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)
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/>. */
22 #include "coretypes.h"
25 #include "stor-layout.h"
27 #include "basic-block.h"
29 #include "tree-ssa-alias.h"
30 #include "internal-fn.h"
31 #include "gimple-expr.h"
34 #include "gimple-iterator.h"
35 #include "gimplify-me.h"
36 #include "gimple-ssa.h"
38 #include "tree-phinodes.h"
39 #include "stringpool.h"
40 #include "tree-ssanames.h"
41 #include "tree-into-ssa.h"
44 #include "gimple-pretty-print.h"
46 #include "tree-pass.h"
48 #include "langhooks.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
66 return n + sum (n - 1);
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
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
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. */
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. */
125 /* The return value of the caller is mult * f + add, where f is the return
126 value of the call. */
129 /* Next tailcall in the chain. */
130 struct tailcall
*next
;
133 /* The variables holding the value of multiplicative and additive
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). */
146 suitable_for_tail_opt_p (void)
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. */
159 suitable_for_tail_call_opt_p (void)
163 /* alloca (until we have stack slot life analysis) inhibits
164 sibling call optimizations, but not tail recursion. */
165 if (cfun
->calls_alloca
)
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 ())
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
)
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
);
185 param
= DECL_CHAIN (param
))
186 if (TREE_ADDRESSABLE (param
))
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. */
199 independent_of_stmt_p (tree expr
, gimple at
, gimple_stmt_iterator gsi
)
201 basic_block bb
, call_bb
, at_bb
;
205 if (is_gimple_min_invariant (expr
))
208 if (TREE_CODE (expr
) != SSA_NAME
)
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
))
220 at
= SSA_NAME_DEF_STMT (expr
);
223 /* The default definition or defined before the chain. */
229 for (; !gsi_end_p (gsi
); gsi_next (&gsi
))
230 if (gsi_stmt (gsi
) == at
)
233 if (!gsi_end_p (gsi
))
238 if (gimple_code (at
) != GIMPLE_PHI
)
244 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
249 expr
= PHI_ARG_DEF_FROM_EDGE (at
, e
);
250 if (TREE_CODE (expr
) != SSA_NAME
)
252 /* The value is a constant. */
257 /* Unmark the blocks. */
258 for (bb
= call_bb
; bb
!= at_bb
; bb
= single_succ (bb
))
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. */
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
288 if (gimple_assign_cast_p (stmt
))
290 if (TYPE_MODE (TREE_TYPE (dest
)) != TYPE_MODE (TREE_TYPE (src_var
)))
293 /* Even if the type modes are the same, if the precision of the
294 type is smaller than mode's precision,
295 reduce_to_bit_field_precision would generate additional code. */
296 if (INTEGRAL_TYPE_P (TREE_TYPE (dest
))
297 && (GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (dest
)))
298 > TYPE_PRECISION (TREE_TYPE (dest
))))
302 if (src_var
!= *ass_var
)
311 case GIMPLE_BINARY_RHS
:
312 op1
= gimple_assign_rhs2 (stmt
);
316 case GIMPLE_UNARY_RHS
:
317 op0
= gimple_assign_rhs1 (stmt
);
324 /* Accumulator optimizations will reverse the order of operations.
