blob: 46f4ffedb4830dd67ae12f8e36c19db234a88723 [file] [log] [blame]
/* Data references and dependences detectors.
Copyright (C) 2003-2021 Free Software Foundation, Inc.
Contributed by Sebastian Pop <pop@cri.ensmp.fr>
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/>. */
/* This pass walks a given loop structure searching for array
references. The information about the array accesses is recorded
in DATA_REFERENCE structures.
The basic test for determining the dependences is:
given two access functions chrec1 and chrec2 to a same array, and
x and y two vectors from the iteration domain, the same element of
the array is accessed twice at iterations x and y if and only if:
| chrec1 (x) == chrec2 (y).
The goals of this analysis are:
- to determine the independence: the relation between two
independent accesses is qualified with the chrec_known (this
information allows a loop parallelization),
- when two data references access the same data, to qualify the
dependence relation with classic dependence representations:
- distance vectors
- direction vectors
- loop carried level dependence
- polyhedron dependence
or with the chains of recurrences based representation,
- to define a knowledge base for storing the data dependence
information,
- to define an interface to access this data.
Definitions:
- subscript: given two array accesses a subscript is the tuple
composed of the access functions for a given dimension. Example:
Given A[f1][f2][f3] and B[g1][g2][g3], there are three subscripts:
(f1, g1), (f2, g2), (f3, g3).
- Diophantine equation: an equation whose coefficients and
solutions are integer constants, for example the equation
| 3*x + 2*y = 1
has an integer solution x = 1 and y = -1.
References:
- "Advanced Compilation for High Performance Computing" by Randy
Allen and Ken Kennedy.
http://citeseer.ist.psu.edu/goff91practical.html
- "Loop Transformations for Restructuring Compilers - The Foundations"
by Utpal Banerjee.
*/
#include "config.h"
#include "system.h"
#include "coretypes.h"
#include "backend.h"
#include "rtl.h"
#include "tree.h"
#include "gimple.h"
#include "gimple-pretty-print.h"
#include "alias.h"
#include "fold-const.h"
#include "expr.h"
#include "gimple-iterator.h"
#include "tree-ssa-loop-niter.h"
#include "tree-ssa-loop.h"
#include "tree-ssa.h"
#include "cfgloop.h"
#include "tree-data-ref.h"
#include "tree-scalar-evolution.h"
#include "dumpfile.h"
#include "tree-affine.h"
#include "builtins.h"
#include "tree-eh.h"
#include "ssa.h"
#include "internal-fn.h"
#include "vr-values.h"
#include "range-op.h"
#include "tree-ssa-loop-ivopts.h"
static struct datadep_stats
{
int num_dependence_tests;
int num_dependence_dependent;
int num_dependence_independent;
int num_dependence_undetermined;
int num_subscript_tests;
int num_subscript_undetermined;
int num_same_subscript_function;
int num_ziv;
int num_ziv_independent;
int num_ziv_dependent;
int num_ziv_unimplemented;
int num_siv;
int num_siv_independent;
int num_siv_dependent;
int num_siv_unimplemented;
int num_miv;
int num_miv_independent;
int num_miv_dependent;
int num_miv_unimplemented;
} dependence_stats;
static bool subscript_dependence_tester_1 (struct data_dependence_relation *,
unsigned int, unsigned int,
class loop *);
/* Returns true iff A divides B. */
static inline bool
tree_fold_divides_p (const_tree a, const_tree b)
{
gcc_assert (TREE_CODE (a) == INTEGER_CST);
gcc_assert (TREE_CODE (b) == INTEGER_CST);
return integer_zerop (int_const_binop (TRUNC_MOD_EXPR, b, a));
}
/* Returns true iff A divides B. */
static inline bool
int_divides_p (lambda_int a, lambda_int b)
{
return ((b % a) == 0);
}
/* Return true if reference REF contains a union access. */
static bool
ref_contains_union_access_p (tree ref)
{
while (handled_component_p (ref))
{
ref = TREE_OPERAND (ref, 0);
if (TREE_CODE (TREE_TYPE (ref)) == UNION_TYPE
|| TREE_CODE (TREE_TYPE (ref)) == QUAL_UNION_TYPE)
return true;
}
return false;
}
/* Dump into FILE all the data references from DATAREFS. */
static void
dump_data_references (FILE *file, vec<data_reference_p> datarefs)
{
for (data_reference *dr : datarefs)
dump_data_reference (file, dr);
}
/* Unified dump into FILE all the data references from DATAREFS. */
DEBUG_FUNCTION void
debug (vec<data_reference_p> &ref)
{
dump_data_references (stderr, ref);
}
DEBUG_FUNCTION void
debug (vec<data_reference_p> *ptr)
{
if (ptr)
debug (*ptr);
else
fprintf (stderr, "<nil>\n");
}
/* Dump into STDERR all the data references from DATAREFS. */
DEBUG_FUNCTION void
debug_data_references (vec<data_reference_p> datarefs)
{
dump_data_references (stderr, datarefs);
}
/* Print to STDERR the data_reference DR. */
DEBUG_FUNCTION void
debug_data_reference (struct data_reference *dr)
{
dump_data_reference (stderr, dr);
}
/* Dump function for a DATA_REFERENCE structure. */
void
dump_data_reference (FILE *outf,
struct data_reference *dr)
{
unsigned int i;
fprintf (outf, "#(Data Ref: \n");
fprintf (outf, "# bb: %d \n", gimple_bb (DR_STMT (dr))->index);
fprintf (outf, "# stmt: ");
print_gimple_stmt (outf, DR_STMT (dr), 0);
fprintf (outf, "# ref: ");
print_generic_stmt (outf, DR_REF (dr));
fprintf (outf, "# base_object: ");
print_generic_stmt (outf, DR_BASE_OBJECT (dr));
for (i = 0; i < DR_NUM_DIMENSIONS (dr); i++)
{
fprintf (outf, "# Access function %d: ", i);
print_generic_stmt (outf, DR_ACCESS_FN (dr, i));
}
fprintf (outf, "#)\n");
}
/* Unified dump function for a DATA_REFERENCE structure. */
DEBUG_FUNCTION void
debug (data_reference &ref)
{
dump_data_reference (stderr, &ref);
}
DEBUG_FUNCTION void
debug (data_reference *ptr)
{
if (ptr)
debug (*ptr);
else
fprintf (stderr, "<nil>\n");
}
/* Dumps the affine function described by FN to the file OUTF. */
DEBUG_FUNCTION void
dump_affine_function (FILE *outf, affine_fn fn)
{
unsigned i;
tree coef;
print_generic_expr (outf, fn[0], TDF_SLIM);
for (i = 1; fn.iterate (i, &coef); i++)
{
fprintf (outf, " + ");
print_generic_expr (outf, coef, TDF_SLIM);
fprintf (outf, " * x_%u", i);
}
}
/* Dumps the conflict function CF to the file OUTF. */
DEBUG_FUNCTION void
dump_conflict_function (FILE *outf, conflict_function *cf)
{
unsigned i;
if (cf->n == NO_DEPENDENCE)
fprintf (outf, "no dependence");
else if (cf->n == NOT_KNOWN)
fprintf (outf, "not known");
else
{
for (i = 0; i < cf->n; i++)
{
if (i != 0)
fprintf (outf, " ");
fprintf (outf, "[");
dump_affine_function (outf, cf->fns[i]);
fprintf (outf, "]");
}
}
}
/* Dump function for a SUBSCRIPT structure. */
DEBUG_FUNCTION void
dump_subscript (FILE *outf, struct subscript *subscript)
{
conflict_function *cf = SUB_CONFLICTS_IN_A (subscript);
fprintf (outf, "\n (subscript \n");
fprintf (outf, " iterations_that_access_an_element_twice_in_A: ");
dump_conflict_function (outf, cf);
if (CF_NONTRIVIAL_P (cf))
{
tree last_iteration = SUB_LAST_CONFLICT (subscript);
fprintf (outf, "\n last_conflict: ");
print_generic_expr (outf, last_iteration);
}
cf = SUB_CONFLICTS_IN_B (subscript);
fprintf (outf, "\n iterations_that_access_an_element_twice_in_B: ");
dump_conflict_function (outf, cf);
if (CF_NONTRIVIAL_P (cf))
{
tree last_iteration = SUB_LAST_CONFLICT (subscript);
fprintf (outf, "\n last_conflict: ");
print_generic_expr (outf, last_iteration);
}
fprintf (outf, "\n (Subscript distance: ");
print_generic_expr (outf, SUB_DISTANCE (subscript));
fprintf (outf, " ))\n");
}
/* Print the classic direction vector DIRV to OUTF. */
DEBUG_FUNCTION void
print_direction_vector (FILE *outf,
lambda_vector dirv,
int length)
{
int eq;
for (eq = 0; eq < length; eq++)
{
enum data_dependence_direction dir = ((enum data_dependence_direction)
dirv[eq]);
switch (dir)
{
case dir_positive:
fprintf (outf, " +");
break;
case dir_negative:
fprintf (outf, " -");
break;
case dir_equal:
fprintf (outf, " =");
break;
case dir_positive_or_equal:
fprintf (outf, " +=");
break;
case dir_positive_or_negative:
fprintf (outf, " +-");
break;
case dir_negative_or_equal:
fprintf (outf, " -=");
break;
case dir_star:
fprintf (outf, " *");
break;
default:
fprintf (outf, "indep");
break;
}
}
fprintf (outf, "\n");
}
/* Print a vector of direction vectors. */
DEBUG_FUNCTION void
print_dir_vectors (FILE *outf, vec<lambda_vector> dir_vects,
int length)
{
for (lambda_vector v : dir_vects)
print_direction_vector (outf, v, length);
}
/* Print out a vector VEC of length N to OUTFILE. */
DEBUG_FUNCTION void
print_lambda_vector (FILE * outfile, lambda_vector vector, int n)
{
int i;
for (i = 0; i < n; i++)
fprintf (outfile, "%3d ", (int)vector[i]);
fprintf (outfile, "\n");
}
/* Print a vector of distance vectors. */
DEBUG_FUNCTION void
print_dist_vectors (FILE *outf, vec<lambda_vector> dist_vects,
int length)
{
for (lambda_vector v : dist_vects)
print_lambda_vector (outf, v, length);
}
/* Dump function for a DATA_DEPENDENCE_RELATION structure. */
DEBUG_FUNCTION void
dump_data_dependence_relation (FILE *outf, const data_dependence_relation *ddr)
{
struct data_reference *dra, *drb;
fprintf (outf, "(Data Dep: \n");
if (!ddr || DDR_ARE_DEPENDENT (ddr) == chrec_dont_know)
{
if (ddr)
{
dra = DDR_A (ddr);
drb = DDR_B (ddr);
if (dra)
dump_data_reference (outf, dra);
else
fprintf (outf, " (nil)\n");
if (drb)
dump_data_reference (outf, drb);
else
fprintf (outf, " (nil)\n");
}
fprintf (outf, " (don't know)\n)\n");
return;
}
dra = DDR_A (ddr);
drb = DDR_B (ddr);
dump_data_reference (outf, dra);
dump_data_reference (outf, drb);
if (DDR_ARE_DEPENDENT (ddr) == chrec_known)
fprintf (outf, " (no dependence)\n");
else if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
{
unsigned int i;
class loop *loopi;
subscript *sub;
FOR_EACH_VEC_ELT (DDR_SUBSCRIPTS (ddr), i, sub)
{
fprintf (outf, " access_fn_A: ");
print_generic_stmt (outf, SUB_ACCESS_FN (sub, 0));
fprintf (outf, " access_fn_B: ");
print_generic_stmt (outf, SUB_ACCESS_FN (sub, 1));
dump_subscript (outf, sub);
}
fprintf (outf, " loop nest: (");
FOR_EACH_VEC_ELT (DDR_LOOP_NEST (ddr), i, loopi)
fprintf (outf, "%d ", loopi->num);
fprintf (outf, ")\n");
for (i = 0; i < DDR_NUM_DIST_VECTS (ddr); i++)
{
fprintf (outf, " distance_vector: ");
print_lambda_vector (outf, DDR_DIST_VECT (ddr, i),
DDR_NB_LOOPS (ddr));
}
for (i = 0; i < DDR_NUM_DIR_VECTS (ddr); i++)
{
fprintf (outf, " direction_vector: ");
print_direction_vector (outf, DDR_DIR_VECT (ddr, i),
DDR_NB_LOOPS (ddr));
}
}
fprintf (outf, ")\n");
}
/* Debug version. */
DEBUG_FUNCTION void
debug_data_dependence_relation (const struct data_dependence_relation *ddr)
{
dump_data_dependence_relation (stderr, ddr);
}
/* Dump into FILE all the dependence relations from DDRS. */
DEBUG_FUNCTION void
dump_data_dependence_relations (FILE *file, const vec<ddr_p> &ddrs)
{
for (auto ddr : ddrs)
dump_data_dependence_relation (file, ddr);
}
DEBUG_FUNCTION void
debug (vec<ddr_p> &ref)
{
dump_data_dependence_relations (stderr, ref);
}
DEBUG_FUNCTION void
debug (vec<ddr_p> *ptr)
{
if (ptr)
debug (*ptr);
else
fprintf (stderr, "<nil>\n");
}
/* Dump to STDERR all the dependence relations from DDRS. */
DEBUG_FUNCTION void
debug_data_dependence_relations (vec<ddr_p> ddrs)
{
dump_data_dependence_relations (stderr, ddrs);
}
/* Dumps the distance and direction vectors in FILE. DDRS contains
the dependence relations, and VECT_SIZE is the size of the
dependence vectors, or in other words the number of loops in the
considered nest. */
DEBUG_FUNCTION void
dump_dist_dir_vectors (FILE *file, vec<ddr_p> ddrs)
{
for (data_dependence_relation *ddr : ddrs)
if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE && DDR_AFFINE_P (ddr))
{
for (lambda_vector v : DDR_DIST_VECTS (ddr))
{
fprintf (file, "DISTANCE_V (");
print_lambda_vector (file, v, DDR_NB_LOOPS (ddr));
fprintf (file, ")\n");
}
for (lambda_vector v : DDR_DIR_VECTS (ddr))
{
fprintf (file, "DIRECTION_V (");
print_direction_vector (file, v, DDR_NB_LOOPS (ddr));
fprintf (file, ")\n");
}
}
fprintf (file, "\n\n");
}
/* Dumps the data dependence relations DDRS in FILE. */
DEBUG_FUNCTION void
dump_ddrs (FILE *file, vec<ddr_p> ddrs)
{
for (data_dependence_relation *ddr : ddrs)
dump_data_dependence_relation (file, ddr);
fprintf (file, "\n\n");
}
DEBUG_FUNCTION void
debug_ddrs (vec<ddr_p> ddrs)
{
dump_ddrs (stderr, ddrs);
}
/* If RESULT_RANGE is nonnull, set *RESULT_RANGE to the range of
OP0 CODE OP1, where:
- OP0 CODE OP1 has integral type TYPE
- the range of OP0 is given by OP0_RANGE and
- the range of OP1 is given by OP1_RANGE.
