clang-format: Enhance list of FOR_EACH macros
[official-gcc.git] / gcc / tree-tailcall.c
blobbbd1b29ae48d423a49353ddbe591977bf6fadbc2
1 /* Tail call optimization on trees.
2 Copyright (C) 2003-2015 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 "backend.h"
24 #include "rtl.h"
25 #include "tree.h"
26 #include "gimple.h"
27 #include "cfghooks.h"
28 #include "tree-pass.h"
29 #include "ssa.h"
30 #include "cgraph.h"
31 #include "gimple-pretty-print.h"
32 #include "fold-const.h"
33 #include "stor-layout.h"
34 #include "gimple-iterator.h"
35 #include "gimplify-me.h"
36 #include "tree-cfg.h"
37 #include "tree-into-ssa.h"
38 #include "tree-dfa.h"
39 #include "except.h"
40 #include "dbgcnt.h"
41 #include "cfgloop.h"
42 #include "common/common-target.h"
43 #include "ipa-utils.h"
45 /* The file implements the tail recursion elimination. It is also used to
46 analyze the tail calls in general, passing the results to the rtl level
47 where they are used for sibcall optimization.
49 In addition to the standard tail recursion elimination, we handle the most
50 trivial cases of making the call tail recursive by creating accumulators.
51 For example the following function
53 int sum (int n)
55 if (n > 0)
56 return n + sum (n - 1);
57 else
58 return 0;
61 is transformed into
63 int sum (int n)
65 int acc = 0;
67 while (n > 0)
68 acc += n--;
70 return acc;
73 To do this, we maintain two accumulators (a_acc and m_acc) that indicate
74 when we reach the return x statement, we should return a_acc + x * m_acc
75 instead. They are initially initialized to 0 and 1, respectively,
76 so the semantics of the function is obviously preserved. If we are
77 guaranteed that the value of the accumulator never change, we
78 omit the accumulator.
80 There are three cases how the function may exit. The first one is
81 handled in adjust_return_value, the other two in adjust_accumulator_values
82 (the second case is actually a special case of the third one and we
83 present it separately just for clarity):
85 1) Just return x, where x is not in any of the remaining special shapes.
86 We rewrite this to a gimple equivalent of return m_acc * x + a_acc.
88 2) return f (...), where f is the current function, is rewritten in a
89 classical tail-recursion elimination way, into assignment of arguments
90 and jump to the start of the function. Values of the accumulators
91 are unchanged.
93 3) return a + m * f(...), where a and m do not depend on call to f.
94 To preserve the semantics described before we want this to be rewritten
95 in such a way that we finally return
97 a_acc + (a + m * f(...)) * m_acc = (a_acc + a * m_acc) + (m * m_acc) * f(...).
99 I.e. we increase a_acc by a * m_acc, multiply m_acc by m and
100 eliminate the tail call to f. Special cases when the value is just
101 added or just multiplied are obtained by setting a = 0 or m = 1.
103 TODO -- it is possible to do similar tricks for other operations. */
105 /* A structure that describes the tailcall. */
107 struct tailcall
109 /* The iterator pointing to the call statement. */
110 gimple_stmt_iterator call_gsi;
112 /* True if it is a call to the current function. */
113 bool tail_recursion;
115 /* The return value of the caller is mult * f + add, where f is the return
116 value of the call. */
117 tree mult, add;
119 /* Next tailcall in the chain. */
120 struct tailcall *next;
123 /* The variables holding the value of multiplicative and additive
124 accumulator. */
125 static tree m_acc, a_acc;
127 static bool optimize_tail_call (struct tailcall *, bool);
128 static void eliminate_tail_call (struct tailcall *);
130 /* Returns false when the function is not suitable for tail call optimization
131 from some reason (e.g. if it takes variable number of arguments). */
133 static bool
134 suitable_for_tail_opt_p (void)
136 if (cfun->stdarg)
137 return false;
139 return true;
141 /* Returns false when the function is not suitable for tail call optimization
142 for some reason (e.g. if it takes variable number of arguments).
