blob: 6c892791bd4c39f672add4e4c22a9d7835e292d6 [file] [log] [blame]
/* Data References Analysis and Manipulation Utilities for Vectorization.
Copyright (C) 2003-2022 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 "predict.h"
#include "memmodel.h"
#include "tm_p.h"
#include "ssa.h"
#include "optabs-tree.h"
#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-ssa-loop-ivopts.h"
#include "tree-ssa-loop-manip.h"
#include "tree-ssa-loop.h"
#include "cfgloop.h"
#include "tree-scalar-evolution.h"
#include "tree-vectorizer.h"
#include "expr.h"
#include "builtins.h"
#include "tree-cfg.h"
#include "tree-hash-traits.h"
#include "vec-perm-indices.h"
#include "internal-fn.h"
#include "gimple-fold.h"
/* Return true if load- or store-lanes optab OPTAB is implemented for
COUNT vectors of type VECTYPE. NAME is the name of OPTAB. */
static bool
vect_lanes_optab_supported_p (const char *name, convert_optab optab,
tree vectype, unsigned HOST_WIDE_INT count)
{
machine_mode mode, array_mode;
bool limit_p;
mode = TYPE_MODE (vectype);
if (!targetm.array_mode (mode, count).exists (&array_mode))
{
poly_uint64 bits = count * GET_MODE_BITSIZE (mode);
limit_p = !targetm.array_mode_supported_p (mode, count);
if (!int_mode_for_size (bits, limit_p).exists (&array_mode))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"no array mode for %s[%wu]\n",
GET_MODE_NAME (mode), count);
return false;
}
}
if (convert_optab_handler (optab, array_mode, mode) == CODE_FOR_nothing)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"cannot use %s<%s><%s>\n", name,
GET_MODE_NAME (array_mode), GET_MODE_NAME (mode));
return false;
}
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"can use %s<%s><%s>\n", name, GET_MODE_NAME (array_mode),
GET_MODE_NAME (mode));
return true;
}
/* Return the smallest scalar part of STMT_INFO.
This is used to determine the vectype of the stmt. We generally set the
vectype according to the type of the result (lhs). For stmts whose
result-type is different than the type of the arguments (e.g., demotion,
promotion), vectype will be reset appropriately (later). Note that we have
to visit the smallest datatype in this function, because that determines the
VF. If the smallest datatype in the loop is present only as the rhs of a
promotion operation - we'd miss it.
Such a case, where a variable of this datatype does not appear in the lhs
anywhere in the loop, can only occur if it's an invariant: e.g.:
'int_x = (int) short_inv', which we'd expect to have been optimized away by
invariant motion. However, we cannot rely on invariant motion to always
take invariants out of the loop, and so in the case of promotion we also
have to check the rhs.
LHS_SIZE_UNIT and RHS_SIZE_UNIT contain the sizes of the corresponding
types. */
tree
vect_get_smallest_scalar_type (stmt_vec_info stmt_info, tree scalar_type)
{
HOST_WIDE_INT lhs, rhs;
/* During the analysis phase, this function is called on arbitrary
statements that might not have scalar results. */
if (!tree_fits_uhwi_p (TYPE_SIZE_UNIT (scalar_type)))
return scalar_type;
lhs = rhs = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (scalar_type));
gassign *assign = dyn_cast <gassign *> (stmt_info->stmt);
if (assign)
{
scalar_type = TREE_TYPE (gimple_assign_lhs (assign));
if (gimple_assign_cast_p (assign)
|| gimple_assign_rhs_code (assign) == DOT_PROD_EXPR
|| gimple_assign_rhs_code (assign) == WIDEN_SUM_EXPR
|| gimple_assign_rhs_code (assign) == WIDEN_MULT_EXPR
|| gimple_assign_rhs_code (assign) == WIDEN_LSHIFT_EXPR
|| gimple_assign_rhs_code (assign) == WIDEN_PLUS_EXPR
|| gimple_assign_rhs_code (assign) == WIDEN_MINUS_EXPR
|| gimple_assign_rhs_code (assign) == FLOAT_EXPR)
{
tree rhs_type = TREE_TYPE (gimple_assign_rhs1 (assign));
rhs = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (rhs_type));
if (rhs < lhs)
scalar_type = rhs_type;
}
}
else if (gcall *call = dyn_cast <gcall *> (stmt_info->stmt))
{
unsigned int i = 0;
if (gimple_call_internal_p (call))
{
internal_fn ifn = gimple_call_internal_fn (call);
if (internal_load_fn_p (ifn))
/* For loads the LHS type does the trick. */
i = ~0U;
else if (internal_store_fn_p (ifn))
{
/* For stores use the tyep of the stored value. */
i = internal_fn_stored_value_index (ifn);
scalar_type = TREE_TYPE (gimple_call_arg (call, i));
i = ~0U;
}
else if (internal_fn_mask_index (ifn) == 0)
i = 1;
}
if (i < gimple_call_num_args (call))
{
tree rhs_type = TREE_TYPE (gimple_call_arg (call, i));
if (tree_fits_uhwi_p (TYPE_SIZE_UNIT (rhs_type)))
{
rhs = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (rhs_type));
if (rhs < lhs)
scalar_type = rhs_type;
}
}
}
return scalar_type;
}
/* Insert DDR into LOOP_VINFO list of ddrs that may alias and need to be
tested at run-time. Return TRUE if DDR was successfully inserted.
Return false if versioning is not supported. */
static opt_result
vect_mark_for_runtime_alias_test (ddr_p ddr, loop_vec_info loop_vinfo)
{
class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
if ((unsigned) param_vect_max_version_for_alias_checks == 0)
return opt_result::failure_at (vect_location,
"will not create alias checks, as"
" --param vect-max-version-for-alias-checks"
" == 0\n");
opt_result res
= runtime_alias_check_p (ddr, loop,
optimize_loop_nest_for_speed_p (loop));
if (!res)
return res;
LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo).safe_push (ddr);
return opt_result::success ();
}
/* Record that loop LOOP_VINFO needs to check that VALUE is nonzero. */
static void
vect_check_nonzero_value (loop_vec_info loop_vinfo, tree value)
{
const vec<tree> &checks = LOOP_VINFO_CHECK_NONZERO (loop_vinfo);
for (unsigned int i = 0; i < checks.length(); ++i)
if (checks[i] == value)
return;
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"need run-time check that %T is nonzero\n",
value);
LOOP_VINFO_CHECK_NONZERO (loop_vinfo).safe_push (value);
}
/* Return true if we know that the order of vectorized DR_INFO_A and
vectorized DR_INFO_B will be the same as the order of DR_INFO_A and
DR_INFO_B. At least one of the accesses is a write. */
static bool
vect_preserves_scalar_order_p (dr_vec_info *dr_info_a, dr_vec_info *dr_info_b)
{
stmt_vec_info stmtinfo_a = dr_info_a->stmt;
stmt_vec_info stmtinfo_b = dr_info_b->stmt;
/* Single statements are always kept in their original order. */
if (!STMT_VINFO_GROUPED_ACCESS (stmtinfo_a)
&& !STMT_VINFO_GROUPED_ACCESS (stmtinfo_b))
return true;
/* STMT_A and STMT_B belong to overlapping groups. All loads are
emitted at the position of the first scalar load.
Stores in a group are emitted at the position of the last scalar store.
Compute that position and check whether the resulting order matches
the current one. */
stmt_vec_info il_a = DR_GROUP_FIRST_ELEMENT (stmtinfo_a);
if (il_a)
{
if (DR_IS_WRITE (STMT_VINFO_DATA_REF (stmtinfo_a)))
for (stmt_vec_info s = DR_GROUP_NEXT_ELEMENT (il_a); s;
s = DR_GROUP_NEXT_ELEMENT (s))
il_a = get_later_stmt (il_a, s);
else /* DR_IS_READ */
for (stmt_vec_info s = DR_GROUP_NEXT_ELEMENT (il_a); s;
s = DR_GROUP_NEXT_ELEMENT (s))
if (get_later_stmt (il_a, s) == il_a)
il_a = s;
}
else
il_a = stmtinfo_a;
stmt_vec_info il_b = DR_GROUP_FIRST_ELEMENT (stmtinfo_b);
if (il_b)
{
if (DR_IS_WRITE (STMT_VINFO_DATA_REF (stmtinfo_b)))
for (stmt_vec_info s = DR_GROUP_NEXT_ELEMENT (il_b); s;
s = DR_GROUP_NEXT_ELEMENT (s))
il_b = get_later_stmt (il_b, s);
else /* DR_IS_READ */
for (stmt_vec_info s = DR_GROUP_NEXT_ELEMENT (il_b); s;
s = DR_GROUP_NEXT_ELEMENT (s))
if (get_later_stmt (il_b, s) == il_b)
il_b = s;
}
else
il_b = stmtinfo_b;
bool a_after_b = (get_later_stmt (stmtinfo_a, stmtinfo_b) == stmtinfo_a);
return (get_later_stmt (il_a, il_b) == il_a) == a_after_b;
}
/* A subroutine of vect_analyze_data_ref_dependence. Handle
DDR_COULD_BE_INDEPENDENT_P ddr DDR that has a known set of dependence
distances. These distances are conservatively correct but they don't
reflect a guaranteed dependence.
Return true if this function does all the work necessary to avoid
an alias or false if the caller should use the dependence distances
to limit the vectorization factor in the usual way. LOOP_DEPTH is
the depth of the loop described by LOOP_VINFO and the other arguments
are as for vect_analyze_data_ref_dependence. */
static bool
vect_analyze_possibly_independent_ddr (data_dependence_relation *ddr,
loop_vec_info loop_vinfo,
int loop_depth, unsigned int *max_vf)
{
class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
for (lambda_vector &dist_v : DDR_DIST_VECTS (ddr))
{
int dist = dist_v[loop_depth];
if (dist != 0 && !(dist > 0 && DDR_REVERSED_P (ddr)))
{
/* If the user asserted safelen >= DIST consecutive iterations
can be executed concurrently, assume independence.
??? An alternative would be to add the alias check even
in this case, and vectorize the fallback loop with the
maximum VF set to safelen. However, if the user has
explicitly given a length, it's less likely that that
would be a win. */
if (loop->safelen >= 2 && abs_hwi (dist) <= loop->safelen)
{
if ((unsigned int) loop->safelen < *max_vf)
*max_vf = loop->safelen;
LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) = false;
continue;
}
/* For dependence distances of 2 or more, we have the option
of limiting VF or checking for an alias at runtime.
Prefer to check at runtime if we can, to avoid limiting
the VF unnecessarily when the bases are in fact independent.
Note that the alias checks will be removed if the VF ends up
being small enough. */
dr_vec_info *dr_info_a = loop_vinfo->lookup_dr (DDR_A (ddr));
dr_vec_info *dr_info_b = loop_vinfo->lookup_dr (DDR_B (ddr));
return (!STMT_VINFO_GATHER_SCATTER_P (dr_info_a->stmt)
&& !STMT_VINFO_GATHER_SCATTER_P (dr_info_b->stmt)
&& vect_mark_for_runtime_alias_test (ddr, loop_vinfo));
}
}
return true;
}
/* Function vect_analyze_data_ref_dependence.
FIXME: I needed to change the sense of the returned flag.
Return FALSE if there (might) exist a dependence between a memory-reference
DRA and a memory-reference DRB. When versioning for alias may check a
dependence at run-time, return TRUE. Adjust *MAX_VF according to
the data dependence. */
static opt_result
vect_analyze_data_ref_dependence (struct data_dependence_relation *ddr,
loop_vec_info loop_vinfo,
unsigned int *max_vf)
{
unsigned int i;
class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
struct data_reference *dra = DDR_A (ddr);
struct data_reference *drb = DDR_B (ddr);
dr_vec_info *dr_info_a = loop_vinfo->lookup_dr (dra);
dr_vec_info *dr_info_b = loop_vinfo->lookup_dr (drb);
stmt_vec_info stmtinfo_a = dr_info_a->stmt;
stmt_vec_info stmtinfo_b = dr_info_b->stmt;
lambda_vector dist_v;
unsigned int loop_depth;
/* If user asserted safelen consecutive iterations can be
executed concurrently, assume independence. */
auto apply_safelen = [&]()
{
if (loop->safelen >= 2)
{
if ((unsigned int) loop->safelen < *max_vf)
*max_vf = loop->safelen;
LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) = false;
return true;
}
return false;
};
/* In loop analysis all data references should be vectorizable. */
if (!STMT_VINFO_VECTORIZABLE (stmtinfo_a)
|| !STMT_VINFO_VECTORIZABLE (stmtinfo_b))
gcc_unreachable ();
/* Independent data accesses. */
if (DDR_ARE_DEPENDENT (ddr) == chrec_known)
return opt_result::success ();
if (dra == drb
|| (DR_IS_READ (dra) && DR_IS_READ (drb)))
return opt_result::success ();
/* We do not have to consider dependences between accesses that belong
to the same group, unless the stride could be smaller than the
group size. */
if (DR_GROUP_FIRST_ELEMENT (stmtinfo_a)
&& (DR_GROUP_FIRST_ELEMENT (stmtinfo_a)
== DR_GROUP_FIRST_ELEMENT (stmtinfo_b))
&& !STMT_VINFO_STRIDED_P (stmtinfo_a))
return opt_result::success ();
/* Even if we have an anti-dependence then, as the vectorized loop covers at
least two scalar iterations, there is always also a true dependence.
