/* Transformation Utilities for Loop Vectorization.
- Copyright (C) 2003,2004,2005,2006 Free Software Foundation, Inc.
+ Copyright (C) 2003, 2004, 2005, 2006, 2007 Free Software Foundation, Inc.
Contributed by Dorit Naishlos <dorit@il.ibm.com>
This file is part of GCC.
#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 "real.h"
/* Utility functions for the code transformation. */
-static bool vect_transform_stmt (tree, block_stmt_iterator *);
+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, 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);
+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 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 void vect_do_peeling_for_alignment (loop_vec_info, struct loops *);
-static void vect_do_peeling_for_loop_bound
- (loop_vec_info, tree *, struct loops *);
static int vect_min_worthwhile_factor (enum tree_code);
+static int
+cost_for_stmt (tree stmt)
+{
+ stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
+
+ switch (STMT_VINFO_TYPE (stmt_info))
+ {
+ case load_vec_info_type:
+ return TARG_SCALAR_LOAD_COST;
+ case store_vec_info_type:
+ return TARG_SCALAR_STORE_COST;
+ case op_vec_info_type:
+ case condition_vec_info_type:
+ case assignment_vec_info_type:
+ case reduc_vec_info_type:
+ case induc_vec_info_type:
+ case type_promotion_vec_info_type:
+ case type_demotion_vec_info_type:
+ case type_conversion_vec_info_type:
+ case call_vec_info_type:
+ return TARG_SCALAR_STMT_COST;
+ case undef_vec_info_type:
+ default:
+ gcc_unreachable ();
+ }
+}
+
+
+/* Function vect_estimate_min_profitable_iters
+
+ Return the number of iterations required for the vector version of the
+ loop to be profitable relative to the cost of the scalar version of the
+ loop.
+
+ TODO: Take profile info into account before making vectorization
+ decisions, if available. */
+
+int
+vect_estimate_min_profitable_iters (loop_vec_info loop_vinfo)
+{
+ int i;
+ int min_profitable_iters;
+ int peel_iters_prologue;
+ int peel_iters_epilogue;
+ int vec_inside_cost = 0;
+ int vec_outside_cost = 0;
+ int scalar_single_iter_cost = 0;
+ int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
+ struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
+ basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo);
+ int nbbs = loop->num_nodes;
+ int byte_misalign;
+
+ /* Cost model disabled. */
+ if (!flag_vect_cost_model)
+ {
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "cost model disabled.");
+ return 0;
+ }
+
+ /* Requires loop versioning tests to handle misalignment.
+ FIXME: Make cost depend on number of stmts in may_misalign list. */
+
+ if (VEC_length (tree, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo)))
+ {
+ vec_outside_cost += TARG_COND_BRANCH_COST;
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "cost model: Adding cost of checks for loop "
+ "versioning.\n");
+ }
+
+ /* Count statements in scalar loop. Using this as scalar cost for a single
+ iteration for now.
+
+ TODO: Add outer loop support.
+
+ TODO: Consider assigning different costs to different scalar
+ statements. */
+
+ for (i = 0; i < nbbs; i++)
+ {
+ block_stmt_iterator si;
+ basic_block bb = bbs[i];
+
+ for (si = bsi_start (bb); !bsi_end_p (si); bsi_next (&si))
+ {
+ tree stmt = bsi_stmt (si);
+ stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
+ if (!STMT_VINFO_RELEVANT_P (stmt_info)
+ && !STMT_VINFO_LIVE_P (stmt_info))
+ continue;
+ scalar_single_iter_cost += cost_for_stmt (stmt);
+ vec_inside_cost += STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info);
+ vec_outside_cost += STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info);
+ }
+ }
+
+ /* Add additional cost for the peeled instructions in prologue and epilogue
+ loop.
+
+ FORNOW: If we dont know the value of peel_iters for prologue or epilogue
+ at compile-time - we assume it's (vf-1)/2 (the worst would be vf-1).
+
+ TODO: Build an expression that represents peel_iters for prologue and
+ epilogue to be used in a run-time test. */
+
+ byte_misalign = LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo);
+
+ if (byte_misalign < 0)
+ {
+ peel_iters_prologue = (vf - 1)/2;
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "cost model: "
+ "prologue peel iters set to (vf-1)/2.");
+
+ /* If peeling for alignment is unknown, loop bound of main loop becomes
+ unknown. */
+ peel_iters_epilogue = (vf - 1)/2;
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "cost model: "
+ "epilogue peel iters set to (vf-1)/2 because "
+ "peeling for alignment is unknown .");
+ }
+ else
+ {
+ if (byte_misalign)
+ {
+ struct data_reference *dr = LOOP_VINFO_UNALIGNED_DR (loop_vinfo);
+ int element_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dr))));
+ tree vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (DR_STMT (dr)));
+ int nelements = TYPE_VECTOR_SUBPARTS (vectype);
+
+ peel_iters_prologue = nelements - (byte_misalign / element_size);
+ }
+ else
+ peel_iters_prologue = 0;
+
+ if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo))
+ {
+ peel_iters_epilogue = (vf - 1)/2;
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "cost model: "
+ "epilogue peel iters set to (vf-1)/2 because "
+ "loop iterations are unknown .");
+ }
+ else
+ {
+ int niters = LOOP_VINFO_INT_NITERS (loop_vinfo);
+ peel_iters_prologue = niters < peel_iters_prologue ?
+ niters : peel_iters_prologue;
+ peel_iters_epilogue = (niters - peel_iters_prologue) % vf;
+ }
+ }
+
+ /* Requires a prologue loop when peeling to handle misalignment. Add cost of
+ two guards, one for the peeled loop and one for the vector loop. */
+
+ if (peel_iters_prologue)
+ {
+ vec_outside_cost += 2 * TARG_COND_BRANCH_COST;
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "cost model: Adding cost of checks for "
+ "prologue.\n");
+ }
+
+ /* Requires an epilogue loop to finish up remaining iterations after vector
+ loop. Add cost of two guards, one for the peeled loop and one for the
+ vector loop. */
+
+ if (peel_iters_epilogue
+ || !LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
+ || LOOP_VINFO_INT_NITERS (loop_vinfo) % vf)
+ {
+ vec_outside_cost += 2 * TARG_COND_BRANCH_COST;
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "cost model : Adding cost of checks for "
+ "epilogue.\n");
+ }
+
+ vec_outside_cost += (peel_iters_prologue * scalar_single_iter_cost)
+ + (peel_iters_epilogue * scalar_single_iter_cost);
+
+ /* Allow targets add additional (outside-of-loop) costs. FORNOW, the only
+ information we provide for the target is whether testing against the
+ threshold involves a runtime test. */
+ if (targetm.vectorize.builtin_vectorization_cost)
+ {
+ bool runtime_test = false;
+
+ /* If the number of iterations is unknown, or the
+ peeling-for-misalignment amount is unknown, we eill have to generate
+ a runtime test to test the loop count agains the threshold. */
+ if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
+ || (byte_misalign < 0))
+ runtime_test = true;
+ vec_outside_cost +=
+ targetm.vectorize.builtin_vectorization_cost (runtime_test);
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "cost model : Adding target out-of-loop cost = %d",
+ targetm.vectorize.builtin_vectorization_cost (runtime_test));
+ }
+
+ /* Calculate number of iterations required to make the vector version
+ profitable, relative to the loop bodies only. The following condition
+ must hold true: ((SIC*VF)-VIC)*niters > VOC*VF, where
+ SIC = scalar iteration cost, VIC = vector iteration cost,
+ VOC = vector outside cost and VF = vectorization factor. */
+
+ if ((scalar_single_iter_cost * vf) > vec_inside_cost)
+ {
+ if (vec_outside_cost == 0)
+ min_profitable_iters = 1;
+ else
+ {
+ min_profitable_iters = (vec_outside_cost * vf)
+ / ((scalar_single_iter_cost * vf)
+ - vec_inside_cost);
+
+ if ((scalar_single_iter_cost * vf * min_profitable_iters)
+ <= ((vec_inside_cost * min_profitable_iters)
+ + (vec_outside_cost * vf)))
+ min_profitable_iters++;
+ }
+ }
+ /* vector version will never be profitable. */
+ else
+ {
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "cost model: vector iteration cost = %d "
+ "is divisible by scalar iteration cost = %d by a factor "
+ "greater than or equal to the vectorization factor = %d .",
+ vec_inside_cost, scalar_single_iter_cost, vf);
+ return -1;
+ }
+
+ if (vect_print_dump_info (REPORT_DETAILS))
+ {
+ fprintf (vect_dump, "Cost model analysis: \n");
+ fprintf (vect_dump, " Vector inside of loop cost: %d\n",
+ vec_inside_cost);
+ fprintf (vect_dump, " Vector outside of loop cost: %d\n",
+ vec_outside_cost);
+ fprintf (vect_dump, " Scalar cost: %d\n", scalar_single_iter_cost);
+ fprintf (vect_dump, " prologue iterations: %d\n",
+ peel_iters_prologue);
+ fprintf (vect_dump, " epilogue iterations: %d\n",
+ peel_iters_epilogue);
+ fprintf (vect_dump, " Calculated minimum iters for profitability: %d\n",
+ min_profitable_iters);
+ fprintf (vect_dump, " Actual minimum iters for profitability: %d\n",
+ min_profitable_iters < vf ? vf : min_profitable_iters);
+ }
+
+ min_profitable_iters =
+ min_profitable_iters < vf ? vf : min_profitable_iters;
+
+ /* Because the condition we create is:
+ if (niters <= min_profitable_iters)
+ then skip the vectorized loop. */
+ min_profitable_iters--;
+ return min_profitable_iters;
+}
+
+
+/* TODO: Close dependency between vect_model_*_cost and vectorizable_*
+ functions. Design better to avoid maintenance issues. */
+
+/* Function vect_model_reduction_cost.
+
+ Models cost for a reduction operation, including the vector ops
+ generated within the strip-mine loop, the initial definition before
+ the loop, and the epilogue code that must be generated. */
+
+static void
+vect_model_reduction_cost (stmt_vec_info stmt_info, enum tree_code reduc_code,
+ int ncopies)
+{
+ int outer_cost = 0;
+ enum tree_code code;
+ optab optab;
+ tree vectype;
+ tree orig_stmt;
+ tree reduction_op;
+ enum machine_mode mode;
+ tree operation = GIMPLE_STMT_OPERAND (STMT_VINFO_STMT (stmt_info), 1);
+ int op_type = TREE_CODE_LENGTH (TREE_CODE (operation));
+
+ /* Cost of reduction op inside loop. */
+ STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info) += ncopies * TARG_VEC_STMT_COST;
+
+ reduction_op = TREE_OPERAND (operation, op_type-1);
+ vectype = get_vectype_for_scalar_type (TREE_TYPE (reduction_op));
+ mode = TYPE_MODE (vectype);
+ orig_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
+
+ if (!orig_stmt)
+ orig_stmt = STMT_VINFO_STMT (stmt_info);
+
+ code = TREE_CODE (GIMPLE_STMT_OPERAND (orig_stmt, 1));
+
+ /* Add in cost for initial definition. */
+ outer_cost += TARG_SCALAR_TO_VEC_COST;
+
+ /* Determine cost of epilogue code.
+
+ We have a reduction operator that will reduce the vector in one statement.
