Branch data Line data Source code
1 : : /* RTL dead zero/sign extension (code) elimination.
2 : : Copyright (C) 2000-2025 Free Software Foundation, Inc.
3 : :
4 : : This file is part of GCC.
5 : :
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
9 : : version.
10 : :
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
14 : : for more details.
15 : :
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/>. */
19 : :
20 : : #include "config.h"
21 : : #include "system.h"
22 : : #include "coretypes.h"
23 : : #include "backend.h"
24 : : #include "rtl.h"
25 : : #include "tree.h"
26 : : #include "memmodel.h"
27 : : #include "insn-config.h"
28 : : #include "emit-rtl.h"
29 : : #include "recog.h"
30 : : #include "cfganal.h"
31 : : #include "tree-pass.h"
32 : : #include "cfgrtl.h"
33 : : #include "rtl-iter.h"
34 : : #include "df.h"
35 : : #include "print-rtl.h"
36 : : #include "dbgcnt.h"
37 : : #include "diagnostic-core.h"
38 : : #include "target.h"
39 : :
40 : : /* These should probably move into a C++ class. */
41 : : static vec<bitmap_head> livein;
42 : : static bitmap all_blocks;
43 : : static bitmap livenow;
44 : : static bitmap changed_pseudos;
45 : : static bool modify;
46 : :
47 : : /* We consider four bit groups for liveness:
48 : : bit 0..7 (least significant byte)
49 : : bit 8..15 (second least significant byte)
50 : : bit 16..31
51 : : bit 32..BITS_PER_WORD-1 */
52 : :
53 : : /* For the given REG, return the number of bit groups implied by the
54 : : size of the REG's mode, up to a maximum of 4 (number of bit groups
55 : : tracked by this pass).
56 : :
57 : : For partial integer and variable sized modes also return 4. This
58 : : could possibly be refined for something like PSI mode, but it
59 : : does not seem worth the effort. */
60 : :
61 : : static int
62 : 232950710 : group_limit (const_rtx reg)
63 : : {
64 : 232950710 : machine_mode mode = GET_MODE (reg);
65 : :
66 : 232950710 : if (!GET_MODE_BITSIZE (mode).is_constant ())
67 : : return 4;
68 : :
69 : 232950710 : int size = GET_MODE_SIZE (mode).to_constant ();
70 : :
71 : 232950710 : size = exact_log2 (size);
72 : :
73 : 232862952 : if (size < 0)
74 : : return 4;
75 : :
76 : 232862952 : size++;
77 : 232862952 : return (size > 4 ? 4 : size);
78 : : }
79 : :
80 : : /* Make all bit groups live for REGNO in bitmap BMAP. For hard regs,
81 : : we assume all groups are live. For a pseudo we consider the size
82 : : of the pseudo to avoid creating unnecessarily live chunks of data. */
83 : :
84 : : static void
85 : 4667768 : make_reg_live (bitmap bmap, int regno)
86 : : {
87 : 4667768 : int limit;
88 : :
89 : : /* For pseudos we can use the mode to limit how many bit groups
90 : : are marked as live since a pseudo only has one mode. Hard
91 : : registers have to be handled more conservatively. */
92 : 4667768 : if (regno > FIRST_PSEUDO_REGISTER)
93 : : {
94 : 909455 : rtx reg = regno_reg_rtx[regno];
95 : 909455 : limit = group_limit (reg);
96 : : }
97 : : else
98 : : limit = 4;
99 : :
100 : 23006816 : for (int i = 0; i < limit; i++)
101 : 18339048 : bitmap_set_bit (bmap, regno * 4 + i);
102 : 4667768 : }
103 : :
104 : : /* Note this pass could be used to narrow memory loads too. It's
105 : : not clear if that's profitable or not in general. */
106 : :
107 : : #define UNSPEC_P(X) (GET_CODE (X) == UNSPEC || GET_CODE (X) == UNSPEC_VOLATILE)
108 : :
109 : : /* If we know the destination of CODE only uses some low bits
110 : : (say just the QI bits of an SI operation), then return true
111 : : if we can propagate the need for just the subset of bits
112 : : from the destination to the sources.
113 : :
114 : : FIXME: This is safe for operands 1 and 2 of an IF_THEN_ELSE, but not
115 : : operand 0. Thus is likely would need some special casing to handle. */
116 : :
117 : : static bool
118 : 142678689 : safe_for_live_propagation (rtx_code code)
119 : : {
120 : : /* First handle rtx classes which as a whole are known to
121 : : be either safe or unsafe. */
122 : 142678689 : switch (GET_RTX_CLASS (code))
123 : : {
124 : : case RTX_OBJ:
125 : : case RTX_CONST_OBJ:
126 : : return true;
127 : :
128 : : case RTX_COMPARE:
129 : : case RTX_COMM_COMPARE:
130 : : case RTX_TERNARY:
131 : : return false;
132 : :
133 : 73649162 : default:
134 : 73649162 : break;
135 : : }
136 : :
137 : : /* What's left are specific codes. We only need to identify those
138 : : which are safe. */
139 : 73649162 : switch (code)
140 : : {
141 : : /* These are trivially safe. */
142 : : case SUBREG:
143 : : case NOT:
144 : : case ZERO_EXTEND:
145 : : case SIGN_EXTEND:
146 : : case TRUNCATE:
147 : : case PLUS:
148 : : case MINUS:
149 : : case MULT:
150 : : case SMUL_HIGHPART:
151 : : case UMUL_HIGHPART:
152 : : case AND:
153 : : case IOR:
154 : : case XOR:
155 : : return true;
156 : :
157 : : /* We can propagate for the shifted operand, but not the shift
158 : : count. The count is handled specially. */
159 : : case ASHIFT:
160 : : case LSHIFTRT:
161 : : case ASHIFTRT:
162 : : case SS_ASHIFT:
163 : : case US_ASHIFT:
164 : : return true;
165 : :
166 : : /* There may be other safe codes. If so they can be added
167 : : individually when discovered. */
168 : : default:
169 : : return false;
170 : : }
171 : : }
172 : :
173 : : /* Clear bits in LIVENOW and set bits in LIVE_TMP for objects
174 : : set/clobbered by OBJ contained in INSN.
175 : :
176 : : Conceptually it is always safe to ignore a particular destination
177 : : here as that will result in more chunks of data being considered
178 : : live. That's what happens when we "continue" the main loop when
179 : : we see something we don't know how to handle such as a vector
180 : : mode destination.
