1/* Decompose multiword subregs.
2 Copyright (C) 2007-2024 Free Software Foundation, Inc.
3 Contributed by Richard Henderson <rth@redhat.com>
4 Ian Lance Taylor <iant@google.com>
5
6This file is part of GCC.
7
8GCC is free software; you can redistribute it and/or modify it under
9the terms of the GNU General Public License as published by the Free
10Software Foundation; either version 3, or (at your option) any later
11version.
12
13GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14WARRANTY; without even the implied warranty of MERCHANTABILITY or
15FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16for more details.
17
18You should have received a copy of the GNU General Public License
19along with GCC; see the file COPYING3. If not see
20<http://www.gnu.org/licenses/>. */
21
22#include "config.h"
23#include "system.h"
24#include "coretypes.h"
25#include "backend.h"
26#include "rtl.h"
27#include "tree.h"
28#include "cfghooks.h"
29#include "df.h"
30#include "memmodel.h"
31#include "tm_p.h"
32#include "expmed.h"
33#include "insn-config.h"
34#include "emit-rtl.h"
35#include "recog.h"
36#include "cfgrtl.h"
37#include "cfgbuild.h"
38#include "dce.h"
39#include "expr.h"
40#include "explow.h"
41#include "tree-pass.h"
42#include "lower-subreg.h"
43#include "rtl-iter.h"
44#include "target.h"
45
46
47/* Decompose multi-word pseudo-registers into individual
48 pseudo-registers when possible and profitable. This is possible
49 when all the uses of a multi-word register are via SUBREG, or are
50 copies of the register to another location. Breaking apart the
51 register permits more CSE and permits better register allocation.
52 This is profitable if the machine does not have move instructions
53 to do this.
54
55 This pass only splits moves with modes that are wider than
56 word_mode and ASHIFTs, LSHIFTRTs, ASHIFTRTs and ZERO_EXTENDs with
57 integer modes that are twice the width of word_mode. The latter
58 could be generalized if there was a need to do this, but the trend in
59 architectures is to not need this.
60
61 There are two useful preprocessor defines for use by maintainers:
62
63 #define LOG_COSTS 1
64
65 if you wish to see the actual cost estimates that are being used
66 for each mode wider than word mode and the cost estimates for zero
67 extension and the shifts. This can be useful when port maintainers
68 are tuning insn rtx costs.
69
70 #define FORCE_LOWERING 1
71
72 if you wish to test the pass with all the transformation forced on.
73 This can be useful for finding bugs in the transformations. */
74
75#define LOG_COSTS 0
76#define FORCE_LOWERING 0
77
78/* Bit N in this bitmap is set if regno N is used in a context in
79 which we can decompose it. */
80static bitmap decomposable_context;
81
82/* Bit N in this bitmap is set if regno N is used in a context in
83 which it cannot be decomposed. */
84static bitmap non_decomposable_context;
85
86/* Bit N in this bitmap is set if regno N is used in a subreg
87 which changes the mode but not the size. This typically happens
88 when the register accessed as a floating-point value; we want to
89 avoid generating accesses to its subwords in integer modes. */
90static bitmap subreg_context;
91
92/* Bit N in the bitmap in element M of this array is set if there is a
93 copy from reg M to reg N. */
94static vec<bitmap> reg_copy_graph;
95
96struct target_lower_subreg default_target_lower_subreg;
97#if SWITCHABLE_TARGET
98struct target_lower_subreg *this_target_lower_subreg
99 = &default_target_lower_subreg;
100#endif
101
102#define twice_word_mode \
103 this_target_lower_subreg->x_twice_word_mode
104#define choices \
105 this_target_lower_subreg->x_choices
106
107/* Return true if MODE is a mode we know how to lower. When returning true,
108 store its byte size in *BYTES and its word size in *WORDS. */
109
110static inline bool
111interesting_mode_p (machine_mode mode, unsigned int *bytes,
112 unsigned int *words)
113{
114 if (!GET_MODE_SIZE (mode).is_constant (const_value: bytes))
115 return false;
116 *words = CEIL (*bytes, UNITS_PER_WORD);
117 return true;
118}
119
120/* RTXes used while computing costs. */
121struct cost_rtxes {
122 /* Source and target registers. */
123 rtx source;
124 rtx target;
125
126 /* A twice_word_mode ZERO_EXTEND of SOURCE. */
127 rtx zext;
128
129 /* A shift of SOURCE. */
130 rtx shift;
131
132 /* A SET of TARGET. */
133 rtx set;
134};
135
136/* Return the cost of a CODE shift in mode MODE by OP1 bits, using the
137 rtxes in RTXES. SPEED_P selects between the speed and size cost. */
138
139static int
140shift_cost (bool speed_p, struct cost_rtxes *rtxes, enum rtx_code code,
141 machine_mode mode, int op1)
142{
143 PUT_CODE (rtxes->shift, code);
144 PUT_MODE (x: rtxes->shift, mode);
145 PUT_MODE (x: rtxes->source, mode);
146 XEXP (rtxes->shift, 1) = gen_int_shift_amount (mode, op1);
147 return set_src_cost (x: rtxes->shift, mode, speed_p);
148}
149
150/* For each X in the range [0, BITS_PER_WORD), set SPLITTING[X]
151 to true if it is profitable to split a double-word CODE shift
152 of X + BITS_PER_WORD bits. SPEED_P says whether we are testing
153 for speed or size profitability.
154
155 Use the rtxes in RTXES to calculate costs. WORD_MOVE_ZERO_COST is
156 the cost of moving zero into a word-mode register. WORD_MOVE_COST
157 is the cost of moving between word registers. */
158
159static void
160compute_splitting_shift (bool speed_p, struct cost_rtxes *rtxes,
161 bool *splitting, enum rtx_code code,
162 int word_move_zero_cost, int word_move_cost)
163{
164 int wide_cost, narrow_cost, upper_cost, i;
165
166 for (i = 0; i < BITS_PER_WORD; i++)
167 {
168 wide_cost = shift_cost (speed_p, rtxes, code, twice_word_mode,
169 op1: i + BITS_PER_WORD);
170 if (i == 0)
171 narrow_cost = word_move_cost;
172 else
173 narrow_cost = shift_cost (speed_p, rtxes, code, mode: word_mode, op1: i);
174
175 if (code != ASHIFTRT)
176 upper_cost = word_move_zero_cost;
177 else if (i == BITS_PER_WORD - 1)
178 upper_cost = word_move_cost;
179 else
180 upper_cost = shift_cost (speed_p, rtxes, code, mode: word_mode,
181 BITS_PER_WORD - 1);
182
183 if (LOG_COSTS)
184 fprintf (stderr, format: "%s %s by %d: original cost %d, split cost %d + %d\n",
185 GET_MODE_NAME (twice_word_mode), GET_RTX_NAME (code),
186 i + BITS_PER_WORD, wide_cost, narrow_cost, upper_cost);
187
188 if (FORCE_LOWERING || wide_cost >= narrow_cost + upper_cost)
189 splitting[i] = true;
190 }
191}
192
193/* Compute what we should do when optimizing for speed or size; SPEED_P
194 selects which. Use RTXES for computing costs. */
195
196static void
197compute_costs (bool speed_p, struct cost_rtxes *rtxes)
198{
199 unsigned int i;
200 int word_move_zero_cost, word_move_cost;
201
202 PUT_MODE (x: rtxes->target, mode: word_mode);
203 SET_SRC (rtxes->set) = CONST0_RTX (word_mode);
204 word_move_zero_cost = set_rtx_cost (x: rtxes->set, speed_p);
205
206 SET_SRC (rtxes->set) = rtxes->source;
207 word_move_cost = set_rtx_cost (x: rtxes->set, speed_p);
208
209 if (LOG_COSTS)
210 fprintf (stderr, format: "%s move: from zero cost %d, from reg cost %d\n",
211 GET_MODE_NAME (word_mode), word_move_zero_cost, word_move_cost);
212
213 for (i = 0; i < MAX_MACHINE_MODE; i++)
214 {
215 machine_mode mode = (machine_mode) i;
216 unsigned int size, factor;
217 if (interesting_mode_p (mode, bytes: &size, words: &factor) && factor > 1)
218 {
219 unsigned int mode_move_cost;
220
221 PUT_MODE (x: rtxes->target, mode);
222 PUT_MODE (x: rtxes->source, mode);
223 mode_move_cost = set_rtx_cost (x: rtxes->set, speed_p);
224
225 if (LOG_COSTS)
226 fprintf (stderr, format: "%s move: original cost %d, split cost %d * %d\n",
227 GET_MODE_NAME (mode), mode_move_cost,
228 word_move_cost, factor);
229
230 if (FORCE_LOWERING || mode_move_cost >= word_move_cost * factor)
231 {
232 choices[speed_p].move_modes_to_split[i] = true;
233 choices[speed_p].something_to_do = true;
234 }
235 }
236 }
237
238 /* For the moves and shifts, the only case that is checked is one
239 where the mode of the target is an integer mode twice the width
240 of the word_mode.