325 We can only do that for floating-point types if we're assuming
326 that addition and multiplication are associative. */
327 if (!flag_associative_math
)
328 if (FLOAT_TYPE_P (TREE_TYPE (DECL_RESULT (current_function_decl
))))
331 if (rhs_class
== GIMPLE_UNARY_RHS
)
333 else if (op0
== *ass_var
334 && (non_ass_var
= independent_of_stmt_p (op1
, stmt
, call
)))
336 else if (op1
== *ass_var
337 && (non_ass_var
= independent_of_stmt_p (op0
, stmt
, call
)))
349 case POINTER_PLUS_EXPR
:
362 *m
= build_minus_one_cst (TREE_TYPE (op0
));
368 *a
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (non_ass_var
), non_ass_var
);
371 *m
= build_minus_one_cst (TREE_TYPE (non_ass_var
));
372 *a
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (non_ass_var
), non_ass_var
);
378 /* TODO -- Handle POINTER_PLUS_EXPR. */
385 /* Propagate VAR through phis on edge E. */
388 propagate_through_phis (tree var
, edge e
)
390 basic_block dest
= e
->dest
;
391 gimple_stmt_iterator gsi
;
393 for (gsi
= gsi_start_phis (dest
); !gsi_end_p (gsi
); gsi_next (&gsi
))
395 gimple phi
= gsi_stmt (gsi
);
396 if (PHI_ARG_DEF_FROM_EDGE (phi
, e
) == var
)
397 return PHI_RESULT (phi
);
402 /* Finds tailcalls falling into basic block BB. The list of found tailcalls is
403 added to the start of RET. */
406 find_tail_calls (basic_block bb
, struct tailcall
**ret
)
408 tree ass_var
= NULL_TREE
, ret_var
, func
, param
;
409 gimple stmt
, call
= NULL
;
410 gimple_stmt_iterator gsi
, agsi
;
419 if (!single_succ_p (bb
))
422 for (gsi
= gsi_last_bb (bb
); !gsi_end_p (gsi
); gsi_prev (&gsi
))
424 stmt
= gsi_stmt (gsi
);
426 /* Ignore labels, returns, clobbers and debug stmts. */
427 if (gimple_code (stmt
) == GIMPLE_LABEL
428 || gimple_code (stmt
) == GIMPLE_RETURN
429 || gimple_clobber_p (stmt
)
430 || is_gimple_debug (stmt
))
433 /* Check for a call. */
434 if (is_gimple_call (stmt
))
437 ass_var
= gimple_call_lhs (stmt
);
441 /* If the statement references memory or volatile operands, fail. */
442 if (gimple_references_memory_p (stmt
)
443 || gimple_has_volatile_ops (stmt
))
450 /* Recurse to the predecessors. */
451 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
452 find_tail_calls (e
->src
, ret
);
457 /* If the LHS of our call is not just a simple register, we can't
458 transform this into a tail or sibling call. This situation happens,
459 in (e.g.) "*p = foo()" where foo returns a struct. In this case
460 we won't have a temporary here, but we need to carry out the side
461 effect anyway, so tailcall is impossible.
463 ??? In some situations (when the struct is returned in memory via
464 invisible argument) we could deal with this, e.g. by passing 'p'
465 itself as that argument to foo, but it's too early to do this here,
466 and expand_call() will not handle it anyway. If it ever can, then
467 we need to revisit this here, to allow that situation. */
468 if (ass_var
&& !is_gimple_reg (ass_var
))
471 /* We found the call, check whether it is suitable. */
472 tail_recursion
= false;
473 func
= gimple_call_fndecl (call
);
475 && !DECL_BUILT_IN (func
)
476 && recursive_call_p (current_function_decl
, func
))
480 for (param
= DECL_ARGUMENTS (func
), idx
= 0;
481 param
&& idx
< gimple_call_num_args (call
);
482 param
= DECL_CHAIN (param
), idx
++)
484 arg
= gimple_call_arg (call
, idx
);