As the vectorizer does not re-order loads and stores we can ignore
the anti-dependence if TBAA can disambiguate both DRs similar to the
case with known negative distance anti-dependences (positive
distance anti-dependences would violate TBAA constraints). */
if (((DR_IS_READ (dra) && DR_IS_WRITE (drb))
|| (DR_IS_WRITE (dra) && DR_IS_READ (drb)))
&& !alias_sets_conflict_p (get_alias_set (DR_REF (dra)),
get_alias_set (DR_REF (drb))))
return opt_result::success ();
if (STMT_VINFO_GATHER_SCATTER_P (stmtinfo_a)
|| STMT_VINFO_GATHER_SCATTER_P (stmtinfo_b))
{
if (apply_safelen ())
return opt_result::success ();
return opt_result::failure_at
(stmtinfo_a->stmt,
"possible alias involving gather/scatter between %T and %T\n",
DR_REF (dra), DR_REF (drb));
}
/* Unknown data dependence. */
if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know)
{
if (apply_safelen ())
return opt_result::success ();
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, stmtinfo_a->stmt,
"versioning for alias required: "
"can't determine dependence between %T and %T\n",
DR_REF (dra), DR_REF (drb));
/* Add to list of ddrs that need to be tested at run-time. */
return vect_mark_for_runtime_alias_test (ddr, loop_vinfo);
}
/* Known data dependence. */
if (DDR_NUM_DIST_VECTS (ddr) == 0)
{
if (apply_safelen ())
return opt_result::success ();
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, stmtinfo_a->stmt,
"versioning for alias required: "
"bad dist vector for %T and %T\n",
DR_REF (dra), DR_REF (drb));
/* Add to list of ddrs that need to be tested at run-time. */
return vect_mark_for_runtime_alias_test (ddr, loop_vinfo);
}
loop_depth = index_in_loop_nest (loop->num, DDR_LOOP_NEST (ddr));
if (DDR_COULD_BE_INDEPENDENT_P (ddr)
&& vect_analyze_possibly_independent_ddr (ddr, loop_vinfo,
loop_depth, max_vf))
return opt_result::success ();
FOR_EACH_VEC_ELT (DDR_DIST_VECTS (ddr), i, dist_v)
{
int dist = dist_v[loop_depth];
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"dependence distance = %d.\n", dist);
if (dist == 0)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"dependence distance == 0 between %T and %T\n",
DR_REF (dra), DR_REF (drb));
/* When we perform grouped accesses and perform implicit CSE
by detecting equal accesses and doing disambiguation with
runtime alias tests like for
.. = a[i];
.. = a[i+1];
a[i] = ..;
a[i+1] = ..;
*p = ..;
.. = a[i];
.. = a[i+1];
where we will end up loading { a[i], a[i+1] } once, make
sure that inserting group loads before the first load and
stores after the last store will do the right thing.
Similar for groups like
a[i] = ...;
... = a[i];
a[i+1] = ...;
where loads from the group interleave with the store. */
if (!vect_preserves_scalar_order_p (dr_info_a, dr_info_b))
return opt_result::failure_at (stmtinfo_a->stmt,
"READ_WRITE dependence"
" in interleaving.\n");
if (loop->safelen < 2)
{
tree indicator = dr_zero_step_indicator (dra);
if (!indicator || integer_zerop (indicator))
return opt_result::failure_at (stmtinfo_a->stmt,
"access also has a zero step\n");
else if (TREE_CODE (indicator) != INTEGER_CST)
vect_check_nonzero_value (loop_vinfo, indicator);
}
continue;
}
if (dist > 0 && DDR_REVERSED_P (ddr))
{
/* If DDR_REVERSED_P the order of the data-refs in DDR was
reversed (to make distance vector positive), and the actual
distance is negative. */
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"dependence distance negative.\n");
/* When doing outer loop vectorization, we need to check if there is
a backward dependence at the inner loop level if the dependence
at the outer loop is reversed. See PR81740. */
if (nested_in_vect_loop_p (loop, stmtinfo_a)
|| nested_in_vect_loop_p (loop, stmtinfo_b))
{
unsigned inner_depth = index_in_loop_nest (loop->inner->num,
DDR_LOOP_NEST (ddr));
if (dist_v[inner_depth] < 0)
return opt_result::failure_at (stmtinfo_a->stmt,
"not vectorized, dependence "
"between data-refs %T and %T\n",
DR_REF (dra), DR_REF (drb));
}
/* Record a negative dependence distance to later limit the
amount of stmt copying / unrolling we can perform.
Only need to handle read-after-write dependence. */
if (DR_IS_READ (drb)
&& (STMT_VINFO_MIN_NEG_DIST (stmtinfo_b) == 0
|| STMT_VINFO_MIN_NEG_DIST (stmtinfo_b) > (unsigned)dist))
STMT_VINFO_MIN_NEG_DIST (stmtinfo_b) = dist;
continue;
}
unsigned int abs_dist = abs (dist);
if (abs_dist >= 2 && abs_dist < *max_vf)
{
/* The dependence distance requires reduction of the maximal
vectorization factor. */
*max_vf = abs_dist;
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"adjusting maximal vectorization factor to %i\n",
*max_vf);
}
if (abs_dist >= *max_vf)
{
/* Dependence distance does not create dependence, as far as
vectorization is concerned, in this case. */
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"dependence distance >= VF.\n");
continue;
}
return opt_result::failure_at (stmtinfo_a->stmt,
"not vectorized, possible dependence "
"between data-refs %T and %T\n",
DR_REF (dra), DR_REF (drb));
}
return opt_result::success ();
}
/* Function vect_analyze_data_ref_dependences.
Examine all the data references in the loop, and make sure there do not
exist any data dependences between them. Set *MAX_VF according to
the maximum vectorization factor the data dependences allow. */
opt_result
vect_analyze_data_ref_dependences (loop_vec_info loop_vinfo,
unsigned int *max_vf)
{
unsigned int i;
struct data_dependence_relation *ddr;
DUMP_VECT_SCOPE ("vect_analyze_data_ref_dependences");
if (!LOOP_VINFO_DDRS (loop_vinfo).exists ())
{
LOOP_VINFO_DDRS (loop_vinfo)
.create (LOOP_VINFO_DATAREFS (loop_vinfo).length ()
* LOOP_VINFO_DATAREFS (loop_vinfo).length ());
/* We do not need read-read dependences. */
bool res = compute_all_dependences (LOOP_VINFO_DATAREFS (loop_vinfo),
&LOOP_VINFO_DDRS (loop_vinfo),
LOOP_VINFO_LOOP_NEST (loop_vinfo),
false);
gcc_assert (res);
}
LOOP_VINFO_NO_DATA_DEPENDENCIES (loop_vinfo) = true;
/* For epilogues we either have no aliases or alias versioning
was applied to original loop. Therefore we may just get max_vf
using VF of original loop. */
if (LOOP_VINFO_EPILOGUE_P (loop_vinfo))
*max_vf = LOOP_VINFO_ORIG_MAX_VECT_FACTOR (loop_vinfo);
else
FOR_EACH_VEC_ELT (LOOP_VINFO_DDRS (loop_vinfo), i, ddr)
{
opt_result res
= vect_analyze_data_ref_dependence (ddr, loop_vinfo, max_vf);
if (!res)
return res;
}
return opt_result::success ();
}
/* Function vect_slp_analyze_data_ref_dependence.
Return TRUE if there (might) exist a dependence between a memory-reference
DRA and a memory-reference DRB for VINFO. When versioning for alias
may check a dependence at run-time, return FALSE. Adjust *MAX_VF
according to the data dependence. */
static bool
vect_slp_analyze_data_ref_dependence (vec_info *vinfo,
struct data_dependence_relation *ddr)
{
struct data_reference *dra = DDR_A (ddr);
struct data_reference *drb = DDR_B (ddr);
dr_vec_info *dr_info_a = vinfo->lookup_dr (dra);
dr_vec_info *dr_info_b = vinfo->lookup_dr (drb);
/* We need to check dependences of statements marked as unvectorizable
as well, they still can prohibit vectorization. */
/* Independent data accesses. */
if (DDR_ARE_DEPENDENT (ddr) == chrec_known)
return false;
if (dra == drb)
return false;
/* Read-read is OK. */
if (DR_IS_READ (dra) && DR_IS_READ (drb))
return false;
/* If dra and drb are part of the same interleaving chain consider
them independent. */
if (STMT_VINFO_GROUPED_ACCESS (dr_info_a->stmt)
&& (DR_GROUP_FIRST_ELEMENT (dr_info_a->stmt)
== DR_GROUP_FIRST_ELEMENT (dr_info_b->stmt)))
return false;
/* Unknown data dependence. */
if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"can't determine dependence between %T and %T\n",
DR_REF (dra), DR_REF (drb));
}
else if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"determined dependence between %T and %T\n",
DR_REF (dra), DR_REF (drb));
return true;
}
/* Analyze dependences involved in the transform of SLP NODE. STORES
contain the vector of scalar stores of this instance if we are
disambiguating the loads. */
static bool
vect_slp_analyze_node_dependences (vec_info *vinfo, slp_tree node,
vec<stmt_vec_info> stores,
stmt_vec_info last_store_info)
{
/* This walks over all stmts involved in the SLP load/store done
in NODE verifying we can sink them up to the last stmt in the
group. */
if (DR_IS_WRITE (STMT_VINFO_DATA_REF (SLP_TREE_REPRESENTATIVE (node))))
{
stmt_vec_info last_access_info = vect_find_last_scalar_stmt_in_slp (node);
for (unsigned k = 0; k < SLP_TREE_SCALAR_STMTS (node).length (); ++k)
{
stmt_vec_info access_info
= vect_orig_stmt (SLP_TREE_SCALAR_STMTS (node)[k]);
if (access_info == last_access_info)
continue;
data_reference *dr_a = STMT_VINFO_DATA_REF (access_info);
ao_ref ref;
bool ref_initialized_p = false;
for (gimple_stmt_iterator gsi = gsi_for_stmt (access_info->stmt);
gsi_stmt (gsi) != last_access_info->stmt; gsi_next (&gsi))
{
gimple *stmt = gsi_stmt (gsi);
if (! gimple_vuse (stmt))
continue;
/* If we couldn't record a (single) data reference for this
stmt we have to resort to the alias oracle. */
stmt_vec_info stmt_info = vinfo->lookup_stmt (stmt);
data_reference *dr_b = STMT_VINFO_DATA_REF (stmt_info);
if (!dr_b)
{
/* We are moving a store - this means
we cannot use TBAA for disambiguation. */
if (!ref_initialized_p)
ao_ref_init (&ref, DR_REF (dr_a));
if (stmt_may_clobber_ref_p_1 (stmt, &ref, false)
|| ref_maybe_used_by_stmt_p (stmt, &ref, false))
return false;
continue;
}
bool dependent = false;
/* If we run into a store of this same instance (we've just
marked those) then delay dependence checking until we run
into the last store because this is where it will have
been sunk to (and we verify if we can do that as well). */
if (gimple_visited_p (stmt))
{
if (stmt_info != last_store_info)
continue;
for (stmt_vec_info &store_info : stores)
{
data_reference *store_dr
= STMT_VINFO_DATA_REF (store_info);
ddr_p ddr = initialize_data_dependence_relation
(dr_a, store_dr, vNULL);
dependent
= vect_slp_analyze_data_ref_dependence (vinfo, ddr);
free_dependence_relation (ddr);
if (dependent)
break;
}
}
else
{
ddr_p ddr = initialize_data_dependence_relation (dr_a,
dr_b, vNULL);
dependent = vect_slp_analyze_data_ref_dependence (vinfo, ddr);
free_dependence_relation (ddr);
}
if (dependent)
return false;
}
}
}
else /* DR_IS_READ */
{
stmt_vec_info first_access_info
= vect_find_first_scalar_stmt_in_slp (node);
for (unsigned k = 0; k < SLP_TREE_SCALAR_STMTS (node).length (); ++k)
{
stmt_vec_info access_info
= vect_orig_stmt (SLP_TREE_SCALAR_STMTS (node)[k]);
if (access_info == first_access_info)
continue;
data_reference *dr_a = STMT_VINFO_DATA_REF (access_info);
ao_ref ref;
bool ref_initialized_p = false;
for (gimple_stmt_iterator gsi = gsi_for_stmt (access_info->stmt);
gsi_stmt (gsi) != first_access_info->stmt; gsi_prev (&gsi))
{
gimple *stmt = gsi_stmt (gsi);
if (! gimple_vdef (stmt))
continue;
/* If we couldn't record a (single) data reference for this
stmt we have to resort to the alias oracle. */
stmt_vec_info stmt_info = vinfo->lookup_stmt (stmt);
data_reference *dr_b = STMT_VINFO_DATA_REF (stmt_info);
/* We are hoisting a load - this means we can use
TBAA for disambiguation. */
if (!ref_initialized_p)
ao_ref_init (&ref, DR_REF (dr_a));
if (stmt_may_clobber_ref_p_1 (stmt, &ref, true))
{
if (!dr_b)
return false;
/* Resort to dependence checking below. */
}
else
/* No dependence. */
continue;
bool dependent = false;
/* If we run into a store of this same instance (we've just
marked those) then delay dependence checking until we run
into the last store because this is where it will have
been sunk to (and we verify if we can do that as well). */
if (gimple_visited_p (stmt))
{
if (stmt_info != last_store_info)
continue;
for (stmt_vec_info &store_info : stores)
{
data_reference *store_dr
= STMT_VINFO_DATA_REF (store_info);
ddr_p ddr = initialize_data_dependence_relation
(dr_a, store_dr, vNULL);
dependent
= vect_slp_analyze_data_ref_dependence (vinfo, ddr);
free_dependence_relation (ddr);
if (dependent)
break;
}
}
else
{
ddr_p ddr = initialize_data_dependence_relation (dr_a,
dr_b, vNULL);
dependent = vect_slp_analyze_data_ref_dependence (vinfo, ddr);
free_dependence_relation (ddr);
}
if (dependent)
return false;
}
}
}
return true;
}
/* Function vect_analyze_data_ref_dependences.
