1 /* High-level loop manipulation functions.
2 Copyright (C) 2004-2014 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by the
8 Free Software Foundation; either version 3, or (at your option) any
11 GCC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
22 #include "coretypes.h"
26 #include "basic-block.h"
27 #include "tree-ssa-alias.h"
28 #include "internal-fn.h"
29 #include "gimple-expr.h"
33 #include "gimple-iterator.h"
34 #include "gimplify-me.h"
35 #include "gimple-ssa.h"
37 #include "tree-phinodes.h"
38 #include "ssa-iterators.h"
39 #include "stringpool.h"
40 #include "tree-ssanames.h"
41 #include "tree-ssa-loop-ivopts.h"
42 #include "tree-ssa-loop-manip.h"
43 #include "tree-ssa-loop-niter.h"
44 #include "tree-ssa-loop.h"
45 #include "tree-into-ssa.h"
48 #include "gimple-pretty-print.h"
50 #include "tree-pass.h" /* ??? for TODO_update_ssa but this isn't a pass. */
51 #include "tree-scalar-evolution.h"
53 #include "tree-inline.h"
54 #include "langhooks.h"
56 /* All bitmaps for rewriting into loop-closed SSA go on this obstack,
57 so that we can free them all at once. */
58 static bitmap_obstack loop_renamer_obstack
;
60 /* Creates an induction variable with value BASE + STEP * iteration in LOOP.
61 It is expected that neither BASE nor STEP are shared with other expressions
62 (unless the sharing rules allow this). Use VAR as a base var_decl for it
63 (if NULL, a new temporary will be created). The increment will occur at
64 INCR_POS (after it if AFTER is true, before it otherwise). INCR_POS and
65 AFTER can be computed using standard_iv_increment_position. The ssa versions
66 of the variable before and after increment will be stored in VAR_BEFORE and
67 VAR_AFTER (unless they are NULL). */
70 create_iv (tree base
, tree step
, tree var
, struct loop
*loop
,
71 gimple_stmt_iterator
*incr_pos
, bool after
,
72 tree
*var_before
, tree
*var_after
)
79 enum tree_code incr_op
= PLUS_EXPR
;
80 edge pe
= loop_preheader_edge (loop
);
84 vb
= make_ssa_name (var
, NULL
);
85 va
= make_ssa_name (var
, NULL
);
89 vb
= make_temp_ssa_name (TREE_TYPE (base
), NULL
, "ivtmp");
90 va
= make_temp_ssa_name (TREE_TYPE (base
), NULL
, "ivtmp");
97 /* For easier readability of the created code, produce MINUS_EXPRs
99 if (TREE_CODE (step
) == INTEGER_CST
)
101 if (TYPE_UNSIGNED (TREE_TYPE (step
)))
103 step1
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
104 if (tree_int_cst_lt (step1
, step
))
106 incr_op
= MINUS_EXPR
;
114 if (!tree_expr_nonnegative_warnv_p (step
, &ovf
)
115 && may_negate_without_overflow_p (step
))
117 incr_op
= MINUS_EXPR
;
118 step
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
122 if (POINTER_TYPE_P (TREE_TYPE (base
)))
124 if (TREE_CODE (base
) == ADDR_EXPR
)
125 mark_addressable (TREE_OPERAND (base
, 0));
126 step
= convert_to_ptrofftype (step
);
127 if (incr_op
== MINUS_EXPR
)
128 step
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
129 incr_op
= POINTER_PLUS_EXPR
;
131 /* Gimplify the step if necessary. We put the computations in front of the
132 loop (i.e. the step should be loop invariant). */
133 step
= force_gimple_operand (step
, &stmts
, true, NULL_TREE
);
135 gsi_insert_seq_on_edge_immediate (pe
, stmts
);
137 stmt
= gimple_build_assign_with_ops (incr_op
, va
, vb
, step
);
139 gsi_insert_after (incr_pos
, stmt
, GSI_NEW_STMT
);
141 gsi_insert_before (incr_pos
, stmt
, GSI_NEW_STMT
);
143 initial
= force_gimple_operand (base
, &stmts
, true, var
);
145 gsi_insert_seq_on_edge_immediate (pe
, stmts
);
147 phi
= create_phi_node (vb
, loop
->header
);
148 add_phi_arg (phi
, initial
, loop_preheader_edge (loop
), UNKNOWN_LOCATION
);
149 add_phi_arg (phi
, va
, loop_latch_edge (loop
), UNKNOWN_LOCATION
);
152 /* Return the innermost superloop LOOP of USE_LOOP that is a superloop of
153 both DEF_LOOP and USE_LOOP. */
155 static inline struct loop
*
156 find_sibling_superloop (struct loop
*use_loop
, struct loop
*def_loop
)
158 unsigned ud
= loop_depth (use_loop
);
159 unsigned dd
= loop_depth (def_loop
);
160 gcc_assert (ud
> 0 && dd
> 0);
162 use_loop
= superloop_at_depth (use_loop
, dd
);
164 def_loop
= superloop_at_depth (def_loop
, ud
);
165 while (loop_outer (use_loop
) != loop_outer (def_loop
))
167 use_loop
= loop_outer (use_loop
);
168 def_loop
= loop_outer (def_loop
);
169 gcc_assert (use_loop
&& def_loop
);
174 /* DEF_BB is a basic block containing a DEF that needs rewriting into
175 loop-closed SSA form. USE_BLOCKS is the set of basic blocks containing
176 uses of DEF that "escape" from the loop containing DEF_BB (i.e. blocks in
177 USE_BLOCKS are dominated by DEF_BB but not in the loop father of DEF_B).
178 ALL_EXITS[I] is the set of all basic blocks that exit loop I.
180 Compute the subset of LOOP_EXITS that exit the loop containing DEF_BB
181 or one of its loop fathers, in which DEF is live. This set is returned
182 in the bitmap LIVE_EXITS.
