1 /* Generic sibling call optimization support
2 Copyright (C) 1999, 2000, 2001 Free Software Foundation, Inc.
4 This file is part of GNU CC.
6 GNU CC 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 2, or (at your option)
11 GNU CC 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 GNU CC; see the file COPYING. If not, write to
18 the Free Software Foundation, 59 Temple Place - Suite 330,
19 Boston, MA 02111-1307, USA. */
27 #include "hard-reg-set.h"
29 #include "insn-config.h"
31 #include "basic-block.h"
35 static int identify_call_return_value PARAMS ((rtx, rtx *, rtx *));
36 static rtx skip_copy_to_return_value PARAMS ((rtx, rtx, rtx));
37 static rtx skip_use_of_return_value PARAMS ((rtx, enum rtx_code));
38 static rtx skip_stack_adjustment PARAMS ((rtx));
39 static rtx skip_pic_restore PARAMS ((rtx));
40 static rtx skip_jump_insn PARAMS ((rtx));
41 static int call_ends_block_p PARAMS ((rtx, rtx));
42 static int uses_addressof PARAMS ((rtx));
43 static int sequence_uses_addressof PARAMS ((rtx));
44 static void purge_reg_equiv_notes PARAMS ((void));
45 static void purge_mem_unchanging_flag PARAMS ((rtx));
47 /* Examine a CALL_PLACEHOLDER pattern and determine where the call's
48 return value is located. P_HARD_RETURN receives the hard register
49 that the function used; P_SOFT_RETURN receives the pseudo register
50 that the sequence used. Return non-zero if the values were located. */
53 identify_call_return_value (cp, p_hard_return, p_soft_return)
55 rtx *p_hard_return, *p_soft_return;
57 rtx insn, set, hard, soft;
60 /* Search backward through the "normal" call sequence to the CALL insn. */
61 while (NEXT_INSN (insn))
62 insn = NEXT_INSN (insn);
63 while (GET_CODE (insn) != CALL_INSN)
64 insn = PREV_INSN (insn);
66 /* Assume the pattern is (set (dest) (call ...)), or that the first
67 member of a parallel is. This is the hard return register used
69 if (GET_CODE (PATTERN (insn)) == SET
70 && GET_CODE (SET_SRC (PATTERN (insn))) == CALL)
71 hard = SET_DEST (PATTERN (insn));
72 else if (GET_CODE (PATTERN (insn)) == PARALLEL
73 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET
74 && GET_CODE (SET_SRC (XVECEXP (PATTERN (insn), 0, 0))) == CALL)
75 hard = SET_DEST (XVECEXP (PATTERN (insn), 0, 0));
79 /* If we didn't get a single hard register (e.g. a parallel), give up. */
80 if (GET_CODE (hard) != REG)
83 /* Stack adjustment done after call may appear here. */
84 insn = skip_stack_adjustment (insn);
88 /* Restore of GP register may appear here. */
89 insn = skip_pic_restore (insn);
93 /* If there's nothing after, there's no soft return value. */
94 insn = NEXT_INSN (insn);
98 /* We're looking for a source of the hard return register. */
99 set = single_set (insn);
100 if (! set || SET_SRC (set) != hard)
103 soft = SET_DEST (set);
104 insn = NEXT_INSN (insn);
106 /* Allow this first destination to be copied to a second register,
107 as might happen if the first register wasn't the particular pseudo
108 we'd been expecting. */
110 && (set = single_set (insn)) != NULL_RTX
111 && SET_SRC (set) == soft)
113 soft = SET_DEST (set);
114 insn = NEXT_INSN (insn);
117 /* Don't fool with anything but pseudo registers. */
118 if (GET_CODE (soft) != REG || REGNO (soft) < FIRST_PSEUDO_REGISTER)
121 /* This value must not be modified before the end of the sequence. */
122 if (reg_set_between_p (soft, insn, NULL_RTX))
125 *p_hard_return = hard;
126 *p_soft_return = soft;
131 /* If the first real insn after ORIG_INSN copies to this function's
132 return value from RETVAL, then return the insn which performs the
133 copy. Otherwise return ORIG_INSN. */
136 skip_copy_to_return_value (orig_insn, hardret, softret)
138 rtx hardret, softret;
140 rtx insn, set = NULL_RTX;
142 insn = next_nonnote_insn (orig_insn);
146 set = single_set (insn);
150 /* The destination must be the same as the called function's return
151 value to ensure that any return value is put in the same place by the
152 current function and the function we're calling.
