IA MCU psABI support: changes to libraries
[gcc.git] / gcc / tree-tailcall.c
1 /* Tail call optimization on trees.
2 Copyright (C) 2003-2015 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 "alias.h"
25 #include "symtab.h"
26 #include "tree.h"
27 #include "fold-const.h"
28 #include "stor-layout.h"
29 #include "tm_p.h"
30 #include "predict.h"
31 #include "hard-reg-set.h"
32 #include "function.h"
33 #include "dominance.h"
34 #include "cfg.h"
35 #include "basic-block.h"
36 #include "tree-ssa-alias.h"
37 #include "internal-fn.h"
38 #include "gimple-expr.h"
39 #include "gimple.h"
40 #include "gimple-iterator.h"
41 #include "gimplify-me.h"
42 #include "gimple-ssa.h"
43 #include "tree-cfg.h"
44 #include "tree-phinodes.h"
45 #include "stringpool.h"
46 #include "tree-ssanames.h"
47 #include "tree-into-ssa.h"
48 #include "rtl.h"
49 #include "flags.h"
50 #include "insn-config.h"
51 #include "expmed.h"
52 #include "dojump.h"
53 #include "explow.h"
54 #include "calls.h"
55 #include "emit-rtl.h"
56 #include "varasm.h"
57 #include "stmt.h"
58 #include "expr.h"
59 #include "tree-dfa.h"
60 #include "gimple-pretty-print.h"
61 #include "except.h"
62 #include "tree-pass.h"
63 #include "langhooks.h"
64 #include "dbgcnt.h"
65 #include "target.h"
66 #include "cfgloop.h"
67 #include "common/common-target.h"
68 #include "cgraph.h"
69 #include "ipa-utils.h"
70
71 /* The file implements the tail recursion elimination. It is also used to
72 analyze the tail calls in general, passing the results to the rtl level
73 where they are used for sibcall optimization.
74
75 In addition to the standard tail recursion elimination, we handle the most
76 trivial cases of making the call tail recursive by creating accumulators.
77 For example the following function
78
79 int sum (int n)
80 {
81 if (n > 0)
82 return n + sum (n - 1);
83 else
84 return 0;
85 }
86
87 is transformed into
88
89 int sum (int n)
90 {
91 int acc = 0;
92
93 while (n > 0)
94 acc += n--;
95
96 return acc;
97 }
98
99 To do this, we maintain two accumulators (a_acc and m_acc) that indicate
100 when we reach the return x statement, we should return a_acc + x * m_acc
101 instead. They are initially initialized to 0 and 1, respectively,
102 so the semantics of the function is obviously preserved. If we are
103 guaranteed that the value of the accumulator never change, we
104 omit the accumulator.
105
106 There are three cases how the function may exit. The first one is
107 handled in adjust_return_value, the other two in adjust_accumulator_values
108 (the second case is actually a special case of the third one and we
109 present it separately just for clarity):
110
111 1) Just return x, where x is not in any of the remaining special shapes.
112 We rewrite this to a gimple equivalent of return m_acc * x + a_acc.
113
114 2) return f (...), where f is the current function, is rewritten in a
115 classical tail-recursion elimination way, into assignment of arguments
116 and jump to the start of the function. Values of the accumulators
117 are unchanged.
118
119 3) return a + m * f(...), where a and m do not depend on call to f.
120 To preserve the semantics described before we want this to be rewritten
121 in such a way that we finally return
122
123 a_acc + (a + m * f(...)) * m_acc = (a_acc + a * m_acc) + (m * m_acc) * f(...).
124
125 I.e. we increase a_acc by a * m_acc, multiply m_acc by m and
126 eliminate the tail call to f. Special cases when the value is just
127 added or just multiplied are obtained by setting a = 0 or m = 1.
128
129 TODO -- it is possible to do similar tricks for other operations. */
130
131 /* A structure that describes the tailcall. */
132
133 struct tailcall
134 {
135 /* The iterator pointing to the call statement. */
136 gimple_stmt_iterator call_gsi;
137
138 /* True if it is a call to the current function. */
139 bool tail_recursion;
140
141 /* The return value of the caller is mult * f + add, where f is the return
142 value of the call. */
143 tree mult, add;
144
145 /* Next tailcall in the chain. */
146 struct tailcall *next;
147 };
148
149 /* The variables holding the value of multiplicative and additive
150 accumulator. */
151 static tree m_acc, a_acc;
152
153 static bool optimize_tail_call (struct tailcall *, bool);
154 static void eliminate_tail_call (struct tailcall *);
155
156 /* Returns false when the function is not suitable for tail call optimization
157 from some reason (e.g. if it takes variable number of arguments). */
158
159 static bool
160 suitable_for_tail_opt_p (void)
161 {
162 if (cfun->stdarg)
163 return false;
164
165 return true;
166 }
167 /* Returns false when the function is not suitable for tail call optimization
168 for some reason (e.g. if it takes variable number of arguments).
