1 /* High-level loop manipulation functions.
2 Copyright (C) 2004-2013 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"
26 #include "basic-block.h"
29 #include "gimple-iterator.h"
30 #include "gimplify-me.h"
31 #include "gimple-ssa.h"
33 #include "tree-phinodes.h"
34 #include "ssa-iterators.h"
35 #include "stringpool.h"
36 #include "tree-ssanames.h"
37 #include "tree-ssa-loop-ivopts.h"
38 #include "tree-ssa-loop-manip.h"
39 #include "tree-ssa-loop-niter.h"
40 #include "tree-ssa-loop.h"
41 #include "tree-into-ssa.h"
44 #include "gimple-pretty-print.h"
46 #include "tree-pass.h" /* ??? for TODO_update_ssa but this isn't a pass. */
47 #include "tree-scalar-evolution.h"
49 #include "tree-inline.h"
50 #include "langhooks.h"
52 /* All bitmaps for rewriting into loop-closed SSA go on this obstack,
53 so that we can free them all at once. */
54 static bitmap_obstack loop_renamer_obstack
;
56 /* Creates an induction variable with value BASE + STEP * iteration in LOOP.
57 It is expected that neither BASE nor STEP are shared with other expressions
58 (unless the sharing rules allow this). Use VAR as a base var_decl for it
59 (if NULL, a new temporary will be created). The increment will occur at
60 INCR_POS (after it if AFTER is true, before it otherwise). INCR_POS and
61 AFTER can be computed using standard_iv_increment_position. The ssa versions
62 of the variable before and after increment will be stored in VAR_BEFORE and
63 VAR_AFTER (unless they are NULL). */
66 create_iv (tree base
, tree step
, tree var
, struct loop
*loop
,
67 gimple_stmt_iterator
*incr_pos
, bool after
,
68 tree
*var_before
, tree
*var_after
)
74 enum tree_code incr_op
= PLUS_EXPR
;
75 edge pe
= loop_preheader_edge (loop
);
79 vb
= make_ssa_name (var
, NULL
);
80 va
= make_ssa_name (var
, NULL
);
84 vb
= make_temp_ssa_name (TREE_TYPE (base
), NULL
, "ivtmp");
85 va
= make_temp_ssa_name (TREE_TYPE (base
), NULL
, "ivtmp");
92 /* For easier readability of the created code, produce MINUS_EXPRs
94 if (TREE_CODE (step
) == INTEGER_CST
)
96 if (TYPE_UNSIGNED (TREE_TYPE (step
)))
98 step1
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
99 if (tree_int_cst_lt (step1
, step
))
101 incr_op
= MINUS_EXPR
;
109 if (!tree_expr_nonnegative_warnv_p (step
, &ovf
)
110 && may_negate_without_overflow_p (step
))
112 incr_op
= MINUS_EXPR
;
113 step
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
117 if (POINTER_TYPE_P (TREE_TYPE (base
)))
119 if (TREE_CODE (base
) == ADDR_EXPR
)
120 mark_addressable (TREE_OPERAND (base
, 0));
121 step
= convert_to_ptrofftype (step
);
122 if (incr_op
== MINUS_EXPR
)
123 step
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
124 incr_op
= POINTER_PLUS_EXPR
;
126 /* Gimplify the step if necessary. We put the computations in front of the
127 loop (i.e. the step should be loop invariant). */
128 step
= force_gimple_operand (step
, &stmts
, true, NULL_TREE
);
130 gsi_insert_seq_on_edge_immediate (pe
, stmts
);
132 stmt
= gimple_build_assign_with_ops (incr_op
, va
, vb
, step
);
134 gsi_insert_after (incr_pos
, stmt
, GSI_NEW_STMT
);
136 gsi_insert_before (incr_pos
, stmt
, GSI_NEW_STMT
);
138 initial
= force_gimple_operand (base
, &stmts
, true, var
);
140 gsi_insert_seq_on_edge_immediate (pe
, stmts
);
142 stmt
= create_phi_node (vb
, loop
->header
);
143 add_phi_arg (stmt
, initial
, loop_preheader_edge (loop
), UNKNOWN_LOCATION
);
144 add_phi_arg (stmt
, va
, loop_latch_edge (loop
), UNKNOWN_LOCATION
);
147 /* Return the innermost superloop LOOP of USE_LOOP that is a superloop of
148 both DEF_LOOP and USE_LOOP. */
150 static inline struct loop
*
151 find_sibling_superloop (struct loop
*use_loop
, struct loop
*def_loop
)
153 unsigned ud
= loop_depth (use_loop
);
154 unsigned dd
= loop_depth (def_loop
);
155 gcc_assert (ud
> 0 && dd
> 0);
157 use_loop
= superloop_at_depth (use_loop
, dd
);
159 def_loop
= superloop_at_depth (def_loop
, ud
);
160 while (loop_outer (use_loop
) != loop_outer (def_loop
))
162 use_loop
= loop_outer (use_loop
);
163 def_loop
= loop_outer (def_loop
);
164 gcc_assert (use_loop
&& def_loop
);
