Google Git
Sign in
gnu / gcc / refs/tags/releases/gcc-8.2.0 / . / gcc / tree-vect-stmts.c
blob: 451448ba1f8ea224aa1a9a909467bb7eb0cd7be6 [file] [log] [blame]
/* Statement Analysis and Transformation for Vectorization
Copyright (C) 2003-2018 Free Software Foundation, Inc.
Contributed by Dorit Naishlos <dorit@il.ibm.com>
and Ira Rosen <irar@il.ibm.com>
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify it under
the terms of the GNU General Public License as published by the Free
Software Foundation; either version 3, or (at your option) any later
version.
GCC is distributed in the hope that it will be useful, but WITHOUT ANY
WARRANTY; without even the implied warranty of MERCHANTABILITY or
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
for more details.
You should have received a copy of the GNU General Public License
along with GCC; see the file COPYING3. If not see
<http://www.gnu.org/licenses/>. */
#include "config.h"
#include "system.h"
#include "coretypes.h"
#include "backend.h"
#include "target.h"
#include "rtl.h"
#include "tree.h"
#include "gimple.h"
#include "ssa.h"
#include "optabs-tree.h"
#include "insn-config.h"
#include "recog.h" /* FIXME: for insn_data */
#include "cgraph.h"
#include "dumpfile.h"
#include "alias.h"
#include "fold-const.h"
#include "stor-layout.h"
#include "tree-eh.h"
#include "gimplify.h"
#include "gimple-iterator.h"
#include "gimplify-me.h"
#include "tree-cfg.h"
#include "tree-ssa-loop-manip.h"
#include "cfgloop.h"
#include "tree-ssa-loop.h"
#include "tree-scalar-evolution.h"
#include "tree-vectorizer.h"
#include "builtins.h"
#include "internal-fn.h"
#include "tree-vector-builder.h"
#include "vec-perm-indices.h"
#include "tree-ssa-loop-niter.h"
#include "gimple-fold.h"
/* For lang_hooks.types.type_for_mode. */
#include "langhooks.h"
/* Return the vectorized type for the given statement. */
tree
stmt_vectype (struct _stmt_vec_info *stmt_info)
{
return STMT_VINFO_VECTYPE (stmt_info);
}
/* Return TRUE iff the given statement is in an inner loop relative to
the loop being vectorized. */
bool
stmt_in_inner_loop_p (struct _stmt_vec_info *stmt_info)
{
gimple *stmt = STMT_VINFO_STMT (stmt_info);
basic_block bb = gimple_bb (stmt);
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
struct loop* loop;
if (!loop_vinfo)
return false;
loop = LOOP_VINFO_LOOP (loop_vinfo);
return (bb->loop_father == loop->inner);
}
/* Record the cost of a statement, either by directly informing the
target model or by saving it in a vector for later processing.
Return a preliminary estimate of the statement's cost. */
unsigned
record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count,
enum vect_cost_for_stmt kind, stmt_vec_info stmt_info,
int misalign, enum vect_cost_model_location where)
{
if ((kind == vector_load || kind == unaligned_load)
&& STMT_VINFO_GATHER_SCATTER_P (stmt_info))
kind = vector_gather_load;
if ((kind == vector_store || kind == unaligned_store)
&& STMT_VINFO_GATHER_SCATTER_P (stmt_info))
kind = vector_scatter_store;
if (body_cost_vec)
{
tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE;
stmt_info_for_cost si = { count, kind,
stmt_info ? STMT_VINFO_STMT (stmt_info) : NULL,
misalign };
body_cost_vec->safe_push (si);
return (unsigned)
(builtin_vectorization_cost (kind, vectype, misalign) * count);
}
else
return add_stmt_cost (stmt_info->vinfo->target_cost_data,
count, kind, stmt_info, misalign, where);
}
/* Return a variable of type ELEM_TYPE[NELEMS]. */
static tree
create_vector_array (tree elem_type, unsigned HOST_WIDE_INT nelems)
{
return create_tmp_var (build_array_type_nelts (elem_type, nelems),
"vect_array");
}
/* ARRAY is an array of vectors created by create_vector_array.
Return an SSA_NAME for the vector in index N. The reference
is part of the vectorization of STMT and the vector is associated
with scalar destination SCALAR_DEST. */
static tree
read_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree scalar_dest,
tree array, unsigned HOST_WIDE_INT n)
{
tree vect_type, vect, vect_name, array_ref;
gimple *new_stmt;
gcc_assert (TREE_CODE (TREE_TYPE (array)) == ARRAY_TYPE);
vect_type = TREE_TYPE (TREE_TYPE (array));
vect = vect_create_destination_var (scalar_dest, vect_type);
array_ref = build4 (ARRAY_REF, vect_type, array,
build_int_cst (size_type_node, n),
NULL_TREE, NULL_TREE);
new_stmt = gimple_build_assign (vect, array_ref);
vect_name = make_ssa_name (vect, new_stmt);
gimple_assign_set_lhs (new_stmt, vect_name);
vect_finish_stmt_generation (stmt, new_stmt, gsi);
return vect_name;
}
/* ARRAY is an array of vectors created by create_vector_array.
Emit code to store SSA_NAME VECT in index N of the array.
The store is part of the vectorization of STMT. */
static void
write_vector_array (gimple *stmt, gimple_stmt_iterator *gsi, tree vect,
tree array, unsigned HOST_WIDE_INT n)
{
tree array_ref;
gimple *new_stmt;
array_ref = build4 (ARRAY_REF, TREE_TYPE (vect), array,
build_int_cst (size_type_node, n),
NULL_TREE, NULL_TREE);
new_stmt = gimple_build_assign (array_ref, vect);
vect_finish_stmt_generation (stmt, new_stmt, gsi);
}
/* PTR is a pointer to an array of type TYPE. Return a representation
of *PTR. The memory reference replaces those in FIRST_DR
(and its group). */
static tree
create_array_ref (tree type, tree ptr, tree alias_ptr_type)
{
tree mem_ref;
mem_ref = build2 (MEM_REF, type, ptr, build_int_cst (alias_ptr_type, 0));
/* Arrays have the same alignment as their type. */
set_ptr_info_alignment (get_ptr_info (ptr), TYPE_ALIGN_UNIT (type), 0);
return mem_ref;
}
/* Utility functions used by vect_mark_stmts_to_be_vectorized. */
/* Function vect_mark_relevant.
Mark STMT as "relevant for vectorization" and add it to WORKLIST. */
static void
vect_mark_relevant (vec<gimple *> *worklist, gimple *stmt,
enum vect_relevant relevant, bool live_p)
{
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info);
bool save_live_p = STMT_VINFO_LIVE_P (stmt_info);
gimple *pattern_stmt;
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"mark relevant %d, live %d: ", relevant, live_p);
dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0);
}
/* If this stmt is an original stmt in a pattern, we might need to mark its
related pattern stmt instead of the original stmt. However, such stmts
may have their own uses that are not in any pattern, in such cases the
stmt itself should be marked. */
if (STMT_VINFO_IN_PATTERN_P (stmt_info))
{
/* This is the last stmt in a sequence that was detected as a
pattern that can potentially be vectorized. Don't mark the stmt
as relevant/live because it's not going to be vectorized.
Instead mark the pattern-stmt that replaces it. */
pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"last stmt in pattern. don't mark"
" relevant/live.\n");
stmt_info = vinfo_for_stmt (pattern_stmt);
gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt);
save_relevant = STMT_VINFO_RELEVANT (stmt_info);
save_live_p = STMT_VINFO_LIVE_P (stmt_info);
stmt = pattern_stmt;
}
STMT_VINFO_LIVE_P (stmt_info) |= live_p;
if (relevant > STMT_VINFO_RELEVANT (stmt_info))
STMT_VINFO_RELEVANT (stmt_info) = relevant;
if (STMT_VINFO_RELEVANT (stmt_info) == save_relevant
&& STMT_VINFO_LIVE_P (stmt_info) == save_live_p)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"already marked relevant/live.\n");
return;
}
worklist->safe_push (stmt);
}
/* Function is_simple_and_all_uses_invariant
Return true if STMT is simple and all uses of it are invariant. */
bool
is_simple_and_all_uses_invariant (gimple *stmt, loop_vec_info loop_vinfo)
{
tree op;
gimple *def_stmt;
ssa_op_iter iter;
if (!is_gimple_assign (stmt))
return false;
FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
{
enum vect_def_type dt = vect_uninitialized_def;
if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"use not simple.\n");
return false;
}
if (dt != vect_external_def && dt != vect_constant_def)
return false;
}
return true;
}
/* Function vect_stmt_relevant_p.
Return true if STMT in loop that is represented by LOOP_VINFO is
"relevant for vectorization".
A stmt is considered "relevant for vectorization" if:
- it has uses outside the loop.
- it has vdefs (it alters memory).
- control stmts in the loop (except for the exit condition).
CHECKME: what other side effects would the vectorizer allow? */
static bool
vect_stmt_relevant_p (gimple *stmt, loop_vec_info loop_vinfo,
enum vect_relevant *relevant, bool *live_p)
{
struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
ssa_op_iter op_iter;
imm_use_iterator imm_iter;
use_operand_p use_p;
def_operand_p def_p;
*relevant = vect_unused_in_scope;
*live_p = false;
/* cond stmt other than loop exit cond. */
if (is_ctrl_stmt (stmt)
&& STMT_VINFO_TYPE (vinfo_for_stmt (stmt))
!= loop_exit_ctrl_vec_info_type)
*relevant = vect_used_in_scope;
/* changing memory. */
if (gimple_code (stmt) != GIMPLE_PHI)
if (gimple_vdef (stmt)
&& !gimple_clobber_p (stmt))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"vec_stmt_relevant_p: stmt has vdefs.\n");
*relevant = vect_used_in_scope;
}
/* uses outside the loop. */
FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF)
{
FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p))
{
basic_block bb = gimple_bb (USE_STMT (use_p));
if (!flow_bb_inside_loop_p (loop, bb))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"vec_stmt_relevant_p: used out of loop.\n");
if (is_gimple_debug (USE_STMT (use_p)))
continue;
/* We expect all such uses to be in the loop exit phis
(because of loop closed form) */
gcc_assert (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI);
gcc_assert (bb == single_exit (loop)->dest);
*live_p = true;
}
}
}
if (*live_p && *relevant == vect_unused_in_scope
&& !is_simple_and_all_uses_invariant (stmt, loop_vinfo))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"vec_stmt_relevant_p: stmt live but not relevant.\n");
*relevant = vect_used_only_live;
}
return (*live_p || *relevant);
}
/* Function exist_non_indexing_operands_for_use_p
USE is one of the uses attached to STMT. Check if USE is
used in STMT for anything other than indexing an array. */
static bool
exist_non_indexing_operands_for_use_p (tree use, gimple *stmt)
{
tree operand;
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
/* USE corresponds to some operand in STMT. If there is no data
reference in STMT, then any operand that corresponds to USE
is not indexing an array. */
if (!STMT_VINFO_DATA_REF (stmt_info))
return true;
/* STMT has a data_ref. FORNOW this means that its of one of
the following forms:
-1- ARRAY_REF = var
-2- var = ARRAY_REF
(This should have been verified in analyze_data_refs).
'var' in the second case corresponds to a def, not a use,
so USE cannot correspond to any operands that are not used
for array indexing.