169 This test must pass in addition to suitable_for_tail_opt_p in order to make
170 tail call discovery happen. */
171
172 static bool
173 suitable_for_tail_call_opt_p (void)
174 {
175 tree param;
176
177 /* alloca (until we have stack slot life analysis) inhibits
178 sibling call optimizations, but not tail recursion. */
179 if (cfun->calls_alloca)
180 return false;
181
182 /* If we are using sjlj exceptions, we may need to add a call to
183 _Unwind_SjLj_Unregister at exit of the function. Which means
184 that we cannot do any sibcall transformations. */
185 if (targetm_common.except_unwind_info (&global_options) == UI_SJLJ
186 && current_function_has_exception_handlers ())
187 return false;
188
189 /* Any function that calls setjmp might have longjmp called from
190 any called function. ??? We really should represent this
191 properly in the CFG so that this needn't be special cased. */
192 if (cfun->calls_setjmp)
193 return false;
194
195 /* ??? It is OK if the argument of a function is taken in some cases,
196 but not in all cases. See PR15387 and PR19616. Revisit for 4.1. */
197 for (param = DECL_ARGUMENTS (current_function_decl);
198 param;
199 param = DECL_CHAIN (param))
200 if (TREE_ADDRESSABLE (param))
201 return false;
202
203 return true;
204 }
205
206 /* Checks whether the expression EXPR in stmt AT is independent of the
207 statement pointed to by GSI (in a sense that we already know EXPR's value
208 at GSI). We use the fact that we are only called from the chain of
209 basic blocks that have only single successor. Returns the expression
210 containing the value of EXPR at GSI. */
211
212 static tree
213 independent_of_stmt_p (tree expr, gimple at, gimple_stmt_iterator gsi)
214 {
215 basic_block bb, call_bb, at_bb;
216 edge e;
217 edge_iterator ei;
218
219 if (is_gimple_min_invariant (expr))
220 return expr;
221
222 if (TREE_CODE (expr) != SSA_NAME)
223 return NULL_TREE;
224
225 /* Mark the blocks in the chain leading to the end. */
226 at_bb = gimple_bb (at);
227 call_bb = gimple_bb (gsi_stmt (gsi));
228 for (bb = call_bb; bb != at_bb; bb = single_succ (bb))
229 bb->aux = &bb->aux;
230 bb->aux = &bb->aux;
231
232 while (1)
233 {
234 at = SSA_NAME_DEF_STMT (expr);
235 bb = gimple_bb (at);
236
237 /* The default definition or defined before the chain. */
238 if (!bb || !bb->aux)
239 break;
240
241 if (bb == call_bb)
242 {
243 for (; !gsi_end_p (gsi); gsi_next (&gsi))
244 if (gsi_stmt (gsi) == at)
245 break;
246
247 if (!gsi_end_p (gsi))
248 expr = NULL_TREE;
249 break;
250 }
251
252 if (gimple_code (at) != GIMPLE_PHI)
253 {
254 expr = NULL_TREE;
255 break;
256 }
257
258 FOR_EACH_EDGE (e, ei, bb->preds)
259 if (e->src->aux)
260 break;
261 gcc_assert (e);
262
263 expr = PHI_ARG_DEF_FROM_EDGE (at, e);
264 if (TREE_CODE (expr) != SSA_NAME)
265 {
266 /* The value is a constant. */
267 break;
268 }
269 }
270
271 /* Unmark the blocks. */
272 for (bb = call_bb; bb != at_bb; bb = single_succ (bb))
273 bb->aux = NULL;
274 bb->aux = NULL;
275
276 return expr;
277 }
278
279 /* Simulates the effect of an assignment STMT on the return value of the tail
280 recursive CALL passed in ASS_VAR. M and A are the multiplicative and the
281 additive factor for the real return value. */
282
283 static bool
284 process_assignment (gassign *stmt, gimple_stmt_iterator call, tree *m,
285 tree *a, tree *ass_var)
286 {
287 tree op0, op1 = NULL_TREE, non_ass_var = NULL_TREE;
288 tree dest = gimple_assign_lhs (stmt);
289 enum tree_code code = gimple_assign_rhs_code (stmt);
290 enum gimple_rhs_class rhs_class = get_gimple_rhs_class (code);
291 tree src_var = gimple_assign_rhs1 (stmt);
292
293 /* See if this is a simple copy operation of an SSA name to the function
294 result. In that case we may have a simple tail call. Ignore type
295 conversions that can never produce extra code between the function
296 call and the function return. */
297 if ((rhs_class == GIMPLE_SINGLE_RHS || gimple_assign_cast_p (stmt))
298 && (TREE_CODE (src_var) == SSA_NAME))
299 {
300 /* Reject a tailcall if the type conversion might need
301 additional code. */
302 if (gimple_assign_cast_p (stmt))
303 {
304 if (TYPE_MODE (TREE_TYPE (dest)) != TYPE_MODE (TREE_TYPE (src_var)))
305 return false;
306
307 /* Even if the type modes are the same, if the precision of the
308 type is smaller than mode's precision,
309 reduce_to_bit_field_precision would generate additional code. */
310 if (INTEGRAL_TYPE_P (TREE_TYPE (dest))
311 && (GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (dest)))
312 > TYPE_PRECISION (TREE_TYPE (dest))))
313 return false;
314 }
315
316 if (src_var != *ass_var)
317 return false;
318
319 *ass_var = dest;
320 return true;
321 }
322
323 switch (rhs_class)
324 {
325 case GIMPLE_BINARY_RHS:
326 op1 = gimple_assign_rhs2 (stmt);
327
328 /* Fall through. */
329
330 case GIMPLE_UNARY_RHS:
331 op0 = gimple_assign_rhs1 (stmt);
332 break;
333
334 default:
335 return false;
336 }
337
338 /* Accumulator optimizations will reverse the order of operations.