Examine all the data references in the basic-block, and make sure there
do not exist any data dependences between them. Set *MAX_VF according to
the maximum vectorization factor the data dependences allow. */
bool
vect_slp_analyze_instance_dependence (vec_info *vinfo, slp_instance instance)
{
DUMP_VECT_SCOPE ("vect_slp_analyze_instance_dependence");
/* The stores of this instance are at the root of the SLP tree. */
slp_tree store = NULL;
if (SLP_INSTANCE_KIND (instance) == slp_inst_kind_store)
store = SLP_INSTANCE_TREE (instance);
/* Verify we can sink stores to the vectorized stmt insert location. */
stmt_vec_info last_store_info = NULL;
if (store)
{
if (! vect_slp_analyze_node_dependences (vinfo, store, vNULL, NULL))
return false;
/* Mark stores in this instance and remember the last one. */
last_store_info = vect_find_last_scalar_stmt_in_slp (store);
for (unsigned k = 0; k < SLP_TREE_SCALAR_STMTS (store).length (); ++k)
gimple_set_visited (SLP_TREE_SCALAR_STMTS (store)[k]->stmt, true);
}
bool res = true;
/* Verify we can sink loads to the vectorized stmt insert location,
special-casing stores of this instance. */
for (slp_tree &load : SLP_INSTANCE_LOADS (instance))
if (! vect_slp_analyze_node_dependences (vinfo, load,
store
? SLP_TREE_SCALAR_STMTS (store)
: vNULL, last_store_info))
{
res = false;
break;
}
/* Unset the visited flag. */
if (store)
for (unsigned k = 0; k < SLP_TREE_SCALAR_STMTS (store).length (); ++k)
gimple_set_visited (SLP_TREE_SCALAR_STMTS (store)[k]->stmt, false);
return res;
}
/* Return the misalignment of DR_INFO accessed in VECTYPE with OFFSET
applied. */
int
dr_misalignment (dr_vec_info *dr_info, tree vectype, poly_int64 offset)
{
HOST_WIDE_INT diff = 0;
/* Alignment is only analyzed for the first element of a DR group,
use that but adjust misalignment by the offset of the access. */
if (STMT_VINFO_GROUPED_ACCESS (dr_info->stmt))
{
dr_vec_info *first_dr
= STMT_VINFO_DR_INFO (DR_GROUP_FIRST_ELEMENT (dr_info->stmt));
/* vect_analyze_data_ref_accesses guarantees that DR_INIT are
INTEGER_CSTs and the first element in the group has the lowest
address. */
diff = (TREE_INT_CST_LOW (DR_INIT (dr_info->dr))
- TREE_INT_CST_LOW (DR_INIT (first_dr->dr)));
gcc_assert (diff >= 0);
dr_info = first_dr;
}
int misalign = dr_info->misalignment;
gcc_assert (misalign != DR_MISALIGNMENT_UNINITIALIZED);
if (misalign == DR_MISALIGNMENT_UNKNOWN)
return misalign;
/* If the access is only aligned for a vector type with smaller alignment
requirement the access has unknown misalignment. */
if (maybe_lt (dr_info->target_alignment * BITS_PER_UNIT,
targetm.vectorize.preferred_vector_alignment (vectype)))
return DR_MISALIGNMENT_UNKNOWN;
/* Apply the offset from the DR group start and the externally supplied
offset which can for example result from a negative stride access. */
poly_int64 misalignment = misalign + diff + offset;
/* vect_compute_data_ref_alignment will have ensured that target_alignment
is constant and otherwise set misalign to DR_MISALIGNMENT_UNKNOWN. */
unsigned HOST_WIDE_INT target_alignment_c
= dr_info->target_alignment.to_constant ();
if (!known_misalignment (misalignment, target_alignment_c, &misalign))
return DR_MISALIGNMENT_UNKNOWN;
return misalign;
}
/* Record the base alignment guarantee given by DRB, which occurs
in STMT_INFO. */
static void
vect_record_base_alignment (vec_info *vinfo, stmt_vec_info stmt_info,
innermost_loop_behavior *drb)
{
bool existed;
std::pair<stmt_vec_info, innermost_loop_behavior *> &entry
= vinfo->base_alignments.get_or_insert (drb->base_address, &existed);
if (!existed || entry.second->base_alignment < drb->base_alignment)
{
entry = std::make_pair (stmt_info, drb);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"recording new base alignment for %T\n"
" alignment: %d\n"
" misalignment: %d\n"
" based on: %G",
drb->base_address,
drb->base_alignment,
drb->base_misalignment,
stmt_info->stmt);
}
}
/* If the region we're going to vectorize is reached, all unconditional
data references occur at least once. We can therefore pool the base
alignment guarantees from each unconditional reference. Do this by
going through all the data references in VINFO and checking whether
the containing statement makes the reference unconditionally. If so,
record the alignment of the base address in VINFO so that it can be
used for all other references with the same base. */
void
vect_record_base_alignments (vec_info *vinfo)
{
loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
class loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL;
for (data_reference *dr : vinfo->shared->datarefs)
{
dr_vec_info *dr_info = vinfo->lookup_dr (dr);
stmt_vec_info stmt_info = dr_info->stmt;
if (!DR_IS_CONDITIONAL_IN_STMT (dr)
&& STMT_VINFO_VECTORIZABLE (stmt_info)
&& !STMT_VINFO_GATHER_SCATTER_P (stmt_info))
{
vect_record_base_alignment (vinfo, stmt_info, &DR_INNERMOST (dr));
/* If DR is nested in the loop that is being vectorized, we can also
record the alignment of the base wrt the outer loop. */
if (loop && nested_in_vect_loop_p (loop, stmt_info))
vect_record_base_alignment
(vinfo, stmt_info, &STMT_VINFO_DR_WRT_VEC_LOOP (stmt_info));
}
}
}
/* Function vect_compute_data_ref_alignment
Compute the misalignment of the data reference DR_INFO when vectorizing
with VECTYPE.
Output:
1. initialized misalignment info for DR_INFO
FOR NOW: No analysis is actually performed. Misalignment is calculated
only for trivial cases. TODO. */
static void
vect_compute_data_ref_alignment (vec_info *vinfo, dr_vec_info *dr_info,
tree vectype)
{
stmt_vec_info stmt_info = dr_info->stmt;
vec_base_alignments *base_alignments = &vinfo->base_alignments;
loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
class loop *loop = NULL;
tree ref = DR_REF (dr_info->dr);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"vect_compute_data_ref_alignment:\n");
if (loop_vinfo)
loop = LOOP_VINFO_LOOP (loop_vinfo);
/* Initialize misalignment to unknown. */
SET_DR_MISALIGNMENT (dr_info, DR_MISALIGNMENT_UNKNOWN);
if (STMT_VINFO_GATHER_SCATTER_P (stmt_info))
return;
innermost_loop_behavior *drb = vect_dr_behavior (vinfo, dr_info);
bool step_preserves_misalignment_p;
poly_uint64 vector_alignment
= exact_div (targetm.vectorize.preferred_vector_alignment (vectype),
BITS_PER_UNIT);
SET_DR_TARGET_ALIGNMENT (dr_info, vector_alignment);
/* If the main loop has peeled for alignment we have no way of knowing
whether the data accesses in the epilogues are aligned. We can't at
compile time answer the question whether we have entered the main loop or
not. Fixes PR 92351. */
if (loop_vinfo)
{
loop_vec_info orig_loop_vinfo = LOOP_VINFO_ORIG_LOOP_INFO (loop_vinfo);
if (orig_loop_vinfo
&& LOOP_VINFO_PEELING_FOR_ALIGNMENT (orig_loop_vinfo) != 0)
return;
}
unsigned HOST_WIDE_INT vect_align_c;
if (!vector_alignment.is_constant (&vect_align_c))
return;
/* No step for BB vectorization. */
if (!loop)
{
gcc_assert (integer_zerop (drb->step));
step_preserves_misalignment_p = true;
}
/* In case the dataref is in an inner-loop of the loop that is being
vectorized (LOOP), we use the base and misalignment information
relative to the outer-loop (LOOP). This is ok only if the misalignment
stays the same throughout the execution of the inner-loop, which is why
we have to check that the stride of the dataref in the inner-loop evenly
divides by the vector alignment. */
else if (nested_in_vect_loop_p (loop, stmt_info))
{
step_preserves_misalignment_p
= (DR_STEP_ALIGNMENT (dr_info->dr) % vect_align_c) == 0;
if (dump_enabled_p ())
{
if (step_preserves_misalignment_p)
dump_printf_loc (MSG_NOTE, vect_location,
"inner step divides the vector alignment.\n");
else
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"inner step doesn't divide the vector"
" alignment.\n");
}
}
/* Similarly we can only use base and misalignment information relative to
an innermost loop if the misalignment stays the same throughout the
execution of the loop. As above, this is the case if the stride of
the dataref evenly divides by the alignment. */
else
{
poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
step_preserves_misalignment_p
= multiple_p (DR_STEP_ALIGNMENT (dr_info->dr) * vf, vect_align_c);
if (!step_preserves_misalignment_p && dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"step doesn't divide the vector alignment.\n");
}
unsigned int base_alignment = drb->base_alignment;
unsigned int base_misalignment = drb->base_misalignment;
/* Calculate the maximum of the pooled base address alignment and the
alignment that we can compute for DR itself. */
std::pair<stmt_vec_info, innermost_loop_behavior *> *entry
= base_alignments->get (drb->base_address);
if (entry
&& base_alignment < (*entry).second->base_alignment
&& (loop_vinfo
|| (dominated_by_p (CDI_DOMINATORS, gimple_bb (stmt_info->stmt),
gimple_bb (entry->first->stmt))
&& (gimple_bb (stmt_info->stmt) != gimple_bb (entry->first->stmt)
|| (entry->first->dr_aux.group <= dr_info->group)))))
{
base_alignment = entry->second->base_alignment;
base_misalignment = entry->second->base_misalignment;
}
if (drb->offset_alignment < vect_align_c
|| !step_preserves_misalignment_p
/* We need to know whether the step wrt the vectorized loop is
negative when computing the starting misalignment below. */
|| TREE_CODE (drb->step) != INTEGER_CST)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Unknown alignment for access: %T\n", ref);
return;
}
if (base_alignment < vect_align_c)
{
unsigned int max_alignment;
tree base = get_base_for_alignment (drb->base_address, &max_alignment);
if (max_alignment < vect_align_c
|| !vect_can_force_dr_alignment_p (base,
vect_align_c * BITS_PER_UNIT))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"can't force alignment of ref: %T\n", ref);
return;
}
/* Force the alignment of the decl.
NOTE: This is the only change to the code we make during
the analysis phase, before deciding to vectorize the loop. */
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"force alignment of %T\n", ref);
dr_info->base_decl = base;
dr_info->base_misaligned = true;
base_misalignment = 0;
}
poly_int64 misalignment
= base_misalignment + wi::to_poly_offset (drb->init).force_shwi ();
unsigned int const_misalignment;
if (!known_misalignment (misalignment, vect_align_c, &const_misalignment))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Non-constant misalignment for access: %T\n", ref);
return;
}
SET_DR_MISALIGNMENT (dr_info, const_misalignment);
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"misalign = %d bytes of ref %T\n",
const_misalignment, ref);
return;
}
/* Return whether DR_INFO, which is related to DR_PEEL_INFO in
that it only differs in DR_INIT, is aligned if DR_PEEL_INFO
is made aligned via peeling. */
static bool
vect_dr_aligned_if_related_peeled_dr_is (dr_vec_info *dr_info,
dr_vec_info *dr_peel_info)
{
if (multiple_p (DR_TARGET_ALIGNMENT (dr_peel_info),
DR_TARGET_ALIGNMENT (dr_info)))
{
poly_offset_int diff
= (wi::to_poly_offset (DR_INIT (dr_peel_info->dr))
- wi::to_poly_offset (DR_INIT (dr_info->dr)));
if (known_eq (diff, 0)
|| multiple_p (diff, DR_TARGET_ALIGNMENT (dr_info)))
return true;
}
return false;
}
/* Return whether DR_INFO is aligned if DR_PEEL_INFO is made
aligned via peeling. */
static bool
vect_dr_aligned_if_peeled_dr_is (dr_vec_info *dr_info,
dr_vec_info *dr_peel_info)
{
if (!operand_equal_p (DR_BASE_ADDRESS (dr_info->dr),
DR_BASE_ADDRESS (dr_peel_info->dr), 0)
|| !operand_equal_p (DR_OFFSET (dr_info->dr),
DR_OFFSET (dr_peel_info->dr), 0)
|| !operand_equal_p (DR_STEP (dr_info->dr),
DR_STEP (dr_peel_info->dr), 0))
return false;
return vect_dr_aligned_if_related_peeled_dr_is (dr_info, dr_peel_info);
}
/* Compute the value for dr_info->misalign so that the access appears
aligned. This is used by peeling to compensate for dr_misalignment
applying the offset for negative step. */
int
vect_dr_misalign_for_aligned_access (dr_vec_info *dr_info)
{
if (tree_int_cst_sgn (DR_STEP (dr_info->dr)) >= 0)
return 0;
tree vectype = STMT_VINFO_VECTYPE (dr_info->stmt);
poly_int64 misalignment
= ((TYPE_VECTOR_SUBPARTS (vectype) - 1)
* TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (vectype))));
unsigned HOST_WIDE_INT target_alignment_c;
int misalign;
if (!dr_info->target_alignment.is_constant (&target_alignment_c)
|| !known_misalignment (misalignment, target_alignment_c, &misalign))
return DR_MISALIGNMENT_UNKNOWN;
return misalign;
}
/* Function vect_update_misalignment_for_peel.
