Google Git
Sign in
gnu / gcc / cd99b30c7cf1e68ef76ed2af7a41ac74c78116fe / . / gcc / tree-vect-transform.c
blob: 294af740d53d55aea372e73f73824e4372326df9 [file] [log] [blame]
/* Transformation Utilities for Loop Vectorization.
Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009
Free Software Foundation, Inc.
Contributed by Dorit Naishlos <dorit@il.ibm.com>
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify it under
the terms of the GNU General Public License as published by the Free
Software Foundation; either version 3, or (at your option) any later
version.
GCC is distributed in the hope that it will be useful, but WITHOUT ANY
WARRANTY; without even the implied warranty of MERCHANTABILITY or
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
for more details.
You should have received a copy of the GNU General Public License
along with GCC; see the file COPYING3. If not see
<http://www.gnu.org/licenses/>. */
#include "config.h"
#include "system.h"
#include "coretypes.h"
#include "tm.h"
#include "ggc.h"
#include "tree.h"
#include "target.h"
#include "rtl.h"
#include "basic-block.h"
#include "diagnostic.h"
#include "tree-flow.h"
#include "tree-dump.h"
#include "timevar.h"
#include "cfgloop.h"
#include "expr.h"
#include "optabs.h"
#include "params.h"
#include "recog.h"
#include "tree-data-ref.h"
#include "tree-chrec.h"
#include "tree-scalar-evolution.h"
#include "tree-vectorizer.h"
#include "langhooks.h"
#include "tree-pass.h"
#include "toplev.h"
#include "real.h"
/* Utility functions for the code transformation. */
static bool vect_transform_stmt (gimple, gimple_stmt_iterator *, bool *,
slp_tree, slp_instance);
static tree vect_create_destination_var (tree, tree);
static tree vect_create_data_ref_ptr
(gimple, struct loop*, tree, tree *, gimple *, bool, bool *, tree);
static tree vect_create_addr_base_for_vector_ref
(gimple, gimple_seq *, tree, struct loop *);
static tree vect_get_new_vect_var (tree, enum vect_var_kind, const char *);
static tree vect_get_vec_def_for_operand (tree, gimple, tree *);
static tree vect_init_vector (gimple, tree, tree, gimple_stmt_iterator *);
static void vect_finish_stmt_generation
(gimple stmt, gimple vec_stmt, gimple_stmt_iterator *);
static bool vect_is_simple_cond (tree, loop_vec_info);
static void vect_create_epilog_for_reduction
(tree, gimple, int, enum tree_code, gimple);
static tree get_initial_def_for_reduction (gimple, tree, tree *);
/* Utility function dealing with loop peeling (not peeling itself). */
static void vect_generate_tmps_on_preheader
(loop_vec_info, tree *, tree *, tree *);
static tree vect_build_loop_niters (loop_vec_info);
static void vect_update_ivs_after_vectorizer (loop_vec_info, tree, edge);
static tree vect_gen_niters_for_prolog_loop (loop_vec_info, tree);
static void vect_update_init_of_dr (struct data_reference *, tree niters);
static void vect_update_inits_of_drs (loop_vec_info, tree);
static int vect_min_worthwhile_factor (enum tree_code);
static int
cost_for_stmt (gimple 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 scalar_outside_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 = LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo);
int peel_guard_costs = 0;
int innerloop_iters = 0, factor;
VEC (slp_instance, heap) *slp_instances;
slp_instance instance;
/* Cost model disabled. */
if (!flag_vect_cost_model)
{
if (vect_print_dump_info (REPORT_COST))
fprintf (vect_dump, "cost model disabled.");
return 0;
}
/* Requires loop versioning tests to handle misalignment. */
if (VEC_length (gimple, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo)))
{
/* FIXME: Make cost depend on complexity of individual check. */
vec_outside_cost +=
VEC_length (gimple, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo));
if (vect_print_dump_info (REPORT_COST))
fprintf (vect_dump, "cost model: Adding cost of checks for loop "
"versioning to treat misalignment.\n");
}
if (VEC_length (ddr_p, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo)))
{
/* FIXME: Make cost depend on complexity of individual check. */
vec_outside_cost +=
VEC_length (ddr_p, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo));
if (vect_print_dump_info (REPORT_COST))
fprintf (vect_dump, "cost model: Adding cost of checks for loop "
"versioning aliasing.\n");
}
if (VEC_length (gimple, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo))
|| VEC_length (ddr_p, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo)))
{
vec_outside_cost += TARG_COND_TAKEN_BRANCH_COST;
}
/* 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. */
/* FORNOW. */
if (loop->inner)
innerloop_iters = 50; /* FIXME */
for (i = 0; i < nbbs; i++)
{
gimple_stmt_iterator si;
basic_block bb = bbs[i];
if (bb->loop_father == loop->inner)
factor = innerloop_iters;
else
factor = 1;
for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
{
gimple stmt = gsi_stmt (si);
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
/* Skip stmts that are not vectorized inside the loop. */
if (!STMT_VINFO_RELEVANT_P (stmt_info)
&& (!STMT_VINFO_LIVE_P (stmt_info)
|| STMT_VINFO_DEF_TYPE (stmt_info) != vect_reduction_def))
continue;
scalar_single_iter_cost += cost_for_stmt (stmt) * factor;
vec_inside_cost += STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info) * factor;
/* FIXME: for stmts in the inner-loop in outer-loop vectorization,
some of the "outside" costs are generated inside the outer-loop. */
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 don't know the value of peel_iters for prologue or epilogue
at compile-time - we assume it's vf/2 (the worst would be vf-1).
TODO: Build an expression that represents peel_iters for prologue and
epilogue to be used in a run-time test. */
if (byte_misalign < 0)
{
peel_iters_prologue = vf/2;
if (vect_print_dump_info (REPORT_COST))
fprintf (vect_dump, "cost model: "
"prologue peel iters set to vf/2.");
/* If peeling for alignment is unknown, loop bound of main loop becomes
unknown. */
peel_iters_epilogue = vf/2;
if (vect_print_dump_info (REPORT_COST))
fprintf (vect_dump, "cost model: "
"epilogue peel iters set to vf/2 because "
"peeling for alignment is unknown .");
/* If peeled iterations are unknown, count a taken branch and a not taken
branch per peeled loop. Even if scalar loop iterations are known,
vector iterations are not known since peeled prologue iterations are
not known. Hence guards remain the same. */
peel_guard_costs += 2 * (TARG_COND_TAKEN_BRANCH_COST
+ TARG_COND_NOT_TAKEN_BRANCH_COST);
}
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/2;
if (vect_print_dump_info (REPORT_COST))
fprintf (vect_dump, "cost model: "
"epilogue peel iters set to vf/2 because "
"loop iterations are unknown .");
/* If peeled iterations are known but number of scalar loop
iterations are unknown, count a taken branch per peeled loop. */
peel_guard_costs += 2 * TARG_COND_TAKEN_BRANCH_COST;
}
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;
}
}
vec_outside_cost += (peel_iters_prologue * scalar_single_iter_cost)
+ (peel_iters_epilogue * scalar_single_iter_cost)
+ peel_guard_costs;
/* FORNOW: The scalar outside cost is incremented in one of the
following ways:
1. The vectorizer checks for alignment and aliasing and generates
a condition that allows dynamic vectorization. A cost model
check is ANDED with the versioning condition. Hence scalar code
path now has the added cost of the versioning check.
if (cost > th & versioning_check)
jmp to vector code
Hence run-time scalar is incremented by not-taken branch cost.
2. The vectorizer then checks if a prologue is required. If the
cost model check was not done before during versioning, it has to
be done before the prologue check.
if (cost <= th)
prologue = scalar_iters
if (prologue == 0)
jmp to vector code
else
execute prologue
if (prologue == num_iters)
go to exit
Hence the run-time scalar cost is incremented by a taken branch,
plus a not-taken branch, plus a taken branch cost.
3. The vectorizer then checks if an epilogue is required. If the
cost model check was not done before during prologue check, it
has to be done with the epilogue check.
if (prologue == 0)
jmp to vector code
else
execute prologue
if (prologue == num_iters)
go to exit
vector code:
if ((cost <= th) | (scalar_iters-prologue-epilogue == 0))
jmp to epilogue
Hence the run-time scalar cost should be incremented by 2 taken
branches.
TODO: The back end may reorder the BBS's differently and reverse
conditions/branch directions. Change the estimates below to
something more reasonable. */
/* If the number of iterations is known and we do not do versioning, we can
decide whether to vectorize at compile time. Hence the scalar version
do not carry cost model guard costs. */
if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
|| VEC_length (gimple, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo))
|| VEC_length (ddr_p, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo)))
{
/* Cost model check occurs at versioning. */
if (VEC_length (gimple, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo))
|| VEC_length (ddr_p, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo)))
scalar_outside_cost += TARG_COND_NOT_TAKEN_BRANCH_COST;
else
{
/* Cost model check occurs at prologue generation. */
if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo) < 0)
scalar_outside_cost += 2 * TARG_COND_TAKEN_BRANCH_COST
+ TARG_COND_NOT_TAKEN_BRANCH_COST;
/* Cost model check occurs at epilogue generation. */
else
scalar_outside_cost += 2 * TARG_COND_TAKEN_BRANCH_COST;
}
}
/* Add SLP costs. */
slp_instances = LOOP_VINFO_SLP_INSTANCES (loop_vinfo);
for (i = 0; VEC_iterate (slp_instance, slp_instances, i, instance); i++)
{
vec_outside_cost += SLP_INSTANCE_OUTSIDE_OF_LOOP_COST (instance);
vec_inside_cost += SLP_INSTANCE_INSIDE_OF_LOOP_COST (instance);
}
/* Calculate number of iterations required to make the vector version
profitable, relative to the loop bodies only. The following condition
must hold true:
SIC * niters + SOC > VIC * ((niters-PL_ITERS-EP_ITERS)/VF) + VOC
where
SIC = scalar iteration cost, VIC = vector iteration cost,
VOC = vector outside cost, VF = vectorization factor,
PL_ITERS = prologue iterations, EP_ITERS= epilogue iterations
SOC = scalar outside cost for run time cost model check. */
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 - scalar_outside_cost) * vf
- vec_inside_cost * peel_iters_prologue
- vec_inside_cost * peel_iters_epilogue)
/ ((scalar_single_iter_cost * vf)
- vec_inside_cost);
if ((scalar_single_iter_cost * vf * min_profitable_iters)
<= ((vec_inside_cost * min_profitable_iters)
+ ((vec_outside_cost - scalar_outside_cost) * vf)))
min_profitable_iters++;
}
}
/* vector version will never be profitable. */
else
{
if (vect_print_dump_info (REPORT_COST))
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_COST))
{
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 iteration cost: %d\n",
scalar_single_iter_cost);
fprintf (vect_dump, " Scalar outside cost: %d\n", scalar_outside_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);
}
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--;
if (vect_print_dump_info (REPORT_COST))
fprintf (vect_dump, " Profitability threshold = %d\n",
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 bool
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;
gimple stmt, orig_stmt;
tree reduction_op;
enum machine_mode mode;
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
/* Cost of reduction op inside loop. */
STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info) += ncopies * TARG_VEC_STMT_COST;
stmt = STMT_VINFO_STMT (stmt_info);
switch (get_gimple_rhs_class (gimple_assign_rhs_code (stmt)))
{
case GIMPLE_SINGLE_RHS:
gcc_assert (TREE_OPERAND_LENGTH (gimple_assign_rhs1 (stmt)) == ternary_op);
reduction_op = TREE_OPERAND (gimple_assign_rhs1 (stmt), 2);
break;
case GIMPLE_UNARY_RHS:
reduction_op = gimple_assign_rhs1 (stmt);
break;
