i386.md (simple LEA peephole2): Add missing mode to zero_extend for zero-extended...
[official-gcc.git] / gcc / tree-tailcall.c
blob37603b2757ce791e689111fcf34dc88859b0393c
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 "stor-layout.h"
26 #include "tm_p.h"
27 #include "basic-block.h"
28 #include "function.h"
29 #include "tree-ssa-alias.h"
30 #include "internal-fn.h"
31 #include "gimple-expr.h"
32 #include "is-a.h"
33 #include "gimple.h"
34 #include "gimple-iterator.h"
35 #include "gimplify-me.h"
36 #include "gimple-ssa.h"
37 #include "tree-cfg.h"
38 #include "tree-phinodes.h"
39 #include "stringpool.h"
40 #include "tree-ssanames.h"
41 #include "tree-into-ssa.h"
42 #include "expr.h"
43 #include "tree-dfa.h"
44 #include "gimple-pretty-print.h"
45 #include "except.h"
46 #include "tree-pass.h"
47 #include "flags.h"
48 #include "langhooks.h"
49 #include "dbgcnt.h"
50 #include "target.h"
51 #include "cfgloop.h"
52 #include "common/common-target.h"
53 #include "ipa-utils.h"
55 /* The file implements the tail recursion elimination. It is also used to
56 analyze the tail calls in general, passing the results to the rtl level
57 where they are used for sibcall optimization.
59 In addition to the standard tail recursion elimination, we handle the most
60 trivial cases of making the call tail recursive by creating accumulators.
61 For example the following function
63 int sum (int n)
65 if (n > 0)
66 return n + sum (n - 1);
67 else
68 return 0;
71 is transformed into
73 int sum (int n)
75 int acc = 0;
77 while (n > 0)
78 acc += n--;
80 return acc;
83 To do this, we maintain two accumulators (a_acc and m_acc) that indicate
84 when we reach the return x statement, we should return a_acc + x * m_acc
85 instead. They are initially initialized to 0 and 1, respectively,
86 so the semantics of the function is obviously preserved. If we are
87 guaranteed that the value of the accumulator never change, we
88 omit the accumulator.
90 There are three cases how the function may exit. The first one is
91 handled in adjust_return_value, the other two in adjust_accumulator_values
92 (the second case is actually a special case of the third one and we
93 present it separately just for clarity):
95 1) Just return x, where x is not in any of the remaining special shapes.
96 We rewrite this to a gimple equivalent of return m_acc * x + a_acc.
98 2) return f (...), where f is the current function, is rewritten in a
99 classical tail-recursion elimination way, into assignment of arguments
100 and jump to the start of the function. Values of the accumulators
101 are unchanged.
103 3) return a + m * f(...), where a and m do not depend on call to f.
104 To preserve the semantics described before we want this to be rewritten
105 in such a way that we finally return
107 a_acc + (a + m * f(...)) * m_acc = (a_acc + a * m_acc) + (m * m_acc) * f(...).
109 I.e. we increase a_acc by a * m_acc, multiply m_acc by m and
110 eliminate the tail call to f. Special cases when the value is just
111 added or just multiplied are obtained by setting a = 0 or m = 1.
113 TODO -- it is possible to do similar tricks for other operations. */
115 /* A structure that describes the tailcall. */
117 struct tailcall
119 /* The iterator pointing to the call statement. */
120 gimple_stmt_iterator call_gsi;
122 /* True if it is a call to the current function. */
123 bool tail_recursion;
125 /* The return value of the caller is mult * f + add, where f is the return
126 value of the call. */
127 tree mult, add;
129 /* Next tailcall in the chain. */
130 struct tailcall *next;
133 /* The variables holding the value of multiplicative and additive
134 accumulator. */
135 static tree m_acc, a_acc;
137 static bool suitable_for_tail_opt_p (void);
138 static bool optimize_tail_call (struct tailcall *, bool);
139 static void eliminate_tail_call (struct tailcall *);
140 static void find_tail_calls (basic_block, struct tailcall **);
142 /* Returns false when the function is not suitable for tail call optimization
143 from some reason (e.g. if it takes variable number of arguments). */
145 static bool
146 suitable_for_tail_opt_p (void)
148 if (cfun->stdarg)
149 return false;
151 return true;
153 /* Returns false when the function is not suitable for tail call optimization
154 from some reason (e.g. if it takes variable number of arguments).
