Google Git
Sign in
gnu / gcc / refs/heads/releases/gcc-4.8 / . / gcc / tree-vect-data-refs.c
blob: 1ff0626da5bb4c1234cf55e1f7a8e7f0e7e25b00 [file] [log] [blame]
/* Data References Analysis and Manipulation Utilities for Vectorization.
Copyright (C) 2003-2013 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 "dumpfile.h"
#include "tm.h"
#include "ggc.h"
#include "tree.h"
#include "tm_p.h"
#include "target.h"
#include "basic-block.h"
#include "gimple-pretty-print.h"
#include "tree-flow.h"
#include "dumpfile.h"
#include "cfgloop.h"
#include "tree-chrec.h"
#include "tree-scalar-evolution.h"
#include "tree-vectorizer.h"
#include "diagnostic-core.h"
/* Need to include rtl.h, expr.h, etc. for optabs. */
#include "expr.h"
#include "optabs.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)
{
enum machine_mode mode, array_mode;
bool limit_p;
mode = TYPE_MODE (vectype);
limit_p = !targetm.array_mode_supported_p (mode, count);
array_mode = mode_for_size (count * GET_MODE_BITSIZE (mode),
MODE_INT, limit_p);
if (array_mode == BLKmode)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"no array mode for %s[" HOST_WIDE_INT_PRINT_DEC "]",
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>", 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>", name, GET_MODE_NAME (array_mode),
GET_MODE_NAME (mode));
return true;
}
/* Return the smallest scalar part of STMT.
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 (gimple stmt, HOST_WIDE_INT *lhs_size_unit,
HOST_WIDE_INT *rhs_size_unit)
{
tree scalar_type = gimple_expr_type (stmt);
HOST_WIDE_INT lhs, rhs;
lhs = rhs = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (scalar_type));
if (is_gimple_assign (stmt)
&& (gimple_assign_cast_p (stmt)
|| gimple_assign_rhs_code (stmt) == WIDEN_MULT_EXPR
|| gimple_assign_rhs_code (stmt) == WIDEN_LSHIFT_EXPR
|| gimple_assign_rhs_code (stmt) == FLOAT_EXPR))
{
tree rhs_type = TREE_TYPE (gimple_assign_rhs1 (stmt));
rhs = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (rhs_type));
if (rhs < lhs)
scalar_type = rhs_type;
}
*lhs_size_unit = lhs;
*rhs_size_unit = rhs;
return scalar_type;
}
/* Find the place of the data-ref in STMT in the interleaving chain that starts
from FIRST_STMT. Return -1 if the data-ref is not a part of the chain. */
int
vect_get_place_in_interleaving_chain (gimple stmt, gimple first_stmt)
{
gimple next_stmt = first_stmt;
int result = 0;
if (first_stmt != GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)))
return -1;
while (next_stmt && next_stmt != stmt)
{
result++;
next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt));
}
if (next_stmt)
return result;
else
return -1;
}
/* Function vect_insert_into_interleaving_chain.
Insert DRA into the interleaving chain of DRB according to DRA's INIT. */
static void
vect_insert_into_interleaving_chain (struct data_reference *dra,
struct data_reference *drb)
{
gimple prev, next;
tree next_init;
stmt_vec_info stmtinfo_a = vinfo_for_stmt (DR_STMT (dra));
stmt_vec_info stmtinfo_b = vinfo_for_stmt (DR_STMT (drb));
prev = GROUP_FIRST_ELEMENT (stmtinfo_b);
next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (prev));
while (next)
{
next_init = DR_INIT (STMT_VINFO_DATA_REF (vinfo_for_stmt (next)));
if (tree_int_cst_compare (next_init, DR_INIT (dra)) > 0)
{
/* Insert here. */
GROUP_NEXT_ELEMENT (vinfo_for_stmt (prev)) = DR_STMT (dra);
GROUP_NEXT_ELEMENT (stmtinfo_a) = next;
return;
}
prev = next;
next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (prev));
}
/* We got to the end of the list. Insert here. */
GROUP_NEXT_ELEMENT (vinfo_for_stmt (prev)) = DR_STMT (dra);
GROUP_NEXT_ELEMENT (stmtinfo_a) = NULL;
}
/* Function vect_update_interleaving_chain.
For two data-refs DRA and DRB that are a part of a chain interleaved data
accesses, update the interleaving chain. DRB's INIT is smaller than DRA's.
There are four possible cases:
1. New stmts - both DRA and DRB are not a part of any chain:
FIRST_DR = DRB
NEXT_DR (DRB) = DRA
2. DRB is a part of a chain and DRA is not:
no need to update FIRST_DR
no need to insert DRB
insert DRA according to init
3. DRA is a part of a chain and DRB is not:
if (init of FIRST_DR > init of DRB)
FIRST_DR = DRB
NEXT(FIRST_DR) = previous FIRST_DR
else
insert DRB according to its init
4. both DRA and DRB are in some interleaving chains:
choose the chain with the smallest init of FIRST_DR
insert the nodes of the second chain into the first one. */
static void
vect_update_interleaving_chain (struct data_reference *drb,
struct data_reference *dra)
{
stmt_vec_info stmtinfo_a = vinfo_for_stmt (DR_STMT (dra));
stmt_vec_info stmtinfo_b = vinfo_for_stmt (DR_STMT (drb));
tree next_init, init_dra_chain, init_drb_chain;
gimple first_a, first_b;
tree node_init;
gimple node, prev, next, first_stmt;
/* 1. New stmts - both DRA and DRB are not a part of any chain. */
if (!GROUP_FIRST_ELEMENT (stmtinfo_a) && !GROUP_FIRST_ELEMENT (stmtinfo_b))
{
GROUP_FIRST_ELEMENT (stmtinfo_a) = DR_STMT (drb);
GROUP_FIRST_ELEMENT (stmtinfo_b) = DR_STMT (drb);
GROUP_NEXT_ELEMENT (stmtinfo_b) = DR_STMT (dra);
return;
}
/* 2. DRB is a part of a chain and DRA is not. */
if (!GROUP_FIRST_ELEMENT (stmtinfo_a) && GROUP_FIRST_ELEMENT (stmtinfo_b))
{
GROUP_FIRST_ELEMENT (stmtinfo_a) = GROUP_FIRST_ELEMENT (stmtinfo_b);
/* Insert DRA into the chain of DRB. */
vect_insert_into_interleaving_chain (dra, drb);
return;
}
/* 3. DRA is a part of a chain and DRB is not. */
if (GROUP_FIRST_ELEMENT (stmtinfo_a) && !GROUP_FIRST_ELEMENT (stmtinfo_b))
{
gimple old_first_stmt = GROUP_FIRST_ELEMENT (stmtinfo_a);
tree init_old = DR_INIT (STMT_VINFO_DATA_REF (vinfo_for_stmt (
old_first_stmt)));
gimple tmp;
if (tree_int_cst_compare (init_old, DR_INIT (drb)) > 0)
{
/* DRB's init is smaller than the init of the stmt previously marked
as the first stmt of the interleaving chain of DRA. Therefore, we
update FIRST_STMT and put DRB in the head of the list. */
GROUP_FIRST_ELEMENT (stmtinfo_b) = DR_STMT (drb);
GROUP_NEXT_ELEMENT (stmtinfo_b) = old_first_stmt;
/* Update all the stmts in the list to point to the new FIRST_STMT. */
tmp = old_first_stmt;
while (tmp)
{
GROUP_FIRST_ELEMENT (vinfo_for_stmt (tmp)) = DR_STMT (drb);
tmp = GROUP_NEXT_ELEMENT (vinfo_for_stmt (tmp));
}
}
else
{
/* Insert DRB in the list of DRA. */
vect_insert_into_interleaving_chain (drb, dra);
GROUP_FIRST_ELEMENT (stmtinfo_b) = GROUP_FIRST_ELEMENT (stmtinfo_a);
}
return;
}
/* 4. both DRA and DRB are in some interleaving chains. */
first_a = GROUP_FIRST_ELEMENT (stmtinfo_a);
first_b = GROUP_FIRST_ELEMENT (stmtinfo_b);
if (first_a == first_b)
return;
init_dra_chain = DR_INIT (STMT_VINFO_DATA_REF (vinfo_for_stmt (first_a)));
init_drb_chain = DR_INIT (STMT_VINFO_DATA_REF (vinfo_for_stmt (first_b)));
if (tree_int_cst_compare (init_dra_chain, init_drb_chain) > 0)
{
/* Insert the nodes of DRA chain into the DRB chain.
After inserting a node, continue from this node of the DRB chain (don't
start from the beginning. */
node = GROUP_FIRST_ELEMENT (stmtinfo_a);
prev = GROUP_FIRST_ELEMENT (stmtinfo_b);
first_stmt = first_b;
}
else
{
/* Insert the nodes of DRB chain into the DRA chain.
After inserting a node, continue from this node of the DRA chain (don't
start from the beginning. */
node = GROUP_FIRST_ELEMENT (stmtinfo_b);
prev = GROUP_FIRST_ELEMENT (stmtinfo_a);
first_stmt = first_a;
}
while (node)
{
node_init = DR_INIT (STMT_VINFO_DATA_REF (vinfo_for_stmt (node)));
next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (prev));
while (next)
{
next_init = DR_INIT (STMT_VINFO_DATA_REF (vinfo_for_stmt (next)));
if (tree_int_cst_compare (next_init, node_init) > 0)
{
/* Insert here. */
GROUP_NEXT_ELEMENT (vinfo_for_stmt (prev)) = node;
GROUP_NEXT_ELEMENT (vinfo_for_stmt (node)) = next;
prev = node;
break;
}
prev = next;
next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (prev));
}
if (!next)
{
/* We got to the end of the list. Insert here. */
GROUP_NEXT_ELEMENT (vinfo_for_stmt (prev)) = node;
GROUP_NEXT_ELEMENT (vinfo_for_stmt (node)) = NULL;
prev = node;
}
GROUP_FIRST_ELEMENT (vinfo_for_stmt (node)) = first_stmt;
node = GROUP_NEXT_ELEMENT (vinfo_for_stmt (node));
}
}
/* Check dependence between DRA and DRB for basic block vectorization.
