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 struct loop
*def_loop
= def_bb
->loop_father
;
191 unsigned def_loop_depth
= loop_depth (def_loop
);
192 bitmap def_loop_exits
;
194 /* Normally the work list size is bounded by the number of basic
195 blocks in the largest loop. We don't know this number, but we
196 can be fairly sure that it will be relatively small. */
197 auto_vec
<basic_block
> worklist (MAX (8, n_basic_blocks_for_fn (cfun
) / 128));
199 EXECUTE_IF_SET_IN_BITMAP (use_blocks
, 0, i
, bi
)
201 basic_block use_bb
= BASIC_BLOCK (i
);
202 struct loop
*use_loop
= use_bb
->loop_father
;
203 gcc_checking_assert (def_loop
!= use_loop
204 && ! flow_loop_nested_p (def_loop
, use_loop
));
205 if (! flow_loop_nested_p (use_loop
, def_loop
))
206 use_bb
= find_sibling_superloop (use_loop
, def_loop
)->header
;
207 if (bitmap_set_bit (live_exits
, use_bb
->index
))
208 worklist
.safe_push (use_bb
);
211 /* Iterate until the worklist is empty. */
212 while (! worklist
.is_empty ())
217 /* Pull a block off the worklist. */
218 basic_block bb
= worklist
.pop ();
220 /* Make sure we have at least enough room in the work list
221 for all predecessors of this block. */
222 worklist
.reserve (EDGE_COUNT (bb
->preds
));
224 /* For each predecessor block. */
225 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
227 basic_block pred
= e
->src
;
228 struct loop
*pred_loop
= pred
->loop_father
;
229 unsigned pred_loop_depth
= loop_depth (pred_loop
);
232 /* We should have met DEF_BB along the way. */
233 gcc_assert (pred
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
));
235 if (pred_loop_depth
>= def_loop_depth
)
237 if (pred_loop_depth
> def_loop_depth
)
238 pred_loop
= superloop_at_depth (pred_loop
, def_loop_depth
);
239 /* If we've reached DEF_LOOP, our train ends here. */
240 if (pred_loop
== def_loop
)
243 else if (! flow_loop_nested_p (pred_loop
, def_loop
))
244 pred
= find_sibling_superloop (pred_loop
, def_loop
)->header
;
246 /* Add PRED to the LIVEIN set. PRED_VISITED is true if
247 we had already added PRED to LIVEIN before. */
248 pred_visited
= !bitmap_set_bit (live_exits
, pred
->index
);
250 /* If we have visited PRED before, don't add it to the worklist.
251 If BB dominates PRED, then we're probably looking at a loop.
252 We're only interested in looking up in the dominance tree
253 because DEF_BB dominates all the uses. */
254 if (pred_visited
|| dominated_by_p (CDI_DOMINATORS
, pred
, bb
))
257 worklist
.quick_push (pred
);
261 def_loop_exits
= BITMAP_ALLOC (&loop_renamer_obstack
);
262 for (struct loop
*loop
= def_loop
;
263 loop
!= current_loops
->tree_root
;
264 loop
= loop_outer (loop
))
265 bitmap_ior_into (def_loop_exits
, loop_exits
[loop
->num
]);
266 bitmap_and_into (live_exits
, def_loop_exits
);
267 BITMAP_FREE (def_loop_exits
);
270 /* Add a loop-closing PHI for VAR in basic block EXIT. */
273 add_exit_phi (basic_block exit
, tree var
)
279 #ifdef ENABLE_CHECKING
280 /* Check that at least one of the edges entering the EXIT block exits
281 the loop, or a superloop of that loop, that VAR is defined in. */
282 gimple def_stmt
= SSA_NAME_DEF_STMT (var
);
283 basic_block def_bb
= gimple_bb (def_stmt
);
284 FOR_EACH_EDGE (e
, ei
, exit
->preds
)
286 struct loop
*aloop
= find_common_loop (def_bb
->loop_father
,
287 e
->src
->loop_father
);
288 if (!flow_bb_inside_loop_p (aloop
, e
->dest
))
292 gcc_checking_assert (e
);
295 phi
= create_phi_node (NULL_TREE
, exit
);
296 create_new_def_for (var
, phi
, gimple_phi_result_ptr (phi
));
297 FOR_EACH_EDGE (e
, ei
, exit
->preds
)
298 add_phi_arg (phi
, var
, e
, UNKNOWN_LOCATION
);
300 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
302 fprintf (dump_file
, ";; Created LCSSA PHI: ");
303 print_gimple_stmt (dump_file
, phi
, 0, dump_flags
);
307 /* Add exit phis for VAR that is used in LIVEIN.
308 Exits of the loops are stored in LOOP_EXITS. */
311 add_exit_phis_var (tree var
, bitmap use_blocks
, bitmap
*loop_exits
)
315 basic_block def_bb
= gimple_bb (SSA_NAME_DEF_STMT (var
));
316 bitmap live_exits
= BITMAP_ALLOC (&loop_renamer_obstack
);
318 gcc_checking_assert (! bitmap_bit_p (use_blocks
, def_bb
->index
));
320 compute_live_loop_exits (live_exits
, use_blocks
, loop_exits
, def_bb
);
322 EXECUTE_IF_SET_IN_BITMAP (live_exits
, 0, index
, bi
)
324 add_exit_phi (BASIC_BLOCK (index
), var
);
327 BITMAP_FREE (live_exits
);
330 /* Add exit phis for the names marked in NAMES_TO_RENAME.
