Fixups after merge
[official-gcc.git] / gcc / tree-tailcall.c
blob9e6e0055b2728d3404caa1b971faeaecb4efeb8c
1 /* Tail call optimization on trees.
2 Copyright (C) 2003-2014 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify
7 it under the terms of the GNU General Public License as published by
8 the Free Software Foundation; either version 3, or (at your option)
9 any later version.
11 GCC is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 GNU General Public License for more details.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "tm.h"
24 #include "tree.h"
25 #include "stor-layout.h"
26 #include "tm_p.h"
27 #include "predict.h"
28 #include "vec.h"
29 #include "hashtab.h"
30 #include "hash-set.h"
31 #include "machmode.h"
32 #include "hard-reg-set.h"
33 #include "input.h"
34 #include "function.h"
35 #include "dominance.h"
36 #include "cfg.h"
37 #include "basic-block.h"
38 #include "tree-ssa-alias.h"
39 #include "internal-fn.h"
40 #include "gimple-expr.h"
41 #include "is-a.h"
42 #include "gimple.h"
43 #include "gimple-iterator.h"
44 #include "gimplify-me.h"
45 #include "gimple-ssa.h"
46 #include "tree-cfg.h"
47 #include "tree-phinodes.h"
48 #include "stringpool.h"
49 #include "tree-ssanames.h"
50 #include "tree-into-ssa.h"
51 #include "expr.h"
52 #include "tree-dfa.h"
53 #include "gimple-pretty-print.h"
54 #include "except.h"
55 #include "tree-pass.h"
56 #include "flags.h"
57 #include "langhooks.h"
58 #include "dbgcnt.h"
59 #include "target.h"
60 #include "cfgloop.h"
61 #include "common/common-target.h"
62 #include "hash-map.h"
63 #include "plugin-api.h"
64 #include "ipa-ref.h"
65 #include "cgraph.h"
66 #include "ipa-utils.h"
68 /* The file implements the tail recursion elimination. It is also used to
69 analyze the tail calls in general, passing the results to the rtl level
70 where they are used for sibcall optimization.
72 In addition to the standard tail recursion elimination, we handle the most
73 trivial cases of making the call tail recursive by creating accumulators.
74 For example the following function
76 int sum (int n)
78 if (n > 0)
79 return n + sum (n - 1);
80 else
81 return 0;
84 is transformed into
86 int sum (int n)
88 int acc = 0;
90 while (n > 0)
91 acc += n--;
93 return acc;
96 To do this, we maintain two accumulators (a_acc and m_acc) that indicate
97 when we reach the return x statement, we should return a_acc + x * m_acc
98 instead. They are initially initialized to 0 and 1, respectively,
99 so the semantics of the function is obviously preserved. If we are
100 guaranteed that the value of the accumulator never change, we
101 omit the accumulator.
103 There are three cases how the function may exit. The first one is
104 handled in adjust_return_value, the other two in adjust_accumulator_values
105 (the second case is actually a special case of the third one and we
106 present it separately just for clarity):
108 1) Just return x, where x is not in any of the remaining special shapes.
109 We rewrite this to a gimple equivalent of return m_acc * x + a_acc.
111 2) return f (...), where f is the current function, is rewritten in a
112 classical tail-recursion elimination way, into assignment of arguments
113 and jump to the start of the function. Values of the accumulators
114 are unchanged.
116 3) return a + m * f(...), where a and m do not depend on call to f.
117 To preserve the semantics described before we want this to be rewritten
118 in such a way that we finally return
120 a_acc + (a + m * f(...)) * m_acc = (a_acc + a * m_acc) + (m * m_acc) * f(...).
122 I.e. we increase a_acc by a * m_acc, multiply m_acc by m and
123 eliminate the tail call to f. Special cases when the value is just
124 added or just multiplied are obtained by setting a = 0 or m = 1.
126 TODO -- it is possible to do similar tricks for other operations. */
128 /* A structure that describes the tailcall. */
130 struct tailcall
132 /* The iterator pointing to the call statement. */
133 gimple_stmt_iterator call_gsi;
135 /* True if it is a call to the current function. */
136 bool tail_recursion;
138 /* The return value of the caller is mult * f + add, where f is the return
139 value of the call. */
140 tree mult, add;
142 /* Next tailcall in the chain. */
143 struct tailcall *next;
146 /* The variables holding the value of multiplicative and additive
147 accumulator. */
148 static tree m_acc, a_acc;
150 static bool suitable_for_tail_opt_p (void);
151 static bool optimize_tail_call (struct tailcall *, bool);
152 static void eliminate_tail_call (struct tailcall *);
153 static void find_tail_calls (basic_block, struct tailcall **);
155 /* Returns false when the function is not suitable for tail call optimization
156 from some reason (e.g. if it takes variable number of arguments). */
158 static bool
159 suitable_for_tail_opt_p (void)
161 if (cfun->stdarg)
162 return false;
164 return true;
166 /* Returns false when the function is not suitable for tail call optimization
167 from some reason (e.g. if it takes variable number of arguments).
