* c-c++-common/ubsan/float-cast-overflow-6.c: Add i?86-*-* target.
[official-gcc.git] / gcc / tree-tailcall.c
blob551ddcbc4a8ab88ee613894f699d4b5b7c6fa3a9
1 /* Tail call optimization on trees.
2 Copyright (C) 2003-2014 Free Software Foundation, Inc.
4 This file is part of GCC.
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.
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.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "tm.h"
24 #include "tree.h"
25 #include "stor-layout.h"
26 #include "tm_p.h"
27 #include "predict.h"
28 #include "vec.h"
29 #include "hashtab.h"
30 #include "hash-set.h"
31 #include "machmode.h"
32 #include "hard-reg-set.h"
33 #include "input.h"
34 #include "function.h"
35 #include "dominance.h"
36 #include "cfg.h"
37 #include "basic-block.h"
38 #include "tree-ssa-alias.h"
39 #include "internal-fn.h"
40 #include "gimple-expr.h"
41 #include "is-a.h"
42 #include "gimple.h"
43 #include "gimple-iterator.h"
44 #include "gimplify-me.h"
45 #include "gimple-ssa.h"
46 #include "tree-cfg.h"
47 #include "tree-phinodes.h"
48 #include "stringpool.h"
49 #include "tree-ssanames.h"
50 #include "tree-into-ssa.h"
51 #include "expr.h"
52 #include "tree-dfa.h"
53 #include "gimple-pretty-print.h"
54 #include "except.h"
55 #include "tree-pass.h"
56 #include "flags.h"
57 #include "langhooks.h"
58 #include "dbgcnt.h"
59 #include "target.h"
60 #include "cfgloop.h"
61 #include "common/common-target.h"
62 #include "hash-map.h"
63 #include "plugin-api.h"
64 #include "ipa-ref.h"
65 #include "cgraph.h"
66 #include "ipa-utils.h"
68 /* The file implements the tail recursion elimination. It is also used to
69 analyze the tail calls in general, passing the results to the rtl level
70 where they are used for sibcall optimization.
72 In addition to the standard tail recursion elimination, we handle the most
73 trivial cases of making the call tail recursive by creating accumulators.
74 For example the following function
76 int sum (int n)
78 if (n > 0)
79 return n + sum (n - 1);
80 else
81 return 0;
84 is transformed into
86 int sum (int n)
88 int acc = 0;
90 while (n > 0)
91 acc += n--;
93 return acc;
96 To do this, we maintain two accumulators (a_acc and m_acc) that indicate
97 when we reach the return x statement, we should return a_acc + x * m_acc
98 instead. They are initially initialized to 0 and 1, respectively,
99 so the semantics of the function is obviously preserved. If we are
100 guaranteed that the value of the accumulator never change, we
101 omit the accumulator.
103 There are three cases how the function may exit. The first one is
104 handled in adjust_return_value, the other two in adjust_accumulator_values
105 (the second case is actually a special case of the third one and we
106 present it separately just for clarity):
108 1) Just return x, where x is not in any of the remaining special shapes.
109 We rewrite this to a gimple equivalent of return m_acc * x + a_acc.
111 2) return f (...), where f is the current function, is rewritten in a
112 classical tail-recursion elimination way, into assignment of arguments
113 and jump to the start of the function. Values of the accumulators
114 are unchanged.
116 3) return a + m * f(...), where a and m do not depend on call to f.
117 To preserve the semantics described before we want this to be rewritten
118 in such a way that we finally return
120 a_acc + (a + m * f(...)) * m_acc = (a_acc + a * m_acc) + (m * m_acc) * f(...).
122 I.e. we increase a_acc by a * m_acc, multiply m_acc by m and
123 eliminate the tail call to f. Special cases when the value is just
124 added or just multiplied are obtained by setting a = 0 or m = 1.
126 TODO -- it is possible to do similar tricks for other operations. */
128 /* A structure that describes the tailcall. */
130 struct tailcall
132 /* The iterator pointing to the call statement. */
133 gimple_stmt_iterator call_gsi;
135 /* True if it is a call to the current function. */
136 bool tail_recursion;
138 /* The return value of the caller is mult * f + add, where f is the return
139 value of the call. */
140 tree mult, add;
142 /* Next tailcall in the chain. */
143 struct tailcall *next;
146 /* The variables holding the value of multiplicative and additive
147 accumulator. */
148 static tree m_acc, a_acc;
150 static bool suitable_for_tail_opt_p (void);
151 static bool optimize_tail_call (struct tailcall *, bool);
152 static void eliminate_tail_call (struct tailcall *);
153 static void find_tail_calls (basic_block, struct tailcall **);
155 /* Returns false when the function is not suitable for tail call optimization
156 from some reason (e.g. if it takes variable number of arguments). */
158 static bool
159 suitable_for_tail_opt_p (void)
161 if (cfun->stdarg)
162 return false;
164 return true;
166 /* Returns false when the function is not suitable for tail call optimization
167 from some reason (e.g. if it takes variable number of arguments).
