]>
Commit | Line | Data |
---|---|---|
e53a16e7 | 1 | /* Decompose multiword subregs. |
d5785e76 JJ |
2 | Copyright (C) 2007, 2008, 2009, 2010, 2011, 2012 |
3 | Free Software Foundation, Inc. | |
e53a16e7 ILT |
4 | Contributed by Richard Henderson <rth@redhat.com> |
5 | Ian Lance Taylor <iant@google.com> | |
6 | ||
7 | This file is part of GCC. | |
8 | ||
9 | GCC is free software; you can redistribute it and/or modify it under | |
10 | the terms of the GNU General Public License as published by the Free | |
9dcd6f09 | 11 | Software Foundation; either version 3, or (at your option) any later |
e53a16e7 ILT |
12 | version. |
13 | ||
14 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
15 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
16 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
17 | for more details. | |
18 | ||
19 | You should have received a copy of the GNU General Public License | |
9dcd6f09 NC |
20 | along with GCC; see the file COPYING3. If not see |
21 | <http://www.gnu.org/licenses/>. */ | |
e53a16e7 ILT |
22 | |
23 | #include "config.h" | |
24 | #include "system.h" | |
25 | #include "coretypes.h" | |
26 | #include "machmode.h" | |
27 | #include "tm.h" | |
28 | #include "rtl.h" | |
29 | #include "tm_p.h" | |
e53a16e7 ILT |
30 | #include "flags.h" |
31 | #include "insn-config.h" | |
32 | #include "obstack.h" | |
33 | #include "basic-block.h" | |
34 | #include "recog.h" | |
35 | #include "bitmap.h" | |
8d074192 | 36 | #include "dce.h" |
e53a16e7 | 37 | #include "expr.h" |
7984c787 | 38 | #include "except.h" |
e53a16e7 ILT |
39 | #include "regs.h" |
40 | #include "tree-pass.h" | |
6fb5fa3c | 41 | #include "df.h" |
af4ba423 | 42 | #include "lower-subreg.h" |
e53a16e7 ILT |
43 | |
44 | #ifdef STACK_GROWS_DOWNWARD | |
45 | # undef STACK_GROWS_DOWNWARD | |
46 | # define STACK_GROWS_DOWNWARD 1 | |
47 | #else | |
48 | # define STACK_GROWS_DOWNWARD 0 | |
49 | #endif | |
50 | ||
51 | DEF_VEC_P (bitmap); | |
52 | DEF_VEC_ALLOC_P (bitmap,heap); | |
53 | ||
54 | /* Decompose multi-word pseudo-registers into individual | |
af4ba423 KZ |
55 | pseudo-registers when possible and profitable. This is possible |
56 | when all the uses of a multi-word register are via SUBREG, or are | |
57 | copies of the register to another location. Breaking apart the | |
58 | register permits more CSE and permits better register allocation. | |
59 | This is profitable if the machine does not have move instructions | |
60 | to do this. | |
61 | ||
62 | This pass only splits moves with modes that are wider than | |
63 | word_mode and ASHIFTs, LSHIFTRTs and ZERO_EXTENDs with integer | |
64 | modes that are twice the width of word_mode. The latter could be | |
65 | generalized if there was a need to do this, but the trend in | |
66 | architectures is to not need this. | |
67 | ||
68 | There are two useful preprocessor defines for use by maintainers: | |
69 | ||
70 | #define LOG_COSTS 1 | |
71 | ||
72 | if you wish to see the actual cost estimates that are being used | |
73 | for each mode wider than word mode and the cost estimates for zero | |
74 | extension and the shifts. This can be useful when port maintainers | |
75 | are tuning insn rtx costs. | |
76 | ||
77 | #define FORCE_LOWERING 1 | |
78 | ||
79 | if you wish to test the pass with all the transformation forced on. | |
80 | This can be useful for finding bugs in the transformations. */ | |
81 | ||
82 | #define LOG_COSTS 0 | |
83 | #define FORCE_LOWERING 0 | |
e53a16e7 ILT |
84 | |
85 | /* Bit N in this bitmap is set if regno N is used in a context in | |
86 | which we can decompose it. */ | |
87 | static bitmap decomposable_context; | |
88 | ||
89 | /* Bit N in this bitmap is set if regno N is used in a context in | |
90 | which it can not be decomposed. */ | |
91 | static bitmap non_decomposable_context; | |
92 | ||
402464a0 BS |
93 | /* Bit N in this bitmap is set if regno N is used in a subreg |
94 | which changes the mode but not the size. This typically happens | |
95 | when the register accessed as a floating-point value; we want to | |
96 | avoid generating accesses to its subwords in integer modes. */ | |
97 | static bitmap subreg_context; | |
98 | ||
e53a16e7 ILT |
99 | /* Bit N in the bitmap in element M of this array is set if there is a |
100 | copy from reg M to reg N. */ | |
101 | static VEC(bitmap,heap) *reg_copy_graph; | |
102 | ||
af4ba423 KZ |
103 | struct target_lower_subreg default_target_lower_subreg; |
104 | #if SWITCHABLE_TARGET | |
105 | struct target_lower_subreg *this_target_lower_subreg | |
106 | = &default_target_lower_subreg; | |
107 | #endif | |
108 | ||
109 | #define twice_word_mode \ | |
110 | this_target_lower_subreg->x_twice_word_mode | |
111 | #define choices \ | |
112 | this_target_lower_subreg->x_choices | |
113 | ||
114 | /* RTXes used while computing costs. */ | |
115 | struct cost_rtxes { | |
116 | /* Source and target registers. */ | |
117 | rtx source; | |
118 | rtx target; | |
119 | ||
120 | /* A twice_word_mode ZERO_EXTEND of SOURCE. */ | |
121 | rtx zext; | |
122 | ||
123 | /* A shift of SOURCE. */ | |
124 | rtx shift; | |
125 | ||
126 | /* A SET of TARGET. */ | |
127 | rtx set; | |
128 | }; | |
129 | ||
130 | /* Return the cost of a CODE shift in mode MODE by OP1 bits, using the | |
131 | rtxes in RTXES. SPEED_P selects between the speed and size cost. */ | |
132 | ||
133 | static int | |
134 | shift_cost (bool speed_p, struct cost_rtxes *rtxes, enum rtx_code code, | |
135 | enum machine_mode mode, int op1) | |
136 | { | |
af4ba423 KZ |
137 | PUT_CODE (rtxes->shift, code); |
138 | PUT_MODE (rtxes->shift, mode); | |
139 | PUT_MODE (rtxes->source, mode); | |
140 | XEXP (rtxes->shift, 1) = GEN_INT (op1); | |
69523a7c | 141 | return set_src_cost (rtxes->shift, speed_p); |
af4ba423 KZ |
142 | } |
143 | ||
144 | /* For each X in the range [0, BITS_PER_WORD), set SPLITTING[X] | |
145 | to true if it is profitable to split a double-word CODE shift | |
146 | of X + BITS_PER_WORD bits. SPEED_P says whether we are testing | |
147 | for speed or size profitability. | |
148 | ||
149 | Use the rtxes in RTXES to calculate costs. WORD_MOVE_ZERO_COST is | |
150 | the cost of moving zero into a word-mode register. WORD_MOVE_COST | |
151 | is the cost of moving between word registers. */ | |
