gcc/ChangeLog:
[official-gcc.git] / gcc / tree-ssa-loop-manip.c
blob39fb48454ae769e3557c2642fbef41ea4f155152
1 /* High-level loop manipulation functions.
2 Copyright (C) 2004-2017 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
9 later version.
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
14 for more details.
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/>. */
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "backend.h"
24 #include "tree.h"
25 #include "gimple.h"
26 #include "cfghooks.h"
27 #include "tree-pass.h" /* ??? for TODO_update_ssa but this isn't a pass. */
28 #include "ssa.h"
29 #include "gimple-pretty-print.h"
30 #include "fold-const.h"
31 #include "cfganal.h"
32 #include "gimplify.h"
33 #include "gimple-iterator.h"
34 #include "gimplify-me.h"
35 #include "tree-cfg.h"
36 #include "tree-ssa-loop-ivopts.h"
37 #include "tree-ssa-loop-manip.h"
38 #include "tree-ssa-loop-niter.h"
39 #include "tree-ssa-loop.h"
40 #include "tree-into-ssa.h"
41 #include "tree-ssa.h"
42 #include "cfgloop.h"
43 #include "tree-scalar-evolution.h"
44 #include "params.h"
45 #include "tree-inline.h"
47 /* All bitmaps for rewriting into loop-closed SSA go on this obstack,
48 so that we can free them all at once. */
49 static bitmap_obstack loop_renamer_obstack;
51 /* Creates an induction variable with value BASE + STEP * iteration in LOOP.
52 It is expected that neither BASE nor STEP are shared with other expressions
53 (unless the sharing rules allow this). Use VAR as a base var_decl for it
54 (if NULL, a new temporary will be created). The increment will occur at
55 INCR_POS (after it if AFTER is true, before it otherwise). INCR_POS and
56 AFTER can be computed using standard_iv_increment_position. The ssa versions
57 of the variable before and after increment will be stored in VAR_BEFORE and
58 VAR_AFTER (unless they are NULL). */
60 void
61 create_iv (tree base, tree step, tree var, struct loop *loop,
62 gimple_stmt_iterator *incr_pos, bool after,
63 tree *var_before, tree *var_after)
65 gassign *stmt;
66 gphi *phi;
67 tree initial, step1;
68 gimple_seq stmts;
69 tree vb, va;
70 enum tree_code incr_op = PLUS_EXPR;
71 edge pe = loop_preheader_edge (loop);
73 if (var != NULL_TREE)
75 vb = make_ssa_name (var);
76 va = make_ssa_name (var);
78 else
80 vb = make_temp_ssa_name (TREE_TYPE (base), NULL, "ivtmp");
81 va = make_temp_ssa_name (TREE_TYPE (base), NULL, "ivtmp");
83 if (var_before)
84 *var_before = vb;
85 if (var_after)
86 *var_after = va;
88 /* For easier readability of the created code, produce MINUS_EXPRs
89 when suitable. */
90 if (TREE_CODE (step) == INTEGER_CST)
92 if (TYPE_UNSIGNED (TREE_TYPE (step)))
94 step1 = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
95 if (tree_int_cst_lt (step1, step))
97 incr_op = MINUS_EXPR;
98 step = step1;
101 else
103 bool ovf;
105 if (!tree_expr_nonnegative_warnv_p (step, &ovf)
106 && may_negate_without_overflow_p (step))
108 incr_op = MINUS_EXPR;
109 step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
113 if (POINTER_TYPE_P (TREE_TYPE (base)))
115 if (TREE_CODE (base) == ADDR_EXPR)
116 mark_addressable (TREE_OPERAND (base, 0));
117 step = convert_to_ptrofftype (step);
118 if (incr_op == MINUS_EXPR)
119 step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
120 incr_op = POINTER_PLUS_EXPR;
122 /* Gimplify the step if necessary. We put the computations in front of the
123 loop (i.e. the step should be loop invariant). */
124 step = force_gimple_operand (step, &stmts, true, NULL_TREE);
125 if (stmts)
126 gsi_insert_seq_on_edge_immediate (pe, stmts);
128 stmt = gimple_build_assign (va, incr_op, vb, step);
129 if (after)
130 gsi_insert_after (incr_pos, stmt, GSI_NEW_STMT);
131 else
132 gsi_insert_before (incr_pos, stmt, GSI_NEW_STMT);
134 initial = force_gimple_operand (base, &stmts, true, var);
135 if (stmts)
136 gsi_insert_seq_on_edge_immediate (pe, stmts);
138 phi = create_phi_node (vb, loop->header);
139 add_phi_arg (phi, initial, loop_preheader_edge (loop), UNKNOWN_LOCATION);
140 add_phi_arg (phi, va, loop_latch_edge (loop), UNKNOWN_LOCATION);
143 /* Return the innermost superloop LOOP of USE_LOOP that is a superloop of
144 both DEF_LOOP and USE_LOOP. */
146 static inline struct loop *
147 find_sibling_superloop (struct loop *use_loop, struct loop *def_loop)
149 unsigned ud = loop_depth (use_loop);
150 unsigned dd = loop_depth (def_loop);
151 gcc_assert (ud > 0 && dd > 0);
152 if (ud > dd)
153 use_loop = superloop_at_depth (use_loop, dd);
154 if (ud < dd)
155 def_loop = superloop_at_depth (def_loop, ud);
156 while (loop_outer (use_loop) != loop_outer (def_loop))
158 use_loop = loop_outer (use_loop);
159 def_loop = loop_outer (def_loop);
160 gcc_assert (use_loop && def_loop);
162 return use_loop;
165 /* DEF_BB is a basic block containing a DEF that needs rewriting into
166 loop-closed SSA form. USE_BLOCKS is the set of basic blocks containing
167 uses of DEF that "escape" from the loop containing DEF_BB (i.e. blocks in
168 USE_BLOCKS are dominated by DEF_BB but not in the loop father of DEF_B).
169 ALL_EXITS[I] is the set of all basic blocks that exit loop I.
171 Compute the subset of LOOP_EXITS that exit the loop containing DEF_BB
172 or one of its loop fathers, in which DEF is live. This set is returned
173 in the bitmap LIVE_EXITS.