168 This test must pass in addition to suitable_for_tail_opt_p in order to make
169 tail call discovery happen. */
171 static bool
172 suitable_for_tail_call_opt_p (void)
174 tree param;
176 /* alloca (until we have stack slot life analysis) inhibits
177 sibling call optimizations, but not tail recursion. */
178 if (cfun->calls_alloca)
179 return false;
181 /* If we are using sjlj exceptions, we may need to add a call to
182 _Unwind_SjLj_Unregister at exit of the function. Which means
183 that we cannot do any sibcall transformations. */
184 if (targetm_common.except_unwind_info (&global_options) == UI_SJLJ
185 && current_function_has_exception_handlers ())
186 return false;
188 /* Any function that calls setjmp might have longjmp called from
189 any called function. ??? We really should represent this
190 properly in the CFG so that this needn't be special cased. */
191 if (cfun->calls_setjmp)
192 return false;
194 /* ??? It is OK if the argument of a function is taken in some cases,
195 but not in all cases. See PR15387 and PR19616. Revisit for 4.1. */
196 for (param = DECL_ARGUMENTS (current_function_decl);
197 param;
198 param = DECL_CHAIN (param))
199 if (TREE_ADDRESSABLE (param))
200 return false;
202 return true;
205 /* Checks whether the expression EXPR in stmt AT is independent of the
206 statement pointed to by GSI (in a sense that we already know EXPR's value
207 at GSI). We use the fact that we are only called from the chain of
208 basic blocks that have only single successor. Returns the expression
209 containing the value of EXPR at GSI. */
211 static tree
212 independent_of_stmt_p (tree expr, gimple at, gimple_stmt_iterator gsi)
214 basic_block bb, call_bb, at_bb;
215 edge e;
216 edge_iterator ei;
218 if (is_gimple_min_invariant (expr))
219 return expr;
221 if (TREE_CODE (expr) != SSA_NAME)
222 return NULL_TREE;
224 /* Mark the blocks in the chain leading to the end. */
225 at_bb = gimple_bb (at);
226 call_bb = gimple_bb (gsi_stmt (gsi));
227 for (bb = call_bb; bb != at_bb; bb = single_succ (bb))
228 bb->aux = &bb->aux;
229 bb->aux = &bb->aux;
231 while (1)
233 at = SSA_NAME_DEF_STMT (expr);
234 bb = gimple_bb (at);
236 /* The default definition or defined before the chain. */
237 if (!bb || !bb->aux)
238 break;
240 if (bb == call_bb)
242 for (; !gsi_end_p (gsi); gsi_next (&gsi))
243 if (gsi_stmt (gsi) == at)
244 break;
246 if (!gsi_end_p (gsi))
247 expr = NULL_TREE;
248 break;
251 if (gimple_code (at) != GIMPLE_PHI)
253 expr = NULL_TREE;
254 break;
257 FOR_EACH_EDGE (e, ei, bb->preds)
258 if (e->src->aux)
259 break;
260 gcc_assert (e);
262 expr = PHI_ARG_DEF_FROM_EDGE (at, e);
263 if (TREE_CODE (expr) != SSA_NAME)
265 /* The value is a constant. */
266 break;
270 /* Unmark the blocks. */
271 for (bb = call_bb; bb != at_bb; bb = single_succ (bb))
272 bb->aux = NULL;
273 bb->aux = NULL;
275 return expr;
278 /* Simulates the effect of an assignment STMT on the return value of the tail
279 recursive CALL passed in ASS_VAR. M and A are the multiplicative and the
280 additive factor for the real return value. */
282 static bool
283 process_assignment (gimple stmt, gimple_stmt_iterator call, tree *m,
284 tree *a, tree *ass_var)
286 tree op0, op1 = NULL_TREE, non_ass_var = NULL_TREE;
287 tree dest = gimple_assign_lhs (stmt);
288 enum tree_code code = gimple_assign_rhs_code (stmt);
289 enum gimple_rhs_class rhs_class = get_gimple_rhs_class (code);
290 tree src_var = gimple_assign_rhs1 (stmt);
292 /* See if this is a simple copy operation of an SSA name to the function
293 result. In that case we may have a simple tail call. Ignore type
294 conversions that can never produce extra code between the function
295 call and the function return. */
296 if ((rhs_class == GIMPLE_SINGLE_RHS || gimple_assign_cast_p (stmt))
297 && (TREE_CODE (src_var) == SSA_NAME))
299 /* Reject a tailcall if the type conversion might need
300 additional code. */
301 if (gimple_assign_cast_p (stmt))
303 if (TYPE_MODE (TREE_TYPE (dest)) != TYPE_MODE (TREE_TYPE (src_var)))
304 return false;
306 /* Even if the type modes are the same, if the precision of the
307 type is smaller than mode's precision,
308 reduce_to_bit_field_precision would generate additional code. */
309 if (INTEGRAL_TYPE_P (TREE_TYPE (dest))
310 && (GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (dest)))
311 > TYPE_PRECISION (TREE_TYPE (dest))))
312 return false;
315 if (src_var != *ass_var)
316 return false;
318 *ass_var = dest;
319 return true;
322 switch (rhs_class)
324 case GIMPLE_BINARY_RHS:
325 op1 = gimple_assign_rhs2 (stmt);
327 /* Fall through. */
329 case GIMPLE_UNARY_RHS:
330 op0 = gimple_assign_rhs1 (stmt);
331 break;
333 default:
334 return false;
337 /* Accumulator optimizations will reverse the order of operations.