184 Instead of computing the complete livein set of the def, we use the loop
185 nesting tree as a form of poor man's structure analysis. This greatly
186 speeds up the analysis, which is important because this function may be
187 called on all SSA names that need rewriting, one at a time. */
190 compute_live_loop_exits (bitmap live_exits
, bitmap use_blocks
,
191 bitmap
*loop_exits
, basic_block def_bb
)
195 struct loop
*def_loop
= def_bb
->loop_father
;
196 unsigned def_loop_depth
= loop_depth (def_loop
);
197 bitmap def_loop_exits
;
199 /* Normally the work list size is bounded by the number of basic
200 blocks in the largest loop. We don't know this number, but we
201 can be fairly sure that it will be relatively small. */
202 auto_vec
<basic_block
> worklist (MAX (8, n_basic_blocks_for_fn (cfun
) / 128));
204 EXECUTE_IF_SET_IN_BITMAP (use_blocks
, 0, i
, bi
)
206 basic_block use_bb
= BASIC_BLOCK_FOR_FN (cfun
, i
);
207 struct loop
*use_loop
= use_bb
->loop_father
;
208 gcc_checking_assert (def_loop
!= use_loop
209 && ! flow_loop_nested_p (def_loop
, use_loop
));
210 if (! flow_loop_nested_p (use_loop
, def_loop
))
211 use_bb
= find_sibling_superloop (use_loop
, def_loop
)->header
;
212 if (bitmap_set_bit (live_exits
, use_bb
->index
))
213 worklist
.safe_push (use_bb
);
216 /* Iterate until the worklist is empty. */
217 while (! worklist
.is_empty ())
222 /* Pull a block off the worklist. */
223 basic_block bb
= worklist
.pop ();
225 /* Make sure we have at least enough room in the work list
226 for all predecessors of this block. */
227 worklist
.reserve (EDGE_COUNT (bb
->preds
));
229 /* For each predecessor block. */
230 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
232 basic_block pred
= e
->src
;
233 struct loop
*pred_loop
= pred
->loop_father
;
234 unsigned pred_loop_depth
= loop_depth (pred_loop
);
237 /* We should have met DEF_BB along the way. */
238 gcc_assert (pred
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
));
240 if (pred_loop_depth
>= def_loop_depth
)
242 if (pred_loop_depth
> def_loop_depth
)
243 pred_loop
= superloop_at_depth (pred_loop
, def_loop_depth
);
244 /* If we've reached DEF_LOOP, our train ends here. */
245 if (pred_loop
== def_loop
)
248 else if (! flow_loop_nested_p (pred_loop
, def_loop
))
249 pred
= find_sibling_superloop (pred_loop
, def_loop
)->header
;
251 /* Add PRED to the LIVEIN set. PRED_VISITED is true if
252 we had already added PRED to LIVEIN before. */
253 pred_visited
= !bitmap_set_bit (live_exits
, pred
->index
);
255 /* If we have visited PRED before, don't add it to the worklist.
256 If BB dominates PRED, then we're probably looking at a loop.
257 We're only interested in looking up in the dominance tree
258 because DEF_BB dominates all the uses. */
259 if (pred_visited
|| dominated_by_p (CDI_DOMINATORS
, pred
, bb
))
262 worklist
.quick_push (pred
);
266 def_loop_exits
= BITMAP_ALLOC (&loop_renamer_obstack
);
267 for (struct loop
*loop
= def_loop
;
268 loop
!= current_loops
->tree_root
;
269 loop
= loop_outer (loop
))
270 bitmap_ior_into (def_loop_exits
, loop_exits
[loop
->num
]);
271 bitmap_and_into (live_exits
, def_loop_exits
);
272 BITMAP_FREE (def_loop_exits
);
275 /* Add a loop-closing PHI for VAR in basic block EXIT. */
278 add_exit_phi (basic_block exit
, tree var
)
284 #ifdef ENABLE_CHECKING
285 /* Check that at least one of the edges entering the EXIT block exits
286 the loop, or a superloop of that loop, that VAR is defined in. */
287 gimple def_stmt
= SSA_NAME_DEF_STMT (var
);
288 basic_block def_bb
= gimple_bb (def_stmt
);
289 FOR_EACH_EDGE (e
, ei
, exit
->preds
)
291 struct loop
*aloop
= find_common_loop (def_bb
->loop_father
,
292 e
->src
->loop_father
);
293 if (!flow_bb_inside_loop_p (aloop
, e
->dest
))
297 gcc_checking_assert (e
);
300 phi
= create_phi_node (NULL_TREE
, exit
);
301 create_new_def_for (var
, phi
, gimple_phi_result_ptr (phi
));
302 FOR_EACH_EDGE (e
, ei
, exit
->preds
)
303 add_phi_arg (phi
, var
, e
, UNKNOWN_LOCATION
);
305 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
307 fprintf (dump_file
, ";; Created LCSSA PHI: ");
308 print_gimple_stmt (dump_file
, phi
, 0, dump_flags
);
312 /* Add exit phis for VAR that is used in LIVEIN.
313 Exits of the loops are stored in LOOP_EXITS. */
316 add_exit_phis_var (tree var
, bitmap use_blocks
, bitmap
*loop_exits
)
320 basic_block def_bb
= gimple_bb (SSA_NAME_DEF_STMT (var
));
321 bitmap live_exits
= BITMAP_ALLOC (&loop_renamer_obstack
);
323 gcc_checking_assert (! bitmap_bit_p (use_blocks
, def_bb
->index
));
325 compute_live_loop_exits (live_exits
, use_blocks
, loop_exits
, def_bb
);
327 EXECUTE_IF_SET_IN_BITMAP (live_exits
, 0, index
, bi
)
329 add_exit_phi (BASIC_BLOCK_FOR_FN (cfun
, index
), var
);
332 BITMAP_FREE (live_exits
);
335 /* Add exit phis for the names marked in NAMES_TO_RENAME.