241
242 If it is not profitable to split a double word move then do not
243 even consider the shifts or the zero extension. */
244 if (choices[speed_p].move_modes_to_split[(int) twice_word_mode])
245 {
246 int zext_cost;
247
248 /* The only case here to check to see if moving the upper part with a
249 zero is cheaper than doing the zext itself. */
250 PUT_MODE (x: rtxes->source, mode: word_mode);
251 zext_cost = set_src_cost (x: rtxes->zext, twice_word_mode, speed_p);
252
253 if (LOG_COSTS)
254 fprintf (stderr, format: "%s %s: original cost %d, split cost %d + %d\n",
255 GET_MODE_NAME (twice_word_mode), GET_RTX_NAME (ZERO_EXTEND),
256 zext_cost, word_move_cost, word_move_zero_cost);
257
258 if (FORCE_LOWERING || zext_cost >= word_move_cost + word_move_zero_cost)
259 choices[speed_p].splitting_zext = true;
260
261 compute_splitting_shift (speed_p, rtxes,
262 choices[speed_p].splitting_ashift, code: ASHIFT,
263 word_move_zero_cost, word_move_cost);
264 compute_splitting_shift (speed_p, rtxes,
265 choices[speed_p].splitting_lshiftrt, code: LSHIFTRT,
266 word_move_zero_cost, word_move_cost);
267 compute_splitting_shift (speed_p, rtxes,
268 choices[speed_p].splitting_ashiftrt, code: ASHIFTRT,
269 word_move_zero_cost, word_move_cost);
270 }
271}
272
273/* Do one-per-target initialisation. This involves determining
274 which operations on the machine are profitable. If none are found,
275 then the pass just returns when called. */
276
277void
278init_lower_subreg (void)
279{
280 struct cost_rtxes rtxes;
281
282 memset (s: this_target_lower_subreg, c: 0, n: sizeof (*this_target_lower_subreg));
283
284 twice_word_mode = GET_MODE_2XWIDER_MODE (m: word_mode).require ();
285
286 rtxes.target = gen_rtx_REG (word_mode, LAST_VIRTUAL_REGISTER + 1);
287 rtxes.source = gen_rtx_REG (word_mode, LAST_VIRTUAL_REGISTER + 2);
288 rtxes.set = gen_rtx_SET (rtxes.target, rtxes.source);
289 rtxes.zext = gen_rtx_ZERO_EXTEND (twice_word_mode, rtxes.source);
290 rtxes.shift = gen_rtx_ASHIFT (twice_word_mode, rtxes.source, const0_rtx);
291
292 if (LOG_COSTS)
293 fprintf (stderr, format: "\nSize costs\n==========\n\n");
294 compute_costs (speed_p: false, rtxes: &rtxes);
295
296 if (LOG_COSTS)
297 fprintf (stderr, format: "\nSpeed costs\n===========\n\n");
298 compute_costs (speed_p: true, rtxes: &rtxes);
299}
300
301static bool
302simple_move_operand (rtx x)
303{
304 if (GET_CODE (x) == SUBREG)
305 x = SUBREG_REG (x);
306
307 if (!OBJECT_P (x))
308 return false;
309
310 if (GET_CODE (x) == LABEL_REF
311 || GET_CODE (x) == SYMBOL_REF
312 || GET_CODE (x) == HIGH
313 || GET_CODE (x) == CONST)
314 return false;
315
316 if (MEM_P (x)
317 && (MEM_VOLATILE_P (x)
318 || mode_dependent_address_p (XEXP (x, 0), MEM_ADDR_SPACE (x))))
319 return false;
320
321 return true;
322}
323
324/* If X is an operator that can be treated as a simple move that we
325 can split, then return the operand that is operated on. */
326
327static rtx
328operand_for_swap_move_operator (rtx x)
329{
330 /* A word sized rotate of a register pair is equivalent to swapping
331 the registers in the register pair. */
332 if (GET_CODE (x) == ROTATE
333 && GET_MODE (x) == twice_word_mode
334 && simple_move_operand (XEXP (x, 0))
335 && CONST_INT_P (XEXP (x, 1))
336 && INTVAL (XEXP (x, 1)) == BITS_PER_WORD)
337 return XEXP (x, 0);
338
339 return NULL_RTX;
340}
341
342/* If INSN is a single set between two objects that we want to split,
343 return the single set. SPEED_P says whether we are optimizing
344 INSN for speed or size.
345
346 INSN should have been passed to recog and extract_insn before this
347 is called. */
348
349static rtx
350simple_move (rtx_insn *insn, bool speed_p)
351{
352 rtx x, op;
353 rtx set;
354 machine_mode mode;
355
356 if (recog_data.n_operands != 2)
357 return NULL_RTX;
358
359 set = single_set (insn);
360 if (!set)
361 return NULL_RTX;
362
363 x = SET_DEST (set);
364 if (x != recog_data.operand[0] && x != recog_data.operand[1])
365 return NULL_RTX;
366 if (!simple_move_operand (x))
367 return NULL_RTX;
368
369 x = SET_SRC (set);
370 if ((op = operand_for_swap_move_operator (x)) != NULL_RTX)
371 x = op;
372
373 if (x != recog_data.operand[0] && x != recog_data.operand[1])
374 return NULL_RTX;
375 /* For the src we can handle ASM_OPERANDS, and it is beneficial for
376 things like x86 rdtsc which returns a DImode value. */
377 if (GET_CODE (x) != ASM_OPERANDS
378 && !simple_move_operand (x))
379 return NULL_RTX;
380
381 /* We try to decompose in integer modes, to avoid generating
382 inefficient code copying between integer and floating point
383 registers. That means that we can't decompose if this is a
384 non-integer mode for which there is no integer mode of the same
385 size. */
386 mode = GET_MODE (SET_DEST (set));
387 scalar_int_mode int_mode;
388 if (!SCALAR_INT_MODE_P (mode)
389 && (!int_mode_for_size (size: GET_MODE_BITSIZE (mode), limit: 0).exists (mode: &int_mode)
390 || !targetm.modes_tieable_p (mode, int_mode)))
391 return NULL_RTX;
392
393 /* Reject PARTIAL_INT modes. They are used for processor specific
394 purposes and it's probably best not to tamper with them. */
395 if (GET_MODE_CLASS (mode) == MODE_PARTIAL_INT)
396 return NULL_RTX;
397
398 if (!choices[speed_p].move_modes_to_split[(int) mode])
399 return NULL_RTX;
400
401 return set;
402}
403
404/* If SET is a copy from one multi-word pseudo-register to another,
405 record that in reg_copy_graph. Return whether it is such a
406 copy. */
407
408static bool
409find_pseudo_copy (rtx set)
410{
411 rtx dest = SET_DEST (set);
412 rtx src = SET_SRC (set);
413 rtx op;
414 unsigned int rd, rs;
415 bitmap b;
416
417 if ((op = operand_for_swap_move_operator (x: src)) != NULL_RTX)
418 src = op;
419
420 if (!REG_P (dest) || !REG_P (src))
421 return false;
422
423 rd = REGNO (dest);
424 rs = REGNO (src);
425 if (HARD_REGISTER_NUM_P (rd) || HARD_REGISTER_NUM_P (rs))
426 return false;
427
428 b = reg_copy_graph[rs];
429 if (b == NULL)
430 {
431 b = BITMAP_ALLOC (NULL);
432 reg_copy_graph[rs] = b;
433 }
434
435 bitmap_set_bit (b, rd);
436
437 return true;
438}
439
440/* Look through the registers in DECOMPOSABLE_CONTEXT. For each case
441 where they are copied to another register, add the register to
442 which they are copied to DECOMPOSABLE_CONTEXT. Use
443 NON_DECOMPOSABLE_CONTEXT to limit this--we don't bother to track
444 copies of registers which are in NON_DECOMPOSABLE_CONTEXT. */
445
446static void
447propagate_pseudo_copies (void)
448{
449 auto_bitmap queue, propagate;
450
451 bitmap_copy (queue, decomposable_context);
452 do
453 {
454 bitmap_iterator iter;
455 unsigned int i;
456
457 bitmap_clear (propagate);
458
459 EXECUTE_IF_SET_IN_BITMAP (queue, 0, i, iter)
460 {
461 bitmap b = reg_copy_graph[i];
462 if (b)
463 bitmap_ior_and_compl_into (A: propagate, B: b, C: non_decomposable_context);
464 }
465
466 bitmap_and_compl (queue, propagate, decomposable_context);
467 bitmap_ior_into (decomposable_context, propagate);
468 }
469 while (!bitmap_empty_p (map: queue));
470}
471
472/* A pointer to one of these values is passed to
473 find_decomposable_subregs. */
474
475enum classify_move_insn
476{
477 /* Not a simple move from one location to another. */
478 NOT_SIMPLE_MOVE,
479 /* A simple move we want to decompose. */
480 DECOMPOSABLE_SIMPLE_MOVE,
481 /* Any other simple move. */
482 SIMPLE_MOVE
483};
484
485/* If we find a SUBREG in *LOC which we could use to decompose a
486 pseudo-register, set a bit in DECOMPOSABLE_CONTEXT. If we find an
487 unadorned register which is not a simple pseudo-register copy,
488 DATA will point at the type of move, and we set a bit in
489 DECOMPOSABLE_CONTEXT or NON_DECOMPOSABLE_CONTEXT as appropriate. */
490
491static void
492find_decomposable_subregs (rtx *loc, enum classify_move_insn *pcmi)
493{
494 subrtx_var_iterator::array_type array;
495 FOR_EACH_SUBRTX_VAR (iter, array, *loc, NONCONST)
496 {
497 rtx x = *iter;
498 if (GET_CODE (x) == SUBREG)
499 {
500 rtx inner = SUBREG_REG (x);
501 unsigned int regno, outer_size, inner_size, outer_words, inner_words;
502
503 if (!REG_P (inner))
504 continue;
505
506 regno = REGNO (inner);
507 if (HARD_REGISTER_NUM_P (regno))
508 {
509 iter.skip_subrtxes ();
510 continue;
511 }
512
513 if (!interesting_mode_p (GET_MODE (x), bytes: &outer_size, words: &outer_words)
514 || !interesting_mode_p (GET_MODE (inner), bytes: &inner_size,
515 words: &inner_words))
516 continue;
517
518 /* We only try to decompose single word subregs of multi-word
519 registers. When we find one, we return -1 to avoid iterating
520 over the inner register.
