Make-lang.in, [...]: Update copyright years.
[gcc.git] / gcc / tree-tailcall.c
1 /* Tail call optimization on trees.
2 Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
3 Free Software Foundation, Inc.
4
5 This file is part of GCC.
6
7 GCC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 3, or (at your option)
10 any later version.
11
12 GCC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
16
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
20
21 #include "config.h"
22 #include "system.h"
23 #include "coretypes.h"
24 #include "tm.h"
25 #include "tree.h"
26 #include "rtl.h"
27 #include "tm_p.h"
28 #include "hard-reg-set.h"
29 #include "basic-block.h"
30 #include "function.h"
31 #include "tree-flow.h"
32 #include "tree-dump.h"
33 #include "diagnostic.h"
34 #include "except.h"
35 #include "tree-pass.h"
36 #include "flags.h"
37 #include "langhooks.h"
38 #include "dbgcnt.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 referenced_var_iterator rvi;
134 tree var;
135
136 if (cfun->stdarg)
137 return false;
138
139 /* No local variable nor structure field should be call-used. */
140 FOR_EACH_REFERENCED_VAR (var, rvi)
141 {
142 if (!is_global_var (var)
143 && is_call_used (var))
144 return false;
145 }
146
147 return true;
148 }
149 /* Returns false when the function is not suitable for tail call optimization
150 from some reason (e.g. if it takes variable number of arguments).
151 This test must pass in addition to suitable_for_tail_opt_p in order to make
152 tail call discovery happen. */
153
154 static bool
155 suitable_for_tail_call_opt_p (void)
156 {
157 tree param;
158
159 /* alloca (until we have stack slot life analysis) inhibits
160 sibling call optimizations, but not tail recursion. */
161 if (cfun->calls_alloca)
162 return false;
163
164 /* If we are using sjlj exceptions, we may need to add a call to
165 _Unwind_SjLj_Unregister at exit of the function. Which means
166 that we cannot do any sibcall transformations. */
167 if (USING_SJLJ_EXCEPTIONS && current_function_has_exception_handlers ())
168 return false;
169
170 /* Any function that calls setjmp might have longjmp called from
171 any called function. ??? We really should represent this
172 properly in the CFG so that this needn't be special cased. */
173 if (cfun->calls_setjmp)
174 return false;
175
176 /* ??? It is OK if the argument of a function is taken in some cases,
177 but not in all cases. See PR15387 and PR19616. Revisit for 4.1. */
178 for (param = DECL_ARGUMENTS (current_function_decl);
179 param;
180 param = TREE_CHAIN (param))
181 if (TREE_ADDRESSABLE (param))
182 return false;
183
184 return true;
185 }
186
187 /* Checks whether the expression EXPR in stmt AT is independent of the
188 statement pointed to by GSI (in a sense that we already know EXPR's value
189 at GSI). We use the fact that we are only called from the chain of
190 basic blocks that have only single successor. Returns the expression
191 containing the value of EXPR at GSI. */
192
193 static tree
194 independent_of_stmt_p (tree expr, gimple at, gimple_stmt_iterator gsi)
195 {
196 basic_block bb, call_bb, at_bb;
197 edge e;
198 edge_iterator ei;
199
200 if (is_gimple_min_invariant (expr))
201 return expr;
202
203 if (TREE_CODE (expr) != SSA_NAME)
204 return NULL_TREE;
205
206 /* Mark the blocks in the chain leading to the end. */
207 at_bb = gimple_bb (at);
208 call_bb = gimple_bb (gsi_stmt (gsi));
209 for (bb = call_bb; bb != at_bb; bb = single_succ (bb))
210 bb->aux = &bb->aux;
211 bb->aux = &bb->aux;
212
