1 /* High-level loop manipulation functions.
2 Copyright (C) 2004-2014 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by the
8 Free Software Foundation; either version 3, or (at your option) any
11 GCC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
22 #include "coretypes.h"
31 #include "hard-reg-set.h"
34 #include "dominance.h"
37 #include "basic-block.h"
38 #include "tree-ssa-alias.h"
39 #include "internal-fn.h"
40 #include "gimple-expr.h"
44 #include "gimple-iterator.h"
45 #include "gimplify-me.h"
46 #include "gimple-ssa.h"
48 #include "tree-phinodes.h"
49 #include "ssa-iterators.h"
50 #include "stringpool.h"
51 #include "tree-ssanames.h"
52 #include "tree-ssa-loop-ivopts.h"
53 #include "tree-ssa-loop-manip.h"
54 #include "tree-ssa-loop-niter.h"
55 #include "tree-ssa-loop.h"
56 #include "tree-into-ssa.h"
59 #include "gimple-pretty-print.h"
61 #include "tree-pass.h" /* ??? for TODO_update_ssa but this isn't a pass. */
62 #include "tree-scalar-evolution.h"
64 #include "tree-inline.h"
65 #include "langhooks.h"
67 /* All bitmaps for rewriting into loop-closed SSA go on this obstack,
68 so that we can free them all at once. */
69 static bitmap_obstack loop_renamer_obstack
;
71 /* Creates an induction variable with value BASE + STEP * iteration in LOOP.
72 It is expected that neither BASE nor STEP are shared with other expressions
73 (unless the sharing rules allow this). Use VAR as a base var_decl for it
74 (if NULL, a new temporary will be created). The increment will occur at
75 INCR_POS (after it if AFTER is true, before it otherwise). INCR_POS and
76 AFTER can be computed using standard_iv_increment_position. The ssa versions
77 of the variable before and after increment will be stored in VAR_BEFORE and
78 VAR_AFTER (unless they are NULL). */
81 create_iv (tree base
, tree step
, tree var
, struct loop
*loop
,
82 gimple_stmt_iterator
*incr_pos
, bool after
,
83 tree
*var_before
, tree
*var_after
)
89 enum tree_code incr_op
= PLUS_EXPR
;
90 edge pe
= loop_preheader_edge (loop
);
94 vb
= make_ssa_name (var
, NULL
);
95 va
= make_ssa_name (var
, NULL
);
99 vb
= make_temp_ssa_name (TREE_TYPE (base
), NULL
, "ivtmp");
100 va
= make_temp_ssa_name (TREE_TYPE (base
), NULL
, "ivtmp");
107 /* For easier readability of the created code, produce MINUS_EXPRs
109 if (TREE_CODE (step
) == INTEGER_CST
)
111 if (TYPE_UNSIGNED (TREE_TYPE (step
)))
113 step1
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
114 if (tree_int_cst_lt (step1
, step
))
116 incr_op
= MINUS_EXPR
;
124 if (!tree_expr_nonnegative_warnv_p (step
, &ovf
)
125 && may_negate_without_overflow_p (step
))
127 incr_op
= MINUS_EXPR
;
128 step
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
132 if (POINTER_TYPE_P (TREE_TYPE (base
)))
134 if (TREE_CODE (base
) == ADDR_EXPR
)
135 mark_addressable (TREE_OPERAND (base
, 0));
136 step
= convert_to_ptrofftype (step
);
137 if (incr_op
== MINUS_EXPR
)
138 step
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
139 incr_op
= POINTER_PLUS_EXPR
;
141 /* Gimplify the step if necessary. We put the computations in front of the
142 loop (i.e. the step should be loop invariant). */
143 step
= force_gimple_operand (step
, &stmts
, true, NULL_TREE
);
145 gsi_insert_seq_on_edge_immediate (pe
, stmts
);
147 stmt
= gimple_build_assign_with_ops (incr_op
, va
, vb
, step
);
149 gsi_insert_after (incr_pos
, stmt
, GSI_NEW_STMT
);
151 gsi_insert_before (incr_pos
, stmt
, GSI_NEW_STMT
);
153 initial
= force_gimple_operand (base
, &stmts
, true, var
);
155 gsi_insert_seq_on_edge_immediate (pe
, stmts
);
157 stmt
= create_phi_node (vb
, loop
->header
);
158 add_phi_arg (stmt
, initial
, loop_preheader_edge (loop
), UNKNOWN_LOCATION
);
159 add_phi_arg (stmt
, va
, loop_latch_edge (loop
), UNKNOWN_LOCATION
);
162 /* Return the innermost superloop LOOP of USE_LOOP that is a superloop of
163 both DEF_LOOP and USE_LOOP. */
165 static inline struct loop
*
166 find_sibling_superloop (struct loop
*use_loop
, struct loop
*def_loop
)
168 unsigned ud
= loop_depth (use_loop
);
169 unsigned dd
= loop_depth (def_loop
);
170 gcc_assert (ud
> 0 && dd
> 0);
172 use_loop
= superloop_at_depth (use_loop
, dd
);
174 def_loop
= superloop_at_depth (def_loop
, ud
);
175 while (loop_outer (use_loop
) != loop_outer (def_loop
))
177 use_loop
= loop_outer (use_loop
);
178 def_loop
= loop_outer (def_loop
);
179 gcc_assert (use_loop
&& def_loop
);
184 /* DEF_BB is a basic block containing a DEF that needs rewriting into
185 loop-closed SSA form. USE_BLOCKS is the set of basic blocks containing
186 uses of DEF that "escape" from the loop containing DEF_BB (i.e. blocks in
187 USE_BLOCKS are dominated by DEF_BB but not in the loop father of DEF_B).
188 ALL_EXITS[I] is the set of all basic blocks that exit loop I.
190 Compute the subset of LOOP_EXITS that exit the loop containing DEF_BB
191 or one of its loop fathers, in which DEF is live. This set is returned
192 in the bitmap LIVE_EXITS.