If the accesses share same bases and offsets, we can compare their initial
constant offsets to decide whether they differ or not. In case of a read-
write dependence we check that the load is before the store to ensure that
vectorization will not change the order of the accesses. */
static bool
vect_drs_dependent_in_basic_block (struct data_reference *dra,
struct data_reference *drb)
{
HOST_WIDE_INT type_size_a, type_size_b, init_a, init_b;
gimple earlier_stmt;
/* We only call this function for pairs of loads and stores, but we verify
it here. */
if (DR_IS_READ (dra) == DR_IS_READ (drb))
{
if (DR_IS_READ (dra))
return false;
else
return true;
}
/* Check that the data-refs have same bases and offsets. If not, we can't
determine if they are dependent. */
if (!operand_equal_p (DR_BASE_ADDRESS (dra), DR_BASE_ADDRESS (drb), 0)
|| !dr_equal_offsets_p (dra, drb))
return true;
/* Check the types. */
type_size_a = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dra))));
type_size_b = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (drb))));
if (type_size_a != type_size_b
|| !types_compatible_p (TREE_TYPE (DR_REF (dra)),
TREE_TYPE (DR_REF (drb))))
return true;
init_a = TREE_INT_CST_LOW (DR_INIT (dra));
init_b = TREE_INT_CST_LOW (DR_INIT (drb));
/* Two different locations - no dependence. */
if (init_a != init_b)
return false;
/* We have a read-write dependence. Check that the load is before the store.
When we vectorize basic blocks, vector load can be only before
corresponding scalar load, and vector store can be only after its
corresponding scalar store. So the order of the acceses is preserved in
case the load is before the store. */
earlier_stmt = get_earlier_stmt (DR_STMT (dra), DR_STMT (drb));
if (DR_IS_READ (STMT_VINFO_DATA_REF (vinfo_for_stmt (earlier_stmt))))
return false;
return true;
}
/* Function vect_check_interleaving.
Check if DRA and DRB are a part of interleaving. In case they are, insert
DRA and DRB in an interleaving chain. */
static bool
vect_check_interleaving (struct data_reference *dra,
struct data_reference *drb)
{
HOST_WIDE_INT type_size_a, type_size_b, diff_mod_size, step, init_a, init_b;
/* Check that the data-refs have same first location (except init) and they
are both either store or load (not load and store). */
if (!operand_equal_p (DR_BASE_ADDRESS (dra), DR_BASE_ADDRESS (drb), 0)
|| !dr_equal_offsets_p (dra, drb)
|| !tree_int_cst_compare (DR_INIT (dra), DR_INIT (drb))
|| DR_IS_READ (dra) != DR_IS_READ (drb))
return false;
/* Check:
1. data-refs are of the same type
2. their steps are equal
3. the step (if greater than zero) is greater than the difference between
data-refs' inits. */
type_size_a = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dra))));
type_size_b = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (drb))));
if (type_size_a != type_size_b
|| tree_int_cst_compare (DR_STEP (dra), DR_STEP (drb))
|| !types_compatible_p (TREE_TYPE (DR_REF (dra)),
TREE_TYPE (DR_REF (drb))))
return false;
init_a = TREE_INT_CST_LOW (DR_INIT (dra));
init_b = TREE_INT_CST_LOW (DR_INIT (drb));
step = TREE_INT_CST_LOW (DR_STEP (dra));
if (init_a > init_b)
{
/* If init_a == init_b + the size of the type * k, we have an interleaving,
and DRB is accessed before DRA. */
diff_mod_size = (init_a - init_b) % type_size_a;
if (step && (init_a - init_b) > step)
return false;
if (diff_mod_size == 0)
{
vect_update_interleaving_chain (drb, dra);
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"Detected interleaving ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dra));
dump_printf (MSG_NOTE, " and ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (drb));
}
return true;
}
}
else
{
/* If init_b == init_a + the size of the type * k, we have an
interleaving, and DRA is accessed before DRB. */
diff_mod_size = (init_b - init_a) % type_size_a;
if (step && (init_b - init_a) > step)
return false;
if (diff_mod_size == 0)
{
vect_update_interleaving_chain (dra, drb);
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"Detected interleaving ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dra));
dump_printf (MSG_NOTE, " and ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (drb));
}
return true;
}
}
return false;
}
/* Check if data references pointed by DR_I and DR_J are same or
belong to same interleaving group. Return FALSE if drs are
different, otherwise return TRUE. */
static bool
vect_same_range_drs (data_reference_p dr_i, data_reference_p dr_j)
{
gimple stmt_i = DR_STMT (dr_i);
gimple stmt_j = DR_STMT (dr_j);
if (operand_equal_p (DR_REF (dr_i), DR_REF (dr_j), 0)
|| (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt_i))
&& GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt_j))
&& (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt_i))
== GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt_j)))))
return true;
else
return false;
}
/* If address ranges represented by DDR_I and DDR_J are equal,
return TRUE, otherwise return FALSE. */
static bool
vect_vfa_range_equal (ddr_p ddr_i, ddr_p ddr_j)
{
if ((vect_same_range_drs (DDR_A (ddr_i), DDR_A (ddr_j))
&& vect_same_range_drs (DDR_B (ddr_i), DDR_B (ddr_j)))
|| (vect_same_range_drs (DDR_A (ddr_i), DDR_B (ddr_j))
&& vect_same_range_drs (DDR_B (ddr_i), DDR_A (ddr_j))))
return true;
else
return false;
}
/* 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 bool
vect_mark_for_runtime_alias_test (ddr_p ddr, loop_vec_info loop_vinfo)
{
struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
if ((unsigned) PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIAS_CHECKS) == 0)
return false;
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"mark for run-time aliasing test between ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (DDR_A (ddr)));
dump_printf (MSG_NOTE, " and ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (DDR_B (ddr)));
}
if (optimize_loop_nest_for_size_p (loop))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"versioning not supported when optimizing for size.");
return false;
}
/* FORNOW: We don't support versioning with outer-loop vectorization. */
if (loop->inner)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"versioning not yet supported for outer-loops.");
return false;
}
/* FORNOW: We don't support creating runtime alias tests for non-constant
step. */
if (TREE_CODE (DR_STEP (DDR_A (ddr))) != INTEGER_CST
|| TREE_CODE (DR_STEP (DDR_B (ddr))) != INTEGER_CST)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"versioning not yet supported for non-constant "
"step");
return false;
}
LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo).safe_push (ddr);
return true;
}
/* Function vect_analyze_data_ref_dependence.
Return TRUE 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 FALSE. Adjust *MAX_VF according to
the data dependence. */
static bool
vect_analyze_data_ref_dependence (struct data_dependence_relation *ddr,
loop_vec_info loop_vinfo, int *max_vf)
{
unsigned int i;
struct loop *loop = NULL;
struct data_reference *dra = DDR_A (ddr);
struct data_reference *drb = DDR_B (ddr);
stmt_vec_info stmtinfo_a = vinfo_for_stmt (DR_STMT (dra));
stmt_vec_info stmtinfo_b = vinfo_for_stmt (DR_STMT (drb));
lambda_vector dist_v;
unsigned int loop_depth;
/* Don't bother to analyze statements marked as unvectorizable. */
if (!STMT_VINFO_VECTORIZABLE (stmtinfo_a)
|| !STMT_VINFO_VECTORIZABLE (stmtinfo_b))
return false;
if (DDR_ARE_DEPENDENT (ddr) == chrec_known)
{
/* Independent data accesses. */
vect_check_interleaving (dra, drb);
return false;
}
if (loop_vinfo)
loop = LOOP_VINFO_LOOP (loop_vinfo);
if ((DR_IS_READ (dra) && DR_IS_READ (drb) && loop_vinfo) || dra == drb)
return false;
if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know)
{
gimple earlier_stmt;
if (loop_vinfo)
{
if (dump_enabled_p ())
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"versioning for alias required: "
"can't determine dependence between ");
dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM,
DR_REF (dra));
dump_printf (MSG_MISSED_OPTIMIZATION, " and ");
dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM,
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);
}
/* When vectorizing a basic block unknown depnedence can still mean
grouped access. */
if (vect_check_interleaving (dra, drb))
return false;
/* Read-read is OK (we need this check here, after checking for
interleaving). */
if (DR_IS_READ (dra) && DR_IS_READ (drb))
return false;
if (dump_enabled_p ())
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"can't determine dependence between ");
dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, DR_REF (dra));
dump_printf (MSG_MISSED_OPTIMIZATION, " and ");
dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, DR_REF (drb));
}
/* We do not vectorize basic blocks with write-write dependencies. */
if (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))
return true;
/* Check that it's not a load-after-store dependence. */
earlier_stmt = get_earlier_stmt (DR_STMT (dra), DR_STMT (drb));
if (DR_IS_WRITE (STMT_VINFO_DATA_REF (vinfo_for_stmt (earlier_stmt))))
return true;
return false;
}
/* Versioning for alias is not yet supported for basic block SLP, and
dependence distance is unapplicable, hence, in case of known data
dependence, basic block vectorization is impossible for now. */
if (!loop_vinfo)
{
if (dra != drb && vect_check_interleaving (dra, drb))
return false;
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"determined dependence between ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dra));
dump_printf (MSG_NOTE, " and ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (drb));
}
/* Do not vectorize basic blcoks with write-write dependences. */
if (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))
return true;
/* Check if this dependence is allowed in basic block vectorization. */
return vect_drs_dependent_in_basic_block (dra, drb);
}
/* Loop-based vectorization and known data dependence. */
if (DDR_NUM_DIST_VECTS (ddr) == 0)
{
if (dump_enabled_p ())
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"versioning for alias required: "
"bad dist vector for ");
dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, DR_REF (dra));
dump_printf (MSG_MISSED_OPTIMIZATION, " and ");
dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, 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));
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.", dist);
if (dist == 0)
{
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"dependence distance == 0 between ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dra));
dump_printf (MSG_NOTE, " and ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (drb));
}
/* For interleaving, mark that there is a read-write dependency if
necessary. We check before that one of the data-refs is store. */
if (DR_IS_READ (dra))
GROUP_READ_WRITE_DEPENDENCE (stmtinfo_a) = true;
else
{
if (DR_IS_READ (drb))
GROUP_READ_WRITE_DEPENDENCE (stmtinfo_b) = true;
}
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_MISSED_OPTIMIZATION, vect_location,
"dependence distance negative.");
/* 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;
}
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",
*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.");
continue;
}
if (dump_enabled_p ())
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not vectorized, possible dependence "
"between data-refs ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dra));
dump_printf (MSG_NOTE, " and ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (drb));
}
return true;
}
return false;
}
/* 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. */
bool
vect_analyze_data_ref_dependences (loop_vec_info loop_vinfo,
bb_vec_info bb_vinfo, int *max_vf)
{
unsigned int i;
vec<ddr_p> ddrs = vNULL;
struct data_dependence_relation *ddr;
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"=== vect_analyze_dependences ===");
if (loop_vinfo)
ddrs = LOOP_VINFO_DDRS (loop_vinfo);
else
ddrs = BB_VINFO_DDRS (bb_vinfo);
FOR_EACH_VEC_ELT (ddrs, i, ddr)
if (vect_analyze_data_ref_dependence (ddr, loop_vinfo, max_vf))
return false;
return true;
}
/* Function vect_compute_data_ref_alignment
Compute the misalignment of the data reference DR.
