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
)
90 enum tree_code incr_op
= PLUS_EXPR
;
91 edge pe
= loop_preheader_edge (loop
);
95 vb
= make_ssa_name (var
);
96 va
= make_ssa_name (var
);
100 vb
= make_temp_ssa_name (TREE_TYPE (base
), NULL
, "ivtmp");
101 va
= make_temp_ssa_name (TREE_TYPE (base
), NULL
, "ivtmp");
108 /* For easier readability of the created code, produce MINUS_EXPRs
110 if (TREE_CODE (step
) == INTEGER_CST
)
112 if (TYPE_UNSIGNED (TREE_TYPE (step
)))
114 step1
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
115 if (tree_int_cst_lt (step1
, step
))
117 incr_op
= MINUS_EXPR
;
125 if (!tree_expr_nonnegative_warnv_p (step
, &ovf
)
126 && may_negate_without_overflow_p (step
))
128 incr_op
= MINUS_EXPR
;
129 step
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
133 if (POINTER_TYPE_P (TREE_TYPE (base
)))
135 if (TREE_CODE (base
) == ADDR_EXPR
)
136 mark_addressable (TREE_OPERAND (base
, 0));
137 step
= convert_to_ptrofftype (step
);
138 if (incr_op
== MINUS_EXPR
)
139 step
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
140 incr_op
= POINTER_PLUS_EXPR
;
142 /* Gimplify the step if necessary. We put the computations in front of the
143 loop (i.e. the step should be loop invariant). */
144 step
= force_gimple_operand (step
, &stmts
, true, NULL_TREE
);
146 gsi_insert_seq_on_edge_immediate (pe
, stmts
);
148 stmt
= gimple_build_assign (va
, incr_op
, vb
, step
);
150 gsi_insert_after (incr_pos
, stmt
, GSI_NEW_STMT
);
152 gsi_insert_before (incr_pos
, stmt
, GSI_NEW_STMT
);
154 initial
= force_gimple_operand (base
, &stmts
, true, var
);
156 gsi_insert_seq_on_edge_immediate (pe
, stmts
);
158 phi
= create_phi_node (vb
, loop
->header
);
159 add_phi_arg (phi
, initial
, loop_preheader_edge (loop
), UNKNOWN_LOCATION
);
160 add_phi_arg (phi
, va
, loop_latch_edge (loop
), UNKNOWN_LOCATION
);
163 /* Return the innermost superloop LOOP of USE_LOOP that is a superloop of
164 both DEF_LOOP and USE_LOOP. */
166 static inline struct loop
*
167 find_sibling_superloop (struct loop
*use_loop
, struct loop
*def_loop
)
169 unsigned ud
= loop_depth (use_loop
);
170 unsigned dd
= loop_depth (def_loop
);
171 gcc_assert (ud
> 0 && dd
> 0);
173 use_loop
= superloop_at_depth (use_loop
, dd
);
175 def_loop
= superloop_at_depth (def_loop
, ud
);
176 while (loop_outer (use_loop
) != loop_outer (def_loop
))
178 use_loop
= loop_outer (use_loop
);
179 def_loop
= loop_outer (def_loop
);
180 gcc_assert (use_loop
&& def_loop
);
185 /* DEF_BB is a basic block containing a DEF that needs rewriting into
186 loop-closed SSA form. USE_BLOCKS is the set of basic blocks containing
187 uses of DEF that "escape" from the loop containing DEF_BB (i.e. blocks in
188 USE_BLOCKS are dominated by DEF_BB but not in the loop father of DEF_B).
189 ALL_EXITS[I] is the set of all basic blocks that exit loop I.
191 Compute the subset of LOOP_EXITS that exit the loop containing DEF_BB
192 or one of its loop fathers, in which DEF is live. This set is returned
193 in the bitmap LIVE_EXITS.
195 Instead of computing the complete livein set of the def, we use the loop
196 nesting tree as a form of poor man's structure analysis. This greatly
197 speeds up the analysis, which is important because this function may be
198 called on all SSA names that need rewriting, one at a time. */
201 compute_live_loop_exits (bitmap live_exits
, bitmap use_blocks
,
202 bitmap
*loop_exits
, basic_block def_bb
)
206 struct loop
*def_loop
= def_bb
->loop_father
;
207 unsigned def_loop_depth
= loop_depth (def_loop
);
208 bitmap def_loop_exits
;
210 /* Normally the work list size is bounded by the number of basic
211 blocks in the largest loop. We don't know this number, but we
212 can be fairly sure that it will be relatively small. */
213 auto_vec
<basic_block
> worklist (MAX (8, n_basic_blocks_for_fn (cfun
) / 128));
215 EXECUTE_IF_SET_IN_BITMAP (use_blocks
, 0, i
, bi
)
217 basic_block use_bb
= BASIC_BLOCK_FOR_FN (cfun
, i
);
218 struct loop
*use_loop
= use_bb
->loop_father
;
219 gcc_checking_assert (def_loop
!= use_loop
220 && ! flow_loop_nested_p (def_loop
, use_loop
));
221 if (! flow_loop_nested_p (use_loop
, def_loop
))
222 use_bb
= find_sibling_superloop (use_loop
, def_loop
)->header
;
223 if (bitmap_set_bit (live_exits
, use_bb
->index
))
224 worklist
.safe_push (use_bb
);
227 /* Iterate until the worklist is empty. */
228 while (! worklist
.is_empty ())
233 /* Pull a block off the worklist. */
234 basic_block bb
= worklist
.pop ();
236 /* Make sure we have at least enough room in the work list
237 for all predecessors of this block. */
238 worklist
.reserve (EDGE_COUNT (bb
->preds
));
240 /* For each predecessor block. */
241 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
243 basic_block pred
= e
->src
;
244 struct loop
*pred_loop
= pred
->loop_father
;
245 unsigned pred_loop_depth
= loop_depth (pred_loop
);
248 /* We should have met DEF_BB along the way. */
249 gcc_assert (pred
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
));
251 if (pred_loop_depth
>= def_loop_depth
)
253 if (pred_loop_depth
> def_loop_depth
)
254 pred_loop
= superloop_at_depth (pred_loop
, def_loop_depth
);
255 /* If we've reached DEF_LOOP, our train ends here. */
256 if (pred_loop
== def_loop
)
259 else if (! flow_loop_nested_p (pred_loop
, def_loop
))
260 pred
= find_sibling_superloop (pred_loop
, def_loop
)->header
;
262 /* Add PRED to the LIVEIN set. PRED_VISITED is true if
263 we had already added PRED to LIVEIN before. */
264 pred_visited
= !bitmap_set_bit (live_exits
, pred
->index
);
266 /* If we have visited PRED before, don't add it to the worklist.