339 We can only do that for floating-point types if we're assuming
340 that addition and multiplication are associative. */
341 if (!flag_associative_math)
342 if (FLOAT_TYPE_P (TREE_TYPE (DECL_RESULT (current_function_decl))))
343 return false;
344
345 if (rhs_class == GIMPLE_UNARY_RHS)
346 ;
347 else if (op0 == *ass_var
348 && (non_ass_var = independent_of_stmt_p (op1, stmt, call)))
349 ;
350 else if (op1 == *ass_var
351 && (non_ass_var = independent_of_stmt_p (op0, stmt, call)))
352 ;
353 else
354 return false;
355
356 switch (code)
357 {
358 case PLUS_EXPR:
359 *a = non_ass_var;
360 *ass_var = dest;
361 return true;
362
363 case POINTER_PLUS_EXPR:
364 if (op0 != *ass_var)
365 return false;
366 *a = non_ass_var;
367 *ass_var = dest;
368 return true;
369
370 case MULT_EXPR:
371 *m = non_ass_var;
372 *ass_var = dest;
373 return true;
374
375 case NEGATE_EXPR:
376 *m = build_minus_one_cst (TREE_TYPE (op0));
377 *ass_var = dest;
378 return true;
379
380 case MINUS_EXPR:
381 if (*ass_var == op0)
382 *a = fold_build1 (NEGATE_EXPR, TREE_TYPE (non_ass_var), non_ass_var);
383 else
384 {
385 *m = build_minus_one_cst (TREE_TYPE (non_ass_var));
386 *a = fold_build1 (NEGATE_EXPR, TREE_TYPE (non_ass_var), non_ass_var);
387 }
388
389 *ass_var = dest;
390 return true;
391
392 /* TODO -- Handle POINTER_PLUS_EXPR. */
393
394 default:
395 return false;
396 }
397 }
398
399 /* Propagate VAR through phis on edge E. */
400
401 static tree
402 propagate_through_phis (tree var, edge e)
403 {
404 basic_block dest = e->dest;
405 gphi_iterator gsi;
406
407 for (gsi = gsi_start_phis (dest); !gsi_end_p (gsi); gsi_next (&gsi))
408 {
409 gphi *phi = gsi.phi ();
410 if (PHI_ARG_DEF_FROM_EDGE (phi, e) == var)
411 return PHI_RESULT (phi);
412 }
413 return var;
414 }
415
416 /* Finds tailcalls falling into basic block BB. The list of found tailcalls is
417 added to the start of RET. */
418
419 static void
420 find_tail_calls (basic_block bb, struct tailcall **ret)
421 {
422 tree ass_var = NULL_TREE, ret_var, func, param;
423 gimple stmt;
424 gcall *call = NULL;
425 gimple_stmt_iterator gsi, agsi;
426 bool tail_recursion;
427 struct tailcall *nw;
428 edge e;
429 tree m, a;
430 basic_block abb;
431 size_t idx;
432 tree var;
433
434 if (!single_succ_p (bb))
435 return;
436
437 for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi); gsi_prev (&gsi))
438 {
439 stmt = gsi_stmt (gsi);
440
441 /* Ignore labels, returns, clobbers and debug stmts. */
442 if (gimple_code (stmt) == GIMPLE_LABEL
443 || gimple_code (stmt) == GIMPLE_RETURN
444 || gimple_clobber_p (stmt)
445 || is_gimple_debug (stmt))
446 continue;
447
448 /* Check for a call. */
449 if (is_gimple_call (stmt))
450 {
451 call = as_a <gcall *> (stmt);
452 ass_var = gimple_call_lhs (call);
453 break;
454 }
455
456 /* If the statement references memory or volatile operands, fail. */
457 if (gimple_references_memory_p (stmt)
458 || gimple_has_volatile_ops (stmt))
459 return;
460 }
461
462 if (gsi_end_p (gsi))
463 {
464 edge_iterator ei;
465 /* Recurse to the predecessors. */
466 FOR_EACH_EDGE (e, ei, bb->preds)
467 find_tail_calls (e->src, ret);
468
469 return;
470 }
471
472 /* If the LHS of our call is not just a simple register, we can't
473 transform this into a tail or sibling call. This situation happens,
474 in (e.g.) "*p = foo()" where foo returns a struct. In this case
475 we won't have a temporary here, but we need to carry out the side
476 effect anyway, so tailcall is impossible.
