PR tree-ssa/57385
[official-gcc.git] / gcc / tree-tailcall.c
blobc6581dcedcdb3619970b1a8a1ce56da6863f648f
1 /* Tail call optimization on trees.
2 Copyright (C) 2003-2013 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 "tm_p.h"
26 #include "basic-block.h"
27 #include "function.h"
28 #include "tree-flow.h"
29 #include "gimple-pretty-print.h"
30 #include "except.h"
31 #include "tree-pass.h"
32 #include "flags.h"
33 #include "langhooks.h"
34 #include "dbgcnt.h"
35 #include "target.h"
36 #include "cfgloop.h"
37 #include "common/common-target.h"
39 /* The file implements the tail recursion elimination. It is also used to
40 analyze the tail calls in general, passing the results to the rtl level
41 where they are used for sibcall optimization.
43 In addition to the standard tail recursion elimination, we handle the most
44 trivial cases of making the call tail recursive by creating accumulators.
45 For example the following function
47 int sum (int n)
49 if (n > 0)
50 return n + sum (n - 1);
51 else
52 return 0;
55 is transformed into
57 int sum (int n)
59 int acc = 0;
61 while (n > 0)
62 acc += n--;
64 return acc;
67 To do this, we maintain two accumulators (a_acc and m_acc) that indicate
68 when we reach the return x statement, we should return a_acc + x * m_acc
69 instead. They are initially initialized to 0 and 1, respectively,
70 so the semantics of the function is obviously preserved. If we are
71 guaranteed that the value of the accumulator never change, we
72 omit the accumulator.
74 There are three cases how the function may exit. The first one is
75 handled in adjust_return_value, the other two in adjust_accumulator_values
76 (the second case is actually a special case of the third one and we
77 present it separately just for clarity):
79 1) Just return x, where x is not in any of the remaining special shapes.
80 We rewrite this to a gimple equivalent of return m_acc * x + a_acc.
82 2) return f (...), where f is the current function, is rewritten in a
83 classical tail-recursion elimination way, into assignment of arguments
84 and jump to the start of the function. Values of the accumulators
85 are unchanged.
87 3) return a + m * f(...), where a and m do not depend on call to f.
88 To preserve the semantics described before we want this to be rewritten
89 in such a way that we finally return
91 a_acc + (a + m * f(...)) * m_acc = (a_acc + a * m_acc) + (m * m_acc) * f(...).
93 I.e. we increase a_acc by a * m_acc, multiply m_acc by m and
94 eliminate the tail call to f. Special cases when the value is just
95 added or just multiplied are obtained by setting a = 0 or m = 1.
97 TODO -- it is possible to do similar tricks for other operations. */
99 /* A structure that describes the tailcall. */
101 struct tailcall
103 /* The iterator pointing to the call statement. */
104 gimple_stmt_iterator call_gsi;
106 /* True if it is a call to the current function. */
107 bool tail_recursion;
109 /* The return value of the caller is mult * f + add, where f is the return
110 value of the call. */
111 tree mult, add;
113 /* Next tailcall in the chain. */
114 struct tailcall *next;
117 /* The variables holding the value of multiplicative and additive
118 accumulator. */
119 static tree m_acc, a_acc;
121 static bool suitable_for_tail_opt_p (void);
122 static bool optimize_tail_call (struct tailcall *, bool);
123 static void eliminate_tail_call (struct tailcall *);
124 static void find_tail_calls (basic_block, struct tailcall **);
126 /* Returns false when the function is not suitable for tail call optimization
127 from some reason (e.g. if it takes variable number of arguments). */
129 static bool
130 suitable_for_tail_opt_p (void)
132 if (cfun->stdarg)
133 return false;
135 return true;
137 /* Returns false when the function is not suitable for tail call optimization
138 from some reason (e.g. if it takes variable number of arguments).