213 while (1)
214 {
215 at = SSA_NAME_DEF_STMT (expr);
216 bb = gimple_bb (at);
217
218 /* The default definition or defined before the chain. */
219 if (!bb || !bb->aux)
220 break;
221
222 if (bb == call_bb)
223 {
224 for (; !gsi_end_p (gsi); gsi_next (&gsi))
225 if (gsi_stmt (gsi) == at)
226 break;
227
228 if (!gsi_end_p (gsi))
229 expr = NULL_TREE;
230 break;
231 }
232
233 if (gimple_code (at) != GIMPLE_PHI)
234 {
235 expr = NULL_TREE;
236 break;
237 }
238
239 FOR_EACH_EDGE (e, ei, bb->preds)
240 if (e->src->aux)
241 break;
242 gcc_assert (e);
243
244 expr = PHI_ARG_DEF_FROM_EDGE (at, e);
245 if (TREE_CODE (expr) != SSA_NAME)
246 {
247 /* The value is a constant. */
248 break;
249 }
250 }
251
252 /* Unmark the blocks. */
253 for (bb = call_bb; bb != at_bb; bb = single_succ (bb))
254 bb->aux = NULL;
255 bb->aux = NULL;
256
257 return expr;
258 }
259
260 /* Simulates the effect of an assignment STMT on the return value of the tail
261 recursive CALL passed in ASS_VAR. M and A are the multiplicative and the
262 additive factor for the real return value. */
263
264 static bool
265 process_assignment (gimple stmt, gimple_stmt_iterator call, tree *m,
266 tree *a, tree *ass_var)
267 {
268 tree op0, op1, non_ass_var;
269 tree dest = gimple_assign_lhs (stmt);
270 enum tree_code code = gimple_assign_rhs_code (stmt);
271 enum gimple_rhs_class rhs_class = get_gimple_rhs_class (code);
272 tree src_var = gimple_assign_rhs1 (stmt);
273
274 /* See if this is a simple copy operation of an SSA name to the function
275 result. In that case we may have a simple tail call. Ignore type
276 conversions that can never produce extra code between the function
277 call and the function return. */
278 if ((rhs_class == GIMPLE_SINGLE_RHS || gimple_assign_cast_p (stmt))
279 && (TREE_CODE (src_var) == SSA_NAME))
280 {
281 /* Reject a tailcall if the type conversion might need
282 additional code. */
283 if (gimple_assign_cast_p (stmt)
284 && TYPE_MODE (TREE_TYPE (dest)) != TYPE_MODE (TREE_TYPE (src_var)))
285 return false;
286
287 if (src_var != *ass_var)
288 return false;
289
290 *ass_var = dest;
291 return true;
292 }
293
294 if (rhs_class != GIMPLE_BINARY_RHS)
295 return false;
296
297 /* Accumulator optimizations will reverse the order of operations.
298 We can only do that for floating-point types if we're assuming
299 that addition and multiplication are associative. */
300 if (!flag_associative_math)
301 if (FLOAT_TYPE_P (TREE_TYPE (DECL_RESULT (current_function_decl))))
302 return false;
303
304 /* We only handle the code like
305
306 x = call ();
307 y = m * x;
308 z = y + a;
309 return z;
310
311 TODO -- Extend it for cases where the linear transformation of the output
312 is expressed in a more complicated way. */
313
314 op0 = gimple_assign_rhs1 (stmt);
315 op1 = gimple_assign_rhs2 (stmt);
316
317 if (op0 == *ass_var
318 && (non_ass_var = independent_of_stmt_p (op1, stmt, call)))
319 ;
320 else if (op1 == *ass_var
321 && (non_ass_var = independent_of_stmt_p (op0, stmt, call)))
322 ;
323 else
324 return false;
325
326 switch (code)
327 {
328 case PLUS_EXPR:
329 *a = non_ass_var;
330 *ass_var = dest;
331 return true;
332
333 case MULT_EXPR:
334 *m = non_ass_var;
335 *ass_var = dest;
336 return true;
337
338 /* TODO -- Handle other codes (NEGATE_EXPR, MINUS_EXPR,
339 POINTER_PLUS_EXPR). */
340
341 default:
342 return false;
343 }
344 }
345
346 /* Propagate VAR through phis on edge E. */
347
348 static tree
349 propagate_through_phis (tree var, edge e)
350 {
351 basic_block dest = e->dest;