521
522 ??? This doesn't allow, e.g., DImode subregs of TImode values
523 on 32-bit targets. We would need to record the way the
524 pseudo-register was used, and only decompose if all the uses
525 were the same number and size of pieces. Hopefully this
526 doesn't happen much. */
527
528 if (outer_words == 1
529 && inner_words > 1
530 /* Don't allow to decompose floating point subregs of
531 multi-word pseudos if the floating point mode does
532 not have word size, because otherwise we'd generate
533 a subreg with that floating mode from a different
534 sized integral pseudo which is not allowed by
535 validate_subreg. */
536 && (!FLOAT_MODE_P (GET_MODE (x))
537 || outer_size == UNITS_PER_WORD))
538 {
539 bitmap_set_bit (decomposable_context, regno);
540 iter.skip_subrtxes ();
541 continue;
542 }
543
544 /* If this is a cast from one mode to another, where the modes
545 have the same size, and they are not tieable, then mark this
546 register as non-decomposable. If we decompose it we are
547 likely to mess up whatever the backend is trying to do. */
548 if (outer_words > 1
549 && outer_size == inner_size
550 && !targetm.modes_tieable_p (GET_MODE (x), GET_MODE (inner)))
551 {
552 bitmap_set_bit (non_decomposable_context, regno);
553 bitmap_set_bit (subreg_context, regno);
554 iter.skip_subrtxes ();
555 continue;
556 }
557 }
558 else if (REG_P (x))
559 {
560 unsigned int regno, size, words;
561
562 /* We will see an outer SUBREG before we see the inner REG, so
563 when we see a plain REG here it means a direct reference to
564 the register.
565
566 If this is not a simple copy from one location to another,
567 then we cannot decompose this register. If this is a simple
568 copy we want to decompose, and the mode is right,
569 then we mark the register as decomposable.
570 Otherwise we don't say anything about this register --
571 it could be decomposed, but whether that would be
572 profitable depends upon how it is used elsewhere.
573
574 We only set bits in the bitmap for multi-word
575 pseudo-registers, since those are the only ones we care about
576 and it keeps the size of the bitmaps down. */
577
578 regno = REGNO (x);
579 if (!HARD_REGISTER_NUM_P (regno)
580 && interesting_mode_p (GET_MODE (x), bytes: &size, words: &words)
581 && words > 1)
582 {
583 switch (*pcmi)
584 {
585 case NOT_SIMPLE_MOVE:
586 bitmap_set_bit (non_decomposable_context, regno);
587 break;
588 case DECOMPOSABLE_SIMPLE_MOVE:
589 if (targetm.modes_tieable_p (GET_MODE (x), word_mode))
590 bitmap_set_bit (decomposable_context, regno);
591 break;
592 case SIMPLE_MOVE:
593 break;
594 default:
595 gcc_unreachable ();
596 }
597 }
598 }
599 else if (MEM_P (x))
600 {
601 enum classify_move_insn cmi_mem = NOT_SIMPLE_MOVE;
602
603 /* Any registers used in a MEM do not participate in a
604 SIMPLE_MOVE or DECOMPOSABLE_SIMPLE_MOVE. Do our own recursion
605 here, and return -1 to block the parent's recursion. */
606 find_decomposable_subregs (loc: &XEXP (x, 0), pcmi: &cmi_mem);
607 iter.skip_subrtxes ();
608 }
609 }
610}
611
612/* Decompose REGNO into word-sized components. We smash the REG node
613 in place. This ensures that (1) something goes wrong quickly if we
614 fail to make some replacement, and (2) the debug information inside
615 the symbol table is automatically kept up to date. */
616
617static void
618decompose_register (unsigned int regno)
619{
620 rtx reg;
621 unsigned int size, words, i;
622 rtvec v;
623
624 reg = regno_reg_rtx[regno];
625
626 regno_reg_rtx[regno] = NULL_RTX;
627
628 if (!interesting_mode_p (GET_MODE (reg), bytes: &size, words: &words))
629 gcc_unreachable ();
630
631 v = rtvec_alloc (words);
632 for (i = 0; i < words; ++i)
633 RTVEC_ELT (v, i) = gen_reg_rtx_offset (reg, word_mode, i * UNITS_PER_WORD);
634
635 PUT_CODE (reg, CONCATN);
636 XVEC (reg, 0) = v;
637
638 if (dump_file)
639 {
640 fprintf (stream: dump_file, format: "; Splitting reg %u ->", regno);
641 for (i = 0; i < words; ++i)
642 fprintf (stream: dump_file, format: " %u", REGNO (XVECEXP (reg, 0, i)));
643 fputc (c: '\n', stream: dump_file);
644 }
645}
646
647/* Get a SUBREG of a CONCATN. */
648
649static rtx
650simplify_subreg_concatn (machine_mode outermode, rtx op, poly_uint64 orig_byte)
651{
652 unsigned int outer_size, outer_words, inner_size, inner_words;
653 machine_mode innermode, partmode;
654 rtx part;
655 unsigned int final_offset;
656 unsigned int byte;
657
658 innermode = GET_MODE (op);
659 if (!interesting_mode_p (mode: outermode, bytes: &outer_size, words: &outer_words)
660 || !interesting_mode_p (mode: innermode, bytes: &inner_size, words: &inner_words))
661 gcc_unreachable ();
662
663 /* Must be constant if interesting_mode_p passes. */
664 byte = orig_byte.to_constant ();
665 gcc_assert (GET_CODE (op) == CONCATN);
666 gcc_assert (byte % outer_size == 0);
667
668 gcc_assert (byte < inner_size);
669 if (outer_size > inner_size)
670 return NULL_RTX;
671
672 inner_size /= XVECLEN (op, 0);
673 part = XVECEXP (op, 0, byte / inner_size);
674 partmode = GET_MODE (part);
675
676 final_offset = byte % inner_size;
677 if (final_offset + outer_size > inner_size)
678 return NULL_RTX;
679
680 /* VECTOR_CSTs in debug expressions are expanded into CONCATN instead of
681 regular CONST_VECTORs. They have vector or integer modes, depending
682 on the capabilities of the target. Cope with them. */
683 if (partmode == VOIDmode && VECTOR_MODE_P (innermode))
684 partmode = GET_MODE_INNER (innermode);
685 else if (partmode == VOIDmode)
686 partmode = mode_for_size (inner_size * BITS_PER_UNIT,
687 GET_MODE_CLASS (innermode), 0).require ();
688
689 return simplify_gen_subreg (outermode, op: part, innermode: partmode, byte: final_offset);
690}
691
692/* Wrapper around simplify_gen_subreg which handles CONCATN. */
693
694static rtx
695simplify_gen_subreg_concatn (machine_mode outermode, rtx op,
696 machine_mode innermode, unsigned int byte)
697{
698 rtx ret;
699
700 /* We have to handle generating a SUBREG of a SUBREG of a CONCATN.