+ Also requires scalar extract. */
+
+ if (reduc_code < NUM_TREE_CODES)
+ outer_cost += TARG_VEC_STMT_COST + TARG_VEC_TO_SCALAR_COST;
+ else
+ {
+ int vec_size_in_bits = tree_low_cst (TYPE_SIZE (vectype), 1);
+ tree bitsize =
+ TYPE_SIZE (TREE_TYPE ( GIMPLE_STMT_OPERAND (orig_stmt, 0)));
+ int element_bitsize = tree_low_cst (bitsize, 1);
+ int nelements = vec_size_in_bits / element_bitsize;
+
+ optab = optab_for_tree_code (code, vectype);
+
+ /* We have a whole vector shift available. */
+ if (VECTOR_MODE_P (mode)
+ && optab->handlers[mode].insn_code != CODE_FOR_nothing
+ && vec_shr_optab->handlers[mode].insn_code != CODE_FOR_nothing)
+ /* Final reduction via vector shifts and the reduction operator. Also
+ requires scalar extract. */
+ outer_cost += ((exact_log2(nelements) * 2) * TARG_VEC_STMT_COST
+ + TARG_VEC_TO_SCALAR_COST);
+ else
+ /* Use extracts and reduction op for final reduction. For N elements,
+ we have N extracts and N-1 reduction ops. */
+ outer_cost += ((nelements + nelements - 1) * TARG_VEC_STMT_COST);
+ }
+
+ STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info) = outer_cost;
+
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "vect_model_reduction_cost: inside_cost = %d, "
+ "outside_cost = %d .", STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info),
+ STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info));
+}
+
+
+/* Function vect_model_induction_cost.
+
+ Models cost for induction operations. */
+
+static void
+vect_model_induction_cost (stmt_vec_info stmt_info, int ncopies)
+{
+ /* loop cost for vec_loop. */
+ STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info) = ncopies * TARG_VEC_STMT_COST;
+ /* prologue cost for vec_init and vec_step. */
+ STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info) = 2 * TARG_SCALAR_TO_VEC_COST;
+
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "vect_model_induction_cost: inside_cost = %d, "
+ "outside_cost = %d .", STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info),
+ STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info));
+}
+
+
+/* Function vect_model_simple_cost.
+
+ Models cost for simple operations, i.e. those that only emit ncopies of a
+ single op. Right now, this does not account for multiple insns that could
+ be generated for the single vector op. We will handle that shortly. */
+
+static void
+vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies, enum vect_def_type *dt)
+{
+ int i;
+
+ STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info) = ncopies * TARG_VEC_STMT_COST;
+
+ /* FORNOW: Assuming maximum 2 args per stmts. */
+ for (i=0; i<2; i++)
+ {
+ if (dt[i] == vect_constant_def || dt[i] == vect_invariant_def)
+ STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info) += TARG_SCALAR_TO_VEC_COST;
+ }
+
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "vect_model_simple_cost: inside_cost = %d, "
+ "outside_cost = %d .", STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info),
+ STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info));
+}
+
+
+/* Function vect_cost_strided_group_size
+
+ For strided load or store, return the group_size only if it is the first
+ load or store of a group, else return 1. This ensures that group size is
+ only returned once per group. */
+
+static int
+vect_cost_strided_group_size (stmt_vec_info stmt_info)
+{
+ tree first_stmt = DR_GROUP_FIRST_DR (stmt_info);
+
+ if (first_stmt == STMT_VINFO_STMT (stmt_info))
+ return DR_GROUP_SIZE (stmt_info);
+
+ return 1;
+}
+
+
+/* Function vect_model_store_cost
+
+ Models cost for stores. In the case of strided accesses, one access
+ has the overhead of the strided access attributed to it. */
+
+static void
+vect_model_store_cost (stmt_vec_info stmt_info, int ncopies, enum vect_def_type dt)
+{
+ int cost = 0;
+ int group_size;
+
+ if (dt == vect_constant_def || dt == vect_invariant_def)
+ STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info) = TARG_SCALAR_TO_VEC_COST;
+
+ /* Strided access? */
+ if (DR_GROUP_FIRST_DR (stmt_info))
+ group_size = vect_cost_strided_group_size (stmt_info);
+ /* Not a strided access. */
+ else
+ group_size = 1;
+
+ /* Is this an access in a group of stores, which provide strided access?
+ If so, add in the cost of the permutes. */
+ if (group_size > 1)
+ {
+ /* Uses a high and low interleave operation for each needed permute. */
+ cost = ncopies * exact_log2(group_size) * group_size
+ * TARG_VEC_STMT_COST;
+
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "vect_model_store_cost: strided group_size = %d .",
+ group_size);
+
+ }
+
+ /* Costs of the stores. */
+ cost += ncopies * TARG_VEC_STORE_COST;
+
+ STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info) = cost;
+
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "vect_model_store_cost: inside_cost = %d, "
+ "outside_cost = %d .", STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info),
+ STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info));
+}
+
+
+/* Function vect_model_load_cost
+
+ Models cost for loads. In the case of strided accesses, the last access
+ has the overhead of the strided access attributed to it. Since unaligned
+ accesses are supported for loads, we also account for the costs of the
+ access scheme chosen. */
+
+static void
+vect_model_load_cost (stmt_vec_info stmt_info, int ncopies)
+
+{
+ int inner_cost = 0;
+ int group_size;
+ int alignment_support_cheme;
+ tree first_stmt;
+ struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr;
+
+ /* Strided accesses? */
+ first_stmt = DR_GROUP_FIRST_DR (stmt_info);
+ if (first_stmt)
+ {
+ group_size = vect_cost_strided_group_size (stmt_info);
+ first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
+ }
+ /* Not a strided access. */
+ else
+ {
+ group_size = 1;
+ first_dr = dr;
+ }
+
+ alignment_support_cheme = vect_supportable_dr_alignment (first_dr);
+
+ /* Is this an access in a group of loads providing strided access?
+ If so, add in the cost of the permutes. */
+ if (group_size > 1)
+ {
+ /* Uses an even and odd extract operations for each needed permute. */
+ inner_cost = ncopies * exact_log2(group_size) * group_size
+ * TARG_VEC_STMT_COST;
+
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "vect_model_load_cost: strided group_size = %d .",
+ group_size);
+
+ }
+
+ /* The loads themselves. */
+ switch (alignment_support_cheme)
+ {
+ case dr_aligned:
+ {
+ inner_cost += ncopies * TARG_VEC_LOAD_COST;
+
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "vect_model_load_cost: aligned.");
+
+ break;
+ }
+ case dr_unaligned_supported:
+ {
+ /* Here, we assign an additional cost for the unaligned load. */
+ inner_cost += ncopies * TARG_VEC_UNALIGNED_LOAD_COST;
+
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "vect_model_load_cost: unaligned supported by "
+ "hardware.");
+
+ break;
+ }
+ case dr_unaligned_software_pipeline:
+ {
+ int outer_cost = 0;
+
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "vect_model_load_cost: unaligned software "
+ "pipelined.");
+
+ /* Unaligned software pipeline has a load of an address, an initial
+ load, and possibly a mask operation to "prime" the loop. However,
+ if this is an access in a group of loads, which provide strided
+ access, then the above cost should only be considered for one
+ access in the group. Inside the loop, there is a load op
+ and a realignment op. */
+
+ if ((!DR_GROUP_FIRST_DR (stmt_info)) || group_size > 1)
+ {
+ outer_cost = 2*TARG_VEC_STMT_COST;
+ if (targetm.vectorize.builtin_mask_for_load)
+ outer_cost += TARG_VEC_STMT_COST;
+ }
+
+ STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info) = outer_cost;
+
+ inner_cost += ncopies * (TARG_VEC_LOAD_COST + TARG_VEC_STMT_COST);
+
+ break;
+ }
+
+ default:
+ gcc_unreachable ();
+ }
+
+ STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info) = inner_cost;
+
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "vect_model_load_cost: inside_cost = %d, "
+ "outside_cost = %d .", STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info),
+ STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info));
+
+}
+
+
/* Function vect_get_new_vect_var.
Returns a name for a new variable. The current naming scheme appends the
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;
}
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 ref = DR_REF (dr);
- tree scalar_type = TREE_TYPE (ref);
- tree scalar_ptr_type = build_pointer_type (scalar_type);
tree vec_stmt;
- tree new_temp;
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);
+ base_offset = fold_convert (sizetype, base_offset);
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);
if (offset)
{
- tree tmp = create_tmp_var (TREE_TYPE (base_offset), "offset");
+ tree tmp = create_tmp_var (sizetype, "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,
- DR_STEP (dr));
+ 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);
+ 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,
+ addr_base = fold_build2 (POINTER_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 (scalar_ptr_type, vect_pointer_var,
+ addr_expr = vect_get_new_vect_var (vect_ptr_type, vect_pointer_var,
get_name (base_name));
add_referenced_var (addr_expr);
- vec_stmt = build2 (MODIFY_EXPR, void_type_node, addr_expr, addr_base);
- new_temp = make_ssa_name (addr_expr, vec_stmt);
- TREE_OPERAND (vec_stmt, 0) = new_temp;
- append_to_statement_list_force (vec_stmt, new_stmt_list);
+ 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 new_temp;
+ return vec_stmt;
}
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
+ 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
- MODIFY_EXPR <name, data-ref> or MODIFY_EXPR <data-ref, name>.
+ GIMPLE_MODIFY_STMT <name, data-ref> or
+ GIMPLE_MODIFY_STMT <data-ref, name>.
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
vect_create_data_ref_ptr (tree stmt,
block_stmt_iterator *bsi ATTRIBUTE_UNUSED,
tree offset, tree *initial_address, tree *ptr_incr,
- bool only_init)
+ bool only_init, tree type)
{
tree base_name;
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
}
/** (1) Create the new vector-pointer variable: **/
-
- vect_ptr_type = build_pointer_type (vectype);
+ 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);
+ tag = DR_SYMBOL_TAG (dr);
gcc_assert (tag);
/* If tag is a variable (and NOT_A_TAG) than a new symbol memory
if (!MTAG_P (tag))
new_type_alias (vect_ptr, tag, DR_REF (dr));
else
- var_ann (vect_ptr)->symbol_mem_tag = tag;
+ set_symbol_mem_tag (vect_ptr, tag);
var_ann (vect_ptr)->subvars = DR_SUBVARS (dr);
/* Create: p = (vectype *) initial_base */
vec_stmt = fold_convert (vect_ptr_type, new_temp);
- vec_stmt = build2 (MODIFY_EXPR, void_type_node, vect_ptr, vec_stmt);
+ vec_stmt = build_gimple_modify_stmt (vect_ptr, vec_stmt);
vect_ptr_init = make_ssa_name (vect_ptr, vec_stmt);
- TREE_OPERAND (vec_stmt, 0) = vect_ptr_init;
+ GIMPLE_STMT_OPERAND (vec_stmt, 0) = vect_ptr_init;
new_bb = bsi_insert_on_edge_immediate (pe, vec_stmt);
gcc_assert (!new_bb);
/* Function bump_vector_ptr
Increment a pointer (to a vector type) by vector-size. Connect the new
- increment stmt to the exising def-use update-chain of the pointer.