169 /* DEF_BB is a basic block containing a DEF that needs rewriting into
170 loop-closed SSA form. USE_BLOCKS is the set of basic blocks containing
171 uses of DEF that "escape" from the loop containing DEF_BB (i.e. blocks in
172 USE_BLOCKS are dominated by DEF_BB but not in the loop father of DEF_B).
173 ALL_EXITS[I] is the set of all basic blocks that exit loop I.
175 Compute the subset of LOOP_EXITS that exit the loop containing DEF_BB
176 or one of its loop fathers, in which DEF is live. This set is returned
177 in the bitmap LIVE_EXITS.
179 Instead of computing the complete livein set of the def, we use the loop
180 nesting tree as a form of poor man's structure analysis. This greatly
181 speeds up the analysis, which is important because this function may be
182 called on all SSA names that need rewriting, one at a time. */
185 compute_live_loop_exits (bitmap live_exits
, bitmap use_blocks
,
186 bitmap
*loop_exits
, basic_block def_bb
)
190 vec
<basic_block
> worklist
;
191 struct loop
*def_loop
= def_bb
->loop_father
;
192 unsigned def_loop_depth
= loop_depth (def_loop
);
193 bitmap def_loop_exits
;
195 /* Normally the work list size is bounded by the number of basic
196 blocks in the largest loop. We don't know this number, but we
197 can be fairly sure that it will be relatively small. */
198 worklist
.create (MAX (8, n_basic_blocks_for_fn (cfun
) / 128));
200 EXECUTE_IF_SET_IN_BITMAP (use_blocks
, 0, i
, bi
)
202 basic_block use_bb
= BASIC_BLOCK (i
);
203 struct loop
*use_loop
= use_bb
->loop_father
;
204 gcc_checking_assert (def_loop
!= use_loop
205 && ! flow_loop_nested_p (def_loop
, use_loop
));
206 if (! flow_loop_nested_p (use_loop
, def_loop
))
207 use_bb
= find_sibling_superloop (use_loop
, def_loop
)->header
;
208 if (bitmap_set_bit (live_exits
, use_bb
->index
))
209 worklist
.safe_push (use_bb
);
212 /* Iterate until the worklist is empty. */
213 while (! worklist
.is_empty ())
218 /* Pull a block off the worklist. */
219 basic_block bb
= worklist
.pop ();
221 /* Make sure we have at least enough room in the work list
222 for all predecessors of this block. */
223 worklist
.reserve (EDGE_COUNT (bb
->preds
));
225 /* For each predecessor block. */
226 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
228 basic_block pred
= e
->src
;
229 struct loop
*pred_loop
= pred
->loop_father
;
230 unsigned pred_loop_depth
= loop_depth (pred_loop
);
233 /* We should have met DEF_BB along the way. */
234 gcc_assert (pred
!= ENTRY_BLOCK_PTR
);
236 if (pred_loop_depth
>= def_loop_depth
)
238 if (pred_loop_depth
> def_loop_depth
)
239 pred_loop
= superloop_at_depth (pred_loop
, def_loop_depth
);
240 /* If we've reached DEF_LOOP, our train ends here. */
241 if (pred_loop
== def_loop
)
244 else if (! flow_loop_nested_p (pred_loop
, def_loop
))
245 pred
= find_sibling_superloop (pred_loop
, def_loop
)->header
;
247 /* Add PRED to the LIVEIN set. PRED_VISITED is true if
248 we had already added PRED to LIVEIN before. */
249 pred_visited
= !bitmap_set_bit (live_exits
, pred
->index
);
251 /* If we have visited PRED before, don't add it to the worklist.
