* cp-tree.h (struct deferred_access_check): Add location.
[official-gcc.git] / gcc / tree-tailcall.c
blobe2e51a45107287a94edac1b044483f7f1e8365d5
1 /* Tail call optimization on trees.
2 Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011, 2012
3 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 3, or (at your option)
10 any later version.
12 GCC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 #include "config.h"
22 #include "system.h"
23 #include "coretypes.h"
24 #include "tm.h"
25 #include "tree.h"
26 #include "tm_p.h"
27 #include "basic-block.h"
28 #include "function.h"
29 #include "tree-flow.h"
30 #include "gimple-pretty-print.h"
31 #include "except.h"
32 #include "tree-pass.h"
33 #include "flags.h"
34 #include "langhooks.h"
35 #include "dbgcnt.h"
36 #include "target.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;
393 referenced_var_iterator rvi;
395 if (!single_succ_p (bb))
396 return;
398 for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi); gsi_prev (&gsi))
400 stmt = gsi_stmt (gsi);
402 /* Ignore labels, returns, clobbers and debug stmts. */
403 if (gimple_code (stmt) == GIMPLE_LABEL
404 || gimple_code (stmt) == GIMPLE_RETURN
405 || gimple_clobber_p (stmt)
406 || is_gimple_debug (stmt))
407 continue;
409 /* Check for a call. */
410 if (is_gimple_call (stmt))
412 call = stmt;
413 ass_var = gimple_call_lhs (stmt);
414 break;
417 /* If the statement references memory or volatile operands, fail. */
418 if (gimple_references_memory_p (stmt)
419 || gimple_has_volatile_ops (stmt))
420 return;
423 if (gsi_end_p (gsi))
425 edge_iterator ei;
426 /* Recurse to the predecessors. */
427 FOR_EACH_EDGE (e, ei, bb->preds)
428 find_tail_calls (e->src, ret);
430 return;
433 /* If the LHS of our call is not just a simple register, we can't
434 transform this into a tail or sibling call. This situation happens,
435 in (e.g.) "*p = foo()" where foo returns a struct. In this case
436 we won't have a temporary here, but we need to carry out the side
437 effect anyway, so tailcall is impossible.
439 ??? In some situations (when the struct is returned in memory via
440 invisible argument) we could deal with this, e.g. by passing 'p'
441 itself as that argument to foo, but it's too early to do this here,
442 and expand_call() will not handle it anyway. If it ever can, then
443 we need to revisit this here, to allow that situation. */
444 if (ass_var && !is_gimple_reg (ass_var))
445 return;
447 /* We found the call, check whether it is suitable. */
448 tail_recursion = false;
449 func = gimple_call_fndecl (call);
450 if (func == current_function_decl)
452 tree arg;
454 for (param = DECL_ARGUMENTS (func), idx = 0;
455 param && idx < gimple_call_num_args (call);
456 param = DECL_CHAIN (param), idx ++)
458 arg = gimple_call_arg (call, idx);
459 if (param != arg)
461 /* Make sure there are no problems with copying. The parameter
462 have a copyable type and the two arguments must have reasonably
463 equivalent types. The latter requirement could be relaxed if
464 we emitted a suitable type conversion statement. */
465 if (!is_gimple_reg_type (TREE_TYPE (param))
466 || !useless_type_conversion_p (TREE_TYPE (param),
467 TREE_TYPE (arg)))
468 break;
470 /* The parameter should be a real operand, so that phi node
471 created for it at the start of the function has the meaning
472 of copying the value. This test implies is_gimple_reg_type
473 from the previous condition, however this one could be
474 relaxed by being more careful with copying the new value
