1 /* Generic sibling call optimization support
2 Copyright (C) 1999, 2000 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_jump_insn PARAMS ((rtx));
40 static int uses_addressof PARAMS ((rtx, int));
41 static int sequence_uses_addressof PARAMS ((rtx));
42 static void purge_reg_equiv_notes PARAMS ((void));
44 /* Examine a CALL_PLACEHOLDER pattern and determine where the call's
45 return value is located. P_HARD_RETURN receives the hard register
46 that the function used; P_SOFT_RETURN receives the pseudo register
47 that the sequence used. Return non-zero if the values were located. */
50 identify_call_return_value (cp, p_hard_return, p_soft_return)
52 rtx *p_hard_return, *p_soft_return;
54 rtx insn, set, hard, soft;
57 /* Search backward through the "normal" call sequence to the CALL insn. */
58 while (NEXT_INSN (insn))
59 insn = NEXT_INSN (insn);
60 while (GET_CODE (insn) != CALL_INSN)
61 insn = PREV_INSN (insn);
63 /* Assume the pattern is (set (dest) (call ...)), or that the first
64 member of a parallel is. This is the hard return register used
66 if (GET_CODE (PATTERN (insn)) == SET
67 && GET_CODE (SET_SRC (PATTERN (insn))) == CALL)
68 hard = SET_DEST (PATTERN (insn));
69 else if (GET_CODE (PATTERN (insn)) == PARALLEL
70 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET
71 && GET_CODE (SET_SRC (XVECEXP (PATTERN (insn), 0, 0))) == CALL)
72 hard = SET_DEST (XVECEXP (PATTERN (insn), 0, 0));
76 /* If we didn't get a single hard register (e.g. a parallel), give up. */
77 if (GET_CODE (hard) != REG)
80 /* Stack adjustment done after call may appear here. */
81 insn = skip_stack_adjustment (insn);
85 /* If there's nothing after, there's no soft return value. */
86 insn = NEXT_INSN (insn);
90 /* We're looking for a source of the hard return register. */
91 set = single_set (insn);
92 if (! set || SET_SRC (set) != hard)
95 soft = SET_DEST (set);
96 insn = NEXT_INSN (insn);
98 /* Allow this first destination to be copied to a second register,
99 as might happen if the first register wasn't the particular pseudo
100 we'd been expecting. */
102 && (set = single_set (insn)) != NULL_RTX
103 && SET_SRC (set) == soft)
105 soft = SET_DEST (set);
106 insn = NEXT_INSN (insn);
109 /* Don't fool with anything but pseudo registers. */
110 if (GET_CODE (soft) != REG || REGNO (soft) < FIRST_PSEUDO_REGISTER)
113 /* This value must not be modified before the end of the sequence. */
114 if (reg_set_between_p (soft, insn, NULL_RTX))
117 *p_hard_return = hard;
118 *p_soft_return = soft;
123 /* If the first real insn after ORIG_INSN copies to this function's
124 return value from RETVAL, then return the insn which performs the
125 copy. Otherwise return ORIG_INSN. */
128 skip_copy_to_return_value (orig_insn, hardret, softret)
130 rtx hardret, softret;
132 rtx insn, set = NULL_RTX;
134 insn = next_nonnote_insn (orig_insn);
138 set = single_set (insn);
142 /* The destination must be the same as the called function's return
143 value to ensure that any return value is put in the same place by the
144 current function and the function we're calling.
146 Further, the source must be the same as the pseudo into which the
147 called function's return value was copied. Otherwise we're returning
150 #ifndef OUTGOING_REGNO
151 #define OUTGOING_REGNO(N) (N)
154 if (SET_DEST (set) == current_function_return_rtx
155 && REG_P (SET_DEST (set))
156 && OUTGOING_REGNO (REGNO (SET_DEST (set))) == REGNO (hardret)
157 && SET_SRC (set) == softret)
160 /* It did not look like a copy of the return value, so return the
161 same insn we were passed. */
165 /* If the first real insn after ORIG_INSN is a CODE of this function's return
166 value, return insn. Otherwise return ORIG_INSN. */
169 skip_use_of_return_value (orig_insn, code)
175 insn = next_nonnote_insn (orig_insn);
178 && GET_CODE (insn) == INSN
179 && GET_CODE (PATTERN (insn)) == code
180 && (XEXP (PATTERN (insn), 0) == current_function_return_rtx
181 || XEXP (PATTERN (insn), 0) == const0_rtx))
187 /* If the first real insn after ORIG_INSN adjusts the stack pointer
188 by a constant, return the insn with the stack pointer adjustment.