336 Exits of the loops are stored in EXITS. Sets of blocks where the ssa
337 names are used are stored in USE_BLOCKS. */
340 add_exit_phis (bitmap names_to_rename
, bitmap
*use_blocks
, bitmap
*loop_exits
)
345 EXECUTE_IF_SET_IN_BITMAP (names_to_rename
, 0, i
, bi
)
347 add_exit_phis_var (ssa_name (i
), use_blocks
[i
], loop_exits
);
351 /* Fill the array of bitmaps LOOP_EXITS with all loop exit edge targets. */
354 get_loops_exits (bitmap
*loop_exits
)
360 FOR_EACH_LOOP (loop
, 0)
362 vec
<edge
> exit_edges
= get_loop_exit_edges (loop
);
363 loop_exits
[loop
->num
] = BITMAP_ALLOC (&loop_renamer_obstack
);
364 FOR_EACH_VEC_ELT (exit_edges
, j
, e
)
365 bitmap_set_bit (loop_exits
[loop
->num
], e
->dest
->index
);
366 exit_edges
.release ();
370 /* For USE in BB, if it is used outside of the loop it is defined in,
371 mark it for rewrite. Record basic block BB where it is used
372 to USE_BLOCKS. Record the ssa name index to NEED_PHIS bitmap. */
375 find_uses_to_rename_use (basic_block bb
, tree use
, bitmap
*use_blocks
,
380 struct loop
*def_loop
;
382 if (TREE_CODE (use
) != SSA_NAME
)
385 ver
= SSA_NAME_VERSION (use
);
386 def_bb
= gimple_bb (SSA_NAME_DEF_STMT (use
));
389 def_loop
= def_bb
->loop_father
;
391 /* If the definition is not inside a loop, it is not interesting. */
392 if (!loop_outer (def_loop
))
395 /* If the use is not outside of the loop it is defined in, it is not
397 if (flow_bb_inside_loop_p (def_loop
, bb
))
400 /* If we're seeing VER for the first time, we still have to allocate
401 a bitmap for its uses. */
402 if (bitmap_set_bit (need_phis
, ver
))
403 use_blocks
[ver
] = BITMAP_ALLOC (&loop_renamer_obstack
);
404 bitmap_set_bit (use_blocks
[ver
], bb
->index
);
407 /* For uses in STMT, mark names that are used outside of the loop they are
408 defined to rewrite. Record the set of blocks in that the ssa
409 names are defined to USE_BLOCKS and the ssa names themselves to
413 find_uses_to_rename_stmt (gimple stmt
, bitmap
*use_blocks
, bitmap need_phis
)
417 basic_block bb
= gimple_bb (stmt
);
419 if (is_gimple_debug (stmt
))
422 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, SSA_OP_USE
)
423 find_uses_to_rename_use (bb
, var
, use_blocks
, need_phis
);
426 /* Marks names that are used in BB and outside of the loop they are
427 defined in for rewrite. Records the set of blocks in that the ssa
428 names are defined to USE_BLOCKS. Record the SSA names that will
429 need exit PHIs in NEED_PHIS. */
432 find_uses_to_rename_bb (basic_block bb
, bitmap
*use_blocks
, bitmap need_phis
)
437 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
438 for (gphi_iterator bsi
= gsi_start_phis (e
->dest
); !gsi_end_p (bsi
);
441 gphi
*phi
= bsi
.phi ();
442 if (! virtual_operand_p (gimple_phi_result (phi
)))
443 find_uses_to_rename_use (bb
, PHI_ARG_DEF_FROM_EDGE (phi
, e
),
444 use_blocks
, need_phis
);
447 for (gimple_stmt_iterator bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
);
449 find_uses_to_rename_stmt (gsi_stmt (bsi
), use_blocks
, need_phis
);
452 /* Marks names that are used outside of the loop they are defined in
453 for rewrite. Records the set of blocks in that the ssa
454 names are defined to USE_BLOCKS. If CHANGED_BBS is not NULL,
455 scan only blocks in this set. */
458 find_uses_to_rename (bitmap changed_bbs
, bitmap
*use_blocks
, bitmap need_phis
)
465 EXECUTE_IF_SET_IN_BITMAP (changed_bbs
, 0, index
, bi
)
466 find_uses_to_rename_bb (BASIC_BLOCK_FOR_FN (cfun
, index
), use_blocks
, need_phis
);
468 FOR_EACH_BB_FN (bb
, cfun
)
469 find_uses_to_rename_bb (bb
, use_blocks
, need_phis
);
472 /* Rewrites the program into a loop closed ssa form -- i.e. inserts extra
473 phi nodes to ensure that no variable is used outside the loop it is
476 This strengthening of the basic ssa form has several advantages:
478 1) Updating it during unrolling/peeling/versioning is trivial, since
479 we do not need to care about the uses outside of the loop.
480 The same applies to virtual operands which are also rewritten into
481 loop closed SSA form. Note that virtual operands are always live
483 2) The behavior of all uses of an induction variable is the same.
484 Without this, you need to distinguish the case when the variable
485 is used outside of the loop it is defined in, for example
487 for (i = 0; i < 100; i++)
489 for (j = 0; j < 100; j++)
497 Looking from the outer loop with the normal SSA form, the first use of k
498 is not well-behaved, while the second one is an induction variable with
501 If CHANGED_BBS is not NULL, we look for uses outside loops only in
502 the basic blocks in this set.