152 | ||
153 | static void | |
154 | compute_splitting_shift (bool speed_p, struct cost_rtxes *rtxes, | |
155 | bool *splitting, enum rtx_code code, | |
156 | int word_move_zero_cost, int word_move_cost) | |
157 | { | |
158 | int wide_cost, narrow_cost, i; | |
159 | ||
160 | for (i = 0; i < BITS_PER_WORD; i++) | |
161 | { | |
162 | wide_cost = shift_cost (speed_p, rtxes, code, twice_word_mode, | |
163 | i + BITS_PER_WORD); | |
164 | if (i == 0) | |
165 | narrow_cost = word_move_cost; | |
166 | else | |
167 | narrow_cost = shift_cost (speed_p, rtxes, code, word_mode, i); | |
168 | ||
169 | if (LOG_COSTS) | |
170 | fprintf (stderr, "%s %s by %d: original cost %d, split cost %d + %d\n", | |
171 | GET_MODE_NAME (twice_word_mode), GET_RTX_NAME (code), | |
172 | i + BITS_PER_WORD, wide_cost, narrow_cost, | |
173 | word_move_zero_cost); | |
174 | ||
175 | if (FORCE_LOWERING || wide_cost >= narrow_cost + word_move_zero_cost) | |
176 | splitting[i] = true; | |
177 | } | |
178 | } | |
179 | ||
180 | /* Compute what we should do when optimizing for speed or size; SPEED_P | |
181 | selects which. Use RTXES for computing costs. */ | |
182 | ||
183 | static void | |
184 | compute_costs (bool speed_p, struct cost_rtxes *rtxes) | |
185 | { | |
186 | unsigned int i; | |
187 | int word_move_zero_cost, word_move_cost; | |
188 | ||
69523a7c | 189 | PUT_MODE (rtxes->target, word_mode); |
af4ba423 | 190 | SET_SRC (rtxes->set) = CONST0_RTX (word_mode); |
69523a7c | 191 | word_move_zero_cost = set_rtx_cost (rtxes->set, speed_p); |
af4ba423 KZ |
192 | |
193 | SET_SRC (rtxes->set) = rtxes->source; | |
69523a7c | 194 | word_move_cost = set_rtx_cost (rtxes->set, speed_p); |
af4ba423 KZ |
195 | |
196 | if (LOG_COSTS) | |
197 | fprintf (stderr, "%s move: from zero cost %d, from reg cost %d\n", | |
198 | GET_MODE_NAME (word_mode), word_move_zero_cost, word_move_cost); | |
199 | ||
200 | for (i = 0; i < MAX_MACHINE_MODE; i++) | |
201 | { | |
202 | enum machine_mode mode = (enum machine_mode) i; | |
203 | int factor = GET_MODE_SIZE (mode) / UNITS_PER_WORD; | |
204 | if (factor > 1) | |
205 | { | |
206 | int mode_move_cost; | |
207 | ||
208 | PUT_MODE (rtxes->target, mode); | |
209 | PUT_MODE (rtxes->source, mode); | |
69523a7c | 210 | mode_move_cost = set_rtx_cost (rtxes->set, speed_p); |
af4ba423 KZ |
211 | |
212 | if (LOG_COSTS) | |
213 | fprintf (stderr, "%s move: original cost %d, split cost %d * %d\n", | |
214 | GET_MODE_NAME (mode), mode_move_cost, | |
215 | word_move_cost, factor); | |
216 | ||
217 | if (FORCE_LOWERING || mode_move_cost >= word_move_cost * factor) | |
218 | { | |
219 | choices[speed_p].move_modes_to_split[i] = true; | |
220 | choices[speed_p].something_to_do = true; | |
221 | } | |
222 | } | |
223 | } | |
224 | ||
225 | /* For the moves and shifts, the only case that is checked is one | |
226 | where the mode of the target is an integer mode twice the width | |
227 | of the word_mode. | |
228 | ||
229 | If it is not profitable to split a double word move then do not | |
230 | even consider the shifts or the zero extension. */ | |
231 | if (choices[speed_p].move_modes_to_split[(int) twice_word_mode]) | |
232 | { | |
233 | int zext_cost; | |
234 | ||
235 | /* The only case here to check to see if moving the upper part with a | |
236 | zero is cheaper than doing the zext itself. */ | |
af4ba423 | 237 | PUT_MODE (rtxes->source, word_mode); |
69523a7c | 238 | zext_cost = set_src_cost (rtxes->zext, speed_p); |
af4ba423 KZ |
239 | |
240 | if (LOG_COSTS) | |
241 | fprintf (stderr, "%s %s: original cost %d, split cost %d + %d\n", | |
242 | GET_MODE_NAME (twice_word_mode), GET_RTX_NAME (ZERO_EXTEND), | |
243 | zext_cost, word_move_cost, word_move_zero_cost); | |
244 | ||
245 | if (FORCE_LOWERING || zext_cost >= word_move_cost + word_move_zero_cost) | |
246 | choices[speed_p].splitting_zext = true; | |
247 | ||
248 | compute_splitting_shift (speed_p, rtxes, | |
249 | choices[speed_p].splitting_ashift, ASHIFT, | |
250 | word_move_zero_cost, word_move_cost); | |
251 | compute_splitting_shift (speed_p, rtxes, | |
252 | choices[speed_p].splitting_lshiftrt, LSHIFTRT, | |
253 | word_move_zero_cost, word_move_cost); | |
254 | } | |
255 | } | |
256 | ||
257 | /* Do one-per-target initialisation. This involves determining | |
258 | which operations on the machine are profitable. If none are found, | |
259 | then the pass just returns when called. */ | |
260 | ||
261 | void | |
262 | init_lower_subreg (void) | |
263 | { | |
264 | struct cost_rtxes rtxes; | |
265 | ||
266 | memset (this_target_lower_subreg, 0, sizeof (*this_target_lower_subreg)); | |
267 | ||
268 | twice_word_mode = GET_MODE_2XWIDER_MODE (word_mode); | |
269 | ||
270 | rtxes.target = gen_rtx_REG (word_mode, FIRST_PSEUDO_REGISTER); | |
271 | rtxes.source = gen_rtx_REG (word_mode, FIRST_PSEUDO_REGISTER + 1); | |
272 | rtxes.set = gen_rtx_SET (VOIDmode, rtxes.target, rtxes.source); | |
273 | rtxes.zext = gen_rtx_ZERO_EXTEND (twice_word_mode, rtxes.source); | |
274 | rtxes.shift = gen_rtx_ASHIFT (twice_word_mode, rtxes.source, const0_rtx); | |
275 | ||
276 | if (LOG_COSTS) | |
277 | fprintf (stderr, "\nSize costs\n==========\n\n"); | |
278 | compute_costs (false, &rtxes); | |
279 | ||
280 | if (LOG_COSTS) | |
281 | fprintf (stderr, "\nSpeed costs\n===========\n\n"); | |
282 | compute_costs (true, &rtxes); | |
283 | } | |
2b54c30f ILT |
284 | |
285 | static bool | |
286 | simple_move_operand (rtx x) | |
287 | { | |
288 | if (GET_CODE (x) == SUBREG) | |
289 | x = SUBREG_REG (x); | |
290 | ||
291 | if (!OBJECT_P (x)) | |
292 | return false; | |
293 | ||
294 | if (GET_CODE (x) == LABEL_REF | |
295 | || GET_CODE (x) == SYMBOL_REF | |
7e0c3f57 ILT |
296 | || GET_CODE (x) == HIGH |
297 | || GET_CODE (x) == CONST) | |
2b54c30f ILT |
298 | return false; |
299 | ||
300 | if (MEM_P (x) | |
301 | && (MEM_VOLATILE_P (x) | |
5bfed9a9 | 302 | || mode_dependent_address_p (XEXP (x, 0), MEM_ADDR_SPACE (x)))) |
2b54c30f ILT |
303 | return false; |
304 | ||
305 | return true; | |
306 | } | |
307 | ||
af4ba423 KZ |
308 | /* If INSN is a single set between two objects that we want to split, |
309 | return the single set. SPEED_P says whether we are optimizing | |
310 | INSN for speed or size. | |
311 | ||