181 : :
182 : : The more accurate we are in identifying what objects (and chunks
183 : : within an object) are set by INSN, the more aggressive the
184 : : optimization phase during use handling will be. */
185 : :
186 : : static bool
187 : 137513902 : ext_dce_process_sets (rtx_insn *insn, rtx obj, bitmap live_tmp)
188 : : {
189 : 137513902 : bool skipped_dest = false;
190 : :
191 : 137513902 : subrtx_iterator::array_type array;
192 : 389329448 : FOR_EACH_SUBRTX (iter, array, obj, NONCONST)
193 : : {
194 : 251815546 : const_rtx x = *iter;
195 : :
196 : : /* An EXPR_LIST (from call fusage) ends in NULL_RTX. */
197 : 251815546 : if (x == NULL_RTX)
198 : 9509248 : continue;
199 : :
200 : 242306298 : if (UNSPEC_P (x))
201 : 566401 : continue;
202 : :
203 : 241739897 : if (GET_CODE (x) == SET || GET_CODE (x) == CLOBBER)
204 : : {
205 : 143315165 : unsigned bit = 0;
206 : 143315165 : x = SET_DEST (x);
207 : :
208 : : /* We don't support vector destinations or destinations
209 : : wider than DImode. */
210 : 143315165 : scalar_mode outer_mode;
211 : 147335739 : if (!is_a <scalar_mode> (GET_MODE (x), &outer_mode)
212 : 91812448 : || GET_MODE_BITSIZE (outer_mode) > HOST_BITS_PER_WIDE_INT)
213 : : {
214 : : /* Skip the subrtxs of this destination. There is
215 : : little value in iterating into the subobjects, so
216 : : just skip them for a bit of efficiency. */
217 : 55523291 : skipped_dest = true;
218 : 55523291 : iter.skip_subrtxes ();
219 : 307338837 : continue;
220 : : }
221 : :
222 : : /* We could have (strict_low_part (subreg ...)). We can not just
223 : : strip the STRICT_LOW_PART as that would result in clearing
224 : : some bits in LIVENOW that are still live. So process the
225 : : STRICT_LOW_PART specially. */
226 : 87791874 : if (GET_CODE (x) == STRICT_LOW_PART)
227 : : {
228 : 0 : x = XEXP (x, 0);
229 : :
230 : : /* The only valid operand of a STRICT_LOW_PART is a non
231 : : paradoxical SUBREG. */
232 : 0 : gcc_assert (SUBREG_P (x)
233 : : && !paradoxical_subreg_p (x)
234 : : && SUBREG_BYTE (x).is_constant ());
235 : :
236 : : /* I think we should always see a REG here. But let's
237 : : be sure. */
238 : 0 : gcc_assert (REG_P (SUBREG_REG (x)));
239 : :
240 : : /* The inner mode might be larger, just punt for
241 : : that case. Remember, we can not just continue to process
242 : : the inner RTXs due to the STRICT_LOW_PART. */
243 : 0 : if (!is_a <scalar_mode> (GET_MODE (SUBREG_REG (x)), &outer_mode)
244 : 0 : || GET_MODE_BITSIZE (outer_mode) > HOST_BITS_PER_WIDE_INT)
245 : : {
246 : : /* Skip the subrtxs of the STRICT_LOW_PART. We can't
247 : : process them because it'll set objects as no longer
248 : : live when they are in fact still live. */
249 : 0 : skipped_dest = true;
250 : 0 : iter.skip_subrtxes ();
251 : 0 : continue;
252 : : }
253 : :
254 : : /* LIVE_TMP contains the set groups that are live-out and set in
255 : : this insn. It is used to narrow the groups live-in for the
256 : : inputs of this insn.
257 : :
258 : : The simple thing to do is mark all the groups as live, but
259 : : that will significantly inhibit optimization.
260 : :
261 : : We also need to be careful in the case where we have an in-out
262 : : operand. If we're not careful we'd clear LIVE_TMP
263 : : incorrectly. */
264 : 0 : HOST_WIDE_INT rn = REGNO (SUBREG_REG (x));
265 : 0 : int limit = group_limit (SUBREG_REG (x));
266 : 0 : for (HOST_WIDE_INT i = 4 * rn; i < 4 * rn + limit; i++)
267 : 0 : if (bitmap_bit_p (livenow, i))
268 : 0 : bitmap_set_bit (live_tmp, i);
269 : :
270 : 0 : if (bitmap_empty_p (live_tmp))
271 : 0 : make_reg_live (live_tmp, rn);
272 : :
273 : : /* The mode of the SUBREG tells us how many bits we can
274 : : clear. */
275 : 0 : machine_mode mode = GET_MODE (x);
276 : 0 : HOST_WIDE_INT size
277 : 0 : = exact_log2 (GET_MODE_SIZE (mode).to_constant ()) + 1;
278 : 0 : bitmap_clear_range (livenow, 4 * rn, size);
279 : :
280 : : /* We have fully processed this destination. */
281 : 0 : iter.skip_subrtxes ();
282 : 0 : continue;
283 : 0 : }
284 : :
285 : : /* Phase one of destination handling. First remove any wrapper
286 : : such as SUBREG or ZERO_EXTRACT. */
287 : 87791874 : unsigned HOST_WIDE_INT mask
288 : 87791874 : = GET_MODE_MASK (GET_MODE_INNER (GET_MODE (x)));
289 : 87791874 : if (SUBREG_P (x))
290 : : {
291 : : /* If we have a SUBREG destination that is too wide, just
292 : : skip the destination rather than continuing this iterator.
293 : : While continuing would be better, we'd need to strip the
294 : : subreg and restart within the SET processing rather than
295 : : the top of the loop which just complicates the flow even
296 : : more. */
297 : 664979 : if (!is_a <scalar_mode> (GET_MODE (SUBREG_REG (x)), &outer_mode)
298 : 547376 : || GET_MODE_BITSIZE (outer_mode) > HOST_BITS_PER_WIDE_INT)
299 : : {
300 : 117603 : skipped_dest = true;
301 : 117603 : iter.skip_subrtxes ();
302 : 117603 : continue;
303 : : }
304 : :
305 : : /* We can safely strip a paradoxical subreg. The inner mode will
306 : : be narrower than the outer mode. We'll clear fewer bits in
307 : : LIVENOW than we'd like, but that's always safe. */
308 : 430203 : if (paradoxical_subreg_p (x))
309 : : x = XEXP (x, 0);
310 : 422761 : else if (SUBREG_BYTE (x).is_constant ())
311 : : {
312 : 422761 : bit = subreg_lsb (x).to_constant ();
313 : 422761 : mask = GET_MODE_MASK (GET_MODE (SUBREG_REG (x))) << bit;
314 : 422761 : gcc_assert (mask);
315 : : x = SUBREG_REG (x);
316 : : }
317 : : else
318 : : gcc_unreachable ();
319 : : }
320 : :
321 : 87674271 : if (GET_CODE (x) == ZERO_EXTRACT)
322 : : {
323 : : /* Unlike a SUBREG destination, a set of a ZERO_EXTRACT only
324 : : modifies the bits referenced in the ZERO_EXTRACT, the rest
325 : : remain the same. Thus we can not continue here, we must
326 : : either figure out what part of the destination is modified
327 : : or skip the sub-rtxs. */
328 : 3446 : skipped_dest = true;
329 : 3446 : iter.skip_subrtxes ();
330 : 3446 : continue;
331 : : }
332 : :
333 : : /* BIT >= 64 indicates something went horribly wrong. */
334 : 87670825 : gcc_assert (bit <= HOST_BITS_PER_WIDE_INT - 1);
335 : :
336 : : /* Now handle the actual object that was changed. */
337 : 87670825 : if (REG_P (x))
338 : : {
339 : : /* LIVE_TMP contains the set groups that are live-out and set in
340 : : this insn. It is used to narrow the groups live-in for the
341 : : inputs of this insn.
