1 /* High-level loop manipulation functions.
2 Copyright (C) 2004-2015 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"
27 #include "fold-const.h"
30 #include "hard-reg-set.h"
32 #include "dominance.h"
35 #include "basic-block.h"
36 #include "tree-ssa-alias.h"
37 #include "internal-fn.h"
38 #include "gimple-expr.h"
41 #include "gimple-iterator.h"
42 #include "gimplify-me.h"
43 #include "gimple-ssa.h"
45 #include "tree-phinodes.h"
46 #include "ssa-iterators.h"
47 #include "stringpool.h"
48 #include "tree-ssanames.h"
49 #include "tree-ssa-loop-ivopts.h"
50 #include "tree-ssa-loop-manip.h"
51 #include "tree-ssa-loop-niter.h"
52 #include "tree-ssa-loop.h"
53 #include "tree-into-ssa.h"
56 #include "gimple-pretty-print.h"
58 #include "tree-pass.h" /* ??? for TODO_update_ssa but this isn't a pass. */
59 #include "tree-scalar-evolution.h"
61 #include "tree-inline.h"
62 #include "langhooks.h"
64 /* All bitmaps for rewriting into loop-closed SSA go on this obstack,
65 so that we can free them all at once. */
66 static bitmap_obstack loop_renamer_obstack
;
68 /* Creates an induction variable with value BASE + STEP * iteration in LOOP.
69 It is expected that neither BASE nor STEP are shared with other expressions
70 (unless the sharing rules allow this). Use VAR as a base var_decl for it
71 (if NULL, a new temporary will be created). The increment will occur at
72 INCR_POS (after it if AFTER is true, before it otherwise). INCR_POS and
73 AFTER can be computed using standard_iv_increment_position. The ssa versions
74 of the variable before and after increment will be stored in VAR_BEFORE and
75 VAR_AFTER (unless they are NULL). */
78 create_iv (tree base
, tree step
, tree var
, struct loop
*loop
,
79 gimple_stmt_iterator
*incr_pos
, bool after
,
80 tree
*var_before
, tree
*var_after
)
87 enum tree_code incr_op
= PLUS_EXPR
;
88 edge pe
= loop_preheader_edge (loop
);
92 vb
= make_ssa_name (var
);
93 va
= make_ssa_name (var
);
97 vb
= make_temp_ssa_name (TREE_TYPE (base
), NULL
, "ivtmp");
98 va
= make_temp_ssa_name (TREE_TYPE (base
), NULL
, "ivtmp");
105 /* For easier readability of the created code, produce MINUS_EXPRs
107 if (TREE_CODE (step
) == INTEGER_CST
)
109 if (TYPE_UNSIGNED (TREE_TYPE (step
)))
111 step1
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
112 if (tree_int_cst_lt (step1
, step
))
114 incr_op
= MINUS_EXPR
;
122 if (!tree_expr_nonnegative_warnv_p (step
, &ovf
)
123 && may_negate_without_overflow_p (step
))
125 incr_op
= MINUS_EXPR
;
126 step
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
130 if (POINTER_TYPE_P (TREE_TYPE (base
)))
132 if (TREE_CODE (base
) == ADDR_EXPR
)
133 mark_addressable (TREE_OPERAND (base
, 0));
134 step
= convert_to_ptrofftype (step
);
135 if (incr_op
== MINUS_EXPR
)
136 step
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
137 incr_op
= POINTER_PLUS_EXPR
;
139 /* Gimplify the step if necessary. We put the computations in front of the
140 loop (i.e. the step should be loop invariant). */
141 step
= force_gimple_operand (step
, &stmts
, true, NULL_TREE
);
143 gsi_insert_seq_on_edge_immediate (pe
, stmts
);
145 stmt
= gimple_build_assign (va
, incr_op
, vb
, step
);
147 gsi_insert_after (incr_pos
, stmt
, GSI_NEW_STMT
);
149 gsi_insert_before (incr_pos
, stmt
, GSI_NEW_STMT
);
151 initial
= force_gimple_operand (base
, &stmts
, true, var
);
153 gsi_insert_seq_on_edge_immediate (pe
, stmts
);
155 phi
= create_phi_node (vb
, loop
->header
);
156 add_phi_arg (phi
, initial
, loop_preheader_edge (loop
), UNKNOWN_LOCATION
);
157 add_phi_arg (phi
, va
, loop_latch_edge (loop
), UNKNOWN_LOCATION
);
160 /* Return the innermost superloop LOOP of USE_LOOP that is a superloop of
161 both DEF_LOOP and USE_LOOP. */
163 static inline struct loop
*
164 find_sibling_superloop (struct loop
*use_loop
, struct loop
*def_loop
)
166 unsigned ud
= loop_depth (use_loop
);
167 unsigned dd
= loop_depth (def_loop
);
168 gcc_assert (ud
> 0 && dd
> 0);
170 use_loop
= superloop_at_depth (use_loop
, dd
);
172 def_loop
= superloop_at_depth (def_loop
, ud
);
173 while (loop_outer (use_loop
) != loop_outer (def_loop
))
175 use_loop
= loop_outer (use_loop
);
176 def_loop
= loop_outer (def_loop
);
177 gcc_assert (use_loop
&& def_loop
);
182 /* DEF_BB is a basic block containing a DEF that needs rewriting into
183 loop-closed SSA form. USE_BLOCKS is the set of basic blocks containing
184 uses of DEF that "escape" from the loop containing DEF_BB (i.e. blocks in
185 USE_BLOCKS are dominated by DEF_BB but not in the loop father of DEF_B).
186 ALL_EXITS[I] is the set of all basic blocks that exit loop I.
188 Compute the subset of LOOP_EXITS that exit the loop containing DEF_BB
189 or one of its loop fathers, in which DEF is live. This set is returned
190 in the bitmap LIVE_EXITS.
