1 /* High-level loop manipulation functions.
2 Copyright (C) 2004-2015 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by the
8 Free Software Foundation; either version 3, or (at your option) any
11 GCC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
22 #include "coretypes.h"
27 #include "hard-reg-set.h"
30 #include "fold-const.h"
33 #include "internal-fn.h"
35 #include "gimple-iterator.h"
36 #include "gimplify-me.h"
38 #include "tree-ssa-loop-ivopts.h"
39 #include "tree-ssa-loop-manip.h"
40 #include "tree-ssa-loop-niter.h"
41 #include "tree-ssa-loop.h"
42 #include "tree-into-ssa.h"
45 #include "gimple-pretty-print.h"
47 #include "tree-pass.h" /* ??? for TODO_update_ssa but this isn't a pass. */
48 #include "tree-scalar-evolution.h"
50 #include "tree-inline.h"
51 #include "langhooks.h"
53 /* All bitmaps for rewriting into loop-closed SSA go on this obstack,
54 so that we can free them all at once. */
55 static bitmap_obstack loop_renamer_obstack
;
57 /* Creates an induction variable with value BASE + STEP * iteration in LOOP.
58 It is expected that neither BASE nor STEP are shared with other expressions
59 (unless the sharing rules allow this). Use VAR as a base var_decl for it
60 (if NULL, a new temporary will be created). The increment will occur at
61 INCR_POS (after it if AFTER is true, before it otherwise). INCR_POS and
62 AFTER can be computed using standard_iv_increment_position. The ssa versions
63 of the variable before and after increment will be stored in VAR_BEFORE and
64 VAR_AFTER (unless they are NULL). */
67 create_iv (tree base
, tree step
, tree var
, struct loop
*loop
,
68 gimple_stmt_iterator
*incr_pos
, bool after
,
69 tree
*var_before
, tree
*var_after
)
76 enum tree_code incr_op
= PLUS_EXPR
;
77 edge pe
= loop_preheader_edge (loop
);
81 vb
= make_ssa_name (var
);
82 va
= make_ssa_name (var
);
86 vb
= make_temp_ssa_name (TREE_TYPE (base
), NULL
, "ivtmp");
87 va
= make_temp_ssa_name (TREE_TYPE (base
), NULL
, "ivtmp");
94 /* For easier readability of the created code, produce MINUS_EXPRs
96 if (TREE_CODE (step
) == INTEGER_CST
)
98 if (TYPE_UNSIGNED (TREE_TYPE (step
)))
100 step1
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
101 if (tree_int_cst_lt (step1
, step
))
103 incr_op
= MINUS_EXPR
;
111 if (!tree_expr_nonnegative_warnv_p (step
, &ovf
)
112 && may_negate_without_overflow_p (step
))
114 incr_op
= MINUS_EXPR
;
115 step
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
119 if (POINTER_TYPE_P (TREE_TYPE (base
)))
121 if (TREE_CODE (base
) == ADDR_EXPR
)
122 mark_addressable (TREE_OPERAND (base
, 0));
123 step
= convert_to_ptrofftype (step
);
124 if (incr_op
== MINUS_EXPR
)
125 step
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
126 incr_op
= POINTER_PLUS_EXPR
;
128 /* Gimplify the step if necessary. We put the computations in front of the
129 loop (i.e. the step should be loop invariant). */
130 step
= force_gimple_operand (step
, &stmts
, true, NULL_TREE
);
132 gsi_insert_seq_on_edge_immediate (pe
, stmts
);
134 stmt
= gimple_build_assign (va
, incr_op
, vb
, step
);
136 gsi_insert_after (incr_pos
, stmt
, GSI_NEW_STMT
);
138 gsi_insert_before (incr_pos
, stmt
, GSI_NEW_STMT
);
140 initial
= force_gimple_operand (base
, &stmts
, true, var
);
142 gsi_insert_seq_on_edge_immediate (pe
, stmts
);
144 phi
= create_phi_node (vb
, loop
->header
);
145 add_phi_arg (phi
, initial
, loop_preheader_edge (loop
), UNKNOWN_LOCATION
);
146 add_phi_arg (phi
, va
, loop_latch_edge (loop
), UNKNOWN_LOCATION
);
149 /* Return the innermost superloop LOOP of USE_LOOP that is a superloop of
150 both DEF_LOOP and USE_LOOP. */
152 static inline struct loop
*
153 find_sibling_superloop (struct loop
*use_loop
, struct loop
*def_loop
)
155 unsigned ud
= loop_depth (use_loop
);
156 unsigned dd
= loop_depth (def_loop
);
157 gcc_assert (ud
> 0 && dd
> 0);
159 use_loop
= superloop_at_depth (use_loop
, dd
);
161 def_loop
= superloop_at_depth (def_loop
, ud
);
162 while (loop_outer (use_loop
) != loop_outer (def_loop
))
164 use_loop
= loop_outer (use_loop
);
165 def_loop
= loop_outer (def_loop
);
166 gcc_assert (use_loop
&& def_loop
);
171 /* DEF_BB is a basic block containing a DEF that needs rewriting into
172 loop-closed SSA form. USE_BLOCKS is the set of basic blocks containing
173 uses of DEF that "escape" from the loop containing DEF_BB (i.e. blocks in
174 USE_BLOCKS are dominated by DEF_BB but not in the loop father of DEF_B).
175 ALL_EXITS[I] is the set of all basic blocks that exit loop I.
177 Compute the subset of LOOP_EXITS that exit the loop containing DEF_BB
178 or one of its loop fathers, in which DEF is live. This set is returned
179 in the bitmap LIVE_EXITS.