143 This test must pass in addition to suitable_for_tail_opt_p in order to make
144 tail call discovery happen. */
146 static bool
147 suitable_for_tail_call_opt_p (void)
149 tree param;
151 /* alloca (until we have stack slot life analysis) inhibits
152 sibling call optimizations, but not tail recursion. */
153 if (cfun->calls_alloca)
154 return false;
156 /* If we are using sjlj exceptions, we may need to add a call to
157 _Unwind_SjLj_Unregister at exit of the function. Which means
158 that we cannot do any sibcall transformations. */
159 if (targetm_common.except_unwind_info (&global_options) == UI_SJLJ
160 && current_function_has_exception_handlers ())
161 return false;
163 /* Any function that calls setjmp might have longjmp called from
164 any called function. ??? We really should represent this
165 properly in the CFG so that this needn't be special cased. */
166 if (cfun->calls_setjmp)
167 return false;
169 /* ??? It is OK if the argument of a function is taken in some cases,
170 but not in all cases. See PR15387 and PR19616. Revisit for 4.1. */
171 for (param = DECL_ARGUMENTS (current_function_decl);
172 param;
173 param = DECL_CHAIN (param))
174 if (TREE_ADDRESSABLE (param))
175 return false;
177 return true;
180 /* Checks whether the expression EXPR in stmt AT is independent of the
181 statement pointed to by GSI (in a sense that we already know EXPR's value
182 at GSI). We use the fact that we are only called from the chain of
183 basic blocks that have only single successor. Returns the expression
184 containing the value of EXPR at GSI. */
186 static tree
187 independent_of_stmt_p (tree expr, gimple *at, gimple_stmt_iterator gsi)
189 basic_block bb, call_bb, at_bb;
190 edge e;
191 edge_iterator ei;
193 if (is_gimple_min_invariant (expr))
194 return expr;
196 if (TREE_CODE (expr) != SSA_NAME)
197 return NULL_TREE;
199 /* Mark the blocks in the chain leading to the end. */
200 at_bb = gimple_bb (at);
201 call_bb = gimple_bb (gsi_stmt (gsi));
202 for (bb = call_bb; bb != at_bb; bb = single_succ (bb))
203 bb->aux = &bb->aux;
204 bb->aux = &bb->aux;
206 while (1)
208 at = SSA_NAME_DEF_STMT (expr);
209 bb = gimple_bb (at);
211 /* The default definition or defined before the chain. */
212 if (!bb || !bb->aux)
213 break;
215 if (bb == call_bb)
217 for (; !gsi_end_p (gsi); gsi_next (&gsi))
218 if (gsi_stmt (gsi) == at)
219 break;
221 if (!gsi_end_p (gsi))
222 expr = NULL_TREE;
223 break;
226 if (gimple_code (at) != GIMPLE_PHI)
228 expr = NULL_TREE;
229 break;
232 FOR_EACH_EDGE (e, ei, bb->preds)
233 if (e->src->aux)
234 break;
235 gcc_assert (e);
237 expr = PHI_ARG_DEF_FROM_EDGE (at, e);
238 if (TREE_CODE (expr) != SSA_NAME)
240 /* The value is a constant. */
241 break;
245 /* Unmark the blocks. */
246 for (bb = call_bb; bb != at_bb; bb = single_succ (bb))
247 bb->aux = NULL;
248 bb->aux = NULL;
250 return expr;
253 /* Simulates the effect of an assignment STMT on the return value of the tail
254 recursive CALL passed in ASS_VAR. M and A are the multiplicative and the
255 additive factor for the real return value. */
257 static bool
258 process_assignment (gassign *stmt, gimple_stmt_iterator call, tree *m,
259 tree *a, tree *ass_var)
261 tree op0, op1 = NULL_TREE, non_ass_var = NULL_TREE;
262 tree dest = gimple_assign_lhs (stmt);
263 enum tree_code code = gimple_assign_rhs_code (stmt);
264 enum gimple_rhs_class rhs_class = get_gimple_rhs_class (code);
265 tree src_var = gimple_assign_rhs1 (stmt);
267 /* See if this is a simple copy operation of an SSA name to the function
268 result. In that case we may have a simple tail call. Ignore type
269 conversions that can never produce extra code between the function
270 call and the function return. */
271 if ((rhs_class == GIMPLE_SINGLE_RHS || gimple_assign_cast_p (stmt))
272 && (TREE_CODE (src_var) == SSA_NAME))
274 /* Reject a tailcall if the type conversion might need
275 additional code. */
276 if (gimple_assign_cast_p (stmt))
278 if (TYPE_MODE (TREE_TYPE (dest)) != TYPE_MODE (TREE_TYPE (src_var)))
279 return false;
281 /* Even if the type modes are the same, if the precision of the
282 type is smaller than mode's precision,
283 reduce_to_bit_field_precision would generate additional code. */
284 if (INTEGRAL_TYPE_P (TREE_TYPE (dest))
285 && (GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (dest)))
286 > TYPE_PRECISION (TREE_TYPE (dest))))
287 return false;
290 if (src_var != *ass_var)
291 return false;
293 *ass_var = dest;
294 return true;
297 switch (rhs_class)
299 case GIMPLE_BINARY_RHS:
300 op1 = gimple_assign_rhs2 (stmt);
302 /* Fall through. */
304 case GIMPLE_UNARY_RHS:
305 op0 = gimple_assign_rhs1 (stmt);
306 break;
308 default:
309 return false;
312 /* Accumulator optimizations will reverse the order of operations.