case GIMPLE_BINARY_RHS:
reduction_op = gimple_assign_rhs2 (stmt);
break;
default:
gcc_unreachable ();
}
vectype = get_vectype_for_scalar_type (TREE_TYPE (reduction_op));
if (!vectype)
{
if (vect_print_dump_info (REPORT_COST))
{
fprintf (vect_dump, "unsupported data-type ");
print_generic_expr (vect_dump, TREE_TYPE (reduction_op), TDF_SLIM);
}
return false;
}
mode = TYPE_MODE (vectype);
orig_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
if (!orig_stmt)
orig_stmt = STMT_VINFO_STMT (stmt_info);
code = gimple_assign_rhs_code (orig_stmt);
/* Add in cost for initial definition. */
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 (!nested_in_vect_loop_p (loop, orig_stmt))
{
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_assign_lhs (orig_stmt)));
int element_bitsize = tree_low_cst (bitsize, 1);
int nelements = vec_size_in_bits / element_bitsize;
optab = optab_for_tree_code (code, vectype, optab_default);
/* We have a whole vector shift available. */
if (VECTOR_MODE_P (mode)
&& optab_handler (optab, mode)->insn_code != CODE_FOR_nothing
&& optab_handler (vec_shr_optab, 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_COST))
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));
return true;
}
/* 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_COST))
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. */
void
vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies,
enum vect_def_type *dt, slp_tree slp_node)
{
int i;
int inside_cost = 0, outside_cost = 0;
/* The SLP costs were already calculated during SLP tree build. */
if (PURE_SLP_STMT (stmt_info))
return;
inside_cost = 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)
outside_cost += TARG_SCALAR_TO_VEC_COST;
}
if (vect_print_dump_info (REPORT_COST))
fprintf (vect_dump, "vect_model_simple_cost: inside_cost = %d, "
"outside_cost = %d .", inside_cost, outside_cost);
/* Set the costs either in STMT_INFO or SLP_NODE (if exists). */
stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost);
stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost);
}
/* 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)
{
gimple 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. */
void
vect_model_store_cost (stmt_vec_info stmt_info, int ncopies,
enum vect_def_type dt, slp_tree slp_node)
{
int group_size;
int inside_cost = 0, outside_cost = 0;
/* The SLP costs were already calculated during SLP tree build. */
if (PURE_SLP_STMT (stmt_info))
return;
if (dt == vect_constant_def || dt == vect_invariant_def)
outside_cost = TARG_SCALAR_TO_VEC_COST;
/* Strided access? */
if (DR_GROUP_FIRST_DR (stmt_info) && !slp_node)
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. */
inside_cost = ncopies * exact_log2(group_size) * group_size
* TARG_VEC_STMT_COST;
if (vect_print_dump_info (REPORT_COST))
fprintf (vect_dump, "vect_model_store_cost: strided group_size = %d .",
group_size);
}
/* Costs of the stores. */
inside_cost += ncopies * TARG_VEC_STORE_COST;
if (vect_print_dump_info (REPORT_COST))
fprintf (vect_dump, "vect_model_store_cost: inside_cost = %d, "
"outside_cost = %d .", inside_cost, outside_cost);
/* Set the costs either in STMT_INFO or SLP_NODE (if exists). */
stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost);
stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost);
}
/* 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. */
void
vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, slp_tree slp_node)
{
int group_size;
int alignment_support_cheme;
gimple first_stmt;
struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr;
int inside_cost = 0, outside_cost = 0;
/* The SLP costs were already calculated during SLP tree build. */
if (PURE_SLP_STMT (stmt_info))
return;
/* Strided accesses? */
first_stmt = DR_GROUP_FIRST_DR (stmt_info);
if (first_stmt && !slp_node)
{
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. */
inside_cost = ncopies * exact_log2(group_size) * group_size
* TARG_VEC_STMT_COST;
if (vect_print_dump_info (REPORT_COST))
fprintf (vect_dump, "vect_model_load_cost: strided group_size = %d .",
group_size);
}
/* The loads themselves. */
switch (alignment_support_cheme)
{
case dr_aligned:
{
inside_cost += ncopies * TARG_VEC_LOAD_COST;
if (vect_print_dump_info (REPORT_COST))
fprintf (vect_dump, "vect_model_load_cost: aligned.");
break;
}
case dr_unaligned_supported:
{
/* Here, we assign an additional cost for the unaligned load. */
inside_cost += ncopies * TARG_VEC_UNALIGNED_LOAD_COST;
if (vect_print_dump_info (REPORT_COST))
fprintf (vect_dump, "vect_model_load_cost: unaligned supported by "
"hardware.");
break;
}
case dr_explicit_realign:
{
inside_cost += ncopies * (2*TARG_VEC_LOAD_COST + TARG_VEC_STMT_COST);
/* FIXME: If the misalignment remains fixed across the iterations of
the containing loop, the following cost should be added to the
outside costs. */
if (targetm.vectorize.builtin_mask_for_load)
inside_cost += TARG_VEC_STMT_COST;
break;
}
case dr_explicit_realign_optimized:
{
if (vect_print_dump_info (REPORT_COST))
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 || slp_node)
{
outside_cost = 2*TARG_VEC_STMT_COST;
if (targetm.vectorize.builtin_mask_for_load)
outside_cost += TARG_VEC_STMT_COST;
}
inside_cost += ncopies * (TARG_VEC_LOAD_COST + TARG_VEC_STMT_COST);
break;
}
default:
gcc_unreachable ();
}
if (vect_print_dump_info (REPORT_COST))
fprintf (vect_dump, "vect_model_load_cost: inside_cost = %d, "
"outside_cost = %d .", inside_cost, outside_cost);
/* Set the costs either in STMT_INFO or SLP_NODE (if exists). */
stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost);
stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost);
}
/* Function vect_get_new_vect_var.
Returns a name for a new variable. The current naming scheme appends the
prefix "vect_" or "vect_p" (depending on the value of VAR_KIND) to
the name of vectorizer generated variables, and appends that to NAME if
provided. */
static tree
vect_get_new_vect_var (tree type, enum vect_var_kind var_kind, const char *name)
{
const char *prefix;
tree new_vect_var;
switch (var_kind)
{
case vect_simple_var:
prefix = "vect_";
break;
case vect_scalar_var:
prefix = "stmp_";
break;
case vect_pointer_var:
prefix = "vect_p";
break;
default:
gcc_unreachable ();
}
if (name)
{
char* tmp = concat (prefix, name, NULL);
new_vect_var = create_tmp_var (type, tmp);
free (tmp);
}
else
new_vect_var = create_tmp_var (type, prefix);
/* Mark vector typed variable as a gimple register variable. */
if (TREE_CODE (type) == VECTOR_TYPE)
DECL_GIMPLE_REG_P (new_vect_var) = true;
return new_vect_var;
}
/* Function vect_create_addr_base_for_vector_ref.
Create an expression that computes the address of the first memory location
that will be accessed for a data reference.
Input:
STMT: The statement containing the data reference.
NEW_STMT_LIST: Must be initialized to NULL_TREE or a statement list.
OFFSET: Optional. If supplied, it is be added to the initial address.
LOOP: Specify relative to which loop-nest should the address be computed.
For example, when the dataref is in an inner-loop nested in an
outer-loop that is now being vectorized, LOOP can be either the
outer-loop, or the inner-loop. The first memory location accessed
by the following dataref ('in' points to short):
for (i=0; i<N; i++)
for (j=0; j<M; j++)
s += in[i+j]
is as follows:
if LOOP=i_loop: &in (relative to i_loop)
if LOOP=j_loop: &in+i*2B (relative to j_loop)
Output:
1. Return an SSA_NAME whose value is the address of the memory location of
the first vector of the data reference.
2. If new_stmt_list is not NULL_TREE after return then the caller must insert
these statement(s) which define the returned SSA_NAME.
FORNOW: We are only handling array accesses with step 1. */
static tree
vect_create_addr_base_for_vector_ref (gimple stmt,
gimple_seq *new_stmt_list,
tree offset,
struct loop *loop)
{
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
struct loop *containing_loop = (gimple_bb (stmt))->loop_father;
tree data_ref_base = unshare_expr (DR_BASE_ADDRESS (dr));
tree base_name;
tree data_ref_base_var;
tree vec_stmt;
tree addr_base, addr_expr;
tree dest;
gimple_seq seq = NULL;
tree base_offset = unshare_expr (DR_OFFSET (dr));
tree init = unshare_expr (DR_INIT (dr));
tree vect_ptr_type, addr_expr2;
tree step = TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr)));
gcc_assert (loop);
if (loop != containing_loop)
{
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
gcc_assert (nested_in_vect_loop_p (loop, stmt));
data_ref_base = unshare_expr (STMT_VINFO_DR_BASE_ADDRESS (stmt_info));
base_offset = unshare_expr (STMT_VINFO_DR_OFFSET (stmt_info));
init = unshare_expr (STMT_VINFO_DR_INIT (stmt_info));
}
/* Create data_ref_base */
base_name = build_fold_indirect_ref (data_ref_base);
data_ref_base_var = create_tmp_var (TREE_TYPE (data_ref_base), "batmp");
add_referenced_var (data_ref_base_var);
data_ref_base = force_gimple_operand (data_ref_base, &seq, true,
data_ref_base_var);
gimple_seq_add_seq (new_stmt_list, seq);
/* Create base_offset */
base_offset = size_binop (PLUS_EXPR,
fold_convert (sizetype, base_offset),
fold_convert (sizetype, init));
dest = create_tmp_var (sizetype, "base_off");
add_referenced_var (dest);
base_offset = force_gimple_operand (base_offset, &seq, true, dest);
gimple_seq_add_seq (new_stmt_list, seq);
if (offset)
{
tree tmp = create_tmp_var (sizetype, "offset");
add_referenced_var (tmp);
offset = fold_build2 (MULT_EXPR, sizetype,
fold_convert (sizetype, offset), step);
base_offset = fold_build2 (PLUS_EXPR, sizetype,
base_offset, offset);
base_offset = force_gimple_operand (base_offset, &seq, false, tmp);
gimple_seq_add_seq (new_stmt_list, seq);
}
/* base + base_offset */
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 (vect_ptr_type, vect_pointer_var,
get_name (base_name));
add_referenced_var (addr_expr);
vec_stmt = fold_convert (vect_ptr_type, addr_base);
addr_expr2 = vect_get_new_vect_var (vect_ptr_type, vect_pointer_var,
get_name (base_name));
add_referenced_var (addr_expr2);
vec_stmt = force_gimple_operand (vec_stmt, &seq, false, addr_expr2);
gimple_seq_add_seq (new_stmt_list, seq);
if (vect_print_dump_info (REPORT_DETAILS))
{
fprintf (vect_dump, "created ");
print_generic_expr (vect_dump, vec_stmt, TDF_SLIM);
}
return vec_stmt;
}
/* Function vect_create_data_ref_ptr.
Create a new pointer to vector type (vp), that points to the first location
accessed in the loop by STMT, along with the def-use update chain to
appropriately advance the pointer through the loop iterations. Also set
aliasing information for the pointer. This vector pointer is used by the
callers to this function to create a memory reference expression for vector
load/store access.
Input:
1. STMT: a stmt that references memory. Expected to be of the form
GIMPLE_ASSIGN <name, data-ref> or
GIMPLE_ASSIGN <data-ref, name>.
2. AT_LOOP: the loop where the vector memref is to be created.
3. OFFSET (optional): an offset to be added to the initial address accessed
by the data-ref in STMT.
4. ONLY_INIT: indicate if vp is to be updated in the loop, or remain
pointing to the initial address.
5. TYPE: if not NULL indicates the required type of the data-ref.
Output:
1. Declare a new ptr to vector_type, and have it point to the base of the
data reference (initial addressed accessed by the data reference).
For example, for vector of type V8HI, the following code is generated:
v8hi *vp;
vp = (v8hi *)initial_address;
if OFFSET is not supplied:
initial_address = &a[init];
if OFFSET is supplied:
initial_address = &a[init + OFFSET];
Return the initial_address in INITIAL_ADDRESS.
2. If ONLY_INIT is true, just return the initial pointer. Otherwise, also
update the pointer in each iteration of the loop.
Return the increment stmt that updates the pointer in PTR_INCR.
3. Set INV_P to true if the access pattern of the data reference in the
vectorized loop is invariant. Set it to false otherwise.
