mips.c (build_mips16_call_stub): On 64-bit targets, combine an SC return value into...
[official-gcc.git] / gcc / tree-ssa-loop-manip.c
blob91aac3eb354410985e1e61f701031cf28e3e49ad
1 /* High-level loop manipulation functions.
2 Copyright (C) 2004, 2005, 2006, 2007 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 "tm.h"
24 #include "tree.h"
25 #include "rtl.h"
26 #include "tm_p.h"
27 #include "hard-reg-set.h"
28 #include "basic-block.h"
29 #include "output.h"
30 #include "diagnostic.h"
31 #include "tree-flow.h"
32 #include "tree-dump.h"
33 #include "timevar.h"
34 #include "cfgloop.h"
35 #include "tree-pass.h"
36 #include "cfglayout.h"
37 #include "tree-scalar-evolution.h"
38 #include "params.h"
39 #include "tree-inline.h"
41 /* Creates an induction variable with value BASE + STEP * iteration in LOOP.
42 It is expected that neither BASE nor STEP are shared with other expressions
43 (unless the sharing rules allow this). Use VAR as a base var_decl for it
44 (if NULL, a new temporary will be created). The increment will occur at
45 INCR_POS (after it if AFTER is true, before it otherwise). INCR_POS and
46 AFTER can be computed using standard_iv_increment_position. The ssa versions
47 of the variable before and after increment will be stored in VAR_BEFORE and
48 VAR_AFTER (unless they are NULL). */
50 void
51 create_iv (tree base, tree step, tree var, struct loop *loop,
52 block_stmt_iterator *incr_pos, bool after,
53 tree *var_before, tree *var_after)
55 tree stmt, initial, step1, stmts;
56 tree vb, va;
57 enum tree_code incr_op = PLUS_EXPR;
58 edge pe = loop_preheader_edge (loop);
60 if (!var)
62 var = create_tmp_var (TREE_TYPE (base), "ivtmp");
63 add_referenced_var (var);
66 vb = make_ssa_name (var, NULL_TREE);
67 if (var_before)
68 *var_before = vb;
69 va = make_ssa_name (var, NULL_TREE);
70 if (var_after)
71 *var_after = va;
73 /* For easier readability of the created code, produce MINUS_EXPRs
74 when suitable. */
75 if (TREE_CODE (step) == INTEGER_CST)
77 if (TYPE_UNSIGNED (TREE_TYPE (step)))
79 step1 = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
80 if (tree_int_cst_lt (step1, step))
82 incr_op = MINUS_EXPR;
83 step = step1;
86 else
88 bool ovf;
90 if (!tree_expr_nonnegative_warnv_p (step, &ovf)
91 && may_negate_without_overflow_p (step))
93 incr_op = MINUS_EXPR;
94 step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
98 if (POINTER_TYPE_P (TREE_TYPE (base)))
100 step = fold_convert (sizetype, step);
101 if (incr_op == MINUS_EXPR)
102 step = fold_build1 (NEGATE_EXPR, sizetype, step);
103 incr_op = POINTER_PLUS_EXPR;
105 /* Gimplify the step if necessary. We put the computations in front of the
106 loop (i.e. the step should be loop invariant). */
107 step = force_gimple_operand (step, &stmts, true, var);
108 if (stmts)
109 bsi_insert_on_edge_immediate (pe, stmts);
111 stmt = build_gimple_modify_stmt (va,
112 build2 (incr_op, TREE_TYPE (base),
113 vb, step));
114 SSA_NAME_DEF_STMT (va) = stmt;
115 if (after)
116 bsi_insert_after (incr_pos, stmt, BSI_NEW_STMT);
117 else
118 bsi_insert_before (incr_pos, stmt, BSI_NEW_STMT);
120 initial = force_gimple_operand (base, &stmts, true, var);
121 if (stmts)
122 bsi_insert_on_edge_immediate (pe, stmts);
124 stmt = create_phi_node (vb, loop->header);
125 SSA_NAME_DEF_STMT (vb) = stmt;
126 add_phi_arg (stmt, initial, loop_preheader_edge (loop));
127 add_phi_arg (stmt, va, loop_latch_edge (loop));
130 /* Add exit phis for the USE on EXIT. */
132 static void
133 add_exit_phis_edge (basic_block exit, tree use)
135 tree phi, def_stmt = SSA_NAME_DEF_STMT (use);
136 basic_block def_bb = bb_for_stmt (def_stmt);
137 struct loop *def_loop;
138 edge e;
139 edge_iterator ei;
141 /* Check that some of the edges entering the EXIT block exits a loop in
142 that USE is defined. */
143 FOR_EACH_EDGE (e, ei, exit->preds)
145 def_loop = find_common_loop (def_bb->loop_father, e->src->loop_father);
146 if (!flow_bb_inside_loop_p (def_loop, e->dest))
147 break;
150 if (!e)
151 return;
153 phi = create_phi_node (use, exit);
154 create_new_def_for (PHI_RESULT (phi), phi, PHI_RESULT_PTR (phi));
155 FOR_EACH_EDGE (e, ei, exit->preds)
156 add_phi_arg (phi, use, e);
159 /* Add exit phis for VAR that is used in LIVEIN.