701 If OP is a SUBREG of a CONCATN, then it must be a simple mode
702 change with the same size and offset 0, or it must extract a
703 part. We shouldn't see anything else here. */
704 if (GET_CODE (op) == SUBREG && GET_CODE (SUBREG_REG (op)) == CONCATN)
705 {
706 rtx op2;
707
708 if (known_eq (GET_MODE_SIZE (GET_MODE (op)),
709 GET_MODE_SIZE (GET_MODE (SUBREG_REG (op))))
710 && known_eq (SUBREG_BYTE (op), 0))
711 return simplify_gen_subreg_concatn (outermode, SUBREG_REG (op),
712 GET_MODE (SUBREG_REG (op)), byte);
713
714 op2 = simplify_subreg_concatn (GET_MODE (op), SUBREG_REG (op),
715 SUBREG_BYTE (op));
716 if (op2 == NULL_RTX)
717 {
718 /* We don't handle paradoxical subregs here. */
719 gcc_assert (!paradoxical_subreg_p (outermode, GET_MODE (op)));
720 gcc_assert (!paradoxical_subreg_p (op));
721 op2 = simplify_subreg_concatn (outermode, SUBREG_REG (op),
722 orig_byte: byte + SUBREG_BYTE (op));
723 gcc_assert (op2 != NULL_RTX);
724 return op2;
725 }
726
727 op = op2;
728 gcc_assert (op != NULL_RTX);
729 gcc_assert (innermode == GET_MODE (op));
730 }
731
732 if (GET_CODE (op) == CONCATN)
733 return simplify_subreg_concatn (outermode, op, orig_byte: byte);
734
735 ret = simplify_gen_subreg (outermode, op, innermode, byte);
736
737 /* If we see an insn like (set (reg:DI) (subreg:DI (reg:SI) 0)) then
738 resolve_simple_move will ask for the high part of the paradoxical
739 subreg, which does not have a value. Just return a zero. */
740 if (ret == NULL_RTX
741 && paradoxical_subreg_p (x: op))
742 return CONST0_RTX (outermode);
743
744 gcc_assert (ret != NULL_RTX);
745 return ret;
746}
747
748/* Return whether we should resolve X into the registers into which it
749 was decomposed. */
750
751static bool
752resolve_reg_p (rtx x)
753{
754 return GET_CODE (x) == CONCATN;
755}
756
757/* Return whether X is a SUBREG of a register which we need to
758 resolve. */
759
760static bool
761resolve_subreg_p (rtx x)
762{
763 if (GET_CODE (x) != SUBREG)
764 return false;
765 return resolve_reg_p (SUBREG_REG (x));
766}
767
768/* Look for SUBREGs in *LOC which need to be decomposed. */
769
770static bool
771resolve_subreg_use (rtx *loc, rtx insn)
772{
773 subrtx_ptr_iterator::array_type array;
774 FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST)
775 {
776 rtx *loc = *iter;
777 rtx x = *loc;
778 if (resolve_subreg_p (x))
779 {
780 x = simplify_subreg_concatn (GET_MODE (x), SUBREG_REG (x),
781 SUBREG_BYTE (x));
782
783 /* It is possible for a note to contain a reference which we can
784 decompose. In this case, return 1 to the caller to indicate
785 that the note must be removed. */
786 if (!x)
787 {
788 gcc_assert (!insn);
789 return true;
790 }
791
792 validate_change (insn, loc, x, 1);
793 iter.skip_subrtxes ();
794 }
795 else if (resolve_reg_p (x))
796 /* Return 1 to the caller to indicate that we found a direct
797 reference to a register which is being decomposed. This can
798 happen inside notes, multiword shift or zero-extend
799 instructions. */
800 return true;
801 }
802
803 return false;
804}
805
806/* Resolve any decomposed registers which appear in register notes on
807 INSN. */
808
809static void
810resolve_reg_notes (rtx_insn *insn)
811{
812 rtx *pnote, note;
813
814 note = find_reg_equal_equiv_note (insn);
815 if (note)
816 {
817 int old_count = num_validated_changes ();
818 if (resolve_subreg_use (loc: &XEXP (note, 0), NULL_RTX))
819 remove_note (insn, note);
820 else
821 if (old_count != num_validated_changes ())
822 df_notes_rescan (insn);
823 }
824
825 pnote = &REG_NOTES (insn);
826 while (*pnote != NULL_RTX)
827 {
828 bool del = false;
829
830 note = *pnote;
831 switch (REG_NOTE_KIND (note))
832 {
833 case REG_DEAD:
834 case REG_UNUSED:
835 if (resolve_reg_p (XEXP (note, 0)))
836 del = true;
837 break;
838
839 default:
840 break;
841 }
842
843 if (del)
844 *pnote = XEXP (note, 1);
845 else
846 pnote = &XEXP (note, 1);
847 }
848}
849
850/* Return whether X can be decomposed into subwords. */
851
852static bool
853can_decompose_p (rtx x)
854{
855 if (REG_P (x))
856 {
857 unsigned int regno = REGNO (x);
858
859 if (HARD_REGISTER_NUM_P (regno))
860 {
861 unsigned int byte, num_bytes, num_words;
862
863 if (!interesting_mode_p (GET_MODE (x), bytes: &num_bytes, words: &num_words))
864 return false;
865 for (byte = 0; byte < num_bytes; byte += UNITS_PER_WORD)
866 if (simplify_subreg_regno (regno, GET_MODE (x), byte, word_mode) < 0)
867 return false;
868 return true;
869 }
870 else
871 return !bitmap_bit_p (subreg_context, regno);
872 }
873
874 return true;
875}
876
877/* OPND is a concatn operand this is used with a simple move operator.
878 Return a new rtx with the concatn's operands swapped. */
879
880static rtx
881resolve_operand_for_swap_move_operator (rtx opnd)
882{
883 gcc_assert (GET_CODE (opnd) == CONCATN);
884 rtx concatn = copy_rtx (opnd);
885 rtx op0 = XVECEXP (concatn, 0, 0);
886 rtx op1 = XVECEXP (concatn, 0, 1);
887 XVECEXP (concatn, 0, 0) = op1;
888 XVECEXP (concatn, 0, 1) = op0;
889 return concatn;
890}
891
892/* Decompose the registers used in a simple move SET within INSN. If
893 we don't change anything, return INSN, otherwise return the start
894 of the sequence of moves. */
895
896static rtx_insn *
897resolve_simple_move (rtx set, rtx_insn *insn)
898{
899 rtx src, dest, real_dest, src_op;
900 rtx_insn *insns;
901 machine_mode orig_mode, dest_mode;
902 unsigned int orig_size, words;
903 bool pushing;
904
905 src = SET_SRC (set);
906 dest = SET_DEST (set);
907 orig_mode = GET_MODE (dest);
908
909 if (!interesting_mode_p (mode: orig_mode, bytes: &orig_size, words: &words))
910 gcc_unreachable ();
911 gcc_assert (words > 1);
912
913 start_sequence ();
914
915 /* We have to handle copying from a SUBREG of a decomposed reg where
916 the SUBREG is larger than word size. Rather than assume that we
917 can take a word_mode SUBREG of the destination, we copy to a new
918 register and then copy that to the destination. */
919
920 real_dest = NULL_RTX;
921
922 if ((src_op = operand_for_swap_move_operator (x: src)) != NULL_RTX)
923 {
924 if (resolve_reg_p (x: dest))
925 {
926 /* DEST is a CONCATN, so swap its operands and strip
927 SRC's operator. */
928 dest = resolve_operand_for_swap_move_operator (opnd: dest);
929 src = src_op;
930 if (resolve_reg_p (x: src))
931 {
932 gcc_assert (GET_CODE (src) == CONCATN);
933 if (reg_overlap_mentioned_p (XVECEXP (dest, 0, 0),
934 XVECEXP (src, 0, 1)))
935 {
936 /* If there is overlap between the first half of the
937 destination and what will be stored to the second one,
938 use a temporary pseudo. See PR114211. */
939 rtx tem = gen_reg_rtx (GET_MODE (XVECEXP (src, 0, 1)));