267 If BB dominates PRED, then we're probably looking at a loop.
268 We're only interested in looking up in the dominance tree
269 because DEF_BB dominates all the uses. */
270 if (pred_visited
|| dominated_by_p (CDI_DOMINATORS
, pred
, bb
))
273 worklist
.quick_push (pred
);
277 def_loop_exits
= BITMAP_ALLOC (&loop_renamer_obstack
);
278 for (struct loop
*loop
= def_loop
;
279 loop
!= current_loops
->tree_root
;
280 loop
= loop_outer (loop
))
281 bitmap_ior_into (def_loop_exits
, loop_exits
[loop
->num
]);
282 bitmap_and_into (live_exits
, def_loop_exits
);
283 BITMAP_FREE (def_loop_exits
);
286 /* Add a loop-closing PHI for VAR in basic block EXIT. */
289 add_exit_phi (basic_block exit
, tree var
)
295 #ifdef ENABLE_CHECKING
296 /* Check that at least one of the edges entering the EXIT block exits
297 the loop, or a superloop of that loop, that VAR is defined in. */
298 gimple def_stmt
= SSA_NAME_DEF_STMT (var
);
299 basic_block def_bb
= gimple_bb (def_stmt
);
300 FOR_EACH_EDGE (e
, ei
, exit
->preds
)
302 struct loop
*aloop
= find_common_loop (def_bb
->loop_father
,
303 e
->src
->loop_father
);
304 if (!flow_bb_inside_loop_p (aloop
, e
->dest
))
308 gcc_checking_assert (e
);
311 phi
= create_phi_node (NULL_TREE
, exit
);
312 create_new_def_for (var
, phi
, gimple_phi_result_ptr (phi
));
313 FOR_EACH_EDGE (e
, ei
, exit
->preds
)
314 add_phi_arg (phi
, var
, e
, UNKNOWN_LOCATION
);
316 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
318 fprintf (dump_file
, ";; Created LCSSA PHI: ");
319 print_gimple_stmt (dump_file
, phi
, 0, dump_flags
);
323 /* Add exit phis for VAR that is used in LIVEIN.
324 Exits of the loops are stored in LOOP_EXITS. */
327 add_exit_phis_var (tree var
, bitmap use_blocks
, bitmap
*loop_exits
)
331 basic_block def_bb
= gimple_bb (SSA_NAME_DEF_STMT (var
));
332 bitmap live_exits
= BITMAP_ALLOC (&loop_renamer_obstack
);
334 gcc_checking_assert (! bitmap_bit_p (use_blocks
, def_bb
->index
));
336 compute_live_loop_exits (live_exits
, use_blocks
, loop_exits
, def_bb
);
338 EXECUTE_IF_SET_IN_BITMAP (live_exits
, 0, index
, bi
)
340 add_exit_phi (BASIC_BLOCK_FOR_FN (cfun
, index
), var
);
343 BITMAP_FREE (live_exits
);
346 /* Add exit phis for the names marked in NAMES_TO_RENAME.
347 Exits of the loops are stored in EXITS. Sets of blocks where the ssa
348 names are used are stored in USE_BLOCKS. */
351 add_exit_phis (bitmap names_to_rename
, bitmap
*use_blocks
, bitmap
*loop_exits
)
356 EXECUTE_IF_SET_IN_BITMAP (names_to_rename
, 0, i
, bi
)
358 add_exit_phis_var (ssa_name (i
), use_blocks
[i
], loop_exits
);
362 /* Fill the array of bitmaps LOOP_EXITS with all loop exit edge targets. */
365 get_loops_exits (bitmap
*loop_exits
)
371 FOR_EACH_LOOP (loop
, 0)
373 vec
<edge
> exit_edges
= get_loop_exit_edges (loop
);
374 loop_exits
[loop
->num
] = BITMAP_ALLOC (&loop_renamer_obstack
);
375 FOR_EACH_VEC_ELT (exit_edges
, j
, e
)
376 bitmap_set_bit (loop_exits
[loop
->num
], e
->dest
->index
);
377 exit_edges
.release ();
381 /* For USE in BB, if it is used outside of the loop it is defined in,
382 mark it for rewrite. Record basic block BB where it is used
383 to USE_BLOCKS. Record the ssa name index to NEED_PHIS bitmap. */
386 find_uses_to_rename_use (basic_block bb
, tree use
, bitmap
*use_blocks
,
391 struct loop
*def_loop
;
393 if (TREE_CODE (use
) != SSA_NAME
)
396 ver
= SSA_NAME_VERSION (use
);
397 def_bb
= gimple_bb (SSA_NAME_DEF_STMT (use
));
400 def_loop
= def_bb
->loop_father
;
402 /* If the definition is not inside a loop, it is not interesting. */
403 if (!loop_outer (def_loop
))
406 /* If the use is not outside of the loop it is defined in, it is not
408 if (flow_bb_inside_loop_p (def_loop
, bb
))
411 /* If we're seeing VER for the first time, we still have to allocate
412 a bitmap for its uses. */
413 if (bitmap_set_bit (need_phis
, ver
))
414 use_blocks
[ver
] = BITMAP_ALLOC (&loop_renamer_obstack
);
415 bitmap_set_bit (use_blocks
[ver
], bb
->index
);
418 /* For uses in STMT, mark names that are used outside of the loop they are
419 defined to rewrite. Record the set of blocks in that the ssa
420 names are defined to USE_BLOCKS and the ssa names themselves to
424 find_uses_to_rename_stmt (gimple stmt
, bitmap
*use_blocks
, bitmap need_phis
)
428 basic_block bb
= gimple_bb (stmt
);
430 if (is_gimple_debug (stmt
))
433 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, SSA_OP_USE
)
434 find_uses_to_rename_use (bb
, var
, use_blocks
, need_phis
);