338 We can only do that for floating-point types if we're assuming
339 that addition and multiplication are associative. */
340 if (!flag_associative_math)
341 if (FLOAT_TYPE_P (TREE_TYPE (DECL_RESULT (current_function_decl))))
342 return false;
344 if (rhs_class == GIMPLE_UNARY_RHS)
346 else if (op0 == *ass_var
347 && (non_ass_var = independent_of_stmt_p (op1, stmt, call)))
349 else if (op1 == *ass_var
350 && (non_ass_var = independent_of_stmt_p (op0, stmt, call)))
352 else
353 return false;
355 switch (code)
357 case PLUS_EXPR:
358 *a = non_ass_var;
359 *ass_var = dest;
360 return true;
362 case POINTER_PLUS_EXPR:
363 if (op0 != *ass_var)
364 return false;
365 *a = non_ass_var;
366 *ass_var = dest;
367 return true;
369 case MULT_EXPR:
370 *m = non_ass_var;
371 *ass_var = dest;
372 return true;
374 case NEGATE_EXPR:
375 *m = build_minus_one_cst (TREE_TYPE (op0));
376 *ass_var = dest;
377 return true;
379 case MINUS_EXPR:
380 if (*ass_var == op0)
381 *a = fold_build1 (NEGATE_EXPR, TREE_TYPE (non_ass_var), non_ass_var);
382 else
384 *m = build_minus_one_cst (TREE_TYPE (non_ass_var));
385 *a = fold_build1 (NEGATE_EXPR, TREE_TYPE (non_ass_var), non_ass_var);
388 *ass_var = dest;
389 return true;
391 /* TODO -- Handle POINTER_PLUS_EXPR. */
393 default:
394 return false;
398 /* Propagate VAR through phis on edge E. */
400 static tree
401 propagate_through_phis (tree var, edge e)
403 basic_block dest = e->dest;
404 gimple_stmt_iterator gsi;
406 for (gsi = gsi_start_phis (dest); !gsi_end_p (gsi); gsi_next (&gsi))
408 gimple phi = gsi_stmt (gsi);
409 if (PHI_ARG_DEF_FROM_EDGE (phi, e) == var)
410 return PHI_RESULT (phi);
412 return var;
415 /* Finds tailcalls falling into basic block BB. The list of found tailcalls is
416 added to the start of RET. */
418 static void
419 find_tail_calls (basic_block bb, struct tailcall **ret)
421 tree ass_var = NULL_TREE, ret_var, func, param;
422 gimple stmt, call = NULL;
423 gimple_stmt_iterator gsi, agsi;
424 bool tail_recursion;
425 struct tailcall *nw;
426 edge e;
427 tree m, a;
428 basic_block abb;
429 size_t idx;
430 tree var;
432 if (!single_succ_p (bb))
433 return;
435 for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi); gsi_prev (&gsi))
437 stmt = gsi_stmt (gsi);
439 /* Ignore labels, returns, clobbers and debug stmts. */
440 if (gimple_code (stmt) == GIMPLE_LABEL
441 || gimple_code (stmt) == GIMPLE_RETURN
442 || gimple_clobber_p (stmt)
443 || is_gimple_debug (stmt))
444 continue;
446 /* Check for a call. */
447 if (is_gimple_call (stmt))
449 call = stmt;
450 ass_var = gimple_call_lhs (stmt);
451 break;
454 /* If the statement references memory or volatile operands, fail. */
455 if (gimple_references_memory_p (stmt)
456 || gimple_has_volatile_ops (stmt))
457 return;
460 if (gsi_end_p (gsi))
462 edge_iterator ei;
463 /* Recurse to the predecessors. */
464 FOR_EACH_EDGE (e, ei, bb->preds)
465 find_tail_calls (e->src, ret);
467 return;
470 /* If the LHS of our call is not just a simple register, we can't
471 transform this into a tail or sibling call. This situation happens,
472 in (e.g.) "*p = foo()" where foo returns a struct. In this case
473 we won't have a temporary here, but we need to carry out the side
474 effect anyway, so tailcall is impossible.