504 UPDATE_FLAG is used in the call to update_ssa. See
505 TODO_update_ssa* for documentation. */
508 rewrite_into_loop_closed_ssa (bitmap changed_bbs
, unsigned update_flag
)
511 bitmap names_to_rename
;
513 loops_state_set (LOOP_CLOSED_SSA
);
514 if (number_of_loops (cfun
) <= 1)
517 /* If the pass has caused the SSA form to be out-of-date, update it
519 update_ssa (update_flag
);
521 bitmap_obstack_initialize (&loop_renamer_obstack
);
523 names_to_rename
= BITMAP_ALLOC (&loop_renamer_obstack
);
525 /* Uses of names to rename. We don't have to initialize this array,
526 because we know that we will only have entries for the SSA names
527 in NAMES_TO_RENAME. */
528 use_blocks
= XNEWVEC (bitmap
, num_ssa_names
);
530 /* Find the uses outside loops. */
531 find_uses_to_rename (changed_bbs
, use_blocks
, names_to_rename
);
533 if (!bitmap_empty_p (names_to_rename
))
535 /* An array of bitmaps where LOOP_EXITS[I] is the set of basic blocks
536 that are the destination of an edge exiting loop number I. */
537 bitmap
*loop_exits
= XNEWVEC (bitmap
, number_of_loops (cfun
));
538 get_loops_exits (loop_exits
);
540 /* Add the PHI nodes on exits of the loops for the names we need to
542 add_exit_phis (names_to_rename
, use_blocks
, loop_exits
);
546 /* Fix up all the names found to be used outside their original
548 update_ssa (TODO_update_ssa
);
551 bitmap_obstack_release (&loop_renamer_obstack
);
555 /* Check invariants of the loop closed ssa form for the USE in BB. */
558 check_loop_closed_ssa_use (basic_block bb
, tree use
)
563 if (TREE_CODE (use
) != SSA_NAME
|| virtual_operand_p (use
))
566 def
= SSA_NAME_DEF_STMT (use
);
567 def_bb
= gimple_bb (def
);
569 || flow_bb_inside_loop_p (def_bb
->loop_father
, bb
));
572 /* Checks invariants of loop closed ssa form in statement STMT in BB. */
575 check_loop_closed_ssa_stmt (basic_block bb
, gimple stmt
)
580 if (is_gimple_debug (stmt
))
583 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, SSA_OP_USE
)
584 check_loop_closed_ssa_use (bb
, var
);
587 /* Checks that invariants of the loop closed ssa form are preserved.
588 Call verify_ssa when VERIFY_SSA_P is true. */
591 verify_loop_closed_ssa (bool verify_ssa_p
)
597 if (number_of_loops (cfun
) <= 1)
601 verify_ssa (false, true);
603 timevar_push (TV_VERIFY_LOOP_CLOSED
);
605 FOR_EACH_BB_FN (bb
, cfun
)
607 for (gphi_iterator bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
);
610 gphi
*phi
= bsi
.phi ();
611 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
612 check_loop_closed_ssa_use (e
->src
,
613 PHI_ARG_DEF_FROM_EDGE (phi
, e
));
616 for (gimple_stmt_iterator bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
);
618 check_loop_closed_ssa_stmt (bb
, gsi_stmt (bsi
));
621 timevar_pop (TV_VERIFY_LOOP_CLOSED
);
624 /* Split loop exit edge EXIT. The things are a bit complicated by a need to
625 preserve the loop closed ssa form. The newly created block is returned. */
628 split_loop_exit_edge (edge exit
)
630 basic_block dest
= exit
->dest
;
631 basic_block bb
= split_edge (exit
);
636 source_location locus
;
638 for (psi
= gsi_start_phis (dest
); !gsi_end_p (psi
); gsi_next (&psi
))
641 op_p
= PHI_ARG_DEF_PTR_FROM_EDGE (phi
, single_succ_edge (bb
));
642 locus
= gimple_phi_arg_location_from_edge (phi
, single_succ_edge (bb
));
644 name
= USE_FROM_PTR (op_p
);
646 /* If the argument of the PHI node is a constant, we do not need
647 to keep it inside loop. */
648 if (TREE_CODE (name
) != SSA_NAME
)
651 /* Otherwise create an auxiliary phi node that will copy the value
652 of the SSA name out of the loop. */
653 new_name
= duplicate_ssa_name (name
, NULL
);
654 new_phi
= create_phi_node (new_name
, bb
);
655 add_phi_arg (new_phi
, name
, exit
, locus
);
656 SET_USE (op_p
, new_name
);
662 /* Returns the basic block in that statements should be emitted for induction
663 variables incremented at the end of the LOOP. */
666 ip_end_pos (struct loop
*loop
)
671 /* Returns the basic block in that statements should be emitted for induction
672 variables incremented just before exit condition of a LOOP. */
675 ip_normal_pos (struct loop
*loop
)
681 if (!single_pred_p (loop
->latch
))
684 bb
= single_pred (loop
->latch
);
685 last
= last_stmt (bb
);
687 || gimple_code (last
) != GIMPLE_COND
)
690 exit
= EDGE_SUCC (bb
, 0);
691 if (exit
->dest
== loop
->latch
)
692 exit
= EDGE_SUCC (bb
, 1);
694 if (flow_bb_inside_loop_p (loop
, exit
->dest
))
700 /* Stores the standard position for induction variable increment in LOOP
701 (just before the exit condition if it is available and latch block is empty,
702 end of the latch block otherwise) to BSI. INSERT_AFTER is set to true if
703 the increment should be inserted after *BSI. */
706 standard_iv_increment_position (struct loop
*loop
, gimple_stmt_iterator
*bsi
,
709 basic_block bb
= ip_normal_pos (loop
), latch
= ip_end_pos (loop
);
710 gimple last
= last_stmt (latch
);
713 || (last
&& gimple_code (last
) != GIMPLE_LABEL
))
715 *bsi
= gsi_last_bb (latch
);