437 /* Marks names that are used in BB and outside of the loop they are
438 defined in for rewrite. Records the set of blocks in that the ssa
439 names are defined to USE_BLOCKS. Record the SSA names that will
440 need exit PHIs in NEED_PHIS. */
443 find_uses_to_rename_bb (basic_block bb
, bitmap
*use_blocks
, bitmap need_phis
)
448 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
449 for (gphi_iterator bsi
= gsi_start_phis (e
->dest
); !gsi_end_p (bsi
);
452 gphi
*phi
= bsi
.phi ();
453 if (! virtual_operand_p (gimple_phi_result (phi
)))
454 find_uses_to_rename_use (bb
, PHI_ARG_DEF_FROM_EDGE (phi
, e
),
455 use_blocks
, need_phis
);
458 for (gimple_stmt_iterator bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
);
460 find_uses_to_rename_stmt (gsi_stmt (bsi
), use_blocks
, need_phis
);
463 /* Marks names that are used outside of the loop they are defined in
464 for rewrite. Records the set of blocks in that the ssa
465 names are defined to USE_BLOCKS. If CHANGED_BBS is not NULL,
466 scan only blocks in this set. */
469 find_uses_to_rename (bitmap changed_bbs
, bitmap
*use_blocks
, bitmap need_phis
)
476 EXECUTE_IF_SET_IN_BITMAP (changed_bbs
, 0, index
, bi
)
477 find_uses_to_rename_bb (BASIC_BLOCK_FOR_FN (cfun
, index
), use_blocks
, need_phis
);
479 FOR_EACH_BB_FN (bb
, cfun
)
480 find_uses_to_rename_bb (bb
, use_blocks
, need_phis
);
483 /* Rewrites the program into a loop closed ssa form -- i.e. inserts extra
484 phi nodes to ensure that no variable is used outside the loop it is
487 This strengthening of the basic ssa form has several advantages:
489 1) Updating it during unrolling/peeling/versioning is trivial, since
490 we do not need to care about the uses outside of the loop.
491 The same applies to virtual operands which are also rewritten into
492 loop closed SSA form. Note that virtual operands are always live
494 2) The behavior of all uses of an induction variable is the same.
495 Without this, you need to distinguish the case when the variable
496 is used outside of the loop it is defined in, for example
498 for (i = 0; i < 100; i++)
500 for (j = 0; j < 100; j++)
508 Looking from the outer loop with the normal SSA form, the first use of k
509 is not well-behaved, while the second one is an induction variable with
512 If CHANGED_BBS is not NULL, we look for uses outside loops only in
513 the basic blocks in this set.
515 UPDATE_FLAG is used in the call to update_ssa. See
516 TODO_update_ssa* for documentation. */
519 rewrite_into_loop_closed_ssa (bitmap changed_bbs
, unsigned update_flag
)
522 bitmap names_to_rename
;
524 loops_state_set (LOOP_CLOSED_SSA
);
525 if (number_of_loops (cfun
) <= 1)
528 /* If the pass has caused the SSA form to be out-of-date, update it
530 update_ssa (update_flag
);
532 bitmap_obstack_initialize (&loop_renamer_obstack
);
534 names_to_rename
= BITMAP_ALLOC (&loop_renamer_obstack
);
536 /* Uses of names to rename. We don't have to initialize this array,
537 because we know that we will only have entries for the SSA names
538 in NAMES_TO_RENAME. */
539 use_blocks
= XNEWVEC (bitmap
, num_ssa_names
);
541 /* Find the uses outside loops. */
542 find_uses_to_rename (changed_bbs
, use_blocks
, names_to_rename
);
544 if (!bitmap_empty_p (names_to_rename
))
546 /* An array of bitmaps where LOOP_EXITS[I] is the set of basic blocks
547 that are the destination of an edge exiting loop number I. */
548 bitmap
*loop_exits
= XNEWVEC (bitmap
, number_of_loops (cfun
));
549 get_loops_exits (loop_exits
);
551 /* Add the PHI nodes on exits of the loops for the names we need to
553 add_exit_phis (names_to_rename
, use_blocks
, loop_exits
);
557 /* Fix up all the names found to be used outside their original
559 update_ssa (TODO_update_ssa
);
562 bitmap_obstack_release (&loop_renamer_obstack
);
566 /* Check invariants of the loop closed ssa form for the USE in BB. */
569 check_loop_closed_ssa_use (basic_block bb
, tree use
)
574 if (TREE_CODE (use
) != SSA_NAME
|| virtual_operand_p (use
))
577 def
= SSA_NAME_DEF_STMT (use
);
578 def_bb
= gimple_bb (def
);
580 || flow_bb_inside_loop_p (def_bb
->loop_father
, bb
));
583 /* Checks invariants of loop closed ssa form in statement STMT in BB. */
586 check_loop_closed_ssa_stmt (basic_block bb
, gimple stmt
)
591 if (is_gimple_debug (stmt
))
594 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, SSA_OP_USE
)
595 check_loop_closed_ssa_use (bb
, var
);
598 /* Checks that invariants of the loop closed ssa form are preserved.