192 Instead of computing the complete livein set of the def, we use the loop
193 nesting tree as a form of poor man's structure analysis. This greatly
194 speeds up the analysis, which is important because this function may be
195 called on all SSA names that need rewriting, one at a time. */
198 compute_live_loop_exits (bitmap live_exits
, bitmap use_blocks
,
199 bitmap
*loop_exits
, basic_block def_bb
)
203 struct loop
*def_loop
= def_bb
->loop_father
;
204 unsigned def_loop_depth
= loop_depth (def_loop
);
205 bitmap def_loop_exits
;
207 /* Normally the work list size is bounded by the number of basic
208 blocks in the largest loop. We don't know this number, but we
209 can be fairly sure that it will be relatively small. */
210 auto_vec
<basic_block
> worklist (MAX (8, n_basic_blocks_for_fn (cfun
) / 128));
212 EXECUTE_IF_SET_IN_BITMAP (use_blocks
, 0, i
, bi
)
214 basic_block use_bb
= BASIC_BLOCK_FOR_FN (cfun
, i
);
215 struct loop
*use_loop
= use_bb
->loop_father
;
216 gcc_checking_assert (def_loop
!= use_loop
217 && ! flow_loop_nested_p (def_loop
, use_loop
));
218 if (! flow_loop_nested_p (use_loop
, def_loop
))
219 use_bb
= find_sibling_superloop (use_loop
, def_loop
)->header
;
220 if (bitmap_set_bit (live_exits
, use_bb
->index
))
221 worklist
.safe_push (use_bb
);
224 /* Iterate until the worklist is empty. */
225 while (! worklist
.is_empty ())
230 /* Pull a block off the worklist. */
231 basic_block bb
= worklist
.pop ();
233 /* Make sure we have at least enough room in the work list
234 for all predecessors of this block. */
235 worklist
.reserve (EDGE_COUNT (bb
->preds
));
237 /* For each predecessor block. */
238 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
240 basic_block pred
= e
->src
;
241 struct loop
*pred_loop
= pred
->loop_father
;
242 unsigned pred_loop_depth
= loop_depth (pred_loop
);
245 /* We should have met DEF_BB along the way. */
246 gcc_assert (pred
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
));
248 if (pred_loop_depth
>= def_loop_depth
)
250 if (pred_loop_depth
> def_loop_depth
)
251 pred_loop
= superloop_at_depth (pred_loop
, def_loop_depth
);
252 /* If we've reached DEF_LOOP, our train ends here. */
253 if (pred_loop
== def_loop
)
256 else if (! flow_loop_nested_p (pred_loop
, def_loop
))
257 pred
= find_sibling_superloop (pred_loop
, def_loop
)->header
;
259 /* Add PRED to the LIVEIN set. PRED_VISITED is true if
260 we had already added PRED to LIVEIN before. */
261 pred_visited
= !bitmap_set_bit (live_exits
, pred
->index
);
263 /* If we have visited PRED before, don't add it to the worklist.
264 If BB dominates PRED, then we're probably looking at a loop.
265 We're only interested in looking up in the dominance tree
266 because DEF_BB dominates all the uses. */
267 if (pred_visited
|| dominated_by_p (CDI_DOMINATORS
, pred
, bb
))
270 worklist
.quick_push (pred
);
274 def_loop_exits
= BITMAP_ALLOC (&loop_renamer_obstack
);
275 for (struct loop
*loop
= def_loop
;
276 loop
!= current_loops
->tree_root
;
277 loop
= loop_outer (loop
))
278 bitmap_ior_into (def_loop_exits
, loop_exits
[loop
->num
]);
279 bitmap_and_into (live_exits
, def_loop_exits
);
280 BITMAP_FREE (def_loop_exits
);
283 /* Add a loop-closing PHI for VAR in basic block EXIT. */
286 add_exit_phi (basic_block exit
, tree var
)
292 #ifdef ENABLE_CHECKING
293 /* Check that at least one of the edges entering the EXIT block exits
294 the loop, or a superloop of that loop, that VAR is defined in. */
295 gimple def_stmt
= SSA_NAME_DEF_STMT (var
);
296 basic_block def_bb
= gimple_bb (def_stmt
);
297 FOR_EACH_EDGE (e
, ei
, exit
->preds
)
299 struct loop
*aloop
= find_common_loop (def_bb
->loop_father
,
300 e
->src
->loop_father
);
301 if (!flow_bb_inside_loop_p (aloop
, e
->dest
))
305 gcc_checking_assert (e
);
308 phi
= create_phi_node (NULL_TREE
, exit
);
309 create_new_def_for (var
, phi
, gimple_phi_result_ptr (phi
));
310 FOR_EACH_EDGE (e
, ei
, exit
->preds
)
311 add_phi_arg (phi
, var
, e
, UNKNOWN_LOCATION
);
313 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
315 fprintf (dump_file
, ";; Created LCSSA PHI: ");
316 print_gimple_stmt (dump_file
, phi
, 0, dump_flags
);
320 /* Add exit phis for VAR that is used in LIVEIN.
