1 /* Tail call optimization on trees.
2 Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011, 2012
3 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 3, or (at your option)
12 GCC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
23 #include "coretypes.h"
27 #include "basic-block.h"
29 #include "tree-flow.h"
30 #include "gimple-pretty-print.h"
32 #include "tree-pass.h"
34 #include "langhooks.h"
37 #include "common/common-target.h"
39 /* The file implements the tail recursion elimination. It is also used to
40 analyze the tail calls in general, passing the results to the rtl level
41 where they are used for sibcall optimization.
43 In addition to the standard tail recursion elimination, we handle the most
44 trivial cases of making the call tail recursive by creating accumulators.
45 For example the following function
50 return n + sum (n - 1);
67 To do this, we maintain two accumulators (a_acc and m_acc) that indicate
68 when we reach the return x statement, we should return a_acc + x * m_acc
69 instead. They are initially initialized to 0 and 1, respectively,
70 so the semantics of the function is obviously preserved. If we are
71 guaranteed that the value of the accumulator never change, we
74 There are three cases how the function may exit. The first one is
75 handled in adjust_return_value, the other two in adjust_accumulator_values
76 (the second case is actually a special case of the third one and we
77 present it separately just for clarity):
79 1) Just return x, where x is not in any of the remaining special shapes.
80 We rewrite this to a gimple equivalent of return m_acc * x + a_acc.
82 2) return f (...), where f is the current function, is rewritten in a
83 classical tail-recursion elimination way, into assignment of arguments
84 and jump to the start of the function. Values of the accumulators
87 3) return a + m * f(...), where a and m do not depend on call to f.
88 To preserve the semantics described before we want this to be rewritten
89 in such a way that we finally return
91 a_acc + (a + m * f(...)) * m_acc = (a_acc + a * m_acc) + (m * m_acc) * f(...).
93 I.e. we increase a_acc by a * m_acc, multiply m_acc by m and
94 eliminate the tail call to f. Special cases when the value is just
95 added or just multiplied are obtained by setting a = 0 or m = 1.
97 TODO -- it is possible to do similar tricks for other operations. */
99 /* A structure that describes the tailcall. */
103 /* The iterator pointing to the call statement. */
104 gimple_stmt_iterator call_gsi
;
106 /* True if it is a call to the current function. */
109 /* The return value of the caller is mult * f + add, where f is the return
110 value of the call. */
113 /* Next tailcall in the chain. */
114 struct tailcall
*next
;
117 /* The variables holding the value of multiplicative and additive
119 static tree m_acc
, a_acc
;
121 static bool suitable_for_tail_opt_p (void);
122 static bool optimize_tail_call (struct tailcall
*, bool);
123 static void eliminate_tail_call (struct tailcall
*);
124 static void find_tail_calls (basic_block
, struct tailcall
**);
126 /* Returns false when the function is not suitable for tail call optimization
127 from some reason (e.g. if it takes variable number of arguments). */
130 suitable_for_tail_opt_p (void)
137 /* Returns false when the function is not suitable for tail call optimization
138 from some reason (e.g. if it takes variable number of arguments).
