gcc/gcc/graphite-isl-ast-to-gimple.c

3361 lines
99 KiB
C

/* Translation of isl AST to Gimple.
Copyright (C) 2014-2016 Free Software Foundation, Inc.
Contributed by Roman Gareev <gareevroman@gmail.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/>. */
#define USES_ISL
#include "config.h"
#ifdef HAVE_isl
#include "system.h"
#include "coretypes.h"
#include "backend.h"
#include "cfghooks.h"
#include "tree.h"
#include "gimple.h"
#include "params.h"
#include "fold-const.h"
#include "gimple-fold.h"
#include "gimple-iterator.h"
#include "gimplify.h"
#include "gimplify-me.h"
#include "tree-eh.h"
#include "tree-ssa-loop.h"
#include "tree-ssa-operands.h"
#include "tree-ssa-propagate.h"
#include "tree-pass.h"
#include "cfgloop.h"
#include "tree-data-ref.h"
#include "tree-ssa-loop-manip.h"
#include "tree-scalar-evolution.h"
#include "gimple-ssa.h"
#include "tree-phinodes.h"
#include "tree-into-ssa.h"
#include "ssa-iterators.h"
#include "tree-cfg.h"
#include "gimple-pretty-print.h"
#include "cfganal.h"
#include "value-prof.h"
#include "graphite.h"
#include <map>
/* We always try to use signed 128 bit types, but fall back to smaller types
in case a platform does not provide types of these sizes. In the future we
should use isl to derive the optimal type for each subexpression. */
static int max_mode_int_precision =
GET_MODE_PRECISION (mode_for_size (MAX_FIXED_MODE_SIZE, MODE_INT, 0));
static int graphite_expression_type_precision = 128 <= max_mode_int_precision ?
128 : max_mode_int_precision;
struct ast_build_info
{
ast_build_info()
: is_parallelizable(false)
{ }
bool is_parallelizable;
};
/* Converts a GMP constant VAL to a tree and returns it. */
static tree
gmp_cst_to_tree (tree type, mpz_t val)
{
tree t = type ? type : integer_type_node;
mpz_t tmp;
mpz_init (tmp);
mpz_set (tmp, val);
wide_int wi = wi::from_mpz (t, tmp, true);
mpz_clear (tmp);
return wide_int_to_tree (t, wi);
}
/* Verifies properties that GRAPHITE should maintain during translation. */
static inline void
graphite_verify (void)
{
checking_verify_loop_structure ();
checking_verify_loop_closed_ssa (true);
}
/* IVS_PARAMS maps isl's scattering and parameter identifiers
to corresponding trees. */
typedef std::map<isl_id *, tree> ivs_params;
/* Free all memory allocated for isl's identifiers. */
void ivs_params_clear (ivs_params &ip)
{
std::map<isl_id *, tree>::iterator it;
for (it = ip.begin ();
it != ip.end (); it++)
{
isl_id_free (it->first);
}
}
#ifdef HAVE_ISL_OPTIONS_SET_SCHEDULE_SERIALIZE_SCCS
/* Set the "separate" option for the schedule node. */
static __isl_give isl_schedule_node *
set_separate_option (__isl_take isl_schedule_node *node, void *user)
{
if (user)
return node;
if (isl_schedule_node_get_type (node) != isl_schedule_node_band)
return node;
/* Set the "separate" option unless it is set earlier to another option. */
if (isl_schedule_node_band_member_get_ast_loop_type (node, 0)
== isl_ast_loop_default)
return isl_schedule_node_band_member_set_ast_loop_type
(node, 0, isl_ast_loop_separate);
return node;
}
#endif
class translate_isl_ast_to_gimple
{
public:
translate_isl_ast_to_gimple (sese_info_p r)
: region (r), codegen_error (false)
{ }
/* Translates an isl AST node NODE to GCC representation in the
context of a SESE. */
edge translate_isl_ast (loop_p context_loop, __isl_keep isl_ast_node *node,
edge next_e, ivs_params &ip);
/* Translates an isl_ast_node_for to Gimple. */
edge translate_isl_ast_node_for (loop_p context_loop,
__isl_keep isl_ast_node *node,
edge next_e, ivs_params &ip);
/* Create the loop for a isl_ast_node_for.
- NEXT_E is the edge where new generated code should be attached. */
edge translate_isl_ast_for_loop (loop_p context_loop,
__isl_keep isl_ast_node *node_for,
edge next_e,
tree type, tree lb, tree ub,
ivs_params &ip);
/* Translates an isl_ast_node_if to Gimple. */
edge translate_isl_ast_node_if (loop_p context_loop,
__isl_keep isl_ast_node *node,
edge next_e, ivs_params &ip);
/* Translates an isl_ast_node_user to Gimple.
FIXME: We should remove iv_map.create (loop->num + 1), if it is
possible. */
edge translate_isl_ast_node_user (__isl_keep isl_ast_node *node,
edge next_e, ivs_params &ip);
/* Translates an isl_ast_node_block to Gimple. */
edge translate_isl_ast_node_block (loop_p context_loop,
__isl_keep isl_ast_node *node,
edge next_e, ivs_params &ip);
/* Converts a unary isl_ast_expr_op expression E to a GCC expression tree of
type TYPE. */
tree unary_op_to_tree (tree type, __isl_take isl_ast_expr *expr,
ivs_params &ip);
/* Converts a binary isl_ast_expr_op expression E to a GCC expression tree of
type TYPE. */
tree binary_op_to_tree (tree type, __isl_take isl_ast_expr *expr,
ivs_params &ip);
/* Converts a ternary isl_ast_expr_op expression E to a GCC expression tree of
type TYPE. */
tree ternary_op_to_tree (tree type, __isl_take isl_ast_expr *expr,
ivs_params &ip);
/* Converts an isl_ast_expr_op expression E with unknown number of arguments
to a GCC expression tree of type TYPE. */
tree nary_op_to_tree (tree type, __isl_take isl_ast_expr *expr,
ivs_params &ip);
/* Converts an isl AST expression E back to a GCC expression tree of
type TYPE. */
tree gcc_expression_from_isl_expression (tree type,
__isl_take isl_ast_expr *,
ivs_params &ip);
/* Return the tree variable that corresponds to the given isl ast identifier
expression (an isl_ast_expr of type isl_ast_expr_id).
FIXME: We should replace blind conversation of id's type with derivation
of the optimal type when we get the corresponding isl support. Blindly
converting type sizes may be problematic when we switch to smaller
types. */
tree gcc_expression_from_isl_ast_expr_id (tree type,
__isl_keep isl_ast_expr *expr_id,
ivs_params &ip);
/* Converts an isl_ast_expr_int expression E to a GCC expression tree of
type TYPE. */
tree gcc_expression_from_isl_expr_int (tree type,
__isl_take isl_ast_expr *expr);
/* Converts an isl_ast_expr_op expression E to a GCC expression tree of
type TYPE. */
tree gcc_expression_from_isl_expr_op (tree type,
__isl_take isl_ast_expr *expr,
ivs_params &ip);
/* Creates a new LOOP corresponding to isl_ast_node_for. Inserts an
induction variable for the new LOOP. New LOOP is attached to CFG
starting at ENTRY_EDGE. LOOP is inserted into the loop tree and
becomes the child loop of the OUTER_LOOP. NEWIVS_INDEX binds
isl's scattering name to the induction variable created for the
loop of STMT. The new induction variable is inserted in the NEWIVS
vector and is of type TYPE. */
struct loop *graphite_create_new_loop (edge entry_edge,
__isl_keep isl_ast_node *node_for,
loop_p outer, tree type,
tree lb, tree ub, ivs_params &ip);
/* All loops generated by create_empty_loop_on_edge have the form of
a post-test loop:
do
{
body of the loop;
} while (lower bound < upper bound);
We create a new if region protecting the loop to be executed, if
the execution count is zero (lower bound > upper bound). */
edge graphite_create_new_loop_guard (edge entry_edge,
__isl_keep isl_ast_node *node_for,
tree *type,
tree *lb, tree *ub, ivs_params &ip);
/* Creates a new if region corresponding to isl's cond. */
edge graphite_create_new_guard (edge entry_edge,
__isl_take isl_ast_expr *if_cond,
ivs_params &ip);
/* Inserts in iv_map a tuple (OLD_LOOP->num, NEW_NAME) for the induction
variables of the loops around GBB in SESE.
FIXME: Instead of using a vec<tree> that maps each loop id to a possible
chrec, we could consider using a map<int, tree> that maps loop ids to the
corresponding tree expressions. */
void build_iv_mapping (vec<tree> iv_map, gimple_poly_bb_p gbb,
__isl_keep isl_ast_expr *user_expr, ivs_params &ip,
sese_l &region);
/* Patch the missing arguments of the phi nodes. */
void translate_pending_phi_nodes (void);
/* Add isl's parameter identifiers and corresponding trees to ivs_params. */
void add_parameters_to_ivs_params (scop_p scop, ivs_params &ip);
/* Get the maximal number of schedule dimensions in the scop SCOP. */
int get_max_schedule_dimensions (scop_p scop);
/* Generates a build, which specifies the constraints on the parameters. */
__isl_give isl_ast_build *generate_isl_context (scop_p scop);
/* Extend the schedule to NB_SCHEDULE_DIMS schedule dimensions.
For schedules with different dimensionality, the isl AST generator can not
define an order and will just randomly choose an order. The solution to
this problem is to extend all schedules to the maximal number of schedule
dimensions (using '0's for the remaining values). */
__isl_give isl_map *extend_schedule (__isl_take isl_map *schedule,
int nb_schedule_dims);
/* Generates a schedule, which specifies an order used to
visit elements in a domain. */
__isl_give isl_union_map *generate_isl_schedule (scop_p scop);
#ifdef HAVE_ISL_OPTIONS_SET_SCHEDULE_SERIALIZE_SCCS
/* Set the "separate" option for all schedules. This helps reducing control
overhead. */
__isl_give isl_schedule *
set_options_for_schedule_tree (__isl_take isl_schedule *schedule);
#endif
/* Set the separate option for all dimensions.
This helps to reduce control overhead. */
__isl_give isl_ast_build * set_options (__isl_take isl_ast_build *control,
__isl_keep isl_union_map *schedule);
/* Generate isl AST from schedule of SCOP. Also, collects IVS_PARAMS in
IP. */
__isl_give isl_ast_node * scop_to_isl_ast (scop_p scop, ivs_params &ip);
/* Return true if RENAME (defined in BB) is a valid use in NEW_BB. The
definition should flow into use, and the use should respect the loop-closed
SSA form. */
bool is_valid_rename (tree rename, basic_block def_bb, basic_block use_bb,
bool loop_phi, tree old_name, basic_block old_bb) const;
/* Returns the expression associated to OLD_NAME (which is used in OLD_BB), in
NEW_BB from RENAME_MAP. LOOP_PHI is true when we want to rename OLD_NAME
within a loop PHI instruction. */
tree get_rename (basic_block new_bb, tree old_name,
basic_block old_bb, bool loop_phi) const;
/* For ops which are scev_analyzeable, we can regenerate a new name from
its scalar evolution around LOOP. */
tree get_rename_from_scev (tree old_name, gimple_seq *stmts, loop_p loop,
basic_block new_bb, basic_block old_bb,
vec<tree> iv_map);
/* Returns a basic block that could correspond to where a constant was defined
in the original code. In the original code OLD_BB had the definition, we
need to find which basic block out of the copies of old_bb, in the new
region, should a definition correspond to if it has to reach BB. */
basic_block get_def_bb_for_const (basic_block bb, basic_block old_bb) const;
/* Get the new name of OP (from OLD_BB) to be used in NEW_BB. LOOP_PHI is
true when we want to rename an OP within a loop PHI instruction. */
tree get_new_name (basic_block new_bb, tree op,
basic_block old_bb, bool loop_phi) const;
/* Collect all the operands of NEW_EXPR by recursively visiting each
operand. */
void collect_all_ssa_names (tree new_expr, vec<tree> *vec_ssa);
/* Copy the PHI arguments from OLD_PHI to the NEW_PHI. The arguments to
NEW_PHI must be found unless they can be POSTPONEd for later. */
bool copy_loop_phi_args (gphi *old_phi, init_back_edge_pair_t &ibp_old_bb,
gphi *new_phi, init_back_edge_pair_t &ibp_new_bb,
bool postpone);
/* Copy loop phi nodes from BB to NEW_BB. */
bool copy_loop_phi_nodes (basic_block bb, basic_block new_bb);
/* Add phi nodes to all merge points of all the diamonds enclosing the loop of
the close phi node PHI. */
bool add_close_phis_to_merge_points (gphi *old_phi, gphi *new_phi,
tree default_value);
tree add_close_phis_to_outer_loops (tree last_merge_name, edge merge_e,
gimple *old_close_phi);
/* Copy all the loop-close phi args from BB to NEW_BB. */
bool copy_loop_close_phi_args (basic_block old_bb, basic_block new_bb,
bool postpone);
/* Copy loop close phi nodes from BB to NEW_BB. */
bool copy_loop_close_phi_nodes (basic_block old_bb, basic_block new_bb);
/* Copy the arguments of cond-phi node PHI, to NEW_PHI in the codegenerated
region. If postpone is true and it isn't possible to copy any arg of PHI,
the PHI is added to the REGION->INCOMPLETE_PHIS to be codegenerated later.
Returns false if the copying was unsuccessful. */
bool copy_cond_phi_args (gphi *phi, gphi *new_phi, vec<tree> iv_map,
bool postpone);
/* Copy cond phi nodes from BB to NEW_BB. A cond-phi node is a basic block
containing phi nodes coming from two predecessors, and none of them are back
edges. */
bool copy_cond_phi_nodes (basic_block bb, basic_block new_bb,
vec<tree> iv_map);
/* Duplicates the statements of basic block BB into basic block NEW_BB
and compute the new induction variables according to the IV_MAP.