Independently of RESULT_RANGE, try to compute:
DELTA = ((sizetype) OP0 CODE (sizetype) OP1)
- (sizetype) (OP0 CODE OP1)
as a constant and subtract DELTA from the ssizetype constant in *OFF.
Return true on success, or false if DELTA is not known at compile time.
Truncation and sign changes are known to distribute over CODE, i.e.
(itype) (A CODE B) == (itype) A CODE (itype) B
for any integral type ITYPE whose precision is no greater than the
precision of A and B. */
static bool
compute_distributive_range (tree type, value_range &op0_range,
tree_code code, value_range &op1_range,
tree *off, value_range *result_range)
{
gcc_assert (INTEGRAL_TYPE_P (type) && !TYPE_OVERFLOW_TRAPS (type));
if (result_range)
{
range_operator *op = range_op_handler (code, type);
op->fold_range (*result_range, type, op0_range, op1_range);
}
/* The distributive property guarantees that if TYPE is no narrower
than SIZETYPE,
(sizetype) (OP0 CODE OP1) == (sizetype) OP0 CODE (sizetype) OP1
and so we can treat DELTA as zero. */
if (TYPE_PRECISION (type) >= TYPE_PRECISION (sizetype))
return true;
/* If overflow is undefined, we can assume that:
X == (ssizetype) OP0 CODE (ssizetype) OP1
is within the range of TYPE, i.e.:
X == (ssizetype) (TYPE) X
Distributing the (TYPE) truncation over X gives:
X == (ssizetype) (OP0 CODE OP1)
Casting both sides to sizetype and distributing the sizetype cast
over X gives:
(sizetype) OP0 CODE (sizetype) OP1 == (sizetype) (OP0 CODE OP1)
and so we can treat DELTA as zero. */
if (TYPE_OVERFLOW_UNDEFINED (type))
return true;
/* Compute the range of:
(ssizetype) OP0 CODE (ssizetype) OP1
The distributive property guarantees that this has the same bitpattern as:
(sizetype) OP0 CODE (sizetype) OP1
but its range is more conducive to analysis. */
range_cast (op0_range, ssizetype);
range_cast (op1_range, ssizetype);
value_range wide_range;
range_operator *op = range_op_handler (code, ssizetype);
bool saved_flag_wrapv = flag_wrapv;
flag_wrapv = 1;
op->fold_range (wide_range, ssizetype, op0_range, op1_range);
flag_wrapv = saved_flag_wrapv;
if (wide_range.num_pairs () != 1 || !range_int_cst_p (&wide_range))
return false;
wide_int lb = wide_range.lower_bound ();
wide_int ub = wide_range.upper_bound ();
/* Calculate the number of times that each end of the range overflows or
underflows TYPE. We can only calculate DELTA if the numbers match. */
unsigned int precision = TYPE_PRECISION (type);
if (!TYPE_UNSIGNED (type))
{
wide_int type_min = wi::mask (precision - 1, true, lb.get_precision ());
lb -= type_min;
ub -= type_min;
}
wide_int upper_bits = wi::mask (precision, true, lb.get_precision ());
lb &= upper_bits;
ub &= upper_bits;
if (lb != ub)
return false;
/* OP0 CODE OP1 overflows exactly arshift (LB, PRECISION) times, with
negative values indicating underflow. The low PRECISION bits of LB
are clear, so DELTA is therefore LB (== UB). */
*off = wide_int_to_tree (ssizetype, wi::to_wide (*off) - lb);
return true;
}
/* Return true if (sizetype) OP == (sizetype) (TO_TYPE) OP,
given that OP has type FROM_TYPE and range RANGE. Both TO_TYPE and
FROM_TYPE are integral types. */
static bool
nop_conversion_for_offset_p (tree to_type, tree from_type, value_range &range)
{
gcc_assert (INTEGRAL_TYPE_P (to_type)
&& INTEGRAL_TYPE_P (from_type)
&& !TYPE_OVERFLOW_TRAPS (to_type)
&& !TYPE_OVERFLOW_TRAPS (from_type));
/* Converting to something no narrower than sizetype and then to sizetype
is equivalent to converting directly to sizetype. */
if (TYPE_PRECISION (to_type) >= TYPE_PRECISION (sizetype))
return true;
/* Check whether TO_TYPE can represent all values that FROM_TYPE can. */
if (TYPE_PRECISION (from_type) < TYPE_PRECISION (to_type)
&& (TYPE_UNSIGNED (from_type) || !TYPE_UNSIGNED (to_type)))
return true;
/* For narrowing conversions, we could in principle test whether
the bits in FROM_TYPE but not in TO_TYPE have a fixed value
and apply a constant adjustment.
For other conversions (which involve a sign change) we could
check that the signs are always equal, and apply a constant
adjustment if the signs are negative.
However, both cases should be rare. */
return range_fits_type_p (&range, TYPE_PRECISION (to_type),
TYPE_SIGN (to_type));
}
static void
split_constant_offset (tree type, tree *var, tree *off,
value_range *result_range,
hash_map<tree, std::pair<tree, tree> > &cache,
unsigned *limit);
/* Helper function for split_constant_offset. If TYPE is a pointer type,
try to express OP0 CODE OP1 as:
POINTER_PLUS <*VAR, (sizetype) *OFF>
where:
- *VAR has type TYPE
- *OFF is a constant of type ssizetype.
If TYPE is an integral type, try to express (sizetype) (OP0 CODE OP1) as:
*VAR + (sizetype) *OFF
where:
- *VAR has type sizetype
- *OFF is a constant of type ssizetype.
In both cases, OP0 CODE OP1 has type TYPE.
Return true on success. A false return value indicates that we can't
do better than set *OFF to zero.
When returning true, set RESULT_RANGE to the range of OP0 CODE OP1,
if RESULT_RANGE is nonnull and if we can do better than assume VR_VARYING.
CACHE caches {*VAR, *OFF} pairs for SSA names that we've previously
visited. LIMIT counts down the number of SSA names that we are
allowed to process before giving up. */
static bool
split_constant_offset_1 (tree type, tree op0, enum tree_code code, tree op1,
tree *var, tree *off, value_range *result_range,
hash_map<tree, std::pair<tree, tree> > &cache,
unsigned *limit)
{
tree var0, var1;
tree off0, off1;
value_range op0_range, op1_range;
*var = NULL_TREE;
*off = NULL_TREE;
if (INTEGRAL_TYPE_P (type) && TYPE_OVERFLOW_TRAPS (type))
return false;
switch (code)
{
case INTEGER_CST:
*var = size_int (0);
*off = fold_convert (ssizetype, op0);
if (result_range)
result_range->set (op0, op0);
return true;
case POINTER_PLUS_EXPR:
split_constant_offset (op0, &var0, &off0, nullptr, cache, limit);
split_constant_offset (op1, &var1, &off1, nullptr, cache, limit);
*var = fold_build2 (POINTER_PLUS_EXPR, type, var0, var1);
*off = size_binop (PLUS_EXPR, off0, off1);
return true;
case PLUS_EXPR:
case MINUS_EXPR:
split_constant_offset (op0, &var0, &off0, &op0_range, cache, limit);
split_constant_offset (op1, &var1, &off1, &op1_range, cache, limit);
*off = size_binop (code, off0, off1);
if (!compute_distributive_range (type, op0_range, code, op1_range,
off, result_range))
return false;
*var = fold_build2 (code, sizetype, var0, var1);
return true;
case MULT_EXPR:
if (TREE_CODE (op1) != INTEGER_CST)
return false;
split_constant_offset (op0, &var0, &off0, &op0_range, cache, limit);
op1_range.set (op1, op1);
*off = size_binop (MULT_EXPR, off0, fold_convert (ssizetype, op1));
if (!compute_distributive_range (type, op0_range, code, op1_range,
off, result_range))
return false;
*var = fold_build2 (MULT_EXPR, sizetype, var0,
fold_convert (sizetype, op1));
return true;
case ADDR_EXPR:
{
tree base, poffset;
poly_int64 pbitsize, pbitpos, pbytepos;
machine_mode pmode;
int punsignedp, preversep, pvolatilep;
op0 = TREE_OPERAND (op0, 0);
base
= get_inner_reference (op0, &pbitsize, &pbitpos, &poffset, &pmode,
&punsignedp, &preversep, &pvolatilep);
if (!multiple_p (pbitpos, BITS_PER_UNIT, &pbytepos))
return false;
base = build_fold_addr_expr (base);
off0 = ssize_int (pbytepos);
if (poffset)
{
split_constant_offset (poffset, &poffset, &off1, nullptr,
cache, limit);
off0 = size_binop (PLUS_EXPR, off0, off1);
base = fold_build_pointer_plus (base, poffset);
}
var0 = fold_convert (type, base);
/* If variable length types are involved, punt, otherwise casts
might be converted into ARRAY_REFs in gimplify_conversion.
To compute that ARRAY_REF's element size TYPE_SIZE_UNIT, which
possibly no longer appears in current GIMPLE, might resurface.
This perhaps could run
if (CONVERT_EXPR_P (var0))
{
gimplify_conversion (&var0);
// Attempt to fill in any within var0 found ARRAY_REF's
// element size from corresponding op embedded ARRAY_REF,
// if unsuccessful, just punt.