313 We can only do that for floating-point types if we're assuming
314 that addition and multiplication are associative. */
315 if (!flag_associative_math)
316 if (FLOAT_TYPE_P (TREE_TYPE (DECL_RESULT (current_function_decl))))
317 return false;
319 if (rhs_class == GIMPLE_UNARY_RHS)
321 else if (op0 == *ass_var
322 && (non_ass_var = independent_of_stmt_p (op1, stmt, call)))
324 else if (op1 == *ass_var
325 && (non_ass_var = independent_of_stmt_p (op0, stmt, call)))
327 else
328 return false;
330 switch (code)
332 case PLUS_EXPR:
333 *a = non_ass_var;
334 *ass_var = dest;
335 return true;
337 case POINTER_PLUS_EXPR:
338 if (op0 != *ass_var)
339 return false;
340 *a = non_ass_var;
341 *ass_var = dest;
342 return true;
344 case MULT_EXPR:
345 *m = non_ass_var;
346 *ass_var = dest;
347 return true;
349 case NEGATE_EXPR:
350 *m = build_minus_one_cst (TREE_TYPE (op0));
351 *ass_var = dest;
352 return true;
354 case MINUS_EXPR:
355 if (*ass_var == op0)
356 *a = fold_build1 (NEGATE_EXPR, TREE_TYPE (non_ass_var), non_ass_var);
357 else
359 *m = build_minus_one_cst (TREE_TYPE (non_ass_var));
360 *a = fold_build1 (NEGATE_EXPR, TREE_TYPE (non_ass_var), non_ass_var);
363 *ass_var = dest;
364 return true;
366 /* TODO -- Handle POINTER_PLUS_EXPR. */
368 default:
369 return false;
373 /* Propagate VAR through phis on edge E. */
375 static tree
376 propagate_through_phis (tree var, edge e)
378 basic_block dest = e->dest;
379 gphi_iterator gsi;
381 for (gsi = gsi_start_phis (dest); !gsi_end_p (gsi); gsi_next (&gsi))
383 gphi *phi = gsi.phi ();
384 if (PHI_ARG_DEF_FROM_EDGE (phi, e) == var)
385 return PHI_RESULT (phi);
387 return var;
390 /* Finds tailcalls falling into basic block BB. The list of found tailcalls is
391 added to the start of RET. */
393 static void
394 find_tail_calls (basic_block bb, struct tailcall **ret)
396 tree ass_var = NULL_TREE, ret_var, func, param;
397 gimple *stmt;
398 gcall *call = NULL;
399 gimple_stmt_iterator gsi, agsi;
400 bool tail_recursion;
401 struct tailcall *nw;
402 edge e;
403 tree m, a;
404 basic_block abb;
405 size_t idx;
406 tree var;
408 if (!single_succ_p (bb))
409 return;
411 for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi); gsi_prev (&gsi))
413 stmt = gsi_stmt (gsi);
415 /* Ignore labels, returns, clobbers and debug stmts. */
416 if (gimple_code (stmt) == GIMPLE_LABEL
417 || gimple_code (stmt) == GIMPLE_RETURN
418 || gimple_clobber_p (stmt)
419 || is_gimple_debug (stmt))
420 continue;
422 /* Check for a call. */
423 if (is_gimple_call (stmt))
425 call = as_a <gcall *> (stmt);
426 ass_var = gimple_call_lhs (call);
427 break;
430 /* If the statement references memory or volatile operands, fail. */
431 if (gimple_references_memory_p (stmt)
432 || gimple_has_volatile_ops (stmt))
433 return;
436 if (gsi_end_p (gsi))
438 edge_iterator ei;
439 /* Recurse to the predecessors. */
440 FOR_EACH_EDGE (e, ei, bb->preds)
441 find_tail_calls (e->src, ret);
443 return;
446 /* If the LHS of our call is not just a simple register, we can't
447 transform this into a tail or sibling call. This situation happens,
448 in (e.g.) "*p = foo()" where foo returns a struct. In this case
449 we won't have a temporary here, but we need to carry out the side
450 effect anyway, so tailcall is impossible.