940 emit_move_insn (tem, XVECEXP (src, 0, 1));
941 src = copy_rtx (src);
942 XVECEXP (src, 0, 1) = tem;
943 }
944 }
945 }
946 else if (resolve_reg_p (x: src_op))
947 {
948 /* SRC is an operation on a CONCATN, so strip the operator and
949 swap the CONCATN's operands. */
950 src = resolve_operand_for_swap_move_operator (opnd: src_op);
951 }
952 }
953
954 if (GET_CODE (src) == SUBREG
955 && resolve_reg_p (SUBREG_REG (src))
956 && (maybe_ne (SUBREG_BYTE (src), b: 0)
957 || maybe_ne (a: orig_size, b: GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))))))
958 {
959 real_dest = dest;
960 dest = gen_reg_rtx (orig_mode);
961 if (REG_P (real_dest))
962 REG_ATTRS (dest) = REG_ATTRS (real_dest);
963 }
964
965 /* Similarly if we are copying to a SUBREG of a decomposed reg where
966 the SUBREG is larger than word size. */
967
968 if (GET_CODE (dest) == SUBREG
969 && resolve_reg_p (SUBREG_REG (dest))
970 && (maybe_ne (SUBREG_BYTE (dest), b: 0)
971 || maybe_ne (a: orig_size,
972 b: GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))))
973 {
974 rtx reg, smove;
975 rtx_insn *minsn;
976
977 reg = gen_reg_rtx (orig_mode);
978 minsn = emit_move_insn (reg, src);
979 smove = single_set (insn: minsn);
980 gcc_assert (smove != NULL_RTX);
981 resolve_simple_move (set: smove, insn: minsn);
982 src = reg;
983 }
984
985 /* If we didn't have any big SUBREGS of decomposed registers, and
986 neither side of the move is a register we are decomposing, then
987 we don't have to do anything here. */
988
989 if (src == SET_SRC (set)
990 && dest == SET_DEST (set)
991 && !resolve_reg_p (x: src)
992 && !resolve_subreg_p (x: src)
993 && !resolve_reg_p (x: dest)
994 && !resolve_subreg_p (x: dest))
995 {
996 end_sequence ();
997 return insn;
998 }
999
1000 /* It's possible for the code to use a subreg of a decomposed
1001 register while forming an address. We need to handle that before
1002 passing the address to emit_move_insn. We pass NULL_RTX as the
1003 insn parameter to resolve_subreg_use because we cannot validate
1004 the insn yet. */
1005 if (MEM_P (src) || MEM_P (dest))
1006 {
1007 int acg;
1008
1009 if (MEM_P (src))
1010 resolve_subreg_use (loc: &XEXP (src, 0), NULL_RTX);
1011 if (MEM_P (dest))
1012 resolve_subreg_use (loc: &XEXP (dest, 0), NULL_RTX);
1013 acg = apply_change_group ();
1014 gcc_assert (acg);
1015 }
1016
1017 /* If SRC is a register which we can't decompose, or has side
1018 effects, we need to move via a temporary register. */
1019
1020 if (!can_decompose_p (x: src)
1021 || side_effects_p (src)
1022 || GET_CODE (src) == ASM_OPERANDS)
1023 {
1024 rtx reg;
1025
1026 reg = gen_reg_rtx (orig_mode);
1027
1028 if (AUTO_INC_DEC)
1029 {
1030 rtx_insn *move = emit_move_insn (reg, src);
1031 if (MEM_P (src))
1032 {
1033 rtx note = find_reg_note (insn, REG_INC, NULL_RTX);
1034 if (note)
1035 add_reg_note (move, REG_INC, XEXP (note, 0));
1036 }
1037 }
1038 else
1039 emit_move_insn (reg, src);
1040
1041 src = reg;
1042 }
1043
1044 /* If DEST is a register which we can't decompose, or has side
1045 effects, we need to first move to a temporary register. We
1046 handle the common case of pushing an operand directly. We also
1047 go through a temporary register if it holds a floating point
1048 value. This gives us better code on systems which can't move
1049 data easily between integer and floating point registers. */
1050
1051 dest_mode = orig_mode;
1052 pushing = push_operand (dest, dest_mode);
1053 if (!can_decompose_p (x: dest)
1054 || (side_effects_p (dest) && !pushing)
1055 || (!SCALAR_INT_MODE_P (dest_mode)
1056 && !resolve_reg_p (x: dest)
1057 && !resolve_subreg_p (x: dest)))
1058 {
1059 if (real_dest == NULL_RTX)
1060 real_dest = dest;
1061 if (!SCALAR_INT_MODE_P (dest_mode))
1062 dest_mode = int_mode_for_mode (dest_mode).require ();
1063 dest = gen_reg_rtx (dest_mode);
1064 if (REG_P (real_dest))
1065 REG_ATTRS (dest) = REG_ATTRS (real_dest);
1066 }
1067
1068 if (pushing)
1069 {
1070 unsigned int i, j, jinc;
1071
1072 gcc_assert (orig_size % UNITS_PER_WORD == 0);
1073 gcc_assert (GET_CODE (XEXP (dest, 0)) != PRE_MODIFY);
1074 gcc_assert (GET_CODE (XEXP (dest, 0)) != POST_MODIFY);
1075
1076 if (WORDS_BIG_ENDIAN == STACK_GROWS_DOWNWARD)
1077 {
1078 j = 0;
1079 jinc = 1;
1080 }
1081 else
1082 {
1083 j = words - 1;
1084 jinc = -1;
1085 }
1086
1087 for (i = 0; i < words; ++i, j += jinc)
1088 {
1089 rtx temp;
1090
1091 temp = copy_rtx (XEXP (dest, 0));
1092 temp = adjust_automodify_address_nv (dest, word_mode, temp,
1093 j * UNITS_PER_WORD);
1094 emit_move_insn (temp,
1095 simplify_gen_subreg_concatn (outermode: word_mode, op: src,
1096 innermode: orig_mode,
1097 byte: j * UNITS_PER_WORD));
1098 }
1099 }
1100 else
1101 {
1102 unsigned int i;
1103
1104 for (i = 0; i < words; ++i)
1105 {
1106 rtx t = simplify_gen_subreg_concatn (outermode: word_mode, op: dest,
1107 innermode: dest_mode,
1108 byte: i * UNITS_PER_WORD);
1109 /* simplify_gen_subreg_concatn can return (const_int 0) for
1110 some sub-objects of paradoxical subregs. As a source operand,
1111 that's fine. As a destination it must be avoided. Those are
1112 supposed to be don't care bits, so we can just drop that store
1113 on the floor. */
1114 if (t != CONST0_RTX (word_mode))
1115 emit_move_insn (t,
1116 simplify_gen_subreg_concatn (outermode: word_mode, op: src,
1117 innermode: orig_mode,
1118 byte: i * UNITS_PER_WORD));
1119 }
1120 }
1121
1122 if (real_dest != NULL_RTX)
1123 {
1124 rtx mdest, smove;
1125 rtx_insn *minsn;
1126
1127 if (dest_mode == orig_mode)
1128 mdest = dest;
1129 else
1130 mdest = simplify_gen_subreg (outermode: orig_mode, op: dest, GET_MODE (dest), byte: 0);
1131 minsn = emit_move_insn (real_dest, mdest);
1132
1133 if (AUTO_INC_DEC && MEM_P (real_dest)
1134 && !(resolve_reg_p (x: real_dest) || resolve_subreg_p (x: real_dest)))
1135 {
1136 rtx note = find_reg_note (insn, REG_INC, NULL_RTX);
1137 if (note)
1138 add_reg_note (minsn, REG_INC, XEXP (note, 0));
1139 }
1140
1141 smove = single_set (insn: minsn);
1142 gcc_assert (smove != NULL_RTX);
1143
1144 resolve_simple_move (set: smove, insn: minsn);
1145 }
1146
1147 insns = get_insns ();
1148 end_sequence ();
1149
1150 copy_reg_eh_region_note_forward (insn, insns, NULL_RTX);
1151
1152 emit_insn_before (insns, insn);
1153
1154 /* If we get here via self-recursion, then INSN is not yet in the insns
1155 chain and delete_insn will fail. We only want to remove INSN from the
1156 current sequence. See PR56738. */
1157 if (in_sequence_p ())
1158 remove_insn (insn);
1159 else
1160 delete_insn (insn);
1161
1162 return insns;
1163}
1164
1165/* Change a CLOBBER of a decomposed register into a CLOBBER of the
1166 component registers. Return whether we changed something. */