139 This test must pass in addition to suitable_for_tail_opt_p in order to make
140 tail call discovery happen. */
143 suitable_for_tail_call_opt_p (void)
147 /* alloca (until we have stack slot life analysis) inhibits
148 sibling call optimizations, but not tail recursion. */
149 if (cfun
->calls_alloca
)
152 /* If we are using sjlj exceptions, we may need to add a call to
153 _Unwind_SjLj_Unregister at exit of the function. Which means
154 that we cannot do any sibcall transformations. */
155 if (targetm_common
.except_unwind_info (&global_options
) == UI_SJLJ
156 && current_function_has_exception_handlers ())
159 /* Any function that calls setjmp might have longjmp called from
160 any called function. ??? We really should represent this
161 properly in the CFG so that this needn't be special cased. */
162 if (cfun
->calls_setjmp
)
165 /* ??? It is OK if the argument of a function is taken in some cases,
166 but not in all cases. See PR15387 and PR19616. Revisit for 4.1. */
167 for (param
= DECL_ARGUMENTS (current_function_decl
);
169 param
= DECL_CHAIN (param
))
170 if (TREE_ADDRESSABLE (param
))
176 /* Checks whether the expression EXPR in stmt AT is independent of the
177 statement pointed to by GSI (in a sense that we already know EXPR's value
178 at GSI). We use the fact that we are only called from the chain of
179 basic blocks that have only single successor. Returns the expression
180 containing the value of EXPR at GSI. */
183 independent_of_stmt_p (tree expr
, gimple at
, gimple_stmt_iterator gsi
)
185 basic_block bb
, call_bb
, at_bb
;
189 if (is_gimple_min_invariant (expr
))
192 if (TREE_CODE (expr
) != SSA_NAME
)
195 /* Mark the blocks in the chain leading to the end. */
196 at_bb
= gimple_bb (at
);
197 call_bb
= gimple_bb (gsi_stmt (gsi
));
198 for (bb
= call_bb
; bb
!= at_bb
; bb
= single_succ (bb
))
204 at
= SSA_NAME_DEF_STMT (expr
);
207 /* The default definition or defined before the chain. */
213 for (; !gsi_end_p (gsi
); gsi_next (&gsi
))
214 if (gsi_stmt (gsi
) == at
)
217 if (!gsi_end_p (gsi
))
222 if (gimple_code (at
) != GIMPLE_PHI
)
228 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
233 expr
= PHI_ARG_DEF_FROM_EDGE (at
, e
);
234 if (TREE_CODE (expr
) != SSA_NAME
)
236 /* The value is a constant. */
241 /* Unmark the blocks. */
242 for (bb
= call_bb
; bb
!= at_bb
; bb
= single_succ (bb
))
249 /* Simulates the effect of an assignment STMT on the return value of the tail
250 recursive CALL passed in ASS_VAR. M and A are the multiplicative and the
251 additive factor for the real return value. */
254 process_assignment (gimple stmt
, gimple_stmt_iterator call
, tree
*m
,
255 tree
*a
, tree
*ass_var
)
257 tree op0
, op1
= NULL_TREE
, non_ass_var
= NULL_TREE
;
258 tree dest
= gimple_assign_lhs (stmt
);
259 enum tree_code code
= gimple_assign_rhs_code (stmt
);
260 enum gimple_rhs_class rhs_class
= get_gimple_rhs_class (code
);
261 tree src_var
= gimple_assign_rhs1 (stmt
);
263 /* See if this is a simple copy operation of an SSA name to the function
264 result. In that case we may have a simple tail call. Ignore type
265 conversions that can never produce extra code between the function
266 call and the function return. */
267 if ((rhs_class
== GIMPLE_SINGLE_RHS
|| gimple_assign_cast_p (stmt
))
268 && (TREE_CODE (src_var
) == SSA_NAME
))
270 /* Reject a tailcall if the type conversion might need
272 if (gimple_assign_cast_p (stmt
)
273 && TYPE_MODE (TREE_TYPE (dest
)) != TYPE_MODE (TREE_TYPE (src_var
)))
276 if (src_var
!= *ass_var
)
285 case GIMPLE_BINARY_RHS
:
286 op1
= gimple_assign_rhs2 (stmt
);
290 case GIMPLE_UNARY_RHS
:
291 op0
= gimple_assign_rhs1 (stmt
);
298 /* Accumulator optimizations will reverse the order of operations.