331 Exits of the loops are stored in EXITS. Sets of blocks where the ssa
332 names are used are stored in USE_BLOCKS. */
335 add_exit_phis (bitmap names_to_rename
, bitmap
*use_blocks
, bitmap
*loop_exits
)
340 EXECUTE_IF_SET_IN_BITMAP (names_to_rename
, 0, i
, bi
)
342 add_exit_phis_var (ssa_name (i
), use_blocks
[i
], loop_exits
);
346 /* Fill the array of bitmaps LOOP_EXITS with all loop exit edge targets. */
349 get_loops_exits (bitmap
*loop_exits
)
355 FOR_EACH_LOOP (loop
, 0)
357 vec
<edge
> exit_edges
= get_loop_exit_edges (loop
);
358 loop_exits
[loop
->num
] = BITMAP_ALLOC (&loop_renamer_obstack
);
359 FOR_EACH_VEC_ELT (exit_edges
, j
, e
)
360 bitmap_set_bit (loop_exits
[loop
->num
], e
->dest
->index
);
361 exit_edges
.release ();
365 /* For USE in BB, if it is used outside of the loop it is defined in,
366 mark it for rewrite. Record basic block BB where it is used
367 to USE_BLOCKS. Record the ssa name index to NEED_PHIS bitmap. */
370 find_uses_to_rename_use (basic_block bb
, tree use
, bitmap
*use_blocks
,
375 struct loop
*def_loop
;
377 if (TREE_CODE (use
) != SSA_NAME
)
380 ver
= SSA_NAME_VERSION (use
);
381 def_bb
= gimple_bb (SSA_NAME_DEF_STMT (use
));
384 def_loop
= def_bb
->loop_father
;
386 /* If the definition is not inside a loop, it is not interesting. */
387 if (!loop_outer (def_loop
))
390 /* If the use is not outside of the loop it is defined in, it is not
392 if (flow_bb_inside_loop_p (def_loop
, bb
))
395 /* If we're seeing VER for the first time, we still have to allocate
396 a bitmap for its uses. */
397 if (bitmap_set_bit (need_phis
, ver
))
398 use_blocks
[ver
] = BITMAP_ALLOC (&loop_renamer_obstack
);
399 bitmap_set_bit (use_blocks
[ver
], bb
->index
);
402 /* For uses in STMT, mark names that are used outside of the loop they are
403 defined to rewrite. Record the set of blocks in that the ssa
404 names are defined to USE_BLOCKS and the ssa names themselves to
408 find_uses_to_rename_stmt (gimple stmt
, bitmap
*use_blocks
, bitmap need_phis
)
412 basic_block bb
= gimple_bb (stmt
);
414 if (is_gimple_debug (stmt
))
417 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, SSA_OP_USE
)
418 find_uses_to_rename_use (bb
, var
, use_blocks
, need_phis
);
421 /* Marks names that are used in BB and outside of the loop they are
422 defined in for rewrite. Records the set of blocks in that the ssa
423 names are defined to USE_BLOCKS. Record the SSA names that will
424 need exit PHIs in NEED_PHIS. */
427 find_uses_to_rename_bb (basic_block bb
, bitmap
*use_blocks
, bitmap need_phis
)
429 gimple_stmt_iterator bsi
;
433 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
434 for (bsi
= gsi_start_phis (e
->dest
); !gsi_end_p (bsi
); gsi_next (&bsi
))
436 gimple phi
= gsi_stmt (bsi
);
437 if (! virtual_operand_p (gimple_phi_result (phi
)))
438 find_uses_to_rename_use (bb
, PHI_ARG_DEF_FROM_EDGE (phi
, e
),
439 use_blocks
, need_phis
);
442 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
443 find_uses_to_rename_stmt (gsi_stmt (bsi
), use_blocks
, need_phis
);
446 /* Marks names that are used outside of the loop they are defined in
447 for rewrite. Records the set of blocks in that the ssa
448 names are defined to USE_BLOCKS. If CHANGED_BBS is not NULL,
449 scan only blocks in this set. */
452 find_uses_to_rename (bitmap changed_bbs
, bitmap
*use_blocks
, bitmap need_phis
)
459 EXECUTE_IF_SET_IN_BITMAP (changed_bbs
, 0, index
, bi
)
460 find_uses_to_rename_bb (BASIC_BLOCK (index
), use_blocks
, need_phis
);
463 find_uses_to_rename_bb (bb
, use_blocks
, need_phis
);
466 /* Rewrites the program into a loop closed ssa form -- i.e. inserts extra
467 phi nodes to ensure that no variable is used outside the loop it is
470 This strengthening of the basic ssa form has several advantages:
472 1) Updating it during unrolling/peeling/versioning is trivial, since
473 we do not need to care about the uses outside of the loop.
474 The same applies to virtual operands which are also rewritten into
475 loop closed SSA form. Note that virtual operands are always live
477 2) The behavior of all uses of an induction variable is the same.
