1 /* Loop autoparallelization.
2 Copyright (C) 2006, 2007, 2008, 2009, 2010, 2011, 2012
3 Free Software Foundation, Inc.
4 Contributed by Sebastian Pop <pop@cri.ensmp.fr>
5 Zdenek Dvorak <dvorakz@suse.cz> and Razya Ladelsky <razya@il.ibm.com>.
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify it under
10 the terms of the GNU General Public License as published by the Free
11 Software Foundation; either version 3, or (at your option) any later
14 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
15 WARRANTY; without even the implied warranty of MERCHANTABILITY or
16 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
25 #include "coretypes.h"
26 #include "tree-flow.h"
28 #include "tree-data-ref.h"
29 #include "tree-scalar-evolution.h"
30 #include "gimple-pretty-print.h"
31 #include "tree-pass.h"
32 #include "langhooks.h"
33 #include "tree-vectorizer.h"
35 /* This pass tries to distribute iterations of loops into several threads.
36 The implementation is straightforward -- for each loop we test whether its
37 iterations are independent, and if it is the case (and some additional
38 conditions regarding profitability and correctness are satisfied), we
39 add GIMPLE_OMP_PARALLEL and GIMPLE_OMP_FOR codes and let omp expansion
42 The most of the complexity is in bringing the code into shape expected
44 -- for GIMPLE_OMP_FOR, ensuring that the loop has only one induction
45 variable and that the exit test is at the start of the loop body
46 -- for GIMPLE_OMP_PARALLEL, replacing the references to local addressable
47 variables by accesses through pointers, and breaking up ssa chains
48 by storing the values incoming to the parallelized loop to a structure
49 passed to the new function as an argument (something similar is done
50 in omp gimplification, unfortunately only a small part of the code
54 -- if there are several parallelizable loops in a function, it may be
55 possible to generate the threads just once (using synchronization to
56 ensure that cross-loop dependences are obeyed).
57 -- handling of common reduction patterns for outer loops.
59 More info can also be found at http://gcc.gnu.org/wiki/AutoParInGCC */
62 currently we use vect_force_simple_reduction() to detect reduction patterns.
63 The code transformation will be introduced by an example.
70 for (i = 0; i < N; i++)
80 # sum_29 = PHI <sum_11(5), 1(3)>
81 # i_28 = PHI <i_12(5), 0(3)>
84 sum_11 = D.1795_8 + sum_29;
92 # sum_21 = PHI <sum_11(4)>
93 printf (&"%d"[0], sum_21);
96 after reduction transformation (only relevant parts):
104 # Storing the initial value given by the user. #
106 .paral_data_store.32.sum.27 = 1;
108 #pragma omp parallel num_threads(4)
110 #pragma omp for schedule(static)
112 # The neutral element corresponding to the particular
113 reduction's operation, e.g. 0 for PLUS_EXPR,
114 1 for MULT_EXPR, etc. replaces the user's initial value. #
116 # sum.27_29 = PHI <sum.27_11, 0>
118 sum.27_11 = D.1827_8 + sum.27_29;
122 # Adding this reduction phi is done at create_phi_for_local_result() #
123 # sum.27_56 = PHI <sum.27_11, 0>
126 # Creating the atomic operation is done at
127 create_call_for_reduction_1() #
129 #pragma omp atomic_load
130 D.1839_59 = *&.paral_data_load.33_51->reduction.23;
131 D.1840_60 = sum.27_56 + D.1839_59;
132 #pragma omp atomic_store (D.1840_60);
136 # collecting the result after the join of the threads is done at
137 create_loads_for_reductions().
138 The value computed by the threads is loaded from the
142 .paral_data_load.33_52 = &.paral_data_store.32;
143 sum_37 = .paral_data_load.33_52->sum.27;
144 sum_43 = D.1795_41 + sum_37;
147 # sum_21 = PHI <sum_43, sum_26>
148 printf (&"%d"[0], sum_21);
156 /* Minimal number of iterations of a loop that should be executed in each
158 #define MIN_PER_THREAD 100
160 /* Element of the hashtable, representing a
161 reduction in the current loop. */
162 struct reduction_info
164 gimple reduc_stmt
; /* reduction statement. */
165 gimple reduc_phi
; /* The phi node defining the reduction. */
166 enum tree_code reduction_code
;/* code for the reduction operation. */
167 unsigned reduc_version
; /* SSA_NAME_VERSION of original reduc_phi
169 gimple keep_res
; /* The PHI_RESULT of this phi is the resulting value
170 of the reduction variable when existing the loop. */
171 tree initial_value
; /* The initial value of the reduction var before entering the loop. */
172 tree field
; /* the name of the field in the parloop data structure intended for reduction. */
173 tree init
; /* reduction initialization value. */
174 gimple new_phi
; /* (helper field) Newly created phi node whose result
175 will be passed to the atomic operation. Represents
176 the local result each thread computed for the reduction
180 /* Equality and hash functions for hashtab code. */
183 reduction_info_eq (const void *aa
, const void *bb
)
185 const struct reduction_info
*a
= (const struct reduction_info
*) aa
;
186 const struct reduction_info
*b
= (const struct reduction_info
*) bb
;
188 return (a
->reduc_phi
== b
->reduc_phi
);
192 reduction_info_hash (const void *aa
)
194 const struct reduction_info
*a
= (const struct reduction_info
*) aa
;
196 return a
->reduc_version
;
199 static struct reduction_info
*
200 reduction_phi (htab_t reduction_list
, gimple phi
)
202 struct reduction_info tmpred
, *red
;
204 if (htab_elements (reduction_list
) == 0 || phi
== NULL
)
207 tmpred
.reduc_phi
= phi
;
208 tmpred
.reduc_version
= gimple_uid (phi
);
209 red
= (struct reduction_info
*) htab_find (reduction_list
, &tmpred
);
214 /* Element of hashtable of names to copy. */
216 struct name_to_copy_elt
218 unsigned version
; /* The version of the name to copy. */
219 tree new_name
; /* The new name used in the copy. */
220 tree field
; /* The field of the structure used to pass the
224 /* Equality and hash functions for hashtab code. */
227 name_to_copy_elt_eq (const void *aa
, const void *bb
)
229 const struct name_to_copy_elt
*a
= (const struct name_to_copy_elt
*) aa
;
230 const struct name_to_copy_elt
*b
= (const struct name_to_copy_elt
*) bb
;
232 return a
->version
== b
->version
;
236 name_to_copy_elt_hash (const void *aa
)
238 const struct name_to_copy_elt
*a
= (const struct name_to_copy_elt
*) aa
;
240 return (hashval_t
) a
->version
;
243 /* A transformation matrix, which is a self-contained ROWSIZE x COLSIZE
244 matrix. Rather than use floats, we simply keep a single DENOMINATOR that
245 represents the denominator for every element in the matrix. */
246 typedef struct lambda_trans_matrix_s
248 lambda_matrix matrix
;
252 } *lambda_trans_matrix
;
253 #define LTM_MATRIX(T) ((T)->matrix)
254 #define LTM_ROWSIZE(T) ((T)->rowsize)
255 #define LTM_COLSIZE(T) ((T)->colsize)
256 #define LTM_DENOMINATOR(T) ((T)->denominator)
258 /* Allocate a new transformation matrix. */
260 static lambda_trans_matrix
261 lambda_trans_matrix_new (int colsize
, int rowsize
,
262 struct obstack
* lambda_obstack
)
264 lambda_trans_matrix ret
;
266 ret
= (lambda_trans_matrix
)
267 obstack_alloc (lambda_obstack
, sizeof (struct lambda_trans_matrix_s
));
268 LTM_MATRIX (ret
) = lambda_matrix_new (rowsize
, colsize
, lambda_obstack
);
269 LTM_ROWSIZE (ret
) = rowsize
;
270 LTM_COLSIZE (ret
) = colsize
;
271 LTM_DENOMINATOR (ret
) = 1;
275 /* Multiply a vector VEC by a matrix MAT.
276 MAT is an M*N matrix, and VEC is a vector with length N. The result
277 is stored in DEST which must be a vector of length M. */
280 lambda_matrix_vector_mult (lambda_matrix matrix
, int m
, int n
,
281 lambda_vector vec
, lambda_vector dest
)
285 lambda_vector_clear (dest
, m
);
286 for (i
= 0; i
< m
; i
++)
287 for (j
= 0; j
< n
; j
++)
288 dest
[i
] += matrix
[i
][j
] * vec
[j
];
291 /* Return true if TRANS is a legal transformation matrix that respects
292 the dependence vectors in DISTS and DIRS. The conservative answer
295 "Wolfe proves that a unimodular transformation represented by the
296 matrix T is legal when applied to a loop nest with a set of
297 lexicographically non-negative distance vectors RDG if and only if
298 for each vector d in RDG, (T.d >= 0) is lexicographically positive.
