1 /* Combine stack adjustments.
2 Copyright (C) 1987-2019 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
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/>. */
20 /* Track stack adjustments and stack memory references. Attempt to
21 reduce the number of stack adjustments by back-propagating across
22 the memory references.
24 This is intended primarily for use with targets that do not define
25 ACCUMULATE_OUTGOING_ARGS. It is of significantly more value to
26 targets that define PREFERRED_STACK_BOUNDARY more aligned than
27 STACK_BOUNDARY (e.g. x86), or if not all registers can be pushed
28 (e.g. x86 fp regs) which would ordinarily have to be implemented
29 as a sub/mov pair due to restrictions in calls.c.
31 Propagation stops when any of the insns that need adjusting are
32 (a) no longer valid because we've exceeded their range, (b) a
33 non-trivial push instruction, or (c) a call instruction.
35 Restriction B is based on the assumption that push instructions
36 are smaller or faster. If a port really wants to remove all
37 pushes, it should have defined ACCUMULATE_OUTGOING_ARGS. The
38 one exception that is made is for an add immediately followed
43 #include "coretypes.h"
47 #include "insn-config.h"
52 #include "tree-pass.h"
56 /* This structure records two kinds of stack references between stack
57 adjusting instructions: stack references in memory addresses for
58 regular insns and all stack references for debug insns. */
62 HOST_WIDE_INT sp_offset
;
65 struct csa_reflist
*next
;
68 static int stack_memref_p (rtx
);
69 static rtx
single_set_for_csa (rtx_insn
*);
70 static void free_csa_reflist (struct csa_reflist
*);
71 static struct csa_reflist
*record_one_stack_ref (rtx_insn
*, rtx
*,
72 struct csa_reflist
*);
73 static int try_apply_stack_adjustment (rtx_insn
*, struct csa_reflist
*,
74 HOST_WIDE_INT
, HOST_WIDE_INT
);
75 static void combine_stack_adjustments_for_block (basic_block
);
78 /* Main entry point for stack adjustment combination. */
81 combine_stack_adjustments (void)
85 FOR_EACH_BB_FN (bb
, cfun
)
86 combine_stack_adjustments_for_block (bb
);
89 /* Recognize a MEM of the form (sp) or (plus sp const). */
92 stack_memref_p (rtx x
)
98 if (x
== stack_pointer_rtx
)
100 if (GET_CODE (x
) == PLUS
101 && XEXP (x
, 0) == stack_pointer_rtx
102 && CONST_INT_P (XEXP (x
, 1)))
108 /* Recognize either normal single_set or the hack in i386.md for
109 tying fp and sp adjustments. */
112 single_set_for_csa (rtx_insn
*insn
)
115 rtx tmp
= single_set (insn
);
119 if (!NONJUMP_INSN_P (insn
)
120 || GET_CODE (PATTERN (insn
)) != PARALLEL
)
123 tmp
= PATTERN (insn
);
124 if (GET_CODE (XVECEXP (tmp
, 0, 0)) != SET
)
127 for (i
= 1; i
< XVECLEN (tmp
, 0); ++i
)
129 rtx this_rtx
= XVECEXP (tmp
, 0, i
);
131 /* The special case is allowing a no-op set. */
132 if (GET_CODE (this_rtx
) == SET
133 && SET_SRC (this_rtx
) == SET_DEST (this_rtx
))
135 else if (GET_CODE (this_rtx
) != CLOBBER
136 && GET_CODE (this_rtx
) != CLOBBER_HIGH
137 && GET_CODE (this_rtx
) != USE
)
141 return XVECEXP (tmp
, 0, 0);
144 /* Free the list of csa_reflist nodes. */
147 free_csa_reflist (struct csa_reflist
*reflist
)
149 struct csa_reflist
*next
;
150 for (; reflist
; reflist
= next
)
152 next
= reflist
->next
;
157 /* Create a new csa_reflist node from the given stack reference.