Output:
1. If during the misalignment computation it is found that the data reference
cannot be vectorized then false is returned.
2. DR_MISALIGNMENT (DR) is defined.
FOR NOW: No analysis is actually performed. Misalignment is calculated
only for trivial cases. TODO. */
static bool
vect_compute_data_ref_alignment (struct data_reference *dr)
{
gimple stmt = DR_STMT (dr);
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
struct loop *loop = NULL;
tree ref = DR_REF (dr);
tree vectype;
tree base, base_addr;
bool base_aligned;
tree misalign;
tree aligned_to, alignment;
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"vect_compute_data_ref_alignment:");
if (loop_vinfo)
loop = LOOP_VINFO_LOOP (loop_vinfo);
/* Initialize misalignment to unknown. */
SET_DR_MISALIGNMENT (dr, -1);
/* Strided loads perform only component accesses, misalignment information
is irrelevant for them. */
if (STMT_VINFO_STRIDE_LOAD_P (stmt_info))
return true;
misalign = DR_INIT (dr);
aligned_to = DR_ALIGNED_TO (dr);
base_addr = DR_BASE_ADDRESS (dr);
vectype = STMT_VINFO_VECTYPE (stmt_info);
/* 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 size. */
if (loop && nested_in_vect_loop_p (loop, stmt))
{
tree step = DR_STEP (dr);
HOST_WIDE_INT dr_step = TREE_INT_CST_LOW (step);
if (dr_step % GET_MODE_SIZE (TYPE_MODE (vectype)) == 0)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"inner step divides the vector-size.");
misalign = STMT_VINFO_DR_INIT (stmt_info);
aligned_to = STMT_VINFO_DR_ALIGNED_TO (stmt_info);
base_addr = STMT_VINFO_DR_BASE_ADDRESS (stmt_info);
}
else
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"inner step doesn't divide the vector-size.");
misalign = NULL_TREE;
}
}
/* Similarly, if we're doing basic-block vectorization, 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 vector size. */
if (!loop)
{
tree step = DR_STEP (dr);
HOST_WIDE_INT dr_step = TREE_INT_CST_LOW (step);
if (dr_step % GET_MODE_SIZE (TYPE_MODE (vectype)) != 0)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"SLP: step doesn't divide the vector-size.");
misalign = NULL_TREE;
}
}
base = build_fold_indirect_ref (base_addr);
alignment = ssize_int (TYPE_ALIGN (vectype)/BITS_PER_UNIT);
if ((aligned_to && tree_int_cst_compare (aligned_to, alignment) < 0)
|| !misalign)
{
if (dump_enabled_p ())
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Unknown alignment for access: ");
dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, base);
}
return true;
}
if ((DECL_P (base)
&& tree_int_cst_compare (ssize_int (DECL_ALIGN_UNIT (base)),
alignment) >= 0)
|| (TREE_CODE (base_addr) == SSA_NAME
&& tree_int_cst_compare (ssize_int (TYPE_ALIGN_UNIT (TREE_TYPE (
TREE_TYPE (base_addr)))),
alignment) >= 0)
|| (get_pointer_alignment (base_addr) >= TYPE_ALIGN (vectype)))
base_aligned = true;
else
base_aligned = false;
if (!base_aligned)
{
/* Do not change the alignment of global variables here if
flag_section_anchors is enabled as we already generated
RTL for other functions. Most global variables should
have been aligned during the IPA increase_alignment pass. */
if (!vect_can_force_dr_alignment_p (base, TYPE_ALIGN (vectype))
|| (TREE_STATIC (base) && flag_section_anchors))
{
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"can't force alignment of ref: ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, ref);
}
return true;
}
/* 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 ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, ref);
}
DECL_ALIGN (base) = TYPE_ALIGN (vectype);
DECL_USER_ALIGN (base) = 1;
}
/* At this point we assume that the base is aligned. */
gcc_assert (base_aligned
|| (TREE_CODE (base) == VAR_DECL
&& DECL_ALIGN (base) >= TYPE_ALIGN (vectype)));
/* If this is a backward running DR then first access in the larger
vectype actually is N-1 elements before the address in the DR.
Adjust misalign accordingly. */
if (tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0)
{
tree offset = ssize_int (TYPE_VECTOR_SUBPARTS (vectype) - 1);
/* DR_STEP(dr) is the same as -TYPE_SIZE of the scalar type,
otherwise we wouldn't be here. */
offset = fold_build2 (MULT_EXPR, ssizetype, offset, DR_STEP (dr));
/* PLUS because DR_STEP was negative. */
misalign = size_binop (PLUS_EXPR, misalign, offset);
}
/* Modulo alignment. */
misalign = size_binop (FLOOR_MOD_EXPR, misalign, alignment);
if (!host_integerp (misalign, 1))
{
/* Negative or overflowed misalignment value. */
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"unexpected misalign value");
return false;
}
SET_DR_MISALIGNMENT (dr, TREE_INT_CST_LOW (misalign));
if (dump_enabled_p ())
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"misalign = %d bytes of ref ", DR_MISALIGNMENT (dr));
dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, ref);
}
return true;
}
/* Function vect_compute_data_refs_alignment
Compute the misalignment of data references in the loop.
Return FALSE if a data reference is found that cannot be vectorized. */
static bool
vect_compute_data_refs_alignment (loop_vec_info loop_vinfo,
bb_vec_info bb_vinfo)
{
vec<data_reference_p> datarefs;
struct data_reference *dr;
unsigned int i;
if (loop_vinfo)
datarefs = LOOP_VINFO_DATAREFS (loop_vinfo);
else
datarefs = BB_VINFO_DATAREFS (bb_vinfo);
FOR_EACH_VEC_ELT (datarefs, i, dr)
if (STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (DR_STMT (dr)))
&& !vect_compute_data_ref_alignment (dr))
{
if (bb_vinfo)
{
/* Mark unsupported statement as unvectorizable. */
STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (DR_STMT (dr))) = false;
continue;
}
else
return false;
}
return true;
}
/* Function vect_update_misalignment_for_peel
DR - the data reference whose misalignment is to be adjusted.
DR_PEEL - 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 is known at compile time. */
static void
vect_update_misalignment_for_peel (struct data_reference *dr,
struct data_reference *dr_peel, int npeel)
{
unsigned int i;
vec<dr_p> same_align_drs;
struct data_reference *current_dr;
int dr_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dr))));
int dr_peel_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dr_peel))));
stmt_vec_info stmt_info = vinfo_for_stmt (DR_STMT (dr));
stmt_vec_info peel_stmt_info = vinfo_for_stmt (DR_STMT (dr_peel));
/* For interleaved data accesses the step in the loop must be multiplied by
the size of the interleaving group. */
if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
dr_size *= GROUP_SIZE (vinfo_for_stmt (GROUP_FIRST_ELEMENT (stmt_info)));
if (STMT_VINFO_GROUPED_ACCESS (peel_stmt_info))
dr_peel_size *= GROUP_SIZE (peel_stmt_info);
/* It can be assumed that the data refs with the same alignment as dr_peel
are aligned in the vector loop. */
same_align_drs
= STMT_VINFO_SAME_ALIGN_REFS (vinfo_for_stmt (DR_STMT (dr_peel)));
FOR_EACH_VEC_ELT (same_align_drs, i, current_dr)
{
if (current_dr != dr)
continue;
gcc_assert (DR_MISALIGNMENT (dr) / dr_size ==
DR_MISALIGNMENT (dr_peel) / dr_peel_size);
SET_DR_MISALIGNMENT (dr, 0);
return;
}
if (known_alignment_for_access_p (dr)
&& known_alignment_for_access_p (dr_peel))
{
bool negative = tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0;
int misal = DR_MISALIGNMENT (dr);
tree vectype = STMT_VINFO_VECTYPE (stmt_info);
misal += negative ? -npeel * dr_size : npeel * dr_size;
misal &= (TYPE_ALIGN (vectype) / BITS_PER_UNIT) - 1;
SET_DR_MISALIGNMENT (dr, misal);
return;
}
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location, "Setting misalignment to -1.");
SET_DR_MISALIGNMENT (dr, -1);
}
/* Function vect_verify_datarefs_alignment
Return TRUE if all data references in the loop can be
handled with respect to alignment. */
bool
vect_verify_datarefs_alignment (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo)
{
vec<data_reference_p> datarefs;
struct data_reference *dr;
enum dr_alignment_support supportable_dr_alignment;
unsigned int i;
if (loop_vinfo)
datarefs = LOOP_VINFO_DATAREFS (loop_vinfo);
else
datarefs = BB_VINFO_DATAREFS (bb_vinfo);
FOR_EACH_VEC_ELT (datarefs, i, dr)
{
gimple stmt = DR_STMT (dr);
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
if (!STMT_VINFO_RELEVANT_P (stmt_info))
continue;
/* For interleaving, only the alignment of the first access matters.