Sets DR_INFO's misalignment
- to 0 if it has the same alignment as DR_PEEL_INFO,
- to the misalignment computed using NPEEL if DR_INFO's salignment is known,
- to -1 (unknown) otherwise.
DR_INFO - the data reference whose misalignment is to be adjusted.
DR_PEEL_INFO - the data reference whose misalignment is being made
zero in the vector loop by the peel.
NPEEL - the number of iterations in the peel loop if the misalignment
of DR_PEEL_INFO is known at compile time. */
static void
vect_update_misalignment_for_peel (dr_vec_info *dr_info,
dr_vec_info *dr_peel_info, int npeel)
{
/* If dr_info is aligned of dr_peel_info is, then mark it so. */
if (vect_dr_aligned_if_peeled_dr_is (dr_info, dr_peel_info))
{
SET_DR_MISALIGNMENT (dr_info,
vect_dr_misalign_for_aligned_access (dr_peel_info));
return;
}
unsigned HOST_WIDE_INT alignment;
if (DR_TARGET_ALIGNMENT (dr_info).is_constant (&alignment)
&& known_alignment_for_access_p (dr_info,
STMT_VINFO_VECTYPE (dr_info->stmt))
&& known_alignment_for_access_p (dr_peel_info,
STMT_VINFO_VECTYPE (dr_peel_info->stmt)))
{
int misal = dr_info->misalignment;
misal += npeel * TREE_INT_CST_LOW (DR_STEP (dr_info->dr));
misal &= alignment - 1;
set_dr_misalignment (dr_info, misal);
return;
}
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location, "Setting misalignment " \
"to unknown (-1).\n");
SET_DR_MISALIGNMENT (dr_info, DR_MISALIGNMENT_UNKNOWN);
}
/* Return true if alignment is relevant for DR_INFO. */
static bool
vect_relevant_for_alignment_p (dr_vec_info *dr_info)
{
stmt_vec_info stmt_info = dr_info->stmt;
if (!STMT_VINFO_RELEVANT_P (stmt_info))
return false;
/* For interleaving, only the alignment of the first access matters. */
if (STMT_VINFO_GROUPED_ACCESS (stmt_info)
&& DR_GROUP_FIRST_ELEMENT (stmt_info) != stmt_info)
return false;
/* Scatter-gather and invariant accesses continue to address individual
scalars, so vector-level alignment is irrelevant. */
if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)
|| integer_zerop (DR_STEP (dr_info->dr)))
return false;
/* Strided accesses perform only component accesses, alignment is
irrelevant for them. */
if (STMT_VINFO_STRIDED_P (stmt_info)
&& !STMT_VINFO_GROUPED_ACCESS (stmt_info))
return false;
return true;
}
/* Given an memory reference EXP return whether its alignment is less
than its size. */
static bool
not_size_aligned (tree exp)
{
if (!tree_fits_uhwi_p (TYPE_SIZE (TREE_TYPE (exp))))
return true;
return (tree_to_uhwi (TYPE_SIZE (TREE_TYPE (exp)))
> get_object_alignment (exp));
}
/* Function vector_alignment_reachable_p
Return true if vector alignment for DR_INFO is reachable by peeling
a few loop iterations. Return false otherwise. */
static bool
vector_alignment_reachable_p (dr_vec_info *dr_info)
{
stmt_vec_info stmt_info = dr_info->stmt;
tree vectype = STMT_VINFO_VECTYPE (stmt_info);
if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
{
/* For interleaved access we peel only if number of iterations in
the prolog loop ({VF - misalignment}), is a multiple of the
number of the interleaved accesses. */
int elem_size, mis_in_elements;
/* FORNOW: handle only known alignment. */
if (!known_alignment_for_access_p (dr_info, vectype))
return false;
poly_uint64 nelements = TYPE_VECTOR_SUBPARTS (vectype);
poly_uint64 vector_size = GET_MODE_SIZE (TYPE_MODE (vectype));
elem_size = vector_element_size (vector_size, nelements);
mis_in_elements = dr_misalignment (dr_info, vectype) / elem_size;
if (!multiple_p (nelements - mis_in_elements, DR_GROUP_SIZE (stmt_info)))
return false;
}
/* If misalignment is known at the compile time then allow peeling
only if natural alignment is reachable through peeling. */
if (known_alignment_for_access_p (dr_info, vectype)
&& !aligned_access_p (dr_info, vectype))
{
HOST_WIDE_INT elmsize =
int_cst_value (TYPE_SIZE_UNIT (TREE_TYPE (vectype)));
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"data size = %wd. misalignment = %d.\n", elmsize,
dr_misalignment (dr_info, vectype));
}
if (dr_misalignment (dr_info, vectype) % elmsize)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"data size does not divide the misalignment.\n");
return false;
}
}
if (!known_alignment_for_access_p (dr_info, vectype))
{
tree type = TREE_TYPE (DR_REF (dr_info->dr));
bool is_packed = not_size_aligned (DR_REF (dr_info->dr));
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Unknown misalignment, %snaturally aligned\n",
is_packed ? "not " : "");
return targetm.vectorize.vector_alignment_reachable (type, is_packed);
}
return true;
}
/* Calculate the cost of the memory access represented by DR_INFO. */
static void
vect_get_data_access_cost (vec_info *vinfo, dr_vec_info *dr_info,
dr_alignment_support alignment_support_scheme,
int misalignment,
unsigned int *inside_cost,
unsigned int *outside_cost,
stmt_vector_for_cost *body_cost_vec,
stmt_vector_for_cost *prologue_cost_vec)
{
stmt_vec_info stmt_info = dr_info->stmt;
loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
int ncopies;
if (PURE_SLP_STMT (stmt_info))
ncopies = 1;
else
ncopies = vect_get_num_copies (loop_vinfo, STMT_VINFO_VECTYPE (stmt_info));
if (DR_IS_READ (dr_info->dr))
vect_get_load_cost (vinfo, stmt_info, ncopies, alignment_support_scheme,
misalignment, true, inside_cost,
outside_cost, prologue_cost_vec, body_cost_vec, false);
else
vect_get_store_cost (vinfo,stmt_info, ncopies, alignment_support_scheme,
misalignment, inside_cost, body_cost_vec);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"vect_get_data_access_cost: inside_cost = %d, "
"outside_cost = %d.\n", *inside_cost, *outside_cost);
}
typedef struct _vect_peel_info
{
dr_vec_info *dr_info;
int npeel;
unsigned int count;
} *vect_peel_info;
typedef struct _vect_peel_extended_info
{
vec_info *vinfo;
struct _vect_peel_info peel_info;
unsigned int inside_cost;
unsigned int outside_cost;
} *vect_peel_extended_info;
/* Peeling hashtable helpers. */
struct peel_info_hasher : free_ptr_hash <_vect_peel_info>
{
static inline hashval_t hash (const _vect_peel_info *);
static inline bool equal (const _vect_peel_info *, const _vect_peel_info *);
};
inline hashval_t
peel_info_hasher::hash (const _vect_peel_info *peel_info)
{
return (hashval_t) peel_info->npeel;
}
inline bool
peel_info_hasher::equal (const _vect_peel_info *a, const _vect_peel_info *b)
{
return (a->npeel == b->npeel);
}
/* Insert DR_INFO into peeling hash table with NPEEL as key. */
static void
vect_peeling_hash_insert (hash_table<peel_info_hasher> *peeling_htab,
loop_vec_info loop_vinfo, dr_vec_info *dr_info,
int npeel, bool supportable_if_not_aligned)
{
struct _vect_peel_info elem, *slot;
_vect_peel_info **new_slot;
elem.npeel = npeel;
slot = peeling_htab->find (&elem);
if (slot)
slot->count++;
else
{
slot = XNEW (struct _vect_peel_info);
slot->npeel = npeel;
slot->dr_info = dr_info;
slot->count = 1;
new_slot = peeling_htab->find_slot (slot, INSERT);
*new_slot = slot;
}
/* If this DR is not supported with unknown misalignment then bias
this slot when the cost model is disabled. */
if (!supportable_if_not_aligned
&& unlimited_cost_model (LOOP_VINFO_LOOP (loop_vinfo)))
slot->count += VECT_MAX_COST;
}
/* Traverse peeling hash table to find peeling option that aligns maximum
number of data accesses. */
int
vect_peeling_hash_get_most_frequent (_vect_peel_info **slot,
_vect_peel_extended_info *max)
{
vect_peel_info elem = *slot;
if (elem->count > max->peel_info.count
|| (elem->count == max->peel_info.count
&& max->peel_info.npeel > elem->npeel))
{
max->peel_info.npeel = elem->npeel;
max->peel_info.count = elem->count;
max->peel_info.dr_info = elem->dr_info;
}
return 1;
}
/* Get the costs of peeling NPEEL iterations for LOOP_VINFO, checking
data access costs for all data refs. If UNKNOWN_MISALIGNMENT is true,
npeel is computed at runtime but DR0_INFO's misalignment will be zero
after peeling. */
static void
vect_get_peeling_costs_all_drs (loop_vec_info loop_vinfo,
dr_vec_info *dr0_info,
unsigned int *inside_cost,
unsigned int *outside_cost,
stmt_vector_for_cost *body_cost_vec,
stmt_vector_for_cost *prologue_cost_vec,
unsigned int npeel)
{
vec<data_reference_p> datarefs = LOOP_VINFO_DATAREFS (loop_vinfo);
bool dr0_alignment_known_p
= (dr0_info
&& known_alignment_for_access_p (dr0_info,
STMT_VINFO_VECTYPE (dr0_info->stmt)));
for (data_reference *dr : datarefs)
{
dr_vec_info *dr_info = loop_vinfo->lookup_dr (dr);
if (!vect_relevant_for_alignment_p (dr_info))
continue;
tree vectype = STMT_VINFO_VECTYPE (dr_info->stmt);
dr_alignment_support alignment_support_scheme;
int misalignment;
unsigned HOST_WIDE_INT alignment;
bool negative = tree_int_cst_compare (DR_STEP (dr_info->dr),
size_zero_node) < 0;
poly_int64 off = 0;
if (negative)
off = ((TYPE_VECTOR_SUBPARTS (vectype) - 1)
* -TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (vectype))));
if (npeel == 0)
misalignment = dr_misalignment (dr_info, vectype, off);
else if (dr_info == dr0_info
|| vect_dr_aligned_if_peeled_dr_is (dr_info, dr0_info))
misalignment = 0;
else if (!dr0_alignment_known_p
|| !known_alignment_for_access_p (dr_info, vectype)
|| !DR_TARGET_ALIGNMENT (dr_info).is_constant (&alignment))
misalignment = DR_MISALIGNMENT_UNKNOWN;
else
{
misalignment = dr_misalignment (dr_info, vectype, off);
misalignment += npeel * TREE_INT_CST_LOW (DR_STEP (dr_info->dr));
misalignment &= alignment - 1;
}
alignment_support_scheme
= vect_supportable_dr_alignment (loop_vinfo, dr_info, vectype,
misalignment);
vect_get_data_access_cost (loop_vinfo, dr_info,
alignment_support_scheme, misalignment,
inside_cost, outside_cost,
body_cost_vec, prologue_cost_vec);
}
}
/* Traverse peeling hash table and calculate cost for each peeling option.
Find the one with the lowest cost. */
int
vect_peeling_hash_get_lowest_cost (_vect_peel_info **slot,
_vect_peel_extended_info *min)
{
vect_peel_info elem = *slot;
int dummy;
unsigned int inside_cost = 0, outside_cost = 0;
loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (min->vinfo);
stmt_vector_for_cost prologue_cost_vec, body_cost_vec,
epilogue_cost_vec;
prologue_cost_vec.create (2);
body_cost_vec.create (2);
epilogue_cost_vec.create (2);
vect_get_peeling_costs_all_drs (loop_vinfo, elem->dr_info, &inside_cost,
&outside_cost, &body_cost_vec,
&prologue_cost_vec, elem->npeel);
body_cost_vec.release ();
outside_cost += vect_get_known_peeling_cost
(loop_vinfo, elem->npeel, &dummy,
&LOOP_VINFO_SCALAR_ITERATION_COST (loop_vinfo),
&prologue_cost_vec, &epilogue_cost_vec);
/* Prologue and epilogue costs are added to the target model later.