342 : :
343 : : The simple thing to do is mark all the groups as live, but
344 : : that will significantly inhibit optimization.
345 : :
346 : : We also need to be careful in the case where we have an in-out
347 : : operand. If we're not careful we'd clear LIVE_TMP
348 : : incorrectly. */
349 : 73717586 : HOST_WIDE_INT rn = REGNO (x);
350 : 73717586 : int limit = group_limit (x);
351 : 329299357 : for (HOST_WIDE_INT i = 4 * rn; i < 4 * rn + limit; i++)
352 : 255581771 : if (bitmap_bit_p (livenow, i))
353 : 248197777 : bitmap_set_bit (live_tmp, i);
354 : :
355 : 73717586 : if (bitmap_empty_p (live_tmp))
356 : 1251110 : make_reg_live (live_tmp, rn);
357 : :
358 : : /* Now clear the bits known written by this instruction.
359 : : Note that BIT need not be a power of two, consider a
360 : : ZERO_EXTRACT destination. */
361 : 73717586 : int start = (bit < 8 ? 0 : bit < 16 ? 1 : bit < 32 ? 2 : 3);
362 : 78774945 : int end = ((mask & ~HOST_WIDE_INT_UC (0xffffffff)) ? 4
363 : 28390816 : : (mask & HOST_WIDE_INT_UC (0xffff0000)) ? 3
364 : 5828554 : : (mask & 0xff00) ? 2 : 1);
365 : 73717586 : bitmap_clear_range (livenow, 4 * rn + start, end - start);
366 : : }
367 : : /* Some ports generate (clobber (const_int)). */
368 : 13953239 : else if (CONST_INT_P (x))
369 : 0 : continue;
370 : : else
371 : 13953239 : gcc_assert (CALL_P (insn)
372 : : || MEM_P (x)
373 : : || x == pc_rtx
374 : : || GET_CODE (x) == SCRATCH);
375 : :
376 : 87670825 : iter.skip_subrtxes ();
377 : 87670825 : }
378 : 98424732 : else if (GET_CODE (x) == COND_EXEC)
379 : : {
380 : : /* This isn't ideal, but may not be so bad in practice. */
381 : 0 : skipped_dest = true;
382 : 0 : iter.skip_subrtxes ();
383 : : }
384 : : }
385 : 137513902 : return skipped_dest;
386 : 137513902 : }
387 : :
388 : : /* INSN has a sign/zero extended source inside SET that we will
389 : : try to turn into a SUBREG. */
390 : : static void
391 : 4796 : ext_dce_try_optimize_insn (rtx_insn *insn, rtx set)
392 : : {
393 : 4796 : rtx src = SET_SRC (set);
394 : 4796 : rtx inner = XEXP (src, 0);
395 : :
396 : : /* Avoid (subreg (mem)) and other constructs which may be valid RTL, but
397 : : not useful for this optimization. */
398 : 4796 : if (!(REG_P (inner) || (SUBREG_P (inner) && REG_P (SUBREG_REG (inner)))))
399 : : return;
400 : :
401 : 2307 : rtx new_pattern;
402 : 2307 : if (dump_file)
403 : : {
404 : 0 : fprintf (dump_file, "Processing insn:\n");
405 : 0 : dump_insn_slim (dump_file, insn);
406 : 0 : fprintf (dump_file, "Trying to simplify pattern:\n");
407 : 0 : print_rtl_single (dump_file, SET_SRC (set));
408 : : }
409 : :
410 : : /* We decided to turn do the optimization but allow it to be rejected for
411 : : bisection purposes. */
412 : 2307 : if (!dbg_cnt (::ext_dce))
413 : : {
414 : 0 : if (dump_file)
415 : 0 : fprintf (dump_file, "Rejected due to debug counter.\n");
416 : 0 : return;
417 : : }
418 : :
419 : 4614 : new_pattern = simplify_gen_subreg (GET_MODE (src), inner,
420 : 2307 : GET_MODE (inner), 0);
421 : : /* simplify_gen_subreg may fail in which case NEW_PATTERN will be NULL.
422 : : We must not pass that as a replacement pattern to validate_change. */
423 : 2307 : if (new_pattern)
424 : : {
425 : 2307 : int ok = validate_change (insn, &SET_SRC (set), new_pattern, false);
426 : :
427 : 2307 : rtx x = SET_DEST (set);
428 : 2307 : while (SUBREG_P (x) || GET_CODE (x) == ZERO_EXTRACT)
429 : 0 : x = XEXP (x, 0);
430 : :
431 : 2307 : gcc_assert (REG_P (x));
432 : 2307 : if (ok)
433 : 2307 : bitmap_set_bit (changed_pseudos, REGNO (x));
434 : :
435 : 2307 : if (dump_file)
436 : : {
437 : 0 : if (ok)
438 : 0 : fprintf (dump_file, "Successfully transformed to:\n");
439 : : else
440 : 0 : fprintf (dump_file, "Failed transformation to:\n");
441 : :
442 : 0 : print_rtl_single (dump_file, new_pattern);
443 : 0 : fprintf (dump_file, "\n");
444 : : }
445 : : }
446 : : else
447 : : {
448 : 0 : if (dump_file)
449 : 0 : fprintf (dump_file, "Unable to generate valid SUBREG expression.\n");
450 : : }
451 : : }
452 : :
453 : : /* Some operators imply that their second operand is fully live,
454 : : regardless of how many bits in the output are live. An example
455 : : would be the shift count on a target without SHIFT_COUNT_TRUNCATED
456 : : defined.
457 : :
458 : : Return TRUE if CODE is such an operator. FALSE otherwise. */
459 : :
460 : : static bool
461 : 77334397 : binop_implies_op2_fully_live (rtx_code code)
462 : : {
463 : 0 : switch (code)
464 : : {
465 : : case ASHIFT:
466 : : case LSHIFTRT:
467 : : case ASHIFTRT:
468 : : case ROTATE:
469 : : case ROTATERT:
470 : : case SS_ASHIFT:
471 : : case US_ASHIFT:
472 : : return !SHIFT_COUNT_TRUNCATED;
473 : :
474 : 0 : default:
475 : 0 : return false;
476 : : }
477 : : }
478 : :
479 : : /* X, with code CODE, is an operation for which safe_for_live_propagation
480 : : holds true, and bits set in MASK are live in the result. Compute a
481 : : mask of (potentially) live bits in the non-constant inputs. In case of
482 : : binop_implies_op2_fully_live (e.g. shifts), the computed mask may
483 : : exclusively pertain to the first operand.