160 Exits of the loops are stored in EXITS. */
162 static void
163 add_exit_phis_var (tree var, bitmap livein, bitmap exits)
165 bitmap def;
166 unsigned index;
167 basic_block def_bb = bb_for_stmt (SSA_NAME_DEF_STMT (var));
168 bitmap_iterator bi;
170 if (is_gimple_reg (var))
171 bitmap_clear_bit (livein, def_bb->index);
172 else
173 bitmap_set_bit (livein, def_bb->index);
175 def = BITMAP_ALLOC (NULL);
176 bitmap_set_bit (def, def_bb->index);
177 compute_global_livein (livein, def);
178 BITMAP_FREE (def);
180 EXECUTE_IF_AND_IN_BITMAP (exits, livein, 0, index, bi)
182 add_exit_phis_edge (BASIC_BLOCK (index), var);
186 /* Add exit phis for the names marked in NAMES_TO_RENAME.
187 Exits of the loops are stored in EXITS. Sets of blocks where the ssa
188 names are used are stored in USE_BLOCKS. */
190 static void
191 add_exit_phis (bitmap names_to_rename, bitmap *use_blocks, bitmap loop_exits)
193 unsigned i;
194 bitmap_iterator bi;
196 EXECUTE_IF_SET_IN_BITMAP (names_to_rename, 0, i, bi)
198 add_exit_phis_var (ssa_name (i), use_blocks[i], loop_exits);
202 /* Returns a bitmap of all loop exit edge targets. */
204 static bitmap
205 get_loops_exits (void)
207 bitmap exits = BITMAP_ALLOC (NULL);
208 basic_block bb;
209 edge e;
210 edge_iterator ei;
212 FOR_EACH_BB (bb)
214 FOR_EACH_EDGE (e, ei, bb->preds)
215 if (e->src != ENTRY_BLOCK_PTR
216 && !flow_bb_inside_loop_p (e->src->loop_father, bb))
218 bitmap_set_bit (exits, bb->index);
219 break;
223 return exits;
226 /* For USE in BB, if it is used outside of the loop it is defined in,
227 mark it for rewrite. Record basic block BB where it is used
228 to USE_BLOCKS. Record the ssa name index to NEED_PHIS bitmap. */
230 static void
231 find_uses_to_rename_use (basic_block bb, tree use, bitmap *use_blocks,
232 bitmap need_phis)
234 unsigned ver;
235 basic_block def_bb;
236 struct loop *def_loop;
238 if (TREE_CODE (use) != SSA_NAME)
239 return;
241 /* We don't need to keep virtual operands in loop-closed form. */
242 if (!is_gimple_reg (use))
243 return;
245 ver = SSA_NAME_VERSION (use);
246 def_bb = bb_for_stmt (SSA_NAME_DEF_STMT (use));
247 if (!def_bb)
248 return;
249 def_loop = def_bb->loop_father;
251 /* If the definition is not inside loop, it is not interesting. */
252 if (!loop_outer (def_loop))
253 return;
255 if (!use_blocks[ver])
256 use_blocks[ver] = BITMAP_ALLOC (NULL);
257 bitmap_set_bit (use_blocks[ver], bb->index);
259 bitmap_set_bit (need_phis, ver);
262 /* For uses in STMT, mark names that are used outside of the loop they are
263 defined to rewrite. Record the set of blocks in that the ssa
264 names are defined to USE_BLOCKS and the ssa names themselves to
265 NEED_PHIS. */
267 static void
268 find_uses_to_rename_stmt (tree stmt, bitmap *use_blocks, bitmap need_phis)
270 ssa_op_iter iter;
271 tree var;
272 basic_block bb = bb_for_stmt (stmt);
274 FOR_EACH_SSA_TREE_OPERAND (var, stmt, iter, SSA_OP_ALL_USES)
275 find_uses_to_rename_use (bb, var, use_blocks, need_phis);
278 /* Marks names that are used in BB and outside of the loop they are
279 defined in for rewrite. Records the set of blocks in that the ssa
280 names are defined to USE_BLOCKS. Record the SSA names that will
281 need exit PHIs in NEED_PHIS. */
283 static void
284 find_uses_to_rename_bb (basic_block bb, bitmap *use_blocks, bitmap need_phis)
286 block_stmt_iterator bsi;
287 edge e;
288 edge_iterator ei;
289 tree phi;
291 FOR_EACH_EDGE (e, ei, bb->succs)
292 for (phi = phi_nodes (e->dest); phi; phi = PHI_CHAIN (phi))
293 find_uses_to_rename_use (bb, PHI_ARG_DEF_FROM_EDGE (phi, e),
294 use_blocks, need_phis);
296 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
297 find_uses_to_rename_stmt (bsi_stmt (bsi), use_blocks, need_phis);
