basic-block.h (edge_list): Prefix member names with "m_".
[gcc.git] / gcc / tree-tailcall.c
1 /* Tail call optimization on trees.
2 Copyright (C) 2003-2013 Free Software Foundation, Inc.
3
4 This file is part of GCC.
5
6 GCC is free software; you can redistribute it and/or modify
7 it under the terms of the GNU General Public License as published by
8 the Free Software Foundation; either version 3, or (at your option)
9 any later version.
10
11 GCC is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 GNU General Public License for more details.
15
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/>. */
19
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "tm.h"
24 #include "tree.h"
25 #include "tm_p.h"
26 #include "basic-block.h"
27 #include "function.h"
28 #include "tree-ssa.h"
29 #include "gimple-pretty-print.h"
30 #include "except.h"
31 #include "tree-pass.h"
32 #include "flags.h"
33 #include "langhooks.h"
34 #include "dbgcnt.h"
35 #include "target.h"
36 #include "cfgloop.h"
37 #include "common/common-target.h"
38 #include "ipa-utils.h"
39
40 /* The file implements the tail recursion elimination. It is also used to
41 analyze the tail calls in general, passing the results to the rtl level
42 where they are used for sibcall optimization.
43
44 In addition to the standard tail recursion elimination, we handle the most
45 trivial cases of making the call tail recursive by creating accumulators.
46 For example the following function
47
48 int sum (int n)
49 {
50 if (n > 0)
51 return n + sum (n - 1);
52 else
53 return 0;
54 }
55
56 is transformed into
57
58 int sum (int n)
59 {
60 int acc = 0;
61
62 while (n > 0)
63 acc += n--;
64
65 return acc;
66 }
67
68 To do this, we maintain two accumulators (a_acc and m_acc) that indicate
69 when we reach the return x statement, we should return a_acc + x * m_acc
70 instead. They are initially initialized to 0 and 1, respectively,
71 so the semantics of the function is obviously preserved. If we are
72 guaranteed that the value of the accumulator never change, we
73 omit the accumulator.
74
75 There are three cases how the function may exit. The first one is
76 handled in adjust_return_value, the other two in adjust_accumulator_values
77 (the second case is actually a special case of the third one and we
78 present it separately just for clarity):
79
80 1) Just return x, where x is not in any of the remaining special shapes.
81 We rewrite this to a gimple equivalent of return m_acc * x + a_acc.
82
83 2) return f (...), where f is the current function, is rewritten in a
84 classical tail-recursion elimination way, into assignment of arguments
85 and jump to the start of the function. Values of the accumulators
86 are unchanged.
87
88 3) return a + m * f(...), where a and m do not depend on call to f.
89 To preserve the semantics described before we want this to be rewritten
90 in such a way that we finally return
91
92 a_acc + (a + m * f(...)) * m_acc = (a_acc + a * m_acc) + (m * m_acc) * f(...).
93
94 I.e. we increase a_acc by a * m_acc, multiply m_acc by m and
95 eliminate the tail call to f. Special cases when the value is just
96 added or just multiplied are obtained by setting a = 0 or m = 1.
97
98 TODO -- it is possible to do similar tricks for other operations. */
99
100 /* A structure that describes the tailcall. */
101
102 struct tailcall
103 {
104 /* The iterator pointing to the call statement. */
105 gimple_stmt_iterator call_gsi;
106
107 /* True if it is a call to the current function. */
108 bool tail_recursion;
109
110 /* The return value of the caller is mult * f + add, where f is the return
111 value of the call. */
112 tree mult, add;
113
114 /* Next tailcall in the chain. */
115 struct tailcall *next;
116 };
117
118 /* The variables holding the value of multiplicative and additive
119 accumulator. */
120 static tree m_acc, a_acc;
121
122 static bool suitable_for_tail_opt_p (void);
123 static bool optimize_tail_call (struct tailcall *, bool);
124 static void eliminate_tail_call (struct tailcall *);
125 static void find_tail_calls (basic_block, struct tailcall **);
126
127 /* Returns false when the function is not suitable for tail call optimization
128 from some reason (e.g. if it takes variable number of arguments). */
129
130 static bool
131 suitable_for_tail_opt_p (void)
132 {
133 if (cfun->stdarg)
134 return false;
135
136 return true;
137 }
138 /* Returns false when the function is not suitable for tail call optimization
139 from some reason (e.g. if it takes variable number of arguments).