Therefore, all we need to check is if STMT falls into the
first case, and whether var corresponds to USE. */
if (!gimple_assign_copy_p (stmt))
{
if (is_gimple_call (stmt)
&& gimple_call_internal_p (stmt))
{
internal_fn ifn = gimple_call_internal_fn (stmt);
int mask_index = internal_fn_mask_index (ifn);
if (mask_index >= 0
&& use == gimple_call_arg (stmt, mask_index))
return true;
int stored_value_index = internal_fn_stored_value_index (ifn);
if (stored_value_index >= 0
&& use == gimple_call_arg (stmt, stored_value_index))
return true;
if (internal_gather_scatter_fn_p (ifn)
&& use == gimple_call_arg (stmt, 1))
return true;
}
return false;
}
if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME)
return false;
operand = gimple_assign_rhs1 (stmt);
if (TREE_CODE (operand) != SSA_NAME)
return false;
if (operand == use)
return true;
return false;
}
/*
Function process_use.
Inputs:
- a USE in STMT in a loop represented by LOOP_VINFO
- RELEVANT - enum value to be set in the STMT_VINFO of the stmt
that defined USE. This is done by calling mark_relevant and passing it
the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant).
- FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't
be performed.
Outputs:
Generally, LIVE_P and RELEVANT are used to define the liveness and
relevance info of the DEF_STMT of this USE:
STMT_VINFO_LIVE_P (DEF_STMT_info) <-- live_p
STMT_VINFO_RELEVANT (DEF_STMT_info) <-- relevant
Exceptions:
- case 1: If USE is used only for address computations (e.g. array indexing),
which does not need to be directly vectorized, then the liveness/relevance
of the respective DEF_STMT is left unchanged.
- case 2: If STMT is a reduction phi and DEF_STMT is a reduction stmt, we
skip DEF_STMT cause it had already been processed.
- case 3: If DEF_STMT and STMT are in different nests, then "relevant" will
be modified accordingly.
Return true if everything is as expected. Return false otherwise. */
static bool
process_use (gimple *stmt, tree use, loop_vec_info loop_vinfo,
enum vect_relevant relevant, vec<gimple *> *worklist,
bool force)
{
struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt);
stmt_vec_info dstmt_vinfo;
basic_block bb, def_bb;
gimple *def_stmt;
enum vect_def_type dt;
/* case 1: we are only interested in uses that need to be vectorized. Uses
that are used for address computation are not considered relevant. */
if (!force && !exist_non_indexing_operands_for_use_p (use, stmt))
return true;
if (!vect_is_simple_use (use, loop_vinfo, &def_stmt, &dt))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not vectorized: unsupported use in stmt.\n");
return false;
}
if (!def_stmt || gimple_nop_p (def_stmt))
return true;
def_bb = gimple_bb (def_stmt);
if (!flow_bb_inside_loop_p (loop, def_bb))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location, "def_stmt is out of loop.\n");
return true;
}
/* case 2: A reduction phi (STMT) defined by a reduction stmt (DEF_STMT).
DEF_STMT must have already been processed, because this should be the
only way that STMT, which is a reduction-phi, was put in the worklist,
as there should be no other uses for DEF_STMT in the loop. So we just
check that everything is as expected, and we are done. */
dstmt_vinfo = vinfo_for_stmt (def_stmt);
bb = gimple_bb (stmt);
if (gimple_code (stmt) == GIMPLE_PHI
&& STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def
&& gimple_code (def_stmt) != GIMPLE_PHI
&& STMT_VINFO_DEF_TYPE (dstmt_vinfo) == vect_reduction_def
&& bb->loop_father == def_bb->loop_father)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"reduc-stmt defining reduc-phi in the same nest.\n");
if (STMT_VINFO_IN_PATTERN_P (dstmt_vinfo))
dstmt_vinfo = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (dstmt_vinfo));
gcc_assert (STMT_VINFO_RELEVANT (dstmt_vinfo) < vect_used_by_reduction);
gcc_assert (STMT_VINFO_LIVE_P (dstmt_vinfo)
|| STMT_VINFO_RELEVANT (dstmt_vinfo) > vect_unused_in_scope);
return true;
}
/* case 3a: outer-loop stmt defining an inner-loop stmt:
outer-loop-header-bb:
d = def_stmt
inner-loop:
stmt # use (d)
outer-loop-tail-bb:
... */
if (flow_loop_nested_p (def_bb->loop_father, bb->loop_father))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"outer-loop def-stmt defining inner-loop stmt.\n");
switch (relevant)
{
case vect_unused_in_scope:
relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_nested_cycle) ?
vect_used_in_scope : vect_unused_in_scope;
break;
case vect_used_in_outer_by_reduction:
gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def);
relevant = vect_used_by_reduction;
break;
case vect_used_in_outer:
gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def);
relevant = vect_used_in_scope;
break;
case vect_used_in_scope:
break;
default:
gcc_unreachable ();
}
}
/* case 3b: inner-loop stmt defining an outer-loop stmt:
outer-loop-header-bb:
...
inner-loop:
d = def_stmt
outer-loop-tail-bb (or outer-loop-exit-bb in double reduction):
stmt # use (d) */
else if (flow_loop_nested_p (bb->loop_father, def_bb->loop_father))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"inner-loop def-stmt defining outer-loop stmt.\n");
switch (relevant)
{
case vect_unused_in_scope:
relevant = (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_reduction_def
|| STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_double_reduction_def) ?
vect_used_in_outer_by_reduction : vect_unused_in_scope;
break;
case vect_used_by_reduction:
case vect_used_only_live:
relevant = vect_used_in_outer_by_reduction;
break;
case vect_used_in_scope:
relevant = vect_used_in_outer;
break;
default:
gcc_unreachable ();
}
}
/* We are also not interested in uses on loop PHI backedges that are
inductions. Otherwise we'll needlessly vectorize the IV increment
and cause hybrid SLP for SLP inductions. Unless the PHI is live
of course. */
else if (gimple_code (stmt) == GIMPLE_PHI
&& STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_induction_def
&& ! STMT_VINFO_LIVE_P (stmt_vinfo)
&& (PHI_ARG_DEF_FROM_EDGE (stmt, loop_latch_edge (bb->loop_father))
== use))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"induction value on backedge.\n");
return true;
}
vect_mark_relevant (worklist, def_stmt, relevant, false);
return true;
}
/* Function vect_mark_stmts_to_be_vectorized.
Not all stmts in the loop need to be vectorized. For example:
for i...
for j...
1. T0 = i + j
2. T1 = a[T0]
3. j = j + 1
Stmt 1 and 3 do not need to be vectorized, because loop control and
addressing of vectorized data-refs are handled differently.
This pass detects such stmts. */
bool
vect_mark_stmts_to_be_vectorized (loop_vec_info loop_vinfo)
{
struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo);
unsigned int nbbs = loop->num_nodes;
gimple_stmt_iterator si;
gimple *stmt;
unsigned int i;
stmt_vec_info stmt_vinfo;
basic_block bb;
gimple *phi;
bool live_p;
enum vect_relevant relevant;
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"=== vect_mark_stmts_to_be_vectorized ===\n");
auto_vec<gimple *, 64> worklist;
/* 1. Init worklist. */
for (i = 0; i < nbbs; i++)
{
bb = bbs[i];
for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si))
{
phi = gsi_stmt (si);
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location, "init: phi relevant? ");
dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0);
}
if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p))
vect_mark_relevant (&worklist, phi, relevant, live_p);
}
for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
{
stmt = gsi_stmt (si);
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location, "init: stmt relevant? ");
dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0);
}
if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p))
vect_mark_relevant (&worklist, stmt, relevant, live_p);
}
}
/* 2. Process_worklist */
while (worklist.length () > 0)
{
use_operand_p use_p;
ssa_op_iter iter;
stmt = worklist.pop ();
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location, "worklist: examine stmt: ");
dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0);
}
/* Examine the USEs of STMT. For each USE, mark the stmt that defines it
(DEF_STMT) as relevant/irrelevant according to the relevance property
of STMT. */
stmt_vinfo = vinfo_for_stmt (stmt);
relevant = STMT_VINFO_RELEVANT (stmt_vinfo);
/* Generally, the relevance property of STMT (in STMT_VINFO_RELEVANT) is
propagated as is to the DEF_STMTs of its USEs.
One exception is when STMT has been identified as defining a reduction
variable; in this case we set the relevance to vect_used_by_reduction.
This is because we distinguish between two kinds of relevant stmts -
those that are used by a reduction computation, and those that are
(also) used by a regular computation. This allows us later on to
identify stmts that are used solely by a reduction, and therefore the
order of the results that they produce does not have to be kept. */
switch (STMT_VINFO_DEF_TYPE (stmt_vinfo))
{
case vect_reduction_def:
gcc_assert (relevant != vect_unused_in_scope);
if (relevant != vect_unused_in_scope
&& relevant != vect_used_in_scope
&& relevant != vect_used_by_reduction
&& relevant != vect_used_only_live)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"unsupported use of reduction.\n");
return false;
}
break;
case vect_nested_cycle:
if (relevant != vect_unused_in_scope
&& relevant != vect_used_in_outer_by_reduction
&& relevant != vect_used_in_outer)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"unsupported use of nested cycle.\n");
return false;
}
break;
case vect_double_reduction_def:
if (relevant != vect_unused_in_scope
&& relevant != vect_used_by_reduction
&& relevant != vect_used_only_live)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"unsupported use of double reduction.\n");
return false;
}
break;
default:
break;
}
if (is_pattern_stmt_p (stmt_vinfo))
{
/* Pattern statements are not inserted into the code, so
FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we
have to scan the RHS or function arguments instead. */
if (is_gimple_assign (stmt))
{
enum tree_code rhs_code = gimple_assign_rhs_code (stmt);
tree op = gimple_assign_rhs1 (stmt);
i = 1;
if (rhs_code == COND_EXPR && COMPARISON_CLASS_P (op))
{
if (!process_use (stmt, TREE_OPERAND (op, 0), loop_vinfo,
relevant, &worklist, false)
|| !process_use (stmt, TREE_OPERAND (op, 1), loop_vinfo,
relevant, &worklist, false))
return false;
i = 2;
}
for (; i < gimple_num_ops (stmt); i++)
{
op = gimple_op (stmt, i);
if (TREE_CODE (op) == SSA_NAME
&& !process_use (stmt, op, loop_vinfo, relevant,
&worklist, false))
return false;
}
}
else if (is_gimple_call (stmt))
{
for (i = 0; i < gimple_call_num_args (stmt); i++)
{
tree arg = gimple_call_arg (stmt, i);
if (!process_use (stmt, arg, loop_vinfo, relevant,
&worklist, false))
return false;
}
}
}
else
FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE)
{
tree op = USE_FROM_PTR (use_p);
if (!process_use (stmt, op, loop_vinfo, relevant,
&worklist, false))
return false;
}
if (STMT_VINFO_GATHER_SCATTER_P (stmt_vinfo))
{
gather_scatter_info gs_info;
if (!vect_check_gather_scatter (stmt, loop_vinfo, &gs_info))
gcc_unreachable ();
if (!process_use (stmt, gs_info.offset, loop_vinfo, relevant,
&worklist, true))
return false;
}
} /* while worklist */
return true;
}
/* 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,
int ndts,
stmt_vector_for_cost *prologue_cost_vec,
stmt_vector_for_cost *body_cost_vec)
{
int i;
int inside_cost = 0, prologue_cost = 0;
/* The SLP costs were already calculated during SLP tree build. */
gcc_assert (!PURE_SLP_STMT (stmt_info));
/* Cost the "broadcast" of a scalar operand in to a vector operand.