194 Instead of computing the complete livein set of the def, we use the loop
195 nesting tree as a form of poor man's structure analysis. This greatly
196 speeds up the analysis, which is important because this function may be
197 called on all SSA names that need rewriting, one at a time. */
200 compute_live_loop_exits (bitmap live_exits
, bitmap use_blocks
,
201 bitmap
*loop_exits
, basic_block def_bb
)
205 struct loop
*def_loop
= def_bb
->loop_father
;
206 unsigned def_loop_depth
= loop_depth (def_loop
);
207 bitmap def_loop_exits
;
209 /* Normally the work list size is bounded by the number of basic
210 blocks in the largest loop. We don't know this number, but we
211 can be fairly sure that it will be relatively small. */
212 auto_vec
<basic_block
> worklist (MAX (8, n_basic_blocks_for_fn (cfun
) / 128));
214 EXECUTE_IF_SET_IN_BITMAP (use_blocks
, 0, i
, bi
)
216 basic_block use_bb
= BASIC_BLOCK_FOR_FN (cfun
, i
);
217 struct loop
*use_loop
= use_bb
->loop_father
;
218 gcc_checking_assert (def_loop
!= use_loop
219 && ! flow_loop_nested_p (def_loop
, use_loop
));
220 if (! flow_loop_nested_p (use_loop
, def_loop
))
221 use_bb
= find_sibling_superloop (use_loop
, def_loop
)->header
;
222 if (bitmap_set_bit (live_exits
, use_bb
->index
))
223 worklist
.safe_push (use_bb
);
226 /* Iterate until the worklist is empty. */
227 while (! worklist
.is_empty ())
232 /* Pull a block off the worklist. */
233 basic_block bb
= worklist
.pop ();
235 /* Make sure we have at least enough room in the work list
236 for all predecessors of this block. */
237 worklist
.reserve (EDGE_COUNT (bb
->preds
));
239 /* For each predecessor block. */
240 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
242 basic_block pred
= e
->src
;
243 struct loop
*pred_loop
= pred
->loop_father
;
244 unsigned pred_loop_depth
= loop_depth (pred_loop
);
247 /* We should have met DEF_BB along the way. */
248 gcc_assert (pred
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
));
250 if (pred_loop_depth
>= def_loop_depth
)
252 if (pred_loop_depth
> def_loop_depth
)
253 pred_loop
= superloop_at_depth (pred_loop
, def_loop_depth
);
254 /* If we've reached DEF_LOOP, our train ends here. */
255 if (pred_loop
== def_loop
)
258 else if (! flow_loop_nested_p (pred_loop
, def_loop
))
259 pred
= find_sibling_superloop (pred_loop
, def_loop
)->header
;
261 /* Add PRED to the LIVEIN set. PRED_VISITED is true if
262 we had already added PRED to LIVEIN before. */
263 pred_visited
= !bitmap_set_bit (live_exits
, pred
->index
);
265 /* If we have visited PRED before, don't add it to the worklist.
266 If BB dominates PRED, then we're probably looking at a loop.
267 We're only interested in looking up in the dominance tree
268 because DEF_BB dominates all the uses. */
269 if (pred_visited
|| dominated_by_p (CDI_DOMINATORS
, pred
, bb
))
272 worklist
.quick_push (pred
);
276 def_loop_exits
= BITMAP_ALLOC (&loop_renamer_obstack
);
277 for (struct loop
*loop
= def_loop
;
278 loop
!= current_loops
->tree_root
;
279 loop
= loop_outer (loop
))
280 bitmap_ior_into (def_loop_exits
, loop_exits
[loop
->num
]);
281 bitmap_and_into (live_exits
, def_loop_exits
);
282 BITMAP_FREE (def_loop_exits
);
285 /* Add a loop-closing PHI for VAR in basic block EXIT. */
288 add_exit_phi (basic_block exit
, tree var
)
294 #ifdef ENABLE_CHECKING
295 /* Check that at least one of the edges entering the EXIT block exits
296 the loop, or a superloop of that loop, that VAR is defined in. */
297 gimple def_stmt
= SSA_NAME_DEF_STMT (var
);
298 basic_block def_bb
= gimple_bb (def_stmt
);
299 FOR_EACH_EDGE (e
, ei
, exit
->preds
)
301 struct loop
*aloop
= find_common_loop (def_bb
->loop_father
,
302 e
->src
->loop_father
);
303 if (!flow_bb_inside_loop_p (aloop
, e
->dest
))
307 gcc_checking_assert (e
);
310 phi
= create_phi_node (NULL_TREE
, exit
);
311 create_new_def_for (var
, phi
, gimple_phi_result_ptr (phi
));
312 FOR_EACH_EDGE (e
, ei
, exit
->preds
)
313 add_phi_arg (phi
, var
, e
, UNKNOWN_LOCATION
);
315 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
317 fprintf (dump_file
, ";; Created LCSSA PHI: ");
318 print_gimple_stmt (dump_file
, phi
, 0, dump_flags
);
322 /* Add exit phis for VAR that is used in LIVEIN.
323 Exits of the loops are stored in LOOP_EXITS. */
326 add_exit_phis_var (tree var
, bitmap use_blocks
, bitmap
*loop_exits
)
330 basic_block def_bb
= gimple_bb (SSA_NAME_DEF_STMT (var
));
331 bitmap live_exits
= BITMAP_ALLOC (&loop_renamer_obstack
);
333 gcc_checking_assert (! bitmap_bit_p (use_blocks
, def_bb
->index
));
335 compute_live_loop_exits (live_exits
, use_blocks
, loop_exits
, def_bb
);
337 EXECUTE_IF_SET_IN_BITMAP (live_exits
, 0, index
, bi
)
339 add_exit_phi (BASIC_BLOCK_FOR_FN (cfun
, index
), var
);
342 BITMAP_FREE (live_exits
);
345 /* Add exit phis for the names marked in NAMES_TO_RENAME.