140 This test must pass in addition to suitable_for_tail_opt_p in order to make
141 tail call discovery happen. */
142
143 static bool
144 suitable_for_tail_call_opt_p (void)
145 {
146 tree param;
147
148 /* alloca (until we have stack slot life analysis) inhibits
149 sibling call optimizations, but not tail recursion. */
150 if (cfun->calls_alloca)
151 return false;
152
153 /* If we are using sjlj exceptions, we may need to add a call to
154 _Unwind_SjLj_Unregister at exit of the function. Which means
155 that we cannot do any sibcall transformations. */
156 if (targetm_common.except_unwind_info (&global_options) == UI_SJLJ
157 && current_function_has_exception_handlers ())
158 return false;
159
160 /* Any function that calls setjmp might have longjmp called from
161 any called function. ??? We really should represent this
162 properly in the CFG so that this needn't be special cased. */
163 if (cfun->calls_setjmp)
164 return false;
165
166 /* ??? It is OK if the argument of a function is taken in some cases,
167 but not in all cases. See PR15387 and PR19616. Revisit for 4.1. */
168 for (param = DECL_ARGUMENTS (current_function_decl);
169 param;
170 param = DECL_CHAIN (param))
171 if (TREE_ADDRESSABLE (param))
172 return false;
173
174 return true;
175 }
176
177 /* Checks whether the expression EXPR in stmt AT is independent of the
178 statement pointed to by GSI (in a sense that we already know EXPR's value
179 at GSI). We use the fact that we are only called from the chain of
180 basic blocks that have only single successor. Returns the expression
181 containing the value of EXPR at GSI. */
182
183 static tree
184 independent_of_stmt_p (tree expr, gimple at, gimple_stmt_iterator gsi)
185 {
186 basic_block bb, call_bb, at_bb;
187 edge e;
188 edge_iterator ei;
189
190 if (is_gimple_min_invariant (expr))
191 return expr;
192
193 if (TREE_CODE (expr) != SSA_NAME)
194 return NULL_TREE;
195
196 /* Mark the blocks in the chain leading to the end. */
197 at_bb = gimple_bb (at);
198 call_bb = gimple_bb (gsi_stmt (gsi));
199 for (bb = call_bb; bb != at_bb; bb = single_succ (bb))
200 bb->aux = &bb->aux;
201 bb->aux = &bb->aux;
202
203 while (1)
204 {
205 at = SSA_NAME_DEF_STMT (expr);
206 bb = gimple_bb (at);
207
208 /* The default definition or defined before the chain. */
209 if (!bb || !bb->aux)
210 break;
211
212 if (bb == call_bb)
213 {
214 for (; !gsi_end_p (gsi); gsi_next (&gsi))
215 if (gsi_stmt (gsi) == at)
216 break;
217
218 if (!gsi_end_p (gsi))
219 expr = NULL_TREE;
220 break;
221 }
222
223 if (gimple_code (at) != GIMPLE_PHI)
224 {
225 expr = NULL_TREE;
226 break;
227 }
228
229 FOR_EACH_EDGE (e, ei, bb->preds)
230 if (e->src->aux)
231 break;
232 gcc_assert (e);
233
234 expr = PHI_ARG_DEF_FROM_EDGE (at, e);
235 if (TREE_CODE (expr) != SSA_NAME)
236 {
237 /* The value is a constant. */
238 break;
239 }
240 }
241
242 /* Unmark the blocks. */
243 for (bb = call_bb; bb != at_bb; bb = single_succ (bb))
244 bb->aux = NULL;
245 bb->aux = NULL;
246
247 return expr;
248 }
249
250 /* Simulates the effect of an assignment STMT on the return value of the tail
251 recursive CALL passed in ASS_VAR. M and A are the multiplicative and the
252 additive factor for the real return value. */
253
254 static bool
255 process_assignment (gimple stmt, gimple_stmt_iterator call, tree *m,
256 tree *a, tree *ass_var)
257 {
258 tree op0, op1 = NULL_TREE, non_ass_var = NULL_TREE;
259 tree dest = gimple_assign_lhs (stmt);
260 enum tree_code code = gimple_assign_rhs_code (stmt);
261 enum gimple_rhs_class rhs_class = get_gimple_rhs_class (code);
262 tree src_var = gimple_assign_rhs1 (stmt);
263
264 /* See if this is a simple copy operation of an SSA name to the function
265 result. In that case we may have a simple tail call. Ignore type
266 conversions that can never produce extra code between the function
267 call and the function return. */
268 if ((rhs_class == GIMPLE_SINGLE_RHS || gimple_assign_cast_p (stmt))
269 && (TREE_CODE (src_var) == SSA_NAME))
270 {
271 /* Reject a tailcall if the type conversion might need
272 additional code. */
273 if (gimple_assign_cast_p (stmt)
274 && TYPE_MODE (TREE_TYPE (dest)) != TYPE_MODE (TREE_TYPE (src_var)))
275 return false;
276
277 if (src_var != *ass_var)
278 return false;
279
280 *ass_var = dest;
281 return true;
282 }
283
284 switch (rhs_class)
285 {
286 case GIMPLE_BINARY_RHS:
287 op1 = gimple_assign_rhs2 (stmt);
288
289 /* Fall through. */
290
291 case GIMPLE_UNARY_RHS:
292 op0 = gimple_assign_rhs1 (stmt);
293 break;
294
295 default:
296 return false;
297 }
298
299 /* Accumulator optimizations will reverse the order of operations.