These costs depend only on the scalar iteration cost, the
number of peeling iterations finally chosen, and the number of
misaligned statements. So discard the information found here. */
prologue_cost_vec.release ();
epilogue_cost_vec.release ();
if (inside_cost < min->inside_cost
|| (inside_cost == min->inside_cost
&& outside_cost < min->outside_cost))
{
min->inside_cost = inside_cost;
min->outside_cost = outside_cost;
min->peel_info.dr_info = elem->dr_info;
min->peel_info.npeel = elem->npeel;
min->peel_info.count = elem->count;
}
return 1;
}
/* Choose best peeling option by traversing peeling hash table and either
choosing an option with the lowest cost (if cost model is enabled) or the
option that aligns as many accesses as possible. */
static struct _vect_peel_extended_info
vect_peeling_hash_choose_best_peeling (hash_table<peel_info_hasher> *peeling_htab,
loop_vec_info loop_vinfo)
{
struct _vect_peel_extended_info res;
res.peel_info.dr_info = NULL;
res.vinfo = loop_vinfo;
if (!unlimited_cost_model (LOOP_VINFO_LOOP (loop_vinfo)))
{
res.inside_cost = INT_MAX;
res.outside_cost = INT_MAX;
peeling_htab->traverse <_vect_peel_extended_info *,
vect_peeling_hash_get_lowest_cost> (&res);
}
else
{
res.peel_info.count = 0;
peeling_htab->traverse <_vect_peel_extended_info *,
vect_peeling_hash_get_most_frequent> (&res);
res.inside_cost = 0;
res.outside_cost = 0;
}
return res;
}
/* Return true if the new peeling NPEEL is supported. */
static bool
vect_peeling_supportable (loop_vec_info loop_vinfo, dr_vec_info *dr0_info,
unsigned npeel)
{
vec<data_reference_p> datarefs = LOOP_VINFO_DATAREFS (loop_vinfo);
enum dr_alignment_support supportable_dr_alignment;
bool dr0_alignment_known_p
= known_alignment_for_access_p (dr0_info,
STMT_VINFO_VECTYPE (dr0_info->stmt));
/* Ensure that all data refs can be vectorized after the peel. */
for (data_reference *dr : datarefs)
{
if (dr == dr0_info->dr)
continue;
dr_vec_info *dr_info = loop_vinfo->lookup_dr (dr);
if (!vect_relevant_for_alignment_p (dr_info)
|| vect_dr_aligned_if_peeled_dr_is (dr_info, dr0_info))
continue;
tree vectype = STMT_VINFO_VECTYPE (dr_info->stmt);
int misalignment;
unsigned HOST_WIDE_INT alignment;
if (!dr0_alignment_known_p
|| !known_alignment_for_access_p (dr_info, vectype)
|| !DR_TARGET_ALIGNMENT (dr_info).is_constant (&alignment))
misalignment = DR_MISALIGNMENT_UNKNOWN;
else
{
misalignment = dr_misalignment (dr_info, vectype);
misalignment += npeel * TREE_INT_CST_LOW (DR_STEP (dr_info->dr));
misalignment &= alignment - 1;
}
supportable_dr_alignment
= vect_supportable_dr_alignment (loop_vinfo, dr_info, vectype,
misalignment);
if (supportable_dr_alignment == dr_unaligned_unsupported)
return false;
}
return true;
}
/* Compare two data-references DRA and DRB to group them into chunks
with related alignment. */
static int
dr_align_group_sort_cmp (const void *dra_, const void *drb_)
{
data_reference_p dra = *(data_reference_p *)const_cast<void *>(dra_);
data_reference_p drb = *(data_reference_p *)const_cast<void *>(drb_);
int cmp;
/* Stabilize sort. */
if (dra == drb)
return 0;
/* Ordering of DRs according to base. */
cmp = data_ref_compare_tree (DR_BASE_ADDRESS (dra),
DR_BASE_ADDRESS (drb));
if (cmp != 0)
return cmp;
/* And according to DR_OFFSET. */
cmp = data_ref_compare_tree (DR_OFFSET (dra), DR_OFFSET (drb));
if (cmp != 0)
return cmp;
/* And after step. */
cmp = data_ref_compare_tree (DR_STEP (dra), DR_STEP (drb));
if (cmp != 0)
return cmp;
/* Then sort after DR_INIT. In case of identical DRs sort after stmt UID. */
cmp = data_ref_compare_tree (DR_INIT (dra), DR_INIT (drb));
if (cmp == 0)
return gimple_uid (DR_STMT (dra)) < gimple_uid (DR_STMT (drb)) ? -1 : 1;
return cmp;
}
/* Function vect_enhance_data_refs_alignment
This pass will use loop versioning and loop peeling in order to enhance
the alignment of data references in the loop.
FOR NOW: we assume that whatever versioning/peeling takes place, only the
original loop is to be vectorized. Any other loops that are created by
the transformations performed in this pass - are not supposed to be
vectorized. This restriction will be relaxed.
This pass will require a cost model to guide it whether to apply peeling
or versioning or a combination of the two. For example, the scheme that
intel uses when given a loop with several memory accesses, is as follows:
choose one memory access ('p') which alignment you want to force by doing
peeling. Then, either (1) generate a loop in which 'p' is aligned and all
other accesses are not necessarily aligned, or (2) use loop versioning to
generate one loop in which all accesses are aligned, and another loop in
which only 'p' is necessarily aligned.
("Automatic Intra-Register Vectorization for the Intel Architecture",
Aart J.C. Bik, Milind Girkar, Paul M. Grey and Ximmin Tian, International
Journal of Parallel Programming, Vol. 30, No. 2, April 2002.)
Devising a cost model is the most critical aspect of this work. It will
guide us on which access to peel for, whether to use loop versioning, how
many versions to create, etc. The cost model will probably consist of
generic considerations as well as target specific considerations (on
powerpc for example, misaligned stores are more painful than misaligned
loads).
Here are the general steps involved in alignment enhancements:
-- original loop, before alignment analysis:
for (i=0; i<N; i++){
x = q[i]; # DR_MISALIGNMENT(q) = unknown
p[i] = y; # DR_MISALIGNMENT(p) = unknown
}
-- After vect_compute_data_refs_alignment:
for (i=0; i<N; i++){
x = q[i]; # DR_MISALIGNMENT(q) = 3
p[i] = y; # DR_MISALIGNMENT(p) = unknown
}
-- Possibility 1: we do loop versioning:
if (p is aligned) {
for (i=0; i<N; i++){ # loop 1A
x = q[i]; # DR_MISALIGNMENT(q) = 3
p[i] = y; # DR_MISALIGNMENT(p) = 0
}
}
else {
for (i=0; i<N; i++){ # loop 1B
x = q[i]; # DR_MISALIGNMENT(q) = 3
p[i] = y; # DR_MISALIGNMENT(p) = unaligned
}
}
-- Possibility 2: we do loop peeling:
for (i = 0; i < 3; i++){ # (scalar loop, not to be vectorized).
x = q[i];
p[i] = y;
}
for (i = 3; i < N; i++){ # loop 2A
x = q[i]; # DR_MISALIGNMENT(q) = 0
p[i] = y; # DR_MISALIGNMENT(p) = unknown
}
-- Possibility 3: combination of loop peeling and versioning:
for (i = 0; i < 3; i++){ # (scalar loop, not to be vectorized).
x = q[i];
p[i] = y;
}
if (p is aligned) {
for (i = 3; i<N; i++){ # loop 3A
x = q[i]; # DR_MISALIGNMENT(q) = 0
p[i] = y; # DR_MISALIGNMENT(p) = 0
}
}
else {
for (i = 3; i<N; i++){ # loop 3B
x = q[i]; # DR_MISALIGNMENT(q) = 0
p[i] = y; # DR_MISALIGNMENT(p) = unaligned
}
}
These loops are later passed to loop_transform to be vectorized. The
vectorizer will use the alignment information to guide the transformation
(whether to generate regular loads/stores, or with special handling for
misalignment). */
opt_result
vect_enhance_data_refs_alignment (loop_vec_info loop_vinfo)
{
class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
dr_vec_info *first_store = NULL;
dr_vec_info *dr0_info = NULL;
struct data_reference *dr;
unsigned int i;
bool do_peeling = false;
bool do_versioning = false;
unsigned int npeel = 0;
bool one_misalignment_known = false;
bool one_misalignment_unknown = false;
bool one_dr_unsupportable = false;
dr_vec_info *unsupportable_dr_info = NULL;
unsigned int dr0_same_align_drs = 0, first_store_same_align_drs = 0;
hash_table<peel_info_hasher> peeling_htab (1);
DUMP_VECT_SCOPE ("vect_enhance_data_refs_alignment");
/* Reset data so we can safely be called multiple times. */
LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo).truncate (0);
LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) = 0;
if (LOOP_VINFO_DATAREFS (loop_vinfo).is_empty ())
return opt_result::success ();
/* Sort the vector of datarefs so DRs that have the same or dependent
alignment are next to each other. */
auto_vec<data_reference_p> datarefs
= LOOP_VINFO_DATAREFS (loop_vinfo).copy ();
datarefs.qsort (dr_align_group_sort_cmp);
/* Compute the number of DRs that become aligned when we peel
a dataref so it becomes aligned. */
auto_vec<unsigned> n_same_align_refs (datarefs.length ());
n_same_align_refs.quick_grow_cleared (datarefs.length ());
unsigned i0;
for (i0 = 0; i0 < datarefs.length (); ++i0)
if (DR_BASE_ADDRESS (datarefs[i0]))
break;
for (i = i0 + 1; i <= datarefs.length (); ++i)
{
if (i == datarefs.length ()
|| !operand_equal_p (DR_BASE_ADDRESS (datarefs[i0]),
DR_BASE_ADDRESS (datarefs[i]), 0)
|| !operand_equal_p (DR_OFFSET (datarefs[i0]),
DR_OFFSET (datarefs[i]), 0)
|| !operand_equal_p (DR_STEP (datarefs[i0]),
DR_STEP (datarefs[i]), 0))
{
/* The subgroup [i0, i-1] now only differs in DR_INIT and
possibly DR_TARGET_ALIGNMENT. Still the whole subgroup
will get known misalignment if we align one of the refs
with the largest DR_TARGET_ALIGNMENT. */
for (unsigned j = i0; j < i; ++j)
{
dr_vec_info *dr_infoj = loop_vinfo->lookup_dr (datarefs[j]);
for (unsigned k = i0; k < i; ++k)
{
if (k == j)
continue;
dr_vec_info *dr_infok = loop_vinfo->lookup_dr (datarefs[k]);
if (vect_dr_aligned_if_related_peeled_dr_is (dr_infok,
dr_infoj))
n_same_align_refs[j]++;
}
}
i0 = i;
}
}
/* While cost model enhancements are expected in the future, the high level
view of the code at this time is as follows:
A) If there is a misaligned access then see if peeling to align
this access can make all data references satisfy
vect_supportable_dr_alignment. If so, update data structures
as needed and return true.
B) If peeling wasn't possible and there is a data reference with an
unknown misalignment that does not satisfy vect_supportable_dr_alignment
then see if loop versioning checks can be used to make all data
references satisfy vect_supportable_dr_alignment. If so, update
data structures as needed and return true.
C) If neither peeling nor versioning were successful then return false if
any data reference does not satisfy vect_supportable_dr_alignment.
D) Return true (all data references satisfy vect_supportable_dr_alignment).
Note, Possibility 3 above (which is peeling and versioning together) is not
being done at this time. */
/* (1) Peeling to force alignment. */
/* (1.1) Decide whether to perform peeling, and how many iterations to peel:
Considerations:
+ How many accesses will become aligned due to the peeling
- How many accesses will become unaligned due to the peeling,
and the cost of misaligned accesses.
- The cost of peeling (the extra runtime checks, the increase
in code size). */
FOR_EACH_VEC_ELT (datarefs, i, dr)
{
dr_vec_info *dr_info = loop_vinfo->lookup_dr (dr);
if (!vect_relevant_for_alignment_p (dr_info))
continue;
stmt_vec_info stmt_info = dr_info->stmt;
tree vectype = STMT_VINFO_VECTYPE (stmt_info);
do_peeling = vector_alignment_reachable_p (dr_info);
if (do_peeling)
{
if (known_alignment_for_access_p (dr_info, vectype))
{
unsigned int npeel_tmp = 0;
bool negative = tree_int_cst_compare (DR_STEP (dr),
size_zero_node) < 0;
/* If known_alignment_for_access_p then we have set
DR_MISALIGNMENT which is only done if we know it at compiler
time, so it is safe to assume target alignment is constant.
*/
unsigned int target_align =
DR_TARGET_ALIGNMENT (dr_info).to_constant ();
unsigned HOST_WIDE_INT dr_size = vect_get_scalar_dr_size (dr_info);
poly_int64 off = 0;
if (negative)
off = (TYPE_VECTOR_SUBPARTS (vectype) - 1) * -dr_size;
unsigned int mis = dr_misalignment (dr_info, vectype, off);
mis = negative ? mis : -mis;
if (mis != 0)
npeel_tmp = (mis & (target_align - 1)) / dr_size;
/* For multiple types, it is possible that the bigger type access
will have more than one peeling option. E.g., a loop with two
types: one of size (vector size / 4), and the other one of
size (vector size / 8). Vectorization factor will 8. If both
accesses are misaligned by 3, the first one needs one scalar
iteration to be aligned, and the second one needs 5. But the
first one will be aligned also by peeling 5 scalar
iterations, and in that case both accesses will be aligned.
Hence, except for the immediate peeling amount, we also want
to try to add full vector size, while we don't exceed
vectorization factor.
We do this automatically for cost model, since we calculate
cost for every peeling option. */
poly_uint64 nscalars = npeel_tmp;
if (unlimited_cost_model (LOOP_VINFO_LOOP (loop_vinfo)))
{
poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
nscalars = (STMT_SLP_TYPE (stmt_info)
? vf * DR_GROUP_SIZE (stmt_info) : vf);
}
/* Save info about DR in the hash table. Also include peeling
amounts according to the explanation above. Indicate
the alignment status when the ref is not aligned.