716 *insert_after
= true;
720 *bsi
= gsi_last_bb (bb
);
721 *insert_after
= false;
725 /* Copies phi node arguments for duplicated blocks. The index of the first
726 duplicated block is FIRST_NEW_BLOCK. */
729 copy_phi_node_args (unsigned first_new_block
)
733 for (i
= first_new_block
; i
< (unsigned) last_basic_block_for_fn (cfun
); i
++)
734 BASIC_BLOCK_FOR_FN (cfun
, i
)->flags
|= BB_DUPLICATED
;
736 for (i
= first_new_block
; i
< (unsigned) last_basic_block_for_fn (cfun
); i
++)
737 add_phi_args_after_copy_bb (BASIC_BLOCK_FOR_FN (cfun
, i
));
739 for (i
= first_new_block
; i
< (unsigned) last_basic_block_for_fn (cfun
); i
++)
740 BASIC_BLOCK_FOR_FN (cfun
, i
)->flags
&= ~BB_DUPLICATED
;
744 /* The same as cfgloopmanip.c:duplicate_loop_to_header_edge, but also
745 updates the PHI nodes at start of the copied region. In order to
746 achieve this, only loops whose exits all lead to the same location
749 Notice that we do not completely update the SSA web after
750 duplication. The caller is responsible for calling update_ssa
751 after the loop has been duplicated. */
754 gimple_duplicate_loop_to_header_edge (struct loop
*loop
, edge e
,
755 unsigned int ndupl
, sbitmap wont_exit
,
756 edge orig
, vec
<edge
> *to_remove
,
759 unsigned first_new_block
;
761 if (!loops_state_satisfies_p (LOOPS_HAVE_SIMPLE_LATCHES
))
763 if (!loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS
))
766 first_new_block
= last_basic_block_for_fn (cfun
);
767 if (!duplicate_loop_to_header_edge (loop
, e
, ndupl
, wont_exit
,
768 orig
, to_remove
, flags
))
771 /* Readd the removed phi args for e. */
772 flush_pending_stmts (e
);
774 /* Copy the phi node arguments. */
775 copy_phi_node_args (first_new_block
);
782 /* Returns true if we can unroll LOOP FACTOR times. Number
783 of iterations of the loop is returned in NITER. */
786 can_unroll_loop_p (struct loop
*loop
, unsigned factor
,
787 struct tree_niter_desc
*niter
)
791 /* Check whether unrolling is possible. We only want to unroll loops
792 for that we are able to determine number of iterations. We also
793 want to split the extra iterations of the loop from its end,
794 therefore we require that the loop has precisely one
797 exit
= single_dom_exit (loop
);
801 if (!number_of_iterations_exit (loop
, exit
, niter
, false)
802 || niter
->cmp
== ERROR_MARK
803 /* Scalar evolutions analysis might have copy propagated
804 the abnormal ssa names into these expressions, hence
805 emitting the computations based on them during loop
806 unrolling might create overlapping life ranges for
807 them, and failures in out-of-ssa. */
808 || contains_abnormal_ssa_name_p (niter
->may_be_zero
)
809 || contains_abnormal_ssa_name_p (niter
->control
.base
)
810 || contains_abnormal_ssa_name_p (niter
->control
.step
)
811 || contains_abnormal_ssa_name_p (niter
->bound
))
814 /* And of course, we must be able to duplicate the loop. */
815 if (!can_duplicate_loop_p (loop
))
818 /* The final loop should be small enough. */
819 if (tree_num_loop_insns (loop
, &eni_size_weights
) * factor
820 > (unsigned) PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS
))
826 /* Determines the conditions that control execution of LOOP unrolled FACTOR
827 times. DESC is number of iterations of LOOP. ENTER_COND is set to
828 condition that must be true if the main loop can be entered.
829 EXIT_BASE, EXIT_STEP, EXIT_CMP and EXIT_BOUND are set to values describing
830 how the exit from the unrolled loop should be controlled. */
833 determine_exit_conditions (struct loop
*loop
, struct tree_niter_desc
*desc
,
834 unsigned factor
, tree
*enter_cond
,
835 tree
*exit_base
, tree
*exit_step
,
836 enum tree_code
*exit_cmp
, tree
*exit_bound
)
839 tree base
= desc
->control
.base
;
840 tree step
= desc
->control
.step
;
841 tree bound
= desc
->bound
;
842 tree type
= TREE_TYPE (step
);
844 tree min
= lower_bound_in_type (type
, type
);
845 tree max
= upper_bound_in_type (type
, type
);
846 enum tree_code cmp
= desc
->cmp
;
847 tree cond
= boolean_true_node
, assum
;
849 /* For pointers, do the arithmetics in the type of step. */
850 base
= fold_convert (type
, base
);
851 bound
= fold_convert (type
, bound
);
853 *enter_cond
= boolean_false_node
;
854 *exit_base
= NULL_TREE
;
855 *exit_step
= NULL_TREE
;
856 *exit_cmp
= ERROR_MARK
;
857 *exit_bound
= NULL_TREE
;
858 gcc_assert (cmp
!= ERROR_MARK
);
860 /* We only need to be correct when we answer question
861 "Do at least FACTOR more iterations remain?" in the unrolled loop.
862 Thus, transforming BASE + STEP * i <> BOUND to
863 BASE + STEP * i < BOUND is ok. */
866 if (tree_int_cst_sign_bit (step
))
871 else if (cmp
== LT_EXPR
)
873 gcc_assert (!tree_int_cst_sign_bit (step
));
875 else if (cmp
== GT_EXPR
)
877 gcc_assert (tree_int_cst_sign_bit (step
));
882 /* The main body of the loop may be entered iff:
884 1) desc->may_be_zero is false.
885 2) it is possible to check that there are at least FACTOR iterations
886 of the loop, i.e., BOUND - step * FACTOR does not overflow.