158 It is already known that the reference is either a MEM satisfying the
159 predicate stack_memref_p or a REG representing the stack pointer. */
161 static struct csa_reflist
*
162 record_one_stack_ref (rtx_insn
*insn
, rtx
*ref
, struct csa_reflist
*next_reflist
)
164 struct csa_reflist
*ml
;
166 ml
= XNEW (struct csa_reflist
);
168 if (REG_P (*ref
) || XEXP (*ref
, 0) == stack_pointer_rtx
)
171 ml
->sp_offset
= INTVAL (XEXP (XEXP (*ref
, 0), 1));
175 ml
->next
= next_reflist
;
180 /* We only know how to adjust the CFA; no other frame-related changes
181 may appear in any insn to be deleted. */
184 no_unhandled_cfa (rtx_insn
*insn
)
186 if (!RTX_FRAME_RELATED_P (insn
))
189 /* No CFA notes at all is a legacy interpretation like
190 FRAME_RELATED_EXPR, and is context sensitive within
191 the prologue state machine. We can't handle that here. */
192 bool has_cfa_adjust
= false;
194 for (rtx link
= REG_NOTES (insn
); link
; link
= XEXP (link
, 1))
195 switch (REG_NOTE_KIND (link
))
199 case REG_CFA_ADJUST_CFA
:
200 has_cfa_adjust
= true;
203 case REG_FRAME_RELATED_EXPR
:
204 case REG_CFA_DEF_CFA
:
206 case REG_CFA_REGISTER
:
207 case REG_CFA_EXPRESSION
:
208 case REG_CFA_RESTORE
:
209 case REG_CFA_SET_VDRAP
:
210 case REG_CFA_WINDOW_SAVE
:
211 case REG_CFA_FLUSH_QUEUE
:
212 case REG_CFA_TOGGLE_RA_MANGLE
:
216 return has_cfa_adjust
;
219 /* Attempt to apply ADJUST to the stack adjusting insn INSN, as well
220 as each of the memories and stack references in REFLIST. Return true
224 try_apply_stack_adjustment (rtx_insn
*insn
, struct csa_reflist
*reflist
,
225 HOST_WIDE_INT new_adjust
, HOST_WIDE_INT delta
)
227 struct csa_reflist
*ml
;
230 set
= single_set_for_csa (insn
);
231 if (MEM_P (SET_DEST (set
)))
232 validate_change (insn
, &SET_DEST (set
),
233 replace_equiv_address (SET_DEST (set
), stack_pointer_rtx
),
236 validate_change (insn
, &XEXP (SET_SRC (set
), 1), GEN_INT (new_adjust
), 1);
238 for (ml
= reflist
; ml
; ml
= ml
->next
)
240 rtx new_addr
= plus_constant (Pmode
, stack_pointer_rtx
,
241 ml
->sp_offset
- delta
);
244 if (MEM_P (*ml
->ref
))
245 new_val
= replace_equiv_address_nv (*ml
->ref
, new_addr
);
246 else if (GET_MODE (*ml
->ref
) == GET_MODE (stack_pointer_rtx
))
249 new_val
= lowpart_subreg (GET_MODE (*ml
->ref
), new_addr
,
250 GET_MODE (new_addr
));
251 validate_change (ml
->insn
, ml
->ref
, new_val
, 1);
254 if (apply_change_group ())
256 /* Succeeded. Update our knowledge of the stack references. */
257 for (ml
= reflist
; ml
; ml
= ml
->next
)
258 ml
->sp_offset
-= delta
;
266 /* For non-debug insns, record all stack memory references in INSN
267 and return true if there were no other (unrecorded) references to the
268 stack pointer. For debug insns, record all stack references regardless
269 of context and unconditionally return true. */
272 record_stack_refs (rtx_insn
*insn
, struct csa_reflist
**reflist
)
274 subrtx_ptr_iterator::array_type array
;
275 FOR_EACH_SUBRTX_PTR (iter
, array
, &PATTERN (insn
), NONCONST
)
279 switch (GET_CODE (x
))
282 if (!reg_mentioned_p (stack_pointer_rtx
, x
))
283 iter
.skip_subrtxes ();
284 /* We are not able to handle correctly all possible memrefs
285 containing stack pointer, so this check is necessary. */
286 else if (stack_memref_p (x
))
288 *reflist
= record_one_stack_ref (insn
, loc
, *reflist
);
289 iter
.skip_subrtxes ();
291 /* Try harder for DEBUG_INSNs, handle e.g.
292 (mem (mem (sp + 16) + 4). */
293 else if (!DEBUG_INSN_P (insn
))
298 /* ??? We want be able to handle non-memory stack pointer
299 references later. For now just discard all insns referring to
300 stack pointer outside mem expressions. We would probably
301 want to teach validate_replace to simplify expressions first.
303 We can't just compare with STACK_POINTER_RTX because the
304 reference to the stack pointer might be in some other mode.
305 In particular, an explicit clobber in an asm statement will
306 result in a QImode clobber.