181 Instead of computing the complete livein set of the def, we use the loop
182 nesting tree as a form of poor man's structure analysis. This greatly
183 speeds up the analysis, which is important because this function may be
184 called on all SSA names that need rewriting, one at a time. */
187 compute_live_loop_exits (bitmap live_exits
, bitmap use_blocks
,
188 bitmap
*loop_exits
, basic_block def_bb
)
192 struct loop
*def_loop
= def_bb
->loop_father
;
193 unsigned def_loop_depth
= loop_depth (def_loop
);
194 bitmap def_loop_exits
;
196 /* Normally the work list size is bounded by the number of basic
197 blocks in the largest loop. We don't know this number, but we
198 can be fairly sure that it will be relatively small. */
199 auto_vec
<basic_block
> worklist (MAX (8, n_basic_blocks_for_fn (cfun
) / 128));
201 EXECUTE_IF_SET_IN_BITMAP (use_blocks
, 0, i
, bi
)
203 basic_block use_bb
= BASIC_BLOCK_FOR_FN (cfun
, i
);
204 struct loop
*use_loop
= use_bb
->loop_father
;
205 gcc_checking_assert (def_loop
!= use_loop
206 && ! flow_loop_nested_p (def_loop
, use_loop
));
207 if (! flow_loop_nested_p (use_loop
, def_loop
))
208 use_bb
= find_sibling_superloop (use_loop
, def_loop
)->header
;
209 if (bitmap_set_bit (live_exits
, use_bb
->index
))
210 worklist
.safe_push (use_bb
);
213 /* Iterate until the worklist is empty. */
214 while (! worklist
.is_empty ())
219 /* Pull a block off the worklist. */
220 basic_block bb
= worklist
.pop ();
222 /* Make sure we have at least enough room in the work list
223 for all predecessors of this block. */
224 worklist
.reserve (EDGE_COUNT (bb
->preds
));
226 /* For each predecessor block. */
227 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
229 basic_block pred
= e
->src
;
230 struct loop
*pred_loop
= pred
->loop_father
;
231 unsigned pred_loop_depth
= loop_depth (pred_loop
);
234 /* We should have met DEF_BB along the way. */
235 gcc_assert (pred
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
));
237 if (pred_loop_depth
>= def_loop_depth
)
239 if (pred_loop_depth
> def_loop_depth
)
240 pred_loop
= superloop_at_depth (pred_loop
, def_loop_depth
);
241 /* If we've reached DEF_LOOP, our train ends here. */
242 if (pred_loop
== def_loop
)
245 else if (! flow_loop_nested_p (pred_loop
, def_loop
))
246 pred
= find_sibling_superloop (pred_loop
, def_loop
)->header
;
248 /* Add PRED to the LIVEIN set. PRED_VISITED is true if
249 we had already added PRED to LIVEIN before. */
250 pred_visited
= !bitmap_set_bit (live_exits
, pred
->index
);
252 /* If we have visited PRED before, don't add it to the worklist.
253 If BB dominates PRED, then we're probably looking at a loop.
254 We're only interested in looking up in the dominance tree
255 because DEF_BB dominates all the uses. */
256 if (pred_visited
|| dominated_by_p (CDI_DOMINATORS
, pred
, bb
))
259 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 /* Check that at least one of the edges entering the EXIT block exits
282 the loop, or a superloop of that loop, that VAR is defined in. */
285 gimple
*def_stmt
= SSA_NAME_DEF_STMT (var
);
286 basic_block def_bb
= gimple_bb (def_stmt
);
287 FOR_EACH_EDGE (e
, ei
, exit
->preds
)
289 struct loop
*aloop
= find_common_loop (def_bb
->loop_father
,
290 e
->src
->loop_father
);
291 if (!flow_bb_inside_loop_p (aloop
, e
->dest
))
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_FOR_FN (cfun
, 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
)
357 FOR_EACH_LOOP (loop
, 0)
359 vec
<edge
> exit_edges
= get_loop_exit_edges (loop
);
360 loop_exits
[loop
->num
] = BITMAP_ALLOC (&loop_renamer_obstack
);
361 FOR_EACH_VEC_ELT (exit_edges
, j
, e
)
362 bitmap_set_bit (loop_exits
[loop
->num
], e
->dest
->index
);
363 exit_edges
.release ();
367 /* For USE in BB, if it is used outside of the loop it is defined in,
368 mark it for rewrite. Record basic block BB where it is used
369 to USE_BLOCKS. Record the ssa name index to NEED_PHIS bitmap.
370 Note that for USEs in phis, BB should be the src of the edge corresponding to
371 the use, rather than the bb containing the phi. */
374 find_uses_to_rename_use (basic_block bb
, tree use
, bitmap
*use_blocks
,
379 struct loop
*def_loop
;
381 if (TREE_CODE (use
) != SSA_NAME
)
384 ver
= SSA_NAME_VERSION (use
);
385 def_bb
= gimple_bb (SSA_NAME_DEF_STMT (use
));
388 def_loop
= def_bb
->loop_father
;
390 /* If the definition is not inside a loop, it is not interesting. */
391 if (!loop_outer (def_loop
))
394 /* If the use is not outside of the loop it is defined in, it is not
396 if (flow_bb_inside_loop_p (def_loop
, bb
))
399 /* If we're seeing VER for the first time, we still have to allocate
400 a bitmap for its uses. */
401 if (bitmap_set_bit (need_phis
, ver
))
402 use_blocks
[ver
] = BITMAP_ALLOC (&loop_renamer_obstack
);
403 bitmap_set_bit (use_blocks
[ver
], bb
->index
);
406 /* For uses matching USE_FLAGS in STMT, mark names that are used outside of the
407 loop they are defined to rewrite. Record the set of blocks in which the ssa
408 names are used to USE_BLOCKS, and the ssa names themselves to NEED_PHIS. */
411 find_uses_to_rename_stmt (gimple
*stmt
, bitmap
*use_blocks
, bitmap need_phis
,
416 basic_block bb
= gimple_bb (stmt
);
418 if (is_gimple_debug (stmt
))
421 /* FOR_EACH_SSA_TREE_OPERAND iterator does not allows SSA_OP_VIRTUAL_USES
423 if (use_flags
== SSA_OP_VIRTUAL_USES
)
425 tree vuse
= gimple_vuse (stmt
);