312 | INSN should have been passed to recog and extract_insn before this | |
313 | is called. */ | |
e53a16e7 ILT |
314 | |
315 | static rtx | |
af4ba423 | 316 | simple_move (rtx insn, bool speed_p) |
e53a16e7 ILT |
317 | { |
318 | rtx x; | |
319 | rtx set; | |
320 | enum machine_mode mode; | |
321 | ||
322 | if (recog_data.n_operands != 2) | |
323 | return NULL_RTX; | |
324 | ||
325 | set = single_set (insn); | |
326 | if (!set) | |
327 | return NULL_RTX; | |
328 | ||
329 | x = SET_DEST (set); | |
330 | if (x != recog_data.operand[0] && x != recog_data.operand[1]) | |
331 | return NULL_RTX; | |
2b54c30f | 332 | if (!simple_move_operand (x)) |
e53a16e7 ILT |
333 | return NULL_RTX; |
334 | ||
335 | x = SET_SRC (set); | |
336 | if (x != recog_data.operand[0] && x != recog_data.operand[1]) | |
337 | return NULL_RTX; | |
2b54c30f ILT |
338 | /* For the src we can handle ASM_OPERANDS, and it is beneficial for |
339 | things like x86 rdtsc which returns a DImode value. */ | |
340 | if (GET_CODE (x) != ASM_OPERANDS | |
341 | && !simple_move_operand (x)) | |
e53a16e7 ILT |
342 | return NULL_RTX; |
343 | ||
344 | /* We try to decompose in integer modes, to avoid generating | |
345 | inefficient code copying between integer and floating point | |
346 | registers. That means that we can't decompose if this is a | |
347 | non-integer mode for which there is no integer mode of the same | |
348 | size. */ | |
349 | mode = GET_MODE (SET_SRC (set)); | |
350 | if (!SCALAR_INT_MODE_P (mode) | |
351 | && (mode_for_size (GET_MODE_SIZE (mode) * BITS_PER_UNIT, MODE_INT, 0) | |
352 | == BLKmode)) | |
353 | return NULL_RTX; | |
354 | ||
1f64a081 ILT |
355 | /* Reject PARTIAL_INT modes. They are used for processor specific |
356 | purposes and it's probably best not to tamper with them. */ | |
357 | if (GET_MODE_CLASS (mode) == MODE_PARTIAL_INT) | |
358 | return NULL_RTX; | |
359 | ||
af4ba423 KZ |
360 | if (!choices[speed_p].move_modes_to_split[(int) mode]) |
361 | return NULL_RTX; | |
362 | ||
e53a16e7 ILT |
363 | return set; |
364 | } | |
365 | ||
366 | /* If SET is a copy from one multi-word pseudo-register to another, | |
367 | record that in reg_copy_graph. Return whether it is such a | |
368 | copy. */ | |
369 | ||
370 | static bool | |
371 | find_pseudo_copy (rtx set) | |
372 | { | |
373 | rtx dest = SET_DEST (set); | |
374 | rtx src = SET_SRC (set); | |
375 | unsigned int rd, rs; | |
376 | bitmap b; | |
377 | ||
378 | if (!REG_P (dest) || !REG_P (src)) | |
379 | return false; | |
380 | ||
381 | rd = REGNO (dest); | |
382 | rs = REGNO (src); | |
383 | if (HARD_REGISTER_NUM_P (rd) || HARD_REGISTER_NUM_P (rs)) | |
384 | return false; | |
385 | ||
e53a16e7 ILT |
386 | b = VEC_index (bitmap, reg_copy_graph, rs); |
387 | if (b == NULL) | |
388 | { | |
389 | b = BITMAP_ALLOC (NULL); | |
390 | VEC_replace (bitmap, reg_copy_graph, rs, b); | |
391 | } | |
392 | ||
393 | bitmap_set_bit (b, rd); | |
394 | ||
395 | return true; | |
396 | } | |
397 | ||
398 | /* Look through the registers in DECOMPOSABLE_CONTEXT. For each case | |
399 | where they are copied to another register, add the register to | |
400 | which they are copied to DECOMPOSABLE_CONTEXT. Use | |
401 | NON_DECOMPOSABLE_CONTEXT to limit this--we don't bother to track | |
402 | copies of registers which are in NON_DECOMPOSABLE_CONTEXT. */ | |
403 | ||
404 | static void | |
405 | propagate_pseudo_copies (void) | |
406 | { | |
407 | bitmap queue, propagate; | |
408 | ||
409 | queue = BITMAP_ALLOC (NULL); | |
410 | propagate = BITMAP_ALLOC (NULL); | |
411 | ||
412 | bitmap_copy (queue, decomposable_context); | |
413 | do | |
414 | { | |
415 | bitmap_iterator iter; | |
416 | unsigned int i; | |
417 | ||
418 | bitmap_clear (propagate); | |
419 | ||
420 | EXECUTE_IF_SET_IN_BITMAP (queue, 0, i, iter) | |
421 | { | |
422 | bitmap b = VEC_index (bitmap, reg_copy_graph, i); | |
423 | if (b) | |
424 | bitmap_ior_and_compl_into (propagate, b, non_decomposable_context); | |
425 | } | |
426 | ||
427 | bitmap_and_compl (queue, propagate, decomposable_context); | |
428 | bitmap_ior_into (decomposable_context, propagate); | |
429 | } | |
430 | while (!bitmap_empty_p (queue)); | |
431 | ||
432 | BITMAP_FREE (queue); | |
433 | BITMAP_FREE (propagate); | |
434 | } | |
435 | ||
436 | /* A pointer to one of these values is passed to | |
437 | find_decomposable_subregs via for_each_rtx. */ | |
438 | ||
439 | enum classify_move_insn | |
440 | { | |
441 | /* Not a simple move from one location to another. */ | |
442 | NOT_SIMPLE_MOVE, | |
c2c47e8f UW |
443 | /* A simple move we want to decompose. */ |
444 | DECOMPOSABLE_SIMPLE_MOVE, | |
445 | /* Any other simple move. */ | |
e53a16e7 ILT |
446 | SIMPLE_MOVE |
447 | }; | |
448 | ||
449 | /* This is called via for_each_rtx. If we find a SUBREG which we | |
450 | could use to decompose a pseudo-register, set a bit in | |
451 | DECOMPOSABLE_CONTEXT. If we find an unadorned register which is | |
452 | not a simple pseudo-register copy, DATA will point at the type of | |
453 | move, and we set a bit in DECOMPOSABLE_CONTEXT or | |
454 | NON_DECOMPOSABLE_CONTEXT as appropriate. */ | |
455 | ||
456 | static int | |
457 | find_decomposable_subregs (rtx *px, void *data) | |
458 | { | |
459 | enum classify_move_insn *pcmi = (enum classify_move_insn *) data; | |
460 | rtx x = *px; | |
461 | ||
462 | if (x == NULL_RTX) | |
463 | return 0; | |
464 | ||
465 | if (GET_CODE (x) == SUBREG) | |
466 | { | |
467 | rtx inner = SUBREG_REG (x); | |
468 | unsigned int regno, outer_size, inner_size, outer_words, inner_words; | |
469 | ||
470 | if (!REG_P (inner)) | |
471 | return 0; | |
472 | ||
473 | regno = REGNO (inner); | |
474 | if (HARD_REGISTER_NUM_P (regno)) | |
475 | return -1; | |
476 | ||
477 | outer_size = GET_MODE_SIZE (GET_MODE (x)); | |
478 | inner_size = GET_MODE_SIZE (GET_MODE (inner)); | |
479 | outer_words = (outer_size + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
480 | inner_words = (inner_size + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
481 | ||
482 | /* We only try to decompose single word subregs of multi-word | |
483 | registers. When we find one, we return -1 to avoid iterating | |
484 | over the inner register. | |
485 | ||
486 | ??? This doesn't allow, e.g., DImode subregs of TImode values | |