189 Otherwise return ORIG_INSN. */
192 skip_stack_adjustment (orig_insn)
195 rtx insn, set = NULL_RTX;
197 insn = next_nonnote_insn (orig_insn);
200 set = single_set (insn);
202 /* The source must be the same as the current function's return value to
203 ensure that any return value is put in the same place by the current
204 function and the function we're calling. The destination register
208 && GET_CODE (SET_SRC (set)) == PLUS
209 && XEXP (SET_SRC (set), 0) == stack_pointer_rtx
210 && GET_CODE (XEXP (SET_SRC (set), 1)) == CONST_INT
211 && SET_DEST (set) == stack_pointer_rtx)
214 /* It did not look like a copy of the return value, so return the
215 same insn we were passed. */
219 /* If the first real insn after ORIG_INSN is a jump, return the JUMP_INSN.
220 Otherwise return ORIG_INSN. */
223 skip_jump_insn (orig_insn)
228 insn = next_nonnote_insn (orig_insn);
231 && GET_CODE (insn) == JUMP_INSN
232 && any_uncondjump_p (insn))
238 /* Scan the rtx X for ADDRESSOF expressions or
239 current_function_internal_arg_pointer registers.
240 INMEM argument should be 1 if we're looking at inner part of some
241 MEM expression, otherwise 0.
242 Return nonzero if an ADDRESSOF expresion is found or if
243 current_function_internal_arg_pointer is found outside of some MEM
244 expression, else return zero. */
247 uses_addressof (x, inmem)
260 if (code == ADDRESSOF)
263 if (x == current_function_internal_arg_pointer && ! inmem)
267 return uses_addressof (XEXP (x, 0), 1);
269 /* Scan all subexpressions. */
270 fmt = GET_RTX_FORMAT (code);
271 for (i = 0; i < GET_RTX_LENGTH (code); i++, fmt++)
275 if (uses_addressof (XEXP (x, i), inmem))
278 else if (*fmt == 'E')
280 for (j = 0; j < XVECLEN (x, i); j++)
281 if (uses_addressof (XVECEXP (x, i, j), inmem))
288 /* Scan the sequence of insns in SEQ to see if any have an ADDRESSOF
289 rtl expression or current_function_internal_arg_pointer occurences
290 not enclosed within a MEM. If an ADDRESSOF expression or
291 current_function_internal_arg_pointer is found, return nonzero, otherwise
294 This function handles CALL_PLACEHOLDERs which contain multiple sequences
298 sequence_uses_addressof (seq)
303 for (insn = seq; insn; insn = NEXT_INSN (insn))
306 /* If this is a CALL_PLACEHOLDER, then recursively call ourselves
307 with each nonempty sequence attached to the CALL_PLACEHOLDER. */
308 if (GET_CODE (insn) == CALL_INSN
309 && GET_CODE (PATTERN (insn)) == CALL_PLACEHOLDER)
311 if (XEXP (PATTERN (insn), 0) != NULL_RTX
312 && sequence_uses_addressof (XEXP (PATTERN (insn), 0)))
314 if (XEXP (PATTERN (insn), 1) != NULL_RTX
315 && sequence_uses_addressof (XEXP (PATTERN (insn), 1)))
317 if (XEXP (PATTERN (insn), 2) != NULL_RTX
318 && sequence_uses_addressof (XEXP (PATTERN (insn), 2)))
321 else if (uses_addressof (PATTERN (insn), 0)
322 || (REG_NOTES (insn) && uses_addressof (REG_NOTES (insn), 0)))
328 /* Remove all REG_EQUIV notes found in the insn chain. */
331 purge_reg_equiv_notes ()
335 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
339 rtx note = find_reg_note (insn, REG_EQUIV, 0);
342 /* Remove the note and keep looking at the notes for
344 remove_note (insn, note);
352 /* Replace the CALL_PLACEHOLDER with one of its children. INSN should be
353 the CALL_PLACEHOLDER insn; USE tells which child to use. */
356 replace_call_placeholder (insn, use)
360 if (use == sibcall_use_tail_recursion)
361 emit_insns_before (XEXP (PATTERN (insn), 2), insn);