??? Rather than using unknown alignment here we should
prune all entries from the peeling hashtable which cause
DRs to be not supported. */
bool supportable_if_not_aligned
= vect_supportable_dr_alignment
(loop_vinfo, dr_info, vectype, DR_MISALIGNMENT_UNKNOWN);
while (known_le (npeel_tmp, nscalars))
{
vect_peeling_hash_insert (&peeling_htab, loop_vinfo,
dr_info, npeel_tmp,
supportable_if_not_aligned);
npeel_tmp += MAX (1, target_align / dr_size);
}
one_misalignment_known = true;
}
else
{
/* If we don't know any misalignment values, we prefer
peeling for data-ref that has the maximum number of data-refs
with the same alignment, unless the target prefers to align
stores over load. */
unsigned same_align_drs = n_same_align_refs[i];
if (!dr0_info
|| dr0_same_align_drs < same_align_drs)
{
dr0_same_align_drs = same_align_drs;
dr0_info = dr_info;
}
/* For data-refs with the same number of related
accesses prefer the one where the misalign
computation will be invariant in the outermost loop. */
else if (dr0_same_align_drs == same_align_drs)
{
class loop *ivloop0, *ivloop;
ivloop0 = outermost_invariant_loop_for_expr
(loop, DR_BASE_ADDRESS (dr0_info->dr));
ivloop = outermost_invariant_loop_for_expr
(loop, DR_BASE_ADDRESS (dr));
if ((ivloop && !ivloop0)
|| (ivloop && ivloop0
&& flow_loop_nested_p (ivloop, ivloop0)))
dr0_info = dr_info;
}
one_misalignment_unknown = true;
/* Check for data refs with unsupportable alignment that
can be peeled. */
enum dr_alignment_support supportable_dr_alignment
= vect_supportable_dr_alignment (loop_vinfo, dr_info, vectype,
DR_MISALIGNMENT_UNKNOWN);
if (supportable_dr_alignment == dr_unaligned_unsupported)
{
one_dr_unsupportable = true;
unsupportable_dr_info = dr_info;
}
if (!first_store && DR_IS_WRITE (dr))
{
first_store = dr_info;
first_store_same_align_drs = same_align_drs;
}
}
}
else
{
if (!aligned_access_p (dr_info, vectype))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"vector alignment may not be reachable\n");
break;
}
}
}
/* Check if we can possibly peel the loop. */
if (!vect_can_advance_ivs_p (loop_vinfo)
|| !slpeel_can_duplicate_loop_p (loop, single_exit (loop))
|| loop->inner)
do_peeling = false;
struct _vect_peel_extended_info peel_for_known_alignment;
struct _vect_peel_extended_info peel_for_unknown_alignment;
struct _vect_peel_extended_info best_peel;
peel_for_unknown_alignment.inside_cost = INT_MAX;
peel_for_unknown_alignment.outside_cost = INT_MAX;
peel_for_unknown_alignment.peel_info.count = 0;
if (do_peeling
&& one_misalignment_unknown)
{
/* Check if the target requires to prefer stores over loads, i.e., if
misaligned stores are more expensive than misaligned loads (taking
drs with same alignment into account). */
unsigned int load_inside_cost = 0;
unsigned int load_outside_cost = 0;
unsigned int store_inside_cost = 0;
unsigned int store_outside_cost = 0;
unsigned int estimated_npeels = vect_vf_for_cost (loop_vinfo) / 2;
stmt_vector_for_cost dummy;
dummy.create (2);
vect_get_peeling_costs_all_drs (loop_vinfo, dr0_info,
&load_inside_cost,
&load_outside_cost,
&dummy, &dummy, estimated_npeels);
dummy.release ();
if (first_store)
{
dummy.create (2);
vect_get_peeling_costs_all_drs (loop_vinfo, first_store,
&store_inside_cost,
&store_outside_cost,
&dummy, &dummy,
estimated_npeels);
dummy.release ();
}
else
{
store_inside_cost = INT_MAX;
store_outside_cost = INT_MAX;
}
if (load_inside_cost > store_inside_cost
|| (load_inside_cost == store_inside_cost
&& load_outside_cost > store_outside_cost))
{
dr0_info = first_store;
dr0_same_align_drs = first_store_same_align_drs;
peel_for_unknown_alignment.inside_cost = store_inside_cost;
peel_for_unknown_alignment.outside_cost = store_outside_cost;
}
else
{
peel_for_unknown_alignment.inside_cost = load_inside_cost;
peel_for_unknown_alignment.outside_cost = load_outside_cost;
}
stmt_vector_for_cost prologue_cost_vec, epilogue_cost_vec;
prologue_cost_vec.create (2);
epilogue_cost_vec.create (2);
int dummy2;
peel_for_unknown_alignment.outside_cost += vect_get_known_peeling_cost
(loop_vinfo, estimated_npeels, &dummy2,
&LOOP_VINFO_SCALAR_ITERATION_COST (loop_vinfo),
&prologue_cost_vec, &epilogue_cost_vec);
prologue_cost_vec.release ();
epilogue_cost_vec.release ();
peel_for_unknown_alignment.peel_info.count = dr0_same_align_drs + 1;
}
peel_for_unknown_alignment.peel_info.npeel = 0;
peel_for_unknown_alignment.peel_info.dr_info = dr0_info;
best_peel = peel_for_unknown_alignment;
peel_for_known_alignment.inside_cost = INT_MAX;
peel_for_known_alignment.outside_cost = INT_MAX;
peel_for_known_alignment.peel_info.count = 0;
peel_for_known_alignment.peel_info.dr_info = NULL;
if (do_peeling && one_misalignment_known)
{
/* Peeling is possible, but there is no data access that is not supported
unless aligned. So we try to choose the best possible peeling from
the hash table. */
peel_for_known_alignment = vect_peeling_hash_choose_best_peeling
(&peeling_htab, loop_vinfo);
}
/* Compare costs of peeling for known and unknown alignment. */
if (peel_for_known_alignment.peel_info.dr_info != NULL
&& peel_for_unknown_alignment.inside_cost
>= peel_for_known_alignment.inside_cost)
{
best_peel = peel_for_known_alignment;
/* If the best peeling for known alignment has NPEEL == 0, perform no
peeling at all except if there is an unsupportable dr that we can
align. */
if (best_peel.peel_info.npeel == 0 && !one_dr_unsupportable)
do_peeling = false;
}
/* If there is an unsupportable data ref, prefer this over all choices so far
since we'd have to discard a chosen peeling except when it accidentally
aligned the unsupportable data ref. */
if (one_dr_unsupportable)
dr0_info = unsupportable_dr_info;
else if (do_peeling)
{
/* Calculate the penalty for no peeling, i.e. leaving everything as-is.
TODO: Use nopeel_outside_cost or get rid of it? */
unsigned nopeel_inside_cost = 0;
unsigned nopeel_outside_cost = 0;
stmt_vector_for_cost dummy;
dummy.create (2);
vect_get_peeling_costs_all_drs (loop_vinfo, NULL, &nopeel_inside_cost,
&nopeel_outside_cost, &dummy, &dummy, 0);
dummy.release ();
/* Add epilogue costs. As we do not peel for alignment here, no prologue
costs will be recorded. */
stmt_vector_for_cost prologue_cost_vec, epilogue_cost_vec;
prologue_cost_vec.create (2);
epilogue_cost_vec.create (2);
int dummy2;
nopeel_outside_cost += vect_get_known_peeling_cost
(loop_vinfo, 0, &dummy2,
&LOOP_VINFO_SCALAR_ITERATION_COST (loop_vinfo),
&prologue_cost_vec, &epilogue_cost_vec);
prologue_cost_vec.release ();
epilogue_cost_vec.release ();
npeel = best_peel.peel_info.npeel;
dr0_info = best_peel.peel_info.dr_info;
/* If no peeling is not more expensive than the best peeling we
have so far, don't perform any peeling. */
if (nopeel_inside_cost <= best_peel.inside_cost)
do_peeling = false;
}
if (do_peeling)
{
stmt_vec_info stmt_info = dr0_info->stmt;
if (known_alignment_for_access_p (dr0_info,
STMT_VINFO_VECTYPE (stmt_info)))
{
bool negative = tree_int_cst_compare (DR_STEP (dr0_info->dr),
size_zero_node) < 0;
if (!npeel)
{
/* Since it's known at compile time, compute the number of
iterations in the peeled loop (the peeling factor) for use in
updating DR_MISALIGNMENT values. The peeling factor is the
vectorization factor minus the misalignment as an element
count. */
tree vectype = STMT_VINFO_VECTYPE (stmt_info);
poly_int64 off = 0;
if (negative)
off = ((TYPE_VECTOR_SUBPARTS (vectype) - 1)
* -TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (vectype))));
unsigned int mis
= dr_misalignment (dr0_info, vectype, off);
mis = negative ? mis : -mis;
/* If known_alignment_for_access_p then we have set
DR_MISALIGNMENT which is only done if we know it at compiler
time, so it is safe to assume target alignment is constant.
*/
unsigned int target_align =
DR_TARGET_ALIGNMENT (dr0_info).to_constant ();
npeel = ((mis & (target_align - 1))
/ vect_get_scalar_dr_size (dr0_info));
}
/* For interleaved data access every iteration accesses all the
members of the group, therefore we divide the number of iterations
by the group size. */
if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
npeel /= DR_GROUP_SIZE (stmt_info);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Try peeling by %d\n", npeel);
}
/* Ensure that all datarefs can be vectorized after the peel. */
if (!vect_peeling_supportable (loop_vinfo, dr0_info, npeel))
do_peeling = false;
/* Check if all datarefs are supportable and log. */
if (do_peeling
&& npeel == 0
&& known_alignment_for_access_p (dr0_info,
STMT_VINFO_VECTYPE (stmt_info)))
return opt_result::success ();
/* Cost model #1 - honor --param vect-max-peeling-for-alignment. */
if (do_peeling)
{
unsigned max_allowed_peel
= param_vect_max_peeling_for_alignment;
if (loop_cost_model (loop) <= VECT_COST_MODEL_CHEAP)
max_allowed_peel = 0;
if (max_allowed_peel != (unsigned)-1)
{
unsigned max_peel = npeel;
if (max_peel == 0)
{
poly_uint64 target_align = DR_TARGET_ALIGNMENT (dr0_info);
unsigned HOST_WIDE_INT target_align_c;
if (target_align.is_constant (&target_align_c))
max_peel =
target_align_c / vect_get_scalar_dr_size (dr0_info) - 1;
else
{
do_peeling = false;
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Disable peeling, max peels set and vector"
" alignment unknown\n");
}
}
if (max_peel > max_allowed_peel)
{
do_peeling = false;
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Disable peeling, max peels reached: %d\n", max_peel);
}
}
}
/* Cost model #2 - if peeling may result in a remaining loop not
iterating enough to be vectorized then do not peel. Since this
is a cost heuristic rather than a correctness decision, use the
most likely runtime value for variable vectorization factors. */
if (do_peeling
&& LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo))
{
unsigned int assumed_vf = vect_vf_for_cost (loop_vinfo);
unsigned int max_peel = npeel == 0 ? assumed_vf - 1 : npeel;
if ((unsigned HOST_WIDE_INT) LOOP_VINFO_INT_NITERS (loop_vinfo)
< assumed_vf + max_peel)
do_peeling = false;
}
if (do_peeling)
{
/* (1.2) Update the DR_MISALIGNMENT of each data reference DR_i.
If the misalignment of DR_i is identical to that of dr0 then set
DR_MISALIGNMENT (DR_i) to zero. If the misalignment of DR_i and
dr0 are known at compile time then increment DR_MISALIGNMENT (DR_i)
by the peeling factor times the element size of DR_i (MOD the
vectorization factor times the size). Otherwise, the
misalignment of DR_i must be set to unknown. */
FOR_EACH_VEC_ELT (datarefs, i, dr)
if (dr != dr0_info->dr)
{
dr_vec_info *dr_info = loop_vinfo->lookup_dr (dr);
if (!vect_relevant_for_alignment_p (dr_info))
continue;
vect_update_misalignment_for_peel (dr_info, dr0_info, npeel);
}
LOOP_VINFO_UNALIGNED_DR (loop_vinfo) = dr0_info;
if (npeel)
LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) = npeel;
else
LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) = -1;
SET_DR_MISALIGNMENT (dr0_info,
vect_dr_misalign_for_aligned_access (dr0_info));
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"Alignment of access forced using peeling.\n");
dump_printf_loc (MSG_NOTE, vect_location,
"Peeling for alignment will be applied.\n");
}
/* The inside-loop cost will be accounted for in vectorizable_load
and vectorizable_store correctly with adjusted alignments.