} */
while (POINTER_TYPE_P (type))
type = TREE_TYPE (type);
if (int_size_in_bytes (type) < 0)
return false;
*var = var0;
*off = off0;
return true;
}
case SSA_NAME:
{
if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (op0))
return false;
gimple *def_stmt = SSA_NAME_DEF_STMT (op0);
enum tree_code subcode;
if (gimple_code (def_stmt) != GIMPLE_ASSIGN)
return false;
subcode = gimple_assign_rhs_code (def_stmt);
/* We are using a cache to avoid un-CSEing large amounts of code. */
bool use_cache = false;
if (!has_single_use (op0)
&& (subcode == POINTER_PLUS_EXPR
|| subcode == PLUS_EXPR
|| subcode == MINUS_EXPR
|| subcode == MULT_EXPR
|| subcode == ADDR_EXPR
|| CONVERT_EXPR_CODE_P (subcode)))
{
use_cache = true;
bool existed;
std::pair<tree, tree> &e = cache.get_or_insert (op0, &existed);
if (existed)
{
if (integer_zerop (e.second))
return false;
*var = e.first;
*off = e.second;
/* The caller sets the range in this case. */
return true;
}
e = std::make_pair (op0, ssize_int (0));
}
if (*limit == 0)
return false;
--*limit;
var0 = gimple_assign_rhs1 (def_stmt);
var1 = gimple_assign_rhs2 (def_stmt);
bool res = split_constant_offset_1 (type, var0, subcode, var1,
var, off, nullptr, cache, limit);
if (res && use_cache)
*cache.get (op0) = std::make_pair (*var, *off);
/* The caller sets the range in this case. */
return res;
}
CASE_CONVERT:
{
/* We can only handle the following conversions:
- Conversions from one pointer type to another pointer type.
- Conversions from one non-trapping integral type to another
non-trapping integral type. In this case, the recursive
call makes sure that:
(sizetype) OP0
can be expressed as a sizetype operation involving VAR and OFF,
and all we need to do is check whether:
(sizetype) OP0 == (sizetype) (TYPE) OP0
- Conversions from a non-trapping sizetype-size integral type to
a like-sized pointer type. In this case, the recursive call
makes sure that:
(sizetype) OP0 == *VAR + (sizetype) *OFF
and we can convert that to:
POINTER_PLUS <(TYPE) *VAR, (sizetype) *OFF>
- Conversions from a sizetype-sized pointer type to a like-sized
non-trapping integral type. In this case, the recursive call
makes sure that:
OP0 == POINTER_PLUS <*VAR, (sizetype) *OFF>
where the POINTER_PLUS and *VAR have the same precision as
TYPE (and the same precision as sizetype). Then:
(sizetype) (TYPE) OP0 == (sizetype) *VAR + (sizetype) *OFF. */
tree itype = TREE_TYPE (op0);
if ((POINTER_TYPE_P (itype)
|| (INTEGRAL_TYPE_P (itype) && !TYPE_OVERFLOW_TRAPS (itype)))
&& (POINTER_TYPE_P (type)
|| (INTEGRAL_TYPE_P (type) && !TYPE_OVERFLOW_TRAPS (type)))
&& (POINTER_TYPE_P (type) == POINTER_TYPE_P (itype)
|| (TYPE_PRECISION (type) == TYPE_PRECISION (sizetype)
&& TYPE_PRECISION (itype) == TYPE_PRECISION (sizetype))))
{
if (POINTER_TYPE_P (type))
{
split_constant_offset (op0, var, off, nullptr, cache, limit);
*var = fold_convert (type, *var);
}
else if (POINTER_TYPE_P (itype))
{
split_constant_offset (op0, var, off, nullptr, cache, limit);
*var = fold_convert (sizetype, *var);
}
else
{
split_constant_offset (op0, var, off, &op0_range,
cache, limit);
if (!nop_conversion_for_offset_p (type, itype, op0_range))
return false;
if (result_range)
{
*result_range = op0_range;
range_cast (*result_range, type);
}
}
return true;
}
return false;
}
default:
return false;
}
}
/* If EXP has pointer type, try to express it as:
POINTER_PLUS <*VAR, (sizetype) *OFF>
where:
- *VAR has the same type as EXP
- *OFF is a constant of type ssizetype.
If EXP has an integral type, try to express (sizetype) EXP as:
*VAR + (sizetype) *OFF
where:
- *VAR has type sizetype
- *OFF is a constant of type ssizetype.
If EXP_RANGE is nonnull, set it to the range of EXP.
CACHE caches {*VAR, *OFF} pairs for SSA names that we've previously
visited. LIMIT counts down the number of SSA names that we are
allowed to process before giving up. */
static void
split_constant_offset (tree exp, tree *var, tree *off, value_range *exp_range,
hash_map<tree, std::pair<tree, tree> > &cache,
unsigned *limit)
{
tree type = TREE_TYPE (exp), op0, op1;
enum tree_code code;
code = TREE_CODE (exp);
if (exp_range)
{
*exp_range = type;
if (code == SSA_NAME)
{
value_range vr;
get_range_query (cfun)->range_of_expr (vr, exp);
if (vr.undefined_p ())
vr.set_varying (TREE_TYPE (exp));
wide_int var_min = wi::to_wide (vr.min ());
wide_int var_max = wi::to_wide (vr.max ());
value_range_kind vr_kind = vr.kind ();
wide_int var_nonzero = get_nonzero_bits (exp);
vr_kind = intersect_range_with_nonzero_bits (vr_kind,
&var_min, &var_max,
var_nonzero,
TYPE_SIGN (type));
/* This check for VR_VARYING is here because the old code
using get_range_info would return VR_RANGE for the entire
domain, instead of VR_VARYING. The new code normalizes
full-domain ranges to VR_VARYING. */
if (vr_kind == VR_RANGE || vr_kind == VR_VARYING)
*exp_range = value_range (type, var_min, var_max);
}
}
if (!tree_is_chrec (exp)
&& get_gimple_rhs_class (TREE_CODE (exp)) != GIMPLE_TERNARY_RHS)
{
extract_ops_from_tree (exp, &code, &op0, &op1);
if (split_constant_offset_1 (type, op0, code, op1, var, off,
exp_range, cache, limit))
return;
}
*var = exp;
if (INTEGRAL_TYPE_P (type))
*var = fold_convert (sizetype, *var);
*off = ssize_int (0);
value_range r;
if (exp_range && code != SSA_NAME
&& get_range_query (cfun)->range_of_expr (r, exp)
&& !r.undefined_p ())
*exp_range = r;
}
/* Expresses EXP as VAR + OFF, where OFF is a constant. VAR has the same
type as EXP while OFF has type ssizetype. */
void
split_constant_offset (tree exp, tree *var, tree *off)
{
unsigned limit = param_ssa_name_def_chain_limit;
static hash_map<tree, std::pair<tree, tree> > *cache;
if (!cache)
cache = new hash_map<tree, std::pair<tree, tree> > (37);
split_constant_offset (exp, var, off, nullptr, *cache, &limit);
*var = fold_convert (TREE_TYPE (exp), *var);
cache->empty ();
}
/* Returns the address ADDR of an object in a canonical shape (without nop
casts, and with type of pointer to the object). */
static tree
canonicalize_base_object_address (tree addr)
{
tree orig = addr;
STRIP_NOPS (addr);
/* The base address may be obtained by casting from integer, in that case
keep the cast. */
if (!POINTER_TYPE_P (TREE_TYPE (addr)))
return orig;
if (TREE_CODE (addr) != ADDR_EXPR)
return addr;
return build_fold_addr_expr (TREE_OPERAND (addr, 0));
}
/* Analyze the behavior of memory reference REF within STMT.
There are two modes:
- BB analysis. In this case we simply split the address into base,
init and offset components, without reference to any containing loop.
The resulting base and offset are general expressions and they can
vary arbitrarily from one iteration of the containing loop to the next.
The step is always zero.
- loop analysis. In this case we analyze the reference both wrt LOOP
and on the basis that the reference occurs (is "used") in LOOP;
see the comment above analyze_scalar_evolution_in_loop for more
information about this distinction. The base, init, offset and
step fields are all invariant in LOOP.
Perform BB analysis if LOOP is null, or if LOOP is the function's
dummy outermost loop. In other cases perform loop analysis.
Return true if the analysis succeeded and store the results in DRB if so.
BB analysis can only fail for bitfield or reversed-storage accesses. */
opt_result
dr_analyze_innermost (innermost_loop_behavior *drb, tree ref,
class loop *loop, const gimple *stmt)
{
poly_int64 pbitsize, pbitpos;
tree base, poffset;
machine_mode pmode;
int punsignedp, preversep, pvolatilep;
affine_iv base_iv, offset_iv;
tree init, dinit, step;
bool in_loop = (loop && loop->num);
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "analyze_innermost: ");
base = get_inner_reference (ref, &pbitsize, &pbitpos, &poffset, &pmode,
&punsignedp, &preversep, &pvolatilep);
gcc_assert (base != NULL_TREE);
poly_int64 pbytepos;
if (!multiple_p (pbitpos, BITS_PER_UNIT, &pbytepos))
return opt_result::failure_at (stmt,
"failed: bit offset alignment.\n");
if (preversep)
return opt_result::failure_at (stmt,
"failed: reverse storage order.\n");
/* Calculate the alignment and misalignment for the inner reference. */
unsigned int HOST_WIDE_INT bit_base_misalignment;
unsigned int bit_base_alignment;
get_object_alignment_1 (base, &bit_base_alignment, &bit_base_misalignment);
/* There are no bitfield references remaining in BASE, so the values
we got back must be whole bytes. */
gcc_assert (bit_base_alignment % BITS_PER_UNIT == 0
&& bit_base_misalignment % BITS_PER_UNIT == 0);
unsigned int base_alignment = bit_base_alignment / BITS_PER_UNIT;
poly_int64 base_misalignment = bit_base_misalignment / BITS_PER_UNIT;
if (TREE_CODE (base) == MEM_REF)
{
if (!integer_zerop (TREE_OPERAND (base, 1)))
{
/* Subtract MOFF from the base and add it to POFFSET instead.
Adjust the misalignment to reflect the amount we subtracted. */
poly_offset_int moff = mem_ref_offset (base);
base_misalignment -= moff.force_shwi ();
tree mofft = wide_int_to_tree (sizetype, moff);
if (!poffset)
poffset = mofft;
else
poffset = size_binop (PLUS_EXPR, poffset, mofft);
}
base = TREE_OPERAND (base, 0);
}
else
base = build_fold_addr_expr (base);
if (in_loop)
{
if (!simple_iv (loop, loop, base, &base_iv, true))
return opt_result::failure_at
(stmt, "failed: evolution of base is not affine.\n");
}
else
{
base_iv.base = base;
base_iv.step = ssize_int (0);
base_iv.no_overflow = true;
}
if (!poffset)
{
offset_iv.base = ssize_int (0);
offset_iv.step = ssize_int (0);
}
else
{
if (!in_loop)
{
offset_iv.base = poffset;
offset_iv.step = ssize_int (0);
}
else if (!simple_iv (loop, loop, poffset, &offset_iv, true))
return opt_result::failure_at
(stmt, "failed: evolution of offset is not affine.\n");
}
init = ssize_int (pbytepos);
/* Subtract any constant component from the base and add it to INIT instead.