175 Instead of computing the complete livein set of the def, we use the loop
176 nesting tree as a form of poor man's structure analysis. This greatly
177 speeds up the analysis, which is important because this function may be
178 called on all SSA names that need rewriting, one at a time. */
180 static void
181 compute_live_loop_exits (bitmap live_exits, bitmap use_blocks,
182 bitmap *loop_exits, basic_block def_bb)
184 unsigned i;
185 bitmap_iterator bi;
186 struct loop *def_loop = def_bb->loop_father;
187 unsigned def_loop_depth = loop_depth (def_loop);
188 bitmap def_loop_exits;
190 /* Normally the work list size is bounded by the number of basic
191 blocks in the largest loop. We don't know this number, but we
192 can be fairly sure that it will be relatively small. */
193 auto_vec<basic_block> worklist (MAX (8, n_basic_blocks_for_fn (cfun) / 128));
195 EXECUTE_IF_SET_IN_BITMAP (use_blocks, 0, i, bi)
197 basic_block use_bb = BASIC_BLOCK_FOR_FN (cfun, i);
198 struct loop *use_loop = use_bb->loop_father;
199 gcc_checking_assert (def_loop != use_loop
200 && ! flow_loop_nested_p (def_loop, use_loop));
201 if (! flow_loop_nested_p (use_loop, def_loop))
202 use_bb = find_sibling_superloop (use_loop, def_loop)->header;
203 if (bitmap_set_bit (live_exits, use_bb->index))
204 worklist.safe_push (use_bb);
207 /* Iterate until the worklist is empty. */
208 while (! worklist.is_empty ())
210 edge e;
211 edge_iterator ei;
213 /* Pull a block off the worklist. */
214 basic_block bb = worklist.pop ();
216 /* Make sure we have at least enough room in the work list
217 for all predecessors of this block. */
218 worklist.reserve (EDGE_COUNT (bb->preds));
220 /* For each predecessor block. */
221 FOR_EACH_EDGE (e, ei, bb->preds)
223 basic_block pred = e->src;
224 struct loop *pred_loop = pred->loop_father;
225 unsigned pred_loop_depth = loop_depth (pred_loop);
226 bool pred_visited;
228 /* We should have met DEF_BB along the way. */
229 gcc_assert (pred != ENTRY_BLOCK_PTR_FOR_FN (cfun));
231 if (pred_loop_depth >= def_loop_depth)
233 if (pred_loop_depth > def_loop_depth)
234 pred_loop = superloop_at_depth (pred_loop, def_loop_depth);
235 /* If we've reached DEF_LOOP, our train ends here. */
236 if (pred_loop == def_loop)
237 continue;
239 else if (! flow_loop_nested_p (pred_loop, def_loop))
240 pred = find_sibling_superloop (pred_loop, def_loop)->header;
242 /* Add PRED to the LIVEIN set. PRED_VISITED is true if
243 we had already added PRED to LIVEIN before. */
244 pred_visited = !bitmap_set_bit (live_exits, pred->index);
246 /* If we have visited PRED before, don't add it to the worklist.
247 If BB dominates PRED, then we're probably looking at a loop.
248 We're only interested in looking up in the dominance tree
249 because DEF_BB dominates all the uses. */
250 if (pred_visited || dominated_by_p (CDI_DOMINATORS, pred, bb))
251 continue;
253 worklist.quick_push (pred);
257 def_loop_exits = BITMAP_ALLOC (&loop_renamer_obstack);
258 for (struct loop *loop = def_loop;
259 loop != current_loops->tree_root;
260 loop = loop_outer (loop))
261 bitmap_ior_into (def_loop_exits, loop_exits[loop->num]);
262 bitmap_and_into (live_exits, def_loop_exits);
263 BITMAP_FREE (def_loop_exits);
266 /* Add a loop-closing PHI for VAR in basic block EXIT. */
268 static void
269 add_exit_phi (basic_block exit, tree var)
271 gphi *phi;
272 edge e;
273 edge_iterator ei;
275 /* Check that at least one of the edges entering the EXIT block exits
276 the loop, or a superloop of that loop, that VAR is defined in. */
277 if (flag_checking)
279 gimple *def_stmt = SSA_NAME_DEF_STMT (var);
280 basic_block def_bb = gimple_bb (def_stmt);
281 FOR_EACH_EDGE (e, ei, exit->preds)
283 struct loop *aloop = find_common_loop (def_bb->loop_father,
284 e->src->loop_father);
285 if (!flow_bb_inside_loop_p (aloop, e->dest))
286 break;
288 gcc_assert (e);
291 phi = create_phi_node (NULL_TREE, exit);
292 create_new_def_for (var, phi, gimple_phi_result_ptr (phi));
293 FOR_EACH_EDGE (e, ei, exit->preds)
294 add_phi_arg (phi, var, e, UNKNOWN_LOCATION);
296 if (dump_file && (dump_flags & TDF_DETAILS))
298 fprintf (dump_file, ";; Created LCSSA PHI: ");
299 print_gimple_stmt (dump_file, phi, 0, dump_flags);
303 /* Add exit phis for VAR that is used in LIVEIN.
304 Exits of the loops are stored in LOOP_EXITS. */
306 static void
307 add_exit_phis_var (tree var, bitmap use_blocks, bitmap *loop_exits)
309 unsigned index;
310 bitmap_iterator bi;
311 basic_block def_bb = gimple_bb (SSA_NAME_DEF_STMT (var));
312 bitmap live_exits = BITMAP_ALLOC (&loop_renamer_obstack);
314 gcc_checking_assert (! bitmap_bit_p (use_blocks, def_bb->index));
316 compute_live_loop_exits (live_exits, use_blocks, loop_exits, def_bb);
318 EXECUTE_IF_SET_IN_BITMAP (live_exits, 0, index, bi)
320 add_exit_phi (BASIC_BLOCK_FOR_FN (cfun, index), var);
323 BITMAP_FREE (live_exits);
326 /* Add exit phis for the names marked in NAMES_TO_RENAME.
327 Exits of the loops are stored in EXITS. Sets of blocks where the ssa
328 names are used are stored in USE_BLOCKS. */
330 static void
331 add_exit_phis (bitmap names_to_rename, bitmap *use_blocks, bitmap *loop_exits)
333 unsigned i;
334 bitmap_iterator bi;
336 EXECUTE_IF_SET_IN_BITMAP (names_to_rename, 0, i, bi)
338 add_exit_phis_var (ssa_name (i), use_blocks[i], loop_exits);
342 /* Fill the array of bitmaps LOOP_EXITS with all loop exit edge targets. */
344 static void
345 get_loops_exits (bitmap *loop_exits)
347 struct loop *loop;
348 unsigned j;
349 edge e;
351 FOR_EACH_LOOP (loop, 0)
353 vec<edge> exit_edges = get_loop_exit_edges (loop);
354 loop_exits[loop->num] = BITMAP_ALLOC (&loop_renamer_obstack);
355 FOR_EACH_VEC_ELT (exit_edges, j, e)
356 bitmap_set_bit (loop_exits[loop->num], e->dest->index);
357 exit_edges.release ();
361 /* For USE in BB, if it is used outside of the loop it is defined in,
362 mark it for rewrite. Record basic block BB where it is used
363 to USE_BLOCKS. Record the ssa name index to NEED_PHIS bitmap.