475 of the parameter (emitting appropriate GIMPLE_ASSIGN and
476 updating the virtual operands). */
477 if (!is_gimple_reg (param))
478 break;
481 if (idx == gimple_call_num_args (call) && !param)
482 tail_recursion = true;
485 /* Make sure the tail invocation of this function does not refer
486 to local variables. */
487 FOR_EACH_REFERENCED_VAR (cfun, var, rvi)
489 if (TREE_CODE (var) != PARM_DECL
490 && auto_var_in_fn_p (var, cfun->decl)
491 && (ref_maybe_used_by_stmt_p (call, var)
492 || call_may_clobber_ref_p (call, var)))
493 return;
496 /* Now check the statements after the call. None of them has virtual
497 operands, so they may only depend on the call through its return
498 value. The return value should also be dependent on each of them,
499 since we are running after dce. */
500 m = NULL_TREE;
501 a = NULL_TREE;
503 abb = bb;
504 agsi = gsi;
505 while (1)
507 tree tmp_a = NULL_TREE;
508 tree tmp_m = NULL_TREE;
509 gsi_next (&agsi);
511 while (gsi_end_p (agsi))
513 ass_var = propagate_through_phis (ass_var, single_succ_edge (abb));
514 abb = single_succ (abb);
515 agsi = gsi_start_bb (abb);
518 stmt = gsi_stmt (agsi);
520 if (gimple_code (stmt) == GIMPLE_LABEL)
521 continue;
523 if (gimple_code (stmt) == GIMPLE_RETURN)
524 break;
526 if (gimple_clobber_p (stmt))
527 continue;
529 if (is_gimple_debug (stmt))
530 continue;
532 if (gimple_code (stmt) != GIMPLE_ASSIGN)
533 return;
535 /* This is a gimple assign. */
536 if (! process_assignment (stmt, gsi, &tmp_m, &tmp_a, &ass_var))
537 return;
539 if (tmp_a)
541 tree type = TREE_TYPE (tmp_a);
542 if (a)
543 a = fold_build2 (PLUS_EXPR, type, fold_convert (type, a), tmp_a);
544 else
545 a = tmp_a;
547 if (tmp_m)
549 tree type = TREE_TYPE (tmp_m);
550 if (m)
551 m = fold_build2 (MULT_EXPR, type, fold_convert (type, m), tmp_m);
552 else
553 m = tmp_m;
555 if (a)
556 a = fold_build2 (MULT_EXPR, type, fold_convert (type, a), tmp_m);
560 /* See if this is a tail call we can handle. */
561 ret_var = gimple_return_retval (stmt);
563 /* We may proceed if there either is no return value, or the return value
564 is identical to the call's return. */
565 if (ret_var
566 && (ret_var != ass_var))
567 return;
569 /* If this is not a tail recursive call, we cannot handle addends or
570 multiplicands. */
571 if (!tail_recursion && (m || a))
572 return;
574 nw = XNEW (struct tailcall);
576 nw->call_gsi = gsi;
578 nw->tail_recursion = tail_recursion;
580 nw->mult = m;
581 nw->add = a;
583 nw->next = *ret;
584 *ret = nw;
587 /* Helper to insert PHI_ARGH to the phi of VAR in the destination of edge E. */
589 static void
590 add_successor_phi_arg (edge e, tree var, tree phi_arg)
592 gimple_stmt_iterator gsi;
594 for (gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi); gsi_next (&gsi))
595 if (PHI_RESULT (gsi_stmt (gsi)) == var)
596 break;
598 gcc_assert (!gsi_end_p (gsi));
599 add_phi_arg (gsi_stmt (gsi), phi_arg, e, UNKNOWN_LOCATION);
602 /* Creates a GIMPLE statement which computes the operation specified by
603 CODE, OP0 and OP1 to a new variable with name LABEL and inserts the
604 statement in the position specified by GSI and UPDATE. Returns the
605 tree node of the statement's result. */
607 static tree
608 adjust_return_value_with_ops (enum tree_code code, const char *label,
609 tree acc, tree op1, gimple_stmt_iterator gsi)
612 tree ret_type = TREE_TYPE (DECL_RESULT (current_function_decl));
613 tree tmp = create_tmp_reg (ret_type, label);
614 gimple stmt;
615 tree result;
617 add_referenced_var (tmp);
619 if (types_compatible_p (TREE_TYPE (acc), TREE_TYPE (op1)))