1167
1168static bool
1169resolve_clobber (rtx pat, rtx_insn *insn)
1170{
1171 rtx reg;
1172 machine_mode orig_mode;
1173 unsigned int orig_size, words, i;
1174 int ret;
1175
1176 reg = XEXP (pat, 0);
1177 /* For clobbers we can look through paradoxical subregs which
1178 we do not handle in simplify_gen_subreg_concatn. */
1179 if (paradoxical_subreg_p (x: reg))
1180 reg = SUBREG_REG (reg);
1181 if (!resolve_reg_p (x: reg) && !resolve_subreg_p (x: reg))
1182 return false;
1183
1184 orig_mode = GET_MODE (reg);
1185 if (!interesting_mode_p (mode: orig_mode, bytes: &orig_size, words: &words))
1186 gcc_unreachable ();
1187
1188 ret = validate_change (NULL_RTX, &XEXP (pat, 0),
1189 simplify_gen_subreg_concatn (outermode: word_mode, op: reg,
1190 innermode: orig_mode, byte: 0),
1191 0);
1192 df_insn_rescan (insn);
1193 gcc_assert (ret != 0);
1194
1195 for (i = words - 1; i > 0; --i)
1196 {
1197 rtx x;
1198
1199 x = simplify_gen_subreg_concatn (outermode: word_mode, op: reg, innermode: orig_mode,
1200 byte: i * UNITS_PER_WORD);
1201 x = gen_rtx_CLOBBER (VOIDmode, x);
1202 emit_insn_after (x, insn);
1203 }
1204
1205 resolve_reg_notes (insn);
1206
1207 return true;
1208}
1209
1210/* A USE of a decomposed register is no longer meaningful. Return
1211 whether we changed something. */
1212
1213static bool
1214resolve_use (rtx pat, rtx_insn *insn)
1215{
1216 if (resolve_reg_p (XEXP (pat, 0)) || resolve_subreg_p (XEXP (pat, 0)))
1217 {
1218 delete_insn (insn);
1219 return true;
1220 }
1221
1222 resolve_reg_notes (insn);
1223
1224 return false;
1225}
1226
1227/* A VAR_LOCATION can be simplified. */
1228
1229static void
1230resolve_debug (rtx_insn *insn)
1231{
1232 subrtx_ptr_iterator::array_type array;
1233 FOR_EACH_SUBRTX_PTR (iter, array, &PATTERN (insn), NONCONST)
1234 {
1235 rtx *loc = *iter;
1236 rtx x = *loc;
1237 if (resolve_subreg_p (x))
1238 {
1239 x = simplify_subreg_concatn (GET_MODE (x), SUBREG_REG (x),
1240 SUBREG_BYTE (x));
1241
1242 if (x)
1243 *loc = x;
1244 else
1245 x = copy_rtx (*loc);
1246 }
1247 if (resolve_reg_p (x))
1248 *loc = copy_rtx (x);
1249 }
1250
1251 df_insn_rescan (insn);
1252
1253 resolve_reg_notes (insn);
1254}
1255
1256/* Check if INSN is a decomposable multiword-shift or zero-extend and
1257 set the decomposable_context bitmap accordingly. SPEED_P is true
1258 if we are optimizing INSN for speed rather than size. Return true
1259 if INSN is decomposable. */
1260
1261static bool
1262find_decomposable_shift_zext (rtx_insn *insn, bool speed_p)
1263{
1264 rtx set;
1265 rtx op;
1266 rtx op_operand;
1267
1268 set = single_set (insn);
1269 if (!set)
1270 return false;
1271
1272 op = SET_SRC (set);
1273 if (GET_CODE (op) != ASHIFT
1274 && GET_CODE (op) != LSHIFTRT
1275 && GET_CODE (op) != ASHIFTRT
1276 && GET_CODE (op) != ZERO_EXTEND)
1277 return false;
1278
1279 op_operand = XEXP (op, 0);
1280 if (!REG_P (SET_DEST (set)) || !REG_P (op_operand)
1281 || HARD_REGISTER_NUM_P (REGNO (SET_DEST (set)))
1282 || HARD_REGISTER_NUM_P (REGNO (op_operand))
1283 || GET_MODE (op) != twice_word_mode)
1284 return false;
1285
1286 if (GET_CODE (op) == ZERO_EXTEND)
1287 {
1288 if (GET_MODE (op_operand) != word_mode
1289 || !choices[speed_p].splitting_zext)
1290 return false;
1291 }
1292 else /* left or right shift */
1293 {
1294 bool *splitting = (GET_CODE (op) == ASHIFT
1295 ? choices[speed_p].splitting_ashift
1296 : GET_CODE (op) == ASHIFTRT
1297 ? choices[speed_p].splitting_ashiftrt
1298 : choices[speed_p].splitting_lshiftrt);
1299 if (!CONST_INT_P (XEXP (op, 1))
1300 || !IN_RANGE (INTVAL (XEXP (op, 1)), BITS_PER_WORD,
1301 2 * BITS_PER_WORD - 1)
1302 || !splitting[INTVAL (XEXP (op, 1)) - BITS_PER_WORD])
1303 return false;
1304
1305 bitmap_set_bit (decomposable_context, REGNO (op_operand));
1306 }
1307
1308 bitmap_set_bit (decomposable_context, REGNO (SET_DEST (set)));
1309
1310 return true;
1311}
1312
1313/* Decompose a more than word wide shift (in INSN) of a multiword
1314 pseudo or a multiword zero-extend of a wordmode pseudo into a move
1315 and 'set to zero' insn. SPEED_P says whether we are optimizing
1316 for speed or size, when checking if a ZERO_EXTEND is preferable.
1317 Return a pointer to the new insn when a replacement was done. */
1318
1319static rtx_insn *
1320resolve_shift_zext (rtx_insn *insn, bool speed_p)
1321{
1322 rtx set;
1323 rtx op;
1324 rtx op_operand;
1325 rtx_insn *insns;
1326 rtx src_reg, dest_reg, dest_upper, upper_src = NULL_RTX;
1327 int src_reg_num, dest_reg_num, offset1, offset2, src_offset;
1328 scalar_int_mode inner_mode;
1329
1330 set = single_set (insn);
1331 if (!set)
1332 return NULL;
1333
1334 op = SET_SRC (set);
1335 if (GET_CODE (op) != ASHIFT
1336 && GET_CODE (op) != LSHIFTRT
1337 && GET_CODE (op) != ASHIFTRT
1338 && GET_CODE (op) != ZERO_EXTEND)
1339 return NULL;
1340
1341 op_operand = XEXP (op, 0);
1342 if (!is_a <scalar_int_mode> (GET_MODE (op_operand), result: &inner_mode))
1343 return NULL;
1344
1345 /* We can tear this operation apart only if the regs were already
1346 torn apart. */
1347 if (!resolve_reg_p (SET_DEST (set)) && !resolve_reg_p (x: op_operand))
1348 return NULL;
1349
1350 /* src_reg_num is the number of the word mode register which we
1351 are operating on. For a left shift and a zero_extend on little
1352 endian machines this is register 0. */
1353 src_reg_num = (GET_CODE (op) == LSHIFTRT || GET_CODE (op) == ASHIFTRT)
1354 ? 1 : 0;
1355
1356 if (WORDS_BIG_ENDIAN && GET_MODE_SIZE (mode: inner_mode) > UNITS_PER_WORD)
1357 src_reg_num = 1 - src_reg_num;
1358
1359 if (GET_CODE (op) == ZERO_EXTEND)
1360 dest_reg_num = WORDS_BIG_ENDIAN ? 1 : 0;
1361 else
1362 dest_reg_num = 1 - src_reg_num;
1363
1364 offset1 = UNITS_PER_WORD * dest_reg_num;
1365 offset2 = UNITS_PER_WORD * (1 - dest_reg_num);
1366 src_offset = UNITS_PER_WORD * src_reg_num;
1367
1368 start_sequence ();
1369
1370 dest_reg = simplify_gen_subreg_concatn (outermode: word_mode, SET_DEST (set),
1371 GET_MODE (SET_DEST (set)),
1372 byte: offset1);
1373 dest_upper = simplify_gen_subreg_concatn (outermode: word_mode, SET_DEST (set),
1374 GET_MODE (SET_DEST (set)),
1375 byte: offset2);
1376 src_reg = simplify_gen_subreg_concatn (outermode: word_mode, op: op_operand,
1377 GET_MODE (op_operand),
1378 byte: src_offset);
1379 if (GET_CODE (op) == ASHIFTRT
1380 && INTVAL (XEXP (op, 1)) != 2 * BITS_PER_WORD - 1)
1381 upper_src = expand_shift (RSHIFT_EXPR, word_mode, copy_rtx (src_reg),
1382 BITS_PER_WORD - 1, NULL_RTX, 0);
1383
1384 if (GET_CODE (op) != ZERO_EXTEND)
1385 {
1386 int shift_count = INTVAL (XEXP (op, 1));
1387 if (shift_count > BITS_PER_WORD)
1388 src_reg = expand_shift (GET_CODE (op) == ASHIFT ?