155 This test must pass in addition to suitable_for_tail_opt_p in order to make
156 tail call discovery happen. */
158 static bool
159 suitable_for_tail_call_opt_p (void)
161 tree param;
163 /* alloca (until we have stack slot life analysis) inhibits
164 sibling call optimizations, but not tail recursion. */
165 if (cfun->calls_alloca)
166 return false;
168 /* If we are using sjlj exceptions, we may need to add a call to
169 _Unwind_SjLj_Unregister at exit of the function. Which means
170 that we cannot do any sibcall transformations. */
171 if (targetm_common.except_unwind_info (&global_options) == UI_SJLJ
172 && current_function_has_exception_handlers ())
173 return false;
175 /* Any function that calls setjmp might have longjmp called from
176 any called function. ??? We really should represent this
177 properly in the CFG so that this needn't be special cased. */
178 if (cfun->calls_setjmp)
179 return false;
181 /* ??? It is OK if the argument of a function is taken in some cases,
182 but not in all cases. See PR15387 and PR19616. Revisit for 4.1. */
183 for (param = DECL_ARGUMENTS (current_function_decl);
184 param;
185 param = DECL_CHAIN (param))
186 if (TREE_ADDRESSABLE (param))
187 return false;
189 return true;
192 /* Checks whether the expression EXPR in stmt AT is independent of the
193 statement pointed to by GSI (in a sense that we already know EXPR's value
194 at GSI). We use the fact that we are only called from the chain of
195 basic blocks that have only single successor. Returns the expression
196 containing the value of EXPR at GSI. */
198 static tree
199 independent_of_stmt_p (tree expr, gimple at, gimple_stmt_iterator gsi)
201 basic_block bb, call_bb, at_bb;
202 edge e;
203 edge_iterator ei;
205 if (is_gimple_min_invariant (expr))
206 return expr;
208 if (TREE_CODE (expr) != SSA_NAME)
209 return NULL_TREE;
211 /* Mark the blocks in the chain leading to the end. */
212 at_bb = gimple_bb (at);
213 call_bb = gimple_bb (gsi_stmt (gsi));
214 for (bb = call_bb; bb != at_bb; bb = single_succ (bb))
215 bb->aux = &bb->aux;
216 bb->aux = &bb->aux;
218 while (1)
220 at = SSA_NAME_DEF_STMT (expr);
221 bb = gimple_bb (at);
223 /* The default definition or defined before the chain. */
224 if (!bb || !bb->aux)
225 break;
227 if (bb == call_bb)
229 for (; !gsi_end_p (gsi); gsi_next (&gsi))
230 if (gsi_stmt (gsi) == at)
231 break;
233 if (!gsi_end_p (gsi))
234 expr = NULL_TREE;
235 break;
238 if (gimple_code (at) != GIMPLE_PHI)
240 expr = NULL_TREE;
241 break;
244 FOR_EACH_EDGE (e, ei, bb->preds)
245 if (e->src->aux)
246 break;
247 gcc_assert (e);
249 expr = PHI_ARG_DEF_FROM_EDGE (at, e);
250 if (TREE_CODE (expr) != SSA_NAME)
252 /* The value is a constant. */
253 break;
257 /* Unmark the blocks. */
258 for (bb = call_bb; bb != at_bb; bb = single_succ (bb))
259 bb->aux = NULL;
260 bb->aux = NULL;
262 return expr;
265 /* Simulates the effect of an assignment STMT on the return value of the tail
266 recursive CALL passed in ASS_VAR. M and A are the multiplicative and the
267 additive factor for the real return value. */
269 static bool
270 process_assignment (gimple stmt, gimple_stmt_iterator call, tree *m,
271 tree *a, tree *ass_var)
273 tree op0, op1 = NULL_TREE, non_ass_var = NULL_TREE;
274 tree dest = gimple_assign_lhs (stmt);
275 enum tree_code code = gimple_assign_rhs_code (stmt);
276 enum gimple_rhs_class rhs_class = get_gimple_rhs_class (code);
277 tree src_var = gimple_assign_rhs1 (stmt);
279 /* See if this is a simple copy operation of an SSA name to the function