Skip statements marked as not vectorizable. */
if ((STMT_VINFO_GROUPED_ACCESS (stmt_info)
&& GROUP_FIRST_ELEMENT (stmt_info) != stmt)
|| !STMT_VINFO_VECTORIZABLE (stmt_info))
continue;
/* Strided loads perform only component accesses, alignment is
irrelevant for them. */
if (STMT_VINFO_STRIDE_LOAD_P (stmt_info))
continue;
supportable_dr_alignment = vect_supportable_dr_alignment (dr, false);
if (!supportable_dr_alignment)
{
if (dump_enabled_p ())
{
if (DR_IS_READ (dr))
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not vectorized: unsupported unaligned load.");
else
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not vectorized: unsupported unaligned "
"store.");
dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM,
DR_REF (dr));
}
return false;
}
if (supportable_dr_alignment != dr_aligned && dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Vectorizing an unaligned access.");
}
return true;
}
/* Given an memory reference EXP return whether its alignment is less
than its size. */
static bool
not_size_aligned (tree exp)
{
if (!host_integerp (TYPE_SIZE (TREE_TYPE (exp)), 1))
return true;
return (TREE_INT_CST_LOW (TYPE_SIZE (TREE_TYPE (exp)))
> get_object_alignment (exp));
}
/* Function vector_alignment_reachable_p
Return true if vector alignment for DR is reachable by peeling
a few loop iterations. Return false otherwise. */
static bool
vector_alignment_reachable_p (struct data_reference *dr)
{
gimple stmt = DR_STMT (dr);
stmt_vec_info stmt_info = vinfo_for_stmt (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;
int nelements = TYPE_VECTOR_SUBPARTS (vectype);
/* FORNOW: handle only known alignment. */
if (!known_alignment_for_access_p (dr))
return false;
elem_size = GET_MODE_SIZE (TYPE_MODE (vectype)) / nelements;
mis_in_elements = DR_MISALIGNMENT (dr) / elem_size;
if ((nelements - mis_in_elements) % 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) && !aligned_access_p (dr))
{
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 =" HOST_WIDE_INT_PRINT_DEC, elmsize);
dump_printf (MSG_NOTE,
". misalignment = %d. ", DR_MISALIGNMENT (dr));
}
if (DR_MISALIGNMENT (dr) % 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))
{
tree type = TREE_TYPE (DR_REF (dr));
bool is_packed = not_size_aligned (DR_REF (dr));
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Unknown misalignment, is_packed = %d",is_packed);
if (targetm.vectorize.vector_alignment_reachable (type, is_packed))
return true;
else
return false;
}
return true;
}
/* Calculate the cost of the memory access represented by DR. */
static void
vect_get_data_access_cost (struct data_reference *dr,
unsigned int *inside_cost,
unsigned int *outside_cost,
stmt_vector_for_cost *body_cost_vec)
{
gimple stmt = DR_STMT (dr);
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
int nunits = TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info));
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
int ncopies = vf / nunits;
if (DR_IS_READ (dr))
vect_get_load_cost (dr, ncopies, true, inside_cost, outside_cost,
NULL, body_cost_vec, false);
else
vect_get_store_cost (dr, ncopies, 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.", *inside_cost, *outside_cost);
}
static hashval_t
vect_peeling_hash (const void *elem)
{
const struct _vect_peel_info *peel_info;
peel_info = (const struct _vect_peel_info *) elem;
return (hashval_t) peel_info->npeel;
}
static int
vect_peeling_hash_eq (const void *elem1, const void *elem2)
{
const struct _vect_peel_info *a, *b;
a = (const struct _vect_peel_info *) elem1;
b = (const struct _vect_peel_info *) elem2;
return (a->npeel == b->npeel);
}
/* Insert DR into peeling hash table with NPEEL as key. */
static void
vect_peeling_hash_insert (loop_vec_info loop_vinfo, struct data_reference *dr,
int npeel)
{
struct _vect_peel_info elem, *slot;
void **new_slot;
bool supportable_dr_alignment = vect_supportable_dr_alignment (dr, true);
elem.npeel = npeel;
slot = (vect_peel_info) htab_find (LOOP_VINFO_PEELING_HTAB (loop_vinfo),
&elem);
if (slot)
slot->count++;
else
{
slot = XNEW (struct _vect_peel_info);
slot->npeel = npeel;
slot->dr = dr;
slot->count = 1;
new_slot = htab_find_slot (LOOP_VINFO_PEELING_HTAB (loop_vinfo), slot,
INSERT);
*new_slot = slot;
}
if (!supportable_dr_alignment && !flag_vect_cost_model)
slot->count += VECT_MAX_COST;
}
/* Traverse peeling hash table to find peeling option that aligns maximum
number of data accesses. */
static int
vect_peeling_hash_get_most_frequent (void **slot, void *data)
{
vect_peel_info elem = (vect_peel_info) *slot;
vect_peel_extended_info max = (vect_peel_extended_info) data;
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 = elem->dr;
}
return 1;
}
/* Traverse peeling hash table and calculate cost for each peeling option.
Find the one with the lowest cost. */
static int
vect_peeling_hash_get_lowest_cost (void **slot, void *data)
{
vect_peel_info elem = (vect_peel_info) *slot;
vect_peel_extended_info min = (vect_peel_extended_info) data;
int save_misalignment, dummy;
unsigned int inside_cost = 0, outside_cost = 0, i;
gimple stmt = DR_STMT (elem->dr);
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
vec<data_reference_p> datarefs = LOOP_VINFO_DATAREFS (loop_vinfo);
struct data_reference *dr;
stmt_vector_for_cost prologue_cost_vec, body_cost_vec, epilogue_cost_vec;
int single_iter_cost;
prologue_cost_vec.create (2);
body_cost_vec.create (2);
epilogue_cost_vec.create (2);
FOR_EACH_VEC_ELT (datarefs, i, dr)
{
stmt = DR_STMT (dr);
stmt_info = vinfo_for_stmt (stmt);
/* For interleaving, only the alignment of the first access
matters. */
if (STMT_VINFO_GROUPED_ACCESS (stmt_info)
&& GROUP_FIRST_ELEMENT (stmt_info) != stmt)
continue;
save_misalignment = DR_MISALIGNMENT (dr);
vect_update_misalignment_for_peel (dr, elem->dr, elem->npeel);
vect_get_data_access_cost (dr, &inside_cost, &outside_cost,
&body_cost_vec);
SET_DR_MISALIGNMENT (dr, save_misalignment);
}
single_iter_cost = vect_get_single_scalar_iteration_cost (loop_vinfo);
outside_cost += vect_get_known_peeling_cost (loop_vinfo, elem->npeel,
&dummy, single_iter_cost,
&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->body_cost_vec.release ();
min->body_cost_vec = body_cost_vec;
min->peel_info.dr = elem->dr;
min->peel_info.npeel = elem->npeel;
}
else
body_cost_vec.release ();
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 data_reference *
vect_peeling_hash_choose_best_peeling (loop_vec_info loop_vinfo,
unsigned int *npeel,
stmt_vector_for_cost *body_cost_vec)
{
struct _vect_peel_extended_info res;
res.peel_info.dr = NULL;
res.body_cost_vec = stmt_vector_for_cost();
if (flag_vect_cost_model)
{
res.inside_cost = INT_MAX;
res.outside_cost = INT_MAX;
htab_traverse (LOOP_VINFO_PEELING_HTAB (loop_vinfo),
vect_peeling_hash_get_lowest_cost, &res);
}
else
{
res.peel_info.count = 0;
htab_traverse (LOOP_VINFO_PEELING_HTAB (loop_vinfo),
vect_peeling_hash_get_most_frequent, &res);
}
*npeel = res.peel_info.npeel;
*body_cost_vec = res.body_cost_vec;
return res.peel_info.dr;
}
/* 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). */
bool
vect_enhance_data_refs_alignment (loop_vec_info loop_vinfo)
{
vec<data_reference_p> datarefs = LOOP_VINFO_DATAREFS (loop_vinfo);
struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
enum dr_alignment_support supportable_dr_alignment;
struct data_reference *dr0 = NULL, *first_store = NULL;
struct data_reference *dr;
unsigned int i, j;
bool do_peeling = false;
bool do_versioning = false;
bool stat;
gimple stmt;
stmt_vec_info stmt_info;
int vect_versioning_for_alias_required;
unsigned int npeel = 0;
bool all_misalignments_unknown = true;
unsigned int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
unsigned possible_npeel_number = 1;
tree vectype;
unsigned int nelements, mis, same_align_drs_max = 0;
stmt_vector_for_cost body_cost_vec = stmt_vector_for_cost();
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"=== vect_enhance_data_refs_alignment ===");
/* 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)
{
stmt = DR_STMT (dr);
stmt_info = vinfo_for_stmt (stmt);
if (!STMT_VINFO_RELEVANT_P (stmt_info))
continue;
/* For interleaving, only the alignment of the first access
matters. */
if (STMT_VINFO_GROUPED_ACCESS (stmt_info)
&& GROUP_FIRST_ELEMENT (stmt_info) != stmt)
continue;
/* For invariant accesses there is nothing to enhance. */
if (integer_zerop (DR_STEP (dr)))