+ 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)
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 update = TYPE_SIZE_UNIT (vectype);
tree incr_stmt;
ssa_op_iter iter;
use_operand_p use_p;
tree new_dataref_ptr;
- incr_stmt = build2 (MODIFY_EXPR, void_type_node, ptr_var,
- build2 (PLUS_EXPR, vptr_type, dataref_ptr, update));
+ incr_stmt = build_gimple_modify_stmt (ptr_var,
+ build2 (POINTER_PLUS_EXPR, vptr_type,
+ dataref_ptr, update));
new_dataref_ptr = make_ssa_name (ptr_var, incr_stmt);
- TREE_OPERAND (incr_stmt, 0) = new_dataref_ptr;
+ 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. */
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);
+ vec_dest = vect_get_new_vect_var (type, kind, new_name);
add_referenced_var (vec_dest);
return vec_dest;
used in the vectorization of STMT. */
static tree
-vect_init_vector (tree stmt, tree vector_var)
+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 vectype = STMT_VINFO_VECTYPE (stmt_vinfo);
tree vec_oprnd;
edge pe;
tree new_temp;
basic_block new_bb;
- new_var = vect_get_new_vect_var (vectype, vect_simple_var, "cst_");
+ new_var = vect_get_new_vect_var (vector_type, vect_simple_var, "cst_");
add_referenced_var (new_var);
- init_stmt = build2 (MODIFY_EXPR, vectype, new_var, vector_var);
+ init_stmt = build_gimple_modify_stmt (new_var, vector_var);
new_temp = make_ssa_name (new_var, init_stmt);
- TREE_OPERAND (init_stmt, 0) = new_temp;
+ GIMPLE_STMT_OPERAND (init_stmt, 0) = new_temp;
pe = loop_preheader_edge (loop);
new_bb = bsi_insert_on_edge_immediate (pe, init_stmt);
print_generic_expr (vect_dump, init_stmt, TDF_SLIM);
}
- vec_oprnd = TREE_OPERAND (init_stmt, 0);
+ vec_oprnd = GIMPLE_STMT_OPERAND (init_stmt, 0);
return vec_oprnd;
}
-/* Function vect_get_vec_def_for_operand.
+/* Function get_initial_def_for_induction
- OP is an operand in STMT. This function returns a (vector) def that will be
- used in the vectorized stmt for STMT.
+ Input:
+ IV_PHI - the initial value of the induction variable
- 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.
+ Output:
+ Return a vector variable, initialized with the first VF values of
+ the induction variable. E.g., for an iv with IV_PHI='X' and
+ evolution S, for a vector of 4 units, we want to return:
+ [X, X + S, X + 2*S, X + 3*S]. */
+
+static tree
+get_initial_def_for_induction (tree iv_phi)
+{
+ stmt_vec_info stmt_vinfo = vinfo_for_stmt (iv_phi);
+ loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
+ struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
+ tree scalar_type = TREE_TYPE (PHI_RESULT_TREE (iv_phi));
+ tree vectype = get_vectype_for_scalar_type (scalar_type);
+ int nunits = TYPE_VECTOR_SUBPARTS (vectype);
+ edge pe = loop_preheader_edge (loop);
+ basic_block new_bb;
+ block_stmt_iterator bsi;
+ tree vec, vec_init, vec_step, t;
+ tree access_fn;
+ tree new_var;
+ tree new_name;
+ tree init_stmt;
+ tree induction_phi, induc_def, new_stmt, vec_def, vec_dest;
+ tree init_expr, step_expr;
+ int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
+ int i;
+ bool ok;
+ int ncopies = vf / nunits;
+ tree expr;
+ stmt_vec_info phi_info = vinfo_for_stmt (iv_phi);
+ tree stmts;
+ tree stmt = NULL_TREE;
+ block_stmt_iterator si;
+ basic_block bb = bb_for_stmt (iv_phi);
+
+ gcc_assert (phi_info);
+ gcc_assert (ncopies >= 1);
+
+ /* Find the first insertion point in the BB. */
+ si = bsi_after_labels (bb);
+ stmt = bsi_stmt (si);
+
+ access_fn = analyze_scalar_evolution (loop, PHI_RESULT (iv_phi));
+ gcc_assert (access_fn);
+ ok = vect_is_simple_iv_evolution (loop->num, access_fn,
+ &init_expr, &step_expr);
+ gcc_assert (ok);
+
+ /* Create the vector that holds the initial_value of the induction. */
+ new_var = vect_get_new_vect_var (scalar_type, vect_scalar_var, "var_");
+ add_referenced_var (new_var);
+
+ new_name = force_gimple_operand (init_expr, &stmts, false, new_var);
+ if (stmts)
+ {
+ new_bb = bsi_insert_on_edge_immediate (pe, stmts);
+ gcc_assert (!new_bb);
+ }
+
+ t = NULL_TREE;
+ t = tree_cons (NULL_TREE, new_name, t);
+ for (i = 1; i < nunits; i++)
+ {
+ tree tmp;
+
+ /* Create: new_name = new_name + step_expr */
+ tmp = fold_build2 (PLUS_EXPR, scalar_type, new_name, step_expr);
+ init_stmt = build_gimple_modify_stmt (new_var, tmp);
+ new_name = make_ssa_name (new_var, init_stmt);
+ GIMPLE_STMT_OPERAND (init_stmt, 0) = new_name;
+
+ 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);
+ }
+ t = tree_cons (NULL_TREE, new_name, t);
+ }
+ vec = build_constructor_from_list (vectype, nreverse (t));
+ vec_init = vect_init_vector (stmt, vec, vectype);
+
+
+ /* Create the vector that holds the step of the induction. */
+ expr = build_int_cst (scalar_type, vf);
+ new_name = fold_build2 (MULT_EXPR, scalar_type, expr, step_expr);
+ t = NULL_TREE;
+ for (i = 0; i < nunits; i++)
+ t = tree_cons (NULL_TREE, unshare_expr (new_name), t);
+ vec = build_constructor_from_list (vectype, t);
+ vec_step = vect_init_vector (stmt, vec, vectype);
+
+
+ /* Create the following def-use cycle:
+ loop prolog:
+ vec_init = [X, X+S, X+2*S, X+3*S]
+ vec_step = [VF*S, VF*S, VF*S, VF*S]
+ loop:
+ vec_iv = PHI <vec_init, vec_loop>
+ ...
+ STMT
+ ...
+ vec_loop = vec_iv + vec_step; */
+
+ /* Create the induction-phi that defines the induction-operand. */
+ vec_dest = vect_get_new_vect_var (vectype, vect_simple_var, "vec_iv_");
+ add_referenced_var (vec_dest);
+ induction_phi = create_phi_node (vec_dest, loop->header);
+ set_stmt_info (get_stmt_ann (induction_phi),
+ new_stmt_vec_info (induction_phi, loop_vinfo));
+ induc_def = PHI_RESULT (induction_phi);
+
+ /* Create the iv update inside the loop */
+ new_stmt = build_gimple_modify_stmt (NULL_TREE,
+ build2 (PLUS_EXPR, vectype,
+ induc_def, vec_step));
+ vec_def = make_ssa_name (vec_dest, new_stmt);
+ GIMPLE_STMT_OPERAND (new_stmt, 0) = vec_def;
+ bsi = bsi_for_stmt (stmt);
+ vect_finish_stmt_generation (stmt, new_stmt, &bsi);
+
+ /* Set the arguments of the phi node: */
+ add_phi_arg (induction_phi, vec_init, loop_preheader_edge (loop));
+ add_phi_arg (induction_phi, vec_def, loop_latch_edge (loop));
+
+
+ /* In case the vectorization factor (VF) is bigger than the number
+ of elements that we can fit in a vectype (nunits), we have to generate
+ more than one vector stmt - i.e - we need to "unroll" the
+ vector stmt by a factor VF/nunits. For more details see documentation
+ in vectorizable_operation. */
+
+ if (ncopies > 1)
+ {
+ stmt_vec_info prev_stmt_vinfo;
+
+ /* Create the vector that holds the step of the induction. */
+ expr = build_int_cst (scalar_type, nunits);
+ new_name = fold_build2 (MULT_EXPR, scalar_type, expr, step_expr);
+ t = NULL_TREE;
+ for (i = 0; i < nunits; i++)
+ t = tree_cons (NULL_TREE, unshare_expr (new_name), t);
+ vec = build_constructor_from_list (vectype, t);
+ vec_step = vect_init_vector (stmt, vec, vectype);
+
+ vec_def = induc_def;
+ prev_stmt_vinfo = vinfo_for_stmt (induction_phi);
+ for (i = 1; i < ncopies; i++)
+ {
+ tree tmp;
+
+ /* vec_i = vec_prev + vec_{step*nunits} */
+ tmp = build2 (PLUS_EXPR, vectype, vec_def, vec_step);
+ new_stmt = build_gimple_modify_stmt (NULL_TREE, tmp);
+ vec_def = make_ssa_name (vec_dest, new_stmt);
+ GIMPLE_STMT_OPERAND (new_stmt, 0) = vec_def;
+ bsi = bsi_for_stmt (stmt);
+ vect_finish_stmt_generation (stmt, new_stmt, &bsi);
+
+ STMT_VINFO_RELATED_STMT (prev_stmt_vinfo) = new_stmt;
+ prev_stmt_vinfo = vinfo_for_stmt (new_stmt);
+ }
+ }
+
+ if (vect_print_dump_info (REPORT_DETAILS))
+ {
+ fprintf (vect_dump, "transform induction: created def-use cycle:");
+ print_generic_expr (vect_dump, induction_phi, TDF_SLIM);
+ fprintf (vect_dump, "\n");
+ print_generic_expr (vect_dump, SSA_NAME_DEF_STMT (vec_def), TDF_SLIM);
+ }
+
+ STMT_VINFO_VEC_STMT (phi_info) = induction_phi;
+ return induc_def;
+}
+
+
+/* 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. */
int i;
enum vect_def_type dt;
bool is_simple_use;
+ tree vector_type;
if (vect_print_dump_info (REPORT_DETAILS))
{
{
t = tree_cons (NULL_TREE, op, t);
}
- vec_cst = build_vector (vectype, t);
- return vect_init_vector (stmt, vec_cst);
+ 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. */
}
/* FIXME: use build_constructor directly. */
- vec_inv = build_constructor_from_list (vectype, t);
- return vect_init_vector (stmt, vec_inv);
+ 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. */
def_stmt_info = vinfo_for_stmt (def_stmt);
vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info);
gcc_assert (vec_stmt);
- vec_oprnd = TREE_OPERAND (vec_stmt, 0);
+ vec_oprnd = GIMPLE_STMT_OPERAND (vec_stmt, 0);
return vec_oprnd;
}
/* 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 */
+ gcc_assert (TREE_CODE (def_stmt) == PHI_NODE);
+
+ /* Get the def before the loop */
+ return get_initial_def_for_induction (def_stmt);
}
default:
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
+ 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.
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 illusration below, in
- which VF=16 and nuniti=4, so the number of copies required is 4):
+ 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
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;
- }
+ /* Do nothing; can reuse same def. */
+ if (dt == vect_invariant_def || dt == vect_constant_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 = TREE_OPERAND (vec_stmt_for_operand, 0);
+ vec_oprnd = GIMPLE_STMT_OPERAND (vec_stmt_for_operand, 0);
return vec_oprnd;
}
}
-#define ADJUST_IN_EPILOG 1
-
/* Function get_initial_def_for_reduction
Input:
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).
+ ADJUSTMENT_DEF - a tree that holds a value to be added to the final result
+ of the reduction (used for adjusting the epilog - see below).
Return a vector variable, initialized according to the operation that STMT
- performs. This vector will be used as the initial value of the
- vector of partial results.
+ 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:
+ 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
+ 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:
or [0,0,0,0] and let the caller know that it needs to adjust
the result at the end by 'init_val'.
- FORNOW: We use the "ADJUST_IN_EPILOG" scheme.
- TODO: Use some cost-model to estimate which scheme is more profitable.
-*/
+ FORNOW, we are using the 'adjust in epilog' scheme, because this way the
+ initialization vector is simpler (same element in all entries).