280 result. In that case we may have a simple tail call. Ignore type
281 conversions that can never produce extra code between the function
282 call and the function return. */
283 if ((rhs_class == GIMPLE_SINGLE_RHS || gimple_assign_cast_p (stmt))
284 && (TREE_CODE (src_var) == SSA_NAME))
286 /* Reject a tailcall if the type conversion might need
287 additional code. */
288 if (gimple_assign_cast_p (stmt)
289 && TYPE_MODE (TREE_TYPE (dest)) != TYPE_MODE (TREE_TYPE (src_var)))
290 return false;
292 if (src_var != *ass_var)
293 return false;
295 *ass_var = dest;
296 return true;
299 switch (rhs_class)
301 case GIMPLE_BINARY_RHS:
302 op1 = gimple_assign_rhs2 (stmt);
304 /* Fall through. */
306 case GIMPLE_UNARY_RHS:
307 op0 = gimple_assign_rhs1 (stmt);
308 break;
310 default:
311 return false;
314 /* Accumulator optimizations will reverse the order of operations.
315 We can only do that for floating-point types if we're assuming
316 that addition and multiplication are associative. */
317 if (!flag_associative_math)
318 if (FLOAT_TYPE_P (TREE_TYPE (DECL_RESULT (current_function_decl))))
319 return false;
321 if (rhs_class == GIMPLE_UNARY_RHS)
323 else if (op0 == *ass_var
324 && (non_ass_var = independent_of_stmt_p (op1, stmt, call)))
326 else if (op1 == *ass_var
327 && (non_ass_var = independent_of_stmt_p (op0, stmt, call)))
329 else
330 return false;
332 switch (code)
334 case PLUS_EXPR:
335 *a = non_ass_var;
336 *ass_var = dest;
337 return true;
339 case POINTER_PLUS_EXPR:
340 if (op0 != *ass_var)
341 return false;
342 *a = non_ass_var;
343 *ass_var = dest;
344 return true;
346 case MULT_EXPR:
347 *m = non_ass_var;
348 *ass_var = dest;
349 return true;
351 case NEGATE_EXPR:
352 *m = build_minus_one_cst (TREE_TYPE (op0));
353 *ass_var = dest;
354 return true;
356 case MINUS_EXPR:
357 if (*ass_var == op0)
358 *a = fold_build1 (NEGATE_EXPR, TREE_TYPE (non_ass_var), non_ass_var);
359 else
361 *m = build_minus_one_cst (TREE_TYPE (non_ass_var));
362 *a = fold_build1 (NEGATE_EXPR, TREE_TYPE (non_ass_var), non_ass_var);
365 *ass_var = dest;
366 return true;
368 /* TODO -- Handle POINTER_PLUS_EXPR. */
370 default:
371 return false;
375 /* Propagate VAR through phis on edge E. */
377 static tree
378 propagate_through_phis (tree var, edge e)
380 basic_block dest = e->dest;
381 gimple_stmt_iterator gsi;
383 for (gsi = gsi_start_phis (dest); !gsi_end_p (gsi); gsi_next (&gsi))
385 gimple phi = gsi_stmt (gsi);
386 if (PHI_ARG_DEF_FROM_EDGE (phi, e) == var)
387 return PHI_RESULT (phi);
389 return var;
392 /* Finds tailcalls falling into basic block BB. The list of found tailcalls is
393 added to the start of RET. */
395 static void
396 find_tail_calls (basic_block bb, struct tailcall **ret)
398 tree ass_var = NULL_TREE, ret_var, func, param;
399 gimple stmt, call = NULL;
400 gimple_stmt_iterator gsi, agsi;
401 bool tail_recursion;
402 struct tailcall *nw;
403 edge e;
404 tree m, a;
405 basic_block abb;
406 size_t idx;
407 tree var;
409 if (!single_succ_p (bb))
410 return;
412 for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi); gsi_prev (&gsi))
414 stmt = gsi_stmt (gsi);
416 /* Ignore labels, returns, clobbers and debug stmts. */
417 if (gimple_code (stmt) == GIMPLE_LABEL
418 || gimple_code (stmt) == GIMPLE_RETURN
419 || gimple_clobber_p (stmt)
420 || is_gimple_debug (stmt))
421 continue;
423 /* Check for a call. */
424 if (is_gimple_call (stmt))
426 call = stmt;
427 ass_var = gimple_call_lhs (stmt);
428 break;
431 /* If the statement references memory or volatile operands, fail. */
432 if (gimple_references_memory_p (stmt)
