1 /* Decompose multiword subregs.
2 Copyright (C) 2007 Free Software Foundation, Inc.
3 Contributed by Richard Henderson <rth@redhat.com>
4 Ian Lance Taylor <iant@google.com>
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
24 #include "coretypes.h"
31 #include "insn-config.h"
33 #include "basic-block.h"
39 #include "tree-pass.h"
42 #ifdef STACK_GROWS_DOWNWARD
43 # undef STACK_GROWS_DOWNWARD
44 # define STACK_GROWS_DOWNWARD 1
46 # define STACK_GROWS_DOWNWARD 0
50 DEF_VEC_ALLOC_P (bitmap,heap);
52 /* Decompose multi-word pseudo-registers into individual
53 pseudo-registers when possible. This is possible when all the uses
54 of a multi-word register are via SUBREG, or are copies of the
55 register to another location. Breaking apart the register permits
56 more CSE and permits better register allocation. */
58 /* Bit N in this bitmap is set if regno N is used in a context in
59 which we can decompose it. */
60 static bitmap decomposable_context;
62 /* Bit N in this bitmap is set if regno N is used in a context in
63 which it can not be decomposed. */
64 static bitmap non_decomposable_context;
66 /* Bit N in the bitmap in element M of this array is set if there is a
67 copy from reg M to reg N. */
68 static VEC(bitmap,heap) *reg_copy_graph;
70 /* Return whether X is a simple object which we can take a word_mode
74 simple_move_operand (rtx x)
76 if (GET_CODE (x) == SUBREG)
82 if (GET_CODE (x) == LABEL_REF
83 || GET_CODE (x) == SYMBOL_REF
84 || GET_CODE (x) == HIGH
85 || GET_CODE (x) == CONST)
89 && (MEM_VOLATILE_P (x)
90 || mode_dependent_address_p (XEXP (x, 0))))
96 /* If INSN is a single set between two objects, return the single set.
97 Such an insn can always be decomposed. INSN should have been
98 passed to recog and extract_insn before this is called. */
101 simple_move (rtx insn)
105 enum machine_mode mode;
107 if (recog_data.n_operands != 2)
110 set = single_set (insn);
115 if (x != recog_data.operand[0] && x != recog_data.operand[1])
117 if (!simple_move_operand (x))
121 if (x != recog_data.operand[0] && x != recog_data.operand[1])
123 /* For the src we can handle ASM_OPERANDS, and it is beneficial for
124 things like x86 rdtsc which returns a DImode value. */
125 if (GET_CODE (x) != ASM_OPERANDS
126 && !simple_move_operand (x))
129 /* We try to decompose in integer modes, to avoid generating
130 inefficient code copying between integer and floating point
131 registers. That means that we can't decompose if this is a
132 non-integer mode for which there is no integer mode of the same
134 mode = GET_MODE (SET_SRC (set));
135 if (!SCALAR_INT_MODE_P (mode)
136 && (mode_for_size (GET_MODE_SIZE (mode) * BITS_PER_UNIT, MODE_INT, 0)
140 /* Reject PARTIAL_INT modes. They are used for processor specific
141 purposes and it's probably best not to tamper with them. */
142 if (GET_MODE_CLASS (mode) == MODE_PARTIAL_INT)
148 /* If SET is a copy from one multi-word pseudo-register to another,
149 record that in reg_copy_graph. Return whether it is such a
153 find_pseudo_copy (rtx set)
155 rtx dest = SET_DEST (set);
156 rtx src = SET_SRC (set);
160 if (!REG_P (dest) || !REG_P (src))
165 if (HARD_REGISTER_NUM_P (rd) || HARD_REGISTER_NUM_P (rs))
168 if (GET_MODE_SIZE (GET_MODE (dest)) <= UNITS_PER_WORD)
171 b = VEC_index (bitmap, reg_copy_graph, rs);
174 b = BITMAP_ALLOC (NULL);
175 VEC_replace (bitmap, reg_copy_graph, rs, b);
178 bitmap_set_bit (b, rd);
183 /* Look through the registers in DECOMPOSABLE_CONTEXT. For each case
184 where they are copied to another register, add the register to
185 which they are copied to DECOMPOSABLE_CONTEXT. Use
186 NON_DECOMPOSABLE_CONTEXT to limit this--we don't bother to track
187 copies of registers which are in NON_DECOMPOSABLE_CONTEXT. */
190 propagate_pseudo_copies (void)
192 bitmap queue, propagate;
194 queue = BITMAP_ALLOC (NULL);
195 propagate = BITMAP_ALLOC (NULL);
197 bitmap_copy (queue, decomposable_context);
200 bitmap_iterator iter;
203 bitmap_clear (propagate);
205 EXECUTE_IF_SET_IN_BITMAP (queue, 0, i, iter)