300 /* Marks names that are used outside of the loop they are defined in
301 for rewrite. Records the set of blocks in that the ssa
302 names are defined to USE_BLOCKS. If CHANGED_BBS is not NULL,
303 scan only blocks in this set. */
305 static void
306 find_uses_to_rename (bitmap changed_bbs, bitmap *use_blocks, bitmap need_phis)
308 basic_block bb;
309 unsigned index;
310 bitmap_iterator bi;
312 if (changed_bbs && !bitmap_empty_p (changed_bbs))
314 EXECUTE_IF_SET_IN_BITMAP (changed_bbs, 0, index, bi)
316 find_uses_to_rename_bb (BASIC_BLOCK (index), use_blocks, need_phis);
319 else
321 FOR_EACH_BB (bb)
323 find_uses_to_rename_bb (bb, use_blocks, need_phis);
328 /* Rewrites the program into a loop closed ssa form -- i.e. inserts extra
329 phi nodes to ensure that no variable is used outside the loop it is
330 defined in.
332 This strengthening of the basic ssa form has several advantages:
334 1) Updating it during unrolling/peeling/versioning is trivial, since
335 we do not need to care about the uses outside of the loop.
336 2) The behavior of all uses of an induction variable is the same.
337 Without this, you need to distinguish the case when the variable
338 is used outside of the loop it is defined in, for example
340 for (i = 0; i < 100; i++)
342 for (j = 0; j < 100; j++)
344 k = i + j;
345 use1 (k);
347 use2 (k);
350 Looking from the outer loop with the normal SSA form, the first use of k
351 is not well-behaved, while the second one is an induction variable with
352 base 99 and step 1.
354 If CHANGED_BBS is not NULL, we look for uses outside loops only in
355 the basic blocks in this set.
357 UPDATE_FLAG is used in the call to update_ssa. See
358 TODO_update_ssa* for documentation. */
360 void
361 rewrite_into_loop_closed_ssa (bitmap changed_bbs, unsigned update_flag)
363 bitmap loop_exits;
364 bitmap *use_blocks;
365 unsigned i, old_num_ssa_names;
366 bitmap names_to_rename;
368 loops_state_set (LOOP_CLOSED_SSA);
369 if (number_of_loops () <= 1)
370 return;
372 loop_exits = get_loops_exits ();
373 names_to_rename = BITMAP_ALLOC (NULL);
375 /* If the pass has caused the SSA form to be out-of-date, update it
376 now. */
377 update_ssa (update_flag);
379 old_num_ssa_names = num_ssa_names;
380 use_blocks = XCNEWVEC (bitmap, old_num_ssa_names);
382 /* Find the uses outside loops. */
383 find_uses_to_rename (changed_bbs, use_blocks, names_to_rename);
385 /* Add the PHI nodes on exits of the loops for the names we need to
386 rewrite. */
387 add_exit_phis (names_to_rename, use_blocks, loop_exits);
389 for (i = 0; i < old_num_ssa_names; i++)
390 BITMAP_FREE (use_blocks[i]);
391 free (use_blocks);
392 BITMAP_FREE (loop_exits);
393 BITMAP_FREE (names_to_rename);
395 /* Fix up all the names found to be used outside their original
396 loops. */
397 update_ssa (TODO_update_ssa);
400 /* Check invariants of the loop closed ssa form for the USE in BB. */
402 static void
403 check_loop_closed_ssa_use (basic_block bb, tree use)
405 tree def;
406 basic_block def_bb;
408 if (TREE_CODE (use) != SSA_NAME || !is_gimple_reg (use))
409 return;
411 def = SSA_NAME_DEF_STMT (use);
412 def_bb = bb_for_stmt (def);
413 gcc_assert (!def_bb
414 || flow_bb_inside_loop_p (def_bb->loop_father, bb));
417 /* Checks invariants of loop closed ssa form in statement STMT in BB. */
419 static void
420 check_loop_closed_ssa_stmt (basic_block bb, tree stmt)
422 ssa_op_iter iter;
423 tree var;
425 FOR_EACH_SSA_TREE_OPERAND (var, stmt, iter, SSA_OP_ALL_USES)
426 check_loop_closed_ssa_use (bb, var);
429 /* Checks that invariants of the loop closed ssa form are preserved. */
431 void
432 verify_loop_closed_ssa (void)
434 basic_block bb;
435 block_stmt_iterator bsi;
436 tree phi;