433 || gimple_has_volatile_ops (stmt))
434 return;
437 if (gsi_end_p (gsi))
439 edge_iterator ei;
440 /* Recurse to the predecessors. */
441 FOR_EACH_EDGE (e, ei, bb->preds)
442 find_tail_calls (e->src, ret);
444 return;
447 /* If the LHS of our call is not just a simple register, we can't
448 transform this into a tail or sibling call. This situation happens,
449 in (e.g.) "*p = foo()" where foo returns a struct. In this case
450 we won't have a temporary here, but we need to carry out the side
451 effect anyway, so tailcall is impossible.
453 ??? In some situations (when the struct is returned in memory via
454 invisible argument) we could deal with this, e.g. by passing 'p'
455 itself as that argument to foo, but it's too early to do this here,
456 and expand_call() will not handle it anyway. If it ever can, then
457 we need to revisit this here, to allow that situation. */
458 if (ass_var && !is_gimple_reg (ass_var))
459 return;
461 /* We found the call, check whether it is suitable. */
462 tail_recursion = false;
463 func = gimple_call_fndecl (call);
464 if (func
465 && !DECL_BUILT_IN (func)
466 && recursive_call_p (current_function_decl, func))
468 tree arg;
470 for (param = DECL_ARGUMENTS (func), idx = 0;
471 param && idx < gimple_call_num_args (call);
472 param = DECL_CHAIN (param), idx ++)
474 arg = gimple_call_arg (call, idx);
475 if (param != arg)
477 /* Make sure there are no problems with copying. The parameter
478 have a copyable type and the two arguments must have reasonably
479 equivalent types. The latter requirement could be relaxed if
480 we emitted a suitable type conversion statement. */
481 if (!is_gimple_reg_type (TREE_TYPE (param))
482 || !useless_type_conversion_p (TREE_TYPE (param),
483 TREE_TYPE (arg)))
484 break;
486 /* The parameter should be a real operand, so that phi node
487 created for it at the start of the function has the meaning
488 of copying the value. This test implies is_gimple_reg_type
489 from the previous condition, however this one could be
490 relaxed by being more careful with copying the new value
491 of the parameter (emitting appropriate GIMPLE_ASSIGN and
492 updating the virtual operands). */
493 if (!is_gimple_reg (param))
494 break;
497 if (idx == gimple_call_num_args (call) && !param)
498 tail_recursion = true;
501 /* Make sure the tail invocation of this function does not refer
502 to local variables. */
503 FOR_EACH_LOCAL_DECL (cfun, idx, var)
505 if (TREE_CODE (var) != PARM_DECL
506 && auto_var_in_fn_p (var, cfun->decl)
507 && (ref_maybe_used_by_stmt_p (call, var)
508 || call_may_clobber_ref_p (call, var)))
509 return;
512 /* Now check the statements after the call. None of them has virtual
513 operands, so they may only depend on the call through its return
514 value. The return value should also be dependent on each of them,
515 since we are running after dce. */
516 m = NULL_TREE;
517 a = NULL_TREE;
519 abb = bb;
520 agsi = gsi;
521 while (1)
523 tree tmp_a = NULL_TREE;
524 tree tmp_m = NULL_TREE;
525 gsi_next (&agsi);
527 while (gsi_end_p (agsi))
529 ass_var = propagate_through_phis (ass_var, single_succ_edge (abb));
530 abb = single_succ (abb);
531 agsi = gsi_start_bb (abb);
534 stmt = gsi_stmt (agsi);
536 if (gimple_code (stmt) == GIMPLE_LABEL)
537 continue;
539 if (gimple_code (stmt) == GIMPLE_RETURN)
540 break;
542 if (gimple_clobber_p (stmt))
543 continue;
545 if (is_gimple_debug (stmt))
546 continue;
548 if (gimple_code (stmt) != GIMPLE_ASSIGN)
549 return;
551 /* This is a gimple assign. */