207 bitmap b = VEC_index (bitmap, reg_copy_graph, i);
209 bitmap_ior_and_compl_into (propagate, b, non_decomposable_context);
212 bitmap_and_compl (queue, propagate, decomposable_context);
213 bitmap_ior_into (decomposable_context, propagate);
215 while (!bitmap_empty_p (queue));
218 BITMAP_FREE (propagate);
221 /* A pointer to one of these values is passed to
222 find_decomposable_subregs via for_each_rtx. */
224 enum classify_move_insn
226 /* Not a simple move from one location to another. */
228 /* A simple move from one pseudo-register to another with no
230 SIMPLE_PSEUDO_REG_MOVE,
231 /* A simple move involving a non-pseudo-register, or from one
232 pseudo-register to another with a REG_RETVAL note. */
236 /* This is called via for_each_rtx. If we find a SUBREG which we
237 could use to decompose a pseudo-register, set a bit in
238 DECOMPOSABLE_CONTEXT. If we find an unadorned register which is
239 not a simple pseudo-register copy, DATA will point at the type of
240 move, and we set a bit in DECOMPOSABLE_CONTEXT or
241 NON_DECOMPOSABLE_CONTEXT as appropriate. */
244 find_decomposable_subregs (rtx *px, void *data)
246 enum classify_move_insn *pcmi = (enum classify_move_insn *) data;
252 if (GET_CODE (x) == SUBREG)
254 rtx inner = SUBREG_REG (x);
255 unsigned int regno, outer_size, inner_size, outer_words, inner_words;
260 regno = REGNO (inner);
261 if (HARD_REGISTER_NUM_P (regno))
264 outer_size = GET_MODE_SIZE (GET_MODE (x));
265 inner_size = GET_MODE_SIZE (GET_MODE (inner));
266 outer_words = (outer_size + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
267 inner_words = (inner_size + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
269 /* We only try to decompose single word subregs of multi-word
270 registers. When we find one, we return -1 to avoid iterating
271 over the inner register.
273 ??? This doesn't allow, e.g., DImode subregs of TImode values
274 on 32-bit targets. We would need to record the way the
275 pseudo-register was used, and only decompose if all the uses
276 were the same number and size of pieces. Hopefully this
277 doesn't happen much. */
279 if (outer_words == 1 && inner_words > 1)
281 bitmap_set_bit (decomposable_context, regno);
285 /* If this is a cast from one mode to another, where the modes
286 have the same size, and they are not tieable, then mark this
287 register as non-decomposable. If we decompose it we are
288 likely to mess up whatever the backend is trying to do. */
290 && outer_size == inner_size
291 && !MODES_TIEABLE_P (GET_MODE (x), GET_MODE (inner)))
293 bitmap_set_bit (non_decomposable_context, regno);
301 /* We will see an outer SUBREG before we see the inner REG, so
302 when we see a plain REG here it means a direct reference to
305 If this is not a simple copy from one location to another,
306 then we can not decompose this register. If this is a simple
307 copy from one pseudo-register to another, with no REG_RETVAL
308 note, and the mode is right, then we mark the register as
309 decomposable. Otherwise we don't say anything about this
310 register--it could be decomposed, but whether that would be
311 profitable depends upon how it is used elsewhere.
313 We only set bits in the bitmap for multi-word
314 pseudo-registers, since those are the only ones we care about
315 and it keeps the size of the bitmaps down. */
318 if (!HARD_REGISTER_NUM_P (regno)
319 && GET_MODE_SIZE (GET_MODE (x)) > UNITS_PER_WORD)
323 case NOT_SIMPLE_MOVE:
324 bitmap_set_bit (non_decomposable_context, regno);
326 case SIMPLE_PSEUDO_REG_MOVE:
327 if (MODES_TIEABLE_P (GET_MODE (x), word_mode))
328 bitmap_set_bit (decomposable_context, regno);
339 enum classify_move_insn cmi_mem = NOT_SIMPLE_MOVE;
341 /* Any registers used in a MEM do not participate in a
342 SIMPLE_MOVE or SIMPLE_PSEUDO_REG_MOVE. Do our own recursion
343 here, and return -1 to block the parent's recursion. */
344 for_each_rtx (&XEXP (x, 0), find_decomposable_subregs, &cmi_mem);
351 /* Decompose REGNO into word-sized components. We smash the REG node