299 i.e.: if and only if it transforms the lexicographically positive
300 distance vectors to lexicographically positive vectors. Note that
301 a unimodular matrix must transform the zero vector (and only it) to
302 the zero vector." S.Muchnick. */
305 lambda_transform_legal_p (lambda_trans_matrix trans
,
307 VEC (ddr_p
, heap
) *dependence_relations
)
310 lambda_vector distres
;
311 struct data_dependence_relation
*ddr
;
313 gcc_assert (LTM_COLSIZE (trans
) == nb_loops
314 && LTM_ROWSIZE (trans
) == nb_loops
);
316 /* When there are no dependences, the transformation is correct. */
317 if (VEC_length (ddr_p
, dependence_relations
) == 0)
320 ddr
= VEC_index (ddr_p
, dependence_relations
, 0);
324 /* When there is an unknown relation in the dependence_relations, we
325 know that it is no worth looking at this loop nest: give up. */
326 if (DDR_ARE_DEPENDENT (ddr
) == chrec_dont_know
)
329 distres
= lambda_vector_new (nb_loops
);
331 /* For each distance vector in the dependence graph. */
332 FOR_EACH_VEC_ELT (ddr_p
, dependence_relations
, i
, ddr
)
334 /* Don't care about relations for which we know that there is no
335 dependence, nor about read-read (aka. output-dependences):
336 these data accesses can happen in any order. */
337 if (DDR_ARE_DEPENDENT (ddr
) == chrec_known
338 || (DR_IS_READ (DDR_A (ddr
)) && DR_IS_READ (DDR_B (ddr
))))
341 /* Conservatively answer: "this transformation is not valid". */
342 if (DDR_ARE_DEPENDENT (ddr
) == chrec_dont_know
)
345 /* If the dependence could not be captured by a distance vector,
346 conservatively answer that the transform is not valid. */
347 if (DDR_NUM_DIST_VECTS (ddr
) == 0)
350 /* Compute trans.dist_vect */
351 for (j
= 0; j
< DDR_NUM_DIST_VECTS (ddr
); j
++)
353 lambda_matrix_vector_mult (LTM_MATRIX (trans
), nb_loops
, nb_loops
,
354 DDR_DIST_VECT (ddr
, j
), distres
);
356 if (!lambda_vector_lexico_pos (distres
, nb_loops
))
363 /* Data dependency analysis. Returns true if the iterations of LOOP
364 are independent on each other (that is, if we can execute them
368 loop_parallel_p (struct loop
*loop
, struct obstack
* parloop_obstack
)
370 VEC (loop_p
, heap
) *loop_nest
;
371 VEC (ddr_p
, heap
) *dependence_relations
;
372 VEC (data_reference_p
, heap
) *datarefs
;
373 lambda_trans_matrix trans
;
376 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
378 fprintf (dump_file
, "Considering loop %d\n", loop
->num
);
380 fprintf (dump_file
, "loop is innermost\n");
382 fprintf (dump_file
, "loop NOT innermost\n");
385 /* Check for problems with dependences. If the loop can be reversed,
386 the iterations are independent. */
387 datarefs
= VEC_alloc (data_reference_p
, heap
, 10);
388 dependence_relations
= VEC_alloc (ddr_p
, heap
, 10 * 10);
389 loop_nest
= VEC_alloc (loop_p
, heap
, 3);
390 if (! compute_data_dependences_for_loop (loop
, true, &loop_nest
, &datarefs
,
391 &dependence_relations
))
393 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
394 fprintf (dump_file
, " FAILED: cannot analyze data dependencies\n");
398 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
399 dump_data_dependence_relations (dump_file
, dependence_relations
);
401 trans
= lambda_trans_matrix_new (1, 1, parloop_obstack
);
402 LTM_MATRIX (trans
)[0][0] = -1;
404 if (lambda_transform_legal_p (trans
, 1, dependence_relations
))
407 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
408 fprintf (dump_file
, " SUCCESS: may be parallelized\n");
410 else if (dump_file
&& (dump_flags
& TDF_DETAILS
))
412 " FAILED: data dependencies exist across iterations\n");
415 VEC_free (loop_p
, heap
, loop_nest
);
416 free_dependence_relations (dependence_relations
);
417 free_data_refs (datarefs
);
422 /* Return true when LOOP contains basic blocks marked with the
423 BB_IRREDUCIBLE_LOOP flag. */
426 loop_has_blocks_with_irreducible_flag (struct loop
*loop
)
429 basic_block
*bbs
= get_loop_body_in_dom_order (loop
);
432 for (i
= 0; i
< loop
->num_nodes
; i
++)
433 if (bbs
[i
]->flags
& BB_IRREDUCIBLE_LOOP
)
442 /* Assigns the address of OBJ in TYPE to an ssa name, and returns this name.
443 The assignment statement is placed on edge ENTRY. DECL_ADDRESS maps decls
444 to their addresses that can be reused. The address of OBJ is known to
445 be invariant in the whole function. Other needed statements are placed
449 take_address_of (tree obj
, tree type
, edge entry
, htab_t decl_address
,
450 gimple_stmt_iterator
*gsi
)
454 struct int_tree_map ielt
, *nielt
;
455 tree
*var_p
, name
, addr
;
459 /* Since the address of OBJ is invariant, the trees may be shared.
460 Avoid rewriting unrelated parts of the code. */
461 obj
= unshare_expr (obj
);
463 handled_component_p (*var_p
);
464 var_p
= &TREE_OPERAND (*var_p
, 0))
467 /* Canonicalize the access to base on a MEM_REF. */
469 *var_p
= build_simple_mem_ref (build_fold_addr_expr (*var_p
));
471 /* Assign a canonical SSA name to the address of the base decl used
472 in the address and share it for all accesses and addresses based
474 uid
= DECL_UID (TREE_OPERAND (TREE_OPERAND (*var_p
, 0), 0));
476 dslot
= htab_find_slot_with_hash (decl_address
, &ielt
, uid
, INSERT
);
481 addr
= TREE_OPERAND (*var_p
, 0);
482 name
= make_temp_ssa_name (TREE_TYPE (addr
), NULL
,
483 get_name (TREE_OPERAND
484 (TREE_OPERAND (*var_p
, 0), 0)));
485 stmt
= gimple_build_assign (name
, addr
);
486 gsi_insert_on_edge_immediate (entry
, stmt
);
488 nielt
= XNEW (struct int_tree_map
);
494 name
= ((struct int_tree_map
*) *dslot
)->to
;
496 /* Express the address in terms of the canonical SSA name. */
497 TREE_OPERAND (*var_p
, 0) = name
;
499 return build_fold_addr_expr_with_type (obj
, type
);
501 name
= force_gimple_operand (build_addr (obj
, current_function_decl
),
502 &stmts
, true, NULL_TREE
);
503 if (!gimple_seq_empty_p (stmts
))
504 gsi_insert_seq_before (gsi
, stmts
, GSI_SAME_STMT
);
506 if (!useless_type_conversion_p (type
, TREE_TYPE (name
)))
508 name
= force_gimple_operand (fold_convert (type
, name
), &stmts
, true,
510 if (!gimple_seq_empty_p (stmts
))
511 gsi_insert_seq_before (gsi
, stmts
, GSI_SAME_STMT
);
517 /* Callback for htab_traverse. Create the initialization statement
518 for reduction described in SLOT, and place it at the preheader of
519 the loop described in DATA. */
522 initialize_reductions (void **slot
, void *data
)
525 tree bvar
, type
, arg
;
528 struct reduction_info
*const reduc
= (struct reduction_info
*) *slot
;
529 struct loop
*loop
= (struct loop
*) data
;
531 /* Create initialization in preheader:
532 reduction_variable = initialization value of reduction. */
534 /* In the phi node at the header, replace the argument coming
535 from the preheader with the reduction initialization value. */
537 /* Create a new variable to initialize the reduction. */
538 type
= TREE_TYPE (PHI_RESULT (reduc
->reduc_phi
));
539 bvar
= create_tmp_var (type
, "reduction");
541 c
= build_omp_clause (gimple_location (reduc
->reduc_stmt
),
542 OMP_CLAUSE_REDUCTION
);
543 OMP_CLAUSE_REDUCTION_CODE (c
) = reduc
->reduction_code
;
544 OMP_CLAUSE_DECL (c
) = SSA_NAME_VAR (gimple_assign_lhs (reduc
->reduc_stmt
));
546 init
= omp_reduction_init (c
, TREE_TYPE (bvar
));
549 /* Replace the argument representing the initialization value
550 with the initialization value for the reduction (neutral
551 element for the particular operation, e.g. 0 for PLUS_EXPR,
552 1 for MULT_EXPR, etc).
553 Keep the old value in a new variable "reduction_initial",
554 that will be taken in consideration after the parallel
555 computing is done. */
557 e
= loop_preheader_edge (loop
);
558 arg
= PHI_ARG_DEF_FROM_EDGE (reduc
->reduc_phi
, e
);
559 /* Create new variable to hold the initial value. */
561 SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE
562 (reduc
->reduc_phi
, loop_preheader_edge (loop
)), init
);
563 reduc
->initial_value
= arg
;
569 struct walk_stmt_info info
;
572 gimple_stmt_iterator
*gsi
;
577 /* Eliminates references to local variables in *TP out of the single
578 entry single exit region starting at DTA->ENTRY.