168 This test must pass in addition to suitable_for_tail_opt_p in order to make
169 tail call discovery happen. */
171 static bool
172 suitable_for_tail_call_opt_p (void)
174 tree param;
176 /* alloca (until we have stack slot life analysis) inhibits
177 sibling call optimizations, but not tail recursion. */
178 if (cfun->calls_alloca)
179 return false;
181 /* If we are using sjlj exceptions, we may need to add a call to
182 _Unwind_SjLj_Unregister at exit of the function. Which means
183 that we cannot do any sibcall transformations. */
184 if (targetm_common.except_unwind_info (&global_options) == UI_SJLJ
185 && current_function_has_exception_handlers ())
186 return false;
188 /* Any function that calls setjmp might have longjmp called from
189 any called function. ??? We really should represent this
190 properly in the CFG so that this needn't be special cased. */
191 if (cfun->calls_setjmp)
192 return false;
194 /* ??? It is OK if the argument of a function is taken in some cases,
195 but not in all cases. See PR15387 and PR19616. Revisit for 4.1. */
196 for (param = DECL_ARGUMENTS (current_function_decl);
197 param;
198 param = DECL_CHAIN (param))
199 if (TREE_ADDRESSABLE (param))
200 return false;
202 return true;
205 /* Checks whether the expression EXPR in stmt AT is independent of the
206 statement pointed to by GSI (in a sense that we already know EXPR's value
207 at GSI). We use the fact that we are only called from the chain of
208 basic blocks that have only single successor. Returns the expression
209 containing the value of EXPR at GSI. */
211 static tree
212 independent_of_stmt_p (tree expr, gimple at, gimple_stmt_iterator gsi)
214 basic_block bb, call_bb, at_bb;
215 edge e;
216 edge_iterator ei;
218 if (is_gimple_min_invariant (expr))
219 return expr;
221 if (TREE_CODE (expr) != SSA_NAME)
222 return NULL_TREE;
224 /* Mark the blocks in the chain leading to the end. */
225 at_bb = gimple_bb (at);
226 call_bb = gimple_bb (gsi_stmt (gsi));
227 for (bb = call_bb; bb != at_bb; bb = single_succ (bb))
228 bb->aux = &bb->aux;
229 bb->aux = &bb->aux;
231 while (1)
233 at = SSA_NAME_DEF_STMT (expr);
234 bb = gimple_bb (at);
236 /* The default definition or defined before the chain. */
237 if (!bb || !bb->aux)
238 break;
240 if (bb == call_bb)
242 for (; !gsi_end_p (gsi); gsi_next (&gsi))
243 if (gsi_stmt (gsi) == at)
244 break;
246 if (!gsi_end_p (gsi))
247 expr = NULL_TREE;
248 break;
251 if (gimple_code (at) != GIMPLE_PHI)
253 expr = NULL_TREE;
254 break;
257 FOR_EACH_EDGE (e, ei, bb->preds)
258 if (e->src->aux)
259 break;
260 gcc_assert (e);
262 expr = PHI_ARG_DEF_FROM_EDGE (at, e);
263 if (TREE_CODE (expr) != SSA_NAME)
265 /* The value is a constant. */
266 break;
270 /* Unmark the blocks. */
271 for (bb = call_bb; bb != at_bb; bb = single_succ (bb))
272 bb->aux = NULL;
273 bb->aux = NULL;
275 return expr;
278 /* Simulates the effect of an assignment STMT on the return value of the tail
279 recursive CALL passed in ASS_VAR. M and A are the multiplicative and the
280 additive factor for the real return value. */
282 static bool
283 process_assignment (gassign *stmt, gimple_stmt_iterator call, tree *m,
284 tree *a, tree *ass_var)
286 tree op0, op1 = NULL_TREE, non_ass_var = NULL_TREE;
287 tree dest = gimple_assign_lhs (stmt);
288 enum tree_code code = gimple_assign_rhs_code (stmt);
289 enum gimple_rhs_class rhs_class = get_gimple_rhs_class (code);
290 tree src_var = gimple_assign_rhs1 (stmt);
292 /* See if this is a simple copy operation of an SSA name to the function
293 result. In that case we may have a simple tail call. Ignore type
294 conversions that can never produce extra code between the function
295 call and the function return. */
296 if ((rhs_class == GIMPLE_SINGLE_RHS || gimple_assign_cast_p (stmt))
297 && (TREE_CODE (src_var) == SSA_NAME))
299 /* Reject a tailcall if the type conversion might need
300 additional code. */
301 if (gimple_assign_cast_p (stmt))
303 if (TYPE_MODE (TREE_TYPE (dest)) != TYPE_MODE (TREE_TYPE (src_var)))
304 return false;
306 /* Even if the type modes are the same, if the precision of the
307 type is smaller than mode's precision,
308 reduce_to_bit_field_precision would generate additional code. */
309 if (INTEGRAL_TYPE_P (TREE_TYPE (dest))
310 && (GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (dest)))
311 > TYPE_PRECISION (TREE_TYPE (dest))))
312 return false;
315 if (src_var != *ass_var)
316 return false;
318 *ass_var = dest;
319 return true;
322 switch (rhs_class)
324 case GIMPLE_BINARY_RHS:
325 op1 = gimple_assign_rhs2 (stmt);
327 /* Fall through. */
329 case GIMPLE_UNARY_RHS:
330 op0 = gimple_assign_rhs1 (stmt);
331 break;
333 default:
334 return false;
337 /* Accumulator optimizations will reverse the order of operations.