352 in place. This ensures that (1) something goes wrong quickly if we
353 fail to make some replacement, and (2) the debug information inside
354 the symbol table is automatically kept up to date. */
357 decompose_register (unsigned int regno)
360 unsigned int words, i;
363 reg = regno_reg_rtx[regno];
365 regno_reg_rtx[regno] = NULL_RTX;
367 words = GET_MODE_SIZE (GET_MODE (reg));
368 words = (words + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
370 v = rtvec_alloc (words);
371 for (i = 0; i < words; ++i)
372 RTVEC_ELT (v, i) = gen_reg_rtx_offset (reg, word_mode, i * UNITS_PER_WORD);
374 PUT_CODE (reg, CONCATN);
379 fprintf (dump_file, "; Splitting reg %u ->", regno);
380 for (i = 0; i < words; ++i)
381 fprintf (dump_file, " %u", REGNO (XVECEXP (reg, 0, i)));
382 fputc ('\n', dump_file);
386 /* Get a SUBREG of a CONCATN. */
389 simplify_subreg_concatn (enum machine_mode outermode, rtx op,
392 unsigned int inner_size;
393 enum machine_mode innermode;
395 unsigned int final_offset;
397 gcc_assert (GET_CODE (op) == CONCATN);
398 gcc_assert (byte % GET_MODE_SIZE (outermode) == 0);
400 innermode = GET_MODE (op);
401 gcc_assert (byte < GET_MODE_SIZE (innermode));
402 gcc_assert (GET_MODE_SIZE (outermode) <= GET_MODE_SIZE (innermode));
404 inner_size = GET_MODE_SIZE (innermode) / XVECLEN (op, 0);
405 part = XVECEXP (op, 0, byte / inner_size);
406 final_offset = byte % inner_size;
407 if (final_offset + GET_MODE_SIZE (outermode) > inner_size)
410 return simplify_gen_subreg (outermode, part, GET_MODE (part), final_offset);
413 /* Wrapper around simplify_gen_subreg which handles CONCATN. */
416 simplify_gen_subreg_concatn (enum machine_mode outermode, rtx op,
417 enum machine_mode innermode, unsigned int byte)
421 /* We have to handle generating a SUBREG of a SUBREG of a CONCATN.
422 If OP is a SUBREG of a CONCATN, then it must be a simple mode
423 change with the same size and offset 0, or it must extract a
424 part. We shouldn't see anything else here. */
425 if (GET_CODE (op) == SUBREG && GET_CODE (SUBREG_REG (op)) == CONCATN)
429 if ((GET_MODE_SIZE (GET_MODE (op))
430 == GET_MODE_SIZE (GET_MODE (SUBREG_REG (op))))
431 && SUBREG_BYTE (op) == 0)
432 return simplify_gen_subreg_concatn (outermode, SUBREG_REG (op),
433 GET_MODE (SUBREG_REG (op)), byte);
435 op2 = simplify_subreg_concatn (GET_MODE (op), SUBREG_REG (op),
439 /* We don't handle paradoxical subregs here. */
440 gcc_assert (GET_MODE_SIZE (outermode)
441 <= GET_MODE_SIZE (GET_MODE (op)));
442 gcc_assert (GET_MODE_SIZE (GET_MODE (op))
443 <= GET_MODE_SIZE (GET_MODE (SUBREG_REG (op))));
444 op2 = simplify_subreg_concatn (outermode, SUBREG_REG (op),
445 byte + SUBREG_BYTE (op));
446 gcc_assert (op2 != NULL_RTX);
451 gcc_assert (op != NULL_RTX);
452 gcc_assert (innermode == GET_MODE (op));
455 if (GET_CODE (op) == CONCATN)
456 return simplify_subreg_concatn (outermode, op, byte);
458 ret = simplify_gen_subreg (outermode, op, innermode, byte);
460 /* If we see an insn like (set (reg:DI) (subreg:DI (reg:SI) 0)) then
461 resolve_simple_move will ask for the high part of the paradoxical
462 subreg, which does not have a value. Just return a zero. */
464 && GET_CODE (op) == SUBREG
465 && SUBREG_BYTE (op) == 0
466 && (GET_MODE_SIZE (innermode)
467 > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op)))))
468 return CONST0_RTX (outermode);
470 gcc_assert (ret != NULL_RTX);
474 /* Return whether we should resolve X into the registers into which it
478 resolve_reg_p (rtx x)
480 return GET_CODE (x) == CONCATN;
483 /* Return whether X is a SUBREG of a register which we need to
487 resolve_subreg_p (rtx x)
489 if (GET_CODE (x) != SUBREG)
491 return resolve_reg_p (SUBREG_REG (x));
494 /* This is called via for_each_rtx. Look for SUBREGs which need to be
498 resolve_subreg_use (rtx *px, void *data)
500 rtx insn = (rtx) data;
506 if (resolve_subreg_p (x))
508 x = simplify_subreg_concatn (GET_MODE (x), SUBREG_REG (x),
511 /* It is possible for a note to contain a reference which we can