452 ??? In some situations (when the struct is returned in memory via
453 invisible argument) we could deal with this, e.g. by passing 'p'
454 itself as that argument to foo, but it's too early to do this here,
455 and expand_call() will not handle it anyway. If it ever can, then
456 we need to revisit this here, to allow that situation. */
457 if (ass_var && !is_gimple_reg (ass_var))
458 return;
460 /* We found the call, check whether it is suitable. */
461 tail_recursion = false;
462 func = gimple_call_fndecl (call);
463 if (func
464 && !DECL_BUILT_IN (func)
465 && recursive_call_p (current_function_decl, func))
467 tree arg;
469 for (param = DECL_ARGUMENTS (func), idx = 0;
470 param && idx < gimple_call_num_args (call);
471 param = DECL_CHAIN (param), idx ++)
473 arg = gimple_call_arg (call, idx);
474 if (param != arg)
476 /* Make sure there are no problems with copying. The parameter
477 have a copyable type and the two arguments must have reasonably
478 equivalent types. The latter requirement could be relaxed if
479 we emitted a suitable type conversion statement. */
480 if (!is_gimple_reg_type (TREE_TYPE (param))
481 || !useless_type_conversion_p (TREE_TYPE (param),
482 TREE_TYPE (arg)))
483 break;
485 /* The parameter should be a real operand, so that phi node
486 created for it at the start of the function has the meaning
487 of copying the value. This test implies is_gimple_reg_type
488 from the previous condition, however this one could be
489 relaxed by being more careful with copying the new value
490 of the parameter (emitting appropriate GIMPLE_ASSIGN and
491 updating the virtual operands). */
492 if (!is_gimple_reg (param))
493 break;
496 if (idx == gimple_call_num_args (call) && !param)
497 tail_recursion = true;
500 /* Make sure the tail invocation of this function does not refer
501 to local variables. */
502 FOR_EACH_LOCAL_DECL (cfun, idx, var)
504 if (TREE_CODE (var) != PARM_DECL
505 && auto_var_in_fn_p (var, cfun->decl)
506 && (ref_maybe_used_by_stmt_p (call, var)
507 || call_may_clobber_ref_p (call, var)))
508 return;
511 /* Now check the statements after the call. None of them has virtual
512 operands, so they may only depend on the call through its return
513 value. The return value should also be dependent on each of them,
514 since we are running after dce. */
515 m = NULL_TREE;
516 a = NULL_TREE;
518 abb = bb;
519 agsi = gsi;
520 while (1)
522 tree tmp_a = NULL_TREE;
523 tree tmp_m = NULL_TREE;
524 gsi_next (&agsi);
526 while (gsi_end_p (agsi))
528 ass_var = propagate_through_phis (ass_var, single_succ_edge (abb));
529 abb = single_succ (abb);
530 agsi = gsi_start_bb (abb);
533 stmt = gsi_stmt (agsi);
535 if (gimple_code (stmt) == GIMPLE_LABEL)
536 continue;
538 if (gimple_code (stmt) == GIMPLE_RETURN)
539 break;
541 if (gimple_clobber_p (stmt))
542 continue;
544 if (is_gimple_debug (stmt))
545 continue;
547 if (gimple_code (stmt) != GIMPLE_ASSIGN)
548 return;
550 /* This is a gimple assign. */
551 if (! process_assignment (as_a <gassign *> (stmt), gsi, &tmp_m,
552 &tmp_a, &ass_var))
553 return;
555 if (tmp_a)
557 tree type = TREE_TYPE (tmp_a);
558 if (a)
559 a = fold_build2 (PLUS_EXPR, type, fold_convert (type, a), tmp_a);
560 else
561 a = tmp_a;
563 if (tmp_m)
565 tree type = TREE_TYPE (tmp_m);
566 if (m)
567 m = fold_build2 (MULT_EXPR, type, fold_convert (type, m), tmp_m);
568 else
569 m = tmp_m;
571 if (a)
572 a = fold_build2 (MULT_EXPR, type, fold_convert (type, a), tmp_m);
576 /* See if this is a tail call we can handle. */
577 ret_var = gimple_return_retval (as_a <greturn *> (stmt));
579 /* We may proceed if there either is no return value, or the return value
580 is identical to the call's return. */
581 if (ret_var
582 && (ret_var != ass_var))
583 return;
585 /* If this is not a tail recursive call, we cannot handle addends or
586 multiplicands. */
587 if (!tail_recursion && (m || a))
588 return;
590 /* For pointers only allow additions. */
591 if (m && POINTER_TYPE_P (TREE_TYPE (DECL_RESULT (current_function_decl))))
592 return;
594 nw = XNEW (struct tailcall);