620 stmt = gimple_build_assign_with_ops (code, tmp, acc, op1);
621 else
623 tree rhs = fold_convert (TREE_TYPE (acc),
624 fold_build2 (code,
625 TREE_TYPE (op1),
626 fold_convert (TREE_TYPE (op1), acc),
627 op1));
628 rhs = force_gimple_operand_gsi (&gsi, rhs,
629 false, NULL, true, GSI_CONTINUE_LINKING);
630 stmt = gimple_build_assign (NULL_TREE, rhs);
633 result = make_ssa_name (tmp, stmt);
634 gimple_assign_set_lhs (stmt, result);
635 update_stmt (stmt);
636 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
637 return result;
640 /* Creates a new GIMPLE statement that adjusts the value of accumulator ACC by
641 the computation specified by CODE and OP1 and insert the statement
642 at the position specified by GSI as a new statement. Returns new SSA name
643 of updated accumulator. */
645 static tree
646 update_accumulator_with_ops (enum tree_code code, tree acc, tree op1,
647 gimple_stmt_iterator gsi)
649 gimple stmt;
650 tree var;
651 if (types_compatible_p (TREE_TYPE (acc), TREE_TYPE (op1)))
652 stmt = gimple_build_assign_with_ops (code, SSA_NAME_VAR (acc), acc, op1);
653 else
655 tree rhs = fold_convert (TREE_TYPE (acc),
656 fold_build2 (code,
657 TREE_TYPE (op1),
658 fold_convert (TREE_TYPE (op1), acc),
659 op1));
660 rhs = force_gimple_operand_gsi (&gsi, rhs,
661 false, NULL, false, GSI_CONTINUE_LINKING);
662 stmt = gimple_build_assign (NULL_TREE, rhs);
664 var = make_ssa_name (SSA_NAME_VAR (acc), stmt);
665 gimple_assign_set_lhs (stmt, var);
666 update_stmt (stmt);
667 gsi_insert_after (&gsi, stmt, GSI_NEW_STMT);
668 return var;
671 /* Adjust the accumulator values according to A and M after GSI, and update
672 the phi nodes on edge BACK. */
674 static void
675 adjust_accumulator_values (gimple_stmt_iterator gsi, tree m, tree a, edge back)
677 tree var, a_acc_arg, m_acc_arg;
679 if (m)
680 m = force_gimple_operand_gsi (&gsi, m, true, NULL, true, GSI_SAME_STMT);
681 if (a)
682 a = force_gimple_operand_gsi (&gsi, a, true, NULL, true, GSI_SAME_STMT);
684 a_acc_arg = a_acc;
685 m_acc_arg = m_acc;
686 if (a)
688 if (m_acc)
690 if (integer_onep (a))
691 var = m_acc;
692 else
693 var = adjust_return_value_with_ops (MULT_EXPR, "acc_tmp", m_acc,
694 a, gsi);
696 else
697 var = a;
699 a_acc_arg = update_accumulator_with_ops (PLUS_EXPR, a_acc, var, gsi);
702 if (m)
703 m_acc_arg = update_accumulator_with_ops (MULT_EXPR, m_acc, m, gsi);
705 if (a_acc)
706 add_successor_phi_arg (back, a_acc, a_acc_arg);
708 if (m_acc)
709 add_successor_phi_arg (back, m_acc, m_acc_arg);
712 /* Adjust value of the return at the end of BB according to M and A
713 accumulators. */
715 static void
716 adjust_return_value (basic_block bb, tree m, tree a)
718 tree retval;
719 gimple ret_stmt = gimple_seq_last_stmt (bb_seq (bb));
720 gimple_stmt_iterator gsi = gsi_last_bb (bb);
722 gcc_assert (gimple_code (ret_stmt) == GIMPLE_RETURN);
724 retval = gimple_return_retval (ret_stmt);
725 if (!retval || retval == error_mark_node)
726 return;
728 if (m)
729 retval = adjust_return_value_with_ops (MULT_EXPR, "mul_tmp", m_acc, retval,
730 gsi);
731 if (a)
732 retval = adjust_return_value_with_ops (PLUS_EXPR, "acc_tmp", a_acc, retval,
733 gsi);
734 gimple_return_set_retval (ret_stmt, retval);
735 update_stmt (ret_stmt);
738 /* Subtract COUNT and FREQUENCY from the basic block and it's
739 outgoing edge. */
740 static void
741 decrease_profile (basic_block bb, gcov_type count, int frequency)
743 edge e;
744 bb->count -= count;
745 if (bb->count < 0)
746 bb->count = 0;
747 bb->frequency -= frequency;