139 This test must pass in addition to suitable_for_tail_opt_p in order to make
140 tail call discovery happen. */
142 static bool
143 suitable_for_tail_call_opt_p (void)
145 tree param;
147 /* alloca (until we have stack slot life analysis) inhibits
148 sibling call optimizations, but not tail recursion. */
149 if (cfun->calls_alloca)
150 return false;
152 /* If we are using sjlj exceptions, we may need to add a call to
153 _Unwind_SjLj_Unregister at exit of the function. Which means
154 that we cannot do any sibcall transformations. */
155 if (targetm_common.except_unwind_info (&global_options) == UI_SJLJ
156 && current_function_has_exception_handlers ())
157 return false;
159 /* Any function that calls setjmp might have longjmp called from
160 any called function. ??? We really should represent this
161 properly in the CFG so that this needn't be special cased. */
162 if (cfun->calls_setjmp)
163 return false;
165 /* ??? It is OK if the argument of a function is taken in some cases,
166 but not in all cases. See PR15387 and PR19616. Revisit for 4.1. */
167 for (param = DECL_ARGUMENTS (current_function_decl);
168 param;
169 param = DECL_CHAIN (param))
170 if (TREE_ADDRESSABLE (param))
171 return false;
173 return true;
176 /* Checks whether the expression EXPR in stmt AT is independent of the
177 statement pointed to by GSI (in a sense that we already know EXPR's value
178 at GSI). We use the fact that we are only called from the chain of
179 basic blocks that have only single successor. Returns the expression
180 containing the value of EXPR at GSI. */
182 static tree
183 independent_of_stmt_p (tree expr, gimple at, gimple_stmt_iterator gsi)
185 basic_block bb, call_bb, at_bb;
186 edge e;
187 edge_iterator ei;
189 if (is_gimple_min_invariant (expr))
190 return expr;
192 if (TREE_CODE (expr) != SSA_NAME)
193 return NULL_TREE;
195 /* Mark the blocks in the chain leading to the end. */
196 at_bb = gimple_bb (at);
197 call_bb = gimple_bb (gsi_stmt (gsi));
198 for (bb = call_bb; bb != at_bb; bb = single_succ (bb))
199 bb->aux = &bb->aux;
200 bb->aux = &bb->aux;
202 while (1)
204 at = SSA_NAME_DEF_STMT (expr);
205 bb = gimple_bb (at);
207 /* The default definition or defined before the chain. */
208 if (!bb || !bb->aux)
209 break;
211 if (bb == call_bb)
213 for (; !gsi_end_p (gsi); gsi_next (&gsi))
214 if (gsi_stmt (gsi) == at)
215 break;
217 if (!gsi_end_p (gsi))
218 expr = NULL_TREE;
219 break;
222 if (gimple_code (at) != GIMPLE_PHI)
224 expr = NULL_TREE;
225 break;
228 FOR_EACH_EDGE (e, ei, bb->preds)
229 if (e->src->aux)
230 break;
231 gcc_assert (e);
233 expr = PHI_ARG_DEF_FROM_EDGE (at, e);
234 if (TREE_CODE (expr) != SSA_NAME)
236 /* The value is a constant. */
237 break;
241 /* Unmark the blocks. */
242 for (bb = call_bb; bb != at_bb; bb = single_succ (bb))
243 bb->aux = NULL;
244 bb->aux = NULL;
246 return expr;
249 /* Simulates the effect of an assignment STMT on the return value of the tail
250 recursive CALL passed in ASS_VAR. M and A are the multiplicative and the
251 additive factor for the real return value. */
253 static bool
254 process_assignment (gimple stmt, gimple_stmt_iterator call, tree *m,
255 tree *a, tree *ass_var)
257 tree op0, op1 = NULL_TREE, non_ass_var = NULL_TREE;
258 tree dest = gimple_assign_lhs (stmt);
259 enum tree_code code = gimple_assign_rhs_code (stmt);
260 enum gimple_rhs_class rhs_class = get_gimple_rhs_class (code);
261 tree src_var = gimple_assign_rhs1 (stmt);
263 /* See if this is a simple copy operation of an SSA name to the function
264 result. In that case we may have a simple tail call. Ignore type
265 conversions that can never produce extra code between the function
266 call and the function return. */
267 if ((rhs_class == GIMPLE_SINGLE_RHS || gimple_assign_cast_p (stmt))
268 && (TREE_CODE (src_var) == SSA_NAME))
270 /* Reject a tailcall if the type conversion might need
271 additional code. */
272 if (gimple_assign_cast_p (stmt)
273 && TYPE_MODE (TREE_TYPE (dest)) != TYPE_MODE (TREE_TYPE (src_var)))
274 return false;
276 if (src_var != *ass_var)
277 return false;
279 *ass_var = dest;
280 return true;
283 switch (rhs_class)
285 case GIMPLE_BINARY_RHS:
286 op1 = gimple_assign_rhs2 (stmt);
288 /* Fall through. */
290 case GIMPLE_UNARY_RHS:
291 op0 = gimple_assign_rhs1 (stmt);
292 break;
294 default:
295 return false;
298 /* Accumulator optimizations will reverse the order of operations.