484 : :
485 : : This looks wrong as we may have some important operations embedded as
486 : : operands of another operation. For example, we might have an extension
487 : : wrapping a shift. It really feels like this needs to be recursing down
488 : : into operands much more often. */
489 : :
490 : : unsigned HOST_WIDE_INT
491 : 72080781 : carry_backpropagate (unsigned HOST_WIDE_INT mask, enum rtx_code code, rtx x)
492 : : {
493 : 73884163 : if (mask == 0)
494 : : return 0;
495 : :
496 : 73884137 : enum machine_mode mode = GET_MODE_INNER (GET_MODE (x));
497 : 73884137 : unsigned HOST_WIDE_INT mmask = GET_MODE_MASK (mode);
498 : :
499 : : /* While we don't try to optimize operations on types larger
500 : : than 64 bits, we do want to make sure not to invoke undefined
501 : : behavior when presented with such operations during use
502 : : processing. The safe thing to do is to just return mmask
503 : : for that scenario indicating every possible chunk is life. */
504 : 73884137 : scalar_int_mode smode;
505 : 73884137 : if (!is_a <scalar_int_mode> (mode, &smode)
506 : 61452239 : || GET_MODE_BITSIZE (smode) > HOST_BITS_PER_WIDE_INT)
507 : : return mmask;
508 : :
509 : 59340759 : switch (code)
510 : : {
511 : 16372900 : case PLUS:
512 : 16372900 : case MINUS:
513 : 16372900 : case MULT:
514 : 16372900 : return (HOST_WIDE_INT_UC (2) << floor_log2 (mask)) - 1;
515 : :
516 : : /* We propagate for the shifted operand, but not the shift
517 : : count. The count is handled specially. */
518 : 1367382 : case ASHIFT:
519 : 1367382 : if (CONST_INT_P (XEXP (x, 1))
520 : 2660970 : && UINTVAL (XEXP (x, 1)) < GET_MODE_BITSIZE (smode))
521 : 1293552 : return (HOST_WIDE_INT) mask >> INTVAL (XEXP (x, 1));
522 : 73830 : return (HOST_WIDE_INT_UC (2) << floor_log2 (mask)) - 1;
523 : :
524 : : /* We propagate for the shifted operand, but not the shift
525 : : count. The count is handled specially. */
526 : 728757 : case LSHIFTRT:
527 : 728757 : if (CONST_INT_P (XEXP (x, 1))
528 : 1423900 : && UINTVAL (XEXP (x, 1)) < GET_MODE_BITSIZE (smode))
529 : 695115 : return mmask & (mask << INTVAL (XEXP (x, 1)));
530 : : return mmask;
531 : :
532 : : /* We propagate for the shifted operand, but not the shift
533 : : count. The count is handled specially. */
534 : 318381 : case ASHIFTRT:
535 : 318381 : if (CONST_INT_P (XEXP (x, 1))
536 : 624924 : && UINTVAL (XEXP (x, 1)) < GET_MODE_BITSIZE (smode))
537 : : {
538 : 306535 : HOST_WIDE_INT sign = 0;
539 : 306535 : if (HOST_BITS_PER_WIDE_INT - clz_hwi (mask) + INTVAL (XEXP (x, 1))
540 : 306535 : > GET_MODE_BITSIZE (smode))
541 : 613070 : sign = HOST_WIDE_INT_1U << (GET_MODE_BITSIZE (smode) - 1);
542 : 306535 : return sign | (mmask & (mask << INTVAL (XEXP (x, 1))));
543 : : }
544 : : return mmask;
545 : :
546 : 61045 : case SMUL_HIGHPART:
547 : 61045 : case UMUL_HIGHPART:
548 : 61045 : if (XEXP (x, 1) == const0_rtx)
549 : : return 0;
550 : 61045 : if (XEXP (x, 1) == const1_rtx)
551 : : return mmask;
552 : 61045 : if (CONST_INT_P (XEXP (x, 1)))
553 : : {
554 : 0 : if (pow2p_hwi (INTVAL (XEXP (x, 1))))
555 : 0 : return mmask & (mask << (GET_MODE_BITSIZE (smode)
556 : 0 : - exact_log2 (INTVAL (XEXP (x, 1)))));
557 : :
558 : 0 : int bits = (HOST_BITS_PER_WIDE_INT + GET_MODE_BITSIZE (smode)
559 : 0 : - clz_hwi (mask) - ctz_hwi (INTVAL (XEXP (x, 1))));
560 : 0 : if (bits < GET_MODE_BITSIZE (smode))
561 : 0 : return (HOST_WIDE_INT_1U << bits) - 1;
562 : : }
563 : : return mmask;
564 : :
565 : 677280 : case SIGN_EXTEND:
566 : 677280 : if (!GET_MODE_BITSIZE (GET_MODE (x)).is_constant ()
567 : 677280 : || !GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))).is_constant ())
568 : : return -1;
569 : :
570 : : /* We want the mode of the inner object. We need to ensure its
571 : : sign bit is on in MASK. */
572 : 677280 : mode = GET_MODE_INNER (GET_MODE (XEXP (x, 0)));
573 : 677280 : if (mask & ~GET_MODE_MASK (mode))
574 : 676823 : mask |= HOST_WIDE_INT_1U << (GET_MODE_BITSIZE (mode).to_constant ()
575 : 676823 : - 1);
576 : :
577 : : /* Recurse into the operand. */
578 : 677280 : return carry_backpropagate (mask, GET_CODE (XEXP (x, 0)), XEXP (x, 0));
579 : :
580 : 1126102 : case ZERO_EXTEND:
581 : 1126102 : if (!GET_MODE_BITSIZE (GET_MODE (x)).is_constant ()
582 : 1126102 : || !GET_MODE_BITSIZE (GET_MODE (XEXP (x, 0))).is_constant ())
583 : : return -1;
584 : :
585 : : /* Recurse into the operand. */
586 : 1126102 : return carry_backpropagate (mask, GET_CODE (XEXP (x, 0)), XEXP (x, 0));
587 : :
588 : : /* We propagate for the shifted operand, but not the shift
589 : : count. The count is handled specially. */
590 : 0 : case SS_ASHIFT:
591 : 0 : case US_ASHIFT:
592 : 0 : if (CONST_INT_P (XEXP (x, 1))
593 : 0 : && UINTVAL (XEXP (x, 1)) < GET_MODE_BITSIZE (smode))
594 : : {
595 : 0 : return ((mmask & ~((unsigned HOST_WIDE_INT) mmask
596 : 0 : >> (INTVAL (XEXP (x, 1))
597 : 0 : + (XEXP (x, 1) != const0_rtx
598 : 0 : && code == SS_ASHIFT))))
599 : 0 : | ((HOST_WIDE_INT) mask >> INTVAL (XEXP (x, 1))));
600 : : }
601 : : return mmask;
602 : :
603 : : default:
604 : : return mask;
605 : : }
606 : : }
607 : :
608 : : /* Process uses in INSN contained in OBJ. Set appropriate bits in LIVENOW
609 : : for any chunks of pseudos that become live, potentially filtering using
610 : : bits from LIVE_TMP.
611 : :
612 : : If MODIFY is true, then optimize sign/zero extensions to SUBREGs when
613 : : the extended bits are never read and mark pseudos which had extensions
614 : : eliminated in CHANGED_PSEUDOS. */
615 : :
616 : : static void
617 : 137513902 : ext_dce_process_uses (rtx_insn *insn, rtx obj,
618 : : bitmap live_tmp, bool skipped_dest)
619 : : {
620 : 137513902 : subrtx_var_iterator::array_type array_var;
621 : 745237773 : FOR_EACH_SUBRTX_VAR (iter, array_var, obj, NONCONST)
622 : : {
623 : : /* An EXPR_LIST (from call fusage) ends in NULL_RTX. */
624 : 607723871 : rtx x = *iter;
625 : 607723871 : if (x == NULL_RTX)
626 : 9509248 : continue;
627 : :
628 : : /* So the basic idea in this FOR_EACH_SUBRTX_VAR loop is to
629 : : handle SETs explicitly, possibly propagating live information
630 : : into the uses.