596 nw->call_gsi = gsi;
598 nw->tail_recursion = tail_recursion;
600 nw->mult = m;
601 nw->add = a;
603 nw->next = *ret;
604 *ret = nw;
607 /* Helper to insert PHI_ARGH to the phi of VAR in the destination of edge E. */
609 static void
610 add_successor_phi_arg (edge e, tree var, tree phi_arg)
612 gphi_iterator gsi;
614 for (gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi); gsi_next (&gsi))
615 if (PHI_RESULT (gsi.phi ()) == var)
616 break;
618 gcc_assert (!gsi_end_p (gsi));
619 add_phi_arg (gsi.phi (), phi_arg, e, UNKNOWN_LOCATION);
622 /* Creates a GIMPLE statement which computes the operation specified by
623 CODE, ACC and OP1 to a new variable with name LABEL and inserts the
624 statement in the position specified by GSI. Returns the
625 tree node of the statement's result. */
627 static tree
628 adjust_return_value_with_ops (enum tree_code code, const char *label,
629 tree acc, tree op1, gimple_stmt_iterator gsi)
632 tree ret_type = TREE_TYPE (DECL_RESULT (current_function_decl));
633 tree result = make_temp_ssa_name (ret_type, NULL, label);
634 gassign *stmt;
636 if (POINTER_TYPE_P (ret_type))
638 gcc_assert (code == PLUS_EXPR && TREE_TYPE (acc) == sizetype);
639 code = POINTER_PLUS_EXPR;
641 if (types_compatible_p (TREE_TYPE (acc), TREE_TYPE (op1))
642 && code != POINTER_PLUS_EXPR)
643 stmt = gimple_build_assign (result, code, acc, op1);
644 else
646 tree tem;
647 if (code == POINTER_PLUS_EXPR)
648 tem = fold_build2 (code, TREE_TYPE (op1), op1, acc);
649 else
650 tem = fold_build2 (code, TREE_TYPE (op1),
651 fold_convert (TREE_TYPE (op1), acc), op1);
652 tree rhs = fold_convert (ret_type, tem);
653 rhs = force_gimple_operand_gsi (&gsi, rhs,
654 false, NULL, true, GSI_SAME_STMT);
655 stmt = gimple_build_assign (result, rhs);
658 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
659 return result;
662 /* Creates a new GIMPLE statement that adjusts the value of accumulator ACC by
663 the computation specified by CODE and OP1 and insert the statement
664 at the position specified by GSI as a new statement. Returns new SSA name
665 of updated accumulator. */
667 static tree
668 update_accumulator_with_ops (enum tree_code code, tree acc, tree op1,
669 gimple_stmt_iterator gsi)
671 gassign *stmt;
672 tree var = copy_ssa_name (acc);
673 if (types_compatible_p (TREE_TYPE (acc), TREE_TYPE (op1)))
674 stmt = gimple_build_assign (var, code, acc, op1);
675 else
677 tree rhs = fold_convert (TREE_TYPE (acc),
678 fold_build2 (code,
679 TREE_TYPE (op1),
680 fold_convert (TREE_TYPE (op1), acc),
681 op1));
682 rhs = force_gimple_operand_gsi (&gsi, rhs,
683 false, NULL, false, GSI_CONTINUE_LINKING);
684 stmt = gimple_build_assign (var, rhs);
686 gsi_insert_after (&gsi, stmt, GSI_NEW_STMT);
687 return var;
690 /* Adjust the accumulator values according to A and M after GSI, and update
691 the phi nodes on edge BACK. */
693 static void
694 adjust_accumulator_values (gimple_stmt_iterator gsi, tree m, tree a, edge back)
696 tree var, a_acc_arg, m_acc_arg;
698 if (m)
699 m = force_gimple_operand_gsi (&gsi, m, true, NULL, true, GSI_SAME_STMT);
700 if (a)
701 a = force_gimple_operand_gsi (&gsi, a, true, NULL, true, GSI_SAME_STMT);
703 a_acc_arg = a_acc;
704 m_acc_arg = m_acc;
705 if (a)
707 if (m_acc)
709 if (integer_onep (a))
710 var = m_acc;
711 else
712 var = adjust_return_value_with_ops (MULT_EXPR, "acc_tmp", m_acc,
713 a, gsi);
715 else
716 var = a;
718 a_acc_arg = update_accumulator_with_ops (PLUS_EXPR, a_acc, var, gsi);
721 if (m)
722 m_acc_arg = update_accumulator_with_ops (MULT_EXPR, m_acc, m, gsi);
724 if (a_acc)
725 add_successor_phi_arg (back, a_acc, a_acc_arg);
727 if (m_acc)
728 add_successor_phi_arg (back, m_acc, m_acc_arg);
731 /* Adjust value of the return at the end of BB according to M and A
732 accumulators. */
734 static void
735 adjust_return_value (basic_block bb, tree m, tree a)
737 tree retval;
738 greturn *ret_stmt = as_a <greturn *> (gimple_seq_last_stmt (bb_seq (bb)));
739 gimple_stmt_iterator gsi = gsi_last_bb (bb);
741 gcc_assert (gimple_code (ret_stmt) == GIMPLE_RETURN);