Drop the body_cst_vec on the floor here. */
return opt_result::success ();
}
}
/* (2) Versioning to force alignment. */
/* Try versioning if:
1) optimize loop for speed and the cost-model is not cheap
2) there is at least one unsupported misaligned data ref with an unknown
misalignment, and
3) all misaligned data refs with a known misalignment are supported, and
4) the number of runtime alignment checks is within reason. */
do_versioning
= (optimize_loop_nest_for_speed_p (loop)
&& !loop->inner /* FORNOW */
&& loop_cost_model (loop) > VECT_COST_MODEL_CHEAP);
if (do_versioning)
{
FOR_EACH_VEC_ELT (datarefs, i, dr)
{
dr_vec_info *dr_info = loop_vinfo->lookup_dr (dr);
if (!vect_relevant_for_alignment_p (dr_info))
continue;
stmt_vec_info stmt_info = dr_info->stmt;
if (STMT_VINFO_STRIDED_P (stmt_info))
{
do_versioning = false;
break;
}
tree vectype = STMT_VINFO_VECTYPE (stmt_info);
bool negative = tree_int_cst_compare (DR_STEP (dr),
size_zero_node) < 0;
poly_int64 off = 0;
if (negative)
off = ((TYPE_VECTOR_SUBPARTS (vectype) - 1)
* -TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (vectype))));
int misalignment;
if ((misalignment = dr_misalignment (dr_info, vectype, off)) == 0)
continue;
enum dr_alignment_support supportable_dr_alignment
= vect_supportable_dr_alignment (loop_vinfo, dr_info, vectype,
misalignment);
if (supportable_dr_alignment == dr_unaligned_unsupported)
{
if (misalignment != DR_MISALIGNMENT_UNKNOWN
|| (LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo).length ()
>= (unsigned) param_vect_max_version_for_alignment_checks))
{
do_versioning = false;
break;
}
/* At present we don't support versioning for alignment
with variable VF, since there's no guarantee that the
VF is a power of two. We could relax this if we added
a way of enforcing a power-of-two size. */
unsigned HOST_WIDE_INT size;
if (!GET_MODE_SIZE (TYPE_MODE (vectype)).is_constant (&size))
{
do_versioning = false;
break;
}
/* Forcing alignment in the first iteration is no good if
we don't keep it across iterations. For now, just disable
versioning in this case.
?? We could actually unroll the loop to achieve the required
overall step alignment, and forcing the alignment could be
done by doing some iterations of the non-vectorized loop. */
if (!multiple_p (LOOP_VINFO_VECT_FACTOR (loop_vinfo)
* DR_STEP_ALIGNMENT (dr),
DR_TARGET_ALIGNMENT (dr_info)))
{
do_versioning = false;
break;
}
/* The rightmost bits of an aligned address must be zeros.
Construct the mask needed for this test. For example,
GET_MODE_SIZE for the vector mode V4SI is 16 bytes so the
mask must be 15 = 0xf. */
int mask = size - 1;
/* FORNOW: use the same mask to test all potentially unaligned
references in the loop. */
if (LOOP_VINFO_PTR_MASK (loop_vinfo)
&& LOOP_VINFO_PTR_MASK (loop_vinfo) != mask)
{
do_versioning = false;
break;
}
LOOP_VINFO_PTR_MASK (loop_vinfo) = mask;
LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo).safe_push (stmt_info);
}
}
/* Versioning requires at least one misaligned data reference. */
if (!LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo))
do_versioning = false;
else if (!do_versioning)
LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo).truncate (0);
}
if (do_versioning)
{
const vec<stmt_vec_info> &may_misalign_stmts
= LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo);
stmt_vec_info stmt_info;
/* It can now be assumed that the data references in the statements
in LOOP_VINFO_MAY_MISALIGN_STMTS will be aligned in the version
of the loop being vectorized. */
FOR_EACH_VEC_ELT (may_misalign_stmts, i, stmt_info)
{
dr_vec_info *dr_info = STMT_VINFO_DR_INFO (stmt_info);
SET_DR_MISALIGNMENT (dr_info,
vect_dr_misalign_for_aligned_access (dr_info));
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Alignment of access forced using versioning.\n");
}
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Versioning for alignment will be applied.\n");
/* Peeling and versioning can't be done together at this time. */
gcc_assert (! (do_peeling && do_versioning));
return opt_result::success ();
}
/* This point is reached if neither peeling nor versioning is being done. */
gcc_assert (! (do_peeling || do_versioning));
return opt_result::success ();
}
/* Function vect_analyze_data_refs_alignment
Analyze the alignment of the data-references in the loop.
Return FALSE if a data reference is found that cannot be vectorized. */
opt_result
vect_analyze_data_refs_alignment (loop_vec_info loop_vinfo)
{
DUMP_VECT_SCOPE ("vect_analyze_data_refs_alignment");
vec<data_reference_p> datarefs = LOOP_VINFO_DATAREFS (loop_vinfo);
struct data_reference *dr;
unsigned int i;
vect_record_base_alignments (loop_vinfo);
FOR_EACH_VEC_ELT (datarefs, i, dr)
{
dr_vec_info *dr_info = loop_vinfo->lookup_dr (dr);
if (STMT_VINFO_VECTORIZABLE (dr_info->stmt))
{
if (STMT_VINFO_GROUPED_ACCESS (dr_info->stmt)
&& DR_GROUP_FIRST_ELEMENT (dr_info->stmt) != dr_info->stmt)
continue;
vect_compute_data_ref_alignment (loop_vinfo, dr_info,
STMT_VINFO_VECTYPE (dr_info->stmt));
}
}
return opt_result::success ();
}
/* Analyze alignment of DRs of stmts in NODE. */
static bool
vect_slp_analyze_node_alignment (vec_info *vinfo, slp_tree node)
{
/* Alignment is maintained in the first element of the group. */
stmt_vec_info first_stmt_info = SLP_TREE_SCALAR_STMTS (node)[0];
first_stmt_info = DR_GROUP_FIRST_ELEMENT (first_stmt_info);
dr_vec_info *dr_info = STMT_VINFO_DR_INFO (first_stmt_info);
tree vectype = SLP_TREE_VECTYPE (node);
poly_uint64 vector_alignment
= exact_div (targetm.vectorize.preferred_vector_alignment (vectype),
BITS_PER_UNIT);
if (dr_info->misalignment == DR_MISALIGNMENT_UNINITIALIZED)
vect_compute_data_ref_alignment (vinfo, dr_info, SLP_TREE_VECTYPE (node));
/* Re-analyze alignment when we're facing a vectorization with a bigger
alignment requirement. */
else if (known_lt (dr_info->target_alignment, vector_alignment))
{
poly_uint64 old_target_alignment = dr_info->target_alignment;
int old_misalignment = dr_info->misalignment;
vect_compute_data_ref_alignment (vinfo, dr_info, SLP_TREE_VECTYPE (node));
/* But keep knowledge about a smaller alignment. */
if (old_misalignment != DR_MISALIGNMENT_UNKNOWN
&& dr_info->misalignment == DR_MISALIGNMENT_UNKNOWN)
{
dr_info->target_alignment = old_target_alignment;
dr_info->misalignment = old_misalignment;
}
}
/* When we ever face unordered target alignments the first one wins in terms
of analyzing and the other will become unknown in dr_misalignment. */
return true;
}
/* Function vect_slp_analyze_instance_alignment
Analyze the alignment of the data-references in the SLP instance.
Return FALSE if a data reference is found that cannot be vectorized. */
bool
vect_slp_analyze_instance_alignment (vec_info *vinfo,
slp_instance instance)
{
DUMP_VECT_SCOPE ("vect_slp_analyze_instance_alignment");
slp_tree node;
unsigned i;
FOR_EACH_VEC_ELT (SLP_INSTANCE_LOADS (instance), i, node)
if (! vect_slp_analyze_node_alignment (vinfo, node))
return false;
if (SLP_INSTANCE_KIND (instance) == slp_inst_kind_store
&& ! vect_slp_analyze_node_alignment
(vinfo, SLP_INSTANCE_TREE (instance)))
return false;
return true;
}
/* Analyze groups of accesses: check that DR_INFO belongs to a group of
accesses of legal size, step, etc. Detect gaps, single element
interleaving, and other special cases. Set grouped access info.
Collect groups of strided stores for further use in SLP analysis.
Worker for vect_analyze_group_access. */
static bool
vect_analyze_group_access_1 (vec_info *vinfo, dr_vec_info *dr_info)
{
data_reference *dr = dr_info->dr;
tree step = DR_STEP (dr);
tree scalar_type = TREE_TYPE (DR_REF (dr));
HOST_WIDE_INT type_size = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (scalar_type));
stmt_vec_info stmt_info = dr_info->stmt;
loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (vinfo);
HOST_WIDE_INT dr_step = -1;
HOST_WIDE_INT groupsize, last_accessed_element = 1;
bool slp_impossible = false;
/* For interleaving, GROUPSIZE is STEP counted in elements, i.e., the
size of the interleaving group (including gaps). */
if (tree_fits_shwi_p (step))
{
dr_step = tree_to_shwi (step);
/* Check that STEP is a multiple of type size. Otherwise there is
a non-element-sized gap at the end of the group which we
cannot represent in DR_GROUP_GAP or DR_GROUP_SIZE.
??? As we can handle non-constant step fine here we should
simply remove uses of DR_GROUP_GAP between the last and first
element and instead rely on DR_STEP. DR_GROUP_SIZE then would
simply not include that gap. */
if ((dr_step % type_size) != 0)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Step %T is not a multiple of the element size"
" for %T\n",
step, DR_REF (dr));
return false;
}
groupsize = absu_hwi (dr_step) / type_size;
}
else
groupsize = 0;
/* Not consecutive access is possible only if it is a part of interleaving. */
if (!DR_GROUP_FIRST_ELEMENT (stmt_info))
{
/* Check if it this DR is a part of interleaving, and is a single
element of the group that is accessed in the loop. */
/* Gaps are supported only for loads. STEP must be a multiple of the type
size. */
if (DR_IS_READ (dr)
&& (dr_step % type_size) == 0
&& groupsize > 0
/* This could be UINT_MAX but as we are generating code in a very
inefficient way we have to cap earlier.
See PR91403 for example. */
&& groupsize <= 4096)
{
DR_GROUP_FIRST_ELEMENT (stmt_info) = stmt_info;
DR_GROUP_SIZE (stmt_info) = groupsize;
DR_GROUP_GAP (stmt_info) = groupsize - 1;
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Detected single element interleaving %T"
" step %T\n",
DR_REF (dr), step);
return true;
}
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not consecutive access %G", stmt_info->stmt);
if (bb_vinfo)
{
/* Mark the statement as unvectorizable. */
STMT_VINFO_VECTORIZABLE (stmt_info) = false;
return true;
}
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location, "using strided accesses\n");
STMT_VINFO_STRIDED_P (stmt_info) = true;
return true;
}
if (DR_GROUP_FIRST_ELEMENT (stmt_info) == stmt_info)
{
/* First stmt in the interleaving chain. Check the chain. */
stmt_vec_info next = DR_GROUP_NEXT_ELEMENT (stmt_info);
struct data_reference *data_ref = dr;
unsigned int count = 1;
tree prev_init = DR_INIT (data_ref);
HOST_WIDE_INT diff, gaps = 0;
/* By construction, all group members have INTEGER_CST DR_INITs. */
while (next)
{
/* We never have the same DR multiple times. */
gcc_assert (tree_int_cst_compare (DR_INIT (data_ref),
DR_INIT (STMT_VINFO_DATA_REF (next))) != 0);
data_ref = STMT_VINFO_DATA_REF (next);
/* All group members have the same STEP by construction. */
gcc_checking_assert (operand_equal_p (DR_STEP (data_ref), step, 0));
/* Check that the distance between two accesses is equal to the type
size. Otherwise, we have gaps. */
diff = (TREE_INT_CST_LOW (DR_INIT (data_ref))
- TREE_INT_CST_LOW (prev_init)) / type_size;
if (diff < 1 || diff > UINT_MAX)
{
/* For artificial testcases with array accesses with large
constant indices we can run into overflow issues which
can end up fooling the groupsize constraint below so
check the individual gaps (which are represented as
unsigned int) as well. */
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"interleaved access with gap larger "
"than representable\n");
return false;
}
if (diff != 1)
{
/* FORNOW: SLP of accesses with gaps is not supported. */
slp_impossible = true;
if (DR_IS_WRITE (data_ref))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"interleaved store with gaps\n");
return false;
}
gaps += diff - 1;
}
last_accessed_element += diff;
/* Store the gap from the previous member of the group. If there is no
gap in the access, DR_GROUP_GAP is always 1. */
DR_GROUP_GAP (next) = diff;
prev_init = DR_INIT (data_ref);
next = DR_GROUP_NEXT_ELEMENT (next);
/* Count the number of data-refs in the chain. */
count++;
}
if (groupsize == 0)
groupsize = count + gaps;
/* This could be UINT_MAX but as we are generating code in a very
inefficient way we have to cap earlier. See PR78699 for example. */
if (groupsize > 4096)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"group is too large\n");
return false;
}
/* Check that the size of the interleaving is equal to count for stores,
i.e., that there are no gaps. */
if (groupsize != count
&& !DR_IS_READ (dr))
{
groupsize = count;
STMT_VINFO_STRIDED_P (stmt_info) = true;
}
/* If there is a gap after the last load in the group it is the
difference between the groupsize and the last accessed
element.