300 We can only do that for floating-point types if we're assuming
301 that addition and multiplication are associative. */
302 if (!flag_associative_math)
303 if (FLOAT_TYPE_P (TREE_TYPE (DECL_RESULT (current_function_decl))))
304 return false;
305
306 if (rhs_class == GIMPLE_UNARY_RHS)
307 ;
308 else if (op0 == *ass_var
309 && (non_ass_var = independent_of_stmt_p (op1, stmt, call)))
310 ;
311 else if (op1 == *ass_var
312 && (non_ass_var = independent_of_stmt_p (op0, stmt, call)))
313 ;
314 else
315 return false;
316
317 switch (code)
318 {
319 case PLUS_EXPR:
320 *a = non_ass_var;
321 *ass_var = dest;
322 return true;
323
324 case POINTER_PLUS_EXPR:
325 if (op0 != *ass_var)
326 return false;
327 *a = non_ass_var;
328 *ass_var = dest;
329 return true;
330
331 case MULT_EXPR:
332 *m = non_ass_var;
333 *ass_var = dest;
334 return true;
335
336 case NEGATE_EXPR:
337 *m = build_minus_one_cst (TREE_TYPE (op0));
338 *ass_var = dest;
339 return true;
340
341 case MINUS_EXPR:
342 if (*ass_var == op0)
343 *a = fold_build1 (NEGATE_EXPR, TREE_TYPE (non_ass_var), non_ass_var);
344 else
345 {
346 *m = build_minus_one_cst (TREE_TYPE (non_ass_var));
347 *a = fold_build1 (NEGATE_EXPR, TREE_TYPE (non_ass_var), non_ass_var);
348 }
349
350 *ass_var = dest;
351 return true;
352
353 /* TODO -- Handle POINTER_PLUS_EXPR. */
354
355 default:
356 return false;
357 }
358 }
359
360 /* Propagate VAR through phis on edge E. */
361
362 static tree
363 propagate_through_phis (tree var, edge e)
364 {
365 basic_block dest = e->dest;
366 gimple_stmt_iterator gsi;
367
368 for (gsi = gsi_start_phis (dest); !gsi_end_p (gsi); gsi_next (&gsi))
369 {
370 gimple phi = gsi_stmt (gsi);
371 if (PHI_ARG_DEF_FROM_EDGE (phi, e) == var)
372 return PHI_RESULT (phi);
373 }
374 return var;
375 }
376
377 /* Finds tailcalls falling into basic block BB. The list of found tailcalls is
378 added to the start of RET. */
379
380 static void
381 find_tail_calls (basic_block bb, struct tailcall **ret)
382 {
383 tree ass_var = NULL_TREE, ret_var, func, param;
384 gimple stmt, call = NULL;
385 gimple_stmt_iterator gsi, agsi;
386 bool tail_recursion;
387 struct tailcall *nw;
388 edge e;
389 tree m, a;
390 basic_block abb;
391 size_t idx;
392 tree var;
393
394 if (!single_succ_p (bb))
395 return;
396
397 for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi); gsi_prev (&gsi))
398 {
399 stmt = gsi_stmt (gsi);
400
401 /* Ignore labels, returns, clobbers and debug stmts. */
402 if (gimple_code (stmt) == GIMPLE_LABEL
403 || gimple_code (stmt) == GIMPLE_RETURN
404 || gimple_clobber_p (stmt)
405 || is_gimple_debug (stmt))
406 continue;
407
408 /* Check for a call. */
409 if (is_gimple_call (stmt))
410 {
411 call = stmt;
412 ass_var = gimple_call_lhs (stmt);
413 break;
414 }
415
416 /* If the statement references memory or volatile operands, fail. */
417 if (gimple_references_memory_p (stmt)
418 || gimple_has_volatile_ops (stmt))
419 return;
420 }
421
422 if (gsi_end_p (gsi))
423 {
424 edge_iterator ei;
425 /* Recurse to the predecessors. */
426 FOR_EACH_EDGE (e, ei, bb->preds)
427 find_tail_calls (e->src, ret);
428
429 return;
430 }
431
432 /* If the LHS of our call is not just a simple register, we can't
433 transform this into a tail or sibling call. This situation happens,
434 in (e.g.) "*p = foo()" where foo returns a struct. In this case
435 we won't have a temporary here, but we need to carry out the side
436 effect anyway, so tailcall is impossible.