Use scalar_to_vec to cost the broadcast, as elsewhere in the vector
cost model. */
for (i = 0; i < ndts; i++)
if (dt[i] == vect_constant_def || dt[i] == vect_external_def)
prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec,
stmt_info, 0, vect_prologue);
/* Pass the inside-of-loop statements to the target-specific cost model. */
inside_cost = record_stmt_cost (body_cost_vec, ncopies, vector_stmt,
stmt_info, 0, vect_body);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"vect_model_simple_cost: inside_cost = %d, "
"prologue_cost = %d .\n", inside_cost, prologue_cost);
}
/* Model cost for type demotion and promotion operations. PWR is normally
zero for single-step promotions and demotions. It will be one if
two-step promotion/demotion is required, and so on. Each additional
step doubles the number of instructions required. */
static void
vect_model_promotion_demotion_cost (stmt_vec_info stmt_info,
enum vect_def_type *dt, int pwr)
{
int i, tmp;
int inside_cost = 0, prologue_cost = 0;
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
void *target_cost_data;
/* The SLP costs were already calculated during SLP tree build. */
gcc_assert (!PURE_SLP_STMT (stmt_info));
if (loop_vinfo)
target_cost_data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo);
else
target_cost_data = BB_VINFO_TARGET_COST_DATA (bb_vinfo);
for (i = 0; i < pwr + 1; i++)
{
tmp = (STMT_VINFO_TYPE (stmt_info) == type_promotion_vec_info_type) ?
(i + 1) : i;
inside_cost += add_stmt_cost (target_cost_data, vect_pow2 (tmp),
vec_promote_demote, stmt_info, 0,
vect_body);
}
/* FORNOW: Assuming maximum 2 args per stmts. */
for (i = 0; i < 2; i++)
if (dt[i] == vect_constant_def || dt[i] == vect_external_def)
prologue_cost += add_stmt_cost (target_cost_data, 1, vector_stmt,
stmt_info, 0, vect_prologue);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"vect_model_promotion_demotion_cost: inside_cost = %d, "
"prologue_cost = %d .\n", inside_cost, prologue_cost);
}
/* Function vect_model_store_cost
Models cost for stores. In the case of grouped accesses, one access
has the overhead of the grouped access attributed to it. */
void
vect_model_store_cost (stmt_vec_info stmt_info, int ncopies,
vect_memory_access_type memory_access_type,
vec_load_store_type vls_type, slp_tree slp_node,
stmt_vector_for_cost *prologue_cost_vec,
stmt_vector_for_cost *body_cost_vec)
{
unsigned int inside_cost = 0, prologue_cost = 0;
struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
gimple *first_stmt = STMT_VINFO_STMT (stmt_info);
bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info);
if (vls_type == VLS_STORE_INVARIANT)
prologue_cost += record_stmt_cost (prologue_cost_vec, 1, scalar_to_vec,
stmt_info, 0, vect_prologue);
/* Grouped stores update all elements in the group at once,
so we want the DR for the first statement. */
if (!slp_node && grouped_access_p)
{
first_stmt = GROUP_FIRST_ELEMENT (stmt_info);
dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
}
/* True if we should include any once-per-group costs as well as
the cost of the statement itself. For SLP we only get called
once per group anyhow. */
bool first_stmt_p = (first_stmt == STMT_VINFO_STMT (stmt_info));
/* We assume that the cost of a single store-lanes instruction is
equivalent to the cost of GROUP_SIZE separate stores. If a grouped
access is instead being provided by a permute-and-store operation,
include the cost of the permutes. */
if (first_stmt_p
&& memory_access_type == VMAT_CONTIGUOUS_PERMUTE)
{
/* Uses a high and low interleave or shuffle operations for each
needed permute. */
int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt));
int nstmts = ncopies * ceil_log2 (group_size) * group_size;
inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm,
stmt_info, 0, vect_body);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"vect_model_store_cost: strided group_size = %d .\n",
group_size);
}
tree vectype = STMT_VINFO_VECTYPE (stmt_info);
/* Costs of the stores. */
if (memory_access_type == VMAT_ELEMENTWISE
|| memory_access_type == VMAT_GATHER_SCATTER)
{
/* N scalar stores plus extracting the elements. */
unsigned int assumed_nunits = vect_nunits_for_cost (vectype);
inside_cost += record_stmt_cost (body_cost_vec,
ncopies * assumed_nunits,
scalar_store, stmt_info, 0, vect_body);
}
else
vect_get_store_cost (dr, ncopies, &inside_cost, body_cost_vec);
if (memory_access_type == VMAT_ELEMENTWISE
|| memory_access_type == VMAT_STRIDED_SLP)
{
/* N scalar stores plus extracting the elements. */
unsigned int assumed_nunits = vect_nunits_for_cost (vectype);
inside_cost += record_stmt_cost (body_cost_vec,
ncopies * assumed_nunits,
vec_to_scalar, stmt_info, 0, vect_body);
}
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"vect_model_store_cost: inside_cost = %d, "
"prologue_cost = %d .\n", inside_cost, prologue_cost);
}
/* Calculate cost of DR's memory access. */
void
vect_get_store_cost (struct data_reference *dr, int ncopies,
unsigned int *inside_cost,
stmt_vector_for_cost *body_cost_vec)
{
int alignment_support_scheme = vect_supportable_dr_alignment (dr, false);
gimple *stmt = DR_STMT (dr);
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
switch (alignment_support_scheme)
{
case dr_aligned:
{
*inside_cost += record_stmt_cost (body_cost_vec, ncopies,
vector_store, stmt_info, 0,
vect_body);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"vect_model_store_cost: aligned.\n");
break;
}
case dr_unaligned_supported:
{
/* Here, we assign an additional cost for the unaligned store. */
*inside_cost += record_stmt_cost (body_cost_vec, ncopies,
unaligned_store, stmt_info,
DR_MISALIGNMENT (dr), vect_body);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"vect_model_store_cost: unaligned supported by "
"hardware.\n");
break;
}
case dr_unaligned_unsupported:
{
*inside_cost = VECT_MAX_COST;
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"vect_model_store_cost: unsupported access.\n");
break;
}
default:
gcc_unreachable ();
}
}
/* Function vect_model_load_cost
Models cost for loads. In the case of grouped accesses, one access has
the overhead of the grouped 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,
vect_memory_access_type memory_access_type,
slp_tree slp_node,
stmt_vector_for_cost *prologue_cost_vec,
stmt_vector_for_cost *body_cost_vec)
{
gimple *first_stmt = STMT_VINFO_STMT (stmt_info);
struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
unsigned int inside_cost = 0, prologue_cost = 0;
bool grouped_access_p = STMT_VINFO_GROUPED_ACCESS (stmt_info);
/* Grouped loads read all elements in the group at once,
so we want the DR for the first statement. */
if (!slp_node && grouped_access_p)
{
first_stmt = GROUP_FIRST_ELEMENT (stmt_info);
dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
}
/* True if we should include any once-per-group costs as well as
the cost of the statement itself. For SLP we only get called
once per group anyhow. */
bool first_stmt_p = (first_stmt == STMT_VINFO_STMT (stmt_info));
/* We assume that the cost of a single load-lanes instruction is
equivalent to the cost of GROUP_SIZE separate loads. If a grouped
access is instead being provided by a load-and-permute operation,
include the cost of the permutes. */
if (first_stmt_p
&& memory_access_type == VMAT_CONTIGUOUS_PERMUTE)
{
/* Uses an even and odd extract operations or shuffle operations
for each needed permute. */
int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt));
int nstmts = ncopies * ceil_log2 (group_size) * group_size;
inside_cost = record_stmt_cost (body_cost_vec, nstmts, vec_perm,
stmt_info, 0, vect_body);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"vect_model_load_cost: strided group_size = %d .\n",
group_size);
}
/* The loads themselves. */
if (memory_access_type == VMAT_ELEMENTWISE
|| memory_access_type == VMAT_GATHER_SCATTER)
{
/* N scalar loads plus gathering them into a vector. */
tree vectype = STMT_VINFO_VECTYPE (stmt_info);
unsigned int assumed_nunits = vect_nunits_for_cost (vectype);
inside_cost += record_stmt_cost (body_cost_vec,
ncopies * assumed_nunits,
scalar_load, stmt_info, 0, vect_body);
}
else
vect_get_load_cost (dr, ncopies, first_stmt_p,
&inside_cost, &prologue_cost,
prologue_cost_vec, body_cost_vec, true);
if (memory_access_type == VMAT_ELEMENTWISE
|| memory_access_type == VMAT_STRIDED_SLP)
inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_construct,
stmt_info, 0, vect_body);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"vect_model_load_cost: inside_cost = %d, "
"prologue_cost = %d .\n", inside_cost, prologue_cost);
}
/* Calculate cost of DR's memory access. */
void
vect_get_load_cost (struct data_reference *dr, int ncopies,
bool add_realign_cost, unsigned int *inside_cost,
unsigned int *prologue_cost,
stmt_vector_for_cost *prologue_cost_vec,
stmt_vector_for_cost *body_cost_vec,
bool record_prologue_costs)
{
int alignment_support_scheme = vect_supportable_dr_alignment (dr, false);
gimple *stmt = DR_STMT (dr);
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
switch (alignment_support_scheme)
{
case dr_aligned:
{
*inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load,
stmt_info, 0, vect_body);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"vect_model_load_cost: aligned.\n");
break;
}
case dr_unaligned_supported:
{
/* Here, we assign an additional cost for the unaligned load. */
*inside_cost += record_stmt_cost (body_cost_vec, ncopies,
unaligned_load, stmt_info,
DR_MISALIGNMENT (dr), vect_body);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"vect_model_load_cost: unaligned supported by "
"hardware.\n");
break;
}
case dr_explicit_realign:
{
*inside_cost += record_stmt_cost (body_cost_vec, ncopies * 2,
vector_load, stmt_info, 0, vect_body);
*inside_cost += record_stmt_cost (body_cost_vec, ncopies,
vec_perm, stmt_info, 0, vect_body);
/* FIXME: If the misalignment remains fixed across the iterations of
the containing loop, the following cost should be added to the
prologue costs. */
if (targetm.vectorize.builtin_mask_for_load)
*inside_cost += record_stmt_cost (body_cost_vec, 1, vector_stmt,
stmt_info, 0, vect_body);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"vect_model_load_cost: explicit realign\n");
break;
}
case dr_explicit_realign_optimized:
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"vect_model_load_cost: unaligned software "
"pipelined.\n");
/* 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 grouped
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 (add_realign_cost && record_prologue_costs)
{
*prologue_cost += record_stmt_cost (prologue_cost_vec, 2,
vector_stmt, stmt_info,
0, vect_prologue);
if (targetm.vectorize.builtin_mask_for_load)
*prologue_cost += record_stmt_cost (prologue_cost_vec, 1,
vector_stmt, stmt_info,
0, vect_prologue);
}
*inside_cost += record_stmt_cost (body_cost_vec, ncopies, vector_load,
stmt_info, 0, vect_body);
*inside_cost += record_stmt_cost (body_cost_vec, ncopies, vec_perm,
stmt_info, 0, vect_body);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"vect_model_load_cost: explicit realign optimized"
"\n");
break;
}
case dr_unaligned_unsupported:
{
*inside_cost = VECT_MAX_COST;
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"vect_model_load_cost: unsupported access.\n");
break;
}
default:
gcc_unreachable ();
}
}
/* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in
the loop preheader for the vectorized stmt STMT. */
static void
vect_init_vector_1 (gimple *stmt, gimple *new_stmt, gimple_stmt_iterator *gsi)
{
if (gsi)
vect_finish_stmt_generation (stmt, new_stmt, gsi);
else
{
stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt);
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
if (loop_vinfo)
{
struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
basic_block new_bb;
edge pe;
if (nested_in_vect_loop_p (loop, stmt))
loop = loop->inner;
pe = loop_preheader_edge (loop);
new_bb = gsi_insert_on_edge_immediate (pe, new_stmt);
gcc_assert (!new_bb);
}
else
{
bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_vinfo);
basic_block bb;
gimple_stmt_iterator gsi_bb_start;
gcc_assert (bb_vinfo);
bb = BB_VINFO_BB (bb_vinfo);
gsi_bb_start = gsi_after_labels (bb);
gsi_insert_before (&gsi_bb_start, new_stmt, GSI_SAME_STMT);
}
}
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"created new init_stmt: ");
dump_gimple_stmt (MSG_NOTE, TDF_SLIM, new_stmt, 0);
}
}
/* Function vect_init_vector.