299 We can only do that for floating-point types if we're assuming
300 that addition and multiplication are associative. */
301 if (!flag_associative_math)
302 if (FLOAT_TYPE_P (TREE_TYPE (DECL_RESULT (current_function_decl))))
303 return false;
305 if (rhs_class == GIMPLE_UNARY_RHS)
307 else if (op0 == *ass_var
308 && (non_ass_var = independent_of_stmt_p (op1, stmt, call)))
310 else if (op1 == *ass_var
311 && (non_ass_var = independent_of_stmt_p (op0, stmt, call)))
313 else
314 return false;
316 switch (code)
318 case PLUS_EXPR:
319 *a = non_ass_var;
320 *ass_var = dest;
321 return true;
323 case MULT_EXPR:
324 *m = non_ass_var;
325 *ass_var = dest;
326 return true;
328 case NEGATE_EXPR:
329 if (FLOAT_TYPE_P (TREE_TYPE (op0)))
330 *m = build_real (TREE_TYPE (op0), dconstm1);
331 else
332 *m = build_int_cst (TREE_TYPE (op0), -1);
334 *ass_var = dest;
335 return true;
337 case MINUS_EXPR:
338 if (*ass_var == op0)
339 *a = fold_build1 (NEGATE_EXPR, TREE_TYPE (non_ass_var), non_ass_var);
340 else
342 if (FLOAT_TYPE_P (TREE_TYPE (non_ass_var)))
343 *m = build_real (TREE_TYPE (non_ass_var), dconstm1);
344 else
345 *m = build_int_cst (TREE_TYPE (non_ass_var), -1);
347 *a = fold_build1 (NEGATE_EXPR, TREE_TYPE (non_ass_var), non_ass_var);
350 *ass_var = dest;
351 return true;
353 /* TODO -- Handle POINTER_PLUS_EXPR. */
355 default:
356 return false;
360 /* Propagate VAR through phis on edge E. */
362 static tree
363 propagate_through_phis (tree var, edge e)
365 basic_block dest = e->dest;
366 gimple_stmt_iterator gsi;
368 for (gsi = gsi_start_phis (dest); !gsi_end_p (gsi); gsi_next (&gsi))
370 gimple phi = gsi_stmt (gsi);
371 if (PHI_ARG_DEF_FROM_EDGE (phi, e) == var)
372 return PHI_RESULT (phi);
374 return var;
377 /* Finds tailcalls falling into basic block BB. The list of found tailcalls is
378 added to the start of RET. */
380 static void
381 find_tail_calls (basic_block bb, struct tailcall **ret)
383 tree ass_var = NULL_TREE, ret_var, func, param;
384 gimple stmt, call = NULL;
385 gimple_stmt_iterator gsi, agsi;
386 bool tail_recursion;
387 struct tailcall *nw;
388 edge e;
389 tree m, a;
390 basic_block abb;
391 size_t idx;
392 tree var;
394 if (!single_succ_p (bb))
395 return;
397 for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi); gsi_prev (&gsi))
399 stmt = gsi_stmt (gsi);
401 /* Ignore labels, returns, clobbers and debug stmts. */
402 if (gimple_code (stmt) == GIMPLE_LABEL
403 || gimple_code (stmt) == GIMPLE_RETURN
404 || gimple_clobber_p (stmt)
405 || is_gimple_debug (stmt))
406 continue;
408 /* Check for a call. */
409 if (is_gimple_call (stmt))
411 call = stmt;
412 ass_var = gimple_call_lhs (stmt);
413 break;
416 /* If the statement references memory or volatile operands, fail. */
417 if (gimple_references_memory_p (stmt)
418 || gimple_has_volatile_ops (stmt))
419 return;
422 if (gsi_end_p (gsi))
424 edge_iterator ei;
425 /* Recurse to the predecessors. */
426 FOR_EACH_EDGE (e, ei, bb->preds)
427 find_tail_calls (e->src, ret);
429 return;
432 /* If the LHS of our call is not just a simple register, we can't
433 transform this into a tail or sibling call. This situation happens,
434 in (e.g.) "*p = foo()" where foo returns a struct. In this case
435 we won't have a temporary here, but we need to carry out the side
436 effect anyway, so tailcall is impossible.