continue;
/* Strided loads perform only component accesses, alignment is
irrelevant for them. */
if (STMT_VINFO_STRIDE_LOAD_P (stmt_info))
continue;
supportable_dr_alignment = vect_supportable_dr_alignment (dr, true);
do_peeling = vector_alignment_reachable_p (dr);
if (do_peeling)
{
if (known_alignment_for_access_p (dr))
{
unsigned int npeel_tmp;
bool negative = tree_int_cst_compare (DR_STEP (dr),
size_zero_node) < 0;
/* Save info about DR in the hash table. */
if (!LOOP_VINFO_PEELING_HTAB (loop_vinfo))
LOOP_VINFO_PEELING_HTAB (loop_vinfo) =
htab_create (1, vect_peeling_hash,
vect_peeling_hash_eq, free);
vectype = STMT_VINFO_VECTYPE (stmt_info);
nelements = TYPE_VECTOR_SUBPARTS (vectype);
mis = DR_MISALIGNMENT (dr) / GET_MODE_SIZE (TYPE_MODE (
TREE_TYPE (DR_REF (dr))));
npeel_tmp = (negative
? (mis - nelements) : (nelements - mis))
& (nelements - 1);
/* 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
access are misaligned by 3, the first one needs one scalar
iteration to be aligned, and the second one needs 5. But the
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 automtically for cost model, since we calculate cost
for every peeling option. */
if (!flag_vect_cost_model)
possible_npeel_number = vf /nelements;
/* Handle the aligned case. We may decide to align some other
access, making DR unaligned. */
if (DR_MISALIGNMENT (dr) == 0)
{
npeel_tmp = 0;
if (!flag_vect_cost_model)
possible_npeel_number++;
}
for (j = 0; j < possible_npeel_number; j++)
{
gcc_assert (npeel_tmp <= vf);
vect_peeling_hash_insert (loop_vinfo, dr, npeel_tmp);
npeel_tmp += nelements;
}
all_misalignments_unknown = false;
/* Data-ref that was chosen for the case that all the
misalignments are unknown is not relevant anymore, since we
have a data-ref with known alignment. */
dr0 = NULL;
}
else
{
/* If we don't know all the misalignment values, we prefer
peeling for data-ref that has maximum number of data-refs
with the same alignment, unless the target prefers to align
stores over load. */
if (all_misalignments_unknown)
{
if (same_align_drs_max
< STMT_VINFO_SAME_ALIGN_REFS (stmt_info).length ()
|| !dr0)
{
same_align_drs_max
= STMT_VINFO_SAME_ALIGN_REFS (stmt_info).length ();
dr0 = dr;
}
if (!first_store && DR_IS_WRITE (dr))
first_store = dr;
}
/* If there are both known and unknown misaligned accesses in the
loop, we choose peeling amount according to the known
accesses. */
if (!supportable_dr_alignment)
{
dr0 = dr;
if (!first_store && DR_IS_WRITE (dr))
first_store = dr;
}
}
}
else
{
if (!aligned_access_p (dr))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"vector alignment may not be reachable");
break;
}
}
}
vect_versioning_for_alias_required
= LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo);
/* Temporarily, if versioning for alias is required, we disable peeling
until we support peeling and versioning. Often peeling for alignment
will require peeling for loop-bound, which in turn requires that we
know how to adjust the loop ivs after the loop. */
if (vect_versioning_for_alias_required
|| !vect_can_advance_ivs_p (loop_vinfo)
|| !slpeel_can_duplicate_loop_p (loop, single_exit (loop)))
do_peeling = false;
if (do_peeling && all_misalignments_unknown
&& vect_supportable_dr_alignment (dr0, false))
{
/* 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). */
if (first_store && DR_IS_READ (dr0))
{
unsigned int load_inside_cost = 0, load_outside_cost = 0;
unsigned int store_inside_cost = 0, store_outside_cost = 0;
unsigned int load_inside_penalty = 0, load_outside_penalty = 0;
unsigned int store_inside_penalty = 0, store_outside_penalty = 0;
stmt_vector_for_cost dummy;
dummy.create (2);
vect_get_data_access_cost (dr0, &load_inside_cost, &load_outside_cost,
&dummy);
vect_get_data_access_cost (first_store, &store_inside_cost,
&store_outside_cost, &dummy);
dummy.release ();
/* Calculate the penalty for leaving FIRST_STORE unaligned (by
aligning the load DR0). */
load_inside_penalty = store_inside_cost;
load_outside_penalty = store_outside_cost;
for (i = 0;
STMT_VINFO_SAME_ALIGN_REFS (vinfo_for_stmt (
DR_STMT (first_store))).iterate (i, &dr);
i++)
if (DR_IS_READ (dr))
{
load_inside_penalty += load_inside_cost;
load_outside_penalty += load_outside_cost;
}
else
{
load_inside_penalty += store_inside_cost;
load_outside_penalty += store_outside_cost;
}
/* Calculate the penalty for leaving DR0 unaligned (by
aligning the FIRST_STORE). */
store_inside_penalty = load_inside_cost;
store_outside_penalty = load_outside_cost;
for (i = 0;
STMT_VINFO_SAME_ALIGN_REFS (vinfo_for_stmt (
DR_STMT (dr0))).iterate (i, &dr);
i++)
if (DR_IS_READ (dr))
{
store_inside_penalty += load_inside_cost;
store_outside_penalty += load_outside_cost;
}
else
{
store_inside_penalty += store_inside_cost;
store_outside_penalty += store_outside_cost;
}
if (load_inside_penalty > store_inside_penalty
|| (load_inside_penalty == store_inside_penalty
&& load_outside_penalty > store_outside_penalty))
dr0 = first_store;
}
/* In case there are only loads with different unknown misalignments, use
peeling only if it may help to align other accesses in the loop. */
if (!first_store
&& !STMT_VINFO_SAME_ALIGN_REFS (
vinfo_for_stmt (DR_STMT (dr0))).length ()
&& vect_supportable_dr_alignment (dr0, false)
!= dr_unaligned_supported)
do_peeling = false;
}
if (do_peeling && !dr0)
{
/* Peeling is possible, but there is no data access that is not supported
unless aligned. So we try to choose the best possible peeling. */
/* We should get here only if there are drs with known misalignment. */
gcc_assert (!all_misalignments_unknown);
/* Choose the best peeling from the hash table. */
dr0 = vect_peeling_hash_choose_best_peeling (loop_vinfo, &npeel,
&body_cost_vec);
if (!dr0 || !npeel)
do_peeling = false;
}
if (do_peeling)
{
stmt = DR_STMT (dr0);
stmt_info = vinfo_for_stmt (stmt);
vectype = STMT_VINFO_VECTYPE (stmt_info);
nelements = TYPE_VECTOR_SUBPARTS (vectype);
if (known_alignment_for_access_p (dr0))
{
bool negative = tree_int_cst_compare (DR_STEP (dr0),
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. */
mis = DR_MISALIGNMENT (dr0);
mis /= GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dr0))));
npeel = ((negative ? mis - nelements : nelements - mis)
& (nelements - 1));
}
/* For interleaved data access every iteration accesses all the
members of the group, therefore we divide the number of iterations
by the group size. */
stmt_info = vinfo_for_stmt (DR_STMT (dr0));
if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
npeel /= GROUP_SIZE (stmt_info);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Try peeling by %d", npeel);
}
/* Ensure that all data refs can be vectorized after the peel. */
FOR_EACH_VEC_ELT (datarefs, i, dr)
{
int save_misalignment;
if (dr == dr0)
continue;
stmt = DR_STMT (dr);
stmt_info = vinfo_for_stmt (stmt);
/* For interleaving, only the alignment of the first access
matters. */
if (STMT_VINFO_GROUPED_ACCESS (stmt_info)
&& GROUP_FIRST_ELEMENT (stmt_info) != stmt)
continue;
/* Strided loads perform only component accesses, alignment is
irrelevant for them. */
if (STMT_VINFO_STRIDE_LOAD_P (stmt_info))
continue;
save_misalignment = DR_MISALIGNMENT (dr);
vect_update_misalignment_for_peel (dr, dr0, npeel);
supportable_dr_alignment = vect_supportable_dr_alignment (dr, false);
SET_DR_MISALIGNMENT (dr, save_misalignment);
if (!supportable_dr_alignment)
{
do_peeling = false;
break;
}
}
if (do_peeling && known_alignment_for_access_p (dr0) && npeel == 0)
{
stat = vect_verify_datarefs_alignment (loop_vinfo, NULL);
if (!stat)
do_peeling = false;
else
{
body_cost_vec.release ();
return stat;
}
}
if (do_peeling)
{
stmt_info_for_cost *si;
void *data = LOOP_VINFO_TARGET_COST_DATA (loop_vinfo);
/* (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)
vect_update_misalignment_for_peel (dr, dr0, npeel);
LOOP_VINFO_UNALIGNED_DR (loop_vinfo) = dr0;
if (npeel)
LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo) = npeel;
else
LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo) = DR_MISALIGNMENT (dr0);
SET_DR_MISALIGNMENT (dr0, 0);
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"Alignment of access forced using peeling.");
dump_printf_loc (MSG_NOTE, vect_location,
"Peeling for alignment will be applied.");
}
/* We've delayed passing the inside-loop peeling costs to the
target cost model until we were sure peeling would happen.