Insert a new stmt (INIT_STMT) that initializes a new variable of type
TYPE with the value VAL. If TYPE is a vector type and VAL does not have
vector type a vector with all elements equal to VAL is created first.
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. */
tree
vect_init_vector (gimple *stmt, tree val, tree type, gimple_stmt_iterator *gsi)
{
gimple *init_stmt;
tree new_temp;
/* We abuse this function to push sth to a SSA name with initial 'val'. */
if (! useless_type_conversion_p (type, TREE_TYPE (val)))
{
gcc_assert (TREE_CODE (type) == VECTOR_TYPE);
if (! types_compatible_p (TREE_TYPE (type), TREE_TYPE (val)))
{
/* Scalar boolean value should be transformed into
all zeros or all ones value before building a vector. */
if (VECTOR_BOOLEAN_TYPE_P (type))
{
tree true_val = build_all_ones_cst (TREE_TYPE (type));
tree false_val = build_zero_cst (TREE_TYPE (type));
if (CONSTANT_CLASS_P (val))
val = integer_zerop (val) ? false_val : true_val;
else
{
new_temp = make_ssa_name (TREE_TYPE (type));
init_stmt = gimple_build_assign (new_temp, COND_EXPR,
val, true_val, false_val);
vect_init_vector_1 (stmt, init_stmt, gsi);
val = new_temp;
}
}
else if (CONSTANT_CLASS_P (val))
val = fold_convert (TREE_TYPE (type), val);
else
{
new_temp = make_ssa_name (TREE_TYPE (type));
if (! INTEGRAL_TYPE_P (TREE_TYPE (val)))
init_stmt = gimple_build_assign (new_temp,
fold_build1 (VIEW_CONVERT_EXPR,
TREE_TYPE (type),
val));
else
init_stmt = gimple_build_assign (new_temp, NOP_EXPR, val);
vect_init_vector_1 (stmt, init_stmt, gsi);
val = new_temp;
}
}
val = build_vector_from_val (type, val);
}
new_temp = vect_get_new_ssa_name (type, vect_simple_var, "cst_");
init_stmt = gimple_build_assign (new_temp, val);
vect_init_vector_1 (stmt, init_stmt, gsi);
return new_temp;
}
/* Function vect_get_vec_def_for_operand_1.
For a defining stmt DEF_STMT of a scalar stmt, return a vector def with type
DT that will be used in the vectorized stmt. */
tree
vect_get_vec_def_for_operand_1 (gimple *def_stmt, enum vect_def_type dt)
{
tree vec_oprnd;
gimple *vec_stmt;
stmt_vec_info def_stmt_info = NULL;
switch (dt)
{
/* operand is a constant or a loop invariant. */
case vect_constant_def:
case vect_external_def:
/* Code should use vect_get_vec_def_for_operand. */
gcc_unreachable ();
/* operand is defined inside the loop. */
case vect_internal_def:
{
/* Get the def from the vectorized stmt. */
def_stmt_info = vinfo_for_stmt (def_stmt);
vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info);
/* Get vectorized pattern statement. */
if (!vec_stmt
&& STMT_VINFO_IN_PATTERN_P (def_stmt_info)
&& !STMT_VINFO_RELEVANT (def_stmt_info))
vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt (
STMT_VINFO_RELATED_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;
}
/* operand is defined by a loop header phi. */
case vect_reduction_def:
case vect_double_reduction_def:
case vect_nested_cycle:
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);
if (gimple_code (vec_stmt) == GIMPLE_PHI)
vec_oprnd = PHI_RESULT (vec_stmt);
else
vec_oprnd = gimple_get_lhs (vec_stmt);
return vec_oprnd;
}
default:
gcc_unreachable ();
}
}
/* 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. VECTYPE may be used to specify a required type for
vector invariant. */
tree
vect_get_vec_def_for_operand (tree op, gimple *stmt, tree vectype)
{
gimple *def_stmt;
enum vect_def_type dt;
bool is_simple_use;
stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt);
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"vect_get_vec_def_for_operand: ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, op);
dump_printf (MSG_NOTE, "\n");
}
is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &dt);
gcc_assert (is_simple_use);
if (def_stmt && dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location, " def_stmt = ");
dump_gimple_stmt (MSG_NOTE, TDF_SLIM, def_stmt, 0);
}
if (dt == vect_constant_def || dt == vect_external_def)
{
tree stmt_vectype = STMT_VINFO_VECTYPE (stmt_vinfo);
tree vector_type;
if (vectype)
vector_type = vectype;
else if (VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (op))
&& VECTOR_BOOLEAN_TYPE_P (stmt_vectype))
vector_type = build_same_sized_truth_vector_type (stmt_vectype);
else
vector_type = get_vectype_for_scalar_type (TREE_TYPE (op));
gcc_assert (vector_type);
return vect_init_vector (stmt, op, vector_type, NULL);
}
else
return vect_get_vec_def_for_operand_1 (def_stmt, dt);
}
/* 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); */
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_external_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);
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. */
void
vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt,
vec<tree> *vec_oprnds0,
vec<tree> *vec_oprnds1)
{
tree vec_oprnd = vec_oprnds0->pop ();
vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd);
vec_oprnds0->quick_push (vec_oprnd);
if (vec_oprnds1 && vec_oprnds1->length ())
{
vec_oprnd = vec_oprnds1->pop ();
vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd);
vec_oprnds1->quick_push (vec_oprnd);
}
}
/* Get vectorized definitions for OP0 and OP1. */
void
vect_get_vec_defs (tree op0, tree op1, gimple *stmt,
vec<tree> *vec_oprnds0,
vec<tree> *vec_oprnds1,
slp_tree slp_node)
{
if (slp_node)
{
int nops = (op1 == NULL_TREE) ? 1 : 2;
auto_vec<tree> ops (nops);
auto_vec<vec<tree> > vec_defs (nops);
ops.quick_push (op0);
if (op1)
ops.quick_push (op1);
vect_get_slp_defs (ops, slp_node, &vec_defs);
*vec_oprnds0 = vec_defs[0];
if (op1)
*vec_oprnds1 = vec_defs[1];
}
else
{
tree vec_oprnd;
vec_oprnds0->create (1);
vec_oprnd = vect_get_vec_def_for_operand (op0, stmt);
vec_oprnds0->quick_push (vec_oprnd);
if (op1)
{
vec_oprnds1->create (1);
vec_oprnd = vect_get_vec_def_for_operand (op1, stmt);
vec_oprnds1->quick_push (vec_oprnd);
}
}
}
/* Helper function called by vect_finish_replace_stmt and
vect_finish_stmt_generation. Set the location of the new
statement and create a stmt_vec_info for it. */
static void
vect_finish_stmt_generation_1 (gimple *stmt, gimple *vec_stmt)
{
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
vec_info *vinfo = stmt_info->vinfo;
set_vinfo_for_stmt (vec_stmt, new_stmt_vec_info (vec_stmt, vinfo));
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location, "add new stmt: ");
dump_gimple_stmt (MSG_NOTE, TDF_SLIM, vec_stmt, 0);
}
gimple_set_location (vec_stmt, gimple_location (stmt));
/* While EH edges will generally prevent vectorization, stmt might
e.g. be in a must-not-throw region. Ensure newly created stmts
that could throw are part of the same region. */
int lp_nr = lookup_stmt_eh_lp (stmt);
if (lp_nr != 0 && stmt_could_throw_p (vec_stmt))
add_stmt_to_eh_lp (vec_stmt, lp_nr);
}
/* Replace the scalar statement STMT with a new vector statement VEC_STMT,
which sets the same scalar result as STMT did. */
void
vect_finish_replace_stmt (gimple *stmt, gimple *vec_stmt)
{
gcc_assert (gimple_get_lhs (stmt) == gimple_get_lhs (vec_stmt));
gimple_stmt_iterator gsi = gsi_for_stmt (stmt);
gsi_replace (&gsi, vec_stmt, false);
vect_finish_stmt_generation_1 (stmt, vec_stmt);
}
/* Function vect_finish_stmt_generation.
Insert a new stmt. */
void
vect_finish_stmt_generation (gimple *stmt, gimple *vec_stmt,
gimple_stmt_iterator *gsi)
{
gcc_assert (gimple_code (stmt) != GIMPLE_LABEL);
if (!gsi_end_p (*gsi)
&& gimple_has_mem_ops (vec_stmt))
{
gimple *at_stmt = gsi_stmt (*gsi);
tree vuse = gimple_vuse (at_stmt);
if (vuse && TREE_CODE (vuse) == SSA_NAME)
{
tree vdef = gimple_vdef (at_stmt);
gimple_set_vuse (vec_stmt, gimple_vuse (at_stmt));
/* If we have an SSA vuse and insert a store, update virtual
SSA form to avoid triggering the renamer. Do so only
if we can easily see all uses - which is what almost always
happens with the way vectorized stmts are inserted. */
if ((vdef && TREE_CODE (vdef) == SSA_NAME)
&& ((is_gimple_assign (vec_stmt)
&& !is_gimple_reg (gimple_assign_lhs (vec_stmt)))
|| (is_gimple_call (vec_stmt)
&& !(gimple_call_flags (vec_stmt)
& (ECF_CONST|ECF_PURE|ECF_NOVOPS)))))
{
tree new_vdef = copy_ssa_name (vuse, vec_stmt);
gimple_set_vdef (vec_stmt, new_vdef);
SET_USE (gimple_vuse_op (at_stmt), new_vdef);
}
}
}
gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT);
vect_finish_stmt_generation_1 (stmt, vec_stmt);
}
/* We want to vectorize a call to combined function CFN with function
decl FNDECL, using VECTYPE_OUT as the type of the output and VECTYPE_IN
as the types of all inputs. Check whether this is possible using
an internal function, returning its code if so or IFN_LAST if not. */
static internal_fn
vectorizable_internal_function (combined_fn cfn, tree fndecl,
tree vectype_out, tree vectype_in)
{
internal_fn ifn;
if (internal_fn_p (cfn))
ifn = as_internal_fn (cfn);
else
ifn = associated_internal_fn (fndecl);
if (ifn != IFN_LAST && direct_internal_fn_p (ifn))
{
const direct_internal_fn_info &info = direct_internal_fn (ifn);
if (info.vectorizable)
{
tree type0 = (info.type0 < 0 ? vectype_out : vectype_in);
tree type1 = (info.type1 < 0 ? vectype_out : vectype_in);
if (direct_internal_fn_supported_p (ifn, tree_pair (type0, type1),
OPTIMIZE_FOR_SPEED))
return ifn;
}
}
return IFN_LAST;
}
static tree permute_vec_elements (tree, tree, tree, gimple *,
gimple_stmt_iterator *);
/* Check whether a load or store statement in the loop described by
LOOP_VINFO is possible in a fully-masked loop. This is testing
whether the vectorizer pass has the appropriate support, as well as
whether the target does.