364 Note that for USEs in phis, BB should be the src of the edge corresponding to
365 the use, rather than the bb containing the phi. */
367 static void
368 find_uses_to_rename_use (basic_block bb, tree use, bitmap *use_blocks,
369 bitmap need_phis)
371 unsigned ver;
372 basic_block def_bb;
373 struct loop *def_loop;
375 if (TREE_CODE (use) != SSA_NAME)
376 return;
378 ver = SSA_NAME_VERSION (use);
379 def_bb = gimple_bb (SSA_NAME_DEF_STMT (use));
380 if (!def_bb)
381 return;
382 def_loop = def_bb->loop_father;
384 /* If the definition is not inside a loop, it is not interesting. */
385 if (!loop_outer (def_loop))
386 return;
388 /* If the use is not outside of the loop it is defined in, it is not
389 interesting. */
390 if (flow_bb_inside_loop_p (def_loop, bb))
391 return;
393 /* If we're seeing VER for the first time, we still have to allocate
394 a bitmap for its uses. */
395 if (bitmap_set_bit (need_phis, ver))
396 use_blocks[ver] = BITMAP_ALLOC (&loop_renamer_obstack);
397 bitmap_set_bit (use_blocks[ver], bb->index);
400 /* For uses matching USE_FLAGS in STMT, mark names that are used outside of the
401 loop they are defined to rewrite. Record the set of blocks in which the ssa
402 names are used to USE_BLOCKS, and the ssa names themselves to NEED_PHIS. */
404 static void
405 find_uses_to_rename_stmt (gimple *stmt, bitmap *use_blocks, bitmap need_phis,
406 int use_flags)
408 ssa_op_iter iter;
409 tree var;
410 basic_block bb = gimple_bb (stmt);
412 if (is_gimple_debug (stmt))
413 return;
415 /* FOR_EACH_SSA_TREE_OPERAND iterator does not allows SSA_OP_VIRTUAL_USES
416 only. */
417 if (use_flags == SSA_OP_VIRTUAL_USES)
419 tree vuse = gimple_vuse (stmt);
420 if (vuse != NULL_TREE)
421 find_uses_to_rename_use (bb, gimple_vuse (stmt), use_blocks, need_phis);
423 else
424 FOR_EACH_SSA_TREE_OPERAND (var, stmt, iter, use_flags)
425 find_uses_to_rename_use (bb, var, use_blocks, need_phis);
428 /* Marks names matching USE_FLAGS that are used in BB and outside of the loop
429 they are defined in for rewrite. Records the set of blocks in which the ssa
430 names are used to USE_BLOCKS. Record the SSA names that will
431 need exit PHIs in NEED_PHIS. */
433 static void
434 find_uses_to_rename_bb (basic_block bb, bitmap *use_blocks, bitmap need_phis,
435 int use_flags)
437 edge e;
438 edge_iterator ei;
439 bool do_virtuals = (use_flags & SSA_OP_VIRTUAL_USES) != 0;
440 bool do_nonvirtuals = (use_flags & SSA_OP_USE) != 0;
442 FOR_EACH_EDGE (e, ei, bb->succs)
443 for (gphi_iterator bsi = gsi_start_phis (e->dest); !gsi_end_p (bsi);
444 gsi_next (&bsi))
446 gphi *phi = bsi.phi ();
447 bool virtual_p = virtual_operand_p (gimple_phi_result (phi));
448 if ((virtual_p && do_virtuals)
449 || (!virtual_p && do_nonvirtuals))
450 find_uses_to_rename_use (bb, PHI_ARG_DEF_FROM_EDGE (phi, e),
451 use_blocks, need_phis);
454 for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi);
455 gsi_next (&bsi))
456 find_uses_to_rename_stmt (gsi_stmt (bsi), use_blocks, need_phis,
457 use_flags);
460 /* Marks names matching USE_FLAGS that are used outside of the loop they are
461 defined in for rewrite. Records the set of blocks in which the ssa names are
462 used to USE_BLOCKS. Record the SSA names that will need exit PHIs in
463 NEED_PHIS. If CHANGED_BBS is not NULL, scan only blocks in this set. */
465 static void
466 find_uses_to_rename (bitmap changed_bbs, bitmap *use_blocks, bitmap need_phis,
467 int use_flags)
469 basic_block bb;
470 unsigned index;
471 bitmap_iterator bi;
473 if (changed_bbs)
474 EXECUTE_IF_SET_IN_BITMAP (changed_bbs, 0, index, bi)
476 bb = BASIC_BLOCK_FOR_FN (cfun, index);
477 if (bb)
478 find_uses_to_rename_bb (bb, use_blocks, need_phis, use_flags);
480 else
481 FOR_EACH_BB_FN (bb, cfun)
482 find_uses_to_rename_bb (bb, use_blocks, need_phis, use_flags);
485 /* Mark uses of DEF that are used outside of the loop they are defined in for
486 rewrite. Record the set of blocks in which the ssa names are used to
487 USE_BLOCKS. Record the SSA names that will need exit PHIs in NEED_PHIS. */
489 static void
490 find_uses_to_rename_def (tree def, bitmap *use_blocks, bitmap need_phis)
492 gimple *use_stmt;
493 imm_use_iterator imm_iter;
495 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
497 basic_block use_bb = gimple_bb (use_stmt);
499 use_operand_p use_p;
500 FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
502 if (gimple_code (use_stmt) == GIMPLE_PHI)
504 edge e = gimple_phi_arg_edge (as_a <gphi *> (use_stmt),
505 PHI_ARG_INDEX_FROM_USE (use_p));
506 use_bb = e->src;
508 find_uses_to_rename_use (use_bb, USE_FROM_PTR (use_p), use_blocks,
509 need_phis);
514 /* Marks names matching USE_FLAGS that are defined in LOOP and used outside of
515 it for rewrite. Records the set of blocks in which the ssa names are used to
516 USE_BLOCKS. Record the SSA names that will need exit PHIs in NEED_PHIS. */
518 static void
519 find_uses_to_rename_in_loop (struct loop *loop, bitmap *use_blocks,
520 bitmap need_phis, int use_flags)
522 bool do_virtuals = (use_flags & SSA_OP_VIRTUAL_USES) != 0;
523 bool do_nonvirtuals = (use_flags & SSA_OP_USE) != 0;
524 int def_flags = ((do_virtuals ? SSA_OP_VIRTUAL_DEFS : 0)
525 | (do_nonvirtuals ? SSA_OP_DEF : 0));
528 basic_block *bbs = get_loop_body (loop);
530 for (unsigned int i = 0; i < loop->num_nodes; i++)
532 basic_block bb = bbs[i];
534 for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi);
535 gsi_next (&bsi))
537 gphi *phi = bsi.phi ();
538 tree res = gimple_phi_result (phi);
539 bool virtual_p = virtual_operand_p (res);
540 if ((virtual_p && do_virtuals)
541 || (!virtual_p && do_nonvirtuals))
542 find_uses_to_rename_def (res, use_blocks, need_phis);
545 for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi);
546 gsi_next (&bsi))
548 gimple *stmt = gsi_stmt (bsi);
549 /* FOR_EACH_SSA_TREE_OPERAND iterator does not allows
550 SSA_OP_VIRTUAL_DEFS only. */
551 if (def_flags == SSA_OP_VIRTUAL_DEFS)
553 tree vdef = gimple_vdef (stmt);
554 if (vdef != NULL)
555 find_uses_to_rename_def (vdef, use_blocks, need_phis);
557 else
559 tree var;
560 ssa_op_iter iter;
561 FOR_EACH_SSA_TREE_OPERAND (var, stmt, iter, def_flags)
562 find_uses_to_rename_def (var, use_blocks, need_phis);
567 XDELETEVEC (bbs);
570 /* Rewrites the program into a loop closed ssa form -- i.e. inserts extra
571 phi nodes to ensure that no variable is used outside the loop it is
572 defined in.