Do so now. */
if (body_cost_vec.exists ())
{
FOR_EACH_VEC_ELT (body_cost_vec, i, si)
{
struct _stmt_vec_info *stmt_info
= si->stmt ? vinfo_for_stmt (si->stmt) : NULL;
(void) add_stmt_cost (data, si->count, si->kind, stmt_info,
si->misalign, vect_body);
}
body_cost_vec.release ();
}
stat = vect_verify_datarefs_alignment (loop_vinfo, NULL);
gcc_assert (stat);
return stat;
}
}
body_cost_vec.release ();
/* (2) Versioning to force alignment. */
/* Try versioning if:
1) flag_tree_vect_loop_version is TRUE
2) optimize loop for speed
3) there is at least one unsupported misaligned data ref with an unknown
misalignment, and
4) all misaligned data refs with a known misalignment are supported, and
5) the number of runtime alignment checks is within reason. */
do_versioning =
flag_tree_vect_loop_version
&& optimize_loop_nest_for_speed_p (loop)
&& (!loop->inner); /* FORNOW */
if (do_versioning)
{
FOR_EACH_VEC_ELT (datarefs, i, dr)
{
stmt = DR_STMT (dr);
stmt_info = vinfo_for_stmt (stmt);
/* For interleaving, only the alignment of the first access
matters. */
if (aligned_access_p (dr)
|| (STMT_VINFO_GROUPED_ACCESS (stmt_info)
&& GROUP_FIRST_ELEMENT (stmt_info) != stmt))
continue;
/* Strided loads perform only component accesses, alignment is
irrelevant for them. */
if (STMT_VINFO_STRIDE_LOAD_P (stmt_info))
continue;
supportable_dr_alignment = vect_supportable_dr_alignment (dr, false);
if (!supportable_dr_alignment)
{
gimple stmt;
int mask;
tree vectype;
if (known_alignment_for_access_p (dr)
|| LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo).length ()
>= (unsigned) PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIGNMENT_CHECKS))
{
do_versioning = false;
break;
}
stmt = DR_STMT (dr);
vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt));
gcc_assert (vectype);
/* 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. */
mask = GET_MODE_SIZE (TYPE_MODE (vectype)) - 1;
/* FORNOW: use the same mask to test all potentially unaligned
references in the loop. The vectorizer currently supports
a single vector size, see the reference to
GET_MODE_NUNITS (TYPE_MODE (vectype)) where the
vectorization factor is computed. */
gcc_assert (!LOOP_VINFO_PTR_MASK (loop_vinfo)
|| LOOP_VINFO_PTR_MASK (loop_vinfo) == mask);
LOOP_VINFO_PTR_MASK (loop_vinfo) = mask;
LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo).safe_push (
DR_STMT (dr));
}
}
/* 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)
{
vec<gimple> may_misalign_stmts
= LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo);
gimple stmt;
/* 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)
{
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
dr = STMT_VINFO_DATA_REF (stmt_info);
SET_DR_MISALIGNMENT (dr, 0);
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Alignment of access forced using versioning.");
}
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Versioning for alignment will be applied.");
/* Peeling and versioning can't be done together at this time. */
gcc_assert (! (do_peeling && do_versioning));
stat = vect_verify_datarefs_alignment (loop_vinfo, NULL);
gcc_assert (stat);
return stat;
}
/* This point is reached if neither peeling nor versioning is being done. */
gcc_assert (! (do_peeling || do_versioning));
stat = vect_verify_datarefs_alignment (loop_vinfo, NULL);
return stat;
}
/* Function vect_find_same_alignment_drs.
Update group and alignment relations according to the chosen
vectorization factor. */
static void
vect_find_same_alignment_drs (struct data_dependence_relation *ddr,
loop_vec_info loop_vinfo)
{
unsigned int i;
struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
int vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
struct data_reference *dra = DDR_A (ddr);
struct data_reference *drb = DDR_B (ddr);
stmt_vec_info stmtinfo_a = vinfo_for_stmt (DR_STMT (dra));
stmt_vec_info stmtinfo_b = vinfo_for_stmt (DR_STMT (drb));
int dra_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dra))));
int drb_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (drb))));
lambda_vector dist_v;
unsigned int loop_depth;
if (DDR_ARE_DEPENDENT (ddr) == chrec_known)
return;
if (dra == drb)
return;
if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know)
return;
/* Loop-based vectorization and known data dependence. */
if (DDR_NUM_DIST_VECTS (ddr) == 0)
return;
/* Data-dependence analysis reports a distance vector of zero
for data-references that overlap only in the first iteration
but have different sign step (see PR45764).
So as a sanity check require equal DR_STEP. */
if (!operand_equal_p (DR_STEP (dra), DR_STEP (drb), 0))
return;
loop_depth = index_in_loop_nest (loop->num, DDR_LOOP_NEST (ddr));
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.", dist);
/* Same loop iteration. */
if (dist == 0
|| (dist % vectorization_factor == 0 && dra_size == drb_size))
{
/* Two references with distance zero have the same alignment. */
STMT_VINFO_SAME_ALIGN_REFS (stmtinfo_a).safe_push (drb);
STMT_VINFO_SAME_ALIGN_REFS (stmtinfo_b).safe_push (dra);
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"accesses have the same alignment.");
dump_printf (MSG_NOTE,
"dependence distance modulo vf == 0 between ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dra));
dump_printf (MSG_NOTE, " and ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (drb));
}
}
}
}
/* 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. */
bool
vect_analyze_data_refs_alignment (loop_vec_info loop_vinfo,
bb_vec_info bb_vinfo)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"=== vect_analyze_data_refs_alignment ===");
/* Mark groups of data references with same alignment using
data dependence information. */
if (loop_vinfo)
{
vec<ddr_p> ddrs = LOOP_VINFO_DDRS (loop_vinfo);
struct data_dependence_relation *ddr;
unsigned int i;
FOR_EACH_VEC_ELT (ddrs, i, ddr)
vect_find_same_alignment_drs (ddr, loop_vinfo);
}
if (!vect_compute_data_refs_alignment (loop_vinfo, bb_vinfo))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not vectorized: can't calculate alignment "
"for data ref.");
return false;
}
return true;
}
/* Analyze groups of accesses: check that DR 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 (struct data_reference *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));
gimple stmt = DR_STMT (dr);
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
HOST_WIDE_INT dr_step = TREE_INT_CST_LOW (step);
HOST_WIDE_INT groupsize, last_accessed_element = 1;
bool slp_impossible = false;
struct loop *loop = NULL;
if (loop_vinfo)
loop = LOOP_VINFO_LOOP (loop_vinfo);
/* For interleaving, GROUPSIZE is STEP counted in elements, i.e., the
size of the interleaving group (including gaps). */
groupsize = dr_step / type_size;
/* Not consecutive access is possible only if it is a part of interleaving. */
if (!GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)))
{
/* 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. The size of the group must be a power of 2. */
if (DR_IS_READ (dr)
&& (dr_step % type_size) == 0
&& groupsize > 0
&& exact_log2 (groupsize) != -1)
{
GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) = stmt;
GROUP_SIZE (vinfo_for_stmt (stmt)) = groupsize;
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"Detected single element interleaving ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dr));
dump_printf (MSG_NOTE, " step ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, step);
}
if (loop_vinfo)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Data access with gaps requires scalar "
"epilogue loop");
if (loop->inner)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Peeling for outer loop is not"
" supported");
return false;
}
LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true;
}
return true;
}
if (dump_enabled_p ())
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not consecutive access ");
dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0);
}
if (bb_vinfo)
{
/* Mark the statement as unvectorizable. */
STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (DR_STMT (dr))) = false;
return true;
}
return false;
}
if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) == stmt)
{
/* First stmt in the interleaving chain. Check the chain. */
gimple next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (stmt));
struct data_reference *data_ref = dr;
unsigned int count = 1;
tree next_step;
tree prev_init = DR_INIT (data_ref);
gimple prev = stmt;
HOST_WIDE_INT diff, count_in_bytes, gaps = 0;
while (next)
{
/* Skip same data-refs. In case that two or more stmts share
data-ref (supported only for loads), we vectorize only the first
stmt, and the rest get their vectorized loads from the first
one. */
if (!tree_int_cst_compare (DR_INIT (data_ref),
DR_INIT (STMT_VINFO_DATA_REF (
vinfo_for_stmt (next)))))
{
if (DR_IS_WRITE (data_ref))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Two store stmts share the same dr.");
return false;
}
/* Check that there is no load-store dependencies for this loads
to prevent a case of load-store-load to the same location. */
if (GROUP_READ_WRITE_DEPENDENCE (vinfo_for_stmt (next))
|| GROUP_READ_WRITE_DEPENDENCE (vinfo_for_stmt (prev)))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"READ_WRITE dependence in interleaving.");
return false;
}
/* For load use the same data-ref load. */
GROUP_SAME_DR_STMT (vinfo_for_stmt (next)) = prev;
prev = next;
next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next));
continue;
}
prev = next;
/* Check that all the accesses have the same STEP. */
next_step = DR_STEP (STMT_VINFO_DATA_REF (vinfo_for_stmt (next)));
if (tree_int_cst_compare (step, next_step))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not consecutive access in interleaving");
return false;
}
data_ref = STMT_VINFO_DATA_REF (vinfo_for_stmt (next));
/* 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)
{
/* 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");
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, GROUP_GAP is always 1. */
GROUP_GAP (vinfo_for_stmt (next)) = diff;
prev_init = DR_INIT (data_ref);
next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next));
/* Count the number of data-refs in the chain. */
count++;
}
/* COUNT is the number of accesses found, we multiply it by the size of
the type to get COUNT_IN_BYTES. */
count_in_bytes = type_size * count;
/* Check that the size of the interleaving (including gaps) is not
greater than STEP. */
if (dr_step && dr_step < count_in_bytes + gaps * type_size)
{
if (dump_enabled_p ())
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"interleaving size is greater than step for ");
dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, DR_REF (dr));
}
return false;
}
/* Check that the size of the interleaving is equal to STEP for stores,
i.e., that there are no gaps. */
if (dr_step && dr_step != count_in_bytes)
{
if (DR_IS_READ (dr))
{
slp_impossible = true;
/* There is a gap after the last load in the group. This gap is a
difference between the groupsize and the number of elements.