CODEGEN_ERROR is set when the code generation cannot continue. */
bool graphite_copy_stmts_from_block (basic_block bb, basic_block new_bb,
vec<tree> iv_map);
/* Copies BB and includes in the copied BB all the statements that can
be reached following the use-def chains from the memory accesses,
and returns the next edge following this new block. codegen_error is
set when the code generation cannot continue. */
edge copy_bb_and_scalar_dependences (basic_block bb, edge next_e,
vec<tree> iv_map);
/* Given a basic block containing close-phi it returns the new basic block
where to insert a copy of the close-phi nodes. All the uses in close phis
should come from a single loop otherwise it returns NULL. */
edge edge_for_new_close_phis (basic_block bb);
/* Add NEW_NAME as the ARGNUM-th arg of NEW_PHI which is in NEW_BB.
DOMINATING_PRED is the predecessor basic block of OLD_BB which dominates
the other pred of OLD_BB as well. If no such basic block exists then it is
NULL. NON_DOMINATING_PRED is a pred which does not dominate OLD_BB, it
cannot be NULL.
Case1: OLD_BB->preds {BB1, BB2} and BB1 does not dominate BB2 and vice
versa. In this case DOMINATING_PRED = NULL.
Case2: OLD_BB->preds {BB1, BB2} and BB1 dominates BB2.
Returns true on successful copy of the args, false otherwise. */
bool add_phi_arg_for_new_expr (tree old_phi_args[2], tree new_phi_args[2],
edge old_bb_dominating_edge,
edge old_bb_non_dominating_edge,
gphi *phi, gphi *new_phi,
basic_block new_bb);
/* Renames the scalar uses of the statement COPY, using the substitution map
RENAME_MAP, inserting the gimplification code at GSI_TGT, for the
translation REGION, with the original copied statement in LOOP, and using
the induction variable renaming map IV_MAP. Returns true when something
has been renamed. codegen_error is set when the code generation cannot
continue. */
bool rename_uses (gimple *copy, gimple_stmt_iterator *gsi_tgt,
basic_block old_bb, loop_p loop, vec<tree> iv_map);
/* Register in RENAME_MAP the rename tuple (OLD_NAME, EXPR).
When OLD_NAME and EXPR are the same we assert. */
void set_rename (tree old_name, tree expr);
/* Create new names for all the definitions created by COPY and add
replacement mappings for each new name. */
void set_rename_for_each_def (gimple *stmt);
/* Insert each statement from SEQ at its earliest insertion p. */
void gsi_insert_earliest (gimple_seq seq);
/* Rename all the operands of NEW_EXPR by recursively visiting each
operand. */
tree rename_all_uses (tree new_expr, basic_block new_bb, basic_block old_bb);
bool codegen_error_p () const
{ return codegen_error; }
/* Prints NODE to FILE. */
void print_isl_ast_node (FILE *file, __isl_keep isl_ast_node *node,
__isl_keep isl_ctx *ctx) const;
/* Return true when OP is a constant tree. */
bool is_constant (tree op) const
{
return TREE_CODE (op) == INTEGER_CST
|| TREE_CODE (op) == REAL_CST
|| TREE_CODE (op) == COMPLEX_CST
|| TREE_CODE (op) == VECTOR_CST;
}
private:
/* The region to be translated. */
sese_info_p region;
/* This flag is set when an error occurred during the translation of isl AST
to Gimple. */
bool codegen_error;
/* A vector of all the edges at if_condition merge points. */
auto_vec<edge, 2> merge_points;
};
/* Return the tree variable that corresponds to the given isl ast identifier
expression (an isl_ast_expr of type isl_ast_expr_id).
FIXME: We should replace blind conversation of id's type with derivation
of the optimal type when we get the corresponding isl support. Blindly
converting type sizes may be problematic when we switch to smaller
types. */
tree
translate_isl_ast_to_gimple::
gcc_expression_from_isl_ast_expr_id (tree type,
__isl_take isl_ast_expr *expr_id,
ivs_params &ip)
{
gcc_assert (isl_ast_expr_get_type (expr_id) == isl_ast_expr_id);
isl_id *tmp_isl_id = isl_ast_expr_get_id (expr_id);
std::map<isl_id *, tree>::iterator res;
res = ip.find (tmp_isl_id);
isl_id_free (tmp_isl_id);
gcc_assert (res != ip.end () &&
"Could not map isl_id to tree expression");
isl_ast_expr_free (expr_id);
tree t = res->second;
return fold_convert (type, t);
}
/* Converts an isl_ast_expr_int expression E to a GCC expression tree of
type TYPE. */
tree
translate_isl_ast_to_gimple::
gcc_expression_from_isl_expr_int (tree type, __isl_take isl_ast_expr *expr)
{
gcc_assert (isl_ast_expr_get_type (expr) == isl_ast_expr_int);
isl_val *val = isl_ast_expr_get_val (expr);
mpz_t val_mpz_t;
mpz_init (val_mpz_t);
tree res;
if (isl_val_get_num_gmp (val, val_mpz_t) == -1)
res = NULL_TREE;
else
res = gmp_cst_to_tree (type, val_mpz_t);
isl_val_free (val);
isl_ast_expr_free (expr);
mpz_clear (val_mpz_t);
return res;
}
/* Converts a binary isl_ast_expr_op expression E to a GCC expression tree of
type TYPE. */
tree
translate_isl_ast_to_gimple::
binary_op_to_tree (tree type, __isl_take isl_ast_expr *expr, ivs_params &ip)
{
isl_ast_expr *arg_expr = isl_ast_expr_get_op_arg (expr, 0);
tree tree_lhs_expr = gcc_expression_from_isl_expression (type, arg_expr, ip);
arg_expr = isl_ast_expr_get_op_arg (expr, 1);
tree tree_rhs_expr = gcc_expression_from_isl_expression (type, arg_expr, ip);
enum isl_ast_op_type expr_type = isl_ast_expr_get_op_type (expr);
isl_ast_expr_free (expr);
if (codegen_error)
return NULL_TREE;
switch (expr_type)
{
case isl_ast_op_add:
return fold_build2 (PLUS_EXPR, type, tree_lhs_expr, tree_rhs_expr);
case isl_ast_op_sub:
return fold_build2 (MINUS_EXPR, type, tree_lhs_expr, tree_rhs_expr);
case isl_ast_op_mul:
return fold_build2 (MULT_EXPR, type, tree_lhs_expr, tree_rhs_expr);
case isl_ast_op_div:
/* As isl operates on arbitrary precision numbers, we may end up with
division by 2^64 that is folded to 0. */
if (integer_zerop (tree_rhs_expr))
{
codegen_error = true;
return NULL_TREE;
}
return fold_build2 (EXACT_DIV_EXPR, type, tree_lhs_expr, tree_rhs_expr);
case isl_ast_op_pdiv_q:
/* As isl operates on arbitrary precision numbers, we may end up with
division by 2^64 that is folded to 0. */
if (integer_zerop (tree_rhs_expr))
{
codegen_error = true;
return NULL_TREE;
}
return fold_build2 (TRUNC_DIV_EXPR, type, tree_lhs_expr, tree_rhs_expr);
#if HAVE_ISL_OPTIONS_SET_SCHEDULE_SERIALIZE_SCCS
/* isl 0.15 or later. */
case isl_ast_op_zdiv_r:
#endif
case isl_ast_op_pdiv_r:
/* As isl operates on arbitrary precision numbers, we may end up with
division by 2^64 that is folded to 0. */
if (integer_zerop (tree_rhs_expr))
{
codegen_error = true;
return NULL_TREE;
}
return fold_build2 (TRUNC_MOD_EXPR, type, tree_lhs_expr, tree_rhs_expr);
case isl_ast_op_fdiv_q:
/* As isl operates on arbitrary precision numbers, we may end up with
division by 2^64 that is folded to 0. */
if (integer_zerop (tree_rhs_expr))
{
codegen_error = true;
return NULL_TREE;
}
return fold_build2 (FLOOR_DIV_EXPR, type, tree_lhs_expr, tree_rhs_expr);
case isl_ast_op_and:
return fold_build2 (TRUTH_ANDIF_EXPR, type,
tree_lhs_expr, tree_rhs_expr);
case isl_ast_op_or:
return fold_build2 (TRUTH_ORIF_EXPR, type, tree_lhs_expr, tree_rhs_expr);
case isl_ast_op_eq:
return fold_build2 (EQ_EXPR, type, tree_lhs_expr, tree_rhs_expr);
case isl_ast_op_le:
return fold_build2 (LE_EXPR, type, tree_lhs_expr, tree_rhs_expr);
case isl_ast_op_lt:
return fold_build2 (LT_EXPR, type, tree_lhs_expr, tree_rhs_expr);
case isl_ast_op_ge:
return fold_build2 (GE_EXPR, type, tree_lhs_expr, tree_rhs_expr);
case isl_ast_op_gt:
return fold_build2 (GT_EXPR, type, tree_lhs_expr, tree_rhs_expr);
default:
gcc_unreachable ();
}
}
/* Converts a ternary isl_ast_expr_op expression E to a GCC expression tree of
type TYPE. */
tree
translate_isl_ast_to_gimple::
ternary_op_to_tree (tree type, __isl_take isl_ast_expr *expr, ivs_params &ip)
{
gcc_assert (isl_ast_expr_get_op_type (expr) == isl_ast_op_minus);
isl_ast_expr *arg_expr = isl_ast_expr_get_op_arg (expr, 0);
tree tree_first_expr
= gcc_expression_from_isl_expression (type, arg_expr, ip);
arg_expr = isl_ast_expr_get_op_arg (expr, 1);
tree tree_second_expr
= gcc_expression_from_isl_expression (type, arg_expr, ip);
arg_expr = isl_ast_expr_get_op_arg (expr, 2);
tree tree_third_expr
= gcc_expression_from_isl_expression (type, arg_expr, ip);
isl_ast_expr_free (expr);
if (codegen_error)
return NULL_TREE;
return fold_build3 (COND_EXPR, type, tree_first_expr,
tree_second_expr, tree_third_expr);
}
/* Converts a unary isl_ast_expr_op expression E to a GCC expression tree of
type TYPE. */
tree
translate_isl_ast_to_gimple::
unary_op_to_tree (tree type, __isl_take isl_ast_expr *expr, ivs_params &ip)
{
gcc_assert (isl_ast_expr_get_op_type (expr) == isl_ast_op_minus);
isl_ast_expr *arg_expr = isl_ast_expr_get_op_arg (expr, 0);
tree tree_expr = gcc_expression_from_isl_expression (type, arg_expr, ip);
isl_ast_expr_free (expr);
return codegen_error ? NULL_TREE : fold_build1 (NEGATE_EXPR, type, tree_expr);
}
/* Converts an isl_ast_expr_op expression E with unknown number of arguments
to a GCC expression tree of type TYPE. */
tree
translate_isl_ast_to_gimple::
nary_op_to_tree (tree type, __isl_take isl_ast_expr *expr, ivs_params &ip)
{
enum tree_code op_code;
switch (isl_ast_expr_get_op_type (expr))
{
case isl_ast_op_max:
op_code = MAX_EXPR;
break;
case isl_ast_op_min:
op_code = MIN_EXPR;
break;
default:
gcc_unreachable ();
}
isl_ast_expr *arg_expr = isl_ast_expr_get_op_arg (expr, 0);
tree res = gcc_expression_from_isl_expression (type, arg_expr, ip);
if (codegen_error)
{
isl_ast_expr_free (expr);
return NULL_TREE;
}
int i;
for (i = 1; i < isl_ast_expr_get_op_n_arg (expr); i++)
{
arg_expr = isl_ast_expr_get_op_arg (expr, i);
tree t = gcc_expression_from_isl_expression (type, arg_expr, ip);
if (codegen_error)
{
isl_ast_expr_free (expr);
return NULL_TREE;
}
res = fold_build2 (op_code, type, res, t);
}
isl_ast_expr_free (expr);
return res;
}
/* Converts an isl_ast_expr_op expression E to a GCC expression tree of
type TYPE. */
tree
translate_isl_ast_to_gimple::
gcc_expression_from_isl_expr_op (tree type, __isl_take isl_ast_expr *expr,
ivs_params &ip)
{
if (codegen_error)
{
isl_ast_expr_free (expr);
return NULL_TREE;
}
gcc_assert (isl_ast_expr_get_type (expr) == isl_ast_expr_op);
switch (isl_ast_expr_get_op_type (expr))
{
/* These isl ast expressions are not supported yet. */
case isl_ast_op_error:
case isl_ast_op_call:
case isl_ast_op_and_then:
case isl_ast_op_or_else:
case isl_ast_op_select:
gcc_unreachable ();
case isl_ast_op_max:
case isl_ast_op_min:
return nary_op_to_tree (type, expr, ip);
case isl_ast_op_add:
case isl_ast_op_sub:
case isl_ast_op_mul:
case isl_ast_op_div:
case isl_ast_op_pdiv_q:
case isl_ast_op_pdiv_r:
case isl_ast_op_fdiv_q:
#if HAVE_ISL_OPTIONS_SET_SCHEDULE_SERIALIZE_SCCS
/* isl 0.15 or later. */
case isl_ast_op_zdiv_r:
#endif
case isl_ast_op_and:
case isl_ast_op_or:
case isl_ast_op_eq:
case isl_ast_op_le:
case isl_ast_op_lt:
case isl_ast_op_ge:
case isl_ast_op_gt:
return binary_op_to_tree (type, expr, ip);
case isl_ast_op_minus:
return unary_op_to_tree (type, expr, ip);
case isl_ast_op_cond:
return ternary_op_to_tree (type, expr, ip);
default:
gcc_unreachable ();
}
return NULL_TREE;
}
/* Converts an isl AST expression E back to a GCC expression tree of
type TYPE. */
tree
translate_isl_ast_to_gimple::
gcc_expression_from_isl_expression (tree type, __isl_take isl_ast_expr *expr,
ivs_params &ip)
{
if (codegen_error)
{
isl_ast_expr_free (expr);
return NULL_TREE;
}
switch (isl_ast_expr_get_type (expr))
{
case isl_ast_expr_id:
return gcc_expression_from_isl_ast_expr_id (type, expr, ip);
case isl_ast_expr_int:
return gcc_expression_from_isl_expr_int (type, expr);
case isl_ast_expr_op:
return gcc_expression_from_isl_expr_op (type, expr, ip);
default:
gcc_unreachable ();
}
return NULL_TREE;
}
/* Creates a new LOOP corresponding to isl_ast_node_for. Inserts an
induction variable for the new LOOP. New LOOP is attached to CFG
starting at ENTRY_EDGE. LOOP is inserted into the loop tree and
becomes the child loop of the OUTER_LOOP. NEWIVS_INDEX binds
isl's scattering name to the induction variable created for the
loop of STMT. The new induction variable is inserted in the NEWIVS
vector and is of type TYPE. */
struct loop *
translate_isl_ast_to_gimple::
graphite_create_new_loop (edge entry_edge, __isl_keep isl_ast_node *node_for,
loop_p outer, tree type, tree lb, tree ub,
ivs_params &ip)
{
isl_ast_expr *for_inc = isl_ast_node_for_get_inc (node_for);
tree stride = gcc_expression_from_isl_expression (type, for_inc, ip);
/* To fail code generation, we generate wrong code until we discard it. */
if (codegen_error)
stride = integer_zero_node;
tree ivvar = create_tmp_var (type, "graphite_IV");
tree iv, iv_after_increment;
loop_p loop = create_empty_loop_on_edge
(entry_edge, lb, stride, ub, ivvar, &iv, &iv_after_increment,
outer ? outer : entry_edge->src->loop_father);
isl_ast_expr *for_iterator = isl_ast_node_for_get_iterator (node_for);
isl_id *id = isl_ast_expr_get_id (for_iterator);
std::map<isl_id *, tree>::iterator res;
res = ip.find (id);
if (ip.count (id))
isl_id_free (res->first);
ip[id] = iv;
isl_ast_expr_free (for_iterator);
return loop;
}
/* Create the loop for a isl_ast_node_for.