487 | on 32-bit targets. We would need to record the way the | |
488 | pseudo-register was used, and only decompose if all the uses | |
489 | were the same number and size of pieces. Hopefully this | |
490 | doesn't happen much. */ | |
491 | ||
492 | if (outer_words == 1 && inner_words > 1) | |
493 | { | |
494 | bitmap_set_bit (decomposable_context, regno); | |
495 | return -1; | |
496 | } | |
03743286 ILT |
497 | |
498 | /* If this is a cast from one mode to another, where the modes | |
499 | have the same size, and they are not tieable, then mark this | |
500 | register as non-decomposable. If we decompose it we are | |
501 | likely to mess up whatever the backend is trying to do. */ | |
502 | if (outer_words > 1 | |
503 | && outer_size == inner_size | |
504 | && !MODES_TIEABLE_P (GET_MODE (x), GET_MODE (inner))) | |
505 | { | |
506 | bitmap_set_bit (non_decomposable_context, regno); | |
402464a0 | 507 | bitmap_set_bit (subreg_context, regno); |
03743286 ILT |
508 | return -1; |
509 | } | |
e53a16e7 | 510 | } |
2b54c30f | 511 | else if (REG_P (x)) |
e53a16e7 ILT |
512 | { |
513 | unsigned int regno; | |
514 | ||
515 | /* We will see an outer SUBREG before we see the inner REG, so | |
516 | when we see a plain REG here it means a direct reference to | |
517 | the register. | |
518 | ||
519 | If this is not a simple copy from one location to another, | |
520 | then we can not decompose this register. If this is a simple | |
c2c47e8f | 521 | copy we want to decompose, and the mode is right, |
4a8cae83 SB |
522 | then we mark the register as decomposable. |
523 | Otherwise we don't say anything about this register -- | |
524 | it could be decomposed, but whether that would be | |
e53a16e7 ILT |
525 | profitable depends upon how it is used elsewhere. |
526 | ||
527 | We only set bits in the bitmap for multi-word | |
528 | pseudo-registers, since those are the only ones we care about | |
529 | and it keeps the size of the bitmaps down. */ | |
530 | ||
531 | regno = REGNO (x); | |
532 | if (!HARD_REGISTER_NUM_P (regno) | |
533 | && GET_MODE_SIZE (GET_MODE (x)) > UNITS_PER_WORD) | |
534 | { | |
535 | switch (*pcmi) | |
536 | { | |
537 | case NOT_SIMPLE_MOVE: | |
538 | bitmap_set_bit (non_decomposable_context, regno); | |
539 | break; | |
c2c47e8f | 540 | case DECOMPOSABLE_SIMPLE_MOVE: |
e53a16e7 ILT |
541 | if (MODES_TIEABLE_P (GET_MODE (x), word_mode)) |
542 | bitmap_set_bit (decomposable_context, regno); | |
543 | break; | |
544 | case SIMPLE_MOVE: | |
545 | break; | |
546 | default: | |
547 | gcc_unreachable (); | |
548 | } | |
549 | } | |
550 | } | |
2b54c30f ILT |
551 | else if (MEM_P (x)) |
552 | { | |
553 | enum classify_move_insn cmi_mem = NOT_SIMPLE_MOVE; | |
554 | ||
555 | /* Any registers used in a MEM do not participate in a | |
c2c47e8f | 556 | SIMPLE_MOVE or DECOMPOSABLE_SIMPLE_MOVE. Do our own recursion |
2b54c30f ILT |
557 | here, and return -1 to block the parent's recursion. */ |
558 | for_each_rtx (&XEXP (x, 0), find_decomposable_subregs, &cmi_mem); | |
559 | return -1; | |
560 | } | |
e53a16e7 ILT |
561 | |
562 | return 0; | |
563 | } | |
564 | ||
565 | /* Decompose REGNO into word-sized components. We smash the REG node | |
566 | in place. This ensures that (1) something goes wrong quickly if we | |
567 | fail to make some replacement, and (2) the debug information inside | |
568 | the symbol table is automatically kept up to date. */ | |
569 | ||
570 | static void | |
571 | decompose_register (unsigned int regno) | |
572 | { | |
573 | rtx reg; | |
574 | unsigned int words, i; | |
575 | rtvec v; | |
576 | ||
577 | reg = regno_reg_rtx[regno]; | |
578 | ||
579 | regno_reg_rtx[regno] = NULL_RTX; | |
e53a16e7 ILT |
580 | |
581 | words = GET_MODE_SIZE (GET_MODE (reg)); | |
582 | words = (words + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
583 | ||
584 | v = rtvec_alloc (words); | |
585 | for (i = 0; i < words; ++i) | |
586 | RTVEC_ELT (v, i) = gen_reg_rtx_offset (reg, word_mode, i * UNITS_PER_WORD); | |
587 | ||
588 | PUT_CODE (reg, CONCATN); | |
589 | XVEC (reg, 0) = v; | |
590 | ||
591 | if (dump_file) | |
592 | { | |
593 | fprintf (dump_file, "; Splitting reg %u ->", regno); | |
594 | for (i = 0; i < words; ++i) | |
595 | fprintf (dump_file, " %u", REGNO (XVECEXP (reg, 0, i))); | |
596 | fputc ('\n', dump_file); | |
597 | } | |
598 | } | |
599 | ||
600 | /* Get a SUBREG of a CONCATN. */ | |
601 | ||
602 | static rtx | |
603 | simplify_subreg_concatn (enum machine_mode outermode, rtx op, | |
604 | unsigned int byte) | |
605 | { | |
606 | unsigned int inner_size; | |
695ae295 | 607 | enum machine_mode innermode, partmode; |
e53a16e7 ILT |
608 | rtx part; |
609 | unsigned int final_offset; | |
610 | ||
611 | gcc_assert (GET_CODE (op) == CONCATN); | |
612 | gcc_assert (byte % GET_MODE_SIZE (outermode) == 0); | |
613 | ||
614 | innermode = GET_MODE (op); | |
615 | gcc_assert (byte < GET_MODE_SIZE (innermode)); | |
616 | gcc_assert (GET_MODE_SIZE (outermode) <= GET_MODE_SIZE (innermode)); | |
617 | ||
618 | inner_size = GET_MODE_SIZE (innermode) / XVECLEN (op, 0); | |
619 | part = XVECEXP (op, 0, byte / inner_size); | |
695ae295 UB |
620 | partmode = GET_MODE (part); |
621 | ||
822a55a0 UB |
622 | /* VECTOR_CSTs in debug expressions are expanded into CONCATN instead of |
623 | regular CONST_VECTORs. They have vector or integer modes, depending | |
624 | on the capabilities of the target. Cope with them. */ | |
625 | if (partmode == VOIDmode && VECTOR_MODE_P (innermode)) | |
626 | partmode = GET_MODE_INNER (innermode); | |
627 | else if (partmode == VOIDmode) | |
695ae295 | 628 | { |
822a55a0 UB |
629 | enum mode_class mclass = GET_MODE_CLASS (innermode); |
630 | partmode = mode_for_size (inner_size * BITS_PER_UNIT, mclass, 0); | |
695ae295 UB |
631 | } |
632 | ||
e53a16e7 ILT |
633 | final_offset = byte % inner_size; |
634 | if (final_offset + GET_MODE_SIZE (outermode) > inner_size) | |
635 | return NULL_RTX; | |
636 | ||
695ae295 | 637 | return simplify_gen_subreg (outermode, part, partmode, final_offset); |
e53a16e7 ILT |
638 | } |
639 | ||
640 | /* Wrapper around simplify_gen_subreg which handles CONCATN. */ | |
641 | ||
642 | static rtx | |
643 | simplify_gen_subreg_concatn (enum machine_mode outermode, rtx op, | |