438 ??? In some situations (when the struct is returned in memory via
439 invisible argument) we could deal with this, e.g. by passing 'p'
440 itself as that argument to foo, but it's too early to do this here,
441 and expand_call() will not handle it anyway. If it ever can, then
442 we need to revisit this here, to allow that situation. */
443 if (ass_var && !is_gimple_reg (ass_var))
444 return;
446 /* We found the call, check whether it is suitable. */
447 tail_recursion = false;
448 func = gimple_call_fndecl (call);
449 if (func == current_function_decl)
451 tree arg;
453 for (param = DECL_ARGUMENTS (func), idx = 0;
454 param && idx < gimple_call_num_args (call);
455 param = DECL_CHAIN (param), idx ++)
457 arg = gimple_call_arg (call, idx);
458 if (param != arg)
460 /* Make sure there are no problems with copying. The parameter
461 have a copyable type and the two arguments must have reasonably
462 equivalent types. The latter requirement could be relaxed if
463 we emitted a suitable type conversion statement. */
464 if (!is_gimple_reg_type (TREE_TYPE (param))
465 || !useless_type_conversion_p (TREE_TYPE (param),
466 TREE_TYPE (arg)))
467 break;
469 /* The parameter should be a real operand, so that phi node
470 created for it at the start of the function has the meaning
471 of copying the value. This test implies is_gimple_reg_type
472 from the previous condition, however this one could be
473 relaxed by being more careful with copying the new value
474 of the parameter (emitting appropriate GIMPLE_ASSIGN and
475 updating the virtual operands). */
476 if (!is_gimple_reg (param))
477 break;
480 if (idx == gimple_call_num_args (call) && !param)
481 tail_recursion = true;
484 /* Make sure the tail invocation of this function does not refer
485 to local variables. */
486 FOR_EACH_LOCAL_DECL (cfun, idx, var)
488 if (TREE_CODE (var) != PARM_DECL
489 && auto_var_in_fn_p (var, cfun->decl)
490 && (ref_maybe_used_by_stmt_p (call, var)
491 || call_may_clobber_ref_p (call, var)))
492 return;
495 /* Now check the statements after the call. None of them has virtual
496 operands, so they may only depend on the call through its return
497 value. The return value should also be dependent on each of them,
498 since we are running after dce. */
499 m = NULL_TREE;
500 a = NULL_TREE;
502 abb = bb;
503 agsi = gsi;
504 while (1)
506 tree tmp_a = NULL_TREE;
507 tree tmp_m = NULL_TREE;
508 gsi_next (&agsi);
510 while (gsi_end_p (agsi))
512 ass_var = propagate_through_phis (ass_var, single_succ_edge (abb));
513 abb = single_succ (abb);
514 agsi = gsi_start_bb (abb);
517 stmt = gsi_stmt (agsi);
519 if (gimple_code (stmt) == GIMPLE_LABEL)
520 continue;
522 if (gimple_code (stmt) == GIMPLE_RETURN)
523 break;
525 if (gimple_clobber_p (stmt))
526 continue;
528 if (is_gimple_debug (stmt))
529 continue;
531 if (gimple_code (stmt) != GIMPLE_ASSIGN)
532 return;
534 /* This is a gimple assign. */
535 if (! process_assignment (stmt, gsi, &tmp_m, &tmp_a, &ass_var))
536 return;
538 if (tmp_a)
540 tree type = TREE_TYPE (tmp_a);
541 if (a)
542 a = fold_build2 (PLUS_EXPR, type, fold_convert (type, a), tmp_a);
543 else
544 a = tmp_a;
546 if (tmp_m)
548 tree type = TREE_TYPE (tmp_m);
549 if (m)
550 m = fold_build2 (MULT_EXPR, type, fold_convert (type, m), tmp_m);
551 else
552 m = tmp_m;
554 if (a)
555 a = fold_build2 (MULT_EXPR, type, fold_convert (type, a), tmp_m);
559 /* See if this is a tail call we can handle. */
560 ret_var = gimple_return_retval (stmt);
562 /* We may proceed if there either is no return value, or the return value