426 if (vuse
!= NULL_TREE
)
427 find_uses_to_rename_use (bb
, gimple_vuse (stmt
), use_blocks
, need_phis
);
430 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, use_flags
)
431 find_uses_to_rename_use (bb
, var
, use_blocks
, need_phis
);
434 /* Marks names matching USE_FLAGS that are used in BB and outside of the loop
435 they are defined in for rewrite. Records the set of blocks in which the ssa
436 names are used to USE_BLOCKS. Record the SSA names that will
437 need exit PHIs in NEED_PHIS. */
440 find_uses_to_rename_bb (basic_block bb
, bitmap
*use_blocks
, bitmap need_phis
,
445 bool do_virtuals
= (use_flags
& SSA_OP_VIRTUAL_USES
) != 0;
446 bool do_nonvirtuals
= (use_flags
& SSA_OP_USE
) != 0;
448 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
449 for (gphi_iterator bsi
= gsi_start_phis (e
->dest
); !gsi_end_p (bsi
);
452 gphi
*phi
= bsi
.phi ();
453 bool virtual_p
= virtual_operand_p (gimple_phi_result (phi
));
454 if ((virtual_p
&& do_virtuals
)
455 || (!virtual_p
&& do_nonvirtuals
))
456 find_uses_to_rename_use (bb
, PHI_ARG_DEF_FROM_EDGE (phi
, e
),
457 use_blocks
, need_phis
);
460 for (gimple_stmt_iterator bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
);
462 find_uses_to_rename_stmt (gsi_stmt (bsi
), use_blocks
, need_phis
,
466 /* Marks names matching USE_FLAGS that are used outside of the loop they are
467 defined in for rewrite. Records the set of blocks in which the ssa names are
468 used to USE_BLOCKS. Record the SSA names that will need exit PHIs in
469 NEED_PHIS. If CHANGED_BBS is not NULL, scan only blocks in this set. */
472 find_uses_to_rename (bitmap changed_bbs
, bitmap
*use_blocks
, bitmap need_phis
,
480 EXECUTE_IF_SET_IN_BITMAP (changed_bbs
, 0, index
, bi
)
481 find_uses_to_rename_bb (BASIC_BLOCK_FOR_FN (cfun
, index
), use_blocks
,
482 need_phis
, use_flags
);
484 FOR_EACH_BB_FN (bb
, cfun
)
485 find_uses_to_rename_bb (bb
, use_blocks
, need_phis
, use_flags
);
488 /* Mark uses of DEF that are used outside of the loop they are defined in for
489 rewrite. Record the set of blocks in which the ssa names are used to
490 USE_BLOCKS. Record the SSA names that will need exit PHIs in NEED_PHIS. */
493 find_uses_to_rename_def (tree def
, bitmap
*use_blocks
, bitmap need_phis
)
496 imm_use_iterator imm_iter
;
498 FOR_EACH_IMM_USE_STMT (use_stmt
, imm_iter
, def
)
500 basic_block use_bb
= gimple_bb (use_stmt
);
503 FOR_EACH_IMM_USE_ON_STMT (use_p
, imm_iter
)
505 if (gimple_code (use_stmt
) == GIMPLE_PHI
)
507 edge e
= gimple_phi_arg_edge (as_a
<gphi
*> (use_stmt
),
508 PHI_ARG_INDEX_FROM_USE (use_p
));
511 find_uses_to_rename_use (use_bb
, USE_FROM_PTR (use_p
), use_blocks
,
517 /* Marks names matching USE_FLAGS that are defined in LOOP and used outside of
518 it for rewrite. Records the set of blocks in which the ssa names are used to
519 USE_BLOCKS. Record the SSA names that will need exit PHIs in NEED_PHIS. */
522 find_uses_to_rename_in_loop (struct loop
*loop
, bitmap
*use_blocks
,
523 bitmap need_phis
, int use_flags
)
525 bool do_virtuals
= (use_flags
& SSA_OP_VIRTUAL_USES
) != 0;
526 bool do_nonvirtuals
= (use_flags
& SSA_OP_USE
) != 0;
527 int def_flags
= ((do_virtuals
? SSA_OP_VIRTUAL_DEFS
: 0)
528 | (do_nonvirtuals
? SSA_OP_DEF
: 0));
531 basic_block
*bbs
= get_loop_body (loop
);
533 for (unsigned int i
= 0; i
< loop
->num_nodes
; i
++)
535 basic_block bb
= bbs
[i
];
537 for (gphi_iterator bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
);
540 gphi
*phi
= bsi
.phi ();
541 tree res
= gimple_phi_result (phi
);
542 bool virtual_p
= virtual_operand_p (res
);
543 if ((virtual_p
&& do_virtuals
)
544 || (!virtual_p
&& do_nonvirtuals
))
545 find_uses_to_rename_def (res
, use_blocks
, need_phis
);
548 for (gimple_stmt_iterator bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
);
551 gimple
*stmt
= gsi_stmt (bsi
);
552 /* FOR_EACH_SSA_TREE_OPERAND iterator does not allows
553 SSA_OP_VIRTUAL_DEFS only. */
554 if (def_flags
== SSA_OP_VIRTUAL_DEFS
)
556 tree vdef
= gimple_vdef (stmt
);
558 find_uses_to_rename_def (vdef
, use_blocks
, need_phis
);
564 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, def_flags
)
565 find_uses_to_rename_def (var
, use_blocks
, need_phis
);
573 /* Rewrites the program into a loop closed ssa form -- i.e. inserts extra
574 phi nodes to ensure that no variable is used outside the loop it is
577 This strengthening of the basic ssa form has several advantages:
579 1) Updating it during unrolling/peeling/versioning is trivial, since
580 we do not need to care about the uses outside of the loop.
581 The same applies to virtual operands which are also rewritten into
582 loop closed SSA form. Note that virtual operands are always live
584 2) The behavior of all uses of an induction variable is the same.
585 Without this, you need to distinguish the case when the variable
586 is used outside of the loop it is defined in, for example
588 for (i = 0; i < 100; i++)
590 for (j = 0; j < 100; j++)
598 Looking from the outer loop with the normal SSA form, the first use of k
599 is not well-behaved, while the second one is an induction variable with
602 If LOOP is non-null, only rewrite uses that have defs in LOOP. Otherwise,
603 if CHANGED_BBS is not NULL, we look for uses outside loops only in the
604 basic blocks in this set.
606 USE_FLAGS allows us to specify whether we want virtual, non-virtual or
607 both variables rewritten.