- NEXT_E is the edge where new generated code should be attached. */
edge
translate_isl_ast_to_gimple::
translate_isl_ast_for_loop (loop_p context_loop,
__isl_keep isl_ast_node *node_for, edge next_e,
tree type, tree lb, tree ub,
ivs_params &ip)
{
gcc_assert (isl_ast_node_get_type (node_for) == isl_ast_node_for);
struct loop *loop = graphite_create_new_loop (next_e, node_for, context_loop,
type, lb, ub, ip);
edge last_e = single_exit (loop);
edge to_body = single_succ_edge (loop->header);
basic_block after = to_body->dest;
/* Translate the body of the loop. */
isl_ast_node *for_body = isl_ast_node_for_get_body (node_for);
next_e = translate_isl_ast (loop, for_body, to_body, ip);
isl_ast_node_free (for_body);
/* Early return if we failed to translate loop body. */
if (!next_e || codegen_error_p ())
return NULL;
if (next_e->dest != after)
redirect_edge_succ_nodup (next_e, after);
set_immediate_dominator (CDI_DOMINATORS, next_e->dest, next_e->src);
if (flag_loop_parallelize_all)
{
isl_id *id = isl_ast_node_get_annotation (node_for);
gcc_assert (id);
ast_build_info *for_info = (ast_build_info *) isl_id_get_user (id);
loop->can_be_parallel = for_info->is_parallelizable;
free (for_info);
isl_id_free (id);
}
return last_e;
}
/* We use this function to get the upper bound because of the form,
which is used by isl to represent loops:
for (iterator = init; cond; iterator += inc)
{
...
}
The loop condition is an arbitrary expression, which contains the
current loop iterator.
(e.g. iterator + 3 < B && C > iterator + A)
We have to know the upper bound of the iterator to generate a loop
in Gimple form. It can be obtained from the special representation
of the loop condition, which is generated by isl,
if the ast_build_atomic_upper_bound option is set. In this case,
isl generates a loop condition that consists of the current loop
iterator, + an operator (< or <=) and an expression not involving
the iterator, which is processed and returned by this function.
(e.g iterator <= upper-bound-expression-without-iterator) */
static __isl_give isl_ast_expr *
get_upper_bound (__isl_keep isl_ast_node *node_for)
{
gcc_assert (isl_ast_node_get_type (node_for) == isl_ast_node_for);
isl_ast_expr *for_cond = isl_ast_node_for_get_cond (node_for);
gcc_assert (isl_ast_expr_get_type (for_cond) == isl_ast_expr_op);
isl_ast_expr *res;
switch (isl_ast_expr_get_op_type (for_cond))
{
case isl_ast_op_le:
res = isl_ast_expr_get_op_arg (for_cond, 1);
break;
case isl_ast_op_lt:
{
/* (iterator < ub) => (iterator <= ub - 1). */
isl_val *one =
isl_val_int_from_si (isl_ast_expr_get_ctx (for_cond), 1);
isl_ast_expr *ub = isl_ast_expr_get_op_arg (for_cond, 1);
res = isl_ast_expr_sub (ub, isl_ast_expr_from_val (one));
break;
}
default:
gcc_unreachable ();
}
isl_ast_expr_free (for_cond);
return res;
}
/* All loops generated by create_empty_loop_on_edge have the form of
a post-test loop:
do
{
body of the loop;
} while (lower bound < upper bound);
We create a new if region protecting the loop to be executed, if
the execution count is zero (lower bound > upper bound). */
edge
translate_isl_ast_to_gimple::
graphite_create_new_loop_guard (edge entry_edge,
__isl_keep isl_ast_node *node_for, tree *type,
tree *lb, tree *ub, ivs_params &ip)
{
gcc_assert (isl_ast_node_get_type (node_for) == isl_ast_node_for);
tree cond_expr;
edge exit_edge;
*type =
build_nonstandard_integer_type (graphite_expression_type_precision, 0);
isl_ast_expr *for_init = isl_ast_node_for_get_init (node_for);
*lb = gcc_expression_from_isl_expression (*type, for_init, ip);
/* To fail code generation, we generate wrong code until we discard it. */
if (codegen_error)
*lb = integer_zero_node;
isl_ast_expr *upper_bound = get_upper_bound (node_for);
*ub = gcc_expression_from_isl_expression (*type, upper_bound, ip);
/* To fail code generation, we generate wrong code until we discard it. */
if (codegen_error)
*ub = integer_zero_node;
/* When ub is simply a constant or a parameter, use lb <= ub. */
if (TREE_CODE (*ub) == INTEGER_CST || TREE_CODE (*ub) == SSA_NAME)
cond_expr = fold_build2 (LE_EXPR, boolean_type_node, *lb, *ub);
else
{
tree one = (POINTER_TYPE_P (*type)
? convert_to_ptrofftype (integer_one_node)
: fold_convert (*type, integer_one_node));
/* Adding +1 and using LT_EXPR helps with loop latches that have a
loop iteration count of "PARAMETER - 1". For PARAMETER == 0 this
becomes 2^k-1 due to integer overflow, and the condition lb <= ub
is true, even if we do not want this. However lb < ub + 1 is false,
as expected. */
tree ub_one = fold_build2 (POINTER_TYPE_P (*type) ? POINTER_PLUS_EXPR
: PLUS_EXPR, *type, *ub, one);
cond_expr = fold_build2 (LT_EXPR, boolean_type_node, *lb, ub_one);
}
if (integer_onep (cond_expr))
exit_edge = entry_edge;
else
exit_edge = create_empty_if_region_on_edge (entry_edge, cond_expr);
return exit_edge;
}
/* Translates an isl_ast_node_for to Gimple. */
edge
translate_isl_ast_to_gimple::
translate_isl_ast_node_for (loop_p context_loop, __isl_keep isl_ast_node *node,
edge next_e, ivs_params &ip)
{
gcc_assert (isl_ast_node_get_type (node) == isl_ast_node_for);
tree type, lb, ub;
edge last_e = graphite_create_new_loop_guard (next_e, node, &type,
&lb, &ub, ip);
if (last_e == next_e)
{
/* There was no guard generated. */
last_e = single_succ_edge (split_edge (last_e));
translate_isl_ast_for_loop (context_loop, node, next_e,
type, lb, ub, ip);
return last_e;
}
edge true_e = get_true_edge_from_guard_bb (next_e->dest);
merge_points.safe_push (last_e);
last_e = single_succ_edge (split_edge (last_e));
translate_isl_ast_for_loop (context_loop, node, true_e, type, lb, ub, ip);
return last_e;
}
/* Inserts in iv_map a tuple (OLD_LOOP->num, NEW_NAME) for the induction
variables of the loops around GBB in SESE.
FIXME: Instead of using a vec<tree> that maps each loop id to a possible
chrec, we could consider using a map<int, tree> that maps loop ids to the
corresponding tree expressions. */
void
translate_isl_ast_to_gimple::
build_iv_mapping (vec<tree> iv_map, gimple_poly_bb_p gbb,
__isl_keep isl_ast_expr *user_expr, ivs_params &ip,
sese_l &region)
{
gcc_assert (isl_ast_expr_get_type (user_expr) == isl_ast_expr_op &&
isl_ast_expr_get_op_type (user_expr) == isl_ast_op_call);
int i;
isl_ast_expr *arg_expr;
for (i = 1; i < isl_ast_expr_get_op_n_arg (user_expr); i++)
{
arg_expr = isl_ast_expr_get_op_arg (user_expr, i);
tree type =
build_nonstandard_integer_type (graphite_expression_type_precision, 0);
tree t = gcc_expression_from_isl_expression (type, arg_expr, ip);
/* To fail code generation, we generate wrong code until we discard it. */
if (codegen_error)
t = integer_zero_node;
loop_p old_loop = gbb_loop_at_index (gbb, region, i - 1);
iv_map[old_loop->num] = t;
}
}
/* Translates an isl_ast_node_user to Gimple.