743 retval = gimple_return_retval (ret_stmt);
744 if (!retval || retval == error_mark_node)
745 return;
747 if (m)
748 retval = adjust_return_value_with_ops (MULT_EXPR, "mul_tmp", m_acc, retval,
749 gsi);
750 if (a)
751 retval = adjust_return_value_with_ops (PLUS_EXPR, "acc_tmp", a_acc, retval,
752 gsi);
753 gimple_return_set_retval (ret_stmt, retval);
754 update_stmt (ret_stmt);
757 /* Subtract COUNT and FREQUENCY from the basic block and it's
758 outgoing edge. */
759 static void
760 decrease_profile (basic_block bb, gcov_type count, int frequency)
762 edge e;
763 bb->count -= count;
764 if (bb->count < 0)
765 bb->count = 0;
766 bb->frequency -= frequency;
767 if (bb->frequency < 0)
768 bb->frequency = 0;
769 if (!single_succ_p (bb))
771 gcc_assert (!EDGE_COUNT (bb->succs));
772 return;
774 e = single_succ_edge (bb);
775 e->count -= count;
776 if (e->count < 0)
777 e->count = 0;
780 /* Returns true if argument PARAM of the tail recursive call needs to be copied
781 when the call is eliminated. */
783 static bool
784 arg_needs_copy_p (tree param)
786 tree def;
788 if (!is_gimple_reg (param))
789 return false;
791 /* Parameters that are only defined but never used need not be copied. */
792 def = ssa_default_def (cfun, param);
793 if (!def)
794 return false;
796 return true;
799 /* Eliminates tail call described by T. TMP_VARS is a list of
800 temporary variables used to copy the function arguments. */
802 static void
803 eliminate_tail_call (struct tailcall *t)
805 tree param, rslt;
806 gimple *stmt, *call;
807 tree arg;
808 size_t idx;
809 basic_block bb, first;
810 edge e;
811 gphi *phi;
812 gphi_iterator gpi;
813 gimple_stmt_iterator gsi;
814 gimple *orig_stmt;
816 stmt = orig_stmt = gsi_stmt (t->call_gsi);
817 bb = gsi_bb (t->call_gsi);
819 if (dump_file && (dump_flags & TDF_DETAILS))
821 fprintf (dump_file, "Eliminated tail recursion in bb %d : ",
822 bb->index);
823 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
824 fprintf (dump_file, "\n");
827 gcc_assert (is_gimple_call (stmt));
829 first = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun));
831 /* Remove the code after call_gsi that will become unreachable. The
832 possibly unreachable code in other blocks is removed later in
833 cfg cleanup. */
834 gsi = t->call_gsi;
835 gimple_stmt_iterator gsi2 = gsi_last_bb (gimple_bb (gsi_stmt (gsi)));
836 while (gsi_stmt (gsi2) != gsi_stmt (gsi))
838 gimple *t = gsi_stmt (gsi2);
839 /* Do not remove the return statement, so that redirect_edge_and_branch
840 sees how the block ends. */
841 if (gimple_code (t) != GIMPLE_RETURN)
843 gimple_stmt_iterator gsi3 = gsi2;
844 gsi_prev (&gsi2);
845 gsi_remove (&gsi3, true);
846 release_defs (t);
848 else
849 gsi_prev (&gsi2);
852 /* Number of executions of function has reduced by the tailcall. */
853 e = single_succ_edge (gsi_bb (t->call_gsi));
854 decrease_profile (EXIT_BLOCK_PTR_FOR_FN (cfun), e->count, EDGE_FREQUENCY (e));
855 decrease_profile (ENTRY_BLOCK_PTR_FOR_FN (cfun), e->count,
856 EDGE_FREQUENCY (e));
857 if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun))
858 decrease_profile (e->dest, e->count, EDGE_FREQUENCY (e));
860 /* Replace the call by a jump to the start of function. */
861 e = redirect_edge_and_branch (single_succ_edge (gsi_bb (t->call_gsi)),
862 first);
863 gcc_assert (e);
864 PENDING_STMT (e) = NULL;
866 /* Add phi node entries for arguments. The ordering of the phi nodes should
867 be the same as the ordering of the arguments. */
868 for (param = DECL_ARGUMENTS (current_function_decl),
869 idx = 0, gpi = gsi_start_phis (first);
870 param;
871 param = DECL_CHAIN (param), idx++)
873 if (!arg_needs_copy_p (param))
874 continue;
876 arg = gimple_call_arg (stmt, idx);
877 phi = gpi.phi ();
878 gcc_assert (param == SSA_NAME_VAR (PHI_RESULT (phi)));
880 add_phi_arg (phi, arg, e, gimple_location (stmt));
881 gsi_next (&gpi);
884 /* Update the values of accumulators. */
885 adjust_accumulator_values (t->call_gsi, t->mult, t->add, e);
887 call = gsi_stmt (t->call_gsi);
888 rslt = gimple_call_lhs (call);