609 UPDATE_FLAG is used in the call to update_ssa. See
610 TODO_update_ssa* for documentation. */
613 rewrite_into_loop_closed_ssa_1 (bitmap changed_bbs
, unsigned update_flag
,
614 int use_flags
, struct loop
*loop
)
617 bitmap names_to_rename
;
619 loops_state_set (LOOP_CLOSED_SSA
);
620 if (number_of_loops (cfun
) <= 1)
623 /* If the pass has caused the SSA form to be out-of-date, update it
625 if (update_flag
== 0)
627 #ifdef ENABLE_CHECKING
628 verify_ssa (true, true);
632 update_ssa (update_flag
);
634 bitmap_obstack_initialize (&loop_renamer_obstack
);
636 names_to_rename
= BITMAP_ALLOC (&loop_renamer_obstack
);
638 /* Uses of names to rename. We don't have to initialize this array,
639 because we know that we will only have entries for the SSA names
640 in NAMES_TO_RENAME. */
641 use_blocks
= XNEWVEC (bitmap
, num_ssa_names
);
645 gcc_assert (changed_bbs
== NULL
);
646 find_uses_to_rename_in_loop (loop
, use_blocks
, names_to_rename
,
651 gcc_assert (loop
== NULL
);
652 find_uses_to_rename (changed_bbs
, use_blocks
, names_to_rename
, use_flags
);
655 if (!bitmap_empty_p (names_to_rename
))
657 /* An array of bitmaps where LOOP_EXITS[I] is the set of basic blocks
658 that are the destination of an edge exiting loop number I. */
659 bitmap
*loop_exits
= XNEWVEC (bitmap
, number_of_loops (cfun
));
660 get_loops_exits (loop_exits
);
662 /* Add the PHI nodes on exits of the loops for the names we need to
664 add_exit_phis (names_to_rename
, use_blocks
, loop_exits
);
668 /* Fix up all the names found to be used outside their original
670 update_ssa (TODO_update_ssa
);
673 bitmap_obstack_release (&loop_renamer_obstack
);
677 /* Rewrites the non-virtual defs and uses into a loop closed ssa form. If
678 CHANGED_BBS is not NULL, we look for uses outside loops only in the basic
679 blocks in this set. UPDATE_FLAG is used in the call to update_ssa. See
680 TODO_update_ssa* for documentation. */
683 rewrite_into_loop_closed_ssa (bitmap changed_bbs
, unsigned update_flag
)
685 rewrite_into_loop_closed_ssa_1 (changed_bbs
, update_flag
, SSA_OP_USE
, NULL
);
688 /* Rewrites virtual defs and uses with def in LOOP into loop closed ssa
692 rewrite_virtuals_into_loop_closed_ssa (struct loop
*loop
)
694 rewrite_into_loop_closed_ssa_1 (NULL
, 0, SSA_OP_VIRTUAL_USES
, loop
);
697 /* Check invariants of the loop closed ssa form for the USE in BB. */
700 check_loop_closed_ssa_use (basic_block bb
, tree use
)
705 if (TREE_CODE (use
) != SSA_NAME
|| virtual_operand_p (use
))
708 def
= SSA_NAME_DEF_STMT (use
);
709 def_bb
= gimple_bb (def
);
711 || flow_bb_inside_loop_p (def_bb
->loop_father
, bb
));
714 /* Checks invariants of loop closed ssa form in statement STMT in BB. */
717 check_loop_closed_ssa_stmt (basic_block bb
, gimple
*stmt
)
722 if (is_gimple_debug (stmt
))
725 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, SSA_OP_USE
)
726 check_loop_closed_ssa_use (bb
, var
);
729 /* Checks that invariants of the loop closed ssa form are preserved.
730 Call verify_ssa when VERIFY_SSA_P is true. */
733 verify_loop_closed_ssa (bool verify_ssa_p
)
739 if (number_of_loops (cfun
) <= 1)
743 verify_ssa (false, true);
745 timevar_push (TV_VERIFY_LOOP_CLOSED
);
747 FOR_EACH_BB_FN (bb
, cfun
)
749 for (gphi_iterator bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
);
752 gphi
*phi
= bsi
.phi ();
753 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
754 check_loop_closed_ssa_use (e
->src
,
755 PHI_ARG_DEF_FROM_EDGE (phi
, e
));
758 for (gimple_stmt_iterator bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
);
760 check_loop_closed_ssa_stmt (bb
, gsi_stmt (bsi
));
763 timevar_pop (TV_VERIFY_LOOP_CLOSED
);
766 /* Split loop exit edge EXIT. The things are a bit complicated by a need to
767 preserve the loop closed ssa form. The newly created block is returned. */
770 split_loop_exit_edge (edge exit
)
772 basic_block dest
= exit
->dest
;
773 basic_block bb
= split_edge (exit
);
778 source_location locus
;
780 for (psi
= gsi_start_phis (dest
); !gsi_end_p (psi
); gsi_next (&psi
))
783 op_p
= PHI_ARG_DEF_PTR_FROM_EDGE (phi
, single_succ_edge (bb
));
784 locus
= gimple_phi_arg_location_from_edge (phi
, single_succ_edge (bb
));
786 name
= USE_FROM_PTR (op_p
);
788 /* If the argument of the PHI node is a constant, we do not need
789 to keep it inside loop. */
790 if (TREE_CODE (name
) != SSA_NAME
)
793 /* Otherwise create an auxiliary phi node that will copy the value
794 of the SSA name out of the loop. */
795 new_name
= duplicate_ssa_name (name
, NULL
);
796 new_phi
= create_phi_node (new_name
, bb
);
797 add_phi_arg (new_phi
, name
, exit
, locus
);
798 SET_USE (op_p
, new_name
);
804 /* Returns the basic block in that statements should be emitted for induction
805 variables incremented at the end of the LOOP. */
808 ip_end_pos (struct loop
*loop
)
813 /* Returns the basic block in that statements should be emitted for induction
814 variables incremented just before exit condition of a LOOP. */
817 ip_normal_pos (struct loop
*loop
)
823 if (!single_pred_p (loop
->latch
))
826 bb
= single_pred (loop
->latch
);
827 last
= last_stmt (bb
);
829 || gimple_code (last
) != GIMPLE_COND
)
832 exit
= EDGE_SUCC (bb
, 0);
833 if (exit
->dest
== loop
->latch
)
834 exit
= EDGE_SUCC (bb
, 1);
836 if (flow_bb_inside_loop_p (loop
, exit
->dest
))
842 /* Stores the standard position for induction variable increment in LOOP
843 (just before the exit condition if it is available and latch block is empty,