FIXME: We should remove iv_map.create (loop->num + 1), if it is possible. */
edge
translate_isl_ast_to_gimple::
translate_isl_ast_node_user (__isl_keep isl_ast_node *node,
edge next_e, ivs_params &ip)
{
gcc_assert (isl_ast_node_get_type (node) == isl_ast_node_user);
isl_ast_expr *user_expr = isl_ast_node_user_get_expr (node);
isl_ast_expr *name_expr = isl_ast_expr_get_op_arg (user_expr, 0);
gcc_assert (isl_ast_expr_get_type (name_expr) == isl_ast_expr_id);
isl_id *name_id = isl_ast_expr_get_id (name_expr);
poly_bb_p pbb = (poly_bb_p) isl_id_get_user (name_id);
gcc_assert (pbb);
gimple_poly_bb_p gbb = PBB_BLACK_BOX (pbb);
isl_ast_expr_free (name_expr);
isl_id_free (name_id);
gcc_assert (GBB_BB (gbb) != ENTRY_BLOCK_PTR_FOR_FN (cfun) &&
"The entry block should not even appear within a scop");
const int nb_loops = number_of_loops (cfun);
vec<tree> iv_map;
iv_map.create (nb_loops);
iv_map.safe_grow_cleared (nb_loops);
build_iv_mapping (iv_map, gbb, user_expr, ip, pbb->scop->scop_info->region);
isl_ast_expr_free (user_expr);
basic_block old_bb = GBB_BB (gbb);
if (dump_file)
{
fprintf (dump_file,
"[codegen] copying from bb_%d on edge (bb_%d, bb_%d)\n",
old_bb->index, next_e->src->index, next_e->dest->index);
print_loops_bb (dump_file, GBB_BB (gbb), 0, 3);
}
next_e = copy_bb_and_scalar_dependences (old_bb, next_e, iv_map);
iv_map.release ();
if (codegen_error_p ())
return NULL;
if (dump_file)
{
fprintf (dump_file, "[codegen] (after copy) new basic block\n");
print_loops_bb (dump_file, next_e->src, 0, 3);
}
return next_e;
}
/* Translates an isl_ast_node_block to Gimple. */
edge
translate_isl_ast_to_gimple::
translate_isl_ast_node_block (loop_p context_loop,
__isl_keep isl_ast_node *node,
edge next_e, ivs_params &ip)
{
gcc_assert (isl_ast_node_get_type (node) == isl_ast_node_block);
isl_ast_node_list *node_list = isl_ast_node_block_get_children (node);
int i;
for (i = 0; i < isl_ast_node_list_n_ast_node (node_list); i++)
{
isl_ast_node *tmp_node = isl_ast_node_list_get_ast_node (node_list, i);
next_e = translate_isl_ast (context_loop, tmp_node, next_e, ip);
isl_ast_node_free (tmp_node);
}
isl_ast_node_list_free (node_list);
return next_e;
}
/* Creates a new if region corresponding to isl's cond. */
edge
translate_isl_ast_to_gimple::
graphite_create_new_guard (edge entry_edge, __isl_take isl_ast_expr *if_cond,
ivs_params &ip)
{
tree type =
build_nonstandard_integer_type (graphite_expression_type_precision, 0);
tree cond_expr = gcc_expression_from_isl_expression (type, if_cond, ip);
/* To fail code generation, we generate wrong code until we discard it. */
if (codegen_error)
cond_expr = integer_zero_node;
edge exit_edge = create_empty_if_region_on_edge (entry_edge, cond_expr);
return exit_edge;
}
/* Translates an isl_ast_node_if to Gimple. */
edge
translate_isl_ast_to_gimple::
translate_isl_ast_node_if (loop_p context_loop,
__isl_keep isl_ast_node *node,
edge next_e, ivs_params &ip)
{
gcc_assert (isl_ast_node_get_type (node) == isl_ast_node_if);
isl_ast_expr *if_cond = isl_ast_node_if_get_cond (node);
edge last_e = graphite_create_new_guard (next_e, if_cond, ip);
edge true_e = get_true_edge_from_guard_bb (next_e->dest);
merge_points.safe_push (last_e);
isl_ast_node *then_node = isl_ast_node_if_get_then (node);
translate_isl_ast (context_loop, then_node, true_e, ip);
isl_ast_node_free (then_node);
edge false_e = get_false_edge_from_guard_bb (next_e->dest);
isl_ast_node *else_node = isl_ast_node_if_get_else (node);
if (isl_ast_node_get_type (else_node) != isl_ast_node_error)
translate_isl_ast (context_loop, else_node, false_e, ip);
isl_ast_node_free (else_node);
return last_e;
}
/* Translates an isl AST node NODE to GCC representation in the
context of a SESE. */
edge
translate_isl_ast_to_gimple::translate_isl_ast (loop_p context_loop,
__isl_keep isl_ast_node *node,
edge next_e, ivs_params &ip)
{
if (codegen_error_p ())
return NULL;
switch (isl_ast_node_get_type (node))
{
case isl_ast_node_error:
gcc_unreachable ();
case isl_ast_node_for:
return translate_isl_ast_node_for (context_loop, node,
next_e, ip);
case isl_ast_node_if:
return translate_isl_ast_node_if (context_loop, node,
next_e, ip);
case isl_ast_node_user:
return translate_isl_ast_node_user (node, next_e, ip);
case isl_ast_node_block:
return translate_isl_ast_node_block (context_loop, node,
next_e, ip);
default:
gcc_unreachable ();
}
}
/* Return true when BB contains loop close phi nodes. A loop close phi node is
at the exit of loop which takes one argument that is the last value of the
variable being used out of the loop. */
static bool
bb_contains_loop_close_phi_nodes (basic_block bb)
{
return single_pred_p (bb)
&& bb->loop_father != single_pred_edge (bb)->src->loop_father;
}
/* Return true when BB contains loop phi nodes. A loop phi node is the loop
header containing phi nodes which has one init-edge and one back-edge. */
static bool
bb_contains_loop_phi_nodes (basic_block bb)
{
gcc_assert (EDGE_COUNT (bb->preds) <= 2);
if (bb->preds->length () == 1)
return false;
unsigned depth = loop_depth (bb->loop_father);
edge preds[2] = { (*bb->preds)[0], (*bb->preds)[1] };
if (depth > loop_depth (preds[0]->src->loop_father)
|| depth > loop_depth (preds[1]->src->loop_father))
return true;
/* When one of the edges correspond to the same loop father and other
doesn't. */
if (bb->loop_father != preds[0]->src->loop_father
&& bb->loop_father == preds[1]->src->loop_father)
return true;
if (bb->loop_father != preds[1]->src->loop_father
&& bb->loop_father == preds[0]->src->loop_father)
return true;
return false;
}
/* Check if USE is defined in a basic block from where the definition of USE can
propagate from all the paths. FIXME: Verify checks for virtual operands. */
static bool
is_loop_closed_ssa_use (basic_block bb, tree use)
{
if (TREE_CODE (use) != SSA_NAME || virtual_operand_p (use))
return true;
/* For close-phi nodes def always comes from a loop which has a back-edge. */
if (bb_contains_loop_close_phi_nodes (bb))
return true;
gimple *def = SSA_NAME_DEF_STMT (use);
basic_block def_bb = gimple_bb (def);
return (!def_bb
|| flow_bb_inside_loop_p (def_bb->loop_father, bb));
}
/* Return the number of phi nodes in BB. */
static int
number_of_phi_nodes (basic_block bb)
{
int num_phis = 0;
for (gphi_iterator psi = gsi_start_phis (bb); !gsi_end_p (psi);
gsi_next (&psi))
num_phis++;
return num_phis;
}
/* Returns true if BB uses name in one of its PHIs. */
static bool
phi_uses_name (basic_block bb, tree name)
{
for (gphi_iterator psi = gsi_start_phis (bb); !gsi_end_p (psi);
gsi_next (&psi))
{
gphi *phi = psi.phi ();
for (unsigned i = 0; i < gimple_phi_num_args (phi); i++)
{
tree use_arg = gimple_phi_arg_def (phi, i);
if (use_arg == name)
return true;
}
}
return false;
}
/* Return true if RENAME (defined in BB) is a valid use in NEW_BB. The
definition should flow into use, and the use should respect the loop-closed
SSA form. */
bool
translate_isl_ast_to_gimple::
is_valid_rename (tree rename, basic_block def_bb, basic_block use_bb,
bool loop_phi, tree old_name, basic_block old_bb) const
{
/* The def of the rename must either dominate the uses or come from a
back-edge. Also the def must respect the loop closed ssa form. */
if (!is_loop_closed_ssa_use (use_bb, rename))
{
if (dump_file)
{
fprintf (dump_file, "[codegen] rename not in loop closed ssa:");
print_generic_expr (dump_file, rename, 0);
fprintf (dump_file, "\n");
}
return false;
}
if (dominated_by_p (CDI_DOMINATORS, use_bb, def_bb))
return true;
if (bb_contains_loop_phi_nodes (use_bb) && loop_phi)
{
/* The loop-header dominates the loop-body. */
if (!dominated_by_p (CDI_DOMINATORS, def_bb, use_bb))
return false;
/* RENAME would be used in loop-phi. */
gcc_assert (number_of_phi_nodes (use_bb));
/* For definitions coming from back edges, we should check that
old_name is used in a loop PHI node.
FIXME: Verify if this is true. */
if (phi_uses_name (old_bb, old_name))
return true;
}
return false;
}
/* Returns the expression associated to OLD_NAME (which is used in OLD_BB), in
NEW_BB from RENAME_MAP. LOOP_PHI is true when we want to rename OLD_NAME
within a loop PHI instruction. */
tree
translate_isl_ast_to_gimple::get_rename (basic_block new_bb,
tree old_name,
basic_block old_bb,
bool loop_phi) const
{
gcc_assert (TREE_CODE (old_name) == SSA_NAME);
vec <tree> *renames = region->rename_map->get (old_name);
if (!renames || renames->is_empty ())
return NULL_TREE;
if (1 == renames->length ())
{
tree rename = (*renames)[0];
if (TREE_CODE (rename) == SSA_NAME)
{
basic_block bb = gimple_bb (SSA_NAME_DEF_STMT (rename));
if (is_valid_rename (rename, bb, new_bb, loop_phi, old_name, old_bb))
return rename;
return NULL_TREE;
}
if (is_constant (rename))
return rename;
return NULL_TREE;
}
/* More than one renames corresponding to the old_name. Find the rename for
which the definition flows into usage at new_bb. */
int i;
tree t1 = NULL_TREE, t2;
basic_block t1_bb = NULL;
FOR_EACH_VEC_ELT (*renames, i, t2)
{
basic_block t2_bb = gimple_bb (SSA_NAME_DEF_STMT (t2));
/* Defined in the same basic block as used. */
if (t2_bb == new_bb)
return t2;
/* NEW_BB and T2_BB are in two unrelated if-clauses. */
if (!dominated_by_p (CDI_DOMINATORS, new_bb, t2_bb))
continue;
/* Compute the nearest dominator. */
if (!t1 || dominated_by_p (CDI_DOMINATORS, t2_bb, t1_bb))
{
t1_bb = t2_bb;
t1 = t2;
}
}
return t1;
}
/* Register in RENAME_MAP the rename tuple (OLD_NAME, EXPR).
When OLD_NAME and EXPR are the same we assert. */
void
translate_isl_ast_to_gimple::set_rename (tree old_name, tree expr)
{
if (dump_file)
{
fprintf (dump_file, "[codegen] setting rename: old_name = ");
print_generic_expr (dump_file, old_name, 0);
fprintf (dump_file, ", new_name = ");
print_generic_expr (dump_file, expr, 0);
fprintf (dump_file, "\n");
}
if (old_name == expr)
return;
vec <tree> *renames = region->rename_map->get (old_name);
if (renames)
renames->safe_push (expr);
else
{
vec<tree> r;
r.create (2);
r.safe_push (expr);
region->rename_map->put (old_name, r);
}
}
/* Return an iterator to the instructions comes last in the execution order.
Either GSI1 and GSI2 should belong to the same basic block or one of their
respective basic blocks should dominate the other. */
gimple_stmt_iterator
later_of_the_two (gimple_stmt_iterator gsi1, gimple_stmt_iterator gsi2)
{
basic_block bb1 = gsi_bb (gsi1);
basic_block bb2 = gsi_bb (gsi2);
/* Find the iterator which is the latest. */
if (bb1 == bb2)
{
/* For empty basic blocks gsis point to the end of the sequence. Since
there is no operator== defined for gimple_stmt_iterator and for gsis
not pointing to a valid statement gsi_next would assert. */
gimple_stmt_iterator gsi = gsi1;
do {
if (gsi_stmt (gsi) == gsi_stmt (gsi2))
return gsi2;
gsi_next (&gsi);
} while (!gsi_end_p (gsi));
return gsi1;
}
/* Find the basic block closest to the basic block which defines stmt. */
if (dominated_by_p (CDI_DOMINATORS, bb1, bb2))
return gsi1;
gcc_assert (dominated_by_p (CDI_DOMINATORS, bb2, bb1));
return gsi2;
}
/* Insert each statement from SEQ at its earliest insertion p. */
void
translate_isl_ast_to_gimple::gsi_insert_earliest (gimple_seq seq)
{
update_modified_stmts (seq);
sese_l &codegen_region = region->if_region->true_region->region;
basic_block begin_bb = get_entry_bb (codegen_region);
/* Inserting the gimple statements in a vector because gimple_seq behave
in strage ways when inserting the stmts from it into different basic
blocks one at a time. */
auto_vec<gimple *, 3> stmts;
for (gimple_stmt_iterator gsi = gsi_start (seq); !gsi_end_p (gsi);
gsi_next (&gsi))
stmts.safe_push (gsi_stmt (gsi));
int i;
gimple *use_stmt;
FOR_EACH_VEC_ELT (stmts, i, use_stmt)
{
gcc_assert (gimple_code (use_stmt) != GIMPLE_PHI);
gimple_stmt_iterator gsi_def_stmt = gsi_start_bb_nondebug (begin_bb);
use_operand_p use_p;
ssa_op_iter op_iter;
FOR_EACH_SSA_USE_OPERAND (use_p, use_stmt, op_iter, SSA_OP_USE)
{
/* Iterator to the current def of use_p. For function parameters or
anything where def is not found, insert at the beginning of the
generated region. */
gimple_stmt_iterator gsi_stmt = gsi_def_stmt;
tree op = USE_FROM_PTR (use_p);
gimple *stmt = SSA_NAME_DEF_STMT (op);
if (stmt && (gimple_code (stmt) != GIMPLE_NOP))
gsi_stmt = gsi_for_stmt (stmt);
/* For region parameters, insert at the beginning of the generated
region. */
if (!bb_in_sese_p (gsi_bb (gsi_stmt), codegen_region))
gsi_stmt = gsi_def_stmt;
gsi_def_stmt = later_of_the_two (gsi_stmt, gsi_def_stmt);
}
if (!gsi_stmt (gsi_def_stmt))
{
gimple_stmt_iterator gsi = gsi_after_labels (gsi_bb (gsi_def_stmt));
gsi_insert_before (&gsi, use_stmt, GSI_NEW_STMT);
}
else if (gimple_code (gsi_stmt (gsi_def_stmt)) == GIMPLE_PHI)
{
gimple_stmt_iterator bsi