631 : :
632 : : We may continue the loop at various points which will cause
633 : : iteration into the next level of RTL. Breaking from the loop
634 : : is never safe as it can lead us to fail to process some of the
635 : : RTL and thus not make objects live when necessary. */
636 : 598214623 : enum rtx_code xcode = GET_CODE (x);
637 : 598214623 : if (xcode == SET)
638 : : {
639 : 121607594 : const_rtx dst = SET_DEST (x);
640 : 121607594 : rtx src = SET_SRC (x);
641 : 121607594 : const_rtx y;
642 : 121607594 : unsigned HOST_WIDE_INT bit = 0;
643 : :
644 : : /* The code of the RHS of a SET. */
645 : 121607594 : enum rtx_code code = GET_CODE (src);
646 : :
647 : : /* ?!? How much of this should mirror SET handling, potentially
648 : : being shared? */
649 : 121607594 : if (SUBREG_P (dst) && SUBREG_BYTE (dst).is_constant ())
650 : : {
651 : 590972 : bit = subreg_lsb (dst).to_constant ();
652 : 590972 : if (bit >= HOST_BITS_PER_WIDE_INT)
653 : : bit = HOST_BITS_PER_WIDE_INT - 1;
654 : 590972 : dst = SUBREG_REG (dst);
655 : : }
656 : 121016622 : else if (GET_CODE (dst) == STRICT_LOW_PART)
657 : 10222 : dst = XEXP (dst, 0);
658 : :
659 : : /* Main processing of the uses. Two major goals here.
660 : :
661 : : First, we want to try and propagate liveness (or the lack
662 : : thereof) from the destination register to the source
663 : : register(s).
664 : :
665 : : Second, if the source is an extension, try to optimize
666 : : it into a SUBREG. The SUBREG form indicates we don't
667 : : care about the upper bits and will usually be copy
668 : : propagated away.
669 : :
670 : : If we fail to handle something in here, the expectation
671 : : is the iterator will dive into the sub-components and
672 : : mark all the chunks in any found REGs as live. */
673 : 121607594 : if (REG_P (dst) && safe_for_live_propagation (code))
674 : : {
675 : : /* Create a mask representing the bits of this output
676 : : operand that are live after this insn. We can use
677 : : this information to refine the live in state of
678 : : inputs to this insn in many cases.
679 : :
680 : : We have to do this on a per SET basis, we might have
681 : : an INSN with multiple SETS, some of which can narrow
682 : : the source operand liveness, some of which may not. */
683 : 72080781 : unsigned HOST_WIDE_INT dst_mask = 0;
684 : 72080781 : HOST_WIDE_INT rn = REGNO (dst);
685 : 72080781 : unsigned HOST_WIDE_INT mask_array[]
686 : : = { 0xff, 0xff00, HOST_WIDE_INT_UC (0xffff0000),
687 : : -HOST_WIDE_INT_UC (0x100000000) };
688 : 360403905 : for (int i = 0; i < 4; i++)
689 : 288323124 : if (bitmap_bit_p (live_tmp, 4 * rn + i))
690 : 232264904 : dst_mask |= mask_array[i];
691 : 72080781 : dst_mask >>= bit;
692 : :
693 : : /* If we ignored a destination during set processing, then
694 : : consider all the bits live. */
695 : 72080781 : if (skipped_dest)
696 : 25011336 : dst_mask = -1;
697 : :
698 : 72080781 : dst_mask = carry_backpropagate (dst_mask, code, src);
699 : :
700 : : /* ??? Could also handle ZERO_EXTRACT / SIGN_EXTRACT
701 : : of the source specially to improve optimization. */
702 : 72080781 : if (code == SIGN_EXTEND || code == ZERO_EXTEND)
703 : : {
704 : 1816041 : rtx inner = XEXP (src, 0);
705 : 1816041 : unsigned HOST_WIDE_INT src_mask
706 : 1816041 : = GET_MODE_MASK (GET_MODE_INNER (GET_MODE (inner)));
707 : :
708 : : /* DST_MASK could be zero if we had something in the SET
709 : : that we couldn't handle. */
710 : 1816041 : if (modify && !skipped_dest && (dst_mask & ~src_mask) == 0)
711 : 4796 : ext_dce_try_optimize_insn (insn, x);
712 : :
713 : : /* Stripping the extension here just seems wrong on multiple
714 : : levels. It's source side handling, so it seems like it
715 : : belongs in the loop below. Stripping here also makes it
716 : : harder than necessary to properly handle live bit groups
717 : : for (ANY_EXTEND (SUBREG)) where the SUBREG has
718 : : SUBREG_PROMOTED state. */
719 : 1816041 : dst_mask &= src_mask;
720 : 1816041 : src = XEXP (src, 0);
721 : 1816041 : code = GET_CODE (src);
722 : : }
723 : :
724 : : /* Optimization is done at this point. We just want to make
725 : : sure everything that should get marked as live is marked
726 : : from here onward. */
727 : :
728 : : /* We will handle the other operand of a binary operator
729 : : at the bottom of the loop by resetting Y. */
730 : 72080781 : if (BINARY_P (src))
731 : 22612423 : y = XEXP (src, 0);
732 : : else
733 : : y = src;
734 : :
735 : : /* We're inside a SET and want to process the source operands
736 : : making things live. Breaking from this loop will cause
737 : : the iterator to work on sub-rtxs, so it is safe to break
738 : : if we see something we don't know how to handle.
739 : :
740 : : This code is just hokey as it really just handles trivial
741 : : unary and binary cases. Otherwise the loop exits and we
742 : : continue iterating on sub-rtxs, but outside the set context. */
743 : : unsigned HOST_WIDE_INT save_mask = dst_mask;
744 : 116144977 : for (;;)
745 : : {
746 : : /* In general we want to restore DST_MASK before each loop
747 : : iteration. The exception is when the opcode implies that
748 : : the other operand is fully live. That's handled by
749 : : changing SAVE_MASK below. */
750 : 94112879 : dst_mask = save_mask;
751 : : /* Strip an outer paradoxical subreg. The bits outside
752 : : the inner mode are don't cares. So we can just strip
753 : : and process the inner object. */
754 : 94112879 : if (paradoxical_subreg_p (y))
755 : : y = XEXP (y, 0);
756 : 94028452 : else if (SUBREG_P (y) && SUBREG_BYTE (y).is_constant ())
757 : : {
758 : : /* We really want to know the outer code here, ie do we
759 : : have (ANY_EXTEND (SUBREG ...)) as we need to know if
760 : : the extension matches the SUBREG_PROMOTED state. In
761 : : that case optimizers can turn the extension into a
762 : : simple copy. Which means that bits outside the
763 : : SUBREG's mode are actually live.