4. Return the pointer. */
static tree
vect_create_data_ref_ptr (gimple stmt, struct loop *at_loop,
tree offset, tree *initial_address, gimple *ptr_incr,
bool only_init, bool *inv_p, tree type)
{
tree base_name;
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
bool nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt);
struct loop *containing_loop = (gimple_bb (stmt))->loop_father;
tree vectype = STMT_VINFO_VECTYPE (stmt_info);
tree vect_ptr_type;
tree vect_ptr;
tree tag;
tree new_temp;
gimple vec_stmt;
gimple_seq new_stmt_list = NULL;
edge pe;
basic_block new_bb;
tree vect_ptr_init;
struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
tree vptr;
gimple_stmt_iterator incr_gsi;
bool insert_after;
tree indx_before_incr, indx_after_incr;
gimple incr;
tree step;
/* Check the step (evolution) of the load in LOOP, and record
whether it's invariant. */
if (nested_in_vect_loop)
step = STMT_VINFO_DR_STEP (stmt_info);
else
step = DR_STEP (STMT_VINFO_DATA_REF (stmt_info));
if (tree_int_cst_compare (step, size_zero_node) == 0)
*inv_p = true;
else
*inv_p = false;
/* Create an expression for the first address accessed by this load
in LOOP. */
base_name = build_fold_indirect_ref (unshare_expr (DR_BASE_ADDRESS (dr)));
if (vect_print_dump_info (REPORT_DETAILS))
{
tree data_ref_base = base_name;
fprintf (vect_dump, "create vector-pointer variable to type: ");
print_generic_expr (vect_dump, vectype, TDF_SLIM);
if (TREE_CODE (data_ref_base) == VAR_DECL)
fprintf (vect_dump, " vectorizing a one dimensional array ref: ");
else if (TREE_CODE (data_ref_base) == ARRAY_REF)
fprintf (vect_dump, " vectorizing a multidimensional array ref: ");
else if (TREE_CODE (data_ref_base) == COMPONENT_REF)
fprintf (vect_dump, " vectorizing a record based array ref: ");
else if (TREE_CODE (data_ref_base) == SSA_NAME)
fprintf (vect_dump, " vectorizing a pointer ref: ");
print_generic_expr (vect_dump, base_name, TDF_SLIM);
}
/** (1) Create the new vector-pointer variable: **/
if (type)
vect_ptr_type = build_pointer_type (type);
else
vect_ptr_type = build_pointer_type (vectype);
if (TREE_CODE (DR_BASE_ADDRESS (dr)) == SSA_NAME
&& TYPE_RESTRICT (TREE_TYPE (DR_BASE_ADDRESS (dr))))
vect_ptr_type = build_qualified_type (vect_ptr_type, TYPE_QUAL_RESTRICT);
vect_ptr = vect_get_new_vect_var (vect_ptr_type, vect_pointer_var,
get_name (base_name));
if (TREE_CODE (DR_BASE_ADDRESS (dr)) == SSA_NAME
&& TYPE_RESTRICT (TREE_TYPE (DR_BASE_ADDRESS (dr))))
{
get_alias_set (base_name);
DECL_POINTER_ALIAS_SET (vect_ptr)
= DECL_POINTER_ALIAS_SET (SSA_NAME_VAR (DR_BASE_ADDRESS (dr)));
}
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_SYMBOL_TAG (dr);
gcc_assert (tag);
/* If tag is a variable (and NOT_A_TAG) than a new symbol memory
tag must be created with tag added to its may alias list. */
if (!MTAG_P (tag))
new_type_alias (vect_ptr, tag, DR_REF (dr));
else
{
set_symbol_mem_tag (vect_ptr, tag);
mark_sym_for_renaming (tag);
}
/** Note: If the dataref is in an inner-loop nested in LOOP, and we are
vectorizing LOOP (i.e. outer-loop vectorization), we need to create two
def-use update cycles for the pointer: One relative to the outer-loop
(LOOP), which is what steps (3) and (4) below do. The other is relative
to the inner-loop (which is the inner-most loop containing the dataref),
and this is done be step (5) below.
When vectorizing inner-most loops, the vectorized loop (LOOP) is also the
inner-most loop, and so steps (3),(4) work the same, and step (5) is
redundant. Steps (3),(4) create the following:
vp0 = &base_addr;
LOOP: vp1 = phi(vp0,vp2)
...
...
vp2 = vp1 + step
goto LOOP
If there is an inner-loop nested in loop, then step (5) will also be
applied, and an additional update in the inner-loop will be created:
vp0 = &base_addr;
LOOP: vp1 = phi(vp0,vp2)
...
inner: vp3 = phi(vp1,vp4)
vp4 = vp3 + inner_step
if () goto inner
...
vp2 = vp1 + step
if () goto LOOP */
/** (3) Calculate the initial address the vector-pointer, and set
the vector-pointer to point to it before the loop: **/
/* Create: (&(base[init_val+offset]) in the loop preheader. */
new_temp = vect_create_addr_base_for_vector_ref (stmt, &new_stmt_list,
offset, loop);
pe = loop_preheader_edge (loop);
if (new_stmt_list)
{
new_bb = gsi_insert_seq_on_edge_immediate (pe, new_stmt_list);
gcc_assert (!new_bb);
}
*initial_address = new_temp;
/* Create: p = (vectype *) initial_base */
vec_stmt = gimple_build_assign (vect_ptr,
fold_convert (vect_ptr_type, new_temp));
vect_ptr_init = make_ssa_name (vect_ptr, vec_stmt);
gimple_assign_set_lhs (vec_stmt, vect_ptr_init);
new_bb = gsi_insert_on_edge_immediate (pe, vec_stmt);
gcc_assert (!new_bb);
/** (4) Handle the updating of the vector-pointer inside the loop.
This is needed when ONLY_INIT is false, and also when AT_LOOP
is the inner-loop nested in LOOP (during outer-loop vectorization).
**/
if (only_init && at_loop == loop) /* No update in loop is required. */
{
/* Copy the points-to information if it exists. */
if (DR_PTR_INFO (dr))
duplicate_ssa_name_ptr_info (vect_ptr_init, DR_PTR_INFO (dr));
vptr = vect_ptr_init;
}
else
{
/* The step of the vector pointer is the Vector Size. */
tree step = TYPE_SIZE_UNIT (vectype);
/* One exception to the above is when the scalar step of the load in
LOOP is zero. In this case the step here is also zero. */
if (*inv_p)
step = size_zero_node;
standard_iv_increment_position (loop, &incr_gsi, &insert_after);
create_iv (vect_ptr_init,
fold_convert (vect_ptr_type, step),
vect_ptr, loop, &incr_gsi, insert_after,
&indx_before_incr, &indx_after_incr);
incr = gsi_stmt (incr_gsi);
set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo));
/* Copy the points-to information if it exists. */
if (DR_PTR_INFO (dr))
{
duplicate_ssa_name_ptr_info (indx_before_incr, DR_PTR_INFO (dr));
duplicate_ssa_name_ptr_info (indx_after_incr, DR_PTR_INFO (dr));
}
merge_alias_info (vect_ptr_init, indx_before_incr);
merge_alias_info (vect_ptr_init, indx_after_incr);
if (ptr_incr)
*ptr_incr = incr;
vptr = indx_before_incr;
}
if (!nested_in_vect_loop || only_init)
return vptr;
/** (5) Handle the updating of the vector-pointer inside the inner-loop
nested in LOOP, if exists: **/
gcc_assert (nested_in_vect_loop);
if (!only_init)
{
standard_iv_increment_position (containing_loop, &incr_gsi,
&insert_after);
create_iv (vptr, fold_convert (vect_ptr_type, DR_STEP (dr)), vect_ptr,
containing_loop, &incr_gsi, insert_after, &indx_before_incr,
&indx_after_incr);
incr = gsi_stmt (incr_gsi);
set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo));
/* Copy the points-to information if it exists. */
if (DR_PTR_INFO (dr))
{
duplicate_ssa_name_ptr_info (indx_before_incr, DR_PTR_INFO (dr));
duplicate_ssa_name_ptr_info (indx_after_incr, DR_PTR_INFO (dr));
}
merge_alias_info (vect_ptr_init, indx_before_incr);
merge_alias_info (vect_ptr_init, indx_after_incr);
if (ptr_incr)
*ptr_incr = incr;
return indx_before_incr;
}
else
gcc_unreachable ();
}
/* Function bump_vector_ptr
Increment a pointer (to a vector type) by vector-size. If requested,
i.e. if PTR-INCR is given, then also connect the new increment stmt
to the existing def-use update-chain of the pointer, by modifying
the PTR_INCR as illustrated below:
The pointer def-use update-chain before this function:
DATAREF_PTR = phi (p_0, p_2)
....
PTR_INCR: p_2 = DATAREF_PTR + step
The pointer def-use update-chain after this function:
DATAREF_PTR = phi (p_0, p_2)
....
NEW_DATAREF_PTR = DATAREF_PTR + BUMP
....
PTR_INCR: p_2 = NEW_DATAREF_PTR + step
Input:
DATAREF_PTR - ssa_name of a pointer (to vector type) that is being updated
in the loop.
PTR_INCR - optional. The stmt that updates the pointer in each iteration of
the loop. The increment amount across iterations is expected
to be vector_size.
BSI - location where the new update stmt is to be placed.
STMT - the original scalar memory-access stmt that is being vectorized.
BUMP - optional. The offset by which to bump the pointer. If not given,
the offset is assumed to be vector_size.
Output: Return NEW_DATAREF_PTR as illustrated above.
*/
static tree
bump_vector_ptr (tree dataref_ptr, gimple ptr_incr, gimple_stmt_iterator *gsi,
gimple stmt, tree bump)
{
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
tree vectype = STMT_VINFO_VECTYPE (stmt_info);
tree ptr_var = SSA_NAME_VAR (dataref_ptr);
tree update = TYPE_SIZE_UNIT (vectype);
gimple incr_stmt;
ssa_op_iter iter;
use_operand_p use_p;
tree new_dataref_ptr;
if (bump)
update = bump;
incr_stmt = gimple_build_assign_with_ops (POINTER_PLUS_EXPR, ptr_var,
dataref_ptr, update);
new_dataref_ptr = make_ssa_name (ptr_var, incr_stmt);
gimple_assign_set_lhs (incr_stmt, new_dataref_ptr);
vect_finish_stmt_generation (stmt, incr_stmt, gsi);
/* Copy the points-to information if it exists. */
if (DR_PTR_INFO (dr))
duplicate_ssa_name_ptr_info (new_dataref_ptr, DR_PTR_INFO (dr));
merge_alias_info (new_dataref_ptr, dataref_ptr);
if (!ptr_incr)
return new_dataref_ptr;
/* Update the vector-pointer's cross-iteration increment. */
FOR_EACH_SSA_USE_OPERAND (use_p, ptr_incr, iter, SSA_OP_USE)
{
tree use = USE_FROM_PTR (use_p);
if (use == dataref_ptr)
SET_USE (use_p, new_dataref_ptr);
else
gcc_assert (tree_int_cst_compare (use, update) == 0);
}
return new_dataref_ptr;
}
/* Function vect_create_destination_var.
Create a new temporary of type VECTYPE. */
static tree
vect_create_destination_var (tree scalar_dest, tree vectype)
{
tree vec_dest;
const char *new_name;
tree type;
enum vect_var_kind kind;
kind = vectype ? vect_simple_var : vect_scalar_var;
type = vectype ? vectype : TREE_TYPE (scalar_dest);
gcc_assert (TREE_CODE (scalar_dest) == SSA_NAME);
new_name = get_name (scalar_dest);
if (!new_name)
new_name = "var_";
vec_dest = vect_get_new_vect_var (type, kind, new_name);
add_referenced_var (vec_dest);
return vec_dest;
}
/* Function vect_init_vector.
Insert a new stmt (INIT_STMT) that initializes a new vector variable with
the vector elements of VECTOR_VAR. Place the initialization at BSI if it
is not NULL. Otherwise, place the initialization at the loop preheader.
Return the DEF of INIT_STMT.