512 decompose. In this case, return 1 to the caller to indicate
513 that the note must be removed. */
520 validate_change (insn, px, x, 1);
524 if (resolve_reg_p (x))
526 /* Return 1 to the caller to indicate that we found a direct
527 reference to a register which is being decomposed. This can
528 happen inside notes, multiword shift or zero-extend
536 /* We are deleting INSN. Move any EH_REGION notes to INSNS. */
539 move_eh_region_note (rtx insn, rtx insns)
543 note = find_reg_note (insn, REG_EH_REGION, NULL_RTX);
544 if (note == NULL_RTX)
547 gcc_assert (CALL_P (insn)
548 || (flag_non_call_exceptions && may_trap_p (PATTERN (insn))));
550 for (p = insns; p != NULL_RTX; p = NEXT_INSN (p))
553 || (flag_non_call_exceptions
555 && may_trap_p (PATTERN (p))))
556 REG_NOTES (p) = gen_rtx_EXPR_LIST (REG_EH_REGION, XEXP (note, 0),
561 /* If there is a REG_LIBCALL note on OLD_START, move it to NEW_START,
562 and link the corresponding REG_RETVAL note to NEW_START. */
565 move_libcall_note (rtx old_start, rtx new_start)
567 rtx note0, note1, end;
569 note0 = find_reg_note (old_start, REG_LIBCALL, NULL);
570 if (note0 == NULL_RTX)
573 remove_note (old_start, note0);
574 end = XEXP (note0, 0);
575 note1 = find_reg_note (end, REG_RETVAL, NULL);
577 XEXP (note0, 1) = REG_NOTES (new_start);
578 REG_NOTES (new_start) = note0;
579 XEXP (note1, 0) = new_start;
582 /* Remove any REG_RETVAL note, the corresponding REG_LIBCALL note, and
583 any markers for a no-conflict block. We have decomposed the
584 registers so the non-conflict is now obvious. */
587 remove_retval_note (rtx insn1)
589 rtx note0, insn0, note1, insn;
591 note1 = find_reg_note (insn1, REG_RETVAL, NULL);
592 if (note1 == NULL_RTX)
595 insn0 = XEXP (note1, 0);
596 note0 = find_reg_note (insn0, REG_LIBCALL, NULL);
598 remove_note (insn0, note0);
599 remove_note (insn1, note1);
601 for (insn = insn0; insn != insn1; insn = NEXT_INSN (insn))
607 note = find_reg_note (insn, REG_NO_CONFLICT, NULL);
608 if (note == NULL_RTX)
610 remove_note (insn, note);
615 /* Resolve any decomposed registers which appear in register notes on
619 resolve_reg_notes (rtx insn)
623 note = find_reg_equal_equiv_note (insn);
626 int old_count = num_validated_changes ();
627 if (for_each_rtx (&XEXP (note, 0), resolve_subreg_use, NULL))
629 remove_note (insn, note);
630 remove_retval_note (insn);
633 if (old_count != num_validated_changes ())
634 df_notes_rescan (insn);
637 pnote = ®_NOTES (insn);
638 while (*pnote != NULL_RTX)
643 switch (REG_NOTE_KIND (note))
645 case REG_NO_CONFLICT:
648 if (resolve_reg_p (XEXP (note, 0)))
657 *pnote = XEXP (note, 1);
659 pnote = &XEXP (note, 1);
663 /* Return whether X can be decomposed into subwords. */
666 can_decompose_p (rtx x)
670 unsigned int regno = REGNO (x);
672 if (HARD_REGISTER_NUM_P (regno))
673 return (validate_subreg (word_mode, GET_MODE (x), x, UNITS_PER_WORD)
674 && HARD_REGNO_MODE_OK (regno, word_mode));
676 return !bitmap_bit_p (non_decomposable_context, regno);
682 /* Decompose the registers used in a simple move SET within INSN. If
683 we don't change anything, return INSN, otherwise return the start
684 of the sequence of moves. */
687 resolve_simple_move (rtx set, rtx insn)
689 rtx src, dest, real_dest, insns;
690 enum machine_mode orig_mode, dest_mode;
695 dest = SET_DEST (set);
696 orig_mode = GET_MODE (dest);
698 words = (GET_MODE_SIZE (orig_mode) + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
704 /* We have to handle copying from a SUBREG of a decomposed reg where
705 the SUBREG is larger than word size. Rather than assume that we
706 can take a word_mode SUBREG of the destination, we copy to a new
707 register and then copy that to the destination. */
709 real_dest = NULL_RTX;
711 if (GET_CODE (src) == SUBREG
712 && resolve_reg_p (SUBREG_REG (src))
713 && (SUBREG_BYTE (src) != 0
714 || (GET_MODE_SIZE (orig_mode)
715 != GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))))))
718 dest = gen_reg_rtx (orig_mode);
719 if (REG_P (real_dest))
720 REG_ATTRS (dest) = REG_ATTRS (real_dest);