574 This strengthening of the basic ssa form has several advantages:
576 1) Updating it during unrolling/peeling/versioning is trivial, since
577 we do not need to care about the uses outside of the loop.
578 The same applies to virtual operands which are also rewritten into
579 loop closed SSA form. Note that virtual operands are always live
580 until function exit.
581 2) The behavior of all uses of an induction variable is the same.
582 Without this, you need to distinguish the case when the variable
583 is used outside of the loop it is defined in, for example
585 for (i = 0; i < 100; i++)
587 for (j = 0; j < 100; j++)
589 k = i + j;
590 use1 (k);
592 use2 (k);
595 Looking from the outer loop with the normal SSA form, the first use of k
596 is not well-behaved, while the second one is an induction variable with
597 base 99 and step 1.
599 If LOOP is non-null, only rewrite uses that have defs in LOOP. Otherwise,
600 if CHANGED_BBS is not NULL, we look for uses outside loops only in the
601 basic blocks in this set.
603 USE_FLAGS allows us to specify whether we want virtual, non-virtual or
604 both variables rewritten.
606 UPDATE_FLAG is used in the call to update_ssa. See
607 TODO_update_ssa* for documentation. */
609 void
610 rewrite_into_loop_closed_ssa_1 (bitmap changed_bbs, unsigned update_flag,
611 int use_flags, struct loop *loop)
613 bitmap *use_blocks;
614 bitmap names_to_rename;
616 loops_state_set (LOOP_CLOSED_SSA);
617 if (number_of_loops (cfun) <= 1)
618 return;
620 /* If the pass has caused the SSA form to be out-of-date, update it
621 now. */
622 if (update_flag != 0)
623 update_ssa (update_flag);
624 else if (flag_checking)
625 verify_ssa (true, true);
627 bitmap_obstack_initialize (&loop_renamer_obstack);
629 names_to_rename = BITMAP_ALLOC (&loop_renamer_obstack);
631 /* Uses of names to rename. We don't have to initialize this array,
632 because we know that we will only have entries for the SSA names
633 in NAMES_TO_RENAME. */
634 use_blocks = XNEWVEC (bitmap, num_ssa_names);
636 if (loop != NULL)
638 gcc_assert (changed_bbs == NULL);
639 find_uses_to_rename_in_loop (loop, use_blocks, names_to_rename,
640 use_flags);
642 else
644 gcc_assert (loop == NULL);
645 find_uses_to_rename (changed_bbs, use_blocks, names_to_rename, use_flags);
648 if (!bitmap_empty_p (names_to_rename))
650 /* An array of bitmaps where LOOP_EXITS[I] is the set of basic blocks
651 that are the destination of an edge exiting loop number I. */
652 bitmap *loop_exits = XNEWVEC (bitmap, number_of_loops (cfun));
653 get_loops_exits (loop_exits);
655 /* Add the PHI nodes on exits of the loops for the names we need to
656 rewrite. */
657 add_exit_phis (names_to_rename, use_blocks, loop_exits);
659 free (loop_exits);
661 /* Fix up all the names found to be used outside their original
662 loops. */
663 update_ssa (TODO_update_ssa);
666 bitmap_obstack_release (&loop_renamer_obstack);
667 free (use_blocks);
670 /* Rewrites the non-virtual defs and uses into a loop closed ssa form. If
671 CHANGED_BBS is not NULL, we look for uses outside loops only in the basic
672 blocks in this set. UPDATE_FLAG is used in the call to update_ssa. See
673 TODO_update_ssa* for documentation. */
675 void
676 rewrite_into_loop_closed_ssa (bitmap changed_bbs, unsigned update_flag)
678 rewrite_into_loop_closed_ssa_1 (changed_bbs, update_flag, SSA_OP_USE, NULL);
681 /* Rewrites virtual defs and uses with def in LOOP into loop closed ssa
682 form. */
684 void
685 rewrite_virtuals_into_loop_closed_ssa (struct loop *loop)
687 rewrite_into_loop_closed_ssa_1 (NULL, 0, SSA_OP_VIRTUAL_USES, loop);
690 /* Check invariants of the loop closed ssa form for the USE in BB. */
692 static void
693 check_loop_closed_ssa_use (basic_block bb, tree use)
695 gimple *def;
696 basic_block def_bb;
698 if (TREE_CODE (use) != SSA_NAME || virtual_operand_p (use))
699 return;
701 def = SSA_NAME_DEF_STMT (use);
702 def_bb = gimple_bb (def);
703 gcc_assert (!def_bb
704 || flow_bb_inside_loop_p (def_bb->loop_father, bb));
707 /* Checks invariants of loop closed ssa form in statement STMT in BB. */
709 static void
710 check_loop_closed_ssa_stmt (basic_block bb, gimple *stmt)
712 ssa_op_iter iter;
713 tree var;
715 if (is_gimple_debug (stmt))
716 return;
718 FOR_EACH_SSA_TREE_OPERAND (var, stmt, iter, SSA_OP_USE)
719 check_loop_closed_ssa_use (bb, var);
722 /* Checks that invariants of the loop closed ssa form are preserved.