477
478 ??? In some situations (when the struct is returned in memory via
479 invisible argument) we could deal with this, e.g. by passing 'p'
480 itself as that argument to foo, but it's too early to do this here,
481 and expand_call() will not handle it anyway. If it ever can, then
482 we need to revisit this here, to allow that situation. */
483 if (ass_var && !is_gimple_reg (ass_var))
484 return;
485
486 /* We found the call, check whether it is suitable. */
487 tail_recursion = false;
488 func = gimple_call_fndecl (call);
489 if (func
490 && !DECL_BUILT_IN (func)
491 && recursive_call_p (current_function_decl, func))
492 {
493 tree arg;
494
495 for (param = DECL_ARGUMENTS (func), idx = 0;
496 param && idx < gimple_call_num_args (call);
497 param = DECL_CHAIN (param), idx ++)
498 {
499 arg = gimple_call_arg (call, idx);
500 if (param != arg)
501 {
502 /* Make sure there are no problems with copying. The parameter
503 have a copyable type and the two arguments must have reasonably
504 equivalent types. The latter requirement could be relaxed if
505 we emitted a suitable type conversion statement. */
506 if (!is_gimple_reg_type (TREE_TYPE (param))
507 || !useless_type_conversion_p (TREE_TYPE (param),
508 TREE_TYPE (arg)))
509 break;
510
511 /* The parameter should be a real operand, so that phi node
512 created for it at the start of the function has the meaning
513 of copying the value. This test implies is_gimple_reg_type
514 from the previous condition, however this one could be
515 relaxed by being more careful with copying the new value
516 of the parameter (emitting appropriate GIMPLE_ASSIGN and
517 updating the virtual operands). */
518 if (!is_gimple_reg (param))
519 break;
520 }
521 }
522 if (idx == gimple_call_num_args (call) && !param)
523 tail_recursion = true;
524 }
525
526 /* Make sure the tail invocation of this function does not refer
527 to local variables. */
528 FOR_EACH_LOCAL_DECL (cfun, idx, var)
529 {
530 if (TREE_CODE (var) != PARM_DECL
531 && auto_var_in_fn_p (var, cfun->decl)
532 && (ref_maybe_used_by_stmt_p (call, var)
533 || call_may_clobber_ref_p (call, var)))
534 return;
535 }
536
537 /* Now check the statements after the call. None of them has virtual
538 operands, so they may only depend on the call through its return
539 value. The return value should also be dependent on each of them,
540 since we are running after dce. */
541 m = NULL_TREE;
542 a = NULL_TREE;
543
544 abb = bb;
545 agsi = gsi;
546 while (1)
547 {
548 tree tmp_a = NULL_TREE;
549 tree tmp_m = NULL_TREE;
550 gsi_next (&agsi);
551
552 while (gsi_end_p (agsi))
553 {
554 ass_var = propagate_through_phis (ass_var, single_succ_edge (abb));
555 abb = single_succ (abb);
556 agsi = gsi_start_bb (abb);
557 }
558
559 stmt = gsi_stmt (agsi);
560
561 if (gimple_code (stmt) == GIMPLE_LABEL)
562 continue;
563
564 if (gimple_code (stmt) == GIMPLE_RETURN)
565 break;
566
567 if (gimple_clobber_p (stmt))
568 continue;
569
570 if (is_gimple_debug (stmt))
571 continue;
572
573 if (gimple_code (stmt) != GIMPLE_ASSIGN)
574 return;
575
576 /* This is a gimple assign. */
577 if (! process_assignment (as_a <gassign *> (stmt), gsi, &tmp_m,
578 &tmp_a, &ass_var))
579 return;
580
581 if (tmp_a)
582 {
583 tree type = TREE_TYPE (tmp_a);
584 if (a)
585 a = fold_build2 (PLUS_EXPR, type, fold_convert (type, a), tmp_a);
586 else
587 a = tmp_a;
588 }
589 if (tmp_m)
590 {
591 tree type = TREE_TYPE (tmp_m);
592 if (m)
593 m = fold_build2 (MULT_EXPR, type, fold_convert (type, m), tmp_m);
594 else
595 m = tmp_m;
596
597 if (a)
598 a = fold_build2 (MULT_EXPR, type, fold_convert (type, a), tmp_m);
599 }
600 }
601
602 /* See if this is a tail call we can handle. */
603 ret_var = gimple_return_retval (as_a <greturn *> (stmt));
604
605 /* We may proceed if there either is no return value, or the return value
606 is identical to the call's return. */
607 if (ret_var
608 && (ret_var != ass_var))
609 return;
610
611 /* If this is not a tail recursive call, we cannot handle addends or
612 multiplicands. */
613 if (!tail_recursion && (m || a))
614 return;
615
616 /* For pointers only allow additions. */
617 if (m && POINTER_TYPE_P (TREE_TYPE (DECL_RESULT (current_function_decl))))
618 return;
619