338 We can only do that for floating-point types if we're assuming
339 that addition and multiplication are associative. */
340 if (!flag_associative_math)
341 if (FLOAT_TYPE_P (TREE_TYPE (DECL_RESULT (current_function_decl))))
342 return false;
344 if (rhs_class == GIMPLE_UNARY_RHS)
346 else if (op0 == *ass_var
347 && (non_ass_var = independent_of_stmt_p (op1, stmt, call)))
349 else if (op1 == *ass_var
350 && (non_ass_var = independent_of_stmt_p (op0, stmt, call)))
352 else
353 return false;
355 switch (code)
357 case PLUS_EXPR:
358 *a = non_ass_var;
359 *ass_var = dest;
360 return true;
362 case POINTER_PLUS_EXPR:
363 if (op0 != *ass_var)
364 return false;
365 *a = non_ass_var;
366 *ass_var = dest;
367 return true;
369 case MULT_EXPR:
370 *m = non_ass_var;
371 *ass_var = dest;
372 return true;
374 case NEGATE_EXPR:
375 *m = build_minus_one_cst (TREE_TYPE (op0));
376 *ass_var = dest;
377 return true;
379 case MINUS_EXPR:
380 if (*ass_var == op0)
381 *a = fold_build1 (NEGATE_EXPR, TREE_TYPE (non_ass_var), non_ass_var);
382 else
384 *m = build_minus_one_cst (TREE_TYPE (non_ass_var));
385 *a = fold_build1 (NEGATE_EXPR, TREE_TYPE (non_ass_var), non_ass_var);
388 *ass_var = dest;
389 return true;
391 /* TODO -- Handle POINTER_PLUS_EXPR. */
393 default:
394 return false;
398 /* Propagate VAR through phis on edge E. */
400 static tree
401 propagate_through_phis (tree var, edge e)
403 basic_block dest = e->dest;
404 gphi_iterator gsi;
406 for (gsi = gsi_start_phis (dest); !gsi_end_p (gsi); gsi_next (&gsi))
408 gphi *phi = gsi.phi ();
409 if (PHI_ARG_DEF_FROM_EDGE (phi, e) == var)
410 return PHI_RESULT (phi);
412 return var;
415 /* Finds tailcalls falling into basic block BB. The list of found tailcalls is
416 added to the start of RET. */
418 static void
419 find_tail_calls (basic_block bb, struct tailcall **ret)
421 tree ass_var = NULL_TREE, ret_var, func, param;
422 gimple stmt;
423 gcall *call = NULL;
424 gimple_stmt_iterator gsi, agsi;
425 bool tail_recursion;
426 struct tailcall *nw;
427 edge e;
428 tree m, a;
429 basic_block abb;
430 size_t idx;
431 tree var;
433 if (!single_succ_p (bb))
434 return;
436 for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi); gsi_prev (&gsi))
438 stmt = gsi_stmt (gsi);
440 /* Ignore labels, returns, clobbers and debug stmts. */
441 if (gimple_code (stmt) == GIMPLE_LABEL
442 || gimple_code (stmt) == GIMPLE_RETURN
443 || gimple_clobber_p (stmt)
444 || is_gimple_debug (stmt))
445 continue;
447 /* Check for a call. */
448 if (is_gimple_call (stmt))
450 call = as_a <gcall *> (stmt);
451 ass_var = gimple_call_lhs (call);
452 break;
455 /* If the statement references memory or volatile operands, fail. */
456 if (gimple_references_memory_p (stmt)
457 || gimple_has_volatile_ops (stmt))
458 return;
461 if (gsi_end_p (gsi))
463 edge_iterator ei;
464 /* Recurse to the predecessors. */
465 FOR_EACH_EDGE (e, ei, bb->preds)
466 find_tail_calls (e->src, ret);
468 return;
471 /* If the LHS of our call is not just a simple register, we can't
472 transform this into a tail or sibling call. This situation happens,
473 in (e.g.) "*p = foo()" where foo returns a struct. In this case
474 we won't have a temporary here, but we need to carry out the side
475 effect anyway, so tailcall is impossible.