1389 LSHIFT_EXPR : RSHIFT_EXPR,
1390 word_mode, src_reg,
1391 shift_count - BITS_PER_WORD,
1392 dest_reg, GET_CODE (op) != ASHIFTRT);
1393 }
1394
1395 /* Consider using ZERO_EXTEND instead of setting DEST_UPPER to zero
1396 if this is considered reasonable. */
1397 if (GET_CODE (op) == LSHIFTRT
1398 && GET_MODE (op) == twice_word_mode
1399 && REG_P (SET_DEST (set))
1400 && !choices[speed_p].splitting_zext)
1401 {
1402 rtx tmp = force_reg (word_mode, copy_rtx (src_reg));
1403 tmp = simplify_gen_unary (code: ZERO_EXTEND, twice_word_mode, op: tmp, op_mode: word_mode);
1404 emit_move_insn (SET_DEST (set), tmp);
1405 }
1406 else
1407 {
1408 if (dest_reg != src_reg)
1409 emit_move_insn (dest_reg, src_reg);
1410 if (GET_CODE (op) != ASHIFTRT)
1411 emit_move_insn (dest_upper, CONST0_RTX (word_mode));
1412 else if (INTVAL (XEXP (op, 1)) == 2 * BITS_PER_WORD - 1)
1413 emit_move_insn (dest_upper, copy_rtx (src_reg));
1414 else
1415 emit_move_insn (dest_upper, upper_src);
1416 }
1417
1418 insns = get_insns ();
1419
1420 end_sequence ();
1421
1422 emit_insn_before (insns, insn);
1423
1424 if (dump_file)
1425 {
1426 rtx_insn *in;
1427 fprintf (stream: dump_file, format: "; Replacing insn: %d with insns: ", INSN_UID (insn));
1428 for (in = insns; in != insn; in = NEXT_INSN (insn: in))
1429 fprintf (stream: dump_file, format: "%d ", INSN_UID (insn: in));
1430 fprintf (stream: dump_file, format: "\n");
1431 }
1432
1433 delete_insn (insn);
1434 return insns;
1435}
1436
1437/* Print to dump_file a description of what we're doing with shift code CODE.
1438 SPLITTING[X] is true if we are splitting shifts by X + BITS_PER_WORD. */
1439
1440static void
1441dump_shift_choices (enum rtx_code code, bool *splitting)
1442{
1443 int i;
1444 const char *sep;
1445
1446 fprintf (stream: dump_file,
1447 format: " Splitting mode %s for %s lowering with shift amounts = ",
1448 GET_MODE_NAME (twice_word_mode), GET_RTX_NAME (code));
1449 sep = "";
1450 for (i = 0; i < BITS_PER_WORD; i++)
1451 if (splitting[i])
1452 {
1453 fprintf (stream: dump_file, format: "%s%d", sep, i + BITS_PER_WORD);
1454 sep = ",";
1455 }
1456 fprintf (stream: dump_file, format: "\n");
1457}
1458
1459/* Print to dump_file a description of what we're doing when optimizing
1460 for speed or size; SPEED_P says which. DESCRIPTION is a description
1461 of the SPEED_P choice. */
1462
1463static void
1464dump_choices (bool speed_p, const char *description)
1465{
1466 unsigned int size, factor, i;
1467
1468 fprintf (stream: dump_file, format: "Choices when optimizing for %s:\n", description);
1469
1470 for (i = 0; i < MAX_MACHINE_MODE; i++)
1471 if (interesting_mode_p (mode: (machine_mode) i, bytes: &size, words: &factor)
1472 && factor > 1)
1473 fprintf (stream: dump_file, format: " %s mode %s for copy lowering.\n",
1474 choices[speed_p].move_modes_to_split[i]
1475 ? "Splitting"
1476 : "Skipping",
1477 GET_MODE_NAME ((machine_mode) i));
1478
1479 fprintf (stream: dump_file, format: " %s mode %s for zero_extend lowering.\n",
1480 choices[speed_p].splitting_zext ? "Splitting" : "Skipping",
1481 GET_MODE_NAME (twice_word_mode));
1482
1483 dump_shift_choices (code: ASHIFT, choices[speed_p].splitting_ashift);
1484 dump_shift_choices (code: LSHIFTRT, choices[speed_p].splitting_lshiftrt);
1485 dump_shift_choices (code: ASHIFTRT, choices[speed_p].splitting_ashiftrt);
1486 fprintf (stream: dump_file, format: "\n");
1487}
1488
1489/* Look for registers which are always accessed via word-sized SUBREGs
1490 or -if DECOMPOSE_COPIES is true- via copies. Decompose these
1491 registers into several word-sized pseudo-registers. */
1492
1493static void
1494decompose_multiword_subregs (bool decompose_copies)
1495{
1496 unsigned int max;
1497 basic_block bb;
1498 bool speed_p;
1499
1500 if (dump_file)
1501 {
1502 dump_choices (speed_p: false, description: "size");
1503 dump_choices (speed_p: true, description: "speed");
1504 }
1505
1506 /* Check if this target even has any modes to consider lowering. */
1507 if (!choices[false].something_to_do && !choices[true].something_to_do)
1508 {
1509 if (dump_file)
1510 fprintf (stream: dump_file, format: "Nothing to do!\n");
1511 return;
1512 }
1513
1514 max = max_reg_num ();
1515
1516 /* First see if there are any multi-word pseudo-registers. If there
1517 aren't, there is nothing we can do. This should speed up this
1518 pass in the normal case, since it should be faster than scanning
1519 all the insns. */
1520 {
1521 unsigned int i;
1522 bool useful_modes_seen = false;
1523
1524 for (i = FIRST_PSEUDO_REGISTER; i < max; ++i)
1525 if (regno_reg_rtx[i] != NULL)
1526 {
1527 machine_mode mode = GET_MODE (regno_reg_rtx[i]);
1528 if (choices[false].move_modes_to_split[(int) mode]
1529 || choices[true].move_modes_to_split[(int) mode])
1530 {
1531 useful_modes_seen = true;
1532 break;
1533 }
1534 }
1535
1536 if (!useful_modes_seen)
1537 {
1538 if (dump_file)
1539 fprintf (stream: dump_file, format: "Nothing to lower in this function.\n");
1540 return;
1541 }
1542 }
1543
1544 if (df)
1545 {
1546 df_set_flags (DF_DEFER_INSN_RESCAN);
1547 run_word_dce ();
1548 }
1549
1550 /* FIXME: It may be possible to change this code to look for each
1551 multi-word pseudo-register and to find each insn which sets or
1552 uses that register. That should be faster than scanning all the
1553 insns. */
1554
1555 decomposable_context = BITMAP_ALLOC (NULL);
1556 non_decomposable_context = BITMAP_ALLOC (NULL);
1557 subreg_context = BITMAP_ALLOC (NULL);
1558
1559 reg_copy_graph.create (nelems: max);
1560 reg_copy_graph.safe_grow_cleared (len: max, exact: true);
1561 memset (s: reg_copy_graph.address (), c: 0, n: sizeof (bitmap) * max);
1562
1563 speed_p = optimize_function_for_speed_p (cfun);
1564 FOR_EACH_BB_FN (bb, cfun)
1565 {
1566 rtx_insn *insn;
1567
1568 FOR_BB_INSNS (bb, insn)
1569 {
1570 rtx set;
1571 enum classify_move_insn cmi;
1572 int i, n;
1573
1574 if (!INSN_P (insn)
1575 || GET_CODE (PATTERN (insn)) == CLOBBER
1576 || GET_CODE (PATTERN (insn)) == USE)
1577 continue;
1578
1579 recog_memoized (insn);
1580
1581 if (find_decomposable_shift_zext (insn, speed_p))
1582 continue;
1583
1584 extract_insn (insn);
1585
1586 set = simple_move (insn, speed_p);
1587
1588 if (!set)
1589 cmi = NOT_SIMPLE_MOVE;
1590 else
1591 {
1592 /* We mark pseudo-to-pseudo copies as decomposable during the
1593 second pass only. The first pass is so early that there is
1594 good chance such moves will be optimized away completely by
1595 subsequent optimizations anyway.
1596
1597 However, we call find_pseudo_copy even during the first pass
1598 so as to properly set up the reg_copy_graph. */
1599 if (find_pseudo_copy (set))
1600 cmi = decompose_copies? DECOMPOSABLE_SIMPLE_MOVE : SIMPLE_MOVE;
1601 else
1602 cmi = SIMPLE_MOVE;
1603 }
1604
1605 n = recog_data.n_operands;
1606 for (i = 0; i < n; ++i)
1607 {
1608 find_decomposable_subregs (loc: &recog_data.operand[i], pcmi: &cmi);
1609
1610 /* We handle ASM_OPERANDS as a special case to support
1611 things like x86 rdtsc which returns a DImode value.