252 If BB dominates PRED, then we're probably looking at a loop.
253 We're only interested in looking up in the dominance tree
254 because DEF_BB dominates all the uses. */
255 if (pred_visited
|| dominated_by_p (CDI_DOMINATORS
, pred
, bb
))
258 worklist
.quick_push (pred
);
263 def_loop_exits
= BITMAP_ALLOC (&loop_renamer_obstack
);
264 for (struct loop
*loop
= def_loop
;
265 loop
!= current_loops
->tree_root
;
266 loop
= loop_outer (loop
))
267 bitmap_ior_into (def_loop_exits
, loop_exits
[loop
->num
]);
268 bitmap_and_into (live_exits
, def_loop_exits
);
269 BITMAP_FREE (def_loop_exits
);
272 /* Add a loop-closing PHI for VAR in basic block EXIT. */
275 add_exit_phi (basic_block exit
, tree var
)
281 #ifdef ENABLE_CHECKING
282 /* Check that at least one of the edges entering the EXIT block exits
283 the loop, or a superloop of that loop, that VAR is defined in. */
284 gimple def_stmt
= SSA_NAME_DEF_STMT (var
);
285 basic_block def_bb
= gimple_bb (def_stmt
);
286 FOR_EACH_EDGE (e
, ei
, exit
->preds
)
288 struct loop
*aloop
= find_common_loop (def_bb
->loop_father
,
289 e
->src
->loop_father
);
290 if (!flow_bb_inside_loop_p (aloop
, e
->dest
))
294 gcc_checking_assert (e
);
297 phi
= create_phi_node (NULL_TREE
, exit
);
298 create_new_def_for (var
, phi
, gimple_phi_result_ptr (phi
));
299 FOR_EACH_EDGE (e
, ei
, exit
->preds
)
300 add_phi_arg (phi
, var
, e
, UNKNOWN_LOCATION
);
302 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
304 fprintf (dump_file
, ";; Created LCSSA PHI: ");
305 print_gimple_stmt (dump_file
, phi
, 0, dump_flags
);
309 /* Add exit phis for VAR that is used in LIVEIN.
310 Exits of the loops are stored in LOOP_EXITS. */
313 add_exit_phis_var (tree var
, bitmap use_blocks
, bitmap
*loop_exits
)
317 basic_block def_bb
= gimple_bb (SSA_NAME_DEF_STMT (var
));
318 bitmap live_exits
= BITMAP_ALLOC (&loop_renamer_obstack
);
320 gcc_checking_assert (! bitmap_bit_p (use_blocks
, def_bb
->index
));
322 compute_live_loop_exits (live_exits
, use_blocks
, loop_exits
, def_bb
);
324 EXECUTE_IF_SET_IN_BITMAP (live_exits
, 0, index
, bi
)
326 add_exit_phi (BASIC_BLOCK (index
), var
);
329 BITMAP_FREE (live_exits
);
332 /* Add exit phis for the names marked in NAMES_TO_RENAME.