599 Call verify_ssa when VERIFY_SSA_P is true. */
602 verify_loop_closed_ssa (bool verify_ssa_p
)
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 (gphi_iterator bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
);
621 gphi
*phi
= bsi
.phi ();
622 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
623 check_loop_closed_ssa_use (e
->src
,
624 PHI_ARG_DEF_FROM_EDGE (phi
, e
));
627 for (gimple_stmt_iterator bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
);
629 check_loop_closed_ssa_stmt (bb
, gsi_stmt (bsi
));
632 timevar_pop (TV_VERIFY_LOOP_CLOSED
);
635 /* Split loop exit edge EXIT. The things are a bit complicated by a need to
636 preserve the loop closed ssa form. The newly created block is returned. */
639 split_loop_exit_edge (edge exit
)
641 basic_block dest
= exit
->dest
;
642 basic_block bb
= split_edge (exit
);
647 source_location locus
;
649 for (psi
= gsi_start_phis (dest
); !gsi_end_p (psi
); gsi_next (&psi
))
652 op_p
= PHI_ARG_DEF_PTR_FROM_EDGE (phi
, single_succ_edge (bb
));
653 locus
= gimple_phi_arg_location_from_edge (phi
, single_succ_edge (bb
));
655 name
= USE_FROM_PTR (op_p
);
657 /* If the argument of the PHI node is a constant, we do not need
658 to keep it inside loop. */
659 if (TREE_CODE (name
) != SSA_NAME
)
662 /* Otherwise create an auxiliary phi node that will copy the value
663 of the SSA name out of the loop. */
664 new_name
= duplicate_ssa_name (name
, NULL
);
665 new_phi
= create_phi_node (new_name
, bb
);
666 add_phi_arg (new_phi
, name
, exit
, locus
);
667 SET_USE (op_p
, new_name
);
673 /* Returns the basic block in that statements should be emitted for induction
674 variables incremented at the end of the LOOP. */
677 ip_end_pos (struct loop
*loop
)
682 /* Returns the basic block in that statements should be emitted for induction
683 variables incremented just before exit condition of a LOOP. */
686 ip_normal_pos (struct loop
*loop
)
692 if (!single_pred_p (loop
->latch
))
695 bb
= single_pred (loop
->latch
);
696 last
= last_stmt (bb
);
698 || gimple_code (last
) != GIMPLE_COND
)
701 exit
= EDGE_SUCC (bb
, 0);
702 if (exit
->dest
== loop
->latch
)
703 exit
= EDGE_SUCC (bb
, 1);
705 if (flow_bb_inside_loop_p (loop
, exit
->dest
))
711 /* Stores the standard position for induction variable increment in LOOP
712 (just before the exit condition if it is available and latch block is empty,
713 end of the latch block otherwise) to BSI. INSERT_AFTER is set to true if
714 the increment should be inserted after *BSI. */
717 standard_iv_increment_position (struct loop
*loop
, gimple_stmt_iterator
*bsi
,
720 basic_block bb
= ip_normal_pos (loop
), latch
= ip_end_pos (loop
);
721 gimple last
= last_stmt (latch
);
724 || (last
&& gimple_code (last
) != GIMPLE_LABEL
))
726 *bsi
= gsi_last_bb (latch
);
727 *insert_after
= true;
731 *bsi
= gsi_last_bb (bb
);
732 *insert_after
= false;
736 /* Copies phi node arguments for duplicated blocks. The index of the first
737 duplicated block is FIRST_NEW_BLOCK. */
740 copy_phi_node_args (unsigned first_new_block
)
744 for (i
= first_new_block
; i
< (unsigned) last_basic_block_for_fn (cfun
); i
++)
745 BASIC_BLOCK_FOR_FN (cfun
, i
)->flags
|= BB_DUPLICATED
;
747 for (i
= first_new_block
; i
< (unsigned) last_basic_block_for_fn (cfun
); i
++)
748 add_phi_args_after_copy_bb (BASIC_BLOCK_FOR_FN (cfun
, i
));
750 for (i
= first_new_block
; i
< (unsigned) last_basic_block_for_fn (cfun
); i
++)
751 BASIC_BLOCK_FOR_FN (cfun
, i
)->flags
&= ~BB_DUPLICATED
;
755 /* The same as cfgloopmanip.c:duplicate_loop_to_header_edge, but also
756 updates the PHI nodes at start of the copied region. In order to
757 achieve this, only loops whose exits all lead to the same location
760 Notice that we do not completely update the SSA web after
761 duplication. The caller is responsible for calling update_ssa
762 after the loop has been duplicated. */
765 gimple_duplicate_loop_to_header_edge (struct loop
*loop
, edge e
,
766 unsigned int ndupl
, sbitmap wont_exit
,
767 edge orig
, vec
<edge
> *to_remove
,
770 unsigned first_new_block
;
772 if (!loops_state_satisfies_p (LOOPS_HAVE_SIMPLE_LATCHES
))
774 if (!loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS
))
777 first_new_block
= last_basic_block_for_fn (cfun
);
778 if (!duplicate_loop_to_header_edge (loop
, e
, ndupl
, wont_exit
,
779 orig
, to_remove
, flags
))
782 /* Readd the removed phi args for e. */
783 flush_pending_stmts (e
);
785 /* Copy the phi node arguments. */
786 copy_phi_node_args (first_new_block
);
793 /* Returns true if we can unroll LOOP FACTOR times. Number
794 of iterations of the loop is returned in NITER. */
797 can_unroll_loop_p (struct loop
*loop
, unsigned factor
,
798 struct tree_niter_desc
*niter
)
802 /* Check whether unrolling is possible. We only want to unroll loops
803 for that we are able to determine number of iterations. We also
804 want to split the extra iterations of the loop from its end,
805 therefore we require that the loop has precisely one
808 exit
= single_dom_exit (loop
);
812 if (!number_of_iterations_exit (loop
, exit
, niter
, false)
813 || niter
->cmp
== ERROR_MARK
814 /* Scalar evolutions analysis might have copy propagated
815 the abnormal ssa names into these expressions, hence
816 emitting the computations based on them during loop
817 unrolling might create overlapping life ranges for
818 them, and failures in out-of-ssa. */
819 || contains_abnormal_ssa_name_p (niter
->may_be_zero
)
820 || contains_abnormal_ssa_name_p (niter
->control
.base
)
821 || contains_abnormal_ssa_name_p (niter
->control
.step
)
822 || contains_abnormal_ssa_name_p (niter
->bound
))
825 /* And of course, we must be able to duplicate the loop. */
826 if (!can_duplicate_loop_p (loop
))
829 /* The final loop should be small enough. */
830 if (tree_num_loop_insns (loop
, &eni_size_weights
) * factor
831 > (unsigned) PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS
))
837 /* Determines the conditions that control execution of LOOP unrolled FACTOR
838 times. DESC is number of iterations of LOOP. ENTER_COND is set to
839 condition that must be true if the main loop can be entered.