VLS_TYPE says whether the statement is a load or store and VECTYPE
is the type of the vector being loaded or stored. MEMORY_ACCESS_TYPE
says how the load or store is going to be implemented and GROUP_SIZE
is the number of load or store statements in the containing group.
If the access is a gather load or scatter store, GS_INFO describes
its arguments.
Clear LOOP_VINFO_CAN_FULLY_MASK_P if a fully-masked loop is not
supported, otherwise record the required mask types. */
static void
check_load_store_masking (loop_vec_info loop_vinfo, tree vectype,
vec_load_store_type vls_type, int group_size,
vect_memory_access_type memory_access_type,
gather_scatter_info *gs_info)
{
/* Invariant loads need no special support. */
if (memory_access_type == VMAT_INVARIANT)
return;
vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo);
machine_mode vecmode = TYPE_MODE (vectype);
bool is_load = (vls_type == VLS_LOAD);
if (memory_access_type == VMAT_LOAD_STORE_LANES)
{
if (is_load
? !vect_load_lanes_supported (vectype, group_size, true)
: !vect_store_lanes_supported (vectype, group_size, true))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"can't use a fully-masked loop because the"
" target doesn't have an appropriate masked"
" load/store-lanes instruction.\n");
LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false;
return;
}
unsigned int ncopies = vect_get_num_copies (loop_vinfo, vectype);
vect_record_loop_mask (loop_vinfo, masks, ncopies, vectype);
return;
}
if (memory_access_type == VMAT_GATHER_SCATTER)
{
internal_fn ifn = (is_load
? IFN_MASK_GATHER_LOAD
: IFN_MASK_SCATTER_STORE);
tree offset_type = TREE_TYPE (gs_info->offset);
if (!internal_gather_scatter_fn_supported_p (ifn, vectype,
gs_info->memory_type,
TYPE_SIGN (offset_type),
gs_info->scale))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"can't use a fully-masked loop because the"
" target doesn't have an appropriate masked"
" gather load or scatter store instruction.\n");
LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false;
return;
}
unsigned int ncopies = vect_get_num_copies (loop_vinfo, vectype);
vect_record_loop_mask (loop_vinfo, masks, ncopies, vectype);
return;
}
if (memory_access_type != VMAT_CONTIGUOUS
&& memory_access_type != VMAT_CONTIGUOUS_PERMUTE)
{
/* Element X of the data must come from iteration i * VF + X of the
scalar loop. We need more work to support other mappings. */
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"can't use a fully-masked loop because an access"
" isn't contiguous.\n");
LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false;
return;
}
machine_mode mask_mode;
if (!(targetm.vectorize.get_mask_mode
(GET_MODE_NUNITS (vecmode),
GET_MODE_SIZE (vecmode)).exists (&mask_mode))
|| !can_vec_mask_load_store_p (vecmode, mask_mode, is_load))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"can't use a fully-masked loop because the target"
" doesn't have the appropriate masked load or"
" store.\n");
LOOP_VINFO_CAN_FULLY_MASK_P (loop_vinfo) = false;
return;
}
/* We might load more scalars than we need for permuting SLP loads.
We checked in get_group_load_store_type that the extra elements
don't leak into a new vector. */
poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
unsigned int nvectors;
if (can_div_away_from_zero_p (group_size * vf, nunits, &nvectors))
vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype);
else
gcc_unreachable ();
}
/* Return the mask input to a masked load or store. VEC_MASK is the vectorized
form of the scalar mask condition and LOOP_MASK, if nonnull, is the mask
that needs to be applied to all loads and stores in a vectorized loop.
Return VEC_MASK if LOOP_MASK is null, otherwise return VEC_MASK & LOOP_MASK.
MASK_TYPE is the type of both masks. If new statements are needed,
insert them before GSI. */
static tree
prepare_load_store_mask (tree mask_type, tree loop_mask, tree vec_mask,
gimple_stmt_iterator *gsi)
{
gcc_assert (useless_type_conversion_p (mask_type, TREE_TYPE (vec_mask)));
if (!loop_mask)
return vec_mask;
gcc_assert (TREE_TYPE (loop_mask) == mask_type);
tree and_res = make_temp_ssa_name (mask_type, NULL, "vec_mask_and");
gimple *and_stmt = gimple_build_assign (and_res, BIT_AND_EXPR,
vec_mask, loop_mask);
gsi_insert_before (gsi, and_stmt, GSI_SAME_STMT);
return and_res;
}
/* Determine whether we can use a gather load or scatter store to vectorize
strided load or store STMT by truncating the current offset to a smaller
width. We need to be able to construct an offset vector:
{ 0, X, X*2, X*3, ... }
without loss of precision, where X is STMT's DR_STEP.
Return true if this is possible, describing the gather load or scatter
store in GS_INFO. MASKED_P is true if the load or store is conditional. */
static bool
vect_truncate_gather_scatter_offset (gimple *stmt, loop_vec_info loop_vinfo,
bool masked_p,
gather_scatter_info *gs_info)
{
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
tree step = DR_STEP (dr);
if (TREE_CODE (step) != INTEGER_CST)
{
/* ??? Perhaps we could use range information here? */
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"cannot truncate variable step.\n");
return false;
}
/* Get the number of bits in an element. */
tree vectype = STMT_VINFO_VECTYPE (stmt_info);
scalar_mode element_mode = SCALAR_TYPE_MODE (TREE_TYPE (vectype));
unsigned int element_bits = GET_MODE_BITSIZE (element_mode);
/* Set COUNT to the upper limit on the number of elements - 1.
Start with the maximum vectorization factor. */
unsigned HOST_WIDE_INT count = vect_max_vf (loop_vinfo) - 1;
/* Try lowering COUNT to the number of scalar latch iterations. */
struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
widest_int max_iters;
if (max_loop_iterations (loop, &max_iters)
&& max_iters < count)
count = max_iters.to_shwi ();
/* Try scales of 1 and the element size. */
int scales[] = { 1, vect_get_scalar_dr_size (dr) };
bool overflow_p = false;
for (int i = 0; i < 2; ++i)
{
int scale = scales[i];
widest_int factor;
if (!wi::multiple_of_p (wi::to_widest (step), scale, SIGNED, &factor))
continue;
/* See whether we can calculate (COUNT - 1) * STEP / SCALE
in OFFSET_BITS bits. */
widest_int range = wi::mul (count, factor, SIGNED, &overflow_p);
if (overflow_p)
continue;
signop sign = range >= 0 ? UNSIGNED : SIGNED;
if (wi::min_precision (range, sign) > element_bits)
{
overflow_p = true;
continue;
}
/* See whether the target supports the operation. */
tree memory_type = TREE_TYPE (DR_REF (dr));
if (!vect_gather_scatter_fn_p (DR_IS_READ (dr), masked_p, vectype,
memory_type, element_bits, sign, scale,
&gs_info->ifn, &gs_info->element_type))
continue;
tree offset_type = build_nonstandard_integer_type (element_bits,
sign == UNSIGNED);
gs_info->decl = NULL_TREE;
/* Logically the sum of DR_BASE_ADDRESS, DR_INIT and DR_OFFSET,
but we don't need to store that here. */
gs_info->base = NULL_TREE;
gs_info->offset = fold_convert (offset_type, step);
gs_info->offset_dt = vect_constant_def;
gs_info->offset_vectype = NULL_TREE;
gs_info->scale = scale;
gs_info->memory_type = memory_type;
return true;
}
if (overflow_p && dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"truncating gather/scatter offset to %d bits"
" might change its value.\n", element_bits);
return false;
}
/* Return true if we can use gather/scatter internal functions to
vectorize STMT, which is a grouped or strided load or store.
MASKED_P is true if load or store is conditional. When returning
true, fill in GS_INFO with the information required to perform the
operation. */
static bool
vect_use_strided_gather_scatters_p (gimple *stmt, loop_vec_info loop_vinfo,
bool masked_p,
gather_scatter_info *gs_info)
{
if (!vect_check_gather_scatter (stmt, loop_vinfo, gs_info)
|| gs_info->decl)
return vect_truncate_gather_scatter_offset (stmt, loop_vinfo,
masked_p, gs_info);
scalar_mode element_mode = SCALAR_TYPE_MODE (gs_info->element_type);
unsigned int element_bits = GET_MODE_BITSIZE (element_mode);
tree offset_type = TREE_TYPE (gs_info->offset);
unsigned int offset_bits = TYPE_PRECISION (offset_type);
/* Enforced by vect_check_gather_scatter. */
gcc_assert (element_bits >= offset_bits);
/* If the elements are wider than the offset, convert the offset to the
same width, without changing its sign. */
if (element_bits > offset_bits)
{
bool unsigned_p = TYPE_UNSIGNED (offset_type);
offset_type = build_nonstandard_integer_type (element_bits, unsigned_p);
gs_info->offset = fold_convert (offset_type, gs_info->offset);
}
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"using gather/scatter for strided/grouped access,"
" scale = %d\n", gs_info->scale);
return true;
}
/* STMT is a non-strided load or store, meaning that it accesses
elements with a known constant step. Return -1 if that step
is negative, 0 if it is zero, and 1 if it is greater than zero. */
static int
compare_step_with_zero (gimple *stmt)
{
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
return tree_int_cst_compare (vect_dr_behavior (dr)->step,
size_zero_node);
}
/* If the target supports a permute mask that reverses the elements in
a vector of type VECTYPE, return that mask, otherwise return null. */
static tree
perm_mask_for_reverse (tree vectype)
{
poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
/* The encoding has a single stepped pattern. */
vec_perm_builder sel (nunits, 1, 3);
for (int i = 0; i < 3; ++i)
sel.quick_push (nunits - 1 - i);
vec_perm_indices indices (sel, 1, nunits);
if (!can_vec_perm_const_p (TYPE_MODE (vectype), indices))
return NULL_TREE;
return vect_gen_perm_mask_checked (vectype, indices);
}
/* STMT is either a masked or unconditional store. Return the value
being stored. */
tree
vect_get_store_rhs (gimple *stmt)
{
if (gassign *assign = dyn_cast <gassign *> (stmt))
{
gcc_assert (gimple_assign_single_p (assign));
return gimple_assign_rhs1 (assign);
}
if (gcall *call = dyn_cast <gcall *> (stmt))
{
internal_fn ifn = gimple_call_internal_fn (call);
int index = internal_fn_stored_value_index (ifn);
gcc_assert (index >= 0);
return gimple_call_arg (stmt, index);
}
gcc_unreachable ();
}
/* A subroutine of get_load_store_type, with a subset of the same
arguments. Handle the case where STMT is part of a grouped load
or store.