299 We can only do that for floating-point types if we're assuming
300 that addition and multiplication are associative. */
301 if (!flag_associative_math
)
302 if (FLOAT_TYPE_P (TREE_TYPE (DECL_RESULT (current_function_decl
))))
305 if (rhs_class
== GIMPLE_UNARY_RHS
)
307 else if (op0
== *ass_var
308 && (non_ass_var
= independent_of_stmt_p (op1
, stmt
, call
)))
310 else if (op1
== *ass_var
311 && (non_ass_var
= independent_of_stmt_p (op0
, stmt
, call
)))
329 if (FLOAT_TYPE_P (TREE_TYPE (op0
)))
330 *m
= build_real (TREE_TYPE (op0
), dconstm1
);
332 *m
= build_int_cst (TREE_TYPE (op0
), -1);
339 *a
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (non_ass_var
), non_ass_var
);
342 if (FLOAT_TYPE_P (TREE_TYPE (non_ass_var
)))
343 *m
= build_real (TREE_TYPE (non_ass_var
), dconstm1
);
345 *m
= build_int_cst (TREE_TYPE (non_ass_var
), -1);
347 *a
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (non_ass_var
), non_ass_var
);
353 /* TODO -- Handle POINTER_PLUS_EXPR. */
360 /* Propagate VAR through phis on edge E. */
363 propagate_through_phis (tree var
, edge e
)
365 basic_block dest
= e
->dest
;
366 gimple_stmt_iterator gsi
;
368 for (gsi
= gsi_start_phis (dest
); !gsi_end_p (gsi
); gsi_next (&gsi
))
370 gimple phi
= gsi_stmt (gsi
);
371 if (PHI_ARG_DEF_FROM_EDGE (phi
, e
) == var
)
372 return PHI_RESULT (phi
);
377 /* Finds tailcalls falling into basic block BB. The list of found tailcalls is
378 added to the start of RET. */
381 find_tail_calls (basic_block bb
, struct tailcall
**ret
)
383 tree ass_var
= NULL_TREE
, ret_var
, func
, param
;
384 gimple stmt
, call
= NULL
;
385 gimple_stmt_iterator gsi
, agsi
;
393 referenced_var_iterator rvi
;
395 if (!single_succ_p (bb
))
398 for (gsi
= gsi_last_bb (bb
); !gsi_end_p (gsi
); gsi_prev (&gsi
))
400 stmt
= gsi_stmt (gsi
);
402 /* Ignore labels, returns, clobbers and debug stmts. */
403 if (gimple_code (stmt
) == GIMPLE_LABEL
404 || gimple_code (stmt
) == GIMPLE_RETURN
405 || gimple_clobber_p (stmt
)
406 || is_gimple_debug (stmt
))
409 /* Check for a call. */
410 if (is_gimple_call (stmt
))
413 ass_var
= gimple_call_lhs (stmt
);
417 /* If the statement references memory or volatile operands, fail. */
418 if (gimple_references_memory_p (stmt
)
419 || gimple_has_volatile_ops (stmt
))
426 /* Recurse to the predecessors. */
427 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
428 find_tail_calls (e
->src
, ret
);
433 /* If the LHS of our call is not just a simple register, we can't
434 transform this into a tail or sibling call. This situation happens,
435 in (e.g.) "*p = foo()" where foo returns a struct. In this case
436 we won't have a temporary here, but we need to carry out the side
437 effect anyway, so tailcall is impossible.