346 Exits of the loops are stored in EXITS. Sets of blocks where the ssa
347 names are used are stored in USE_BLOCKS. */
350 add_exit_phis (bitmap names_to_rename
, bitmap
*use_blocks
, bitmap
*loop_exits
)
355 EXECUTE_IF_SET_IN_BITMAP (names_to_rename
, 0, i
, bi
)
357 add_exit_phis_var (ssa_name (i
), use_blocks
[i
], loop_exits
);
361 /* Fill the array of bitmaps LOOP_EXITS with all loop exit edge targets. */
364 get_loops_exits (bitmap
*loop_exits
)
370 FOR_EACH_LOOP (loop
, 0)
372 vec
<edge
> exit_edges
= get_loop_exit_edges (loop
);
373 loop_exits
[loop
->num
] = BITMAP_ALLOC (&loop_renamer_obstack
);
374 FOR_EACH_VEC_ELT (exit_edges
, j
, e
)
375 bitmap_set_bit (loop_exits
[loop
->num
], e
->dest
->index
);
376 exit_edges
.release ();
380 /* For USE in BB, if it is used outside of the loop it is defined in,
381 mark it for rewrite. Record basic block BB where it is used
382 to USE_BLOCKS. Record the ssa name index to NEED_PHIS bitmap. */
385 find_uses_to_rename_use (basic_block bb
, tree use
, bitmap
*use_blocks
,
390 struct loop
*def_loop
;
392 if (TREE_CODE (use
) != SSA_NAME
)
395 ver
= SSA_NAME_VERSION (use
);
396 def_bb
= gimple_bb (SSA_NAME_DEF_STMT (use
));
399 def_loop
= def_bb
->loop_father
;
401 /* If the definition is not inside a loop, it is not interesting. */
402 if (!loop_outer (def_loop
))
405 /* If the use is not outside of the loop it is defined in, it is not
407 if (flow_bb_inside_loop_p (def_loop
, bb
))
410 /* If we're seeing VER for the first time, we still have to allocate
411 a bitmap for its uses. */
412 if (bitmap_set_bit (need_phis
, ver
))
413 use_blocks
[ver
] = BITMAP_ALLOC (&loop_renamer_obstack
);
414 bitmap_set_bit (use_blocks
[ver
], bb
->index
);
417 /* For uses in STMT, mark names that are used outside of the loop they are
418 defined to rewrite. Record the set of blocks in that the ssa
419 names are defined to USE_BLOCKS and the ssa names themselves to
423 find_uses_to_rename_stmt (gimple stmt
, bitmap
*use_blocks
, bitmap need_phis
)
427 basic_block bb
= gimple_bb (stmt
);
429 if (is_gimple_debug (stmt
))
432 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, SSA_OP_USE
)
433 find_uses_to_rename_use (bb
, var
, use_blocks
, need_phis
);
436 /* Marks names that are used in BB and outside of the loop they are
437 defined in for rewrite. Records the set of blocks in that the ssa
438 names are defined to USE_BLOCKS. Record the SSA names that will
439 need exit PHIs in NEED_PHIS. */
442 find_uses_to_rename_bb (basic_block bb
, bitmap
*use_blocks
, bitmap need_phis
)
444 gimple_stmt_iterator bsi
;
448 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
449 for (bsi
= gsi_start_phis (e
->dest
); !gsi_end_p (bsi
); gsi_next (&bsi
))
451 gimple phi
= gsi_stmt (bsi
);
452 if (! virtual_operand_p (gimple_phi_result (phi
)))
453 find_uses_to_rename_use (bb
, PHI_ARG_DEF_FROM_EDGE (phi
, e
),
454 use_blocks
, need_phis
);
457 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
458 find_uses_to_rename_stmt (gsi_stmt (bsi
), use_blocks
, need_phis
);
461 /* Marks names that are used outside of the loop they are defined in
462 for rewrite. Records the set of blocks in that the ssa
463 names are defined to USE_BLOCKS. If CHANGED_BBS is not NULL,
464 scan only blocks in this set. */
467 find_uses_to_rename (bitmap changed_bbs
, bitmap
*use_blocks
, bitmap need_phis
)
474 EXECUTE_IF_SET_IN_BITMAP (changed_bbs
, 0, index
, bi
)
475 find_uses_to_rename_bb (BASIC_BLOCK_FOR_FN (cfun
, index
), use_blocks
, need_phis
);
477 FOR_EACH_BB_FN (bb
, cfun
)
478 find_uses_to_rename_bb (bb
, use_blocks
, need_phis
);
481 /* Rewrites the program into a loop closed ssa form -- i.e. inserts extra
482 phi nodes to ensure that no variable is used outside the loop it is
485 This strengthening of the basic ssa form has several advantages:
487 1) Updating it during unrolling/peeling/versioning is trivial, since
488 we do not need to care about the uses outside of the loop.
489 The same applies to virtual operands which are also rewritten into
490 loop closed SSA form. Note that virtual operands are always live
492 2) The behavior of all uses of an induction variable is the same.
493 Without this, you need to distinguish the case when the variable
494 is used outside of the loop it is defined in, for example
496 for (i = 0; i < 100; i++)
498 for (j = 0; j < 100; j++)
506 Looking from the outer loop with the normal SSA form, the first use of k
507 is not well-behaved, while the second one is an induction variable with
510 If CHANGED_BBS is not NULL, we look for uses outside loops only in
511 the basic blocks in this set.