477 ??? In some situations (when the struct is returned in memory via
478 invisible argument) we could deal with this, e.g. by passing 'p'
479 itself as that argument to foo, but it's too early to do this here,
480 and expand_call() will not handle it anyway. If it ever can, then
481 we need to revisit this here, to allow that situation. */
482 if (ass_var && !is_gimple_reg (ass_var))
483 return;
485 /* We found the call, check whether it is suitable. */
486 tail_recursion = false;
487 func = gimple_call_fndecl (call);
488 if (func
489 && !DECL_BUILT_IN (func)
490 && recursive_call_p (current_function_decl, func))
492 tree arg;
494 for (param = DECL_ARGUMENTS (func), idx = 0;
495 param && idx < gimple_call_num_args (call);
496 param = DECL_CHAIN (param), idx ++)
498 arg = gimple_call_arg (call, idx);
499 if (param != arg)
501 /* Make sure there are no problems with copying. The parameter
502 have a copyable type and the two arguments must have reasonably
503 equivalent types. The latter requirement could be relaxed if
504 we emitted a suitable type conversion statement. */
505 if (!is_gimple_reg_type (TREE_TYPE (param))
506 || !useless_type_conversion_p (TREE_TYPE (param),
507 TREE_TYPE (arg)))
508 break;
510 /* The parameter should be a real operand, so that phi node
511 created for it at the start of the function has the meaning
512 of copying the value. This test implies is_gimple_reg_type
513 from the previous condition, however this one could be
514 relaxed by being more careful with copying the new value
515 of the parameter (emitting appropriate GIMPLE_ASSIGN and
516 updating the virtual operands). */
517 if (!is_gimple_reg (param))
518 break;
521 if (idx == gimple_call_num_args (call) && !param)
522 tail_recursion = true;
525 /* Make sure the tail invocation of this function does not refer
526 to local variables. */
527 FOR_EACH_LOCAL_DECL (cfun, idx, var)
529 if (TREE_CODE (var) != PARM_DECL
530 && auto_var_in_fn_p (var, cfun->decl)
531 && (ref_maybe_used_by_stmt_p (call, var)
532 || call_may_clobber_ref_p (call, var)))
533 return;
536 /* Now check the statements after the call. None of them has virtual
537 operands, so they may only depend on the call through its return
538 value. The return value should also be dependent on each of them,
539 since we are running after dce. */
540 m = NULL_TREE;
541 a = NULL_TREE;
543 abb = bb;
544 agsi = gsi;
545 while (1)
547 tree tmp_a = NULL_TREE;
548 tree tmp_m = NULL_TREE;
549 gsi_next (&agsi);
551 while (gsi_end_p (agsi))
553 ass_var = propagate_through_phis (ass_var, single_succ_edge (abb));
554 abb = single_succ (abb);
555 agsi = gsi_start_bb (abb);
558 stmt = gsi_stmt (agsi);
560 if (gimple_code (stmt) == GIMPLE_LABEL)
561 continue;
563 if (gimple_code (stmt) == GIMPLE_RETURN)
564 break;
566 if (gimple_clobber_p (stmt))
567 continue;
569 if (is_gimple_debug (stmt))
570 continue;
572 if (gimple_code (stmt) != GIMPLE_ASSIGN)
573 return;
575 /* This is a gimple assign. */
576 if (! process_assignment (as_a <gassign *> (stmt), gsi, &tmp_m,
577 &tmp_a, &ass_var))
578 return;
580 if (tmp_a)
582 tree type = TREE_TYPE (tmp_a);
583 if (a)
584 a = fold_build2 (PLUS_EXPR, type, fold_convert (type, a), tmp_a);
585 else
586 a = tmp_a;
588 if (tmp_m)
590 tree type = TREE_TYPE (tmp_m);
591 if (m)
592 m = fold_build2 (MULT_EXPR, type, fold_convert (type, m), tmp_m);
593 else
594 m = tmp_m;
596 if (a)
597 a = fold_build2 (MULT_EXPR, type, fold_convert (type, a), tmp_m);
601 /* See if this is a tail call we can handle. */
602 ret_var = gimple_return_retval (as_a <greturn *> (stmt));
604 /* We may proceed if there either is no return value, or the return value
605 is identical to the call's return. */
606 if (ret_var
607 && (ret_var != ass_var))
608 return;
610 /* If this is not a tail recursive call, we cannot handle addends or
611 multiplicands. */
612 if (!tail_recursion && (m || a))
613 return;
615 /* For pointers only allow additions. */
616 if (m && POINTER_TYPE_P (TREE_TYPE (DECL_RESULT (current_function_decl))))
617 return;