439 ??? In some situations (when the struct is returned in memory via
440 invisible argument) we could deal with this, e.g. by passing 'p'
441 itself as that argument to foo, but it's too early to do this here,
442 and expand_call() will not handle it anyway. If it ever can, then
443 we need to revisit this here, to allow that situation. */
444 if (ass_var
&& !is_gimple_reg (ass_var
))
447 /* We found the call, check whether it is suitable. */
448 tail_recursion
= false;
449 func
= gimple_call_fndecl (call
);
450 if (func
== current_function_decl
)
454 for (param
= DECL_ARGUMENTS (func
), idx
= 0;
455 param
&& idx
< gimple_call_num_args (call
);
456 param
= DECL_CHAIN (param
), idx
++)
458 arg
= gimple_call_arg (call
, idx
);
461 /* Make sure there are no problems with copying. The parameter
462 have a copyable type and the two arguments must have reasonably
463 equivalent types. The latter requirement could be relaxed if
464 we emitted a suitable type conversion statement. */
465 if (!is_gimple_reg_type (TREE_TYPE (param
))
466 || !useless_type_conversion_p (TREE_TYPE (param
),
470 /* The parameter should be a real operand, so that phi node
471 created for it at the start of the function has the meaning
472 of copying the value. This test implies is_gimple_reg_type
473 from the previous condition, however this one could be
474 relaxed by being more careful with copying the new value
475 of the parameter (emitting appropriate GIMPLE_ASSIGN and
476 updating the virtual operands). */
477 if (!is_gimple_reg (param
))
481 if (idx
== gimple_call_num_args (call
) && !param
)
482 tail_recursion
= true;
485 /* Make sure the tail invocation of this function does not refer
486 to local variables. */
487 FOR_EACH_REFERENCED_VAR (cfun
, var
, rvi
)
489 if (TREE_CODE (var
) != PARM_DECL
490 && auto_var_in_fn_p (var
, cfun
->decl
)
491 && (ref_maybe_used_by_stmt_p (call
, var
)
492 || call_may_clobber_ref_p (call
, var
)))
496 /* Now check the statements after the call. None of them has virtual
497 operands, so they may only depend on the call through its return
498 value. The return value should also be dependent on each of them,
499 since we are running after dce. */
507 tree tmp_a
= NULL_TREE
;
508 tree tmp_m
= NULL_TREE
;
511 while (gsi_end_p (agsi
))
513 ass_var
= propagate_through_phis (ass_var
, single_succ_edge (abb
));
514 abb
= single_succ (abb
);
515 agsi
= gsi_start_bb (abb
);
518 stmt
= gsi_stmt (agsi
);
520 if (gimple_code (stmt
) == GIMPLE_LABEL
)
523 if (gimple_code (stmt
) == GIMPLE_RETURN
)
526 if (gimple_clobber_p (stmt
))
529 if (is_gimple_debug (stmt
))
532 if (gimple_code (stmt
) != GIMPLE_ASSIGN
)
535 /* This is a gimple assign. */
536 if (! process_assignment (stmt
, gsi
, &tmp_m
, &tmp_a
, &ass_var
))
541 tree type
= TREE_TYPE (tmp_a
);
543 a
= fold_build2 (PLUS_EXPR
, type
, fold_convert (type
, a
), tmp_a
);
549 tree type
= TREE_TYPE (tmp_m
);
551 m
= fold_build2 (MULT_EXPR
, type
, fold_convert (type
, m
), tmp_m
);
556 a
= fold_build2 (MULT_EXPR
, type
, fold_convert (type
, a
), tmp_m
);
560 /* See if this is a tail call we can handle. */
561 ret_var
= gimple_return_retval (stmt
);
563 /* We may proceed if there either is no return value, or the return value
564 is identical to the call's return. */
566 && (ret_var
!= ass_var
))
569 /* If this is not a tail recursive call, we cannot handle addends or
571 if (!tail_recursion
&& (m
|| a
))
574 nw
= XNEW (struct tailcall
);
578 nw
->tail_recursion
= tail_recursion
;
