last stmt in the group. */ if (DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt)) < DR_GROUP_SIZE (vinfo_for_stmt (first_stmt))) { *vec_stmt = NULL_TREE; return true; } } else { first_stmt = stmt; first_dr = dr; group_size = 1; } dr_chain = VEC_alloc (tree, heap, group_size); oprnds = VEC_alloc (tree, heap, group_size); alignment_support_cheme = vect_supportable_dr_alignment (first_dr); gcc_assert (alignment_support_cheme); gcc_assert (alignment_support_cheme == dr_aligned); /* FORNOW */ /* In case the vectorization factor (VF) is bigger than the number of elements that we can fit in a vectype (nunits), we have to generate more than one vector stmt - i.e - we need to "unroll" the vector stmt by a factor VF/nunits. For more details see documentation in vect_get_vec_def_for_copy_stmt. */ /* In case of interleaving (non-unit strided access): S1: &base + 2 = x2 S2: &base = x0 S3: &base + 1 = x1 S4: &base + 3 = x3 We create vectorized stores starting from base address (the access of the first stmt in the chain (S2 in the above example), when the last store stmt of the chain (S4) is reached: VS1: &base = vx2 VS2: &base + vec_size*1 = vx0 VS3: &base + vec_size*2 = vx1 VS4: &base + vec_size*3 = vx3 Then permutation statements are generated: VS5: vx5 = VEC_INTERLEAVE_HIGH_EXPR < vx0, vx3 > VS6: vx6 = VEC_INTERLEAVE_LOW_EXPR < vx0, vx3 > ... And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts (the order of the data-refs in the output of vect_permute_store_chain corresponds to the order of scalar stmts in the interleaving chain - see the documentation of vect_permute_store_chain()). In case of both multiple types and interleaving, above vector stores and permutation stmts are created for every copy. The result vector stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding STMT_VINFO_RELATED_STMT for the next copies. */ prev_stmt_info = NULL; for (j = 0; j < ncopies; j++) { tree new_stmt; tree ptr_incr; if (j == 0) { /* For interleaved stores we collect vectorized defs for all the stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then used as an input to vect_permute_store_chain(), and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the next copy. If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and OPRNDS are of size 1. */ next_stmt = first_stmt; for (i = 0; i < group_size; i++) { /* Since gaps are not supported for interleaved stores, GROUP_SIZE is the exact number of stmts in the chain. Therefore, NEXT_STMT can't be NULL_TREE. In case that there is no interleaving, GROUP_SIZE is 1, and only one iteration of the loop will be executed. */ gcc_assert (next_stmt); op = GIMPLE_STMT_OPERAND (next_stmt, 1); vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt, NULL); VEC_quick_push(tree, dr_chain, vec_oprnd); VEC_quick_push(tree, oprnds, vec_oprnd); next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt)); } dataref_ptr = vect_create_data_ref_ptr (first_stmt, bsi, NULL_TREE, &dummy, &ptr_incr, false, TREE_TYPE (vec_oprnd)); } else { /* For interleaved stores we created vectorized defs for all the defs stored in OPRNDS in the previous iteration (previous copy). DR_CHAIN is then used as an input to vect_permute_store_chain(), and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the next copy. If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and OPRNDS are of size 1. */ for (i = 0; i < group_size; i++) { vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, VEC_index (tree, oprnds, i)); VEC_replace(tree, dr_chain, i, vec_oprnd); VEC_replace(tree, oprnds, i, vec_oprnd); } dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, bsi, stmt); } if (strided_store) { result_chain = VEC_alloc (tree, heap, group_size); /* Permute. */ if (!vect_permute_store_chain (dr_chain, group_size, stmt, bsi, &result_chain)) return false; } next_stmt = first_stmt; for (i = 0; i < group_size; i++) { /* For strided stores vectorized defs are interleaved in vect_permute_store_chain(). */ if (strided_store) vec_oprnd = VEC_index(tree, result_chain, i); data_ref = build_fold_indirect_ref (dataref_ptr); /* Arguments are ready. Create the new vector stmt. */ new_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, data_ref, vec_oprnd); vect_finish_stmt_generation (stmt, new_stmt, bsi); /* Set the VDEFs for the vector pointer. If this virtual def has a use outside the loop and a loop peel is performed then the def may be renamed by the peel. Mark it for renaming so the later use will also be renamed. */ copy_virtual_operands (new_stmt, next_stmt); if (j == 0) { /* The original store is deleted so the same SSA_NAMEs can be used. */ FOR_EACH_SSA_TREE_OPERAND (def, next_stmt, iter, SSA_OP_VDEF) { SSA_NAME_DEF_STMT (def) = new_stmt; mark_sym_for_renaming (SSA_NAME_VAR (def)); } STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; } else { /* Create new names for all the definitions created by COPY and add replacement mappings for each new name. */ FOR_EACH_SSA_DEF_OPERAND (def_p, new_stmt, iter, SSA_OP_VDEF) { create_new_def_for (DEF_FROM_PTR (def_p), new_stmt, def_p); mark_sym_for_renaming (SSA_NAME_VAR (DEF_FROM_PTR (def_p))); } STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; } prev_stmt_info = vinfo_for_stmt (new_stmt); next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt)); if (!next_stmt) break; /* Bump the vector pointer. */ dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, bsi, stmt); } } return true; } /* Function vect_setup_realignment This function is called when vectorizing an unaligned load using the dr_unaligned_software_pipeline scheme. This function generates the following code at the loop prolog: p = initial_addr; msq_init = *(floor(p)); # prolog load realignment_token = call target_builtin; loop: msq = phi (msq_init, ---) 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). 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); 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. 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. */ static tree vect_setup_realignment (tree stmt, block_stmt_iterator *bsi, tree *realignment_token) { 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 loop *loop = LOOP_VINFO_LOOP (loop_vinfo); edge pe = loop_preheader_edge (loop); tree scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0); tree vec_dest; tree init_addr; tree inc; tree ptr; tree data_ref; tree new_stmt; basic_block new_bb; tree msq_init; tree new_temp; tree phi_stmt; tree msq; /* 1. Create msq_init = *(floor(p1)) in the loop preheader */ vec_dest = vect_create_destination_var (scalar_dest, vectype); ptr = vect_create_data_ref_ptr (stmt, bsi, NULL_TREE, &init_addr, &inc, true, NULL_TREE); data_ref = build1 (ALIGN_INDIRECT_REF, vectype, ptr); new_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, vec_dest, data_ref); new_temp = make_ssa_name (vec_dest, new_stmt); GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp; new_bb = bsi_insert_on_edge_immediate (pe, new_stmt); gcc_assert (!new_bb); msq_init = GIMPLE_STMT_OPERAND (new_stmt, 0); copy_virtual_operands (new_stmt, stmt); update_vuses_to_preheader (new_stmt, loop); /* 2. Create permutation mask, if required, in loop preheader. */ if (targetm.vectorize.builtin_mask_for_load) { tree builtin_decl; tree params = build_tree_list (NULL_TREE, init_addr); builtin_decl = targetm.vectorize.builtin_mask_for_load (); new_stmt = build_function_call_expr (builtin_decl, params); vec_dest = vect_create_destination_var (scalar_dest, TREE_TYPE (new_stmt)); new_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, vec_dest, new_stmt); new_temp = make_ssa_name (vec_dest, new_stmt); GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp; new_bb = bsi_insert_on_edge_immediate (pe, new_stmt); gcc_assert (!new_bb); *realignment_token = GIMPLE_STMT_OPERAND (new_stmt, 0); /* 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)); } /* 3. Create msq = phi in loop */ vec_dest = vect_create_destination_var (scalar_dest, vectype); msq = make_ssa_name (vec_dest, NULL_TREE); phi_stmt = create_phi_node (msq, loop->header); SSA_NAME_DEF_STMT (msq) = phi_stmt; add_phi_arg (phi_stmt, msq_init, loop_preheader_edge (loop)); return msq; } /* Function vect_strided_load_supported. Returns TRUE is EXTRACT_EVEN and EXTRACT_ODD operations are supported, and FALSE otherwise. */ static bool vect_strided_load_supported (tree vectype) { optab perm_even_optab, perm_odd_optab; int mode; mode = (int) TYPE_MODE (vectype); perm_even_optab = optab_for_tree_code (VEC_EXTRACT_EVEN_EXPR, vectype); if (!perm_even_optab) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "no optab for perm_even."); return false; } if (perm_even_optab->handlers[mode].insn_code == CODE_FOR_nothing) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "perm_even op not supported by target."); return false; } perm_odd_optab = optab_for_tree_code (VEC_EXTRACT_ODD_EXPR, vectype); if (!perm_odd_optab) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "no optab for perm_odd."); return false; } if (perm_odd_optab->handlers[mode].insn_code == CODE_FOR_nothing) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "perm_odd op not supported by target."); return false; } return true; } /* 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 bool vect_permute_load_chain (VEC(tree,heap) *dr_chain, unsigned int length, tree stmt, block_stmt_iterator *bsi, VEC(tree,heap) **result_chain) { tree perm_dest, perm_stmt, data_ref, first_vect, second_vect; tree vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt)); int i; unsigned int j; /* Check that the operation is supported. */ if (!vect_strided_load_supported (vectype)) return false; *result_chain = VEC_copy (tree, heap, dr_chain); for (i = 0; i < exact_log2 (length); i++) { for (j = 0; j < length; j +=2) { first_vect = VEC_index (tree, dr_chain, j); second_vect = VEC_index (tree, dr_chain, j+1); /* data_ref = permute_even (first_data_ref, second_data_ref); */ perm_dest = create_tmp_var (vectype, "vect_perm_even"); DECL_GIMPLE_REG_P (perm_dest) = 1; add_referenced_var (perm_dest); perm_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, perm_dest, build2 (VEC_EXTRACT_EVEN_EXPR, vectype, first_vect, second_vect)); data_ref = make_ssa_name (perm_dest, perm_stmt); GIMPLE_STMT_OPERAND (perm_stmt, 0) = data_ref; vect_finish_stmt_generation (stmt, perm_stmt, bsi); mark_symbols_for_renaming (perm_stmt); VEC_replace (tree, *result_chain, j/2, data_ref); /* data_ref = permute_odd (first_data_ref, second_data_ref); */ perm_dest = create_tmp_var (vectype, "vect_perm_odd"); DECL_GIMPLE_REG_P (perm_dest) = 1; add_referenced_var (perm_dest); perm_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, perm_dest, build2 (VEC_EXTRACT_ODD_EXPR, vectype, first_vect, second_vect)); data_ref = make_ssa_name (perm_dest, perm_stmt); GIMPLE_STMT_OPERAND (perm_stmt, 0) = data_ref; vect_finish_stmt_generation (stmt, perm_stmt, bsi); mark_symbols_for_renaming (perm_stmt); VEC_replace (tree, *result_chain, j/2+length/2, data_ref); } dr_chain = VEC_copy (tree, heap, *result_chain); } return true; } /* Function vect_transform_strided_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. */ static bool vect_transform_strided_load (tree stmt, VEC(tree,heap) *dr_chain, int size, block_stmt_iterator *bsi) { stmt_vec_info stmt_info = vinfo_for_stmt (stmt); tree first_stmt = DR_GROUP_FIRST_DR (stmt_info); tree next_stmt, new_stmt; VEC(tree,heap) *result_chain = NULL; unsigned int i, gap_count; tree tmp_data_ref; /* 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 = VEC_alloc (tree, heap, size); /* Permute. */ if (!vect_permute_load_chain (dr_chain, size, stmt, bsi, &result_chain)) return false; /* 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 (i = 0; VEC_iterate(tree, result_chain, i, tmp_data_ref); i++) { if (!next_stmt) break; /* Skip the gaps. Loads created for the gaps will be removed by dead code elimination pass later. DR_GROUP_GAP is the number of steps in elements from the previous access (if there is no gap DR_GROUP_GAP is 1). We skip loads that correspond to the gaps. */ if (gap_count < DR_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 { tree prev_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt (next_stmt)); tree 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 = DR_GROUP_NEXT_DR (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 || !DR_GROUP_SAME_DR_STMT (vinfo_for_stmt (next_stmt))) break; } } return true; } /* vectorizable_load. Check if STMT reads a non scalar data-ref (array/pointer/structure) that can be vectorized. If VEC_STMT is also passed, vectorize the STMT: create a vectorized stmt to replace it, put it in VEC_STMT, and insert it at BSI. Return FALSE if not a vectorizable STMT, TRUE otherwise. */ bool vectorizable_load (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt) { tree scalar_dest; tree vec_dest = NULL; tree data_ref = NULL; tree op; stmt_vec_info stmt_info = vinfo_for_stmt (stmt); stmt_vec_info prev_stmt_info; loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr; tree vectype = STMT_VINFO_VECTYPE (stmt_info); tree new_temp; int mode; tree new_stmt = NULL_TREE; tree dummy; enum dr_alignment_support alignment_support_cheme; tree dataref_ptr = NULL_TREE; tree ptr_incr; int nunits = TYPE_VECTOR_SUBPARTS (vectype); int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; int i, j, group_size; tree msq = NULL_TREE, lsq; tree offset = NULL_TREE; tree realignment_token = NULL_TREE; tree phi_stmt = NULL_TREE; VEC(tree,heap) *dr_chain = NULL; bool strided_load = false; tree first_stmt; /* Is vectorizable load? */ if (!STMT_VINFO_RELEVANT_P (stmt_info)) return false; gcc_assert (STMT_VINFO_DEF_TYPE (stmt_info) == vect_loop_def); if (STMT_VINFO_LIVE_P (stmt_info)) { /* FORNOW: not yet supported. */ if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "value used after loop."); return false; } if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT) return false; scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0); if (TREE_CODE (scalar_dest) != SSA_NAME) return false; op = GIMPLE_STMT_OPERAND (stmt, 1); if (TREE_CODE (op) != ARRAY_REF && TREE_CODE (op) != INDIRECT_REF && !DR_GROUP_FIRST_DR (stmt_info)) return false; if (!STMT_VINFO_DATA_REF (stmt_info)) return false; mode = (int) TYPE_MODE (vectype); /* FORNOW. In some cases can vectorize even if data-type not supported (e.g. - data copies). */ if (mov_optab->handlers[mode].insn_code == CODE_FOR_nothing) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "Aligned load, but unsupported type."); return false; } /* Check if the load is a part of an interleaving chain. */ if (DR_GROUP_FIRST_DR (stmt_info)) { strided_load = true; /* Check if interleaving is