Adjust the misalignment to reflect the amount we subtracted. */
split_constant_offset (base_iv.base, &base_iv.base, &dinit);
init = size_binop (PLUS_EXPR, init, dinit);
base_misalignment -= TREE_INT_CST_LOW (dinit);
split_constant_offset (offset_iv.base, &offset_iv.base, &dinit);
init = size_binop (PLUS_EXPR, init, dinit);
step = size_binop (PLUS_EXPR,
fold_convert (ssizetype, base_iv.step),
fold_convert (ssizetype, offset_iv.step));
base = canonicalize_base_object_address (base_iv.base);
/* See if get_pointer_alignment can guarantee a higher alignment than
the one we calculated above. */
unsigned int HOST_WIDE_INT alt_misalignment;
unsigned int alt_alignment;
get_pointer_alignment_1 (base, &alt_alignment, &alt_misalignment);
/* As above, these values must be whole bytes. */
gcc_assert (alt_alignment % BITS_PER_UNIT == 0
&& alt_misalignment % BITS_PER_UNIT == 0);
alt_alignment /= BITS_PER_UNIT;
alt_misalignment /= BITS_PER_UNIT;
if (base_alignment < alt_alignment)
{
base_alignment = alt_alignment;
base_misalignment = alt_misalignment;
}
drb->base_address = base;
drb->offset = fold_convert (ssizetype, offset_iv.base);
drb->init = init;
drb->step = step;
if (known_misalignment (base_misalignment, base_alignment,
&drb->base_misalignment))
drb->base_alignment = base_alignment;
else
{
drb->base_alignment = known_alignment (base_misalignment);
drb->base_misalignment = 0;
}
drb->offset_alignment = highest_pow2_factor (offset_iv.base);
drb->step_alignment = highest_pow2_factor (step);
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "success.\n");
return opt_result::success ();
}
/* Return true if OP is a valid component reference for a DR access
function. This accepts a subset of what handled_component_p accepts. */
static bool
access_fn_component_p (tree op)
{
switch (TREE_CODE (op))
{
case REALPART_EXPR:
case IMAGPART_EXPR:
case ARRAY_REF:
return true;
case COMPONENT_REF:
return TREE_CODE (TREE_TYPE (TREE_OPERAND (op, 0))) == RECORD_TYPE;
default:
return false;
}
}
/* Returns whether BASE can have a access_fn_component_p with BASE
as base. */
static bool
base_supports_access_fn_components_p (tree base)
{
switch (TREE_CODE (TREE_TYPE (base)))
{
case COMPLEX_TYPE:
case ARRAY_TYPE:
case RECORD_TYPE:
return true;
default:
return false;
}
}
/* Determines the base object and the list of indices of memory reference
DR, analyzed in LOOP and instantiated before NEST. */
static void
dr_analyze_indices (struct indices *dri, tree ref, edge nest, loop_p loop)
{
/* If analyzing a basic-block there are no indices to analyze
and thus no access functions. */
if (!nest)
{
dri->base_object = ref;
dri->access_fns.create (0);
return;
}
vec<tree> access_fns = vNULL;
/* REALPART_EXPR and IMAGPART_EXPR can be handled like accesses
into a two element array with a constant index. The base is
then just the immediate underlying object. */
if (TREE_CODE (ref) == REALPART_EXPR)
{
ref = TREE_OPERAND (ref, 0);
access_fns.safe_push (integer_zero_node);
}
else if (TREE_CODE (ref) == IMAGPART_EXPR)
{
ref = TREE_OPERAND (ref, 0);
access_fns.safe_push (integer_one_node);
}
/* Analyze access functions of dimensions we know to be independent.
The list of component references handled here should be kept in
sync with access_fn_component_p. */
while (handled_component_p (ref))
{
if (TREE_CODE (ref) == ARRAY_REF)
{
tree op = TREE_OPERAND (ref, 1);
tree access_fn = analyze_scalar_evolution (loop, op);
access_fn = instantiate_scev (nest, loop, access_fn);
access_fns.safe_push (access_fn);
}
else if (TREE_CODE (ref) == COMPONENT_REF
&& TREE_CODE (TREE_TYPE (TREE_OPERAND (ref, 0))) == RECORD_TYPE)
{
/* For COMPONENT_REFs of records (but not unions!) use the
FIELD_DECL offset as constant access function so we can
disambiguate a[i].f1 and a[i].f2. */
tree off = component_ref_field_offset (ref);
off = size_binop (PLUS_EXPR,
size_binop (MULT_EXPR,
fold_convert (bitsizetype, off),
bitsize_int (BITS_PER_UNIT)),
DECL_FIELD_BIT_OFFSET (TREE_OPERAND (ref, 1)));
access_fns.safe_push (off);
}
else
/* If we have an unhandled component we could not translate
to an access function stop analyzing. We have determined
our base object in this case. */
break;
ref = TREE_OPERAND (ref, 0);
}
/* If the address operand of a MEM_REF base has an evolution in the
analyzed nest, add it as an additional independent access-function. */
if (TREE_CODE (ref) == MEM_REF)
{
tree op = TREE_OPERAND (ref, 0);
tree access_fn = analyze_scalar_evolution (loop, op);
access_fn = instantiate_scev (nest, loop, access_fn);
STRIP_NOPS (access_fn);
if (TREE_CODE (access_fn) == POLYNOMIAL_CHREC)
{
tree memoff = TREE_OPERAND (ref, 1);
tree base = initial_condition (access_fn);
tree orig_type = TREE_TYPE (base);
STRIP_USELESS_TYPE_CONVERSION (base);
tree off;
split_constant_offset (base, &base, &off);
STRIP_USELESS_TYPE_CONVERSION (base);
/* Fold the MEM_REF offset into the evolutions initial
value to make more bases comparable. */
if (!integer_zerop (memoff))
{
off = size_binop (PLUS_EXPR, off,
fold_convert (ssizetype, memoff));
memoff = build_int_cst (TREE_TYPE (memoff), 0);
}
/* Adjust the offset so it is a multiple of the access type
size and thus we separate bases that can possibly be used
to produce partial overlaps (which the access_fn machinery
cannot handle). */
wide_int rem;
if (TYPE_SIZE_UNIT (TREE_TYPE (ref))
&& TREE_CODE (TYPE_SIZE_UNIT (TREE_TYPE (ref))) == INTEGER_CST
&& !integer_zerop (TYPE_SIZE_UNIT (TREE_TYPE (ref))))
rem = wi::mod_trunc
(wi::to_wide (off),
wi::to_wide (TYPE_SIZE_UNIT (TREE_TYPE (ref))),
SIGNED);
else
/* If we can't compute the remainder simply force the initial
condition to zero. */
rem = wi::to_wide (off);
off = wide_int_to_tree (ssizetype, wi::to_wide (off) - rem);
memoff = wide_int_to_tree (TREE_TYPE (memoff), rem);
/* And finally replace the initial condition. */
access_fn = chrec_replace_initial_condition
(access_fn, fold_convert (orig_type, off));
/* ??? This is still not a suitable base object for
dr_may_alias_p - the base object needs to be an
access that covers the object as whole. With
an evolution in the pointer this cannot be
guaranteed.
As a band-aid, mark the access so we can special-case
it in dr_may_alias_p. */
tree old = ref;
ref = fold_build2_loc (EXPR_LOCATION (ref),
MEM_REF, TREE_TYPE (ref),
base, memoff);
MR_DEPENDENCE_CLIQUE (ref) = MR_DEPENDENCE_CLIQUE (old);
MR_DEPENDENCE_BASE (ref) = MR_DEPENDENCE_BASE (old);
dri->unconstrained_base = true;
access_fns.safe_push (access_fn);
}
}
else if (DECL_P (ref))
{
/* Canonicalize DR_BASE_OBJECT to MEM_REF form. */
ref = build2 (MEM_REF, TREE_TYPE (ref),
build_fold_addr_expr (ref),
build_int_cst (reference_alias_ptr_type (ref), 0));
}
dri->base_object = ref;
dri->access_fns = access_fns;
}
/* Extracts the alias analysis information from the memory reference DR. */
static void
dr_analyze_alias (struct data_reference *dr)
{
tree ref = DR_REF (dr);
tree base = get_base_address (ref), addr;
if (INDIRECT_REF_P (base)
|| TREE_CODE (base) == MEM_REF)
{
addr = TREE_OPERAND (base, 0);
if (TREE_CODE (addr) == SSA_NAME)
DR_PTR_INFO (dr) = SSA_NAME_PTR_INFO (addr);
}
}
/* Frees data reference DR. */
void
free_data_ref (data_reference_p dr)
{
DR_ACCESS_FNS (dr).release ();
if (dr->alt_indices.base_object)
dr->alt_indices.access_fns.release ();
free (dr);
}
/* Analyze memory reference MEMREF, which is accessed in STMT.
The reference is a read if IS_READ is true, otherwise it is a write.
IS_CONDITIONAL_IN_STMT indicates that the reference is conditional
within STMT, i.e. that it might not occur even if STMT is executed
and runs to completion.
Return the data_reference description of MEMREF. NEST is the outermost
loop in which the reference should be instantiated, LOOP is the loop
in which the data reference should be analyzed. */
struct data_reference *
create_data_ref (edge nest, loop_p loop, tree memref, gimple *stmt,
bool is_read, bool is_conditional_in_stmt)
{
struct data_reference *dr;
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "Creating dr for ");
print_generic_expr (dump_file, memref, TDF_SLIM);
fprintf (dump_file, "\n");
}
dr = XCNEW (struct data_reference);
DR_STMT (dr) = stmt;
DR_REF (dr) = memref;
DR_IS_READ (dr) = is_read;
DR_IS_CONDITIONAL_IN_STMT (dr) = is_conditional_in_stmt;
dr_analyze_innermost (&DR_INNERMOST (dr), memref,
nest != NULL ? loop : NULL, stmt);
dr_analyze_indices (&dr->indices, DR_REF (dr), nest, loop);
dr_analyze_alias (dr);
if (dump_file && (dump_flags & TDF_DETAILS))
{
unsigned i;
fprintf (dump_file, "\tbase_address: ");
print_generic_expr (dump_file, DR_BASE_ADDRESS (dr), TDF_SLIM);
fprintf (dump_file, "\n\toffset from base address: ");
print_generic_expr (dump_file, DR_OFFSET (dr), TDF_SLIM);
fprintf (dump_file, "\n\tconstant offset from base address: ");
print_generic_expr (dump_file, DR_INIT (dr), TDF_SLIM);
fprintf (dump_file, "\n\tstep: ");
print_generic_expr (dump_file, DR_STEP (dr), TDF_SLIM);
fprintf (dump_file, "\n\tbase alignment: %d", DR_BASE_ALIGNMENT (dr));
fprintf (dump_file, "\n\tbase misalignment: %d",
DR_BASE_MISALIGNMENT (dr));
fprintf (dump_file, "\n\toffset alignment: %d",
DR_OFFSET_ALIGNMENT (dr));
fprintf (dump_file, "\n\tstep alignment: %d", DR_STEP_ALIGNMENT (dr));
fprintf (dump_file, "\n\tbase_object: ");
print_generic_expr (dump_file, DR_BASE_OBJECT (dr), TDF_SLIM);
fprintf (dump_file, "\n");
for (i = 0; i < DR_NUM_DIMENSIONS (dr); i++)
{
fprintf (dump_file, "\tAccess function %d: ", i);
print_generic_stmt (dump_file, DR_ACCESS_FN (dr, i), TDF_SLIM);
}
}
return dr;
}
/* A helper function computes order between two tree expressions T1 and T2.