563 is identical to the call's return. */
564 if (ret_var
565 && (ret_var != ass_var))
566 return;
568 /* If this is not a tail recursive call, we cannot handle addends or
569 multiplicands. */
570 if (!tail_recursion && (m || a))
571 return;
573 nw = XNEW (struct tailcall);
575 nw->call_gsi = gsi;
577 nw->tail_recursion = tail_recursion;
579 nw->mult = m;
580 nw->add = a;
582 nw->next = *ret;
583 *ret = nw;
586 /* Helper to insert PHI_ARGH to the phi of VAR in the destination of edge E. */
588 static void
589 add_successor_phi_arg (edge e, tree var, tree phi_arg)
591 gimple_stmt_iterator gsi;
593 for (gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi); gsi_next (&gsi))
594 if (PHI_RESULT (gsi_stmt (gsi)) == var)
595 break;
597 gcc_assert (!gsi_end_p (gsi));
598 add_phi_arg (gsi_stmt (gsi), phi_arg, e, UNKNOWN_LOCATION);
601 /* Creates a GIMPLE statement which computes the operation specified by
602 CODE, ACC and OP1 to a new variable with name LABEL and inserts the
603 statement in the position specified by GSI. Returns the
604 tree node of the statement's result. */
606 static tree
607 adjust_return_value_with_ops (enum tree_code code, const char *label,
608 tree acc, tree op1, gimple_stmt_iterator gsi)
611 tree ret_type = TREE_TYPE (DECL_RESULT (current_function_decl));
612 tree result = make_temp_ssa_name (ret_type, NULL, label);
613 gimple stmt;
615 if (types_compatible_p (TREE_TYPE (acc), TREE_TYPE (op1)))
616 stmt = gimple_build_assign_with_ops (code, result, acc, op1);
617 else
619 tree rhs = fold_convert (TREE_TYPE (acc),
620 fold_build2 (code,
621 TREE_TYPE (op1),
622 fold_convert (TREE_TYPE (op1), acc),
623 op1));
624 rhs = force_gimple_operand_gsi (&gsi, rhs,
625 false, NULL, true, GSI_SAME_STMT);
626 stmt = gimple_build_assign (result, rhs);
629 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
630 return result;
633 /* Creates a new GIMPLE statement that adjusts the value of accumulator ACC by
634 the computation specified by CODE and OP1 and insert the statement
635 at the position specified by GSI as a new statement. Returns new SSA name
636 of updated accumulator. */
638 static tree
639 update_accumulator_with_ops (enum tree_code code, tree acc, tree op1,
640 gimple_stmt_iterator gsi)
642 gimple stmt;
643 tree var = copy_ssa_name (acc, NULL);
644 if (types_compatible_p (TREE_TYPE (acc), TREE_TYPE (op1)))
645 stmt = gimple_build_assign_with_ops (code, var, acc, op1);
646 else
648 tree rhs = fold_convert (TREE_TYPE (acc),
649 fold_build2 (code,
650 TREE_TYPE (op1),
651 fold_convert (TREE_TYPE (op1), acc),
652 op1));
653 rhs = force_gimple_operand_gsi (&gsi, rhs,
654 false, NULL, false, GSI_CONTINUE_LINKING);
655 stmt = gimple_build_assign (var, rhs);
657 gsi_insert_after (&gsi, stmt, GSI_NEW_STMT);
658 return var;
661 /* Adjust the accumulator values according to A and M after GSI, and update
662 the phi nodes on edge BACK. */
664 static void
665 adjust_accumulator_values (gimple_stmt_iterator gsi, tree m, tree a, edge back)
667 tree var, a_acc_arg, m_acc_arg;
669 if (m)
670 m = force_gimple_operand_gsi (&gsi, m, true, NULL, true, GSI_SAME_STMT);
671 if (a)
672 a = force_gimple_operand_gsi (&gsi, a, true, NULL, true, GSI_SAME_STMT);
674 a_acc_arg = a_acc;
675 m_acc_arg = m_acc;
676 if (a)
678 if (m_acc)
680 if (integer_onep (a))
681 var = m_acc;
682 else
683 var = adjust_return_value_with_ops (MULT_EXPR, "acc_tmp", m_acc,
684 a, gsi);
686 else
687 var = a;
689 a_acc_arg = update_accumulator_with_ops (PLUS_EXPR, a_acc, var, gsi);
692 if (m)