352 gimple_stmt_iterator gsi;
353
354 for (gsi = gsi_start_phis (dest); !gsi_end_p (gsi); gsi_next (&gsi))
355 {
356 gimple phi = gsi_stmt (gsi);
357 if (PHI_ARG_DEF_FROM_EDGE (phi, e) == var)
358 return PHI_RESULT (phi);
359 }
360 return var;
361 }
362
363 /* Finds tailcalls falling into basic block BB. The list of found tailcalls is
364 added to the start of RET. */
365
366 static void
367 find_tail_calls (basic_block bb, struct tailcall **ret)
368 {
369 tree ass_var = NULL_TREE, ret_var, func, param;
370 gimple stmt, call = NULL;
371 gimple_stmt_iterator gsi, agsi;
372 bool tail_recursion;
373 struct tailcall *nw;
374 edge e;
375 tree m, a;
376 basic_block abb;
377 size_t idx;
378
379 if (!single_succ_p (bb))
380 return;
381
382 for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi); gsi_prev (&gsi))
383 {
384 stmt = gsi_stmt (gsi);
385
386 /* Ignore labels. */
387 if (gimple_code (stmt) == GIMPLE_LABEL || is_gimple_debug (stmt))
388 continue;
389
390 /* Check for a call. */
391 if (is_gimple_call (stmt))
392 {
393 call = stmt;
394 ass_var = gimple_call_lhs (stmt);
395 break;
396 }
397
398 /* If the statement references memory or volatile operands, fail. */
399 if (gimple_references_memory_p (stmt)
400 || gimple_has_volatile_ops (stmt))
401 return;
402 }
403
404 if (gsi_end_p (gsi))
405 {
406 edge_iterator ei;
407 /* Recurse to the predecessors. */
408 FOR_EACH_EDGE (e, ei, bb->preds)
409 find_tail_calls (e->src, ret);
410
411 return;
412 }
413
414 /* If the LHS of our call is not just a simple register, we can't
415 transform this into a tail or sibling call. This situation happens,
416 in (e.g.) "*p = foo()" where foo returns a struct. In this case
417 we won't have a temporary here, but we need to carry out the side
418 effect anyway, so tailcall is impossible.
419
420 ??? In some situations (when the struct is returned in memory via
421 invisible argument) we could deal with this, e.g. by passing 'p'
422 itself as that argument to foo, but it's too early to do this here,
423 and expand_call() will not handle it anyway. If it ever can, then
424 we need to revisit this here, to allow that situation. */
425 if (ass_var && !is_gimple_reg (ass_var))
426 return;
427
428 /* We found the call, check whether it is suitable. */
429 tail_recursion = false;
430 func = gimple_call_fndecl (call);
431 if (func == current_function_decl)
432 {
433 tree arg;
434 for (param = DECL_ARGUMENTS (func), idx = 0;
435 param && idx < gimple_call_num_args (call);
436 param = TREE_CHAIN (param), idx ++)
437 {
438 arg = gimple_call_arg (call, idx);
439 if (param != arg)
440 {
441 /* Make sure there are no problems with copying. The parameter
442 have a copyable type and the two arguments must have reasonably
443 equivalent types. The latter requirement could be relaxed if
444 we emitted a suitable type conversion statement. */
445 if (!is_gimple_reg_type (TREE_TYPE (param))
446 || !useless_type_conversion_p (TREE_TYPE (param),
447 TREE_TYPE (arg)))
448 break;
449
450 /* The parameter should be a real operand, so that phi node
451 created for it at the start of the function has the meaning
452 of copying the value. This test implies is_gimple_reg_type
453 from the previous condition, however this one could be
454 relaxed by being more careful with copying the new value
455 of the parameter (emitting appropriate GIMPLE_ASSIGN and
456 updating the virtual operands). */
457 if (!is_gimple_reg (param))
458 break;
459 }
460 }
461 if (idx == gimple_call_num_args (call) && !param)
462 tail_recursion = true;