644 | enum machine_mode innermode, unsigned int byte) | |
645 | { | |
0e6c5b58 ILT |
646 | rtx ret; |
647 | ||
e53a16e7 ILT |
648 | /* We have to handle generating a SUBREG of a SUBREG of a CONCATN. |
649 | If OP is a SUBREG of a CONCATN, then it must be a simple mode | |
650 | change with the same size and offset 0, or it must extract a | |
651 | part. We shouldn't see anything else here. */ | |
652 | if (GET_CODE (op) == SUBREG && GET_CODE (SUBREG_REG (op)) == CONCATN) | |
653 | { | |
654 | rtx op2; | |
655 | ||
656 | if ((GET_MODE_SIZE (GET_MODE (op)) | |
657 | == GET_MODE_SIZE (GET_MODE (SUBREG_REG (op)))) | |
658 | && SUBREG_BYTE (op) == 0) | |
659 | return simplify_gen_subreg_concatn (outermode, SUBREG_REG (op), | |
660 | GET_MODE (SUBREG_REG (op)), byte); | |
661 | ||
662 | op2 = simplify_subreg_concatn (GET_MODE (op), SUBREG_REG (op), | |
663 | SUBREG_BYTE (op)); | |
664 | if (op2 == NULL_RTX) | |
665 | { | |
666 | /* We don't handle paradoxical subregs here. */ | |
667 | gcc_assert (GET_MODE_SIZE (outermode) | |
668 | <= GET_MODE_SIZE (GET_MODE (op))); | |
669 | gcc_assert (GET_MODE_SIZE (GET_MODE (op)) | |
670 | <= GET_MODE_SIZE (GET_MODE (SUBREG_REG (op)))); | |
671 | op2 = simplify_subreg_concatn (outermode, SUBREG_REG (op), | |
672 | byte + SUBREG_BYTE (op)); | |
673 | gcc_assert (op2 != NULL_RTX); | |
674 | return op2; | |
675 | } | |
676 | ||
677 | op = op2; | |
678 | gcc_assert (op != NULL_RTX); | |
679 | gcc_assert (innermode == GET_MODE (op)); | |
680 | } | |
0e6c5b58 | 681 | |
e53a16e7 ILT |
682 | if (GET_CODE (op) == CONCATN) |
683 | return simplify_subreg_concatn (outermode, op, byte); | |
0e6c5b58 ILT |
684 | |
685 | ret = simplify_gen_subreg (outermode, op, innermode, byte); | |
686 | ||
687 | /* If we see an insn like (set (reg:DI) (subreg:DI (reg:SI) 0)) then | |
688 | resolve_simple_move will ask for the high part of the paradoxical | |
689 | subreg, which does not have a value. Just return a zero. */ | |
690 | if (ret == NULL_RTX | |
691 | && GET_CODE (op) == SUBREG | |
692 | && SUBREG_BYTE (op) == 0 | |
693 | && (GET_MODE_SIZE (innermode) | |
694 | > GET_MODE_SIZE (GET_MODE (SUBREG_REG (op))))) | |
695 | return CONST0_RTX (outermode); | |
696 | ||
697 | gcc_assert (ret != NULL_RTX); | |
698 | return ret; | |
e53a16e7 ILT |
699 | } |
700 | ||
701 | /* Return whether we should resolve X into the registers into which it | |
702 | was decomposed. */ | |
703 | ||
704 | static bool | |
705 | resolve_reg_p (rtx x) | |
706 | { | |
707 | return GET_CODE (x) == CONCATN; | |
708 | } | |
709 | ||
710 | /* Return whether X is a SUBREG of a register which we need to | |
711 | resolve. */ | |
712 | ||
713 | static bool | |
714 | resolve_subreg_p (rtx x) | |
715 | { | |
716 | if (GET_CODE (x) != SUBREG) | |
717 | return false; | |
718 | return resolve_reg_p (SUBREG_REG (x)); | |
719 | } | |
720 | ||
721 | /* This is called via for_each_rtx. Look for SUBREGs which need to be | |
722 | decomposed. */ | |
723 | ||
724 | static int | |
725 | resolve_subreg_use (rtx *px, void *data) | |
726 | { | |
727 | rtx insn = (rtx) data; | |
728 | rtx x = *px; | |
729 | ||
730 | if (x == NULL_RTX) | |
731 | return 0; | |
732 | ||
733 | if (resolve_subreg_p (x)) | |
734 | { | |
735 | x = simplify_subreg_concatn (GET_MODE (x), SUBREG_REG (x), | |
736 | SUBREG_BYTE (x)); | |
737 | ||
738 | /* It is possible for a note to contain a reference which we can | |
739 | decompose. In this case, return 1 to the caller to indicate | |
740 | that the note must be removed. */ | |
741 | if (!x) | |
742 | { | |
30d18db4 | 743 | gcc_assert (!insn); |
e53a16e7 ILT |
744 | return 1; |
745 | } | |
746 | ||
747 | validate_change (insn, px, x, 1); | |
748 | return -1; | |
749 | } | |
750 | ||
751 | if (resolve_reg_p (x)) | |
752 | { | |
753 | /* Return 1 to the caller to indicate that we found a direct | |
754 | reference to a register which is being decomposed. This can | |
e0892570 AK |
755 | happen inside notes, multiword shift or zero-extend |
756 | instructions. */ | |
e53a16e7 ILT |
757 | return 1; |
758 | } | |
759 | ||
760 | return 0; | |
761 | } | |
762 | ||
b5b8b0ac AO |
763 | /* This is called via for_each_rtx. Look for SUBREGs which can be |
764 | decomposed and decomposed REGs that need copying. */ | |
765 | ||
766 | static int | |
767 | adjust_decomposed_uses (rtx *px, void *data ATTRIBUTE_UNUSED) | |
768 | { | |
769 | rtx x = *px; | |
770 | ||
771 | if (x == NULL_RTX) | |
772 | return 0; | |
773 | ||
774 | if (resolve_subreg_p (x)) | |
775 | { | |
776 | x = simplify_subreg_concatn (GET_MODE (x), SUBREG_REG (x), | |
777 | SUBREG_BYTE (x)); | |
778 | ||
779 | if (x) | |
780 | *px = x; | |
781 | else | |
782 | x = copy_rtx (*px); | |
783 | } | |
784 | ||
785 | if (resolve_reg_p (x)) | |
786 | *px = copy_rtx (x); | |
787 | ||
788 | return 0; | |
789 | } | |
790 | ||
e53a16e7 ILT |
791 | /* Resolve any decomposed registers which appear in register notes on |
792 | INSN. */ | |
793 | ||
794 | static void | |
795 | resolve_reg_notes (rtx insn) | |
796 | { | |
797 | rtx *pnote, note; | |
798 | ||
799 | note = find_reg_equal_equiv_note (insn); | |
800 | if (note) | |
801 | { | |
6fb5fa3c | 802 | int old_count = num_validated_changes (); |
e53a16e7 | 803 | if (for_each_rtx (&XEXP (note, 0), resolve_subreg_use, NULL)) |
4a8cae83 | 804 | remove_note (insn, note); |
6fb5fa3c DB |
805 | else |
806 | if (old_count != num_validated_changes ()) | |
807 | df_notes_rescan (insn); | |
e53a16e7 ILT |
808 | } |
809 | ||
810 | pnote = ®_NOTES (insn); | |
811 | while (*pnote != NULL_RTX) | |
812 | { | |
60564289 | 813 | bool del = false; |
e53a16e7 ILT |
814 | |
815 | note = *pnote; | |
816 | switch (REG_NOTE_KIND (note)) | |
817 | { | |
6fb5fa3c DB |
818 | case REG_DEAD: |
819 | case REG_UNUSED: | |
e53a16e7 | 820 | if (resolve_reg_p (XEXP (note, 0))) |
60564289 | 821 | del = true; |
e53a16e7 ILT |
822 | break; |
823 | ||
824 | default: | |
825 | break; | |
826 | } | |
827 | ||
60564289 | 828 | if (del) |
e53a16e7 ILT |
829 | *pnote = XEXP (note, 1); |
830 | else | |
831 | pnote = &XEXP (note, 1); | |
832 | } | |
833 | } | |
834 | ||
2b54c30f | 835 | /* Return whether X can be decomposed into subwords. */ |
e53a16e7 ILT |
836 | |
837 | static bool | |
2b54c30f | 838 | can_decompose_p (rtx x) |
e53a16e7 ILT |