723 Call verify_ssa when VERIFY_SSA_P is true. */
725 DEBUG_FUNCTION void
726 verify_loop_closed_ssa (bool verify_ssa_p)
728 basic_block bb;
729 edge e;
730 edge_iterator ei;
732 if (number_of_loops (cfun) <= 1)
733 return;
735 if (verify_ssa_p)
736 verify_ssa (false, true);
738 timevar_push (TV_VERIFY_LOOP_CLOSED);
740 FOR_EACH_BB_FN (bb, cfun)
742 for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi);
743 gsi_next (&bsi))
745 gphi *phi = bsi.phi ();
746 FOR_EACH_EDGE (e, ei, bb->preds)
747 check_loop_closed_ssa_use (e->src,
748 PHI_ARG_DEF_FROM_EDGE (phi, e));
751 for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi);
752 gsi_next (&bsi))
753 check_loop_closed_ssa_stmt (bb, gsi_stmt (bsi));
756 timevar_pop (TV_VERIFY_LOOP_CLOSED);
759 /* Split loop exit edge EXIT. The things are a bit complicated by a need to
760 preserve the loop closed ssa form. The newly created block is returned. */
762 basic_block
763 split_loop_exit_edge (edge exit)
765 basic_block dest = exit->dest;
766 basic_block bb = split_edge (exit);
767 gphi *phi, *new_phi;
768 tree new_name, name;
769 use_operand_p op_p;
770 gphi_iterator psi;
771 source_location locus;
773 for (psi = gsi_start_phis (dest); !gsi_end_p (psi); gsi_next (&psi))
775 phi = psi.phi ();
776 op_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, single_succ_edge (bb));
777 locus = gimple_phi_arg_location_from_edge (phi, single_succ_edge (bb));
779 name = USE_FROM_PTR (op_p);
781 /* If the argument of the PHI node is a constant, we do not need
782 to keep it inside loop. */
783 if (TREE_CODE (name) != SSA_NAME)
784 continue;
786 /* Otherwise create an auxiliary phi node that will copy the value
787 of the SSA name out of the loop. */
788 new_name = duplicate_ssa_name (name, NULL);
789 new_phi = create_phi_node (new_name, bb);
790 add_phi_arg (new_phi, name, exit, locus);
791 SET_USE (op_p, new_name);
794 return bb;
797 /* Returns the basic block in that statements should be emitted for induction
798 variables incremented at the end of the LOOP. */
800 basic_block
801 ip_end_pos (struct loop *loop)
803 return loop->latch;
806 /* Returns the basic block in that statements should be emitted for induction
807 variables incremented just before exit condition of a LOOP. */
809 basic_block
810 ip_normal_pos (struct loop *loop)
812 gimple *last;
813 basic_block bb;
814 edge exit;
816 if (!single_pred_p (loop->latch))
817 return NULL;
819 bb = single_pred (loop->latch);
820 last = last_stmt (bb);
821 if (!last
822 || gimple_code (last) != GIMPLE_COND)
823 return NULL;
825 exit = EDGE_SUCC (bb, 0);
826 if (exit->dest == loop->latch)
827 exit = EDGE_SUCC (bb, 1);
829 if (flow_bb_inside_loop_p (loop, exit->dest))
830 return NULL;
832 return bb;
835 /* Stores the standard position for induction variable increment in LOOP
836 (just before the exit condition if it is available and latch block is empty,
837 end of the latch block otherwise) to BSI. INSERT_AFTER is set to true if
838 the increment should be inserted after *BSI. */
840 void
841 standard_iv_increment_position (struct loop *loop, gimple_stmt_iterator *bsi,
842 bool *insert_after)
844 basic_block bb = ip_normal_pos (loop), latch = ip_end_pos (loop);
845 gimple *last = last_stmt (latch);
847 if (!bb
848 || (last && gimple_code (last) != GIMPLE_LABEL))
850 *bsi = gsi_last_bb (latch);
851 *insert_after = true;
853 else
855 *bsi = gsi_last_bb (bb);
856 *insert_after = false;
860 /* Copies phi node arguments for duplicated blocks. The index of the first
861 duplicated block is FIRST_NEW_BLOCK. */
863 static void
864 copy_phi_node_args (unsigned first_new_block)
866 unsigned i;
868 for (i = first_new_block; i < (unsigned) last_basic_block_for_fn (cfun); i++)
869 BASIC_BLOCK_FOR_FN (cfun, i)->flags |= BB_DUPLICATED;
871 for (i = first_new_block; i < (unsigned) last_basic_block_for_fn (cfun); i++)
872 add_phi_args_after_copy_bb (BASIC_BLOCK_FOR_FN (cfun, i));
874 for (i = first_new_block; i < (unsigned) last_basic_block_for_fn (cfun); i++)
875 BASIC_BLOCK_FOR_FN (cfun, i)->flags &= ~BB_DUPLICATED;
879 /* The same as cfgloopmanip.c:duplicate_loop_to_header_edge, but also
880 updates the PHI nodes at start of the copied region. In order to
881 achieve this, only loops whose exits all lead to the same location
882 are handled.
884 Notice that we do not completely update the SSA web after
885 duplication. The caller is responsible for calling update_ssa
886 after the loop has been duplicated. */
888 bool
889 gimple_duplicate_loop_to_header_edge (struct loop *loop, edge e,
890 unsigned int ndupl, sbitmap wont_exit,
891 edge orig, vec<edge> *to_remove,
892 int flags)
894 unsigned first_new_block;
896 if (!loops_state_satisfies_p (LOOPS_HAVE_SIMPLE_LATCHES))
897 return false;
898 if (!loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS))
899 return false;
901 first_new_block = last_basic_block_for_fn (cfun);
902 if (!duplicate_loop_to_header_edge (loop, e, ndupl, wont_exit,
903 orig, to_remove, flags))
904 return false;
906 /* Readd the removed phi args for e. */
907 flush_pending_stmts (e);
909 /* Copy the phi node arguments. */
910 copy_phi_node_args (first_new_block);
912 scev_reset ();
914 return true;
917 /* Returns true if we can unroll LOOP FACTOR times. Number
918 of iterations of the loop is returned in NITER. */
920 bool
921 can_unroll_loop_p (struct loop *loop, unsigned factor,
922 struct tree_niter_desc *niter)
924 edge exit;
926 /* Check whether unrolling is possible. We only want to unroll loops
927 for that we are able to determine number of iterations. We also
928 want to split the extra iterations of the loop from its end,
929 therefore we require that the loop has precisely one
930 exit. */
932 exit = single_dom_exit (loop);
933 if (!exit)
934 return false;
936 if (!number_of_iterations_exit (loop, exit, niter, false)
937 || niter->cmp == ERROR_MARK
938 /* Scalar evolutions analysis might have copy propagated
939 the abnormal ssa names into these expressions, hence
940 emitting the computations based on them during loop
941 unrolling might create overlapping life ranges for
942 them, and failures in out-of-ssa. */
943 || contains_abnormal_ssa_name_p (niter->may_be_zero)
944 || contains_abnormal_ssa_name_p (niter->control.base)
945 || contains_abnormal_ssa_name_p (niter->control.step)
946 || contains_abnormal_ssa_name_p (niter->bound))
947 return false;
949 /* And of course, we must be able to duplicate the loop. */
950 if (!can_duplicate_loop_p (loop))
951 return false;
953 /* The final loop should be small enough. */
954 if (tree_num_loop_insns (loop, &eni_size_weights) * factor
955 > (unsigned) PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS))
956 return false;
958 return true;
961 /* Determines the conditions that control execution of LOOP unrolled FACTOR
962 times. DESC is number of iterations of LOOP. ENTER_COND is set to
963 condition that must be true if the main loop can be entered.
964 EXIT_BASE, EXIT_STEP, EXIT_CMP and EXIT_BOUND are set to values describing
965 how the exit from the unrolled loop should be controlled. */
967 static void
968 determine_exit_conditions (struct loop *loop, struct tree_niter_desc *desc,
969 unsigned factor, tree *enter_cond,
970 tree *exit_base, tree *exit_step,
971 enum tree_code *exit_cmp, tree *exit_bound)
973 gimple_seq stmts;
974 tree base = desc->control.base;
975 tree step = desc->control.step;
976 tree bound = desc->bound;
977 tree type = TREE_TYPE (step);
978 tree bigstep, delta;
979 tree min = lower_bound_in_type (type, type);
980 tree max = upper_bound_in_type (type, type);
981 enum tree_code cmp = desc->cmp;
982 tree cond = boolean_true_node, assum;
984 /* For pointers, do the arithmetics in the type of step. */
985 base = fold_convert (type, base);
986 bound = fold_convert (type, bound);
988 *enter_cond = boolean_false_node;
989 *exit_base = NULL_TREE;
990 *exit_step = NULL_TREE;
991 *exit_cmp = ERROR_MARK;
992 *exit_bound = NULL_TREE;
993 gcc_assert (cmp != ERROR_MARK);
995 /* We only need to be correct when we answer question
996 "Do at least FACTOR more iterations remain?" in the unrolled loop.