579 DECL_ADDRESS contains addresses of the references that had their
580 address taken already. If the expression is changed, CHANGED is
581 set to true. Callback for walk_tree. */
584 eliminate_local_variables_1 (tree
*tp
, int *walk_subtrees
, void *data
)
586 struct elv_data
*const dta
= (struct elv_data
*) data
;
587 tree t
= *tp
, var
, addr
, addr_type
, type
, obj
;
593 if (!SSA_VAR_P (t
) || DECL_EXTERNAL (t
))
596 type
= TREE_TYPE (t
);
597 addr_type
= build_pointer_type (type
);
598 addr
= take_address_of (t
, addr_type
, dta
->entry
, dta
->decl_address
,
600 if (dta
->gsi
== NULL
&& addr
== NULL_TREE
)
606 *tp
= build_simple_mem_ref (addr
);
612 if (TREE_CODE (t
) == ADDR_EXPR
)
614 /* ADDR_EXPR may appear in two contexts:
615 -- as a gimple operand, when the address taken is a function invariant
616 -- as gimple rhs, when the resulting address in not a function
618 We do not need to do anything special in the latter case (the base of
619 the memory reference whose address is taken may be replaced in the
620 DECL_P case). The former case is more complicated, as we need to
621 ensure that the new address is still a gimple operand. Thus, it
622 is not sufficient to replace just the base of the memory reference --
623 we need to move the whole computation of the address out of the
625 if (!is_gimple_val (t
))
629 obj
= TREE_OPERAND (t
, 0);
630 var
= get_base_address (obj
);
631 if (!var
|| !SSA_VAR_P (var
) || DECL_EXTERNAL (var
))
634 addr_type
= TREE_TYPE (t
);
635 addr
= take_address_of (obj
, addr_type
, dta
->entry
, dta
->decl_address
,
637 if (dta
->gsi
== NULL
&& addr
== NULL_TREE
)
654 /* Moves the references to local variables in STMT at *GSI out of the single
655 entry single exit region starting at ENTRY. DECL_ADDRESS contains
656 addresses of the references that had their address taken
660 eliminate_local_variables_stmt (edge entry
, gimple_stmt_iterator
*gsi
,
664 gimple stmt
= gsi_stmt (*gsi
);
666 memset (&dta
.info
, '\0', sizeof (dta
.info
));
668 dta
.decl_address
= decl_address
;
672 if (gimple_debug_bind_p (stmt
))
675 walk_tree (gimple_debug_bind_get_value_ptr (stmt
),
676 eliminate_local_variables_1
, &dta
.info
, NULL
);
679 gimple_debug_bind_reset_value (stmt
);
686 walk_gimple_op (stmt
, eliminate_local_variables_1
, &dta
.info
);
693 /* Eliminates the references to local variables from the single entry
694 single exit region between the ENTRY and EXIT edges.
697 1) Taking address of a local variable -- these are moved out of the
698 region (and temporary variable is created to hold the address if
701 2) Dereferencing a local variable -- these are replaced with indirect
705 eliminate_local_variables (edge entry
, edge exit
)
708 VEC (basic_block
, heap
) *body
= VEC_alloc (basic_block
, heap
, 3);
710 gimple_stmt_iterator gsi
;
711 bool has_debug_stmt
= false;
712 htab_t decl_address
= htab_create (10, int_tree_map_hash
, int_tree_map_eq
,
714 basic_block entry_bb
= entry
->src
;
715 basic_block exit_bb
= exit
->dest
;
717 gather_blocks_in_sese_region (entry_bb
, exit_bb
, &body
);
719 FOR_EACH_VEC_ELT (basic_block
, body
, i
, bb
)
720 if (bb
!= entry_bb
&& bb
!= exit_bb
)
721 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
722 if (is_gimple_debug (gsi_stmt (gsi
)))
724 if (gimple_debug_bind_p (gsi_stmt (gsi
)))
725 has_debug_stmt
= true;
728 eliminate_local_variables_stmt (entry
, &gsi
, decl_address
);
731 FOR_EACH_VEC_ELT (basic_block
, body
, i
, bb
)
732 if (bb
!= entry_bb
&& bb
!= exit_bb
)
733 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
734 if (gimple_debug_bind_p (gsi_stmt (gsi
)))
735 eliminate_local_variables_stmt (entry
, &gsi
, decl_address
);
737 htab_delete (decl_address
);
738 VEC_free (basic_block
, heap
, body
);
741 /* Returns true if expression EXPR is not defined between ENTRY and
742 EXIT, i.e. if all its operands are defined outside of the region. */
745 expr_invariant_in_region_p (edge entry
, edge exit
, tree expr
)
747 basic_block entry_bb
= entry
->src
;
748 basic_block exit_bb
= exit
->dest
;
751 if (is_gimple_min_invariant (expr
))
754 if (TREE_CODE (expr
) == SSA_NAME
)
756 def_bb
= gimple_bb (SSA_NAME_DEF_STMT (expr
));
758 && dominated_by_p (CDI_DOMINATORS
, def_bb
, entry_bb
)
759 && !dominated_by_p (CDI_DOMINATORS
, def_bb
, exit_bb
))
768 /* If COPY_NAME_P is true, creates and returns a duplicate of NAME.
769 The copies are stored to NAME_COPIES, if NAME was already duplicated,
770 its duplicate stored in NAME_COPIES is returned.
772 Regardless of COPY_NAME_P, the decl used as a base of the ssa name is also
773 duplicated, storing the copies in DECL_COPIES. */
776 separate_decls_in_region_name (tree name
,
777 htab_t name_copies
, htab_t decl_copies
,
780 tree copy
, var
, var_copy
;
781 unsigned idx
, uid
, nuid
;
782 struct int_tree_map ielt
, *nielt
;
783 struct name_to_copy_elt elt
, *nelt
;
784 void **slot
, **dslot
;
786 if (TREE_CODE (name
) != SSA_NAME
)
789 idx
= SSA_NAME_VERSION (name
);
791 slot
= htab_find_slot_with_hash (name_copies
, &elt
, idx
,
792 copy_name_p
? INSERT
: NO_INSERT
);
794 return ((struct name_to_copy_elt
*) *slot
)->new_name
;
798 copy
= duplicate_ssa_name (name
, NULL
);
799 nelt
= XNEW (struct name_to_copy_elt
);
801 nelt
->new_name
= copy
;
802 nelt
->field
= NULL_TREE
;
811 var
= SSA_NAME_VAR (name
);
815 uid
= DECL_UID (var
);
817 dslot
= htab_find_slot_with_hash (decl_copies
, &ielt
, uid
, INSERT
);
820 var_copy
= create_tmp_var (TREE_TYPE (var
), get_name (var
));
821 DECL_GIMPLE_REG_P (var_copy
) = DECL_GIMPLE_REG_P (var
);
822 nielt
= XNEW (struct int_tree_map
);
824 nielt
->to
= var_copy
;
827 /* Ensure that when we meet this decl next time, we won't duplicate
829 nuid
= DECL_UID (var_copy
);
831 dslot
= htab_find_slot_with_hash (decl_copies
, &ielt
, nuid
, INSERT
);
832 gcc_assert (!*dslot
);
833 nielt
= XNEW (struct int_tree_map
);
835 nielt
->to
= var_copy
;
839 var_copy
= ((struct int_tree_map
*) *dslot
)->to
;
841 replace_ssa_name_symbol (copy
, var_copy
);
845 /* Finds the ssa names used in STMT that are defined outside the
846 region between ENTRY and EXIT and replaces such ssa names with
847 their duplicates. The duplicates are stored to NAME_COPIES. Base
848 decls of all ssa names used in STMT (including those defined in
849 LOOP) are replaced with the new temporary variables; the
850 replacement decls are stored in DECL_COPIES. */
853 separate_decls_in_region_stmt (edge entry
, edge exit
, gimple stmt
,
854 htab_t name_copies
, htab_t decl_copies
)
862 FOR_EACH_PHI_OR_STMT_DEF (def
, stmt
, oi
, SSA_OP_DEF
)
864 name
= DEF_FROM_PTR (def
);
865 gcc_assert (TREE_CODE (name
) == SSA_NAME
);
866 copy
= separate_decls_in_region_name (name
, name_copies
, decl_copies
,
868 gcc_assert (copy
== name
);
871 FOR_EACH_PHI_OR_STMT_USE (use
, stmt
, oi
, SSA_OP_USE
)
873 name
= USE_FROM_PTR (use
);
874 if (TREE_CODE (name
) != SSA_NAME
)
877 copy_name_p
= expr_invariant_in_region_p (entry
, exit
, name
);
878 copy
= separate_decls_in_region_name (name
, name_copies
, decl_copies
,
884 /* Finds the ssa names used in STMT that are defined outside the
885 region between ENTRY and EXIT and replaces such ssa names with
886 their duplicates. The duplicates are stored to NAME_COPIES. Base
887 decls of all ssa names used in STMT (including those defined in
888 LOOP) are replaced with the new temporary variables; the
889 replacement decls are stored in DECL_COPIES. */
892 separate_decls_in_region_debug (gimple stmt
, htab_t name_copies
,
898 struct int_tree_map ielt
;
899 struct name_to_copy_elt elt
;
900 void **slot
, **dslot
;
902 if (gimple_debug_bind_p (stmt
))
903 var
= gimple_debug_bind_get_var (stmt
);
904 else if (gimple_debug_source_bind_p (stmt
))
905 var
= gimple_debug_source_bind_get_var (stmt
);
908 if (TREE_CODE (var
) == DEBUG_EXPR_DECL
|| TREE_CODE (var
) == LABEL_DECL
)
910 gcc_assert (DECL_P (var
) && SSA_VAR_P (var
));
911 ielt
.uid
= DECL_UID (var
);
912 dslot
= htab_find_slot_with_hash (decl_copies
, &ielt
, ielt
.uid
, NO_INSERT
);
915 if (gimple_debug_bind_p (stmt
))
916 gimple_debug_bind_set_var (stmt
, ((struct int_tree_map
*) *dslot
)->to
);
917 else if (gimple_debug_source_bind_p (stmt
))
918 gimple_debug_source_bind_set_var (stmt
, ((struct int_tree_map
*) *dslot
)->to
);
920 FOR_EACH_PHI_OR_STMT_USE (use
, stmt
, oi
, SSA_OP_USE
)
922 name
= USE_FROM_PTR (use
);
923 if (TREE_CODE (name
) != SSA_NAME
)
926 elt
.version
= SSA_NAME_VERSION (name
);
927 slot
= htab_find_slot_with_hash (name_copies
, &elt
, elt
.version
, NO_INSERT
);
930 gimple_debug_bind_reset_value (stmt
);
935 SET_USE (use
, ((struct name_to_copy_elt
*) *slot
)->new_name
);
941 /* Callback for htab_traverse. Adds a field corresponding to the reduction
942 specified in SLOT. The type is passed in DATA. */
945 add_field_for_reduction (void **slot
, void *data
)
948 struct reduction_info
*const red
= (struct reduction_info
*) *slot
;
949 tree
const type
= (tree
) data
;
950 tree var
= SSA_NAME_VAR (gimple_assign_lhs (red
->reduc_stmt
));
951 tree field
= build_decl (gimple_location (red
->reduc_stmt
),
952 FIELD_DECL
, DECL_NAME (var
), TREE_TYPE (var
));
954 insert_field_into_struct (type
, field
);
961 /* Callback for htab_traverse. Adds a field corresponding to a ssa name
962 described in SLOT. The type is passed in DATA. */
965 add_field_for_name (void **slot
, void *data
)
967 struct name_to_copy_elt
*const elt
= (struct name_to_copy_elt
*) *slot
;
968 tree type
= (tree
) data
;
969 tree name
= ssa_name (elt
->version
);
970 tree field
= build_decl (UNKNOWN_LOCATION
,
971 FIELD_DECL
, SSA_NAME_IDENTIFIER (name
),
974 insert_field_into_struct (type
, field
);
980 /* Callback for htab_traverse. A local result is the intermediate result
982 thread, or the initial value in case no iteration was executed.