619 nw = XNEW (struct tailcall);
621 nw->call_gsi = gsi;
623 nw->tail_recursion = tail_recursion;
625 nw->mult = m;
626 nw->add = a;
628 nw->next = *ret;
629 *ret = nw;
632 /* Helper to insert PHI_ARGH to the phi of VAR in the destination of edge E. */
634 static void
635 add_successor_phi_arg (edge e, tree var, tree phi_arg)
637 gphi_iterator gsi;
639 for (gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi); gsi_next (&gsi))
640 if (PHI_RESULT (gsi.phi ()) == var)
641 break;
643 gcc_assert (!gsi_end_p (gsi));
644 add_phi_arg (gsi.phi (), phi_arg, e, UNKNOWN_LOCATION);
647 /* Creates a GIMPLE statement which computes the operation specified by
648 CODE, ACC and OP1 to a new variable with name LABEL and inserts the
649 statement in the position specified by GSI. Returns the
650 tree node of the statement's result. */
652 static tree
653 adjust_return_value_with_ops (enum tree_code code, const char *label,
654 tree acc, tree op1, gimple_stmt_iterator gsi)
657 tree ret_type = TREE_TYPE (DECL_RESULT (current_function_decl));
658 tree result = make_temp_ssa_name (ret_type, NULL, label);
659 gassign *stmt;
661 if (POINTER_TYPE_P (ret_type))
663 gcc_assert (code == PLUS_EXPR && TREE_TYPE (acc) == sizetype);
664 code = POINTER_PLUS_EXPR;
666 if (types_compatible_p (TREE_TYPE (acc), TREE_TYPE (op1))
667 && code != POINTER_PLUS_EXPR)
668 stmt = gimple_build_assign_with_ops (code, result, acc, op1);
669 else
671 tree tem;
672 if (code == POINTER_PLUS_EXPR)
673 tem = fold_build2 (code, TREE_TYPE (op1), op1, acc);
674 else
675 tem = fold_build2 (code, TREE_TYPE (op1),
676 fold_convert (TREE_TYPE (op1), acc), op1);
677 tree rhs = fold_convert (ret_type, tem);
678 rhs = force_gimple_operand_gsi (&gsi, rhs,
679 false, NULL, true, GSI_SAME_STMT);
680 stmt = gimple_build_assign (result, rhs);
683 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
684 return result;
687 /* Creates a new GIMPLE statement that adjusts the value of accumulator ACC by
688 the computation specified by CODE and OP1 and insert the statement
689 at the position specified by GSI as a new statement. Returns new SSA name
690 of updated accumulator. */
692 static tree
693 update_accumulator_with_ops (enum tree_code code, tree acc, tree op1,
694 gimple_stmt_iterator gsi)
696 gassign *stmt;
697 tree var = copy_ssa_name (acc, NULL);
698 if (types_compatible_p (TREE_TYPE (acc), TREE_TYPE (op1)))
699 stmt = gimple_build_assign_with_ops (code, var, acc, op1);
700 else
702 tree rhs = fold_convert (TREE_TYPE (acc),
703 fold_build2 (code,
704 TREE_TYPE (op1),
705 fold_convert (TREE_TYPE (op1), acc),
706 op1));
707 rhs = force_gimple_operand_gsi (&gsi, rhs,
708 false, NULL, false, GSI_CONTINUE_LINKING);
709 stmt = gimple_build_assign (var, rhs);
711 gsi_insert_after (&gsi, stmt, GSI_NEW_STMT);
712 return var;
715 /* Adjust the accumulator values according to A and M after GSI, and update
716 the phi nodes on edge BACK. */
718 static void
719 adjust_accumulator_values (gimple_stmt_iterator gsi, tree m, tree a, edge back)
721 tree var, a_acc_arg, m_acc_arg;
723 if (m)
724 m = force_gimple_operand_gsi (&gsi, m, true, NULL, true, GSI_SAME_STMT);
725 if (a)
726 a = force_gimple_operand_gsi (&gsi, a, true, NULL, true, GSI_SAME_STMT);
728 a_acc_arg = a_acc;
729 m_acc_arg = m_acc;
730 if (a)
732 if (m_acc)
734 if (integer_onep (a))
735 var = m_acc;
736 else
737 var = adjust_return_value_with_ops (MULT_EXPR, "acc_tmp", m_acc,
738 a, gsi);
740 else
741 var = a;
743 a_acc_arg = update_accumulator_with_ops (PLUS_EXPR, a_acc, var, gsi);
746 if (m)
747 m_acc_arg = update_accumulator_with_ops (MULT_EXPR, m_acc, m, gsi);
749 if (a_acc)
750 add_successor_phi_arg (back, a_acc, a_acc_arg);
752 if (m_acc)
753 add_successor_phi_arg (back, m_acc, m_acc_arg);
756 /* Adjust value of the return at the end of BB according to M and A
757 accumulators. */
759 static void
760 adjust_return_value (basic_block bb, tree m, tree a)
762 tree retval;
763 greturn *ret_stmt =
764 as_a <greturn *> (gimple_seq_last_stmt (bb_seq (bb)));