= gsi_start_bb_nondebug (gsi_bb (gsi_def_stmt));
/* Insert right after the PHI statements. */
gsi_insert_before (&bsi, use_stmt, GSI_NEW_STMT);
}
else
gsi_insert_after (&gsi_def_stmt, use_stmt, GSI_NEW_STMT);
if (dump_file)
{
fprintf (dump_file, "[codegen] inserting statement: ");
print_gimple_stmt (dump_file, use_stmt, 0, TDF_VOPS | TDF_MEMSYMS);
print_loops_bb (dump_file, gimple_bb (use_stmt), 0, 3);
}
}
}
/* Collect all the operands of NEW_EXPR by recursively visiting each
operand. */
void
translate_isl_ast_to_gimple::collect_all_ssa_names (tree new_expr,
vec<tree> *vec_ssa)
{
/* Rename all uses in new_expr. */
if (TREE_CODE (new_expr) == SSA_NAME)
{
vec_ssa->safe_push (new_expr);
return;
}
/* Iterate over SSA_NAMES in NEW_EXPR. */
for (int i = 0; i < (TREE_CODE_LENGTH (TREE_CODE (new_expr))); i++)
{
tree op = TREE_OPERAND (new_expr, i);
collect_all_ssa_names (op, vec_ssa);
}
}
/* This is abridged version of the function copied from:
tree.c:substitute_in_expr (tree exp, tree f, tree r). */
static tree
substitute_ssa_name (tree exp, tree f, tree r)
{
enum tree_code code = TREE_CODE (exp);
tree op0, op1, op2, op3;
tree new_tree;
/* We handle TREE_LIST and COMPONENT_REF separately. */
if (code == TREE_LIST)
{
op0 = substitute_ssa_name (TREE_CHAIN (exp), f, r);
op1 = substitute_ssa_name (TREE_VALUE (exp), f, r);
if (op0 == TREE_CHAIN (exp) && op1 == TREE_VALUE (exp))
return exp;
return tree_cons (TREE_PURPOSE (exp), op1, op0);
}
else if (code == COMPONENT_REF)
{
tree inner;
/* If this expression is getting a value from a PLACEHOLDER_EXPR
and it is the right field, replace it with R. */
for (inner = TREE_OPERAND (exp, 0);
REFERENCE_CLASS_P (inner);
inner = TREE_OPERAND (inner, 0))
;
/* The field. */
op1 = TREE_OPERAND (exp, 1);
if (TREE_CODE (inner) == PLACEHOLDER_EXPR && op1 == f)
return r;
/* If this expression hasn't been completed let, leave it alone. */
if (TREE_CODE (inner) == PLACEHOLDER_EXPR && !TREE_TYPE (inner))
return exp;
op0 = substitute_ssa_name (TREE_OPERAND (exp, 0), f, r);
if (op0 == TREE_OPERAND (exp, 0))
return exp;
new_tree
= fold_build3 (COMPONENT_REF, TREE_TYPE (exp), op0, op1, NULL_TREE);
}
else
switch (TREE_CODE_CLASS (code))
{
case tcc_constant:
return exp;
case tcc_declaration:
if (exp == f)
return r;
else
return exp;
case tcc_expression:
if (exp == f)
return r;
/* Fall through... */
case tcc_exceptional:
case tcc_unary:
case tcc_binary:
case tcc_comparison:
case tcc_reference:
switch (TREE_CODE_LENGTH (code))
{
case 0:
if (exp == f)
return r;
return exp;
case 1:
op0 = substitute_ssa_name (TREE_OPERAND (exp, 0), f, r);
if (op0 == TREE_OPERAND (exp, 0))
return exp;
new_tree = fold_build1 (code, TREE_TYPE (exp), op0);
break;
case 2:
op0 = substitute_ssa_name (TREE_OPERAND (exp, 0), f, r);
op1 = substitute_ssa_name (TREE_OPERAND (exp, 1), f, r);
if (op0 == TREE_OPERAND (exp, 0) && op1 == TREE_OPERAND (exp, 1))
return exp;
new_tree = fold_build2 (code, TREE_TYPE (exp), op0, op1);
break;
case 3:
op0 = substitute_ssa_name (TREE_OPERAND (exp, 0), f, r);
op1 = substitute_ssa_name (TREE_OPERAND (exp, 1), f, r);
op2 = substitute_ssa_name (TREE_OPERAND (exp, 2), f, r);
if (op0 == TREE_OPERAND (exp, 0) && op1 == TREE_OPERAND (exp, 1)
&& op2 == TREE_OPERAND (exp, 2))
return exp;
new_tree = fold_build3 (code, TREE_TYPE (exp), op0, op1, op2);
break;
case 4:
op0 = substitute_ssa_name (TREE_OPERAND (exp, 0), f, r);
op1 = substitute_ssa_name (TREE_OPERAND (exp, 1), f, r);
op2 = substitute_ssa_name (TREE_OPERAND (exp, 2), f, r);
op3 = substitute_ssa_name (TREE_OPERAND (exp, 3), f, r);
if (op0 == TREE_OPERAND (exp, 0) && op1 == TREE_OPERAND (exp, 1)
&& op2 == TREE_OPERAND (exp, 2)
&& op3 == TREE_OPERAND (exp, 3))
return exp;
new_tree
= fold (build4 (code, TREE_TYPE (exp), op0, op1, op2, op3));
break;
default:
gcc_unreachable ();
}
break;
case tcc_vl_exp:
default:
gcc_unreachable ();
}
TREE_READONLY (new_tree) |= TREE_READONLY (exp);
if (code == INDIRECT_REF || code == ARRAY_REF || code == ARRAY_RANGE_REF)
TREE_THIS_NOTRAP (new_tree) |= TREE_THIS_NOTRAP (exp);
return new_tree;
}
/* Rename all the operands of NEW_EXPR by recursively visiting each operand. */
tree
translate_isl_ast_to_gimple::rename_all_uses (tree new_expr, basic_block new_bb,
basic_block old_bb)
{
auto_vec<tree, 2> ssa_names;
collect_all_ssa_names (new_expr, &ssa_names);
tree t;
int i;
FOR_EACH_VEC_ELT (ssa_names, i, t)
if (tree r = get_rename (new_bb, t, old_bb, false))
new_expr = substitute_ssa_name (new_expr, t, r);
return new_expr;
}
/* For ops which are scev_analyzeable, we can regenerate a new name from its
scalar evolution around LOOP. */
tree
translate_isl_ast_to_gimple::
get_rename_from_scev (tree old_name, gimple_seq *stmts, loop_p loop,
basic_block new_bb, basic_block old_bb,
vec<tree> iv_map)
{
tree scev = scalar_evolution_in_region (region->region, loop, old_name);
/* At this point we should know the exact scev for each
scalar SSA_NAME used in the scop: all the other scalar
SSA_NAMEs should have been translated out of SSA using
arrays with one element. */
tree new_expr;
if (chrec_contains_undetermined (scev))
{
codegen_error = true;
return build_zero_cst (TREE_TYPE (old_name));
}
new_expr = chrec_apply_map (scev, iv_map);
/* The apply should produce an expression tree containing
the uses of the new induction variables. We should be
able to use new_expr instead of the old_name in the newly
generated loop nest. */
if (chrec_contains_undetermined (new_expr)
|| tree_contains_chrecs (new_expr, NULL))
{
codegen_error = true;
return build_zero_cst (TREE_TYPE (old_name));
}
/* We should check all the operands and all of them should dominate the use at
new_expr. */
if (TREE_CODE (new_expr) == SSA_NAME)
{
basic_block bb = gimple_bb (SSA_NAME_DEF_STMT (new_expr));
if (bb && !dominated_by_p (CDI_DOMINATORS, new_bb, bb))
{
codegen_error = true;
return build_zero_cst (TREE_TYPE (old_name));
}
}
new_expr = rename_all_uses (new_expr, new_bb, old_bb);
/* We check all the operands and all of them should dominate the use at
new_expr. */
auto_vec <tree, 2> new_ssa_names;
collect_all_ssa_names (new_expr, &new_ssa_names);
int i;
tree new_ssa_name;
FOR_EACH_VEC_ELT (new_ssa_names, i, new_ssa_name)
{
if (TREE_CODE (new_ssa_name) == SSA_NAME)
{
basic_block bb = gimple_bb (SSA_NAME_DEF_STMT (new_ssa_name));
if (bb && !dominated_by_p (CDI_DOMINATORS, new_bb, bb))
{
codegen_error = true;
return build_zero_cst (TREE_TYPE (old_name));
}
}
}
/* Replace the old_name with the new_expr. */
return force_gimple_operand (unshare_expr (new_expr), stmts,
true, NULL_TREE);
}
/* Renames the scalar uses of the statement COPY, using the
substitution map RENAME_MAP, inserting the gimplification code at
GSI_TGT, for the translation REGION, with the original copied
statement in LOOP, and using the induction variable renaming map
IV_MAP. Returns true when something has been renamed. codegen_error
is set when the code generation cannot continue. */
bool
translate_isl_ast_to_gimple::rename_uses (gimple *copy,
gimple_stmt_iterator *gsi_tgt,
basic_block old_bb,
loop_p loop, vec<tree> iv_map)
{
bool changed = false;
if (is_gimple_debug (copy))
{
if (gimple_debug_bind_p (copy))
gimple_debug_bind_reset_value (copy);
else if (gimple_debug_source_bind_p (copy))
return false;
else
gcc_unreachable ();
return false;
}
if (dump_file)
{
fprintf (dump_file, "[codegen] renaming uses of stmt: ");
print_gimple_stmt (dump_file, copy, 0, 0);
}
use_operand_p use_p;
ssa_op_iter op_iter;
FOR_EACH_SSA_USE_OPERAND (use_p, copy, op_iter, SSA_OP_USE)
{
tree old_name = USE_FROM_PTR (use_p);
if (dump_file)
{
fprintf (dump_file, "[codegen] renaming old_name = ");
print_generic_expr (dump_file, old_name, 0);
fprintf (dump_file, "\n");
}
if (TREE_CODE (old_name) != SSA_NAME
|| SSA_NAME_IS_DEFAULT_DEF (old_name))
continue;
changed = true;
tree new_expr = get_rename (gsi_tgt->bb, old_name,
old_bb, false);
if (new_expr)
{
tree type_old_name = TREE_TYPE (old_name);
tree type_new_expr = TREE_TYPE (new_expr);
if (dump_file)
{
fprintf (dump_file, "[codegen] from rename_map: new_name = ");
print_generic_expr (dump_file, new_expr, 0);
fprintf (dump_file, "\n");
}
if (type_old_name != type_new_expr
|| TREE_CODE (new_expr) != SSA_NAME)
{
tree var = create_tmp_var (type_old_name, "var");
if (!useless_type_conversion_p (type_old_name, type_new_expr))
new_expr = fold_convert (type_old_name, new_expr);
gimple_seq stmts;
new_expr = force_gimple_operand (new_expr, &stmts, true, var);
gsi_insert_earliest (stmts);
}
replace_exp (use_p, new_expr);
continue;
}
gimple_seq stmts;
new_expr = get_rename_from_scev (old_name, &stmts, loop, gimple_bb (copy),
old_bb, iv_map);
if (!new_expr || codegen_error_p ())
return false;
if (dump_file)
{
fprintf (dump_file, "[codegen] not in rename map, scev: ");
print_generic_expr (dump_file, new_expr, 0);
fprintf (dump_file, "\n");
}
gsi_insert_earliest (stmts);
replace_exp (use_p, new_expr);
if (TREE_CODE (new_expr) == INTEGER_CST
&& is_gimple_assign (copy))
{
tree rhs = gimple_assign_rhs1 (copy);
if (TREE_CODE (rhs) == ADDR_EXPR)
recompute_tree_invariant_for_addr_expr (rhs);
}
set_rename (old_name, new_expr);
}
return changed;
}
/* Returns a basic block that could correspond to where a constant was defined
in the original code. In the original code OLD_BB had the definition, we
need to find which basic block out of the copies of old_bb, in the new
region, should a definition correspond to if it has to reach BB. */
basic_block
translate_isl_ast_to_gimple::get_def_bb_for_const (basic_block bb,
basic_block old_bb) const
{
vec <basic_block> *bbs = region->copied_bb_map->get (old_bb);
if (!bbs || bbs->is_empty ())
return NULL;
if (1 == bbs->length ())
return (*bbs)[0];
int i;
basic_block b1 = NULL, b2;
FOR_EACH_VEC_ELT (*bbs, i, b2)
{
if (b2 == bb)
return bb;
/* BB and B2 are in two unrelated if-clauses. */
if (!dominated_by_p (CDI_DOMINATORS, bb, b2))
continue;
/* Compute the nearest dominator. */
if (!b1 || dominated_by_p (CDI_DOMINATORS, b2, b1))
b1 = b2;
}
gcc_assert (b1);
return b1;
}
/* Get the new name of OP (from OLD_BB) to be used in NEW_BB. LOOP_PHI is true
when we want to rename an OP within a loop PHI instruction. */
tree
translate_isl_ast_to_gimple::
get_new_name (basic_block new_bb, tree op,
basic_block old_bb, bool loop_phi) const
{
/* For constants the names are the same. */
if (is_constant (op))
return op;
return get_rename (new_bb, op, old_bb, loop_phi);
}
/* Return a debug location for OP. */
static location_t
get_loc (tree op)
{
location_t loc = UNKNOWN_LOCATION;
if (TREE_CODE (op) == SSA_NAME)
loc = gimple_location (SSA_NAME_DEF_STMT (op));
return loc;
}
/* Returns the incoming edges of basic_block BB in the pair. The first edge is
the init edge (from outside the loop) and the second one is the back edge
from the same loop. */
std::pair<edge, edge>
get_edges (basic_block bb)
{
std::pair<edge, edge> edges;
edge e;
edge_iterator ei;
FOR_EACH_EDGE (e, ei, bb->preds)
if (bb->loop_father != e->src->loop_father)
edges.first = e;
else
edges.second = e;
return edges;
}
/* Copy the PHI arguments from OLD_PHI to the NEW_PHI. The arguments to NEW_PHI
must be found unless they can be POSTPONEd for later. */
bool
translate_isl_ast_to_gimple::
copy_loop_phi_args (gphi *old_phi, init_back_edge_pair_t &ibp_old_bb,
gphi *new_phi, init_back_edge_pair_t &ibp_new_bb,
bool postpone)
{
gcc_assert (gimple_phi_num_args (old_phi) == gimple_phi_num_args (new_phi));
basic_block new_bb = gimple_bb (new_phi);
for (unsigned i = 0; i < gimple_phi_num_args (old_phi); i++)
{
edge e;
if (gimple_phi_arg_edge (old_phi, i) == ibp_old_bb.first)
e = ibp_new_bb.first;
else
e = ibp_new_bb.second;
tree old_name = gimple_phi_arg_def (old_phi, i);
tree new_name = get_new_name (new_bb, old_name,
gimple_bb (old_phi), true);
if (new_name)
{
add_phi_arg (new_phi, new_name, e, get_loc (old_name));
continue;
}
gimple *old_def_stmt = SSA_NAME_DEF_STMT (old_name);
if (!old_def_stmt || gimple_code (old_def_stmt) == GIMPLE_NOP)
/* If the phi arg was a function arg, or wasn't defined, just use the
old name. */
add_phi_arg (new_phi, old_name, e, get_loc (old_name));
else if (postpone)
{
/* Postpone code gen for later for those back-edges we don't have the
names yet. */
region->incomplete_phis.safe_push (std::make_pair (old_phi, new_phi));
if (dump_file)
fprintf (dump_file, "[codegen] postpone loop phi nodes.\n");
}
else
/* Either we should add the arg to phi or, we should postpone. */
return false;
}
return true;
}
/* Copy loop phi nodes from BB to NEW_BB. */
bool
translate_isl_ast_to_gimple::copy_loop_phi_nodes (basic_block bb,
basic_block new_bb)
{
if (dump_file)
fprintf (dump_file, "[codegen] copying loop phi nodes in bb_%d.\n",
new_bb->index);
/* Loop phi nodes should have only two arguments. */
