1 /* Generic sibling call optimization support
2 Copyright (C) 1999, 2000, 2001, 2002, 2003
3 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 2, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to the Free
19 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
24 #include "coretypes.h"
30 #include "hard-reg-set.h"
32 #include "insn-config.h"
34 #include "basic-block.h"
39 /* In case alternate_exit_block contains copy from pseudo, to return value,
40 record the pseudo here. In such case the pseudo must be set to function
41 return in the sibcall sequence. */
42 static rtx return_value_pseudo;
44 static int identify_call_return_value (rtx, rtx *, rtx *);
45 static rtx skip_copy_to_return_value (rtx);
46 static rtx skip_use_of_return_value (rtx, enum rtx_code);
47 static rtx skip_stack_adjustment (rtx);
48 static rtx skip_pic_restore (rtx);
49 static rtx skip_jump_insn (rtx);
50 static int call_ends_block_p (rtx, rtx);
51 static int uses_addressof (rtx);
52 static int sequence_uses_addressof (rtx);
53 static void purge_reg_equiv_notes (void);
54 static void purge_mem_unchanging_flag (rtx);
55 static rtx skip_unreturned_value (rtx);
57 /* Examine a CALL_PLACEHOLDER pattern and determine where the call's
58 return value is located. P_HARD_RETURN receives the hard register
59 that the function used; P_SOFT_RETURN receives the pseudo register
60 that the sequence used. Return nonzero if the values were located. */
63 identify_call_return_value (rtx cp, rtx *p_hard_return, rtx *p_soft_return)
65 rtx insn, set, hard, soft;
68 /* Search backward through the "normal" call sequence to the CALL insn. */
69 while (NEXT_INSN (insn))
70 insn = NEXT_INSN (insn);
71 while (GET_CODE (insn) != CALL_INSN)
72 insn = PREV_INSN (insn);
74 /* Assume the pattern is (set (dest) (call ...)), or that the first
75 member of a parallel is. This is the hard return register used
77 if (GET_CODE (PATTERN (insn)) == SET
78 && GET_CODE (SET_SRC (PATTERN (insn))) == CALL)
79 hard = SET_DEST (PATTERN (insn));
80 else if (GET_CODE (PATTERN (insn)) == PARALLEL
81 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET
82 && GET_CODE (SET_SRC (XVECEXP (PATTERN (insn), 0, 0))) == CALL)
83 hard = SET_DEST (XVECEXP (PATTERN (insn), 0, 0));
87 /* If we didn't get a single hard register (e.g. a parallel), give up. */
88 if (GET_CODE (hard) != REG)
91 /* Stack adjustment done after call may appear here. */
92 insn = skip_stack_adjustment (insn);
96 /* Restore of GP register may appear here. */
97 insn = skip_pic_restore (insn);
101 /* If there's nothing after, there's no soft return value. */
102 insn = NEXT_INSN (insn);
106 /* We're looking for a source of the hard return register. */
107 set = single_set (insn);
108 if (! set || SET_SRC (set) != hard)
111 soft = SET_DEST (set);
112 insn = NEXT_INSN (insn);
114 /* Allow this first destination to be copied to a second register,
115 as might happen if the first register wasn't the particular pseudo
116 we'd been expecting. */
118 && (set = single_set (insn)) != NULL_RTX
119 && SET_SRC (set) == soft)
121 soft = SET_DEST (set);
122 insn = NEXT_INSN (insn);
125 /* Don't fool with anything but pseudo registers. */
126 if (GET_CODE (soft) != REG || REGNO (soft) < FIRST_PSEUDO_REGISTER)
129 /* This value must not be modified before the end of the sequence. */
130 if (reg_set_between_p (soft, insn, NULL_RTX))
133 *p_hard_return = hard;
134 *p_soft_return = soft;
139 /* If the first real insn after ORIG_INSN copies to this function's
140 return value from RETVAL, then return the insn which performs the
141 copy. Otherwise return ORIG_INSN. */
144 skip_copy_to_return_value (rtx orig_insn)
146 rtx insn, set = NULL_RTX;
147 rtx hardret, softret;
149 /* If there is no return value, we have nothing to do. */
150 if (! identify_call_return_value (PATTERN (orig_insn), &hardret, &softret))
153 insn = next_nonnote_insn (orig_insn);
157 set = single_set (insn);
161 if (return_value_pseudo)
163 if (SET_DEST (set) == return_value_pseudo
164 && SET_SRC (set) == softret)
169 /* The destination must be the same as the called function's return
170 value to ensure that any return value is put in the same place by the
171 current function and the function we're calling.