765 gimple_stmt_iterator gsi = gsi_last_bb (bb);
767 gcc_assert (gimple_code (ret_stmt) == GIMPLE_RETURN);
769 retval = gimple_return_retval (ret_stmt);
770 if (!retval || retval == error_mark_node)
771 return;
773 if (m)
774 retval = adjust_return_value_with_ops (MULT_EXPR, "mul_tmp", m_acc, retval,
775 gsi);
776 if (a)
777 retval = adjust_return_value_with_ops (PLUS_EXPR, "acc_tmp", a_acc, retval,
778 gsi);
779 gimple_return_set_retval (ret_stmt, retval);
780 update_stmt (ret_stmt);
783 /* Subtract COUNT and FREQUENCY from the basic block and it's
784 outgoing edge. */
785 static void
786 decrease_profile (basic_block bb, gcov_type count, int frequency)
788 edge e;
789 bb->count -= count;
790 if (bb->count < 0)
791 bb->count = 0;
792 bb->frequency -= frequency;
793 if (bb->frequency < 0)
794 bb->frequency = 0;
795 if (!single_succ_p (bb))
797 gcc_assert (!EDGE_COUNT (bb->succs));
798 return;
800 e = single_succ_edge (bb);
801 e->count -= count;
802 if (e->count < 0)
803 e->count = 0;
806 /* Returns true if argument PARAM of the tail recursive call needs to be copied
807 when the call is eliminated. */
809 static bool
810 arg_needs_copy_p (tree param)
812 tree def;
814 if (!is_gimple_reg (param))
815 return false;
817 /* Parameters that are only defined but never used need not be copied. */
818 def = ssa_default_def (cfun, param);
819 if (!def)
820 return false;
822 return true;
825 /* Eliminates tail call described by T. TMP_VARS is a list of
826 temporary variables used to copy the function arguments. */
828 static void
829 eliminate_tail_call (struct tailcall *t)
831 tree param, rslt;
832 gimple stmt, call;
833 tree arg;
834 size_t idx;
835 basic_block bb, first;
836 edge e;
837 gphi *phi;
838 gphi_iterator gpi;
839 gimple_stmt_iterator gsi;
840 gimple orig_stmt;
842 stmt = orig_stmt = gsi_stmt (t->call_gsi);
843 bb = gsi_bb (t->call_gsi);
845 if (dump_file && (dump_flags & TDF_DETAILS))
847 fprintf (dump_file, "Eliminated tail recursion in bb %d : ",
848 bb->index);
849 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
850 fprintf (dump_file, "\n");
853 gcc_assert (is_gimple_call (stmt));
855 first = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun));
857 /* Remove the code after call_gsi that will become unreachable. The
858 possibly unreachable code in other blocks is removed later in
859 cfg cleanup. */
860 gsi = t->call_gsi;
861 gsi_next (&gsi);
862 while (!gsi_end_p (gsi))
864 gimple t = gsi_stmt (gsi);
865 /* Do not remove the return statement, so that redirect_edge_and_branch
866 sees how the block ends. */
867 if (gimple_code (t) == GIMPLE_RETURN)
868 break;
870 gsi_remove (&gsi, true);
871 release_defs (t);
874 /* Number of executions of function has reduced by the tailcall. */
875 e = single_succ_edge (gsi_bb (t->call_gsi));
876 decrease_profile (EXIT_BLOCK_PTR_FOR_FN (cfun), e->count, EDGE_FREQUENCY (e));
877 decrease_profile (ENTRY_BLOCK_PTR_FOR_FN (cfun), e->count,
878 EDGE_FREQUENCY (e));
879 if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun))
880 decrease_profile (e->dest, e->count, EDGE_FREQUENCY (e));
882 /* Replace the call by a jump to the start of function. */
883 e = redirect_edge_and_branch (single_succ_edge (gsi_bb (t->call_gsi)),
884 first);
885 gcc_assert (e);
886 PENDING_STMT (e) = NULL;
888 /* Add phi node entries for arguments. The ordering of the phi nodes should
889 be the same as the ordering of the arguments. */
890 for (param = DECL_ARGUMENTS (current_function_decl),
891 idx = 0, gpi = gsi_start_phis (first);
892 param;
893 param = DECL_CHAIN (param), idx++)
895 if (!arg_needs_copy_p (param))
896 continue;
898 arg = gimple_call_arg (stmt, idx);
899 phi = gpi.phi ();
900 gcc_assert (param == SSA_NAME_VAR (PHI_RESULT (phi)));
902 add_phi_arg (phi, arg, e, gimple_location (stmt));
903 gsi_next (&gpi);
906 /* Update the values of accumulators. */
907 adjust_accumulator_values (t->call_gsi, t->mult, t->add, e);
909 call = gsi_stmt (t->call_gsi);
910 rslt = gimple_call_lhs (call);
911 if (rslt != NULL_TREE)