887 3) # of iterations is at least FACTOR */
889 if (!integer_zerop (desc
->may_be_zero
))
890 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
891 invert_truthvalue (desc
->may_be_zero
),
894 bigstep
= fold_build2 (MULT_EXPR
, type
, step
,
895 build_int_cst_type (type
, factor
));
896 delta
= fold_build2 (MINUS_EXPR
, type
, bigstep
, step
);
898 assum
= fold_build2 (GE_EXPR
, boolean_type_node
,
900 fold_build2 (PLUS_EXPR
, type
, min
, delta
));
902 assum
= fold_build2 (LE_EXPR
, boolean_type_node
,
904 fold_build2 (PLUS_EXPR
, type
, max
, delta
));
905 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
, assum
, cond
);
907 bound
= fold_build2 (MINUS_EXPR
, type
, bound
, delta
);
908 assum
= fold_build2 (cmp
, boolean_type_node
, base
, bound
);
909 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
, assum
, cond
);
911 cond
= force_gimple_operand (unshare_expr (cond
), &stmts
, false, NULL_TREE
);
913 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
914 /* cond now may be a gimple comparison, which would be OK, but also any
915 other gimple rhs (say a && b). In this case we need to force it to
917 if (!is_gimple_condexpr (cond
))
919 cond
= force_gimple_operand (cond
, &stmts
, true, NULL_TREE
);
921 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
925 base
= force_gimple_operand (unshare_expr (base
), &stmts
, true, NULL_TREE
);
927 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
928 bound
= force_gimple_operand (unshare_expr (bound
), &stmts
, true, NULL_TREE
);
930 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
933 *exit_step
= bigstep
;
938 /* Scales the frequencies of all basic blocks in LOOP that are strictly
939 dominated by BB by NUM/DEN. */
942 scale_dominated_blocks_in_loop (struct loop
*loop
, basic_block bb
,
950 for (son
= first_dom_son (CDI_DOMINATORS
, bb
);
952 son
= next_dom_son (CDI_DOMINATORS
, son
))
954 if (!flow_bb_inside_loop_p (loop
, son
))
956 scale_bbs_frequencies_int (&son
, 1, num
, den
);
957 scale_dominated_blocks_in_loop (loop
, son
, num
, den
);
961 /* Unroll LOOP FACTOR times. DESC describes number of iterations of LOOP.
962 EXIT is the exit of the loop to that DESC corresponds.
964 If N is number of iterations of the loop and MAY_BE_ZERO is the condition
965 under that loop exits in the first iteration even if N != 0,
969 x = phi (init, next);
977 becomes (with possibly the exit conditions formulated a bit differently,
978 avoiding the need to create a new iv):
980 if (MAY_BE_ZERO || N < FACTOR)
985 x = phi (init, next);
996 } while (N >= FACTOR);
999 init' = phi (init, x);
1003 x = phi (init', next);
1011 Before the loop is unrolled, TRANSFORM is called for it (only for the
1012 unrolled loop, but not for its versioned copy). DATA is passed to
1015 /* Probability in % that the unrolled loop is entered. Just a guess. */
1016 #define PROB_UNROLLED_LOOP_ENTERED 90
1019 tree_transform_and_unroll_loop (struct loop
*loop
, unsigned factor
,
1020 edge exit
, struct tree_niter_desc
*desc
,
1021 transform_callback transform
,
1025 tree ctr_before
, ctr_after
;
1026 tree enter_main_cond
, exit_base
, exit_step
, exit_bound
;
1027 enum tree_code exit_cmp
;
1028 gphi
*phi_old_loop
, *phi_new_loop
, *phi_rest
;
1029 gphi_iterator psi_old_loop
, psi_new_loop
;
1030 tree init
, next
, new_init
;
1031 struct loop
*new_loop
;
1032 basic_block rest
, exit_bb
;
1033 edge old_entry
, new_entry
, old_latch
, precond_edge
, new_exit
;
1034 edge new_nonexit
, e
;
1035 gimple_stmt_iterator bsi
;
1038 unsigned est_niter
, prob_entry
, scale_unrolled
, scale_rest
, freq_e
, freq_h
;
1039 unsigned new_est_niter
, i
, prob
;
1040 unsigned irr
= loop_preheader_edge (loop
)->flags
& EDGE_IRREDUCIBLE_LOOP
;
1042 auto_vec
<edge
> to_remove
;
1044 est_niter
= expected_loop_iterations (loop
);
1045 determine_exit_conditions (loop
, desc
, factor
,
1046 &enter_main_cond
, &exit_base
, &exit_step
,
1047 &exit_cmp
, &exit_bound
);
1049 /* Let us assume that the unrolled loop is quite likely to be entered. */
1050 if (integer_nonzerop (enter_main_cond
))
1051 prob_entry
= REG_BR_PROB_BASE
;
1053 prob_entry
= PROB_UNROLLED_LOOP_ENTERED
* REG_BR_PROB_BASE
/ 100;
1055 /* The values for scales should keep profile consistent, and somewhat close
1058 TODO: The current value of SCALE_REST makes it appear that the loop that
1059 is created by splitting the remaining iterations of the unrolled loop is
1060 executed the same number of times as the original loop, and with the same
1061 frequencies, which is obviously wrong. This does not appear to cause
1062 problems, so we do not bother with fixing it for now. To make the profile
1063 correct, we would need to change the probability of the exit edge of the
1064 loop, and recompute the distribution of frequencies in its body because
1065 of this change (scale the frequencies of blocks before and after the exit
1066 by appropriate factors). */
1067 scale_unrolled
= prob_entry
;
1068 scale_rest
= REG_BR_PROB_BASE
;
1070 new_loop
= loop_version (loop
, enter_main_cond
, NULL
,
1071 prob_entry
, scale_unrolled
, scale_rest
, true);