587 /* Helper to insert PHI_ARGH to the phi of VAR in the destination of edge E. */
590 add_successor_phi_arg (edge e
, tree var
, tree phi_arg
)
592 gimple_stmt_iterator gsi
;
594 for (gsi
= gsi_start_phis (e
->dest
); !gsi_end_p (gsi
); gsi_next (&gsi
))
595 if (PHI_RESULT (gsi_stmt (gsi
)) == var
)
598 gcc_assert (!gsi_end_p (gsi
));
599 add_phi_arg (gsi_stmt (gsi
), phi_arg
, e
, UNKNOWN_LOCATION
);
602 /* Creates a GIMPLE statement which computes the operation specified by
603 CODE, OP0 and OP1 to a new variable with name LABEL and inserts the
604 statement in the position specified by GSI and UPDATE. Returns the
605 tree node of the statement's result. */
608 adjust_return_value_with_ops (enum tree_code code
, const char *label
,
609 tree acc
, tree op1
, gimple_stmt_iterator gsi
)
612 tree ret_type
= TREE_TYPE (DECL_RESULT (current_function_decl
));
613 tree tmp
= create_tmp_reg (ret_type
, label
);
617 add_referenced_var (tmp
);
619 if (types_compatible_p (TREE_TYPE (acc
), TREE_TYPE (op1
)))
620 stmt
= gimple_build_assign_with_ops (code
, tmp
, acc
, op1
);
623 tree rhs
= fold_convert (TREE_TYPE (acc
),
626 fold_convert (TREE_TYPE (op1
), acc
),
628 rhs
= force_gimple_operand_gsi (&gsi
, rhs
,
629 false, NULL
, true, GSI_CONTINUE_LINKING
);
630 stmt
= gimple_build_assign (NULL_TREE
, rhs
);
633 result
= make_ssa_name (tmp
, stmt
);
634 gimple_assign_set_lhs (stmt
, result
);
636 gsi_insert_before (&gsi
, stmt
, GSI_NEW_STMT
);
640 /* Creates a new GIMPLE statement that adjusts the value of accumulator ACC by
641 the computation specified by CODE and OP1 and insert the statement
642 at the position specified by GSI as a new statement. Returns new SSA name
643 of updated accumulator. */
646 update_accumulator_with_ops (enum tree_code code
, tree acc
, tree op1
,
647 gimple_stmt_iterator gsi
)
651 if (types_compatible_p (TREE_TYPE (acc
), TREE_TYPE (op1
)))
652 stmt
= gimple_build_assign_with_ops (code
, SSA_NAME_VAR (acc
), acc
, op1
);
655 tree rhs
= fold_convert (TREE_TYPE (acc
),
658 fold_convert (TREE_TYPE (op1
), acc
),
660 rhs
= force_gimple_operand_gsi (&gsi
, rhs
,
661 false, NULL
, false, GSI_CONTINUE_LINKING
);
662 stmt
= gimple_build_assign (NULL_TREE
, rhs
);
664 var
= make_ssa_name (SSA_NAME_VAR (acc
), stmt
);
665 gimple_assign_set_lhs (stmt
, var
);
667 gsi_insert_after (&gsi
, stmt
, GSI_NEW_STMT
);
671 /* Adjust the accumulator values according to A and M after GSI, and update
672 the phi nodes on edge BACK. */
675 adjust_accumulator_values (gimple_stmt_iterator gsi
, tree m
, tree a
, edge back
)
677 tree var
, a_acc_arg
, m_acc_arg
;
680 m
= force_gimple_operand_gsi (&gsi
, m
, true, NULL
, true, GSI_SAME_STMT
);
682 a
= force_gimple_operand_gsi (&gsi
, a
, true, NULL
, true, GSI_SAME_STMT
);
690 if (integer_onep (a
))
693 var
= adjust_return_value_with_ops (MULT_EXPR
, "acc_tmp", m_acc
,
699 a_acc_arg
= update_accumulator_with_ops (PLUS_EXPR
, a_acc
, var
, gsi
);
703 m_acc_arg
= update_accumulator_with_ops (MULT_EXPR
, m_acc
, m
, gsi
);
706 add_successor_phi_arg (back
, a_acc
, a_acc_arg
);
709 add_successor_phi_arg (back
, m_acc
, m_acc_arg
);
712 /* Adjust value of the return at the end of BB according to M and A
716 adjust_return_value (basic_block bb
, tree m
, tree a
)
719 gimple ret_stmt
= gimple_seq_last_stmt (bb_seq (bb
));
720 gimple_stmt_iterator gsi
= gsi_last_bb (bb
);
722 gcc_assert (gimple_code (ret_stmt
) == GIMPLE_RETURN
);
724 retval
= gimple_return_retval (ret_stmt
);
725 if (!retval
|| retval