748 if (bb->frequency < 0)
749 bb->frequency = 0;
750 if (!single_succ_p (bb))
752 gcc_assert (!EDGE_COUNT (bb->succs));
753 return;
755 e = single_succ_edge (bb);
756 e->count -= count;
757 if (e->count < 0)
758 e->count = 0;
761 /* Returns true if argument PARAM of the tail recursive call needs to be copied
762 when the call is eliminated. */
764 static bool
765 arg_needs_copy_p (tree param)
767 tree def;
769 if (!is_gimple_reg (param) || !var_ann (param))
770 return false;
772 /* Parameters that are only defined but never used need not be copied. */
773 def = gimple_default_def (cfun, param);
774 if (!def)
775 return false;
777 return true;
780 /* Eliminates tail call described by T. TMP_VARS is a list of
781 temporary variables used to copy the function arguments. */
783 static void
784 eliminate_tail_call (struct tailcall *t)
786 tree param, rslt;
787 gimple stmt, call;
788 tree arg;
789 size_t idx;
790 basic_block bb, first;
791 edge e;
792 gimple phi;
793 gimple_stmt_iterator gsi;
794 gimple orig_stmt;
796 stmt = orig_stmt = gsi_stmt (t->call_gsi);
797 bb = gsi_bb (t->call_gsi);
799 if (dump_file && (dump_flags & TDF_DETAILS))
801 fprintf (dump_file, "Eliminated tail recursion in bb %d : ",
802 bb->index);
803 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
804 fprintf (dump_file, "\n");
807 gcc_assert (is_gimple_call (stmt));
809 first = single_succ (ENTRY_BLOCK_PTR);
811 /* Remove the code after call_gsi that will become unreachable. The
812 possibly unreachable code in other blocks is removed later in
813 cfg cleanup. */
814 gsi = t->call_gsi;
815 gsi_next (&gsi);
816 while (!gsi_end_p (gsi))
818 gimple t = gsi_stmt (gsi);
819 /* Do not remove the return statement, so that redirect_edge_and_branch
820 sees how the block ends. */
821 if (gimple_code (t) == GIMPLE_RETURN)
822 break;
824 gsi_remove (&gsi, true);
825 release_defs (t);
828 /* Number of executions of function has reduced by the tailcall. */
829 e = single_succ_edge (gsi_bb (t->call_gsi));
830 decrease_profile (EXIT_BLOCK_PTR, e->count, EDGE_FREQUENCY (e));
831 decrease_profile (ENTRY_BLOCK_PTR, e->count, EDGE_FREQUENCY (e));
832 if (e->dest != EXIT_BLOCK_PTR)
833 decrease_profile (e->dest, e->count, EDGE_FREQUENCY (e));
835 /* Replace the call by a jump to the start of function. */
836 e = redirect_edge_and_branch (single_succ_edge (gsi_bb (t->call_gsi)),
837 first);
838 gcc_assert (e);
839 PENDING_STMT (e) = NULL;
841 /* Add phi node entries for arguments. The ordering of the phi nodes should
842 be the same as the ordering of the arguments. */
843 for (param = DECL_ARGUMENTS (current_function_decl),
844 idx = 0, gsi = gsi_start_phis (first);
845 param;
846 param = DECL_CHAIN (param), idx++)
848 if (!arg_needs_copy_p (param))
849 continue;
851 arg = gimple_call_arg (stmt, idx);
852 phi = gsi_stmt (gsi);
853 gcc_assert (param == SSA_NAME_VAR (PHI_RESULT (phi)));
855 add_phi_arg (phi, arg, e, gimple_location (stmt));
856 gsi_next (&gsi);
859 /* Update the values of accumulators. */
860 adjust_accumulator_values (t->call_gsi, t->mult, t->add, e);
862 call = gsi_stmt (t->call_gsi);
863 rslt = gimple_call_lhs (call);
864 if (rslt != NULL_TREE)
866 /* Result of the call will no longer be defined. So adjust the
867 SSA_NAME_DEF_STMT accordingly. */
868 SSA_NAME_DEF_STMT (rslt) = gimple_build_nop ();
871 gsi_remove (&t->call_gsi, true);
872 release_defs (call);
875 /* Add phi nodes for the virtual operands defined in the function to the
876 header of the loop created by tail recursion elimination.