This is used in comparator functions sorting objects based on the order
of tree expressions. The function returns -1, 0, or 1. */
int
data_ref_compare_tree (tree t1, tree t2)
{
int i, cmp;
enum tree_code code;
char tclass;
if (t1 == t2)
return 0;
if (t1 == NULL)
return -1;
if (t2 == NULL)
return 1;
STRIP_USELESS_TYPE_CONVERSION (t1);
STRIP_USELESS_TYPE_CONVERSION (t2);
if (t1 == t2)
return 0;
if (TREE_CODE (t1) != TREE_CODE (t2)
&& ! (CONVERT_EXPR_P (t1) && CONVERT_EXPR_P (t2)))
return TREE_CODE (t1) < TREE_CODE (t2) ? -1 : 1;
code = TREE_CODE (t1);
switch (code)
{
case INTEGER_CST:
return tree_int_cst_compare (t1, t2);
case STRING_CST:
if (TREE_STRING_LENGTH (t1) != TREE_STRING_LENGTH (t2))
return TREE_STRING_LENGTH (t1) < TREE_STRING_LENGTH (t2) ? -1 : 1;
return memcmp (TREE_STRING_POINTER (t1), TREE_STRING_POINTER (t2),
TREE_STRING_LENGTH (t1));
case SSA_NAME:
if (SSA_NAME_VERSION (t1) != SSA_NAME_VERSION (t2))
return SSA_NAME_VERSION (t1) < SSA_NAME_VERSION (t2) ? -1 : 1;
break;
default:
if (POLY_INT_CST_P (t1))
return compare_sizes_for_sort (wi::to_poly_widest (t1),
wi::to_poly_widest (t2));
tclass = TREE_CODE_CLASS (code);
/* For decls, compare their UIDs. */
if (tclass == tcc_declaration)
{
if (DECL_UID (t1) != DECL_UID (t2))
return DECL_UID (t1) < DECL_UID (t2) ? -1 : 1;
break;
}
/* For expressions, compare their operands recursively. */
else if (IS_EXPR_CODE_CLASS (tclass))
{
for (i = TREE_OPERAND_LENGTH (t1) - 1; i >= 0; --i)
{
cmp = data_ref_compare_tree (TREE_OPERAND (t1, i),
TREE_OPERAND (t2, i));
if (cmp != 0)
return cmp;
}
}
else
gcc_unreachable ();
}
return 0;
}
/* Return TRUE it's possible to resolve data dependence DDR by runtime alias
check. */
opt_result
runtime_alias_check_p (ddr_p ddr, class loop *loop, bool speed_p)
{
if (dump_enabled_p ())
dump_printf (MSG_NOTE,
"consider run-time aliasing test between %T and %T\n",
DR_REF (DDR_A (ddr)), DR_REF (DDR_B (ddr)));
if (!speed_p)
return opt_result::failure_at (DR_STMT (DDR_A (ddr)),
"runtime alias check not supported when"
" optimizing for size.\n");
/* FORNOW: We don't support versioning with outer-loop in either
vectorization or loop distribution. */
if (loop != NULL && loop->inner != NULL)
return opt_result::failure_at (DR_STMT (DDR_A (ddr)),
"runtime alias check not supported for"
" outer loop.\n");
return opt_result::success ();
}
/* Operator == between two dr_with_seg_len objects.
This equality operator is used to make sure two data refs
are the same one so that we will consider to combine the
aliasing checks of those two pairs of data dependent data
refs. */
static bool
operator == (const dr_with_seg_len& d1,
const dr_with_seg_len& d2)
{
return (operand_equal_p (DR_BASE_ADDRESS (d1.dr),
DR_BASE_ADDRESS (d2.dr), 0)
&& data_ref_compare_tree (DR_OFFSET (d1.dr), DR_OFFSET (d2.dr)) == 0
&& data_ref_compare_tree (DR_INIT (d1.dr), DR_INIT (d2.dr)) == 0
&& data_ref_compare_tree (d1.seg_len, d2.seg_len) == 0
&& known_eq (d1.access_size, d2.access_size)
&& d1.align == d2.align);
}
/* Comparison function for sorting objects of dr_with_seg_len_pair_t
so that we can combine aliasing checks in one scan. */
static int
comp_dr_with_seg_len_pair (const void *pa_, const void *pb_)
{
const dr_with_seg_len_pair_t* pa = (const dr_with_seg_len_pair_t *) pa_;
const dr_with_seg_len_pair_t* pb = (const dr_with_seg_len_pair_t *) pb_;
const dr_with_seg_len &a1 = pa->first, &a2 = pa->second;
const dr_with_seg_len &b1 = pb->first, &b2 = pb->second;
/* For DR pairs (a, b) and (c, d), we only consider to merge the alias checks
if a and c have the same basic address snd step, and b and d have the same
address and step. Therefore, if any a&c or b&d don't have the same address
and step, we don't care the order of those two pairs after sorting. */
int comp_res;
if ((comp_res = data_ref_compare_tree (DR_BASE_ADDRESS (a1.dr),
DR_BASE_ADDRESS (b1.dr))) != 0)
return comp_res;
if ((comp_res = data_ref_compare_tree (DR_BASE_ADDRESS (a2.dr),
DR_BASE_ADDRESS (b2.dr))) != 0)
return comp_res;
if ((comp_res = data_ref_compare_tree (DR_STEP (a1.dr),
DR_STEP (b1.dr))) != 0)
return comp_res;
if ((comp_res = data_ref_compare_tree (DR_STEP (a2.dr),
DR_STEP (b2.dr))) != 0)
return comp_res;
if ((comp_res = data_ref_compare_tree (DR_OFFSET (a1.dr),
DR_OFFSET (b1.dr))) != 0)
return comp_res;
if ((comp_res = data_ref_compare_tree (DR_INIT (a1.dr),
DR_INIT (b1.dr))) != 0)
return comp_res;
if ((comp_res = data_ref_compare_tree (DR_OFFSET (a2.dr),
DR_OFFSET (b2.dr))) != 0)
return comp_res;
if ((comp_res = data_ref_compare_tree (DR_INIT (a2.dr),
DR_INIT (b2.dr))) != 0)
return comp_res;
return 0;
}
/* Dump information about ALIAS_PAIR, indenting each line by INDENT. */
static void
dump_alias_pair (dr_with_seg_len_pair_t *alias_pair, const char *indent)
{
dump_printf (MSG_NOTE, "%sreference: %T vs. %T\n", indent,
DR_REF (alias_pair->first.dr),
DR_REF (alias_pair->second.dr));
dump_printf (MSG_NOTE, "%ssegment length: %T", indent,
alias_pair->first.seg_len);
if (!operand_equal_p (alias_pair->first.seg_len,
alias_pair->second.seg_len, 0))
dump_printf (MSG_NOTE, " vs. %T", alias_pair->second.seg_len);
dump_printf (MSG_NOTE, "\n%saccess size: ", indent);
dump_dec (MSG_NOTE, alias_pair->first.access_size);
if (maybe_ne (alias_pair->first.access_size, alias_pair->second.access_size))
{
dump_printf (MSG_NOTE, " vs. ");
dump_dec (MSG_NOTE, alias_pair->second.access_size);
}
dump_printf (MSG_NOTE, "\n%salignment: %d", indent,
alias_pair->first.align);
if (alias_pair->first.align != alias_pair->second.align)
dump_printf (MSG_NOTE, " vs. %d", alias_pair->second.align);
dump_printf (MSG_NOTE, "\n%sflags: ", indent);
if (alias_pair->flags & DR_ALIAS_RAW)
dump_printf (MSG_NOTE, " RAW");
if (alias_pair->flags & DR_ALIAS_WAR)
dump_printf (MSG_NOTE, " WAR");
if (alias_pair->flags & DR_ALIAS_WAW)
dump_printf (MSG_NOTE, " WAW");
if (alias_pair->flags & DR_ALIAS_ARBITRARY)
dump_printf (MSG_NOTE, " ARBITRARY");
if (alias_pair->flags & DR_ALIAS_SWAPPED)
dump_printf (MSG_NOTE, " SWAPPED");
if (alias_pair->flags & DR_ALIAS_UNSWAPPED)
dump_printf (MSG_NOTE, " UNSWAPPED");
if (alias_pair->flags & DR_ALIAS_MIXED_STEPS)
dump_printf (MSG_NOTE, " MIXED_STEPS");
if (alias_pair->flags == 0)
dump_printf (MSG_NOTE, " <none>");
dump_printf (MSG_NOTE, "\n");
}
/* Merge alias checks recorded in ALIAS_PAIRS and remove redundant ones.
FACTOR is number of iterations that each data reference is accessed.
Basically, for each pair of dependent data refs store_ptr_0 & load_ptr_0,
we create an expression:
((store_ptr_0 + store_segment_length_0) <= load_ptr_0)
|| (load_ptr_0 + load_segment_length_0) <= store_ptr_0))
for aliasing checks. However, in some cases we can decrease the number
of checks by combining two checks into one. For example, suppose we have
another pair of data refs store_ptr_0 & load_ptr_1, and if the following
condition is satisfied:
load_ptr_0 < load_ptr_1 &&
load_ptr_1 - load_ptr_0 - load_segment_length_0 < store_segment_length_0
(this condition means, in each iteration of vectorized loop, the accessed
memory of store_ptr_0 cannot be between the memory of load_ptr_0 and
load_ptr_1.)
we then can use only the following expression to finish the alising checks
between store_ptr_0 & load_ptr_0 and store_ptr_0 & load_ptr_1:
((store_ptr_0 + store_segment_length_0) <= load_ptr_0)
|| (load_ptr_1 + load_segment_length_1 <= store_ptr_0))
Note that we only consider that load_ptr_0 and load_ptr_1 have the same
basic address. */
void
prune_runtime_alias_test_list (vec<dr_with_seg_len_pair_t> *alias_pairs,
poly_uint64)
{
if (alias_pairs->is_empty ())
return;
/* Canonicalize each pair so that the base components are ordered wrt
data_ref_compare_tree. This allows the loop below to merge more
cases. */
unsigned int i;
dr_with_seg_len_pair_t *alias_pair;
FOR_EACH_VEC_ELT (*alias_pairs, i, alias_pair)
{
data_reference_p dr_a = alias_pair->first.dr;
data_reference_p dr_b = alias_pair->second.dr;
int comp_res = data_ref_compare_tree (DR_BASE_ADDRESS (dr_a),
DR_BASE_ADDRESS (dr_b));
if (comp_res == 0)
comp_res = data_ref_compare_tree (DR_OFFSET (dr_a), DR_OFFSET (dr_b));
if (comp_res == 0)
comp_res = data_ref_compare_tree (DR_INIT (dr_a), DR_INIT (dr_b));
if (comp_res > 0)
{
std::swap (alias_pair->first, alias_pair->second);
alias_pair->flags |= DR_ALIAS_SWAPPED;
}
else
alias_pair->flags |= DR_ALIAS_UNSWAPPED;
}
/* Sort the collected data ref pairs so that we can scan them once to
combine all possible aliasing checks. */
alias_pairs->qsort (comp_dr_with_seg_len_pair);
/* Scan the sorted dr pairs and check if we can combine alias checks
of two neighboring dr pairs. */
unsigned int last = 0;
for (i = 1; i < alias_pairs->length (); ++i)
{
/* Deal with two ddrs (dr_a1, dr_b1) and (dr_a2, dr_b2). */
dr_with_seg_len_pair_t *alias_pair1 = &(*alias_pairs)[last];
dr_with_seg_len_pair_t *alias_pair2 = &(*alias_pairs)[i];
dr_with_seg_len *dr_a1 = &alias_pair1->first;
dr_with_seg_len *dr_b1 = &alias_pair1->second;
dr_with_seg_len *dr_a2 = &alias_pair2->first;
dr_with_seg_len *dr_b2 = &alias_pair2->second;
/* Remove duplicate data ref pairs. */
if (*dr_a1 == *dr_a2 && *dr_b1 == *dr_b2)
{
if (dump_enabled_p ())
dump_printf (MSG_NOTE, "found equal ranges %T, %T and %T, %T\n",
DR_REF (dr_a1->dr), DR_REF (dr_b1->dr),
DR_REF (dr_a2->dr), DR_REF (dr_b2->dr));
alias_pair1->flags |= alias_pair2->flags;
continue;
}
/* Assume that we won't be able to merge the pairs, then correct
if we do. */
last += 1;
if (last != i)
(*alias_pairs)[last] = (*alias_pairs)[i];
if (*dr_a1 == *dr_a2 || *dr_b1 == *dr_b2)
{
/* We consider the case that DR_B1 and DR_B2 are same memrefs,
and DR_A1 and DR_A2 are two consecutive memrefs. */
if (*dr_a1 == *dr_a2)
{
std::swap (dr_a1, dr_b1);
std::swap (dr_a2, dr_b2);
}
poly_int64 init_a1, init_a2;
/* Only consider cases in which the distance between the initial
DR_A1 and the initial DR_A2 is known at compile time. */
if (!operand_equal_p (DR_BASE_ADDRESS (dr_a1->dr),
DR_BASE_ADDRESS (dr_a2->dr), 0)
|| !operand_equal_p (DR_OFFSET (dr_a1->dr),
DR_OFFSET (dr_a2->dr), 0)
|| !poly_int_tree_p (DR_INIT (dr_a1->dr), &init_a1)
|| !poly_int_tree_p (DR_INIT (dr_a2->dr), &init_a2))
continue;
/* Don't combine if we can't tell which one comes first. */
if (!ordered_p (init_a1, init_a2))
continue;
/* Work out what the segment length would be if we did combine
DR_A1 and DR_A2:
- If DR_A1 and DR_A2 have equal lengths, that length is
also the combined length.