When there is no gap, this difference should be 0. */
DR_GROUP_GAP (stmt_info) = groupsize - last_accessed_element;
DR_GROUP_SIZE (stmt_info) = groupsize;
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"Detected interleaving ");
if (DR_IS_READ (dr))
dump_printf (MSG_NOTE, "load ");
else if (STMT_VINFO_STRIDED_P (stmt_info))
dump_printf (MSG_NOTE, "strided store ");
else
dump_printf (MSG_NOTE, "store ");
dump_printf (MSG_NOTE, "of size %u\n",
(unsigned)groupsize);
dump_printf_loc (MSG_NOTE, vect_location, "\t%G", stmt_info->stmt);
next = DR_GROUP_NEXT_ELEMENT (stmt_info);
while (next)
{
if (DR_GROUP_GAP (next) != 1)
dump_printf_loc (MSG_NOTE, vect_location,
"\t<gap of %d elements>\n",
DR_GROUP_GAP (next) - 1);
dump_printf_loc (MSG_NOTE, vect_location, "\t%G", next->stmt);
next = DR_GROUP_NEXT_ELEMENT (next);
}
if (DR_GROUP_GAP (stmt_info) != 0)
dump_printf_loc (MSG_NOTE, vect_location,
"\t<gap of %d elements>\n",
DR_GROUP_GAP (stmt_info));
}
/* SLP: create an SLP data structure for every interleaving group of
stores for further analysis in vect_analyse_slp. */
if (DR_IS_WRITE (dr) && !slp_impossible)
{
if (loop_vinfo)
LOOP_VINFO_GROUPED_STORES (loop_vinfo).safe_push (stmt_info);
if (bb_vinfo)
BB_VINFO_GROUPED_STORES (bb_vinfo).safe_push (stmt_info);
}
}
return true;
}
/* Analyze groups of accesses: check that DR_INFO belongs to a group of
accesses of legal size, step, etc. Detect gaps, single element
interleaving, and other special cases. Set grouped access info.
Collect groups of strided stores for further use in SLP analysis. */
static bool
vect_analyze_group_access (vec_info *vinfo, dr_vec_info *dr_info)
{
if (!vect_analyze_group_access_1 (vinfo, dr_info))
{
/* Dissolve the group if present. */
stmt_vec_info stmt_info = DR_GROUP_FIRST_ELEMENT (dr_info->stmt);
while (stmt_info)
{
stmt_vec_info next = DR_GROUP_NEXT_ELEMENT (stmt_info);
DR_GROUP_FIRST_ELEMENT (stmt_info) = NULL;
DR_GROUP_NEXT_ELEMENT (stmt_info) = NULL;
stmt_info = next;
}
return false;
}
return true;
}
/* Analyze the access pattern of the data-reference DR_INFO.
In case of non-consecutive accesses call vect_analyze_group_access() to
analyze groups of accesses. */
static bool
vect_analyze_data_ref_access (vec_info *vinfo, dr_vec_info *dr_info)
{
data_reference *dr = dr_info->dr;
tree step = DR_STEP (dr);
tree scalar_type = TREE_TYPE (DR_REF (dr));
stmt_vec_info stmt_info = dr_info->stmt;
loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
class loop *loop = NULL;
if (STMT_VINFO_GATHER_SCATTER_P (stmt_info))
return true;
if (loop_vinfo)
loop = LOOP_VINFO_LOOP (loop_vinfo);
if (loop_vinfo && !step)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"bad data-ref access in loop\n");
return false;
}
/* Allow loads with zero step in inner-loop vectorization. */
if (loop_vinfo && integer_zerop (step))
{
DR_GROUP_FIRST_ELEMENT (stmt_info) = NULL;
if (!nested_in_vect_loop_p (loop, stmt_info))
return DR_IS_READ (dr);
/* Allow references with zero step for outer loops marked
with pragma omp simd only - it guarantees absence of
loop-carried dependencies between inner loop iterations. */
if (loop->safelen < 2)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"zero step in inner loop of nest\n");
return false;
}
}
if (loop && nested_in_vect_loop_p (loop, stmt_info))
{
/* Interleaved accesses are not yet supported within outer-loop
vectorization for references in the inner-loop. */
DR_GROUP_FIRST_ELEMENT (stmt_info) = NULL;
/* For the rest of the analysis we use the outer-loop step. */
step = STMT_VINFO_DR_STEP (stmt_info);
if (integer_zerop (step))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"zero step in outer loop.\n");
return DR_IS_READ (dr);
}
}
/* Consecutive? */
if (TREE_CODE (step) == INTEGER_CST)
{
HOST_WIDE_INT dr_step = TREE_INT_CST_LOW (step);
if (!tree_int_cst_compare (step, TYPE_SIZE_UNIT (scalar_type))
|| (dr_step < 0
&& !compare_tree_int (TYPE_SIZE_UNIT (scalar_type), -dr_step)))
{
/* Mark that it is not interleaving. */
DR_GROUP_FIRST_ELEMENT (stmt_info) = NULL;
return true;
}
}
if (loop && nested_in_vect_loop_p (loop, stmt_info))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"grouped access in outer loop.\n");
return false;
}
/* Assume this is a DR handled by non-constant strided load case. */
if (TREE_CODE (step) != INTEGER_CST)
return (STMT_VINFO_STRIDED_P (stmt_info)
&& (!STMT_VINFO_GROUPED_ACCESS (stmt_info)
|| vect_analyze_group_access (vinfo, dr_info)));
/* Not consecutive access - check if it's a part of interleaving group. */
return vect_analyze_group_access (vinfo, dr_info);
}
/* Compare two data-references DRA and DRB to group them into chunks
suitable for grouping. */
static int
dr_group_sort_cmp (const void *dra_, const void *drb_)
{
dr_vec_info *dra_info = *(dr_vec_info **)const_cast<void *>(dra_);
dr_vec_info *drb_info = *(dr_vec_info **)const_cast<void *>(drb_);
data_reference_p dra = dra_info->dr;
data_reference_p drb = drb_info->dr;
int cmp;
/* Stabilize sort. */
if (dra == drb)
return 0;
/* Different group IDs lead never belong to the same group. */
if (dra_info->group != drb_info->group)
return dra_info->group < drb_info->group ? -1 : 1;
/* Ordering of DRs according to base. */
cmp = data_ref_compare_tree (DR_BASE_ADDRESS (dra),
DR_BASE_ADDRESS (drb));
if (cmp != 0)
return cmp;
/* And according to DR_OFFSET. */
cmp = data_ref_compare_tree (DR_OFFSET (dra), DR_OFFSET (drb));
if (cmp != 0)
return cmp;
/* Put reads before writes. */
if (DR_IS_READ (dra) != DR_IS_READ (drb))
return DR_IS_READ (dra) ? -1 : 1;
/* Then sort after access size. */
cmp = data_ref_compare_tree (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dra))),
TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (drb))));
if (cmp != 0)
return cmp;
/* And after step. */
cmp = data_ref_compare_tree (DR_STEP (dra), DR_STEP (drb));
if (cmp != 0)
return cmp;
/* Then sort after DR_INIT. In case of identical DRs sort after stmt UID. */
cmp = data_ref_compare_tree (DR_INIT (dra), DR_INIT (drb));
if (cmp == 0)
return gimple_uid (DR_STMT (dra)) < gimple_uid (DR_STMT (drb)) ? -1 : 1;
return cmp;
}
/* If OP is the result of a conversion, return the unconverted value,
otherwise return null. */
static tree
strip_conversion (tree op)
{
if (TREE_CODE (op) != SSA_NAME)
return NULL_TREE;
gimple *stmt = SSA_NAME_DEF_STMT (op);
if (!is_gimple_assign (stmt)
|| !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt)))
return NULL_TREE;
return gimple_assign_rhs1 (stmt);
}
/* Return true if vectorizable_* routines can handle statements STMT1_INFO
and STMT2_INFO being in a single group. When ALLOW_SLP_P, masked loads can
be grouped in SLP mode. */
static bool
can_group_stmts_p (stmt_vec_info stmt1_info, stmt_vec_info stmt2_info,
bool allow_slp_p)
{
if (gimple_assign_single_p (stmt1_info->stmt))
return gimple_assign_single_p (stmt2_info->stmt);
gcall *call1 = dyn_cast <gcall *> (stmt1_info->stmt);
if (call1 && gimple_call_internal_p (call1))
{
/* Check for two masked loads or two masked stores. */
gcall *call2 = dyn_cast <gcall *> (stmt2_info->stmt);
if (!call2 || !gimple_call_internal_p (call2))
return false;
internal_fn ifn = gimple_call_internal_fn (call1);
if (ifn != IFN_MASK_LOAD && ifn != IFN_MASK_STORE)
return false;
if (ifn != gimple_call_internal_fn (call2))
return false;
/* Check that the masks are the same. Cope with casts of masks,
like those created by build_mask_conversion. */
tree mask1 = gimple_call_arg (call1, 2);
tree mask2 = gimple_call_arg (call2, 2);
if (!operand_equal_p (mask1, mask2, 0)
&& (ifn == IFN_MASK_STORE || !allow_slp_p))
{
mask1 = strip_conversion (mask1);
if (!mask1)
return false;
mask2 = strip_conversion (mask2);
if (!mask2)
return false;
if (!operand_equal_p (mask1, mask2, 0))
return false;
}
return true;
}
return false;
}
/* Function vect_analyze_data_ref_accesses.
Analyze the access pattern of all the data references in the loop.
FORNOW: the only access pattern that is considered vectorizable is a
simple step 1 (consecutive) access.
FORNOW: handle only arrays and pointer accesses. */
opt_result
vect_analyze_data_ref_accesses (vec_info *vinfo,
vec<int> *dataref_groups)
{
unsigned int i;
vec<data_reference_p> datarefs = vinfo->shared->datarefs;
DUMP_VECT_SCOPE ("vect_analyze_data_ref_accesses");
if (datarefs.is_empty ())
return opt_result::success ();
/* Sort the array of datarefs to make building the interleaving chains
linear. Don't modify the original vector's order, it is needed for
determining what dependencies are reversed. */
vec<dr_vec_info *> datarefs_copy;
datarefs_copy.create (datarefs.length ());
for (unsigned i = 0; i < datarefs.length (); i++)
{
dr_vec_info *dr_info = vinfo->lookup_dr (datarefs[i]);
/* If the caller computed DR grouping use that, otherwise group by
basic blocks. */
if (dataref_groups)
dr_info->group = (*dataref_groups)[i];
else
dr_info->group = gimple_bb (DR_STMT (datarefs[i]))->index;
datarefs_copy.quick_push (dr_info);
}
datarefs_copy.qsort (dr_group_sort_cmp);
hash_set<stmt_vec_info> to_fixup;
/* Build the interleaving chains. */
for (i = 0; i < datarefs_copy.length () - 1;)
{
dr_vec_info *dr_info_a = datarefs_copy[i];
data_reference_p dra = dr_info_a->dr;
int dra_group_id = dr_info_a->group;
stmt_vec_info stmtinfo_a = dr_info_a->stmt;
stmt_vec_info lastinfo = NULL;
if (!STMT_VINFO_VECTORIZABLE (stmtinfo_a)
|| STMT_VINFO_GATHER_SCATTER_P (stmtinfo_a))
{
++i;
continue;
}
for (i = i + 1; i < datarefs_copy.length (); ++i)
{
dr_vec_info *dr_info_b = datarefs_copy[i];
data_reference_p drb = dr_info_b->dr;
int drb_group_id = dr_info_b->group;
stmt_vec_info stmtinfo_b = dr_info_b->stmt;
if (!STMT_VINFO_VECTORIZABLE (stmtinfo_b)
|| STMT_VINFO_GATHER_SCATTER_P (stmtinfo_b))
break;
/* ??? Imperfect sorting (non-compatible types, non-modulo
accesses, same accesses) can lead to a group to be artificially
split here as we don't just skip over those. If it really
matters we can push those to a worklist and re-iterate
over them. The we can just skip ahead to the next DR here. */
/* DRs in a different DR group should not be put into the same
interleaving group. */
if (dra_group_id != drb_group_id)
break;
/* Check that the data-refs have same first location (except init)
and they are both either store or load (not load and store,
not masked loads or stores). */
if (DR_IS_READ (dra) != DR_IS_READ (drb)
|| data_ref_compare_tree (DR_BASE_ADDRESS (dra),
DR_BASE_ADDRESS (drb)) != 0
|| data_ref_compare_tree (DR_OFFSET (dra), DR_OFFSET (drb)) != 0
|| !can_group_stmts_p (stmtinfo_a, stmtinfo_b, true))
break;
/* Check that the data-refs have the same constant size. */
tree sza = TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dra)));
tree szb = TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (drb)));
if (!tree_fits_uhwi_p (sza)
|| !tree_fits_uhwi_p (szb)
|| !tree_int_cst_equal (sza, szb))
break;
/* Check that the data-refs have the same step. */
if (data_ref_compare_tree (DR_STEP (dra), DR_STEP (drb)) != 0)
break;
/* Check the types are compatible.
??? We don't distinguish this during sorting. */
if (!types_compatible_p (TREE_TYPE (DR_REF (dra)),
TREE_TYPE (DR_REF (drb))))
break;
/* Check that the DR_INITs are compile-time constants. */
if (!tree_fits_shwi_p (DR_INIT (dra))
|| !tree_fits_shwi_p (DR_INIT (drb)))
break;
/* Different .GOMP_SIMD_LANE calls still give the same lane,
just hold extra information. */
if (STMT_VINFO_SIMD_LANE_ACCESS_P (stmtinfo_a)
&& STMT_VINFO_SIMD_LANE_ACCESS_P (stmtinfo_b)
&& data_ref_compare_tree (DR_INIT (dra), DR_INIT (drb)) == 0)
break;
/* Sorting has ensured that DR_INIT (dra) <= DR_INIT (drb). */
HOST_WIDE_INT init_a = TREE_INT_CST_LOW (DR_INIT (dra));
HOST_WIDE_INT init_b = TREE_INT_CST_LOW (DR_INIT (drb));
HOST_WIDE_INT init_prev