620 nw = XNEW (struct tailcall);
621
622 nw->call_gsi = gsi;
623
624 nw->tail_recursion = tail_recursion;
625
626 nw->mult = m;
627 nw->add = a;
628
629 nw->next = *ret;
630 *ret = nw;
631 }
632
633 /* Helper to insert PHI_ARGH to the phi of VAR in the destination of edge E. */
634
635 static void
636 add_successor_phi_arg (edge e, tree var, tree phi_arg)
637 {
638 gphi_iterator gsi;
639
640 for (gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi); gsi_next (&gsi))
641 if (PHI_RESULT (gsi.phi ()) == var)
642 break;
643
644 gcc_assert (!gsi_end_p (gsi));
645 add_phi_arg (gsi.phi (), phi_arg, e, UNKNOWN_LOCATION);
646 }
647
648 /* Creates a GIMPLE statement which computes the operation specified by
649 CODE, ACC and OP1 to a new variable with name LABEL and inserts the
650 statement in the position specified by GSI. Returns the
651 tree node of the statement's result. */
652
653 static tree
654 adjust_return_value_with_ops (enum tree_code code, const char *label,
655 tree acc, tree op1, gimple_stmt_iterator gsi)
656 {
657
658 tree ret_type = TREE_TYPE (DECL_RESULT (current_function_decl));
659 tree result = make_temp_ssa_name (ret_type, NULL, label);
660 gassign *stmt;
661
662 if (POINTER_TYPE_P (ret_type))
663 {
664 gcc_assert (code == PLUS_EXPR && TREE_TYPE (acc) == sizetype);
665 code = POINTER_PLUS_EXPR;
666 }
667 if (types_compatible_p (TREE_TYPE (acc), TREE_TYPE (op1))
668 && code != POINTER_PLUS_EXPR)
669 stmt = gimple_build_assign (result, code, acc, op1);
670 else
671 {
672 tree tem;
673 if (code == POINTER_PLUS_EXPR)
674 tem = fold_build2 (code, TREE_TYPE (op1), op1, acc);
675 else
676 tem = fold_build2 (code, TREE_TYPE (op1),
677 fold_convert (TREE_TYPE (op1), acc), op1);
678 tree rhs = fold_convert (ret_type, tem);
679 rhs = force_gimple_operand_gsi (&gsi, rhs,
680 false, NULL, true, GSI_SAME_STMT);
681 stmt = gimple_build_assign (result, rhs);
682 }
683
684 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
685 return result;
686 }
687
688 /* Creates a new GIMPLE statement that adjusts the value of accumulator ACC by
689 the computation specified by CODE and OP1 and insert the statement
690 at the position specified by GSI as a new statement. Returns new SSA name
691 of updated accumulator. */
692
693 static tree
694 update_accumulator_with_ops (enum tree_code code, tree acc, tree op1,
695 gimple_stmt_iterator gsi)
696 {
697 gassign *stmt;
698 tree var = copy_ssa_name (acc);
699 if (types_compatible_p (TREE_TYPE (acc), TREE_TYPE (op1)))
700 stmt = gimple_build_assign (var, code, acc, op1);
701 else
702 {
703 tree rhs = fold_convert (TREE_TYPE (acc),
704 fold_build2 (code,
705 TREE_TYPE (op1),
706 fold_convert (TREE_TYPE (op1), acc),
707 op1));
708 rhs = force_gimple_operand_gsi (&gsi, rhs,
709 false, NULL, false, GSI_CONTINUE_LINKING);
710 stmt = gimple_build_assign (var, rhs);
711 }
712 gsi_insert_after (&gsi, stmt, GSI_NEW_STMT);
713 return var;
714 }
715
716 /* Adjust the accumulator values according to A and M after GSI, and update
717 the phi nodes on edge BACK. */
718
719 static void
720 adjust_accumulator_values (gimple_stmt_iterator gsi, tree m, tree a, edge back)
721 {
722 tree var, a_acc_arg, m_acc_arg;
723
724 if (m)
725 m = force_gimple_operand_gsi (&gsi, m, true, NULL, true, GSI_SAME_STMT);
726 if (a)
727 a = force_gimple_operand_gsi (&gsi, a, true, NULL, true, GSI_SAME_STMT);
728
729 a_acc_arg = a_acc;
730 m_acc_arg = m_acc;
731 if (a)
732 {
733 if (m_acc)
734 {
735 if (integer_onep (a))
736 var = m_acc;
737 else
738 var = adjust_return_value_with_ops (MULT_EXPR, "acc_tmp", m_acc,
739 a, gsi);
740 }
741 else
742 var = a;
743
744 a_acc_arg = update_accumulator_with_ops (PLUS_EXPR, a_acc, var, gsi);
745 }
746
747 if (m)
748 m_acc_arg = update_accumulator_with_ops (MULT_EXPR, m_acc, m, gsi);
749
750 if (a_acc)
751 add_successor_phi_arg (back, a_acc, a_acc_arg);
752
753 if (m_acc)
754 add_successor_phi_arg (back, m_acc, m_acc_arg);
755 }
756
757 /* Adjust value of the return at the end of BB according to M and A
758 accumulators. */
759
760 static void
761 adjust_return_value (basic_block bb, tree m, tree a)
762 {
763 tree retval;
764 greturn *ret_stmt = as_a <greturn *> (gimple_seq_last_stmt (bb_seq (bb)));
765 gimple_stmt_iterator gsi = gsi_last_bb (bb);