487 /* Make sure there are no problems with copying. The parameter
488 have a copyable type and the two arguments must have reasonably
489 equivalent types. The latter requirement could be relaxed if
490 we emitted a suitable type conversion statement. */
491 if (!is_gimple_reg_type (TREE_TYPE (param
))
492 || !useless_type_conversion_p (TREE_TYPE (param
),
496 /* The parameter should be a real operand, so that phi node
497 created for it at the start of the function has the meaning
498 of copying the value. This test implies is_gimple_reg_type
499 from the previous condition, however this one could be
500 relaxed by being more careful with copying the new value
501 of the parameter (emitting appropriate GIMPLE_ASSIGN and
502 updating the virtual operands). */
503 if (!is_gimple_reg (param
))
507 if (idx
== gimple_call_num_args (call
) && !param
)
508 tail_recursion
= true;
511 /* Make sure the tail invocation of this function does not refer
512 to local variables. */
513 FOR_EACH_LOCAL_DECL (cfun
, idx
, var
)
515 if (TREE_CODE (var
) != PARM_DECL
516 && auto_var_in_fn_p (var
, cfun
->decl
)
517 && (ref_maybe_used_by_stmt_p (call
, var
)
518 || call_may_clobber_ref_p (call
, var
)))
522 /* Now check the statements after the call. None of them has virtual
523 operands, so they may only depend on the call through its return
524 value. The return value should also be dependent on each of them,
525 since we are running after dce. */
533 tree tmp_a
= NULL_TREE
;
534 tree tmp_m
= NULL_TREE
;
537 while (gsi_end_p (agsi
))
539 ass_var
= propagate_through_phis (ass_var
, single_succ_edge (abb
));
540 abb
= single_succ (abb
);
541 agsi
= gsi_start_bb (abb
);
544 stmt
= gsi_stmt (agsi
);
546 if (gimple_code (stmt
) == GIMPLE_LABEL
)
549 if (gimple_code (stmt
) == GIMPLE_RETURN
)
552 if (gimple_clobber_p (stmt
))
555 if (is_gimple_debug (stmt
))
558 if (gimple_code (stmt
) != GIMPLE_ASSIGN
)
561 /* This is a gimple assign. */
562 if (! process_assignment (stmt
, gsi
, &tmp_m
, &tmp_a
, &ass_var
))
567 tree type
= TREE_TYPE (tmp_a
);
569 a
= fold_build2 (PLUS_EXPR
, type
, fold_convert (type
, a
), tmp_a
);
575 tree type
= TREE_TYPE (tmp_m
);
577 m
= fold_build2 (MULT_EXPR
, type
, fold_convert (type
, m
), tmp_m
);
582 a
= fold_build2 (MULT_EXPR
, type
, fold_convert (type
, a
), tmp_m
);
586 /* See if this is a tail call we can handle. */
587 ret_var
= gimple_return_retval (stmt
);
589 /* We may proceed if there either is no return value, or the return value
590 is identical to the call's return. */
592 && (ret_var
!= ass_var
))
595 /* If this is not a tail recursive call, we cannot handle addends or
597 if (!tail_recursion
&& (m
|| a
))
600 /* For pointers only allow additions. */
601 if (m
&& POINTER_TYPE_P (TREE_TYPE (DECL_RESULT (current_function_decl
))))
604 nw
= XNEW (struct tailcall
);
608 nw
->tail_recursion
= tail_recursion
;
617 /* Helper to insert PHI_ARGH to the phi of VAR in the destination of edge E. */
620 add_successor_phi_arg (edge e
, tree var
, tree phi_arg
)
622 gimple_stmt_iterator gsi
;
624 for (gsi
= gsi_start_phis (e
->dest
); !gsi_end_p (gsi
); gsi_next (&gsi
))
625 if (PHI_RESULT (gsi_stmt (gsi
)) == var
)
628 gcc_assert (!gsi_end_p (gsi
));
629 add_phi_arg (gsi_stmt (gsi
), phi_arg
, e
, UNKNOWN_LOCATION
);
632 /* Creates a GIMPLE statement which computes the operation specified by