When there is no gap, this difference should be 0. */
GROUP_GAP (vinfo_for_stmt (stmt)) = groupsize - count;
}
else
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"interleaved store with gaps");
return false;
}
}
/* Check that STEP is a multiple of type size. */
if (dr_step && (dr_step % type_size) != 0)
{
if (dump_enabled_p ())
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"step is not a multiple of type size: step ");
dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM, step);
dump_printf (MSG_MISSED_OPTIMIZATION, " size ");
dump_generic_expr (MSG_MISSED_OPTIMIZATION, TDF_SLIM,
TYPE_SIZE_UNIT (scalar_type));
}
return false;
}
if (groupsize == 0)
groupsize = count;
GROUP_SIZE (vinfo_for_stmt (stmt)) = groupsize;
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"Detected interleaving of size %d", (int)groupsize);
/* 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);
if (bb_vinfo)
BB_VINFO_GROUPED_STORES (bb_vinfo).safe_push (stmt);
}
/* There is a gap in the end of the group. */
if (groupsize - last_accessed_element > 0 && loop_vinfo)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Data access with gaps requires scalar "
"epilogue loop");
if (loop->inner)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"Peeling for outer loop is not supported");
return false;
}
LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true;
}
}
return true;
}
/* Analyze the access pattern of the data-reference DR.
In case of non-consecutive accesses call vect_analyze_group_access() to
analyze groups of accesses. */
static bool
vect_analyze_data_ref_access (struct data_reference *dr)
{
tree step = DR_STEP (dr);
tree scalar_type = TREE_TYPE (DR_REF (dr));
gimple stmt = DR_STMT (dr);
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
struct loop *loop = NULL;
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");
return false;
}
/* Allow invariant loads in not nested loops. */
if (loop_vinfo && integer_zerop (step))
{
GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) = NULL;
if (nested_in_vect_loop_p (loop, stmt))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"zero step in inner loop of nest");
return false;
}
return DR_IS_READ (dr);
}
if (loop && nested_in_vect_loop_p (loop, stmt))
{
/* Interleaved accesses are not yet supported within outer-loop
vectorization for references in the inner-loop. */
GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) = 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.");
if (DR_IS_READ (dr))
return true;
else
return false;
}
}
/* 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. */
GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) = NULL;
return true;
}
}
if (loop && nested_in_vect_loop_p (loop, stmt))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"grouped access in outer loop.");
return false;
}
/* Assume this is a DR handled by non-constant strided load case. */
if (TREE_CODE (step) != INTEGER_CST)
return STMT_VINFO_STRIDE_LOAD_P (stmt_info);
/* Not consecutive access - check if it's a part of interleaving group. */
return vect_analyze_group_access (dr);
}
/* 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. */
bool
vect_analyze_data_ref_accesses (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo)
{
unsigned int i;
vec<data_reference_p> datarefs;
struct data_reference *dr;
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"=== vect_analyze_data_ref_accesses ===");
if (loop_vinfo)
datarefs = LOOP_VINFO_DATAREFS (loop_vinfo);
else
datarefs = BB_VINFO_DATAREFS (bb_vinfo);
FOR_EACH_VEC_ELT (datarefs, i, dr)
if (STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (DR_STMT (dr)))
&& !vect_analyze_data_ref_access (dr))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not vectorized: complicated access pattern.");
if (bb_vinfo)
{
/* Mark the statement as not vectorizable. */
STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (DR_STMT (dr))) = false;
continue;
}
else
return false;
}
return true;
}
/* Function vect_prune_runtime_alias_test_list.
Prune a list of ddrs to be tested at run-time by versioning for alias.
Return FALSE if resulting list of ddrs is longer then allowed by
PARAM_VECT_MAX_VERSION_FOR_ALIAS_CHECKS, otherwise return TRUE. */
bool
vect_prune_runtime_alias_test_list (loop_vec_info loop_vinfo)
{
vec<ddr_p> ddrs =
LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo);
unsigned i, j;
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"=== vect_prune_runtime_alias_test_list ===");
for (i = 0; i < ddrs.length (); )
{
bool found;
ddr_p ddr_i;
ddr_i = ddrs[i];
found = false;
for (j = 0; j < i; j++)
{
ddr_p ddr_j = ddrs[j];
if (vect_vfa_range_equal (ddr_i, ddr_j))
{
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"found equal ranges ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (DDR_A (ddr_i)));
dump_printf (MSG_NOTE, ", ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (DDR_B (ddr_i)));
dump_printf (MSG_NOTE, " and ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (DDR_A (ddr_j)));
dump_printf (MSG_NOTE, ", ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (DDR_B (ddr_j)));
}
found = true;
break;
}
}
if (found)
{
ddrs.ordered_remove (i);
continue;
}
i++;
}
if (ddrs.length () >
(unsigned) PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIAS_CHECKS))
{
if (dump_enabled_p ())
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"disable versioning for alias - max number of "
"generated checks exceeded.");
}
LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo).truncate (0);
return false;
}
return true;
}
/* Check whether a non-affine read in stmt is suitable for gather load
and if so, return a builtin decl for that operation. */
tree
vect_check_gather (gimple stmt, loop_vec_info loop_vinfo, tree *basep,
tree *offp, int *scalep)
{
HOST_WIDE_INT scale = 1, pbitpos, pbitsize;
struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
tree offtype = NULL_TREE;
tree decl, base, off;
enum machine_mode pmode;
int punsignedp, pvolatilep;
/* The gather builtins need address of the form
loop_invariant + vector * {1, 2, 4, 8}
or
loop_invariant + sign_extend (vector) * { 1, 2, 4, 8 }.
Unfortunately DR_BASE_ADDRESS/DR_OFFSET can be a mixture
of loop invariants/SSA_NAMEs defined in the loop, with casts,
multiplications and additions in it. To get a vector, we need
a single SSA_NAME that will be defined in the loop and will
contain everything that is not loop invariant and that can be
vectorized. The following code attempts to find such a preexistng
SSA_NAME OFF and put the loop invariants into a tree BASE
that can be gimplified before the loop. */
base = get_inner_reference (DR_REF (dr), &pbitsize, &pbitpos, &off,
&pmode, &punsignedp, &pvolatilep, false);
gcc_assert (base != NULL_TREE && (pbitpos % BITS_PER_UNIT) == 0);
if (TREE_CODE (base) == MEM_REF)
{
if (!integer_zerop (TREE_OPERAND (base, 1)))
{
if (off == NULL_TREE)
{
double_int moff = mem_ref_offset (base);
off = double_int_to_tree (sizetype, moff);
}
else
off = size_binop (PLUS_EXPR, off,
fold_convert (sizetype, TREE_OPERAND (base, 1)));
}
base = TREE_OPERAND (base, 0);
}
else
base = build_fold_addr_expr (base);
if (off == NULL_TREE)
off = size_zero_node;
/* If base is not loop invariant, either off is 0, then we start with just
the constant offset in the loop invariant BASE and continue with base
as OFF, otherwise give up.
We could handle that case by gimplifying the addition of base + off
into some SSA_NAME and use that as off, but for now punt. */
if (!expr_invariant_in_loop_p (loop, base))
{
if (!integer_zerop (off))
return NULL_TREE;
off = base;
base = size_int (pbitpos / BITS_PER_UNIT);
}
/* Otherwise put base + constant offset into the loop invariant BASE
and continue with OFF. */
else
{
base = fold_convert (sizetype, base);
base = size_binop (PLUS_EXPR, base, size_int (pbitpos / BITS_PER_UNIT));
}
/* OFF at this point may be either a SSA_NAME or some tree expression
from get_inner_reference. Try to peel off loop invariants from it
into BASE as long as possible. */
STRIP_NOPS (off);
while (offtype == NULL_TREE)
{
enum tree_code code;
tree op0, op1, add = NULL_TREE;
if (TREE_CODE (off) == SSA_NAME)
{
gimple def_stmt = SSA_NAME_DEF_STMT (off);
if (expr_invariant_in_loop_p (loop, off))
return NULL_TREE;
if (gimple_code (def_stmt) != GIMPLE_ASSIGN)
break;
op0 = gimple_assign_rhs1 (def_stmt);
code = gimple_assign_rhs_code (def_stmt);
op1 = gimple_assign_rhs2 (def_stmt);
}
else
{
if (get_gimple_rhs_class (TREE_CODE (off)) == GIMPLE_TERNARY_RHS)
return NULL_TREE;
code = TREE_CODE (off);
extract_ops_from_tree (off, &code, &op0, &op1);
}
switch (code)
{
case POINTER_PLUS_EXPR:
case PLUS_EXPR:
if (expr_invariant_in_loop_p (loop, op0))
{
add = op0;
off = op1;
do_add:
add = fold_convert (sizetype, add);
if (scale != 1)
add = size_binop (MULT_EXPR, add, size_int (scale));
base = size_binop (PLUS_EXPR, base, add);
continue;
}
if (expr_invariant_in_loop_p (loop, op1))
{
add = op1;
off = op0;
goto do_add;
}
break;
case MINUS_EXPR:
if (expr_invariant_in_loop_p (loop, op1))
{
add = fold_convert (sizetype, op1);
add = size_binop (MINUS_EXPR, size_zero_node, add);
off = op0;
goto do_add;
}
break;
case MULT_EXPR:
if (scale == 1 && host_integerp (op1, 0))
{
scale = tree_low_cst (op1, 0);
off = op0;
continue;
}
break;
case SSA_NAME:
off = op0;
continue;
CASE_CONVERT:
if (!POINTER_TYPE_P (TREE_TYPE (op0))
&& !INTEGRAL_TYPE_P (TREE_TYPE (op0)))
break;
if (TYPE_PRECISION (TREE_TYPE (op0))
== TYPE_PRECISION (TREE_TYPE (off)))
{
off = op0;
continue;
}
if (TYPE_PRECISION (TREE_TYPE (op0))
< TYPE_PRECISION (TREE_TYPE (off)))
{
off = op0;
offtype = TREE_TYPE (off);
STRIP_NOPS (off);
continue;
}
break;
default:
break;
}
break;
}
/* If at the end OFF still isn't a SSA_NAME or isn't
defined in the loop, punt. */
if (TREE_CODE (off) != SSA_NAME
|| expr_invariant_in_loop_p (loop, off))
return NULL_TREE;
if (offtype == NULL_TREE)
offtype = TREE_TYPE (off);
decl = targetm.vectorize.builtin_gather (STMT_VINFO_VECTYPE (stmt_info),
offtype, scale);
if (decl == NULL_TREE)
return NULL_TREE;
if (basep)
*basep = base;
if (offp)
*offp = off;
if (scalep)
*scalep = scale;
return decl;
}
/* Check wether a non-affine load in STMT (being in the loop referred to
in LOOP_VINFO) is suitable for handling as strided load. That is the case
if its address is a simple induction variable. If so return the base
of that induction variable in *BASEP and the (loop-invariant) step
in *STEPP, both only when that pointer is non-zero.