437
438 ??? In some situations (when the struct is returned in memory via
439 invisible argument) we could deal with this, e.g. by passing 'p'
440 itself as that argument to foo, but it's too early to do this here,
441 and expand_call() will not handle it anyway. If it ever can, then
442 we need to revisit this here, to allow that situation. */
443 if (ass_var && !is_gimple_reg (ass_var))
444 return;
445
446 /* We found the call, check whether it is suitable. */
447 tail_recursion = false;
448 func = gimple_call_fndecl (call);
449 if (func && recursive_call_p (current_function_decl, func))
450 {
451 tree arg;
452
453 for (param = DECL_ARGUMENTS (func), idx = 0;
454 param && idx < gimple_call_num_args (call);
455 param = DECL_CHAIN (param), idx ++)
456 {
457 arg = gimple_call_arg (call, idx);
458 if (param != arg)
459 {
460 /* Make sure there are no problems with copying. The parameter
461 have a copyable type and the two arguments must have reasonably
462 equivalent types. The latter requirement could be relaxed if
463 we emitted a suitable type conversion statement. */
464 if (!is_gimple_reg_type (TREE_TYPE (param))
465 || !useless_type_conversion_p (TREE_TYPE (param),
466 TREE_TYPE (arg)))
467 break;
468
469 /* The parameter should be a real operand, so that phi node
470 created for it at the start of the function has the meaning
471 of copying the value. This test implies is_gimple_reg_type
472 from the previous condition, however this one could be
473 relaxed by being more careful with copying the new value
474 of the parameter (emitting appropriate GIMPLE_ASSIGN and
475 updating the virtual operands). */
476 if (!is_gimple_reg (param))
477 break;
478 }
479 }
480 if (idx == gimple_call_num_args (call) && !param)
481 tail_recursion = true;
482 }
483
484 /* Make sure the tail invocation of this function does not refer
485 to local variables. */
486 FOR_EACH_LOCAL_DECL (cfun, idx, var)
487 {
488 if (TREE_CODE (var) != PARM_DECL
489 && auto_var_in_fn_p (var, cfun->decl)
490 && (ref_maybe_used_by_stmt_p (call, var)
491 || call_may_clobber_ref_p (call, var)))
492 return;
493 }
494
495 /* Now check the statements after the call. None of them has virtual
496 operands, so they may only depend on the call through its return
497 value. The return value should also be dependent on each of them,
498 since we are running after dce. */
499 m = NULL_TREE;
500 a = NULL_TREE;
501
502 abb = bb;
503 agsi = gsi;
504 while (1)
505 {
506 tree tmp_a = NULL_TREE;
507 tree tmp_m = NULL_TREE;
508 gsi_next (&agsi);
509
510 while (gsi_end_p (agsi))
511 {
512 ass_var = propagate_through_phis (ass_var, single_succ_edge (abb));
513 abb = single_succ (abb);
514 agsi = gsi_start_bb (abb);
515 }
516
517 stmt = gsi_stmt (agsi);
518
519 if (gimple_code (stmt) == GIMPLE_LABEL)
520 continue;
521
522 if (gimple_code (stmt) == GIMPLE_RETURN)
523 break;
524
525 if (gimple_clobber_p (stmt))
526 continue;
527
528 if (is_gimple_debug (stmt))
529 continue;
530
531 if (gimple_code (stmt) != GIMPLE_ASSIGN)
532 return;
533
534 /* This is a gimple assign. */
535 if (! process_assignment (stmt, gsi, &tmp_m, &tmp_a, &ass_var))
536 return;
537
538 if (tmp_a)
539 {
540 tree type = TREE_TYPE (tmp_a);
541 if (a)
542 a = fold_build2 (PLUS_EXPR, type, fold_convert (type, a), tmp_a);
543 else
544 a = tmp_a;
545 }
546 if (tmp_m)
547 {
548 tree type = TREE_TYPE (tmp_m);
549 if (m)
550 m = fold_build2 (MULT_EXPR, type, fold_convert (type, m), tmp_m);
551 else
552 m = tmp_m;
553
554 if (a)
555 a = fold_build2 (MULT_EXPR, type, fold_convert (type, a), tmp_m);
556 }
557 }
558
559 /* See if this is a tail call we can handle. */
560 ret_var = gimple_return_retval (stmt);
561
562 /* We may proceed if there either is no return value, or the return value
563 is identical to the call's return. */
564 if (ret_var
565 && (ret_var != ass_var))
566 return;
567
568 /* If this is not a tail recursive call, we cannot handle addends or
569 multiplicands. */
570 if (!tail_recursion && (m || a))
571 return;
572
573 /* For pointers only allow additions. */
574 if (m && POINTER_TYPE_P (TREE_TYPE (DECL_RESULT (current_function_decl))))
575 return;
576