476 ??? In some situations (when the struct is returned in memory via
477 invisible argument) we could deal with this, e.g. by passing 'p'
478 itself as that argument to foo, but it's too early to do this here,
479 and expand_call() will not handle it anyway. If it ever can, then
480 we need to revisit this here, to allow that situation. */
481 if (ass_var && !is_gimple_reg (ass_var))
482 return;
484 /* We found the call, check whether it is suitable. */
485 tail_recursion = false;
486 func = gimple_call_fndecl (call);
487 if (func
488 && !DECL_BUILT_IN (func)
489 && recursive_call_p (current_function_decl, func))
491 tree arg;
493 for (param = DECL_ARGUMENTS (func), idx = 0;
494 param && idx < gimple_call_num_args (call);
495 param = DECL_CHAIN (param), idx ++)
497 arg = gimple_call_arg (call, idx);
498 if (param != arg)
500 /* Make sure there are no problems with copying. The parameter
501 have a copyable type and the two arguments must have reasonably
502 equivalent types. The latter requirement could be relaxed if
503 we emitted a suitable type conversion statement. */
504 if (!is_gimple_reg_type (TREE_TYPE (param))
505 || !useless_type_conversion_p (TREE_TYPE (param),
506 TREE_TYPE (arg)))
507 break;
509 /* The parameter should be a real operand, so that phi node
510 created for it at the start of the function has the meaning
511 of copying the value. This test implies is_gimple_reg_type
512 from the previous condition, however this one could be
513 relaxed by being more careful with copying the new value
514 of the parameter (emitting appropriate GIMPLE_ASSIGN and
515 updating the virtual operands). */
516 if (!is_gimple_reg (param))
517 break;
520 if (idx == gimple_call_num_args (call) && !param)
521 tail_recursion = true;
524 /* Make sure the tail invocation of this function does not refer
525 to local variables. */
526 FOR_EACH_LOCAL_DECL (cfun, idx, var)
528 if (TREE_CODE (var) != PARM_DECL
529 && auto_var_in_fn_p (var, cfun->decl)
530 && (ref_maybe_used_by_stmt_p (call, var)
531 || call_may_clobber_ref_p (call, var)))
532 return;
535 /* Now check the statements after the call. None of them has virtual
536 operands, so they may only depend on the call through its return
537 value. The return value should also be dependent on each of them,
538 since we are running after dce. */
539 m = NULL_TREE;
540 a = NULL_TREE;
542 abb = bb;
543 agsi = gsi;
544 while (1)
546 tree tmp_a = NULL_TREE;
547 tree tmp_m = NULL_TREE;
548 gsi_next (&agsi);
550 while (gsi_end_p (agsi))
552 ass_var = propagate_through_phis (ass_var, single_succ_edge (abb));
553 abb = single_succ (abb);
554 agsi = gsi_start_bb (abb);
557 stmt = gsi_stmt (agsi);
559 if (gimple_code (stmt) == GIMPLE_LABEL)
560 continue;
562 if (gimple_code (stmt) == GIMPLE_RETURN)
563 break;
565 if (gimple_clobber_p (stmt))
566 continue;
568 if (is_gimple_debug (stmt))
569 continue;
571 if (gimple_code (stmt) != GIMPLE_ASSIGN)
572 return;
574 /* This is a gimple assign. */
575 if (! process_assignment (stmt, gsi, &tmp_m, &tmp_a, &ass_var))
576 return;
578 if (tmp_a)
580 tree type = TREE_TYPE (tmp_a);
581 if (a)
582 a = fold_build2 (PLUS_EXPR, type, fold_convert (type, a), tmp_a);
583 else
584 a = tmp_a;
586 if (tmp_m)
588 tree type = TREE_TYPE (tmp_m);
589 if (m)
590 m = fold_build2 (MULT_EXPR, type, fold_convert (type, m), tmp_m);
591 else
592 m = tmp_m;
594 if (a)
595 a = fold_build2 (MULT_EXPR, type, fold_convert (type, a), tmp_m);
599 /* See if this is a tail call we can handle. */
600 ret_var = gimple_return_retval (stmt);
602 /* We may proceed if there either is no return value, or the return value
603 is identical to the call's return. */
604 if (ret_var
605 && (ret_var != ass_var))
606 return;
608 /* If this is not a tail recursive call, we cannot handle addends or
609 multiplicands. */
610 if (!tail_recursion && (m || a))
611 return;
613 /* For pointers only allow additions. */
614 if (m && POINTER_TYPE_P (TREE_TYPE (DECL_RESULT (current_function_decl))))
615 return;
617 nw = XNEW (struct tailcall);