723 /* Similarly if we are copying to a SUBREG of a decomposed reg where
724 the SUBREG is larger than word size. */
726 if (GET_CODE (dest) == SUBREG
727 && resolve_reg_p (SUBREG_REG (dest))
728 && (SUBREG_BYTE (dest) != 0
729 || (GET_MODE_SIZE (orig_mode)
730 != GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))))
732 rtx reg, minsn, smove;
734 reg = gen_reg_rtx (orig_mode);
735 minsn = emit_move_insn (reg, src);
736 smove = single_set (minsn);
737 gcc_assert (smove != NULL_RTX);
738 resolve_simple_move (smove, minsn);
742 /* If we didn't have any big SUBREGS of decomposed registers, and
743 neither side of the move is a register we are decomposing, then
744 we don't have to do anything here. */
746 if (src == SET_SRC (set)
747 && dest == SET_DEST (set)
748 && !resolve_reg_p (src)
749 && !resolve_subreg_p (src)
750 && !resolve_reg_p (dest)
751 && !resolve_subreg_p (dest))
757 /* It's possible for the code to use a subreg of a decomposed
758 register while forming an address. We need to handle that before
759 passing the address to emit_move_insn. We pass NULL_RTX as the
760 insn parameter to resolve_subreg_use because we can not validate
762 if (MEM_P (src) || MEM_P (dest))
767 for_each_rtx (&XEXP (src, 0), resolve_subreg_use, NULL_RTX);
769 for_each_rtx (&XEXP (dest, 0), resolve_subreg_use, NULL_RTX);
770 acg = apply_change_group ();
774 /* If SRC is a register which we can't decompose, or has side
775 effects, we need to move via a temporary register. */
777 if (!can_decompose_p (src)
778 || side_effects_p (src)
779 || GET_CODE (src) == ASM_OPERANDS)
783 reg = gen_reg_rtx (orig_mode);
784 emit_move_insn (reg, src);
788 /* If DEST is a register which we can't decompose, or has side
789 effects, we need to first move to a temporary register. We
790 handle the common case of pushing an operand directly. We also
791 go through a temporary register if it holds a floating point
792 value. This gives us better code on systems which can't move
793 data easily between integer and floating point registers. */
795 dest_mode = orig_mode;
796 pushing = push_operand (dest, dest_mode);
797 if (!can_decompose_p (dest)
798 || (side_effects_p (dest) && !pushing)
799 || (!SCALAR_INT_MODE_P (dest_mode)
800 && !resolve_reg_p (dest)
801 && !resolve_subreg_p (dest)))
803 if (real_dest == NULL_RTX)
805 if (!SCALAR_INT_MODE_P (dest_mode))
807 dest_mode = mode_for_size (GET_MODE_SIZE (dest_mode) * BITS_PER_UNIT,
809 gcc_assert (dest_mode != BLKmode);
811 dest = gen_reg_rtx (dest_mode);
812 if (REG_P (real_dest))
813 REG_ATTRS (dest) = REG_ATTRS (real_dest);
818 unsigned int i, j, jinc;
820 gcc_assert (GET_MODE_SIZE (orig_mode) % UNITS_PER_WORD == 0);
821 gcc_assert (GET_CODE (XEXP (dest, 0)) != PRE_MODIFY);
822 gcc_assert (GET_CODE (XEXP (dest, 0)) != POST_MODIFY);
824 if (WORDS_BIG_ENDIAN == STACK_GROWS_DOWNWARD)
835 for (i = 0; i < words; ++i, j += jinc)
839 temp = copy_rtx (XEXP (dest, 0));
840 temp = adjust_automodify_address_nv (dest, word_mode, temp,
842 emit_move_insn (temp,
843 simplify_gen_subreg_concatn (word_mode, src,
845 j * UNITS_PER_WORD));
852 if (REG_P (dest) && !HARD_REGISTER_NUM_P (REGNO (dest)))
853 emit_insn (gen_rtx_CLOBBER (VOIDmode, dest));
855 for (i = 0; i < words; ++i)
856 emit_move_insn (simplify_gen_subreg_concatn (word_mode, dest,
859 simplify_gen_subreg_concatn (word_mode, src,
861 i * UNITS_PER_WORD));
864 if (real_dest != NULL_RTX)
866 rtx mdest, minsn, smove;
868 if (dest_mode == orig_mode)
871 mdest = simplify_gen_subreg (orig_mode, dest, GET_MODE (dest), 0);
872 minsn = emit_move_insn (real_dest, mdest);
874 smove = single_set (minsn);
875 gcc_assert (smove != NULL_RTX);
877 resolve_simple_move (smove, minsn);
880 insns = get_insns ();
883 move_eh_region_note (insn, insns);
885 emit_insn_before (insns, insn);
887 move_libcall_note (insn, insns);
888 remove_retval_note (insn);
894 /* Change a CLOBBER of a decomposed register into a CLOBBER of the
895 component registers. Return whether we changed something. */