839 | { |
840 | if (REG_P (x)) | |
841 | { | |
842 | unsigned int regno = REGNO (x); | |
843 | ||
844 | if (HARD_REGISTER_NUM_P (regno)) | |
488c8379 RS |
845 | { |
846 | unsigned int byte, num_bytes; | |
847 | ||
848 | num_bytes = GET_MODE_SIZE (GET_MODE (x)); | |
849 | for (byte = 0; byte < num_bytes; byte += UNITS_PER_WORD) | |
850 | if (simplify_subreg_regno (regno, GET_MODE (x), byte, word_mode) < 0) | |
851 | return false; | |
852 | return true; | |
853 | } | |
e53a16e7 | 854 | else |
402464a0 | 855 | return !bitmap_bit_p (subreg_context, regno); |
e53a16e7 ILT |
856 | } |
857 | ||
2b54c30f | 858 | return true; |
e53a16e7 ILT |
859 | } |
860 | ||
861 | /* Decompose the registers used in a simple move SET within INSN. If | |
862 | we don't change anything, return INSN, otherwise return the start | |
863 | of the sequence of moves. */ | |
864 | ||
865 | static rtx | |
866 | resolve_simple_move (rtx set, rtx insn) | |
867 | { | |
868 | rtx src, dest, real_dest, insns; | |
869 | enum machine_mode orig_mode, dest_mode; | |
870 | unsigned int words; | |
871 | bool pushing; | |
872 | ||
873 | src = SET_SRC (set); | |
874 | dest = SET_DEST (set); | |
875 | orig_mode = GET_MODE (dest); | |
876 | ||
877 | words = (GET_MODE_SIZE (orig_mode) + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
af4ba423 | 878 | gcc_assert (words > 1); |
e53a16e7 ILT |
879 | |
880 | start_sequence (); | |
881 | ||
882 | /* We have to handle copying from a SUBREG of a decomposed reg where | |
883 | the SUBREG is larger than word size. Rather than assume that we | |
884 | can take a word_mode SUBREG of the destination, we copy to a new | |
885 | register and then copy that to the destination. */ | |
886 | ||
887 | real_dest = NULL_RTX; | |
888 | ||
889 | if (GET_CODE (src) == SUBREG | |
890 | && resolve_reg_p (SUBREG_REG (src)) | |
891 | && (SUBREG_BYTE (src) != 0 | |
892 | || (GET_MODE_SIZE (orig_mode) | |
893 | != GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))))) | |
894 | { | |
895 | real_dest = dest; | |
896 | dest = gen_reg_rtx (orig_mode); | |
897 | if (REG_P (real_dest)) | |
898 | REG_ATTRS (dest) = REG_ATTRS (real_dest); | |
899 | } | |
900 | ||
901 | /* Similarly if we are copying to a SUBREG of a decomposed reg where | |
902 | the SUBREG is larger than word size. */ | |
903 | ||
904 | if (GET_CODE (dest) == SUBREG | |
905 | && resolve_reg_p (SUBREG_REG (dest)) | |
906 | && (SUBREG_BYTE (dest) != 0 | |
907 | || (GET_MODE_SIZE (orig_mode) | |
908 | != GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest)))))) | |
909 | { | |
910 | rtx reg, minsn, smove; | |
911 | ||
912 | reg = gen_reg_rtx (orig_mode); | |
913 | minsn = emit_move_insn (reg, src); | |
914 | smove = single_set (minsn); | |
915 | gcc_assert (smove != NULL_RTX); | |
916 | resolve_simple_move (smove, minsn); | |
917 | src = reg; | |
918 | } | |
919 | ||
920 | /* If we didn't have any big SUBREGS of decomposed registers, and | |
921 | neither side of the move is a register we are decomposing, then | |
922 | we don't have to do anything here. */ | |
923 | ||
924 | if (src == SET_SRC (set) | |
925 | && dest == SET_DEST (set) | |
926 | && !resolve_reg_p (src) | |
927 | && !resolve_subreg_p (src) | |
928 | && !resolve_reg_p (dest) | |
929 | && !resolve_subreg_p (dest)) | |
930 | { | |
931 | end_sequence (); | |
932 | return insn; | |
933 | } | |
934 | ||
30d18db4 ILT |
935 | /* It's possible for the code to use a subreg of a decomposed |
936 | register while forming an address. We need to handle that before | |
937 | passing the address to emit_move_insn. We pass NULL_RTX as the | |
938 | insn parameter to resolve_subreg_use because we can not validate | |
939 | the insn yet. */ | |
940 | if (MEM_P (src) || MEM_P (dest)) | |
941 | { | |
942 | int acg; | |
943 | ||
944 | if (MEM_P (src)) | |
945 | for_each_rtx (&XEXP (src, 0), resolve_subreg_use, NULL_RTX); | |
946 | if (MEM_P (dest)) | |
947 | for_each_rtx (&XEXP (dest, 0), resolve_subreg_use, NULL_RTX); | |
948 | acg = apply_change_group (); | |
949 | gcc_assert (acg); | |
950 | } | |
951 | ||
e53a16e7 ILT |
952 | /* If SRC is a register which we can't decompose, or has side |
953 | effects, we need to move via a temporary register. */ | |
954 | ||
2b54c30f | 955 | if (!can_decompose_p (src) |
e53a16e7 ILT |
956 | || side_effects_p (src) |
957 | || GET_CODE (src) == ASM_OPERANDS) | |
958 | { | |
959 | rtx reg; | |
960 | ||
961 | reg = gen_reg_rtx (orig_mode); | |
962 | emit_move_insn (reg, src); | |
963 | src = reg; | |
964 | } | |
965 | ||
966 | /* If DEST is a register which we can't decompose, or has side | |
967 | effects, we need to first move to a temporary register. We | |
968 | handle the common case of pushing an operand directly. We also | |
969 | go through a temporary register if it holds a floating point | |
970 | value. This gives us better code on systems which can't move | |
971 | data easily between integer and floating point registers. */ | |
972 | ||
973 | dest_mode = orig_mode; | |
974 | pushing = push_operand (dest, dest_mode); | |
2b54c30f | 975 | if (!can_decompose_p (dest) |
e53a16e7 ILT |
976 | || (side_effects_p (dest) && !pushing) |
977 | || (!SCALAR_INT_MODE_P (dest_mode) | |
978 | && !resolve_reg_p (dest) | |
979 | && !resolve_subreg_p (dest))) | |
980 | { | |
981 | if (real_dest == NULL_RTX) | |
982 | real_dest = dest; | |
983 | if (!SCALAR_INT_MODE_P (dest_mode)) | |
984 | { | |
985 | dest_mode = mode_for_size (GET_MODE_SIZE (dest_mode) * BITS_PER_UNIT, | |
986 | MODE_INT, 0); | |
987 | gcc_assert (dest_mode != BLKmode); | |
988 | } | |
989 | dest = gen_reg_rtx (dest_mode); | |
990 | if (REG_P (real_dest)) | |
991 | REG_ATTRS (dest) = REG_ATTRS (real_dest); | |
992 | } | |
993 | ||
994 | if (pushing) | |
995 | { | |
996 | unsigned int i, j, jinc; | |
997 | ||
998 | gcc_assert (GET_MODE_SIZE (orig_mode) % UNITS_PER_WORD == 0); | |
999 | gcc_assert (GET_CODE (XEXP (dest, 0)) != PRE_MODIFY); | |
1000 | gcc_assert (GET_CODE (XEXP (dest, 0)) != POST_MODIFY); | |
1001 | ||
1002 | if (WORDS_BIG_ENDIAN == STACK_GROWS_DOWNWARD) | |
1003 | { | |
1004 | j = 0; | |
1005 | jinc = 1; | |
1006 | } | |
1007 | else | |
1008 | { | |
1009 | j = words - 1; | |
1010 | jinc = -1; | |
1011 | } | |
1012 | ||
1013 | for (i = 0; i < words; ++i, j += jinc) | |