362 else if (use == sibcall_use_sibcall)
363 emit_insns_before (XEXP (PATTERN (insn), 1), insn);
364 else if (use == sibcall_use_normal)
365 emit_insns_before (XEXP (PATTERN (insn), 0), insn);
369 /* Turn off LABEL_PRESERVE_P for the tail recursion label if it
370 exists. We only had to set it long enough to keep the jump
371 pass above from deleting it as unused. */
372 if (XEXP (PATTERN (insn), 3))
373 LABEL_PRESERVE_P (XEXP (PATTERN (insn), 3)) = 0;
375 /* "Delete" the placeholder insn. */
376 PUT_CODE (insn, NOTE);
377 NOTE_SOURCE_FILE (insn) = 0;
378 NOTE_LINE_NUMBER (insn) = NOTE_INSN_DELETED;
381 /* Given a (possibly empty) set of potential sibling or tail recursion call
382 sites, determine if optimization is possible.
384 Potential sibling or tail recursion calls are marked with CALL_PLACEHOLDER
385 insns. The CALL_PLACEHOLDER insn holds chains of insns to implement a
386 normal call, sibling call or tail recursive call.
388 Replace the CALL_PLACEHOLDER with an appropriate insn chain. */
391 optimize_sibling_and_tail_recursive_calls ()
394 basic_block alternate_exit = EXIT_BLOCK_PTR;
395 int current_function_uses_addressof;
396 int successful_sibling_call = 0;
397 int replaced_call_placeholder = 0;
400 insns = get_insns ();
402 /* We do not perform these calls when flag_exceptions is true, so this
403 is probably a NOP at the current time. However, we may want to support
404 sibling and tail recursion optimizations in the future, so let's plan
405 ahead and find all the EH labels. */
406 find_exception_handler_labels ();
408 /* Run a jump optimization pass to clean up the CFG. We primarily want
409 this to thread jumps so that it is obvious which blocks jump to the
411 jump_optimize_minimal (insns);
413 /* We need cfg information to determine which blocks are succeeded
414 only by the epilogue. */
415 find_basic_blocks (insns, max_reg_num (), 0);
418 /* If there are no basic blocks, then there is nothing to do. */
419 if (n_basic_blocks == 0)
422 /* Find the exit block.
424 It is possible that we have blocks which can reach the exit block
425 directly. However, most of the time a block will jump (or fall into)
426 N_BASIC_BLOCKS - 1, which in turn falls into the exit block. */
427 for (e = EXIT_BLOCK_PTR->pred;
428 e && alternate_exit == EXIT_BLOCK_PTR;
433 if (e->dest != EXIT_BLOCK_PTR || e->succ_next != NULL)
436 /* Walk forwards through the last normal block and see if it
437 does nothing except fall into the exit block. */
438 for (insn = BLOCK_HEAD (n_basic_blocks - 1);
440 insn = NEXT_INSN (insn))
442 /* This should only happen once, at the start of this block. */
443 if (GET_CODE (insn) == CODE_LABEL)
446 if (GET_CODE (insn) == NOTE)
449 if (GET_CODE (insn) == INSN
450 && GET_CODE (PATTERN (insn)) == USE)
456 /* If INSN is zero, then the search walked all the way through the
457 block without hitting anything interesting. This block is a
458 valid alternate exit block. */
460 alternate_exit = e->src;
463 /* If the function uses ADDRESSOF, we can't (easily) determine
464 at this point if the value will end up on the stack. */
465 current_function_uses_addressof = sequence_uses_addressof (insns);
467 /* Walk the insn chain and find any CALL_PLACEHOLDER insns. We need to
468 select one of the insn sequences attached to each CALL_PLACEHOLDER.
470 The different sequences represent different ways to implement the call,
471 ie, tail recursion, sibling call or normal call.