898 resolve_clobber (rtx pat, rtx insn)
901 enum machine_mode orig_mode;
902 unsigned int words, i;
906 if (!resolve_reg_p (reg) && !resolve_subreg_p (reg))
909 /* If this clobber has a REG_LIBCALL note, then it is the initial
910 clobber added by emit_no_conflict_block. We were able to
911 decompose the register, so we no longer need the clobber. */
912 if (find_reg_note (insn, REG_LIBCALL, NULL_RTX) != NULL_RTX)
918 orig_mode = GET_MODE (reg);
919 words = GET_MODE_SIZE (orig_mode);
920 words = (words + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
922 ret = validate_change (NULL_RTX, &XEXP (pat, 0),
923 simplify_gen_subreg_concatn (word_mode, reg,
926 df_insn_rescan (insn);
927 gcc_assert (ret != 0);
929 for (i = words - 1; i > 0; --i)
933 x = simplify_gen_subreg_concatn (word_mode, reg, orig_mode,
935 x = gen_rtx_CLOBBER (VOIDmode, x);
936 emit_insn_after (x, insn);
942 /* A USE of a decomposed register is no longer meaningful. Return
943 whether we changed something. */
946 resolve_use (rtx pat, rtx insn)
948 if (resolve_reg_p (XEXP (pat, 0)) || resolve_subreg_p (XEXP (pat, 0)))
956 /* Checks if INSN is a decomposable multiword-shift or zero-extend and
957 sets the decomposable_context bitmap accordingly. A non-zero value
958 is returned if a decomposable insn has been found. */
961 find_decomposable_shift_zext (rtx insn)
967 set = single_set (insn);
972 if (GET_CODE (op) != ASHIFT
973 && GET_CODE (op) != LSHIFTRT
974 && GET_CODE (op) != ZERO_EXTEND)
977 op_operand = XEXP (op, 0);
978 if (!REG_P (SET_DEST (set)) || !REG_P (op_operand)
979 || HARD_REGISTER_NUM_P (REGNO (SET_DEST (set)))
980 || HARD_REGISTER_NUM_P (REGNO (op_operand))
981 || !SCALAR_INT_MODE_P (GET_MODE (op)))
984 if (GET_CODE (op) == ZERO_EXTEND)
986 if (GET_MODE (op_operand) != word_mode
987 || GET_MODE_BITSIZE (GET_MODE (op)) != 2 * BITS_PER_WORD)
990 else /* left or right shift */
992 if (GET_CODE (XEXP (op, 1)) != CONST_INT
993 || INTVAL (XEXP (op, 1)) < BITS_PER_WORD
994 || GET_MODE_BITSIZE (GET_MODE (op_operand)) != 2 * BITS_PER_WORD)
998 bitmap_set_bit (decomposable_context, REGNO (SET_DEST (set)));
1000 if (GET_CODE (op) != ZERO_EXTEND)
1001 bitmap_set_bit (decomposable_context, REGNO (op_operand));
1006 /* Decompose a more than word wide shift (in INSN) of a multiword
1007 pseudo or a multiword zero-extend of a wordmode pseudo into a move
1008 and 'set to zero' insn. Return a pointer to the new insn when a
1009 replacement was done. */
1012 resolve_shift_zext (rtx insn)
1018 rtx src_reg, dest_reg, dest_zero;
1019 int src_reg_num, dest_reg_num, offset1, offset2, src_offset;
1021 set = single_set (insn);
1026 if (GET_CODE (op) != ASHIFT
1027 && GET_CODE (op) != LSHIFTRT
1028 && GET_CODE (op) != ZERO_EXTEND)
1031 op_operand = XEXP (op, 0);
1033 if (!resolve_reg_p (SET_DEST (set)) && !resolve_reg_p (op_operand))
1036 /* src_reg_num is the number of the word mode register which we
1037 are operating on. For a left shift and a zero_extend on little
1038 endian machines this is register 0. */
1039 src_reg_num = GET_CODE (op) == LSHIFTRT ? 1 : 0;
1041 if (WORDS_BIG_ENDIAN)
1042 src_reg_num = 1 - src_reg_num;
1044 if (GET_CODE (op) == ZERO_EXTEND)
1045 dest_reg_num = src_reg_num;
1047 dest_reg_num = 1 - src_reg_num;
1049 offset1 = UNITS_PER_WORD * dest_reg_num;
1050 offset2 = UNITS_PER_WORD * (1 - dest_reg_num);
1051 src_offset = UNITS_PER_WORD * src_reg_num;
1053 if (WORDS_BIG_ENDIAN != BYTES_BIG_ENDIAN)
1055 offset1 += UNITS_PER_WORD - 1;
1056 offset2 += UNITS_PER_WORD - 1;
1057 src_offset += UNITS_PER_WORD - 1;
1062 dest_reg = simplify_gen_subreg_concatn (word_mode, SET_DEST (set),
1063 GET_MODE (SET_DEST (set)),
1065 dest_zero = simplify_gen_subreg_concatn (word_mode, SET_DEST (set),
1066 GET_MODE (SET_DEST (set)),
1068 src_reg = simplify_gen_subreg_concatn (word_mode, op_operand,
1069 GET_MODE (op_operand),
1071 if (GET_CODE (op) != ZERO_EXTEND)
1073 int shift_count = INTVAL (XEXP (op, 1));
1074 if (shift_count > BITS_PER_WORD)
1075 src_reg = expand_shift (GET_CODE (op) == ASHIFT ?