321 Exits of the loops are stored in LOOP_EXITS. */
324 add_exit_phis_var (tree var
, bitmap use_blocks
, bitmap
*loop_exits
)
328 basic_block def_bb
= gimple_bb (SSA_NAME_DEF_STMT (var
));
329 bitmap live_exits
= BITMAP_ALLOC (&loop_renamer_obstack
);
331 gcc_checking_assert (! bitmap_bit_p (use_blocks
, def_bb
->index
));
333 compute_live_loop_exits (live_exits
, use_blocks
, loop_exits
, def_bb
);
335 EXECUTE_IF_SET_IN_BITMAP (live_exits
, 0, index
, bi
)
337 add_exit_phi (BASIC_BLOCK_FOR_FN (cfun
, index
), var
);
340 BITMAP_FREE (live_exits
);
343 /* Add exit phis for the names marked in NAMES_TO_RENAME.
344 Exits of the loops are stored in EXITS. Sets of blocks where the ssa
345 names are used are stored in USE_BLOCKS. */
348 add_exit_phis (bitmap names_to_rename
, bitmap
*use_blocks
, bitmap
*loop_exits
)
353 EXECUTE_IF_SET_IN_BITMAP (names_to_rename
, 0, i
, bi
)
355 add_exit_phis_var (ssa_name (i
), use_blocks
[i
], loop_exits
);
359 /* Fill the array of bitmaps LOOP_EXITS with all loop exit edge targets. */
362 get_loops_exits (bitmap
*loop_exits
)
368 FOR_EACH_LOOP (loop
, 0)
370 vec
<edge
> exit_edges
= get_loop_exit_edges (loop
);
371 loop_exits
[loop
->num
] = BITMAP_ALLOC (&loop_renamer_obstack
);
372 FOR_EACH_VEC_ELT (exit_edges
, j
, e
)
373 bitmap_set_bit (loop_exits
[loop
->num
], e
->dest
->index
);
374 exit_edges
.release ();
378 /* For USE in BB, if it is used outside of the loop it is defined in,
379 mark it for rewrite. Record basic block BB where it is used
380 to USE_BLOCKS. Record the ssa name index to NEED_PHIS bitmap. */
383 find_uses_to_rename_use (basic_block bb
, tree use
, bitmap
*use_blocks
,
388 struct loop
*def_loop
;
390 if (TREE_CODE (use
) != SSA_NAME
)
393 ver
= SSA_NAME_VERSION (use
);
394 def_bb
= gimple_bb (SSA_NAME_DEF_STMT (use
));
397 def_loop
= def_bb
->loop_father
;
399 /* If the definition is not inside a loop, it is not interesting. */
400 if (!loop_outer (def_loop
))
403 /* If the use is not outside of the loop it is defined in, it is not
405 if (flow_bb_inside_loop_p (def_loop
, bb
))
408 /* If we're seeing VER for the first time, we still have to allocate
409 a bitmap for its uses. */
410 if (bitmap_set_bit (need_phis
, ver
))
411 use_blocks
[ver
] = BITMAP_ALLOC (&loop_renamer_obstack
);
412 bitmap_set_bit (use_blocks
[ver
], bb
->index
);
415 /* For uses in STMT, mark names that are used outside of the loop they are
416 defined to rewrite. Record the set of blocks in that the ssa
417 names are defined to USE_BLOCKS and the ssa names themselves to
421 find_uses_to_rename_stmt (gimple stmt
, bitmap
*use_blocks
, bitmap need_phis
)
425 basic_block bb
= gimple_bb (stmt
);
427 if (is_gimple_debug (stmt
))
430 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, SSA_OP_USE
)
431 find_uses_to_rename_use (bb
, var
, use_blocks
, need_phis
);
434 /* Marks names that are used in BB and outside of the loop they are
435 defined in for rewrite. Records the set of blocks in that the ssa
436 names are defined to USE_BLOCKS. Record the SSA names that will
437 need exit PHIs in NEED_PHIS. */
440 find_uses_to_rename_bb (basic_block bb
, bitmap
*use_blocks
, bitmap need_phis
)
445 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
446 for (gphi_iterator bsi
= gsi_start_phis (e
->dest
); !gsi_end_p (bsi
);
449 gphi
*phi
= bsi
.phi ();
450 if (! virtual_operand_p (gimple_phi_result (phi
)))
451 find_uses_to_rename_use (bb
, PHI_ARG_DEF_FROM_EDGE (phi
, e
),
452 use_blocks
, need_phis
);
455 for (gimple_stmt_iterator bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
);
457 find_uses_to_rename_stmt (gsi_stmt (bsi
), use_blocks
, need_phis
);
460 /* Marks names that are used outside of the loop they are defined in
461 for rewrite. Records the set of blocks in that the ssa
462 names are defined to USE_BLOCKS. If CHANGED_BBS is not NULL,
463 scan only blocks in this set. */
466 find_uses_to_rename (bitmap changed_bbs
, bitmap
*use_blocks
, bitmap need_phis
)
473 EXECUTE_IF_SET_IN_BITMAP (changed_bbs
, 0, index
, bi
)
474 find_uses_to_rename_bb (BASIC_BLOCK_FOR_FN (cfun
, index
), use_blocks
, need_phis
);
476 FOR_EACH_BB_FN (bb
, cfun
)
477 find_uses_to_rename_bb (bb
, use_blocks
, need_phis
);
480 /* Rewrites the program into a loop closed ssa form -- i.e. inserts extra
481 phi nodes to ensure that no variable is used outside the loop it is
484 This strengthening of the basic ssa form has several advantages:
486 1) Updating it during unrolling/peeling/versioning is trivial, since
487 we do not need to care about the uses outside of the loop.