577 nw = XNEW (struct tailcall);
578
579 nw->call_gsi = gsi;
580
581 nw->tail_recursion = tail_recursion;
582
583 nw->mult = m;
584 nw->add = a;
585
586 nw->next = *ret;
587 *ret = nw;
588 }
589
590 /* Helper to insert PHI_ARGH to the phi of VAR in the destination of edge E. */
591
592 static void
593 add_successor_phi_arg (edge e, tree var, tree phi_arg)
594 {
595 gimple_stmt_iterator gsi;
596
597 for (gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi); gsi_next (&gsi))
598 if (PHI_RESULT (gsi_stmt (gsi)) == var)
599 break;
600
601 gcc_assert (!gsi_end_p (gsi));
602 add_phi_arg (gsi_stmt (gsi), phi_arg, e, UNKNOWN_LOCATION);
603 }
604
605 /* Creates a GIMPLE statement which computes the operation specified by
606 CODE, ACC and OP1 to a new variable with name LABEL and inserts the
607 statement in the position specified by GSI. Returns the
608 tree node of the statement's result. */
609
610 static tree
611 adjust_return_value_with_ops (enum tree_code code, const char *label,
612 tree acc, tree op1, gimple_stmt_iterator gsi)
613 {
614
615 tree ret_type = TREE_TYPE (DECL_RESULT (current_function_decl));
616 tree result = make_temp_ssa_name (ret_type, NULL, label);
617 gimple stmt;
618
619 if (POINTER_TYPE_P (ret_type))
620 {
621 gcc_assert (code == PLUS_EXPR && TREE_TYPE (acc) == sizetype);
622 code = POINTER_PLUS_EXPR;
623 }
624 if (types_compatible_p (TREE_TYPE (acc), TREE_TYPE (op1))
625 && code != POINTER_PLUS_EXPR)
626 stmt = gimple_build_assign_with_ops (code, result, acc, op1);
627 else
628 {
629 tree tem;
630 if (code == POINTER_PLUS_EXPR)
631 tem = fold_build2 (code, TREE_TYPE (op1), op1, acc);
632 else
633 tem = fold_build2 (code, TREE_TYPE (op1),
634 fold_convert (TREE_TYPE (op1), acc), op1);
635 tree rhs = fold_convert (ret_type, tem);
636 rhs = force_gimple_operand_gsi (&gsi, rhs,
637 false, NULL, true, GSI_SAME_STMT);
638 stmt = gimple_build_assign (result, rhs);
639 }
640
641 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
642 return result;
643 }
644
645 /* Creates a new GIMPLE statement that adjusts the value of accumulator ACC by
646 the computation specified by CODE and OP1 and insert the statement
647 at the position specified by GSI as a new statement. Returns new SSA name
648 of updated accumulator. */
649
650 static tree
651 update_accumulator_with_ops (enum tree_code code, tree acc, tree op1,
652 gimple_stmt_iterator gsi)
653 {
654 gimple stmt;
655 tree var = copy_ssa_name (acc, NULL);
656 if (types_compatible_p (TREE_TYPE (acc), TREE_TYPE (op1)))
657 stmt = gimple_build_assign_with_ops (code, var, acc, op1);
658 else
659 {
660 tree rhs = fold_convert (TREE_TYPE (acc),
661 fold_build2 (code,
662 TREE_TYPE (op1),
663 fold_convert (TREE_TYPE (op1), acc),
664 op1));
665 rhs = force_gimple_operand_gsi (&gsi, rhs,
666 false, NULL, false, GSI_CONTINUE_LINKING);
667 stmt = gimple_build_assign (var, rhs);
668 }
669 gsi_insert_after (&gsi, stmt, GSI_NEW_STMT);
670 return var;
671 }
672
673 /* Adjust the accumulator values according to A and M after GSI, and update
674 the phi nodes on edge BACK. */
675
676 static void
677 adjust_accumulator_values (gimple_stmt_iterator gsi, tree m, tree a, edge back)
678 {
679 tree var, a_acc_arg, m_acc_arg;
680
681 if (m)
682 m = force_gimple_operand_gsi (&gsi, m, true, NULL, true, GSI_SAME_STMT);
683 if (a)
684 a = force_gimple_operand_gsi (&gsi, a, true, NULL, true, GSI_SAME_STMT);
685
686 a_acc_arg = a_acc;
687 m_acc_arg = m_acc;
688 if (a)
689 {
690 if (m_acc)
691 {
692 if (integer_onep (a))
693 var = m_acc;
694 else
695 var = adjust_return_value_with_ops (MULT_EXPR, "acc_tmp", m_acc,
696 a, gsi);
697 }
698 else
699 var = a;
700
701 a_acc_arg = update_accumulator_with_ops (PLUS_EXPR, a_acc, var, gsi);
702 }
703
704 if (m)
705 m_acc_arg = update_accumulator_with_ops (MULT_EXPR, m_acc, m, gsi);
706
707 if (a_acc)
708 add_successor_phi_arg (back, a_acc, a_acc_arg);
709
710 if (m_acc)
711 add_successor_phi_arg (back, m_acc, m_acc_arg);
712 }
713
714 /* Adjust value of the return at the end of BB according to M and A
715 accumulators. */
716
717 static void
718 adjust_return_value (basic_block bb, tree m, tree a)
719 {
720 tree retval;