333 Exits of the loops are stored in EXITS. Sets of blocks where the ssa
334 names are used are stored in USE_BLOCKS. */
337 add_exit_phis (bitmap names_to_rename
, bitmap
*use_blocks
, bitmap
*loop_exits
)
342 EXECUTE_IF_SET_IN_BITMAP (names_to_rename
, 0, i
, bi
)
344 add_exit_phis_var (ssa_name (i
), use_blocks
[i
], loop_exits
);
348 /* Fill the array of bitmaps LOOP_EXITS with all loop exit edge targets. */
351 get_loops_exits (bitmap
*loop_exits
)
358 FOR_EACH_LOOP (li
, loop
, 0)
360 vec
<edge
> exit_edges
= get_loop_exit_edges (loop
);
361 loop_exits
[loop
->num
] = BITMAP_ALLOC (&loop_renamer_obstack
);
362 FOR_EACH_VEC_ELT (exit_edges
, j
, e
)
363 bitmap_set_bit (loop_exits
[loop
->num
], e
->dest
->index
);
364 exit_edges
.release ();
368 /* For USE in BB, if it is used outside of the loop it is defined in,
369 mark it for rewrite. Record basic block BB where it is used
370 to USE_BLOCKS. Record the ssa name index to NEED_PHIS bitmap. */
373 find_uses_to_rename_use (basic_block bb
, tree use
, bitmap
*use_blocks
,
378 struct loop
*def_loop
;
380 if (TREE_CODE (use
) != SSA_NAME
)
383 ver
= SSA_NAME_VERSION (use
);
384 def_bb
= gimple_bb (SSA_NAME_DEF_STMT (use
));
387 def_loop
= def_bb
->loop_father
;
389 /* If the definition is not inside a loop, it is not interesting. */
390 if (!loop_outer (def_loop
))
393 /* If the use is not outside of the loop it is defined in, it is not
395 if (flow_bb_inside_loop_p (def_loop
, bb
))
398 /* If we're seeing VER for the first time, we still have to allocate
399 a bitmap for its uses. */
400 if (bitmap_set_bit (need_phis
, ver
))
401 use_blocks
[ver
] = BITMAP_ALLOC (&loop_renamer_obstack
);
402 bitmap_set_bit (use_blocks
[ver
], bb
->index
);
405 /* For uses in STMT, mark names that are used outside of the loop they are
406 defined to rewrite. Record the set of blocks in that the ssa
407 names are defined to USE_BLOCKS and the ssa names themselves to
411 find_uses_to_rename_stmt (gimple stmt
, bitmap
*use_blocks
, bitmap need_phis
)
415 basic_block bb
= gimple_bb (stmt
);
417 if (is_gimple_debug (stmt
))
420 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, SSA_OP_USE
)
421 find_uses_to_rename_use (bb
, var
, use_blocks
, need_phis
);
424 /* Marks names that are used in BB and outside of the loop they are
425 defined in for rewrite. Records the set of blocks in that the ssa
426 names are defined to USE_BLOCKS. Record the SSA names that will
427 need exit PHIs in NEED_PHIS. */
430 find_uses_to_rename_bb (basic_block bb
, bitmap
*use_blocks
, bitmap need_phis
)
432 gimple_stmt_iterator bsi
;
436 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
437 for (bsi
= gsi_start_phis (e
->dest
); !gsi_end_p (bsi
); gsi_next (&bsi
))
439 gimple phi
= gsi_stmt (bsi
);
440 if (! virtual_operand_p (gimple_phi_result (phi
)))
441 find_uses_to_rename_use (bb
, PHI_ARG_DEF_FROM_EDGE (phi
, e
),
442 use_blocks
, need_phis
);
445 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
446 find_uses_to_rename_stmt (gsi_stmt (bsi
), use_blocks
, need_phis
);
449 /* Marks names that are used outside of the loop they are defined in
450 for rewrite. Records the set of blocks in that the ssa
451 names are defined to USE_BLOCKS. If CHANGED_BBS is not NULL,
452 scan only blocks in this set. */
455 find_uses_to_rename (bitmap changed_bbs
, bitmap
*use_blocks
, bitmap need_phis
)
462 EXECUTE_IF_SET_IN_BITMAP (changed_bbs
, 0, index
, bi
)
463 find_uses_to_rename_bb (BASIC_BLOCK (index
), use_blocks
, need_phis
);
466 find_uses_to_rename_bb (bb
, use_blocks
, need_phis
);
469 /* Rewrites the program into a loop closed ssa form -- i.e. inserts extra
470 phi nodes to ensure that no variable is used outside the loop it is
473 This strengthening of the basic ssa form has several advantages:
475 1) Updating it during unrolling/peeling/versioning is trivial, since
476 we do not need to care about the uses outside of the loop.
477 The same applies to virtual operands which are also rewritten into
478 loop closed SSA form. Note that virtual operands are always live
480 2) The behavior of all uses of an induction variable is the same.
481 Without this, you need to distinguish the case when the variable
482 is used outside of the loop it is defined in, for example
484 for (i = 0; i < 100; i++)
486 for (j = 0; j < 100; j++)
494 Looking from the outer loop with the normal SSA form, the first use of k
495 is not well-behaved, while the second one is an induction variable with
498 If CHANGED_BBS is not NULL, we look for uses outside loops only in
499 the basic blocks in this set.