437 unsigned i;
439 if (number_of_loops () <= 1)
440 return;
442 verify_ssa (false);
444 FOR_EACH_BB (bb)
446 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
447 for (i = 0; i < (unsigned) PHI_NUM_ARGS (phi); i++)
448 check_loop_closed_ssa_use (PHI_ARG_EDGE (phi, i)->src,
449 PHI_ARG_DEF (phi, i));
451 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
452 check_loop_closed_ssa_stmt (bb, bsi_stmt (bsi));
456 /* Split loop exit edge EXIT. The things are a bit complicated by a need to
457 preserve the loop closed ssa form. The newly created block is returned. */
459 basic_block
460 split_loop_exit_edge (edge exit)
462 basic_block dest = exit->dest;
463 basic_block bb = split_edge (exit);
464 tree phi, new_phi, new_name, name;
465 use_operand_p op_p;
467 for (phi = phi_nodes (dest); phi; phi = PHI_CHAIN (phi))
469 op_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, single_succ_edge (bb));
471 name = USE_FROM_PTR (op_p);
473 /* If the argument of the PHI node is a constant, we do not need
474 to keep it inside loop. */
475 if (TREE_CODE (name) != SSA_NAME)
476 continue;
478 /* Otherwise create an auxiliary phi node that will copy the value
479 of the SSA name out of the loop. */
480 new_name = duplicate_ssa_name (name, NULL);
481 new_phi = create_phi_node (new_name, bb);
482 SSA_NAME_DEF_STMT (new_name) = new_phi;
483 add_phi_arg (new_phi, name, exit);
484 SET_USE (op_p, new_name);
487 return bb;
490 /* Returns the basic block in that statements should be emitted for induction
491 variables incremented at the end of the LOOP. */
493 basic_block
494 ip_end_pos (struct loop *loop)
496 return loop->latch;
499 /* Returns the basic block in that statements should be emitted for induction
500 variables incremented just before exit condition of a LOOP. */
502 basic_block
503 ip_normal_pos (struct loop *loop)
505 tree last;
506 basic_block bb;
507 edge exit;
509 if (!single_pred_p (loop->latch))
510 return NULL;
512 bb = single_pred (loop->latch);
513 last = last_stmt (bb);
514 if (!last
515 || TREE_CODE (last) != COND_EXPR)
516 return NULL;
518 exit = EDGE_SUCC (bb, 0);
519 if (exit->dest == loop->latch)
520 exit = EDGE_SUCC (bb, 1);
522 if (flow_bb_inside_loop_p (loop, exit->dest))
523 return NULL;
525 return bb;
528 /* Stores the standard position for induction variable increment in LOOP
529 (just before the exit condition if it is available and latch block is empty,
530 end of the latch block otherwise) to BSI. INSERT_AFTER is set to true if
531 the increment should be inserted after *BSI. */
533 void
534 standard_iv_increment_position (struct loop *loop, block_stmt_iterator *bsi,
535 bool *insert_after)
537 basic_block bb = ip_normal_pos (loop), latch = ip_end_pos (loop);
538 tree last = last_stmt (latch);
540 if (!bb
541 || (last && TREE_CODE (last) != LABEL_EXPR))
543 *bsi = bsi_last (latch);
544 *insert_after = true;
546 else
548 *bsi = bsi_last (bb);
549 *insert_after = false;
553 /* Copies phi node arguments for duplicated blocks. The index of the first
554 duplicated block is FIRST_NEW_BLOCK. */
556 static void
557 copy_phi_node_args (unsigned first_new_block)
559 unsigned i;
561 for (i = first_new_block; i < (unsigned) last_basic_block; i++)
562 BASIC_BLOCK (i)->flags |= BB_DUPLICATED;
564 for (i = first_new_block; i < (unsigned) last_basic_block; i++)
565 add_phi_args_after_copy_bb (BASIC_BLOCK (i));
567 for (i = first_new_block; i < (unsigned) last_basic_block; i++)
568 BASIC_BLOCK (i)->flags &= ~BB_DUPLICATED;
572 /* The same as cfgloopmanip.c:duplicate_loop_to_header_edge, but also
573 updates the PHI nodes at start of the copied region. In order to
574 achieve this, only loops whose exits all lead to the same location
575 are handled.