It will be used in the vectorization of STMT. */
static tree
vect_init_vector (gimple stmt, tree vector_var, tree vector_type,
gimple_stmt_iterator *gsi)
{
stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt);
tree new_var;
gimple init_stmt;
tree vec_oprnd;
edge pe;
tree new_temp;
basic_block new_bb;
new_var = vect_get_new_vect_var (vector_type, vect_simple_var, "cst_");
add_referenced_var (new_var);
init_stmt = gimple_build_assign (new_var, vector_var);
new_temp = make_ssa_name (new_var, init_stmt);
gimple_assign_set_lhs (init_stmt, new_temp);
if (gsi)
vect_finish_stmt_generation (stmt, init_stmt, gsi);
else
{
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
if (nested_in_vect_loop_p (loop, stmt))
loop = loop->inner;
pe = loop_preheader_edge (loop);
new_bb = gsi_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_gimple_stmt (vect_dump, init_stmt, 0, TDF_SLIM);
}
vec_oprnd = gimple_assign_lhs (init_stmt);
return vec_oprnd;
}
/* For constant and loop invariant defs of SLP_NODE this function returns
(vector) defs (VEC_OPRNDS) that will be used in the vectorized stmts.
OP_NUM determines if we gather defs for operand 0 or operand 1 of the scalar
stmts. NUMBER_OF_VECTORS is the number of vector defs to create. */
static void
vect_get_constant_vectors (slp_tree slp_node, VEC(tree,heap) **vec_oprnds,
unsigned int op_num, unsigned int number_of_vectors)
{
VEC (gimple, heap) *stmts = SLP_TREE_SCALAR_STMTS (slp_node);
gimple stmt = VEC_index (gimple, stmts, 0);
stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt);
tree vectype = STMT_VINFO_VECTYPE (stmt_vinfo);
int nunits;
tree vec_cst;
tree t = NULL_TREE;
int j, number_of_places_left_in_vector;
tree vector_type;
tree op, vop;
int group_size = VEC_length (gimple, stmts);
unsigned int vec_num, i;
int number_of_copies = 1;
VEC (tree, heap) *voprnds = VEC_alloc (tree, heap, number_of_vectors);
bool constant_p, is_store;
if (STMT_VINFO_DATA_REF (stmt_vinfo))
{
is_store = true;
op = gimple_assign_rhs1 (stmt);
}
else
{
is_store = false;
op = gimple_op (stmt, op_num + 1);
}
if (CONSTANT_CLASS_P (op))
{
vector_type = vectype;
constant_p = true;
}
else
{
vector_type = get_vectype_for_scalar_type (TREE_TYPE (op));
gcc_assert (vector_type);
constant_p = false;
}
nunits = TYPE_VECTOR_SUBPARTS (vector_type);
/* NUMBER_OF_COPIES is the number of times we need to use the same values in
created vectors. It is greater than 1 if unrolling is performed.
For example, we have two scalar operands, s1 and s2 (e.g., group of
strided accesses of size two), while NUNITS is four (i.e., four scalars
of this type can be packed in a vector). The output vector will contain
two copies of each scalar operand: {s1, s2, s1, s2}. (NUMBER_OF_COPIES
will be 2).
If GROUP_SIZE > NUNITS, the scalars will be split into several vectors
containing the operands.
For example, NUNITS is four as before, and the group size is 8
(s1, s2, ..., s8). We will create two vectors {s1, s2, s3, s4} and
{s5, s6, s7, s8}. */
number_of_copies = least_common_multiple (nunits, group_size) / group_size;
number_of_places_left_in_vector = nunits;
for (j = 0; j < number_of_copies; j++)
{
for (i = group_size - 1; VEC_iterate (gimple, stmts, i, stmt); i--)
{
if (is_store)
op = gimple_assign_rhs1 (stmt);
else
op = gimple_op (stmt, op_num + 1);
/* Create 'vect_ = {op0,op1,...,opn}'. */
t = tree_cons (NULL_TREE, op, t);
number_of_places_left_in_vector--;
if (number_of_places_left_in_vector == 0)
{
number_of_places_left_in_vector = nunits;
if (constant_p)
vec_cst = build_vector (vector_type, t);
else
vec_cst = build_constructor_from_list (vector_type, t);
VEC_quick_push (tree, voprnds,
vect_init_vector (stmt, vec_cst, vector_type, NULL));
t = NULL_TREE;
}
}
}
/* Since the vectors are created in the reverse order, we should invert
them. */
vec_num = VEC_length (tree, voprnds);
for (j = vec_num - 1; j >= 0; j--)
{
vop = VEC_index (tree, voprnds, j);
VEC_quick_push (tree, *vec_oprnds, vop);
}
VEC_free (tree, heap, voprnds);
/* In case that VF is greater than the unrolling factor needed for the SLP
group of stmts, NUMBER_OF_VECTORS to be created is greater than
NUMBER_OF_SCALARS/NUNITS or NUNITS/NUMBER_OF_SCALARS, and hence we have
to replicate the vectors. */
while (number_of_vectors > VEC_length (tree, *vec_oprnds))
{
for (i = 0; VEC_iterate (tree, *vec_oprnds, i, vop) && i < vec_num; i++)
VEC_quick_push (tree, *vec_oprnds, vop);
}
}
/* Get vectorized definitions from SLP_NODE that contains corresponding
vectorized def-stmts. */
static void
vect_get_slp_vect_defs (slp_tree slp_node, VEC (tree,heap) **vec_oprnds)
{
tree vec_oprnd;
gimple vec_def_stmt;
unsigned int i;
gcc_assert (SLP_TREE_VEC_STMTS (slp_node));
for (i = 0;
VEC_iterate (gimple, SLP_TREE_VEC_STMTS (slp_node), i, vec_def_stmt);
i++)
{
gcc_assert (vec_def_stmt);
vec_oprnd = gimple_get_lhs (vec_def_stmt);
VEC_quick_push (tree, *vec_oprnds, vec_oprnd);
}
}
/* Get vectorized definitions for SLP_NODE.
If the scalar definitions are loop invariants or constants, collect them and
call vect_get_constant_vectors() to create vector stmts.
Otherwise, the def-stmts must be already vectorized and the vectorized stmts
must be stored in the LEFT/RIGHT node of SLP_NODE, and we call
vect_get_slp_vect_defs() to retrieve them.
If VEC_OPRNDS1 is NULL, don't get vector defs for the second operand (from
the right node. This is used when the second operand must remain scalar. */
static void
vect_get_slp_defs (slp_tree slp_node, VEC (tree,heap) **vec_oprnds0,
VEC (tree,heap) **vec_oprnds1)
{
gimple first_stmt;
enum tree_code code;
int number_of_vects;
HOST_WIDE_INT lhs_size_unit, rhs_size_unit;
first_stmt = VEC_index (gimple, SLP_TREE_SCALAR_STMTS (slp_node), 0);
/* The number of vector defs is determined by the number of vector statements
in the node from which we get those statements. */
if (SLP_TREE_LEFT (slp_node))
number_of_vects = SLP_TREE_NUMBER_OF_VEC_STMTS (SLP_TREE_LEFT (slp_node));
else
{
number_of_vects = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node);
/* Number of vector stmts was calculated according to LHS in
vect_schedule_slp_instance(), fix it by replacing LHS with RHS, if
necessary. See vect_get_smallest_scalar_type() for details. */
vect_get_smallest_scalar_type (first_stmt, &lhs_size_unit,
&rhs_size_unit);
if (rhs_size_unit != lhs_size_unit)
{
number_of_vects *= rhs_size_unit;
number_of_vects /= lhs_size_unit;
}
}
/* Allocate memory for vectorized defs. */
*vec_oprnds0 = VEC_alloc (tree, heap, number_of_vects);
/* SLP_NODE corresponds either to a group of stores or to a group of
unary/binary operations. We don't call this function for loads. */
if (SLP_TREE_LEFT (slp_node))
/* The defs are already vectorized. */
vect_get_slp_vect_defs (SLP_TREE_LEFT (slp_node), vec_oprnds0);
else
/* Build vectors from scalar defs. */
vect_get_constant_vectors (slp_node, vec_oprnds0, 0, number_of_vects);
if (STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)))
/* Since we don't call this function with loads, this is a group of
stores. */
return;
code = gimple_assign_rhs_code (first_stmt);
if (get_gimple_rhs_class (code) != GIMPLE_BINARY_RHS || !vec_oprnds1)
return;
/* The number of vector defs is determined by the number of vector statements
in the node from which we get those statements. */
if (SLP_TREE_RIGHT (slp_node))
number_of_vects = SLP_TREE_NUMBER_OF_VEC_STMTS (SLP_TREE_RIGHT (slp_node));
else
number_of_vects = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node);
*vec_oprnds1 = VEC_alloc (tree, heap, number_of_vects);
if (SLP_TREE_RIGHT (slp_node))
/* The defs are already vectorized. */
vect_get_slp_vect_defs (SLP_TREE_RIGHT (slp_node), vec_oprnds1);
else
/* Build vectors from scalar defs. */
vect_get_constant_vectors (slp_node, vec_oprnds1, 1, number_of_vects);
}
/* Function get_initial_def_for_induction
Input:
STMT - a stmt that performs an induction operation in the loop.
IV_PHI - the initial value of the induction variable
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 (gimple 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 (gimple_phi_result (iv_phi));
tree vectype;
int nunits;
edge pe = loop_preheader_edge (loop);
struct loop *iv_loop;
basic_block new_bb;
tree vec, vec_init, vec_step, t;
tree access_fn;
tree new_var;
tree new_name;
gimple init_stmt, induction_phi, new_stmt;
tree induc_def, vec_def, vec_dest;
tree init_expr, step_expr;
int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
int i;
bool ok;
int ncopies;
tree expr;
stmt_vec_info phi_info = vinfo_for_stmt (iv_phi);
bool nested_in_vect_loop = false;
gimple_seq stmts = NULL;
imm_use_iterator imm_iter;
use_operand_p use_p;
gimple exit_phi;
edge latch_e;
tree loop_arg;
gimple_stmt_iterator si;
basic_block bb = gimple_bb (iv_phi);
vectype = get_vectype_for_scalar_type (scalar_type);
gcc_assert (vectype);
nunits = TYPE_VECTOR_SUBPARTS (vectype);
ncopies = vf / nunits;
gcc_assert (phi_info);
gcc_assert (ncopies >= 1);
/* Find the first insertion point in the BB. */
si = gsi_after_labels (bb);
if (INTEGRAL_TYPE_P (scalar_type) || POINTER_TYPE_P (scalar_type))
step_expr = build_int_cst (scalar_type, 0);
else
step_expr = build_real (scalar_type, dconst0);
/* Is phi in an inner-loop, while vectorizing an enclosing outer-loop? */
if (nested_in_vect_loop_p (loop, iv_phi))
{
nested_in_vect_loop = true;
iv_loop = loop->inner;
}
else
iv_loop = loop;
gcc_assert (iv_loop == (gimple_bb (iv_phi))->loop_father);
latch_e = loop_latch_edge (iv_loop);
loop_arg = PHI_ARG_DEF_FROM_EDGE (iv_phi, latch_e);
access_fn = analyze_scalar_evolution (iv_loop, PHI_RESULT (iv_phi));
gcc_assert (access_fn);
ok = vect_is_simple_iv_evolution (iv_loop->num, access_fn,
&init_expr, &step_expr);
gcc_assert (ok);
pe = loop_preheader_edge (iv_loop);
/* Create the vector that holds the initial_value of the induction. */
if (nested_in_vect_loop)
{
/* iv_loop is nested in the loop to be vectorized. init_expr had already
been created during vectorization of previous stmts; We obtain it from
the STMT_VINFO_VEC_STMT of the defining stmt. */
tree iv_def = PHI_ARG_DEF_FROM_EDGE (iv_phi, loop_preheader_edge (iv_loop));
vec_init = vect_get_vec_def_for_operand (iv_def, iv_phi, NULL);
}
else
{
/* iv_loop is the loop to be vectorized. Create:
vec_init = [X, X+S, X+2*S, X+3*S] (S = step_expr, X = init_expr) */
new_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 = gsi_insert_seq_on_edge_immediate (pe, stmts);
gcc_assert (!new_bb);
}
t = NULL_TREE;
t = tree_cons (NULL_TREE, init_expr, t);
for (i = 1; i < nunits; i++)
{
/* Create: new_name_i = new_name + step_expr */
enum tree_code code = POINTER_TYPE_P (scalar_type)
? POINTER_PLUS_EXPR : PLUS_EXPR;
init_stmt = gimple_build_assign_with_ops (code, new_var,
new_name, step_expr);
new_name = make_ssa_name (new_var, init_stmt);
gimple_assign_set_lhs (init_stmt, new_name);
new_bb = gsi_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_gimple_stmt (vect_dump, init_stmt, 0, TDF_SLIM);
}
t = tree_cons (NULL_TREE, new_name, t);
}
/* Create a vector from [new_name_0, new_name_1, ..., new_name_nunits-1] */
vec = build_constructor_from_list (vectype, nreverse (t));
vec_init = vect_init_vector (iv_phi, vec, vectype, NULL);
}
/* Create the vector that holds the step of the induction. */
if (nested_in_vect_loop)
/* iv_loop is nested in the loop to be vectorized. Generate:
vec_step = [S, S, S, S] */
new_name = step_expr;
else
{
/* iv_loop is the loop to be vectorized. Generate:
vec_step = [VF*S, VF*S, VF*S, VF*S] */
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);
gcc_assert (CONSTANT_CLASS_P (new_name));
vec = build_vector (vectype, t);
vec_step = vect_init_vector (iv_phi, vec, vectype, NULL);
/* Create the following def-use cycle:
loop prolog:
vec_init = ...
vec_step = ...
loop:
vec_iv = PHI <vec_init, vec_loop>
...