1076 LSHIFT_EXPR : RSHIFT_EXPR,
1078 build_int_cst (NULL_TREE,
1079 shift_count - BITS_PER_WORD),
1083 if (dest_reg != src_reg)
1084 emit_move_insn (dest_reg, src_reg);
1085 emit_move_insn (dest_zero, CONST0_RTX (word_mode));
1086 insns = get_insns ();
1090 emit_insn_before (insns, insn);
1095 fprintf (dump_file, "; Replacing insn: %d with insns: ", INSN_UID (insn));
1096 for (in = insns; in != insn; in = NEXT_INSN (in))
1097 fprintf (dump_file, "%d ", INSN_UID (in));
1098 fprintf (dump_file, "\n");
1105 /* Look for registers which are always accessed via word-sized SUBREGs
1106 or via copies. Decompose these registers into several word-sized
1107 pseudo-registers. */
1110 decompose_multiword_subregs (void)
1116 df_set_flags (DF_DEFER_INSN_RESCAN);
1118 max = max_reg_num ();
1120 /* First see if there are any multi-word pseudo-registers. If there
1121 aren't, there is nothing we can do. This should speed up this
1122 pass in the normal case, since it should be faster than scanning
1127 for (i = FIRST_PSEUDO_REGISTER; i < max; ++i)
1129 if (regno_reg_rtx[i] != NULL
1130 && GET_MODE_SIZE (GET_MODE (regno_reg_rtx[i])) > UNITS_PER_WORD)
1137 /* FIXME: When the dataflow branch is merged, we can change this
1138 code to look for each multi-word pseudo-register and to find each
1139 insn which sets or uses that register. That should be faster
1140 than scanning all the insns. */
1142 decomposable_context = BITMAP_ALLOC (NULL);
1143 non_decomposable_context = BITMAP_ALLOC (NULL);
1145 reg_copy_graph = VEC_alloc (bitmap, heap, max);
1146 VEC_safe_grow (bitmap, heap, reg_copy_graph, max);
1147 memset (VEC_address (bitmap, reg_copy_graph), 0, sizeof (bitmap) * max);
1153 FOR_BB_INSNS (bb, insn)
1156 enum classify_move_insn cmi;
1160 || GET_CODE (PATTERN (insn)) == CLOBBER
1161 || GET_CODE (PATTERN (insn)) == USE)
1164 if (find_decomposable_shift_zext (insn))
1167 recog_memoized (insn);
1168 extract_insn (insn);
1170 set = simple_move (insn);
1173 cmi = NOT_SIMPLE_MOVE;
1178 retval = find_reg_note (insn, REG_RETVAL, NULL_RTX) != NULL_RTX;
1180 if (find_pseudo_copy (set) && !retval)
1181 cmi = SIMPLE_PSEUDO_REG_MOVE;
1183 && REG_P (SET_SRC (set))
1184 && HARD_REGISTER_P (SET_SRC (set)))
1188 /* We don't want to decompose an assignment which
1189 copies the value returned by a libcall to a
1190 pseudo-register. Doing that will lose the RETVAL
1191 note with no real gain. */
1192 cmi = NOT_SIMPLE_MOVE;
1194 /* If we have a RETVAL note, there should be an
1195 EQUAL note. We don't want to decompose any
1196 registers which that EQUAL note refers to
1197 directly. If we do, we will no longer know the
1198 value of the libcall. */
1199 note = find_reg_equal_equiv_note (insn);
1200 if (note != NULL_RTX)
1201 for_each_rtx (&XEXP (note, 0), find_decomposable_subregs,
1208 n = recog_data.n_operands;
1209 for (i = 0; i < n; ++i)
1211 for_each_rtx (&recog_data.operand[i],
1212 find_decomposable_subregs,
1215 /* We handle ASM_OPERANDS as a special case to support
1216 things like x86 rdtsc which returns a DImode value.
1217 We can decompose the output, which will certainly be
1218 operand 0, but not the inputs. */
1220 if (cmi == SIMPLE_MOVE
1221 && GET_CODE (SET_SRC (set)) == ASM_OPERANDS)
1223 gcc_assert (i == 0);
1224 cmi = NOT_SIMPLE_MOVE;
1230 bitmap_and_compl_into (decomposable_context, non_decomposable_context);
1231 if (!bitmap_empty_p (decomposable_context))
1235 sbitmap_iterator sbi;
1236 bitmap_iterator iter;
1239 propagate_pseudo_copies ();
1241 sub_blocks = sbitmap_alloc (last_basic_block);
1242 sbitmap_zero (sub_blocks);
1244 EXECUTE_IF_SET_IN_BITMAP (decomposable_context, 0, regno, iter)
1245 decompose_register (regno);
1251 FOR_BB_INSNS (bb, insn)
1259 next = NEXT_INSN (insn);
1262 pat = PATTERN (insn);
1263 if (GET_CODE (pat) == CLOBBER)
1265 if (resolve_clobber (pat, insn))
1268 else if (GET_CODE (pat) == USE)
1270 if (resolve_use (pat, insn))
1278 recog_memoized (insn);
1279 extract_insn (insn);
1281 set = simple_move (insn);
1284 rtx orig_insn = insn;
1285 bool cfi = control_flow_insn_p (insn);
1287 /* We can end up splitting loads to multi-word pseudos
1288 into separate loads to machine word size pseudos.