173 Further, the source must be the same as the pseudo into which the
174 called function's return value was copied. Otherwise we're returning
177 #ifndef OUTGOING_REGNO
178 #define OUTGOING_REGNO(N) (N)
181 if (SET_DEST (set) == current_function_return_rtx
182 && REG_P (SET_DEST (set))
183 && OUTGOING_REGNO (REGNO (SET_DEST (set))) == REGNO (hardret)
184 && SET_SRC (set) == softret)
187 /* Recognize the situation when the called function's return value
188 is copied in two steps: first into an intermediate pseudo, then
189 the into the calling functions return value register. */
191 if (REG_P (SET_DEST (set))
192 && SET_SRC (set) == softret)
194 rtx x = SET_DEST (set);
196 insn = next_nonnote_insn (insn);
200 set = single_set (insn);
204 if (SET_DEST (set) == current_function_return_rtx
205 && REG_P (SET_DEST (set))
206 && OUTGOING_REGNO (REGNO (SET_DEST (set))) == REGNO (hardret)
207 && SET_SRC (set) == x)
211 /* It did not look like a copy of the return value, so return the
212 same insn we were passed. */
216 /* If the first real insn after ORIG_INSN is a CODE of this function's return
217 value, return insn. Otherwise return ORIG_INSN. */
220 skip_use_of_return_value (rtx orig_insn, enum rtx_code code)
224 insn = next_nonnote_insn (orig_insn);
227 && GET_CODE (insn) == INSN
228 && GET_CODE (PATTERN (insn)) == code
229 && (XEXP (PATTERN (insn), 0) == current_function_return_rtx
230 || XEXP (PATTERN (insn), 0) == const0_rtx))
236 /* In case function does not return value, we get clobber of pseudo followed
237 by set to hard return value. */
239 skip_unreturned_value (rtx orig_insn)
241 rtx insn = next_nonnote_insn (orig_insn);
243 /* Skip possible clobber of pseudo return register. */
245 && GET_CODE (insn) == INSN
246 && GET_CODE (PATTERN (insn)) == CLOBBER
247 && REG_P (XEXP (PATTERN (insn), 0))
248 && (REGNO (XEXP (PATTERN (insn), 0)) >= FIRST_PSEUDO_REGISTER))
250 rtx set_insn = next_nonnote_insn (insn);
254 set = single_set (set_insn);
256 || SET_SRC (set) != XEXP (PATTERN (insn), 0)
257 || SET_DEST (set) != current_function_return_rtx)
264 /* If the first real insn after ORIG_INSN adjusts the stack pointer
265 by a constant, return the insn with the stack pointer adjustment.
266 Otherwise return ORIG_INSN. */
269 skip_stack_adjustment (rtx orig_insn)
271 rtx insn, set = NULL_RTX;
273 insn = next_nonnote_insn (orig_insn);
276 set = single_set (insn);
280 && GET_CODE (SET_SRC (set)) == PLUS
281 && XEXP (SET_SRC (set), 0) == stack_pointer_rtx
282 && GET_CODE (XEXP (SET_SRC (set), 1)) == CONST_INT
283 && SET_DEST (set) == stack_pointer_rtx)
289 /* If the first real insn after ORIG_INSN sets the pic register,
290 return it. Otherwise return ORIG_INSN. */
293 skip_pic_restore (rtx orig_insn)
295 rtx insn, set = NULL_RTX;
297 insn = next_nonnote_insn (orig_insn);
300 set = single_set (insn);
302 if (insn && set && SET_DEST (set) == pic_offset_table_rtx)
308 /* If the first real insn after ORIG_INSN is a jump, return the JUMP_INSN.