552 if (! process_assignment (stmt, gsi, &tmp_m, &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 (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 gimple_stmt_iterator gsi;
614 for (gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi); gsi_next (&gsi))
615 if (PHI_RESULT (gsi_stmt (gsi)) == var)
616 break;
618 gcc_assert (!gsi_end_p (gsi));
619 add_phi_arg (gsi_stmt (gsi), 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 gimple 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_with_ops (code, result, 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 gimple stmt;
672 tree var = copy_ssa_name (acc, NULL);
673 if (types_compatible_p (TREE_TYPE (acc), TREE_TYPE (op1)))
674 stmt = gimple_build_assign_with_ops (code, var, 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 gimple ret_stmt = 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 gimple phi;
812 gimple_stmt_iterator gsi;
813 gimple orig_stmt;
815 stmt = orig_stmt = gsi_stmt (t->call_gsi);
816 bb = gsi_bb (t->call_gsi);
818 if (dump_file && (dump_flags & TDF_DETAILS))
820 fprintf (dump_file, "Eliminated tail recursion in bb %d : ",
821 bb->index);
822 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
823 fprintf (dump_file, "\n");
826 gcc_assert (is_gimple_call (stmt));
828 first = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun));
830 /* Remove the code after call_gsi that will become unreachable. The
831 possibly unreachable code in other blocks is removed later in
832 cfg cleanup. */
833 gsi = t->call_gsi;
834 gsi_next (&gsi);
835 while (!gsi_end_p (gsi))
837 gimple t = gsi_stmt (gsi);
838 /* Do not remove the return statement, so that redirect_edge_and_branch
839 sees how the block ends. */
840 if (gimple_code (t) == GIMPLE_RETURN)
841 break;
843 gsi_remove (&gsi, true);
844 release_defs (t);
847 /* Number of executions of function has reduced by the tailcall. */
848 e = single_succ_edge (gsi_bb (t->call_gsi));
849 decrease_profile (EXIT_BLOCK_PTR_FOR_FN (cfun), e->count, EDGE_FREQUENCY (e));
850 decrease_profile (ENTRY_BLOCK_PTR_FOR_FN (cfun), e->count,
851 EDGE_FREQUENCY (e));
852 if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun))
853 decrease_profile (e->dest, e->count, EDGE_FREQUENCY (e));
855 /* Replace the call by a jump to the start of function. */
856 e = redirect_edge_and_branch (single_succ_edge (gsi_bb (t->call_gsi)),
857 first);
858 gcc_assert (e);
859 PENDING_STMT (e) = NULL;
861 /* Add phi node entries for arguments. The ordering of the phi nodes should
862 be the same as the ordering of the arguments. */
863 for (param = DECL_ARGUMENTS (current_function_decl),
864 idx = 0, gsi = gsi_start_phis (first);
865 param;
866 param = DECL_CHAIN (param), idx++)
868 if (!arg_needs_copy_p (param))
869 continue;
871 arg = gimple_call_arg (stmt, idx);
872 phi = gsi_stmt (gsi);
873 gcc_assert (param == SSA_NAME_VAR (PHI_RESULT (phi)));
875 add_phi_arg (phi, arg, e, gimple_location (stmt));
876 gsi_next (&gsi);
879 /* Update the values of accumulators. */
880 adjust_accumulator_values (t->call_gsi, t->mult, t->add, e);
882 call = gsi_stmt (t->call_gsi);
883 rslt = gimple_call_lhs (call);
884 if (rslt != NULL_TREE)
886 /* Result of the call will no longer be defined. So adjust the
887 SSA_NAME_DEF_STMT accordingly. */
888 SSA_NAME_DEF_STMT (rslt) = gimple_build_nop ();
891 gsi_remove (&t->call_gsi, true);
892 release_defs (call);
895 /* Optimizes the tailcall described by T. If OPT_TAILCALLS is true, also
896 mark the tailcalls for the sibcall optimization. */
898 static bool
899 optimize_tail_call (struct tailcall *t, bool opt_tailcalls)