STMT
...
vec_loop = vec_iv + vec_step; */
/* Create the induction-phi that defines the induction-operand. */
vec_dest = vect_get_new_vect_var (vectype, vect_simple_var, "vec_iv_");
add_referenced_var (vec_dest);
induction_phi = create_phi_node (vec_dest, iv_loop->header);
set_vinfo_for_stmt (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 = gimple_build_assign_with_ops (PLUS_EXPR, vec_dest,
induc_def, vec_step);
vec_def = make_ssa_name (vec_dest, new_stmt);
gimple_assign_set_lhs (new_stmt, vec_def);
gsi_insert_before (&si, new_stmt, GSI_SAME_STMT);
set_vinfo_for_stmt (new_stmt, new_stmt_vec_info (new_stmt, loop_vinfo));
/* Set the arguments of the phi node: */
add_phi_arg (induction_phi, vec_init, pe);
add_phi_arg (induction_phi, vec_def, loop_latch_edge (iv_loop));
/* In case that vectorization factor (VF) is bigger than the number
of elements that we can fit in a vectype (nunits), we have to generate
more than one vector stmt - i.e - we need to "unroll" the
vector stmt by a factor VF/nunits. For more details see documentation
in vectorizable_operation. */
if (ncopies > 1)
{
stmt_vec_info prev_stmt_vinfo;
/* FORNOW. This restriction should be relaxed. */
gcc_assert (!nested_in_vect_loop);
/* 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);
gcc_assert (CONSTANT_CLASS_P (new_name));
vec = build_vector (vectype, t);
vec_step = vect_init_vector (iv_phi, vec, vectype, NULL);
vec_def = induc_def;
prev_stmt_vinfo = vinfo_for_stmt (induction_phi);
for (i = 1; i < ncopies; i++)
{
/* vec_i = vec_prev + vec_step */
new_stmt = gimple_build_assign_with_ops (PLUS_EXPR, vec_dest,
vec_def, vec_step);
vec_def = make_ssa_name (vec_dest, new_stmt);
gimple_assign_set_lhs (new_stmt, vec_def);
gsi_insert_before (&si, new_stmt, GSI_SAME_STMT);
set_vinfo_for_stmt (new_stmt,
new_stmt_vec_info (new_stmt, loop_vinfo));
STMT_VINFO_RELATED_STMT (prev_stmt_vinfo) = new_stmt;
prev_stmt_vinfo = vinfo_for_stmt (new_stmt);
}
}
if (nested_in_vect_loop)
{
/* Find the loop-closed exit-phi of the induction, and record
the final vector of induction results: */
exit_phi = NULL;
FOR_EACH_IMM_USE_FAST (use_p, imm_iter, loop_arg)
{
if (!flow_bb_inside_loop_p (iv_loop, gimple_bb (USE_STMT (use_p))))
{
exit_phi = USE_STMT (use_p);
break;
}
}
if (exit_phi)
{
stmt_vec_info stmt_vinfo = vinfo_for_stmt (exit_phi);
/* FORNOW. Currently not supporting the case that an inner-loop induction
is not used in the outer-loop (i.e. only outside the outer-loop). */
gcc_assert (STMT_VINFO_RELEVANT_P (stmt_vinfo)
&& !STMT_VINFO_LIVE_P (stmt_vinfo));
STMT_VINFO_VEC_STMT (stmt_vinfo) = new_stmt;
if (vect_print_dump_info (REPORT_DETAILS))
{
fprintf (vect_dump, "vector of inductions after inner-loop:");
print_gimple_stmt (vect_dump, new_stmt, 0, TDF_SLIM);
}
}
}
if (vect_print_dump_info (REPORT_DETAILS))
{
fprintf (vect_dump, "transform induction: created def-use cycle: ");
print_gimple_stmt (vect_dump, induction_phi, 0, TDF_SLIM);
fprintf (vect_dump, "\n");
print_gimple_stmt (vect_dump, SSA_NAME_DEF_STMT (vec_def), 0, 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. */
static tree
vect_get_vec_def_for_operand (tree op, gimple stmt, tree *scalar_def)
{
tree vec_oprnd;
gimple vec_stmt;
gimple def_stmt;
stmt_vec_info def_stmt_info = NULL;
stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt);
tree vectype = STMT_VINFO_VECTYPE (stmt_vinfo);
unsigned int nunits = TYPE_VECTOR_SUBPARTS (vectype);
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
tree vec_inv;
tree vec_cst;
tree t = NULL_TREE;
tree def;
int i;
enum vect_def_type dt;
bool is_simple_use;
tree vector_type;
if (vect_print_dump_info (REPORT_DETAILS))
{
fprintf (vect_dump, "vect_get_vec_def_for_operand: ");
print_generic_expr (vect_dump, op, TDF_SLIM);
}
is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt);
gcc_assert (is_simple_use);
if (vect_print_dump_info (REPORT_DETAILS))
{
if (def)
{
fprintf (vect_dump, "def = ");
print_generic_expr (vect_dump, def, TDF_SLIM);
}
if (def_stmt)
{
fprintf (vect_dump, " def_stmt = ");
print_gimple_stmt (vect_dump, def_stmt, 0, TDF_SLIM);
}
}
switch (dt)
{
/* Case 1: operand is a constant. */
case vect_constant_def:
{
if (scalar_def)
*scalar_def = op;
/* Create 'vect_cst_ = {cst,cst,...,cst}' */
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "Create vector_cst. nunits = %d", nunits);
for (i = nunits - 1; i >= 0; --i)
{
t = tree_cons (NULL_TREE, op, t);
}
vec_cst = build_vector (vectype, t);
return vect_init_vector (stmt, vec_cst, vectype, NULL);
}
/* Case 2: operand is defined outside the loop - loop invariant. */
case vect_invariant_def:
{
vector_type = get_vectype_for_scalar_type (TREE_TYPE (def));
gcc_assert (vector_type);
nunits = TYPE_VECTOR_SUBPARTS (vector_type);
if (scalar_def)
*scalar_def = def;
/* Create 'vec_inv = {inv,inv,..,inv}' */
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "Create vector_inv.");
for (i = nunits - 1; i >= 0; --i)
{
t = tree_cons (NULL_TREE, def, t);
}
/* FIXME: use build_constructor directly. */
vec_inv = build_constructor_from_list (vector_type, t);
return vect_init_vector (stmt, vec_inv, vector_type, NULL);
}
/* Case 3: operand is defined inside the loop. */
case vect_loop_def:
{
if (scalar_def)
*scalar_def = NULL/* FIXME tuples: def_stmt*/;
/* Get the def from the vectorized stmt. */
def_stmt_info = vinfo_for_stmt (def_stmt);
vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info);
gcc_assert (vec_stmt);
if (gimple_code (vec_stmt) == GIMPLE_PHI)
vec_oprnd = PHI_RESULT (vec_stmt);
else if (is_gimple_call (vec_stmt))
vec_oprnd = gimple_call_lhs (vec_stmt);
else
vec_oprnd = gimple_assign_lhs (vec_stmt);
return vec_oprnd;
}
/* Case 4: operand is defined by a loop header phi - reduction */
case vect_reduction_def:
{
struct loop *loop;
gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI);
loop = (gimple_bb (def_stmt))->loop_father;
/* Get the def before the loop */
op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop));
return get_initial_def_for_reduction (stmt, op, scalar_def);
}
/* Case 5: operand is defined by loop-header phi - induction. */
case vect_induction_def:
{
gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI);
/* Get the def from the vectorized stmt. */
def_stmt_info = vinfo_for_stmt (def_stmt);
vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info);
gcc_assert (vec_stmt && gimple_code (vec_stmt) == GIMPLE_PHI);
vec_oprnd = PHI_RESULT (vec_stmt);
return vec_oprnd;
}
default:
gcc_unreachable ();
}
}
/* Function vect_get_vec_def_for_stmt_copy
Return a vector-def for an operand. This function is used when the
vectorized stmt to be created (by the caller to this function) is a "copy"
created in case the vectorized result cannot fit in one vector, and several
copies of the vector-stmt are required. In this case the vector-def is
retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field
of the stmt that defines VEC_OPRND.
DT is the type of the vector def VEC_OPRND.
Context:
In case the vectorization factor (VF) is bigger than the number
of elements that can fit in a vectype (nunits), we have to generate
more than one vector stmt to vectorize the scalar stmt. This situation
arises when there are multiple data-types operated upon in the loop; the
smallest data-type determines the VF, and as a result, when vectorizing
stmts operating on wider types we need to create 'VF/nunits' "copies" of the
vector stmt (each computing a vector of 'nunits' results, and together
computing 'VF' results in each iteration). This function is called when
vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in
which VF=16 and nunits=4, so the number of copies required is 4):
scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT
S1: x = load VS1.0: vx.0 = memref0 VS1.1
VS1.1: vx.1 = memref1 VS1.2
VS1.2: vx.2 = memref2 VS1.3
VS1.3: vx.3 = memref3
S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1
VSnew.1: vz1 = vx.1 + ... VSnew.2
VSnew.2: vz2 = vx.2 + ... VSnew.3
VSnew.3: vz3 = vx.3 + ...
The vectorization of S1 is explained in vectorizable_load.
The vectorization of S2:
To create the first vector-stmt out of the 4 copies - VSnew.0 -
the function 'vect_get_vec_def_for_operand' is called to
get the relevant vector-def for each operand of S2. For operand x it
returns the vector-def 'vx.0'.
To create the remaining copies of the vector-stmt (VSnew.j), this
function is called to get the relevant vector-def for each operand. It is
obtained from the respective VS1.j stmt, which is recorded in the
STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND.
For example, to obtain the vector-def 'vx.1' in order to create the
vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'.
Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the
STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1',
and return its def ('vx.1').
Overall, to create the above sequence this function will be called 3 times:
vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0);
vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1);
vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */
static tree
vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd)
{
gimple vec_stmt_for_operand;
stmt_vec_info def_stmt_info;
/* 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 = gimple_get_lhs (vec_stmt_for_operand);
if (gimple_code (vec_stmt_for_operand) == GIMPLE_PHI)
vec_oprnd = PHI_RESULT (vec_stmt_for_operand);
else
vec_oprnd = gimple_get_lhs (vec_stmt_for_operand);
return vec_oprnd;
}
/* Get vectorized definitions for the operands to create a copy of an original
stmt. See vect_get_vec_def_for_stmt_copy() for details. */
static void
vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt,
VEC(tree,heap) **vec_oprnds0,
VEC(tree,heap) **vec_oprnds1)
{
tree vec_oprnd = VEC_pop (tree, *vec_oprnds0);
vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd);
VEC_quick_push (tree, *vec_oprnds0, vec_oprnd);
if (vec_oprnds1 && *vec_oprnds1)
{
vec_oprnd = VEC_pop (tree, *vec_oprnds1);
vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd);
VEC_quick_push (tree, *vec_oprnds1, vec_oprnd);
}
}
/* Get vectorized definitions for OP0 and OP1, or SLP_NODE if it is not NULL. */
static void
vect_get_vec_defs (tree op0, tree op1, gimple stmt,
VEC(tree,heap) **vec_oprnds0, VEC(tree,heap) **vec_oprnds1,
slp_tree slp_node)
{
if (slp_node)
vect_get_slp_defs (slp_node, vec_oprnds0, vec_oprnds1);
else
{
tree vec_oprnd;
*vec_oprnds0 = VEC_alloc (tree, heap, 1);
vec_oprnd = vect_get_vec_def_for_operand (op0, stmt, NULL);
VEC_quick_push (tree, *vec_oprnds0, vec_oprnd);
if (op1)
{
*vec_oprnds1 = VEC_alloc (tree, heap, 1);
vec_oprnd = vect_get_vec_def_for_operand (op1, stmt, NULL);
VEC_quick_push (tree, *vec_oprnds1, vec_oprnd);
}
}
}
/* Function vect_finish_stmt_generation.