983 This function creates a phi node reflecting these values.
984 The phi's result will be stored in NEW_PHI field of the
985 reduction's data structure. */
988 create_phi_for_local_result (void **slot
, void *data
)
990 struct reduction_info
*const reduc
= (struct reduction_info
*) *slot
;
991 const struct loop
*const loop
= (const struct loop
*) data
;
994 basic_block store_bb
;
996 source_location locus
;
998 /* STORE_BB is the block where the phi
999 should be stored. It is the destination of the loop exit.
1000 (Find the fallthru edge from GIMPLE_OMP_CONTINUE). */
1001 store_bb
= FALLTHRU_EDGE (loop
->latch
)->dest
;
1003 /* STORE_BB has two predecessors. One coming from the loop
1004 (the reduction's result is computed at the loop),
1005 and another coming from a block preceding the loop,
1007 are executed (the initial value should be taken). */
1008 if (EDGE_PRED (store_bb
, 0) == FALLTHRU_EDGE (loop
->latch
))
1009 e
= EDGE_PRED (store_bb
, 1);
1011 e
= EDGE_PRED (store_bb
, 0);
1012 local_res
= copy_ssa_name (gimple_assign_lhs (reduc
->reduc_stmt
), NULL
);
1013 locus
= gimple_location (reduc
->reduc_stmt
);
1014 new_phi
= create_phi_node (local_res
, store_bb
);
1015 add_phi_arg (new_phi
, reduc
->init
, e
, locus
);
1016 add_phi_arg (new_phi
, gimple_assign_lhs (reduc
->reduc_stmt
),
1017 FALLTHRU_EDGE (loop
->latch
), locus
);
1018 reduc
->new_phi
= new_phi
;
1028 basic_block store_bb
;
1029 basic_block load_bb
;
1032 /* Callback for htab_traverse. Create an atomic instruction for the
1033 reduction described in SLOT.
1034 DATA annotates the place in memory the atomic operation relates to,
1035 and the basic block it needs to be generated in. */
1038 create_call_for_reduction_1 (void **slot
, void *data
)
1040 struct reduction_info
*const reduc
= (struct reduction_info
*) *slot
;
1041 struct clsn_data
*const clsn_data
= (struct clsn_data
*) data
;
1042 gimple_stmt_iterator gsi
;
1043 tree type
= TREE_TYPE (PHI_RESULT (reduc
->reduc_phi
));
1048 tree t
, addr
, ref
, x
;
1049 tree tmp_load
, name
;
1052 load_struct
= build_simple_mem_ref (clsn_data
->load
);
1053 t
= build3 (COMPONENT_REF
, type
, load_struct
, reduc
->field
, NULL_TREE
);
1055 addr
= build_addr (t
, current_function_decl
);
1057 /* Create phi node. */
1058 bb
= clsn_data
->load_bb
;
1060 e
= split_block (bb
, t
);
1063 tmp_load
= create_tmp_var (TREE_TYPE (TREE_TYPE (addr
)), NULL
);
1064 tmp_load
= make_ssa_name (tmp_load
, NULL
);
1065 load
= gimple_build_omp_atomic_load (tmp_load
, addr
);
1066 SSA_NAME_DEF_STMT (tmp_load
) = load
;
1067 gsi
= gsi_start_bb (new_bb
);
1068 gsi_insert_after (&gsi
, load
, GSI_NEW_STMT
);
1070 e
= split_block (new_bb
, load
);
1072 gsi
= gsi_start_bb (new_bb
);
1074 x
= fold_build2 (reduc
->reduction_code
,
1075 TREE_TYPE (PHI_RESULT (reduc
->new_phi
)), ref
,
1076 PHI_RESULT (reduc
->new_phi
));
1078 name
= force_gimple_operand_gsi (&gsi
, x
, true, NULL_TREE
, true,
1079 GSI_CONTINUE_LINKING
);
1081 gsi_insert_after (&gsi
, gimple_build_omp_atomic_store (name
), GSI_NEW_STMT
);
1085 /* Create the atomic operation at the join point of the threads.
1086 REDUCTION_LIST describes the reductions in the LOOP.
1087 LD_ST_DATA describes the shared data structure where
1088 shared data is stored in and loaded from. */
1090 create_call_for_reduction (struct loop
*loop
, htab_t reduction_list
,
1091 struct clsn_data
*ld_st_data
)
1093 htab_traverse (reduction_list
, create_phi_for_local_result
, loop
);
1094 /* Find the fallthru edge from GIMPLE_OMP_CONTINUE. */
1095 ld_st_data
->load_bb
= FALLTHRU_EDGE (loop
->latch
)->dest
;
1096 htab_traverse (reduction_list
, create_call_for_reduction_1
, ld_st_data
);
1099 /* Callback for htab_traverse. Loads the final reduction value at the
1100 join point of all threads, and inserts it in the right place. */
1103 create_loads_for_reductions (void **slot
, void *data
)
1105 struct reduction_info
*const red
= (struct reduction_info
*) *slot
;
1106 struct clsn_data
*const clsn_data
= (struct clsn_data
*) data
;
1108 gimple_stmt_iterator gsi
;
1109 tree type
= TREE_TYPE (gimple_assign_lhs (red
->reduc_stmt
));
1114 gsi
= gsi_after_labels (clsn_data
->load_bb
);
1115 load_struct
= build_simple_mem_ref (clsn_data
->load
);
1116 load_struct
= build3 (COMPONENT_REF
, type
, load_struct
, red
->field
,
1120 name
= PHI_RESULT (red
->keep_res
);
1121 stmt
= gimple_build_assign (name
, x
);
1122 SSA_NAME_DEF_STMT (name
) = stmt
;
1124 gsi_insert_after (&gsi
, stmt
, GSI_NEW_STMT
);
1126 for (gsi
= gsi_start_phis (gimple_bb (red
->keep_res
));
1127 !gsi_end_p (gsi
); gsi_next (&gsi
))
1128 if (gsi_stmt (gsi
) == red
->keep_res
)
1130 remove_phi_node (&gsi
, false);
1136 /* Load the reduction result that was stored in LD_ST_DATA.
1137 REDUCTION_LIST describes the list of reductions that the
1138 loads should be generated for. */
1140 create_final_loads_for_reduction (htab_t reduction_list
,
1141 struct clsn_data
*ld_st_data
)
1143 gimple_stmt_iterator gsi
;
1147 gsi
= gsi_after_labels (ld_st_data
->load_bb
);
1148 t
= build_fold_addr_expr (ld_st_data
->store
);
1149 stmt
= gimple_build_assign (ld_st_data
->load
, t
);
1151 gsi_insert_before (&gsi
, stmt
, GSI_NEW_STMT
);
1152 SSA_NAME_DEF_STMT (ld_st_data
->load
) = stmt
;
1154 htab_traverse (reduction_list
, create_loads_for_reductions
, ld_st_data
);
1158 /* Callback for htab_traverse. Store the neutral value for the
1159 particular reduction's operation, e.g. 0 for PLUS_EXPR,
1160 1 for MULT_EXPR, etc. into the reduction field.