513 UPDATE_FLAG is used in the call to update_ssa. See
514 TODO_update_ssa* for documentation. */
517 rewrite_into_loop_closed_ssa (bitmap changed_bbs
, unsigned update_flag
)
520 bitmap names_to_rename
;
522 loops_state_set (LOOP_CLOSED_SSA
);
523 if (number_of_loops (cfun
) <= 1)
526 /* If the pass has caused the SSA form to be out-of-date, update it
528 update_ssa (update_flag
);
530 bitmap_obstack_initialize (&loop_renamer_obstack
);
532 names_to_rename
= BITMAP_ALLOC (&loop_renamer_obstack
);
534 /* Uses of names to rename. We don't have to initialize this array,
535 because we know that we will only have entries for the SSA names
536 in NAMES_TO_RENAME. */
537 use_blocks
= XNEWVEC (bitmap
, num_ssa_names
);
539 /* Find the uses outside loops. */
540 find_uses_to_rename (changed_bbs
, use_blocks
, names_to_rename
);
542 if (!bitmap_empty_p (names_to_rename
))
544 /* An array of bitmaps where LOOP_EXITS[I] is the set of basic blocks
545 that are the destination of an edge exiting loop number I. */
546 bitmap
*loop_exits
= XNEWVEC (bitmap
, number_of_loops (cfun
));
547 get_loops_exits (loop_exits
);
549 /* Add the PHI nodes on exits of the loops for the names we need to
551 add_exit_phis (names_to_rename
, use_blocks
, loop_exits
);
555 /* Fix up all the names found to be used outside their original
557 update_ssa (TODO_update_ssa
);
560 bitmap_obstack_release (&loop_renamer_obstack
);
564 /* Check invariants of the loop closed ssa form for the USE in BB. */
567 check_loop_closed_ssa_use (basic_block bb
, tree use
)
572 if (TREE_CODE (use
) != SSA_NAME
|| virtual_operand_p (use
))
575 def
= SSA_NAME_DEF_STMT (use
);
576 def_bb
= gimple_bb (def
);
578 || flow_bb_inside_loop_p (def_bb
->loop_father
, bb
));
581 /* Checks invariants of loop closed ssa form in statement STMT in BB. */
584 check_loop_closed_ssa_stmt (basic_block bb
, gimple stmt
)
589 if (is_gimple_debug (stmt
))
592 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, SSA_OP_USE
)
593 check_loop_closed_ssa_use (bb
, var
);
596 /* Checks that invariants of the loop closed ssa form are preserved.
597 Call verify_ssa when VERIFY_SSA_P is true. */
600 verify_loop_closed_ssa (bool verify_ssa_p
)
603 gimple_stmt_iterator bsi
;
608 if (number_of_loops (cfun
) <= 1)
612 verify_ssa (false, true);
614 timevar_push (TV_VERIFY_LOOP_CLOSED
);
616 FOR_EACH_BB_FN (bb
, cfun
)
618 for (bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
620 phi
= gsi_stmt (bsi
);
621 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
622 check_loop_closed_ssa_use (e
->src
,
623 PHI_ARG_DEF_FROM_EDGE (phi
, e
));
626 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
627 check_loop_closed_ssa_stmt (bb
, gsi_stmt (bsi
));
630 timevar_pop (TV_VERIFY_LOOP_CLOSED
);
633 /* Split loop exit edge EXIT. The things are a bit complicated by a need to
634 preserve the loop closed ssa form. The newly created block is returned. */
637 split_loop_exit_edge (edge exit
)
639 basic_block dest
= exit
->dest
;
640 basic_block bb
= split_edge (exit
);
644 gimple_stmt_iterator psi
;
645 source_location locus
;
647 for (psi
= gsi_start_phis (dest
); !gsi_end_p (psi
); gsi_next (&psi
))
649 phi
= gsi_stmt (psi
);
650 op_p
= PHI_ARG_DEF_PTR_FROM_EDGE (phi
, single_succ_edge (bb
));
651 locus
= gimple_phi_arg_location_from_edge (phi
, single_succ_edge (bb
));
653 name
= USE_FROM_PTR (op_p
);
655 /* If the argument of the PHI node is a constant, we do not need
656 to keep it inside loop. */
657 if (TREE_CODE (name
) != SSA_NAME
)
660 /* Otherwise create an auxiliary phi node that will copy the value
661 of the SSA name out of the loop. */
662 new_name
= duplicate_ssa_name (name
, NULL
);
663 new_phi
= create_phi_node (new_name
, bb
);
664 add_phi_arg (new_phi
, name
, exit
, locus
);
665 SET_USE (op_p
, new_name
);
671 /* Returns the basic block in that statements should be emitted for induction
672 variables incremented at the end of the LOOP. */
675 ip_end_pos (struct loop
*loop
)
680 /* Returns the basic block in that statements should be emitted for induction
681 variables incremented just before exit condition of a LOOP. */
684 ip_normal_pos (struct loop
*loop
)
690 if (!single_pred_p (loop
->latch
))
693 bb
= single_pred (loop
->latch
);
694 last
= last_stmt (bb
);
696 || gimple_code (last
) != GIMPLE_COND
)
699 exit
= EDGE_SUCC (bb
, 0);
700 if (exit
->dest
== loop
->latch
)
701 exit
= EDGE_SUCC (bb
, 1);
703 if (flow_bb_inside_loop_p (loop
, exit
->dest
))
709 /* Stores the standard position for induction variable increment in LOOP
710 (just before the exit condition if it is available and latch block is empty,
711 end of the latch block otherwise) to BSI. INSERT_AFTER is set to true if
712 the increment should be inserted after *BSI. */
715 standard_iv_increment_position (struct loop
*loop
, gimple_stmt_iterator
*bsi
,
718 basic_block bb
= ip_normal_pos (loop
), latch
= ip_end_pos (loop
);
719 gimple last
= last_stmt (latch
);
722 || (last
&& gimple_code (last
) != GIMPLE_LABEL
))
724 *bsi