633 CODE, ACC and OP1 to a new variable with name LABEL and inserts the
634 statement in the position specified by GSI. Returns the
635 tree node of the statement's result. */
638 adjust_return_value_with_ops (enum tree_code code
, const char *label
,
639 tree acc
, tree op1
, gimple_stmt_iterator gsi
)
642 tree ret_type
= TREE_TYPE (DECL_RESULT (current_function_decl
));
643 tree result
= make_temp_ssa_name (ret_type
, NULL
, label
);
646 if (POINTER_TYPE_P (ret_type
))
648 gcc_assert (code
== PLUS_EXPR
&& TREE_TYPE (acc
) == sizetype
);
649 code
= POINTER_PLUS_EXPR
;
651 if (types_compatible_p (TREE_TYPE (acc
), TREE_TYPE (op1
))
652 && code
!= POINTER_PLUS_EXPR
)
653 stmt
= gimple_build_assign_with_ops (code
, result
, acc
, op1
);
657 if (code
== POINTER_PLUS_EXPR
)
658 tem
= fold_build2 (code
, TREE_TYPE (op1
), op1
, acc
);
660 tem
= fold_build2 (code
, TREE_TYPE (op1
),
661 fold_convert (TREE_TYPE (op1
), acc
), op1
);
662 tree rhs
= fold_convert (ret_type
, tem
);
663 rhs
= force_gimple_operand_gsi (&gsi
, rhs
,
664 false, NULL
, true, GSI_SAME_STMT
);
665 stmt
= gimple_build_assign (result
, rhs
);
668 gsi_insert_before (&gsi
, stmt
, GSI_NEW_STMT
);
672 /* Creates a new GIMPLE statement that adjusts the value of accumulator ACC by
673 the computation specified by CODE and OP1 and insert the statement
674 at the position specified by GSI as a new statement. Returns new SSA name
675 of updated accumulator. */
678 update_accumulator_with_ops (enum tree_code code
, tree acc
, tree op1
,
679 gimple_stmt_iterator gsi
)
682 tree var
= copy_ssa_name (acc
, NULL
);
683 if (types_compatible_p (TREE_TYPE (acc
), TREE_TYPE (op1
)))
684 stmt
= gimple_build_assign_with_ops (code
, var
, acc
, op1
);
687 tree rhs
= fold_convert (TREE_TYPE (acc
),
690 fold_convert (TREE_TYPE (op1
), acc
),
692 rhs
= force_gimple_operand_gsi (&gsi
, rhs
,
693 false, NULL
, false, GSI_CONTINUE_LINKING
);
694 stmt
= gimple_build_assign (var
, rhs
);
696 gsi_insert_after (&gsi
, stmt
, GSI_NEW_STMT
);
700 /* Adjust the accumulator values according to A and M after GSI, and update
701 the phi nodes on edge BACK. */
704 adjust_accumulator_values (gimple_stmt_iterator gsi
, tree m
, tree a
, edge back
)
706 tree var
, a_acc_arg
, m_acc_arg
;
709 m
= force_gimple_operand_gsi (&gsi
, m
, true, NULL
, true, GSI_SAME_STMT
);
711 a
= force_gimple_operand_gsi (&gsi
, a
, true, NULL
, true, GSI_SAME_STMT
);
719 if (integer_onep (a
))
722 var
= adjust_return_value_with_ops (MULT_EXPR
, "acc_tmp", m_acc
,
728 a_acc_arg
= update_accumulator_with_ops (PLUS_EXPR
, a_acc
, var
, gsi
);
732 m_acc_arg
= update_accumulator_with_ops (MULT_EXPR
, m_acc
, m
, gsi
);
735 add_successor_phi_arg (back
, a_acc
, a_acc_arg
);
738 add_successor_phi_arg (back
, m_acc
, m_acc_arg
);
741 /* Adjust value of the return at the end of BB according to M and A
745 adjust_return_value (basic_block bb
, tree m
, tree a
)
748 gimple ret_stmt
= gimple_seq_last_stmt (bb_seq (bb
));
749 gimple_stmt_iterator gsi
= gsi_last_bb (bb
);
751 gcc_assert (gimple_code (ret_stmt
) == GIMPLE_RETURN
);
753 retval
= gimple_return_retval (ret_stmt
);
754 if (!retval
|| retval
== error_mark_node
)
758 retval
= adjust_return_value_with_ops (MULT_EXPR
, "mul_tmp", m_acc
, retval
,
761 retval
= adjust_return_value_with_ops (PLUS_EXPR
, "acc_tmp", a_acc