693 m_acc_arg = update_accumulator_with_ops (MULT_EXPR, m_acc, m, gsi);
695 if (a_acc)
696 add_successor_phi_arg (back, a_acc, a_acc_arg);
698 if (m_acc)
699 add_successor_phi_arg (back, m_acc, m_acc_arg);
702 /* Adjust value of the return at the end of BB according to M and A
703 accumulators. */
705 static void
706 adjust_return_value (basic_block bb, tree m, tree a)
708 tree retval;
709 gimple ret_stmt = gimple_seq_last_stmt (bb_seq (bb));
710 gimple_stmt_iterator gsi = gsi_last_bb (bb);
712 gcc_assert (gimple_code (ret_stmt) == GIMPLE_RETURN);
714 retval = gimple_return_retval (ret_stmt);
715 if (!retval || retval == error_mark_node)
716 return;
718 if (m)
719 retval = adjust_return_value_with_ops (MULT_EXPR, "mul_tmp", m_acc, retval,
720 gsi);
721 if (a)
722 retval = adjust_return_value_with_ops (PLUS_EXPR, "acc_tmp", a_acc, retval,
723 gsi);
724 gimple_return_set_retval (ret_stmt, retval);
725 update_stmt (ret_stmt);
728 /* Subtract COUNT and FREQUENCY from the basic block and it's
729 outgoing edge. */
730 static void
731 decrease_profile (basic_block bb, gcov_type count, int frequency)
733 edge e;
734 bb->count -= count;
735 if (bb->count < 0)
736 bb->count = 0;
737 bb->frequency -= frequency;
738 if (bb->frequency < 0)
739 bb->frequency = 0;
740 if (!single_succ_p (bb))
742 gcc_assert (!EDGE_COUNT (bb->succs));
743 return;
745 e = single_succ_edge (bb);
746 e->count -= count;
747 if (e->count < 0)
748 e->count = 0;
751 /* Returns true if argument PARAM of the tail recursive call needs to be copied
752 when the call is eliminated. */
754 static bool
755 arg_needs_copy_p (tree param)
757 tree def;
759 if (!is_gimple_reg (param))
760 return false;
762 /* Parameters that are only defined but never used need not be copied. */
763 def = ssa_default_def (cfun, param);
764 if (!def)
765 return false;
767 return true;
770 /* Eliminates tail call described by T. TMP_VARS is a list of
771 temporary variables used to copy the function arguments. */
773 static void
774 eliminate_tail_call (struct tailcall *t)
776 tree param, rslt;
777 gimple stmt, call;
778 tree arg;
779 size_t idx;
780 basic_block bb, first;
781 edge e;
782 gimple phi;
783 gimple_stmt_iterator gsi;
784 gimple orig_stmt;
786 stmt = orig_stmt = gsi_stmt (t->call_gsi);
787 bb = gsi_bb (t->call_gsi);
789 if (dump_file && (dump_flags & TDF_DETAILS))
791 fprintf (dump_file, "Eliminated tail recursion in bb %d : ",
792 bb->index);
793 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
794 fprintf (dump_file, "\n");
797 gcc_assert (is_gimple_call (stmt));
799 first = single_succ (ENTRY_BLOCK_PTR);
801 /* Remove the code after call_gsi that will become unreachable. The
802 possibly unreachable code in other blocks is removed later in
803 cfg cleanup. */
804 gsi = t->call_gsi;
805 gsi_next (&gsi);
806 while (!gsi_end_p (gsi))
808 gimple t = gsi_stmt (gsi);
809 /* Do not remove the return statement, so that redirect_edge_and_branch
810 sees how the block ends. */
811 if (gimple_code (t) == GIMPLE_RETURN)
812 break;
814 gsi_remove (&gsi, true);
815 release_defs (t);
818 /* Number of executions of function has reduced by the tailcall. */
819 e = single_succ_edge (gsi_bb (t->call_gsi));
820 decrease_profile (EXIT_BLOCK_PTR, e->count, EDGE_FREQUENCY (e));
821 decrease_profile (ENTRY_BLOCK_PTR, e->count, EDGE_FREQUENCY (e));
822 if (e->dest != EXIT_BLOCK_PTR)
823 decrease_profile (e->dest, e->count, EDGE_FREQUENCY (e));
825 /* Replace the call by a jump to the start of function. */
826 e = redirect_edge_and_branch (single_succ_edge (gsi_bb (t->call_gsi)),