488 The same applies to virtual operands which are also rewritten into
489 loop closed SSA form. Note that virtual operands are always live
491 2) The behavior of all uses of an induction variable is the same.
492 Without this, you need to distinguish the case when the variable
493 is used outside of the loop it is defined in, for example
495 for (i = 0; i < 100; i++)
497 for (j = 0; j < 100; j++)
505 Looking from the outer loop with the normal SSA form, the first use of k
506 is not well-behaved, while the second one is an induction variable with
509 If CHANGED_BBS is not NULL, we look for uses outside loops only in
510 the basic blocks in this set.
512 UPDATE_FLAG is used in the call to update_ssa. See
513 TODO_update_ssa* for documentation. */
516 rewrite_into_loop_closed_ssa (bitmap changed_bbs
, unsigned update_flag
)
519 bitmap names_to_rename
;
521 loops_state_set (LOOP_CLOSED_SSA
);
522 if (number_of_loops (cfun
) <= 1)
525 /* If the pass has caused the SSA form to be out-of-date, update it
527 update_ssa (update_flag
);
529 bitmap_obstack_initialize (&loop_renamer_obstack
);
531 names_to_rename
= BITMAP_ALLOC (&loop_renamer_obstack
);
533 /* Uses of names to rename. We don't have to initialize this array,
534 because we know that we will only have entries for the SSA names
535 in NAMES_TO_RENAME. */
536 use_blocks
= XNEWVEC (bitmap
, num_ssa_names
);
538 /* Find the uses outside loops. */
539 find_uses_to_rename (changed_bbs
, use_blocks
, names_to_rename
);
541 if (!bitmap_empty_p (names_to_rename
))
543 /* An array of bitmaps where LOOP_EXITS[I] is the set of basic blocks
544 that are the destination of an edge exiting loop number I. */
545 bitmap
*loop_exits
= XNEWVEC (bitmap
, number_of_loops (cfun
));
546 get_loops_exits (loop_exits
);
548 /* Add the PHI nodes on exits of the loops for the names we need to
550 add_exit_phis (names_to_rename
, use_blocks
, loop_exits
);
554 /* Fix up all the names found to be used outside their original
556 update_ssa (TODO_update_ssa
);
559 bitmap_obstack_release (&loop_renamer_obstack
);
563 /* Check invariants of the loop closed ssa form for the USE in BB. */
566 check_loop_closed_ssa_use (basic_block bb
, tree use
)
571 if (TREE_CODE (use
) != SSA_NAME
|| virtual_operand_p (use
))
574 def
= SSA_NAME_DEF_STMT (use
);
575 def_bb
= gimple_bb (def
);
577 || flow_bb_inside_loop_p (def_bb
->loop_father
, bb
));
580 /* Checks invariants of loop closed ssa form in statement STMT in BB. */
583 check_loop_closed_ssa_stmt (basic_block bb
, gimple stmt
)
588 if (is_gimple_debug (stmt
))
591 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, SSA_OP_USE
)
592 check_loop_closed_ssa_use (bb
, var
);
595 /* Checks that invariants of the loop closed ssa form are preserved.
596 Call verify_ssa when VERIFY_SSA_P is true. */
599 verify_loop_closed_ssa (bool verify_ssa_p
)
605 if (number_of_loops (cfun
) <= 1)
609 verify_ssa (false, true);
611 timevar_push (TV_VERIFY_LOOP_CLOSED
);
613 FOR_EACH_BB_FN (bb
, cfun
)
615 for (gphi_iterator bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
);
618 gphi
*phi
= bsi
.phi ();
619 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
620 check_loop_closed_ssa_use (e
->src
,
621 PHI_ARG_DEF_FROM_EDGE (phi
, e
));
624 for (gimple_stmt_iterator bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
);
626 check_loop_closed_ssa_stmt (bb
, gsi_stmt (bsi
));
629 timevar_pop (TV_VERIFY_LOOP_CLOSED
);
632 /* Split loop exit edge EXIT. The things are a bit complicated by a need to
633 preserve the loop closed ssa form. The newly created block is returned. */
636 split_loop_exit_edge (edge exit
)
638 basic_block dest
= exit
->dest
;
639 basic_block bb
= split_edge (exit
);
644 source_location locus
;
646 for (psi
= gsi_start_phis (dest
); !gsi_end_p (psi
); gsi_next (&psi
))
649 op_p
= PHI_ARG_DEF_PTR_FROM_EDGE (phi
, single_succ_edge (bb
));
650 locus
= gimple_phi_arg_location_from_edge (phi
, single_succ_edge (bb
));
652 name
= USE_FROM_PTR (op_p
);
654 /* If the argument of the PHI node is a constant, we do not need
655 to keep it inside loop. */
656 if (TREE_CODE (name
) != SSA_NAME
)
659 /* Otherwise create an auxiliary phi node that will copy the value
660 of the SSA name out of the loop. */
661 new_name
= duplicate_ssa_name (name
, NULL
);
662 new_phi
= create_phi_node (new_name
, bb
);
663 add_phi_arg (new_phi
, name
, exit
, locus
);
664 SET_USE (op_p
, new_name
);
670 /* Returns the basic block in that statements should be emitted for induction
671 variables incremented at the end of the LOOP. */
674 ip_end_pos (struct loop
*loop
)
679 /* Returns the basic block in that statements should be emitted for induction
680 variables incremented just before exit condition of a LOOP. */
683 ip_normal_pos (struct loop
*loop
)
689 if (!single_pred_p (loop
->latch
))
692 bb
= single_pred (loop
->latch
);
693 last
= last_stmt (bb
);
695 || gimple_code (last
) != GIMPLE_COND
)
698 exit
= EDGE_SUCC (bb
, 0);
699 if (exit
->dest
== loop
->latch
)
700 exit
= EDGE_SUCC (bb
, 1);
702 if (flow_bb_inside_loop_p (loop
, exit