478 Without this, you need to distinguish the case when the variable
479 is used outside of the loop it is defined in, for example
481 for (i = 0; i < 100; i++)
483 for (j = 0; j < 100; j++)
491 Looking from the outer loop with the normal SSA form, the first use of k
492 is not well-behaved, while the second one is an induction variable with
495 If CHANGED_BBS is not NULL, we look for uses outside loops only in
496 the basic blocks in this set.
498 UPDATE_FLAG is used in the call to update_ssa. See
499 TODO_update_ssa* for documentation. */
502 rewrite_into_loop_closed_ssa (bitmap changed_bbs
, unsigned update_flag
)
505 bitmap names_to_rename
;
507 loops_state_set (LOOP_CLOSED_SSA
);
508 if (number_of_loops (cfun
) <= 1)
511 /* If the pass has caused the SSA form to be out-of-date, update it
513 update_ssa (update_flag
);
515 bitmap_obstack_initialize (&loop_renamer_obstack
);
517 names_to_rename
= BITMAP_ALLOC (&loop_renamer_obstack
);
519 /* Uses of names to rename. We don't have to initialize this array,
520 because we know that we will only have entries for the SSA names
521 in NAMES_TO_RENAME. */
522 use_blocks
= XNEWVEC (bitmap
, num_ssa_names
);
524 /* Find the uses outside loops. */
525 find_uses_to_rename (changed_bbs
, use_blocks
, names_to_rename
);
527 if (!bitmap_empty_p (names_to_rename
))
529 /* An array of bitmaps where LOOP_EXITS[I] is the set of basic blocks
530 that are the destination of an edge exiting loop number I. */
531 bitmap
*loop_exits
= XNEWVEC (bitmap
, number_of_loops (cfun
));
532 get_loops_exits (loop_exits
);
534 /* Add the PHI nodes on exits of the loops for the names we need to
536 add_exit_phis (names_to_rename
, use_blocks
, loop_exits
);
540 /* Fix up all the names found to be used outside their original
542 update_ssa (TODO_update_ssa
);
545 bitmap_obstack_release (&loop_renamer_obstack
);
549 /* Check invariants of the loop closed ssa form for the USE in BB. */
552 check_loop_closed_ssa_use (basic_block bb
, tree use
)
557 if (TREE_CODE (use
) != SSA_NAME
|| virtual_operand_p (use
))
560 def
= SSA_NAME_DEF_STMT (use
);
561 def_bb
= gimple_bb (def
);
563 || flow_bb_inside_loop_p (def_bb
->loop_father
, bb
));
566 /* Checks invariants of loop closed ssa form in statement STMT in BB. */
569 check_loop_closed_ssa_stmt (basic_block bb
, gimple stmt
)
574 if (is_gimple_debug (stmt
))
577 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, SSA_OP_USE
)
578 check_loop_closed_ssa_use (bb
, var
);
581 /* Checks that invariants of the loop closed ssa form are preserved.
582 Call verify_ssa when VERIFY_SSA_P is true. */
585 verify_loop_closed_ssa (bool verify_ssa_p
)
588 gimple_stmt_iterator bsi
;
593 if (number_of_loops (cfun
) <= 1)
599 timevar_push (TV_VERIFY_LOOP_CLOSED
);
603 for (bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
605 phi
= gsi_stmt (bsi
);
606 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
607 check_loop_closed_ssa_use (e
->src
,
608 PHI_ARG_DEF_FROM_EDGE (phi
, e
));
611 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
612 check_loop_closed_ssa_stmt (bb
, gsi_stmt (bsi
));
615 timevar_pop (TV_VERIFY_LOOP_CLOSED
);
618 /* Split loop exit edge EXIT. The things are a bit complicated by a need to
619 preserve the loop closed ssa form. The newly created block is returned. */
622 split_loop_exit_edge (edge exit
)
624 basic_block dest
= exit
->dest
;
625 basic_block bb
= split_edge (exit
);
629 gimple_stmt_iterator psi
;
630 source_location locus
;
632 for (psi
= gsi_start_phis (dest
); !gsi_end_p (psi
); gsi_next (&psi
))
634 phi
= gsi_stmt (psi
);
635 op_p
= PHI_ARG_DEF_PTR_FROM_EDGE (phi
, single_succ_edge (bb
));
636 locus
= gimple_phi_arg_location_from_edge (phi
, single_succ_edge (bb
));
638 name
= USE_FROM_PTR (op_p
);
640 /* If the argument of the PHI node is a constant, we do not need
641 to keep it inside loop. */
642 if (TREE_CODE (name
) != SSA_NAME
)
645 /* Otherwise create an auxiliary phi node that will copy the value
646 of the SSA name out of the loop. */
647 new_name
= duplicate_ssa_name (name
, NULL
);
648 new_phi
= create_phi_node (new_name
, bb
);
649 add_phi_arg (new_phi
, name
, exit
, locus
);
650 SET_USE (op_p
, new_name
);
656 /* Returns the basic block in that statements should be emitted for induction
657 variables incremented at the end of the LOOP. */
660 ip_end_pos (struct loop
*loop
)
665 /* Returns the basic block in that statements should be emitted for induction
666 variables incremented just before exit condition of a LOOP. */
669 ip_normal_pos (struct loop
*loop
)
675 if (!single_pred_p (loop