766
767 gcc_assert (gimple_code (ret_stmt) == GIMPLE_RETURN);
768
769 retval = gimple_return_retval (ret_stmt);
770 if (!retval || retval == error_mark_node)
771 return;
772
773 if (m)
774 retval = adjust_return_value_with_ops (MULT_EXPR, "mul_tmp", m_acc, retval,
775 gsi);
776 if (a)
777 retval = adjust_return_value_with_ops (PLUS_EXPR, "acc_tmp", a_acc, retval,
778 gsi);
779 gimple_return_set_retval (ret_stmt, retval);
780 update_stmt (ret_stmt);
781 }
782
783 /* Subtract COUNT and FREQUENCY from the basic block and it's
784 outgoing edge. */
785 static void
786 decrease_profile (basic_block bb, gcov_type count, int frequency)
787 {
788 edge e;
789 bb->count -= count;
790 if (bb->count < 0)
791 bb->count = 0;
792 bb->frequency -= frequency;
793 if (bb->frequency < 0)
794 bb->frequency = 0;
795 if (!single_succ_p (bb))
796 {
797 gcc_assert (!EDGE_COUNT (bb->succs));
798 return;
799 }
800 e = single_succ_edge (bb);
801 e->count -= count;
802 if (e->count < 0)
803 e->count = 0;
804 }
805
806 /* Returns true if argument PARAM of the tail recursive call needs to be copied
807 when the call is eliminated. */
808
809 static bool
810 arg_needs_copy_p (tree param)
811 {
812 tree def;
813
814 if (!is_gimple_reg (param))
815 return false;
816
817 /* Parameters that are only defined but never used need not be copied. */
818 def = ssa_default_def (cfun, param);
819 if (!def)
820 return false;
821
822 return true;
823 }
824
825 /* Eliminates tail call described by T. TMP_VARS is a list of
826 temporary variables used to copy the function arguments. */
827
828 static void
829 eliminate_tail_call (struct tailcall *t)
830 {
831 tree param, rslt;
832 gimple stmt, call;
833 tree arg;
834 size_t idx;
835 basic_block bb, first;
836 edge e;
837 gphi *phi;
838 gphi_iterator gpi;
839 gimple_stmt_iterator gsi;
840 gimple orig_stmt;
841
842 stmt = orig_stmt = gsi_stmt (t->call_gsi);
843 bb = gsi_bb (t->call_gsi);
844
845 if (dump_file && (dump_flags & TDF_DETAILS))
846 {
847 fprintf (dump_file, "Eliminated tail recursion in bb %d : ",
848 bb->index);
849 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
850 fprintf (dump_file, "\n");
851 }
852
853 gcc_assert (is_gimple_call (stmt));
854
855 first = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun));
856
857 /* Remove the code after call_gsi that will become unreachable. The
858 possibly unreachable code in other blocks is removed later in
859 cfg cleanup. */
860 gsi = t->call_gsi;
861 gsi_next (&gsi);
862 while (!gsi_end_p (gsi))
863 {
864 gimple t = gsi_stmt (gsi);
865 /* Do not remove the return statement, so that redirect_edge_and_branch
866 sees how the block ends. */
867 if (gimple_code (t) == GIMPLE_RETURN)
868 break;
869
870 gsi_remove (&gsi, true);
871 release_defs (t);
872 }
873
874 /* Number of executions of function has reduced by the tailcall. */
875 e = single_succ_edge (gsi_bb (t->call_gsi));
876 decrease_profile (EXIT_BLOCK_PTR_FOR_FN (cfun), e->count, EDGE_FREQUENCY (e));
877 decrease_profile (ENTRY_BLOCK_PTR_FOR_FN (cfun), e->count,
878 EDGE_FREQUENCY (e));
879 if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun))
880 decrease_profile (e->dest, e->count, EDGE_FREQUENCY (e));
881
882 /* Replace the call by a jump to the start of function. */
883 e = redirect_edge_and_branch (single_succ_edge (gsi_bb (t->call_gsi)),
884 first);
885 gcc_assert (e);
886 PENDING_STMT (e) = NULL;
887
888 /* Add phi node entries for arguments. The ordering of the phi nodes should
889 be the same as the ordering of the arguments. */
890 for (param = DECL_ARGUMENTS (current_function_decl),
891 idx = 0, gpi = gsi_start_phis (first);
892 param;
893 param = DECL_CHAIN (param), idx++)
894 {
895 if (!arg_needs_copy_p (param))
896 continue;
897
898 arg = gimple_call_arg (stmt, idx);
899 phi = gpi.phi ();
900 gcc_assert (param == SSA_NAME_VAR (PHI_RESULT (phi)));
901
902 add_phi_arg (phi, arg, e, gimple_location (stmt));
903 gsi_next (&gpi);
904 }
905
906 /* Update the values of accumulators. */
907 adjust_accumulator_values (t->call_gsi, t->mult, t->add, e);
908
909 call = gsi_stmt (t->call_gsi);
910 rslt = gimple_call_lhs (call);
911 if (rslt != NULL_TREE)
912 {
913 /* Result of the call will no longer be defined. So adjust the