- If DR_A1 and DR_A2 both have negative "lengths", the combined
length is the lower bound on those lengths.
- If DR_A1 and DR_A2 both have positive lengths, the combined
length is the upper bound on those lengths.
Other cases are unlikely to give a useful combination.
The lengths both have sizetype, so the sign is taken from
the step instead. */
poly_uint64 new_seg_len = 0;
bool new_seg_len_p = !operand_equal_p (dr_a1->seg_len,
dr_a2->seg_len, 0);
if (new_seg_len_p)
{
poly_uint64 seg_len_a1, seg_len_a2;
if (!poly_int_tree_p (dr_a1->seg_len, &seg_len_a1)
|| !poly_int_tree_p (dr_a2->seg_len, &seg_len_a2))
continue;
tree indicator_a = dr_direction_indicator (dr_a1->dr);
if (TREE_CODE (indicator_a) != INTEGER_CST)
continue;
tree indicator_b = dr_direction_indicator (dr_a2->dr);
if (TREE_CODE (indicator_b) != INTEGER_CST)
continue;
int sign_a = tree_int_cst_sgn (indicator_a);
int sign_b = tree_int_cst_sgn (indicator_b);
if (sign_a <= 0 && sign_b <= 0)
new_seg_len = lower_bound (seg_len_a1, seg_len_a2);
else if (sign_a >= 0 && sign_b >= 0)
new_seg_len = upper_bound (seg_len_a1, seg_len_a2);
else
continue;
}
/* At this point we're committed to merging the refs. */
/* Make sure dr_a1 starts left of dr_a2. */
if (maybe_gt (init_a1, init_a2))
{
std::swap (*dr_a1, *dr_a2);
std::swap (init_a1, init_a2);
}
/* The DR_Bs are equal, so only the DR_As can introduce
mixed steps. */
if (!operand_equal_p (DR_STEP (dr_a1->dr), DR_STEP (dr_a2->dr), 0))
alias_pair1->flags |= DR_ALIAS_MIXED_STEPS;
if (new_seg_len_p)
{
dr_a1->seg_len = build_int_cst (TREE_TYPE (dr_a1->seg_len),
new_seg_len);
dr_a1->align = MIN (dr_a1->align, known_alignment (new_seg_len));
}
/* This is always positive due to the swap above. */
poly_uint64 diff = init_a2 - init_a1;
/* The new check will start at DR_A1. Make sure that its access
size encompasses the initial DR_A2. */
if (maybe_lt (dr_a1->access_size, diff + dr_a2->access_size))
{
dr_a1->access_size = upper_bound (dr_a1->access_size,
diff + dr_a2->access_size);
unsigned int new_align = known_alignment (dr_a1->access_size);
dr_a1->align = MIN (dr_a1->align, new_align);
}
if (dump_enabled_p ())
dump_printf (MSG_NOTE, "merging ranges for %T, %T and %T, %T\n",
DR_REF (dr_a1->dr), DR_REF (dr_b1->dr),
DR_REF (dr_a2->dr), DR_REF (dr_b2->dr));
alias_pair1->flags |= alias_pair2->flags;
last -= 1;
}
}
alias_pairs->truncate (last + 1);
/* Try to restore the original dr_with_seg_len order within each
dr_with_seg_len_pair_t. If we ended up combining swapped and
unswapped pairs into the same check, we have to invalidate any
RAW, WAR and WAW information for it. */
if (dump_enabled_p ())
dump_printf (MSG_NOTE, "merged alias checks:\n");
FOR_EACH_VEC_ELT (*alias_pairs, i, alias_pair)
{
unsigned int swap_mask = (DR_ALIAS_SWAPPED | DR_ALIAS_UNSWAPPED);
unsigned int swapped = (alias_pair->flags & swap_mask);
if (swapped == DR_ALIAS_SWAPPED)
std::swap (alias_pair->first, alias_pair->second);
else if (swapped != DR_ALIAS_UNSWAPPED)
alias_pair->flags |= DR_ALIAS_ARBITRARY;
alias_pair->flags &= ~swap_mask;
if (dump_enabled_p ())
dump_alias_pair (alias_pair, " ");
}
}
/* A subroutine of create_intersect_range_checks, with a subset of the
same arguments. Try to use IFN_CHECK_RAW_PTRS and IFN_CHECK_WAR_PTRS
to optimize cases in which the references form a simple RAW, WAR or
WAR dependence. */
static bool
create_ifn_alias_checks (tree *cond_expr,
const dr_with_seg_len_pair_t &alias_pair)
{
const dr_with_seg_len& dr_a = alias_pair.first;
const dr_with_seg_len& dr_b = alias_pair.second;
/* Check for cases in which:
(a) we have a known RAW, WAR or WAR dependence
(b) the accesses are well-ordered in both the original and new code
(see the comment above the DR_ALIAS_* flags for details); and
(c) the DR_STEPs describe all access pairs covered by ALIAS_PAIR. */
if (alias_pair.flags & ~(DR_ALIAS_RAW | DR_ALIAS_WAR | DR_ALIAS_WAW))
return false;
/* Make sure that both DRs access the same pattern of bytes,
with a constant length and step. */
poly_uint64 seg_len;
if (!operand_equal_p (dr_a.seg_len, dr_b.seg_len, 0)
|| !poly_int_tree_p (dr_a.seg_len, &seg_len)
|| maybe_ne (dr_a.access_size, dr_b.access_size)
|| !operand_equal_p (DR_STEP (dr_a.dr), DR_STEP (dr_b.dr), 0)
|| !tree_fits_uhwi_p (DR_STEP (dr_a.dr)))
return false;
unsigned HOST_WIDE_INT bytes = tree_to_uhwi (DR_STEP (dr_a.dr));
tree addr_a = DR_BASE_ADDRESS (dr_a.dr);
tree addr_b = DR_BASE_ADDRESS (dr_b.dr);
/* See whether the target suports what we want to do. WAW checks are
equivalent to WAR checks here. */
internal_fn ifn = (alias_pair.flags & DR_ALIAS_RAW
? IFN_CHECK_RAW_PTRS
: IFN_CHECK_WAR_PTRS);
unsigned int align = MIN (dr_a.align, dr_b.align);
poly_uint64 full_length = seg_len + bytes;
if (!internal_check_ptrs_fn_supported_p (ifn, TREE_TYPE (addr_a),
full_length, align))
{
full_length = seg_len + dr_a.access_size;
if (!internal_check_ptrs_fn_supported_p (ifn, TREE_TYPE (addr_a),
full_length, align))
return false;
}
/* Commit to using this form of test. */
addr_a = fold_build_pointer_plus (addr_a, DR_OFFSET (dr_a.dr));
addr_a = fold_build_pointer_plus (addr_a, DR_INIT (dr_a.dr));
addr_b = fold_build_pointer_plus (addr_b, DR_OFFSET (dr_b.dr));
addr_b = fold_build_pointer_plus (addr_b, DR_INIT (dr_b.dr));
*cond_expr = build_call_expr_internal_loc (UNKNOWN_LOCATION,
ifn, boolean_type_node,
4, addr_a, addr_b,
size_int (full_length),
size_int (align));
if (dump_enabled_p ())
{
if (ifn == IFN_CHECK_RAW_PTRS)
dump_printf (MSG_NOTE, "using an IFN_CHECK_RAW_PTRS test\n");
else
dump_printf (MSG_NOTE, "using an IFN_CHECK_WAR_PTRS test\n");
}
return true;
}
/* Try to generate a runtime condition that is true if ALIAS_PAIR is
free of aliases, using a condition based on index values instead
of a condition based on addresses. Return true on success,
storing the condition in *COND_EXPR.
This can only be done if the two data references in ALIAS_PAIR access
the same array object and the index is the only difference. For example,
if the two data references are DR_A and DR_B:
DR_A DR_B
data-ref arr[i] arr[j]
base_object arr arr
index {i_0, +, 1}_loop {j_0, +, 1}_loop
The addresses and their index are like:
|<- ADDR_A ->| |<- ADDR_B ->|
------------------------------------------------------->
| | | | | | | | | |
------------------------------------------------------->
i_0 ... i_0+4 j_0 ... j_0+4
We can create expression based on index rather than address:
(unsigned) (i_0 - j_0 + 3) <= 6
i.e. the indices are less than 4 apart.
Note evolution step of index needs to be considered in comparison. */
static bool
create_intersect_range_checks_index (class loop *loop, tree *cond_expr,
const dr_with_seg_len_pair_t &alias_pair)
{
const dr_with_seg_len &dr_a = alias_pair.first;
const dr_with_seg_len &dr_b = alias_pair.second;
if ((alias_pair.flags & DR_ALIAS_MIXED_STEPS)
|| integer_zerop (DR_STEP (dr_a.dr))
|| integer_zerop (DR_STEP (dr_b.dr))
|| DR_NUM_DIMENSIONS (dr_a.dr) != DR_NUM_DIMENSIONS (dr_b.dr))
return false;
poly_uint64 seg_len1, seg_len2;
if (!poly_int_tree_p (dr_a.seg_len, &seg_len1)
|| !poly_int_tree_p (dr_b.seg_len, &seg_len2))
return false;
if (!tree_fits_shwi_p (DR_STEP (dr_a.dr)))
return false;
if (!operand_equal_p (DR_BASE_OBJECT (dr_a.dr), DR_BASE_OBJECT (dr_b.dr), 0))
return false;
if (!operand_equal_p (DR_STEP (dr_a.dr), DR_STEP (dr_b.dr), 0))
return false;
gcc_assert (TREE_CODE (DR_STEP (dr_a.dr)) == INTEGER_CST);
bool neg_step = tree_int_cst_compare (DR_STEP (dr_a.dr), size_zero_node) < 0;
unsigned HOST_WIDE_INT abs_step = tree_to_shwi (DR_STEP (dr_a.dr));
if (neg_step)
{
abs_step = -abs_step;
seg_len1 = (-wi::to_poly_wide (dr_a.seg_len)).force_uhwi ();
seg_len2 = (-wi::to_poly_wide (dr_b.seg_len)).force_uhwi ();
}
/* Infer the number of iterations with which the memory segment is accessed
by DR. In other words, alias is checked if memory segment accessed by
DR_A in some iterations intersect with memory segment accessed by DR_B
in the same amount iterations.