577 Notice that we do not completely update the SSA web after
578 duplication. The caller is responsible for calling update_ssa
579 after the loop has been duplicated. */
581 bool
582 tree_duplicate_loop_to_header_edge (struct loop *loop, edge e,
583 unsigned int ndupl, sbitmap wont_exit,
584 edge orig, VEC (edge, heap) **to_remove,
585 int flags)
587 unsigned first_new_block;
589 if (!loops_state_satisfies_p (LOOPS_HAVE_SIMPLE_LATCHES))
590 return false;
591 if (!loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS))
592 return false;
594 #ifdef ENABLE_CHECKING
595 verify_loop_closed_ssa ();
596 #endif
598 first_new_block = last_basic_block;
599 if (!duplicate_loop_to_header_edge (loop, e, ndupl, wont_exit,
600 orig, to_remove, flags))
601 return false;
603 /* Readd the removed phi args for e. */
604 flush_pending_stmts (e);
606 /* Copy the phi node arguments. */
607 copy_phi_node_args (first_new_block);
609 scev_reset ();
611 return true;
614 /* Returns true if we can unroll LOOP FACTOR times. Number
615 of iterations of the loop is returned in NITER. */
617 bool
618 can_unroll_loop_p (struct loop *loop, unsigned factor,
619 struct tree_niter_desc *niter)
621 edge exit;
623 /* Check whether unrolling is possible. We only want to unroll loops
624 for that we are able to determine number of iterations. We also
625 want to split the extra iterations of the loop from its end,
626 therefore we require that the loop has precisely one
627 exit. */
629 exit = single_dom_exit (loop);
630 if (!exit)
631 return false;
633 if (!number_of_iterations_exit (loop, exit, niter, false)
634 || niter->cmp == ERROR_MARK
635 /* Scalar evolutions analysis might have copy propagated
636 the abnormal ssa names into these expressions, hence
637 emitting the computations based on them during loop
638 unrolling might create overlapping life ranges for
639 them, and failures in out-of-ssa. */
640 || contains_abnormal_ssa_name_p (niter->may_be_zero)
641 || contains_abnormal_ssa_name_p (niter->control.base)
642 || contains_abnormal_ssa_name_p (niter->control.step)
643 || contains_abnormal_ssa_name_p (niter->bound))
644 return false;
646 /* And of course, we must be able to duplicate the loop. */
647 if (!can_duplicate_loop_p (loop))
648 return false;
650 /* The final loop should be small enough. */
651 if (tree_num_loop_insns (loop, &eni_size_weights) * factor
652 > (unsigned) PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS))
653 return false;
655 return true;
658 /* Determines the conditions that control execution of LOOP unrolled FACTOR
659 times. DESC is number of iterations of LOOP. ENTER_COND is set to
660 condition that must be true if the main loop can be entered.
661 EXIT_BASE, EXIT_STEP, EXIT_CMP and EXIT_BOUND are set to values describing
662 how the exit from the unrolled loop should be controlled. */
664 static void
665 determine_exit_conditions (struct loop *loop, struct tree_niter_desc *desc,
666 unsigned factor, tree *enter_cond,
667 tree *exit_base, tree *exit_step,
668 enum tree_code *exit_cmp, tree *exit_bound)
670 tree stmts;
671 tree base = desc->control.base;
672 tree step = desc->control.step;
673 tree bound = desc->bound;
674 tree type = TREE_TYPE (step);
675 tree bigstep, delta;
676 tree min = lower_bound_in_type (type, type);
677 tree max = upper_bound_in_type (type, type);
678 enum tree_code cmp = desc->cmp;
679 tree cond = boolean_true_node, assum;
681 /* For pointers, do the arithmetics in the type of step (sizetype). */
682 base = fold_convert (type, base);
683 bound = fold_convert (type, bound);
685 *enter_cond = boolean_false_node;
686 *exit_base = NULL_TREE;
687 *exit_step = NULL_TREE;
688 *exit_cmp = ERROR_MARK;
689 *exit_bound = NULL_TREE;
690 gcc_assert (cmp != ERROR_MARK);
692 /* We only need to be correct when we answer question
693 "Do at least FACTOR more iterations remain?" in the unrolled loop.
694 Thus, transforming BASE + STEP * i <> BOUND to
695 BASE + STEP * i < BOUND is ok. */
696 if (cmp == NE_EXPR)
698 if (tree_int_cst_sign_bit (step))
699 cmp = GT_EXPR;
700 else
701 cmp = LT_EXPR;
703 else if (cmp == LT_EXPR)
705 gcc_assert (!tree_int_cst_sign_bit (step));
707 else if (cmp == GT_EXPR)
709 gcc_assert (tree_int_cst_sign_bit (step));
711 else
712 gcc_unreachable ();
714 /* The main body of the loop may be entered iff:
716 1) desc->may_be_zero is false.
717 2) it is possible to check that there are at least FACTOR iterations
718 of the loop, i.e., BOUND - step * FACTOR does not overflow.