844 end of the latch block otherwise) to BSI. INSERT_AFTER is set to true if
845 the increment should be inserted after *BSI. */
848 standard_iv_increment_position (struct loop
*loop
, gimple_stmt_iterator
*bsi
,
851 basic_block bb
= ip_normal_pos (loop
), latch
= ip_end_pos (loop
);
852 gimple
*last
= last_stmt (latch
);
855 || (last
&& gimple_code (last
) != GIMPLE_LABEL
))
857 *bsi
= gsi_last_bb (latch
);
858 *insert_after
= true;
862 *bsi
= gsi_last_bb (bb
);
863 *insert_after
= false;
867 /* Copies phi node arguments for duplicated blocks. The index of the first
868 duplicated block is FIRST_NEW_BLOCK. */
871 copy_phi_node_args (unsigned first_new_block
)
875 for (i
= first_new_block
; i
< (unsigned) last_basic_block_for_fn (cfun
); i
++)
876 BASIC_BLOCK_FOR_FN (cfun
, i
)->flags
|= BB_DUPLICATED
;
878 for (i
= first_new_block
; i
< (unsigned) last_basic_block_for_fn (cfun
); i
++)
879 add_phi_args_after_copy_bb (BASIC_BLOCK_FOR_FN (cfun
, i
));
881 for (i
= first_new_block
; i
< (unsigned) last_basic_block_for_fn (cfun
); i
++)
882 BASIC_BLOCK_FOR_FN (cfun
, i
)->flags
&= ~BB_DUPLICATED
;
886 /* The same as cfgloopmanip.c:duplicate_loop_to_header_edge, but also
887 updates the PHI nodes at start of the copied region. In order to
888 achieve this, only loops whose exits all lead to the same location
891 Notice that we do not completely update the SSA web after
892 duplication. The caller is responsible for calling update_ssa
893 after the loop has been duplicated. */
896 gimple_duplicate_loop_to_header_edge (struct loop
*loop
, edge e
,
897 unsigned int ndupl
, sbitmap wont_exit
,
898 edge orig
, vec
<edge
> *to_remove
,
901 unsigned first_new_block
;
903 if (!loops_state_satisfies_p (LOOPS_HAVE_SIMPLE_LATCHES
))
905 if (!loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS
))
908 first_new_block
= last_basic_block_for_fn (cfun
);
909 if (!duplicate_loop_to_header_edge (loop
, e
, ndupl
, wont_exit
,
910 orig
, to_remove
, flags
))
913 /* Readd the removed phi args for e. */
914 flush_pending_stmts (e
);
916 /* Copy the phi node arguments. */
917 copy_phi_node_args (first_new_block
);
924 /* Returns true if we can unroll LOOP FACTOR times. Number
925 of iterations of the loop is returned in NITER. */
928 can_unroll_loop_p (struct loop
*loop
, unsigned factor
,
929 struct tree_niter_desc
*niter
)
933 /* Check whether unrolling is possible. We only want to unroll loops
934 for that we are able to determine number of iterations. We also
935 want to split the extra iterations of the loop from its end,
936 therefore we require that the loop has precisely one
939 exit
= single_dom_exit (loop
);
943 if (!number_of_iterations_exit (loop
, exit
, niter
, false)
944 || niter
->cmp
== ERROR_MARK
945 /* Scalar evolutions analysis might have copy propagated
946 the abnormal ssa names into these expressions, hence
947 emitting the computations based on them during loop
948 unrolling might create overlapping life ranges for
949 them, and failures in out-of-ssa. */
950 || contains_abnormal_ssa_name_p (niter
->may_be_zero
)
951 || contains_abnormal_ssa_name_p (niter
->control
.base
)
952 || contains_abnormal_ssa_name_p (niter
->control
.step
)
953 || contains_abnormal_ssa_name_p (niter
->bound
))
956 /* And of course, we must be able to duplicate the loop. */
957 if (!can_duplicate_loop_p (loop
))
960 /* The final loop should be small enough. */
961 if (tree_num_loop_insns (loop
, &eni_size_weights
) * factor
962 > (unsigned) PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS
))
968 /* Determines the conditions that control execution of LOOP unrolled FACTOR
969 times. DESC is number of iterations of LOOP. ENTER_COND is set to
970 condition that must be true if the main loop can be entered.
971 EXIT_BASE, EXIT_STEP, EXIT_CMP and EXIT_BOUND are set to values describing
972 how the exit from the unrolled loop should be controlled. */
975 determine_exit_conditions (struct loop
*loop
, struct tree_niter_desc
*desc
,
976 unsigned factor
, tree
*enter_cond
,
977 tree
*exit_base
, tree
*exit_step
,
978 enum tree_code
*exit_cmp
, tree
*exit_bound
)
981 tree base
= desc
->control
.base
;
982 tree step
= desc
->control
.step
;
983 tree bound
= desc
->bound
;
984 tree type
= TREE_TYPE (step
);
986 tree min
= lower_bound_in_type (type
, type
);
987 tree max
= upper_bound_in_type (type
, type
);
988 enum tree_code cmp
= desc
->cmp
;
989 tree cond
= boolean_true_node
, assum
;
991 /* For pointers, do the arithmetics in the type of step. */
992 base
= fold_convert (type
, base
);
993 bound
= fold_convert (type
, bound
);
995 *enter_cond
= boolean_false_node
;
996 *exit_base
= NULL_TREE
;
997 *exit_step
= NULL_TREE
;
998 *exit_cmp
= ERROR_MARK
;
999 *exit_bound
= NULL_TREE
;
1000 gcc_assert (cmp
!= ERROR_MARK
);
1002 /* We only need to be correct when we answer question
1003 "Do at least FACTOR more iterations remain?" in the unrolled loop.
1004 Thus, transforming BASE + STEP * i <> BOUND to
1005 BASE + STEP * i < BOUND is ok. */
1008 if (tree_int_cst_sign_bit (step
))
1013 else if (cmp
== LT_EXPR
)
1015 gcc_assert (!tree_int_cst_sign_bit (step
));
1017 else if (cmp
== GT_EXPR
)
1019 gcc_assert (tree_int_cst_sign_bit (step
));
1024 /* The main body of the loop may be entered iff:
1026 1) desc->may_be_zero is false.
1027 2) it is possible to check that there are at least FACTOR iterations
1028 of the loop, i.e., BOUND - step * FACTOR does not overflow.