840 EXIT_BASE, EXIT_STEP, EXIT_CMP and EXIT_BOUND are set to values describing
841 how the exit from the unrolled loop should be controlled. */
844 determine_exit_conditions (struct loop
*loop
, struct tree_niter_desc
*desc
,
845 unsigned factor
, tree
*enter_cond
,
846 tree
*exit_base
, tree
*exit_step
,
847 enum tree_code
*exit_cmp
, tree
*exit_bound
)
850 tree base
= desc
->control
.base
;
851 tree step
= desc
->control
.step
;
852 tree bound
= desc
->bound
;
853 tree type
= TREE_TYPE (step
);
855 tree min
= lower_bound_in_type (type
, type
);
856 tree max
= upper_bound_in_type (type
, type
);
857 enum tree_code cmp
= desc
->cmp
;
858 tree cond
= boolean_true_node
, assum
;
860 /* For pointers, do the arithmetics in the type of step. */
861 base
= fold_convert (type
, base
);
862 bound
= fold_convert (type
, bound
);
864 *enter_cond
= boolean_false_node
;
865 *exit_base
= NULL_TREE
;
866 *exit_step
= NULL_TREE
;
867 *exit_cmp
= ERROR_MARK
;
868 *exit_bound
= NULL_TREE
;
869 gcc_assert (cmp
!= ERROR_MARK
);
871 /* We only need to be correct when we answer question
872 "Do at least FACTOR more iterations remain?" in the unrolled loop.
873 Thus, transforming BASE + STEP * i <> BOUND to
874 BASE + STEP * i < BOUND is ok. */
877 if (tree_int_cst_sign_bit (step
))
882 else if (cmp
== LT_EXPR
)
884 gcc_assert (!tree_int_cst_sign_bit (step
));
886 else if (cmp
== GT_EXPR
)
888 gcc_assert (tree_int_cst_sign_bit (step
));
893 /* The main body of the loop may be entered iff:
895 1) desc->may_be_zero is false.
896 2) it is possible to check that there are at least FACTOR iterations
897 of the loop, i.e., BOUND - step * FACTOR does not overflow.
898 3) # of iterations is at least FACTOR */
900 if (!integer_zerop (desc
->may_be_zero
))
901 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
902 invert_truthvalue (desc
->may_be_zero
),
905 bigstep
= fold_build2 (MULT_EXPR
, type
, step
,
906 build_int_cst_type (type
, factor
));
907 delta
= fold_build2 (MINUS_EXPR
, type
, bigstep
, step
);
909 assum
= fold_build2 (GE_EXPR
, boolean_type_node
,
911 fold_build2 (PLUS_EXPR
, type
, min
, delta
));
913 assum
= fold_build2 (LE_EXPR
, boolean_type_node
,
915 fold_build2 (PLUS_EXPR
, type
, max
, delta
));
916 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
, assum
, cond
);
918 bound
= fold_build2 (MINUS_EXPR
, type
, bound
, delta
);
919 assum
= fold_build2 (cmp
, boolean_type_node
, base
, bound
);
920 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
, assum
, cond
);
922 cond
= force_gimple_operand (unshare_expr (cond
), &stmts
, false, NULL_TREE
);
924 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
925 /* cond now may be a gimple comparison, which would be OK, but also any
926 other gimple rhs (say a && b). In this case we need to force it to
928 if (!is_gimple_condexpr (cond
))
930 cond
= force_gimple_operand (cond
, &stmts
, true, NULL_TREE
);
932 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
936 base
= force_gimple_operand (unshare_expr (base
), &stmts
, true, NULL_TREE
);
938 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
939 bound
= force_gimple_operand (unshare_expr (bound
), &stmts
, true, NULL_TREE
);
941 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
944 *exit_step
= bigstep
;
949 /* Scales the frequencies of all basic blocks in LOOP that are strictly
950 dominated by BB by NUM/DEN. */
953 scale_dominated_blocks_in_loop (struct loop
*loop
, basic_block bb
,
961 for (son
= first_dom_son (CDI_DOMINATORS
, bb
);
963 son
= next_dom_son (CDI_DOMINATORS
, son
))
965 if (!flow_bb_inside_loop_p (loop
, son
))
967 scale_bbs_frequencies_int (&son
, 1, num
, den
);
968 scale_dominated_blocks_in_loop (loop
, son
, num
, den
);
972 /* Unroll LOOP FACTOR times. DESC describes number of iterations of LOOP.