719 3) # of iterations is at least FACTOR */
721 if (!integer_zerop (desc->may_be_zero))
722 cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
723 invert_truthvalue (desc->may_be_zero),
724 cond);
726 bigstep = fold_build2 (MULT_EXPR, type, step,
727 build_int_cst_type (type, factor));
728 delta = fold_build2 (MINUS_EXPR, type, bigstep, step);
729 if (cmp == LT_EXPR)
730 assum = fold_build2 (GE_EXPR, boolean_type_node,
731 bound,
732 fold_build2 (PLUS_EXPR, type, min, delta));
733 else
734 assum = fold_build2 (LE_EXPR, boolean_type_node,
735 bound,
736 fold_build2 (PLUS_EXPR, type, max, delta));
737 cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, assum, cond);
739 bound = fold_build2 (MINUS_EXPR, type, bound, delta);
740 assum = fold_build2 (cmp, boolean_type_node, base, bound);
741 cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, assum, cond);
743 cond = force_gimple_operand (unshare_expr (cond), &stmts, false, NULL_TREE);
744 if (stmts)
745 bsi_insert_on_edge_immediate (loop_preheader_edge (loop), stmts);
746 /* cond now may be a gimple comparison, which would be OK, but also any
747 other gimple rhs (say a && b). In this case we need to force it to
748 operand. */
749 if (!is_gimple_condexpr (cond))
751 cond = force_gimple_operand (cond, &stmts, true, NULL_TREE);
752 if (stmts)
753 bsi_insert_on_edge_immediate (loop_preheader_edge (loop), stmts);
755 *enter_cond = cond;
757 base = force_gimple_operand (unshare_expr (base), &stmts, true, NULL_TREE);
758 if (stmts)
759 bsi_insert_on_edge_immediate (loop_preheader_edge (loop), stmts);
760 bound = force_gimple_operand (unshare_expr (bound), &stmts, true, NULL_TREE);
761 if (stmts)
762 bsi_insert_on_edge_immediate (loop_preheader_edge (loop), stmts);
764 *exit_base = base;
765 *exit_step = bigstep;
766 *exit_cmp = cmp;
767 *exit_bound = bound;
770 /* Scales the frequencies of all basic blocks in LOOP that are strictly
771 dominated by BB by NUM/DEN. */
773 static void
774 scale_dominated_blocks_in_loop (struct loop *loop, basic_block bb,
775 int num, int den)
777 basic_block son;
779 if (den == 0)
780 return;
782 for (son = first_dom_son (CDI_DOMINATORS, bb);
783 son;
784 son = next_dom_son (CDI_DOMINATORS, son))
786 if (!flow_bb_inside_loop_p (loop, son))
787 continue;
788 scale_bbs_frequencies_int (&son, 1, num, den);
789 scale_dominated_blocks_in_loop (loop, son, num, den);
793 /* Unroll LOOP FACTOR times. DESC describes number of iterations of LOOP.
794 EXIT is the exit of the loop to that DESC corresponds.
796 If N is number of iterations of the loop and MAY_BE_ZERO is the condition
797 under that loop exits in the first iteration even if N != 0,
799 while (1)
801 x = phi (init, next);
803 pre;
804 if (st)
805 break;
806 post;
809 becomes (with possibly the exit conditions formulated a bit differently,
810 avoiding the need to create a new iv):
812 if (MAY_BE_ZERO || N < FACTOR)
813 goto rest;
817 x = phi (init, next);
819 pre;
820 post;
821 pre;
822 post;
824 pre;
825 post;
826 N -= FACTOR;
828 } while (N >= FACTOR);
830 rest:
831 init' = phi (init, x);
833 while (1)
835 x = phi (init', next);
837 pre;
838 if (st)
839 break;
840 post;
843 Before the loop is unrolled, TRANSFORM is called for it (only for the
844 unrolled loop, but not for its versioned copy). DATA is passed to
845 TRANSFORM. */
847 /* Probability in % that the unrolled loop is entered. Just a guess. */
848 #define PROB_UNROLLED_LOOP_ENTERED 90
850 void
851 tree_transform_and_unroll_loop (struct loop *loop, unsigned factor,
852 edge exit, struct tree_niter_desc *desc,
853 transform_callback transform,
854 void *data)
856 tree exit_if, ctr_before, ctr_after;
857 tree enter_main_cond, exit_base, exit_step, exit_bound;
858 enum tree_code exit_cmp;
859 tree phi_old_loop, phi_new_loop, phi_rest, init, next, new_init, var;
860 struct loop *new_loop;
861 basic_block rest, exit_bb;
862 edge old_entry, new_entry, old_latch, precond_edge, new_exit;
863 edge new_nonexit, e;
864 block_stmt_iterator bsi;
865 use_operand_p op;
866 bool ok;
867 unsigned est_niter, prob_entry, scale_unrolled, scale_rest, freq_e, freq_h;
868 unsigned new_est_niter, i, prob;
869 unsigned irr = loop_preheader_edge (loop)->flags & EDGE_IRREDUCIBLE_LOOP;
870 sbitmap wont_exit;
871 VEC (edge, heap) *to_remove = NULL;
873 est_niter = expected_loop_iterations (loop);
874 determine_exit_conditions (loop, desc, factor,
875 &enter_main_cond, &exit_base, &exit_step,
876 &exit_cmp, &exit_bound);
878 /* Let us assume that the unrolled loop is quite likely to be entered. */
879 if (integer_nonzerop (enter_main_cond))
880 prob_entry = REG_BR_PROB_BASE;
881 else
882 prob_entry = PROB_UNROLLED_LOOP_ENTERED * REG_BR_PROB_BASE / 100;
884 /* The values for scales should keep profile consistent, and somewhat close
885 to correct.