914 SSA_NAME_DEF_STMT accordingly. */
915 SSA_NAME_DEF_STMT (rslt) = gimple_build_nop ();
916 }
917
918 gsi_remove (&t->call_gsi, true);
919 release_defs (call);
920 }
921
922 /* Optimizes the tailcall described by T. If OPT_TAILCALLS is true, also
923 mark the tailcalls for the sibcall optimization. */
924
925 static bool
926 optimize_tail_call (struct tailcall *t, bool opt_tailcalls)
927 {
928 if (t->tail_recursion)
929 {
930 eliminate_tail_call (t);
931 return true;
932 }
933
934 if (opt_tailcalls)
935 {
936 gcall *stmt = as_a <gcall *> (gsi_stmt (t->call_gsi));
937
938 gimple_call_set_tail (stmt, true);
939 cfun->tail_call_marked = true;
940 if (dump_file && (dump_flags & TDF_DETAILS))
941 {
942 fprintf (dump_file, "Found tail call ");
943 print_gimple_stmt (dump_file, stmt, 0, dump_flags);
944 fprintf (dump_file, " in bb %i\n", (gsi_bb (t->call_gsi))->index);
945 }
946 }
947
948 return false;
949 }
950
951 /* Creates a tail-call accumulator of the same type as the return type of the
952 current function. LABEL is the name used to creating the temporary
953 variable for the accumulator. The accumulator will be inserted in the
954 phis of a basic block BB with single predecessor with an initial value
955 INIT converted to the current function return type. */
956
957 static tree
958 create_tailcall_accumulator (const char *label, basic_block bb, tree init)
959 {
960 tree ret_type = TREE_TYPE (DECL_RESULT (current_function_decl));
961 if (POINTER_TYPE_P (ret_type))
962 ret_type = sizetype;
963
964 tree tmp = make_temp_ssa_name (ret_type, NULL, label);
965 gphi *phi;
966
967 phi = create_phi_node (tmp, bb);
968 /* RET_TYPE can be a float when -ffast-maths is enabled. */
969 add_phi_arg (phi, fold_convert (ret_type, init), single_pred_edge (bb),
970 UNKNOWN_LOCATION);
971 return PHI_RESULT (phi);
972 }
973
974 /* Optimizes tail calls in the function, turning the tail recursion
975 into iteration. */
976
977 static unsigned int
978 tree_optimize_tail_calls_1 (bool opt_tailcalls)
979 {
980 edge e;
981 bool phis_constructed = false;
982 struct tailcall *tailcalls = NULL, *act, *next;
983 bool changed = false;
984 basic_block first = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun));
985 tree param;
986 gimple stmt;
987 edge_iterator ei;
988
989 if (!suitable_for_tail_opt_p ())
990 return 0;
991 if (opt_tailcalls)
992 opt_tailcalls = suitable_for_tail_call_opt_p ();
993
994 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds)
995 {
996 /* Only traverse the normal exits, i.e. those that end with return
997 statement. */
998 stmt = last_stmt (e->src);
999
1000 if (stmt
1001 && gimple_code (stmt) == GIMPLE_RETURN)
1002 find_tail_calls (e->src, &tailcalls);
1003 }
1004
1005 /* Construct the phi nodes and accumulators if necessary. */
1006 a_acc = m_acc = NULL_TREE;
1007 for (act = tailcalls; act; act = act->next)
1008 {
1009 if (!act->tail_recursion)
1010 continue;
1011
1012 if (!phis_constructed)
1013 {
1014 /* Ensure that there is only one predecessor of the block
1015 or if there are existing degenerate PHI nodes. */
1016 if (!single_pred_p (first)
1017 || !gimple_seq_empty_p (phi_nodes (first)))
1018 first =
1019 split_edge (single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun)));
1020
1021 /* Copy the args if needed. */
1022 for (param = DECL_ARGUMENTS (current_function_decl);
1023 param;
1024 param = DECL_CHAIN (param))
1025 if (arg_needs_copy_p (param))
1026 {
1027 tree name = ssa_default_def (cfun, param);
1028 tree new_name = make_ssa_name (param, SSA_NAME_DEF_STMT (name));
1029 gphi *phi;
1030
1031 set_ssa_default_def (cfun, param, new_name);
1032 phi = create_phi_node (name, first);
1033 add_phi_arg (phi, new_name, single_pred_edge (first),
1034 EXPR_LOCATION (param));
1035 }
1036 phis_constructed = true;
1037 }
1038
1039 if (act->add && !a_acc)
1040 a_acc = create_tailcall_accumulator ("add_acc", first,
1041 integer_zero_node);
1042
1043 if (act->mult && !m_acc)
1044 m_acc = create_tailcall_accumulator ("mult_acc", first,
1045 integer_one_node);
1046 }
1047
1048 if (a_acc || m_acc)
1049 {
1050 /* When the tail call elimination using accumulators is performed,
1051 statements adding the accumulated value are inserted at all exits.