721 gimple ret_stmt = gimple_seq_last_stmt (bb_seq (bb));
722 gimple_stmt_iterator gsi = gsi_last_bb (bb);
723
724 gcc_assert (gimple_code (ret_stmt) == GIMPLE_RETURN);
725
726 retval = gimple_return_retval (ret_stmt);
727 if (!retval || retval == error_mark_node)
728 return;
729
730 if (m)
731 retval = adjust_return_value_with_ops (MULT_EXPR, "mul_tmp", m_acc, retval,
732 gsi);
733 if (a)
734 retval = adjust_return_value_with_ops (PLUS_EXPR, "acc_tmp", a_acc, retval,
735 gsi);
736 gimple_return_set_retval (ret_stmt, retval);
737 update_stmt (ret_stmt);
738 }
739
740 /* Subtract COUNT and FREQUENCY from the basic block and it's
741 outgoing edge. */
742 static void
743 decrease_profile (basic_block bb, gcov_type count, int frequency)
744 {
745 edge e;
746 bb->count -= count;
747 if (bb->count < 0)
748 bb->count = 0;
749 bb->frequency -= frequency;
750 if (bb->frequency < 0)
751 bb->frequency = 0;
752 if (!single_succ_p (bb))
753 {
754 gcc_assert (!EDGE_COUNT (bb->succs));
755 return;
756 }
757 e = single_succ_edge (bb);
758 e->count -= count;
759 if (e->count < 0)
760 e->count = 0;
761 }
762
763 /* Returns true if argument PARAM of the tail recursive call needs to be copied
764 when the call is eliminated. */
765
766 static bool
767 arg_needs_copy_p (tree param)
768 {
769 tree def;
770
771 if (!is_gimple_reg (param))
772 return false;
773
774 /* Parameters that are only defined but never used need not be copied. */
775 def = ssa_default_def (cfun, param);
776 if (!def)
777 return false;
778
779 return true;
780 }
781
782 /* Eliminates tail call described by T. TMP_VARS is a list of
783 temporary variables used to copy the function arguments. */
784
785 static void
786 eliminate_tail_call (struct tailcall *t)
787 {
788 tree param, rslt;
789 gimple stmt, call;
790 tree arg;
791 size_t idx;
792 basic_block bb, first;
793 edge e;
794 gimple phi;
795 gimple_stmt_iterator gsi;
796 gimple orig_stmt;
797
798 stmt = orig_stmt = gsi_stmt (t->call_gsi);
799 bb = gsi_bb (t->call_gsi);
800
801 if (dump_file && (dump_flags & TDF_DETAILS))
802 {
803 fprintf (dump_file, "Eliminated tail recursion in bb %d : ",
804 bb->index);
805 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
806 fprintf (dump_file, "\n");
807 }
808
809 gcc_assert (is_gimple_call (stmt));
810
811 first = single_succ (ENTRY_BLOCK_PTR);
812
813 /* Remove the code after call_gsi that will become unreachable. The
814 possibly unreachable code in other blocks is removed later in
815 cfg cleanup. */
816 gsi = t->call_gsi;
817 gsi_next (&gsi);
818 while (!gsi_end_p (gsi))
819 {
820 gimple t = gsi_stmt (gsi);
821 /* Do not remove the return statement, so that redirect_edge_and_branch
822 sees how the block ends. */
823 if (gimple_code (t) == GIMPLE_RETURN)
824 break;
825
826 gsi_remove (&gsi, true);
827 release_defs (t);
828 }
829
830 /* Number of executions of function has reduced by the tailcall. */
831 e = single_succ_edge (gsi_bb (t->call_gsi));
832 decrease_profile (EXIT_BLOCK_PTR, e->count, EDGE_FREQUENCY (e));
833 decrease_profile (ENTRY_BLOCK_PTR, e->count, EDGE_FREQUENCY (e));
834 if (e->dest != EXIT_BLOCK_PTR)
835 decrease_profile (e->dest, e->count, EDGE_FREQUENCY (e));
836
837 /* Replace the call by a jump to the start of function. */
838 e = redirect_edge_and_branch (single_succ_edge (gsi_bb (t->call_gsi)),
839 first);
840 gcc_assert (e);
841 PENDING_STMT (e) = NULL;
842
843 /* Add phi node entries for arguments. The ordering of the phi nodes should
844 be the same as the ordering of the arguments. */
845 for (param = DECL_ARGUMENTS (current_function_decl),
846 idx = 0, gsi = gsi_start_phis (first);
847 param;
848 param = DECL_CHAIN (param), idx++)
849 {
850 if (!arg_needs_copy_p (param))
851 continue;
852
853 arg = gimple_call_arg (stmt, idx);
854 phi = gsi_stmt (gsi);
855 gcc_assert (param == SSA_NAME_VAR (PHI_RESULT (phi)));
856
857 add_phi_arg (phi, arg, e, gimple_location (stmt));
858 gsi_next (&gsi);
859 }
860
861 /* Update the values of accumulators. */
862 adjust_accumulator_values (t->call_gsi, t->mult, t->add, e);
863
864 call = gsi_stmt (t->call_gsi);
865 rslt = gimple_call_lhs (call);
866 if (rslt != NULL_TREE)