913 /* Result of the call will no longer be defined. So adjust the
914 SSA_NAME_DEF_STMT accordingly. */
915 SSA_NAME_DEF_STMT (rslt) = gimple_build_nop ();
918 gsi_remove (&t->call_gsi, true);
919 release_defs (call);
922 /* Optimizes the tailcall described by T. If OPT_TAILCALLS is true, also
923 mark the tailcalls for the sibcall optimization. */
925 static bool
926 optimize_tail_call (struct tailcall *t, bool opt_tailcalls)
928 if (t->tail_recursion)
930 eliminate_tail_call (t);
931 return true;
934 if (opt_tailcalls)
936 gcall *stmt = as_a <gcall *> (gsi_stmt (t->call_gsi));
938 gimple_call_set_tail (stmt, true);
939 cfun->tail_call_marked = true;
940 if (dump_file && (dump_flags & TDF_DETAILS))
942 fprintf (dump_file, "Found tail call ");
943 print_gimple_stmt (dump_file, stmt, 0, dump_flags);
944 fprintf (dump_file, " in bb %i\n", (gsi_bb (t->call_gsi))->index);
948 return false;
951 /* Creates a tail-call accumulator of the same type as the return type of the
952 current function. LABEL is the name used to creating the temporary
953 variable for the accumulator. The accumulator will be inserted in the
954 phis of a basic block BB with single predecessor with an initial value
955 INIT converted to the current function return type. */
957 static tree
958 create_tailcall_accumulator (const char *label, basic_block bb, tree init)
960 tree ret_type = TREE_TYPE (DECL_RESULT (current_function_decl));
961 if (POINTER_TYPE_P (ret_type))
962 ret_type = sizetype;
964 tree tmp = make_temp_ssa_name (ret_type, NULL, label);
965 gphi *phi;
967 phi = create_phi_node (tmp, bb);
968 /* RET_TYPE can be a float when -ffast-maths is enabled. */
969 add_phi_arg (phi, fold_convert (ret_type, init), single_pred_edge (bb),
970 UNKNOWN_LOCATION);
971 return PHI_RESULT (phi);
974 /* Optimizes tail calls in the function, turning the tail recursion
975 into iteration. */
977 static unsigned int
978 tree_optimize_tail_calls_1 (bool opt_tailcalls)
980 edge e;
981 bool phis_constructed = false;
982 struct tailcall *tailcalls = NULL, *act, *next;
983 bool changed = false;
984 basic_block first = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun));
985 tree param;
986 gimple stmt;
987 edge_iterator ei;
989 if (!suitable_for_tail_opt_p ())
990 return 0;
991 if (opt_tailcalls)
992 opt_tailcalls = suitable_for_tail_call_opt_p ();
994 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds)
996 /* Only traverse the normal exits, i.e. those that end with return
997 statement. */
998 stmt = last_stmt (e->src);
1000 if (stmt
1001 && gimple_code (stmt) == GIMPLE_RETURN)
1002 find_tail_calls (e->src, &tailcalls);
1005 /* Construct the phi nodes and accumulators if necessary. */
1006 a_acc = m_acc = NULL_TREE;
1007 for (act = tailcalls; act; act = act->next)
1009 if (!act->tail_recursion)
1010 continue;
1012 if (!phis_constructed)
1014 /* Ensure that there is only one predecessor of the block
1015 or if there are existing degenerate PHI nodes. */
1016 if (!single_pred_p (first)
1017 || !gimple_seq_empty_p (phi_nodes (first)))
1018 first =
1019 split_edge (single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun)));
1021 /* Copy the args if needed. */
1022 for (param = DECL_ARGUMENTS (current_function_decl);
1023 param;
1024 param = DECL_CHAIN (param))
1025 if (arg_needs_copy_p (param))
1027 tree name = ssa_default_def (cfun, param);
1028 tree new_name = make_ssa_name (param, SSA_NAME_DEF_STMT (name));
1029 gphi *phi;
1031 set_ssa_default_def (cfun, param, new_name);
1032 phi = create_phi_node (name, first);
1033 add_phi_arg (phi, new_name, single_pred_edge (first),
1034 EXPR_LOCATION (param));
1036 phis_constructed = true;
1039 if (act->add && !a_acc)
1040 a_acc = create_tailcall_accumulator ("add_acc", first,
1041 integer_zero_node);
1043 if (act->mult && !m_acc)
1044 m_acc = create_tailcall_accumulator ("mult_acc", first,
1045 integer_one_node);
1048 if (a_acc || m_acc)
1050 /* When the tail call elimination using accumulators is performed,
1051 statements adding the accumulated value are inserted at all exits.