827 first);
828 gcc_assert (e);
829 PENDING_STMT (e) = NULL;
831 /* Add phi node entries for arguments. The ordering of the phi nodes should
832 be the same as the ordering of the arguments. */
833 for (param = DECL_ARGUMENTS (current_function_decl),
834 idx = 0, gsi = gsi_start_phis (first);
835 param;
836 param = DECL_CHAIN (param), idx++)
838 if (!arg_needs_copy_p (param))
839 continue;
841 arg = gimple_call_arg (stmt, idx);
842 phi = gsi_stmt (gsi);
843 gcc_assert (param == SSA_NAME_VAR (PHI_RESULT (phi)));
845 add_phi_arg (phi, arg, e, gimple_location (stmt));
846 gsi_next (&gsi);
849 /* Update the values of accumulators. */
850 adjust_accumulator_values (t->call_gsi, t->mult, t->add, e);
852 call = gsi_stmt (t->call_gsi);
853 rslt = gimple_call_lhs (call);
854 if (rslt != NULL_TREE)
856 /* Result of the call will no longer be defined. So adjust the
857 SSA_NAME_DEF_STMT accordingly. */
858 SSA_NAME_DEF_STMT (rslt) = gimple_build_nop ();
861 gsi_remove (&t->call_gsi, true);
862 release_defs (call);
865 /* Optimizes the tailcall described by T. If OPT_TAILCALLS is true, also
866 mark the tailcalls for the sibcall optimization. */
868 static bool
869 optimize_tail_call (struct tailcall *t, bool opt_tailcalls)
871 if (t->tail_recursion)
873 eliminate_tail_call (t);
874 return true;
877 if (opt_tailcalls)
879 gimple stmt = gsi_stmt (t->call_gsi);
881 gimple_call_set_tail (stmt, true);
882 if (dump_file && (dump_flags & TDF_DETAILS))
884 fprintf (dump_file, "Found tail call ");
885 print_gimple_stmt (dump_file, stmt, 0, dump_flags);
886 fprintf (dump_file, " in bb %i\n", (gsi_bb (t->call_gsi))->index);
890 return false;
893 /* Creates a tail-call accumulator of the same type as the return type of the
894 current function. LABEL is the name used to creating the temporary
895 variable for the accumulator. The accumulator will be inserted in the
896 phis of a basic block BB with single predecessor with an initial value
897 INIT converted to the current function return type. */
899 static tree
900 create_tailcall_accumulator (const char *label, basic_block bb, tree init)
902 tree ret_type = TREE_TYPE (DECL_RESULT (current_function_decl));
903 tree tmp = make_temp_ssa_name (ret_type, NULL, label);
904 gimple phi;
906 phi = create_phi_node (tmp, bb);
907 /* RET_TYPE can be a float when -ffast-maths is enabled. */
908 add_phi_arg (phi, fold_convert (ret_type, init), single_pred_edge (bb),
909 UNKNOWN_LOCATION);
910 return PHI_RESULT (phi);
913 /* Optimizes tail calls in the function, turning the tail recursion
914 into iteration. */
916 static unsigned int
917 tree_optimize_tail_calls_1 (bool opt_tailcalls)
919 edge e;
920 bool phis_constructed = false;
921 struct tailcall *tailcalls = NULL, *act, *next;
922 bool changed = false;
923 basic_block first = single_succ (ENTRY_BLOCK_PTR);
924 tree param;
925 gimple stmt;
926 edge_iterator ei;
928 if (!suitable_for_tail_opt_p ())
929 return 0;
930 if (opt_tailcalls)
931 opt_tailcalls = suitable_for_tail_call_opt_p ();
933 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds)
935 /* Only traverse the normal exits, i.e. those that end with return
936 statement. */
937 stmt = last_stmt (e->src);
939 if (stmt
940 && gimple_code (stmt) == GIMPLE_RETURN)
941 find_tail_calls (e->src, &tailcalls);
944 /* Construct the phi nodes and accumulators if necessary. */
945 a_acc = m_acc = NULL_TREE;
946 for (act = tailcalls; act; act = act->next)
948 if (!act->tail_recursion)
949 continue;
951 if (!phis_constructed)