== error_mark_node
)
729 retval
= adjust_return_value_with_ops (MULT_EXPR
, "mul_tmp", m_acc
, retval
,
732 retval
= adjust_return_value_with_ops (PLUS_EXPR
, "acc_tmp", a_acc
, retval
,
734 gimple_return_set_retval (ret_stmt
, retval
);
735 update_stmt (ret_stmt
);
738 /* Subtract COUNT and FREQUENCY from the basic block and it's
741 decrease_profile (basic_block bb
, gcov_type count
, int frequency
)
747 bb
->frequency
-= frequency
;
748 if (bb
->frequency
< 0)
750 if (!single_succ_p (bb
))
752 gcc_assert (!EDGE_COUNT (bb
->succs
));
755 e
= single_succ_edge (bb
);
761 /* Returns true if argument PARAM of the tail recursive call needs to be copied
762 when the call is eliminated. */
765 arg_needs_copy_p (tree param
)
769 if (!is_gimple_reg (param
) || !var_ann (param
))
772 /* Parameters that are only defined but never used need not be copied. */
773 def
= gimple_default_def (cfun
, param
);
780 /* Eliminates tail call described by T. TMP_VARS is a list of
781 temporary variables used to copy the function arguments. */
784 eliminate_tail_call (struct tailcall
*t
)
790 basic_block bb
, first
;
793 gimple_stmt_iterator gsi
;
796 stmt
= orig_stmt
= gsi_stmt (t
->call_gsi
);
797 bb
= gsi_bb (t
->call_gsi
);
799 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
801 fprintf (dump_file
, "Eliminated tail recursion in bb %d : ",
803 print_gimple_stmt (dump_file
, stmt
, 0, TDF_SLIM
);
804 fprintf (dump_file
, "\n");
807 gcc_assert (is_gimple_call (stmt
));
809 first
= single_succ (ENTRY_BLOCK_PTR
);
811 /* Remove the code after call_gsi that will become unreachable. The
812 possibly unreachable code in other blocks is removed later in
816 while (!gsi_end_p (gsi
))
818 gimple t
= gsi_stmt (gsi
);
819 /* Do not remove the return statement, so that redirect_edge_and_branch
820 sees how the block ends. */
821 if (gimple_code (t
) == GIMPLE_RETURN
)
824 gsi_remove (&gsi
, true);
828 /* Number of executions of function has reduced by the tailcall. */
829 e
= single_succ_edge (gsi_bb (t
->call_gsi
));
830 decrease_profile (EXIT_BLOCK_PTR
, e
->count
, EDGE_FREQUENCY (e
));
831 decrease_profile (ENTRY_BLOCK_PTR
, e
->count
, EDGE_FREQUENCY (e
));
832 if (e
->dest
!= EXIT_BLOCK_PTR
)
833 decrease_profile (e
->dest
, e
->count
, EDGE_FREQUENCY (e
));
835 /* Replace the call by a jump to the start of function. */
836 e
= redirect_edge_and_branch (single_succ_edge (gsi_bb (t
->call_gsi
)),
839 PENDING_STMT (e
) = NULL
;
841 /* Add phi node entries for arguments. The ordering of the phi nodes should
842 be the same as the ordering of the arguments. */
843 for (param
= DECL_ARGUMENTS (current_function_decl
),
844 idx
= 0, gsi
= gsi_start_phis (first
);
846 param
= DECL_CHAIN (param
), idx
++)
848 if (!arg_needs_copy_p (param
))
851 arg
= gimple_call_arg (stmt
, idx
);
852 phi
= gsi_stmt (gsi
);
853 gcc_assert (param
== SSA_NAME_VAR (PHI_RESULT (phi
)));
855 add_phi_arg (phi
, arg
, e
, gimple_location (stmt
));
859 /* Update the values of accumulators. */
860 adjust_accumulator_values (t
->call_gsi
, t
->mult
, t
->add
, e
);
862 call
= gsi_stmt (t
->call_gsi
);
863 rslt
= gimple_call_lhs (call
);
864 if (rslt
!= NULL_TREE
)
866 /* Result of the call will no longer be defined. So adjust the
867 SSA_NAME_DEF_STMT accordingly. */
868 SSA_NAME_DEF_STMT (rslt
) = gimple_build_nop ();
871 gsi_remove (&t
->call_gsi
, true);
875 /* Add phi nodes for the virtual operands defined in the function to the
876 header of the loop created by tail recursion elimination.