764 : :
765 : : We don't want to mark those bits live unnecessarily
766 : : as that inhibits extension elimination in important
767 : : cases such as those in Coremark. So we need that
768 : : outer code.
769 : :
770 : : But if !TRULY_NOOP_TRUNCATION_MODES_P, the mode
771 : : change performed by Y would normally need to be a
772 : : TRUNCATE rather than a SUBREG. It is probably the
773 : : guarantee provided by SUBREG_PROMOTED_VAR_P that
774 : : allows the SUBREG in Y as an exception. We must
775 : : therefore preserve that guarantee and treat the
776 : : upper bits of the inner register as live
777 : : regardless of the outer code. See PR 120050. */
778 : 1979411 : if (!REG_P (SUBREG_REG (y))
779 : 1979411 : || (SUBREG_PROMOTED_VAR_P (y)
780 : 11204 : && ((GET_CODE (SET_SRC (x)) == SIGN_EXTEND
781 : 1159 : && SUBREG_PROMOTED_SIGNED_P (y))
782 : 11204 : || (GET_CODE (SET_SRC (x)) == ZERO_EXTEND
783 : 0 : && SUBREG_PROMOTED_UNSIGNED_P (y))
784 : 12150 : || !TRULY_NOOP_TRUNCATION_MODES_P (
785 : : GET_MODE (y),
786 : : GET_MODE (SUBREG_REG (y))))))
787 : : break;
788 : :
789 : 1978465 : bit = subreg_lsb (y).to_constant ();
790 : :
791 : : /* If this is a wide object (more bits than we can fit
792 : : in a HOST_WIDE_INT), then just break from the SET
793 : : context. That will cause the iterator to walk down
794 : : into the subrtx and if we land on a REG we'll mark
795 : : the whole think live. */
796 : 1978465 : if (bit >= HOST_BITS_PER_WIDE_INT)
797 : : break;
798 : :
799 : : /* The SUBREG's mode determines the live width. */
800 : 1738231 : if (dst_mask)
801 : : {
802 : 1738231 : dst_mask <<= bit;
803 : 1738231 : if (!dst_mask)
804 : 0 : dst_mask = -HOST_WIDE_INT_UC (0x100000000);
805 : : }
806 : 1738231 : y = SUBREG_REG (y);
807 : : }
808 : :
809 : 93871699 : if (REG_P (y))
810 : : {
811 : : /* We have found the use of a register. We need to mark
812 : : the appropriate chunks of the register live. The mode
813 : : of the REG is a starting point. We may refine that
814 : : based on what chunks in the output were live. */
815 : 49855752 : rn = 4 * REGNO (y);
816 : 49855752 : unsigned HOST_WIDE_INT tmp_mask = dst_mask;
817 : :
818 : : /* If the RTX code for the SET_SRC is not one we can
819 : : propagate destination liveness through, then just
820 : : set the mask to the mode's mask. */
821 : 49855752 : if (!safe_for_live_propagation (code))
822 : 29708 : tmp_mask
823 : 59416 : = GET_MODE_MASK (GET_MODE_INNER (GET_MODE (y)));
824 : :
825 : 49855752 : if (tmp_mask & 0xff)
826 : 49373349 : bitmap_set_bit (livenow, rn);
827 : 49855752 : if (tmp_mask & 0xff00)
828 : 47883899 : bitmap_set_bit (livenow, rn + 1);
829 : 49855752 : if (tmp_mask & HOST_WIDE_INT_UC (0xffff0000))
830 : 47641759 : bitmap_set_bit (livenow, rn + 2);
831 : 49855752 : if (tmp_mask & -HOST_WIDE_INT_UC (0x100000000))
832 : 41221206 : bitmap_set_bit (livenow, rn + 3);
833 : : }
834 : 44015947 : else if (!CONSTANT_P (y))
835 : : break;
836 : :
837 : : /* We might have (ashift (const_int 1) (reg...))
838 : : By setting dst_mask we can continue iterating on the
839 : : the next operand and it will be considered fully live.
840 : :
841 : : Note that since we restore DST_MASK from SAVE_MASK at the
842 : : top of the loop, we have to change SAVE_MASK to get the
843 : : semantics we want. */
844 : 77334397 : if (binop_implies_op2_fully_live (GET_CODE (src)))
845 : 2521407 : save_mask = -1;
846 : :
847 : : /* If this was anything but a binary operand, break the inner
848 : : loop. This is conservatively correct as it will cause the
849 : : iterator to look at the sub-rtxs outside the SET context. */
850 : 77334397 : if (!BINARY_P (src))
851 : : break;
852 : :
853 : : /* We processed the first operand of a binary operator. Now
854 : : handle the second. */
855 : 22032098 : y = XEXP (src, 1), src = pc_rtx;
856 : 22032098 : }
857 : :
858 : : /* These are leaf nodes, no need to iterate down into them. */
859 : 72080781 : if (REG_P (y) || CONSTANT_P (y))
860 : 55302299 : iter.skip_subrtxes ();
861 : : }
862 : : }
863 : : /* If we are reading the low part of a SUBREG, then we can
864 : : refine liveness of the input register, otherwise let the
865 : : iterator continue into SUBREG_REG. */
866 : 476607029 : else if (SUBREG_P (x)
867 : 1334384 : && REG_P (SUBREG_REG (x))
868 : 1332627 : && !paradoxical_subreg_p (x)
869 : 1310995 : && subreg_lowpart_p (x)
870 : 997655 : && GET_MODE_BITSIZE (GET_MODE (x)).is_constant ()
871 : 478602339 : && GET_MODE_BITSIZE (GET_MODE (x)).to_constant () <= 32)
872 : : {
873 : 513638 : HOST_WIDE_INT size = GET_MODE_BITSIZE (GET_MODE (x)).to_constant ();
874 : 513638 : HOST_WIDE_INT rn = 4 * REGNO (SUBREG_REG (x));
875 : :
876 : : /* If this is a promoted subreg, then more of it may be live than
877 : : is otherwise obvious. */
878 : 513638 : if (SUBREG_PROMOTED_VAR_P (x))
879 : 3978 : size = GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x))).to_constant ();
880 : :
881 : 513638 : bitmap_set_bit (livenow, rn);
882 : 513638 : if (size > 8)
883 : 335438 : bitmap_set_bit (livenow, rn + 1);
884 : 335438 : if (size > 16)
885 : 293973 : bitmap_set_bit (livenow, rn + 2);
886 : 293973 : if (size >= 32)
887 : 293973 : bitmap_set_bit (livenow, rn + 3);
888 : 513638 : iter.skip_subrtxes ();
889 : : }
890 : : /* If we have a register reference that is not otherwise handled,
891 : : just assume all the chunks are live. */
892 : 476093391 : else if (REG_P (x))
893 : 158323669 : bitmap_set_range (livenow, REGNO (x) * 4, group_limit (x));
894 : : }
895 : 137513902 : }
896 : :
897 : : /* Process a single basic block BB with current liveness information
898 : : in LIVENOW, returning updated liveness information.