->dest
))
708 /* Stores the standard position for induction variable increment in LOOP
709 (just before the exit condition if it is available and latch block is empty,
710 end of the latch block otherwise) to BSI. INSERT_AFTER is set to true if
711 the increment should be inserted after *BSI. */
714 standard_iv_increment_position (struct loop
*loop
, gimple_stmt_iterator
*bsi
,
717 basic_block bb
= ip_normal_pos (loop
), latch
= ip_end_pos (loop
);
718 gimple last
= last_stmt (latch
);
721 || (last
&& gimple_code (last
) != GIMPLE_LABEL
))
723 *bsi
= gsi_last_bb (latch
);
724 *insert_after
= true;
728 *bsi
= gsi_last_bb (bb
);
729 *insert_after
= false;
733 /* Copies phi node arguments for duplicated blocks. The index of the first
734 duplicated block is FIRST_NEW_BLOCK. */
737 copy_phi_node_args (unsigned first_new_block
)
741 for (i
= first_new_block
; i
< (unsigned) last_basic_block_for_fn (cfun
); i
++)
742 BASIC_BLOCK_FOR_FN (cfun
, i
)->flags
|= BB_DUPLICATED
;
744 for (i
= first_new_block
; i
< (unsigned) last_basic_block_for_fn (cfun
); i
++)
745 add_phi_args_after_copy_bb (BASIC_BLOCK_FOR_FN (cfun
, i
));
747 for (i
= first_new_block
; i
< (unsigned) last_basic_block_for_fn (cfun
); i
++)
748 BASIC_BLOCK_FOR_FN (cfun
, i
)->flags
&= ~BB_DUPLICATED
;
752 /* The same as cfgloopmanip.c:duplicate_loop_to_header_edge, but also
753 updates the PHI nodes at start of the copied region. In order to
754 achieve this, only loops whose exits all lead to the same location
757 Notice that we do not completely update the SSA web after
758 duplication. The caller is responsible for calling update_ssa
759 after the loop has been duplicated. */
762 gimple_duplicate_loop_to_header_edge (struct loop
*loop
, edge e
,
763 unsigned int ndupl
, sbitmap wont_exit
,
764 edge orig
, vec
<edge
> *to_remove
,
767 unsigned first_new_block
;
769 if (!loops_state_satisfies_p (LOOPS_HAVE_SIMPLE_LATCHES
))
771 if (!loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS
))
774 first_new_block
= last_basic_block_for_fn (cfun
);
775 if (!duplicate_loop_to_header_edge (loop
, e
, ndupl
, wont_exit
,
776 orig
, to_remove
, flags
))
779 /* Readd the removed phi args for e. */
780 flush_pending_stmts (e
);
782 /* Copy the phi node arguments. */
783 copy_phi_node_args (first_new_block
);
790 /* Returns true if we can unroll LOOP FACTOR times. Number
791 of iterations of the loop is returned in NITER. */
794 can_unroll_loop_p (struct loop
*loop
, unsigned factor
,
795 struct tree_niter_desc
*niter
)
799 /* Check whether unrolling is possible. We only want to unroll loops
800 for that we are able to determine number of iterations. We also
801 want to split the extra iterations of the loop from its end,
802 therefore we require that the loop has precisely one
805 exit
= single_dom_exit (loop
);
809 if (!number_of_iterations_exit (loop
, exit
, niter
, false)
810 || niter
->cmp
== ERROR_MARK
811 /* Scalar evolutions analysis might have copy propagated
812 the abnormal ssa names into these expressions, hence
813 emitting the computations based on them during loop
814 unrolling might create overlapping life ranges for
815 them, and failures in out-of-ssa. */
816 || contains_abnormal_ssa_name_p (niter
->may_be_zero
)
817 || contains_abnormal_ssa_name_p (niter
->control
.base
)
818 || contains_abnormal_ssa_name_p (niter
->control
.step
)
819 || contains_abnormal_ssa_name_p (niter
->bound
))
822 /* And of course, we must be able to duplicate the loop. */
823 if (!can_duplicate_loop_p (loop
))
826 /* The final loop should be small enough. */
827 if (tree_num_loop_insns (loop
, &eni_size_weights
) * factor
828 > (unsigned) PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS
))
834 /* Determines the conditions that control execution of LOOP unrolled FACTOR
835 times. DESC is number of iterations of LOOP. ENTER_COND is set to
836 condition that must be true if the main loop can be entered.
837 EXIT_BASE, EXIT_STEP, EXIT_CMP and EXIT_BOUND are set to values describing
838 how the exit from the unrolled loop should be controlled. */
841 determine_exit_conditions (struct loop
*loop
, struct tree_niter_desc
*desc
,
842 unsigned factor
, tree
*enter_cond
,
843 tree
*exit_base
, tree
*exit_step
,
844 enum tree_code
*exit_cmp
, tree
*exit_bound
)
847 tree base
= desc
->control
.base
;
848 tree step
= desc
->control
.step
;
849 tree bound
= desc
->bound
;
850 tree type
= TREE_TYPE (step
);
852 tree min
= lower_bound_in_type (type
, type
);
853 tree max
= upper_bound_in_type (type
, type
);
854 enum tree_code cmp
= desc
->cmp
;
855 tree cond
= boolean_true_node
, assum
;
857 /* For pointers, do the arithmetics in the type of step. */
858 base
= fold_convert (type
, base
);
859 bound
= fold_convert (type
, bound
);
861 *enter_cond
= boolean_false_node
;
862 *exit_base
= NULL_TREE
;
863 *exit_step
= NULL_TREE
;
864 *exit_cmp
= ERROR_MARK
;
865 *exit_bound
= NULL_TREE
;
866 gcc_assert (cmp
!= ERROR_MARK
);
868 /* We only need to be correct when we answer question
869 "Do at least FACTOR more iterations remain?" in the unrolled loop.