619 nw->call_gsi = gsi;
621 nw->tail_recursion = tail_recursion;
623 nw->mult = m;
624 nw->add = a;
626 nw->next = *ret;
627 *ret = nw;
630 /* Helper to insert PHI_ARGH to the phi of VAR in the destination of edge E. */
632 static void
633 add_successor_phi_arg (edge e, tree var, tree phi_arg)
635 gimple_stmt_iterator gsi;
637 for (gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi); gsi_next (&gsi))
638 if (PHI_RESULT (gsi_stmt (gsi)) == var)
639 break;
641 gcc_assert (!gsi_end_p (gsi));
642 add_phi_arg (gsi_stmt (gsi), phi_arg, e, UNKNOWN_LOCATION);
645 /* Creates a GIMPLE statement which computes the operation specified by
646 CODE, ACC and OP1 to a new variable with name LABEL and inserts the
647 statement in the position specified by GSI. Returns the
648 tree node of the statement's result. */
650 static tree
651 adjust_return_value_with_ops (enum tree_code code, const char *label,
652 tree acc, tree op1, gimple_stmt_iterator gsi)
655 tree ret_type = TREE_TYPE (DECL_RESULT (current_function_decl));
656 tree result = make_temp_ssa_name (ret_type, NULL, label);
657 gimple stmt;
659 if (POINTER_TYPE_P (ret_type))
661 gcc_assert (code == PLUS_EXPR && TREE_TYPE (acc) == sizetype);
662 code = POINTER_PLUS_EXPR;
664 if (types_compatible_p (TREE_TYPE (acc), TREE_TYPE (op1))
665 && code != POINTER_PLUS_EXPR)
666 stmt = gimple_build_assign_with_ops (code, result, acc, op1);
667 else
669 tree tem;
670 if (code == POINTER_PLUS_EXPR)
671 tem = fold_build2 (code, TREE_TYPE (op1), op1, acc);
672 else
673 tem = fold_build2 (code, TREE_TYPE (op1),
674 fold_convert (TREE_TYPE (op1), acc), op1);
675 tree rhs = fold_convert (ret_type, tem);
676 rhs = force_gimple_operand_gsi (&gsi, rhs,
677 false, NULL, true, GSI_SAME_STMT);
678 stmt = gimple_build_assign (result, rhs);
681 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
682 return result;
685 /* Creates a new GIMPLE statement that adjusts the value of accumulator ACC by
686 the computation specified by CODE and OP1 and insert the statement
687 at the position specified by GSI as a new statement. Returns new SSA name
688 of updated accumulator. */
690 static tree
691 update_accumulator_with_ops (enum tree_code code, tree acc, tree op1,
692 gimple_stmt_iterator gsi)
694 gimple stmt;
695 tree var = copy_ssa_name (acc, NULL);
696 if (types_compatible_p (TREE_TYPE (acc), TREE_TYPE (op1)))
697 stmt = gimple_build_assign_with_ops (code, var, acc, op1);
698 else
700 tree rhs = fold_convert (TREE_TYPE (acc),
701 fold_build2 (code,
702 TREE_TYPE (op1),
703 fold_convert (TREE_TYPE (op1), acc),
704 op1));
705 rhs = force_gimple_operand_gsi (&gsi, rhs,
706 false, NULL, false, GSI_CONTINUE_LINKING);
707 stmt = gimple_build_assign (var, rhs);
709 gsi_insert_after (&gsi, stmt, GSI_NEW_STMT);
710 return var;
713 /* Adjust the accumulator values according to A and M after GSI, and update
714 the phi nodes on edge BACK. */
716 static void
717 adjust_accumulator_values (gimple_stmt_iterator gsi, tree m, tree a, edge back)
719 tree var, a_acc_arg, m_acc_arg;
721 if (m)
722 m = force_gimple_operand_gsi (&gsi, m, true, NULL, true, GSI_SAME_STMT);
723 if (a)
724 a = force_gimple_operand_gsi (&gsi, a, true, NULL, true, GSI_SAME_STMT);
726 a_acc_arg = a_acc;
727 m_acc_arg = m_acc;
728 if (a)
730 if (m_acc)
732 if (integer_onep (a))
733 var = m_acc;
734 else
735 var = adjust_return_value_with_ops (MULT_EXPR, "acc_tmp", m_acc,
736 a, gsi);
738 else
739 var = a;
741 a_acc_arg = update_accumulator_with_ops (PLUS_EXPR, a_acc, var, gsi);
744 if (m)
745 m_acc_arg = update_accumulator_with_ops (MULT_EXPR, m_acc, m, gsi);
747 if (a_acc)
748 add_successor_phi_arg (back, a_acc, a_acc_arg);
750 if (m_acc)
751 add_successor_phi_arg (back, m_acc, m_acc_arg);
754 /* Adjust value of the return at the end of BB according to M and A
755 accumulators. */
757 static void
758 adjust_return_value (basic_block bb, tree m, tree a)
760 tree retval;
761 gimple ret_stmt = gimple_seq_last_stmt (bb_seq (bb));
762 gimple_stmt_iterator gsi = gsi_last_bb (bb);