This handles ARRAY_REFs (with variant index) and MEM_REFs (with variant
base pointer) only. */
static bool
vect_check_strided_load (gimple stmt, loop_vec_info loop_vinfo)
{
struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
tree base, off;
affine_iv iv;
if (!DR_IS_READ (dr))
return false;
base = DR_REF (dr);
if (TREE_CODE (base) == ARRAY_REF)
{
off = TREE_OPERAND (base, 1);
base = TREE_OPERAND (base, 0);
}
else if (TREE_CODE (base) == MEM_REF)
{
off = TREE_OPERAND (base, 0);
base = TREE_OPERAND (base, 1);
}
else
return false;
if (TREE_CODE (off) != SSA_NAME)
return false;
if (!expr_invariant_in_loop_p (loop, base)
|| !simple_iv (loop, loop_containing_stmt (stmt), off, &iv, true))
return false;
return true;
}
/* Function vect_analyze_data_refs.
Find all the data references in the loop or basic block.
The general structure of the analysis of data refs in the vectorizer is as
follows:
1- vect_analyze_data_refs(loop/bb): call
compute_data_dependences_for_loop/bb to find and analyze all data-refs
in the loop/bb and their dependences.
2- vect_analyze_dependences(): apply dependence testing using ddrs.
3- vect_analyze_drs_alignment(): check that ref_stmt.alignment is ok.
4- vect_analyze_drs_access(): check that ref_stmt.step is ok.
*/
bool
vect_analyze_data_refs (loop_vec_info loop_vinfo,
bb_vec_info bb_vinfo,
int *min_vf)
{
struct loop *loop = NULL;
basic_block bb = NULL;
unsigned int i;
vec<data_reference_p> datarefs;
struct data_reference *dr;
tree scalar_type;
bool res, stop_bb_analysis = false;
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"=== vect_analyze_data_refs ===\n");
if (loop_vinfo)
{
loop = LOOP_VINFO_LOOP (loop_vinfo);
res = compute_data_dependences_for_loop
(loop, true,
&LOOP_VINFO_LOOP_NEST (loop_vinfo),
&LOOP_VINFO_DATAREFS (loop_vinfo),
&LOOP_VINFO_DDRS (loop_vinfo));
if (!res)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not vectorized: loop contains function calls"
" or data references that cannot be analyzed");
return false;
}
datarefs = LOOP_VINFO_DATAREFS (loop_vinfo);
}
else
{
gimple_stmt_iterator gsi;
bb = BB_VINFO_BB (bb_vinfo);
for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
{
gimple stmt = gsi_stmt (gsi);
if (!find_data_references_in_stmt (NULL, stmt,
&BB_VINFO_DATAREFS (bb_vinfo)))
{
/* Mark the rest of the basic-block as unvectorizable. */
for (; !gsi_end_p (gsi); gsi_next (&gsi))
{
stmt = gsi_stmt (gsi);
STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (stmt)) = false;
}
break;
}
}
if (!compute_all_dependences (BB_VINFO_DATAREFS (bb_vinfo),
&BB_VINFO_DDRS (bb_vinfo),
vNULL, true))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not vectorized: basic block contains function"
" calls or data references that cannot be"
" analyzed");
return false;
}
datarefs = BB_VINFO_DATAREFS (bb_vinfo);
}
/* Go through the data-refs, check that the analysis succeeded. Update
pointer from stmt_vec_info struct to DR and vectype. */
FOR_EACH_VEC_ELT (datarefs, i, dr)
{
gimple stmt;
stmt_vec_info stmt_info;
tree base, offset, init;
bool gather = false;
int vf;
if (!dr || !DR_REF (dr))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not vectorized: unhandled data-ref ");
return false;
}
stmt = DR_STMT (dr);
stmt_info = vinfo_for_stmt (stmt);
if (stop_bb_analysis)
{
STMT_VINFO_VECTORIZABLE (stmt_info) = false;
continue;
}
/* Check that analysis of the data-ref succeeded. */
if (!DR_BASE_ADDRESS (dr) || !DR_OFFSET (dr) || !DR_INIT (dr)
|| !DR_STEP (dr))
{
/* If target supports vector gather loads, see if they can't
be used. */
if (loop_vinfo
&& DR_IS_READ (dr)
&& !TREE_THIS_VOLATILE (DR_REF (dr))
&& targetm.vectorize.builtin_gather != NULL
&& !nested_in_vect_loop_p (loop, stmt))
{
struct data_reference *newdr
= create_data_ref (NULL, loop_containing_stmt (stmt),
DR_REF (dr), stmt, true);
gcc_assert (newdr != NULL && DR_REF (newdr));
if (DR_BASE_ADDRESS (newdr)
&& DR_OFFSET (newdr)
&& DR_INIT (newdr)
&& DR_STEP (newdr)
&& integer_zerop (DR_STEP (newdr)))
{
dr = newdr;
gather = true;
}
else
free_data_ref (newdr);
}
if (!gather)
{
if (dump_enabled_p ())
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not vectorized: data ref analysis "
"failed ");
dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0);
}
if (bb_vinfo)
{
STMT_VINFO_VECTORIZABLE (stmt_info) = false;
stop_bb_analysis = true;
continue;
}
return false;
}
}
if (TREE_CODE (DR_BASE_ADDRESS (dr)) == INTEGER_CST)
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not vectorized: base addr of dr is a "
"constant");
if (bb_vinfo)
{
STMT_VINFO_VECTORIZABLE (stmt_info) = false;
stop_bb_analysis = true;
continue;
}
if (gather)
free_data_ref (dr);
return false;
}
if (TREE_THIS_VOLATILE (DR_REF (dr)))
{
if (dump_enabled_p ())
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not vectorized: volatile type ");
dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0);
}
if (bb_vinfo)
{
STMT_VINFO_VECTORIZABLE (stmt_info) = false;
stop_bb_analysis = true;
continue;
}
return false;
}
if (stmt_can_throw_internal (stmt))
{
if (dump_enabled_p ())
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not vectorized: statement can throw an "
"exception ");
dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0);
}
if (bb_vinfo)
{
STMT_VINFO_VECTORIZABLE (stmt_info) = false;
stop_bb_analysis = true;
continue;
}
if (gather)
free_data_ref (dr);
return false;
}
if (TREE_CODE (DR_REF (dr)) == COMPONENT_REF
&& DECL_BIT_FIELD (TREE_OPERAND (DR_REF (dr), 1)))
{
if (dump_enabled_p ())
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not vectorized: statement is bitfield "
"access ");
dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0);
}
if (bb_vinfo)
{
STMT_VINFO_VECTORIZABLE (stmt_info) = false;
stop_bb_analysis = true;
continue;
}
if (gather)
free_data_ref (dr);
return false;
}
base = unshare_expr (DR_BASE_ADDRESS (dr));
offset = unshare_expr (DR_OFFSET (dr));
init = unshare_expr (DR_INIT (dr));
if (is_gimple_call (stmt))
{
if (dump_enabled_p ())
{
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
"not vectorized: dr in a call ");
dump_gimple_stmt (MSG_MISSED_OPTIMIZATION, TDF_SLIM, stmt, 0);
}
if (bb_vinfo)
{
STMT_VINFO_VECTORIZABLE (stmt_info) = false;
stop_bb_analysis = true;
continue;
}
if (gather)
free_data_ref (dr);
return false;
}
/* Update DR field in stmt_vec_info struct. */
/* If the dataref is in an inner-loop of the loop that is considered for
for vectorization, we also want to analyze the access relative to
the outer-loop (DR contains information only relative to the
inner-most enclosing loop). We do that by building a reference to the
first location accessed by the inner-loop, and analyze it relative to
the outer-loop. */
if (loop && nested_in_vect_loop_p (loop, stmt))
{
tree outer_step, outer_base, outer_init;
HOST_WIDE_INT pbitsize, pbitpos;
tree poffset;
enum machine_mode pmode;
int punsignedp, pvolatilep;
affine_iv base_iv, offset_iv;
tree dinit;
/* Build a reference to the first location accessed by the
inner-loop: *(BASE+INIT). (The first location is actually
BASE+INIT+OFFSET, but we add OFFSET separately later). */
tree inner_base = build_fold_indirect_ref
(fold_build_pointer_plus (base, init));
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"analyze in outer-loop: ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, inner_base);
}
outer_base = get_inner_reference (inner_base, &pbitsize, &pbitpos,
&poffset, &pmode, &punsignedp, &pvolatilep, false);
gcc_assert (outer_base != NULL_TREE);
if (pbitpos % BITS_PER_UNIT != 0)