supported. */ if (!vect_strided_load_supported (vectype)) return false; } if (!vec_stmt) /* transformation not required. */ { STMT_VINFO_TYPE (stmt_info) = load_vec_info_type; return true; } /** Transform. **/ if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "transform load."); if (strided_load) { first_stmt = DR_GROUP_FIRST_DR (stmt_info); /* Check if the chain of loads is already vectorized. */ if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt))) { *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); return true; } first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt)); dr_chain = VEC_alloc (tree, heap, group_size); } else { first_stmt = stmt; first_dr = dr; group_size = 1; } alignment_support_cheme = vect_supportable_dr_alignment (first_dr); gcc_assert (alignment_support_cheme); /* In case the vectorization factor (VF) is bigger than the number of elements that we can fit in a vectype (nunits), we have to generate more than one vector stmt - i.e - we need to "unroll" the vector stmt by a factor VF/nunits. In doing so, we record a pointer from one copy of the vector stmt to the next, in the field STMT_VINFO_RELATED_STMT. This is necessary in order to allow following stages to find the correct vector defs to be used when vectorizing stmts that use the defs of the current stmt. The example below illustrates the vectorization process when VF=16 and nunits=4 (i.e - we need to create 4 vectorized stmts): before vectorization: RELATED_STMT VEC_STMT S1: x = memref - - S2: z = x + 1 - - step 1: vectorize stmt S1: We first create the vector stmt VS1_0, and, as usual, record a pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1. Next, we create the vector stmt VS1_1, and record a pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0. Similarly, for VS1_2 and VS1_3. This is the resulting chain of stmts and pointers: RELATED_STMT VEC_STMT VS1_0: vx0 = memref0 VS1_1 - VS1_1: vx1 = memref1 VS1_2 - VS1_2: vx2 = memref2 VS1_3 - VS1_3: vx3 = memref3 - - S1: x = load - VS1_0 S2: z = x + 1 - - See in documentation in vect_get_vec_def_for_stmt_copy for how the information we recorded in RELATED_STMT field is used to vectorize stmt S2. */ /* In case of interleaving (non-unit strided access): S1: x2 = &base + 2 S2: x0 = &base S3: x1 = &base + 1 S4: x3 = &base + 3 Vectorized loads are created in the order of memory accesses starting from the access of the first stmt of the chain: VS1: vx0 = &base VS2: vx1 = &base + vec_size*1 VS3: vx3 = &base + vec_size*2 VS4: vx4 = &base + vec_size*3 Then permutation statements are generated: VS5: vx5 = VEC_EXTRACT_EVEN_EXPR < vx0, vx1 > VS6: vx6 = VEC_EXTRACT_ODD_EXPR < vx0, vx1 > ... And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts (the order of the data-refs in the output of vect_permute_load_chain corresponds to the order of scalar stmts in the interleaving chain - see the documentation of vect_permute_load_chain()). The generation of permutation stmts and recording them in STMT_VINFO_VEC_STMT is done in vect_transform_strided_load(). In case of both multiple types and interleaving, the vector loads and permutation stmts above are created for every copy. The result vector stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding STMT_VINFO_RELATED_STMT for the next copies. */ /* If the data reference is aligned (dr_aligned) or potentially unaligned on a target that supports unaligned accesses (dr_unaligned_supported) we generate the following code: p = initial_addr; indx = 0; loop { p = p + indx * vectype_size; vec_dest = *(p); indx = indx + 1; } Otherwise, the data reference is potentially unaligned on a target that does not support unaligned accesses (dr_unaligned_software_pipeline) - then generate the following code, in which the data in each iteration is obtained by two vector loads, one from the previous iteration, and one from the current iteration: p1 = initial_addr; msq_init = *(floor(p1)) p2 = initial_addr + VS - 1; realignment_token = call target_builtin; indx = 0; loop { p2 = p2 + indx * vectype_size lsq = *(floor(p2)) vec_dest = realign_load (msq, lsq, realignment_token) indx = indx + 1; msq = lsq; } */ if (alignment_support_cheme == dr_unaligned_software_pipeline) { msq = vect_setup_realignment (first_stmt, bsi, &realignment_token); phi_stmt = SSA_NAME_DEF_STMT (msq); offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1); } prev_stmt_info = NULL; for (j = 0; j < ncopies; j++) { /* 1. Create the vector pointer update chain. */ if (j == 0) dataref_ptr = vect_create_data_ref_ptr (first_stmt, bsi, offset, &dummy, &ptr_incr, false, NULL_TREE); else dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, bsi, stmt); for (i = 0; i < group_size; i++) { /* 2. Create the vector-load in the loop. */ switch (alignment_support_cheme) { case dr_aligned: gcc_assert (aligned_access_p (first_dr)); data_ref = build_fold_indirect_ref (dataref_ptr); break; case dr_unaligned_supported: { int mis = DR_MISALIGNMENT (first_dr); tree tmis = (mis == -1 ? size_zero_node : size_int (mis)); gcc_assert (!aligned_access_p (first_dr)); tmis = size_binop (MULT_EXPR, tmis, size_int(BITS_PER_UNIT)); data_ref = build2 (MISALIGNED_INDIRECT_REF, vectype, dataref_ptr, tmis); break; } case dr_unaligned_software_pipeline: gcc_assert (!aligned_access_p (first_dr)); data_ref = build1 (ALIGN_INDIRECT_REF, vectype, dataref_ptr); break; default: gcc_unreachable (); } vec_dest = vect_create_destination_var (scalar_dest, vectype); new_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, vec_dest, data_ref); new_temp = make_ssa_name (vec_dest, new_stmt); GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp; vect_finish_stmt_generation (stmt, new_stmt, bsi); copy_virtual_operands (new_stmt, stmt); mark_symbols_for_renaming (new_stmt); /* 3. Handle explicit realignment if necessary/supported. */ if (alignment_support_cheme == dr_unaligned_software_pipeline) { /* Create in loop: */ lsq = GIMPLE_STMT_OPERAND (new_stmt, 0); if (!realignment_token) realignment_token = dataref_ptr; vec_dest = vect_create_destination_var (scalar_dest, vectype); new_stmt = build3 (REALIGN_LOAD_EXPR, vectype, msq, lsq, realignment_token); new_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, vec_dest, new_stmt); new_temp = make_ssa_name (vec_dest, new_stmt); GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp; vect_finish_stmt_generation (stmt, new_stmt, bsi); if (i == group_size - 1 && j == ncopies - 1) add_phi_arg (phi_stmt, lsq, loop_latch_edge (loop)); msq = lsq; } if (strided_load) VEC_quick_push (tree, dr_chain, new_temp); if (i < group_size - 1) dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, bsi, stmt); } if (strided_load) { if (!vect_transform_strided_load (stmt, dr_chain, group_size, bsi)) return false; *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); dr_chain = VEC_alloc (tree, heap, group_size); } else { if (j == 0) STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; else STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; prev_stmt_info = vinfo_for_stmt (new_stmt); } } return true; } /* Function vectorizable_live_operation. STMT computes a value that is used outside the loop. Check if it can be supported. */ bool vectorizable_live_operation (tree stmt, block_stmt_iterator *bsi ATTRIBUTE_UNUSED, tree *vec_stmt ATTRIBUTE_UNUSED) { tree operation; stmt_vec_info stmt_info = vinfo_for_stmt (stmt); loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); int i; enum tree_code code; int op_type; tree op; tree def, def_stmt; enum vect_def_type dt; if (!STMT_VINFO_LIVE_P (stmt_info)) return false; if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT) return false; if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) != SSA_NAME) return false; operation = GIMPLE_STMT_OPERAND (stmt, 1); code = TREE_CODE (operation); op_type = TREE_CODE_LENGTH (code); /* FORNOW: support only if all uses are invariant. This means that the scalar operations can remain in place, unvectorized. The original last scalar value that they compute will be used. */ for (i = 0; i < op_type; i++) { op = TREE_OPERAND (operation, i); if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt)) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "use not simple."); return false; } if (dt != vect_invariant_def && dt != vect_constant_def) return false; } /* No transformation is required for the cases we currently support. */ return true; } /* Function vect_is_simple_cond. Input: LOOP - the loop that is being vectorized. COND - Condition that is checked for simple use. Returns whether a COND can be vectorized. Checks whether condition operands are supportable using vec_is_simple_use. */ static bool vect_is_simple_cond (tree cond, loop_vec_info loop_vinfo) { tree lhs, rhs; tree def; enum vect_def_type dt; if (!COMPARISON_CLASS_P (cond)) return false; lhs = TREE_OPERAND (cond, 0); rhs = TREE_OPERAND (cond, 1); if (TREE_CODE (lhs) == SSA_NAME) { tree lhs_def_stmt = SSA_NAME_DEF_STMT (lhs); if (!vect_is_simple_use (lhs, loop_vinfo, &lhs_def_stmt, &def, &dt)) return false; } else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST) return false; if (TREE_CODE (rhs) == SSA_NAME) { tree rhs_def_stmt = SSA_NAME_DEF_STMT (rhs); if (!vect_is_simple_use (rhs, loop_vinfo, &rhs_def_stmt, &def, &dt)) return false; } else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST) return false; return true; } /* vectorizable_condition. Check if STMT is conditional modify expression that can be vectorized. If VEC_STMT is also passed, vectorize the STMT: create a vectorized stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it at BSI. Return FALSE if not a vectorizable STMT, TRUE otherwise. */ bool vectorizable_condition (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt) { tree scalar_dest = NULL_TREE; tree vec_dest = NULL_TREE; tree op = NULL_TREE; tree cond_expr, then_clause, else_clause; stmt_vec_info stmt_info = vinfo_for_stmt (stmt); tree vectype = STMT_VINFO_VECTYPE (stmt_info); tree vec_cond_lhs, vec_cond_rhs, vec_then_clause, vec_else_clause; tree vec_compare, vec_cond_expr; tree new_temp; loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); enum machine_mode vec_mode; tree def; enum vect_def_type dt; int nunits = TYPE_VECTOR_SUBPARTS (vectype); int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; gcc_assert (ncopies >= 1); if (ncopies > 1) return false; /* FORNOW */ if (!STMT_VINFO_RELEVANT_P (stmt_info)) return false; gcc_assert (STMT_VINFO_DEF_TYPE (stmt_info) == vect_loop_def); if (STMT_VINFO_LIVE_P (stmt_info)) { /* FORNOW: not yet supported. */ if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "value used after loop."); return false; } if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT) return false; op = GIMPLE_STMT_OPERAND (stmt, 1); if (TREE_CODE (op) != COND_EXPR) return false; cond_expr = TREE_OPERAND (op, 0); then_clause = TREE_OPERAND (op, 1); else_clause = TREE_OPERAND (op, 2); if (!vect_is_simple_cond (cond_expr, loop_vinfo)) return false; /* We do not handle two different vector types for the condition and the values. */ if (TREE_TYPE (TREE_OPERAND (cond_expr, 0)) != TREE_TYPE (vectype)) return false; if (TREE_CODE (then_clause) == SSA_NAME) { tree then_def_stmt = SSA_NAME_DEF_STMT (then_clause); if (!vect_is_simple_use (then_clause, loop_vinfo, &then_def_stmt, &def, &dt)) return false; } else if (TREE_CODE (then_clause) != INTEGER_CST && TREE_CODE (then_clause) != REAL_CST) return false; if (TREE_CODE (else_clause) == SSA_NAME) { tree else_def_stmt = SSA_NAME_DEF_STMT (else_clause); if (!vect_is_simple_use (else_clause, loop_vinfo, &else_def_stmt, &def, &dt)) return false; } else if (TREE_CODE (else_clause) != INTEGER_CST && TREE_CODE (else_clause) != REAL_CST) return false; vec_mode = TYPE_MODE (vectype); if (!vec_stmt) { STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type; return expand_vec_cond_expr_p (op, vec_mode); } /* Transform */ /* Handle def. */ scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0); vec_dest = vect_create_destination_var (scalar_dest, vectype); /* Handle cond expr. */ vec_cond_lhs = vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), stmt, NULL); vec_cond_rhs = vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), stmt, NULL); vec_then_clause = vect_get_vec_def_for_operand (then_clause, stmt, NULL); vec_else_clause = vect_get_vec_def_for_operand (else_clause, stmt, NULL); /* Arguments are ready. create the new vector stmt. */ vec_compare = build2 (TREE_CODE (cond_expr), vectype, vec_cond_lhs, vec_cond_rhs); vec_cond_expr = build3 (VEC_COND_EXPR, vectype, vec_compare, vec_then_clause, vec_else_clause); *vec_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, vec_dest, vec_cond_expr); new_temp = make_ssa_name (vec_dest, *vec_stmt); GIMPLE_STMT_OPERAND (*vec_stmt, 0) = new_temp; vect_finish_stmt_generation (stmt, *vec_stmt, bsi); return true; } /* Function vect_transform_stmt. Create a vectorized stmt to replace STMT, and insert it at BSI. */ bool vect_transform_stmt (tree stmt, block_stmt_iterator *bsi, bool *strided_store) { bool is_store = false; tree vec_stmt = NULL_TREE; stmt_vec_info stmt_info = vinfo_for_stmt (stmt); tree orig_stmt_in_pattern; bool done; if (STMT_VINFO_RELEVANT_P (stmt_info)) { switch (STMT_VINFO_TYPE (stmt_info)) { case type_demotion_vec_info_type: done = vectorizable_type_demotion (stmt, bsi, &vec_stmt); gcc_assert (done); break; case type_promotion_vec_info_type: done = vectorizable_type_promotion (stmt, bsi, &vec_stmt); gcc_assert (done); break; case op_vec_info_type: done = vectorizable_operation (stmt, bsi, &vec_stmt); gcc_assert (done); break; case assignment_vec_info_type: done = vectorizable_assignment (stmt, bsi, &vec_stmt); gcc_assert (done); break; case load_vec_info_type: done = vectorizable_load (stmt, bsi, &vec_stmt); gcc_assert (done); break; case store_vec_info_type: done = vectorizable_store (stmt, bsi, &vec_stmt); gcc_assert (done); if (DR_GROUP_FIRST_DR (stmt_info)) { /* In case of interleaving, the whole chain is vectorized when the last store in the chain is reached. Store stmts before the last one are skipped, and there vec_stmt_info shouldn't be freed meanwhile. */ *strided_store = true; if (STMT_VINFO_VEC_STMT (stmt_info)) is_store = true; } else is_store = true; break; case condition_vec_info_type: done = vectorizable_condition (stmt, bsi, &vec_stmt); gcc_assert (done); break; case call_vec_info_type: done = vectorizable_call (stmt, bsi, &vec_stmt); break; default: if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "stmt not supported."); gcc_unreachable (); } gcc_assert (vec_stmt || *strided_store); if (vec_stmt) { STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt; orig_stmt_in_pattern = STMT_VINFO_RELATED_STMT (stmt_info); if (orig_stmt_in_pattern) { stmt_vec_info stmt_vinfo = vinfo_for_stmt (orig_stmt_in_pattern); if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo)) { gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo) == stmt); /* STMT was inserted by the vectorizer to replace a computation idiom. ORIG_STMT_IN_PATTERN is a stmt in the original sequence that computed this idiom. We need to record a pointer to VEC_STMT in the stmt_info of ORIG_STMT_IN_PATTERN. See more details in the documentation of vect_pattern_recog. */ STMT_VINFO_VEC_STMT (stmt_vinfo) = vec_stmt; } } } } if (STMT_VINFO_LIVE_P (stmt_info)) { switch (STMT_VINFO_TYPE (stmt_info)) { case reduc_vec_info_type: done = vectorizable_reduction (stmt, bsi, &vec_stmt); gcc_assert (done); break; default: done = vectorizable_live_operation (stmt, bsi, &vec_stmt); gcc_assert (done); } } return is_store; } /* This function builds ni_name = number of iterations loop executes on the loop preheader. */ static tree vect_build_loop_niters (loop_vec_info loop_vinfo) { tree ni_name, stmt, var; edge pe; struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); tree ni = unshare_expr (LOOP_VINFO_NITERS (loop_vinfo)); var = create_tmp_var (TREE_TYPE (ni), "niters"); add_referenced_var (var); ni_name = force_gimple_operand (ni, &stmt, false, var); pe = loop_preheader_edge (loop); if (stmt) { basic_block new_bb = bsi_insert_on_edge_immediate (pe, stmt); gcc_assert (!new_bb); } return ni_name; } /* This function generates the following statements: ni_name = number of iterations loop executes ratio = ni_name / vf ratio_mult_vf_name = ratio * vf and places them at the loop preheader edge. */ static void vect_generate_tmps_on_preheader (loop_vec_info loop_vinfo, tree *ni_name_ptr, tree *ratio_mult_vf_name_ptr, tree *ratio_name_ptr) { edge pe; basic_block new_bb; tree stmt, ni_name; tree var; tree ratio_name; tree ratio_mult_vf_name; struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); tree ni = LOOP_VINFO_NITERS (loop_vinfo); int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); tree log_vf; pe = loop_preheader_edge (loop); /* Generate temporary variable that contains number of iterations loop executes. */ ni_name = vect_build_loop_niters (loop_vinfo); log_vf = build_int_cst (TREE_TYPE (ni), exact_log2 (vf)); /* Create: ratio = ni >> log2(vf) */ ratio_name = fold_build2 (RSHIFT_EXPR, TREE_TYPE (ni_name), ni_name, log_vf); if (!is_gimple_val (ratio_name)) { var = create_tmp_var (TREE_TYPE (ni), "bnd"); add_referenced_var (var); ratio_name = force_gimple_operand (ratio_name, &stmt, true, var); pe = loop_preheader_edge (loop); new_bb = bsi_insert_on_edge_immediate (pe, stmt); gcc_assert (!new_bb); } /* Create: ratio_mult_vf = ratio << log2 (vf). */ ratio_mult_vf_name = fold_build2 (LSHIFT_EXPR, TREE_TYPE (ratio_name), ratio_name, log_vf); if (!is_gimple_val (ratio_mult_vf_name)) { var = create_tmp_var (TREE_TYPE (ni), "ratio_mult_vf"); add_referenced_var (var); ratio_mult_vf_name = force_gimple_operand (ratio_mult_vf_name, &stmt, true, var); pe = loop_preheader_edge (loop); new_bb = bsi_insert_on_edge_immediate (pe, stmt); gcc_assert (!new_bb); } *ni_name_ptr = ni_name; *ratio_mult_vf_name_ptr = ratio_mult_vf_name; *ratio_name_ptr = ratio_name; return; } /* Function update_vuses_to_preheader. Input: STMT - a statement with potential VUSEs. LOOP - the loop whose preheader will contain STMT. It's possible to vectorize a loop even though an SSA_NAME from a VUSE appears to be defined in a VDEF in another statement in a loop. One such case is when the VUSE is at the dereference of a __restricted__ pointer in a load and the VDEF is at the dereference of a different __restricted__ pointer in a store. Vectorization may result in copy_virtual_uses being called to copy the problematic VUSE to a new statement that is being inserted in the loop preheader. This procedure is called to change the SSA_NAME in the new statement's VUSE from the SSA_NAME updated in the loop to the related SSA_NAME available on the path entering the loop. When this function is called, we have the following situation: # vuse S1: vload do { # name1 = phi < name0 , name2> # vuse S2: vload # name2 = vdef S3: vstore }while... Stmt S1 was created in the loop preheader block as part of misaligned-load handling. This function fixes the name of the vuse of S1 from 'name1' to 'name0'. */ static void update_vuses_to_preheader (tree stmt, struct loop *loop) { basic_block header_bb = loop->header; edge preheader_e = loop_preheader_edge (loop); ssa_op_iter iter; use_operand_p use_p; FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_VUSE) { tree ssa_name = USE_FROM_PTR (use_p); tree def_stmt = SSA_NAME_DEF_STMT (ssa_name); tree name_var = SSA_NAME_VAR (ssa_name); basic_block bb = bb_for_stmt (def_stmt); /* For a use before any definitions, def_stmt is a NOP_EXPR. */ if (!IS_EMPTY_STMT (def_stmt) && flow_bb_inside_loop_p (loop, bb)) { /* If the block containing the statement defining the SSA_NAME is in the loop then it's necessary to find the definition outside the loop using the PHI nodes of the header. */ tree phi; bool updated = false; for (phi = phi_nodes (header_bb); phi; phi = TREE_CHAIN (phi)) { if (SSA_NAME_VAR (PHI_RESULT (phi)) == name_var) { SET_USE (use_p, PHI_ARG_DEF (phi, preheader_e->dest_idx)); updated = true; break; } } gcc_assert (updated); } } } /* Function vect_update_ivs_after_vectorizer. "Advance" the induction variables of LOOP to the value they should take after the execution of LOOP. This is currently necessary because the vectorizer does not handle induction variables that are used after the loop. Such a situation occurs when the last iterations of LOOP are peeled, because: 1. We introduced new uses after LOOP for IVs that were not originally used after LOOP: the IVs of LOOP are now used by an epilog loop. 2. LOOP is going to be vectorized; this means that it will iterate N/VF times, whereas the loop IVs should be bumped N times. Input: - LOOP - a loop that is going to be vectorized. The last few iterations of LOOP were peeled. - NITERS - the number of iterations that LOOP executes (before it is vectorized). i.e, the number of times the ivs should be bumped. - UPDATE_E - a successor edge of LOOP->exit that is on the (only) path coming out from LOOP on which there are uses of the LOOP ivs (this is the path from LOOP->exit to epilog_loop->preheader). The new definitions of the ivs are placed in LOOP->exit. The phi args associated with the edge UPDATE_E in the bb UPDATE_E->dest are updated accordingly. Assumption 1: Like the rest of the vectorizer, this function assumes a single loop exit that has a single predecessor. Assumption 2: The phi nodes in the LOOP header and in update_bb are organized in the same order. Assumption 3: The access function of the ivs is simple enough (see vect_can_advance_ivs_p). This assumption will be relaxed in the future. Assumption 4: Exactly one of the successors of LOOP exit-bb is on a path coming out of LOOP on which the ivs of LOOP are used (this is the path that leads to the epilog loop; other paths skip the epilog loop). This path starts with the edge UPDATE_E, and its destination (denoted update_bb) needs to have its phis updated. */ static void vect_update_ivs_after_vectorizer (loop_vec_info loop_vinfo, tree