901 if (t->tail_recursion)
903 eliminate_tail_call (t);
904 return true;
907 if (opt_tailcalls)
909 gimple stmt = gsi_stmt (t->call_gsi);
911 gimple_call_set_tail (stmt, true);
912 if (dump_file && (dump_flags & TDF_DETAILS))
914 fprintf (dump_file, "Found tail call ");
915 print_gimple_stmt (dump_file, stmt, 0, dump_flags);
916 fprintf (dump_file, " in bb %i\n", (gsi_bb (t->call_gsi))->index);
920 return false;
923 /* Creates a tail-call accumulator of the same type as the return type of the
924 current function. LABEL is the name used to creating the temporary
925 variable for the accumulator. The accumulator will be inserted in the
926 phis of a basic block BB with single predecessor with an initial value
927 INIT converted to the current function return type. */
929 static tree
930 create_tailcall_accumulator (const char *label, basic_block bb, tree init)
932 tree ret_type = TREE_TYPE (DECL_RESULT (current_function_decl));
933 if (POINTER_TYPE_P (ret_type))
934 ret_type = sizetype;
936 tree tmp = make_temp_ssa_name (ret_type, NULL, label);
937 gimple phi;
939 phi = create_phi_node (tmp, bb);
940 /* RET_TYPE can be a float when -ffast-maths is enabled. */
941 add_phi_arg (phi, fold_convert (ret_type, init), single_pred_edge (bb),
942 UNKNOWN_LOCATION);
943 return PHI_RESULT (phi);
946 /* Optimizes tail calls in the function, turning the tail recursion
947 into iteration. */
949 static unsigned int
950 tree_optimize_tail_calls_1 (bool opt_tailcalls)
952 edge e;
953 bool phis_constructed = false;
954 struct tailcall *tailcalls = NULL, *act, *next;
955 bool changed = false;
956 basic_block first = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun));
957 tree param;
958 gimple stmt;
959 edge_iterator ei;
961 if (!suitable_for_tail_opt_p ())
962 return 0;
963 if (opt_tailcalls)
964 opt_tailcalls = suitable_for_tail_call_opt_p ();
966 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds)
968 /* Only traverse the normal exits, i.e. those that end with return
969 statement. */
970 stmt = last_stmt (e->src);
972 if (stmt
973 && gimple_code (stmt) == GIMPLE_RETURN)
974 find_tail_calls (e->src, &tailcalls);
977 /* Construct the phi nodes and accumulators if necessary. */
978 a_acc = m_acc = NULL_TREE;
979 for (act = tailcalls; act; act = act->next)
981 if (!act->tail_recursion)
982 continue;
984 if (!phis_constructed)
986 /* Ensure that there is only one predecessor of the block
987 or if there are existing degenerate PHI nodes. */
988 if (!single_pred_p (first)
989 || !gimple_seq_empty_p (phi_nodes (first)))
990 first =
991 split_edge (single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun)));
993 /* Copy the args if needed. */
994 for (param = DECL_ARGUMENTS (current_function_decl);
995 param;
996 param = DECL_CHAIN (param))
997 if (arg_needs_copy_p (param))
999 tree name = ssa_default_def (cfun, param);
1000 tree new_name = make_ssa_name (param, SSA_NAME_DEF_STMT (name));
1001 gimple phi;
1003 set_ssa_default_def (cfun, param, new_name);
1004 phi = create_phi_node (name, first);
1005 add_phi_arg (phi, new_name, single_pred_edge (first),
1006 EXPR_LOCATION (param));
1008 phis_constructed = true;
1011 if (act->add && !a_acc)
1012 a_acc = create_tailcall_accumulator ("add_acc", first,
1013 integer_zero_node);
1015 if (act->mult && !m_acc)
1016 m_acc = create_tailcall_accumulator ("mult_acc", first,
1017 integer_one_node);
1020 if (a_acc || m_acc)
1022 /* When the tail call elimination using accumulators is performed,
1023 statements adding the accumulated value are inserted at all exits.