For stores, the statements in the group are all consecutive
and there is no gap at the end. For loads, the statements in the
group might not be consecutive; there can be gaps between statements
as well as at the end. */
static bool
get_group_load_store_type (gimple *stmt, tree vectype, bool slp,
bool masked_p, vec_load_store_type vls_type,
vect_memory_access_type *memory_access_type,
gather_scatter_info *gs_info)
{
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
vec_info *vinfo = stmt_info->vinfo;
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL;
gimple *first_stmt = GROUP_FIRST_ELEMENT (stmt_info);
data_reference *first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
unsigned int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt));
bool single_element_p = (stmt == first_stmt
&& !GROUP_NEXT_ELEMENT (stmt_info));
unsigned HOST_WIDE_INT gap = GROUP_GAP (vinfo_for_stmt (first_stmt));
poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
/* True if the vectorized statements would access beyond the last
statement in the group. */
bool overrun_p = false;
/* True if we can cope with such overrun by peeling for gaps, so that
there is at least one final scalar iteration after the vector loop. */
bool can_overrun_p = (!masked_p
&& vls_type == VLS_LOAD
&& loop_vinfo
&& !loop->inner);
/* There can only be a gap at the end of the group if the stride is
known at compile time. */
gcc_assert (!STMT_VINFO_STRIDED_P (stmt_info) || gap == 0);
/* Stores can't yet have gaps. */
gcc_assert (slp || vls_type == VLS_LOAD || gap == 0);
if (slp)
{
if (STMT_VINFO_STRIDED_P (stmt_info))
{
/* Try to use consecutive accesses of GROUP_SIZE elements,
separated by the stride, until we have a complete vector.
Fall back to scalar accesses if that isn't possible. */
if (multiple_p (nunits, group_size))
*memory_access_type = VMAT_STRIDED_SLP;
else
*memory_access_type = VMAT_ELEMENTWISE;
}
else
{
overrun_p = loop_vinfo && gap != 0;
if (overrun_p && vls_type != VLS_LOAD)
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Grouped store with gaps requires"
" non-consecutive accesses\n");
return false;
}
/* An overrun is fine if the trailing elements are smaller
than the alignment boundary B. Every vector access will
be a multiple of B and so we are guaranteed to access a
non-gap element in the same B-sized block. */
if (overrun_p
&& gap < (vect_known_alignment_in_bytes (first_dr)
/ vect_get_scalar_dr_size (first_dr)))
overrun_p = false;
if (overrun_p && !can_overrun_p)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Peeling for outer loop is not supported\n");
return false;
}
*memory_access_type = VMAT_CONTIGUOUS;
}
}
else
{
/* We can always handle this case using elementwise accesses,
but see if something more efficient is available. */
*memory_access_type = VMAT_ELEMENTWISE;
/* If there is a gap at the end of the group then these optimizations
would access excess elements in the last iteration. */
bool would_overrun_p = (gap != 0);
/* An overrun is fine if the trailing elements are smaller than the
alignment boundary B. Every vector access will be a multiple of B
and so we are guaranteed to access a non-gap element in the
same B-sized block. */
if (would_overrun_p
&& !masked_p
&& gap < (vect_known_alignment_in_bytes (first_dr)
/ vect_get_scalar_dr_size (first_dr)))
would_overrun_p = false;
if (!STMT_VINFO_STRIDED_P (stmt_info)
&& (can_overrun_p || !would_overrun_p)
&& compare_step_with_zero (stmt) > 0)
{
/* First cope with the degenerate case of a single-element
vector. */
if (known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U))
*memory_access_type = VMAT_CONTIGUOUS;
/* Otherwise try using LOAD/STORE_LANES. */
if (*memory_access_type == VMAT_ELEMENTWISE
&& (vls_type == VLS_LOAD
? vect_load_lanes_supported (vectype, group_size, masked_p)
: vect_store_lanes_supported (vectype, group_size,
masked_p)))
{
*memory_access_type = VMAT_LOAD_STORE_LANES;
overrun_p = would_overrun_p;
}
/* If that fails, try using permuting loads. */
if (*memory_access_type == VMAT_ELEMENTWISE
&& (vls_type == VLS_LOAD
? vect_grouped_load_supported (vectype, single_element_p,
group_size)
: vect_grouped_store_supported (vectype, group_size)))
{
*memory_access_type = VMAT_CONTIGUOUS_PERMUTE;
overrun_p = would_overrun_p;
}
}
/* As a last resort, trying using a gather load or scatter store.
??? Although the code can handle all group sizes correctly,
it probably isn't a win to use separate strided accesses based
on nearby locations. Or, even if it's a win over scalar code,
it might not be a win over vectorizing at a lower VF, if that
allows us to use contiguous accesses. */
if (*memory_access_type == VMAT_ELEMENTWISE
&& single_element_p
&& loop_vinfo
&& vect_use_strided_gather_scatters_p (stmt, loop_vinfo,
masked_p, gs_info))
*memory_access_type = VMAT_GATHER_SCATTER;
}
if (vls_type != VLS_LOAD && first_stmt == stmt)
{
/* STMT is the leader of the group. Check the operands of all the
stmts of the group. */
gimple *next_stmt = GROUP_NEXT_ELEMENT (stmt_info);
while (next_stmt)
{
tree op = vect_get_store_rhs (next_stmt);
gimple *def_stmt;
enum vect_def_type dt;
if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"use not simple.\n");
return false;
}
next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt));
}
}
if (overrun_p)
{
gcc_assert (can_overrun_p);
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Data access with gaps requires scalar "
"epilogue loop\n");
LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true;
}
return true;
}
/* A subroutine of get_load_store_type, with a subset of the same
arguments. Handle the case where STMT is a load or store that
accesses consecutive elements with a negative step. */
static vect_memory_access_type
get_negative_load_store_type (gimple *stmt, tree vectype,
vec_load_store_type vls_type,
unsigned int ncopies)
{
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
dr_alignment_support alignment_support_scheme;
if (ncopies > 1)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"multiple types with negative step.\n");
return VMAT_ELEMENTWISE;
}
alignment_support_scheme = vect_supportable_dr_alignment (dr, false);
if (alignment_support_scheme != dr_aligned
&& alignment_support_scheme != dr_unaligned_supported)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"negative step but alignment required.\n");
return VMAT_ELEMENTWISE;
}
if (vls_type == VLS_STORE_INVARIANT)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"negative step with invariant source;"
" no permute needed.\n");
return VMAT_CONTIGUOUS_DOWN;
}
if (!perm_mask_for_reverse (vectype))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"negative step and reversing not supported.\n");
return VMAT_ELEMENTWISE;
}
return VMAT_CONTIGUOUS_REVERSE;
}
/* Analyze load or store statement STMT of type VLS_TYPE. Return true
if there is a memory access type that the vectorized form can use,
storing it in *MEMORY_ACCESS_TYPE if so. If we decide to use gathers
or scatters, fill in GS_INFO accordingly.
SLP says whether we're performing SLP rather than loop vectorization.
MASKED_P is true if the statement is conditional on a vectorized mask.
VECTYPE is the vector type that the vectorized statements will use.
NCOPIES is the number of vector statements that will be needed. */
static bool
get_load_store_type (gimple *stmt, tree vectype, bool slp, bool masked_p,
vec_load_store_type vls_type, unsigned int ncopies,
vect_memory_access_type *memory_access_type,
gather_scatter_info *gs_info)
{
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
vec_info *vinfo = stmt_info->vinfo;
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
if (STMT_VINFO_GATHER_SCATTER_P (stmt_info))
{
*memory_access_type = VMAT_GATHER_SCATTER;
gimple *def_stmt;
if (!vect_check_gather_scatter (stmt, loop_vinfo, gs_info))
gcc_unreachable ();
else if (!vect_is_simple_use (gs_info->offset, vinfo, &def_stmt,
&gs_info->offset_dt,
&gs_info->offset_vectype))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"%s index use not simple.\n",
vls_type == VLS_LOAD ? "gather" : "scatter");
return false;
}
}
else if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
{
if (!get_group_load_store_type (stmt, vectype, slp, masked_p, vls_type,
memory_access_type, gs_info))
return false;
}
else if (STMT_VINFO_STRIDED_P (stmt_info))
{
gcc_assert (!slp);
if (loop_vinfo
&& vect_use_strided_gather_scatters_p (stmt, loop_vinfo,
masked_p, gs_info))
*memory_access_type = VMAT_GATHER_SCATTER;
else
*memory_access_type = VMAT_ELEMENTWISE;
}
else
{
int cmp = compare_step_with_zero (stmt);
if (cmp < 0)
*memory_access_type = get_negative_load_store_type
(stmt, vectype, vls_type, ncopies);
else if (cmp == 0)
{
gcc_assert (vls_type == VLS_LOAD);
*memory_access_type = VMAT_INVARIANT;
}
else
*memory_access_type = VMAT_CONTIGUOUS;
}
if ((*memory_access_type == VMAT_ELEMENTWISE
|| *memory_access_type == VMAT_STRIDED_SLP)
&& !nunits.is_constant ())
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Not using elementwise accesses due to variable "
"vectorization factor.\n");
return false;
}
/* FIXME: At the moment the cost model seems to underestimate the
cost of using elementwise accesses. This check preserves the
traditional behavior until that can be fixed. */
if (*memory_access_type == VMAT_ELEMENTWISE
&& !STMT_VINFO_STRIDED_P (stmt_info)
&& !(stmt == GROUP_FIRST_ELEMENT (stmt_info)
&& !GROUP_NEXT_ELEMENT (stmt_info)
&& !pow2p_hwi (GROUP_SIZE (stmt_info))))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not falling back to elementwise accesses\n");
return false;
}
return true;
}
/* Return true if boolean argument MASK is suitable for vectorizing
conditional load or store STMT. When returning true, store the type
of the definition in *MASK_DT_OUT and the type of the vectorized mask
in *MASK_VECTYPE_OUT. */
static bool
vect_check_load_store_mask (gimple *stmt, tree mask,
vect_def_type *mask_dt_out,
tree *mask_vectype_out)
{
if (!VECT_SCALAR_BOOLEAN_TYPE_P (TREE_TYPE (mask)))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"mask argument is not a boolean.\n");
return false;
}
if (TREE_CODE (mask) != SSA_NAME)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"mask argument is not an SSA name.\n");
return false;
}
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
gimple *def_stmt;
enum vect_def_type mask_dt;
tree mask_vectype;
if (!vect_is_simple_use (mask, stmt_info->vinfo, &def_stmt, &mask_dt,
&mask_vectype))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"mask use not simple.\n");
return false;
}
tree vectype = STMT_VINFO_VECTYPE (stmt_info);
if (!mask_vectype)
mask_vectype = get_mask_type_for_scalar_type (TREE_TYPE (vectype));
if (!mask_vectype || !VECTOR_BOOLEAN_TYPE_P (mask_vectype))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"could not find an appropriate vector mask type.\n");
return false;
}
if (maybe_ne (TYPE_VECTOR_SUBPARTS (mask_vectype),
TYPE_VECTOR_SUBPARTS (vectype)))
{
if (dump_enabled_p ())
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"vector mask type ");
dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, mask_vectype);
dump_printf (MSG_MISSED_OPTIMIZATION,
" does not match vector data type ");
dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, vectype);
dump_printf (MSG_MISSED_OPTIMIZATION, ".\n");
}
return false;
}
*mask_dt_out = mask_dt;
*mask_vectype_out = mask_vectype;
return true;
}
/* Return true if stored value RHS is suitable for vectorizing store
statement STMT. When returning true, store the type of the
definition in *RHS_DT_OUT, the type of the vectorized store value in
*RHS_VECTYPE_OUT and the type of the store in *VLS_TYPE_OUT. */
static bool
vect_check_store_rhs (gimple *stmt, tree rhs, vect_def_type *rhs_dt_out,
tree *rhs_vectype_out, vec_load_store_type *vls_type_out)
{
/* In the case this is a store from a constant make sure
native_encode_expr can handle it. */
if (CONSTANT_CLASS_P (rhs) && native_encode_expr (rhs, NULL, 64) == 0)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"cannot encode constant as a byte sequence.\n");
return false;
}
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
gimple *def_stmt;
enum vect_def_type rhs_dt;
tree rhs_vectype;
if (!vect_is_simple_use (rhs, stmt_info->vinfo, &def_stmt, &rhs_dt,
&rhs_vectype))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"use not simple.\n");
return false;
}
tree vectype = STMT_VINFO_VECTYPE (stmt_info);
if (rhs_vectype && !useless_type_conversion_p (vectype, rhs_vectype))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"incompatible vector types.\n");
return false;
}
*rhs_dt_out = rhs_dt;
*rhs_vectype_out = rhs_vectype;
if (rhs_dt == vect_constant_def || rhs_dt == vect_external_def)
*vls_type_out = VLS_STORE_INVARIANT;
else
*vls_type_out = VLS_STORE;
return true;
}
/* Build an all-ones vector mask of type MASKTYPE while vectorizing STMT.