309 Otherwise return ORIG_INSN. */
312 skip_jump_insn (rtx orig_insn)
316 insn = next_nonnote_insn (orig_insn);
319 && GET_CODE (insn) == JUMP_INSN
320 && any_uncondjump_p (insn))
326 /* Using the above functions, see if INSN, skipping any of the above,
327 goes all the way to END, the end of a basic block. Return 1 if so. */
330 call_ends_block_p (rtx insn, rtx end)
334 /* END might be a note, so get the last nonnote insn of the block. */
336 end = prev_nonnote_insn (end);
338 /* If the call was the end of the block, then we're OK. */
342 /* Skip over copying from the call's return value pseudo into
343 this function's hard return register and if that's the end
344 of the block, we're OK. */
345 new_insn = skip_copy_to_return_value (insn);
347 /* In case we return value in pseudo, we must set the pseudo to
348 return value of called function, otherwise we are returning
350 if (return_value_pseudo && insn == new_insn)
357 /* Skip any stack adjustment. */
358 insn = skip_stack_adjustment (insn);
362 /* Skip over a CLOBBER of the return value as a hard reg. */
363 insn = skip_use_of_return_value (insn, CLOBBER);
367 /* Skip over a CLOBBER of the return value as a hard reg. */
368 insn = skip_unreturned_value (insn);
372 /* Skip over a USE of the return value (as a hard reg). */
373 insn = skip_use_of_return_value (insn, USE);
377 /* Skip over a JUMP_INSN at the end of the block. If that doesn't end the
378 block, the original CALL_INSN didn't. */
379 insn = skip_jump_insn (insn);
383 /* Scan the rtx X for ADDRESSOF expressions or
384 current_function_internal_arg_pointer registers.
385 Return nonzero if an ADDRESSOF or current_function_internal_arg_pointer
386 is found outside of some MEM expression, else return zero. */
389 uses_addressof (rtx x)
400 if (code == ADDRESSOF || x == current_function_internal_arg_pointer)
406 /* Scan all subexpressions. */
407 fmt = GET_RTX_FORMAT (code);
408 for (i = 0; i < GET_RTX_LENGTH (code); i++, fmt++)
412 if (uses_addressof (XEXP (x, i)))
415 else if (*fmt == 'E')
417 for (j = 0; j < XVECLEN (x, i); j++)
418 if (uses_addressof (XVECEXP (x, i, j)))
425 /* Scan the sequence of insns in SEQ to see if any have an ADDRESSOF
426 rtl expression or current_function_internal_arg_pointer occurrences
427 not enclosed within a MEM. If an ADDRESSOF expression or
428 current_function_internal_arg_pointer is found, return nonzero, otherwise
431 This function handles CALL_PLACEHOLDERs which contain multiple sequences
435 sequence_uses_addressof (rtx seq)
439 for (insn = seq; insn; insn = NEXT_INSN (insn))
442 /* If this is a CALL_PLACEHOLDER, then recursively call ourselves
443 with each nonempty sequence attached to the CALL_PLACEHOLDER. */
444 if (GET_CODE (insn) == CALL_INSN
445 && GET_CODE (PATTERN (insn)) == CALL_PLACEHOLDER)
447 if (XEXP (PATTERN (insn), 0) != NULL_RTX
448 && sequence_uses_addressof (XEXP (PATTERN (insn), 0)))
450 if (XEXP (PATTERN (insn), 1) != NULL_RTX
451 && sequence_uses_addressof (XEXP (PATTERN (insn), 1)))
453 if (XEXP (PATTERN (insn), 2) != NULL_RTX
454 && sequence_uses_addressof (XEXP (PATTERN (insn), 2)))
457 else if (uses_addressof (PATTERN (insn))
458 || (REG_NOTES (insn) && uses_addressof (REG_NOTES (insn))))
464 /* Remove all REG_EQUIV notes found in the insn chain. */
467 purge_reg_equiv_notes (void)
471 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
475 rtx note = find_reg_note (insn, REG_EQUIV, 0);
478 /* Remove the note and keep looking at the notes for
480 remove_note (insn, note);
488 /* Clear RTX_UNCHANGING_P flag of incoming argument MEMs. */
491 purge_mem_unchanging_flag (rtx x)
504 if (RTX_UNCHANGING_P (x)
505 && (XEXP (x, 0) == current_function_internal_arg_pointer
506 || (GET_CODE (XEXP (x, 0)) == PLUS
507 && XEXP (XEXP (x, 0), 0) ==
508 current_function_internal_arg_pointer
509 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT)))
510 RTX_UNCHANGING_P (x) = 0;