1014 | { | |
1015 | rtx temp; | |
1016 | ||
1017 | temp = copy_rtx (XEXP (dest, 0)); | |
1018 | temp = adjust_automodify_address_nv (dest, word_mode, temp, | |
1019 | j * UNITS_PER_WORD); | |
1020 | emit_move_insn (temp, | |
1021 | simplify_gen_subreg_concatn (word_mode, src, | |
1022 | orig_mode, | |
1023 | j * UNITS_PER_WORD)); | |
1024 | } | |
1025 | } | |
1026 | else | |
1027 | { | |
1028 | unsigned int i; | |
1029 | ||
1030 | if (REG_P (dest) && !HARD_REGISTER_NUM_P (REGNO (dest))) | |
c41c1387 | 1031 | emit_clobber (dest); |
e53a16e7 ILT |
1032 | |
1033 | for (i = 0; i < words; ++i) | |
1034 | emit_move_insn (simplify_gen_subreg_concatn (word_mode, dest, | |
1035 | dest_mode, | |
1036 | i * UNITS_PER_WORD), | |
1037 | simplify_gen_subreg_concatn (word_mode, src, | |
1038 | orig_mode, | |
1039 | i * UNITS_PER_WORD)); | |
1040 | } | |
1041 | ||
1042 | if (real_dest != NULL_RTX) | |
1043 | { | |
1044 | rtx mdest, minsn, smove; | |
1045 | ||
1046 | if (dest_mode == orig_mode) | |
1047 | mdest = dest; | |
1048 | else | |
1049 | mdest = simplify_gen_subreg (orig_mode, dest, GET_MODE (dest), 0); | |
1050 | minsn = emit_move_insn (real_dest, mdest); | |
1051 | ||
1052 | smove = single_set (minsn); | |
1053 | gcc_assert (smove != NULL_RTX); | |
1054 | ||
1055 | resolve_simple_move (smove, minsn); | |
1056 | } | |
1057 | ||
1058 | insns = get_insns (); | |
1059 | end_sequence (); | |
1060 | ||
1d65f45c | 1061 | copy_reg_eh_region_note_forward (insn, insns, NULL_RTX); |
73663bb7 | 1062 | |
e53a16e7 ILT |
1063 | emit_insn_before (insns, insn); |
1064 | ||
e53a16e7 ILT |
1065 | delete_insn (insn); |
1066 | ||
1067 | return insns; | |
1068 | } | |
1069 | ||
1070 | /* Change a CLOBBER of a decomposed register into a CLOBBER of the | |
1071 | component registers. Return whether we changed something. */ | |
1072 | ||
1073 | static bool | |
1074 | resolve_clobber (rtx pat, rtx insn) | |
1075 | { | |
d70dcf29 | 1076 | rtx reg; |
e53a16e7 ILT |
1077 | enum machine_mode orig_mode; |
1078 | unsigned int words, i; | |
7e0c3f57 | 1079 | int ret; |
e53a16e7 ILT |
1080 | |
1081 | reg = XEXP (pat, 0); | |
9a5a8e58 | 1082 | if (!resolve_reg_p (reg) && !resolve_subreg_p (reg)) |
e53a16e7 ILT |
1083 | return false; |
1084 | ||
1085 | orig_mode = GET_MODE (reg); | |
1086 | words = GET_MODE_SIZE (orig_mode); | |
1087 | words = (words + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
1088 | ||
7e0c3f57 ILT |
1089 | ret = validate_change (NULL_RTX, &XEXP (pat, 0), |
1090 | simplify_gen_subreg_concatn (word_mode, reg, | |
1091 | orig_mode, 0), | |
1092 | 0); | |
6fb5fa3c | 1093 | df_insn_rescan (insn); |
7e0c3f57 ILT |
1094 | gcc_assert (ret != 0); |
1095 | ||
e53a16e7 ILT |
1096 | for (i = words - 1; i > 0; --i) |
1097 | { | |
1098 | rtx x; | |
1099 | ||
9a5a8e58 ILT |
1100 | x = simplify_gen_subreg_concatn (word_mode, reg, orig_mode, |
1101 | i * UNITS_PER_WORD); | |
e53a16e7 ILT |
1102 | x = gen_rtx_CLOBBER (VOIDmode, x); |
1103 | emit_insn_after (x, insn); | |
1104 | } | |
1105 | ||
d4fd3465 ILT |
1106 | resolve_reg_notes (insn); |
1107 | ||
e53a16e7 ILT |
1108 | return true; |
1109 | } | |
1110 | ||
1111 | /* A USE of a decomposed register is no longer meaningful. Return | |
1112 | whether we changed something. */ | |
1113 | ||
1114 | static bool | |
1115 | resolve_use (rtx pat, rtx insn) | |
1116 | { | |
1117 | if (resolve_reg_p (XEXP (pat, 0)) || resolve_subreg_p (XEXP (pat, 0))) | |
1118 | { | |
1119 | delete_insn (insn); | |
1120 | return true; | |
1121 | } | |
d4fd3465 ILT |
1122 | |
1123 | resolve_reg_notes (insn); | |
1124 | ||
e53a16e7 ILT |
1125 | return false; |
1126 | } | |
1127 | ||
b5b8b0ac AO |
1128 | /* A VAR_LOCATION can be simplified. */ |
1129 | ||
1130 | static void | |
1131 | resolve_debug (rtx insn) | |
1132 | { | |
1133 | for_each_rtx (&PATTERN (insn), adjust_decomposed_uses, NULL_RTX); | |
1134 | ||
1135 | df_insn_rescan (insn); | |
1136 | ||
1137 | resolve_reg_notes (insn); | |
1138 | } | |
1139 | ||
af4ba423 KZ |
1140 | /* Check if INSN is a decomposable multiword-shift or zero-extend and |
1141 | set the decomposable_context bitmap accordingly. SPEED_P is true | |
1142 | if we are optimizing INSN for speed rather than size. Return true | |
1143 | if INSN is decomposable. */ | |
e0892570 | 1144 | |
af4ba423 KZ |
1145 | static bool |
1146 | find_decomposable_shift_zext (rtx insn, bool speed_p) | |
e0892570 AK |
1147 | { |
1148 | rtx set; | |
1149 | rtx op; | |
1150 | rtx op_operand; | |
1151 | ||
1152 | set = single_set (insn); | |
1153 | if (!set) | |
af4ba423 | 1154 | return false; |
e0892570 AK |
1155 | |
1156 | op = SET_SRC (set); | |
1157 | if (GET_CODE (op) != ASHIFT | |
1158 | && GET_CODE (op) != LSHIFTRT | |
1159 | && GET_CODE (op) != ZERO_EXTEND) | |
af4ba423 | 1160 | return false; |
e0892570 AK |
1161 | |
1162 | op_operand = XEXP (op, 0); | |
1163 | if (!REG_P (SET_DEST (set)) || !REG_P (op_operand) | |
1164 | || HARD_REGISTER_NUM_P (REGNO (SET_DEST (set))) | |
1165 | || HARD_REGISTER_NUM_P (REGNO (op_operand)) | |
af4ba423 KZ |
1166 | || GET_MODE (op) != twice_word_mode) |
1167 | return false; | |
e0892570 AK |
1168 | |
1169 | if (GET_CODE (op) == ZERO_EXTEND) | |
1170 | { | |
1171 | if (GET_MODE (op_operand) != word_mode | |
af4ba423 KZ |
1172 | || !choices[speed_p].splitting_zext) |
1173 | return false; | |
e0892570 AK |
1174 | } |
1175 | else /* left or right shift */ | |
1176 | { | |
af4ba423 KZ |
1177 | bool *splitting = (GET_CODE (op) == ASHIFT |
1178 | ? choices[speed_p].splitting_ashift | |
1179 | : choices[speed_p].splitting_lshiftrt); | |
481683e1 | 1180 | if (!CONST_INT_P (XEXP (op, 1)) |
af4ba423 KZ |
1181 | || !IN_RANGE (INTVAL (XEXP (op, 1)), BITS_PER_WORD, |
1182 | 2 * BITS_PER_WORD - 1) | |
1183 | || !splitting[INTVAL (XEXP (op, 1)) - BITS_PER_WORD]) | |
1184 | return false; | |
1185 | ||
1186 | bitmap_set_bit (decomposable_context, REGNO (op_operand)); | |
e0892570 AK |
1187 | } |
1188 | ||
1189 | bitmap_set_bit (decomposable_context, REGNO (SET_DEST (set))); | |
1190 | ||
af4ba423 | 1191 | return true; |
e0892570 AK |
1192 | } |
1193 | ||
1194 | /* Decompose a more than word wide shift (in INSN) of a multiword | |
1195 | pseudo or a multiword zero-extend of a wordmode pseudo into a move | |