1161 The reduction is specified in SLOT. The store information is
1165 create_stores_for_reduction (void **slot
, void *data
)
1167 struct reduction_info
*const red
= (struct reduction_info
*) *slot
;
1168 struct clsn_data
*const clsn_data
= (struct clsn_data
*) data
;
1171 gimple_stmt_iterator gsi
;
1172 tree type
= TREE_TYPE (gimple_assign_lhs (red
->reduc_stmt
));
1174 gsi
= gsi_last_bb (clsn_data
->store_bb
);
1175 t
= build3 (COMPONENT_REF
, type
, clsn_data
->store
, red
->field
, NULL_TREE
);
1176 stmt
= gimple_build_assign (t
, red
->initial_value
);
1177 gsi_insert_after (&gsi
, stmt
, GSI_NEW_STMT
);
1182 /* Callback for htab_traverse. Creates loads to a field of LOAD in LOAD_BB and
1183 store to a field of STORE in STORE_BB for the ssa name and its duplicate
1184 specified in SLOT. */
1187 create_loads_and_stores_for_name (void **slot
, void *data
)
1189 struct name_to_copy_elt
*const elt
= (struct name_to_copy_elt
*) *slot
;
1190 struct clsn_data
*const clsn_data
= (struct clsn_data
*) data
;
1193 gimple_stmt_iterator gsi
;
1194 tree type
= TREE_TYPE (elt
->new_name
);
1197 gsi
= gsi_last_bb (clsn_data
->store_bb
);
1198 t
= build3 (COMPONENT_REF
, type
, clsn_data
->store
, elt
->field
, NULL_TREE
);
1199 stmt
= gimple_build_assign (t
, ssa_name (elt
->version
));
1200 gsi_insert_after (&gsi
, stmt
, GSI_NEW_STMT
);
1202 gsi
= gsi_last_bb (clsn_data
->load_bb
);
1203 load_struct
= build_simple_mem_ref (clsn_data
->load
);
1204 t
= build3 (COMPONENT_REF
, type
, load_struct
, elt
->field
, NULL_TREE
);
1205 stmt
= gimple_build_assign (elt
->new_name
, t
);
1206 SSA_NAME_DEF_STMT (elt
->new_name
) = stmt
;
1207 gsi_insert_after (&gsi
, stmt
, GSI_NEW_STMT
);
1212 /* Moves all the variables used in LOOP and defined outside of it (including
1213 the initial values of loop phi nodes, and *PER_THREAD if it is a ssa
1214 name) to a structure created for this purpose. The code
1222 is transformed this way:
1237 `old' is stored to *ARG_STRUCT and `new' is stored to NEW_ARG_STRUCT. The
1238 pointer `new' is intentionally not initialized (the loop will be split to a
1239 separate function later, and `new' will be initialized from its arguments).
1240 LD_ST_DATA holds information about the shared data structure used to pass
1241 information among the threads. It is initialized here, and
1242 gen_parallel_loop will pass it to create_call_for_reduction that
1243 needs this information. REDUCTION_LIST describes the reductions
1247 separate_decls_in_region (edge entry
, edge exit
, htab_t reduction_list
,
1248 tree
*arg_struct
, tree
*new_arg_struct
,
1249 struct clsn_data
*ld_st_data
)
1252 basic_block bb1
= split_edge (entry
);
1253 basic_block bb0
= single_pred (bb1
);
1254 htab_t name_copies
= htab_create (10, name_to_copy_elt_hash
,
1255 name_to_copy_elt_eq
, free
);
1256 htab_t decl_copies
= htab_create (10, int_tree_map_hash
, int_tree_map_eq
,
1259 tree type
, type_name
, nvar
;
1260 gimple_stmt_iterator gsi
;
1261 struct clsn_data clsn_data
;
1262 VEC (basic_block
, heap
) *body
= VEC_alloc (basic_block
, heap
, 3);
1264 basic_block entry_bb
= bb1
;
1265 basic_block exit_bb
= exit
->dest
;
1266 bool has_debug_stmt
= false;
1268 entry
= single_succ_edge (entry_bb
);
1269 gather_blocks_in_sese_region (entry_bb
, exit_bb
, &body
);
1271 FOR_EACH_VEC_ELT (basic_block
, body
, i
, bb
)
1273 if (bb
!= entry_bb
&& bb
!= exit_bb
)
1275 for (gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1276 separate_decls_in_region_stmt (entry
, exit
, gsi_stmt (gsi
),
1277 name_copies
, decl_copies
);
1279 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1281 gimple stmt
= gsi_stmt (gsi
);
1283 if (is_gimple_debug (stmt
))
1284 has_debug_stmt
= true;
1286 separate_decls_in_region_stmt (entry
, exit
, stmt
,
1287 name_copies
, decl_copies
);
1292 /* Now process debug bind stmts. We must not create decls while
1293 processing debug stmts, so we defer their processing so as to
1294 make sure we will have debug info for as many variables as
1295 possible (all of those that were dealt with in the loop above),
1296 and discard those for which we know there's nothing we can
1299 FOR_EACH_VEC_ELT (basic_block
, body
, i
, bb
)
1300 if (bb
!= entry_bb
&& bb
!= exit_bb
)
1302 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
);)
1304 gimple stmt
= gsi_stmt (gsi
);
1306 if (is_gimple_debug (stmt
))
1308 if (separate_decls_in_region_debug (stmt
, name_copies
,
1311 gsi_remove (&gsi
, true);
1320 VEC_free (basic_block
, heap
, body
);
1322 if (htab_elements (name_copies
) == 0 && htab_elements (reduction_list
) == 0)
1324 /* It may happen that there is nothing to copy (if there are only
1325 loop carried and external variables in the loop). */
1327 *new_arg_struct
= NULL
;
1331 /* Create the type for the structure to store the ssa names to. */
1332 type
= lang_hooks
.types
.make_type (RECORD_TYPE
);
1333 type_name
= build_decl (UNKNOWN_LOCATION
,
1334 TYPE_DECL
, create_tmp_var_name (".paral_data"),
1336 TYPE_NAME (type
) = type_name
;
1338 htab_traverse (name_copies
, add_field_for_name
, type
);
1339 if (reduction_list
&& htab_elements (reduction_list
) > 0)
1341 /* Create the fields for reductions. */
1342 htab_traverse (reduction_list
, add_field_for_reduction
,
1347 /* Create the loads and stores. */
1348 *arg_struct
= create_tmp_var (type
, ".paral_data_store");
1349 nvar
= create_tmp_var (build_pointer_type (type
), ".paral_data_load");
1350 *new_arg_struct
= make_ssa_name (nvar
, NULL
);
1352 ld_st_data
->store
= *arg_struct
;
1353 ld_st_data
->load
= *new_arg_struct
;
1354 ld_st_data
->store_bb
= bb0
;
1355 ld_st_data
->load_bb
= bb1
;
1357 htab_traverse (name_copies
, create_loads_and_stores_for_name
,
1360 /* Load the calculation from memory (after the join of the threads). */
1362 if (reduction_list
&& htab_elements (reduction_list
) > 0)
1364 htab_traverse (reduction_list
, create_stores_for_reduction
,
1366 clsn_data
.load
= make_ssa_name (nvar
, NULL
);
1367 clsn_data
.load_bb
= exit
->dest
;
1368 clsn_data
.store
= ld_st_data
->store
;
1369 create_final_loads_for_reduction (reduction_list
, &clsn_data
);
1373 htab_delete (decl_copies
);
1374 htab_delete (name_copies
);
1377 /* Bitmap containing uids of functions created by parallelization. We cannot
1378 allocate it from the default obstack, as it must live across compilation
1379 of several functions; we make it gc allocated instead. */
1381 static GTY(()) bitmap parallelized_functions
;
1383 /* Returns true if FN was created by create_loop_fn. */
1386 parallelized_function_p (tree fn
)
1388 if (!parallelized_functions
|| !DECL_ARTIFICIAL (fn
))
1391 return bitmap_bit_p (parallelized_functions
, DECL_UID (fn
));
1394 /* Creates and returns an empty function that will receive the body of
1395 a parallelized loop. */
1398 create_loop_fn (location_t loc
)
1402 tree decl
, type
, name
, t
;
1403 struct function
*act_cfun
= cfun
;
1404 static unsigned loopfn_num
;
1406 snprintf (buf
, 100, "%s.$loopfn", current_function_name ());
1407 ASM_FORMAT_PRIVATE_NAME (tname
, buf
, loopfn_num
++);
1408 clean_symbol_name (tname
);
1409 name
= get_identifier (tname
);
1410 type
= build_function_type_list (void_type_node
, ptr_type_node
, NULL_TREE
);
1412 decl
= build_decl (loc
, FUNCTION_DECL
, name
, type
);
1413 if (!parallelized_functions
)
1414 parallelized_functions
= BITMAP_GGC_ALLOC ();
1415 bitmap_set_bit (parallelized_functions
, DECL_UID (decl
));
1417 TREE_STATIC (decl
) = 1;
1418 TREE_USED (decl
) = 1;
1419 DECL_ARTIFICIAL (decl
) = 1;
1420 DECL_IGNORED_P (decl
) = 0;
1421 TREE_PUBLIC (decl
) = 0;
1422 DECL_UNINLINABLE (decl
) = 1;
1423 DECL_EXTERNAL (decl
) = 0;
1424 DECL_CONTEXT (decl
) = NULL_TREE
;
1425 DECL_INITIAL (decl
) = make_node (BLOCK
);
1427 t
= build_decl (loc
, RESULT_DECL
, NULL_TREE
, void_type_node
);
1428 DECL_ARTIFICIAL (t
) = 1;
1429 DECL_IGNORED_P (t
) = 1;
1430 DECL_RESULT (decl
) = t
;
1432 t
= build_decl (loc
, PARM_DECL
, get_identifier (".paral_data_param"),
1434 DECL_ARTIFICIAL (t
) = 1;
1435 DECL_ARG_TYPE (t
) = ptr_type_node
;
1436 DECL_CONTEXT (t
) = decl
;
1438 DECL_ARGUMENTS (decl
) = t
;
1440 allocate_struct_function (decl
, false);
1442 /* The call to allocate_struct_function clobbers CFUN, so we need to restore
1444 set_cfun (act_cfun
);
1449 /* Moves the exit condition of LOOP to the beginning of its header, and
1450 duplicates the part of the last iteration that gets disabled to the
1451 exit of the loop. NIT is the number of iterations of the loop
1452 (used to initialize the variables in the duplicated part).