+ A cost model should help decide between these two schemes. */
static tree
-get_initial_def_for_reduction (tree stmt, tree init_val, tree *scalar_def)
+get_initial_def_for_reduction (tree stmt, tree init_val, tree *adjustment_def)
{
stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt);
tree vectype = STMT_VINFO_VECTYPE (stmt_vinfo);
- int nunits = GET_MODE_NUNITS (TYPE_MODE (vectype));
- int nelements;
- enum tree_code code = TREE_CODE (TREE_OPERAND (stmt, 1));
+ int nunits = TYPE_VECTOR_SUBPARTS (vectype);
+ enum tree_code code = TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 1));
tree type = TREE_TYPE (init_val);
- tree def;
- tree vec, t = NULL_TREE;
- bool need_epilog_adjust;
+ tree vecdef;
+ tree def_for_init;
+ tree init_def;
+ tree t = NULL_TREE;
int i;
+ tree vector_type;
gcc_assert (INTEGRAL_TYPE_P (type) || SCALAR_FLOAT_TYPE_P (type));
+ vecdef = vect_get_vec_def_for_operand (init_val, stmt, NULL);
switch (code)
{
case WIDEN_SUM_EXPR:
case DOT_PROD_EXPR:
case PLUS_EXPR:
+ *adjustment_def = init_val;
+ /* Create a vector of zeros for init_def. */
if (INTEGRAL_TYPE_P (type))
- def = build_int_cst (type, 0);
+ def_for_init = build_int_cst (type, 0);
else
- def = build_real (type, dconst0);
-
-#ifdef ADJUST_IN_EPILOG
- /* All the 'nunits' elements are set to 0. The final result will be
- adjusted by 'init_val' at the loop epilog. */
- nelements = nunits;
- need_epilog_adjust = true;
-#else
- /* 'nunits - 1' elements are set to 0; The last element is set to
- 'init_val'. No further adjustments at the epilog are needed. */
- nelements = nunits - 1;
- need_epilog_adjust = false;
-#endif
+ def_for_init = build_real (type, dconst0);
+ for (i = nunits - 1; i >= 0; --i)
+ t = tree_cons (NULL_TREE, def_for_init, t);
+ vector_type = get_vectype_for_scalar_type (TREE_TYPE (def_for_init));
+ init_def = build_vector (vector_type, t);
break;
case MIN_EXPR:
case MAX_EXPR:
- def = init_val;
- nelements = nunits;
- need_epilog_adjust = false;
+ *adjustment_def = NULL_TREE;
+ init_def = vecdef;
break;
default:
gcc_unreachable ();
}
- for (i = nelements - 1; 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);
-
- if (TREE_CODE (init_val) == INTEGER_CST || TREE_CODE (init_val) == REAL_CST)
- vec = build_vector (vectype, t);
- else
- vec = build_constructor_from_list (vectype, t);
-
- if (!need_epilog_adjust)
- *scalar_def = NULL_TREE;
- else
- *scalar_def = init_val;
-
- return vect_init_vector (stmt, vec);
+ return init_def;
}
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
+ 1. Creates the reduction def-use cycle: sets 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.
loop:
vec_def = phi <null, null> # REDUCTION_PHI
- VECT_DEF = vector_stmt # vectorized form of STMT
+ VECT_DEF = vector_stmt # vectorized form of STMT
s_loop = scalar_stmt # (scalar) STMT
loop_exit:
s_out0 = phi <s_loop> # (scalar) EXIT_PHI
tree epilog_stmt;
tree new_scalar_dest, exit_phi;
tree bitsize, bitpos, bytesize;
- enum tree_code code = TREE_CODE (TREE_OPERAND (stmt, 1));
+ enum tree_code code = TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 1));
tree scalar_initial_def;
tree vec_initial_def;
tree orig_name;
tree reduction_op;
tree orig_stmt;
tree use_stmt;
- tree operation = TREE_OPERAND (stmt, 1);
+ tree operation = GIMPLE_STMT_OPERAND (stmt, 1);
int op_type;
- op_type = TREE_CODE_LENGTH (TREE_CODE (operation));
+ op_type = TREE_OPERAND_LENGTH (operation);
reduction_op = TREE_OPERAND (operation, op_type-1);
vectype = get_vectype_for_scalar_type (TREE_TYPE (reduction_op));
mode = TYPE_MODE (vectype);
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);
+ exit_bsi = bsi_after_labels (exit_bb);
/* 2.2 Get the relevant tree-code to use in the epilog for schemes 2,3
- (i.e. when reduc_code is not available) and in the final adjustment code
- (if needed). Also get the original scalar reduction variable as
+ (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.
gcc_assert (STMT_VINFO_IN_PATTERN_P (stmt_vinfo));
gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo) == stmt);
}
- code = TREE_CODE (TREE_OPERAND (orig_stmt, 1));
- scalar_dest = TREE_OPERAND (orig_stmt, 0);
+ 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);
if (reduc_code < NUM_TREE_CODES)
{
+ tree tmp;
+
/*** Case 1: Create:
v_out2 = reduc_expr <v_out1> */
fprintf (vect_dump, "Reduce using direct vector reduction.");
vec_dest = vect_create_destination_var (scalar_dest, vectype);
- epilog_stmt = build2 (MODIFY_EXPR, vectype, vec_dest,
- build1 (reduc_code, vectype, PHI_RESULT (new_phi)));
+ tmp = build1 (reduc_code, vectype, PHI_RESULT (new_phi));
+ epilog_stmt = build_gimple_modify_stmt (vec_dest, tmp);
new_temp = make_ssa_name (vec_dest, epilog_stmt);
- TREE_OPERAND (epilog_stmt, 0) = new_temp;
- bsi_insert_after (&exit_bsi, epilog_stmt, BSI_NEW_STMT);
+ GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_temp;
+ bsi_insert_before (&exit_bsi, epilog_stmt, BSI_SAME_STMT);
extract_scalar_result = true;
}
bit_offset /= 2)
{
tree bitpos = size_int (bit_offset);
-
- epilog_stmt = build2 (MODIFY_EXPR, vectype, vec_dest,
- build2 (shift_code, vectype,
- new_temp, bitpos));
+ tree tmp = build2 (shift_code, vectype, new_temp, bitpos);
+ epilog_stmt = build_gimple_modify_stmt (vec_dest, tmp);
new_name = make_ssa_name (vec_dest, epilog_stmt);
- TREE_OPERAND (epilog_stmt, 0) = new_name;
- bsi_insert_after (&exit_bsi, epilog_stmt, BSI_NEW_STMT);
+ GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_name;
+ bsi_insert_before (&exit_bsi, epilog_stmt, BSI_SAME_STMT);
- epilog_stmt = build2 (MODIFY_EXPR, vectype, vec_dest,
- build2 (code, vectype,
- new_name, new_temp));
+ tmp = build2 (code, vectype, new_name, new_temp);
+ epilog_stmt = build_gimple_modify_stmt (vec_dest, tmp);
new_temp = make_ssa_name (vec_dest, epilog_stmt);
- TREE_OPERAND (epilog_stmt, 0) = new_temp;
- bsi_insert_after (&exit_bsi, epilog_stmt, BSI_NEW_STMT);
+ GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_temp;
+ bsi_insert_before (&exit_bsi, epilog_stmt, BSI_SAME_STMT);
}
extract_scalar_result = true;
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 (MODIFY_EXPR, scalar_type, new_scalar_dest, rhs);
+ epilog_stmt = build_gimple_modify_stmt (new_scalar_dest, rhs);
new_temp = make_ssa_name (new_scalar_dest, epilog_stmt);
- TREE_OPERAND (epilog_stmt, 0) = new_temp;
- bsi_insert_after (&exit_bsi, epilog_stmt, BSI_NEW_STMT);
+ GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_temp;
+ bsi_insert_before (&exit_bsi, epilog_stmt, BSI_SAME_STMT);
for (bit_offset = element_bitsize;
bit_offset < vec_size_in_bits;
bit_offset += element_bitsize)
{
+ tree tmp;
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 (MODIFY_EXPR, scalar_type, new_scalar_dest,
- rhs);
+ epilog_stmt = build_gimple_modify_stmt (new_scalar_dest, rhs);
new_name = make_ssa_name (new_scalar_dest, epilog_stmt);
- TREE_OPERAND (epilog_stmt, 0) = new_name;
- bsi_insert_after (&exit_bsi, epilog_stmt, BSI_NEW_STMT);
+ GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_name;
+ bsi_insert_before (&exit_bsi, epilog_stmt, BSI_SAME_STMT);
- epilog_stmt = build2 (MODIFY_EXPR, scalar_type, new_scalar_dest,
- build2 (code, scalar_type, new_name, new_temp));
+ tmp = build2 (code, scalar_type, new_name, new_temp);
+ epilog_stmt = build_gimple_modify_stmt (new_scalar_dest, tmp);
new_temp = make_ssa_name (new_scalar_dest, epilog_stmt);
- TREE_OPERAND (epilog_stmt, 0) = new_temp;
- bsi_insert_after (&exit_bsi, epilog_stmt, BSI_NEW_STMT);
+ GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_temp;
+ bsi_insert_before (&exit_bsi, epilog_stmt, BSI_SAME_STMT);
}
extract_scalar_result = false;
rhs = build3 (BIT_FIELD_REF, scalar_type, new_temp, bitsize, bitpos);
BIT_FIELD_REF_UNSIGNED (rhs) = TYPE_UNSIGNED (scalar_type);
- epilog_stmt = build2 (MODIFY_EXPR, scalar_type, new_scalar_dest, rhs);
+ epilog_stmt = build_gimple_modify_stmt (new_scalar_dest, rhs);
new_temp = make_ssa_name (new_scalar_dest, epilog_stmt);
- TREE_OPERAND (epilog_stmt, 0) = new_temp;
- bsi_insert_after (&exit_bsi, epilog_stmt, BSI_NEW_STMT);
+ GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_temp;
+ bsi_insert_before (&exit_bsi, epilog_stmt, BSI_SAME_STMT);
}
/* 2.4 Adjust the final result by the initial value of the reduction
if (scalar_initial_def)
{
- epilog_stmt = build2 (MODIFY_EXPR, scalar_type, new_scalar_dest,
- build2 (code, scalar_type, new_temp, scalar_initial_def));
+ tree tmp = build2 (code, scalar_type, new_temp, scalar_initial_def);
+ epilog_stmt = build_gimple_modify_stmt (new_scalar_dest, tmp);
new_temp = make_ssa_name (new_scalar_dest, epilog_stmt);
- TREE_OPERAND (epilog_stmt, 0) = new_temp;
- bsi_insert_after (&exit_bsi, epilog_stmt, BSI_NEW_STMT);
+ GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_temp;
+ bsi_insert_before (&exit_bsi, epilog_stmt, BSI_SAME_STMT);
}
/* 2.6 Replace uses of s_out0 with uses of s_out3 */
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
+ 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
+ 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.
+ 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).
+ 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*
+ 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:
+ 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
+ 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. */
inside the loop body. The last operand is the reduction variable,
which is defined by the loop-header-phi. */
- gcc_assert (TREE_CODE (stmt) == MODIFY_EXPR);
+ gcc_assert (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT);
- operation = TREE_OPERAND (stmt, 1);
+ operation = GIMPLE_STMT_OPERAND (stmt, 1);
code = TREE_CODE (operation);
- op_type = TREE_CODE_LENGTH (code);
+ op_type = TREE_OPERAND_LENGTH (operation);
if (op_type != binary_op && op_type != ternary_op)
return false;
- scalar_dest = TREE_OPERAND (stmt, 0);
+ 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.