= gsi_last_bb (latch
);
725 *insert_after
= true;
729 *bsi
= gsi_last_bb (bb
);
730 *insert_after
= false;
734 /* Copies phi node arguments for duplicated blocks. The index of the first
735 duplicated block is FIRST_NEW_BLOCK. */
738 copy_phi_node_args (unsigned first_new_block
)
742 for (i
= first_new_block
; i
< (unsigned) last_basic_block_for_fn (cfun
); i
++)
743 BASIC_BLOCK_FOR_FN (cfun
, i
)->flags
|= BB_DUPLICATED
;
745 for (i
= first_new_block
; i
< (unsigned) last_basic_block_for_fn (cfun
); i
++)
746 add_phi_args_after_copy_bb (BASIC_BLOCK_FOR_FN (cfun
, i
));
748 for (i
= first_new_block
; i
< (unsigned) last_basic_block_for_fn (cfun
); i
++)
749 BASIC_BLOCK_FOR_FN (cfun
, i
)->flags
&= ~BB_DUPLICATED
;
753 /* The same as cfgloopmanip.c:duplicate_loop_to_header_edge, but also
754 updates the PHI nodes at start of the copied region. In order to
755 achieve this, only loops whose exits all lead to the same location
758 Notice that we do not completely update the SSA web after
759 duplication. The caller is responsible for calling update_ssa
760 after the loop has been duplicated. */
763 gimple_duplicate_loop_to_header_edge (struct loop
*loop
, edge e
,
764 unsigned int ndupl
, sbitmap wont_exit
,
765 edge orig
, vec
<edge
> *to_remove
,
768 unsigned first_new_block
;
770 if (!loops_state_satisfies_p (LOOPS_HAVE_SIMPLE_LATCHES
))
772 if (!loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS
))
775 first_new_block
= last_basic_block_for_fn (cfun
);
776 if (!duplicate_loop_to_header_edge (loop
, e
, ndupl
, wont_exit
,
777 orig
, to_remove
, flags
))
780 /* Readd the removed phi args for e. */
781 flush_pending_stmts (e
);
783 /* Copy the phi node arguments. */
784 copy_phi_node_args (first_new_block
);
791 /* Returns true if we can unroll LOOP FACTOR times. Number
792 of iterations of the loop is returned in NITER. */
795 can_unroll_loop_p (struct loop
*loop
, unsigned factor
,
796 struct tree_niter_desc
*niter
)
800 /* Check whether unrolling is possible. We only want to unroll loops
801 for that we are able to determine number of iterations. We also
802 want to split the extra iterations of the loop from its end,
803 therefore we require that the loop has precisely one
806 exit
= single_dom_exit (loop
);
810 if (!number_of_iterations_exit (loop
, exit
, niter
, false)
811 || niter
->cmp
== ERROR_MARK
812 /* Scalar evolutions analysis might have copy propagated
813 the abnormal ssa names into these expressions, hence
814 emitting the computations based on them during loop
815 unrolling might create overlapping life ranges for
816 them, and failures in out-of-ssa. */
817 || contains_abnormal_ssa_name_p (niter
->may_be_zero
)
818 || contains_abnormal_ssa_name_p (niter
->control
.base
)
819 || contains_abnormal_ssa_name_p (niter
->control
.step
)
820 || contains_abnormal_ssa_name_p (niter
->bound
))
823 /* And of course, we must be able to duplicate the loop. */
824 if (!can_duplicate_loop_p (loop
))
827 /* The final loop should be small enough. */
828 if (tree_num_loop_insns (loop
, &eni_size_weights
) * factor
829 > (unsigned) PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS
))
835 /* Determines the conditions that control execution of LOOP unrolled FACTOR
836 times. DESC is number of iterations of LOOP. ENTER_COND is set to
837 condition that must be true if the main loop can be entered.
838 EXIT_BASE, EXIT_STEP, EXIT_CMP and EXIT_BOUND are set to values describing
839 how the exit from the unrolled loop should be controlled. */
842 determine_exit_conditions (struct loop
*loop
, struct tree_niter_desc
*desc
,
843 unsigned factor
, tree
*enter_cond
,
844 tree
*exit_base
, tree
*exit_step
,
845 enum tree_code
*exit_cmp
, tree
*exit_bound
)
848 tree base
= desc
->control
.base
;
849 tree step
= desc
->control
.step
;
850 tree bound
= desc
->bound
;
851 tree type
= TREE_TYPE (step
);
853 tree min
= lower_bound_in_type (type
, type
);
854 tree max
= upper_bound_in_type (type
, type
);
855 enum tree_code cmp
= desc
->cmp
;
856 tree cond
= boolean_true_node
, assum
;
858 /* For pointers, do the arithmetics in the type of step. */
859 base
= fold_convert (type
, base
);
860 bound
= fold_convert (type
, bound
);
862 *enter_cond
= boolean_false_node
;
863 *exit_base
= NULL_TREE
;
864 *exit_step
= NULL_TREE
;
865 *exit_cmp
= ERROR_MARK
;
866 *exit_bound
= NULL_TREE
;
867 gcc_assert (cmp
!= ERROR_MARK
);
869 /* We only need to be correct when we answer question
870 "Do at least FACTOR more iterations remain?" in the unrolled loop.
871 Thus, transforming BASE + STEP * i <> BOUND to
872 BASE + STEP * i < BOUND is ok. */
875 if (tree_int_cst_sign_bit (step
))
880 else if (cmp
== LT_EXPR
)
882 gcc_assert (!tree_int_cst_sign_bit (step
));
884 else if (cmp
== GT_EXPR
)
886 gcc_assert (tree_int_cst_sign_bit (step
));
891 /* The main body of the loop may be entered iff:
893 1) desc->may_be_zero is false.
894 2) it is possible to check that there are at least FACTOR iterations
895 of the loop, i.e., BOUND - step * FACTOR does not overflow.