154 Further, the source must be the same as the pseudo into which the
155 called function's return value was copied. Otherwise we're returning
158 #ifndef OUTGOING_REGNO
159 #define OUTGOING_REGNO(N) (N)
162 if (SET_DEST (set) == current_function_return_rtx
163 && REG_P (SET_DEST (set))
164 && OUTGOING_REGNO (REGNO (SET_DEST (set))) == REGNO (hardret)
165 && SET_SRC (set) == softret)
168 /* It did not look like a copy of the return value, so return the
169 same insn we were passed. */
173 /* If the first real insn after ORIG_INSN is a CODE of this function's return
174 value, return insn. Otherwise return ORIG_INSN. */
177 skip_use_of_return_value (orig_insn, code)
183 insn = next_nonnote_insn (orig_insn);
186 && GET_CODE (insn) == INSN
187 && GET_CODE (PATTERN (insn)) == code
188 && (XEXP (PATTERN (insn), 0) == current_function_return_rtx
189 || XEXP (PATTERN (insn), 0) == const0_rtx))
195 /* If the first real insn after ORIG_INSN adjusts the stack pointer
196 by a constant, return the insn with the stack pointer adjustment.
197 Otherwise return ORIG_INSN. */
200 skip_stack_adjustment (orig_insn)
203 rtx insn, set = NULL_RTX;
205 insn = next_nonnote_insn (orig_insn);
208 set = single_set (insn);
212 && GET_CODE (SET_SRC (set)) == PLUS
213 && XEXP (SET_SRC (set), 0) == stack_pointer_rtx
214 && GET_CODE (XEXP (SET_SRC (set), 1)) == CONST_INT
215 && SET_DEST (set) == stack_pointer_rtx)
221 /* If the first real insn after ORIG_INSN sets the pic register,
222 return it. Otherwise return ORIG_INSN. */
225 skip_pic_restore (orig_insn)
228 rtx insn, set = NULL_RTX;
230 insn = next_nonnote_insn (orig_insn);
233 set = single_set (insn);
235 if (insn && set && SET_DEST (set) == pic_offset_table_rtx)
241 /* If the first real insn after ORIG_INSN is a jump, return the JUMP_INSN.
242 Otherwise return ORIG_INSN. */
245 skip_jump_insn (orig_insn)
250 insn = next_nonnote_insn (orig_insn);
253 && GET_CODE (insn) == JUMP_INSN
254 && any_uncondjump_p (insn))
260 /* Using the above functions, see if INSN, skipping any of the above,
261 goes all the way to END, the end of a basic block. Return 1 if so. */
264 call_ends_block_p (insn, end)
268 rtx hardret, softret;
270 /* END might be a note, so get the last nonnote insn of the block. */
271 end = next_nonnote_insn (PREV_INSN (end));
273 /* If the call was the end of the block, then we're OK. */
277 /* Skip over copying from the call's return value pseudo into
278 this function's hard return register and if that's the end
279 of the block, we're OK. */
280 identify_call_return_value (PATTERN (insn), &hardret, &softret);
281 insn = skip_copy_to_return_value (insn, hardret, softret);
285 /* Skip any stack adjustment. */
286 insn = skip_stack_adjustment (insn);
290 /* Skip over a CLOBBER of the return value as a hard reg. */
291 insn = skip_use_of_return_value (insn, CLOBBER);
295 /* Skip over a USE of the return value (as a hard reg). */
296 insn = skip_use_of_return_value (insn, USE);
300 /* Skip over a JUMP_INSN at the end of the block. If that doesn't end the
301 block, the original CALL_INSN didn't. */
302 insn = skip_jump_insn (insn);
306 /* Scan the rtx X for ADDRESSOF expressions or
307 current_function_internal_arg_pointer registers.