899 : :
900 : : If MODIFY is true, then this is the last pass and unnecessary
901 : : extensions should be eliminated when possible. If an extension
902 : : is removed, the source pseudo is marked in CHANGED_PSEUDOS. */
903 : :
904 : : static void
905 : 22785259 : ext_dce_process_bb (basic_block bb)
906 : : {
907 : 22785259 : rtx_insn *insn;
908 : :
909 : 300891510 : FOR_BB_INSNS_REVERSE (bb, insn)
910 : : {
911 : 428207848 : if (!NONDEBUG_INSN_P (insn))
912 : 150101597 : continue;
913 : :
914 : : /* Live-out state of the destination of this insn. We can
915 : : use this to refine the live-in state of the sources of
916 : : this insn in many cases. */
917 : 128004654 : bitmap live_tmp = BITMAP_ALLOC (NULL);
918 : :
919 : : /* First process any sets/clobbers in INSN. */
920 : 128004654 : bool skipped_dest = ext_dce_process_sets (insn, PATTERN (insn), live_tmp);
921 : :
922 : : /* CALL_INSNs need processing their fusage data. */
923 : 128004654 : if (CALL_P (insn))
924 : 9509248 : skipped_dest |= ext_dce_process_sets (insn,
925 : : CALL_INSN_FUNCTION_USAGE (insn),
926 : : live_tmp);
927 : :
928 : : /* And now uses, optimizing away SIGN/ZERO extensions as we go. */
929 : 128004654 : ext_dce_process_uses (insn, PATTERN (insn), live_tmp, skipped_dest);
930 : :
931 : : /* A nonlocal goto implicitly uses the frame pointer. */
932 : 128004654 : if (JUMP_P (insn) && find_reg_note (insn, REG_NON_LOCAL_GOTO, NULL_RTX))
933 : : {
934 : 1130 : bitmap_set_range (livenow, FRAME_POINTER_REGNUM * 4, 4);
935 : 1130 : if (!HARD_FRAME_POINTER_IS_FRAME_POINTER)
936 : 1130 : bitmap_set_range (livenow, HARD_FRAME_POINTER_REGNUM * 4, 4);
937 : : }
938 : :
939 : : /* And process fusage data for the use as well. */
940 : 128004654 : if (CALL_P (insn))
941 : : {
942 : 9509248 : if (!FAKE_CALL_P (insn))
943 : 9509188 : bitmap_set_range (livenow, STACK_POINTER_REGNUM * 4, 4);
944 : :
945 : : /* If this is not a call to a const fucntion, then assume it
946 : : can read any global register. */
947 : 9509248 : if (!RTL_CONST_CALL_P (insn))
948 : 853864248 : for (unsigned i = 0; i < FIRST_PSEUDO_REGISTER; i++)
949 : 844682912 : if (global_regs[i])
950 : 230 : bitmap_set_range (livenow, i * 4, 4);
951 : :
952 : 9509248 : ext_dce_process_uses (insn, CALL_INSN_FUNCTION_USAGE (insn), live_tmp, false);
953 : : }
954 : :
955 : 128004654 : BITMAP_FREE (live_tmp);
956 : : }
957 : 22785259 : }
958 : :
959 : : /* SUBREG_PROMOTED_VAR_P is set by the gimple->rtl optimizers and
960 : : is usually helpful. However, in some cases setting the value when
961 : : it not strictly needed can cause this pass to miss optimizations.
962 : :
963 : : Specifically consider (set (mem) (subreg (reg))). If set in that
964 : : case it will cause more bit groups to be live for REG than would
965 : : be strictly necessary which in turn can inhibit extension removal.
966 : :
967 : : So do a pass over the IL wiping the SUBREG_PROMOTED_VAR_P when it
968 : : is obviously not needed. */
969 : :
970 : : static void
971 : 947921 : maybe_clear_subreg_promoted_p (void)
972 : : {
973 : 118830534 : for (rtx_insn *insn = get_insns(); insn; insn = NEXT_INSN (insn))
974 : : {
975 : 117882613 : if (!NONDEBUG_INSN_P (insn))
976 : 63092171 : continue;
977 : :
978 : 54790442 : rtx set = single_set (insn);
979 : 54790442 : if (!set)
980 : 3638247 : continue;
981 : :
982 : : /* There may be other cases where we should clear, but for
983 : : now, this is the only known case where it causes problems. */
984 : 51152195 : if (MEM_P (SET_DEST (set)) && SUBREG_P (SET_SRC (set))
985 : 70729 : && GET_MODE (SET_DEST (set)) <= GET_MODE (SUBREG_REG (SET_SRC (set))))
986 : 61859 : SUBREG_PROMOTED_VAR_P (SET_SRC (set)) = 0;
987 : : }
988 : 947921 : }
989 : :
990 : :
991 : : /* We optimize away sign/zero extensions in this pass and replace
992 : : them with SUBREGs indicating certain bits are don't cares.
993 : :
994 : : This changes the SUBREG_PROMOTED_VAR_P state of the object.
995 : : It is fairly painful to fix this on the fly, so we have
996 : : recorded which pseudos are affected and we look for SUBREGs
997 : : of those pseudos and fix them up. */
998 : :
999 : : static void
1000 : 947921 : reset_subreg_promoted_p (void)
1001 : : {
1002 : : /* If we removed an extension, that changed the promoted state
1003 : : of the destination of that extension. Thus we need to go
1004 : : find any SUBREGs that reference that pseudo and adjust their
1005 : : SUBREG_PROMOTED_P state. */
1006 : 118830534 : for (rtx_insn *insn = get_insns(); insn; insn = NEXT_INSN (insn))
1007 : : {
1008 : 117882613 : if (!NONDEBUG_INSN_P (insn))
1009 : 63092171 : continue;
1010 : :
1011 : 54790442 : rtx pat = PATTERN (insn);
1012 : 54790442 : subrtx_var_iterator::array_type array;
1013 : 349185454 : FOR_EACH_SUBRTX_VAR (iter, array, pat, NONCONST)
1014 : : {
1015 : 294395012 : rtx sub = *iter;
1016 : :
1017 : : /* We only care about SUBREGs. */
1018 : 294395012 : if (GET_CODE (sub) != SUBREG)
1019 : 292853152 : continue;
1020 : :
1021 : 1541860 : const_rtx x = SUBREG_REG (sub);
1022 : :
1023 : : /* We only care if the inner object is a REG. */
1024 : 1541860 : if (!REG_P (x))
1025 : 732 : continue;
1026 : :
1027 : : /* And only if the SUBREG is a promoted var. */
1028 : 1541128 : if (!SUBREG_PROMOTED_VAR_P (sub))
1029 : 1535981 : continue;
1030 : :
1031 : 5147 : if (bitmap_bit_p (changed_pseudos, REGNO (x)))
1032 : 0 : SUBREG_PROMOTED_VAR_P (sub) = 0;
1033 : : }
1034 : 54790442 : }
1035 : 947921 : }
1036 : :
1037 : : /* Initialization of the ext-dce pass. Primarily this means
1038 : : setting up the various bitmaps we utilize. */
1039 : :
1040 : : static void