867 {
868 /* Result of the call will no longer be defined. So adjust the
869 SSA_NAME_DEF_STMT accordingly. */
870 SSA_NAME_DEF_STMT (rslt) = gimple_build_nop ();
871 }
872
873 gsi_remove (&t->call_gsi, true);
874 release_defs (call);
875 }
876
877 /* Optimizes the tailcall described by T. If OPT_TAILCALLS is true, also
878 mark the tailcalls for the sibcall optimization. */
879
880 static bool
881 optimize_tail_call (struct tailcall *t, bool opt_tailcalls)
882 {
883 if (t->tail_recursion)
884 {
885 eliminate_tail_call (t);
886 return true;
887 }
888
889 if (opt_tailcalls)
890 {
891 gimple stmt = gsi_stmt (t->call_gsi);
892
893 gimple_call_set_tail (stmt, true);
894 if (dump_file && (dump_flags & TDF_DETAILS))
895 {
896 fprintf (dump_file, "Found tail call ");
897 print_gimple_stmt (dump_file, stmt, 0, dump_flags);
898 fprintf (dump_file, " in bb %i\n", (gsi_bb (t->call_gsi))->index);
899 }
900 }
901
902 return false;
903 }
904
905 /* Creates a tail-call accumulator of the same type as the return type of the
906 current function. LABEL is the name used to creating the temporary
907 variable for the accumulator. The accumulator will be inserted in the
908 phis of a basic block BB with single predecessor with an initial value
909 INIT converted to the current function return type. */
910
911 static tree
912 create_tailcall_accumulator (const char *label, basic_block bb, tree init)
913 {
914 tree ret_type = TREE_TYPE (DECL_RESULT (current_function_decl));
915 if (POINTER_TYPE_P (ret_type))
916 ret_type = sizetype;
917
918 tree tmp = make_temp_ssa_name (ret_type, NULL, label);
919 gimple phi;
920
921 phi = create_phi_node (tmp, bb);
922 /* RET_TYPE can be a float when -ffast-maths is enabled. */
923 add_phi_arg (phi, fold_convert (ret_type, init), single_pred_edge (bb),
924 UNKNOWN_LOCATION);
925 return PHI_RESULT (phi);
926 }
927
928 /* Optimizes tail calls in the function, turning the tail recursion
929 into iteration. */
930
931 static unsigned int
932 tree_optimize_tail_calls_1 (bool opt_tailcalls)
933 {
934 edge e;
935 bool phis_constructed = false;
936 struct tailcall *tailcalls = NULL, *act, *next;
937 bool changed = false;
938 basic_block first = single_succ (ENTRY_BLOCK_PTR);
939 tree param;
940 gimple stmt;
941 edge_iterator ei;
942
943 if (!suitable_for_tail_opt_p ())
944 return 0;
945 if (opt_tailcalls)
946 opt_tailcalls = suitable_for_tail_call_opt_p ();
947
948 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds)
949 {
950 /* Only traverse the normal exits, i.e. those that end with return
951 statement. */
952 stmt = last_stmt (e->src);
953
954 if (stmt
955 && gimple_code (stmt) == GIMPLE_RETURN)
956 find_tail_calls (e->src, &tailcalls);
957 }
958
959 /* Construct the phi nodes and accumulators if necessary. */
960 a_acc = m_acc = NULL_TREE;
961 for (act = tailcalls; act; act = act->next)
962 {
963 if (!act->tail_recursion)
964 continue;
965
966 if (!phis_constructed)
967 {
968 /* Ensure that there is only one predecessor of the block
969 or if there are existing degenerate PHI nodes. */
970 if (!single_pred_p (first)
971 || !gimple_seq_empty_p (phi_nodes (first)))
972 first = split_edge (single_succ_edge (ENTRY_BLOCK_PTR));
973
974 /* Copy the args if needed. */
975 for (param = DECL_ARGUMENTS (current_function_decl);
976 param;
977 param = DECL_CHAIN (param))
978 if (arg_needs_copy_p (param))
979 {
980 tree name = ssa_default_def (cfun, param);
981 tree new_name = make_ssa_name (param, SSA_NAME_DEF_STMT (name));
982 gimple phi;
983
984 set_ssa_default_def (cfun, param, new_name);
985 phi = create_phi_node (name, first);
986 add_phi_arg (phi, new_name, single_pred_edge (first),
987 EXPR_LOCATION (param));
988 }
989 phis_constructed = true;
990 }
991
992 if (act->add && !a_acc)
993 a_acc = create_tailcall_accumulator ("add_acc", first,
994 integer_zero_node);
995
996 if (act->mult && !m_acc)
997 m_acc = create_tailcall_accumulator ("mult_acc", first,
998 integer_one_node);
999 }
1000
1001 if (a_acc || m_acc)
1002 {
1003 /* When the tail call elimination using accumulators is performed,
1004 statements adding the accumulated value are inserted at all exits.