gcc_assert (2 == EDGE_COUNT (bb->preds));
/* First edge is the init edge and second is the back edge. */
init_back_edge_pair_t ibp_old_bb = get_edges (bb);
/* First edge is the init edge and second is the back edge. */
init_back_edge_pair_t ibp_new_bb = get_edges (new_bb);
for (gphi_iterator psi = gsi_start_phis (bb); !gsi_end_p (psi);
gsi_next (&psi))
{
gphi *phi = psi.phi ();
tree res = gimple_phi_result (phi);
if (virtual_operand_p (res))
continue;
if (is_gimple_reg (res) && scev_analyzable_p (res, region->region))
continue;
gphi *new_phi = create_phi_node (SSA_NAME_VAR (res), new_bb);
tree new_res = create_new_def_for (res, new_phi,
gimple_phi_result_ptr (new_phi));
set_rename (res, new_res);
codegen_error = !copy_loop_phi_args (phi, ibp_old_bb, new_phi,
ibp_new_bb, true);
update_stmt (new_phi);
if (dump_file)
{
fprintf (dump_file, "[codegen] creating loop-phi node: ");
print_gimple_stmt (dump_file, new_phi, 0, 0);
}
}
return true;
}
/* Return the init value of PHI, the value coming from outside the loop. */
static tree
get_loop_init_value (gphi *phi)
{
loop_p loop = gimple_bb (phi)->loop_father;
edge e;
edge_iterator ei;
FOR_EACH_EDGE (e, ei, gimple_bb (phi)->preds)
if (e->src->loop_father != loop)
return gimple_phi_arg_def (phi, e->dest_idx);
return NULL_TREE;
}
/* Find the init value (the value which comes from outside the loop), of one of
the operands of DEF which is defined by a loop phi. */
static tree
find_init_value (gimple *def)
{
if (gimple_code (def) == GIMPLE_PHI)
return get_loop_init_value (as_a <gphi*> (def));
if (gimple_vuse (def))
return NULL_TREE;
ssa_op_iter iter;
use_operand_p use_p;
FOR_EACH_SSA_USE_OPERAND (use_p, def, iter, SSA_OP_USE)
{
tree use = USE_FROM_PTR (use_p);
if (TREE_CODE (use) == SSA_NAME)
{
if (tree res = find_init_value (SSA_NAME_DEF_STMT (use)))
return res;
}
}
return NULL_TREE;
}
/* Return the init value, the value coming from outside the loop. */
static tree
find_init_value_close_phi (gphi *phi)
{
gcc_assert (gimple_phi_num_args (phi) == 1);
tree use_arg = gimple_phi_arg_def (phi, 0);
gimple *def = SSA_NAME_DEF_STMT (use_arg);
return find_init_value (def);
}
tree translate_isl_ast_to_gimple::
add_close_phis_to_outer_loops (tree last_merge_name, edge last_e,
gimple *old_close_phi)
{
sese_l &codegen_region = region->if_region->true_region->region;
gimple *stmt = SSA_NAME_DEF_STMT (last_merge_name);
basic_block bb = gimple_bb (stmt);
if (!bb_in_sese_p (bb, codegen_region))
return last_merge_name;
loop_p loop = bb->loop_father;
if (!loop_in_sese_p (loop, codegen_region))
return last_merge_name;
edge e = single_exit (loop);
if (dominated_by_p (CDI_DOMINATORS, e->dest, last_e->src))
return last_merge_name;
tree old_name = gimple_phi_arg_def (old_close_phi, 0);
tree old_close_phi_name = gimple_phi_result (old_close_phi);
bb = e->dest;
if (!bb_contains_loop_close_phi_nodes (bb) || !single_succ_p (bb))
bb = split_edge (e);
gphi *close_phi = create_phi_node (SSA_NAME_VAR (last_merge_name), bb);
tree res = create_new_def_for (last_merge_name, close_phi,
gimple_phi_result_ptr (close_phi));
set_rename (old_close_phi_name, res);
add_phi_arg (close_phi, last_merge_name, e, get_loc (old_name));
last_merge_name = res;
return add_close_phis_to_outer_loops (last_merge_name, last_e, old_close_phi);
}
/* Add phi nodes to all merge points of all the diamonds enclosing the loop of
the close phi node PHI. */
bool translate_isl_ast_to_gimple::
add_close_phis_to_merge_points (gphi *old_close_phi, gphi *new_close_phi,
tree default_value)
{
sese_l &codegen_region = region->if_region->true_region->region;
basic_block default_value_bb = get_entry_bb (codegen_region);
if (SSA_NAME == TREE_CODE (default_value))
{
gimple *stmt = SSA_NAME_DEF_STMT (default_value);
if (!stmt || gimple_code (stmt) == GIMPLE_NOP)
return false;
default_value_bb = gimple_bb (stmt);
}
basic_block new_close_phi_bb = gimple_bb (new_close_phi);
tree old_close_phi_name = gimple_phi_result (old_close_phi);
tree new_close_phi_name = gimple_phi_result (new_close_phi);
tree last_merge_name = new_close_phi_name;
tree old_name = gimple_phi_arg_def (old_close_phi, 0);
int i;
edge merge_e;
FOR_EACH_VEC_ELT_REVERSE (merge_points, i, merge_e)
{
basic_block new_merge_bb = merge_e->src;
if (!dominated_by_p (CDI_DOMINATORS, new_merge_bb, default_value_bb))
continue;
last_merge_name = add_close_phis_to_outer_loops (last_merge_name, merge_e,
old_close_phi);
gphi *merge_phi = create_phi_node (SSA_NAME_VAR (old_close_phi_name), new_merge_bb);
tree merge_res = create_new_def_for (old_close_phi_name, merge_phi,
gimple_phi_result_ptr (merge_phi));
set_rename (old_close_phi_name, merge_res);
edge from_loop = NULL, from_default_value = NULL;
edge e;
edge_iterator ei;
FOR_EACH_EDGE (e, ei, new_merge_bb->preds)
if (dominated_by_p (CDI_DOMINATORS, e->src, new_close_phi_bb))
from_loop = e;
else
from_default_value = e;
/* Because CDI_POST_DOMINATORS are not updated, we only rely on
CDI_DOMINATORS, which may not handle all cases where new_close_phi_bb
is contained in another condition. */
if (!from_default_value || !from_loop)
return false;
add_phi_arg (merge_phi, last_merge_name, from_loop, get_loc (old_name));
add_phi_arg (merge_phi, default_value, from_default_value, get_loc (old_name));
if (dump_file)
{
fprintf (dump_file, "[codegen] Adding guard-phi: ");
print_gimple_stmt (dump_file, merge_phi, 0, 0);
}
update_stmt (merge_phi);
last_merge_name = merge_res;
}
return true;
}
/* Copy all the loop-close phi args from BB to NEW_BB. */
bool
translate_isl_ast_to_gimple::copy_loop_close_phi_args (basic_block old_bb,
basic_block new_bb,
bool postpone)
{
for (gphi_iterator psi = gsi_start_phis (old_bb); !gsi_end_p (psi);
gsi_next (&psi))
{
gphi *old_close_phi = psi.phi ();
tree res = gimple_phi_result (old_close_phi);
if (virtual_operand_p (res))
continue;
if (is_gimple_reg (res) && scev_analyzable_p (res, region->region))
/* Loop close phi nodes should not be scev_analyzable_p. */
gcc_unreachable ();
gphi *new_close_phi = create_phi_node (SSA_NAME_VAR (res), new_bb);
tree new_res = create_new_def_for (res, new_close_phi,
gimple_phi_result_ptr (new_close_phi));
set_rename (res, new_res);
tree old_name = gimple_phi_arg_def (old_close_phi, 0);
tree new_name = get_new_name (new_bb, old_name, old_bb, false);
/* Predecessor basic blocks of a loop close phi should have been code
generated before. FIXME: This is fixable by merging PHIs from inner
loops as well. See: gfortran.dg/graphite/interchange-3.f90. */
if (!new_name)
return false;
add_phi_arg (new_close_phi, new_name, single_pred_edge (new_bb),
get_loc (old_name));
if (dump_file)
{
fprintf (dump_file, "[codegen] Adding loop close phi: ");
print_gimple_stmt (dump_file, new_close_phi, 0, 0);
}
update_stmt (new_close_phi);
/* When there is no loop guard around this codegenerated loop, there is no
need to collect the close-phi arg. */
if (merge_points.is_empty ())
continue;
/* Add a PHI in the succ_new_bb for each close phi of the loop. */
tree default_value = find_init_value_close_phi (new_close_phi);
/* A close phi must come from a loop-phi having a default value. */
if (!default_value)
{
if (!postpone)
return false;
region->incomplete_phis.safe_push (std::make_pair (old_close_phi,
new_close_phi));
if (dump_file)
{
fprintf (dump_file, "[codegen] postpone close phi nodes: ");
print_gimple_stmt (dump_file, new_close_phi, 0, 0);
}
continue;
}
if (!add_close_phis_to_merge_points (old_close_phi, new_close_phi,
default_value))
return false;
}
return true;
}
/* Copy loop close phi nodes from BB to NEW_BB. */
bool
translate_isl_ast_to_gimple::copy_loop_close_phi_nodes (basic_block old_bb,
basic_block new_bb)
{
if (dump_file)
fprintf (dump_file, "[codegen] copying loop close phi nodes in bb_%d.\n",
new_bb->index);
/* Loop close phi nodes should have only one argument. */
gcc_assert (1 == EDGE_COUNT (old_bb->preds));
return copy_loop_close_phi_args (old_bb, new_bb, true);
}
/* Add NEW_NAME as the ARGNUM-th arg of NEW_PHI which is in NEW_BB.
DOMINATING_PRED is the predecessor basic block of OLD_BB which dominates the
other pred of OLD_BB as well. If no such basic block exists then it is NULL.
NON_DOMINATING_PRED is a pred which does not dominate OLD_BB, it cannot be
NULL.
Case1: OLD_BB->preds {BB1, BB2} and BB1 does not dominate BB2 and vice versa.
In this case DOMINATING_PRED = NULL.
Case2: OLD_BB->preds {BB1, BB2} and BB1 dominates BB2.
Returns true on successful copy of the args, false otherwise. */
bool
translate_isl_ast_to_gimple::
add_phi_arg_for_new_expr (tree old_phi_args[2], tree new_phi_args[2],
edge old_bb_dominating_edge,
edge old_bb_non_dominating_edge,
gphi *phi, gphi *new_phi,
basic_block new_bb)
{
basic_block def_pred[2] = { NULL, NULL };
int not_found_bb_index = -1;
for (int i = 0; i < 2; i++)
{
/* If the corresponding def_bb could not be found the entry will be
NULL. */
if (TREE_CODE (old_phi_args[i]) == INTEGER_CST)
def_pred[i] = get_def_bb_for_const (new_bb,
gimple_phi_arg_edge (phi, i)->src);
else if (new_phi_args[i] && (TREE_CODE (new_phi_args[i]) == SSA_NAME))
def_pred[i] = gimple_bb (SSA_NAME_DEF_STMT (new_phi_args[i]));
if (!def_pred[i])
{
/* When non are available bail out. */
if (not_found_bb_index != -1)
return false;
not_found_bb_index = i;
}
}
/* Here we are pattern matching on the structure of CFG w.r.t. old one. */
if (old_bb_dominating_edge)
{
if (not_found_bb_index != -1)
return false;
basic_block new_pred1 = (*new_bb->preds)[0]->src;
basic_block new_pred2 = (*new_bb->preds)[1]->src;
vec <basic_block> *bbs
= region->copied_bb_map->get (old_bb_non_dominating_edge->src);
/* Could not find a mapping. */
if (!bbs)
return false;
basic_block new_pred = NULL;
basic_block b;
int i;
FOR_EACH_VEC_ELT (*bbs, i, b)
{
if (dominated_by_p (CDI_DOMINATORS, new_pred1, b))
{
/* FIXME: If we have already found new_pred then we have to
disambiguate, bail out for now. */
if (new_pred)
return false;
new_pred = new_pred1;
}
if (dominated_by_p (CDI_DOMINATORS, new_pred2, b))
{
/* FIXME: If we have already found new_pred then we have to either
it dominates both or we have to disambiguate, bail out. */
if (new_pred)
return false;
new_pred = new_pred2;
}
}
if (!new_pred)
return false;
edge new_non_dominating_edge = find_edge (new_pred, new_bb);
gcc_assert (new_non_dominating_edge);
/* FIXME: Validate each args just like in loop-phis. */
/* By the process of elimination we first insert insert phi-edge for
non-dominating pred which is computed above and then we insert the
remaining one. */
int inserted_edge = 0;
for (; inserted_edge < 2; inserted_edge++)
{
edge new_bb_pred_edge = gimple_phi_arg_edge (new_phi, inserted_edge);
if (new_non_dominating_edge == new_bb_pred_edge)
{
add_phi_arg (new_phi, new_phi_args[inserted_edge],
new_non_dominating_edge,
get_loc (old_phi_args[inserted_edge]));
break;
}
}
if (inserted_edge == 2)
return false;
int edge_dominating = inserted_edge == 0 ? 1 : 0;
edge new_dominating_edge = NULL;
for (inserted_edge = 0; inserted_edge < 2; inserted_edge++)
{
edge e = gimple_phi_arg_edge (new_phi, inserted_edge);
if (e != new_non_dominating_edge)
{
new_dominating_edge = e;
add_phi_arg (new_phi, new_phi_args[edge_dominating],
new_dominating_edge,
get_loc (old_phi_args[inserted_edge]));
break;
}
}
gcc_assert (new_dominating_edge);
}
else
{
/* Classic diamond structure: both edges are non-dominating. We need to
find one unique edge then the other can be found be elimination. If
any definition (def_pred) dominates both the preds of new_bb then we
bail out. Entries of def_pred maybe NULL, in that case we must
uniquely find pred with help of only one entry. */
edge new_e[2] = { NULL, NULL };
for (int i = 0; i < 2; i++)
{
edge e;
edge_iterator ei;
FOR_EACH_EDGE (e, ei, new_bb->preds)
if (def_pred[i]
&& dominated_by_p (CDI_DOMINATORS, e->src, def_pred[i]))
{
if (new_e[i])
/* We do not know how to handle the case when def_pred
dominates more than a predecessor. */
return false;
new_e[i] = e;
}
}
gcc_assert (new_e[0] || new_e[1]);
/* Find the other edge by process of elimination. */
if (not_found_bb_index != -1)
{
gcc_assert (!new_e[not_found_bb_index]);
int found_bb_index = not_found_bb_index == 1 ? 0 : 1;
edge e;
edge_iterator ei;
FOR_EACH_EDGE (e, ei, new_bb->preds)
{
if (new_e[found_bb_index] == e)
continue;
new_e[not_found_bb_index] = e;
}
}
/* Add edges to phi args. */
for (int i = 0; i < 2; i++)
add_phi_arg (new_phi, new_phi_args[i], new_e[i],
get_loc (old_phi_args[i]));
}
return true;
}
/* Copy the arguments of cond-phi node PHI, to NEW_PHI in the codegenerated
region. If postpone is true and it isn't possible to copy any arg of PHI,
the PHI is added to the REGION->INCOMPLETE_PHIS to be codegenerated later.