->latch
))
678 bb
= single_pred (loop
->latch
);
679 last
= last_stmt (bb
);
681 || gimple_code (last
) != GIMPLE_COND
)
684 exit
= EDGE_SUCC (bb
, 0);
685 if (exit
->dest
== loop
->latch
)
686 exit
= EDGE_SUCC (bb
, 1);
688 if (flow_bb_inside_loop_p (loop
, exit
->dest
))
694 /* Stores the standard position for induction variable increment in LOOP
695 (just before the exit condition if it is available and latch block is empty,
696 end of the latch block otherwise) to BSI. INSERT_AFTER is set to true if
697 the increment should be inserted after *BSI. */
700 standard_iv_increment_position (struct loop
*loop
, gimple_stmt_iterator
*bsi
,
703 basic_block bb
= ip_normal_pos (loop
), latch
= ip_end_pos (loop
);
704 gimple last
= last_stmt (latch
);
707 || (last
&& gimple_code (last
) != GIMPLE_LABEL
))
709 *bsi
= gsi_last_bb (latch
);
710 *insert_after
= true;
714 *bsi
= gsi_last_bb (bb
);
715 *insert_after
= false;
719 /* Copies phi node arguments for duplicated blocks. The index of the first
720 duplicated block is FIRST_NEW_BLOCK. */
723 copy_phi_node_args (unsigned first_new_block
)
727 for (i
= first_new_block
; i
< (unsigned) last_basic_block
; i
++)
728 BASIC_BLOCK (i
)->flags
|= BB_DUPLICATED
;
730 for (i
= first_new_block
; i
< (unsigned) last_basic_block
; i
++)
731 add_phi_args_after_copy_bb (BASIC_BLOCK (i
));
733 for (i
= first_new_block
; i
< (unsigned) last_basic_block
; i
++)
734 BASIC_BLOCK (i
)->flags
&= ~BB_DUPLICATED
;
738 /* The same as cfgloopmanip.c:duplicate_loop_to_header_edge, but also
739 updates the PHI nodes at start of the copied region. In order to
740 achieve this, only loops whose exits all lead to the same location
743 Notice that we do not completely update the SSA web after
744 duplication. The caller is responsible for calling update_ssa
745 after the loop has been duplicated. */
748 gimple_duplicate_loop_to_header_edge (struct loop
*loop
, edge e
,
749 unsigned int ndupl
, sbitmap wont_exit
,
750 edge orig
, vec
<edge
> *to_remove
,
753 unsigned first_new_block
;
755 if (!loops_state_satisfies_p (LOOPS_HAVE_SIMPLE_LATCHES
))
757 if (!loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS
))
760 #ifdef ENABLE_CHECKING
761 /* ??? This forces needless update_ssa calls after processing each
762 loop instead of just once after processing all loops. We should
763 instead verify that loop-closed SSA form is up-to-date for LOOP
764 only (and possibly SSA form). For now just skip verifying if
765 there are to-be renamed variables. */
766 if (!need_ssa_update_p (cfun
)
767 && loops_state_satisfies_p (LOOP_CLOSED_SSA
))
768 verify_loop_closed_ssa (true);
771 first_new_block
= last_basic_block
;
772 if (!duplicate_loop_to_header_edge (loop
, e
, ndupl
, wont_exit
,
773 orig
, to_remove
, flags
))
776 /* Readd the removed phi args for e. */
777 flush_pending_stmts (e
);
779 /* Copy the phi node arguments. */
780 copy_phi_node_args (first_new_block
);
787 /* Returns true if we can unroll LOOP FACTOR times. Number
788 of iterations of the loop is returned in NITER. */
791 can_unroll_loop_p (struct loop
*loop
, unsigned factor
,
792 struct tree_niter_desc
*niter
)
796 /* Check whether unrolling is possible. We only want to unroll loops
797 for that we are able to determine number of iterations. We also
798 want to split the extra iterations of the loop from its end,
799 therefore we require that the loop has precisely one
802 exit
= single_dom_exit (loop
);
806 if (!number_of_iterations_exit (loop
, exit
, niter
, false)
807 || niter
->cmp
== ERROR_MARK
808 /* Scalar evolutions analysis might have copy propagated
809 the abnormal ssa names into these expressions, hence
810 emitting the computations based on them during loop
811 unrolling might create overlapping life ranges for
812 them, and failures in out-of-ssa. */
813 || contains_abnormal_ssa_name_p (niter
->may_be_zero
)
814 || contains_abnormal_ssa_name_p (niter
->control
.base
)
815 || contains_abnormal_ssa_name_p (niter
->control
.step
)
816 || contains_abnormal_ssa_name_p (niter
->bound
))
819 /* And of course, we must be able to duplicate the loop. */
820 if (!can_duplicate_loop_p (loop
))
823 /* The final loop should be small enough. */
824 if (tree_num_loop_insns (loop
, &eni_size_weights
) * factor
825 > (unsigned) PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS
))
831 /* Determines the conditions that control execution of LOOP unrolled FACTOR
832 times. DESC is number of iterations of LOOP. ENTER_COND is set to
833 condition that must be true if the main loop can be entered.