870 Thus, transforming BASE + STEP * i <> BOUND to
871 BASE + STEP * i < BOUND is ok. */
874 if (tree_int_cst_sign_bit (step
))
879 else if (cmp
== LT_EXPR
)
881 gcc_assert (!tree_int_cst_sign_bit (step
));
883 else if (cmp
== GT_EXPR
)
885 gcc_assert (tree_int_cst_sign_bit (step
));
890 /* The main body of the loop may be entered iff:
892 1) desc->may_be_zero is false.
893 2) it is possible to check that there are at least FACTOR iterations
894 of the loop, i.e., BOUND - step * FACTOR does not overflow.
895 3) # of iterations is at least FACTOR */
897 if (!integer_zerop (desc
->may_be_zero
))
898 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
899 invert_truthvalue (desc
->may_be_zero
),
902 bigstep
= fold_build2 (MULT_EXPR
, type
, step
,
903 build_int_cst_type (type
, factor
));
904 delta
= fold_build2 (MINUS_EXPR
, type
, bigstep
, step
);
906 assum
= fold_build2 (GE_EXPR
, boolean_type_node
,
908 fold_build2 (PLUS_EXPR
, type
, min
, delta
));
910 assum
= fold_build2 (LE_EXPR
, boolean_type_node
,
912 fold_build2 (PLUS_EXPR
, type
, max
, delta
));
913 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
, assum
, cond
);
915 bound
= fold_build2 (MINUS_EXPR
, type
, bound
, delta
);
916 assum
= fold_build2 (cmp
, boolean_type_node
, base
, bound
);
917 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
, assum
, cond
);
919 cond
= force_gimple_operand (unshare_expr (cond
), &stmts
, false, NULL_TREE
);
921 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
922 /* cond now may be a gimple comparison, which would be OK, but also any
923 other gimple rhs (say a && b). In this case we need to force it to
925 if (!is_gimple_condexpr (cond
))
927 cond
= force_gimple_operand (cond
, &stmts
, true, NULL_TREE
);
929 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
933 base
= force_gimple_operand (unshare_expr (base
), &stmts
, true, NULL_TREE
);
935 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
936 bound
= force_gimple_operand (unshare_expr (bound
), &stmts
, true, NULL_TREE
);
938 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
941 *exit_step
= bigstep
;
946 /* Scales the frequencies of all basic blocks in LOOP that are strictly
947 dominated by BB by NUM/DEN. */
950 scale_dominated_blocks_in_loop (struct loop
*loop
, basic_block bb
,
958 for (son
= first_dom_son (CDI_DOMINATORS
, bb
);
960 son
= next_dom_son (CDI_DOMINATORS
, son
))
962 if (!flow_bb_inside_loop_p (loop
, son
))
964 scale_bbs_frequencies_int (&son
, 1, num
, den
);
965 scale_dominated_blocks_in_loop (loop
, son
, num
, den
);
969 /* Unroll LOOP FACTOR times. DESC describes number of iterations of LOOP.
970 EXIT is the exit of the loop to that DESC corresponds.
972 If N is number of iterations of the loop and MAY_BE_ZERO is the condition
973 under that loop exits in the first iteration even if N != 0,
977 x = phi (init, next);
985 becomes (with possibly the exit conditions formulated a bit differently,
986 avoiding the need to create a new iv):
988 if (MAY_BE_ZERO || N < FACTOR)
993 x = phi (init, next);
1004 } while (N >= FACTOR);
1007 init' = phi (init, x);
1011 x = phi (init', next);
1019 Before the loop is unrolled, TRANSFORM is called for it (only for the
1020 unrolled loop, but not for its versioned copy). DATA is passed to
1023 /* Probability in % that the unrolled loop is entered. Just a guess. */
1024 #define PROB_UNROLLED_LOOP_ENTERED 90
1027 tree_transform_and_unroll_loop (struct loop
*loop
, unsigned factor
,
1028 edge exit
, struct tree_niter_desc
*desc
,
1029 transform_callback transform
,
1033 tree ctr_before
, ctr_after
;
1034 tree enter_main_cond
, exit_base
, exit_step
, exit_bound
;
1035 enum tree_code exit_cmp
;
1036 gphi
*phi_old_loop
, *phi_new_loop
, *phi_rest
;
1037 gphi_iterator psi_old_loop
, psi_new_loop
;
1038 tree init
, next
, new_init
;
1039 struct loop
*new_loop
;
1040 basic_block rest
, exit_bb
;
1041 edge old_entry
, new_entry
, old_latch
, precond_edge
, new_exit
;
1042 edge new_nonexit
, e
;
1043 gimple_stmt_iterator bsi
;
1046 unsigned est_niter
, prob_entry
, scale_unrolled
, scale_rest
, freq_e
, freq_h
;
1047 unsigned new_est_niter
, i
, prob
;
1048 unsigned irr
= loop_preheader_edge (loop
)->flags
& EDGE_IRREDUCIBLE_LOOP
;
1050 auto_vec