473 Since we do not create nested CALL_PLACEHOLDERs, the scan
474 continues with the insn that was after a replaced CALL_PLACEHOLDER;
475 we don't rescan the replacement insns. */
476 for (insn = insns; insn; insn = NEXT_INSN (insn))
478 if (GET_CODE (insn) == CALL_INSN
479 && GET_CODE (PATTERN (insn)) == CALL_PLACEHOLDER)
481 int sibcall = (XEXP (PATTERN (insn), 1) != NULL_RTX);
482 int tailrecursion = (XEXP (PATTERN (insn), 2) != NULL_RTX);
483 basic_block succ_block, call_block;
484 rtx temp, hardret, softret;
486 /* We must be careful with stack slots which are live at
487 potential optimization sites.
489 ?!? This test is overly conservative and will be replaced. */
493 /* alloca (until we have stack slot life analysis) inhibits
494 sibling call optimizations, but not tail recursion.
496 Similarly if we have ADDRESSOF expressions.
498 Similarly if we use varargs or stdarg since they implicitly
499 may take the address of an argument. */
500 if (current_function_calls_alloca || current_function_uses_addressof
501 || current_function_varargs || current_function_stdarg)
504 call_block = BLOCK_FOR_INSN (insn);
506 /* If the block has more than one successor, then we can not
507 perform sibcall or tail recursion optimizations. */
508 if (call_block->succ == NULL
509 || call_block->succ->succ_next != NULL)
512 /* If the single successor is not the exit block, then we can not
513 perform sibcall or tail recursion optimizations.
515 Note that this test combined with the previous is sufficient
516 to prevent tail call optimization in the presense of active
517 exception handlers. */
518 succ_block = call_block->succ->dest;
519 if (succ_block != EXIT_BLOCK_PTR && succ_block != alternate_exit)
522 /* If the call was the end of the block, then we're OK. */
524 if (temp == call_block->end)
527 /* Skip over copying from the call's return value pseudo into
528 this function's hard return register. */
529 if (identify_call_return_value (PATTERN (insn), &hardret, &softret))
531 temp = skip_copy_to_return_value (temp, hardret, softret);
532 if (temp == call_block->end)
536 /* Skip any stack adjustment. */
537 temp = skip_stack_adjustment (temp);
538 if (temp == call_block->end)
541 /* Skip over a CLOBBER of the return value (as a hard reg). */
542 temp = skip_use_of_return_value (temp, CLOBBER);
543 if (temp == call_block->end)
546 /* Skip over a USE of the return value (as a hard reg). */
547 temp = skip_use_of_return_value (temp, USE);
548 if (temp == call_block->end)
551 /* Skip over the JUMP_INSN at the end of the block. */
552 temp = skip_jump_insn (temp);
553 if (GET_CODE (temp) == NOTE)
554 temp = next_nonnote_insn (temp);
555 if (temp == call_block->end)
558 /* There are operations at the end of the block which we must
559 execute after returning from the function call. So this call
560 can not be optimized. */
562 sibcall = 0, tailrecursion = 0;
565 /* Select a set of insns to implement the call and emit them.
566 Tail recursion is the most efficient, so select it over
567 a tail/sibling call. */
570 successful_sibling_call = 1;
571 replaced_call_placeholder = 1;
572 replace_call_placeholder (insn,
574 ? sibcall_use_tail_recursion
576 ? sibcall_use_sibcall
577 : sibcall_use_normal);
581 /* A sibling call sequence invalidates any REG_EQUIV notes made for
582 this function's incoming arguments.
584 At the start of RTL generation we know the only REG_EQUIV notes
585 in the rtl chain are those for incoming arguments, so we can safely
586 flush any REG_EQUIV note.
588 This is (slight) overkill. We could keep track of the highest argument
589 we clobber and be more selective in removing notes, but it does not
590 seem to be worth the effort. */
591 if (successful_sibling_call)
592 purge_reg_equiv_notes ();
594 /* There may have been NOTE_INSN_BLOCK_{BEGIN,END} notes in the
595 CALL_PLACEHOLDER alternatives that we didn't emit. Rebuild the
596 lexical block tree to correspond to the notes that still exist. */
597 if (replaced_call_placeholder)
600 /* This information will be invalid after inline expansion. Kill it now. */
601 free_basic_block_vars (0);