896 3) # of iterations is at least FACTOR */
898 if (!integer_zerop (desc
->may_be_zero
))
899 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
900 invert_truthvalue (desc
->may_be_zero
),
903 bigstep
= fold_build2 (MULT_EXPR
, type
, step
,
904 build_int_cst_type (type
, factor
));
905 delta
= fold_build2 (MINUS_EXPR
, type
, bigstep
, step
);
907 assum
= fold_build2 (GE_EXPR
, boolean_type_node
,
909 fold_build2 (PLUS_EXPR
, type
, min
, delta
));
911 assum
= fold_build2 (LE_EXPR
, boolean_type_node
,
913 fold_build2 (PLUS_EXPR
, type
, max
, delta
));
914 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
, assum
, cond
);
916 bound
= fold_build2 (MINUS_EXPR
, type
, bound
, delta
);
917 assum
= fold_build2 (cmp
, boolean_type_node
, base
, bound
);
918 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
, assum
, cond
);
920 cond
= force_gimple_operand (unshare_expr (cond
), &stmts
, false, NULL_TREE
);
922 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
923 /* cond now may be a gimple comparison, which would be OK, but also any
924 other gimple rhs (say a && b). In this case we need to force it to
926 if (!is_gimple_condexpr (cond
))
928 cond
= force_gimple_operand (cond
, &stmts
, true, NULL_TREE
);
930 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
934 base
= force_gimple_operand (unshare_expr (base
), &stmts
, true, NULL_TREE
);
936 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
937 bound
= force_gimple_operand (unshare_expr (bound
), &stmts
, true, NULL_TREE
);
939 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
942 *exit_step
= bigstep
;
947 /* Scales the frequencies of all basic blocks in LOOP that are strictly
948 dominated by BB by NUM/DEN. */
951 scale_dominated_blocks_in_loop (struct loop
*loop
, basic_block bb
,
959 for (son
= first_dom_son (CDI_DOMINATORS
, bb
);
961 son
= next_dom_son (CDI_DOMINATORS
, son
))
963 if (!flow_bb_inside_loop_p (loop
, son
))
965 scale_bbs_frequencies_int (&son
, 1, num
, den
);
966 scale_dominated_blocks_in_loop (loop
, son
, num
, den
);
970 /* Unroll LOOP FACTOR times. DESC describes number of iterations of LOOP.
971 EXIT is the exit of the loop to that DESC corresponds.
973 If N is number of iterations of the loop and MAY_BE_ZERO is the condition
974 under that loop exits in the first iteration even if N != 0,
978 x = phi (init, next);
986 becomes (with possibly the exit conditions formulated a bit differently,
987 avoiding the need to create a new iv):
989 if (MAY_BE_ZERO || N < FACTOR)
994 x = phi (init, next);
1005 } while (N >= FACTOR);
1008 init' = phi (init, x);
1012 x = phi (init', next);
1020 Before the loop is unrolled, TRANSFORM is called for it (only for the
1021 unrolled loop, but not for its versioned copy). DATA is passed to
1024 /* Probability in % that the unrolled loop is entered. Just a guess. */
1025 #define PROB_UNROLLED_LOOP_ENTERED 90
1028 tree_transform_and_unroll_loop (struct loop
*loop
, unsigned factor
,
1029 edge exit
, struct tree_niter_desc
*desc
,
1030 transform_callback transform
,
1034 tree ctr_before
, ctr_after
;
1035 tree enter_main_cond
, exit_base
, exit_step
, exit_bound
;
1036 enum tree_code exit_cmp
;
1037 gimple phi_old_loop
, phi_new_loop
, phi_rest
;
1038 gimple_stmt_iterator psi_old_loop
, psi_new_loop
;
1039 tree init
, next
, new_init
;
1040 struct loop
*new_loop
;
1041 basic_block rest
, exit_bb
;
1042 edge old_entry
, new_entry
, old_latch
, precond_edge
, new_exit
;
1043 edge new_nonexit
, e
;
1044 gimple_stmt_iterator bsi
;
1047 unsigned est_niter
, prob_entry
, scale_unrolled
, scale_rest
, freq_e
, freq_h
;
1048 unsigned new_est_niter
, i
, prob
;
1049 unsigned irr
= loop_preheader_edge (loop
)->flags
& EDGE_IRREDUCIBLE_LOOP
;
1051 auto_vec
<edge
> to_remove
;
1053 est_niter
= expected_loop_iterations (loop
);
1054 determine_exit_conditions (loop
, desc
, factor
,
1055 &enter_main_cond
, &exit_base
, &exit_step
,
1056 &exit_cmp
, &exit_bound
);
1058 /* Let us assume that the unrolled loop is quite likely to be entered. */
1059 if (integer_nonzerop (enter_main_cond
))
1060 prob_entry
= REG_BR_PROB_BASE
;
1062 prob_entry
= PROB_UNROLLED_LOOP_ENTERED
* REG_BR_PROB_BASE
/ 100;
1064 /* The values for scales should keep profile consistent, and somewhat close
1067 TODO: The current value of SCALE_REST makes it appear that the loop that
1068 is created by splitting the remaining iterations of the unrolled loop is
1069 executed the same number of times as the original loop, and with the same
1070 frequencies, which is obviously wrong. This does not appear to cause
1071 problems, so we do not bother with fixing it for now. To make the profile
1072 correct, we would need to change the probability of the exit edge of the
1073 loop, and recompute the distribution of frequencies in its body because
1074 of this change (scale the frequencies of blocks before and after the exit
1075 by appropriate factors). */
1076 scale_unrolled
= prob_entry
;
1077 scale_rest
= REG_BR_PROB_BASE
;
1079 new_loop
= loop_version (loop
, enter_main_cond
, NULL
,
1080 prob_entry
, scale_unrolled
, scale_rest
, true);
1081 gcc_assert (new_loop