Insert a new stmt. */
static void
vect_finish_stmt_generation (gimple stmt, gimple vec_stmt,
gimple_stmt_iterator *gsi)
{
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
gcc_assert (gimple_code (stmt) != GIMPLE_LABEL);
gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT);
set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, loop_vinfo));
if (vect_print_dump_info (REPORT_DETAILS))
{
fprintf (vect_dump, "add new stmt: ");
print_gimple_stmt (vect_dump, vec_stmt, 0, TDF_SLIM);
}
gimple_set_location (vec_stmt, gimple_location (gsi_stmt (*gsi)));
}
/* Function get_initial_def_for_reduction
Input:
STMT - a stmt that performs a reduction operation in the loop.
INIT_VAL - the initial value of the reduction variable
Output:
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.
Option1 (adjust in epilog): Initialize the vector as follows:
add: [0,0,...,0,0]
mult: [1,1,...,1,1]
min/max: [init_val,init_val,..,init_val,init_val]
bit and/or: [init_val,init_val,..,init_val,init_val]
and when necessary (e.g. add/mult case) let the caller know
that it needs to adjust the result by init_val.
Option2: Initialize the vector as follows:
add: [0,0,...,0,init_val]
mult: [1,1,...,1,init_val]
min/max: [init_val,init_val,...,init_val]
bit and/or: [init_val,init_val,...,init_val]
and no adjustments are needed.
For example, for the following code:
s = init_val;
for (i=0;i<n;i++)
s = s + a[i];
STMT is 's = s + a[i]', and the reduction variable is 's'.
For a vector of 4 units, we want to return either [0,0,0,init_val],
or [0,0,0,0] and let the caller know that it needs to adjust
the result at the end by 'init_val'.
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 (gimple stmt, tree init_val, tree *adjustment_def)
{
stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt);
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
tree vectype = STMT_VINFO_VECTYPE (stmt_vinfo);
int nunits = TYPE_VECTOR_SUBPARTS (vectype);
tree scalar_type = TREE_TYPE (vectype);
enum tree_code code = gimple_assign_rhs_code (stmt);
tree type = TREE_TYPE (init_val);
tree vecdef;
tree def_for_init;
tree init_def;
tree t = NULL_TREE;
int i;
bool nested_in_vect_loop = false;
gcc_assert (POINTER_TYPE_P (type) || INTEGRAL_TYPE_P (type) || SCALAR_FLOAT_TYPE_P (type));
if (nested_in_vect_loop_p (loop, stmt))
nested_in_vect_loop = true;
else
gcc_assert (loop == (gimple_bb (stmt))->loop_father);
vecdef = vect_get_vec_def_for_operand (init_val, stmt, NULL);
switch (code)
{
case WIDEN_SUM_EXPR:
case DOT_PROD_EXPR:
case PLUS_EXPR:
if (nested_in_vect_loop)
*adjustment_def = vecdef;
else
*adjustment_def = init_val;
/* Create a vector of zeros for init_def. */
if (SCALAR_FLOAT_TYPE_P (scalar_type))
def_for_init = build_real (scalar_type, dconst0);
else
def_for_init = build_int_cst (scalar_type, 0);
for (i = nunits - 1; i >= 0; --i)
t = tree_cons (NULL_TREE, def_for_init, t);
init_def = build_vector (vectype, t);
break;
case MIN_EXPR:
case MAX_EXPR:
*adjustment_def = NULL_TREE;
init_def = vecdef;
break;
default:
gcc_unreachable ();
}
return init_def;
}
/* Function vect_create_epilog_for_reduction
Create code at the loop-epilog to finalize the result of a reduction
computation.
VECT_DEF is a vector of partial results.
REDUC_CODE is the tree-code for the epilog reduction.
NCOPIES is > 1 in case the vectorization factor (VF) is bigger than the
number of elements that we can fit in a vectype (nunits). In this case
we have to generate more than one vector stmt - i.e - we need to "unroll"
the vector stmt by a factor VF/nunits. For more details see documentation
in vectorizable_operation.
STMT is the scalar reduction stmt that is being vectorized.
REDUCTION_PHI is the phi-node that carries the reduction computation.
This function:
1. Creates the reduction def-use cycle: sets the arguments for
REDUCTION_PHI:
The loop-entry argument is the vectorized initial-value of the reduction.
The loop-latch argument is VECT_DEF - the vector of partial sums.
2. "Reduces" the vector of partial results VECT_DEF into a single result,
by applying the operation specified by REDUC_CODE if available, or by
other means (whole-vector shifts or a scalar loop).
The function also creates a new phi node at the loop exit to preserve
loop-closed form, as illustrated below.
The flow at the entry to this function:
loop:
vec_def = phi <null, null> # REDUCTION_PHI
VECT_DEF = vector_stmt # vectorized form of STMT
s_loop = scalar_stmt # (scalar) STMT
loop_exit:
s_out0 = phi <s_loop> # (scalar) EXIT_PHI
use <s_out0>
use <s_out0>
The above is transformed by this function into:
loop:
vec_def = phi <vec_init, VECT_DEF> # REDUCTION_PHI
VECT_DEF = vector_stmt # vectorized form of STMT
s_loop = scalar_stmt # (scalar) STMT
loop_exit:
s_out0 = phi <s_loop> # (scalar) EXIT_PHI
v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
v_out2 = reduce <v_out1>
s_out3 = extract_field <v_out2, 0>
s_out4 = adjust_result <s_out3>
use <s_out4>
use <s_out4>
*/
static void
vect_create_epilog_for_reduction (tree vect_def, gimple stmt,
int ncopies,
enum tree_code reduc_code,
gimple reduction_phi)
{
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
stmt_vec_info prev_phi_info;
tree vectype;
enum machine_mode mode;
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
basic_block exit_bb;
tree scalar_dest;
tree scalar_type;
gimple new_phi = NULL, phi;
gimple_stmt_iterator exit_gsi;
tree vec_dest;
tree new_temp = NULL_TREE;
tree new_name;
gimple epilog_stmt = NULL;
tree new_scalar_dest, new_dest;
gimple exit_phi;
tree bitsize, bitpos, bytesize;
enum tree_code code = gimple_assign_rhs_code (stmt);
tree adjustment_def;
tree vec_initial_def, def;
tree orig_name;
imm_use_iterator imm_iter;
use_operand_p use_p;
bool extract_scalar_result = false;
tree reduction_op, expr;
gimple orig_stmt;
gimple use_stmt;
bool nested_in_vect_loop = false;
VEC(gimple,heap) *phis = NULL;
enum vect_def_type dt = vect_unknown_def_type;
int j, i;
if (nested_in_vect_loop_p (loop, stmt))
{
loop = loop->inner;
nested_in_vect_loop = true;
}
switch (get_gimple_rhs_class (gimple_assign_rhs_code (stmt)))
{
case GIMPLE_SINGLE_RHS:
gcc_assert (TREE_OPERAND_LENGTH (gimple_assign_rhs1 (stmt)) == ternary_op);
reduction_op = TREE_OPERAND (gimple_assign_rhs1 (stmt), 2);
break;
case GIMPLE_UNARY_RHS:
reduction_op = gimple_assign_rhs1 (stmt);
break;
case GIMPLE_BINARY_RHS:
reduction_op = gimple_assign_rhs2 (stmt);
break;
default:
gcc_unreachable ();
}
vectype = get_vectype_for_scalar_type (TREE_TYPE (reduction_op));
gcc_assert (vectype);
mode = TYPE_MODE (vectype);
/*** 1. Create the reduction def-use cycle ***/
/* For the case of reduction, vect_get_vec_def_for_operand returns
the scalar def before the loop, that defines the initial value
of the reduction variable. */
vec_initial_def = vect_get_vec_def_for_operand (reduction_op, stmt,
&adjustment_def);
phi = reduction_phi;
def = vect_def;
for (j = 0; j < ncopies; j++)
{
/* 1.1 set the loop-entry arg of the reduction-phi: */
add_phi_arg (phi, vec_initial_def, loop_preheader_edge (loop));
/* 1.2 set the loop-latch arg for the reduction-phi: */
if (j > 0)
def = vect_get_vec_def_for_stmt_copy (dt, def);
add_phi_arg (phi, def, loop_latch_edge (loop));
if (vect_print_dump_info (REPORT_DETAILS))
{
fprintf (vect_dump, "transform reduction: created def-use cycle: ");
print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM);
fprintf (vect_dump, "\n");
print_gimple_stmt (vect_dump, SSA_NAME_DEF_STMT (def), 0, TDF_SLIM);
}
phi = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (phi));
}
/*** 2. Create epilog code
The reduction epilog code operates across the elements of the vector
of partial results computed by the vectorized loop.
The reduction epilog code consists of:
step 1: compute the scalar result in a vector (v_out2)
step 2: extract the scalar result (s_out3) from the vector (v_out2)
step 3: adjust the scalar result (s_out3) if needed.
Step 1 can be accomplished using one the following three schemes:
(scheme 1) using reduc_code, if available.
(scheme 2) using whole-vector shifts, if available.
(scheme 3) using a scalar loop. In this case steps 1+2 above are
combined.
The overall epilog code looks like this:
s_out0 = phi <s_loop> # original EXIT_PHI
v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
v_out2 = reduce <v_out1> # step 1
s_out3 = extract_field <v_out2, 0> # step 2
s_out4 = adjust_result <s_out3> # step 3
(step 3 is optional, and steps 1 and 2 may be combined).
Lastly, the uses of s_out0 are replaced by s_out4.
***/
/* 2.1 Create new loop-exit-phi to preserve loop-closed form:
v_out1 = phi <v_loop> */
exit_bb = single_exit (loop)->dest;
def = vect_def;
prev_phi_info = NULL;
for (j = 0; j < ncopies; j++)
{
phi = create_phi_node (SSA_NAME_VAR (vect_def), exit_bb);
set_vinfo_for_stmt (phi, new_stmt_vec_info (phi, loop_vinfo));
if (j == 0)
new_phi = phi;
else
{
def = vect_get_vec_def_for_stmt_copy (dt, def);
STMT_VINFO_RELATED_STMT (prev_phi_info) = phi;
}
SET_PHI_ARG_DEF (phi, single_exit (loop)->dest_idx, def);
prev_phi_info = vinfo_for_stmt (phi);
}
exit_gsi = gsi_after_labels (exit_bb);
/* 2.2 Get the relevant tree-code to use in the epilog for schemes 2,3
(i.e. when reduc_code is not available) and in the final adjustment
code (if needed). Also get the original scalar reduction variable as
defined in the loop. In case STMT is a "pattern-stmt" (i.e. - it
represents a reduction pattern), the tree-code and scalar-def are
taken from the original stmt that the pattern-stmt (STMT) replaces.