973 EXIT is the exit of the loop to that DESC corresponds.
975 If N is number of iterations of the loop and MAY_BE_ZERO is the condition
976 under that loop exits in the first iteration even if N != 0,
980 x = phi (init, next);
988 becomes (with possibly the exit conditions formulated a bit differently,
989 avoiding the need to create a new iv):
991 if (MAY_BE_ZERO || N < FACTOR)
996 x = phi (init, next);
1007 } while (N >= FACTOR);
1010 init' = phi (init, x);
1014 x = phi (init', next);
1022 Before the loop is unrolled, TRANSFORM is called for it (only for the
1023 unrolled loop, but not for its versioned copy). DATA is passed to
1026 /* Probability in % that the unrolled loop is entered. Just a guess. */
1027 #define PROB_UNROLLED_LOOP_ENTERED 90
1030 tree_transform_and_unroll_loop (struct loop
*loop
, unsigned factor
,
1031 edge exit
, struct tree_niter_desc
*desc
,
1032 transform_callback transform
,
1036 tree ctr_before
, ctr_after
;
1037 tree enter_main_cond
, exit_base
, exit_step
, exit_bound
;
1038 enum tree_code exit_cmp
;
1039 gphi
*phi_old_loop
, *phi_new_loop
, *phi_rest
;
1040 gphi_iterator psi_old_loop
, psi_new_loop
;
1041 tree init
, next
, new_init
;
1042 struct loop
*new_loop
;
1043 basic_block rest
, exit_bb
;
1044 edge old_entry
, new_entry
, old_latch
, precond_edge
, new_exit
;
1045 edge new_nonexit
, e
;
1046 gimple_stmt_iterator bsi
;
1049 unsigned est_niter
, prob_entry
, scale_unrolled
, scale_rest
, freq_e
, freq_h
;
1050 unsigned new_est_niter
, i
, prob
;
1051 unsigned irr
= loop_preheader_edge (loop
)->flags
& EDGE_IRREDUCIBLE_LOOP
;
1053 auto_vec
<edge
> to_remove
;
1055 est_niter
= expected_loop_iterations (loop
);
1056 determine_exit_conditions (loop
, desc
, factor
,
1057 &enter_main_cond
, &exit_base
, &exit_step
,
1058 &exit_cmp
, &exit_bound
);
1060 /* Let us assume that the unrolled loop is quite likely to be entered. */
1061 if (integer_nonzerop (enter_main_cond
))
1062 prob_entry
= REG_BR_PROB_BASE
;
1064 prob_entry
= PROB_UNROLLED_LOOP_ENTERED
* REG_BR_PROB_BASE
/ 100;
1066 /* The values for scales should keep profile consistent, and somewhat close
1069 TODO: The current value of SCALE_REST makes it appear that the loop that
1070 is created by splitting the remaining iterations of the unrolled loop is
1071 executed the same number of times as the original loop, and with the same
1072 frequencies, which is obviously wrong. This does not appear to cause
1073 problems, so we do not bother with fixing it for now. To make the profile
1074 correct, we would need to change the probability of the exit edge of the
1075 loop, and recompute the distribution of frequencies in its body because
1076 of this change (scale the frequencies of blocks before and after the exit
1077 by appropriate factors). */
1078 scale_unrolled
= prob_entry
;
1079 scale_rest
= REG_BR_PROB_BASE
;
1081 new_loop
= loop_version (loop
, enter_main_cond
, NULL
,
1082 prob_entry
, scale_unrolled
, scale_rest
, true);
1083 gcc_assert (new_loop
!= NULL
);
1084 update_ssa (TODO_update_ssa
);
1086 /* Determine the probability of the exit edge of the unrolled loop. */
1087 new_est_niter
= est_niter
/ factor
;
1089 /* Without profile feedback, loops for that we do not know a better estimate
1090 are assumed to roll 10 times. When we unroll such loop, it appears to
1091 roll too little, and it may even seem to be cold. To avoid this, we
1092 ensure that the created loop appears to roll at least 5 times (but at
1093 most as many times as before unrolling). */
1094 if (new_est_niter
< 5)
1097 new_est_niter
= est_niter
;
1102 /* Prepare the cfg and update the phi nodes. Move the loop exit to the
1103 loop latch (and make its condition dummy, for the moment). */
1104 rest
= loop_preheader_edge (new_loop
)->src
;
1105 precond_edge
= single_pred_edge (rest
);
1106 split_edge (loop_latch_edge (loop
));
1107 exit_bb
= single_pred (loop
->latch
);
1109 /* Since the exit edge will be removed, the frequency of all the blocks
1110 in the loop that are dominated by it must be scaled by
1111 1 / (1 - exit->probability). */
1112 scale_dominated_blocks_in_loop (loop
, exit
->src
,
1114 REG_BR_PROB_BASE
- exit
->probability
);
1116 bsi
= gsi_last_bb (exit_bb
);
1117 exit_if
= gimple_build_cond (EQ_EXPR
, integer_zero_node
,
1119 NULL_TREE
, NULL_TREE
);
1121 gsi_insert_after (&bsi
, exit_if
, GSI_NEW_STMT
);
1122 new_exit
= make_edge (exit_bb