1041 : 947921 : ext_dce_init (void)
1042 : : {
1043 : 947921 : livein.create (last_basic_block_for_fn (cfun));
1044 : 947921 : livein.quick_grow_cleared (last_basic_block_for_fn (cfun));
1045 : 12593493 : for (int i = 0; i < last_basic_block_for_fn (cfun); i++)
1046 : 11645572 : bitmap_initialize (&livein[i], &bitmap_default_obstack);
1047 : :
1048 : 947921 : auto_bitmap refs (&bitmap_default_obstack);
1049 : 947921 : df_get_exit_block_use_set (refs);
1050 : :
1051 : 947921 : unsigned i;
1052 : 947921 : bitmap_iterator bi;
1053 : 4364579 : EXECUTE_IF_SET_IN_BITMAP (refs, 0, i, bi)
1054 : 3416658 : make_reg_live (&livein[EXIT_BLOCK], i);
1055 : :
1056 : 947921 : livenow = BITMAP_ALLOC (NULL);
1057 : 947921 : all_blocks = BITMAP_ALLOC (NULL);
1058 : 947921 : changed_pseudos = BITMAP_ALLOC (NULL);
1059 : :
1060 : 12593493 : for (int i = 0; i < last_basic_block_for_fn (cfun); i++)
1061 : 11645572 : if (i != ENTRY_BLOCK && i != EXIT_BLOCK)
1062 : 9749730 : bitmap_set_bit (all_blocks, i);
1063 : :
1064 : 947921 : modify = false;
1065 : 947921 : }
1066 : :
1067 : : /* Finalization of the ext-dce pass. Primarily this means
1068 : : releasing up the various bitmaps we utilize. */
1069 : :
1070 : : static void
1071 : 947921 : ext_dce_finish (void)
1072 : : {
1073 : 12593493 : for (unsigned i = 0; i < livein.length (); i++)
1074 : 11645572 : bitmap_clear (&livein[i]);
1075 : 947921 : livein.release ();
1076 : :
1077 : 947921 : BITMAP_FREE (livenow);
1078 : 947921 : BITMAP_FREE (changed_pseudos);
1079 : 947921 : BITMAP_FREE (all_blocks);
1080 : 947921 : }
1081 : :
1082 : : /* Process block number BB_INDEX as part of the backward
1083 : : simple dataflow analysis. Return TRUE if something in
1084 : : this block changed or FALSE otherwise. */
1085 : :
1086 : : static bool
1087 : 26576943 : ext_dce_rd_transfer_n (int bb_index)
1088 : : {
1089 : : /* The ENTRY/EXIT blocks never change. */
1090 : 26576943 : if (bb_index == ENTRY_BLOCK || bb_index == EXIT_BLOCK)
1091 : : return false;
1092 : :
1093 : 22785259 : basic_block bb = BASIC_BLOCK_FOR_FN (cfun, bb_index);
1094 : :
1095 : : /* Make everything live that's live in the successors. */
1096 : 22785259 : bitmap_clear (livenow);
1097 : 22785259 : edge_iterator ei;
1098 : 22785259 : edge e;
1099 : :
1100 : 57332212 : FOR_EACH_EDGE (e, ei, bb->succs)
1101 : 34546953 : bitmap_ior_into (livenow, &livein[e->dest->index]);
1102 : :
1103 : 22785259 : ext_dce_process_bb (bb);
1104 : :
1105 : : /* We only allow widening the set of objects live at the start
1106 : : of a block. Otherwise we run the risk of not converging. */
1107 : 22785259 : return bitmap_ior_into (&livein[bb_index], livenow);
1108 : : }
1109 : :
1110 : : /* Dummy function for the df_simple_dataflow API. */
1111 : 33164354 : static bool ext_dce_rd_confluence_n (edge) { return true; }
1112 : :
1113 : : /* Use lifetime analyis to identify extensions that set bits that
1114 : : are never read. Turn such extensions into SUBREGs instead which
1115 : : can often be propagated away. */
1116 : :
1117 : : void
1118 : 947921 : ext_dce_execute (void)
1119 : : {
1120 : : /* Limit the amount of memory we use for livein, with 4 bits per
1121 : : reg per basic-block including overhead that maps to one byte
1122 : : per reg per basic-block. */
1123 : 947921 : uint64_t memory_request
1124 : 947921 : = (uint64_t)n_basic_blocks_for_fn (cfun) * max_reg_num ();
1125 : 947921 : if (memory_request / 1024 > (uint64_t)param_max_gcse_memory)
1126 : : {
1127 : 0 : warning (OPT_Wdisabled_optimization,
1128 : : "ext-dce disabled: %d basic blocks and %d registers; "
1129 : : "increase %<--param max-gcse-memory%> above %wu",
1130 : 0 : n_basic_blocks_for_fn (cfun), max_reg_num (),
1131 : : memory_request / 1024);
1132 : 0 : return;
1133 : : }
1134 : :
1135 : : /* Some settings of SUBREG_PROMOTED_VAR_P are actively harmful
1136 : : to this pass. Clear it for those cases. */
1137 : 947921 : maybe_clear_subreg_promoted_p ();
1138 : 947921 : df_analyze ();
1139 : 947921 : ext_dce_init ();
1140 : :
1141 : 3791684 : do
1142 : : {
1143 : 1895842 : df_simple_dataflow (DF_BACKWARD, NULL, NULL,
1144 : : ext_dce_rd_confluence_n, ext_dce_rd_transfer_n,
1145 : : all_blocks, df_get_postorder (DF_BACKWARD),
1146 : : df_get_n_blocks (DF_BACKWARD));
1147 : 1895842 : modify = !modify;
1148 : : }
1149 : : while (modify);
1150 : :
1151 : 947921 : reset_subreg_promoted_p ();
1152 : :
1153 : 947921 : ext_dce_finish ();
1154 : : }
1155 : :
1156 : :
1157 : : namespace {
1158 : :
1159 : : const pass_data pass_data_ext_dce =
1160 : : {
1161 : : RTL_PASS, /* type */
1162 : : "ext_dce", /* name */
1163 : : OPTGROUP_NONE, /* optinfo_flags */
1164 : : TV_EXT_DCE, /* tv_id */
1165 : : PROP_cfglayout, /* properties_required */
1166 : : 0, /* properties_provided */
1167 : : 0, /* properties_destroyed */
1168 : : 0, /* todo_flags_start */
1169 : : TODO_df_finish, /* todo_flags_finish */
1170 : : };
1171 : :
1172 : : class pass_ext_dce : public rtl_opt_pass
1173 : : {
1174 : : public:
1175 : 285081 : pass_ext_dce (gcc::context *ctxt)
1176 : 570162 : : rtl_opt_pass (pass_data_ext_dce, ctxt)
1177 : : {}
1178 : :
1179 : : /* opt_pass methods: */
1180 : 1449863 : virtual bool gate (function *) { return flag_ext_dce && optimize > 0; }
1181 : 947921 : virtual unsigned int execute (function *)
1182 : : {
1183 : 947921 : ext_dce_execute ();
1184 : 947921 : return 0;
1185 : : }
1186 : :
1187 : : }; // class pass_combine
1188 : :
1189 : : } // anon namespace
1190 : :
1191 : : rtl_opt_pass *
1192 : 285081 : make_pass_ext_dce (gcc::context *ctxt)
1193 : : {
1194 : 285081 : return new pass_ext_dce (ctxt);
1195 : : }
|