Returns false if the copying was unsuccessful. */
bool
translate_isl_ast_to_gimple::copy_cond_phi_args (gphi *phi, gphi *new_phi,
vec<tree> iv_map,
bool postpone)
{
if (dump_file)
fprintf (dump_file, "[codegen] copying cond phi args.\n");
gcc_assert (2 == gimple_phi_num_args (phi));
basic_block new_bb = gimple_bb (new_phi);
loop_p loop = gimple_bb (phi)->loop_father;
basic_block old_bb = gimple_bb (phi);
edge old_bb_non_dominating_edge = NULL, old_bb_dominating_edge = NULL;
edge e;
edge_iterator ei;
FOR_EACH_EDGE (e, ei, old_bb->preds)
if (!dominated_by_p (CDI_DOMINATORS, old_bb, e->src))
old_bb_non_dominating_edge = e;
else
old_bb_dominating_edge = e;
gcc_assert (!dominated_by_p (CDI_DOMINATORS, old_bb,
old_bb_non_dominating_edge->src));
tree new_phi_args[2];
tree old_phi_args[2];
for (unsigned i = 0; i < gimple_phi_num_args (phi); i++)
{
tree old_name = gimple_phi_arg_def (phi, i);
tree new_name = get_new_name (new_bb, old_name, old_bb, false);
old_phi_args[i] = old_name;
if (new_name)
{
new_phi_args [i] = new_name;
continue;
}
/* If the phi-arg was a parameter. */
if (vec_find (region->params, old_name) != -1)
{
new_phi_args [i] = old_name;
if (dump_file)
{
fprintf (dump_file,
"[codegen] parameter argument to phi, new_expr: ");
print_generic_expr (dump_file, new_phi_args[i], 0);
fprintf (dump_file, "\n");
}
continue;
}
gimple *old_def_stmt = SSA_NAME_DEF_STMT (old_name);
if (!old_def_stmt || gimple_code (old_def_stmt) == GIMPLE_NOP)
/* FIXME: If the phi arg was a function arg, or wasn't defined, just use
the old name. */
return false;
if (postpone)
{
/* If the phi-arg is scev-analyzeable but only in the first stage. */
if (is_gimple_reg (old_name)
&& scev_analyzable_p (old_name, region->region))
{
gimple_seq stmts;
tree new_expr = get_rename_from_scev (old_name, &stmts, loop,
new_bb, old_bb, iv_map);
if (codegen_error_p ())
return false;
gcc_assert (new_expr);
if (dump_file)
{
fprintf (dump_file,
"[codegen] scev analyzeable, new_expr: ");
print_generic_expr (dump_file, new_expr, 0);
fprintf (dump_file, "\n");
}
gsi_insert_earliest (stmts);
new_phi_args [i] = new_name;
continue;
}
/* Postpone code gen for later for back-edges. */
region->incomplete_phis.safe_push (std::make_pair (phi, new_phi));
if (dump_file)
{
fprintf (dump_file, "[codegen] postpone cond phi nodes: ");
print_gimple_stmt (dump_file, new_phi, 0, 0);
}
new_phi_args [i] = NULL_TREE;
continue;
}
else
/* Either we should add the arg to phi or, we should postpone. */
return false;
}
/* If none of the args have been determined in the first stage then wait until
later. */
if (postpone && !new_phi_args[0] && !new_phi_args[1])
return true;
return add_phi_arg_for_new_expr (old_phi_args, new_phi_args,
old_bb_dominating_edge,
old_bb_non_dominating_edge,
phi, new_phi, new_bb);
}
/* Copy cond phi nodes from BB to NEW_BB. A cond-phi node is a basic block
containing phi nodes coming from two predecessors, and none of them are back
edges. */
bool
translate_isl_ast_to_gimple::copy_cond_phi_nodes (basic_block bb,
basic_block new_bb,
vec<tree> iv_map)
{
gcc_assert (!bb_contains_loop_close_phi_nodes (bb));
if (dump_file)
fprintf (dump_file, "[codegen] copying cond phi nodes in bb_%d.\n",
new_bb->index);
/* Cond phi nodes should have exactly two arguments. */
gcc_assert (2 == EDGE_COUNT (bb->preds));
for (gphi_iterator psi = gsi_start_phis (bb); !gsi_end_p (psi);
gsi_next (&psi))
{
gphi *phi = psi.phi ();
tree res = gimple_phi_result (phi);
if (virtual_operand_p (res))
continue;
if (is_gimple_reg (res) && scev_analyzable_p (res, region->region))
/* Cond phi nodes should not be scev_analyzable_p. */
gcc_unreachable ();
gphi *new_phi = create_phi_node (SSA_NAME_VAR (res), new_bb);
tree new_res = create_new_def_for (res, new_phi,
gimple_phi_result_ptr (new_phi));
set_rename (res, new_res);
if (!copy_cond_phi_args (phi, new_phi, iv_map, true))
return false;
update_stmt (new_phi);
}
return true;
}
/* Return true if STMT should be copied from region to the new code-generated
region. LABELs, CONDITIONS, induction-variables and region parameters need
not be copied. */
static bool
should_copy_to_new_region (gimple *stmt, sese_info_p region)
{
/* Do not copy labels or conditions. */
if (gimple_code (stmt) == GIMPLE_LABEL
|| gimple_code (stmt) == GIMPLE_COND)
return false;
tree lhs;
/* Do not copy induction variables. */
if (is_gimple_assign (stmt)
&& (lhs = gimple_assign_lhs (stmt))
&& TREE_CODE (lhs) == SSA_NAME
&& is_gimple_reg (lhs)
&& scev_analyzable_p (lhs, region->region))
return false;
return true;
}
/* Create new names for all the definitions created by COPY and add replacement
mappings for each new name. */
void
translate_isl_ast_to_gimple::set_rename_for_each_def (gimple *stmt)
{
def_operand_p def_p;
ssa_op_iter op_iter;
FOR_EACH_SSA_DEF_OPERAND (def_p, stmt, op_iter, SSA_OP_ALL_DEFS)
{
tree old_name = DEF_FROM_PTR (def_p);
tree new_name = create_new_def_for (old_name, stmt, def_p);
set_rename (old_name, new_name);
}
}
/* Duplicates the statements of basic block BB into basic block NEW_BB
and compute the new induction variables according to the IV_MAP.
CODEGEN_ERROR is set when the code generation cannot continue. */
bool
translate_isl_ast_to_gimple::graphite_copy_stmts_from_block (basic_block bb,
basic_block new_bb,
vec<tree> iv_map)
{
/* Iterator poining to the place where new statement (s) will be inserted. */
gimple_stmt_iterator gsi_tgt = gsi_last_bb (new_bb);
for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi);
gsi_next (&gsi))
{
gimple *stmt = gsi_stmt (gsi);
if (!should_copy_to_new_region (stmt, region))
continue;
/* Create a new copy of STMT and duplicate STMT's virtual
operands. */
gimple *copy = gimple_copy (stmt);
gsi_insert_after (&gsi_tgt, copy, GSI_NEW_STMT);
if (dump_file)
{
fprintf (dump_file, "[codegen] inserting statement: ");
print_gimple_stmt (dump_file, copy, 0, 0);
}
maybe_duplicate_eh_stmt (copy, stmt);
gimple_duplicate_stmt_histograms (cfun, copy, cfun, stmt);
/* Crete new names for each def in the copied stmt. */
set_rename_for_each_def (copy);
loop_p loop = bb->loop_father;
if (rename_uses (copy, &gsi_tgt, bb, loop, iv_map))
{
fold_stmt_inplace (&gsi_tgt);
gcc_assert (gsi_stmt (gsi_tgt) == copy);
}
if (codegen_error_p ())
return false;
update_stmt (copy);
}
return true;
}
/* Given a basic block containing close-phi it returns the new basic block where
to insert a copy of the close-phi nodes. All the uses in close phis should
come from a single loop otherwise it returns NULL. */
edge
translate_isl_ast_to_gimple::edge_for_new_close_phis (basic_block bb)
{
/* Make sure that NEW_BB is the new_loop->exit->dest. We find the definition
of close phi in the original code and then find the mapping of basic block
defining that variable. If there are multiple close-phis and they are
defined in different loops (in the original or in the new code) because of
loop splitting, then we bail out. */
loop_p new_loop = NULL;
for (gphi_iterator psi = gsi_start_phis (bb); !gsi_end_p (psi);
gsi_next (&psi))
{
gphi *phi = psi.phi ();
tree name = gimple_phi_arg_def (phi, 0);
basic_block old_loop_bb = gimple_bb (SSA_NAME_DEF_STMT (name));
vec <basic_block> *bbs = region->copied_bb_map->get (old_loop_bb);
if (!bbs || bbs->length () != 1)
/* This is one of the places which shows preserving original structure
is not always possible, as we may need to insert close PHI for a loop
where the latch does not have any mapping, or the mapping is
ambiguous. */
return NULL;
if (!new_loop)
new_loop = (*bbs)[0]->loop_father;
else if (new_loop != (*bbs)[0]->loop_father)
return NULL;
}
if (!new_loop)
return NULL;
return single_exit (new_loop);
}
/* Copies BB and includes in the copied BB all the statements that can
be reached following the use-def chains from the memory accesses,
and returns the next edge following this new block. codegen_error is
set when the code generation cannot continue. */
edge
translate_isl_ast_to_gimple::copy_bb_and_scalar_dependences (basic_block bb,
edge next_e,
vec<tree> iv_map)
{
int num_phis = number_of_phi_nodes (bb);
if (region->copied_bb_map->get (bb))
{
/* FIXME: we should be able to handle phi nodes with args coming from
outside the region. */
if (num_phis)
{
codegen_error = true;
return NULL;
}
}
basic_block new_bb = NULL;
if (bb_contains_loop_close_phi_nodes (bb))
{
if (dump_file)
fprintf (dump_file, "[codegen] bb_%d contains close phi nodes.\n",
bb->index);
edge e = edge_for_new_close_phis (bb);
if (!e)
{
codegen_error = true;
return NULL;
}
basic_block phi_bb = e->dest;
if (!bb_contains_loop_close_phi_nodes (phi_bb) || !single_succ_p (phi_bb))
phi_bb = split_edge (e);
gcc_assert (single_pred_edge (phi_bb)->src->loop_father
!= single_pred_edge (phi_bb)->dest->loop_father);
if (!copy_loop_close_phi_nodes (bb, phi_bb))
{
codegen_error = true;
return NULL;
}
if (e == next_e)
new_bb = phi_bb;
else
new_bb = split_edge (next_e);
}
else
{
new_bb = split_edge (next_e);
if (num_phis > 0 && bb_contains_loop_phi_nodes (bb))
{
basic_block phi_bb = next_e->dest->loop_father->header;
/* At this point we are unable to codegenerate by still preserving the SSA
structure because maybe the loop is completely unrolled and the PHIs
and cross-bb scalar dependencies are untrackable w.r.t. the original
code. See gfortran.dg/graphite/pr29832.f90. */
if (EDGE_COUNT (bb->preds) != EDGE_COUNT (phi_bb->preds))
{
codegen_error = true;
return NULL;
}
/* In case isl did some loop peeling, like this:
S_8(0);
for (int c1 = 1; c1 <= 5; c1 += 1) {
S_8(c1);
}
S_8(6);
there should be no loop-phi nodes in S_8(0).