!= NULL
);
1082 update_ssa (TODO_update_ssa
);
1084 /* Determine the probability of the exit edge of the unrolled loop. */
1085 new_est_niter
= est_niter
/ factor
;
1087 /* Without profile feedback, loops for that we do not know a better estimate
1088 are assumed to roll 10 times. When we unroll such loop, it appears to
1089 roll too little, and it may even seem to be cold. To avoid this, we
1090 ensure that the created loop appears to roll at least 5 times (but at
1091 most as many times as before unrolling). */
1092 if (new_est_niter
< 5)
1095 new_est_niter
= est_niter
;
1100 /* Prepare the cfg and update the phi nodes. Move the loop exit to the
1101 loop latch (and make its condition dummy, for the moment). */
1102 rest
= loop_preheader_edge (new_loop
)->src
;
1103 precond_edge
= single_pred_edge (rest
);
1104 split_edge (loop_latch_edge (loop
));
1105 exit_bb
= single_pred (loop
->latch
);
1107 /* Since the exit edge will be removed, the frequency of all the blocks
1108 in the loop that are dominated by it must be scaled by
1109 1 / (1 - exit->probability). */
1110 scale_dominated_blocks_in_loop (loop
, exit
->src
,
1112 REG_BR_PROB_BASE
- exit
->probability
);
1114 bsi
= gsi_last_bb (exit_bb
);
1115 exit_if
= gimple_build_cond (EQ_EXPR
, integer_zero_node
,
1117 NULL_TREE
, NULL_TREE
);
1119 gsi_insert_after (&bsi
, exit_if
, GSI_NEW_STMT
);
1120 new_exit
= make_edge (exit_bb
, rest
, EDGE_FALSE_VALUE
| irr
);
1121 rescan_loop_exit (new_exit
, true, false);
1123 /* Set the probability of new exit to the same of the old one. Fix
1124 the frequency of the latch block, by scaling it back by
1125 1 - exit->probability. */
1126 new_exit
->count
= exit
->count
;
1127 new_exit
->probability
= exit
->probability
;
1128 new_nonexit
= single_pred_edge (loop
->latch
);
1129 new_nonexit
->probability
= REG_BR_PROB_BASE
- exit
->probability
;
1130 new_nonexit
->flags
= EDGE_TRUE_VALUE
;
1131 new_nonexit
->count
-= exit
->count
;
1132 if (new_nonexit
->count
< 0)
1133 new_nonexit
->count
= 0;
1134 scale_bbs_frequencies_int (&loop
->latch
, 1, new_nonexit
->probability
,
1137 old_entry
= loop_preheader_edge (loop
);
1138 new_entry
= loop_preheader_edge (new_loop
);
1139 old_latch
= loop_latch_edge (loop
);
1140 for (psi_old_loop
= gsi_start_phis (loop
->header
),
1141 psi_new_loop
= gsi_start_phis (new_loop
->header
);
1142 !gsi_end_p (psi_old_loop
);
1143 gsi_next (&psi_old_loop
), gsi_next (&psi_new_loop
))
1145 phi_old_loop
= gsi_stmt (psi_old_loop
);
1146 phi_new_loop
= gsi_stmt (psi_new_loop
);
1148 init
= PHI_ARG_DEF_FROM_EDGE (phi_old_loop
, old_entry
);
1149 op
= PHI_ARG_DEF_PTR_FROM_EDGE (phi_new_loop
, new_entry
);
1150 gcc_assert (operand_equal_for_phi_arg_p (init
, USE_FROM_PTR (op
)));
1151 next
= PHI_ARG_DEF_FROM_EDGE (phi_old_loop
, old_latch
);
1153 /* Prefer using original variable as a base for the new ssa name.
1154 This is necessary for virtual ops, and useful in order to avoid
1155 losing debug info for real ops. */
1156 if (TREE_CODE (next
) == SSA_NAME
1157 && useless_type_conversion_p (TREE_TYPE (next
),
1159 new_init
= copy_ssa_name (next
, NULL
);
1160 else if (TREE_CODE (init
) == SSA_NAME
1161 && useless_type_conversion_p (TREE_TYPE (init
),
1163 new_init
= copy_ssa_name (init
, NULL
);
1164 else if (useless_type_conversion_p (TREE_TYPE (next
), TREE_TYPE (init
)))
1165 new_init
= make_temp_ssa_name (TREE_TYPE (next
), NULL
, "unrinittmp");
1167 new_init
= make_temp_ssa_name (TREE_TYPE (init
), NULL
, "unrinittmp");
1169 phi_rest
= create_phi_node (new_init
, rest
);
1171 add_phi_arg (phi_rest
, init
, precond_edge
, UNKNOWN_LOCATION
);
1172 add_phi_arg (phi_rest
, next
, new_exit
, UNKNOWN_LOCATION
);
1173 SET_USE (op
, new_init
);
1178 /* Transform the loop. */
1180 (*transform
) (loop
, data
);
1182 /* Unroll the loop and remove the exits in all iterations except for the
1184 wont_exit
= sbitmap_alloc (factor
);
1185 bitmap_ones (wont_exit
);
1186 bitmap_clear_bit (wont_exit
, factor
- 1);
1188 ok
= gimple_duplicate_loop_to_header_edge
1189 (loop
, loop_latch_edge (loop
), factor
- 1,
1190 wont_exit
, new_exit
, &to_remove
, DLTHE_FLAG_UPDATE_FREQ
);
1194 FOR_EACH_VEC_ELT (to_remove
, i
, e
)
1196 ok
= remove_path (e
);
1199 update_ssa (TODO_update_ssa
);
1201 /* Ensure that the frequencies in the loop match the new estimated
1202 number of iterations, and change the probability of the new
1204 freq_h
= loop
->header
->frequency
;
1205 freq_e
= EDGE_FREQUENCY (loop_preheader_edge (loop
));
1207 scale_loop_frequencies (loop
, freq_e
* (new_est_niter
+ 1), freq_h
);
1209 exit_bb
= single_pred (loop
->latch
);
1210 new_exit
= find_edge (exit_bb
, rest
);
1211 new_exit
->count
= loop_preheader_edge (loop
)->count
;
1212 new_exit
->probability
= REG_BR_PROB_BASE
/ (new_est_niter
+ 1);
1214 rest
->count
+= new_exit
->count
;
1215 rest
->frequency
+= EDGE_FREQUENCY (new_exit
);
1217 new_nonexit
= single_pred_edge (loop
->latch
);
1218 prob
= new_nonexit
->probability
;
1219 new_nonexit
->probability
= REG_BR_PROB_BASE
- new_exit
->probability
;
1220 new_nonexit
->count
= exit_bb
->count
- new_exit
->count
;
1221 if (new_nonexit
->count