878 Originally, we used to add phi nodes only for call clobbered variables,
879 as the value of the non-call clobbered ones obviously cannot be used
880 or changed within the recursive call. However, the local variables
881 from multiple calls now share the same location, so the virtual ssa form
882 requires us to say that the location dies on further iterations of the loop,
883 which requires adding phi nodes.
885 static void
886 add_virtual_phis (void)
888 referenced_var_iterator rvi;
889 tree var;
891 /* The problematic part is that there is no way how to know what
892 to put into phi nodes (there in fact does not have to be such
893 ssa name available). A solution would be to have an artificial
894 use/kill for all virtual operands in EXIT node. Unless we have
895 this, we cannot do much better than to rebuild the ssa form for
896 possibly affected virtual ssa names from scratch. */
898 FOR_EACH_REFERENCED_VAR (cfun, var, rvi)
900 if (!is_gimple_reg (var) && gimple_default_def (cfun, var) != NULL_TREE)
901 mark_sym_for_renaming (var);
905 /* Optimizes the tailcall described by T. If OPT_TAILCALLS is true, also
906 mark the tailcalls for the sibcall optimization. */
908 static bool
909 optimize_tail_call (struct tailcall *t, bool opt_tailcalls)
911 if (t->tail_recursion)
913 eliminate_tail_call (t);
914 return true;
917 if (opt_tailcalls)
919 gimple stmt = gsi_stmt (t->call_gsi);
921 gimple_call_set_tail (stmt, true);
922 if (dump_file && (dump_flags & TDF_DETAILS))
924 fprintf (dump_file, "Found tail call ");
925 print_gimple_stmt (dump_file, stmt, 0, dump_flags);
926 fprintf (dump_file, " in bb %i\n", (gsi_bb (t->call_gsi))->index);
930 return false;
933 /* Creates a tail-call accumulator of the same type as the return type of the
934 current function. LABEL is the name used to creating the temporary
935 variable for the accumulator. The accumulator will be inserted in the
936 phis of a basic block BB with single predecessor with an initial value
937 INIT converted to the current function return type. */
939 static tree
940 create_tailcall_accumulator (const char *label, basic_block bb, tree init)
942 tree ret_type = TREE_TYPE (DECL_RESULT (current_function_decl));
943 tree tmp = create_tmp_reg (ret_type, label);
944 gimple phi;
946 add_referenced_var (tmp);
947 phi = create_phi_node (tmp, bb);
948 /* RET_TYPE can be a float when -ffast-maths is enabled. */
949 add_phi_arg (phi, fold_convert (ret_type, init), single_pred_edge (bb),
950 UNKNOWN_LOCATION);
951 return PHI_RESULT (phi);
954 /* Optimizes tail calls in the function, turning the tail recursion
955 into iteration. */
957 static unsigned int
958 tree_optimize_tail_calls_1 (bool opt_tailcalls)
960 edge e;
961 bool phis_constructed = false;
962 struct tailcall *tailcalls = NULL, *act, *next;
963 bool changed = false;
964 basic_block first = single_succ (ENTRY_BLOCK_PTR);
965 tree param;
966 gimple stmt;
967 edge_iterator ei;
969 if (!suitable_for_tail_opt_p ())
970 return 0;
971 if (opt_tailcalls)
972 opt_tailcalls = suitable_for_tail_call_opt_p ();
974 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds)
976 /* Only traverse the normal exits, i.e. those that end with return
977 statement. */
978 stmt = last_stmt (e->src);
980 if (stmt
981 && gimple_code (stmt) == GIMPLE_RETURN)