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);
+ if (dt != vect_loop_def
+ && dt != vect_invariant_def
+ && dt != vect_constant_def
+ && dt != vect_induction_def)
+ return false;
}
op = TREE_OPERAND (operation, i);
{
/* 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 (TREE_OPERAND (orig_stmt, 1));
+ orig_code = TREE_CODE (GIMPLE_STMT_OPERAND (orig_stmt, 1));
vectype = get_vectype_for_scalar_type (TREE_TYPE (def));
vec_mode = TYPE_MODE (vectype);
}
if (!vec_stmt) /* transformation not required. */
{
STMT_VINFO_TYPE (stmt_info) = reduc_vec_info_type;
+ vect_model_reduction_cost (stmt_info, epilog_reduc_code, ncopies);
return true;
}
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);
}
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 = TREE_OPERAND (new_stmt ,0);
+ 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 (MODIFY_EXPR, void_type_node, vec_dest, expr);
+ reduc_def);
+ new_stmt = build_gimple_modify_stmt (vec_dest, expr);
new_temp = make_ssa_name (vec_dest, new_stmt);
- TREE_OPERAND (new_stmt, 0) = new_temp;
+ 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);
+ 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
+ a function declaration if the target has a vectorized version
+ of the function, or NULL_TREE if the function cannot be vectorized. */
+
+tree
+vectorizable_function (tree call, tree vectype_out, tree vectype_in)
+{
+ tree fndecl = get_callee_fndecl (call);
+ enum built_in_function code;
+
+ /* 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 NULL_TREE;
+
+ if (!fndecl
+ || TREE_CODE (fndecl) != FUNCTION_DECL
+ || !DECL_BUILT_IN (fndecl))
+ return NULL_TREE;
+
+ code = DECL_FUNCTION_CODE (fndecl);
+ return targetm.vectorize.builtin_vectorized_function (code, vectype_out,
+ vectype_in);
+}
+
+/* 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, type;
+ tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE;
+ stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info;
+ tree vectype_out, vectype_in;
+ int nunits_in;
+ int nunits_out;
+ loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
+ tree fndecl, rhs, new_temp, def, def_stmt, rhs_type, lhs_type;
+ enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
+ tree new_stmt;
+ int ncopies, j, nargs;
+ call_expr_arg_iterator iter;
+ tree vargs;
+ enum { NARROW, NONE, WIDEN } modifier;
+
+ if (!STMT_VINFO_RELEVANT_P (stmt_info))
+ return false;
+
+ if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
+ return false;
+
+ /* FORNOW: not yet supported. */
+ if (STMT_VINFO_LIVE_P (stmt_info))
+ {
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "value used after loop.");
+ return false;
+ }
+
+ /* 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;
+
+ /* Process function arguments. */
+ rhs_type = NULL_TREE;
+ nargs = 0;
+ FOR_EACH_CALL_EXPR_ARG (op, iter, operation)
+ {
+ /* Bail out if the function has more than two arguments, we
+ do not have interesting builtin functions to vectorize with
+ more than two arguments. */
+ if (nargs >= 2)
+ return false;
+
+ /* We can only handle calls with arguments of the same type. */
+ if (rhs_type
+ && rhs_type != TREE_TYPE (op))
+ {
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "argument types differ.");
+ return false;
+ }
+ rhs_type = TREE_TYPE (op);
+
+ if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt[nargs]))
+ {
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "use not simple.");
+ return false;
+ }
+
+ ++nargs;
+ }
+
+ /* No arguments is also not good. */
+ if (nargs == 0)
+ return false;
+
+ vectype_in = get_vectype_for_scalar_type (rhs_type);
+ nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
+
+ lhs_type = TREE_TYPE (GIMPLE_STMT_OPERAND (stmt, 0));
+ vectype_out = get_vectype_for_scalar_type (lhs_type);
+ nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
+
+ /* FORNOW */
+ if (nunits_in == nunits_out / 2)
+ modifier = NARROW;
+ else if (nunits_out == nunits_in)
+ modifier = NONE;
+ else if (nunits_out == nunits_in / 2)
+ modifier = WIDEN;
+ else
+ 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. */
+ fndecl = vectorizable_function (operation, vectype_out, vectype_in);
+ if (fndecl == NULL_TREE)
+ {
+ 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));
+
+ if (modifier == NARROW)
+ ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
+ else
+ ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
+
+ /* Sanity check: make sure that at least one copy of the vectorized stmt
+ needs to be generated. */
+ gcc_assert (ncopies >= 1);
+
+ if (!vec_stmt) /* transformation not required. */
+ {
+ STMT_VINFO_TYPE (stmt_info) = call_vec_info_type;
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "=== vectorizable_call ===");
+ vect_model_simple_cost (stmt_info, ncopies, dt);
+ 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_out);
+
+ prev_stmt_info = NULL;
+ switch (modifier)
+ {
+ case NONE:
+ for (j = 0; j < ncopies; ++j)
+ {
+ /* Build argument list for the vectorized call. */
+ /* FIXME: Rewrite this so that it doesn't
+ construct a temporary list. */
+ vargs = NULL_TREE;
+ nargs = 0;
+ FOR_EACH_CALL_EXPR_ARG (op, iter, operation)
+ {
+ if (j == 0)
+ vec_oprnd0
+ = vect_get_vec_def_for_operand (op, stmt, NULL);
+ else
+ vec_oprnd0
+ = vect_get_vec_def_for_stmt_copy (dt[nargs], vec_oprnd0);
+
+ vargs = tree_cons (NULL_TREE, vec_oprnd0, vargs);
+
+ ++nargs;
+ }
+ vargs = nreverse (vargs);
+
+ rhs = build_function_call_expr (fndecl, vargs);
+ new_stmt = build_gimple_modify_stmt (vec_dest, rhs);
+ 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);
+ }
+
+ break;
+
+ case NARROW:
+ for (j = 0; j < ncopies; ++j)
+ {
+ /* Build argument list for the vectorized call. */
+ /* FIXME: Rewrite this so that it doesn't
+ construct a temporary list. */
+ vargs = NULL_TREE;
+ nargs = 0;
+ FOR_EACH_CALL_EXPR_ARG (op, iter, operation)
+ {
+ if (j == 0)
+ {
+ vec_oprnd0
+ = vect_get_vec_def_for_operand (op, stmt, NULL);
+ vec_oprnd1
+ = vect_get_vec_def_for_stmt_copy (dt[nargs], vec_oprnd0);
+ }
+ else
+ {
+ vec_oprnd0
+ = vect_get_vec_def_for_stmt_copy (dt[nargs], vec_oprnd1);
+ vec_oprnd1
+ = vect_get_vec_def_for_stmt_copy (dt[nargs], vec_oprnd0);
+ }
+
+ vargs = tree_cons (NULL_TREE, vec_oprnd0, vargs);
+ vargs = tree_cons (NULL_TREE, vec_oprnd1, vargs);
+
+ ++nargs;
+ }
+ vargs = nreverse (vargs);
+
+ rhs = build_function_call_expr (fndecl, vargs);
+ new_stmt = build_gimple_modify_stmt (vec_dest, rhs);
+ 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);
+
+ break;
+
+ case WIDEN:
+ /* No current target implements this case. */
+ return false;
+ }
+
+ /* 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);
+ update_stmt (stmt);
+
+ 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 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 */
+ if (op_type == binary_op)
+ expr = build_call_expr (decl, 2, vec_oprnd0, vec_oprnd1);
+ else
+ expr = build_call_expr (decl, 1, vec_oprnd0);
+ }
+ 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 = build_gimple_modify_stmt (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_conversion.
+
+Check if STMT performs a conversion 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_conversion (tree stmt, block_stmt_iterator * bsi,
+ tree * vec_stmt)
+{
+ tree vec_dest;
+ tree scalar_dest;
+ tree operation;
+ tree op0;
+ tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE;
+ 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 = ERROR_MARK, code2 = ERROR_MARK;
+ tree decl1 = NULL_TREE, decl2 = NULL_TREE;
+ 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, vectype_in;
+ int ncopies, j;
+ tree expr;
+ tree rhs_type, lhs_type;
+ tree builtin_decl;
+ enum { NARROW, NONE, WIDEN } modifier;
+
+ /* Is STMT a vectorizable conversion? */
+
+ if (!STMT_VINFO_RELEVANT_P (stmt_info))
+ return false;
+
+ if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
+ return false;
+
+ 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 != FIX_TRUNC_EXPR && code != FLOAT_EXPR)
+ return false;
+
+ /* Check types of lhs and rhs */
+ op0 = TREE_OPERAND (operation, 0);
+ rhs_type = TREE_TYPE (op0);
+ vectype_in = get_vectype_for_scalar_type (rhs_type);
+ nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
+
+ scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
+ lhs_type = TREE_TYPE (scalar_dest);
+ vectype_out = get_vectype_for_scalar_type (lhs_type);
+ nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
+
+ /* FORNOW */
+ if (nunits_in == nunits_out / 2)
+ modifier = NARROW;
+ else if (nunits_out == nunits_in)
+ modifier = NONE;
+ else if (nunits_out == nunits_in / 2)
+ modifier = WIDEN;
+ else
+ return false;
+
+ if (modifier == NONE)
+ gcc_assert (STMT_VINFO_VECTYPE (stmt_info) == vectype_out);
+
+ /* Bail out if the types are both integral or non-integral */
+ if ((INTEGRAL_TYPE_P (rhs_type) && INTEGRAL_TYPE_P (lhs_type))
+ || (!INTEGRAL_TYPE_P (rhs_type) && !INTEGRAL_TYPE_P (lhs_type)))
+ return false;
+
+ if (modifier == NARROW)
+ ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
+ else
+ ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
+
+ /* Sanity check: make sure that at least one copy of the vectorized stmt
+ needs to be generated. */
+ gcc_assert (ncopies >= 1);
+
+ /* 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? */
+ if ((modifier == NONE
+ && !targetm.vectorize.builtin_conversion (code, vectype_in))
+ || (modifier == WIDEN
+ && !supportable_widening_operation (code, stmt, vectype_in,
+ &decl1, &decl2,
+ &code1, &code2))
+ || (modifier == NARROW
+ && !supportable_narrowing_operation (code, stmt, vectype_in,
+ &code1)))
+ {
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "op not supported by target.");
+ return false;
+ }
+
+ if (modifier != NONE)
+ STMT_VINFO_VECTYPE (stmt_info) = vectype_in;
+
+ if (!vec_stmt) /* transformation not required. */
+ {
+ STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type;
+ return true;
+ }
+
+ /** Transform. **/
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "transform conversion.");
+
+ /* Handle def. */
+ vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
+
+ prev_stmt_info = NULL;
+ switch (modifier)
+ {
+ case NONE:
+ for (j = 0; j < ncopies; j++)
+ {
+ tree sym;
+ ssa_op_iter iter;
+
+ if (j == 0)
+ vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
+ else
+ vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt0, vec_oprnd0);
+
+ builtin_decl =
+ targetm.vectorize.builtin_conversion (code, vectype_in);
+ new_stmt = build_call_expr (builtin_decl, 1, vec_oprnd0);
+
+ /* Arguments are ready. create the new vector stmt. */
+ new_stmt = build_gimple_modify_stmt (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);
+ 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);
+ }
+
+ 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);
+ }
+ break;
+
+ case WIDEN:
+ /* 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 (j = 0; j < ncopies; j++)
+ {
+ if (j == 0)
+ vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
+ else
+ vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt0, vec_oprnd0);
+
+ STMT_VINFO_VECTYPE (stmt_info) = vectype_in;
+
+ /* Generate first half of the widened result: */
+ new_stmt
+ = vect_gen_widened_results_half (code1, vectype_out, decl1,
+ vec_oprnd0, vec_oprnd1,
+ unary_op, 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,
+ unary_op, vec_dest, bsi, stmt);
+ STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
+ prev_stmt_info = vinfo_for_stmt (new_stmt);
+ }
+ break;
+
+ case NARROW:
+ /* 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 (j = 0; j < ncopies; j++)
+ {
+ /* Handle uses. */
+ if (j == 0)
+ {
+ vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
+ vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt0, 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 (code1, vectype_out, vec_oprnd0, vec_oprnd1);
+ new_stmt = build_gimple_modify_stmt (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;
}
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
tree new_temp;
tree def, def_stmt;
- enum vect_def_type dt;
+ enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
int nunits = TYPE_VECTOR_SUBPARTS (vectype);
int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
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 (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
+ return false;
- if (TREE_CODE (stmt) != MODIFY_EXPR)
+ /* FORNOW: not yet supported. */
+ if (STMT_VINFO_LIVE_P (stmt_info))
+ {
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "value used after loop.");
+ return false;
+ }
+
+ /* Is vectorizable assignment? */
+ if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
return false;
- scalar_dest = TREE_OPERAND (stmt, 0);
+ scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
if (TREE_CODE (scalar_dest) != SSA_NAME)
return false;
- op = TREE_OPERAND (stmt, 1);
- if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt))
+ op = GIMPLE_STMT_OPERAND (stmt, 1);
+ if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt[0]))
{
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "use not simple.");
if (!vec_stmt) /* transformation not required. */
{
STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type;
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "=== vectorizable_assignment ===");
+ vect_model_simple_cost (stmt_info, ncopies, dt);
return true;
}
vec_dest = vect_create_destination_var (scalar_dest, vectype);
/* Handle use. */
- op = TREE_OPERAND (stmt, 1);
+ 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 (MODIFY_EXPR, vectype, vec_dest, vec_oprnd);
+ *vec_stmt = build_gimple_modify_stmt (vec_dest, vec_oprnd);
new_temp = make_ssa_name (vec_dest, *vec_stmt);
- TREE_OPERAND (*vec_stmt, 0) = new_temp;
+ GIMPLE_STMT_OPERAND (*vec_stmt, 0) = new_temp;
vect_finish_stmt_generation (stmt, *vec_stmt, bsi);
return true;
{
switch (code)
{
- case PLUS_EXPR:
- case MINUS_EXPR:
- case NEGATE_EXPR:
- return 4;
+ 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_induction
+
+ Check if PHI performs an induction computation that can be vectorized.