764 gcc_assert (gimple_code (ret_stmt) == GIMPLE_RETURN);
766 retval = gimple_return_retval (ret_stmt);
767 if (!retval || retval == error_mark_node)
768 return;
770 if (m)
771 retval = adjust_return_value_with_ops (MULT_EXPR, "mul_tmp", m_acc, retval,
772 gsi);
773 if (a)
774 retval = adjust_return_value_with_ops (PLUS_EXPR, "acc_tmp", a_acc, retval,
775 gsi);
776 gimple_return_set_retval (ret_stmt, retval);
777 update_stmt (ret_stmt);
780 /* Subtract COUNT and FREQUENCY from the basic block and it's
781 outgoing edge. */
782 static void
783 decrease_profile (basic_block bb, gcov_type count, int frequency)
785 edge e;
786 bb->count -= count;
787 if (bb->count < 0)
788 bb->count = 0;
789 bb->frequency -= frequency;
790 if (bb->frequency < 0)
791 bb->frequency = 0;
792 if (!single_succ_p (bb))
794 gcc_assert (!EDGE_COUNT (bb->succs));
795 return;
797 e = single_succ_edge (bb);
798 e->count -= count;
799 if (e->count < 0)
800 e->count = 0;
803 /* Returns true if argument PARAM of the tail recursive call needs to be copied
804 when the call is eliminated. */
806 static bool
807 arg_needs_copy_p (tree param)
809 tree def;
811 if (!is_gimple_reg (param))
812 return false;
814 /* Parameters that are only defined but never used need not be copied. */
815 def = ssa_default_def (cfun, param);
816 if (!def)
817 return false;
819 return true;
822 /* Eliminates tail call described by T. TMP_VARS is a list of
823 temporary variables used to copy the function arguments. */
825 static void
826 eliminate_tail_call (struct tailcall *t)
828 tree param, rslt;
829 gimple stmt, call;
830 tree arg;
831 size_t idx;
832 basic_block bb, first;
833 edge e;
834 gimple phi;
835 gimple_stmt_iterator gsi;
836 gimple orig_stmt;
838 stmt = orig_stmt = gsi_stmt (t->call_gsi);
839 bb = gsi_bb (t->call_gsi);
841 if (dump_file && (dump_flags & TDF_DETAILS))
843 fprintf (dump_file, "Eliminated tail recursion in bb %d : ",
844 bb->index);
845 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
846 fprintf (dump_file, "\n");
849 gcc_assert (is_gimple_call (stmt));
851 first = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun));
853 /* Remove the code after call_gsi that will become unreachable. The
854 possibly unreachable code in other blocks is removed later in
855 cfg cleanup. */
856 gsi = t->call_gsi;
857 gsi_next (&gsi);
858 while (!gsi_end_p (gsi))
860 gimple t = gsi_stmt (gsi);
861 /* Do not remove the return statement, so that redirect_edge_and_branch
862 sees how the block ends. */
863 if (gimple_code (t) == GIMPLE_RETURN)
864 break;
866 gsi_remove (&gsi, true);
867 release_defs (t);
870 /* Number of executions of function has reduced by the tailcall. */
871 e = single_succ_edge (gsi_bb (t->call_gsi));
872 decrease_profile (EXIT_BLOCK_PTR_FOR_FN (cfun), e->count, EDGE_FREQUENCY (e));
873 decrease_profile (ENTRY_BLOCK_PTR_FOR_FN (cfun), e->count,
874 EDGE_FREQUENCY (e));
875 if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun))
876 decrease_profile (e->dest, e->count, EDGE_FREQUENCY (e));
878 /* Replace the call by a jump to the start of function. */
879 e = redirect_edge_and_branch (single_succ_edge (gsi_bb (t->call_gsi)),
880 first);
881 gcc_assert (e);
882 PENDING_STMT (e) = NULL;
884 /* Add phi node entries for arguments. The ordering of the phi nodes should
885 be the same as the ordering of the arguments. */
886 for (param = DECL_ARGUMENTS (current_function_decl),
887 idx = 0, gsi = gsi_start_phis (first);
888 param;
889 param = DECL_CHAIN (param), idx++)
891 if (!arg_needs_copy_p (param))
892 continue;
894 arg = gimple_call_arg (stmt, idx);
895 phi = gsi_stmt (gsi);
896 gcc_assert (param == SSA_NAME_VAR (PHI_RESULT (phi)));
898 add_phi_arg (phi, arg, e, gimple_location (stmt));
899 gsi_next (&gsi);
902 /* Update the values of accumulators. */
903 adjust_accumulator_values (t->call_gsi, t->mult, t->add, e);
905 call = gsi_stmt (t->call_gsi);
906 rslt = gimple_call_lhs (call);
907 if (rslt != NULL_TREE)