997 Thus, transforming BASE + STEP * i <> BOUND to
998 BASE + STEP * i < BOUND is ok. */
999 if (cmp == NE_EXPR)
1001 if (tree_int_cst_sign_bit (step))
1002 cmp = GT_EXPR;
1003 else
1004 cmp = LT_EXPR;
1006 else if (cmp == LT_EXPR)
1008 gcc_assert (!tree_int_cst_sign_bit (step));
1010 else if (cmp == GT_EXPR)
1012 gcc_assert (tree_int_cst_sign_bit (step));
1014 else
1015 gcc_unreachable ();
1017 /* The main body of the loop may be entered iff:
1019 1) desc->may_be_zero is false.
1020 2) it is possible to check that there are at least FACTOR iterations
1021 of the loop, i.e., BOUND - step * FACTOR does not overflow.
1022 3) # of iterations is at least FACTOR */
1024 if (!integer_zerop (desc->may_be_zero))
1025 cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
1026 invert_truthvalue (desc->may_be_zero),
1027 cond);
1029 bigstep = fold_build2 (MULT_EXPR, type, step,
1030 build_int_cst_type (type, factor));
1031 delta = fold_build2 (MINUS_EXPR, type, bigstep, step);
1032 if (cmp == LT_EXPR)
1033 assum = fold_build2 (GE_EXPR, boolean_type_node,
1034 bound,
1035 fold_build2 (PLUS_EXPR, type, min, delta));
1036 else
1037 assum = fold_build2 (LE_EXPR, boolean_type_node,
1038 bound,
1039 fold_build2 (PLUS_EXPR, type, max, delta));
1040 cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, assum, cond);
1042 bound = fold_build2 (MINUS_EXPR, type, bound, delta);
1043 assum = fold_build2 (cmp, boolean_type_node, base, bound);
1044 cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, assum, cond);
1046 cond = force_gimple_operand (unshare_expr (cond), &stmts, false, NULL_TREE);
1047 if (stmts)
1048 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
1049 /* cond now may be a gimple comparison, which would be OK, but also any
1050 other gimple rhs (say a && b). In this case we need to force it to
1051 operand. */
1052 if (!is_gimple_condexpr (cond))
1054 cond = force_gimple_operand (cond, &stmts, true, NULL_TREE);
1055 if (stmts)
1056 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
1058 *enter_cond = cond;
1060 base = force_gimple_operand (unshare_expr (base), &stmts, true, NULL_TREE);
1061 if (stmts)
1062 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
1063 bound = force_gimple_operand (unshare_expr (bound), &stmts, true, NULL_TREE);
1064 if (stmts)
1065 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
1067 *exit_base = base;
1068 *exit_step = bigstep;
1069 *exit_cmp = cmp;
1070 *exit_bound = bound;
1073 /* Scales the frequencies of all basic blocks in LOOP that are strictly
1074 dominated by BB by NUM/DEN. */
1076 static void
1077 scale_dominated_blocks_in_loop (struct loop *loop, basic_block bb,
1078 int num, int den)
1080 basic_block son;
1082 if (den == 0)
1083 return;
1085 for (son = first_dom_son (CDI_DOMINATORS, bb);
1086 son;
1087 son = next_dom_son (CDI_DOMINATORS, son))
1089 if (!flow_bb_inside_loop_p (loop, son))
1090 continue;
1091 scale_bbs_frequencies_int (&son, 1, num, den);
1092 scale_dominated_blocks_in_loop (loop, son, num, den);
1096 /* Unroll LOOP FACTOR times. DESC describes number of iterations of LOOP.
1097 EXIT is the exit of the loop to that DESC corresponds.
1099 If N is number of iterations of the loop and MAY_BE_ZERO is the condition
1100 under that loop exits in the first iteration even if N != 0,
1102 while (1)
1104 x = phi (init, next);
1106 pre;
1107 if (st)
1108 break;
1109 post;
1112 becomes (with possibly the exit conditions formulated a bit differently,
1113 avoiding the need to create a new iv):
1115 if (MAY_BE_ZERO || N < FACTOR)
1116 goto rest;
1120 x = phi (init, next);
1122 pre;
1123 post;
1124 pre;
1125 post;
1127 pre;
1128 post;
1129 N -= FACTOR;
1131 } while (N >= FACTOR);
1133 rest:
1134 init' = phi (init, x);
1136 while (1)
1138 x = phi (init', next);
1140 pre;
1141 if (st)
1142 break;
1143 post;
1146 Before the loop is unrolled, TRANSFORM is called for it (only for the
1147 unrolled loop, but not for its versioned copy). DATA is passed to
1148 TRANSFORM. */
1150 /* Probability in % that the unrolled loop is entered. Just a guess. */
1151 #define PROB_UNROLLED_LOOP_ENTERED 90
1153 void
1154 tree_transform_and_unroll_loop (struct loop *loop, unsigned factor,
1155 edge exit, struct tree_niter_desc *desc,
1156 transform_callback transform,
1157 void *data)
1159 gcond *exit_if;
1160 tree ctr_before, ctr_after;
1161 tree enter_main_cond, exit_base, exit_step, exit_bound;
1162 enum tree_code exit_cmp;
1163 gphi *phi_old_loop, *phi_new_loop, *phi_rest;
1164 gphi_iterator psi_old_loop, psi_new_loop;
1165 tree init, next, new_init;
1166 struct loop *new_loop;
1167 basic_block rest, exit_bb;
1168 edge old_entry, new_entry, old_latch, precond_edge, new_exit;
1169 edge new_nonexit, e;
1170 gimple_stmt_iterator bsi;
1171 use_operand_p op;
1172 bool ok;
1173 unsigned est_niter, prob_entry, scale_unrolled, scale_rest, freq_e, freq_h;
1174 unsigned new_est_niter, i, prob;
1175 unsigned irr = loop_preheader_edge (loop)->flags & EDGE_IRREDUCIBLE_LOOP;
1176 auto_vec<edge> to_remove;
1178 est_niter = expected_loop_iterations (loop);
1179 determine_exit_conditions (loop, desc, factor,
1180 &enter_main_cond, &exit_base, &exit_step,
1181 &exit_cmp, &exit_bound);
1183 /* Let us assume that the unrolled loop is quite likely to be entered. */
1184 if (integer_nonzerop (enter_main_cond))
1185 prob_entry = REG_BR_PROB_BASE;
1186 else
1187 prob_entry = PROB_UNROLLED_LOOP_ENTERED * REG_BR_PROB_BASE / 100;
1189 /* The values for scales should keep profile consistent, and somewhat close
1190 to correct.