Otherwise (it is a regular reduction) - the tree-code and scalar-def
are taken from STMT. */
orig_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
if (!orig_stmt)
{
/* Regular reduction */
orig_stmt = stmt;
}
else
{
/* Reduction pattern */
stmt_vec_info stmt_vinfo = vinfo_for_stmt (orig_stmt);
gcc_assert (STMT_VINFO_IN_PATTERN_P (stmt_vinfo));
gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo) == stmt);
}
code = gimple_assign_rhs_code (orig_stmt);
scalar_dest = gimple_assign_lhs (orig_stmt);
scalar_type = TREE_TYPE (scalar_dest);
new_scalar_dest = vect_create_destination_var (scalar_dest, NULL);
bitsize = TYPE_SIZE (scalar_type);
bytesize = TYPE_SIZE_UNIT (scalar_type);
/* In case this is a reduction in an inner-loop while vectorizing an outer
loop - we don't need to extract a single scalar result at the end of the
inner-loop. The final vector of partial results will be used in the
vectorized outer-loop, or reduced to a scalar result at the end of the
outer-loop. */
if (nested_in_vect_loop)
goto vect_finalize_reduction;
/* FORNOW */
gcc_assert (ncopies == 1);
/* 2.3 Create the reduction code, using one of the three schemes described
above. */
if (reduc_code < NUM_TREE_CODES)
{
tree tmp;
/*** Case 1: Create:
v_out2 = reduc_expr <v_out1> */
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "Reduce using direct vector reduction.");
vec_dest = vect_create_destination_var (scalar_dest, vectype);
tmp = build1 (reduc_code, vectype, PHI_RESULT (new_phi));
epilog_stmt = gimple_build_assign (vec_dest, tmp);
new_temp = make_ssa_name (vec_dest, epilog_stmt);
gimple_assign_set_lhs (epilog_stmt, new_temp);
gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
extract_scalar_result = true;
}
else
{
enum tree_code shift_code = 0;
bool have_whole_vector_shift = true;
int bit_offset;
int element_bitsize = tree_low_cst (bitsize, 1);
int vec_size_in_bits = tree_low_cst (TYPE_SIZE (vectype), 1);
tree vec_temp;
if (optab_handler (vec_shr_optab, mode)->insn_code != CODE_FOR_nothing)
shift_code = VEC_RSHIFT_EXPR;
else
have_whole_vector_shift = false;
/* Regardless of whether we have a whole vector shift, if we're
emulating the operation via tree-vect-generic, we don't want
to use it. Only the first round of the reduction is likely
to still be profitable via emulation. */
/* ??? It might be better to emit a reduction tree code here, so that
tree-vect-generic can expand the first round via bit tricks. */
if (!VECTOR_MODE_P (mode))
have_whole_vector_shift = false;
else
{
optab optab = optab_for_tree_code (code, vectype, optab_default);
if (optab_handler (optab, mode)->insn_code == CODE_FOR_nothing)
have_whole_vector_shift = false;
}
if (have_whole_vector_shift)
{
/*** Case 2: Create:
for (offset = VS/2; offset >= element_size; offset/=2)
{
Create: va' = vec_shift <va, offset>
Create: va = vop <va, va'>
} */
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "Reduce using vector shifts");
vec_dest = vect_create_destination_var (scalar_dest, vectype);
new_temp = PHI_RESULT (new_phi);
for (bit_offset = vec_size_in_bits/2;
bit_offset >= element_bitsize;
bit_offset /= 2)
{
tree bitpos = size_int (bit_offset);
epilog_stmt = gimple_build_assign_with_ops (shift_code, vec_dest,
new_temp, bitpos);
new_name = make_ssa_name (vec_dest, epilog_stmt);
gimple_assign_set_lhs (epilog_stmt, new_name);
gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
epilog_stmt = gimple_build_assign_with_ops (code, vec_dest,
new_name, new_temp);
new_temp = make_ssa_name (vec_dest, epilog_stmt);
gimple_assign_set_lhs (epilog_stmt, new_temp);
gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
}
extract_scalar_result = true;
}
else
{
tree rhs;
/*** Case 3: Create:
s = extract_field <v_out2, 0>
for (offset = element_size;
offset < vector_size;
offset += element_size;)
{
Create: s' = extract_field <v_out2, offset>
Create: s = op <s, s'>
} */
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "Reduce using scalar code. ");
vec_temp = PHI_RESULT (new_phi);
vec_size_in_bits = tree_low_cst (TYPE_SIZE (vectype), 1);
rhs = build3 (BIT_FIELD_REF, scalar_type, vec_temp, bitsize,
bitsize_zero_node);
epilog_stmt = gimple_build_assign (new_scalar_dest, rhs);
new_temp = make_ssa_name (new_scalar_dest, epilog_stmt);
gimple_assign_set_lhs (epilog_stmt, new_temp);
gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
for (bit_offset = element_bitsize;
bit_offset < vec_size_in_bits;
bit_offset += element_bitsize)
{
tree bitpos = bitsize_int (bit_offset);
tree rhs = build3 (BIT_FIELD_REF, scalar_type, vec_temp, bitsize,
bitpos);
epilog_stmt = gimple_build_assign (new_scalar_dest, rhs);
new_name = make_ssa_name (new_scalar_dest, epilog_stmt);
gimple_assign_set_lhs (epilog_stmt, new_name);
gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
epilog_stmt = gimple_build_assign_with_ops (code,
new_scalar_dest,
new_name, new_temp);
new_temp = make_ssa_name (new_scalar_dest, epilog_stmt);
gimple_assign_set_lhs (epilog_stmt, new_temp);
gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
}
extract_scalar_result = false;
}
}
/* 2.4 Extract the final scalar result. Create:
s_out3 = extract_field <v_out2, bitpos> */
if (extract_scalar_result)
{
tree rhs;
gcc_assert (!nested_in_vect_loop);
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "extract scalar result");
if (BYTES_BIG_ENDIAN)
bitpos = size_binop (MULT_EXPR,
bitsize_int (TYPE_VECTOR_SUBPARTS (vectype) - 1),
TYPE_SIZE (scalar_type));
else
bitpos = bitsize_zero_node;
rhs = build3 (BIT_FIELD_REF, scalar_type, new_temp, bitsize, bitpos);
epilog_stmt = gimple_build_assign (new_scalar_dest, rhs);
new_temp = make_ssa_name (new_scalar_dest, epilog_stmt);
gimple_assign_set_lhs (epilog_stmt, new_temp);
gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
}
vect_finalize_reduction:
/* 2.5 Adjust the final result by the initial value of the reduction
variable. (When such adjustment is not needed, then
'adjustment_def' is zero). For example, if code is PLUS we create:
new_temp = loop_exit_def + adjustment_def */
if (adjustment_def)
{
if (nested_in_vect_loop)
{
gcc_assert (TREE_CODE (TREE_TYPE (adjustment_def)) == VECTOR_TYPE);
expr = build2 (code, vectype, PHI_RESULT (new_phi), adjustment_def);
new_dest = vect_create_destination_var (scalar_dest, vectype);
}
else
{
gcc_assert (TREE_CODE (TREE_TYPE (adjustment_def)) != VECTOR_TYPE);
expr = build2 (code, scalar_type, new_temp, adjustment_def);
new_dest = vect_create_destination_var (scalar_dest, scalar_type);
}
epilog_stmt = gimple_build_assign (new_dest, expr);
new_temp = make_ssa_name (new_dest, epilog_stmt);
gimple_assign_set_lhs (epilog_stmt, new_temp);
SSA_NAME_DEF_STMT (new_temp) = epilog_stmt;
gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
}
/* 2.6 Handle the loop-exit phi */
/* Replace uses of s_out0 with uses of s_out3:
Find the loop-closed-use at the loop exit of the original scalar result.
(The reduction result is expected to have two immediate uses - one at the
latch block, and one at the loop exit). */
phis = VEC_alloc (gimple, heap, 10);
FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest)
{
if (!flow_bb_inside_loop_p (loop, gimple_bb (USE_STMT (use_p))))
{
exit_phi = USE_STMT (use_p);
VEC_quick_push (gimple, phis, exit_phi);
}
}
/* We expect to have found an exit_phi because of loop-closed-ssa form. */
gcc_assert (!VEC_empty (gimple, phis));
for (i = 0; VEC_iterate (gimple, phis, i, exit_phi); i++)
{
if (nested_in_vect_loop)
{
stmt_vec_info stmt_vinfo = vinfo_for_stmt (exit_phi);
/* FORNOW. Currently not supporting the case that an inner-loop
reduction is not used in the outer-loop (but only outside the
outer-loop). */
gcc_assert (STMT_VINFO_RELEVANT_P (stmt_vinfo)
&& !STMT_VINFO_LIVE_P (stmt_vinfo));
epilog_stmt = adjustment_def ? epilog_stmt : new_phi;
STMT_VINFO_VEC_STMT (stmt_vinfo) = epilog_stmt;
set_vinfo_for_stmt (epilog_stmt,
new_stmt_vec_info (epilog_stmt, loop_vinfo));
if (adjustment_def)
STMT_VINFO_RELATED_STMT (vinfo_for_stmt (epilog_stmt)) =
STMT_VINFO_RELATED_STMT (vinfo_for_stmt (new_phi));
continue;
}
/* Replace the uses: */
orig_name = PHI_RESULT (exit_phi);
FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, orig_name)
FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
SET_USE (use_p, new_temp);
}
VEC_free (gimple, heap, phis);
}
/* Function vectorizable_reduction.
Check if STMT performs a reduction operation that can be vectorized.
If VEC_STMT is also passed, vectorize the STMT: create a vectorized
stmt to replace it, put it in VEC_STMT, and insert it at BSI.
Return FALSE if not a vectorizable STMT, TRUE otherwise.
This function also handles reduction idioms (patterns) that have been
recognized in advance during vect_pattern_recog. In this case, STMT may be
of this form:
X = pattern_expr (arg0, arg1, ..., X)
and it's STMT_VINFO_RELATED_STMT points to the last stmt in the original
sequence that had been detected and replaced by the pattern-stmt (STMT).
In some cases of reduction patterns, the type of the reduction variable X is
different than the type of the other arguments of STMT.
In such cases, the vectype that is used when transforming STMT into a vector
stmt is different than the vectype that is used to determine the
vectorization factor, because it consists of a different number of elements
than the actual number of elements that are being operated upon in parallel.
For example, consider an accumulation of shorts into an int accumulator.
On some targets it's possible to vectorize this pattern operating on 8
shorts at a time (hence, the vectype for purposes of determining the
vectorization factor should be V8HI); on the other hand, the vectype that
is used to create the vector form is actually V4SI (the type of the result).
Upon entry to this function, STMT_VINFO_VECTYPE records the vectype that
indicates what is the actual level of parallelism (V8HI in the example), so
that the right vectorization factor would be derived. This vectype
corresponds to the type of arguments to the reduction stmt, and should *NOT*
be used to create the vectorized stmt. The right vectype for the vectorized
stmt is obtained from the type of the result X:
get_vectype_for_scalar_type (TREE_TYPE (X))
This means that, contrary to "regular" reductions (or "regular" stmts in
general), the following equation:
STMT_VINFO_VECTYPE == get_vectype_for_scalar_type (TREE_TYPE (X))
does *NOT* necessarily hold for reduction patterns. */
bool
vectorizable_reduction (gimple stmt, gimple_stmt_iterator *gsi,
gimple *vec_stmt)
{
tree vec_dest;
tree scalar_dest;
tree loop_vec_def0 = NULL_TREE, loop_vec_def1 = NULL_TREE;
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
tree vectype = STMT_VINFO_VECTYPE (stmt_info);
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
enum tree_code code, orig_code, epilog_reduc_code = 0;
enum machine_mode vec_mode;
int op_type;
optab optab, reduc_optab;
tree new_temp = NULL_TREE;
tree def;
gimple def_stmt;
enum vect_def_type dt;
gimple new_phi = NULL;
tree scalar_type;
bool is_simple_use;
gimple orig_stmt;
stmt_vec_info orig_stmt_info;
tree expr = NULL_TREE;
int i;
int nunits = TYPE_VECTOR_SUBPARTS (vectype);
int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
int epilog_copies;
stmt_vec_info prev_stmt_info, prev_phi_info;
gimple first_phi = NULL;
bool single_defuse_cycle = false;
tree reduc_def;
gimple new_stmt = NULL;
int j;
tree ops[3];
if (nested_in_vect_loop_p (loop, stmt))
loop = loop->inner;
gcc_assert (ncopies >= 1);
/* FORNOW: SLP not supported. */
if (STMT_SLP_TYPE (stmt_info))
return false;
/* 1. Is vectorizable reduction? */
/* Not supportable if the reduction variable is used in the loop. */
if (STMT_VINFO_RELEVANT (stmt_info) > vect_used_in_outer)
return false;
/* Reductions that are not used even in an enclosing outer-loop,
are expected to be "live" (used out of the loop). */
if (STMT_VINFO_RELEVANT (stmt_info) == vect_unused_in_loop
&& !STMT_VINFO_LIVE_P (stmt_info))
return false;
/* Make sure it was already recognized as a reduction computation. */
if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_reduction_def)
return false;
/* 2. Has this been recognized as a reduction pattern?