Note segnment length is a linear function of number of iterations with
DR_STEP as the coefficient. */
poly_uint64 niter_len1, niter_len2;
if (!can_div_trunc_p (seg_len1 + abs_step - 1, abs_step, &niter_len1)
|| !can_div_trunc_p (seg_len2 + abs_step - 1, abs_step, &niter_len2))
return false;
/* Divide each access size by the byte step, rounding up. */
poly_uint64 niter_access1, niter_access2;
if (!can_div_trunc_p (dr_a.access_size + abs_step - 1,
abs_step, &niter_access1)
|| !can_div_trunc_p (dr_b.access_size + abs_step - 1,
abs_step, &niter_access2))
return false;
bool waw_or_war_p = (alias_pair.flags & ~(DR_ALIAS_WAR | DR_ALIAS_WAW)) == 0;
int found = -1;
for (unsigned int i = 0; i < DR_NUM_DIMENSIONS (dr_a.dr); i++)
{
tree access1 = DR_ACCESS_FN (dr_a.dr, i);
tree access2 = DR_ACCESS_FN (dr_b.dr, i);
/* Two indices must be the same if they are not scev, or not scev wrto
current loop being vecorized. */
if (TREE_CODE (access1) != POLYNOMIAL_CHREC
|| TREE_CODE (access2) != POLYNOMIAL_CHREC
|| CHREC_VARIABLE (access1) != (unsigned)loop->num
|| CHREC_VARIABLE (access2) != (unsigned)loop->num)
{
if (operand_equal_p (access1, access2, 0))
continue;
return false;
}
if (found >= 0)
return false;
found = i;
}
/* Ought not to happen in practice, since if all accesses are equal then the
alias should be decidable at compile time. */
if (found < 0)
return false;
/* The two indices must have the same step. */
tree access1 = DR_ACCESS_FN (dr_a.dr, found);
tree access2 = DR_ACCESS_FN (dr_b.dr, found);
if (!operand_equal_p (CHREC_RIGHT (access1), CHREC_RIGHT (access2), 0))
return false;
tree idx_step = CHREC_RIGHT (access1);
/* Index must have const step, otherwise DR_STEP won't be constant. */
gcc_assert (TREE_CODE (idx_step) == INTEGER_CST);
/* Index must evaluate in the same direction as DR. */
gcc_assert (!neg_step || tree_int_cst_sign_bit (idx_step) == 1);
tree min1 = CHREC_LEFT (access1);
tree min2 = CHREC_LEFT (access2);
if (!types_compatible_p (TREE_TYPE (min1), TREE_TYPE (min2)))
return false;
/* Ideally, alias can be checked against loop's control IV, but we
need to prove linear mapping between control IV and reference
index. Although that should be true, we check against (array)
index of data reference. Like segment length, index length is
linear function of the number of iterations with index_step as
the coefficient, i.e, niter_len * idx_step. */
offset_int abs_idx_step = offset_int::from (wi::to_wide (idx_step),
SIGNED);
if (neg_step)
abs_idx_step = -abs_idx_step;
poly_offset_int idx_len1 = abs_idx_step * niter_len1;
poly_offset_int idx_len2 = abs_idx_step * niter_len2;
poly_offset_int idx_access1 = abs_idx_step * niter_access1;
poly_offset_int idx_access2 = abs_idx_step * niter_access2;
gcc_assert (known_ge (idx_len1, 0)
&& known_ge (idx_len2, 0)
&& known_ge (idx_access1, 0)
&& known_ge (idx_access2, 0));
/* Each access has the following pattern, with lengths measured
in units of INDEX:
<-- idx_len -->
<--- A: -ve step --->
+-----+-------+-----+-------+-----+
| n-1 | ..... | 0 | ..... | n-1 |
+-----+-------+-----+-------+-----+
<--- B: +ve step --->
<-- idx_len -->
|
min
where "n" is the number of scalar iterations covered by the segment
and where each access spans idx_access units.
A is the range of bytes accessed when the step is negative,
B is the range when the step is positive.
When checking for general overlap, we need to test whether
the range:
[min1 + low_offset1, min1 + high_offset1 + idx_access1 - 1]
overlaps:
[min2 + low_offset2, min2 + high_offset2 + idx_access2 - 1]
where:
low_offsetN = +ve step ? 0 : -idx_lenN;
high_offsetN = +ve step ? idx_lenN : 0;
This is equivalent to testing whether:
min1 + low_offset1 <= min2 + high_offset2 + idx_access2 - 1
&& min2 + low_offset2 <= min1 + high_offset1 + idx_access1 - 1
Converting this into a single test, there is an overlap if:
0 <= min2 - min1 + bias <= limit
where bias = high_offset2 + idx_access2 - 1 - low_offset1
limit = (high_offset1 - low_offset1 + idx_access1 - 1)
+ (high_offset2 - low_offset2 + idx_access2 - 1)
i.e. limit = idx_len1 + idx_access1 - 1 + idx_len2 + idx_access2 - 1
Combining the tests requires limit to be computable in an unsigned
form of the index type; if it isn't, we fall back to the usual
pointer-based checks.
We can do better if DR_B is a write and if DR_A and DR_B are
well-ordered in both the original and the new code (see the
comment above the DR_ALIAS_* flags for details). In this case
we know that for each i in [0, n-1], the write performed by
access i of DR_B occurs after access numbers j<=i of DR_A in
both the original and the new code. Any write or anti
dependencies wrt those DR_A accesses are therefore maintained.
We just need to make sure that each individual write in DR_B does not
overlap any higher-indexed access in DR_A; such DR_A accesses happen
after the DR_B access in the original code but happen before it in
the new code.
We know the steps for both accesses are equal, so by induction, we
just need to test whether the first write of DR_B overlaps a later
access of DR_A. In other words, we need to move min1 along by
one iteration:
min1' = min1 + idx_step
and use the ranges:
[min1' + low_offset1', min1' + high_offset1' + idx_access1 - 1]
and:
[min2, min2 + idx_access2 - 1]
where:
low_offset1' = +ve step ? 0 : -(idx_len1 - |idx_step|)
high_offset1' = +ve_step ? idx_len1 - |idx_step| : 0. */
if (waw_or_war_p)
idx_len1 -= abs_idx_step;
poly_offset_int limit = idx_len1 + idx_access1 - 1 + idx_access2 - 1;
if (!waw_or_war_p)
limit += idx_len2;
tree utype = unsigned_type_for (TREE_TYPE (min1));
if (!wi::fits_to_tree_p (limit, utype))
return false;
poly_offset_int low_offset1 = neg_step ? -idx_len1 : 0;
poly_offset_int high_offset2 = neg_step || waw_or_war_p ? 0 : idx_len2;
poly_offset_int bias = high_offset2 + idx_access2 - 1 - low_offset1;
/* Equivalent to adding IDX_STEP to MIN1. */
if (waw_or_war_p)
bias -= wi::to_offset (idx_step);
tree subject = fold_build2 (MINUS_EXPR, utype,
fold_convert (utype, min2),
fold_convert (utype, min1));
subject = fold_build2 (PLUS_EXPR, utype, subject,
wide_int_to_tree (utype, bias));
tree part_cond_expr = fold_build2 (GT_EXPR, boolean_type_node, subject,
wide_int_to_tree (utype, limit));
if (*cond_expr)
*cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
*cond_expr, part_cond_expr);
else
*cond_expr = part_cond_expr;
if (dump_enabled_p ())
{
if (waw_or_war_p)
dump_printf (MSG_NOTE, "using an index-based WAR/WAW test\n");
else
dump_printf (MSG_NOTE, "using an index-based overlap test\n");
}
return true;
}
/* A subroutine of create_intersect_range_checks, with a subset of the
same arguments. Try to optimize cases in which the second access
is a write and in which some overlap is valid. */
static bool
create_waw_or_war_checks (tree *cond_expr,
const dr_with_seg_len_pair_t &alias_pair)
{
const dr_with_seg_len& dr_a = alias_pair.first;
const dr_with_seg_len& dr_b = alias_pair.second;
/* Check for cases in which:
(a) DR_B is always a write;
(b) the accesses are well-ordered in both the original and new code
(see the comment above the DR_ALIAS_* flags for details); and
(c) the DR_STEPs describe all access pairs covered by ALIAS_PAIR. */
if (alias_pair.flags & ~(DR_ALIAS_WAR | DR_ALIAS_WAW))
return false;
/* Check for equal (but possibly variable) steps. */
tree step = DR_STEP (dr_a.dr);
if (!operand_equal_p (step, DR_STEP (dr_b.dr)))
return false;
/* Make sure that we can operate on sizetype without loss of precision. */
tree addr_type = TREE_TYPE (DR_BASE_ADDRESS (dr_a.dr));
if (TYPE_PRECISION (addr_type) != TYPE_PRECISION (sizetype))
return false;
/* All addresses involved are known to have a common alignment ALIGN.
We can therefore subtract ALIGN from an exclusive endpoint to get
an inclusive endpoint. In the best (and common) case, ALIGN is the
same as the access sizes of both DRs, and so subtracting ALIGN
cancels out the addition of an access size. */
unsigned int align = MIN (dr_a.align, dr_b.align);
poly_uint64 last_chunk_a = dr_a.access_size - align;
poly_uint64 last_chunk_b = dr_b.access_size - align;
/* Get a boolean expression that is true when the step is negative. */
tree indicator = dr_direction_indicator (dr_a.dr);
tree neg_step = fold_build2 (LT_EXPR, boolean_type_node,
fold_convert (ssizetype, indicator),
ssize_int (0));
/* Get lengths in sizetype. */
tree seg_len_a
= fold_convert (sizetype, rewrite_to_non_trapping_overflow (dr_a.seg_len));
step = fold_convert (sizetype, rewrite_to_non_trapping_overflow (step));
/* Each access has the following pattern:
<- |seg_len| ->
<--- A: -ve step --->
+-----+-------+-----+-------+-----+
| n-1 | ..... | 0 | ..... | n-1 |
+-----+-------+-----+-------+-----+
<--- B: +ve step --->
<- |seg_len| ->
|
base address
where "n" is the number of scalar iterations covered by the segment.
A is the range of bytes accessed when the step is negative,
B is the range when the step is positive.
We know that DR_B is a write. We also know (from checking that
DR_A and DR_B are well-ordered) that for each i in [0, n-1],
the write performed by access i of DR_B occurs after access numbers
j<=i of DR_A in both the original and the new code. Any write or
anti dependencies wrt those DR_A accesses are therefore maintained.
We just need to make sure that each individual write in DR_B does not
overlap any higher-indexed access in DR_A; such DR_A accesses happen
after the DR_B access in the original code but happen before it in
the new code.
We know the steps for both accesses are equal, so by induction, we
just need to test whether the first write of DR_B overlaps a later
access of DR_A. In other words, we need to move addr_a along by
one iteration:
addr_a' = addr_a + step
and check whether:
[addr_b, addr_b + last_chunk_b]
overlaps:
[addr_a' + low_offset_a, addr_a' + high_offset_a + last_chunk_a]
where [low_offset_a, high_offset_a] spans accesses [1, n-1]. I.e.:
low_offset_a = +ve step ? 0 : seg_len_a - step
high_offset_a = +ve step ? seg_len_a - step : 0
This is equivalent to testing whether:
addr_a' + low_offset_a <= addr_b + last_chunk_b
&& addr_b <= addr_a' + high_offset_a + last_chunk_a
Converting this into a single test, there is an overlap if:
0 <= addr_b + last_chunk_b - addr_a' - low_offset_a <= limit
where limit = high_offset_a - low_offset_a + last_chunk_a + last_chunk_b
If DR_A is performed, limit + |step| - last_chunk_b is known to be
less than the size of the object underlying DR_A. We also know
that last_chunk_b <= |step|; this is checked elsewhere if it isn't
guaranteed at compile time. There can therefore be no overflow if
"limit" is calculated in an unsigned type with pointer precision. */
tree addr_a = fold_build_pointer_plus (DR_BASE_ADDRESS (dr_a.dr),
DR_OFFSET (dr_a.dr));
addr_a = fold_build_pointer_plus (addr_a, DR_INIT (dr_a.dr));
tree addr_b = fold_build_pointer_plus (DR_BASE_ADDRESS (dr_b.dr),
DR_OFFSET (dr_b.dr));
addr_b = fold_build_pointer_plus (addr_b, DR_INIT (dr_b.dr));
/* Advance ADDR_A by one iteration and adjust the length to compensate. */
addr_a = fold_build_pointer_plus (addr_a, step);
tree seg_len_a_minus_step = fold_build2 (MINUS_EXPR, sizetype,
seg_len_a, step);
if (!CONSTANT_CLASS_P (seg_len_a_minus_step))
seg_len_a_minus_step = build1 (SAVE_EXPR, sizetype, seg_len_a_minus_step);
tree low_offset_a = fold_build3 (COND_EXPR, sizetype, neg_step,
seg_len_a_minus_step, size_zero_node);
if (!CONSTANT_CLASS_P (low_offset_a))
low_offset_a = build1 (SAVE_EXPR, sizetype, low_offset_a);
/* We could use COND_EXPR <neg_step, size_zero_node, seg_len_a_minus_step>,
but it's usually more efficient to reuse the LOW_OFFSET_A result. */
tree high_offset_a = fold_build2 (MINUS_EXPR, sizetype, seg_len_a_minus_step,
low_offset_a);
/* The amount added to addr_b - addr_a'. */
tree bias = fold_build2 (MINUS_EXPR, sizetype,
size_int (last_chunk_b), low_offset_a);
tree limit = fold_build2 (MINUS_EXPR, sizetype, high_offset_a, low_offset_a);
limit = fold_build2 (PLUS_EXPR, sizetype, limit,
size_int (last_chunk_a + last_chunk_b));
tree subject = fold_build2 (POINTER_DIFF_EXPR, ssizetype, addr_b, addr_a);
subject = fold_build2 (PLUS_EXPR, sizetype,
fold_convert (sizetype, subject), bias);
*cond_expr = fold_build2 (GT_EXPR, boolean_type_node, subject, limit);
if (dump_enabled_p ())
dump_printf (MSG_NOTE, "using an address-based WAR/WAW test\n");
return true;
}
/* If ALIGN is nonzero, set up *SEQ_MIN_OUT and *SEQ_MAX_OUT so that for
every address ADDR accessed by D:
*SEQ_MIN_OUT <= ADDR (== ADDR & -ALIGN) <= *SEQ_MAX_OUT
In this case, every element accessed by D is aligned to at least
ALIGN bytes.