463 }
464
465 /* Now check the statements after the call. None of them has virtual
466 operands, so they may only depend on the call through its return
467 value. The return value should also be dependent on each of them,
468 since we are running after dce. */
469 m = NULL_TREE;
470 a = NULL_TREE;
471
472 abb = bb;
473 agsi = gsi;
474 while (1)
475 {
476 tree tmp_a = NULL_TREE;
477 tree tmp_m = NULL_TREE;
478 gsi_next (&agsi);
479
480 while (gsi_end_p (agsi))
481 {
482 ass_var = propagate_through_phis (ass_var, single_succ_edge (abb));
483 abb = single_succ (abb);
484 agsi = gsi_start_bb (abb);
485 }
486
487 stmt = gsi_stmt (agsi);
488
489 if (gimple_code (stmt) == GIMPLE_LABEL)
490 continue;
491
492 if (gimple_code (stmt) == GIMPLE_RETURN)
493 break;
494
495 if (is_gimple_debug (stmt))
496 continue;
497
498 if (gimple_code (stmt) != GIMPLE_ASSIGN)
499 return;
500
501 /* This is a gimple assign. */
502 if (! process_assignment (stmt, gsi, &tmp_m, &tmp_a, &ass_var))
503 return;
504
505 if (tmp_a)
506 {
507 if (a)
508 a = fold_build2 (PLUS_EXPR, TREE_TYPE (tmp_a), a, tmp_a);
509 else
510 a = tmp_a;
511 }
512 if (tmp_m)
513 {
514 if (m)
515 m = fold_build2 (MULT_EXPR, TREE_TYPE (tmp_m), m, tmp_m);
516 else
517 m = tmp_m;
518
519 if (a)
520 a = fold_build2 (MULT_EXPR, TREE_TYPE (tmp_m), a, tmp_m);
521 }
522 }
523
524 /* See if this is a tail call we can handle. */
525 ret_var = gimple_return_retval (stmt);
526
527 /* We may proceed if there either is no return value, or the return value
528 is identical to the call's return. */
529 if (ret_var
530 && (ret_var != ass_var))
531 return;
532
533 /* If this is not a tail recursive call, we cannot handle addends or
534 multiplicands. */
535 if (!tail_recursion && (m || a))
536 return;
537
538 nw = XNEW (struct tailcall);
539
540 nw->call_gsi = gsi;
541
542 nw->tail_recursion = tail_recursion;
543
544 nw->mult = m;
545 nw->add = a;
546
547 nw->next = *ret;
548 *ret = nw;
549 }
550
551 /* Helper to insert PHI_ARGH to the phi of VAR in the destination of edge E. */
552
553 static void
554 add_successor_phi_arg (edge e, tree var, tree phi_arg)
555 {
556 gimple_stmt_iterator gsi;
557
558 for (gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi); gsi_next (&gsi))
559 if (PHI_RESULT (gsi_stmt (gsi)) == var)
560 break;
561
562 gcc_assert (!gsi_end_p (gsi));
563 add_phi_arg (gsi_stmt (gsi), phi_arg, e, UNKNOWN_LOCATION);
564 }
565
566 /* Creates a GIMPLE statement which computes the operation specified by
567 CODE, OP0 and OP1 to a new variable with name LABEL and inserts the
568 statement in the position specified by GSI and UPDATE. Returns the
569 tree node of the statement's result. */
570
571 static tree
572 adjust_return_value_with_ops (enum tree_code code, const char *label,
573 tree acc, tree op1, gimple_stmt_iterator gsi)
574 {
575
576 tree ret_type = TREE_TYPE (DECL_RESULT (current_function_decl));
577 tree tmp = create_tmp_var (ret_type, label);
578 gimple stmt;
579 tree result;
580
581 if (TREE_CODE (ret_type) == COMPLEX_TYPE
582 || TREE_CODE (ret_type) == VECTOR_TYPE)
583 DECL_GIMPLE_REG_P (tmp) = 1;
584 add_referenced_var (tmp);
585
586 if (types_compatible_p (TREE_TYPE (acc), TREE_TYPE (op1)))
587 stmt = gimple_build_assign_with_ops (code, tmp, acc, op1);
588 else
589 {
590 tree rhs = fold_convert (TREE_TYPE (acc),
591 fold_build2 (code,
592 TREE_TYPE (op1),
593 fold_convert (TREE_TYPE (op1), acc),
594 op1));
595 rhs = force_gimple_operand_gsi (&gsi, rhs,
596 false, NULL, true, GSI_CONTINUE_LINKING);