1454 TODO: the common case is that latch of the loop is empty and immediately
1455 follows the loop exit. In this case, it would be better not to copy the
1456 body of the loop, but only move the entry of the loop directly before the
1457 exit check and increase the number of iterations of the loop by one.
1458 This may need some additional preconditioning in case NIT = ~0.
1459 REDUCTION_LIST describes the reductions in LOOP. */
1462 transform_to_exit_first_loop (struct loop
*loop
, htab_t reduction_list
, tree nit
)
1464 basic_block
*bbs
, *nbbs
, ex_bb
, orig_header
;
1467 edge exit
= single_dom_exit (loop
), hpred
;
1468 tree control
, control_name
, res
, t
;
1469 gimple phi
, nphi
, cond_stmt
, stmt
, cond_nit
;
1470 gimple_stmt_iterator gsi
;
1473 split_block_after_labels (loop
->header
);
1474 orig_header
= single_succ (loop
->header
);
1475 hpred
= single_succ_edge (loop
->header
);
1477 cond_stmt
= last_stmt (exit
->src
);
1478 control
= gimple_cond_lhs (cond_stmt
);
1479 gcc_assert (gimple_cond_rhs (cond_stmt
) == nit
);
1481 /* Make sure that we have phi nodes on exit for all loop header phis
1482 (create_parallel_loop requires that). */
1483 for (gsi
= gsi_start_phis (loop
->header
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1485 phi
= gsi_stmt (gsi
);
1486 res
= PHI_RESULT (phi
);
1487 t
= copy_ssa_name (res
, phi
);
1488 SET_PHI_RESULT (phi
, t
);
1489 nphi
= create_phi_node (res
, orig_header
);
1490 add_phi_arg (nphi
, t
, hpred
, UNKNOWN_LOCATION
);
1494 gimple_cond_set_lhs (cond_stmt
, t
);
1495 update_stmt (cond_stmt
);
1500 bbs
= get_loop_body_in_dom_order (loop
);
1502 for (n
= 0; bbs
[n
] != exit
->src
; n
++)
1504 nbbs
= XNEWVEC (basic_block
, n
);
1505 ok
= gimple_duplicate_sese_tail (single_succ_edge (loop
->header
), exit
,
1512 /* Other than reductions, the only gimple reg that should be copied
1513 out of the loop is the control variable. */
1514 exit
= single_dom_exit (loop
);
1515 control_name
= NULL_TREE
;
1516 for (gsi
= gsi_start_phis (ex_bb
); !gsi_end_p (gsi
); )
1518 phi
= gsi_stmt (gsi
);
1519 res
= PHI_RESULT (phi
);
1520 if (virtual_operand_p (res
))
1526 /* Check if it is a part of reduction. If it is,
1527 keep the phi at the reduction's keep_res field. The
1528 PHI_RESULT of this phi is the resulting value of the reduction
1529 variable when exiting the loop. */
1531 if (htab_elements (reduction_list
) > 0)
1533 struct reduction_info
*red
;
1535 tree val
= PHI_ARG_DEF_FROM_EDGE (phi
, exit
);
1536 red
= reduction_phi (reduction_list
, SSA_NAME_DEF_STMT (val
));
1539 red
->keep_res
= phi
;
1544 gcc_assert (control_name
== NULL_TREE
1545 && SSA_NAME_VAR (res
) == SSA_NAME_VAR (control
));
1547 remove_phi_node (&gsi
, false);
1549 gcc_assert (control_name
!= NULL_TREE
);
1551 /* Initialize the control variable to number of iterations
1552 according to the rhs of the exit condition. */
1553 gsi
= gsi_after_labels (ex_bb
);
1554 cond_nit
= last_stmt (exit
->src
);
1555 nit_1
= gimple_cond_rhs (cond_nit
);
1556 nit_1
= force_gimple_operand_gsi (&gsi
,
1557 fold_convert (TREE_TYPE (control_name
), nit_1
),
1558 false, NULL_TREE
, false, GSI_SAME_STMT
);
1559 stmt
= gimple_build_assign (control_name
, nit_1
);
1560 gsi_insert_before (&gsi
, stmt
, GSI_NEW_STMT
);
1561 SSA_NAME_DEF_STMT (control_name
) = stmt
;
1564 /* Create the parallel constructs for LOOP as described in gen_parallel_loop.
1565 LOOP_FN and DATA are the arguments of GIMPLE_OMP_PARALLEL.
1566 NEW_DATA is the variable that should be initialized from the argument
1567 of LOOP_FN. N_THREADS is the requested number of threads. Returns the
1568 basic block containing GIMPLE_OMP_PARALLEL tree. */
1571 create_parallel_loop (struct loop
*loop
, tree loop_fn
, tree data
,
1572 tree new_data
, unsigned n_threads
, location_t loc
)
1574 gimple_stmt_iterator gsi
;
1575 basic_block bb
, paral_bb
, for_bb
, ex_bb
;
1577 gimple stmt
, for_stmt
, phi
, cond_stmt
;
1578 tree cvar
, cvar_init
, initvar
, cvar_next
, cvar_base
, type
;
1579 edge exit
, nexit
, guard
, end
, e
;
1581 /* Prepare the GIMPLE_OMP_PARALLEL statement. */
1582 bb
= loop_preheader_edge (loop
)->src
;
1583 paral_bb
= single_pred (bb
);
1584 gsi
= gsi_last_bb (paral_bb
);
1586 t
= build_omp_clause (loc
, OMP_CLAUSE_NUM_THREADS
);
1587 OMP_CLAUSE_NUM_THREADS_EXPR (t
)
1588 = build_int_cst (integer_type_node
, n_threads
);
1589 stmt
= gimple_build_omp_parallel (NULL
, t
, loop_fn
, data
);
1590 gimple_set_location (stmt
, loc
);
1592 gsi_insert_after (&gsi
, stmt
, GSI_NEW_STMT
);
1594 /* Initialize NEW_DATA. */
1597 gsi
= gsi_after_labels (bb
);
1599 param
= make_ssa_name (DECL_ARGUMENTS (loop_fn
), NULL
);
1600 stmt
= gimple_build_assign (param
, build_fold_addr_expr (data
));
1601 gsi_insert_before (&gsi
, stmt
, GSI_SAME_STMT
);
1602 SSA_NAME_DEF_STMT (param
) = stmt
;
1604 stmt
= gimple_build_assign (new_data
,
1605 fold_convert (TREE_TYPE (new_data
), param
));
1606 gsi_insert_before (&gsi
, stmt
, GSI_SAME_STMT
);
1607 SSA_NAME_DEF_STMT (new_data
) = stmt
;
1610 /* Emit GIMPLE_OMP_RETURN for GIMPLE_OMP_PARALLEL. */
1611 bb
= split_loop_exit_edge (single_dom_exit (loop
));
1612 gsi
= gsi_last_bb (bb
);
1613 stmt
= gimple_build_omp_return (false);
1614 gimple_set_location (stmt
, loc
);
1615 gsi_insert_after (&gsi
, stmt
, GSI_NEW_STMT
);
1617 /* Extract data for GIMPLE_OMP_FOR. */
1618 gcc_assert (loop
->header
== single_dom_exit (loop
)->src
);
1619 cond_stmt
= last_stmt (loop
->header
);
1621 cvar
= gimple_cond_lhs (cond_stmt
);
1622 cvar_base
= SSA_NAME_VAR (cvar
);
1623 phi
= SSA_NAME_DEF_STMT (cvar
);
1624 cvar_init
= PHI_ARG_DEF_FROM_EDGE (phi
, loop_preheader_edge (loop
));
1625 initvar
= copy_ssa_name (cvar
, NULL
);
1626 SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi
, loop_preheader_edge (loop
)),
1628 cvar_next
= PHI_ARG_DEF_FROM_EDGE (phi
, loop_latch_edge (loop
));
1630 gsi
= gsi_last_nondebug_bb (loop
->latch
);
1631 gcc_assert (gsi_stmt (gsi
) == SSA_NAME_DEF_STMT (cvar_next
));
1632 gsi_remove (&gsi
, true);
1635 for_bb
= split_edge (loop_preheader_edge (loop
));
1636 ex_bb
= split_loop_exit_edge (single_dom_exit (loop
));
1637 extract_true_false_edges_from_block (loop
->header
, &nexit
, &exit
);
1638 gcc_assert (exit
== single_dom_exit (loop
));
1640 guard
= make_edge (for_bb
, ex_bb
, 0);
1641 single_succ_edge (loop
->latch
)->flags
= 0;
1642 end
= make_edge (loop
->latch
, ex_bb
, EDGE_FALLTHRU
);
1643 for (gsi
= gsi_start_phis (ex_bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1645 source_location locus
;
1647 phi
= gsi_stmt (gsi
);
1648 stmt
= SSA_NAME_DEF_STMT (PHI_ARG_DEF_FROM_EDGE (phi
, exit
));
1650 def
= PHI_ARG_DEF_FROM_EDGE (stmt
, loop_preheader_edge (loop
));
1651 locus
= gimple_phi_arg_location_from_edge (stmt
,
1652 loop_preheader_edge (loop
));
1653 add_phi_arg (phi
, def
, guard
, locus
);
1655 def
= PHI_ARG_DEF_FROM_EDGE (stmt
, loop_latch_edge (loop
));
1656 locus
= gimple_phi_arg_location_from_edge (stmt
, loop_latch_edge (loop
));
1657 add_phi_arg (phi
, def
, end
, locus
);
1659 e
= redirect_edge_and_branch (exit
, nexit
->dest
);
1660 PENDING_STMT (e
) = NULL
;
1662 /* Emit GIMPLE_OMP_FOR. */
1663 gimple_cond_set_lhs (cond_stmt
, cvar_base
);
1664 type
= TREE_TYPE (cvar
);
1665 t
= build_omp_clause (loc
, OMP_CLAUSE_SCHEDULE
);
1666 OMP_CLAUSE_SCHEDULE_KIND (t
) = OMP_CLAUSE_SCHEDULE_STATIC
;
1668 for_stmt
= gimple_build_omp_for (NULL
, t
, 1, NULL
);
1669 gimple_set_location (for_stmt
, loc
);
1670 gimple_omp_for_set_index (for_stmt
, 0, initvar
);