501 UPDATE_FLAG is used in the call to update_ssa. See
502 TODO_update_ssa* for documentation. */
505 rewrite_into_loop_closed_ssa (bitmap changed_bbs
, unsigned update_flag
)
508 bitmap names_to_rename
;
510 loops_state_set (LOOP_CLOSED_SSA
);
511 if (number_of_loops (cfun
) <= 1)
514 /* If the pass has caused the SSA form to be out-of-date, update it
516 update_ssa (update_flag
);
518 bitmap_obstack_initialize (&loop_renamer_obstack
);
520 names_to_rename
= BITMAP_ALLOC (&loop_renamer_obstack
);
522 /* Uses of names to rename. We don't have to initialize this array,
523 because we know that we will only have entries for the SSA names
524 in NAMES_TO_RENAME. */
525 use_blocks
= XNEWVEC (bitmap
, num_ssa_names
);
527 /* Find the uses outside loops. */
528 find_uses_to_rename (changed_bbs
, use_blocks
, names_to_rename
);
530 if (!bitmap_empty_p (names_to_rename
))
532 /* An array of bitmaps where LOOP_EXITS[I] is the set of basic blocks
533 that are the destination of an edge exiting loop number I. */
534 bitmap
*loop_exits
= XNEWVEC (bitmap
, number_of_loops (cfun
));
535 get_loops_exits (loop_exits
);
537 /* Add the PHI nodes on exits of the loops for the names we need to
539 add_exit_phis (names_to_rename
, use_blocks
, loop_exits
);
543 /* Fix up all the names found to be used outside their original
545 update_ssa (TODO_update_ssa
);
548 bitmap_obstack_release (&loop_renamer_obstack
);
552 /* Check invariants of the loop closed ssa form for the USE in BB. */
555 check_loop_closed_ssa_use (basic_block bb
, tree use
)
560 if (TREE_CODE (use
) != SSA_NAME
|| virtual_operand_p (use
))
563 def
= SSA_NAME_DEF_STMT (use
);
564 def_bb
= gimple_bb (def
);
566 || flow_bb_inside_loop_p (def_bb
->loop_father
, bb
));
569 /* Checks invariants of loop closed ssa form in statement STMT in BB. */
572 check_loop_closed_ssa_stmt (basic_block bb
, gimple stmt
)
577 if (is_gimple_debug (stmt
))
580 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, SSA_OP_USE
)
581 check_loop_closed_ssa_use (bb
, var
);
584 /* Checks that invariants of the loop closed ssa form are preserved.
585 Call verify_ssa when VERIFY_SSA_P is true. */
588 verify_loop_closed_ssa (bool verify_ssa_p
)
591 gimple_stmt_iterator bsi
;
596 if (number_of_loops (cfun
) <= 1)
602 timevar_push (TV_VERIFY_LOOP_CLOSED
);
606 for (bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
608 phi
= gsi_stmt (bsi
);
609 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
610 check_loop_closed_ssa_use (e
->src
,
611 PHI_ARG_DEF_FROM_EDGE (phi
, e
));
614 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
615 check_loop_closed_ssa_stmt (bb
, gsi_stmt (bsi
));
618 timevar_pop (TV_VERIFY_LOOP_CLOSED
);
621 /* Split loop exit edge EXIT. The things are a bit complicated by a need to
622 preserve the loop closed ssa form. The newly created block is returned. */
625 split_loop_exit_edge (edge exit
)
627 basic_block dest
= exit
->dest
;
628 basic_block bb
= split_edge (exit
);
632 gimple_stmt_iterator psi
;
633 source_location locus
;
635 for (psi
= gsi_start_phis (dest
); !gsi_end_p (psi
); gsi_next (&psi
))
637 phi
= gsi_stmt (psi
);
638 op_p
= PHI_ARG_DEF_PTR_FROM_EDGE (phi
, single_succ_edge (bb
));
639 locus
= gimple_phi_arg_location_from_edge (phi
, single_succ_edge (bb
));
641 name
= USE_FROM_PTR (op_p
);
643 /* If the argument of the PHI node is a constant, we do not need
644 to keep it inside loop. */
645 if (TREE_CODE (name
) != SSA_NAME
)
648 /* Otherwise create an auxiliary phi node that will copy the value
649 of the SSA name out of the loop. */
650 new_name
= duplicate_ssa_name (name
, NULL
);
651 new_phi
= create_phi_node (new_name
, bb
);
652 add_phi_arg (new_phi
, name
, exit
, locus
);
653 SET_USE (op_p
, new_name