597 stmt = gimple_build_assign (NULL_TREE, rhs);
598 }
599
600 result = make_ssa_name (tmp, stmt);
601 gimple_assign_set_lhs (stmt, result);
602 update_stmt (stmt);
603 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
604 return result;
605 }
606
607 /* Creates a new GIMPLE statement that adjusts the value of accumulator ACC by
608 the computation specified by CODE and OP1 and insert the statement
609 at the position specified by GSI as a new statement. Returns new SSA name
610 of updated accumulator. */
611
612 static tree
613 update_accumulator_with_ops (enum tree_code code, tree acc, tree op1,
614 gimple_stmt_iterator gsi)
615 {
616 gimple stmt;
617 tree var;
618 if (types_compatible_p (TREE_TYPE (acc), TREE_TYPE (op1)))
619 stmt = gimple_build_assign_with_ops (code, SSA_NAME_VAR (acc), acc, op1);
620 else
621 {
622 tree rhs = fold_convert (TREE_TYPE (acc),
623 fold_build2 (code,
624 TREE_TYPE (op1),
625 fold_convert (TREE_TYPE (op1), acc),
626 op1));
627 rhs = force_gimple_operand_gsi (&gsi, rhs,
628 false, NULL, false, GSI_CONTINUE_LINKING);
629 stmt = gimple_build_assign (NULL_TREE, rhs);
630 }
631 var = make_ssa_name (SSA_NAME_VAR (acc), stmt);
632 gimple_assign_set_lhs (stmt, var);
633 update_stmt (stmt);
634 gsi_insert_after (&gsi, stmt, GSI_NEW_STMT);
635 return var;
636 }
637
638 /* Adjust the accumulator values according to A and M after GSI, and update
639 the phi nodes on edge BACK. */
640
641 static void
642 adjust_accumulator_values (gimple_stmt_iterator gsi, tree m, tree a, edge back)
643 {
644 tree var, a_acc_arg, m_acc_arg;
645
646 if (m)
647 m = force_gimple_operand_gsi (&gsi, m, true, NULL, true, GSI_SAME_STMT);
648 if (a)
649 a = force_gimple_operand_gsi (&gsi, a, true, NULL, true, GSI_SAME_STMT);
650
651 a_acc_arg = a_acc;
652 m_acc_arg = m_acc;
653 if (a)
654 {
655 if (m_acc)
656 {
657 if (integer_onep (a))
658 var = m_acc;
659 else
660 var = adjust_return_value_with_ops (MULT_EXPR, "acc_tmp", m_acc,
661 a, gsi);
662 }
663 else
664 var = a;
665
666 a_acc_arg = update_accumulator_with_ops (PLUS_EXPR, a_acc, var, gsi);
667 }
668
669 if (m)
670 m_acc_arg = update_accumulator_with_ops (MULT_EXPR, m_acc, m, gsi);
671
672 if (a_acc)
673 add_successor_phi_arg (back, a_acc, a_acc_arg);
674
675 if (m_acc)
676 add_successor_phi_arg (back, m_acc, m_acc_arg);
677 }
678
679 /* Adjust value of the return at the end of BB according to M and A
680 accumulators. */
681
682 static void
683 adjust_return_value (basic_block bb, tree m, tree a)
684 {
685 tree retval;
686 gimple ret_stmt = gimple_seq_last_stmt (bb_seq (bb));
687 gimple_stmt_iterator gsi = gsi_last_bb (bb);
688
689 gcc_assert (gimple_code (ret_stmt) == GIMPLE_RETURN);
690
691 retval = gimple_return_retval (ret_stmt);
692 if (!retval || retval == error_mark_node)
693 return;
694
695 if (m)
696 retval = adjust_return_value_with_ops (MULT_EXPR, "mul_tmp", m_acc, retval,
697 gsi);
698 if (a)
699 retval = adjust_return_value_with_ops (PLUS_EXPR, "acc_tmp", a_acc, retval,
700 gsi);
701 gimple_return_set_retval (ret_stmt, retval);
702 update_stmt (ret_stmt);
703 }
704
705 /* Subtract COUNT and FREQUENCY from the basic block and it's
706 outgoing edge. */
707 static void
708 decrease_profile (basic_block bb, gcov_type count, int frequency)
709 {
710 edge e;
711 bb->count -= count;
712 if (bb->count < 0)
713 bb->count = 0;
714 bb->frequency -= frequency;
715 if (bb->frequency < 0)
716 bb->frequency = 0;
717 if (!single_succ_p (bb))
718 {
719 gcc_assert (!EDGE_COUNT (bb->succs));
720 return;
721 }
722 e = single_succ_edge (bb);