<edge
> to_remove
;
1052 est_niter
= expected_loop_iterations (loop
);
1053 determine_exit_conditions (loop
, desc
, factor
,
1054 &enter_main_cond
, &exit_base
, &exit_step
,
1055 &exit_cmp
, &exit_bound
);
1057 /* Let us assume that the unrolled loop is quite likely to be entered. */
1058 if (integer_nonzerop (enter_main_cond
))
1059 prob_entry
= REG_BR_PROB_BASE
;
1061 prob_entry
= PROB_UNROLLED_LOOP_ENTERED
* REG_BR_PROB_BASE
/ 100;
1063 /* The values for scales should keep profile consistent, and somewhat close
1066 TODO: The current value of SCALE_REST makes it appear that the loop that
1067 is created by splitting the remaining iterations of the unrolled loop is
1068 executed the same number of times as the original loop, and with the same
1069 frequencies, which is obviously wrong. This does not appear to cause
1070 problems, so we do not bother with fixing it for now. To make the profile
1071 correct, we would need to change the probability of the exit edge of the
1072 loop, and recompute the distribution of frequencies in its body because
1073 of this change (scale the frequencies of blocks before and after the exit
1074 by appropriate factors). */
1075 scale_unrolled
= prob_entry
;
1076 scale_rest
= REG_BR_PROB_BASE
;
1078 new_loop
= loop_version (loop
, enter_main_cond
, NULL
,
1079 prob_entry
, scale_unrolled
, scale_rest
, true);
1080 gcc_assert (new_loop
!= NULL
);
1081 update_ssa (TODO_update_ssa
);
1083 /* Determine the probability of the exit edge of the unrolled loop. */
1084 new_est_niter
= est_niter
/ factor
;
1086 /* Without profile feedback, loops for that we do not know a better estimate
1087 are assumed to roll 10 times. When we unroll such loop, it appears to
1088 roll too little, and it may even seem to be cold. To avoid this, we
1089 ensure that the created loop appears to roll at least 5 times (but at
1090 most as many times as before unrolling). */
1091 if (new_est_niter
< 5)
1094 new_est_niter
= est_niter
;
1099 /* Prepare the cfg and update the phi nodes. Move the loop exit to the
1100 loop latch (and make its condition dummy, for the moment). */
1101 rest
= loop_preheader_edge (new_loop
)->src
;
1102 precond_edge
= single_pred_edge (rest
);
1103 split_edge (loop_latch_edge (loop
));
1104 exit_bb
= single_pred (loop
->latch
);
1106 /* Since the exit edge will be removed, the frequency of all the blocks
1107 in the loop that are dominated by it must be scaled by
1108 1 / (1 - exit->probability). */
1109 scale_dominated_blocks_in_loop (loop
, exit
->src
,
1111 REG_BR_PROB_BASE
- exit
->probability
);
1113 bsi
= gsi_last_bb (exit_bb
);
1114 exit_if
= gimple_build_cond (EQ_EXPR
, integer_zero_node
,
1116 NULL_TREE
, NULL_TREE
);
1118 gsi_insert_after (&bsi
, exit_if
, GSI_NEW_STMT
);
1119 new_exit
= make_edge (exit_bb
, rest
, EDGE_FALSE_VALUE
| irr
);
1120 rescan_loop_exit (new_exit
, true, false);
1122 /* Set the probability of new exit to the same of the old one. Fix
1123 the frequency of the latch block, by scaling it back by
1124 1 - exit->probability. */
1125 new_exit
->count
= exit
->count
;
1126 new_exit
->probability
= exit
->probability
;
1127 new_nonexit
= single_pred_edge (loop
->latch
);
1128 new_nonexit
->probability
= REG_BR_PROB_BASE
- exit
->probability
;
1129 new_nonexit
->flags
= EDGE_TRUE_VALUE
;
1130 new_nonexit
->count
-= exit
->count
;
1131 if (new_nonexit
->count
< 0)
1132 new_nonexit
->count
= 0;
1133 scale_bbs_frequencies_int (&loop
->latch
, 1, new_nonexit
->probability
,
1136 old_entry
= loop_preheader_edge (loop
);
1137 new_entry
= loop_preheader_edge (new_loop
);
1138 old_latch
= loop_latch_edge (loop
);
1139 for (psi_old_loop
= gsi_start_phis (loop
->header
),
1140 psi_new_loop
= gsi_start_phis (new_loop
->header
);
1141 !gsi_end_p (psi_old_loop
);
1142 gsi_next (&psi_old_loop
), gsi_next (&psi_new_loop
))
1144 phi_old_loop
= psi_old_loop
.phi ();
1145 phi_new_loop
= psi_new_loop
.phi ();
1147 init
= PHI_ARG_DEF_FROM_EDGE (phi_old_loop
, old_entry
);
1148 op
= PHI_ARG_DEF_PTR_FROM_EDGE (phi_new_loop
, new_entry
);
1149 gcc_assert (operand_equal_for_phi_arg_p (init
, USE_FROM_PTR (op
)));
1150 next
= PHI_ARG_DEF_FROM_EDGE (phi_old_loop
, old_latch
);
1152 /* Prefer using original variable as a base for the new ssa name.