1072 gcc_assert (new_loop
!= NULL
);
1073 update_ssa (TODO_update_ssa
);
1075 /* Determine the probability of the exit edge of the unrolled loop. */
1076 new_est_niter
= est_niter
/ factor
;
1078 /* Without profile feedback, loops for that we do not know a better estimate
1079 are assumed to roll 10 times. When we unroll such loop, it appears to
1080 roll too little, and it may even seem to be cold. To avoid this, we
1081 ensure that the created loop appears to roll at least 5 times (but at
1082 most as many times as before unrolling). */
1083 if (new_est_niter
< 5)
1086 new_est_niter
= est_niter
;
1091 /* Prepare the cfg and update the phi nodes. Move the loop exit to the
1092 loop latch (and make its condition dummy, for the moment). */
1093 rest
= loop_preheader_edge (new_loop
)->src
;
1094 precond_edge
= single_pred_edge (rest
);
1095 split_edge (loop_latch_edge (loop
));
1096 exit_bb
= single_pred (loop
->latch
);
1098 /* Since the exit edge will be removed, the frequency of all the blocks
1099 in the loop that are dominated by it must be scaled by
1100 1 / (1 - exit->probability). */
1101 scale_dominated_blocks_in_loop (loop
, exit
->src
,
1103 REG_BR_PROB_BASE
- exit
->probability
);
1105 bsi
= gsi_last_bb (exit_bb
);
1106 exit_if
= gimple_build_cond (EQ_EXPR
, integer_zero_node
,
1108 NULL_TREE
, NULL_TREE
);
1110 gsi_insert_after (&bsi
, exit_if
, GSI_NEW_STMT
);
1111 new_exit
= make_edge (exit_bb
, rest
, EDGE_FALSE_VALUE
| irr
);
1112 rescan_loop_exit (new_exit
, true, false);
1114 /* Set the probability of new exit to the same of the old one. Fix
1115 the frequency of the latch block, by scaling it back by
1116 1 - exit->probability. */
1117 new_exit
->count
= exit
->count
;
1118 new_exit
->probability
= exit
->probability
;
1119 new_nonexit
= single_pred_edge (loop
->latch
);
1120 new_nonexit
->probability
= REG_BR_PROB_BASE
- exit
->probability
;
1121 new_nonexit
->flags
= EDGE_TRUE_VALUE
;
1122 new_nonexit
->count
-= exit
->count
;
1123 if (new_nonexit
->count
< 0)
1124 new_nonexit
->count
= 0;
1125 scale_bbs_frequencies_int (&loop
->latch
, 1, new_nonexit
->probability
,
1128 old_entry
= loop_preheader_edge (loop
);
1129 new_entry
= loop_preheader_edge (new_loop
);
1130 old_latch
= loop_latch_edge (loop
);
1131 for (psi_old_loop
= gsi_start_phis (loop
->header
),
1132 psi_new_loop
= gsi_start_phis (new_loop
->header
);
1133 !gsi_end_p (psi_old_loop
);
1134 gsi_next (&psi_old_loop
), gsi_next (&psi_new_loop
))
1136 phi_old_loop
= psi_old_loop
.phi ();
1137 phi_new_loop
= psi_new_loop
.phi ();
1139 init
= PHI_ARG_DEF_FROM_EDGE (phi_old_loop
, old_entry
);
1140 op
= PHI_ARG_DEF_PTR_FROM_EDGE (phi_new_loop
, new_entry
);
1141 gcc_assert (operand_equal_for_phi_arg_p (init
, USE_FROM_PTR (op
)));
1142 next
= PHI_ARG_DEF_FROM_EDGE (phi_old_loop
, old_latch
);
1144 /* Prefer using original variable as a base for the new ssa name.
1145 This is necessary for virtual ops, and useful in order to avoid
1146 losing debug info for real ops. */
1147 if (TREE_CODE (next
) == SSA_NAME
1148 && useless_type_conversion_p (TREE_TYPE (next
),
1150 new_init
= copy_ssa_name (next
, NULL
);
1151 else if (TREE_CODE (init
) == SSA_NAME
1152 && useless_type_conversion_p (TREE_TYPE (init
),
1154 new_init
= copy_ssa_name (init
, NULL
);
1155 else if (useless_type_conversion_p (TREE_TYPE (next
), TREE_TYPE (init
)))
1156 new_init
= make_temp_ssa_name (TREE_TYPE (next
), NULL
, "unrinittmp");
1158 new_init
= make_temp_ssa_name (TREE_TYPE (init
), NULL
, "unrinittmp");
1160 phi_rest
= create_phi_node (new_init
, rest
);
1162 add_phi_arg (phi_rest
, init
, precond_edge
, UNKNOWN_LOCATION
);
1163 add_phi_arg (phi_rest
, next
, new_exit
, UNKNOWN_LOCATION
);
1164 SET_USE (op
, new_init
);
1169 /* Transform the loop. */
1171 (*transform
) (loop
, data
);
1173 /* Unroll the loop and remove the exits in all iterations except for the
1175 wont_exit
= sbitmap_alloc (factor
);
1176 bitmap_ones (wont_exit
);
1177 bitmap_clear_bit (wont_exit
, factor
- 1);
1179 ok
= gimple_duplicate_loop_to_header_edge
1180 (loop
, loop_latch_edge (loop
), factor
- 1,
1181 wont_exit
, new_exit
, &to_remove
, DLTHE_FLAG_UPDATE_FREQ
);
1185 FOR_EACH_VEC_ELT (to_remove
, i
, e
)
1187 ok
= remove_path (e
);
1190 update_ssa (TODO_update_ssa
);
1192 /* Ensure that the frequencies in the loop match the new estimated
1193 number of iterations, and change the probability of the new
1195 freq_h
= loop
->header
->frequency
;
1196 freq_e
= EDGE_FREQUENCY (loop_preheader_edge (loop
));
1198 scale_loop_frequencies (loop
, freq_e
* (new_est_niter
+ 1), freq_h
);
1200 exit_bb
= single_pred (loop
->latch
);
1201 new_exit
= find_edge (exit_bb
, rest
);
1202 new_exit
->count
= loop_preheader_edge (loop
)->count
;
1203 new_exit
->probability
= REG_BR_PROB_BASE
/ (new_est_niter
+ 1);
1205 rest
->count
+= new_exit
->count
;
1206 rest
->frequency
+= EDGE_FREQUENCY (new_exit
);
1208 new_nonexit
= single_pred_edge (loop
->latch
);
1209 prob
= new_nonexit
->probability
;
1210 new_nonexit
->probability
= REG_BR_PROB_BASE
- new_exit
->probability
;
1211 new_nonexit
->count
= exit_bb
->count
- new_exit
->count