, retval
,
763 gimple_return_set_retval (ret_stmt
, retval
);
764 update_stmt (ret_stmt
);
767 /* Subtract COUNT and FREQUENCY from the basic block and it's
770 decrease_profile (basic_block bb
, gcov_type count
, int frequency
)
776 bb
->frequency
-= frequency
;
777 if (bb
->frequency
< 0)
779 if (!single_succ_p (bb
))
781 gcc_assert (!EDGE_COUNT (bb
->succs
));
784 e
= single_succ_edge (bb
);
790 /* Returns true if argument PARAM of the tail recursive call needs to be copied
791 when the call is eliminated. */
794 arg_needs_copy_p (tree param
)
798 if (!is_gimple_reg (param
))
801 /* Parameters that are only defined but never used need not be copied. */
802 def
= ssa_default_def (cfun
, param
);
809 /* Eliminates tail call described by T. TMP_VARS is a list of
810 temporary variables used to copy the function arguments. */
813 eliminate_tail_call (struct tailcall
*t
)
819 basic_block bb
, first
;
822 gimple_stmt_iterator gsi
;
825 stmt
= orig_stmt
= gsi_stmt (t
->call_gsi
);
826 bb
= gsi_bb (t
->call_gsi
);
828 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
830 fprintf (dump_file
, "Eliminated tail recursion in bb %d : ",
832 print_gimple_stmt (dump_file
, stmt
, 0, TDF_SLIM
);
833 fprintf (dump_file
, "\n");
836 gcc_assert (is_gimple_call (stmt
));
838 first
= single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun
));
840 /* Remove the code after call_gsi that will become unreachable. The
841 possibly unreachable code in other blocks is removed later in
845 while (!gsi_end_p (gsi
))
847 gimple t
= gsi_stmt (gsi
);
848 /* Do not remove the return statement, so that redirect_edge_and_branch
849 sees how the block ends. */
850 if (gimple_code (t
) == GIMPLE_RETURN
)
853 gsi_remove (&gsi
, true);
857 /* Number of executions of function has reduced by the tailcall. */
858 e
= single_succ_edge (gsi_bb (t
->call_gsi
));
859 decrease_profile (EXIT_BLOCK_PTR_FOR_FN (cfun
), e
->count
, EDGE_FREQUENCY (e
));
860 decrease_profile (ENTRY_BLOCK_PTR_FOR_FN (cfun
), e
->count
,
862 if (e
->dest
!= EXIT_BLOCK_PTR_FOR_FN (cfun
))
863 decrease_profile (e
->dest
, e
->count
, EDGE_FREQUENCY (e
));
865 /* Replace the call by a jump to the start of function. */
866 e
= redirect_edge_and_branch (single_succ_edge (gsi_bb (t
->call_gsi
)),
869 PENDING_STMT (e
) = NULL
;
871 /* Add phi node entries for arguments. The ordering of the phi nodes should
872 be the same as the ordering of the arguments. */
873 for (param
= DECL_ARGUMENTS (current_function_decl
),
874 idx
= 0, gsi
= gsi_start_phis (first
);
876 param
= DECL_CHAIN (param
), idx
++)
878 if (!arg_needs_copy_p (param
))
881 arg
= gimple_call_arg (stmt
, idx
);
882 phi
= gsi_stmt (gsi
);
883 gcc_assert (param
== SSA_NAME_VAR (PHI_RESULT (phi
)));
885 add_phi_arg (phi
, arg
, e
, gimple_location (stmt
));
889 /* Update the values of accumulators. */
890 adjust_accumulator_values (t
->call_gsi
, t
->mult
, t
->add
, e
);
892 call
= gsi_stmt (t
->call_gsi
);
893 rslt
= gimple_call_lhs (call
);
894 if (rslt
!= NULL_TREE
)
896 /* Result of the call will no longer be defined. So adjust the
897 SSA_NAME_DEF_STMT accordingly. */
898 SSA_NAME_DEF_STMT (rslt
) = gimple_build_nop ();
901 gsi_remove (&t
->call_gsi
, true);
905 /* Optimizes the tailcall described by T. If OPT_TAILCALLS is true, also
906 mark the tailcalls for the sibcall optimization. */