1029 3) # of iterations is at least FACTOR */
1031 if (!integer_zerop (desc
->may_be_zero
))
1032 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
1033 invert_truthvalue (desc
->may_be_zero
),
1036 bigstep
= fold_build2 (MULT_EXPR
, type
, step
,
1037 build_int_cst_type (type
, factor
));
1038 delta
= fold_build2 (MINUS_EXPR
, type
, bigstep
, step
);
1040 assum
= fold_build2 (GE_EXPR
, boolean_type_node
,
1042 fold_build2 (PLUS_EXPR
, type
, min
, delta
));
1044 assum
= fold_build2 (LE_EXPR
, boolean_type_node
,
1046 fold_build2 (PLUS_EXPR
, type
, max
, delta
));
1047 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
, assum
, cond
);
1049 bound
= fold_build2 (MINUS_EXPR
, type
, bound
, delta
);
1050 assum
= fold_build2 (cmp
, boolean_type_node
, base
, bound
);
1051 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
, assum
, cond
);
1053 cond
= force_gimple_operand (unshare_expr (cond
), &stmts
, false, NULL_TREE
);
1055 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
1056 /* cond now may be a gimple comparison, which would be OK, but also any
1057 other gimple rhs (say a && b). In this case we need to force it to
1059 if (!is_gimple_condexpr (cond
))
1061 cond
= force_gimple_operand (cond
, &stmts
, true, NULL_TREE
);
1063 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
1067 base
= force_gimple_operand (unshare_expr (base
), &stmts
, true, NULL_TREE
);
1069 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
1070 bound
= force_gimple_operand (unshare_expr (bound
), &stmts
, true, NULL_TREE
);
1072 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
1075 *exit_step
= bigstep
;
1077 *exit_bound
= bound
;
1080 /* Scales the frequencies of all basic blocks in LOOP that are strictly
1081 dominated by BB by NUM/DEN. */
1084 scale_dominated_blocks_in_loop (struct loop
*loop
, basic_block bb
,
1092 for (son
= first_dom_son (CDI_DOMINATORS
, bb
);
1094 son
= next_dom_son (CDI_DOMINATORS
, son
))
1096 if (!flow_bb_inside_loop_p (loop
, son
))
1098 scale_bbs_frequencies_int (&son
, 1, num
, den
);
1099 scale_dominated_blocks_in_loop (loop
, son
, num
, den
);
1103 /* Unroll LOOP FACTOR times. DESC describes number of iterations of LOOP.
1104 EXIT is the exit of the loop to that DESC corresponds.
1106 If N is number of iterations of the loop and MAY_BE_ZERO is the condition
1107 under that loop exits in the first iteration even if N != 0,
1111 x = phi (init, next);
1119 becomes (with possibly the exit conditions formulated a bit differently,
1120 avoiding the need to create a new iv):
1122 if (MAY_BE_ZERO || N < FACTOR)
1127 x = phi (init, next);
1138 } while (N >= FACTOR);
1141 init' = phi (init, x);
1145 x = phi (init', next);
1153 Before the loop is unrolled, TRANSFORM is called for it (only for the
1154 unrolled loop, but not for its versioned copy). DATA is passed to
1157 /* Probability in % that the unrolled loop is entered. Just a guess. */
1158 #define PROB_UNROLLED_LOOP_ENTERED 90
1161 tree_transform_and_unroll_loop (struct loop
*loop
, unsigned factor
,
1162 edge exit
, struct tree_niter_desc
*desc
,
1163 transform_callback transform
,
1167 tree ctr_before
, ctr_after
;
1168 tree enter_main_cond
, exit_base
, exit_step
, exit_bound
;
1169 enum tree_code exit_cmp
;
1170 gphi
*phi_old_loop
, *phi_new_loop
, *phi_rest
;
1171 gphi_iterator psi_old_loop
, psi_new_loop
;
1172 tree init
, next
, new_init
;
1173 struct loop
*new_loop
;
1174 basic_block rest
, exit_bb
;
1175 edge old_entry
, new_entry
, old_latch
, precond_edge
, new_exit
;
1176 edge new_nonexit
, e
;
1177 gimple_stmt_iterator bsi
;
1180 unsigned est_niter
, prob_entry
, scale_unrolled
, scale_rest
, freq_e
, freq_h
;
1181 unsigned new_est_niter
, i
, prob
;
1182 unsigned irr
= loop_preheader_edge (loop
)->flags
& EDGE_IRREDUCIBLE_LOOP
;
1184 auto_vec
<edge
> to_remove
;
1186 est_niter
= expected_loop_iterations (loop
);
1187 determine_exit_conditions (loop
, desc
, factor
,
1188 &enter_main_cond
, &exit_base
, &exit_step
,
1189 &exit_cmp
, &exit_bound
);
1191 /* Let us assume that the unrolled loop is quite likely to be entered. */
1192 if (integer_nonzerop (enter_main_cond
))
1193 prob_entry
= REG_BR_PROB_BASE
;
1195 prob_entry
= PROB_UNROLLED_LOOP_ENTERED
* REG_BR_PROB_BASE
/ 100;
1197 /* The values for scales should keep profile consistent, and somewhat close
1200 TODO: The current value of SCALE_REST makes it appear that the loop that
1201 is created by splitting the remaining iterations of the unrolled loop is
1202 executed the same number of times as the original loop, and with the same
1203 frequencies, which is obviously wrong. This does not appear to cause
1204 problems, so we do not bother with fixing it for now. To make the profile
1205 correct, we would need to change the probability of the exit edge of the
1206 loop, and recompute the distribution of frequencies in its body because
1207 of this change (scale the frequencies of blocks before and after the exit
1208 by appropriate factors). */
1209 scale_unrolled
= prob_entry
;
1210 scale_rest
= REG_BR_PROB_BASE
;
1212 new_loop
= loop_version (loop
, enter_main_cond
, NULL
,