Note that we support masks with floating-point type, in which case the
floats are interpreted as a bitmask. */
static tree
vect_build_all_ones_mask (gimple *stmt, tree masktype)
{
if (TREE_CODE (masktype) == INTEGER_TYPE)
return build_int_cst (masktype, -1);
else if (TREE_CODE (TREE_TYPE (masktype)) == INTEGER_TYPE)
{
tree mask = build_int_cst (TREE_TYPE (masktype), -1);
mask = build_vector_from_val (masktype, mask);
return vect_init_vector (stmt, mask, masktype, NULL);
}
else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (masktype)))
{
REAL_VALUE_TYPE r;
long tmp[6];
for (int j = 0; j < 6; ++j)
tmp[j] = -1;
real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (masktype)));
tree mask = build_real (TREE_TYPE (masktype), r);
mask = build_vector_from_val (masktype, mask);
return vect_init_vector (stmt, mask, masktype, NULL);
}
gcc_unreachable ();
}
/* Build an all-zero merge value of type VECTYPE while vectorizing
STMT as a gather load. */
static tree
vect_build_zero_merge_argument (gimple *stmt, tree vectype)
{
tree merge;
if (TREE_CODE (TREE_TYPE (vectype)) == INTEGER_TYPE)
merge = build_int_cst (TREE_TYPE (vectype), 0);
else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (vectype)))
{
REAL_VALUE_TYPE r;
long tmp[6];
for (int j = 0; j < 6; ++j)
tmp[j] = 0;
real_from_target (&r, tmp, TYPE_MODE (TREE_TYPE (vectype)));
merge = build_real (TREE_TYPE (vectype), r);
}
else
gcc_unreachable ();
merge = build_vector_from_val (vectype, merge);
return vect_init_vector (stmt, merge, vectype, NULL);
}
/* Build a gather load call while vectorizing STMT. Insert new instructions
before GSI and add them to VEC_STMT. GS_INFO describes the gather load
operation. If the load is conditional, MASK is the unvectorized
condition and MASK_DT is its definition type, otherwise MASK is null. */
static void
vect_build_gather_load_calls (gimple *stmt, gimple_stmt_iterator *gsi,
gimple **vec_stmt, gather_scatter_info *gs_info,
tree mask, vect_def_type mask_dt)
{
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
tree vectype = STMT_VINFO_VECTYPE (stmt_info);
poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
int ncopies = vect_get_num_copies (loop_vinfo, vectype);
edge pe = loop_preheader_edge (loop);
enum { NARROW, NONE, WIDEN } modifier;
poly_uint64 gather_off_nunits
= TYPE_VECTOR_SUBPARTS (gs_info->offset_vectype);
tree arglist = TYPE_ARG_TYPES (TREE_TYPE (gs_info->decl));
tree rettype = TREE_TYPE (TREE_TYPE (gs_info->decl));
tree srctype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist);
tree ptrtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist);
tree idxtype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist);
tree masktype = TREE_VALUE (arglist); arglist = TREE_CHAIN (arglist);
tree scaletype = TREE_VALUE (arglist);
gcc_checking_assert (types_compatible_p (srctype, rettype)
&& (!mask || types_compatible_p (srctype, masktype)));
tree perm_mask = NULL_TREE;
tree mask_perm_mask = NULL_TREE;
if (known_eq (nunits, gather_off_nunits))
modifier = NONE;
else if (known_eq (nunits * 2, gather_off_nunits))
{
modifier = WIDEN;
/* Currently widening gathers and scatters are only supported for
fixed-length vectors. */
int count = gather_off_nunits.to_constant ();
vec_perm_builder sel (count, count, 1);
for (int i = 0; i < count; ++i)
sel.quick_push (i | (count / 2));
vec_perm_indices indices (sel, 1, count);
perm_mask = vect_gen_perm_mask_checked (gs_info->offset_vectype,
indices);
}
else if (known_eq (nunits, gather_off_nunits * 2))
{
modifier = NARROW;
/* Currently narrowing gathers and scatters are only supported for
fixed-length vectors. */
int count = nunits.to_constant ();
vec_perm_builder sel (count, count, 1);
sel.quick_grow (count);
for (int i = 0; i < count; ++i)
sel[i] = i < count / 2 ? i : i + count / 2;
vec_perm_indices indices (sel, 2, count);
perm_mask = vect_gen_perm_mask_checked (vectype, indices);
ncopies *= 2;
if (mask)
{
for (int i = 0; i < count; ++i)
sel[i] = i | (count / 2);
indices.new_vector (sel, 2, count);
mask_perm_mask = vect_gen_perm_mask_checked (masktype, indices);
}
}
else
gcc_unreachable ();
tree vec_dest = vect_create_destination_var (gimple_get_lhs (stmt),
vectype);
tree ptr = fold_convert (ptrtype, gs_info->base);
if (!is_gimple_min_invariant (ptr))
{
gimple_seq seq;
ptr = force_gimple_operand (ptr, &seq, true, NULL_TREE);
basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq);
gcc_assert (!new_bb);
}
tree scale = build_int_cst (scaletype, gs_info->scale);
tree vec_oprnd0 = NULL_TREE;
tree vec_mask = NULL_TREE;
tree src_op = NULL_TREE;
tree mask_op = NULL_TREE;
tree prev_res = NULL_TREE;
stmt_vec_info prev_stmt_info = NULL;
if (!mask)
{
src_op = vect_build_zero_merge_argument (stmt, rettype);
mask_op = vect_build_all_ones_mask (stmt, masktype);
}
for (int j = 0; j < ncopies; ++j)
{
tree op, var;
gimple *new_stmt;
if (modifier == WIDEN && (j & 1))
op = permute_vec_elements (vec_oprnd0, vec_oprnd0,
perm_mask, stmt, gsi);
else if (j == 0)
op = vec_oprnd0
= vect_get_vec_def_for_operand (gs_info->offset, stmt);
else
op = vec_oprnd0
= vect_get_vec_def_for_stmt_copy (gs_info->offset_dt, vec_oprnd0);
if (!useless_type_conversion_p (idxtype, TREE_TYPE (op)))
{
gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (op)),
TYPE_VECTOR_SUBPARTS (idxtype)));
var = vect_get_new_ssa_name (idxtype, vect_simple_var);
op = build1 (VIEW_CONVERT_EXPR, idxtype, op);
new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op);
vect_finish_stmt_generation (stmt, new_stmt, gsi);
op = var;
}
if (mask)
{
if (mask_perm_mask && (j & 1))
mask_op = permute_vec_elements (mask_op, mask_op,
mask_perm_mask, stmt, gsi);
else
{
if (j == 0)
vec_mask = vect_get_vec_def_for_operand (mask, stmt);
else
vec_mask = vect_get_vec_def_for_stmt_copy (mask_dt, vec_mask);
mask_op = vec_mask;
if (!useless_type_conversion_p (masktype, TREE_TYPE (vec_mask)))
{
gcc_assert
(known_eq (TYPE_VECTOR_SUBPARTS (TREE_TYPE (mask_op)),
TYPE_VECTOR_SUBPARTS (masktype)));
var = vect_get_new_ssa_name (masktype, vect_simple_var);
mask_op = build1 (VIEW_CONVERT_EXPR, masktype, mask_op);
new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR,
mask_op);
vect_finish_stmt_generation (stmt, new_stmt, gsi);
mask_op = var;
}
}
src_op = mask_op;
}
new_stmt = gimple_build_call (gs_info->decl, 5, src_op, ptr, op,
mask_op, scale);
if (!useless_type_conversion_p (vectype, rettype))
{
gcc_assert (known_eq (TYPE_VECTOR_SUBPARTS (vectype),
TYPE_VECTOR_SUBPARTS (rettype)));
op = vect_get_new_ssa_name (rettype, vect_simple_var);
gimple_call_set_lhs (new_stmt, op);
vect_finish_stmt_generation (stmt, new_stmt, gsi);
var = make_ssa_name (vec_dest);
op = build1 (VIEW_CONVERT_EXPR, vectype, op);
new_stmt = gimple_build_assign (var, VIEW_CONVERT_EXPR, op);
}
else
{
var = make_ssa_name (vec_dest, new_stmt);
gimple_call_set_lhs (new_stmt, var);
}
vect_finish_stmt_generation (stmt, new_stmt, gsi);
if (modifier == NARROW)
{
if ((j & 1) == 0)
{
prev_res = var;
continue;
}
var = permute_vec_elements (prev_res, var, perm_mask, stmt, gsi);
new_stmt = SSA_NAME_DEF_STMT (var);
}
if (prev_stmt_info == NULL)
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);
}
}
/* Prepare the base and offset in GS_INFO for vectorization.
Set *DATAREF_PTR to the loop-invariant base address and *VEC_OFFSET
to the vectorized offset argument for the first copy of STMT. STMT
is the statement described by GS_INFO and LOOP is the containing loop. */
static void
vect_get_gather_scatter_ops (struct loop *loop, gimple *stmt,
gather_scatter_info *gs_info,
tree *dataref_ptr, tree *vec_offset)
{
gimple_seq stmts = NULL;
*dataref_ptr = force_gimple_operand (gs_info->base, &stmts, true, NULL_TREE);
if (stmts != NULL)
{
basic_block new_bb;
edge pe = loop_preheader_edge (loop);
new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts);
gcc_assert (!new_bb);
}
tree offset_type = TREE_TYPE (gs_info->offset);
tree offset_vectype = get_vectype_for_scalar_type (offset_type);
*vec_offset = vect_get_vec_def_for_operand (gs_info->offset, stmt,
offset_vectype);
}
/* Prepare to implement a grouped or strided load or store using
the gather load or scatter store operation described by GS_INFO.
STMT is the load or store statement.