878 Originally, we used to add phi nodes only for call clobbered variables,
879 as the value of the non-call clobbered ones obviously cannot be used
880 or changed within the recursive call. However, the local variables
881 from multiple calls now share the same location, so the virtual ssa form
882 requires us to say that the location dies on further iterations of the loop,
883 which requires adding phi nodes.
886 add_virtual_phis (void)
888 referenced_var_iterator rvi
;
891 /* The problematic part is that there is no way how to know what
892 to put into phi nodes (there in fact does not have to be such
893 ssa name available). A solution would be to have an artificial
894 use/kill for all virtual operands in EXIT node. Unless we have
895 this, we cannot do much better than to rebuild the ssa form for
896 possibly affected virtual ssa names from scratch. */
898 FOR_EACH_REFERENCED_VAR (cfun
, var
, rvi
)
900 if (!is_gimple_reg (var
) && gimple_default_def (cfun
, var
) != NULL_TREE
)
901 mark_sym_for_renaming (var
);
905 /* Optimizes the tailcall described by T. If OPT_TAILCALLS is true, also
906 mark the tailcalls for the sibcall optimization. */
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 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
924 fprintf (dump_file
, "Found tail call ");
925 print_gimple_stmt (dump_file
, stmt
, 0, dump_flags
);
926 fprintf (dump_file
, " in bb %i\n", (gsi_bb (t
->call_gsi
))->index
);
933 /* Creates a tail-call accumulator of the same type as the return type of the
934 current function. LABEL is the name used to creating the temporary
935 variable for the accumulator. The accumulator will be inserted in the
936 phis of a basic block BB with single predecessor with an initial value
937 INIT converted to the current function return type. */
940 create_tailcall_accumulator (const char *label
, basic_block bb
, tree init
)
942 tree ret_type
= TREE_TYPE (DECL_RESULT (current_function_decl
));
943 tree tmp
= create_tmp_reg (ret_type
, label
);
946 add_referenced_var (tmp
);
947 phi
= create_phi_node (tmp
, bb
);
948 /* RET_TYPE can be a float when -ffast-maths is enabled. */
949 add_phi_arg (phi
, fold_convert (ret_type
, init
), single_pred_edge (bb
),
951 return PHI_RESULT (phi
);
954 /* Optimizes tail calls in the function, turning the tail recursion
958 tree_optimize_tail_calls_1 (bool opt_tailcalls
)
961 bool phis_constructed
= false;
962 struct tailcall
*tailcalls
= NULL
, *act
, *next
;
963 bool changed
= false;
964 basic_block first
= single_succ (ENTRY_BLOCK_PTR
);
969 if (!suitable_for_tail_opt_p ())
972 opt_tailcalls
= suitable_for_tail_call_opt_p ();
974 FOR_EACH_EDGE (e
, ei
, EXIT_BLOCK_PTR
->preds
)
976 /* Only traverse the normal exits, i.e. those that end with return
978 stmt
= last_stmt (e
->src
);
981 && gimple_code (stmt
) == GIMPLE_RETURN
)
982 find_tail_calls (e
->src
, &tailcalls
);
985 /* Construct the phi nodes and accumulators if necessary. */