1024 This turns all other tail calls to non-tail ones. */
1025 opt_tailcalls = false;
1028 for (; tailcalls; tailcalls = next)
1030 next = tailcalls->next;
1031 changed |= optimize_tail_call (tailcalls, opt_tailcalls);
1032 free (tailcalls);
1035 if (a_acc || m_acc)
1037 /* Modify the remaining return statements. */
1038 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds)
1040 stmt = last_stmt (e->src);
1042 if (stmt
1043 && gimple_code (stmt) == GIMPLE_RETURN)
1044 adjust_return_value (e->src, m_acc, a_acc);
1048 if (changed)
1050 /* We may have created new loops. Make them magically appear. */
1051 if (current_loops)
1052 loops_state_set (LOOPS_NEED_FIXUP);
1053 free_dominance_info (CDI_DOMINATORS);
1056 /* Add phi nodes for the virtual operands defined in the function to the
1057 header of the loop created by tail recursion elimination. Do so
1058 by triggering the SSA renamer. */
1059 if (phis_constructed)
1060 mark_virtual_operands_for_renaming (cfun);
1062 if (changed)
1063 return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals;
1064 return 0;
1067 static unsigned int
1068 execute_tail_recursion (void)
1070 return tree_optimize_tail_calls_1 (false);
1073 static bool
1074 gate_tail_calls (void)
1076 return flag_optimize_sibling_calls != 0 && dbg_cnt (tail_call);
1079 static unsigned int
1080 execute_tail_calls (void)
1082 return tree_optimize_tail_calls_1 (true);
1085 namespace {
1087 const pass_data pass_data_tail_recursion =
1089 GIMPLE_PASS, /* type */
1090 "tailr", /* name */
1091 OPTGROUP_NONE, /* optinfo_flags */
1092 true, /* has_gate */
1093 true, /* has_execute */
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 */
1102 class pass_tail_recursion : public gimple_opt_pass
1104 public:
1105 pass_tail_recursion (gcc::context *ctxt)
1106 : gimple_opt_pass (pass_data_tail_recursion, ctxt)
1109 /* opt_pass methods: */
1110 opt_pass * clone () { return new pass_tail_recursion (m_ctxt); }
1111 bool gate () { return gate_tail_calls (); }
1112 unsigned int execute () { return execute_tail_recursion (); }
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 true, /* has_gate */
1132 true, /* has_execute */
1133 TV_NONE, /* tv_id */
1134 ( PROP_cfg | PROP_ssa ), /* properties_required */
1135 0, /* properties_provided */
1136 0, /* properties_destroyed */
1137 0, /* todo_flags_start */
1138 TODO_verify_ssa, /* todo_flags_finish */
1141 class pass_tail_calls : public gimple_opt_pass
1143 public:
1144 pass_tail_calls (gcc::context *ctxt)
1145 : gimple_opt_pass (pass_data_tail_calls, ctxt)
1148 /* opt_pass methods: */
1149 bool gate () { return gate_tail_calls (); }
1150 unsigned int execute () { return execute_tail_calls (); }
1152 }; // class pass_tail_calls
1154 } // anon namespace
1156 gimple_opt_pass *
1157 make_pass_tail_calls (gcc::context *ctxt)
1159 return new pass_tail_calls (ctxt);