514 /* Scan all subexpressions. */
515 fmt = GET_RTX_FORMAT (code);
516 for (i = 0; i < GET_RTX_LENGTH (code); i++, fmt++)
519 purge_mem_unchanging_flag (XEXP (x, i));
520 else if (*fmt == 'E')
521 for (j = 0; j < XVECLEN (x, i); j++)
522 purge_mem_unchanging_flag (XVECEXP (x, i, j));
526 /* Replace the CALL_PLACEHOLDER with one of its children. INSN should be
527 the CALL_PLACEHOLDER insn; USE tells which child to use. */
530 replace_call_placeholder (rtx insn, sibcall_use_t use)
532 if (use == sibcall_use_tail_recursion)
533 emit_insn_before (XEXP (PATTERN (insn), 2), insn);
534 else if (use == sibcall_use_sibcall)
535 emit_insn_before (XEXP (PATTERN (insn), 1), insn);
536 else if (use == sibcall_use_normal)
537 emit_insn_before (XEXP (PATTERN (insn), 0), insn);
541 /* Turn off LABEL_PRESERVE_P for the tail recursion label if it
542 exists. We only had to set it long enough to keep the jump
543 pass above from deleting it as unused. */
544 if (XEXP (PATTERN (insn), 3))
545 LABEL_PRESERVE_P (XEXP (PATTERN (insn), 3)) = 0;
547 /* "Delete" the placeholder insn. */
551 /* Given a (possibly empty) set of potential sibling or tail recursion call
552 sites, determine if optimization is possible.
554 Potential sibling or tail recursion calls are marked with CALL_PLACEHOLDER
555 insns. The CALL_PLACEHOLDER insn holds chains of insns to implement a
556 normal call, sibling call or tail recursive call.
558 Replace the CALL_PLACEHOLDER with an appropriate insn chain. */
561 optimize_sibling_and_tail_recursive_calls (void)
564 basic_block alternate_exit = EXIT_BLOCK_PTR;
565 bool no_sibcalls_this_function = false;
566 bool successful_replacement = false;
567 bool replaced_call_placeholder = false;
570 insns = get_insns ();
572 cleanup_cfg (CLEANUP_PRE_SIBCALL | CLEANUP_PRE_LOOP);
574 /* If there are no basic blocks, then there is nothing to do. */
575 if (n_basic_blocks == 0)
578 /* If we are using sjlj exceptions, we may need to add a call to
579 _Unwind_SjLj_Unregister at exit of the function. Which means
580 that we cannot do any sibcall transformations. */
581 if (USING_SJLJ_EXCEPTIONS && current_function_has_exception_handlers ())
582 no_sibcalls_this_function = true;
584 return_value_pseudo = NULL_RTX;
586 /* Find the exit block.
588 It is possible that we have blocks which can reach the exit block
589 directly. However, most of the time a block will jump (or fall into)
590 N_BASIC_BLOCKS - 1, which in turn falls into the exit block. */
591 for (e = EXIT_BLOCK_PTR->pred;
592 e && alternate_exit == EXIT_BLOCK_PTR;
597 if (e->dest != EXIT_BLOCK_PTR || e->succ_next != NULL)
600 /* Walk forwards through the last normal block and see if it
601 does nothing except fall into the exit block. */
602 for (insn = BB_HEAD (EXIT_BLOCK_PTR->prev_bb);
604 insn = NEXT_INSN (insn))
607 /* This should only happen once, at the start of this block. */
608 if (GET_CODE (insn) == CODE_LABEL)
611 if (GET_CODE (insn) == NOTE)
614 if (GET_CODE (insn) == INSN
615 && GET_CODE (PATTERN (insn)) == USE)
618 /* Exit block also may contain copy from pseudo containing
619 return value to hard register. */
620 if (GET_CODE (insn) == INSN
621 && (set = single_set (insn))
622 && SET_DEST (set) == current_function_return_rtx
623 && REG_P (SET_SRC (set))
624 && !return_value_pseudo)
626 return_value_pseudo = SET_SRC (set);
633 /* If INSN is zero, then the search walked all the way through the
634 block without hitting anything interesting. This block is a
635 valid alternate exit block. */
637 alternate_exit = e->src;
639 return_value_pseudo = NULL;
642 /* If the function uses ADDRESSOF, we can't (easily) determine
643 at this point if the value will end up on the stack. */
644 no_sibcalls_this_function |= sequence_uses_addressof (insns);
646 /* Walk the insn chain and find any CALL_PLACEHOLDER insns. We need to
647 select one of the insn sequences attached to each CALL_PLACEHOLDER.