887 TODO: The current value of SCALE_REST makes it appear that the loop that
888 is created by splitting the remaining iterations of the unrolled loop is
889 executed the same number of times as the original loop, and with the same
890 frequencies, which is obviously wrong. This does not appear to cause
891 problems, so we do not bother with fixing it for now. To make the profile
892 correct, we would need to change the probability of the exit edge of the
893 loop, and recompute the distribution of frequencies in its body because
894 of this change (scale the frequencies of blocks before and after the exit
895 by appropriate factors). */
896 scale_unrolled = prob_entry;
897 scale_rest = REG_BR_PROB_BASE;
899 new_loop = loop_version (loop, enter_main_cond, NULL,
900 prob_entry, scale_unrolled, scale_rest, true);
901 gcc_assert (new_loop != NULL);
902 update_ssa (TODO_update_ssa);
904 /* Determine the probability of the exit edge of the unrolled loop. */
905 new_est_niter = est_niter / factor;
907 /* Without profile feedback, loops for that we do not know a better estimate
908 are assumed to roll 10 times. When we unroll such loop, it appears to
909 roll too little, and it may even seem to be cold. To avoid this, we
910 ensure that the created loop appears to roll at least 5 times (but at
911 most as many times as before unrolling). */
912 if (new_est_niter < 5)
914 if (est_niter < 5)
915 new_est_niter = est_niter;
916 else
917 new_est_niter = 5;
920 /* Prepare the cfg and update the phi nodes. Move the loop exit to the
921 loop latch (and make its condition dummy, for the moment). */
922 rest = loop_preheader_edge (new_loop)->src;
923 precond_edge = single_pred_edge (rest);
924 split_edge (loop_latch_edge (loop));
925 exit_bb = single_pred (loop->latch);
927 /* Since the exit edge will be removed, the frequency of all the blocks
928 in the loop that are dominated by it must be scaled by
929 1 / (1 - exit->probability). */
930 scale_dominated_blocks_in_loop (loop, exit->src,
931 REG_BR_PROB_BASE,
932 REG_BR_PROB_BASE - exit->probability);
934 bsi = bsi_last (exit_bb);
935 exit_if = build3 (COND_EXPR, void_type_node, boolean_true_node,
936 NULL_TREE, NULL_TREE);
938 bsi_insert_after (&bsi, exit_if, BSI_NEW_STMT);
939 new_exit = make_edge (exit_bb, rest, EDGE_FALSE_VALUE | irr);
940 rescan_loop_exit (new_exit, true, false);
942 /* Set the probability of new exit to the same of the old one. Fix
943 the frequency of the latch block, by scaling it back by
944 1 - exit->probability. */
945 new_exit->count = exit->count;
946 new_exit->probability = exit->probability;
947 new_nonexit = single_pred_edge (loop->latch);
948 new_nonexit->probability = REG_BR_PROB_BASE - exit->probability;
949 new_nonexit->flags = EDGE_TRUE_VALUE;
950 new_nonexit->count -= exit->count;
951 if (new_nonexit->count < 0)
952 new_nonexit->count = 0;
953 scale_bbs_frequencies_int (&loop->latch, 1, new_nonexit->probability,
954 REG_BR_PROB_BASE);
956 old_entry = loop_preheader_edge (loop);
957 new_entry = loop_preheader_edge (new_loop);
958 old_latch = loop_latch_edge (loop);
959 for (phi_old_loop = phi_nodes (loop->header),
960 phi_new_loop = phi_nodes (new_loop->header);
961 phi_old_loop;
962 phi_old_loop = PHI_CHAIN (phi_old_loop),
963 phi_new_loop = PHI_CHAIN (phi_new_loop))
965 init = PHI_ARG_DEF_FROM_EDGE (phi_old_loop, old_entry);
966 op = PHI_ARG_DEF_PTR_FROM_EDGE (phi_new_loop, new_entry);
967 gcc_assert (operand_equal_for_phi_arg_p (init, USE_FROM_PTR (op)));
968 next = PHI_ARG_DEF_FROM_EDGE (phi_old_loop, old_latch);
970 /* Prefer using original variable as a base for the new ssa name.