308 In DEBUG_INSNs, we want to replace all occurrences, otherwise
309 they will cause -fcompare-debug failures. */
310 if (REGNO (x
) == STACK_POINTER_REGNUM
)
312 if (!DEBUG_INSN_P (insn
))
314 *reflist
= record_one_stack_ref (insn
, loc
, *reflist
);
325 /* If INSN has a REG_ARGS_SIZE note, move it to LAST.
326 AFTER is true iff LAST follows INSN in the instruction stream. */
329 maybe_move_args_size_note (rtx_insn
*last
, rtx_insn
*insn
, bool after
)
333 note
= find_reg_note (insn
, REG_ARGS_SIZE
, NULL_RTX
);
337 last_note
= find_reg_note (last
, REG_ARGS_SIZE
, NULL_RTX
);
340 /* The ARGS_SIZE notes are *not* cumulative. They represent an
341 absolute value, and the "most recent" note wins. */
343 XEXP (last_note
, 0) = XEXP (note
, 0);
346 add_reg_note (last
, REG_ARGS_SIZE
, XEXP (note
, 0));
349 /* Merge any REG_CFA_ADJUST_CFA note from SRC into DST.
350 AFTER is true iff DST follows SRC in the instruction stream. */
353 maybe_merge_cfa_adjust (rtx_insn
*dst
, rtx_insn
*src
, bool after
)
355 rtx snote
= NULL
, dnote
= NULL
;
359 if (RTX_FRAME_RELATED_P (src
))
360 snote
= find_reg_note (src
, REG_CFA_ADJUST_CFA
, NULL_RTX
);
363 sexp
= XEXP (snote
, 0);
365 if (RTX_FRAME_RELATED_P (dst
))
366 dnote
= find_reg_note (dst
, REG_CFA_ADJUST_CFA
, NULL_RTX
);
369 add_reg_note (dst
, REG_CFA_ADJUST_CFA
, sexp
);
372 dexp
= XEXP (dnote
, 0);
374 gcc_assert (GET_CODE (sexp
) == SET
);
375 gcc_assert (GET_CODE (dexp
) == SET
);
378 exp1
= dexp
, exp2
= sexp
;
380 exp1
= sexp
, exp2
= dexp
;
382 SET_SRC (exp1
) = simplify_replace_rtx (SET_SRC (exp1
), SET_DEST (exp2
),
384 XEXP (dnote
, 0) = exp1
;
387 /* Return the next (or previous) active insn within BB. */
390 prev_active_insn_bb (basic_block bb
, rtx_insn
*insn
)
392 for (insn
= PREV_INSN (insn
);
393 insn
!= PREV_INSN (BB_HEAD (bb
));
394 insn
= PREV_INSN (insn
))
395 if (active_insn_p (insn
))
401 next_active_insn_bb (basic_block bb
, rtx_insn
*insn
)
403 for (insn
= NEXT_INSN (insn
);
404 insn
!= NEXT_INSN (BB_END (bb
));
405 insn
= NEXT_INSN (insn
))
406 if (active_insn_p (insn
))
411 /* If INSN has a REG_ARGS_SIZE note, if possible move it to PREV. Otherwise
412 search for a nearby candidate within BB where we can stick the note. */
415 force_move_args_size_note (basic_block bb
, rtx_insn
*prev
, rtx_insn
*insn
)
418 rtx_insn
*test
, *next_candidate
, *prev_candidate
;
420 /* If PREV exists, tail-call to the logic in the other function. */
423 maybe_move_args_size_note (prev
, insn
, false);
427 /* First, make sure there's anything that needs doing. */
428 note
= find_reg_note (insn
, REG_ARGS_SIZE
, NULL_RTX
);
432 /* We need to find a spot between the previous and next exception points
433 where we can place the note and "properly" deallocate the arguments. */
434 next_candidate
= prev_candidate
= NULL
;
436 /* It is often the case that we have insns in the order:
438 add sp (previous deallocation)
439 sub sp (align for next arglist)
441 and the add/sub cancel. Therefore we begin by searching forward. */
444 while ((test
= next_active_insn_bb (bb
, test
)) != NULL
)
446 /* Found an existing note: nothing to do. */
447 if (find_reg_note (test
, REG_ARGS_SIZE
, NULL_RTX
))
449 /* Found something that affects unwinding. Stop searching. */
450 if (CALL_P (test
) || !insn_nothrow_p (test
))
452 if (next_candidate
== NULL
)
453 next_candidate
= test
;
457 while ((test
= prev_active_insn_bb (bb
, test
)) != NULL
)
460 /* Found a place that seems logical to adjust the stack. */
461 tnote
= find_reg_note (test
, REG_ARGS_SIZE
, NULL_RTX
);
464 XEXP (tnote
, 0) = XEXP (note
, 0);
467 if (prev_candidate
== NULL
)
468 prev_candidate
= test
;
469 /* Found something that affects unwinding. Stop searching. */
470 if (CALL_P (test
) || !insn_nothrow_p (test
))
475 test
= prev_candidate
;
476 else if (next_candidate
)
477 test
= next_candidate
;
480 /* ??? We *must* have a place, lest we ICE on the lost adjustment.