1192 TODO: The current value of SCALE_REST makes it appear that the loop that
1193 is created by splitting the remaining iterations of the unrolled loop is
1194 executed the same number of times as the original loop, and with the same
1195 frequencies, which is obviously wrong. This does not appear to cause
1196 problems, so we do not bother with fixing it for now. To make the profile
1197 correct, we would need to change the probability of the exit edge of the
1198 loop, and recompute the distribution of frequencies in its body because
1199 of this change (scale the frequencies of blocks before and after the exit
1200 by appropriate factors). */
1201 scale_unrolled = prob_entry;
1202 scale_rest = REG_BR_PROB_BASE;
1204 new_loop = loop_version (loop, enter_main_cond, NULL,
1205 prob_entry, scale_unrolled, scale_rest, true);
1206 gcc_assert (new_loop != NULL);
1207 update_ssa (TODO_update_ssa);
1209 /* Determine the probability of the exit edge of the unrolled loop. */
1210 new_est_niter = est_niter / factor;
1212 /* Without profile feedback, loops for that we do not know a better estimate
1213 are assumed to roll 10 times. When we unroll such loop, it appears to
1214 roll too little, and it may even seem to be cold. To avoid this, we
1215 ensure that the created loop appears to roll at least 5 times (but at
1216 most as many times as before unrolling). */
1217 if (new_est_niter < 5)
1219 if (est_niter < 5)
1220 new_est_niter = est_niter;
1221 else
1222 new_est_niter = 5;
1225 /* Prepare the cfg and update the phi nodes. Move the loop exit to the
1226 loop latch (and make its condition dummy, for the moment). */
1227 rest = loop_preheader_edge (new_loop)->src;
1228 precond_edge = single_pred_edge (rest);
1229 split_edge (loop_latch_edge (loop));
1230 exit_bb = single_pred (loop->latch);
1232 /* Since the exit edge will be removed, the frequency of all the blocks
1233 in the loop that are dominated by it must be scaled by
1234 1 / (1 - exit->probability). */
1235 scale_dominated_blocks_in_loop (loop, exit->src,
1236 REG_BR_PROB_BASE,
1237 REG_BR_PROB_BASE - exit->probability);
1239 bsi = gsi_last_bb (exit_bb);
1240 exit_if = gimple_build_cond (EQ_EXPR, integer_zero_node,
1241 integer_zero_node,
1242 NULL_TREE, NULL_TREE);
1244 gsi_insert_after (&bsi, exit_if, GSI_NEW_STMT);
1245 new_exit = make_edge (exit_bb, rest, EDGE_FALSE_VALUE | irr);
1246 rescan_loop_exit (new_exit, true, false);
1248 /* Set the probability of new exit to the same of the old one. Fix
1249 the frequency of the latch block, by scaling it back by
1250 1 - exit->probability. */
1251 new_exit->count = exit->count;
1252 new_exit->probability = exit->probability;
1253 new_nonexit = single_pred_edge (loop->latch);
1254 new_nonexit->probability = REG_BR_PROB_BASE - exit->probability;
1255 new_nonexit->flags = EDGE_TRUE_VALUE;
1256 new_nonexit->count -= exit->count;
1257 if (new_nonexit->count < 0)
1258 new_nonexit->count = 0;
1259 scale_bbs_frequencies_int (&loop->latch, 1, new_nonexit->probability,
1260 REG_BR_PROB_BASE);
1262 old_entry = loop_preheader_edge (loop);
1263 new_entry = loop_preheader_edge (new_loop);
1264 old_latch = loop_latch_edge (loop);
1265 for (psi_old_loop = gsi_start_phis (loop->header),
1266 psi_new_loop = gsi_start_phis (new_loop->header);
1267 !gsi_end_p (psi_old_loop);
1268 gsi_next (&psi_old_loop), gsi_next (&psi_new_loop))
1270 phi_old_loop = psi_old_loop.phi ();
1271 phi_new_loop = psi_new_loop.phi ();
1273 init = PHI_ARG_DEF_FROM_EDGE (phi_old_loop, old_entry);
1274 op = PHI_ARG_DEF_PTR_FROM_EDGE (phi_new_loop, new_entry);
1275 gcc_assert (operand_equal_for_phi_arg_p (init, USE_FROM_PTR (op)));
1276 next = PHI_ARG_DEF_FROM_EDGE (phi_old_loop, old_latch);
1278 /* Prefer using original variable as a base for the new ssa name.
1279 This is necessary for virtual ops, and useful in order to avoid
1280 losing debug info for real ops. */
1281 if (TREE_CODE (next) == SSA_NAME
1282 && useless_type_conversion_p (TREE_TYPE (next),
1283 TREE_TYPE (init)))
1284 new_init = copy_ssa_name (next);
1285 else if (TREE_CODE (init) == SSA_NAME
1286 && useless_type_conversion_p (TREE_TYPE (init),
1287 TREE_TYPE (next)))
1288 new_init = copy_ssa_name (init);
1289 else if (useless_type_conversion_p (TREE_TYPE (next), TREE_TYPE (init)))
1290 new_init = make_temp_ssa_name (TREE_TYPE (next), NULL, "unrinittmp");
1291 else
1292 new_init = make_temp_ssa_name (TREE_TYPE (init), NULL, "unrinittmp");
1294 phi_rest = create_phi_node (new_init, rest);
1296 add_phi_arg (phi_rest, init, precond_edge, UNKNOWN_LOCATION);
1297 add_phi_arg (phi_rest, next, new_exit, UNKNOWN_LOCATION);
1298 SET_USE (op, new_init);
1301 remove_path (exit);
1303 /* Transform the loop. */
1304 if (transform)
1305 (*transform) (loop, data);
1307 /* Unroll the loop and remove the exits in all iterations except for the
1308 last one. */
1309 auto_sbitmap wont_exit (factor);
1310 bitmap_ones (wont_exit);
1311 bitmap_clear_bit (wont_exit, factor - 1);
1313 ok = gimple_duplicate_loop_to_header_edge
1314 (loop, loop_latch_edge (loop), factor - 1,
1315 wont_exit, new_exit, &to_remove, DLTHE_FLAG_UPDATE_FREQ);
1316 gcc_assert (ok);
1318 FOR_EACH_VEC_ELT (to_remove, i, e)
1320 ok = remove_path (e);
1321 gcc_assert (ok);
1323 update_ssa (TODO_update_ssa);
1325 /* Ensure that the frequencies in the loop match the new estimated
1326 number of iterations, and change the probability of the new
1327 exit edge. */
1328 freq_h = loop->header->frequency;
1329 freq_e = EDGE_FREQUENCY (loop_preheader_edge (loop));
1330 if (freq_h != 0)
1331 scale_loop_frequencies (loop, freq_e * (new_est_niter + 1), freq_h);
1333 exit_bb = single_pred (loop->latch);
1334 new_exit = find_edge (exit_bb, rest);
1335 new_exit->count = loop_preheader_edge (loop)->count;
1336 new_exit->probability = REG_BR_PROB_BASE / (new_est_niter + 1);
1338 rest->count += new_exit->count;