982 find_tail_calls (e->src, &tailcalls);
985 /* Construct the phi nodes and accumulators if necessary. */
986 a_acc = m_acc = NULL_TREE;
987 for (act = tailcalls; act; act = act->next)
989 if (!act->tail_recursion)
990 continue;
992 if (!phis_constructed)
994 /* Ensure that there is only one predecessor of the block
995 or if there are existing degenerate PHI nodes. */
996 if (!single_pred_p (first)
997 || !gimple_seq_empty_p (phi_nodes (first)))
998 first = split_edge (single_succ_edge (ENTRY_BLOCK_PTR));
1000 /* Copy the args if needed. */
1001 for (param = DECL_ARGUMENTS (current_function_decl);
1002 param;
1003 param = DECL_CHAIN (param))
1004 if (arg_needs_copy_p (param))
1006 tree name = gimple_default_def (cfun, param);
1007 tree new_name = make_ssa_name (param, SSA_NAME_DEF_STMT (name));
1008 gimple phi;
1010 set_default_def (param, new_name);
1011 phi = create_phi_node (name, first);
1012 SSA_NAME_DEF_STMT (name) = phi;
1013 add_phi_arg (phi, new_name, single_pred_edge (first),
1014 EXPR_LOCATION (param));
1016 phis_constructed = true;
1019 if (act->add && !a_acc)
1020 a_acc = create_tailcall_accumulator ("add_acc", first,
1021 integer_zero_node);
1023 if (act->mult && !m_acc)
1024 m_acc = create_tailcall_accumulator ("mult_acc", first,
1025 integer_one_node);
1028 if (a_acc || m_acc)
1030 /* When the tail call elimination using accumulators is performed,
1031 statements adding the accumulated value are inserted at all exits.
1032 This turns all other tail calls to non-tail ones. */
1033 opt_tailcalls = false;
1036 for (; tailcalls; tailcalls = next)
1038 next = tailcalls->next;
1039 changed |= optimize_tail_call (tailcalls, opt_tailcalls);
1040 free (tailcalls);
1043 if (a_acc || m_acc)
1045 /* Modify the remaining return statements. */
1046 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds)
1048 stmt = last_stmt (e->src);
1050 if (stmt
1051 && gimple_code (stmt) == GIMPLE_RETURN)
1052 adjust_return_value (e->src, m_acc, a_acc);
1056 if (changed)
1057 free_dominance_info (CDI_DOMINATORS);
1059 if (phis_constructed)
1060 add_virtual_phis ();
1061 if (changed)
1062 return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals;
1063 return 0;
1066 static unsigned int
1067 execute_tail_recursion (void)
1069 return tree_optimize_tail_calls_1 (false);
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 struct gimple_opt_pass pass_tail_recursion =
1087 GIMPLE_PASS,
1088 "tailr", /* name */
1089 gate_tail_calls, /* gate */
1090 execute_tail_recursion, /* execute */
1091 NULL, /* sub */
1092 NULL, /* next */
1093 0, /* static_pass_number */
1094 TV_NONE, /* tv_id */
1095 PROP_cfg | PROP_ssa, /* properties_required */
1096 0, /* properties_provided */
1097 0, /* properties_destroyed */
1098 0, /* todo_flags_start */
1099 TODO_verify_ssa /* todo_flags_finish */
1103 struct gimple_opt_pass pass_tail_calls =
1106 GIMPLE_PASS,
1107 "tailc", /* name */
1108 gate_tail_calls, /* gate */
1109 execute_tail_calls, /* execute */
1110 NULL, /* sub */
1111 NULL, /* next */
1112 0, /* static_pass_number */
1113 TV_NONE, /* tv_id */
1114 PROP_cfg | PROP_ssa, /* properties_required */
1115 0, /* properties_provided */
1116 0, /* properties_destroyed */
1117 0, /* todo_flags_start */
1118 TODO_verify_ssa /* todo_flags_finish */