481 Options are: dummy clobber insn, nop, or prevent the removal of
483 /* TODO: Find another way to indicate to the dwarf2 code that we
484 have not in fact lost an adjustment. */
485 test
= emit_insn_before (gen_rtx_CLOBBER (VOIDmode
, const0_rtx
), insn
);
487 add_reg_note (test
, REG_ARGS_SIZE
, XEXP (note
, 0));
490 /* Subroutine of combine_stack_adjustments, called for each basic block. */
493 combine_stack_adjustments_for_block (basic_block bb
)
495 HOST_WIDE_INT last_sp_adjust
= 0;
496 rtx_insn
*last_sp_set
= NULL
;
497 rtx_insn
*last2_sp_set
= NULL
;
498 struct csa_reflist
*reflist
= NULL
;
499 rtx_insn
*insn
, *next
;
501 bool end_of_block
= false;
503 for (insn
= BB_HEAD (bb
); !end_of_block
; insn
= next
)
505 end_of_block
= insn
== BB_END (bb
);
506 next
= NEXT_INSN (insn
);
511 set
= single_set_for_csa (insn
);
512 if (set
&& find_reg_note (insn
, REG_STACK_CHECK
, NULL_RTX
))
516 rtx dest
= SET_DEST (set
);
517 rtx src
= SET_SRC (set
);
519 /* Find constant additions to the stack pointer. */
520 if (dest
== stack_pointer_rtx
521 && GET_CODE (src
) == PLUS
522 && XEXP (src
, 0) == stack_pointer_rtx
523 && CONST_INT_P (XEXP (src
, 1)))
525 HOST_WIDE_INT this_adjust
= INTVAL (XEXP (src
, 1));
527 /* If we've not seen an adjustment previously, record
528 it now and continue. */
532 last_sp_adjust
= this_adjust
;
536 /* If not all recorded refs can be adjusted, or the
537 adjustment is now too large for a constant addition,
538 we cannot merge the two stack adjustments.
540 Also we need to be careful to not move stack pointer
541 such that we create stack accesses outside the allocated
542 area. We can combine an allocation into the first insn,
543 or a deallocation into the second insn. We cannot
544 combine an allocation followed by a deallocation.
546 The only somewhat frequent occurrence of the later is when
547 a function allocates a stack frame but does not use it.
548 For this case, we would need to analyze rtl stream to be
549 sure that allocated area is really unused. This means not
550 only checking the memory references, but also all registers
551 or global memory references possibly containing a stack
554 Perhaps the best way to address this problem is to teach
555 gcc not to allocate stack for objects never used. */
557 /* Combine an allocation into the first instruction. */
558 if (STACK_GROWS_DOWNWARD
? this_adjust
<= 0 : this_adjust
>= 0)
560 if (no_unhandled_cfa (insn
)
561 && try_apply_stack_adjustment (last_sp_set
, reflist
,
567 maybe_move_args_size_note (last_sp_set
, insn
, false);
568 maybe_merge_cfa_adjust (last_sp_set
, insn
, false);
570 last_sp_adjust
+= this_adjust
;
575 /* Otherwise we have a deallocation. Do not combine with
576 a previous allocation. Combine into the second insn. */
577 else if (STACK_GROWS_DOWNWARD
578 ? last_sp_adjust
>= 0 : last_sp_adjust
<= 0)
580 if (no_unhandled_cfa (last_sp_set
)
581 && try_apply_stack_adjustment (insn
, reflist
,
587 maybe_move_args_size_note (insn
, last_sp_set
, true);
588 maybe_merge_cfa_adjust (insn
, last_sp_set
, true);
589 delete_insn (last_sp_set
);
591 last_sp_adjust
+= this_adjust
;
592 free_csa_reflist (reflist
);
598 /* Combination failed. Restart processing from here. If
599 deallocation+allocation conspired to cancel, we can
600 delete the old deallocation insn. */
603 if (last_sp_adjust
== 0 && no_unhandled_cfa (last_sp_set