889 if (rslt != NULL_TREE)
891 /* Result of the call will no longer be defined. So adjust the
892 SSA_NAME_DEF_STMT accordingly. */
893 SSA_NAME_DEF_STMT (rslt) = gimple_build_nop ();
896 gsi_remove (&t->call_gsi, true);
897 release_defs (call);
900 /* Optimizes the tailcall described by T. If OPT_TAILCALLS is true, also
901 mark the tailcalls for the sibcall optimization. */
903 static bool
904 optimize_tail_call (struct tailcall *t, bool opt_tailcalls)
906 if (t->tail_recursion)
908 eliminate_tail_call (t);
909 return true;
912 if (opt_tailcalls)
914 gcall *stmt = as_a <gcall *> (gsi_stmt (t->call_gsi));
916 gimple_call_set_tail (stmt, true);
917 cfun->tail_call_marked = true;
918 if (dump_file && (dump_flags & TDF_DETAILS))
920 fprintf (dump_file, "Found tail call ");
921 print_gimple_stmt (dump_file, stmt, 0, dump_flags);
922 fprintf (dump_file, " in bb %i\n", (gsi_bb (t->call_gsi))->index);
926 return false;
929 /* Creates a tail-call accumulator of the same type as the return type of the
930 current function. LABEL is the name used to creating the temporary
931 variable for the accumulator. The accumulator will be inserted in the
932 phis of a basic block BB with single predecessor with an initial value
933 INIT converted to the current function return type. */
935 static tree
936 create_tailcall_accumulator (const char *label, basic_block bb, tree init)
938 tree ret_type = TREE_TYPE (DECL_RESULT (current_function_decl));
939 if (POINTER_TYPE_P (ret_type))
940 ret_type = sizetype;
942 tree tmp = make_temp_ssa_name (ret_type, NULL, label);
943 gphi *phi;
945 phi = create_phi_node (tmp, bb);
946 /* RET_TYPE can be a float when -ffast-maths is enabled. */
947 add_phi_arg (phi, fold_convert (ret_type, init), single_pred_edge (bb),
948 UNKNOWN_LOCATION);
949 return PHI_RESULT (phi);
952 /* Optimizes tail calls in the function, turning the tail recursion
953 into iteration. */
955 static unsigned int
956 tree_optimize_tail_calls_1 (bool opt_tailcalls)
958 edge e;
959 bool phis_constructed = false;
960 struct tailcall *tailcalls = NULL, *act, *next;
961 bool changed = false;
962 basic_block first = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun));
963 tree param;
964 gimple *stmt;
965 edge_iterator ei;
967 if (!suitable_for_tail_opt_p ())
968 return 0;
969 if (opt_tailcalls)
970 opt_tailcalls = suitable_for_tail_call_opt_p ();
972 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds)
974 /* Only traverse the normal exits, i.e. those that end with return
975 statement. */
976 stmt = last_stmt (e->src);
978 if (stmt
979 && gimple_code (stmt) == GIMPLE_RETURN)
980 find_tail_calls (e->src, &tailcalls);
983 /* Construct the phi nodes and accumulators if necessary. */
984 a_acc = m_acc = NULL_TREE;
985 for (act = tailcalls; act; act = act->next)
987 if (!act->tail_recursion)
988 continue;
990 if (!phis_constructed)
992 /* Ensure that there is only one predecessor of the block
993 or if there are existing degenerate PHI nodes. */
994 if (!single_pred_p (first)
995 || !gimple_seq_empty_p (phi_nodes (first)))
996 first =
997 split_edge (single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun)));
999 /* Copy the args if needed. */
1000 for (param = DECL_ARGUMENTS (current_function_decl);
1001 param;
1002 param = DECL_CHAIN (param))
1003 if (arg_needs_copy_p (param))
1005 tree name = ssa_default_def (cfun, param);
1006 tree new_name = make_ssa_name (param, SSA_NAME_DEF_STMT (name));
1007 gphi *phi;
1009 set_ssa_default_def (cfun, param, new_name);
1010 phi = create_phi_node (name, first);
1011 add_phi_arg (phi, new_name, single_pred_edge (first),
1012 EXPR_LOCATION (param));
1014 phis_constructed = true;
1017 if (act->add && !a_acc)
1018 a_acc = create_tailcall_accumulator ("add_acc", first,
1019 integer_zero_node);
1021 if (act->mult && !m_acc)
1022 m_acc = create_tailcall_accumulator ("mult_acc", first,
1023 integer_one_node);
1026 if (a_acc || m_acc)
1028 /* When the tail call elimination using accumulators is performed,
1029 statements adding the accumulated value are inserted at all exits.