;
1212 if (new_nonexit
->count
< 0)
1213 new_nonexit
->count
= 0;
1215 scale_bbs_frequencies_int (&loop
->latch
, 1, new_nonexit
->probability
,
1218 /* Finally create the new counter for number of iterations and add the new
1219 exit instruction. */
1220 bsi
= gsi_last_nondebug_bb (exit_bb
);
1221 exit_if
= as_a
<gcond
*> (gsi_stmt (bsi
));
1222 create_iv (exit_base
, exit_step
, NULL_TREE
, loop
,
1223 &bsi
, false, &ctr_before
, &ctr_after
);
1224 gimple_cond_set_code (exit_if
, exit_cmp
);
1225 gimple_cond_set_lhs (exit_if
, ctr_after
);
1226 gimple_cond_set_rhs (exit_if
, exit_bound
);
1227 update_stmt (exit_if
);
1229 #ifdef ENABLE_CHECKING
1230 verify_flow_info ();
1231 verify_loop_structure ();
1232 verify_loop_closed_ssa (true);
1236 /* Wrapper over tree_transform_and_unroll_loop for case we do not
1237 want to transform the loop before unrolling. The meaning
1238 of the arguments is the same as for tree_transform_and_unroll_loop. */
1241 tree_unroll_loop (struct loop
*loop
, unsigned factor
,
1242 edge exit
, struct tree_niter_desc
*desc
)
1244 tree_transform_and_unroll_loop (loop
, factor
, exit
, desc
,
1248 /* Rewrite the phi node at position PSI in function of the main
1249 induction variable MAIN_IV and insert the generated code at GSI. */
1252 rewrite_phi_with_iv (loop_p loop
,
1254 gimple_stmt_iterator
*gsi
,
1259 gphi
*phi
= psi
->phi ();
1260 tree atype
, mtype
, val
, res
= PHI_RESULT (phi
);
1262 if (virtual_operand_p (res
) || res
== main_iv
)
1268 if (!simple_iv (loop
, loop
, res
, &iv
, true))
1274 remove_phi_node (psi
, false);
1276 atype
= TREE_TYPE (res
);
1277 mtype
= POINTER_TYPE_P (atype
) ? sizetype
: atype
;
1278 val
= fold_build2 (MULT_EXPR
, mtype
, unshare_expr (iv
.step
),
1279 fold_convert (mtype
, main_iv
));
1280 val
= fold_build2 (POINTER_TYPE_P (atype
)
1281 ? POINTER_PLUS_EXPR
: PLUS_EXPR
,
1282 atype
, unshare_expr (iv
.base
), val
);
1283 val
= force_gimple_operand_gsi (gsi
, val
, false, NULL_TREE
, true,
1285 stmt
= gimple_build_assign (res
, val
);
1286 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
1289 /* Rewrite all the phi nodes of LOOP in function of the main induction
1290 variable MAIN_IV. */
1293 rewrite_all_phi_nodes_with_iv (loop_p loop
, tree main_iv
)
1296 basic_block
*bbs
= get_loop_body_in_dom_order (loop
);
1299 for (i
= 0; i
< loop
->num_nodes
; i
++)
1301 basic_block bb
= bbs
[i
];
1302 gimple_stmt_iterator gsi
= gsi_after_labels (bb
);
1304 if (bb
->loop_father
!= loop
)
1307 for (psi
= gsi_start_phis (bb
); !gsi_end_p (psi
); )
1308 rewrite_phi_with_iv (loop
, &psi
, &gsi
, main_iv
);
1314 /* Bases all the induction variables in LOOP on a single induction
1315 variable (unsigned with base 0 and step 1), whose final value is
1316 compared with *NIT. When the IV type precision has to be larger
1317 than *NIT type precision, *NIT is converted to the larger type, the
1318 conversion code is inserted before the loop, and *NIT is updated to
1319 the new definition. When BUMP_IN_LATCH is true, the induction
1320 variable is incremented in the loop latch, otherwise it is
1321 incremented in the loop header. Return the induction variable that
1325 canonicalize_loop_ivs (struct loop
*loop
, tree
*nit
, bool bump_in_latch
)
1327 unsigned precision
= TYPE_PRECISION (TREE_TYPE (*nit
));
1328 unsigned original_precision
= precision
;
1329 tree type
, var_before
;
1330 gimple_stmt_iterator gsi
;
1333 edge exit
= single_dom_exit (loop
);
1335 enum machine_mode mode
;
1336 bool unsigned_p
= false;
1338 for (psi
= gsi_start_phis (loop
->header
);
1339 !gsi_end_p (psi
); gsi_next (&psi
))
1341 gphi
*phi
= psi
.phi ();
1342 tree res
= PHI_RESULT (phi
);
1345 type
= TREE_TYPE (res
);
1346 if (virtual_operand_p (res
)
1347 || (!INTEGRAL_TYPE_P (type
)
1348 && !POINTER_TYPE_P (type
))
1349 || TYPE_PRECISION (type
) < precision
)
1352 uns
= POINTER_TYPE_P (type
) | TYPE_UNSIGNED (type
);
1354 if (TYPE_PRECISION (type
) > precision
)
1359 precision
= TYPE_PRECISION (type
);
1362 mode
= smallest_mode_for_size (precision
, MODE_INT
);
1363 precision
= GET_MODE_PRECISION (mode
);
1364 type
= build_nonstandard_integer_type (precision
, unsigned_p
);
1366 if (original_precision
!= precision
)
1368 *nit
= fold_convert (type
, *nit
);
1369 *nit
= force_gimple_operand (*nit
, &stmts
, true, NULL_TREE
);
1371 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
1375 gsi
= gsi_last_bb (loop
->latch
);
1377 gsi
= gsi_last_nondebug_bb (loop
->header
);
1378 create_iv (build_int_cst_type (type
, 0), build_int_cst (type
, 1), NULL_TREE
,
1379 loop
, &gsi
, bump_in_latch
, &var_before
, NULL
);
1381 rewrite_all_phi_nodes_with_iv (loop
, var_before
);
1383 stmt
= as_a
<gcond
*> (last_stmt (exit
->src
));
1384 /* Make the loop exit if the control condition is not satisfied. */
1385 if (exit
->flags
& EDGE_TRUE_VALUE
)
1389 extract_true_false_edges_from_block (exit
->src
, &te
, &fe
);
1390 te
->flags
= EDGE_FALSE_VALUE
;
1391 fe
->flags
= EDGE_TRUE_VALUE
;
1393 gimple_cond_set_code (stmt
, LT_EXPR
);
1394 gimple_cond_set_lhs (stmt
, var_before
);
1395 gimple_cond_set_rhs (stmt
, *nit
);