+ If VEC_STMT is also passed, vectorize the induction PHI: create a vectorized
+ phi to replace it, put it in VEC_STMT, and add it to the same basic block.
+ Return FALSE if not a vectorizable STMT, TRUE otherwise. */
+
+bool
+vectorizable_induction (tree phi, block_stmt_iterator *bsi ATTRIBUTE_UNUSED,
+ tree *vec_stmt)
+{
+ stmt_vec_info stmt_info = vinfo_for_stmt (phi);
+ tree vectype = STMT_VINFO_VECTYPE (stmt_info);
+ loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
+ int nunits = TYPE_VECTOR_SUBPARTS (vectype);
+ int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
+ tree vec_def;
+
+ gcc_assert (ncopies >= 1);
+
+ if (!STMT_VINFO_RELEVANT_P (stmt_info))
+ return false;
+
+ gcc_assert (STMT_VINFO_DEF_TYPE (stmt_info) == vect_induction_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;
+ }
- case BIT_AND_EXPR:
- case BIT_IOR_EXPR:
- case BIT_XOR_EXPR:
- case BIT_NOT_EXPR:
- return 2;
+ if (TREE_CODE (phi) != PHI_NODE)
+ return false;
- default:
- return INT_MAX;
+ if (!vec_stmt) /* transformation not required. */
+ {
+ STMT_VINFO_TYPE (stmt_info) = induc_vec_info_type;
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "=== vectorizable_induction ===");
+ vect_model_induction_cost (stmt_info, ncopies);
+ return true;
}
+
+ /** Transform. **/
+
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "transform induction phi.");
+
+ vec_def = get_initial_def_for_induction (phi);
+ *vec_stmt = SSA_NAME_DEF_STMT (vec_def);
+ return true;
}
int icode;
enum machine_mode optab_op2_mode;
tree def, def_stmt;
- enum vect_def_type dt0, dt1;
+ enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
tree new_stmt;
stmt_vec_info prev_stmt_info;
int nunits_in = TYPE_VECTOR_SUBPARTS (vectype);
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_DEF_TYPE (stmt_info) != vect_loop_def)
+ return false;
+ /* FORNOW: not yet supported. */
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) != MODIFY_EXPR)
+ /* Is STMT a vectorizable binary/unary operation? */
+ if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
return false;
- if (TREE_CODE (TREE_OPERAND (stmt, 0)) != SSA_NAME)
+ if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) != SSA_NAME)
return false;
- scalar_dest = TREE_OPERAND (stmt, 0);
+ 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 = TREE_OPERAND (stmt, 1);
+ operation = GIMPLE_STMT_OPERAND (stmt, 1);
code = TREE_CODE (operation);
+
+ /* For pointer addition, we should use the normal plus for
+ the vector addition. */
+ if (code == POINTER_PLUS_EXPR)
+ code = PLUS_EXPR;
+
optab = optab_for_tree_code (code, vectype);
/* Support only unary or binary operations. */
- op_type = TREE_CODE_LENGTH (code);
+ op_type = TREE_OPERAND_LENGTH (operation);
if (op_type != unary_op && op_type != binary_op)
{
if (vect_print_dump_info (REPORT_DETAILS))
}
op0 = TREE_OPERAND (operation, 0);
- if (!vect_is_simple_use (op0, loop_vinfo, &def_stmt, &def, &dt0))
+ if (!vect_is_simple_use (op0, loop_vinfo, &def_stmt, &def, &dt[0]))
{
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_is_simple_use (op1, loop_vinfo, &def_stmt, &def, &dt[1]))
{
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "use not simple.");
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))
+ || dt[1] == vect_constant_def
+ || dt[1] == vect_invariant_def))
{
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "operand mode requires invariant argument.");
if (!vec_stmt) /* transformation not required. */
{
STMT_VINFO_TYPE (stmt_info) = op_vec_info_type;
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "=== vectorizable_operation ===");
+ vect_model_simple_cost (stmt_info, ncopies, dt);
return true;
}
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_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
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++)
{
}
else
{
- vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt0, vec_oprnd0);
+ vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
if (op_type == binary_op)
- vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt1, vec_oprnd1);
+ vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd1);
}
/* Arguments are ready. create the new vector stmt. */
-
+
if (op_type == binary_op)
- new_stmt = build2 (MODIFY_EXPR, void_type_node, vec_dest,
+ new_stmt = build_gimple_modify_stmt (vec_dest,
build2 (code, vectype, vec_oprnd0, vec_oprnd1));
else
- new_stmt = build2 (MODIFY_EXPR, void_type_node, vec_dest,
+ new_stmt = build_gimple_modify_stmt (vec_dest,
build1 (code, vectype, vec_oprnd0));
new_temp = make_ssa_name (vec_dest, new_stmt);
- TREE_OPERAND (new_stmt, 0) = new_temp;
+ 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
/* 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_stmt)
{
tree vec_dest;
tree scalar_dest;
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;
+ enum tree_code code, code1 = ERROR_MARK;
tree new_temp;
tree def, def_stmt;
- enum vect_def_type dt0;
+ enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
tree new_stmt;
stmt_vec_info prev_stmt_info;
int nunits_in;
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_DEF_TYPE (stmt_info) != vect_loop_def)
+ return false;
+
+ /* FORNOW: not yet supported. */
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) != MODIFY_EXPR)
+
+ /* Is STMT a vectorizable type-demotion operation? */
+ if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
return false;
-
- if (TREE_CODE (TREE_OPERAND (stmt, 0)) != SSA_NAME)
+
+ if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) != SSA_NAME)
return false;
-
- operation = TREE_OPERAND (stmt, 1);
+
+ 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 = TREE_OPERAND (stmt, 0);
- scalar_type = TREE_TYPE (scalar_dest);
- vectype_out = get_vectype_for_scalar_type (scalar_type);
+
+ 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_in != nunits_out / 2) /* FORNOW */
return false;
-
+
ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
gcc_assert (ncopies >= 1);
-
+
+ if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest))
+ && INTEGRAL_TYPE_P (TREE_TYPE (op0)))
+ || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest))
+ && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0))
+ && (code == NOP_EXPR || code == CONVERT_EXPR))))
+ return false;
+
/* Check the operands of the operation. */
- if (!vect_is_simple_use (op0, loop_vinfo, &def_stmt, &def, &dt0))
+ if (!vect_is_simple_use (op0, loop_vinfo, &def_stmt, &def, &dt[0]))
{
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)
+ if (!supportable_narrowing_operation (code, stmt, vectype_in, &code1))
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;
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "=== vectorizable_demotion ===");
+ vect_model_simple_cost (stmt_info, ncopies, dt);
return true;
}
-
+
/** Transform. **/
-
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "transform type demotion operation. ncopies = %d.",
- ncopies);
-
+ ncopies);
+
/* Handle def. */
vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
/* 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);
+ vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], 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);
+ vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd1);
+ vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
}
-
+
/* Arguments are ready. Create the new vector stmt. */
- expr = build2 (code, vectype_out, vec_oprnd0, vec_oprnd1);
- new_stmt = build2 (MODIFY_EXPR, void_type_node, vec_dest, expr);
+ expr = build2 (code1, vectype_out, vec_oprnd0, vec_oprnd1);
+ new_stmt = build_gimple_modify_stmt (vec_dest, expr);
new_temp = make_ssa_name (vec_dest, new_stmt);
- TREE_OPERAND (new_stmt, 0) = new_temp;
+ 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 (MODIFY_EXPR, void_type_node, vec_dest, expr);
- new_temp = make_ssa_name (vec_dest, new_stmt);
- TREE_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
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;
+ enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK;
tree decl1 = NULL_TREE, decl2 = NULL_TREE;
int op_type;
tree def, def_stmt;
- enum vect_def_type dt0, dt1;
+ enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
tree new_stmt;
stmt_vec_info prev_stmt_info;
int nunits_in;
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_DEF_TYPE (stmt_info) != vect_loop_def)
+ return false;
+ /* FORNOW: not yet supported. */
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.");
+ fprintf (vect_dump, "value used after loop.");
return false;
}
- if (TREE_CODE (stmt) != MODIFY_EXPR)
+ /* Is STMT a vectorizable type-promotion operation? */
+ if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
return false;
- if (TREE_CODE (TREE_OPERAND (stmt, 0)) != SSA_NAME)
+ if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) != SSA_NAME)
return false;
- operation = TREE_OPERAND (stmt, 1);
+ operation = GIMPLE_STMT_OPERAND (stmt, 1);
code = TREE_CODE (operation);
- if (code != NOP_EXPR && code != WIDEN_MULT_EXPR)
+ if (code != NOP_EXPR && code != CONVERT_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 = TREE_OPERAND (stmt, 0);
+ 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;
+ ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
+ gcc_assert (ncopies >= 1);
+
+ if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest))
+ && INTEGRAL_TYPE_P (TREE_TYPE (op0)))
+ || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest))
+ && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0))
+ && (code == CONVERT_EXPR || code == NOP_EXPR))))
+ return false;
+
/* Check the operands of the operation. */
- if (!vect_is_simple_use (op0, loop_vinfo, &def_stmt, &def, &dt0))
+ if (!vect_is_simple_use (op0, loop_vinfo, &def_stmt, &def, &dt[0]))
{
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "use not simple.");
if (op_type == binary_op)
{
op1 = TREE_OPERAND (operation, 1);
- if (!vect_is_simple_use (op1, loop_vinfo, &def_stmt, &def, &dt1))
+ if (!vect_is_simple_use (op1, loop_vinfo, &def_stmt, &def, &dt[1]))
{
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "use not simple.");
if (!vec_stmt) /* transformation not required. */
{
STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type;
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "=== vectorizable_promotion ===");
+ vect_model_simple_cost (stmt_info, 2*ncopies, dt);
return true;
}
}
else
{
- vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt0, vec_oprnd0);
+ vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
if (op_type == binary_op)
- vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt1, vec_oprnd1);
+ vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd1);