1213 prob_entry
, scale_unrolled
, scale_rest
, true);
1214 gcc_assert (new_loop
!= NULL
);
1215 update_ssa (TODO_update_ssa
);
1217 /* Determine the probability of the exit edge of the unrolled loop. */
1218 new_est_niter
= est_niter
/ factor
;
1220 /* Without profile feedback, loops for that we do not know a better estimate
1221 are assumed to roll 10 times. When we unroll such loop, it appears to
1222 roll too little, and it may even seem to be cold. To avoid this, we
1223 ensure that the created loop appears to roll at least 5 times (but at
1224 most as many times as before unrolling). */
1225 if (new_est_niter
< 5)
1228 new_est_niter
= est_niter
;
1233 /* Prepare the cfg and update the phi nodes. Move the loop exit to the
1234 loop latch (and make its condition dummy, for the moment). */
1235 rest
= loop_preheader_edge (new_loop
)->src
;
1236 precond_edge
= single_pred_edge (rest
);
1237 split_edge (loop_latch_edge (loop
));
1238 exit_bb
= single_pred (loop
->latch
);
1240 /* Since the exit edge will be removed, the frequency of all the blocks
1241 in the loop that are dominated by it must be scaled by
1242 1 / (1 - exit->probability). */
1243 scale_dominated_blocks_in_loop (loop
, exit
->src
,
1245 REG_BR_PROB_BASE
- exit
->probability
);
1247 bsi
= gsi_last_bb (exit_bb
);
1248 exit_if
= gimple_build_cond (EQ_EXPR
, integer_zero_node
,
1250 NULL_TREE
, NULL_TREE
);
1252 gsi_insert_after (&bsi
, exit_if
, GSI_NEW_STMT
);
1253 new_exit
= make_edge (exit_bb
, rest
, EDGE_FALSE_VALUE
| irr
);
1254 rescan_loop_exit (new_exit
, true, false);
1256 /* Set the probability of new exit to the same of the old one. Fix
1257 the frequency of the latch block, by scaling it back by
1258 1 - exit->probability. */
1259 new_exit
->count
= exit
->count
;
1260 new_exit
->probability
= exit
->probability
;
1261 new_nonexit
= single_pred_edge (loop
->latch
);
1262 new_nonexit
->probability
= REG_BR_PROB_BASE
- exit
->probability
;
1263 new_nonexit
->flags
= EDGE_TRUE_VALUE
;
1264 new_nonexit
->count
-= exit
->count
;
1265 if (new_nonexit
->count
< 0)
1266 new_nonexit
->count
= 0;
1267 scale_bbs_frequencies_int (&loop
->latch
, 1, new_nonexit
->probability
,
1270 old_entry
= loop_preheader_edge (loop
);
1271 new_entry
= loop_preheader_edge (new_loop
);
1272 old_latch
= loop_latch_edge (loop
);
1273 for (psi_old_loop
= gsi_start_phis (loop
->header
),
1274 psi_new_loop
= gsi_start_phis (new_loop
->header
);
1275 !gsi_end_p (psi_old_loop
);
1276 gsi_next (&psi_old_loop
), gsi_next (&psi_new_loop
))
1278 phi_old_loop
= psi_old_loop
.phi ();
1279 phi_new_loop
= psi_new_loop
.phi ();
1281 init
= PHI_ARG_DEF_FROM_EDGE (phi_old_loop
, old_entry
);
1282 op
= PHI_ARG_DEF_PTR_FROM_EDGE (phi_new_loop
, new_entry
);
1283 gcc_assert (operand_equal_for_phi_arg_p (init
, USE_FROM_PTR (op
)));
1284 next
= PHI_ARG_DEF_FROM_EDGE (phi_old_loop
, old_latch
);
1286 /* Prefer using original variable as a base for the new ssa name.
1287 This is necessary for virtual ops, and useful in order to avoid
1288 losing debug info for real ops. */
1289 if (TREE_CODE (next
) == SSA_NAME
1290 && useless_type_conversion_p (TREE_TYPE (next
),
1292 new_init
= copy_ssa_name (next
);
1293 else if (TREE_CODE (init
) == SSA_NAME
1294 && useless_type_conversion_p (TREE_TYPE (init
),
1296 new_init
= copy_ssa_name (init
);
1297 else if (useless_type_conversion_p (TREE_TYPE (next
), TREE_TYPE (init
)))
1298 new_init
= make_temp_ssa_name (TREE_TYPE (next
), NULL
, "unrinittmp");
1300 new_init
= make_temp_ssa_name (TREE_TYPE (init
), NULL
, "unrinittmp");
1302 phi_rest
= create_phi_node (new_init
, rest
);
1304 add_phi_arg (phi_rest
, init
, precond_edge
, UNKNOWN_LOCATION
);
1305 add_phi_arg (phi_rest
, next
, new_exit
, UNKNOWN_LOCATION
);
1306 SET_USE (op
, new_init
);
1311 /* Transform the loop. */
1313 (*transform
) (loop
, data
);
1315 /* Unroll the loop and remove the exits in all iterations except for the
1317 wont_exit
= sbitmap_alloc (factor
);
1318 bitmap_ones (wont_exit
);
1319 bitmap_clear_bit (wont_exit
, factor
- 1);
1321 ok
= gimple_duplicate_loop_to_header_edge
1322 (loop
, loop_latch_edge (loop
), factor
- 1,
1323 wont_exit
, new_exit
, &to_remove
, DLTHE_FLAG_UPDATE_FREQ
);
1327 FOR_EACH_VEC_ELT (to_remove
, i
, e
)
1329 ok
= remove_path (e
);
1332 update_ssa (TODO_update_ssa
);
1334 /* Ensure that the frequencies in the loop match the new estimated
1335 number of iterations, and change the probability of the new
1337 freq_h
= loop
->header
->frequency
;
1338 freq_e
= EDGE_FREQUENCY (loop_preheader_edge (loop
));
1340 scale_loop_frequencies (loop
, freq_e
* (new_est_niter
+ 1), freq_h
);
1342 exit_bb
= single_pred (loop
->latch
);
1343 new_exit
= find_edge (exit_bb
, rest
);
1344 new_exit
->count
= loop_preheader_edge (loop
)->count
;
1345 new_exit
->probability
= REG_BR_PROB_BASE
/ (new_est_niter
+ 1);
1347 rest
->count
+= new_exit
->count
;
1348 rest
->frequency
+= EDGE_FREQUENCY (new_exit
);
1350 new_nonexit
= single_pred_edge (loop
->latch
);
1351 prob
= new_nonexit
->probability
;
1352 new_nonexit
->probability
= REG_BR_PROB_BASE
- new_exit
->probability
;
1353 new_nonexit