Set *DATAREF_BUMP to the amount that should be added to the base
address after each copy of the vectorized statement. Set *VEC_OFFSET
to an invariant offset vector in which element I has the value
I * DR_STEP / SCALE. */
static void
vect_get_strided_load_store_ops (gimple *stmt, loop_vec_info loop_vinfo,
gather_scatter_info *gs_info,
tree *dataref_bump, tree *vec_offset)
{
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
tree vectype = STMT_VINFO_VECTYPE (stmt_info);
gimple_seq stmts;
tree bump = size_binop (MULT_EXPR,
fold_convert (sizetype, DR_STEP (dr)),
size_int (TYPE_VECTOR_SUBPARTS (vectype)));
*dataref_bump = force_gimple_operand (bump, &stmts, true, NULL_TREE);
if (stmts)
gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
/* The offset given in GS_INFO can have pointer type, so use the element
type of the vector instead. */
tree offset_type = TREE_TYPE (gs_info->offset);
tree offset_vectype = get_vectype_for_scalar_type (offset_type);
offset_type = TREE_TYPE (offset_vectype);
/* Calculate X = DR_STEP / SCALE and convert it to the appropriate type. */
tree step = size_binop (EXACT_DIV_EXPR, DR_STEP (dr),
ssize_int (gs_info->scale));
step = fold_convert (offset_type, step);
step = force_gimple_operand (step, &stmts, true, NULL_TREE);
/* Create {0, X, X*2, X*3, ...}. */
*vec_offset = gimple_build (&stmts, VEC_SERIES_EXPR, offset_vectype,
build_zero_cst (offset_type), step);
if (stmts)
gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
}
/* Return the amount that should be added to a vector pointer to move
to the next or previous copy of AGGR_TYPE. DR is the data reference
being vectorized and MEMORY_ACCESS_TYPE describes the type of
vectorization. */
static tree
vect_get_data_ptr_increment (data_reference *dr, tree aggr_type,
vect_memory_access_type memory_access_type)
{
if (memory_access_type == VMAT_INVARIANT)
return size_zero_node;
tree iv_step = TYPE_SIZE_UNIT (aggr_type);
tree step = vect_dr_behavior (dr)->step;
if (tree_int_cst_sgn (step) == -1)
iv_step = fold_build1 (NEGATE_EXPR, TREE_TYPE (iv_step), iv_step);
return iv_step;
}
/* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64}. */
static bool
vectorizable_bswap (gimple *stmt, gimple_stmt_iterator *gsi,
gimple **vec_stmt, slp_tree slp_node,
tree vectype_in, enum vect_def_type *dt)
{
tree op, vectype;
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
unsigned ncopies;
unsigned HOST_WIDE_INT nunits, num_bytes;
op = gimple_call_arg (stmt, 0);
vectype = STMT_VINFO_VECTYPE (stmt_info);
if (!TYPE_VECTOR_SUBPARTS (vectype).is_constant (&nunits))
return false;
/* Multiple types in SLP are handled by creating the appropriate number of
vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
case of SLP. */
if (slp_node)
ncopies = 1;
else
ncopies = vect_get_num_copies (loop_vinfo, vectype);
gcc_assert (ncopies >= 1);
tree char_vectype = get_same_sized_vectype (char_type_node, vectype_in);
if (! char_vectype)
return false;
if (!TYPE_VECTOR_SUBPARTS (char_vectype).is_constant (&num_bytes))
return false;
unsigned word_bytes = num_bytes / nunits;
/* The encoding uses one stepped pattern for each byte in the word. */
vec_perm_builder elts (num_bytes, word_bytes, 3);
for (unsigned i = 0; i < 3; ++i)
for (unsigned j = 0; j < word_bytes; ++j)
elts.quick_push ((i + 1) * word_bytes - j - 1);
vec_perm_indices indices (elts, 1, num_bytes);
if (!can_vec_perm_const_p (TYPE_MODE (char_vectype), indices))
return false;
if (! vec_stmt)
{
STMT_VINFO_TYPE (stmt_info) = call_vec_info_type;
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location, "=== vectorizable_bswap ==="
"\n");
if (! slp_node)
{
add_stmt_cost (stmt_info->vinfo->target_cost_data,
1, vector_stmt, stmt_info, 0, vect_prologue);
add_stmt_cost (stmt_info->vinfo->target_cost_data,
ncopies, vec_perm, stmt_info, 0, vect_body);
}
return true;
}
tree bswap_vconst = vec_perm_indices_to_tree (char_vectype, indices);
/* Transform. */
vec<tree> vec_oprnds = vNULL;
gimple *new_stmt = NULL;
stmt_vec_info prev_stmt_info = NULL;
for (unsigned j = 0; j < ncopies; j++)
{
/* Handle uses. */
if (j == 0)
vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node);
else
vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds, NULL);
/* Arguments are ready. create the new vector stmt. */
unsigned i;
tree vop;
FOR_EACH_VEC_ELT (vec_oprnds, i, vop)
{
tree tem = make_ssa_name (char_vectype);
new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR,
char_vectype, vop));
vect_finish_stmt_generation (stmt, new_stmt, gsi);
tree tem2 = make_ssa_name (char_vectype);
new_stmt = gimple_build_assign (tem2, VEC_PERM_EXPR,
tem, tem, bswap_vconst);
vect_finish_stmt_generation (stmt, new_stmt, gsi);
tem = make_ssa_name (vectype);
new_stmt = gimple_build_assign (tem, build1 (VIEW_CONVERT_EXPR,
vectype, tem2));
vect_finish_stmt_generation (stmt, new_stmt, gsi);
if (slp_node)
SLP_TREE_VEC_STMTS (slp_node).quick_push (new_stmt);
}
if (slp_node)
continue;
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);
}
vec_oprnds.release ();
return true;
}
/* Return true if vector types VECTYPE_IN and VECTYPE_OUT have
integer elements and if we can narrow VECTYPE_IN to VECTYPE_OUT
in a single step. On success, store the binary pack code in
*CONVERT_CODE. */
static bool
simple_integer_narrowing (tree vectype_out, tree vectype_in,
tree_code *convert_code)
{
if (!INTEGRAL_TYPE_P (TREE_TYPE (vectype_out))
|| !INTEGRAL_TYPE_P (TREE_TYPE (vectype_in)))
return false;
tree_code code;
int multi_step_cvt = 0;
auto_vec <tree, 8> interm_types;
if (!supportable_narrowing_operation (NOP_EXPR, vectype_out, vectype_in,
&code, &multi_step_cvt,
&interm_types)
|| multi_step_cvt)
return false;
*convert_code = code;
return true;
}
/* Function vectorizable_call.
Check if GS 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. */
static bool
vectorizable_call (gimple *gs, gimple_stmt_iterator *gsi, gimple **vec_stmt,
slp_tree slp_node)
{
gcall *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 (gs), prev_stmt_info;
tree vectype_out, vectype_in;
poly_uint64 nunits_in;
poly_uint64 nunits_out;
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
vec_info *vinfo = stmt_info->vinfo;
tree fndecl, new_temp, rhs_type;
gimple *def_stmt;
enum vect_def_type dt[3]
= {vect_unknown_def_type, vect_unknown_def_type, vect_unknown_def_type};
int ndts = 3;
gimple *new_stmt = NULL;
int ncopies, j;
vec<tree> vargs = vNULL;
enum { NARROW, NONE, WIDEN } modifier;
size_t i, nargs;
tree lhs;
if (!STMT_VINFO_RELEVANT_P (stmt_info) && !bb_vinfo)
return false;
if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_internal_def
&& ! vec_stmt)
return false;
/* Is GS a vectorizable call? */
stmt = dyn_cast <gcall *> (gs);
if (!stmt)
return false;
if (gimple_call_internal_p (stmt)
&& (internal_load_fn_p (gimple_call_internal_fn (stmt))
|| internal_store_fn_p (gimple_call_internal_fn (stmt))))
/* Handled by vectorizable_load and vectorizable_store. */
return false;
if (gimple_call_lhs (stmt) == NULL_TREE
|| TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME)
return false;
gcc_checking_assert (!stmt_can_throw_internal (stmt));
vectype_out = STMT_VINFO_VECTYPE (stmt_info);
/* Process function arguments. */
rhs_type = NULL_TREE;
vectype_in = NULL_TREE;
nargs = gimple_call_num_args (stmt);
/* Bail out if the function has more than three arguments, we do not have
interesting builtin functions to vectorize with more than two arguments
except for fma. No arguments is also not good. */
if (nargs == 0 || nargs > 3)
return false;
/* Ignore the argument of IFN_GOMP_SIMD_LANE, it is magic. */
if (gimple_call_internal_p (stmt)
&& gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE)
{
nargs = 0;
rhs_type = unsigned_type_node;
}
for (i = 0; i < nargs; i++)
{
tree opvectype;
op = gimple_call_arg (stmt, i);
/* We can only handle calls with arguments of the same type. */
if (rhs_type
&& !types_compatible_p (rhs_type, TREE_TYPE (op)))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"argument types differ.\n");
return false;
}
if (!rhs_type)
rhs_type = TREE_TYPE (op);
if (!vect_is_simple_use (op, vinfo, &def_stmt, &dt[i], &opvectype))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"use not simple.\n");
return false;
}
if (!vectype_in)
vectype_in = opvectype;
else if (opvectype
&& opvectype != vectype_in)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"argument vector types differ.\n");
return false;
}
}
/* If all arguments are external or constant defs use a vector type with
the same size as the output vector type. */
if (!vectype_in)
vectype_in = get_same_sized_vectype (rhs_type, vectype_out);
if (vec_stmt)
gcc_assert (vectype_in);
if (!vectype_in)
{
if (dump_enabled_p ())
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"no vectype for scalar type ");
dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, rhs_type);
dump_printf (MSG_MISSED_OPTIMIZATION, "\n");
}
return false;
}
/* FORNOW */
nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
if (known_eq (nunits_in * 2, nunits_out))
modifier = NARROW;
else if (known_eq (nunits_out, nunits_in))
modifier = NONE;
else if (known_eq (nunits_out * 2, nunits_in))
modifier = WIDEN;
else
return false;
/* We only handle functions that do not read or clobber memory. */
if (gimple_vuse (stmt))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"function reads from or writes to memory.\n");
return false;
}
/* For now, we only vectorize functions if a target specific builtin
is available. TODO -- in some cases, it might be profitable to
insert the calls for pieces of the vector, in order to be able
to vectorize other operations in the loop. */
fndecl = NULL_TREE;
internal_fn ifn = IFN_LAST;
combined_fn cfn = gimple_call_combined_fn (stmt);
tree callee = gimple_call_fndecl (stmt);
/* First try using an internal function. */
tree_code convert_code = ERROR_MARK;
if (cfn != CFN_LAST
&& (modifier == NONE
|| (modifier == NARROW
&& simple_integer_narrowing (vectype_out, vectype_in,
&convert_code))))
ifn = vectorizable_internal_function (cfn, callee, vectype_out,
vectype_in);
/* If that fails, try asking for a target-specific built-in function. */
if (ifn == IFN_LAST)
{
if (cfn != CFN_LAST)
fndecl = targetm.vectorize.builtin_vectorized_function
(cfn, vectype_out, vectype_in);
else if (callee)
fndecl = targetm.vectorize.builtin_md_vectorized_function
(callee, vectype_out, vectype_in);
}
if (ifn == IFN_LAST && !fndecl)
{
if (cfn == CFN_GOMP_SIMD_LANE
&& !slp_node
&& loop_vinfo
&& LOOP_VINFO_LOOP (loop_vinfo)->simduid
&& TREE_CODE (gimple_call_arg (stmt, 0)) == SSA_NAME
&& LOOP_VINFO_LOOP (loop_vinfo)->simduid
== SSA_NAME_VAR (gimple_call_arg (stmt, 0)))
{
/* We can handle IFN_GOMP_SIMD_LANE by returning a
{ 0, 1, 2, ... vf - 1 } vector. */
gcc_assert (nargs == 0);
}
else if (modifier == NONE
&& (gimple_call_builtin_p (stmt, BUILT_IN_BSWAP16)
|| gimple_call_builtin_p (stmt, BUILT_IN_BSWAP32)
|| gimple_call_builtin_p (stmt, BUILT_IN_BSWAP64)))
return vectorizable_bswap (stmt, gsi, vec_stmt, slp_node,