834 EXIT_BASE, EXIT_STEP, EXIT_CMP and EXIT_BOUND are set to values describing
835 how the exit from the unrolled loop should be controlled. */
838 determine_exit_conditions (struct loop
*loop
, struct tree_niter_desc
*desc
,
839 unsigned factor
, tree
*enter_cond
,
840 tree
*exit_base
, tree
*exit_step
,
841 enum tree_code
*exit_cmp
, tree
*exit_bound
)
844 tree base
= desc
->control
.base
;
845 tree step
= desc
->control
.step
;
846 tree bound
= desc
->bound
;
847 tree type
= TREE_TYPE (step
);
849 tree min
= lower_bound_in_type (type
, type
);
850 tree max
= upper_bound_in_type (type
, type
);
851 enum tree_code cmp
= desc
->cmp
;
852 tree cond
= boolean_true_node
, assum
;
854 /* For pointers, do the arithmetics in the type of step. */
855 base
= fold_convert (type
, base
);
856 bound
= fold_convert (type
, bound
);
858 *enter_cond
= boolean_false_node
;
859 *exit_base
= NULL_TREE
;
860 *exit_step
= NULL_TREE
;
861 *exit_cmp
= ERROR_MARK
;
862 *exit_bound
= NULL_TREE
;
863 gcc_assert (cmp
!= ERROR_MARK
);
865 /* We only need to be correct when we answer question
866 "Do at least FACTOR more iterations remain?" in the unrolled loop.
867 Thus, transforming BASE + STEP * i <> BOUND to
868 BASE + STEP * i < BOUND is ok. */
871 if (tree_int_cst_sign_bit (step
))
876 else if (cmp
== LT_EXPR
)
878 gcc_assert (!tree_int_cst_sign_bit (step
));
880 else if (cmp
== GT_EXPR
)
882 gcc_assert (tree_int_cst_sign_bit (step
));
887 /* The main body of the loop may be entered iff:
889 1) desc->may_be_zero is false.
890 2) it is possible to check that there are at least FACTOR iterations
891 of the loop, i.e., BOUND - step * FACTOR does not overflow.
892 3) # of iterations is at least FACTOR */
894 if (!integer_zerop (desc
->may_be_zero
))
895 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
896 invert_truthvalue (desc
->may_be_zero
),
899 bigstep
= fold_build2 (MULT_EXPR
, type
, step
,
900 build_int_cst_type (type
, factor
));
901 delta
= fold_build2 (MINUS_EXPR
, type
, bigstep
, step
);
903 assum
= fold_build2 (GE_EXPR
, boolean_type_node
,
905 fold_build2 (PLUS_EXPR
, type
, min
, delta
));
907 assum
= fold_build2 (LE_EXPR
, boolean_type_node
,
909 fold_build2 (PLUS_EXPR
, type
, max
, delta
));
910 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
, assum
, cond
);
912 bound
= fold_build2 (MINUS_EXPR
, type
, bound
, delta
);
913 assum
= fold_build2 (cmp
, boolean_type_node
, base
, bound
);
914 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
, assum
, cond
);
916 cond
= force_gimple_operand (unshare_expr (cond
), &stmts
, false, NULL_TREE
);
918 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
919 /* cond now may be a gimple comparison, which would be OK, but also any
920 other gimple rhs (say a && b). In this case we need to force it to
922 if (!is_gimple_condexpr (cond
))
924 cond
= force_gimple_operand (cond
, &stmts
, true, NULL_TREE
);
926 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
930 base
= force_gimple_operand (unshare_expr (base
), &stmts
, true, NULL_TREE
);
932 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
933 bound
= force_gimple_operand (unshare_expr (bound
), &stmts
, true, NULL_TREE
);
935 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
938 *exit_step
= bigstep
;
943 /* Scales the frequencies of all basic blocks in LOOP that are strictly
944 dominated by BB by NUM/DEN. */
947 scale_dominated_blocks_in_loop (struct loop
*loop
, basic_block bb
,
955 for (son
= first_dom_son (CDI_DOMINATORS
, bb
);
957 son
= next_dom_son (CDI_DOMINATORS
, son
))
959 if (!flow_bb_inside_loop_p (loop
, son
))
961 scale_bbs_frequencies_int (&son
, 1, num
, den
);
962 scale_dominated_blocks_in_loop (loop
, son
, num
, den
);
966 /* Unroll LOOP FACTOR times. DESC describes number of iterations of LOOP.