If ALIGN is zero then instead set *SEG_MAX_OUT so that:
*SEQ_MIN_OUT <= ADDR < *SEQ_MAX_OUT. */
static void
get_segment_min_max (const dr_with_seg_len &d, tree *seg_min_out,
tree *seg_max_out, HOST_WIDE_INT align)
{
/* Each access has the following pattern:
<- |seg_len| ->
<--- A: -ve step --->
+-----+-------+-----+-------+-----+
| n-1 | ,.... | 0 | ..... | n-1 |
+-----+-------+-----+-------+-----+
<--- B: +ve step --->
<- |seg_len| ->
|
base address
where "n" is the number of scalar iterations covered by the segment.
(This should be VF for a particular pair if we know that both steps
are the same, otherwise it will be the full number of scalar loop
iterations.)
A is the range of bytes accessed when the step is negative,
B is the range when the step is positive.
If the access size is "access_size" bytes, the lowest addressed byte is:
base + (step < 0 ? seg_len : 0) [LB]
and the highest addressed byte is always below:
base + (step < 0 ? 0 : seg_len) + access_size [UB]
Thus:
LB <= ADDR < UB
If ALIGN is nonzero, all three values are aligned to at least ALIGN
bytes, so:
LB <= ADDR <= UB - ALIGN
where "- ALIGN" folds naturally with the "+ access_size" and often
cancels it out.
We don't try to simplify LB and UB beyond this (e.g. by using
MIN and MAX based on whether seg_len rather than the stride is
negative) because it is possible for the absolute size of the
segment to overflow the range of a ssize_t.
Keeping the pointer_plus outside of the cond_expr should allow
the cond_exprs to be shared with other alias checks. */
tree indicator = dr_direction_indicator (d.dr);
tree neg_step = fold_build2 (LT_EXPR, boolean_type_node,
fold_convert (ssizetype, indicator),
ssize_int (0));
tree addr_base = fold_build_pointer_plus (DR_BASE_ADDRESS (d.dr),
DR_OFFSET (d.dr));
addr_base = fold_build_pointer_plus (addr_base, DR_INIT (d.dr));
tree seg_len
= fold_convert (sizetype, rewrite_to_non_trapping_overflow (d.seg_len));
tree min_reach = fold_build3 (COND_EXPR, sizetype, neg_step,
seg_len, size_zero_node);
tree max_reach = fold_build3 (COND_EXPR, sizetype, neg_step,
size_zero_node, seg_len);
max_reach = fold_build2 (PLUS_EXPR, sizetype, max_reach,
size_int (d.access_size - align));
*seg_min_out = fold_build_pointer_plus (addr_base, min_reach);
*seg_max_out = fold_build_pointer_plus (addr_base, max_reach);
}
/* Generate a runtime condition that is true if ALIAS_PAIR is free of aliases,
storing the condition in *COND_EXPR. The fallback is to generate a
a test that the two accesses do not overlap:
end_a <= start_b || end_b <= start_a. */
static void
create_intersect_range_checks (class loop *loop, tree *cond_expr,
const dr_with_seg_len_pair_t &alias_pair)
{
const dr_with_seg_len& dr_a = alias_pair.first;
const dr_with_seg_len& dr_b = alias_pair.second;
*cond_expr = NULL_TREE;
if (create_intersect_range_checks_index (loop, cond_expr, alias_pair))
return;
if (create_ifn_alias_checks (cond_expr, alias_pair))
return;
if (create_waw_or_war_checks (cond_expr, alias_pair))
return;
unsigned HOST_WIDE_INT min_align;
tree_code cmp_code;
/* We don't have to check DR_ALIAS_MIXED_STEPS here, since both versions
are equivalent. This is just an optimization heuristic. */
if (TREE_CODE (DR_STEP (dr_a.dr)) == INTEGER_CST
&& TREE_CODE (DR_STEP (dr_b.dr)) == INTEGER_CST)
{
/* In this case adding access_size to seg_len is likely to give
a simple X * step, where X is either the number of scalar
iterations or the vectorization factor. We're better off
keeping that, rather than subtracting an alignment from it.
In this case the maximum values are exclusive and so there is
no alias if the maximum of one segment equals the minimum
of another. */
min_align = 0;
cmp_code = LE_EXPR;
}
else
{
/* Calculate the minimum alignment shared by all four pointers,
then arrange for this alignment to be subtracted from the
exclusive maximum values to get inclusive maximum values.
This "- min_align" is cumulative with a "+ access_size"
in the calculation of the maximum values. In the best
(and common) case, the two cancel each other out, leaving
us with an inclusive bound based only on seg_len. In the
worst case we're simply adding a smaller number than before.
Because the maximum values are inclusive, there is an alias
if the maximum value of one segment is equal to the minimum
value of the other. */
min_align = MIN (dr_a.align, dr_b.align);
cmp_code = LT_EXPR;
}
tree seg_a_min, seg_a_max, seg_b_min, seg_b_max;
get_segment_min_max (dr_a, &seg_a_min, &seg_a_max, min_align);
get_segment_min_max (dr_b, &seg_b_min, &seg_b_max, min_align);
*cond_expr
= fold_build2 (TRUTH_OR_EXPR, boolean_type_node,
fold_build2 (cmp_code, boolean_type_node, seg_a_max, seg_b_min),
fold_build2 (cmp_code, boolean_type_node, seg_b_max, seg_a_min));
if (dump_enabled_p ())
dump_printf (MSG_NOTE, "using an address-based overlap test\n");
}
/* Create a conditional expression that represents the run-time checks for
overlapping of address ranges represented by a list of data references
pairs passed in ALIAS_PAIRS. Data references are in LOOP. The returned
COND_EXPR is the conditional expression to be used in the if statement
that controls which version of the loop gets executed at runtime. */
void
create_runtime_alias_checks (class loop *loop,
const vec<dr_with_seg_len_pair_t> *alias_pairs,
tree * cond_expr)
{
tree part_cond_expr;
fold_defer_overflow_warnings ();
for (const dr_with_seg_len_pair_t &alias_pair : alias_pairs)
{
gcc_assert (alias_pair.flags);
if (dump_enabled_p ())
dump_printf (MSG_NOTE,
"create runtime check for data references %T and %T\n",
DR_REF (alias_pair.first.dr),
DR_REF (alias_pair.second.dr));
/* Create condition expression for each pair data references. */
create_intersect_range_checks (loop, &part_cond_expr, alias_pair);
if (*cond_expr)
*cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
*cond_expr, part_cond_expr);
else
*cond_expr = part_cond_expr;
}
fold_undefer_and_ignore_overflow_warnings ();
}
/* Check if OFFSET1 and OFFSET2 (DR_OFFSETs of some data-refs) are identical
expressions. */
static bool
dr_equal_offsets_p1 (tree offset1, tree offset2)
{
bool res;
STRIP_NOPS (offset1);
STRIP_NOPS (offset2);
if (offset1 == offset2)
return true;
if (TREE_CODE (offset1) != TREE_CODE (offset2)
|| (!BINARY_CLASS_P (offset1) && !UNARY_CLASS_P (offset1)))
return false;
res = dr_equal_offsets_p1 (TREE_OPERAND (offset1, 0),
TREE_OPERAND (offset2, 0));
if (!res || !BINARY_CLASS_P (offset1))
return res;
res = dr_equal_offsets_p1 (TREE_OPERAND (offset1, 1),
TREE_OPERAND (offset2, 1));
return res;
}
/* Check if DRA and DRB have equal offsets. */
bool
dr_equal_offsets_p (struct data_reference *dra,
struct data_reference *drb)
{
tree offset1, offset2;
offset1 = DR_OFFSET (dra);
offset2 = DR_OFFSET (drb);
return dr_equal_offsets_p1 (offset1, offset2);
}
/* Returns true if FNA == FNB. */
static bool
affine_function_equal_p (affine_fn fna, affine_fn fnb)
{
unsigned i, n = fna.length ();
if (n != fnb.length ())
return false;
for (i = 0; i < n; i++)
if (!operand_equal_p (fna[i], fnb[i], 0))
return false;
return true;
}
/* If all the functions in CF are the same, returns one of them,
otherwise returns NULL. */
static affine_fn
common_affine_function (conflict_function *cf)
{
unsigned i;
affine_fn comm;
if (!CF_NONTRIVIAL_P (cf))
return affine_fn ();
comm = cf->fns[0];
for (i = 1; i < cf->n; i++)
if (!affine_function_equal_p (comm, cf->fns[i]))
return affine_fn ();
return comm;
}
/* Returns the base of the affine function FN. */
static tree
affine_function_base (affine_fn fn)
{
return fn[0];
}
/* Returns true if FN is a constant. */
static bool
affine_function_constant_p (affine_fn fn)
{
unsigned i;
tree coef;
for (i = 1; fn.iterate (i, &coef); i++)
if (!integer_zerop (coef))
return false;
return true;
}
/* Returns true if FN is the zero constant function. */
static bool
affine_function_zero_p (affine_fn fn)
{
return (integer_zerop (affine_function_base (fn))
&& affine_function_constant_p (fn));
}
/* Returns a signed integer type with the largest precision from TA
and TB. */
static tree
signed_type_for_types (tree ta, tree tb)
{
if (TYPE_PRECISION (ta) > TYPE_PRECISION (tb))
return signed_type_for (ta);
else
return signed_type_for (tb);
}
/* Applies operation OP on affine functions FNA and FNB, and returns the
result. */
static affine_fn
affine_fn_op (enum tree_code op, affine_fn fna, affine_fn fnb)
{
unsigned i, n, m;
affine_fn ret;
tree coef;
if (fnb.length () > fna.length ())
{
n = fna.length ();
m = fnb.length ();
}
else
{
n = fnb.length ();
m = fna.length ();
}
ret.create (m);
for (i = 0; i < n; i++)
{
tree type = signed_type_for_types (TREE_TYPE (fna[i]),
TREE_TYPE (fnb[i]));
ret.quick_push (fold_build2 (op, type, fna[i], fnb[i]));
}
for (; fna.iterate (i, &coef); i++)
ret.quick_push (fold_build2 (op, signed_type_for (TREE_TYPE (coef)),
coef, integer_zero_node));
for (; fnb.iterate (i, &coef); i++)
ret.quick_push (fold_build2 (op, signed_type_for (TREE_TYPE (coef)),
integer_zero_node, coef));
return ret;
}
/* Returns the sum of affine functions FNA and FNB. */
static affine_fn
affine_fn_plus (affine_fn fna, affine_fn fnb)
{
return affine_fn_op (PLUS_EXPR, fna, fnb);
}
/* Returns the difference of affine functions FNA and FNB. */
static affine_fn
affine_fn_minus (affine_fn fna, affine_fn fnb)
{
return affine_fn_op (MINUS_EXPR, fna, fnb);
}
/* Frees affine function FN. */
static void
affine_fn_free (affine_fn fn)
{
fn.release ();
}
/* Determine for each subscript in the data dependence relation DDR
the distance. */
static void
compute_subscript_distance (struct data_dependence_relation *ddr)
{
conflict_function *cf_a, *cf_b;
affine_fn fn_a, fn_b, diff;
if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
{
unsigned int i;
for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
{