1612 We can decompose the output, which will certainly be
1613 operand 0, but not the inputs. */
1614
1615 if (cmi == SIMPLE_MOVE
1616 && GET_CODE (SET_SRC (set)) == ASM_OPERANDS)
1617 {
1618 gcc_assert (i == 0);
1619 cmi = NOT_SIMPLE_MOVE;
1620 }
1621 }
1622 }
1623 }
1624
1625 bitmap_and_compl_into (decomposable_context, non_decomposable_context);
1626 if (!bitmap_empty_p (map: decomposable_context))
1627 {
1628 unsigned int i;
1629 sbitmap_iterator sbi;
1630 bitmap_iterator iter;
1631 unsigned int regno;
1632
1633 propagate_pseudo_copies ();
1634
1635 auto_sbitmap sub_blocks (last_basic_block_for_fn (cfun));
1636 bitmap_clear (sub_blocks);
1637
1638 EXECUTE_IF_SET_IN_BITMAP (decomposable_context, 0, regno, iter)
1639 decompose_register (regno);
1640
1641 FOR_EACH_BB_FN (bb, cfun)
1642 {
1643 rtx_insn *insn;
1644
1645 FOR_BB_INSNS (bb, insn)
1646 {
1647 rtx pat;
1648
1649 if (!INSN_P (insn))
1650 continue;
1651
1652 pat = PATTERN (insn);
1653 if (GET_CODE (pat) == CLOBBER)
1654 resolve_clobber (pat, insn);
1655 else if (GET_CODE (pat) == USE)
1656 resolve_use (pat, insn);
1657 else if (DEBUG_INSN_P (insn))
1658 resolve_debug (insn);
1659 else
1660 {
1661 rtx set;
1662 int i;
1663
1664 recog_memoized (insn);
1665 extract_insn (insn);
1666
1667 set = simple_move (insn, speed_p);
1668 if (set)
1669 {
1670 rtx_insn *orig_insn = insn;
1671 bool cfi = control_flow_insn_p (insn);
1672
1673 /* We can end up splitting loads to multi-word pseudos
1674 into separate loads to machine word size pseudos.
1675 When this happens, we first had one load that can
1676 throw, and after resolve_simple_move we'll have a
1677 bunch of loads (at least two). All those loads may
1678 trap if we can have non-call exceptions, so they
1679 all will end the current basic block. We split the
1680 block after the outer loop over all insns, but we
1681 make sure here that we will be able to split the
1682 basic block and still produce the correct control
1683 flow graph for it. */
1684 gcc_assert (!cfi
1685 || (cfun->can_throw_non_call_exceptions
1686 && can_throw_internal (insn)));
1687
1688 insn = resolve_simple_move (set, insn);
1689 if (insn != orig_insn)
1690 {
1691 recog_memoized (insn);
1692 extract_insn (insn);
1693
1694 if (cfi)
1695 bitmap_set_bit (map: sub_blocks, bitno: bb->index);
1696 }
1697 }
1698 else
1699 {
1700 rtx_insn *decomposed_shift;
1701
1702 decomposed_shift = resolve_shift_zext (insn, speed_p);
1703 if (decomposed_shift != NULL_RTX)
1704 {
1705 insn = decomposed_shift;
1706 recog_memoized (insn);
1707 extract_insn (insn);
1708 }
1709 }
1710
1711 for (i = recog_data.n_operands - 1; i >= 0; --i)
1712 resolve_subreg_use (loc: recog_data.operand_loc[i], insn);
1713
1714 resolve_reg_notes (insn);
1715
1716 if (num_validated_changes () > 0)
1717 {
1718 for (i = recog_data.n_dups - 1; i >= 0; --i)
1719 {
1720 rtx *pl = recog_data.dup_loc[i];
1721 int dup_num = recog_data.dup_num[i];
1722 rtx *px = recog_data.operand_loc[dup_num];
1723
1724 validate_unshare_change (insn, pl, *px, 1);
1725 }
1726
1727 i = apply_change_group ();
1728 gcc_assert (i);
1729 }
1730 }
1731 }
1732 }
1733
1734 /* If we had insns to split that caused control flow insns in the middle
1735 of a basic block, split those blocks now. Note that we only handle
1736 the case where splitting a load has caused multiple possibly trapping
1737 loads to appear. */
1738 EXECUTE_IF_SET_IN_BITMAP (sub_blocks, 0, i, sbi)
1739 {
1740 rtx_insn *insn, *end;
1741 edge fallthru;
1742
1743 bb = BASIC_BLOCK_FOR_FN (cfun, i);
1744 insn = BB_HEAD (bb);
1745 end = BB_END (bb);
1746
1747 while (insn != end)
1748 {
1749 if (control_flow_insn_p (insn))
1750 {
1751 /* Split the block after insn. There will be a fallthru
1752 edge, which is OK so we keep it. We have to create the
1753 exception edges ourselves. */
1754 fallthru = split_block (bb, insn);
1755 rtl_make_eh_edge (NULL, bb, BB_END (bb));
1756 bb = fallthru->dest;
1757 insn = BB_HEAD (bb);
1758 }
1759 else
1760 insn = NEXT_INSN (insn);
1761 }
1762 }
1763 }
1764
1765 for (bitmap b : reg_copy_graph)
1766 if (b)
1767 BITMAP_FREE (b);
1768
1769 reg_copy_graph.release ();
1770
1771 BITMAP_FREE (decomposable_context);
1772 BITMAP_FREE (non_decomposable_context);
1773 BITMAP_FREE (subreg_context);
1774}
1775
1776/* Implement first lower subreg pass. */
1777
1778namespace {
1779
1780const pass_data pass_data_lower_subreg =
1781{
1782 .type: RTL_PASS, /* type */
1783 .name: "subreg1", /* name */
1784 .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */
1785 .tv_id: TV_LOWER_SUBREG, /* tv_id */
1786 .properties_required: 0, /* properties_required */
1787 .properties_provided: 0, /* properties_provided */
1788 .properties_destroyed: 0, /* properties_destroyed */
1789 .todo_flags_start: 0, /* todo_flags_start */
1790 .todo_flags_finish: 0, /* todo_flags_finish */
1791};
1792
1793class pass_lower_subreg : public rtl_opt_pass
1794{
1795public:
1796 pass_lower_subreg (gcc::context *ctxt)
1797 : rtl_opt_pass (pass_data_lower_subreg, ctxt)
1798 {}
1799
1800 /* opt_pass methods: */
1801 bool gate (function *) final override { return flag_split_wide_types != 0; }
1802 unsigned int execute (function *) final override
1803 {
1804 decompose_multiword_subregs (decompose_copies: false);
1805 return 0;
1806 }
1807
1808}; // class pass_lower_subreg
1809
1810} // anon namespace
1811
1812rtl_opt_pass *
1813make_pass_lower_subreg (gcc::context *ctxt)
1814{
1815 return new pass_lower_subreg (ctxt);
1816}
1817
1818/* Implement second lower subreg pass. */
1819
1820namespace {
1821
1822const pass_data pass_data_lower_subreg2 =
1823{
1824 .type: RTL_PASS, /* type */
1825 .name: "subreg2", /* name */
1826 .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */
1827 .tv_id: TV_LOWER_SUBREG, /* tv_id */
1828 .properties_required: 0, /* properties_required */
1829 .properties_provided: 0, /* properties_provided */
1830 .properties_destroyed: 0, /* properties_destroyed */
1831 .todo_flags_start: 0, /* todo_flags_start */
1832 TODO_df_finish, /* todo_flags_finish */
1833};
1834
1835class pass_lower_subreg2 : public rtl_opt_pass
1836{
1837public:
1838 pass_lower_subreg2 (gcc::context *ctxt)
1839 : rtl_opt_pass (pass_data_lower_subreg2, ctxt)
1840 {}
1841
1842 /* opt_pass methods: */
1843 bool gate (function *) final override
1844 {
1845 return flag_split_wide_types && flag_split_wide_types_early;
1846 }
1847 unsigned int execute (function *) final override
1848 {
1849 decompose_multiword_subregs (decompose_copies: true);
1850 return 0;
1851 }
1852
1853}; // class pass_lower_subreg2
1854
1855} // anon namespace
1856
1857rtl_opt_pass *
1858make_pass_lower_subreg2 (gcc::context *ctxt)
1859{
1860 return new pass_lower_subreg2 (ctxt);
1861}
1862
1863/* Implement third lower subreg pass. */
1864
1865namespace {
1866
1867const pass_data pass_data_lower_subreg3 =
1868{
1869 .type: RTL_PASS, /* type */
1870 .name: "subreg3", /* name */
1871 .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */
1872 .tv_id: TV_LOWER_SUBREG, /* tv_id */
1873 .properties_required: 0, /* properties_required */
1874 .properties_provided: 0, /* properties_provided */
1875 .properties_destroyed: 0, /* properties_destroyed */
1876 .todo_flags_start: 0, /* todo_flags_start */
1877 TODO_df_finish, /* todo_flags_finish */
1878};
1879
1880class pass_lower_subreg3 : public rtl_opt_pass
1881{
1882public:
1883 pass_lower_subreg3 (gcc::context *ctxt)
1884 : rtl_opt_pass (pass_data_lower_subreg3, ctxt)
1885 {}
1886
1887 /* opt_pass methods: */
1888 bool gate (function *) final override { return flag_split_wide_types; }
1889 unsigned int execute (function *) final override
1890 {
1891 decompose_multiword_subregs (decompose_copies: true);
1892 return 0;
1893 }
1894
1895}; // class pass_lower_subreg3
1896
1897} // anon namespace
1898
1899rtl_opt_pass *
1900make_pass_lower_subreg3 (gcc::context *ctxt)
1901{
1902 return new pass_lower_subreg3 (ctxt);
1903}
1904

source code of gcc/lower-subreg.cc