723 e->count -= count;
724 if (e->count < 0)
725 e->count = 0;
726 }
727
728 /* Returns true if argument PARAM of the tail recursive call needs to be copied
729 when the call is eliminated. */
730
731 static bool
732 arg_needs_copy_p (tree param)
733 {
734 tree def;
735
736 if (!is_gimple_reg (param) || !var_ann (param))
737 return false;
738
739 /* Parameters that are only defined but never used need not be copied. */
740 def = gimple_default_def (cfun, param);
741 if (!def)
742 return false;
743
744 return true;
745 }
746
747 /* Eliminates tail call described by T. TMP_VARS is a list of
748 temporary variables used to copy the function arguments. */
749
750 static void
751 eliminate_tail_call (struct tailcall *t)
752 {
753 tree param, rslt;
754 gimple stmt, call;
755 tree arg;
756 size_t idx;
757 basic_block bb, first;
758 edge e;
759 gimple phi;
760 gimple_stmt_iterator gsi;
761 gimple orig_stmt;
762
763 stmt = orig_stmt = gsi_stmt (t->call_gsi);
764 bb = gsi_bb (t->call_gsi);
765
766 if (dump_file && (dump_flags & TDF_DETAILS))
767 {
768 fprintf (dump_file, "Eliminated tail recursion in bb %d : ",
769 bb->index);
770 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
771 fprintf (dump_file, "\n");
772 }
773
774 gcc_assert (is_gimple_call (stmt));
775
776 first = single_succ (ENTRY_BLOCK_PTR);
777
778 /* Remove the code after call_gsi that will become unreachable. The
779 possibly unreachable code in other blocks is removed later in
780 cfg cleanup. */
781 gsi = t->call_gsi;
782 gsi_next (&gsi);
783 while (!gsi_end_p (gsi))
784 {
785 gimple t = gsi_stmt (gsi);
786 /* Do not remove the return statement, so that redirect_edge_and_branch
787 sees how the block ends. */
788 if (gimple_code (t) == GIMPLE_RETURN)
789 break;
790
791 gsi_remove (&gsi, true);
792 release_defs (t);
793 }
794
795 /* Number of executions of function has reduced by the tailcall. */
796 e = single_succ_edge (gsi_bb (t->call_gsi));
797 decrease_profile (EXIT_BLOCK_PTR, e->count, EDGE_FREQUENCY (e));
798 decrease_profile (ENTRY_BLOCK_PTR, e->count, EDGE_FREQUENCY (e));
799 if (e->dest != EXIT_BLOCK_PTR)
800 decrease_profile (e->dest, e->count, EDGE_FREQUENCY (e));
801
802 /* Replace the call by a jump to the start of function. */
803 e = redirect_edge_and_branch (single_succ_edge (gsi_bb (t->call_gsi)),
804 first);
805 gcc_assert (e);
806 PENDING_STMT (e) = NULL;
807
808 /* Add phi node entries for arguments. The ordering of the phi nodes should
809 be the same as the ordering of the arguments. */
810 for (param = DECL_ARGUMENTS (current_function_decl),
811 idx = 0, gsi = gsi_start_phis (first);
812 param;
813 param = TREE_CHAIN (param), idx++)
814 {
815 if (!arg_needs_copy_p (param))
816 continue;
817
818 arg = gimple_call_arg (stmt, idx);
819 phi = gsi_stmt (gsi);
820 gcc_assert (param == SSA_NAME_VAR (PHI_RESULT (phi)));
821
822 add_phi_arg (phi, arg, e, gimple_location (stmt));
823 gsi_next (&gsi);
824 }
825
826 /* Update the values of accumulators. */
827 adjust_accumulator_values (t->call_gsi, t->mult, t->add, e);
828
829 call = gsi_stmt (t->call_gsi);
830 rslt = gimple_call_lhs (call);
831 if (rslt != NULL_TREE)
832 {
833 /* Result of the call will no longer be defined. So adjust the
834 SSA_NAME_DEF_STMT accordingly. */
835 SSA_NAME_DEF_STMT (rslt) = gimple_build_nop ();
836 }
837
838 gsi_remove (&t->call_gsi, true);
839 release_defs (call);
840 }
841
842 /* Add phi nodes for the virtual operands defined in the function to the
843 header of the loop created by tail recursion elimination.