, rest
, EDGE_FALSE_VALUE
| irr
);
1123 rescan_loop_exit (new_exit
, true, false);
1125 /* Set the probability of new exit to the same of the old one. Fix
1126 the frequency of the latch block, by scaling it back by
1127 1 - exit->probability. */
1128 new_exit
->count
= exit
->count
;
1129 new_exit
->probability
= exit
->probability
;
1130 new_nonexit
= single_pred_edge (loop
->latch
);
1131 new_nonexit
->probability
= REG_BR_PROB_BASE
- exit
->probability
;
1132 new_nonexit
->flags
= EDGE_TRUE_VALUE
;
1133 new_nonexit
->count
-= exit
->count
;
1134 if (new_nonexit
->count
< 0)
1135 new_nonexit
->count
= 0;
1136 scale_bbs_frequencies_int (&loop
->latch
, 1, new_nonexit
->probability
,
1139 old_entry
= loop_preheader_edge (loop
);
1140 new_entry
= loop_preheader_edge (new_loop
);
1141 old_latch
= loop_latch_edge (loop
);
1142 for (psi_old_loop
= gsi_start_phis (loop
->header
),
1143 psi_new_loop
= gsi_start_phis (new_loop
->header
);
1144 !gsi_end_p (psi_old_loop
);
1145 gsi_next (&psi_old_loop
), gsi_next (&psi_new_loop
))
1147 phi_old_loop
= psi_old_loop
.phi ();
1148 phi_new_loop
= psi_new_loop
.phi ();
1150 init
= PHI_ARG_DEF_FROM_EDGE (phi_old_loop
, old_entry
);
1151 op
= PHI_ARG_DEF_PTR_FROM_EDGE (phi_new_loop
, new_entry
);
1152 gcc_assert (operand_equal_for_phi_arg_p (init
, USE_FROM_PTR (op
)));
1153 next
= PHI_ARG_DEF_FROM_EDGE (phi_old_loop
, old_latch
);
1155 /* Prefer using original variable as a base for the new ssa name.
1156 This is necessary for virtual ops, and useful in order to avoid
1157 losing debug info for real ops. */
1158 if (TREE_CODE (next
) == SSA_NAME
1159 && useless_type_conversion_p (TREE_TYPE (next
),
1161 new_init
= copy_ssa_name (next
);
1162 else if (TREE_CODE (init
) == SSA_NAME
1163 && useless_type_conversion_p (TREE_TYPE (init
),
1165 new_init
= copy_ssa_name (init
);
1166 else if (useless_type_conversion_p (TREE_TYPE (next
), TREE_TYPE (init
)))
1167 new_init
= make_temp_ssa_name (TREE_TYPE (next
), NULL
, "unrinittmp");
1169 new_init
= make_temp_ssa_name (TREE_TYPE (init
), NULL
, "unrinittmp");
1171 phi_rest
= create_phi_node (new_init
, rest
);
1173 add_phi_arg (phi_rest
, init
, precond_edge
, UNKNOWN_LOCATION
);
1174 add_phi_arg (phi_rest
, next
, new_exit
, UNKNOWN_LOCATION
);
1175 SET_USE (op
, new_init
);
1180 /* Transform the loop. */
1182 (*transform
) (loop
, data
);
1184 /* Unroll the loop and remove the exits in all iterations except for the
1186 wont_exit
= sbitmap_alloc (factor
);
1187 bitmap_ones (wont_exit
);
1188 bitmap_clear_bit (wont_exit
, factor
- 1);
1190 ok
= gimple_duplicate_loop_to_header_edge
1191 (loop
, loop_latch_edge (loop
), factor
- 1,
1192 wont_exit
, new_exit
, &to_remove
, DLTHE_FLAG_UPDATE_FREQ
);
1196 FOR_EACH_VEC_ELT (to_remove
, i
, e
)
1198 ok
= remove_path (e
);
1201 update_ssa (TODO_update_ssa
);
1203 /* Ensure that the frequencies in the loop match the new estimated
1204 number of iterations, and change the probability of the new
1206 freq_h
= loop
->header
->frequency
;
1207 freq_e
= EDGE_FREQUENCY (loop_preheader_edge (loop
));
1209 scale_loop_frequencies (loop
, freq_e
* (new_est_niter
+ 1), freq_h
);
1211 exit_bb
= single_pred (loop
->latch
);
1212 new_exit
= find_edge (exit_bb
, rest
);
1213 new_exit
->count
= loop_preheader_edge (loop
)->count
;
1214 new_exit
->probability
= REG_BR_PROB_BASE
/ (new_est_niter
+ 1);
1216 rest
->count
+= new_exit
->count
;
1217 rest
->frequency
+= EDGE_FREQUENCY (new_exit
);
1219 new_nonexit
= single_pred_edge (loop
->latch
);
1220 prob
= new_nonexit
->probability
;
1221 new_nonexit
->probability
= REG_BR_PROB_BASE
- new_exit
->probability
;
1222 new_nonexit
->count
= exit_bb
->count
- new_exit
->count
;
1223 if (new_nonexit
->count
< 0)
1224 new_nonexit
->count
= 0;
1226 scale_bbs_frequencies_int (&loop
->latch
, 1, new_nonexit
->probability
,
1229 /* Finally create the new counter for number of iterations and add the new
1230 exit instruction. */
1231 bsi
= gsi_last_nondebug_bb (exit_bb
);
1232 exit_if
= as_a
<gcond
*> (gsi_stmt (bsi
));
1233 create_iv (exit_base
, exit_step
, NULL_TREE
, loop
,
1234 &bsi
, false, &ctr_before
, &ctr_after
);
1235 gimple_cond_set_code (exit_if
, exit_cmp
);
1236 gimple_cond_set_lhs (exit_if
, ctr_after
);
1237 gimple_cond_set_rhs (exit_if
, exit_bound
);
1238 update_stmt (exit_if
);
1240 #ifdef ENABLE_CHECKING
1241 verify_flow_info ();