953 /* Ensure that there is only one predecessor of the block
954 or if there are existing degenerate PHI nodes. */
955 if (!single_pred_p (first)
956 || !gimple_seq_empty_p (phi_nodes (first)))
957 first = split_edge (single_succ_edge (ENTRY_BLOCK_PTR));
959 /* Copy the args if needed. */
960 for (param = DECL_ARGUMENTS (current_function_decl);
961 param;
962 param = DECL_CHAIN (param))
963 if (arg_needs_copy_p (param))
965 tree name = ssa_default_def (cfun, param);
966 tree new_name = make_ssa_name (param, SSA_NAME_DEF_STMT (name));
967 gimple phi;
969 set_ssa_default_def (cfun, param, new_name);
970 phi = create_phi_node (name, first);
971 add_phi_arg (phi, new_name, single_pred_edge (first),
972 EXPR_LOCATION (param));
974 phis_constructed = true;
977 if (act->add && !a_acc)
978 a_acc = create_tailcall_accumulator ("add_acc", first,
979 integer_zero_node);
981 if (act->mult && !m_acc)
982 m_acc = create_tailcall_accumulator ("mult_acc", first,
983 integer_one_node);
986 if (a_acc || m_acc)
988 /* When the tail call elimination using accumulators is performed,
989 statements adding the accumulated value are inserted at all exits.
990 This turns all other tail calls to non-tail ones. */
991 opt_tailcalls = false;
994 for (; tailcalls; tailcalls = next)
996 next = tailcalls->next;
997 changed |= optimize_tail_call (tailcalls, opt_tailcalls);
998 free (tailcalls);
1001 if (a_acc || m_acc)
1003 /* Modify the remaining return statements. */
1004 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds)
1006 stmt = last_stmt (e->src);
1008 if (stmt
1009 && gimple_code (stmt) == GIMPLE_RETURN)
1010 adjust_return_value (e->src, m_acc, a_acc);
1014 if (changed)
1016 /* We may have created new loops. Make them magically appear. */
1017 if (current_loops)
1018 loops_state_set (LOOPS_NEED_FIXUP);
1019 free_dominance_info (CDI_DOMINATORS);
1022 /* Add phi nodes for the virtual operands defined in the function to the
1023 header of the loop created by tail recursion elimination. Do so
1024 by triggering the SSA renamer. */
1025 if (phis_constructed)
1026 mark_virtual_operands_for_renaming (cfun);
1028 if (changed)
1029 return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals;
1030 return 0;
1033 static unsigned int
1034 execute_tail_recursion (void)
1036 return tree_optimize_tail_calls_1 (false);
1039 static bool
1040 gate_tail_calls (void)
1042 return flag_optimize_sibling_calls != 0 && dbg_cnt (tail_call);
1045 static unsigned int
1046 execute_tail_calls (void)
1048 return tree_optimize_tail_calls_1 (true);
1051 struct gimple_opt_pass pass_tail_recursion =
1054 GIMPLE_PASS,
1055 "tailr", /* name */
1056 OPTGROUP_NONE, /* optinfo_flags */
1057 gate_tail_calls, /* gate */
1058 execute_tail_recursion, /* execute */
1059 NULL, /* sub */
1060 NULL, /* next */
1061 0, /* static_pass_number */
1062 TV_NONE, /* tv_id */
1063 PROP_cfg | PROP_ssa, /* properties_required */
1064 0, /* properties_provided */
1065 0, /* properties_destroyed */
1066 0, /* todo_flags_start */
1067 TODO_verify_ssa /* todo_flags_finish */
1071 struct gimple_opt_pass pass_tail_calls =
1074 GIMPLE_PASS,
1075 "tailc", /* name */
1076 OPTGROUP_NONE, /* optinfo_flags */
1077 gate_tail_calls, /* gate */
1078 execute_tail_calls, /* execute */
1079 NULL, /* sub */
1080 NULL, /* next */
1081 0, /* static_pass_number */
1082 TV_NONE, /* tv_id */
1083 PROP_cfg | PROP_ssa, /* properties_required */
1084 0, /* properties_provided */
1085 0, /* properties_destroyed */
1086 0, /* todo_flags_start */
1087 TODO_verify_ssa /* todo_flags_finish */