1052 This turns all other tail calls to non-tail ones. */
1053 opt_tailcalls = false;
1056 for (; tailcalls; tailcalls = next)
1058 next = tailcalls->next;
1059 changed |= optimize_tail_call (tailcalls, opt_tailcalls);
1060 free (tailcalls);
1063 if (a_acc || m_acc)
1065 /* Modify the remaining return statements. */
1066 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds)
1068 stmt = last_stmt (e->src);
1070 if (stmt
1071 && gimple_code (stmt) == GIMPLE_RETURN)
1072 adjust_return_value (e->src, m_acc, a_acc);
1076 if (changed)
1078 /* We may have created new loops. Make them magically appear. */
1079 loops_state_set (LOOPS_NEED_FIXUP);
1080 free_dominance_info (CDI_DOMINATORS);
1083 /* Add phi nodes for the virtual operands defined in the function to the
1084 header of the loop created by tail recursion elimination. Do so
1085 by triggering the SSA renamer. */
1086 if (phis_constructed)
1087 mark_virtual_operands_for_renaming (cfun);
1089 if (changed)
1090 return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals;
1091 return 0;
1094 static bool
1095 gate_tail_calls (void)
1097 return flag_optimize_sibling_calls != 0 && dbg_cnt (tail_call);
1100 static unsigned int
1101 execute_tail_calls (void)
1103 return tree_optimize_tail_calls_1 (true);
1106 namespace {
1108 const pass_data pass_data_tail_recursion =
1110 GIMPLE_PASS, /* type */
1111 "tailr", /* name */
1112 OPTGROUP_NONE, /* optinfo_flags */
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 0, /* todo_flags_finish */
1121 class pass_tail_recursion : public gimple_opt_pass
1123 public:
1124 pass_tail_recursion (gcc::context *ctxt)
1125 : gimple_opt_pass (pass_data_tail_recursion, ctxt)
1128 /* opt_pass methods: */
1129 opt_pass * clone () { return new pass_tail_recursion (m_ctxt); }
1130 virtual bool gate (function *) { return gate_tail_calls (); }
1131 virtual unsigned int execute (function *)
1133 return tree_optimize_tail_calls_1 (false);
1136 }; // class pass_tail_recursion
1138 } // anon namespace
1140 gimple_opt_pass *
1141 make_pass_tail_recursion (gcc::context *ctxt)
1143 return new pass_tail_recursion (ctxt);
1146 namespace {
1148 const pass_data pass_data_tail_calls =
1150 GIMPLE_PASS, /* type */
1151 "tailc", /* name */
1152 OPTGROUP_NONE, /* optinfo_flags */
1153 TV_NONE, /* tv_id */
1154 ( PROP_cfg | PROP_ssa ), /* properties_required */
1155 0, /* properties_provided */
1156 0, /* properties_destroyed */
1157 0, /* todo_flags_start */
1158 0, /* todo_flags_finish */
1161 class pass_tail_calls : public gimple_opt_pass
1163 public:
1164 pass_tail_calls (gcc::context *ctxt)
1165 : gimple_opt_pass (pass_data_tail_calls, ctxt)
1168 /* opt_pass methods: */
1169 virtual bool gate (function *) { return gate_tail_calls (); }
1170 virtual unsigned int execute (function *) { return execute_tail_calls (); }
1172 }; // class pass_tail_calls
1174 } // anon namespace
1176 gimple_opt_pass *
1177 make_pass_tail_calls (gcc::context *ctxt)
1179 return new pass_tail_calls (ctxt);