1242 verify_loop_structure ();
1243 verify_loop_closed_ssa (true);
1247 /* Wrapper over tree_transform_and_unroll_loop for case we do not
1248 want to transform the loop before unrolling. The meaning
1249 of the arguments is the same as for tree_transform_and_unroll_loop. */
1252 tree_unroll_loop (struct loop
*loop
, unsigned factor
,
1253 edge exit
, struct tree_niter_desc
*desc
)
1255 tree_transform_and_unroll_loop (loop
, factor
, exit
, desc
,
1259 /* Rewrite the phi node at position PSI in function of the main
1260 induction variable MAIN_IV and insert the generated code at GSI. */
1263 rewrite_phi_with_iv (loop_p loop
,
1265 gimple_stmt_iterator
*gsi
,
1270 gphi
*phi
= psi
->phi ();
1271 tree atype
, mtype
, val
, res
= PHI_RESULT (phi
);
1273 if (virtual_operand_p (res
) || res
== main_iv
)
1279 if (!simple_iv (loop
, loop
, res
, &iv
, true))
1285 remove_phi_node (psi
, false);
1287 atype
= TREE_TYPE (res
);
1288 mtype
= POINTER_TYPE_P (atype
) ? sizetype
: atype
;
1289 val
= fold_build2 (MULT_EXPR
, mtype
, unshare_expr (iv
.step
),
1290 fold_convert (mtype
, main_iv
));
1291 val
= fold_build2 (POINTER_TYPE_P (atype
)
1292 ? POINTER_PLUS_EXPR
: PLUS_EXPR
,
1293 atype
, unshare_expr (iv
.base
), val
);
1294 val
= force_gimple_operand_gsi (gsi
, val
, false, NULL_TREE
, true,
1296 stmt
= gimple_build_assign (res
, val
);
1297 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
1300 /* Rewrite all the phi nodes of LOOP in function of the main induction
1301 variable MAIN_IV. */
1304 rewrite_all_phi_nodes_with_iv (loop_p loop
, tree main_iv
)
1307 basic_block
*bbs
= get_loop_body_in_dom_order (loop
);
1310 for (i
= 0; i
< loop
->num_nodes
; i
++)
1312 basic_block bb
= bbs
[i
];
1313 gimple_stmt_iterator gsi
= gsi_after_labels (bb
);
1315 if (bb
->loop_father
!= loop
)
1318 for (psi
= gsi_start_phis (bb
); !gsi_end_p (psi
); )
1319 rewrite_phi_with_iv (loop
, &psi
, &gsi
, main_iv
);
1325 /* Bases all the induction variables in LOOP on a single induction
1326 variable (unsigned with base 0 and step 1), whose final value is
1327 compared with *NIT. When the IV type precision has to be larger
1328 than *NIT type precision, *NIT is converted to the larger type, the
1329 conversion code is inserted before the loop, and *NIT is updated to
1330 the new definition. When BUMP_IN_LATCH is true, the induction
1331 variable is incremented in the loop latch, otherwise it is
1332 incremented in the loop header. Return the induction variable that
1336 canonicalize_loop_ivs (struct loop
*loop
, tree
*nit
, bool bump_in_latch
)
1338 unsigned precision
= TYPE_PRECISION (TREE_TYPE (*nit
));
1339 unsigned original_precision
= precision
;
1340 tree type
, var_before
;
1341 gimple_stmt_iterator gsi
;
1344 edge exit
= single_dom_exit (loop
);
1347 bool unsigned_p
= false;
1349 for (psi
= gsi_start_phis (loop
->header
);
1350 !gsi_end_p (psi
); gsi_next (&psi
))
1352 gphi
*phi
= psi
.phi ();
1353 tree res
= PHI_RESULT (phi
);
1356 type
= TREE_TYPE (res
);
1357 if (virtual_operand_p (res
)
1358 || (!INTEGRAL_TYPE_P (type
)
1359 && !POINTER_TYPE_P (type
))
1360 || TYPE_PRECISION (type
) < precision
)
1363 uns
= POINTER_TYPE_P (type
) | TYPE_UNSIGNED (type
);
1365 if (TYPE_PRECISION (type
) > precision
)
1370 precision
= TYPE_PRECISION (type
);
1373 mode
= smallest_mode_for_size (precision
, MODE_INT
);
1374 precision
= GET_MODE_PRECISION (mode
);
1375 type
= build_nonstandard_integer_type (precision
, unsigned_p
);
1377 if (original_precision
!= precision
)
1379 *nit
= fold_convert (type
, *nit
);
1380 *nit
= force_gimple_operand (*nit
, &stmts
, true, NULL_TREE
);
1382 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
1386 gsi
= gsi_last_bb (loop
->latch
);
1388 gsi
= gsi_last_nondebug_bb (loop
->header
);
1389 create_iv (build_int_cst_type (type
, 0), build_int_cst (type
, 1), NULL_TREE
,
1390 loop
, &gsi
, bump_in_latch
, &var_before
, NULL
);
1392 rewrite_all_phi_nodes_with_iv (loop
, var_before
);
1394 stmt
= as_a
<gcond
*> (last_stmt (exit
->src
));
1395 /* Make the loop exit if the control condition is not satisfied. */
1396 if (exit
->flags
& EDGE_TRUE_VALUE
)
1400 extract_true_false_edges_from_block (exit
->src
, &te
, &fe
);
1401 te
->flags
= EDGE_FALSE_VALUE
;
1402 fe
->flags
= EDGE_TRUE_VALUE
;
1404 gimple_cond_set_code (stmt
, LT_EXPR
);
1405 gimple_cond_set_lhs (stmt
, var_before
);
1406 gimple_cond_set_rhs (stmt
, *nit
);