Check if STMT represents a pattern that has been recognized
in earlier analysis stages. For stmts that represent a pattern,
the STMT_VINFO_RELATED_STMT field records the last stmt in
the original sequence that constitutes the pattern. */
orig_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
if (orig_stmt)
{
orig_stmt_info = vinfo_for_stmt (orig_stmt);
gcc_assert (STMT_VINFO_RELATED_STMT (orig_stmt_info) == stmt);
gcc_assert (STMT_VINFO_IN_PATTERN_P (orig_stmt_info));
gcc_assert (!STMT_VINFO_IN_PATTERN_P (stmt_info));
}
/* 3. Check the operands of the operation. The first operands are defined
inside the loop body. The last operand is the reduction variable,
which is defined by the loop-header-phi. */
gcc_assert (is_gimple_assign (stmt));
/* Flatten RHS */
switch (get_gimple_rhs_class (gimple_assign_rhs_code (stmt)))
{
case GIMPLE_SINGLE_RHS:
op_type = TREE_OPERAND_LENGTH (gimple_assign_rhs1 (stmt));
if (op_type == ternary_op)
{
tree rhs = gimple_assign_rhs1 (stmt);
ops[0] = TREE_OPERAND (rhs, 0);
ops[1] = TREE_OPERAND (rhs, 1);
ops[2] = TREE_OPERAND (rhs, 2);
code = TREE_CODE (rhs);
}
else
return false;
break;
case GIMPLE_BINARY_RHS:
code = gimple_assign_rhs_code (stmt);
op_type = TREE_CODE_LENGTH (code);
gcc_assert (op_type == binary_op);
ops[0] = gimple_assign_rhs1 (stmt);
ops[1] = gimple_assign_rhs2 (stmt);
break;
case GIMPLE_UNARY_RHS:
return false;
default:
gcc_unreachable ();
}
scalar_dest = gimple_assign_lhs (stmt);
scalar_type = TREE_TYPE (scalar_dest);
if (!POINTER_TYPE_P (scalar_type) && !INTEGRAL_TYPE_P (scalar_type)
&& !SCALAR_FLOAT_TYPE_P (scalar_type))
return false;
/* All uses but the last are expected to be defined in the loop.
The last use is the reduction variable. */
for (i = 0; i < op_type-1; i++)
{
is_simple_use = vect_is_simple_use (ops[i], loop_vinfo, &def_stmt,
&def, &dt);
gcc_assert (is_simple_use);
if (dt != vect_loop_def
&& dt != vect_invariant_def
&& dt != vect_constant_def
&& dt != vect_induction_def)
return false;
}
is_simple_use = vect_is_simple_use (ops[i], loop_vinfo, &def_stmt, &def, &dt);
gcc_assert (is_simple_use);
gcc_assert (dt == vect_reduction_def);
gcc_assert (gimple_code (def_stmt) == GIMPLE_PHI);
if (orig_stmt)
gcc_assert (orig_stmt == vect_is_simple_reduction (loop_vinfo, def_stmt));
else
gcc_assert (stmt == vect_is_simple_reduction (loop_vinfo, def_stmt));
if (STMT_VINFO_LIVE_P (vinfo_for_stmt (def_stmt)))
return false;
/* 4. Supportable by target? */
/* 4.1. check support for the operation in the loop */
optab = optab_for_tree_code (code, vectype, optab_default);
if (!optab)
{
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "no optab.");
return false;
}
vec_mode = TYPE_MODE (vectype);
if (optab_handler (optab, vec_mode)->insn_code == CODE_FOR_nothing)
{
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "op not supported by target.");
if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD
|| LOOP_VINFO_VECT_FACTOR (loop_vinfo)
< vect_min_worthwhile_factor (code))
return false;
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "proceeding using word mode.");
}
/* Worthwhile without SIMD support? */
if (!VECTOR_MODE_P (TYPE_MODE (vectype))
&& LOOP_VINFO_VECT_FACTOR (loop_vinfo)
< vect_min_worthwhile_factor (code))
{
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "not worthwhile without SIMD support.");
return false;
}
/* 4.2. Check support for the epilog operation.
If STMT represents a reduction pattern, then the type of the
reduction variable may be different than the type of the rest
of the arguments. For example, consider the case of accumulation
of shorts into an int accumulator; The original code:
S1: int_a = (int) short_a;
orig_stmt-> S2: int_acc = plus <int_a ,int_acc>;
was replaced with:
STMT: int_acc = widen_sum <short_a, int_acc>
This means that:
1. The tree-code that is used to create the vector operation in the
epilog code (that reduces the partial results) is not the
tree-code of STMT, but is rather the tree-code of the original
stmt from the pattern that STMT is replacing. I.e, in the example
above we want to use 'widen_sum' in the loop, but 'plus' in the
epilog.
2. The type (mode) we use to check available target support
for the vector operation to be created in the *epilog*, is
determined by the type of the reduction variable (in the example
above we'd check this: plus_optab[vect_int_mode]).
However the type (mode) we use to check available target support
for the vector operation to be created *inside the loop*, is
determined by the type of the other arguments to STMT (in the
example we'd check this: widen_sum_optab[vect_short_mode]).
This is contrary to "regular" reductions, in which the types of all
the arguments are the same as the type of the reduction variable.
For "regular" reductions we can therefore use the same vector type
(and also the same tree-code) when generating the epilog code and
when generating the code inside the loop. */
if (orig_stmt)
{
/* This is a reduction pattern: get the vectype from the type of the
reduction variable, and get the tree-code from orig_stmt. */
orig_code = gimple_assign_rhs_code (orig_stmt);
vectype = get_vectype_for_scalar_type (TREE_TYPE (def));
if (!vectype)
{
if (vect_print_dump_info (REPORT_DETAILS))
{
fprintf (vect_dump, "unsupported data-type ");
print_generic_expr (vect_dump, TREE_TYPE (def), TDF_SLIM);
}
return false;
}
vec_mode = TYPE_MODE (vectype);
}
else
{
/* Regular reduction: use the same vectype and tree-code as used for
the vector code inside the loop can be used for the epilog code. */
orig_code = code;
}
if (!reduction_code_for_scalar_code (orig_code, &epilog_reduc_code))
return false;
reduc_optab = optab_for_tree_code (epilog_reduc_code, vectype, optab_default);
if (!reduc_optab)
{
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "no optab for reduction.");
epilog_reduc_code = NUM_TREE_CODES;
}
if (optab_handler (reduc_optab, vec_mode)->insn_code == CODE_FOR_nothing)
{
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "reduc op not supported by target.");
epilog_reduc_code = NUM_TREE_CODES;
}
if (!vec_stmt) /* transformation not required. */
{
STMT_VINFO_TYPE (stmt_info) = reduc_vec_info_type;
if (!vect_model_reduction_cost (stmt_info, epilog_reduc_code, ncopies))
return false;
return true;
}
/** Transform. **/
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "transform reduction.");
/* Create the destination vector */
vec_dest = vect_create_destination_var (scalar_dest, vectype);
/* In case the vectorization factor (VF) is bigger than the number
of elements that we can fit in a vectype (nunits), we have to generate
more than one vector stmt - i.e - we need to "unroll" the
vector stmt by a factor VF/nunits. For more details see documentation
in vectorizable_operation. */
/* If the reduction is used in an outer loop we need to generate
VF intermediate results, like so (e.g. for ncopies=2):
r0 = phi (init, r0)
r1 = phi (init, r1)
r0 = x0 + r0;
r1 = x1 + r1;
(i.e. we generate VF results in 2 registers).
In this case we have a separate def-use cycle for each copy, and therefore
for each copy we get the vector def for the reduction variable from the
respective phi node created for this copy.
Otherwise (the reduction is unused in the loop nest), we can combine
together intermediate results, like so (e.g. for ncopies=2):
r = phi (init, r)
r = x0 + r;
r = x1 + r;
(i.e. we generate VF/2 results in a single register).
In this case for each copy we get the vector def for the reduction variable
from the vectorized reduction operation generated in the previous iteration.
*/
if (STMT_VINFO_RELEVANT (stmt_info) == vect_unused_in_loop)
{
single_defuse_cycle = true;
epilog_copies = 1;
}
else
epilog_copies = ncopies;
prev_stmt_info = NULL;
prev_phi_info = NULL;
for (j = 0; j < ncopies; j++)
{
if (j == 0 || !single_defuse_cycle)
{
/* Create the reduction-phi that defines the reduction-operand. */
new_phi = create_phi_node (vec_dest, loop->header);
set_vinfo_for_stmt (new_phi, new_stmt_vec_info (new_phi, loop_vinfo));
}
/* Handle uses. */
if (j == 0)
{
loop_vec_def0 = vect_get_vec_def_for_operand (ops[0], stmt, NULL);
if (op_type == ternary_op)
{
loop_vec_def1 = vect_get_vec_def_for_operand (ops[1], stmt, NULL);
}
/* Get the vector def for the reduction variable from the phi node */
reduc_def = PHI_RESULT (new_phi);
first_phi = new_phi;
}
else
{
enum vect_def_type dt = vect_unknown_def_type; /* Dummy */
loop_vec_def0 = vect_get_vec_def_for_stmt_copy (dt, loop_vec_def0);
if (op_type == ternary_op)
loop_vec_def1 = vect_get_vec_def_for_stmt_copy (dt, loop_vec_def1);
if (single_defuse_cycle)
reduc_def = gimple_assign_lhs (new_stmt);
else
reduc_def = PHI_RESULT (new_phi);
STMT_VINFO_RELATED_STMT (prev_phi_info) = new_phi;
}
/* 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 = gimple_build_assign (vec_dest, expr);
new_temp = make_ssa_name (vec_dest, new_stmt);
gimple_assign_set_lhs (new_stmt, new_temp);
vect_finish_stmt_generation (stmt, new_stmt, gsi);
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);
prev_phi_info = vinfo_for_stmt (new_phi);
}
/* Finalize the reduction-phi (set its arguments) and create the
epilog reduction code. */
if (!single_defuse_cycle)
new_temp = gimple_assign_lhs (*vec_stmt);
vect_create_epilog_for_reduction (new_temp, stmt, epilog_copies,
epilog_reduc_code, first_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 (gimple call, tree vectype_out, tree vectype_in)
{
tree fndecl = gimple_call_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 (!(gimple_call_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 (gimple stmt, gimple_stmt_iterator *gsi, gimple *vec_stmt)
{
tree vec_dest;
tree scalar_dest;
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, new_temp, def, rhs_type, lhs_type;
gimple def_stmt;
enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
gimple new_stmt;
int ncopies, j;
VEC(tree, heap) *vargs = NULL;
enum { NARROW, NONE, WIDEN } modifier;
size_t i, nargs;
if (!STMT_VINFO_RELEVANT_P (stmt_info))
return false;
if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
return false;
/* FORNOW: SLP not supported. */
if (STMT_SLP_TYPE (stmt_info))
return false;
/* Is STMT a vectorizable call? */
if (!is_gimple_call (stmt))
return false;
if (TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME)
return false;
/* Process function arguments. */
rhs_type = NULL_TREE;
nargs = gimple_call_num_args (stmt);
/* Bail out if the function has more than two arguments, we
do not have interesting builtin functions to vectorize with
more than two arguments. No arguments is also not good. */
if (nargs == 0 || nargs > 2)
return false;
for (i = 0; i < nargs; i++)
{
op = gimple_call_arg (stmt, i);
/* 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[i]))
{
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "use not simple.");
return false;
}
}
vectype_in = get_vectype_for_scalar_type (rhs_type);
if (!vectype_in)
return false;
nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
lhs_type = TREE_TYPE (gimple_call_lhs (stmt));
vectype_out = get_vectype_for_scalar_type (lhs_type);
if (!vectype_out)
return false;