}
/* Arguments are ready. Create the new vector stmt. We are creating
}
+/* 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, tmp;
+ 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)
+ tmp = build2 (VEC_INTERLEAVE_HIGH_EXPR, vectype, vect1, vect2);
+ else
+ tmp = build2 (VEC_INTERLEAVE_LOW_EXPR, vectype, vect1, vect2);
+ perm_stmt = build_gimple_modify_stmt (perm_dest, tmp);
+ 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)
+ tmp = build2 (VEC_INTERLEAVE_LOW_EXPR, vectype, vect1, vect2);
+ else
+ tmp = build2 (VEC_INTERLEAVE_HIGH_EXPR, vectype, vect1, vect2);
+ perm_stmt = build_gimple_modify_stmt (perm_dest, tmp);
+ 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
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);
+ 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;
def_operand_p def_p;
tree def, def_stmt;
enum vect_def_type dt;
- stmt_vec_info prev_stmt_info;
+ 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);
+ if (!STMT_VINFO_RELEVANT_P (stmt_info))
+ return false;
+
+ if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
+ return false;
+
+ if (STMT_VINFO_LIVE_P (stmt_info))
+ {
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "value used after loop.");
+ return false;
+ }
+
/* Is vectorizable store? */
- if (TREE_CODE (stmt) != MODIFY_EXPR)
+ if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
return false;
- scalar_dest = TREE_OPERAND (stmt, 0);
+ scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
if (TREE_CODE (scalar_dest) != ARRAY_REF
- && TREE_CODE (scalar_dest) != INDIRECT_REF)
+ && TREE_CODE (scalar_dest) != INDIRECT_REF
+ && !DR_GROUP_FIRST_DR (stmt_info))
return false;
- op = TREE_OPERAND (stmt, 1);
+ 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))
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;
+ vect_model_store_cost (stmt_info, ncopies, dt);
return true;
}
/** Transform. **/
+ 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;
+ }
+
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "transform store. ncopies = %d",ncopies);
- alignment_support_cheme = vect_supportable_dr_alignment (dr);
+ 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 */
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++)
{
if (j == 0)
{
- vec_oprnd = vect_get_vec_def_for_operand (op, stmt, NULL);
- dataref_ptr = vect_create_data_ref_ptr (stmt, bsi, NULL_TREE, &dummy,
- &ptr_incr, false);
+ /* 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
{
- vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, vec_oprnd);
+ /* 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);
}
- /* Arguments are ready. create the new vector stmt. */
- data_ref = build_fold_indirect_ref (dataref_ptr);
- new_stmt = build2 (MODIFY_EXPR, vectype, data_ref, vec_oprnd);
- vect_finish_stmt_generation (stmt, new_stmt, bsi);
+ 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;
+ }
- /* Set the V_MAY_DEFS 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, stmt);
- if (j == 0)
+ next_stmt = first_stmt;
+ for (i = 0; i < group_size; i++)
{
- /* The original store is deleted so the same SSA_NAMEs can be used.
- */
- FOR_EACH_SSA_TREE_OPERAND (def, stmt, iter, SSA_OP_VMAYDEF)
+ /* 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 = build_gimple_modify_stmt (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)
{
- SSA_NAME_DEF_STMT (def) = new_stmt;
- mark_sym_for_renaming (SSA_NAME_VAR (def));
+ /* 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;
}
-
- 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_VMAYDEF)
+ else
{
- 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)));
+ /* 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;
}
- STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
- }
-
- prev_stmt_info = vinfo_for_stmt (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;
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.
-*/
+ Return value - the result of the loop-header phi node. */
static tree
vect_setup_realignment (tree stmt, block_stmt_iterator *bsi,
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 = TREE_OPERAND (stmt, 0);
+ tree scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
tree vec_dest;
tree init_addr;
tree inc;
/* 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);
+ 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 (MODIFY_EXPR, vectype, vec_dest, data_ref);
+ new_stmt = build_gimple_modify_stmt (vec_dest, data_ref);
new_temp = make_ssa_name (vec_dest, new_stmt);
- TREE_OPERAND (new_stmt, 0) = new_temp;
+ GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
new_bb = bsi_insert_on_edge_immediate (pe, new_stmt);
gcc_assert (!new_bb);
- msq_init = TREE_OPERAND (new_stmt, 0);
+ msq_init = GIMPLE_STMT_OPERAND (new_stmt, 0);
copy_virtual_operands (new_stmt, stmt);
update_vuses_to_preheader (new_stmt, loop);
if (targetm.vectorize.builtin_mask_for_load)
{
tree builtin_decl;
- tree params = build_tree_list (NULL_TREE, init_addr);
- vec_dest = vect_create_destination_var (scalar_dest,
- TREE_TYPE (new_stmt));
builtin_decl = targetm.vectorize.builtin_mask_for_load ();
- new_stmt = build_function_call_expr (builtin_decl, params);
- new_stmt = build2 (MODIFY_EXPR, void_type_node, vec_dest, new_stmt);
+ new_stmt = build_call_expr (builtin_decl, 1, init_addr);
+ vec_dest = vect_create_destination_var (scalar_dest,
+ TREE_TYPE (new_stmt));
+ new_stmt = build_gimple_modify_stmt (vec_dest, new_stmt);
new_temp = make_ssa_name (vec_dest, new_stmt);
- TREE_OPERAND (new_stmt, 0) = new_temp;
+ GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
new_bb = bsi_insert_on_edge_immediate (pe, new_stmt);
gcc_assert (!new_bb);
- *realignment_token = TREE_OPERAND (new_stmt, 0);
+ *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
}
+/* 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));
+ tree tmp;
+ 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);
+
+ tmp = build2 (VEC_EXTRACT_EVEN_EXPR, vectype,
+ first_vect, second_vect);
+ perm_stmt = build_gimple_modify_stmt (perm_dest, tmp);
+
+ 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);
+
+ tmp = build2 (VEC_EXTRACT_ODD_EXPR, vectype,
+ first_vect, second_vect);
+ perm_stmt = build_gimple_modify_stmt (perm_dest, tmp);
+ 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
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);
+ 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;
+ 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 j;
+ 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_DEF_TYPE (stmt_info) != vect_loop_def)
+ return false;
+ /* FORNOW: not yet supported. */
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) != MODIFY_EXPR)
+ /* Is vectorizable load? */
+ if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
return false;
- scalar_dest = TREE_OPERAND (stmt, 0);
+ scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
if (TREE_CODE (scalar_dest) != SSA_NAME)
return false;
- op = TREE_OPERAND (stmt, 1);
- if (TREE_CODE (op) != ARRAY_REF && TREE_CODE (op) != INDIRECT_REF)
+ 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;
}
+ /* 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;
+ vect_model_load_cost (stmt_info, ncopies);
return true;
}
- /** Transform. **/
-
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "transform load.");
- alignment_support_cheme = vect_supportable_dr_alignment (dr);
+ /** Transform. **/
+
+ 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
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:
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 (stmt, bsi, &realignment_token);
+ msq = vect_setup_realignment (first_stmt, bsi, &realignment_token);
phi_stmt = SSA_NAME_DEF_STMT (msq);
offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1);
}
{
/* 1. Create the vector pointer update chain. */
if (j == 0)
- dataref_ptr = vect_create_data_ref_ptr (stmt, bsi, offset,
- &dummy, &ptr_incr, false);
+ 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);
- /* 2. Create the vector-load in the loop. */
- switch (alignment_support_cheme)
- {
- case dr_aligned:
- gcc_assert (aligned_access_p (dr));
- data_ref = build_fold_indirect_ref (dataref_ptr);
- break;
- case dr_unaligned_supported:
- {
- int mis = DR_MISALIGNMENT (dr);
- tree tmis = (mis == -1 ? size_zero_node : size_int (mis));
-
- gcc_assert (!aligned_access_p (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 (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 (MODIFY_EXPR, vectype, vec_dest, data_ref);
- new_temp = make_ssa_name (vec_dest, new_stmt);
- TREE_OPERAND (new_stmt, 0) = new_temp;
- vect_finish_stmt_generation (stmt, new_stmt, bsi);
- copy_virtual_operands (new_stmt, stmt);
- mark_new_vars_to_rename (new_stmt);
-
- /* 3. Handle explicit realignment if necessary/supported. */
- if (alignment_support_cheme == dr_unaligned_software_pipeline)
- {
- /* Create in loop:
- <vec_dest = realign_load (msq, lsq, realignment_token)> */
- lsq = TREE_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 (MODIFY_EXPR, vectype, vec_dest, new_stmt);
- new_temp = make_ssa_name (vec_dest, new_stmt);
- TREE_OPERAND (new_stmt, 0) = new_temp;
- vect_finish_stmt_generation (stmt, new_stmt, bsi);
- if (j == ncopies - 1)
- add_phi_arg (phi_stmt, lsq, loop_latch_edge (loop));
- msq = lsq;
- }
+ 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 = build_gimple_modify_stmt (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:
+ <vec_dest = realign_load (msq, lsq, realignment_token)> */
+ 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 = build_gimple_modify_stmt (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 (j == 0)
- STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_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
- STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
- prev_stmt_info = vinfo_for_stmt (new_stmt);
+ {
+ 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;
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;
+ gcc_assert (STMT_VINFO_LIVE_P (stmt_info));
- if (TREE_CODE (stmt) != MODIFY_EXPR)
+ if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)
return false;
- if (TREE_CODE (TREE_OPERAND (stmt, 0)) != SSA_NAME)
+ if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
return false;
- operation = TREE_OPERAND (stmt, 1);
- code = TREE_CODE (operation);
+ if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) != SSA_NAME)
+ return false;
- op_type = TREE_CODE_LENGTH (code);
+ operation = GIMPLE_STMT_OPERAND (stmt, 1);
+ op_type = TREE_OPERAND_LENGTH (operation);
/* FORNOW: support only if all uses are invariant. This means
that the scalar operations can remain in place, unvectorized.