< 0)
1222 new_nonexit
->count
= 0;
1224 scale_bbs_frequencies_int (&loop
->latch
, 1, new_nonexit
->probability
,
1227 /* Finally create the new counter for number of iterations and add the new
1228 exit instruction. */
1229 bsi
= gsi_last_nondebug_bb (exit_bb
);
1230 exit_if
= gsi_stmt (bsi
);
1231 create_iv (exit_base
, exit_step
, NULL_TREE
, loop
,
1232 &bsi
, false, &ctr_before
, &ctr_after
);
1233 gimple_cond_set_code (exit_if
, exit_cmp
);
1234 gimple_cond_set_lhs (exit_if
, ctr_after
);
1235 gimple_cond_set_rhs (exit_if
, exit_bound
);
1236 update_stmt (exit_if
);
1238 #ifdef ENABLE_CHECKING
1239 verify_flow_info ();
1240 verify_loop_structure ();
1241 verify_loop_closed_ssa (true);
1245 /* Wrapper over tree_transform_and_unroll_loop for case we do not
1246 want to transform the loop before unrolling. The meaning
1247 of the arguments is the same as for tree_transform_and_unroll_loop. */
1250 tree_unroll_loop (struct loop
*loop
, unsigned factor
,
1251 edge exit
, struct tree_niter_desc
*desc
)
1253 tree_transform_and_unroll_loop (loop
, factor
, exit
, desc
,
1257 /* Rewrite the phi node at position PSI in function of the main
1258 induction variable MAIN_IV and insert the generated code at GSI. */
1261 rewrite_phi_with_iv (loop_p loop
,
1262 gimple_stmt_iterator
*psi
,
1263 gimple_stmt_iterator
*gsi
,
1267 gimple stmt
, phi
= gsi_stmt (*psi
);
1268 tree atype
, mtype
, val
, res
= PHI_RESULT (phi
);
1270 if (virtual_operand_p (res
) || res
== main_iv
)
1276 if (!simple_iv (loop
, loop
, res
, &iv
, true))
1282 remove_phi_node (psi
, false);
1284 atype
= TREE_TYPE (res
);
1285 mtype
= POINTER_TYPE_P (atype
) ? sizetype
: atype
;
1286 val
= fold_build2 (MULT_EXPR
, mtype
, unshare_expr (iv
.step
),
1287 fold_convert (mtype
, main_iv
));
1288 val
= fold_build2 (POINTER_TYPE_P (atype
)
1289 ? POINTER_PLUS_EXPR
: PLUS_EXPR
,
1290 atype
, unshare_expr (iv
.base
), val
);
1291 val
= force_gimple_operand_gsi (gsi
, val
, false, NULL_TREE
, true,
1293 stmt
= gimple_build_assign (res
, val
);
1294 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
1297 /* Rewrite all the phi nodes of LOOP in function of the main induction
1298 variable MAIN_IV. */
1301 rewrite_all_phi_nodes_with_iv (loop_p loop
, tree main_iv
)
1304 basic_block
*bbs
= get_loop_body_in_dom_order (loop
);
1305 gimple_stmt_iterator psi
;
1307 for (i
= 0; i
< loop
->num_nodes
; i
++)
1309 basic_block bb
= bbs
[i
];
1310 gimple_stmt_iterator gsi
= gsi_after_labels (bb
);
1312 if (bb
->loop_father
!= loop
)
1315 for (psi
= gsi_start_phis (bb
); !gsi_end_p (psi
); )
1316 rewrite_phi_with_iv (loop
, &psi
, &gsi
, main_iv
);
1322 /* Bases all the induction variables in LOOP on a single induction
1323 variable (unsigned with base 0 and step 1), whose final value is
1324 compared with *NIT. When the IV type precision has to be larger
1325 than *NIT type precision, *NIT is converted to the larger type, the
1326 conversion code is inserted before the loop, and *NIT is updated to
1327 the new definition. When BUMP_IN_LATCH is true, the induction
1328 variable is incremented in the loop latch, otherwise it is
1329 incremented in the loop header. Return the induction variable that
1333 canonicalize_loop_ivs (struct loop
*loop
, tree
*nit
, bool bump_in_latch
)
1335 unsigned precision
= TYPE_PRECISION (TREE_TYPE (*nit
));
1336 unsigned original_precision
= precision
;
1337 tree type
, var_before
;
1338 gimple_stmt_iterator gsi
, psi
;
1340 edge exit
= single_dom_exit (loop
);
1342 enum machine_mode mode
;
1343 bool unsigned_p
= false;
1345 for (psi
= gsi_start_phis (loop
->header
);
1346 !gsi_end_p (psi
); gsi_next (&psi
))
1348 gimple phi
= gsi_stmt (psi
);
1349 tree res
= PHI_RESULT (phi
);
1352 type
= TREE_TYPE (res
);
1353 if (virtual_operand_p (res
)
1354 || (!INTEGRAL_TYPE_P (type
)
1355 && !POINTER_TYPE_P (type
))
1356 || TYPE_PRECISION (type
) < precision
)
1359 uns
= POINTER_TYPE_P (type
) | TYPE_UNSIGNED (type
);
1361 if (TYPE_PRECISION (type
) > precision
)
1366 precision
= TYPE_PRECISION (type
);
1369 mode
= smallest_mode_for_size (precision
, MODE_INT
);
1370 precision
= GET_MODE_PRECISION (mode
);
1371 type
= build_nonstandard_integer_type (precision
, unsigned_p
);
1373 if (original_precision
!= precision
)
1375 *nit
= fold_convert (type
, *nit
);
1376 *nit
= force_gimple_operand (*nit
, &stmts
, true, NULL_TREE
);
1378 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
1382 gsi
= gsi_last_bb (loop
->latch
);
1384 gsi
= gsi_last_nondebug_bb (loop
->header
);
1385 create_iv (build_int_cst_type (type
, 0), build_int_cst (type
, 1), NULL_TREE
,
1386 loop
, &gsi
, bump_in_latch
, &var_before
, NULL
);
1388 rewrite_all_phi_nodes_with_iv (loop
, var_before
);
1390 stmt
= last_stmt (exit
->src
);
1391 /* Make the loop exit if the control condition is not satisfied. */
1392 if (exit
->flags
& EDGE_TRUE_VALUE
)
1396 extract_true_false_edges_from_block (exit
->src
, &te
, &fe
);
1397 te
->flags
= EDGE_FALSE_VALUE
;
1398 fe
->flags
= EDGE_TRUE_VALUE
;
1400 gimple_cond_set_code (stmt
, LT_EXPR
);
1401 gimple_cond_set_lhs (stmt
, var_before
);
1402 gimple_cond_set_rhs (stmt
, *nit
);