->count
= exit_bb
->count
- new_exit
->count
;
1354 if (new_nonexit
->count
< 0)
1355 new_nonexit
->count
= 0;
1357 scale_bbs_frequencies_int (&loop
->latch
, 1, new_nonexit
->probability
,
1360 /* Finally create the new counter for number of iterations and add the new
1361 exit instruction. */
1362 bsi
= gsi_last_nondebug_bb (exit_bb
);
1363 exit_if
= as_a
<gcond
*> (gsi_stmt (bsi
));
1364 create_iv (exit_base
, exit_step
, NULL_TREE
, loop
,
1365 &bsi
, false, &ctr_before
, &ctr_after
);
1366 gimple_cond_set_code (exit_if
, exit_cmp
);
1367 gimple_cond_set_lhs (exit_if
, ctr_after
);
1368 gimple_cond_set_rhs (exit_if
, exit_bound
);
1369 update_stmt (exit_if
);
1371 checking_verify_flow_info ();
1372 checking_verify_loop_structure ();
1373 checking_verify_loop_closed_ssa (true);
1376 /* Wrapper over tree_transform_and_unroll_loop for case we do not
1377 want to transform the loop before unrolling. The meaning
1378 of the arguments is the same as for tree_transform_and_unroll_loop. */
1381 tree_unroll_loop (struct loop
*loop
, unsigned factor
,
1382 edge exit
, struct tree_niter_desc
*desc
)
1384 tree_transform_and_unroll_loop (loop
, factor
, exit
, desc
,
1388 /* Rewrite the phi node at position PSI in function of the main
1389 induction variable MAIN_IV and insert the generated code at GSI. */
1392 rewrite_phi_with_iv (loop_p loop
,
1394 gimple_stmt_iterator
*gsi
,
1399 gphi
*phi
= psi
->phi ();
1400 tree atype
, mtype
, val
, res
= PHI_RESULT (phi
);
1402 if (virtual_operand_p (res
) || res
== main_iv
)
1408 if (!simple_iv (loop
, loop
, res
, &iv
, true))
1414 remove_phi_node (psi
, false);
1416 atype
= TREE_TYPE (res
);
1417 mtype
= POINTER_TYPE_P (atype
) ? sizetype
: atype
;
1418 val
= fold_build2 (MULT_EXPR
, mtype
, unshare_expr (iv
.step
),
1419 fold_convert (mtype
, main_iv
));
1420 val
= fold_build2 (POINTER_TYPE_P (atype
)
1421 ? POINTER_PLUS_EXPR
: PLUS_EXPR
,
1422 atype
, unshare_expr (iv
.base
), val
);
1423 val
= force_gimple_operand_gsi (gsi
, val
, false, NULL_TREE
, true,
1425 stmt
= gimple_build_assign (res
, val
);
1426 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
1429 /* Rewrite all the phi nodes of LOOP in function of the main induction
1430 variable MAIN_IV. */
1433 rewrite_all_phi_nodes_with_iv (loop_p loop
, tree main_iv
)
1436 basic_block
*bbs
= get_loop_body_in_dom_order (loop
);
1439 for (i
= 0; i
< loop
->num_nodes
; i
++)
1441 basic_block bb
= bbs
[i
];
1442 gimple_stmt_iterator gsi
= gsi_after_labels (bb
);
1444 if (bb
->loop_father
!= loop
)
1447 for (psi
= gsi_start_phis (bb
); !gsi_end_p (psi
); )
1448 rewrite_phi_with_iv (loop
, &psi
, &gsi
, main_iv
);
1454 /* Bases all the induction variables in LOOP on a single induction variable
1455 (with base 0 and step 1), whose final value is compared with *NIT. When the
1456 IV type precision has to be larger than *NIT type precision, *NIT is
1457 converted to the larger type, the conversion code is inserted before the
1458 loop, and *NIT is updated to the new definition. When BUMP_IN_LATCH is true,
1459 the induction variable is incremented in the loop latch, otherwise it is
1460 incremented in the loop header. Return the induction variable that was
1464 canonicalize_loop_ivs (struct loop
*loop
, tree
*nit
, bool bump_in_latch
)
1466 unsigned precision
= TYPE_PRECISION (TREE_TYPE (*nit
));
1467 unsigned original_precision
= precision
;
1468 tree type
, var_before
;
1469 gimple_stmt_iterator gsi
;
1472 edge exit
= single_dom_exit (loop
);
1475 bool unsigned_p
= false;
1477 for (psi
= gsi_start_phis (loop
->header
);
1478 !gsi_end_p (psi
); gsi_next (&psi
))
1480 gphi
*phi
= psi
.phi ();
1481 tree res
= PHI_RESULT (phi
);
1484 type
= TREE_TYPE (res
);
1485 if (virtual_operand_p (res
)
1486 || (!INTEGRAL_TYPE_P (type
)
1487 && !POINTER_TYPE_P (type
))
1488 || TYPE_PRECISION (type
) < precision
)
1491 uns
= POINTER_TYPE_P (type
) | TYPE_UNSIGNED (type
);
1493 if (TYPE_PRECISION (type
) > precision
)
1498 precision
= TYPE_PRECISION (type
);
1501 mode
= smallest_mode_for_size (precision
, MODE_INT
);
1502 precision
= GET_MODE_PRECISION (mode
);
1503 type
= build_nonstandard_integer_type (precision
, unsigned_p
);
1505 if (original_precision
!= precision
)
1507 *nit
= fold_convert (type
, *nit
);
1508 *nit
= force_gimple_operand (*nit
, &stmts
, true, NULL_TREE
);
1510 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
1514 gsi
= gsi_last_bb (loop
->latch
);
1516 gsi
= gsi_last_nondebug_bb (loop
->header
);
1517 create_iv (build_int_cst_type (type
, 0), build_int_cst (type
, 1), NULL_TREE
,
1518 loop
, &gsi
, bump_in_latch
, &var_before
, NULL
);
1520 rewrite_all_phi_nodes_with_iv (loop
, var_before
);
1522 stmt
= as_a
<gcond
*> (last_stmt (exit
->src
));
1523 /* Make the loop exit if the control condition is not satisfied. */
1524 if (exit
->flags
& EDGE_TRUE_VALUE
)
1528 extract_true_false_edges_from_block (exit
->src
, &te
, &fe
);
1529 te
->flags
= EDGE_FALSE_VALUE
;
1530 fe
->flags
= EDGE_TRUE_VALUE
;
1532 gimple_cond_set_code (stmt
, LT_EXPR
);
1533 gimple_cond_set_lhs (stmt
, var_before
);
1534 gimple_cond_set_rhs (stmt
, *nit
);