FIXME: We need to reason about dynamic instances of S_8, i.e., the
values of all scalar variables: for the moment we instantiate only
SCEV analyzable expressions on the iteration domain, and we need to
extend that to reductions that cannot be analyzed by SCEV. */
if (!bb_in_sese_p (phi_bb, region->if_region->true_region->region))
{
codegen_error = true;
return NULL;
}
if (dump_file)
fprintf (dump_file, "[codegen] bb_%d contains loop phi nodes.\n",
bb->index);
if (!copy_loop_phi_nodes (bb, phi_bb))
{
codegen_error = true;
return NULL;
}
}
else if (num_phis > 0)
{
if (dump_file)
fprintf (dump_file, "[codegen] bb_%d contains cond phi nodes.\n",
bb->index);
basic_block phi_bb = single_pred (new_bb);
loop_p loop_father = new_bb->loop_father;
/* Move back until we find the block with two predecessors. */
while (single_pred_p (phi_bb))
phi_bb = single_pred_edge (phi_bb)->src;
/* If a corresponding merge-point was not found, then abort codegen. */
if (phi_bb->loop_father != loop_father
|| !bb_in_sese_p (phi_bb, region->if_region->true_region->region)
|| !copy_cond_phi_nodes (bb, phi_bb, iv_map))
{
codegen_error = true;
return NULL;
}
}
}
if (dump_file)
fprintf (dump_file, "[codegen] copying from bb_%d to bb_%d.\n",
bb->index, new_bb->index);
vec <basic_block> *copied_bbs = region->copied_bb_map->get (bb);
if (copied_bbs)
copied_bbs->safe_push (new_bb);
else
{
vec<basic_block> bbs;
bbs.create (2);
bbs.safe_push (new_bb);
region->copied_bb_map->put (bb, bbs);
}
if (!graphite_copy_stmts_from_block (bb, new_bb, iv_map))
{
codegen_error = true;
return NULL;
}
return single_succ_edge (new_bb);
}
/* Patch the missing arguments of the phi nodes. */
void
translate_isl_ast_to_gimple::translate_pending_phi_nodes ()
{
int i;
phi_rename *rename;
FOR_EACH_VEC_ELT (region->incomplete_phis, i, rename)
{
gphi *old_phi = rename->first;
gphi *new_phi = rename->second;
basic_block old_bb = gimple_bb (old_phi);
basic_block new_bb = gimple_bb (new_phi);
/* First edge is the init edge and second is the back edge. */
init_back_edge_pair_t ibp_old_bb = get_edges (old_bb);
init_back_edge_pair_t ibp_new_bb = get_edges (new_bb);
if (dump_file)
{
fprintf (dump_file, "[codegen] translating pending old-phi: ");
print_gimple_stmt (dump_file, old_phi, 0, 0);
}
auto_vec <tree, 1> iv_map;
if (bb_contains_loop_phi_nodes (new_bb))
codegen_error = !copy_loop_phi_args (old_phi, ibp_old_bb, new_phi,
ibp_new_bb, false);
else if (bb_contains_loop_close_phi_nodes (new_bb))
codegen_error = !copy_loop_close_phi_args (old_bb, new_bb, false);
else
codegen_error = !copy_cond_phi_args (old_phi, new_phi, iv_map, false);
if (dump_file)
{
fprintf (dump_file, "[codegen] to new-phi: ");
print_gimple_stmt (dump_file, new_phi, 0, 0);
}
if (codegen_error)
return;
}
}
/* Prints NODE to FILE. */
void
translate_isl_ast_to_gimple::print_isl_ast_node (FILE *file,
__isl_keep isl_ast_node *node,
__isl_keep isl_ctx *ctx) const
{
isl_printer *prn = isl_printer_to_file (ctx, file);
prn = isl_printer_set_output_format (prn, ISL_FORMAT_C);
prn = isl_printer_print_ast_node (prn, node);
prn = isl_printer_print_str (prn, "\n");
isl_printer_free (prn);
}
/* Add isl's parameter identifiers and corresponding trees to ivs_params. */
void
translate_isl_ast_to_gimple::add_parameters_to_ivs_params (scop_p scop,
ivs_params &ip)
{
sese_info_p region = scop->scop_info;
unsigned nb_parameters = isl_set_dim (scop->param_context, isl_dim_param);
gcc_assert (nb_parameters == region->params.length ());
unsigned i;
for (i = 0; i < nb_parameters; i++)
{
isl_id *tmp_id = isl_set_get_dim_id (scop->param_context,
isl_dim_param, i);
ip[tmp_id] = region->params[i];
}
}
/* Generates a build, which specifies the constraints on the parameters. */
__isl_give isl_ast_build *
translate_isl_ast_to_gimple::generate_isl_context (scop_p scop)
{
isl_set *context_isl = isl_set_params (isl_set_copy (scop->param_context));
return isl_ast_build_from_context (context_isl);
}
/* Get the maximal number of schedule dimensions in the scop SCOP. */
int
translate_isl_ast_to_gimple::get_max_schedule_dimensions (scop_p scop)
{
int i;
poly_bb_p pbb;
int schedule_dims = 0;
FOR_EACH_VEC_ELT (scop->pbbs, i, pbb)
{
int pbb_schedule_dims = isl_map_dim (pbb->transformed, isl_dim_out);
if (pbb_schedule_dims > schedule_dims)
schedule_dims = pbb_schedule_dims;
}
return schedule_dims;
}
/* Extend the schedule to NB_SCHEDULE_DIMS schedule dimensions.
For schedules with different dimensionality, the isl AST generator can not
define an order and will just randomly choose an order. The solution to this
problem is to extend all schedules to the maximal number of schedule
dimensions (using '0's for the remaining values). */
__isl_give isl_map *
translate_isl_ast_to_gimple::extend_schedule (__isl_take isl_map *schedule,
int nb_schedule_dims)
{
int tmp_dims = isl_map_dim (schedule, isl_dim_out);
schedule =
isl_map_add_dims (schedule, isl_dim_out, nb_schedule_dims - tmp_dims);
isl_val *zero =
isl_val_int_from_si (isl_map_get_ctx (schedule), 0);
int i;
for (i = tmp_dims; i < nb_schedule_dims; i++)
{
schedule
= isl_map_fix_val (schedule, isl_dim_out, i, isl_val_copy (zero));
}
isl_val_free (zero);
return schedule;
}
/* Generates a schedule, which specifies an order used to
visit elements in a domain. */
__isl_give isl_union_map *
translate_isl_ast_to_gimple::generate_isl_schedule (scop_p scop)
{
int nb_schedule_dims = get_max_schedule_dimensions (scop);
int i;
poly_bb_p pbb;
isl_union_map *schedule_isl =
isl_union_map_empty (isl_set_get_space (scop->param_context));
FOR_EACH_VEC_ELT (scop->pbbs, i, pbb)
{
/* Dead code elimination: when the domain of a PBB is empty,
don't generate code for the PBB. */
if (isl_set_is_empty (pbb->domain))
continue;
isl_map *bb_schedule = isl_map_copy (pbb->transformed);
bb_schedule = isl_map_intersect_domain (bb_schedule,
isl_set_copy (pbb->domain));
bb_schedule = extend_schedule (bb_schedule, nb_schedule_dims);
schedule_isl
= isl_union_map_union (schedule_isl,
isl_union_map_from_map (bb_schedule));
}
return schedule_isl;
}
/* This method is executed before the construction of a for node. */
__isl_give isl_id *
ast_build_before_for (__isl_keep isl_ast_build *build, void *user)
{
isl_union_map *dependences = (isl_union_map *) user;
ast_build_info *for_info = XNEW (struct ast_build_info);
isl_union_map *schedule = isl_ast_build_get_schedule (build);
isl_space *schedule_space = isl_ast_build_get_schedule_space (build);
int dimension = isl_space_dim (schedule_space, isl_dim_out);
for_info->is_parallelizable =
!carries_deps (schedule, dependences, dimension);
isl_union_map_free (schedule);
isl_space_free (schedule_space);
isl_id *id = isl_id_alloc (isl_ast_build_get_ctx (build), "", for_info);
return id;
}
#ifdef HAVE_ISL_OPTIONS_SET_SCHEDULE_SERIALIZE_SCCS
/* Set the separate option for all schedules. This helps reducing control
overhead. */
__isl_give isl_schedule *
translate_isl_ast_to_gimple::set_options_for_schedule_tree
(__isl_take isl_schedule *schedule)
{
return isl_schedule_map_schedule_node_bottom_up
(schedule, set_separate_option, NULL);
}
#endif
/* Set the separate option for all dimensions.
This helps to reduce control overhead. */
__isl_give isl_ast_build *
translate_isl_ast_to_gimple::set_options (__isl_take isl_ast_build *control,
__isl_keep isl_union_map *schedule)
{
isl_ctx *ctx = isl_union_map_get_ctx (schedule);
isl_space *range_space = isl_space_set_alloc (ctx, 0, 1);
range_space =
isl_space_set_tuple_name (range_space, isl_dim_set, "separate");
isl_union_set *range =
isl_union_set_from_set (isl_set_universe (range_space));
isl_union_set *domain = isl_union_map_range (isl_union_map_copy (schedule));
domain = isl_union_set_universe (domain);
isl_union_map *options = isl_union_map_from_domain_and_range (domain, range);
return isl_ast_build_set_options (control, options);
}
/* Generate isl AST from schedule of SCOP. Also, collects IVS_PARAMS in IP. */
__isl_give isl_ast_node *
translate_isl_ast_to_gimple::scop_to_isl_ast (scop_p scop, ivs_params &ip)
{
isl_ast_node *ast_isl = NULL;
/* Generate loop upper bounds that consist of the current loop iterator, an
operator (< or <=) and an expression not involving the iterator. If this
option is not set, then the current loop iterator may appear several times
in the upper bound. See the isl manual for more details. */
isl_options_set_ast_build_atomic_upper_bound (scop->isl_context, true);
add_parameters_to_ivs_params (scop, ip);
isl_union_map *schedule_isl = generate_isl_schedule (scop);
isl_ast_build *context_isl = generate_isl_context (scop);
context_isl = set_options (context_isl, schedule_isl);
if (flag_loop_parallelize_all)
{
isl_union_map *dependence = scop_get_dependences (scop);
context_isl =
isl_ast_build_set_before_each_for (context_isl, ast_build_before_for,
dependence);
}
#ifdef HAVE_ISL_OPTIONS_SET_SCHEDULE_SERIALIZE_SCCS
if (scop->schedule)
{
scop->schedule = set_options_for_schedule_tree (scop->schedule);
ast_isl = isl_ast_build_node_from_schedule (context_isl, scop->schedule);
isl_union_map_free(schedule_isl);
}
else
ast_isl = isl_ast_build_ast_from_schedule (context_isl, schedule_isl);
#else
ast_isl = isl_ast_build_ast_from_schedule (context_isl, schedule_isl);
isl_schedule_free (scop->schedule);
#endif
isl_ast_build_free (context_isl);
return ast_isl;
}
/* GIMPLE Loop Generator: generates loops from STMT in GIMPLE form for
the given SCOP. Return true if code generation succeeded.
FIXME: This is not yet a full implementation of the code generator
with isl ASTs. Generation of GIMPLE code has to be completed. */
bool
graphite_regenerate_ast_isl (scop_p scop)
{
sese_info_p region = scop->scop_info;
translate_isl_ast_to_gimple t (region);
ifsese if_region = NULL;
isl_ast_node *root_node;
ivs_params ip;
timevar_push (TV_GRAPHITE_CODE_GEN);
root_node = t.scop_to_isl_ast (scop, ip);
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "AST generated by isl: \n");
t.print_isl_ast_node (dump_file, root_node, scop->isl_context);
}
recompute_all_dominators ();
graphite_verify ();
if_region = move_sese_in_condition (region);
region->if_region = if_region;
recompute_all_dominators ();
loop_p context_loop = region->region.entry->src->loop_father;
edge e = single_succ_edge (if_region->true_region->region.entry->dest);
basic_block bb = split_edge (e);
/* Update the true_region exit edge. */
region->if_region->true_region->region.exit = single_succ_edge (bb);
t.translate_isl_ast (context_loop, root_node, e, ip);
if (t.codegen_error_p ())
{
if (dump_file)
fprintf (dump_file, "[codegen] unsuccessful,"
" reverting back to the original code.\n");
set_ifsese_condition (if_region, integer_zero_node);
}
else
{
t.translate_pending_phi_nodes ();
if (!t.codegen_error_p ())
{
sese_insert_phis_for_liveouts (region,
if_region->region->region.exit->src,
if_region->false_region->region.exit,
if_region->true_region->region.exit);
mark_virtual_operands_for_renaming (cfun);
update_ssa (TODO_update_ssa);
graphite_verify ();
scev_reset ();
recompute_all_dominators ();
graphite_verify ();
}
else
{
if (dump_file)
fprintf (dump_file, "[codegen] unsuccessful in translating"
" pending phis, reverting back to the original code.\n");
set_ifsese_condition (if_region, integer_zero_node);
}
}
free (if_region->true_region);
free (if_region->region);
free (if_region);
ivs_params_clear (ip);
isl_ast_node_free (root_node);
timevar_pop (TV_GRAPHITE_CODE_GEN);
if (dump_file && (dump_flags & TDF_DETAILS))
{
loop_p loop;
int num_no_dependency = 0;
FOR_EACH_LOOP (loop, 0)
if (loop->can_be_parallel)
num_no_dependency++;
fprintf (dump_file, "%d loops carried no dependency.\n",
num_no_dependency);
}
return !t.codegen_error_p ();
}
#endif /* HAVE_isl */