308 Return nonzero if an ADDRESSOF or current_function_internal_arg_pointer
309 is found outside of some MEM expression, else return zero. */
324 if (code == ADDRESSOF || x == current_function_internal_arg_pointer)
330 /* Scan all subexpressions. */
331 fmt = GET_RTX_FORMAT (code);
332 for (i = 0; i < GET_RTX_LENGTH (code); i++, fmt++)
336 if (uses_addressof (XEXP (x, i)))
339 else if (*fmt == 'E')
341 for (j = 0; j < XVECLEN (x, i); j++)
342 if (uses_addressof (XVECEXP (x, i, j)))
349 /* Scan the sequence of insns in SEQ to see if any have an ADDRESSOF
350 rtl expression or current_function_internal_arg_pointer occurences
351 not enclosed within a MEM. If an ADDRESSOF expression or
352 current_function_internal_arg_pointer is found, return nonzero, otherwise
355 This function handles CALL_PLACEHOLDERs which contain multiple sequences
359 sequence_uses_addressof (seq)
364 for (insn = seq; insn; insn = NEXT_INSN (insn))
367 /* If this is a CALL_PLACEHOLDER, then recursively call ourselves
368 with each nonempty sequence attached to the CALL_PLACEHOLDER. */
369 if (GET_CODE (insn) == CALL_INSN
370 && GET_CODE (PATTERN (insn)) == CALL_PLACEHOLDER)
372 if (XEXP (PATTERN (insn), 0) != NULL_RTX
373 && sequence_uses_addressof (XEXP (PATTERN (insn), 0)))
375 if (XEXP (PATTERN (insn), 1) != NULL_RTX
376 && sequence_uses_addressof (XEXP (PATTERN (insn), 1)))
378 if (XEXP (PATTERN (insn), 2) != NULL_RTX
379 && sequence_uses_addressof (XEXP (PATTERN (insn), 2)))
382 else if (uses_addressof (PATTERN (insn))
383 || (REG_NOTES (insn) && uses_addressof (REG_NOTES (insn))))
389 /* Remove all REG_EQUIV notes found in the insn chain. */
392 purge_reg_equiv_notes ()
396 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
400 rtx note = find_reg_note (insn, REG_EQUIV, 0);
403 /* Remove the note and keep looking at the notes for
405 remove_note (insn, note);
413 /* Clear RTX_UNCHANGING_P flag of incoming argument MEMs. */
416 purge_mem_unchanging_flag (x)
430 if (RTX_UNCHANGING_P (x)
431 && (XEXP (x, 0) == current_function_internal_arg_pointer
432 || (GET_CODE (XEXP (x, 0)) == PLUS
433 && XEXP (XEXP (x, 0), 0) ==
434 current_function_internal_arg_pointer
435 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT)))
436 RTX_UNCHANGING_P (x) = 0;
440 /* Scan all subexpressions. */
441 fmt = GET_RTX_FORMAT (code);
442 for (i = 0; i < GET_RTX_LENGTH (code); i++, fmt++)
445 purge_mem_unchanging_flag (XEXP (x, i));
446 else if (*fmt == 'E')
447 for (j = 0; j < XVECLEN (x, i); j++)
448 purge_mem_unchanging_flag (XVECEXP (x, i, j));
452 /* Replace the CALL_PLACEHOLDER with one of its children. INSN should be
453 the CALL_PLACEHOLDER insn; USE tells which child to use. */
456 replace_call_placeholder (insn, use)
460 if (use == sibcall_use_tail_recursion)
461 emit_insns_before (XEXP (PATTERN (insn), 2), insn);
462 else if (use == sibcall_use_sibcall)
463 emit_insns_before (XEXP (PATTERN (insn), 1), insn);
464 else if (use == sibcall_use_normal)
465 emit_insns_before (XEXP (PATTERN (insn), 0), insn);
469 /* Turn off LABEL_PRESERVE_P for the tail recursion label if it
470 exists. We only had to set it long enough to keep the jump
471 pass above from deleting it as unused. */
472 if (XEXP (PATTERN (insn), 3))
473 LABEL_PRESERVE_P (XEXP (PATTERN (insn), 3)) = 0;
475 /* "Delete" the placeholder insn. */
476 PUT_CODE (insn, NOTE);
477 NOTE_SOURCE_FILE (insn) = 0;
478 NOTE_LINE_NUMBER (insn) = NOTE_INSN_DELETED;
481 /* Given a (possibly empty) set of potential sibling or tail recursion call
482 sites, determine if optimization is possible.
484 Potential sibling or tail recursion calls are marked with CALL_PLACEHOLDER
485 insns. The CALL_PLACEHOLDER insn holds chains of insns to implement a
486 normal call, sibling call or tail recursive call.
488 Replace the CALL_PLACEHOLDER with an appropriate insn chain. */
491 optimize_sibling_and_tail_recursive_calls ()
494 basic_block alternate_exit = EXIT_BLOCK_PTR;
495 int current_function_uses_addressof;
496 int successful_sibling_call = 0;
497 int replaced_call_placeholder = 0;
500 insns = get_insns ();
502 /* We do not perform these calls when flag_exceptions is true, so this
503 is probably a NOP at the current time. However, we may want to support
504 sibling and tail recursion optimizations in the future, so let's plan
505 ahead and find all the EH labels. */
506 find_exception_handler_labels ();
508 /* Run a jump optimization pass to clean up the CFG. We primarily want
509 this to thread jumps so that it is obvious which blocks jump to the
511 jump_optimize_minimal (insns);
513 /* We need cfg information to determine which blocks are succeeded
514 only by the epilogue. */
515 find_basic_blocks (insns, max_reg_num (), 0);
518 /* If there are no basic blocks, then there is nothing to do. */
519 if (n_basic_blocks == 0)
522 /* Find the exit block.