909 /* Result of the call will no longer be defined. So adjust the
910 SSA_NAME_DEF_STMT accordingly. */
911 SSA_NAME_DEF_STMT (rslt) = gimple_build_nop ();
914 gsi_remove (&t->call_gsi, true);
915 release_defs (call);
918 /* Optimizes the tailcall described by T. If OPT_TAILCALLS is true, also
919 mark the tailcalls for the sibcall optimization. */
921 static bool
922 optimize_tail_call (struct tailcall *t, bool opt_tailcalls)
924 if (t->tail_recursion)
926 eliminate_tail_call (t);
927 return true;
930 if (opt_tailcalls)
932 gimple stmt = gsi_stmt (t->call_gsi);
934 gimple_call_set_tail (stmt, true);
935 cfun->tail_call_marked = true;
936 if (dump_file && (dump_flags & TDF_DETAILS))
938 fprintf (dump_file, "Found tail call ");
939 print_gimple_stmt (dump_file, stmt, 0, dump_flags);
940 fprintf (dump_file, " in bb %i\n", (gsi_bb (t->call_gsi))->index);
944 return false;
947 /* Creates a tail-call accumulator of the same type as the return type of the
948 current function. LABEL is the name used to creating the temporary
949 variable for the accumulator. The accumulator will be inserted in the
950 phis of a basic block BB with single predecessor with an initial value
951 INIT converted to the current function return type. */
953 static tree
954 create_tailcall_accumulator (const char *label, basic_block bb, tree init)
956 tree ret_type = TREE_TYPE (DECL_RESULT (current_function_decl));
957 if (POINTER_TYPE_P (ret_type))
958 ret_type = sizetype;
960 tree tmp = make_temp_ssa_name (ret_type, NULL, label);
961 gimple phi;
963 phi = create_phi_node (tmp, bb);
964 /* RET_TYPE can be a float when -ffast-maths is enabled. */
965 add_phi_arg (phi, fold_convert (ret_type, init), single_pred_edge (bb),
966 UNKNOWN_LOCATION);
967 return PHI_RESULT (phi);
970 /* Optimizes tail calls in the function, turning the tail recursion
971 into iteration. */
973 static unsigned int
974 tree_optimize_tail_calls_1 (bool opt_tailcalls)
976 edge e;
977 bool phis_constructed = false;
978 struct tailcall *tailcalls = NULL, *act, *next;
979 bool changed = false;
980 basic_block first = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun));
981 tree param;
982 gimple stmt;
983 edge_iterator ei;
985 if (!suitable_for_tail_opt_p ())
986 return 0;
987 if (opt_tailcalls)
988 opt_tailcalls = suitable_for_tail_call_opt_p ();
990 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds)
992 /* Only traverse the normal exits, i.e. those that end with return
993 statement. */
994 stmt = last_stmt (e->src);
996 if (stmt
997 && gimple_code (stmt) == GIMPLE_RETURN)
998 find_tail_calls (e->src, &tailcalls);
1001 /* Construct the phi nodes and accumulators if necessary. */
1002 a_acc = m_acc = NULL_TREE;
1003 for (act = tailcalls; act; act = act->next)
1005 if (!act->tail_recursion)
1006 continue;
1008 if (!phis_constructed)
1010 /* Ensure that there is only one predecessor of the block
1011 or if there are existing degenerate PHI nodes. */
1012 if (!single_pred_p (first)
1013 || !gimple_seq_empty_p (phi_nodes (first)))
1014 first =
1015 split_edge (single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun)));
1017 /* Copy the args if needed. */
1018 for (param = DECL_ARGUMENTS (current_function_decl);
1019 param;
1020 param = DECL_CHAIN (param))
1021 if (arg_needs_copy_p (param))
1023 tree name = ssa_default_def (cfun, param);
1024 tree new_name = make_ssa_name (param, SSA_NAME_DEF_STMT (name));
1025 gimple phi;
1027 set_ssa_default_def (cfun, param, new_name);
1028 phi = create_phi_node (name, first);
1029 add_phi_arg (phi, new_name, single_pred_edge (first),
1030 EXPR_LOCATION (param));
1032 phis_constructed = true;
1035 if (act->add && !a_acc)
1036 a_acc = create_tailcall_accumulator ("add_acc", first,
1037 integer_zero_node);
1039 if (act->mult && !m_acc)
1040 m_acc = create_tailcall_accumulator ("mult_acc", first,
1041 integer_one_node);
1044 if (a_acc || m_acc)
1046 /* When the tail call elimination using accumulators is performed,
1047 statements adding the accumulated value are inserted at all exits.