649 The different sequences represent different ways to implement the call,
650 ie, tail recursion, sibling call or normal call.
652 Since we do not create nested CALL_PLACEHOLDERs, the scan
653 continues with the insn that was after a replaced CALL_PLACEHOLDER;
654 we don't rescan the replacement insns. */
655 for (insn = insns; insn; insn = NEXT_INSN (insn))
657 if (GET_CODE (insn) == CALL_INSN
658 && GET_CODE (PATTERN (insn)) == CALL_PLACEHOLDER)
660 int sibcall = (XEXP (PATTERN (insn), 1) != NULL_RTX);
661 int tailrecursion = (XEXP (PATTERN (insn), 2) != NULL_RTX);
662 basic_block call_block = BLOCK_FOR_INSN (insn);
664 /* alloca (until we have stack slot life analysis) inhibits
665 sibling call optimizations, but not tail recursion.
666 Similarly if we use varargs or stdarg since they implicitly
667 may take the address of an argument. */
668 if (current_function_calls_alloca || current_function_stdarg)
671 /* See if there are any reasons we can't perform either sibling or
672 tail call optimizations. We must be careful with stack slots
673 which are live at potential optimization sites. */
674 if (no_sibcalls_this_function
675 /* ??? Overly conservative. */
677 /* Any function that calls setjmp might have longjmp called from
678 any called function. ??? We really should represent this
679 properly in the CFG so that this needn't be special cased. */
680 || current_function_calls_setjmp
681 /* Can't if more than one successor or single successor is not
682 exit block. These two tests prevent tail call optimization
683 in the presence of active exception handlers. */
684 || call_block->succ == NULL
685 || call_block->succ->succ_next != NULL
686 || (call_block->succ->dest != EXIT_BLOCK_PTR
687 && call_block->succ->dest != alternate_exit)
688 /* If this call doesn't end the block, there are operations at
689 the end of the block which we must execute after returning. */
690 || ! call_ends_block_p (insn, BB_END (call_block)))
691 sibcall = 0, tailrecursion = 0;
693 /* Select a set of insns to implement the call and emit them.
694 Tail recursion is the most efficient, so select it over
695 a tail/sibling call. */
697 if (sibcall || tailrecursion)
698 successful_replacement = true;
699 replaced_call_placeholder = true;
701 replace_call_placeholder (insn,
703 ? sibcall_use_tail_recursion
705 ? sibcall_use_sibcall
706 : sibcall_use_normal);
710 if (successful_replacement)
715 /* A sibling call sequence invalidates any REG_EQUIV notes made for
716 this function's incoming arguments.
718 At the start of RTL generation we know the only REG_EQUIV notes
719 in the rtl chain are those for incoming arguments, so we can safely
720 flush any REG_EQUIV note.
722 This is (slight) overkill. We could keep track of the highest
723 argument we clobber and be more selective in removing notes, but it
724 does not seem to be worth the effort. */
725 purge_reg_equiv_notes ();
727 /* A sibling call sequence also may invalidate RTX_UNCHANGING_P
728 flag of some incoming arguments MEM RTLs, because it can write into
729 those slots. We clear all those bits now.
731 This is (slight) overkill, we could keep track of which arguments
732 we actually write into. */
733 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
736 purge_mem_unchanging_flag (PATTERN (insn));
739 /* Similarly, invalidate RTX_UNCHANGING_P for any incoming
740 arguments passed in registers. */
741 for (arg = DECL_ARGUMENTS (current_function_decl);
743 arg = TREE_CHAIN (arg))
745 if (REG_P (DECL_RTL (arg)))
746 RTX_UNCHANGING_P (DECL_RTL (arg)) = false;
750 /* There may have been NOTE_INSN_BLOCK_{BEGIN,END} notes in the
751 CALL_PLACEHOLDER alternatives that we didn't emit. Rebuild the
752 lexical block tree to correspond to the notes that still exist. */
753 if (replaced_call_placeholder)
756 /* This information will be invalid after inline expansion. Kill it now. */
757 free_basic_block_vars (0);
758 free_EXPR_LIST_list (&tail_recursion_label_list);