967 EXIT is the exit of the loop to that DESC corresponds.
969 If N is number of iterations of the loop and MAY_BE_ZERO is the condition
970 under that loop exits in the first iteration even if N != 0,
974 x = phi (init, next);
982 becomes (with possibly the exit conditions formulated a bit differently,
983 avoiding the need to create a new iv):
985 if (MAY_BE_ZERO || N < FACTOR)
990 x = phi (init, next);
1001 } while (N >= FACTOR);
1004 init' = phi (init, x);
1008 x = phi (init', next);
1016 Before the loop is unrolled, TRANSFORM is called for it (only for the
1017 unrolled loop, but not for its versioned copy). DATA is passed to
1020 /* Probability in % that the unrolled loop is entered. Just a guess. */
1021 #define PROB_UNROLLED_LOOP_ENTERED 90
1024 tree_transform_and_unroll_loop (struct loop
*loop
, unsigned factor
,
1025 edge exit
, struct tree_niter_desc
*desc
,
1026 transform_callback transform
,
1030 tree ctr_before
, ctr_after
;
1031 tree enter_main_cond
, exit_base
, exit_step
, exit_bound
;
1032 enum tree_code exit_cmp
;
1033 gimple phi_old_loop
, phi_new_loop
, phi_rest
;
1034 gimple_stmt_iterator psi_old_loop
, psi_new_loop
;
1035 tree init
, next
, new_init
;
1036 struct loop
*new_loop
;
1037 basic_block rest
, exit_bb
;
1038 edge old_entry
, new_entry
, old_latch
, precond_edge
, new_exit
;
1039 edge new_nonexit
, e
;
1040 gimple_stmt_iterator bsi
;
1043 unsigned est_niter
, prob_entry
, scale_unrolled
, scale_rest
, freq_e
, freq_h
;
1044 unsigned new_est_niter
, i
, prob
;
1045 unsigned irr
= loop_preheader_edge (loop
)->flags
& EDGE_IRREDUCIBLE_LOOP
;
1047 auto_vec
<edge
> to_remove
;
1049 est_niter
= expected_loop_iterations (loop
);
1050 determine_exit_conditions (loop
, desc
, factor
,
1051 &enter_main_cond
, &exit_base
, &exit_step
,
1052 &exit_cmp
, &exit_bound
);
1054 /* Let us assume that the unrolled loop is quite likely to be entered. */
1055 if (integer_nonzerop (enter_main_cond
))
1056 prob_entry
= REG_BR_PROB_BASE
;
1058 prob_entry
= PROB_UNROLLED_LOOP_ENTERED
* REG_BR_PROB_BASE
/ 100;
1060 /* The values for scales should keep profile consistent, and somewhat close
1063 TODO: The current value of SCALE_REST makes it appear that the loop that
1064 is created by splitting the remaining iterations of the unrolled loop is
1065 executed the same number of times as the original loop, and with the same
1066 frequencies, which is obviously wrong. This does not appear to cause
1067 problems, so we do not bother with fixing it for now. To make the profile
1068 correct, we would need to change the probability of the exit edge of the
1069 loop, and recompute the distribution of frequencies in its body because
1070 of this change (scale the frequencies of blocks before and after the exit
1071 by appropriate factors). */
1072 scale_unrolled
= prob_entry
;
1073 scale_rest
= REG_BR_PROB_BASE
;
1075 new_loop
= loop_version (loop
, enter_main_cond
, NULL
,
1076 prob_entry
, scale_unrolled
, scale_rest
, true);
1077 gcc_assert (new_loop
!= NULL
);
1078 update_ssa (TODO_update_ssa
);
1080 /* Determine the probability of the exit edge of the unrolled loop. */
1081 new_est_niter
= est_niter
/ factor
;
1083 /* Without profile feedback, loops for that we do not know a better estimate
1084 are assumed to roll 10 times. When we unroll such loop, it appears to
1085 roll too little, and it may even seem to be cold. To avoid this, we
1086 ensure that the created loop appears to roll at least 5 times (but at
1087 most as many times as before unrolling). */
1088 if (new_est_niter
< 5)
1091 new_est_niter
= est_niter
;
1096 /* Prepare the cfg and update the phi nodes. Move the loop exit to the
1097 loop latch (and make its condition dummy, for the moment). */
1098 rest
= loop_preheader_edge (new_loop
)->src
;
1099 precond_edge
= single_pred_edge (rest
);
1100 split_edge (loop_latch_edge (loop
));
1101 exit_bb
= single_pred (loop
->latch
);
1103 /* Since the exit edge will be removed, the frequency of all the blocks
1104 in the loop that are dominated by it must be scaled by
1105 1 / (1 - exit->probability). */
1106 scale_dominated_blocks_in_loop (loop
, exit
->src
,
1108 REG_BR_PROB_BASE
- exit
->probability
);
1110 bsi
= gsi_last_bb (exit_bb
);
1111 exit_if
= gimple_build_cond (EQ_EXPR
, integer_zero_node
,
1113 NULL_TREE
, NULL_TREE
);
1115 gsi_insert_after (&bsi
, exit_if
, GSI_NEW_STMT
);
1116 new_exit
= make_edge (exit_bb
, rest
, EDGE_FALSE_VALUE
| irr
);
1117 rescan_loop_exit (new_exit
, true, false);
1119 /* Set the probability of new exit to the same of the old one. Fix
1120 the frequency of the latch block, by scaling it back by
1121 1 - exit->probability. */
1122 new_exit
->count
= exit
->count
;
1123 new_exit
->probability
= exit
->probability
;
1124 new_nonexit
= single_pred_edge (loop
->latch
);
1125 new_nonexit
->probability
= REG_BR_PROB_BASE
- exit
->probability
;
1126 new_nonexit
->flags
= EDGE_TRUE_VALUE
;
1127 new_nonexit
->count
-= exit
->count
;
1128 if (new_nonexit
->count
< 0)
1129 new_nonexit
->count
= 0;
1130 scale_bbs_frequencies_int (&loop
->latch
, 1, new_nonexit
->probability
,
1133 old_entry
= loop_preheader_edge (loop
);
1134 new_entry
= loop_preheader_edge (new_loop
);
1135 old_latch
= loop_latch_edge (loop
);
1136 for (psi_old_loop
= gsi_start_phis (loop
->header
),
1137 psi_new_loop
= gsi_start_phis (new_loop
->header
);
1138 !gsi_end_p (psi_old_loop
);
1139 gsi_next (&psi_old_loop
), gsi_next (&psi_new_loop
))
1141 phi_old_loop
= gsi_stmt (psi_old_loop
);
1142 phi_new_loop
= gsi_stmt (psi_new_loop
);
1144 init
= PHI_ARG_DEF_FROM_EDGE (phi_old_loop
, old_entry
);
1145 op
= PHI_ARG_DEF_PTR_FROM_EDGE (phi_new_loop
, new_entry
);
1146 gcc_assert (operand_equal_for_phi_arg_p (init
, USE_FROM_PTR (op
)));
1147 next
= PHI_ARG_DEF_FROM_EDGE (phi_old_loop
, old_latch
);
1149 /* Prefer using original variable as a base for the new ssa name.