1289 When this happens, we first had one load that can
1290 throw, and after resolve_simple_move we'll have a
1291 bunch of loads (at least two). All those loads may
1292 trap if we can have non-call exceptions, so they
1293 all will end the current basic block. We split the
1294 block after the outer loop over all insns, but we
1295 make sure here that we will be able to split the
1296 basic block and still produce the correct control
1297 flow graph for it. */
1299 || (flag_non_call_exceptions
1300 && can_throw_internal (insn)));
1302 insn = resolve_simple_move (set, insn);
1303 if (insn != orig_insn)
1307 remove_retval_note (insn);
1309 recog_memoized (insn);
1310 extract_insn (insn);
1313 SET_BIT (sub_blocks, bb->index);
1318 rtx decomposed_shift;
1320 decomposed_shift = resolve_shift_zext (insn);
1321 if (decomposed_shift != NULL_RTX)
1324 insn = decomposed_shift;
1325 recog_memoized (insn);
1326 extract_insn (insn);
1330 for (i = recog_data.n_operands - 1; i >= 0; --i)
1331 for_each_rtx (recog_data.operand_loc[i],
1335 resolve_reg_notes (insn);
1337 if (num_validated_changes () > 0)
1339 for (i = recog_data.n_dups - 1; i >= 0; --i)
1341 rtx *pl = recog_data.dup_loc[i];
1342 int dup_num = recog_data.dup_num[i];
1343 rtx *px = recog_data.operand_loc[dup_num];
1345 validate_change (insn, pl, *px, 1);
1348 i = apply_change_group ();
1351 remove_retval_note (insn);
1359 /* If we had insns to split that caused control flow insns in the middle
1360 of a basic block, split those blocks now. Note that we only handle
1361 the case where splitting a load has caused multiple possibly trapping
1363 EXECUTE_IF_SET_IN_SBITMAP (sub_blocks, 0, i, sbi)
1368 bb = BASIC_BLOCK (i);
1369 insn = BB_HEAD (bb);
1374 if (control_flow_insn_p (insn))
1376 /* Split the block after insn. There will be a fallthru
1377 edge, which is OK so we keep it. We have to create the
1378 exception edges ourselves. */
1379 fallthru = split_block (bb, insn);
1380 rtl_make_eh_edge (NULL, bb, BB_END (bb));
1381 bb = fallthru->dest;
1382 insn = BB_HEAD (bb);
1385 insn = NEXT_INSN (insn);
1389 sbitmap_free (sub_blocks);
1396 for (i = 0; VEC_iterate (bitmap, reg_copy_graph, i, b); ++i)
1401 VEC_free (bitmap, heap, reg_copy_graph);
1403 BITMAP_FREE (decomposable_context);
1404 BITMAP_FREE (non_decomposable_context);
1407 /* Gate function for lower subreg pass. */
1410 gate_handle_lower_subreg (void)
1412 return flag_split_wide_types != 0;
1415 /* Implement first lower subreg pass. */
1418 rest_of_handle_lower_subreg (void)
1420 decompose_multiword_subregs ();
1424 /* Implement second lower subreg pass. */
1427 rest_of_handle_lower_subreg2 (void)
1429 decompose_multiword_subregs ();
1433 struct tree_opt_pass pass_lower_subreg =
1435 "subreg", /* name */
1436 gate_handle_lower_subreg, /* gate */
1437 rest_of_handle_lower_subreg, /* execute */
1440 0, /* static_pass_number */
1441 TV_LOWER_SUBREG, /* tv_id */
1442 0, /* properties_required */
1443 0, /* properties_provided */
1444 0, /* properties_destroyed */
1445 0, /* todo_flags_start */
1448 TODO_verify_flow, /* todo_flags_finish */
1452 struct tree_opt_pass pass_lower_subreg2 =
1454 "subreg2", /* name */
1455 gate_handle_lower_subreg, /* gate */
1456 rest_of_handle_lower_subreg2, /* execute */
1459 0, /* static_pass_number */
1460 TV_LOWER_SUBREG, /* tv_id */
1461 0, /* properties_required */
1462 0, /* properties_provided */
1463 0, /* properties_destroyed */
1464 0, /* todo_flags_start */
1468 TODO_verify_flow, /* todo_flags_finish */