909 optimize_tail_call (struct tailcall
*t
, bool opt_tailcalls
)
911 if (t
->tail_recursion
)
913 eliminate_tail_call (t
);
919 gimple stmt
= gsi_stmt (t
->call_gsi
);
921 gimple_call_set_tail (stmt
, true);
922 cfun
->tail_call_marked
= true;
923 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
925 fprintf (dump_file
, "Found tail call ");
926 print_gimple_stmt (dump_file
, stmt
, 0, dump_flags
);
927 fprintf (dump_file
, " in bb %i\n", (gsi_bb (t
->call_gsi
))->index
);
934 /* Creates a tail-call accumulator of the same type as the return type of the
935 current function. LABEL is the name used to creating the temporary
936 variable for the accumulator. The accumulator will be inserted in the
937 phis of a basic block BB with single predecessor with an initial value
938 INIT converted to the current function return type. */
941 create_tailcall_accumulator (const char *label
, basic_block bb
, tree init
)
943 tree ret_type
= TREE_TYPE (DECL_RESULT (current_function_decl
));
944 if (POINTER_TYPE_P (ret_type
))
947 tree tmp
= make_temp_ssa_name (ret_type
, NULL
, label
);
950 phi
= create_phi_node (tmp
, bb
);
951 /* RET_TYPE can be a float when -ffast-maths is enabled. */
952 add_phi_arg (phi
, fold_convert (ret_type
, init
), single_pred_edge (bb
),
954 return PHI_RESULT (phi
);
957 /* Optimizes tail calls in the function, turning the tail recursion
961 tree_optimize_tail_calls_1 (bool opt_tailcalls
)
964 bool phis_constructed
= false;
965 struct tailcall
*tailcalls
= NULL
, *act
, *next
;
966 bool changed
= false;
967 basic_block first
= single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun
));
972 if (!suitable_for_tail_opt_p ())
975 opt_tailcalls
= suitable_for_tail_call_opt_p ();
977 FOR_EACH_EDGE (e
, ei
, EXIT_BLOCK_PTR_FOR_FN (cfun
)->preds
)
979 /* Only traverse the normal exits, i.e. those that end with return
981 stmt
= last_stmt (e
->src
);
984 && gimple_code (stmt
) == GIMPLE_RETURN
)
985 find_tail_calls (e
->src
, &tailcalls
);
988 /* Construct the phi nodes and accumulators if necessary. */
989 a_acc
= m_acc
= NULL_TREE
;
990 for (act
= tailcalls
; act
; act
= act
->next
)
992 if (!act
->tail_recursion
)
995 if (!phis_constructed
)
997 /* Ensure that there is only one predecessor of the block
998 or if there are existing degenerate PHI nodes. */
999 if (!single_pred_p (first
)
1000 || !gimple_seq_empty_p (phi_nodes (first
)))
1002 split_edge (single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun
)));
1004 /* Copy the args if needed. */
1005 for (param
= DECL_ARGUMENTS (current_function_decl
);
1007 param
= DECL_CHAIN (param
))
1008 if (arg_needs_copy_p (param
))
1010 tree name
= ssa_default_def (cfun
, param
);
1011 tree new_name
= make_ssa_name (param
, SSA_NAME_DEF_STMT (name
));
1014 set_ssa_default_def (cfun
, param
, new_name
);
1015 phi
= create_phi_node (name
, first
);
1016 add_phi_arg (phi
, new_name
, single_pred_edge (first
),
1017 EXPR_LOCATION (param
));
1019 phis_constructed
= true;
1022 if (act
->add
&& !a_acc
)
1023 a_acc
= create_tailcall_accumulator ("add_acc", first
,
1026 if (act
->mult
&& !m_acc
)
1027 m_acc
= create_tailcall_accumulator ("mult_acc", first
,
1033 /* When the tail call elimination using accumulators is performed,
1034 statements adding the accumulated value are inserted at all exits.