986 a_acc
= m_acc
= NULL_TREE
;
987 for (act
= tailcalls
; act
; act
= act
->next
)
989 if (!act
->tail_recursion
)
992 if (!phis_constructed
)
994 /* Ensure that there is only one predecessor of the block
995 or if there are existing degenerate PHI nodes. */
996 if (!single_pred_p (first
)
997 || !gimple_seq_empty_p (phi_nodes (first
)))
998 first
= split_edge (single_succ_edge (ENTRY_BLOCK_PTR
));
1000 /* Copy the args if needed. */
1001 for (param
= DECL_ARGUMENTS (current_function_decl
);
1003 param
= DECL_CHAIN (param
))
1004 if (arg_needs_copy_p (param
))
1006 tree name
= gimple_default_def (cfun
, param
);
1007 tree new_name
= make_ssa_name (param
, SSA_NAME_DEF_STMT (name
));
1010 set_default_def (param
, new_name
);
1011 phi
= create_phi_node (name
, first
);
1012 SSA_NAME_DEF_STMT (name
) = phi
;
1013 add_phi_arg (phi
, new_name
, single_pred_edge (first
),
1014 EXPR_LOCATION (param
));
1016 phis_constructed
= true;
1019 if (act
->add
&& !a_acc
)
1020 a_acc
= create_tailcall_accumulator ("add_acc", first
,
1023 if (act
->mult
&& !m_acc
)
1024 m_acc
= create_tailcall_accumulator ("mult_acc", first
,
1030 /* When the tail call elimination using accumulators is performed,
1031 statements adding the accumulated value are inserted at all exits.
1032 This turns all other tail calls to non-tail ones. */
1033 opt_tailcalls
= false;
1036 for (; tailcalls
; tailcalls
= next
)
1038 next
= tailcalls
->next
;
1039 changed
|= optimize_tail_call (tailcalls
, opt_tailcalls
);
1045 /* Modify the remaining return statements. */
1046 FOR_EACH_EDGE (e
, ei
, EXIT_BLOCK_PTR
->preds
)
1048 stmt
= last_stmt (e
->src
);
1051 && gimple_code (stmt
) == GIMPLE_RETURN
)
1052 adjust_return_value (e
->src
, m_acc
, a_acc
);
1057 free_dominance_info (CDI_DOMINATORS
);
1059 if (phis_constructed
)
1060 add_virtual_phis ();
1062 return TODO_cleanup_cfg
| TODO_update_ssa_only_virtuals
;
1067 execute_tail_recursion (void)
1069 return tree_optimize_tail_calls_1 (false);
1073 gate_tail_calls (void)
1075 return flag_optimize_sibling_calls
!= 0 && dbg_cnt (tail_call
);
1079 execute_tail_calls (void)
1081 return tree_optimize_tail_calls_1 (true);
1084 struct gimple_opt_pass pass_tail_recursion
=
1089 gate_tail_calls
, /* gate */
1090 execute_tail_recursion
, /* execute */
1093 0, /* static_pass_number */
1094 TV_NONE
, /* tv_id */
1095 PROP_cfg
| PROP_ssa
, /* properties_required */
1096 0, /* properties_provided */
1097 0, /* properties_destroyed */
1098 0, /* todo_flags_start */
1099 TODO_verify_ssa
/* todo_flags_finish */
1103 struct gimple_opt_pass pass_tail_calls
=
1108 gate_tail_calls
, /* gate */
1109 execute_tail_calls
, /* execute */
1112 0, /* static_pass_number */
1113 TV_NONE
, /* tv_id */
1114 PROP_cfg
| PROP_ssa
, /* properties_required */
1115 0, /* properties_provided */
1116 0, /* properties_destroyed */
1117 0, /* todo_flags_start */
1118 TODO_verify_ssa
/* todo_flags_finish */