if (!STMT_VINFO_RELEVANT_P (stmt_info))
return false;
- gcc_assert (STMT_VINFO_DEF_TYPE (stmt_info) == vect_loop_def);
+ if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
+ return false;
+ /* FORNOW: not yet supported. */
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) != MODIFY_EXPR)
+ /* Is vectorizable conditional operation? */
+ if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
return false;
- op = TREE_OPERAND (stmt, 1);
+ op = GIMPLE_STMT_OPERAND (stmt, 1);
if (TREE_CODE (op) != COND_EXPR)
return false;
/* Transform */
/* Handle def. */
- scalar_dest = TREE_OPERAND (stmt, 0);
+ scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
vec_dest = vect_create_destination_var (scalar_dest, vectype);
/* Handle cond expr. */
vec_cond_expr = build3 (VEC_COND_EXPR, vectype,
vec_compare, vec_then_clause, vec_else_clause);
- *vec_stmt = build2 (MODIFY_EXPR, vectype, vec_dest, vec_cond_expr);
+ *vec_stmt = build_gimple_modify_stmt (vec_dest, vec_cond_expr);
new_temp = make_ssa_name (vec_dest, *vec_stmt);
- TREE_OPERAND (*vec_stmt, 0) = new_temp;
+ GIMPLE_STMT_OPERAND (*vec_stmt, 0) = new_temp;
vect_finish_stmt_generation (stmt, *vec_stmt, bsi);
return true;
Create a vectorized stmt to replace STMT, and insert it at BSI. */
bool
-vect_transform_stmt (tree stmt, block_stmt_iterator *bsi)
+vect_transform_stmt (tree stmt, block_stmt_iterator *bsi, bool *strided_store)
{
bool is_store = false;
tree vec_stmt = NULL_TREE;
tree orig_stmt_in_pattern;
bool done;
- if (STMT_VINFO_RELEVANT_P (stmt_info))
+ switch (STMT_VINFO_TYPE (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);
+ 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 type_conversion_vec_info_type:
+ done = vectorizable_conversion (stmt, bsi, &vec_stmt);
+ gcc_assert (done);
+ break;
+
+ case induc_vec_info_type:
+ done = vectorizable_induction (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;
+ break;
- default:
- if (vect_print_dump_info (REPORT_DETAILS))
- fprintf (vect_dump, "stmt not supported.");
- gcc_unreachable ();
- }
+ case condition_vec_info_type:
+ done = vectorizable_condition (stmt, bsi, &vec_stmt);
+ gcc_assert (done);
+ break;
- gcc_assert (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);
+ case call_vec_info_type:
+ done = vectorizable_call (stmt, bsi, &vec_stmt);
+ break;
- /* 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
- detail in the documentation of vect_pattern_recog. */
+ case reduc_vec_info_type:
+ done = vectorizable_reduction (stmt, bsi, &vec_stmt);
+ gcc_assert (done);
+ break;
- STMT_VINFO_VEC_STMT (stmt_vinfo) = vec_stmt;
- }
- }
+ default:
+ if (!STMT_VINFO_LIVE_P (stmt_info))
+ {
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "stmt not supported.");
+ gcc_unreachable ();
+ }
}
- if (STMT_VINFO_LIVE_P (stmt_info))
+ if (STMT_VINFO_LIVE_P (stmt_info)
+ && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type)
{
- switch (STMT_VINFO_TYPE (stmt_info))
- {
- case reduc_vec_info_type:
- done = vectorizable_reduction (stmt, bsi, &vec_stmt);
- gcc_assert (done);
- break;
+ done = vectorizable_live_operation (stmt, bsi, &vec_stmt);
+ gcc_assert (done);
+ }
- default:
- done = vectorizable_live_operation (stmt, bsi, &vec_stmt);
- gcc_assert (done);
- }
+ 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);
+ /* 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. */
+ if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo))
+ {
+ gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo) == stmt);
+ STMT_VINFO_VEC_STMT (stmt_vinfo) = vec_stmt;
+ }
+ }
}
return is_store;
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 V_MAY_DEF in another statement in a loop.
+ 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 V_MAY_DEF is at the dereference of a different
+ 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
tree phi;
bool updated = false;
- for (phi = phi_nodes (header_bb); phi; phi = TREE_CHAIN (phi))
+ for (phi = phi_nodes (header_bb); phi; phi = PHI_CHAIN (phi))
{
if (SSA_NAME_VAR (PHI_RESULT (phi)) == name_var)
{
tree evolution_part;
tree init_expr;
tree step_expr;
- tree var, stmt, ni, ni_name;
+ tree var, ni, ni_name;
block_stmt_iterator last_bsi;
if (vect_print_dump_info (REPORT_DETAILS))
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 (niters),
- niters, step_expr), init_expr);
+ if (POINTER_TYPE_P (TREE_TYPE (init_expr)))
+ ni = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (init_expr),
+ init_expr,
+ fold_convert (sizetype,
+ fold_build2 (MULT_EXPR, TREE_TYPE (niters),
+ niters, step_expr)));
+ else
+ 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);
-
+ ni_name = force_gimple_operand_bsi (&last_bsi, ni, false, var,
+ true, BSI_SAME_STMT);
+
/* Fix phi expressions in the successor bb. */
SET_PHI_ARG_DEF (phi1, update_e->dest_idx, ni_name);
}
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,
- struct loops *loops)
+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);
edge update_e;
basic_block preheader;
int loop_num;
+ unsigned int th;
+ int min_scalar_loop_bound;
+ int min_profitable_iters;
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "=== vect_do_peeling_for_loop_bound ===");
&ratio_mult_vf_name, ratio);
loop_num = loop->num;
- new_loop = slpeel_tree_peel_loop_to_edge (loop, loops, single_exit (loop),
- ratio_mult_vf_name, ni_name, false);
+
+ /* Analyze cost to set threshhold for vectorized loop. */
+ min_profitable_iters = LOOP_VINFO_COST_MODEL_MIN_ITERS (loop_vinfo);
+ min_scalar_loop_bound = (PARAM_VALUE (PARAM_MIN_VECT_LOOP_BOUND))
+ * LOOP_VINFO_VECT_FACTOR (loop_vinfo);
+
+ /* Use the cost model only if it is more conservative than user specified
+ threshold. */
+
+ th = (unsigned) min_scalar_loop_bound;
+ if (min_profitable_iters
+ && (!min_scalar_loop_bound
+ || min_profitable_iters > min_scalar_loop_bound))
+ th = (unsigned) min_profitable_iters;
+
+ if (min_profitable_iters
+ && !LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
+ && vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "vectorization may not be profitable.");
+
+ 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
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) */
+ 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.
+
+ The above formulas assume that VF == number of elements in the vector. This
+ may not hold when there are multiple-types in the loop.
+ In this case, for some data-references in the loop the VF does not represent
+ the number of elements that fit in the vector. Therefore, instead of VF we
+ use TYPE_VECTOR_SUBPARTS. */
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;
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))));
+ int nelements = TYPE_VECTOR_SUBPARTS (vectype);
+
+ 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 element_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dr))));
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-1));
+ iters = build_int_cst (niters_type,
+ (nelements - elem_misalign)&(nelements/group_size-1));
}
else
{
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);
+ build_int_cst (type, exact_log2 (vectype_align/nelements));
+ tree nelements_minus_1 = build_int_cst (type, nelements - 1);
+ tree nelements_tree = build_int_cst (type, nelements);
tree byte_misalign;
tree elem_misalign;
/* Create: byte_misalign = addr & (vectype_size - 1) */
byte_misalign =
- fold_build2 (BIT_AND_EXPR, type, start_addr, vectype_size_minus_1);
+ fold_build2 (BIT_AND_EXPR, type, fold_convert (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);
+ /* Create: (niters_type) (nelements - elem_misalign)&(nelements - 1) */
+ iters = fold_build2 (MINUS_EXPR, type, nelements_tree, elem_misalign);
+ iters = fold_build2 (BIT_AND_EXPR, type, iters, nelements_minus_1);
iters = fold_convert (niters_type, iters);
}
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. */
+ 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)
VEC (data_reference_p, heap) *datarefs = LOOP_VINFO_DATAREFS (loop_vinfo);
struct data_reference *dr;
- if (vect_dump && (dump_flags & TDF_DETAILS))
+ if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "=== vect_update_inits_of_dr ===");
for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++)
peeling is recorded in LOOP_VINFO_UNALIGNED_DR. */
static void
-vect_do_peeling_for_alignment (loop_vec_info loop_vinfo, struct loops *loops)
+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;
/* Peel the prolog loop and iterate it niters_of_prolog_loop. */
new_loop =
- slpeel_tree_peel_loop_to_edge (loop, loops, loop_preheader_edge (loop),
- niters_of_prolog_loop, ni_name, true);
+ 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);
{
VEC(tree,heap) *may_misalign_stmts
= LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo);
- tree ref_stmt;
+ tree ref_stmt, tmp;
int mask = LOOP_VINFO_PTR_MASK (loop_vinfo);
tree mask_cst;
unsigned int i;
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 (MODIFY_EXPR, void_type_node,
- addr_tmp_name, addr_stmt);
+ addr_stmt = build_gimple_modify_stmt (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);
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 (MODIFY_EXPR, void_type_node, new_or_tmp_name,
- build2 (BIT_IOR_EXPR, int_ptrsize_type,
- or_tmp_name,
- addr_tmp_name));
+ tmp = build2 (BIT_IOR_EXPR, int_ptrsize_type,
+ or_tmp_name, addr_tmp_name);
+ or_stmt = build_gimple_modify_stmt (new_or_tmp_name, tmp);
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;
add_referenced_var (and_tmp);
and_tmp_name = make_ssa_name (and_tmp, NULL_TREE);
- and_stmt = build2 (MODIFY_EXPR, void_type_node,
- and_tmp_name,
- build2 (BIT_AND_EXPR, int_ptrsize_type,
- or_tmp_name, mask_cst));
+ tmp = build2 (BIT_AND_EXPR, int_ptrsize_type, or_tmp_name, mask_cst);
+ and_stmt = build_gimple_modify_stmt (and_tmp_name, tmp);
SSA_NAME_DEF_STMT (and_tmp_name) = and_stmt;
append_to_statement_list_force (and_stmt, cond_expr_stmt_list);
stmts in the loop, and update the loop exit condition. */
void
-vect_transform_loop (loop_vec_info loop_vinfo,
- struct loops *loops ATTRIBUTE_UNUSED)
+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;
+ block_stmt_iterator si, next_si;
int i;
tree ratio = NULL;
int vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
- bitmap_iterator bi;
- unsigned int j;
+ bool strided_store;
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "=== vec_transform_loop ===");
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 (loops, loop, cond_expr, &condition_bb, true);
+ 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
/* CHECKME: we wouldn't need this if we called update_ssa once
for all loops. */
- bitmap_zero (vect_vnames_to_rename);
+ 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, loops);
+ 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
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, loops);
+ 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);
for (i = 0; i < nbbs; i++)
{
basic_block bb = bbs[i];
+ stmt_vec_info stmt_info;
+ tree phi;
+
+ for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
+ {
+ if (vect_print_dump_info (REPORT_DETAILS))
+ {
+ fprintf (vect_dump, "------>vectorizing phi: ");
+ print_generic_expr (vect_dump, phi, TDF_SLIM);
+ }
+ stmt_info = vinfo_for_stmt (phi);
+ if (!stmt_info)
+ continue;
+ if (!STMT_VINFO_RELEVANT_P (stmt_info)
+ && !STMT_VINFO_LIVE_P (stmt_info))
+ 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.");
+
+ if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_induction_def)
+ {
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "transform phi.");
+ vect_transform_stmt (phi, NULL, NULL);
+ }
+ }
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))
continue;
}
+ gcc_assert (STMT_VINFO_VECTYPE (stmt_info));
if ((TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info))
!= (unsigned HOST_WIDE_INT) vectorization_factor)
&& vect_print_dump_info (REPORT_DETAILS))
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "transform statement.");
- is_store = vect_transform_stmt (stmt, &si);
- if (is_store)
- {
- /* Free the attached stmt_vec_info and remove the stmt. */
- stmt_ann_t ann = stmt_ann (stmt);
- free (stmt_info);
- set_stmt_info (ann, NULL);
- bsi_remove (&si, true);
- continue;
+ 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_si = bsi_for_stmt (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);
+ bsi_remove (&next_si, true);
+ 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;
+ }
}
-
bsi_next (&si);
} /* stmts in BB */
} /* BBs in loop */
slpeel_make_loop_iterate_ntimes (loop, ratio);
- EXECUTE_IF_SET_IN_BITMAP (vect_vnames_to_rename, 0, j, bi)
- mark_sym_for_renaming (SSA_NAME_VAR (ssa_name (j)));
+ 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