1671 gimple_omp_for_set_initial (for_stmt
, 0, cvar_init
);
1672 gimple_omp_for_set_final (for_stmt
, 0, gimple_cond_rhs (cond_stmt
));
1673 gimple_omp_for_set_cond (for_stmt
, 0, gimple_cond_code (cond_stmt
));
1674 gimple_omp_for_set_incr (for_stmt
, 0, build2 (PLUS_EXPR
, type
,
1676 build_int_cst (type
, 1)));
1678 gsi
= gsi_last_bb (for_bb
);
1679 gsi_insert_after (&gsi
, for_stmt
, GSI_NEW_STMT
);
1680 SSA_NAME_DEF_STMT (initvar
) = for_stmt
;
1682 /* Emit GIMPLE_OMP_CONTINUE. */
1683 gsi
= gsi_last_bb (loop
->latch
);
1684 stmt
= gimple_build_omp_continue (cvar_next
, cvar
);
1685 gimple_set_location (stmt
, loc
);
1686 gsi_insert_after (&gsi
, stmt
, GSI_NEW_STMT
);
1687 SSA_NAME_DEF_STMT (cvar_next
) = stmt
;
1689 /* Emit GIMPLE_OMP_RETURN for GIMPLE_OMP_FOR. */
1690 gsi
= gsi_last_bb (ex_bb
);
1691 stmt
= gimple_build_omp_return (true);
1692 gimple_set_location (stmt
, loc
);
1693 gsi_insert_after (&gsi
, stmt
, GSI_NEW_STMT
);
1695 /* After the above dom info is hosed. Re-compute it. */
1696 free_dominance_info (CDI_DOMINATORS
);
1697 calculate_dominance_info (CDI_DOMINATORS
);
1702 /* Generates code to execute the iterations of LOOP in N_THREADS
1703 threads in parallel.
1705 NITER describes number of iterations of LOOP.
1706 REDUCTION_LIST describes the reductions existent in the LOOP. */
1709 gen_parallel_loop (struct loop
*loop
, htab_t reduction_list
,
1710 unsigned n_threads
, struct tree_niter_desc
*niter
)
1713 tree many_iterations_cond
, type
, nit
;
1714 tree arg_struct
, new_arg_struct
;
1716 basic_block parallel_head
;
1718 struct clsn_data clsn_data
;
1722 unsigned int m_p_thread
=2;
1726 ---------------------------------------------------------------------
1729 IV = phi (INIT, IV + STEP)
1735 ---------------------------------------------------------------------
1737 with # of iterations NITER (possibly with MAY_BE_ZERO assumption),
1738 we generate the following code:
1740 ---------------------------------------------------------------------
1743 || NITER < MIN_PER_THREAD * N_THREADS)
1747 store all local loop-invariant variables used in body of the loop to DATA.
1748 GIMPLE_OMP_PARALLEL (OMP_CLAUSE_NUM_THREADS (N_THREADS), LOOPFN, DATA);
1749 load the variables from DATA.
1750 GIMPLE_OMP_FOR (IV = INIT; COND; IV += STEP) (OMP_CLAUSE_SCHEDULE (static))
1753 GIMPLE_OMP_CONTINUE;
1754 GIMPLE_OMP_RETURN -- GIMPLE_OMP_FOR
1755 GIMPLE_OMP_RETURN -- GIMPLE_OMP_PARALLEL
1761 IV = phi (INIT, IV + STEP)
1772 /* Create two versions of the loop -- in the old one, we know that the
1773 number of iterations is large enough, and we will transform it into the
1774 loop that will be split to loop_fn, the new one will be used for the
1775 remaining iterations. */
1777 /* We should compute a better number-of-iterations value for outer loops.
1780 for (i = 0; i < n; ++i)
1781 for (j = 0; j < m; ++j)
1784 we should compute nit = n * m, not nit = n.
1785 Also may_be_zero handling would need to be adjusted. */
1787 type
= TREE_TYPE (niter
->niter
);
1788 nit
= force_gimple_operand (unshare_expr (niter
->niter
), &stmts
, true,
1791 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
1796 m_p_thread
=MIN_PER_THREAD
;
1798 many_iterations_cond
=
1799 fold_build2 (GE_EXPR
, boolean_type_node
,
1800 nit
, build_int_cst (type
, m_p_thread
* n_threads
));
1802 many_iterations_cond
1803 = fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
1804 invert_truthvalue (unshare_expr (niter
->may_be_zero
)),
1805 many_iterations_cond
);
1806 many_iterations_cond
1807 = force_gimple_operand (many_iterations_cond
, &stmts
, false, NULL_TREE
);
1809 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
1810 if (!is_gimple_condexpr (many_iterations_cond
))
1812 many_iterations_cond
1813 = force_gimple_operand (many_iterations_cond
, &stmts
,
1816 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
1819 initialize_original_copy_tables ();
1821 /* We assume that the loop usually iterates a lot. */
1822 prob
= 4 * REG_BR_PROB_BASE
/ 5;
1823 loop_version (loop
, many_iterations_cond
, NULL
,
1824 prob
, prob
, REG_BR_PROB_BASE
- prob
, true);
1825 update_ssa (TODO_update_ssa
);
1826 free_original_copy_tables ();
1828 /* Base all the induction variables in LOOP on a single control one. */
1829 canonicalize_loop_ivs (loop
, &nit
, true);
1831 /* Ensure that the exit condition is the first statement in the loop. */
1832 transform_to_exit_first_loop (loop
, reduction_list
, nit
);
1834 /* Generate initializations for reductions. */
1835 if (htab_elements (reduction_list
) > 0)
1836 htab_traverse (reduction_list
, initialize_reductions
, loop
);
1838 /* Eliminate the references to local variables from the loop. */
1839 gcc_assert (single_exit (loop
));
1840 entry
= loop_preheader_edge (loop
);
1841 exit
= single_dom_exit (loop
);
1843 eliminate_local_variables (entry
, exit
);
1844 /* In the old loop, move all variables non-local to the loop to a structure
1845 and back, and create separate decls for the variables used in loop. */
1846 separate_decls_in_region (entry
, exit
, reduction_list
, &arg_struct
,
1847 &new_arg_struct
, &clsn_data
);
1849 /* Create the parallel constructs. */
1850 loc
= UNKNOWN_LOCATION
;
1851 cond_stmt
= last_stmt (loop
->header
);
1853 loc
= gimple_location (cond_stmt
);
1854 parallel_head
= create_parallel_loop (loop
, create_loop_fn (loc
), arg_struct
,
1855 new_arg_struct
, n_threads
, loc
);
1856 if (htab_elements (reduction_list
) > 0)
1857 create_call_for_reduction (loop
, reduction_list
, &clsn_data
);
1861 /* Cancel the loop (it is simpler to do it here rather than to teach the
1862 expander to do it). */
1863 cancel_loop_tree (loop
);
1865 /* Free loop bound estimations that could contain references to
1866 removed statements. */
1867 FOR_EACH_LOOP (li
, loop
, 0)
1868 free_numbers_of_iterations_estimates_loop (loop
);
1870 /* Expand the parallel constructs. We do it directly here instead of running
1871 a separate expand_omp pass, since it is more efficient, and less likely to
1872 cause troubles with further analyses not being able to deal with the
1875 omp_expand_local (parallel_head
);
1878 /* Returns true when LOOP contains vector phi nodes. */
1881 loop_has_vector_phi_nodes (struct loop
*loop ATTRIBUTE_UNUSED
)
1884 basic_block
*bbs
= get_loop_body_in_dom_order (loop
);
1885 gimple_stmt_iterator gsi
;
1888 for (i
= 0; i
< loop
->num_nodes
; i
++)
1889 for (gsi
= gsi_start_phis (bbs
[i
]); !gsi_end_p (gsi
); gsi_next (&gsi
))
1890 if (TREE_CODE (TREE_TYPE (PHI_RESULT (gsi_stmt (gsi
)))) == VECTOR_TYPE
)
1899 /* Create a reduction_info struct, initialize it with REDUC_STMT
1900 and PHI, insert it to the REDUCTION_LIST. */
1903 build_new_reduction (htab_t reduction_list
, gimple reduc_stmt
, gimple phi
)
1906 struct reduction_info
*new_reduction
;
1908 gcc_assert (reduc_stmt
);
1910 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1913 "Detected reduction. reduction stmt is: \n");
1914 print_gimple_stmt (dump_file
, reduc_stmt
, 0, 0);
1915 fprintf (dump_file
, "\n");
1918 new_reduction
= XCNEW (struct reduction_info
);
1920 new_reduction
->reduc_stmt
= reduc_stmt
;
1921 new_reduction
->reduc_phi
= phi
;
1922 new_reduction
->reduc_version
= SSA_NAME_VERSION (gimple_phi_result (phi
));
1923 new_reduction
->reduction_code
= gimple_assign_rhs_code (reduc_stmt
);
1924 slot
= htab_find_slot (reduction_list
, new_reduction
, INSERT
);
1925 *slot
= new_reduction
;
1928 /* Callback for htab_traverse. Sets gimple_uid of reduc_phi stmts. */
1931 set_reduc_phi_uids (void **slot