1035 This turns all other tail calls to non-tail ones. */
1036 opt_tailcalls
= false;
1039 for (; tailcalls
; tailcalls
= next
)
1041 next
= tailcalls
->next
;
1042 changed
|= optimize_tail_call (tailcalls
, opt_tailcalls
);
1048 /* Modify the remaining return statements. */
1049 FOR_EACH_EDGE (e
, ei
, EXIT_BLOCK_PTR_FOR_FN (cfun
)->preds
)
1051 stmt
= last_stmt (e
->src
);
1054 && gimple_code (stmt
) == GIMPLE_RETURN
)
1055 adjust_return_value (e
->src
, m_acc
, a_acc
);
1061 /* We may have created new loops. Make them magically appear. */
1062 loops_state_set (LOOPS_NEED_FIXUP
);
1063 free_dominance_info (CDI_DOMINATORS
);
1066 /* Add phi nodes for the virtual operands defined in the function to the
1067 header of the loop created by tail recursion elimination. Do so
1068 by triggering the SSA renamer. */
1069 if (phis_constructed
)
1070 mark_virtual_operands_for_renaming (cfun
);
1073 return TODO_cleanup_cfg
| TODO_update_ssa_only_virtuals
;
1078 gate_tail_calls (void)
1080 return flag_optimize_sibling_calls
!= 0 && dbg_cnt (tail_call
);
1084 execute_tail_calls (void)
1086 return tree_optimize_tail_calls_1 (true);
1091 const pass_data pass_data_tail_recursion
=
1093 GIMPLE_PASS
, /* type */
1095 OPTGROUP_NONE
, /* optinfo_flags */
1096 TV_NONE
, /* tv_id */
1097 ( PROP_cfg
| PROP_ssa
), /* properties_required */
1098 0, /* properties_provided */
1099 0, /* properties_destroyed */
1100 0, /* todo_flags_start */
1101 0, /* todo_flags_finish */
1104 class pass_tail_recursion
: public gimple_opt_pass
1107 pass_tail_recursion (gcc::context
*ctxt
)
1108 : gimple_opt_pass (pass_data_tail_recursion
, ctxt
)
1111 /* opt_pass methods: */
1112 opt_pass
* clone () { return new pass_tail_recursion (m_ctxt
); }
1113 virtual bool gate (function
*) { return gate_tail_calls (); }
1114 virtual unsigned int execute (function
*)
1116 return tree_optimize_tail_calls_1 (false);
1119 }; // class pass_tail_recursion
1124 make_pass_tail_recursion (gcc::context
*ctxt
)
1126 return new pass_tail_recursion (ctxt
);
1131 const pass_data pass_data_tail_calls
=
1133 GIMPLE_PASS
, /* type */
1135 OPTGROUP_NONE
, /* optinfo_flags */
1136 TV_NONE
, /* tv_id */
1137 ( PROP_cfg
| PROP_ssa
), /* properties_required */
1138 0, /* properties_provided */
1139 0, /* properties_destroyed */
1140 0, /* todo_flags_start */
1141 0, /* todo_flags_finish */
1144 class pass_tail_calls
: public gimple_opt_pass
1147 pass_tail_calls (gcc::context
*ctxt
)
1148 : gimple_opt_pass (pass_data_tail_calls
, ctxt
)
1151 /* opt_pass methods: */
1152 virtual bool gate (function
*) { return gate_tail_calls (); }
1153 virtual unsigned int execute (function
*) { return execute_tail_calls (); }
1155 }; // class pass_tail_calls
1160 make_pass_tail_calls (gcc::context
*ctxt
)
1162 return new pass_tail_calls (ctxt
);