1153 This is necessary for virtual ops, and useful in order to avoid
1154 losing debug info for real ops. */
1155 if (TREE_CODE (next
) == SSA_NAME
1156 && useless_type_conversion_p (TREE_TYPE (next
),
1158 new_init
= copy_ssa_name (next
);
1159 else if (TREE_CODE (init
) == SSA_NAME
1160 && useless_type_conversion_p (TREE_TYPE (init
),
1162 new_init
= copy_ssa_name (init
);
1163 else if (useless_type_conversion_p (TREE_TYPE (next
), TREE_TYPE (init
)))
1164 new_init
= make_temp_ssa_name (TREE_TYPE (next
), NULL
, "unrinittmp");
1166 new_init
= make_temp_ssa_name (TREE_TYPE (init
), NULL
, "unrinittmp");
1168 phi_rest
= create_phi_node (new_init
, rest
);
1170 add_phi_arg (phi_rest
, init
, precond_edge
, UNKNOWN_LOCATION
);
1171 add_phi_arg (phi_rest
, next
, new_exit
, UNKNOWN_LOCATION
);
1172 SET_USE (op
, new_init
);
1177 /* Transform the loop. */
1179 (*transform
) (loop
, data
);
1181 /* Unroll the loop and remove the exits in all iterations except for the
1183 wont_exit
= sbitmap_alloc (factor
);
1184 bitmap_ones (wont_exit
);
1185 bitmap_clear_bit (wont_exit
, factor
- 1);
1187 ok
= gimple_duplicate_loop_to_header_edge
1188 (loop
, loop_latch_edge (loop
), factor
- 1,
1189 wont_exit
, new_exit
, &to_remove
, DLTHE_FLAG_UPDATE_FREQ
);
1193 FOR_EACH_VEC_ELT (to_remove
, i
, e
)
1195 ok
= remove_path (e
);
1198 update_ssa (TODO_update_ssa
);
1200 /* Ensure that the frequencies in the loop match the new estimated
1201 number of iterations, and change the probability of the new
1203 freq_h
= loop
->header
->frequency
;
1204 freq_e
= EDGE_FREQUENCY (loop_preheader_edge (loop
));
1206 scale_loop_frequencies (loop
, freq_e
* (new_est_niter
+ 1), freq_h
);
1208 exit_bb
= single_pred (loop
->latch
);
1209 new_exit
= find_edge (exit_bb
, rest
);
1210 new_exit
->count
= loop_preheader_edge (loop
)->count
;
1211 new_exit
->probability
= REG_BR_PROB_BASE
/ (new_est_niter
+ 1);
1213 rest
->count
+= new_exit
->count
;
1214 rest
->frequency
+= EDGE_FREQUENCY (new_exit
);
1216 new_nonexit
= single_pred_edge (loop
->latch
);
1217 prob
= new_nonexit
->probability
;
1218 new_nonexit
->probability
= REG_BR_PROB_BASE
- new_exit
->probability
;
1219 new_nonexit
->count
= exit_bb
->count
- new_exit
->count
;
1220 if (new_nonexit
->count
< 0)
1221 new_nonexit
->count
= 0;
1223 scale_bbs_frequencies_int (&loop
->latch
, 1, new_nonexit
->probability
,
1226 /* Finally create the new counter for number of iterations and add the new
1227 exit instruction. */
1228 bsi
= gsi_last_nondebug_bb (exit_bb
);
1229 exit_if
= as_a
<gcond
*> (gsi_stmt (bsi
));
1230 create_iv (exit_base
, exit_step
, NULL_TREE
, loop
,
1231 &bsi
, false, &ctr_before
, &ctr_after
);
1232 gimple_cond_set_code (exit_if
, exit_cmp
);
1233 gimple_cond_set_lhs (exit_if
, ctr_after
);
1234 gimple_cond_set_rhs (exit_if
, exit_bound
);
1235 update_stmt (exit_if
);
1237 #ifdef ENABLE_CHECKING
1238 verify_flow_info ();
1239 verify_loop_structure ();
1240 verify_loop_closed_ssa (true);
1244 /* Wrapper over tree_transform_and_unroll_loop for case we do not
1245 want to transform the loop before unrolling. The meaning
1246 of the arguments is the same as for tree_transform_and_unroll_loop. */
1249 tree_unroll_loop (struct loop
*loop
, unsigned factor
,
1250 edge exit
, struct tree_niter_desc
*desc
)
1252 tree_transform_and_unroll_loop (loop
, factor
, exit
, desc
,
1256 /* Rewrite the phi node at position PSI in function of the main
1257 induction variable MAIN_IV and insert the generated code at GSI. */
1260 rewrite_phi_with_iv (loop_p loop
,
1262 gimple_stmt_iterator
*gsi
,
1267 gphi
*phi
= psi
->phi ();
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
);
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 variable
1323 (with base 0 and step 1), whose final value is compared with *NIT. When the
1324 IV type precision has to be larger than *NIT type precision, *NIT is
1325 converted to the larger type, the conversion code is inserted before the
1326 loop, and *NIT is updated to the new definition. When BUMP_IN_LATCH is true,
1327 the induction variable is incremented in the loop latch, otherwise it is
1328 incremented in the loop header. Return the induction variable that was
1332 canonicalize_loop_ivs (struct loop
*loop
, tree
*nit
, bool bump_in_latch
)
1334 unsigned precision
= TYPE_PRECISION (TREE_TYPE (*nit
));
1335 unsigned original_precision
= precision
;
1336 tree type
, var_before
;
1337 gimple_stmt_iterator gsi
;
1340 edge exit
= single_dom_exit (loop
);
1343 bool unsigned_p
= false;
1345 for (psi
= gsi_start_phis (loop
->header
);
1346 !gsi_end_p (psi
); gsi_next (&psi
))
1348 gphi
*phi
= psi
.phi ();
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
= as_a
<gcond
*> (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
);