1005 This turns all other tail calls to non-tail ones. */
1006 opt_tailcalls = false;
1007 }
1008
1009 for (; tailcalls; tailcalls = next)
1010 {
1011 next = tailcalls->next;
1012 changed |= optimize_tail_call (tailcalls, opt_tailcalls);
1013 free (tailcalls);
1014 }
1015
1016 if (a_acc || m_acc)
1017 {
1018 /* Modify the remaining return statements. */
1019 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds)
1020 {
1021 stmt = last_stmt (e->src);
1022
1023 if (stmt
1024 && gimple_code (stmt) == GIMPLE_RETURN)
1025 adjust_return_value (e->src, m_acc, a_acc);
1026 }
1027 }
1028
1029 if (changed)
1030 {
1031 /* We may have created new loops. Make them magically appear. */
1032 if (current_loops)
1033 loops_state_set (LOOPS_NEED_FIXUP);
1034 free_dominance_info (CDI_DOMINATORS);
1035 }
1036
1037 /* Add phi nodes for the virtual operands defined in the function to the
1038 header of the loop created by tail recursion elimination. Do so
1039 by triggering the SSA renamer. */
1040 if (phis_constructed)
1041 mark_virtual_operands_for_renaming (cfun);
1042
1043 if (changed)
1044 return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals;
1045 return 0;
1046 }
1047
1048 static unsigned int
1049 execute_tail_recursion (void)
1050 {
1051 return tree_optimize_tail_calls_1 (false);
1052 }
1053
1054 static bool
1055 gate_tail_calls (void)
1056 {
1057 return flag_optimize_sibling_calls != 0 && dbg_cnt (tail_call);
1058 }
1059
1060 static unsigned int
1061 execute_tail_calls (void)
1062 {
1063 return tree_optimize_tail_calls_1 (true);
1064 }
1065
1066 namespace {
1067
1068 const pass_data pass_data_tail_recursion =
1069 {
1070 GIMPLE_PASS, /* type */
1071 "tailr", /* name */
1072 OPTGROUP_NONE, /* optinfo_flags */
1073 true, /* has_gate */
1074 true, /* has_execute */
1075 TV_NONE, /* tv_id */
1076 ( PROP_cfg | PROP_ssa ), /* properties_required */
1077 0, /* properties_provided */
1078 0, /* properties_destroyed */
1079 0, /* todo_flags_start */
1080 TODO_verify_ssa, /* todo_flags_finish */
1081 };
1082
1083 class pass_tail_recursion : public gimple_opt_pass
1084 {
1085 public:
1086 pass_tail_recursion (gcc::context *ctxt)
1087 : gimple_opt_pass (pass_data_tail_recursion, ctxt)
1088 {}
1089
1090 /* opt_pass methods: */
1091 opt_pass * clone () { return new pass_tail_recursion (m_ctxt); }
1092 bool gate () { return gate_tail_calls (); }
1093 unsigned int execute () { return execute_tail_recursion (); }
1094
1095 }; // class pass_tail_recursion
1096
1097 } // anon namespace
1098
1099 gimple_opt_pass *
1100 make_pass_tail_recursion (gcc::context *ctxt)
1101 {
1102 return new pass_tail_recursion (ctxt);
1103 }
1104
1105 namespace {
1106
1107 const pass_data pass_data_tail_calls =
1108 {
1109 GIMPLE_PASS, /* type */
1110 "tailc", /* name */
1111 OPTGROUP_NONE, /* optinfo_flags */
1112 true, /* has_gate */
1113 true, /* has_execute */
1114 TV_NONE, /* tv_id */
1115 ( PROP_cfg | PROP_ssa ), /* properties_required */
1116 0, /* properties_provided */
1117 0, /* properties_destroyed */
1118 0, /* todo_flags_start */
1119 TODO_verify_ssa, /* todo_flags_finish */
1120 };
1121
1122 class pass_tail_calls : public gimple_opt_pass
1123 {
1124 public:
1125 pass_tail_calls (gcc::context *ctxt)
1126 : gimple_opt_pass (pass_data_tail_calls, ctxt)
1127 {}
1128
1129 /* opt_pass methods: */
1130 bool gate () { return gate_tail_calls (); }
1131 unsigned int execute () { return execute_tail_calls (); }
1132
1133 }; // class pass_tail_calls
1134
1135 } // anon namespace
1136
1137 gimple_opt_pass *
1138 make_pass_tail_calls (gcc::context *ctxt)
1139 {
1140 return new pass_tail_calls (ctxt);
1141 }