1052 This turns all other tail calls to non-tail ones. */
1053 opt_tailcalls = false;
1054 }
1055
1056 for (; tailcalls; tailcalls = next)
1057 {
1058 next = tailcalls->next;
1059 changed |= optimize_tail_call (tailcalls, opt_tailcalls);
1060 free (tailcalls);
1061 }
1062
1063 if (a_acc || m_acc)
1064 {
1065 /* Modify the remaining return statements. */
1066 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds)
1067 {
1068 stmt = last_stmt (e->src);
1069
1070 if (stmt
1071 && gimple_code (stmt) == GIMPLE_RETURN)
1072 adjust_return_value (e->src, m_acc, a_acc);
1073 }
1074 }
1075
1076 if (changed)
1077 {
1078 /* We may have created new loops. Make them magically appear. */
1079 loops_state_set (LOOPS_NEED_FIXUP);
1080 free_dominance_info (CDI_DOMINATORS);
1081 }
1082
1083 /* Add phi nodes for the virtual operands defined in the function to the
1084 header of the loop created by tail recursion elimination. Do so
1085 by triggering the SSA renamer. */
1086 if (phis_constructed)
1087 mark_virtual_operands_for_renaming (cfun);
1088
1089 if (changed)
1090 return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals;
1091 return 0;
1092 }
1093
1094 static bool
1095 gate_tail_calls (void)
1096 {
1097 return flag_optimize_sibling_calls != 0 && dbg_cnt (tail_call);
1098 }
1099
1100 static unsigned int
1101 execute_tail_calls (void)
1102 {
1103 return tree_optimize_tail_calls_1 (true);
1104 }
1105
1106 namespace {
1107
1108 const pass_data pass_data_tail_recursion =
1109 {
1110 GIMPLE_PASS, /* type */
1111 "tailr", /* name */
1112 OPTGROUP_NONE, /* optinfo_flags */
1113 TV_NONE, /* tv_id */
1114 ( PROP_cfg | PROP_ssa ), /* properties_required */
1115 0, /* properties_provided */
1116 0, /* properties_destroyed */
1117 0, /* todo_flags_start */
1118 0, /* todo_flags_finish */
1119 };
1120
1121 class pass_tail_recursion : public gimple_opt_pass
1122 {
1123 public:
1124 pass_tail_recursion (gcc::context *ctxt)
1125 : gimple_opt_pass (pass_data_tail_recursion, ctxt)
1126 {}
1127
1128 /* opt_pass methods: */
1129 opt_pass * clone () { return new pass_tail_recursion (m_ctxt); }
1130 virtual bool gate (function *) { return gate_tail_calls (); }
1131 virtual unsigned int execute (function *)
1132 {
1133 return tree_optimize_tail_calls_1 (false);
1134 }
1135
1136 }; // class pass_tail_recursion
1137
1138 } // anon namespace
1139
1140 gimple_opt_pass *
1141 make_pass_tail_recursion (gcc::context *ctxt)
1142 {
1143 return new pass_tail_recursion (ctxt);
1144 }
1145
1146 namespace {
1147
1148 const pass_data pass_data_tail_calls =
1149 {
1150 GIMPLE_PASS, /* type */
1151 "tailc", /* name */
1152 OPTGROUP_NONE, /* optinfo_flags */
1153 TV_NONE, /* tv_id */
1154 ( PROP_cfg | PROP_ssa ), /* properties_required */
1155 0, /* properties_provided */
1156 0, /* properties_destroyed */
1157 0, /* todo_flags_start */
1158 0, /* todo_flags_finish */
1159 };
1160
1161 class pass_tail_calls : public gimple_opt_pass
1162 {
1163 public:
1164 pass_tail_calls (gcc::context *ctxt)
1165 : gimple_opt_pass (pass_data_tail_calls, ctxt)
1166 {}
1167
1168 /* opt_pass methods: */
1169 virtual bool gate (function *) { return gate_tail_calls (); }
1170 virtual unsigned int execute (function *) { return execute_tail_calls (); }
1171
1172 }; // class pass_tail_calls
1173
1174 } // anon namespace
1175
1176 gimple_opt_pass *
1177 make_pass_tail_calls (gcc::context *ctxt)
1178 {
1179 return new pass_tail_calls (ctxt);
1180 }