1339 rest->frequency += EDGE_FREQUENCY (new_exit);
1341 new_nonexit = single_pred_edge (loop->latch);
1342 prob = new_nonexit->probability;
1343 new_nonexit->probability = REG_BR_PROB_BASE - new_exit->probability;
1344 new_nonexit->count = exit_bb->count - new_exit->count;
1345 if (new_nonexit->count < 0)
1346 new_nonexit->count = 0;
1347 if (prob > 0)
1348 scale_bbs_frequencies_int (&loop->latch, 1, new_nonexit->probability,
1349 prob);
1351 /* Finally create the new counter for number of iterations and add the new
1352 exit instruction. */
1353 bsi = gsi_last_nondebug_bb (exit_bb);
1354 exit_if = as_a <gcond *> (gsi_stmt (bsi));
1355 create_iv (exit_base, exit_step, NULL_TREE, loop,
1356 &bsi, false, &ctr_before, &ctr_after);
1357 gimple_cond_set_code (exit_if, exit_cmp);
1358 gimple_cond_set_lhs (exit_if, ctr_after);
1359 gimple_cond_set_rhs (exit_if, exit_bound);
1360 update_stmt (exit_if);
1362 checking_verify_flow_info ();
1363 checking_verify_loop_structure ();
1364 checking_verify_loop_closed_ssa (true);
1367 /* Wrapper over tree_transform_and_unroll_loop for case we do not
1368 want to transform the loop before unrolling. The meaning
1369 of the arguments is the same as for tree_transform_and_unroll_loop. */
1371 void
1372 tree_unroll_loop (struct loop *loop, unsigned factor,
1373 edge exit, struct tree_niter_desc *desc)
1375 tree_transform_and_unroll_loop (loop, factor, exit, desc,
1376 NULL, NULL);
1379 /* Rewrite the phi node at position PSI in function of the main
1380 induction variable MAIN_IV and insert the generated code at GSI. */
1382 static void
1383 rewrite_phi_with_iv (loop_p loop,
1384 gphi_iterator *psi,
1385 gimple_stmt_iterator *gsi,
1386 tree main_iv)
1388 affine_iv iv;
1389 gassign *stmt;
1390 gphi *phi = psi->phi ();
1391 tree atype, mtype, val, res = PHI_RESULT (phi);
1393 if (virtual_operand_p (res) || res == main_iv)
1395 gsi_next (psi);
1396 return;
1399 if (!simple_iv (loop, loop, res, &iv, true))
1401 gsi_next (psi);
1402 return;
1405 remove_phi_node (psi, false);
1407 atype = TREE_TYPE (res);
1408 mtype = POINTER_TYPE_P (atype) ? sizetype : atype;
1409 val = fold_build2 (MULT_EXPR, mtype, unshare_expr (iv.step),
1410 fold_convert (mtype, main_iv));
1411 val = fold_build2 (POINTER_TYPE_P (atype)
1412 ? POINTER_PLUS_EXPR : PLUS_EXPR,
1413 atype, unshare_expr (iv.base), val);
1414 val = force_gimple_operand_gsi (gsi, val, false, NULL_TREE, true,
1415 GSI_SAME_STMT);
1416 stmt = gimple_build_assign (res, val);
1417 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
1420 /* Rewrite all the phi nodes of LOOP in function of the main induction
1421 variable MAIN_IV. */
1423 static void
1424 rewrite_all_phi_nodes_with_iv (loop_p loop, tree main_iv)
1426 unsigned i;
1427 basic_block *bbs = get_loop_body_in_dom_order (loop);
1428 gphi_iterator psi;
1430 for (i = 0; i < loop->num_nodes; i++)
1432 basic_block bb = bbs[i];
1433 gimple_stmt_iterator gsi = gsi_after_labels (bb);
1435 if (bb->loop_father != loop)
1436 continue;
1438 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); )
1439 rewrite_phi_with_iv (loop, &psi, &gsi, main_iv);
1442 free (bbs);
1445 /* Bases all the induction variables in LOOP on a single induction variable
1446 (with base 0 and step 1), whose final value is compared with *NIT. When the
1447 IV type precision has to be larger than *NIT type precision, *NIT is
1448 converted to the larger type, the conversion code is inserted before the
1449 loop, and *NIT is updated to the new definition. When BUMP_IN_LATCH is true,
1450 the induction variable is incremented in the loop latch, otherwise it is
1451 incremented in the loop header. Return the induction variable that was
1452 created. */
1454 tree
1455 canonicalize_loop_ivs (struct loop *loop, tree *nit, bool bump_in_latch)
1457 unsigned precision = TYPE_PRECISION (TREE_TYPE (*nit));
1458 unsigned original_precision = precision;
1459 tree type, var_before;
1460 gimple_stmt_iterator gsi;
1461 gphi_iterator psi;
1462 gcond *stmt;
1463 edge exit = single_dom_exit (loop);
1464 gimple_seq stmts;
1465 machine_mode mode;
1466 bool unsigned_p = false;
1468 for (psi = gsi_start_phis (loop->header);
1469 !gsi_end_p (psi); gsi_next (&psi))
1471 gphi *phi = psi.phi ();
1472 tree res = PHI_RESULT (phi);
1473 bool uns;
1475 type = TREE_TYPE (res);
1476 if (virtual_operand_p (res)
1477 || (!INTEGRAL_TYPE_P (type)
1478 && !POINTER_TYPE_P (type))
1479 || TYPE_PRECISION (type) < precision)
1480 continue;
1482 uns = POINTER_TYPE_P (type) | TYPE_UNSIGNED (type);
1484 if (TYPE_PRECISION (type) > precision)
1485 unsigned_p = uns;
1486 else
1487 unsigned_p |= uns;
1489 precision = TYPE_PRECISION (type);
1492 mode = smallest_mode_for_size (precision, MODE_INT);
1493 precision = GET_MODE_PRECISION (mode);
1494 type = build_nonstandard_integer_type (precision, unsigned_p);
1496 if (original_precision != precision)
1498 *nit = fold_convert (type, *nit);
1499 *nit = force_gimple_operand (*nit, &stmts, true, NULL_TREE);
1500 if (stmts)
1501 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
1504 if (bump_in_latch)
1505 gsi = gsi_last_bb (loop->latch);
1506 else
1507 gsi = gsi_last_nondebug_bb (loop->header);
1508 create_iv (build_int_cst_type (type, 0), build_int_cst (type, 1), NULL_TREE,
1509 loop, &gsi, bump_in_latch, &var_before, NULL);
1511 rewrite_all_phi_nodes_with_iv (loop, var_before);
1513 stmt = as_a <gcond *> (last_stmt (exit->src));
1514 /* Make the loop exit if the control condition is not satisfied. */
1515 if (exit->flags & EDGE_TRUE_VALUE)
1517 edge te, fe;
1519 extract_true_false_edges_from_block (exit->src, &te, &fe);
1520 te->flags = EDGE_FALSE_VALUE;
1521 fe->flags = EDGE_TRUE_VALUE;
1523 gimple_cond_set_code (stmt, LT_EXPR);
1524 gimple_cond_set_lhs (stmt, var_before);
1525 gimple_cond_set_rhs (stmt, *nit);
1526 update_stmt (stmt);
1528 return var_before;