);
659 /* Returns the basic block in that statements should be emitted for induction
660 variables incremented at the end of the LOOP. */
663 ip_end_pos (struct loop
*loop
)
668 /* Returns the basic block in that statements should be emitted for induction
669 variables incremented just before exit condition of a LOOP. */
672 ip_normal_pos (struct loop
*loop
)
678 if (!single_pred_p (loop
->latch
))
681 bb
= single_pred (loop
->latch
);
682 last
= last_stmt (bb
);
684 || gimple_code (last
) != GIMPLE_COND
)
687 exit
= EDGE_SUCC (bb
, 0);
688 if (exit
->dest
== loop
->latch
)
689 exit
= EDGE_SUCC (bb
, 1);
691 if (flow_bb_inside_loop_p (loop
, exit
->dest
))
697 /* Stores the standard position for induction variable increment in LOOP
698 (just before the exit condition if it is available and latch block is empty,
699 end of the latch block otherwise) to BSI. INSERT_AFTER is set to true if
700 the increment should be inserted after *BSI. */
703 standard_iv_increment_position (struct loop
*loop
, gimple_stmt_iterator
*bsi
,
706 basic_block bb
= ip_normal_pos (loop
), latch
= ip_end_pos (loop
);
707 gimple last
= last_stmt (latch
);
710 || (last
&& gimple_code (last
) != GIMPLE_LABEL
))
712 *bsi
= gsi_last_bb (latch
);
713 *insert_after
= true;
717 *bsi
= gsi_last_bb (bb
);
718 *insert_after
= false;
722 /* Copies phi node arguments for duplicated blocks. The index of the first
723 duplicated block is FIRST_NEW_BLOCK. */
726 copy_phi_node_args (unsigned first_new_block
)
730 for (i
= first_new_block
; i
< (unsigned) last_basic_block
; i
++)
731 BASIC_BLOCK (i
)->flags
|= BB_DUPLICATED
;
733 for (i
= first_new_block
; i
< (unsigned) last_basic_block
; i
++)
734 add_phi_args_after_copy_bb (BASIC_BLOCK (i
));
736 for (i
= first_new_block
; i
< (unsigned) last_basic_block
; i
++)
737 BASIC_BLOCK (i
)->flags
&= ~BB_DUPLICATED
;
741 /* The same as cfgloopmanip.c:duplicate_loop_to_header_edge, but also
742 updates the PHI nodes at start of the copied region. In order to
743 achieve this, only loops whose exits all lead to the same location
746 Notice that we do not completely update the SSA web after
747 duplication. The caller is responsible for calling update_ssa
748 after the loop has been duplicated. */
751 gimple_duplicate_loop_to_header_edge (struct loop
*loop
, edge e
,
752 unsigned int ndupl
, sbitmap wont_exit
,
753 edge orig
, vec
<edge
> *to_remove
,
756 unsigned first_new_block
;
758 if (!loops_state_satisfies_p (LOOPS_HAVE_SIMPLE_LATCHES
))
760 if (!loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS
))
763 #ifdef ENABLE_CHECKING
764 /* ??? This forces needless update_ssa calls after processing each
765 loop instead of just once after processing all loops. We should
766 instead verify that loop-closed SSA form is up-to-date for LOOP
767 only (and possibly SSA form). For now just skip verifying if
768 there are to-be renamed variables. */
769 if (!need_ssa_update_p (cfun
)
770 && loops_state_satisfies_p (LOOP_CLOSED_SSA
))
771 verify_loop_closed_ssa (true);
774 first_new_block
= last_basic_block
;
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 gimple phi_old_loop
, phi_new_loop
, phi_rest
;
1037 gimple_stmt_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 vec
<edge
> to_remove
= vNULL
;
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
= gsi_stmt (psi_old_loop
);
1145 phi_new_loop
= gsi_stmt (psi_new_loop
);
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
, NULL
);
1159 else if (TREE_CODE (init
) == SSA_NAME
1160 && useless_type_conversion_p (TREE_TYPE (init
),
1162 new_init
= copy_ssa_name (init
, NULL
);
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 to_remove
.release ();
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
);