1150 This is necessary for virtual ops, and useful in order to avoid
1151 losing debug info for real ops. */
1152 if (TREE_CODE (next
) == SSA_NAME
1153 && useless_type_conversion_p (TREE_TYPE (next
),
1155 new_init
= copy_ssa_name (next
, NULL
);
1156 else if (TREE_CODE (init
) == SSA_NAME
1157 && useless_type_conversion_p (TREE_TYPE (init
),
1159 new_init
= copy_ssa_name (init
, NULL
);
1160 else if (useless_type_conversion_p (TREE_TYPE (next
), TREE_TYPE (init
)))
1161 new_init
= make_temp_ssa_name (TREE_TYPE (next
), NULL
, "unrinittmp");
1163 new_init
= make_temp_ssa_name (TREE_TYPE (init
), NULL
, "unrinittmp");
1165 phi_rest
= create_phi_node (new_init
, rest
);
1167 add_phi_arg (phi_rest
, init
, precond_edge
, UNKNOWN_LOCATION
);
1168 add_phi_arg (phi_rest
, next
, new_exit
, UNKNOWN_LOCATION
);
1169 SET_USE (op
, new_init
);
1174 /* Transform the loop. */
1176 (*transform
) (loop
, data
);
1178 /* Unroll the loop and remove the exits in all iterations except for the
1180 wont_exit
= sbitmap_alloc (factor
);
1181 bitmap_ones (wont_exit
);
1182 bitmap_clear_bit (wont_exit
, factor
- 1);
1184 ok
= gimple_duplicate_loop_to_header_edge
1185 (loop
, loop_latch_edge (loop
), factor
- 1,
1186 wont_exit
, new_exit
, &to_remove
, DLTHE_FLAG_UPDATE_FREQ
);
1190 FOR_EACH_VEC_ELT (to_remove
, i
, e
)
1192 ok
= remove_path (e
);
1195 update_ssa (TODO_update_ssa
);
1197 /* Ensure that the frequencies in the loop match the new estimated
1198 number of iterations, and change the probability of the new
1200 freq_h
= loop
->header
->frequency
;
1201 freq_e
= EDGE_FREQUENCY (loop_preheader_edge (loop
));
1203 scale_loop_frequencies (loop
, freq_e
* (new_est_niter
+ 1), freq_h
);
1205 exit_bb
= single_pred (loop
->latch
);
1206 new_exit
= find_edge (exit_bb
, rest
);
1207 new_exit
->count
= loop_preheader_edge (loop
)->count
;
1208 new_exit
->probability
= REG_BR_PROB_BASE
/ (new_est_niter
+ 1);
1210 rest
->count
+= new_exit
->count
;
1211 rest
->frequency
+= EDGE_FREQUENCY (new_exit
);
1213 new_nonexit
= single_pred_edge (loop
->latch
);
1214 prob
= new_nonexit
->probability
;
1215 new_nonexit
->probability
= REG_BR_PROB_BASE
- new_exit
->probability
;
1216 new_nonexit
->count
= exit_bb
->count
- new_exit
->count
;
1217 if (new_nonexit
->count
< 0)
1218 new_nonexit
->count
= 0;
1220 scale_bbs_frequencies_int (&loop
->latch
, 1, new_nonexit
->probability
,
1223 /* Finally create the new counter for number of iterations and add the new
1224 exit instruction. */
1225 bsi
= gsi_last_nondebug_bb (exit_bb
);
1226 exit_if
= gsi_stmt (bsi
);
1227 create_iv (exit_base
, exit_step
, NULL_TREE
, loop
,
1228 &bsi
, false, &ctr_before
, &ctr_after
);
1229 gimple_cond_set_code (exit_if
, exit_cmp
);
1230 gimple_cond_set_lhs (exit_if
, ctr_after
);
1231 gimple_cond_set_rhs (exit_if
, exit_bound
);
1232 update_stmt (exit_if
);
1234 #ifdef ENABLE_CHECKING
1235 verify_flow_info ();
1236 verify_loop_structure ();
1237 verify_loop_closed_ssa (true);
1241 /* Wrapper over tree_transform_and_unroll_loop for case we do not
1242 want to transform the loop before unrolling. The meaning
1243 of the arguments is the same as for tree_transform_and_unroll_loop. */
1246 tree_unroll_loop (struct loop
*loop
, unsigned factor
,
1247 edge exit
, struct tree_niter_desc
*desc
)
1249 tree_transform_and_unroll_loop (loop
, factor
, exit
, desc
,
1253 /* Rewrite the phi node at position PSI in function of the main
1254 induction variable MAIN_IV and insert the generated code at GSI. */
1257 rewrite_phi_with_iv (loop_p loop
,
1258 gimple_stmt_iterator
*psi
,
1259 gimple_stmt_iterator
*gsi
,
1263 gimple stmt
, phi
= gsi_stmt (*psi
);
1264 tree atype
, mtype
, val
, res
= PHI_RESULT (phi