))
605 maybe_move_args_size_note (insn
, last_sp_set
, true);
606 maybe_merge_cfa_adjust (insn
, last_sp_set
, true);
607 delete_insn (last_sp_set
);
610 last2_sp_set
= last_sp_set
;
612 free_csa_reflist (reflist
);
615 last_sp_adjust
= this_adjust
;
619 /* Find a store with pre-(dec|inc)rement or pre-modify of exactly
620 the previous adjustment and turn it into a simple store. This
621 is equivalent to anticipating the stack adjustment so this must
624 && ((STACK_GROWS_DOWNWARD
625 ? (GET_CODE (XEXP (dest
, 0)) == PRE_DEC
626 && known_eq (last_sp_adjust
,
627 GET_MODE_SIZE (GET_MODE (dest
))))
628 : (GET_CODE (XEXP (dest
, 0)) == PRE_INC
629 && known_eq (-last_sp_adjust
,
630 GET_MODE_SIZE (GET_MODE (dest
)))))
631 || ((STACK_GROWS_DOWNWARD
632 ? last_sp_adjust
>= 0 : last_sp_adjust
<= 0)
633 && GET_CODE (XEXP (dest
, 0)) == PRE_MODIFY
634 && GET_CODE (XEXP (XEXP (dest
, 0), 1)) == PLUS
635 && XEXP (XEXP (XEXP (dest
, 0), 1), 0)
637 && GET_CODE (XEXP (XEXP (XEXP (dest
, 0), 1), 1))
639 && INTVAL (XEXP (XEXP (XEXP (dest
, 0), 1), 1))
641 && XEXP (XEXP (dest
, 0), 0) == stack_pointer_rtx
642 && !reg_mentioned_p (stack_pointer_rtx
, src
)
643 && memory_address_p (GET_MODE (dest
), stack_pointer_rtx
)
644 && try_apply_stack_adjustment (insn
, reflist
, 0,
648 maybe_move_args_size_note (last2_sp_set
, last_sp_set
, false);
650 maybe_move_args_size_note (insn
, last_sp_set
, true);
651 delete_insn (last_sp_set
);
652 free_csa_reflist (reflist
);
660 if (!CALL_P (insn
) && last_sp_set
661 && record_stack_refs (insn
, &reflist
))
664 /* Otherwise, we were not able to process the instruction.
665 Do not continue collecting data across such a one. */
668 || reg_mentioned_p (stack_pointer_rtx
, PATTERN (insn
))))
670 if (last_sp_set
&& last_sp_adjust
== 0)
672 force_move_args_size_note (bb
, last2_sp_set
, last_sp_set
);
673 delete_insn (last_sp_set
);
675 free_csa_reflist (reflist
);
683 if (last_sp_set
&& last_sp_adjust
== 0)
685 force_move_args_size_note (bb
, last2_sp_set
, last_sp_set
);
686 delete_insn (last_sp_set
);
690 free_csa_reflist (reflist
);
694 rest_of_handle_stack_adjustments (void)
696 df_note_add_problem ();
698 combine_stack_adjustments ();
704 const pass_data pass_data_stack_adjustments
=
708 OPTGROUP_NONE
, /* optinfo_flags */
709 TV_COMBINE_STACK_ADJUST
, /* tv_id */
710 0, /* properties_required */
711 0, /* properties_provided */
712 0, /* properties_destroyed */
713 0, /* todo_flags_start */
714 TODO_df_finish
, /* todo_flags_finish */
717 class pass_stack_adjustments
: public rtl_opt_pass
720 pass_stack_adjustments (gcc::context
*ctxt
)
721 : rtl_opt_pass (pass_data_stack_adjustments
, ctxt
)
724 /* opt_pass methods: */
725 virtual bool gate (function
*);
726 virtual unsigned int execute (function
*)
728 return rest_of_handle_stack_adjustments ();
731 }; // class pass_stack_adjustments
734 pass_stack_adjustments::gate (function
*)
736 /* This is kind of a heuristic. We need to run combine_stack_adjustments
737 even for machines with possibly nonzero TARGET_RETURN_POPS_ARGS
738 and ACCUMULATE_OUTGOING_ARGS. We expect that only ports having
739 push instructions will have popping returns. */
740 #ifndef PUSH_ROUNDING
741 if (ACCUMULATE_OUTGOING_ARGS
)
744 return flag_combine_stack_adjustments
;
750 make_pass_stack_adjustments (gcc::context
*ctxt
)
752 return new pass_stack_adjustments (ctxt
);