844
845 Originally, we used to add phi nodes only for call clobbered variables,
846 as the value of the non-call clobbered ones obviously cannot be used
847 or changed within the recursive call. However, the local variables
848 from multiple calls now share the same location, so the virtual ssa form
849 requires us to say that the location dies on further iterations of the loop,
850 which requires adding phi nodes.
851 */
852 static void
853 add_virtual_phis (void)
854 {
855 referenced_var_iterator rvi;
856 tree var;
857
858 /* The problematic part is that there is no way how to know what
859 to put into phi nodes (there in fact does not have to be such
860 ssa name available). A solution would be to have an artificial
861 use/kill for all virtual operands in EXIT node. Unless we have
862 this, we cannot do much better than to rebuild the ssa form for
863 possibly affected virtual ssa names from scratch. */
864
865 FOR_EACH_REFERENCED_VAR (var, rvi)
866 {
867 if (!is_gimple_reg (var) && gimple_default_def (cfun, var) != NULL_TREE)
868 mark_sym_for_renaming (var);
869 }
870 }
871
872 /* Optimizes the tailcall described by T. If OPT_TAILCALLS is true, also
873 mark the tailcalls for the sibcall optimization. */
874
875 static bool
876 optimize_tail_call (struct tailcall *t, bool opt_tailcalls)
877 {
878 if (t->tail_recursion)
879 {
880 eliminate_tail_call (t);
881 return true;
882 }
883
884 if (opt_tailcalls)
885 {
886 gimple stmt = gsi_stmt (t->call_gsi);
887
888 gimple_call_set_tail (stmt, true);
889 if (dump_file && (dump_flags & TDF_DETAILS))
890 {
891 fprintf (dump_file, "Found tail call ");
892 print_gimple_stmt (dump_file, stmt, 0, dump_flags);
893 fprintf (dump_file, " in bb %i\n", (gsi_bb (t->call_gsi))->index);
894 }
895 }
896
897 return false;
898 }
899
900 /* Creates a tail-call accumulator of the same type as the return type of the
901 current function. LABEL is the name used to creating the temporary
902 variable for the accumulator. The accumulator will be inserted in the
903 phis of a basic block BB with single predecessor with an initial value
904 INIT converted to the current function return type. */
905
906 static tree
907 create_tailcall_accumulator (const char *label, basic_block bb, tree init)
908 {
909 tree ret_type = TREE_TYPE (DECL_RESULT (current_function_decl));
910 tree tmp = create_tmp_var (ret_type, label);
911 gimple phi;
912
913 if (TREE_CODE (ret_type) == COMPLEX_TYPE
914 || TREE_CODE (ret_type) == VECTOR_TYPE)
915 DECL_GIMPLE_REG_P (tmp) = 1;
916 add_referenced_var (tmp);
917 phi = create_phi_node (tmp, bb);
918 /* RET_TYPE can be a float when -ffast-maths is enabled. */
919 add_phi_arg (phi, fold_convert (ret_type, init), single_pred_edge (bb),
920 UNKNOWN_LOCATION);
921 return PHI_RESULT (phi);
922 }
923
924 /* Optimizes tail calls in the function, turning the tail recursion
925 into iteration. */
926
927 static unsigned int
928 tree_optimize_tail_calls_1 (bool opt_tailcalls)
929 {
930 edge e;
931 bool phis_constructed = false;
932 struct tailcall *tailcalls = NULL, *act, *next;
933 bool changed = false;
934 basic_block first = single_succ (ENTRY_BLOCK_PTR);
935 tree param;
936 gimple stmt;
937 edge_iterator ei;
938
939 if (!suitable_for_tail_opt_p ())
940 return 0;
941 if (opt_tailcalls)
942 opt_tailcalls = suitable_for_tail_call_opt_p ();
943
944 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds)
945 {
946 /* Only traverse the normal exits, i.e. those that end with return