);
1266 if (virtual_operand_p (res
) || res
== main_iv
)
1272 if (!simple_iv (loop
, loop
, res
, &iv
, true))
1278 remove_phi_node (psi
, false);
1280 atype
= TREE_TYPE (res
);
1281 mtype
= POINTER_TYPE_P (atype
) ? sizetype
: atype
;
1282 val
= fold_build2 (MULT_EXPR
, mtype
, unshare_expr (iv
.step
),
1283 fold_convert (mtype
, main_iv
));
1284 val
= fold_build2 (POINTER_TYPE_P (atype
)
1285 ? POINTER_PLUS_EXPR
: PLUS_EXPR
,
1286 atype
, unshare_expr (iv
.base
), val
);
1287 val
= force_gimple_operand_gsi (gsi
, val
, false, NULL_TREE
, true,
1289 stmt
= gimple_build_assign (res
, val
);
1290 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
1293 /* Rewrite all the phi nodes of LOOP in function of the main induction
1294 variable MAIN_IV. */
1297 rewrite_all_phi_nodes_with_iv (loop_p loop
, tree main_iv
)
1300 basic_block
*bbs
= get_loop_body_in_dom_order (loop
);
1301 gimple_stmt_iterator psi
;
1303 for (i
= 0; i
< loop
->num_nodes
; i
++)
1305 basic_block bb
= bbs
[i
];
1306 gimple_stmt_iterator gsi
= gsi_after_labels (bb
);
1308 if (bb
->loop_father
!= loop
)
1311 for (psi
= gsi_start_phis (bb
); !gsi_end_p (psi
); )
1312 rewrite_phi_with_iv (loop
, &psi
, &gsi
, main_iv
);
1318 /* Bases all the induction variables in LOOP on a single induction
1319 variable (unsigned with base 0 and step 1), whose final value is
1320 compared with *NIT. When the IV type precision has to be larger
1321 than *NIT type precision, *NIT is converted to the larger type, the
1322 conversion code is inserted before the loop, and *NIT is updated to
1323 the new definition. When BUMP_IN_LATCH is true, the induction
1324 variable is incremented in the loop latch, otherwise it is
1325 incremented in the loop header. Return the induction variable that
1329 canonicalize_loop_ivs (struct loop
*loop
, tree
*nit
, bool bump_in_latch
)
1331 unsigned precision
= TYPE_PRECISION (TREE_TYPE (*nit
));
1332 unsigned original_precision
= precision
;
1333 tree type
, var_before
;
1334 gimple_stmt_iterator gsi
, psi
;
1336 edge exit
= single_dom_exit (loop
);
1338 enum machine_mode mode
;
1339 bool unsigned_p
= false;
1341 for (psi
= gsi_start_phis (loop
->header
);
1342 !gsi_end_p (psi
); gsi_next (&psi
))
1344 gimple phi
= gsi_stmt (psi
);
1345 tree res
= PHI_RESULT (phi
);
1348 type
= TREE_TYPE (res
);
1349 if (virtual_operand_p (res
)
1350 || (!INTEGRAL_TYPE_P (type
)
1351 && !POINTER_TYPE_P (type
))
1352 || TYPE_PRECISION (type
) < precision
)
1355 uns
= POINTER_TYPE_P (type
) | TYPE_UNSIGNED (type
);
1357 if (TYPE_PRECISION (type
) > precision
)
1362 precision
= TYPE_PRECISION (type
);
1365 mode
= smallest_mode_for_size (precision
, MODE_INT
);
1366 precision
= GET_MODE_PRECISION (mode
);
1367 type
= build_nonstandard_integer_type (precision
, unsigned_p
);
1369 if (original_precision
!= precision
)
1371 *nit
= fold_convert (type
, *nit
);
1372 *nit
= force_gimple_operand (*nit
, &stmts
, true, NULL_TREE
);
1374 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
1378 gsi
= gsi_last_bb (loop
->latch
);
1380 gsi
= gsi_last_nondebug_bb (loop
->header
);
1381 create_iv (build_int_cst_type (type
, 0), build_int_cst (type
, 1), NULL_TREE
,
1382 loop
, &gsi
, bump_in_latch
, &var_before
, NULL
);
1384 rewrite_all_phi_nodes_with_iv (loop
, var_before
);
1386 stmt
= last_stmt (exit
->src
);
1387 /* Make the loop exit if the control condition is not satisfied. */
1388 if (exit
->flags
& EDGE_TRUE_VALUE
)
1392 extract_true_false_edges_from_block (exit
->src
, &te
, &fe
);
1393 te
->flags
= EDGE_FALSE_VALUE
;
1394 fe
->flags
= EDGE_TRUE_VALUE
;
1396 gimple_cond_set_code (stmt
, LT_EXPR
);
1397 gimple_cond_set_lhs (stmt
, var_before
);
1398 gimple_cond_set_rhs (stmt
, *nit
);