, void *data ATTRIBUTE_UNUSED
)
1933 struct reduction_info
*const red
= (struct reduction_info
*) *slot
;
1934 gimple_set_uid (red
->reduc_phi
, red
->reduc_version
);
1938 /* Detect all reductions in the LOOP, insert them into REDUCTION_LIST. */
1941 gather_scalar_reductions (loop_p loop
, htab_t reduction_list
)
1943 gimple_stmt_iterator gsi
;
1944 loop_vec_info simple_loop_info
;
1947 simple_loop_info
= vect_analyze_loop_form (loop
);
1949 for (gsi
= gsi_start_phis (loop
->header
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1951 gimple phi
= gsi_stmt (gsi
);
1953 tree res
= PHI_RESULT (phi
);
1956 if (virtual_operand_p (res
))
1959 if (!simple_iv (loop
, loop
, res
, &iv
, true)
1960 && simple_loop_info
)
1962 gimple reduc_stmt
= vect_force_simple_reduction (simple_loop_info
,
1965 if (reduc_stmt
&& !double_reduc
)
1966 build_new_reduction (reduction_list
, reduc_stmt
, phi
);
1969 destroy_loop_vec_info (simple_loop_info
, true);
1971 /* As gimple_uid is used by the vectorizer in between vect_analyze_loop_form
1972 and destroy_loop_vec_info, we can set gimple_uid of reduc_phi stmts
1974 htab_traverse (reduction_list
, set_reduc_phi_uids
, NULL
);
1977 /* Try to initialize NITER for code generation part. */
1980 try_get_loop_niter (loop_p loop
, struct tree_niter_desc
*niter
)
1982 edge exit
= single_dom_exit (loop
);
1986 /* We need to know # of iterations, and there should be no uses of values
1987 defined inside loop outside of it, unless the values are invariants of
1989 if (!number_of_iterations_exit (loop
, exit
, niter
, false))
1991 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1992 fprintf (dump_file
, " FAILED: number of iterations not known\n");
1999 /* Try to initialize REDUCTION_LIST for code generation part.
2000 REDUCTION_LIST describes the reductions. */
2003 try_create_reduction_list (loop_p loop
, htab_t reduction_list
)
2005 edge exit
= single_dom_exit (loop
);
2006 gimple_stmt_iterator gsi
;
2010 gather_scalar_reductions (loop
, reduction_list
);
2013 for (gsi
= gsi_start_phis (exit
->dest
); !gsi_end_p (gsi
); gsi_next (&gsi
))
2015 gimple phi
= gsi_stmt (gsi
);
2016 struct reduction_info
*red
;
2017 imm_use_iterator imm_iter
;
2018 use_operand_p use_p
;
2020 tree val
= PHI_ARG_DEF_FROM_EDGE (phi
, exit
);
2022 if (!virtual_operand_p (val
))
2024 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2026 fprintf (dump_file
, "phi is ");
2027 print_gimple_stmt (dump_file
, phi
, 0, 0);
2028 fprintf (dump_file
, "arg of phi to exit: value ");
2029 print_generic_expr (dump_file
, val
, 0);
2030 fprintf (dump_file
, " used outside loop\n");
2032 " checking if it a part of reduction pattern: \n");
2034 if (htab_elements (reduction_list
) == 0)
2036 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2038 " FAILED: it is not a part of reduction.\n");
2042 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, val
)
2044 if (!gimple_debug_bind_p (USE_STMT (use_p
))
2045 && flow_bb_inside_loop_p (loop
, gimple_bb (USE_STMT (use_p
))))
2047 reduc_phi
= USE_STMT (use_p
);
2051 red
= reduction_phi (reduction_list
, reduc_phi
);
2054 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2056 " FAILED: it is not a part of reduction.\n");
2059 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2061 fprintf (dump_file
, "reduction phi is ");
2062 print_gimple_stmt (dump_file
, red
->reduc_phi
, 0, 0);
2063 fprintf (dump_file
, "reduction stmt is ");
2064 print_gimple_stmt (dump_file
, red
->reduc_stmt
, 0, 0);
2069 /* The iterations of the loop may communicate only through bivs whose
2070 iteration space can be distributed efficiently. */
2071 for (gsi
= gsi_start_phis (loop
->header
); !gsi_end_p (gsi
); gsi_next (&gsi
))
2073 gimple phi
= gsi_stmt (gsi
);
2074 tree def
= PHI_RESULT (phi
);
2077 if (!virtual_operand_p (def
) && !simple_iv (loop
, loop
, def
, &iv
, true))
2079 struct reduction_info
*red
;
2081 red
= reduction_phi (reduction_list
, phi
);
2084 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2086 " FAILED: scalar dependency between iterations\n");
2096 /* Detect parallel loops and generate parallel code using libgomp
2097 primitives. Returns true if some loop was parallelized, false
2101 parallelize_loops (void)
2103 unsigned n_threads
= flag_tree_parallelize_loops
;
2104 bool changed
= false;
2106 struct tree_niter_desc niter_desc
;
2108 htab_t reduction_list
;
2109 struct obstack parloop_obstack
;
2110 HOST_WIDE_INT estimated
;
2113 /* Do not parallelize loops in the functions created by parallelization. */
2114 if (parallelized_function_p (cfun
->decl
))
2116 if (cfun
->has_nonlocal_label
)
2119 gcc_obstack_init (&parloop_obstack
);
2120 reduction_list
= htab_create (10, reduction_info_hash
,
2121 reduction_info_eq
, free
);
2122 init_stmt_vec_info_vec ();
2124 FOR_EACH_LOOP (li
, loop
, 0)
2126 htab_empty (reduction_list
);
2127 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2129 fprintf (dump_file
, "Trying loop %d as candidate\n",loop
->num
);
2131 fprintf (dump_file
, "loop %d is not innermost\n",loop
->num
);
2133 fprintf (dump_file
, "loop %d is innermost\n",loop
->num
);
2136 /* If we use autopar in graphite pass, we use its marked dependency
2137 checking results. */
2138 if (flag_loop_parallelize_all
&& !loop
->can_be_parallel
)
2140 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2141 fprintf (dump_file
, "loop is not parallel according to graphite\n");
2145 if (!single_dom_exit (loop
))
2148 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2149 fprintf (dump_file
, "loop is !single_dom_exit\n");
2154 if (/* And of course, the loop must be parallelizable. */
2155 !can_duplicate_loop_p (loop
)
2156 || loop_has_blocks_with_irreducible_flag (loop
)
2157 || (loop_preheader_edge (loop
)->src
->flags
& BB_IRREDUCIBLE_LOOP
)
2158 /* FIXME: the check for vector phi nodes could be removed. */
2159 || loop_has_vector_phi_nodes (loop
))
2162 estimated
= estimated_stmt_executions_int (loop
);
2163 if (estimated
== -1)
2164 estimated
= max_stmt_executions_int (loop
);
2165 /* FIXME: Bypass this check as graphite doesn't update the
2166 count and frequency correctly now. */
2167 if (!flag_loop_parallelize_all
2168 && ((estimated
!= -1
2169 && estimated
<= (HOST_WIDE_INT
) n_threads
* MIN_PER_THREAD
)
2170 /* Do not bother with loops in cold areas. */
2171 || optimize_loop_nest_for_size_p (loop
)))
2174 if (!try_get_loop_niter (loop
, &niter_desc
))
2177 if (!try_create_reduction_list (loop
, reduction_list
))
2180 if (!flag_loop_parallelize_all
2181 && !loop_parallel_p (loop
, &parloop_obstack
))
2185 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2188 fprintf (dump_file
, "parallelizing outer loop %d\n",loop
->header
->index
);
2190 fprintf (dump_file
, "parallelizing inner loop %d\n",loop
->header
->index
);
2191 loop_loc
= find_loop_location (loop
);
2192 if (loop_loc
!= UNKNOWN_LOC
)
2193 fprintf (dump_file
, "\nloop at %s:%d: ",
2194 LOC_FILE (loop_loc
), LOC_LINE (loop_loc
));
2196 gen_parallel_loop (loop
, reduction_list
,
2197 n_threads
, &niter_desc
);
2198 #ifdef ENABLE_CHECKING
2199 verify_flow_info ();
2200 verify_loop_structure ();
2201 verify_loop_closed_ssa (true);
2205 free_stmt_vec_info_vec ();
2206 htab_delete (reduction_list
);
2207 obstack_free (&parloop_obstack
, NULL
);
2209 /* Parallelization will cause new function calls to be inserted through
2210 which local variables will escape. Reset the points-to solution
2213 pt_solution_reset (&cfun
->gimple_df
->escaped
);
2218 #include "gt-tree-parloops.h"