1048 This turns all other tail calls to non-tail ones. */
1049 opt_tailcalls = false;
1052 for (; tailcalls; tailcalls = next)
1054 next = tailcalls->next;
1055 changed |= optimize_tail_call (tailcalls, opt_tailcalls);
1056 free (tailcalls);
1059 if (a_acc || m_acc)
1061 /* Modify the remaining return statements. */
1062 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds)
1064 stmt = last_stmt (e->src);
1066 if (stmt
1067 && gimple_code (stmt) == GIMPLE_RETURN)
1068 adjust_return_value (e->src, m_acc, a_acc);
1072 if (changed)
1074 /* We may have created new loops. Make them magically appear. */
1075 loops_state_set (LOOPS_NEED_FIXUP);
1076 free_dominance_info (CDI_DOMINATORS);
1079 /* Add phi nodes for the virtual operands defined in the function to the
1080 header of the loop created by tail recursion elimination. Do so
1081 by triggering the SSA renamer. */
1082 if (phis_constructed)
1083 mark_virtual_operands_for_renaming (cfun);
1085 if (changed)
1086 return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals;
1087 return 0;
1090 static bool
1091 gate_tail_calls (void)
1093 return flag_optimize_sibling_calls != 0 && dbg_cnt (tail_call);
1096 static unsigned int
1097 execute_tail_calls (void)
1099 return tree_optimize_tail_calls_1 (true);
1102 namespace {
1104 const pass_data pass_data_tail_recursion =
1106 GIMPLE_PASS, /* type */
1107 "tailr", /* name */
1108 OPTGROUP_NONE, /* optinfo_flags */
1109 TV_NONE, /* tv_id */
1110 ( PROP_cfg | PROP_ssa ), /* properties_required */
1111 0, /* properties_provided */
1112 0, /* properties_destroyed */
1113 0, /* todo_flags_start */
1114 0, /* todo_flags_finish */
1117 class pass_tail_recursion : public gimple_opt_pass
1119 public:
1120 pass_tail_recursion (gcc::context *ctxt)
1121 : gimple_opt_pass (pass_data_tail_recursion, ctxt)
1124 /* opt_pass methods: */
1125 opt_pass * clone () { return new pass_tail_recursion (m_ctxt); }
1126 virtual bool gate (function *) { return gate_tail_calls (); }
1127 virtual unsigned int execute (function *)
1129 return tree_optimize_tail_calls_1 (false);
1132 }; // class pass_tail_recursion
1134 } // anon namespace
1136 gimple_opt_pass *
1137 make_pass_tail_recursion (gcc::context *ctxt)
1139 return new pass_tail_recursion (ctxt);
1142 namespace {
1144 const pass_data pass_data_tail_calls =
1146 GIMPLE_PASS, /* type */
1147 "tailc", /* name */
1148 OPTGROUP_NONE, /* optinfo_flags */
1149 TV_NONE, /* tv_id */
1150 ( PROP_cfg | PROP_ssa ), /* properties_required */
1151 0, /* properties_provided */
1152 0, /* properties_destroyed */
1153 0, /* todo_flags_start */
1154 0, /* todo_flags_finish */
1157 class pass_tail_calls : public gimple_opt_pass
1159 public:
1160 pass_tail_calls (gcc::context *ctxt)
1161 : gimple_opt_pass (pass_data_tail_calls, ctxt)
1164 /* opt_pass methods: */
1165 virtual bool gate (function *) { return gate_tail_calls (); }
1166 virtual unsigned int execute (function *) { return execute_tail_calls (); }
1168 }; // class pass_tail_calls
1170 } // anon namespace
1172 gimple_opt_pass *
1173 make_pass_tail_calls (gcc::context *ctxt)
1175 return new pass_tail_calls (ctxt);