1030 This turns all other tail calls to non-tail ones. */
1031 opt_tailcalls = false;
1034 for (; tailcalls; tailcalls = next)
1036 next = tailcalls->next;
1037 changed |= optimize_tail_call (tailcalls, opt_tailcalls);
1038 free (tailcalls);
1041 if (a_acc || m_acc)
1043 /* Modify the remaining return statements. */
1044 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds)
1046 stmt = last_stmt (e->src);
1048 if (stmt
1049 && gimple_code (stmt) == GIMPLE_RETURN)
1050 adjust_return_value (e->src, m_acc, a_acc);
1054 if (changed)
1056 /* We may have created new loops. Make them magically appear. */
1057 loops_state_set (LOOPS_NEED_FIXUP);
1058 free_dominance_info (CDI_DOMINATORS);
1061 /* Add phi nodes for the virtual operands defined in the function to the
1062 header of the loop created by tail recursion elimination. Do so
1063 by triggering the SSA renamer. */
1064 if (phis_constructed)
1065 mark_virtual_operands_for_renaming (cfun);
1067 if (changed)
1068 return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals;
1069 return 0;
1072 static bool
1073 gate_tail_calls (void)
1075 return flag_optimize_sibling_calls != 0 && dbg_cnt (tail_call);
1078 static unsigned int
1079 execute_tail_calls (void)
1081 return tree_optimize_tail_calls_1 (true);
1084 namespace {
1086 const pass_data pass_data_tail_recursion =
1088 GIMPLE_PASS, /* type */
1089 "tailr", /* name */
1090 OPTGROUP_NONE, /* optinfo_flags */
1091 TV_NONE, /* tv_id */
1092 ( PROP_cfg | PROP_ssa ), /* properties_required */
1093 0, /* properties_provided */
1094 0, /* properties_destroyed */
1095 0, /* todo_flags_start */
1096 0, /* todo_flags_finish */
1099 class pass_tail_recursion : public gimple_opt_pass
1101 public:
1102 pass_tail_recursion (gcc::context *ctxt)
1103 : gimple_opt_pass (pass_data_tail_recursion, ctxt)
1106 /* opt_pass methods: */
1107 opt_pass * clone () { return new pass_tail_recursion (m_ctxt); }
1108 virtual bool gate (function *) { return gate_tail_calls (); }
1109 virtual unsigned int execute (function *)
1111 return tree_optimize_tail_calls_1 (false);
1114 }; // class pass_tail_recursion
1116 } // anon namespace
1118 gimple_opt_pass *
1119 make_pass_tail_recursion (gcc::context *ctxt)
1121 return new pass_tail_recursion (ctxt);
1124 namespace {
1126 const pass_data pass_data_tail_calls =
1128 GIMPLE_PASS, /* type */
1129 "tailc", /* name */
1130 OPTGROUP_NONE, /* optinfo_flags */
1131 TV_NONE, /* tv_id */
1132 ( PROP_cfg | PROP_ssa ), /* properties_required */
1133 0, /* properties_provided */
1134 0, /* properties_destroyed */
1135 0, /* todo_flags_start */
1136 0, /* todo_flags_finish */
1139 class pass_tail_calls : public gimple_opt_pass
1141 public:
1142 pass_tail_calls (gcc::context *ctxt)
1143 : gimple_opt_pass (pass_data_tail_calls, ctxt)
1146 /* opt_pass methods: */
1147 virtual bool gate (function *) { return gate_tail_calls (); }
1148 virtual unsigned int execute (function *) { return execute_tail_calls (); }
1150 }; // class pass_tail_calls
1152 } // anon namespace
1154 gimple_opt_pass *
1155 make_pass_tail_calls (gcc::context *ctxt)
1157 return new pass_tail_calls (ctxt);