524 It is possible that we have blocks which can reach the exit block
525 directly. However, most of the time a block will jump (or fall into)
526 N_BASIC_BLOCKS - 1, which in turn falls into the exit block. */
527 for (e = EXIT_BLOCK_PTR->pred;
528 e && alternate_exit == EXIT_BLOCK_PTR;
533 if (e->dest != EXIT_BLOCK_PTR || e->succ_next != NULL)
536 /* Walk forwards through the last normal block and see if it
537 does nothing except fall into the exit block. */
538 for (insn = BLOCK_HEAD (n_basic_blocks - 1);
540 insn = NEXT_INSN (insn))
542 /* This should only happen once, at the start of this block. */
543 if (GET_CODE (insn) == CODE_LABEL)
546 if (GET_CODE (insn) == NOTE)
549 if (GET_CODE (insn) == INSN
550 && GET_CODE (PATTERN (insn)) == USE)
556 /* If INSN is zero, then the search walked all the way through the
557 block without hitting anything interesting. This block is a
558 valid alternate exit block. */
560 alternate_exit = e->src;
563 /* If the function uses ADDRESSOF, we can't (easily) determine
564 at this point if the value will end up on the stack. */
565 current_function_uses_addressof = sequence_uses_addressof (insns);
567 /* Walk the insn chain and find any CALL_PLACEHOLDER insns. We need to
568 select one of the insn sequences attached to each CALL_PLACEHOLDER.
570 The different sequences represent different ways to implement the call,
571 ie, tail recursion, sibling call or normal call.
573 Since we do not create nested CALL_PLACEHOLDERs, the scan
574 continues with the insn that was after a replaced CALL_PLACEHOLDER;
575 we don't rescan the replacement insns. */
576 for (insn = insns; insn; insn = NEXT_INSN (insn))
578 if (GET_CODE (insn) == CALL_INSN
579 && GET_CODE (PATTERN (insn)) == CALL_PLACEHOLDER)
581 int sibcall = (XEXP (PATTERN (insn), 1) != NULL_RTX);
582 int tailrecursion = (XEXP (PATTERN (insn), 2) != NULL_RTX);
583 basic_block call_block = BLOCK_FOR_INSN (insn);
585 /* alloca (until we have stack slot life analysis) inhibits
586 sibling call optimizations, but not tail recursion.
587 Similarly if we use varargs or stdarg since they implicitly
588 may take the address of an argument. */
589 if (current_function_calls_alloca
590 || current_function_varargs || current_function_stdarg)
593 /* See if there are any reasons we can't perform either sibling or
594 tail call optimizations. We must be careful with stack slots
595 which are live at potential optimization sites. ?!? The first
596 test is overly conservative and should be replaced. */
598 /* Can't take address of local var if used by recursive call. */
599 || current_function_uses_addressof
600 /* Can't if more than one successor or single successor is not
601 exit block. These two tests prevent tail call optimization
602 in the presense of active exception handlers. */
603 || call_block->succ == NULL
604 || call_block->succ->succ_next != NULL
605 || (call_block->succ->dest != EXIT_BLOCK_PTR
606 && call_block->succ->dest != alternate_exit)
607 /* If this call doesn't end the block, there are operations at
608 the end of the block which we must execute after returning. */
609 || ! call_ends_block_p (insn, call_block->end))
610 sibcall = 0, tailrecursion = 0;
612 /* Select a set of insns to implement the call and emit them.
613 Tail recursion is the most efficient, so select it over
614 a tail/sibling call. */
616 successful_sibling_call = 1;
618 replaced_call_placeholder = 1;
619 replace_call_placeholder (insn,
621 ? sibcall_use_tail_recursion
623 ? sibcall_use_sibcall
624 : sibcall_use_normal);
628 if (successful_sibling_call)
632 /* A sibling call sequence invalidates any REG_EQUIV notes made for
633 this function's incoming arguments.
635 At the start of RTL generation we know the only REG_EQUIV notes
636 in the rtl chain are those for incoming arguments, so we can safely
637 flush any REG_EQUIV note.
639 This is (slight) overkill. We could keep track of the highest
640 argument we clobber and be more selective in removing notes, but it
641 does not seem to be worth the effort. */
642 purge_reg_equiv_notes ();
644 /* A sibling call sequence also may invalidate RTX_UNCHANGING_P
645 flag of some incoming arguments MEM RTLs, because it can write into
646 those slots. We clear all those bits now.
648 This is (slight) overkill, we could keep track of which arguments
649 we actually write into. */
650 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
652 if (GET_CODE (insn) == NOTE)
654 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_FUNCTION_BEG)
657 else if (INSN_P (insn))
658 purge_mem_unchanging_flag (PATTERN (insn));
662 /* There may have been NOTE_INSN_BLOCK_{BEGIN,END} notes in the
663 CALL_PLACEHOLDER alternatives that we didn't emit. Rebuild the
664 lexical block tree to correspond to the notes that still exist. */
665 if (replaced_call_placeholder)
668 /* This information will be invalid after inline expansion. Kill it now. */
669 free_basic_block_vars (0);