971 This is necessary for virtual ops, and useful in order to avoid
972 losing debug info for real ops. */
973 if (TREE_CODE (next) == SSA_NAME)
974 var = SSA_NAME_VAR (next);
975 else if (TREE_CODE (init) == SSA_NAME)
976 var = SSA_NAME_VAR (init);
977 else
979 var = create_tmp_var (TREE_TYPE (init), "unrinittmp");
980 add_referenced_var (var);
983 new_init = make_ssa_name (var, NULL_TREE);
984 phi_rest = create_phi_node (new_init, rest);
985 SSA_NAME_DEF_STMT (new_init) = phi_rest;
987 add_phi_arg (phi_rest, init, precond_edge);
988 add_phi_arg (phi_rest, next, new_exit);
989 SET_USE (op, new_init);
992 remove_path (exit);
994 /* Transform the loop. */
995 if (transform)
996 (*transform) (loop, data);
998 /* Unroll the loop and remove the exits in all iterations except for the
999 last one. */
1000 wont_exit = sbitmap_alloc (factor);
1001 sbitmap_ones (wont_exit);
1002 RESET_BIT (wont_exit, factor - 1);
1004 ok = tree_duplicate_loop_to_header_edge
1005 (loop, loop_latch_edge (loop), factor - 1,
1006 wont_exit, new_exit, &to_remove, DLTHE_FLAG_UPDATE_FREQ);
1007 free (wont_exit);
1008 gcc_assert (ok);
1010 for (i = 0; VEC_iterate (edge, to_remove, i, e); i++)
1012 ok = remove_path (e);
1013 gcc_assert (ok);
1015 VEC_free (edge, heap, to_remove);
1016 update_ssa (TODO_update_ssa);
1018 /* Ensure that the frequencies in the loop match the new estimated
1019 number of iterations, and change the probability of the new
1020 exit edge. */
1021 freq_h = loop->header->frequency;
1022 freq_e = EDGE_FREQUENCY (loop_preheader_edge (loop));
1023 if (freq_h != 0)
1024 scale_loop_frequencies (loop, freq_e * (new_est_niter + 1), freq_h);
1026 exit_bb = single_pred (loop->latch);
1027 new_exit = find_edge (exit_bb, rest);
1028 new_exit->count = loop_preheader_edge (loop)->count;
1029 new_exit->probability = REG_BR_PROB_BASE / (new_est_niter + 1);
1031 rest->count += new_exit->count;
1032 rest->frequency += EDGE_FREQUENCY (new_exit);
1034 new_nonexit = single_pred_edge (loop->latch);
1035 prob = new_nonexit->probability;
1036 new_nonexit->probability = REG_BR_PROB_BASE - new_exit->probability;
1037 new_nonexit->count = exit_bb->count - new_exit->count;
1038 if (new_nonexit->count < 0)
1039 new_nonexit->count = 0;
1040 if (prob > 0)
1041 scale_bbs_frequencies_int (&loop->latch, 1, new_nonexit->probability,
1042 prob);
1044 /* Finally create the new counter for number of iterations and add the new
1045 exit instruction. */
1046 bsi = bsi_last (exit_bb);
1047 exit_if = bsi_stmt (bsi);
1048 create_iv (exit_base, exit_step, NULL_TREE, loop,
1049 &bsi, false, &ctr_before, &ctr_after);
1050 COND_EXPR_COND (exit_if) = build2 (exit_cmp, boolean_type_node, ctr_after,
1051 exit_bound);
1052 update_stmt (exit_if);
1054 #ifdef ENABLE_CHECKING
1055 verify_flow_info ();
1056 verify_dominators (CDI_DOMINATORS);
1057 verify_loop_structure ();
1058 verify_loop_closed_ssa ();
1059 #endif
1062 /* Wrapper over tree_transform_and_unroll_loop for case we do not
1063 want to transform the loop before unrolling. The meaning
1064 of the arguments is the same as for tree_transform_and_unroll_loop. */
1066 void
1067 tree_unroll_loop (struct loop *loop, unsigned factor,
1068 edge exit, struct tree_niter_desc *desc)
1070 tree_transform_and_unroll_loop (loop, factor, exit, desc,
1071 NULL, NULL);