947 statement. */
948 stmt = last_stmt (e->src);
949
950 if (stmt
951 && gimple_code (stmt) == GIMPLE_RETURN)
952 find_tail_calls (e->src, &tailcalls);
953 }
954
955 /* Construct the phi nodes and accumulators if necessary. */
956 a_acc = m_acc = NULL_TREE;
957 for (act = tailcalls; act; act = act->next)
958 {
959 if (!act->tail_recursion)
960 continue;
961
962 if (!phis_constructed)
963 {
964 /* Ensure that there is only one predecessor of the block
965 or if there are existing degenerate PHI nodes. */
966 if (!single_pred_p (first)
967 || !gimple_seq_empty_p (phi_nodes (first)))
968 first = split_edge (single_succ_edge (ENTRY_BLOCK_PTR));
969
970 /* Copy the args if needed. */
971 for (param = DECL_ARGUMENTS (current_function_decl);
972 param;
973 param = TREE_CHAIN (param))
974 if (arg_needs_copy_p (param))
975 {
976 tree name = gimple_default_def (cfun, param);
977 tree new_name = make_ssa_name (param, SSA_NAME_DEF_STMT (name));
978 gimple phi;
979
980 set_default_def (param, new_name);
981 phi = create_phi_node (name, first);
982 SSA_NAME_DEF_STMT (name) = phi;
983 add_phi_arg (phi, new_name, single_pred_edge (first),
984 EXPR_LOCATION (param));
985 }
986 phis_constructed = true;
987 }
988
989 if (act->add && !a_acc)
990 a_acc = create_tailcall_accumulator ("add_acc", first,
991 integer_zero_node);
992
993 if (act->mult && !m_acc)
994 m_acc = create_tailcall_accumulator ("mult_acc", first,
995 integer_one_node);
996 }
997
998 for (; tailcalls; tailcalls = next)
999 {
1000 next = tailcalls->next;
1001 changed |= optimize_tail_call (tailcalls, opt_tailcalls);
1002 free (tailcalls);
1003 }
1004
1005 if (a_acc || m_acc)
1006 {
1007 /* Modify the remaining return statements. */
1008 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds)
1009 {
1010 stmt = last_stmt (e->src);
1011
1012 if (stmt
1013 && gimple_code (stmt) == GIMPLE_RETURN)
1014 adjust_return_value (e->src, m_acc, a_acc);
1015 }
1016 }
1017
1018 if (changed)
1019 free_dominance_info (CDI_DOMINATORS);
1020
1021 if (phis_constructed)
1022 add_virtual_phis ();
1023 if (changed)
1024 return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals;
1025 return 0;
1026 }
1027
1028 static unsigned int
1029 execute_tail_recursion (void)
1030 {
1031 return tree_optimize_tail_calls_1 (false);
1032 }
1033
1034 static bool
1035 gate_tail_calls (void)
1036 {
1037 return flag_optimize_sibling_calls != 0 && dbg_cnt (tail_call);
1038 }
1039
1040 static unsigned int
1041 execute_tail_calls (void)
1042 {
1043 return tree_optimize_tail_calls_1 (true);
1044 }
1045
1046 struct gimple_opt_pass pass_tail_recursion =
1047 {
1048 {
1049 GIMPLE_PASS,
1050 "tailr", /* name */
1051 gate_tail_calls, /* gate */
1052 execute_tail_recursion, /* execute */
1053 NULL, /* sub */
1054 NULL, /* next */
1055 0, /* static_pass_number */
1056 TV_NONE, /* tv_id */
1057 PROP_cfg | PROP_ssa, /* properties_required */
1058 0, /* properties_provided */
1059 0, /* properties_destroyed */
1060 0, /* todo_flags_start */
1061 TODO_dump_func | TODO_verify_ssa /* todo_flags_finish */
1062 }
1063 };
1064
1065 struct gimple_opt_pass pass_tail_calls =
1066 {
1067 {
1068 GIMPLE_PASS,
1069 "tailc", /* name */
1070 gate_tail_calls, /* gate */
1071 execute_tail_calls, /* execute */
1072 NULL, /* sub */
1073 NULL, /* next */
1074 0, /* static_pass_number */
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_dump_func | TODO_verify_ssa /* todo_flags_finish */
1081 }
1082 };