niters, edge update_e) { struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); basic_block exit_bb = single_exit (loop)->dest; tree phi, phi1; basic_block update_bb = update_e->dest; /* gcc_assert (vect_can_advance_ivs_p (loop_vinfo)); */ /* Make sure there exists a single-predecessor exit bb: */ gcc_assert (single_pred_p (exit_bb)); for (phi = phi_nodes (loop->header), phi1 = phi_nodes (update_bb); phi && phi1; phi = PHI_CHAIN (phi), phi1 = PHI_CHAIN (phi1)) { tree access_fn = NULL; tree evolution_part; tree init_expr; tree step_expr; tree var, stmt, ni, ni_name; block_stmt_iterator last_bsi; if (vect_print_dump_info (REPORT_DETAILS)) { fprintf (vect_dump, "vect_update_ivs_after_vectorizer: phi: "); print_generic_expr (vect_dump, phi, TDF_SLIM); } /* Skip virtual phi's. */ if (!is_gimple_reg (SSA_NAME_VAR (PHI_RESULT (phi)))) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "virtual phi. skip."); continue; } /* Skip reduction phis. */ if (STMT_VINFO_DEF_TYPE (vinfo_for_stmt (phi)) == vect_reduction_def) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "reduc phi. skip."); continue; } access_fn = analyze_scalar_evolution (loop, PHI_RESULT (phi)); gcc_assert (access_fn); evolution_part = unshare_expr (evolution_part_in_loop_num (access_fn, loop->num)); gcc_assert (evolution_part != NULL_TREE); /* FORNOW: We do not support IVs whose evolution function is a polynomial of degree >= 2 or exponential. */ gcc_assert (!tree_is_chrec (evolution_part)); step_expr = evolution_part; init_expr = unshare_expr (initial_condition_in_loop_num (access_fn, loop->num)); ni = fold_build2 (PLUS_EXPR, TREE_TYPE (init_expr), fold_build2 (MULT_EXPR, TREE_TYPE (init_expr), fold_convert (TREE_TYPE (init_expr), niters), step_expr), init_expr); var = create_tmp_var (TREE_TYPE (init_expr), "tmp"); add_referenced_var (var); ni_name = force_gimple_operand (ni, &stmt, false, var); /* Insert stmt into exit_bb. */ last_bsi = bsi_last (exit_bb); if (stmt) bsi_insert_before (&last_bsi, stmt, BSI_SAME_STMT); /* Fix phi expressions in the successor bb. */ SET_PHI_ARG_DEF (phi1, update_e->dest_idx, ni_name); } } /* Function vect_do_peeling_for_loop_bound Peel the last iterations of the loop represented by LOOP_VINFO. The peeled iterations form a new epilog loop. Given that the loop now iterates NITERS times, the new epilog loop iterates NITERS % VECTORIZATION_FACTOR times. The original loop will later be made to iterate NITERS / VECTORIZATION_FACTOR times (this value is placed into RATIO). */ static void vect_do_peeling_for_loop_bound (loop_vec_info loop_vinfo, tree *ratio) { tree ni_name, ratio_mult_vf_name; struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); struct loop *new_loop; edge update_e; basic_block preheader; int loop_num; unsigned int th; if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "=== vect_do_peeling_for_loop_bound ==="); initialize_original_copy_tables (); /* Generate the following variables on the preheader of original loop: ni_name = number of iteration the original loop executes ratio = ni_name / vf ratio_mult_vf_name = ratio * vf */ vect_generate_tmps_on_preheader (loop_vinfo, &ni_name, &ratio_mult_vf_name, ratio); loop_num = loop->num; /* Threshold for vectorized loop. */ th = (PARAM_VALUE (PARAM_MIN_VECT_LOOP_BOUND)) * LOOP_VINFO_VECT_FACTOR (loop_vinfo); new_loop = slpeel_tree_peel_loop_to_edge (loop, single_exit (loop), ratio_mult_vf_name, ni_name, false, th); gcc_assert (new_loop); gcc_assert (loop_num == loop->num); #ifdef ENABLE_CHECKING slpeel_verify_cfg_after_peeling (loop, new_loop); #endif /* A guard that controls whether the new_loop is to be executed or skipped is placed in LOOP->exit. LOOP->exit therefore has two successors - one is the preheader of NEW_LOOP, where the IVs from LOOP are used. The other is a bb after NEW_LOOP, where these IVs are not used. Find the edge that is on the path where the LOOP IVs are used and need to be updated. */ preheader = loop_preheader_edge (new_loop)->src; if (EDGE_PRED (preheader, 0)->src == single_exit (loop)->dest) update_e = EDGE_PRED (preheader, 0); else update_e = EDGE_PRED (preheader, 1); /* Update IVs of original loop as if they were advanced by ratio_mult_vf_name steps. */ vect_update_ivs_after_vectorizer (loop_vinfo, ratio_mult_vf_name, update_e); /* After peeling we have to reset scalar evolution analyzer. */ scev_reset (); free_original_copy_tables (); } /* Function vect_gen_niters_for_prolog_loop Set the number of iterations for the loop represented by LOOP_VINFO to the minimum between LOOP_NITERS (the original iteration count of the loop) and the misalignment of DR - the data reference recorded in LOOP_VINFO_UNALIGNED_DR (LOOP_VINFO). As a result, after the execution of this loop, the data reference DR will refer to an aligned location. The following computation is generated: If the misalignment of DR is known at compile time: addr_mis = int mis = DR_MISALIGNMENT (dr); Else, compute address misalignment in bytes: addr_mis = addr & (vectype_size - 1) prolog_niters = min ( LOOP_NITERS , (VF - addr_mis/elem_size)&(VF-1) ) (elem_size = element type size; an element is the scalar element whose type is the inner type of the vectype) For interleaving, prolog_niters = min ( LOOP_NITERS , (VF/group_size - addr_mis/elem_size)&(VF/group_size-1) ) where group_size is the size of the interleaved group. */ static tree vect_gen_niters_for_prolog_loop (loop_vec_info loop_vinfo, tree loop_niters) { struct data_reference *dr = LOOP_VINFO_UNALIGNED_DR (loop_vinfo); int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); tree var, stmt; tree iters, iters_name; edge pe; basic_block new_bb; tree dr_stmt = DR_STMT (dr); stmt_vec_info stmt_info = vinfo_for_stmt (dr_stmt); tree vectype = STMT_VINFO_VECTYPE (stmt_info); int vectype_align = TYPE_ALIGN (vectype) / BITS_PER_UNIT; tree niters_type = TREE_TYPE (loop_niters); int group_size = 1; int element_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dr)))); if (DR_GROUP_FIRST_DR (stmt_info)) { /* For interleaved access element size must be multiplied by the size of the interleaved group. */ group_size = DR_GROUP_SIZE (vinfo_for_stmt ( DR_GROUP_FIRST_DR (stmt_info))); element_size *= group_size; } pe = loop_preheader_edge (loop); if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo) > 0) { int byte_misalign = LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo); int elem_misalign = byte_misalign / element_size; if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "known alignment = %d.", byte_misalign); iters = build_int_cst (niters_type, (vf - elem_misalign)&(vf/group_size-1)); } else { tree new_stmts = NULL_TREE; tree start_addr = vect_create_addr_base_for_vector_ref (dr_stmt, &new_stmts, NULL_TREE); tree ptr_type = TREE_TYPE (start_addr); tree size = TYPE_SIZE (ptr_type); tree type = lang_hooks.types.type_for_size (tree_low_cst (size, 1), 1); tree vectype_size_minus_1 = build_int_cst (type, vectype_align - 1); tree elem_size_log = build_int_cst (type, exact_log2 (vectype_align/vf)); tree vf_minus_1 = build_int_cst (type, vf - 1); tree vf_tree = build_int_cst (type, vf); tree byte_misalign; tree elem_misalign; new_bb = bsi_insert_on_edge_immediate (pe, new_stmts); gcc_assert (!new_bb); /* Create: byte_misalign = addr & (vectype_size - 1) */ byte_misalign = fold_build2 (BIT_AND_EXPR, type, start_addr, vectype_size_minus_1); /* Create: elem_misalign = byte_misalign / element_size */ elem_misalign = fold_build2 (RSHIFT_EXPR, type, byte_misalign, elem_size_log); /* Create: (niters_type) (VF - elem_misalign)&(VF - 1) */ iters = fold_build2 (MINUS_EXPR, type, vf_tree, elem_misalign); iters = fold_build2 (BIT_AND_EXPR, type, iters, vf_minus_1); iters = fold_convert (niters_type, iters); } /* Create: prolog_loop_niters = min (iters, loop_niters) */ /* If the loop bound is known at compile time we already verified that it is greater than vf; since the misalignment ('iters') is at most vf, there's no need to generate the MIN_EXPR in this case. */ if (TREE_CODE (loop_niters) != INTEGER_CST) iters = fold_build2 (MIN_EXPR, niters_type, iters, loop_niters); if (vect_print_dump_info (REPORT_DETAILS)) { fprintf (vect_dump, "niters for prolog loop: "); print_generic_expr (vect_dump, iters, TDF_SLIM); } var = create_tmp_var (niters_type, "prolog_loop_niters"); add_referenced_var (var); iters_name = force_gimple_operand (iters, &stmt, false, var); /* Insert stmt on loop preheader edge. */ if (stmt) { basic_block new_bb = bsi_insert_on_edge_immediate (pe, stmt); gcc_assert (!new_bb); } return iters_name; } /* Function vect_update_init_of_dr NITERS iterations were peeled from LOOP. DR represents a data reference in LOOP. This function updates the information recorded in DR to account for the fact that the first NITERS iterations had already been executed. Specifically, it updates the OFFSET field of DR. */ static void vect_update_init_of_dr (struct data_reference *dr, tree niters) { tree offset = DR_OFFSET (dr); niters = fold_build2 (MULT_EXPR, TREE_TYPE (niters), niters, DR_STEP (dr)); offset = fold_build2 (PLUS_EXPR, TREE_TYPE (offset), offset, niters); DR_OFFSET (dr) = offset; } /* Function vect_update_inits_of_drs NITERS iterations were peeled from the loop represented by LOOP_VINFO. This function updates the information recorded for the data references in the loop to account for the fact that the first NITERS iterations had already been executed. Specifically, it updates the initial_condition of the access_function of all the data_references in the loop. */ static void vect_update_inits_of_drs (loop_vec_info loop_vinfo, tree niters) { unsigned int i; VEC (data_reference_p, heap) *datarefs = LOOP_VINFO_DATAREFS (loop_vinfo); struct data_reference *dr; if (vect_dump && (dump_flags & TDF_DETAILS)) fprintf (vect_dump, "=== vect_update_inits_of_dr ==="); for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++) vect_update_init_of_dr (dr, niters); } /* Function vect_do_peeling_for_alignment Peel the first 'niters' iterations of the loop represented by LOOP_VINFO. 'niters' is set to the misalignment of one of the data references in the loop, thereby forcing it to refer to an aligned location at the beginning of the execution of this loop. The data reference for which we are peeling is recorded in LOOP_VINFO_UNALIGNED_DR. */ static void vect_do_peeling_for_alignment (loop_vec_info loop_vinfo) { struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); tree niters_of_prolog_loop, ni_name; tree n_iters; struct loop *new_loop; if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "=== vect_do_peeling_for_alignment ==="); initialize_original_copy_tables (); ni_name = vect_build_loop_niters (loop_vinfo); niters_of_prolog_loop = vect_gen_niters_for_prolog_loop (loop_vinfo, ni_name); /* Peel the prolog loop and iterate it niters_of_prolog_loop. */ new_loop = slpeel_tree_peel_loop_to_edge (loop, loop_preheader_edge (loop), niters_of_prolog_loop, ni_name, true, 0); gcc_assert (new_loop); #ifdef ENABLE_CHECKING slpeel_verify_cfg_after_peeling (new_loop, loop); #endif /* Update number of times loop executes. */ n_iters = LOOP_VINFO_NITERS (loop_vinfo); LOOP_VINFO_NITERS (loop_vinfo) = fold_build2 (MINUS_EXPR, TREE_TYPE (n_iters), n_iters, niters_of_prolog_loop); /* Update the init conditions of the access functions of all data refs. */ vect_update_inits_of_drs (loop_vinfo, niters_of_prolog_loop); /* After peeling we have to reset scalar evolution analyzer. */ scev_reset (); free_original_copy_tables (); } /* Function vect_create_cond_for_align_checks. Create a conditional expression that represents the alignment checks for all of data references (array element references) whose alignment must be checked at runtime. Input: LOOP_VINFO - two fields of the loop information are used. LOOP_VINFO_PTR_MASK is the mask used to check the alignment. LOOP_VINFO_MAY_MISALIGN_STMTS contains the refs to be checked. Output: COND_EXPR_STMT_LIST - statements needed to construct the conditional expression. The returned value is the conditional expression to be used in the if statement that controls which version of the loop gets executed at runtime. The algorithm makes two assumptions: 1) The number of bytes "n" in a vector is a power of 2. 2) An address "a" is aligned if a%n is zero and that this test can be done as a&(n-1) == 0. For example, for 16 byte vectors the test is a&0xf == 0. */ static tree vect_create_cond_for_align_checks (loop_vec_info loop_vinfo, tree *cond_expr_stmt_list) { VEC(tree,heap) *may_misalign_stmts = LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo); tree ref_stmt; int mask = LOOP_VINFO_PTR_MASK (loop_vinfo); tree mask_cst; unsigned int i; tree psize; tree int_ptrsize_type; char tmp_name[20]; tree or_tmp_name = NULL_TREE; tree and_tmp, and_tmp_name, and_stmt; tree ptrsize_zero; /* Check that mask is one less than a power of 2, i.e., mask is all zeros followed by all ones. */ gcc_assert ((mask != 0) && ((mask & (mask+1)) == 0)); /* CHECKME: what is the best integer or unsigned type to use to hold a cast from a pointer value? */ psize = TYPE_SIZE (ptr_type_node); int_ptrsize_type = lang_hooks.types.type_for_size (tree_low_cst (psize, 1), 0); /* Create expression (mask & (dr_1 || ... || dr_n)) where dr_i is the address of the first vector of the i'th data reference. */ for (i = 0; VEC_iterate (tree, may_misalign_stmts, i, ref_stmt); i++) { tree new_stmt_list = NULL_TREE; tree addr_base; tree addr_tmp, addr_tmp_name, addr_stmt; tree or_tmp, new_or_tmp_name, or_stmt; /* create: addr_tmp = (int)(address_of_first_vector) */ addr_base = vect_create_addr_base_for_vector_ref (ref_stmt, &new_stmt_list, NULL_TREE); if (new_stmt_list != NULL_TREE) append_to_statement_list_force (new_stmt_list, cond_expr_stmt_list); sprintf (tmp_name, "%s%d", "addr2int", i); addr_tmp = create_tmp_var (int_ptrsize_type, tmp_name); add_referenced_var (addr_tmp); addr_tmp_name = make_ssa_name (addr_tmp, NULL_TREE); addr_stmt = fold_convert (int_ptrsize_type, addr_base); addr_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, addr_tmp_name, addr_stmt); SSA_NAME_DEF_STMT (addr_tmp_name) = addr_stmt; append_to_statement_list_force (addr_stmt, cond_expr_stmt_list); /* The addresses are OR together. */ if (or_tmp_name != NULL_TREE) { /* create: or_tmp = or_tmp | addr_tmp */ sprintf (tmp_name, "%s%d", "orptrs", i); or_tmp = create_tmp_var (int_ptrsize_type, tmp_name); add_referenced_var (or_tmp); new_or_tmp_name = make_ssa_name (or_tmp, NULL_TREE); or_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, new_or_tmp_name, build2 (BIT_IOR_EXPR, int_ptrsize_type, or_tmp_name, addr_tmp_name)); SSA_NAME_DEF_STMT (new_or_tmp_name) = or_stmt; append_to_statement_list_force (or_stmt, cond_expr_stmt_list); or_tmp_name = new_or_tmp_name; } else or_tmp_name = addr_tmp_name; } /* end for i */ mask_cst = build_int_cst (int_ptrsize_type, mask); /* create: and_tmp = or_tmp & mask */ and_tmp = create_tmp_var (int_ptrsize_type, "andmask" ); add_referenced_var (and_tmp); and_tmp_name = make_ssa_name (and_tmp, NULL_TREE); and_stmt = build2 (GIMPLE_MODIFY_STMT, void_type_node, and_tmp_name, build2 (BIT_AND_EXPR, int_ptrsize_type, or_tmp_name, mask_cst)); SSA_NAME_DEF_STMT (and_tmp_name) = and_stmt; append_to_statement_list_force (and_stmt, cond_expr_stmt_list); /* Make and_tmp the left operand of the conditional test against zero. if and_tmp has a nonzero bit then some address is unaligned. */ ptrsize_zero = build_int_cst (int_ptrsize_type, 0); return build2 (EQ_EXPR, boolean_type_node, and_tmp_name, ptrsize_zero); } /* Function vect_transform_loop. The analysis phase has determined that the loop is vectorizable. Vectorize the loop - created vectorized stmts to replace the scalar stmts in the loop, and update the loop exit condition. */ void vect_transform_loop (loop_vec_info loop_vinfo) { struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); int nbbs = loop->num_nodes; block_stmt_iterator si; int i; tree ratio = NULL; int vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo); bool strided_store; if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "=== vec_transform_loop ==="); /* If the loop has data references that may or may not be aligned then two versions of the loop need to be generated, one which is vectorized and one which isn't. A test is then generated to control which of the loops is executed. The test checks for the alignment of all of the data references that may or may not be aligned. */ if (VEC_length (tree, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo))) { struct loop *nloop; tree cond_expr; tree cond_expr_stmt_list = NULL_TREE; basic_block condition_bb; block_stmt_iterator cond_exp_bsi; basic_block merge_bb; basic_block new_exit_bb; edge new_exit_e, e; tree orig_phi, new_phi, arg; unsigned prob = 4 * REG_BR_PROB_BASE / 5; cond_expr = vect_create_cond_for_align_checks (loop_vinfo, &cond_expr_stmt_list); initialize_original_copy_tables (); nloop = loop_version (loop, cond_expr, &condition_bb, prob, prob, REG_BR_PROB_BASE - prob, true); free_original_copy_tables(); /** Loop versioning violates an assumption we try to maintain during vectorization - that the loop exit block has a single predecessor. After versioning, the exit block of both loop versions is the same basic block (i.e. it has two predecessors). Just in order to simplify following transformations in the vectorizer, we fix this situation here by adding a new (empty) block on the exit-edge of the loop, with the proper loop-exit phis to maintain loop-closed-form. **/ merge_bb = single_exit (loop)->dest; gcc_assert (EDGE_COUNT (merge_bb->preds) == 2); new_exit_bb = split_edge (single_exit (loop)); new_exit_e = single_exit (loop); e = EDGE_SUCC (new_exit_bb, 0); for (orig_phi = phi_nodes (merge_bb); orig_phi; orig_phi = PHI_CHAIN (orig_phi)) { new_phi = create_phi_node (SSA_NAME_VAR (PHI_RESULT (orig_phi)), new_exit_bb); arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, e); add_phi_arg (new_phi, arg, new_exit_e); SET_PHI_ARG_DEF (orig_phi, e->dest_idx, PHI_RESULT (new_phi)); } /** end loop-exit-fixes after versioning **/ update_ssa (TODO_update_ssa); cond_exp_bsi = bsi_last (condition_bb); bsi_insert_before (&cond_exp_bsi, cond_expr_stmt_list, BSI_SAME_STMT); } /* CHECKME: we wouldn't need this if we called update_ssa once for all loops. */ bitmap_zero (vect_memsyms_to_rename); /* Peel the loop if there are data refs with unknown alignment. Only one data ref with unknown store is allowed. */ if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo)) vect_do_peeling_for_alignment (loop_vinfo); /* If the loop has a symbolic number of iterations 'n' (i.e. it's not a compile time constant), or it is a constant that doesn't divide by the vectorization factor, then an epilog loop needs to be created. We therefore duplicate the loop: the original loop will be vectorized, and will compute the first (n/VF) iterations. The second copy of the loop will remain scalar and will compute the remaining (n%VF) iterations. (VF is the vectorization factor). */ if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) || (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) && LOOP_VINFO_INT_NITERS (loop_vinfo) % vectorization_factor != 0)) vect_do_peeling_for_loop_bound (loop_vinfo, &ratio); else ratio = build_int_cst (TREE_TYPE (LOOP_VINFO_NITERS (loop_vinfo)), LOOP_VINFO_INT_NITERS (loop_vinfo) / vectorization_factor); /* 1) Make sure the loop header has exactly two entries 2) Make sure we have a preheader basic block. */ gcc_assert (EDGE_COUNT (loop->header->preds) == 2); split_edge (loop_preheader_edge (loop)); /* FORNOW: the vectorizer supports only loops which body consist of one basic block (header + empty latch). When the vectorizer will support more involved loop forms, the order by which the BBs are traversed need to be reconsidered. */ for (i = 0; i < nbbs; i++) { basic_block bb = bbs[i]; for (si = bsi_start (bb); !bsi_end_p (si);) { tree stmt = bsi_stmt (si); stmt_vec_info stmt_info; bool is_store; if (vect_print_dump_info (REPORT_DETAILS)) { fprintf (vect_dump, "------>vectorizing statement: "); print_generic_expr (vect_dump, stmt, TDF_SLIM); } stmt_info = vinfo_for_stmt (stmt); gcc_assert (stmt_info); if (!STMT_VINFO_RELEVANT_P (stmt_info) && !STMT_VINFO_LIVE_P (stmt_info)) { bsi_next (&si); continue; } if ((TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info)) != (unsigned HOST_WIDE_INT) vectorization_factor) && vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "multiple-types."); /* -------- vectorize statement ------------ */ if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "transform statement."); strided_store = false; is_store = vect_transform_stmt (stmt, &si, &strided_store); if (is_store) { stmt_ann_t ann; if (DR_GROUP_FIRST_DR (stmt_info)) { /* Interleaving. If IS_STORE is TRUE, the vectorization of the interleaving chain was completed - free all the stores in the chain. */ tree next = DR_GROUP_FIRST_DR (stmt_info); tree tmp; stmt_vec_info next_stmt_info; while (next) { next_stmt_info = vinfo_for_stmt (next); /* Free the attached stmt_vec_info and remove the stmt. */ ann = stmt_ann (next); tmp = DR_GROUP_NEXT_DR (next_stmt_info); free (next_stmt_info); set_stmt_info (ann, NULL); next = tmp; } bsi_remove (&si, true); continue; } else { /* Free the attached stmt_vec_info and remove the stmt. */ ann = stmt_ann (stmt); free (stmt_info); set_stmt_info (ann, NULL); bsi_remove (&si, true); continue; } } else { if (strided_store) { /* This is case of skipped interleaved store. We don't free its stmt_vec_info. */ bsi_remove (&si, true); continue; } } bsi_next (&si); } /* stmts in BB */ } /* BBs in loop */ slpeel_make_loop_iterate_ntimes (loop, ratio); mark_set_for_renaming (vect_memsyms_to_rename); /* The memory tags and pointers in vectorized statements need to have their SSA forms updated. FIXME, why can't this be delayed until all the loops have been transformed? */ update_ssa (TODO_update_ssa); if (vect_print_dump_info (REPORT_VECTORIZED_LOOPS)) fprintf (vect_dump, "LOOP VECTORIZED."); }