1196 | and 'set to zero' insn. Return a pointer to the new insn when a | |
1197 | replacement was done. */ | |
1198 | ||
1199 | static rtx | |
1200 | resolve_shift_zext (rtx insn) | |
1201 | { | |
1202 | rtx set; | |
1203 | rtx op; | |
1204 | rtx op_operand; | |
1205 | rtx insns; | |
1206 | rtx src_reg, dest_reg, dest_zero; | |
1207 | int src_reg_num, dest_reg_num, offset1, offset2, src_offset; | |
1208 | ||
1209 | set = single_set (insn); | |
1210 | if (!set) | |
1211 | return NULL_RTX; | |
1212 | ||
1213 | op = SET_SRC (set); | |
1214 | if (GET_CODE (op) != ASHIFT | |
1215 | && GET_CODE (op) != LSHIFTRT | |
1216 | && GET_CODE (op) != ZERO_EXTEND) | |
1217 | return NULL_RTX; | |
1218 | ||
1219 | op_operand = XEXP (op, 0); | |
1220 | ||
af4ba423 KZ |
1221 | /* We can tear this operation apart only if the regs were already |
1222 | torn apart. */ | |
e0892570 AK |
1223 | if (!resolve_reg_p (SET_DEST (set)) && !resolve_reg_p (op_operand)) |
1224 | return NULL_RTX; | |
1225 | ||
1226 | /* src_reg_num is the number of the word mode register which we | |
1227 | are operating on. For a left shift and a zero_extend on little | |
1228 | endian machines this is register 0. */ | |
1229 | src_reg_num = GET_CODE (op) == LSHIFTRT ? 1 : 0; | |
1230 | ||
acbe5496 AK |
1231 | if (WORDS_BIG_ENDIAN |
1232 | && GET_MODE_SIZE (GET_MODE (op_operand)) > UNITS_PER_WORD) | |
e0892570 AK |
1233 | src_reg_num = 1 - src_reg_num; |
1234 | ||
1235 | if (GET_CODE (op) == ZERO_EXTEND) | |
acbe5496 | 1236 | dest_reg_num = WORDS_BIG_ENDIAN ? 1 : 0; |
e0892570 AK |
1237 | else |
1238 | dest_reg_num = 1 - src_reg_num; | |
1239 | ||
1240 | offset1 = UNITS_PER_WORD * dest_reg_num; | |
1241 | offset2 = UNITS_PER_WORD * (1 - dest_reg_num); | |
1242 | src_offset = UNITS_PER_WORD * src_reg_num; | |
1243 | ||
e0892570 AK |
1244 | start_sequence (); |
1245 | ||
1246 | dest_reg = simplify_gen_subreg_concatn (word_mode, SET_DEST (set), | |
1247 | GET_MODE (SET_DEST (set)), | |
1248 | offset1); | |
1249 | dest_zero = simplify_gen_subreg_concatn (word_mode, SET_DEST (set), | |
1250 | GET_MODE (SET_DEST (set)), | |
1251 | offset2); | |
1252 | src_reg = simplify_gen_subreg_concatn (word_mode, op_operand, | |
1253 | GET_MODE (op_operand), | |
1254 | src_offset); | |
1255 | if (GET_CODE (op) != ZERO_EXTEND) | |
1256 | { | |
1257 | int shift_count = INTVAL (XEXP (op, 1)); | |
1258 | if (shift_count > BITS_PER_WORD) | |
1259 | src_reg = expand_shift (GET_CODE (op) == ASHIFT ? | |
1260 | LSHIFT_EXPR : RSHIFT_EXPR, | |
1261 | word_mode, src_reg, | |
eb6c3df1 | 1262 | shift_count - BITS_PER_WORD, |
e0892570 AK |
1263 | dest_reg, 1); |
1264 | } | |
1265 | ||
1266 | if (dest_reg != src_reg) | |
1267 | emit_move_insn (dest_reg, src_reg); | |
1268 | emit_move_insn (dest_zero, CONST0_RTX (word_mode)); | |
1269 | insns = get_insns (); | |
1270 | ||
1271 | end_sequence (); | |
1272 | ||
1273 | emit_insn_before (insns, insn); | |
1274 | ||
1275 | if (dump_file) | |
1276 | { | |
1277 | rtx in; | |
1278 | fprintf (dump_file, "; Replacing insn: %d with insns: ", INSN_UID (insn)); | |
1279 | for (in = insns; in != insn; in = NEXT_INSN (in)) | |
1280 | fprintf (dump_file, "%d ", INSN_UID (in)); | |
1281 | fprintf (dump_file, "\n"); | |
1282 | } | |
1283 | ||
1284 | delete_insn (insn); | |
1285 | return insns; | |
1286 | } | |
1287 | ||
af4ba423 KZ |
1288 | /* Print to dump_file a description of what we're doing with shift code CODE. |
1289 | SPLITTING[X] is true if we are splitting shifts by X + BITS_PER_WORD. */ | |
1290 | ||
1291 | static void | |
1292 | dump_shift_choices (enum rtx_code code, bool *splitting) | |
1293 | { | |
1294 | int i; | |
1295 | const char *sep; | |
1296 | ||
1297 | fprintf (dump_file, | |
1298 | " Splitting mode %s for %s lowering with shift amounts = ", | |
1299 | GET_MODE_NAME (twice_word_mode), GET_RTX_NAME (code)); | |
1300 | sep = ""; | |
1301 | for (i = 0; i < BITS_PER_WORD; i++) | |
1302 | if (splitting[i]) | |
1303 | { | |
1304 | fprintf (dump_file, "%s%d", sep, i + BITS_PER_WORD); | |
1305 | sep = ","; | |
1306 | } | |
1307 | fprintf (dump_file, "\n"); | |
1308 | } | |
1309 | ||
1310 | /* Print to dump_file a description of what we're doing when optimizing | |
1311 | for speed or size; SPEED_P says which. DESCRIPTION is a description | |
1312 | of the SPEED_P choice. */ | |
1313 | ||
1314 | static void | |
1315 | dump_choices (bool speed_p, const char *description) | |
1316 | { | |
1317 | unsigned int i; | |
1318 | ||
1319 | fprintf (dump_file, "Choices when optimizing for %s:\n", description); | |
1320 | ||
1321 | for (i = 0; i < MAX_MACHINE_MODE; i++) | |
1322 | if (GET_MODE_SIZE (i) > UNITS_PER_WORD) | |
1323 | fprintf (dump_file, " %s mode %s for copy lowering.\n", | |
1324 | choices[speed_p].move_modes_to_split[i] | |
1325 | ? "Splitting" | |
1326 | : "Skipping", | |
1327 | GET_MODE_NAME ((enum machine_mode) i)); | |
1328 | ||
1329 | fprintf (dump_file, " %s mode %s for zero_extend lowering.\n", | |
1330 | choices[speed_p].splitting_zext ? "Splitting" : "Skipping", | |
1331 | GET_MODE_NAME (twice_word_mode)); | |
1332 | ||
1333 | dump_shift_choices (ASHIFT, choices[speed_p].splitting_ashift); | |
1334 | dump_shift_choices (LSHIFTRT, choices[speed_p].splitting_ashift); | |
1335 | fprintf (dump_file, "\n"); | |
1336 | } | |
1337 | ||
e53a16e7 | 1338 | /* Look for registers which are always accessed via word-sized SUBREGs |
c2c47e8f UW |
1339 | or -if DECOMPOSE_COPIES is true- via copies. Decompose these |
1340 | registers into several word-sized pseudo-registers. */ | |
e53a16e7 ILT |
1341 | |
1342 | static void | |
c2c47e8f | 1343 | decompose_multiword_subregs (bool decompose_copies) |
e53a16e7 ILT |
1344 | { |
1345 | unsigned int max; | |
1346 | basic_block bb; | |
af4ba423 | 1347 | bool speed_p; |
e53a16e7 | 1348 | |
af4ba423 KZ |
1349 | if (dump_file) |
1350 | { | |
1351 | dump_choices (false, "size"); | |
1352 | dump_choices (true, "speed"); | |
1353 | } | |
1354 | ||
1355 | /* Check if this target even has any modes to consider lowering. */ | |
1356 | if (!choices[false].something_to_do && !choices[true].something_to_do) | |
1357 | { | |
1358 | if (dump_file) | |
1359 | fprintf (dump_file, "Nothing to do!\n"); | |
1360 | return; | |
1361 | } | |
6fb5fa3c | 1362 | |
e53a16e7 ILT |
1363 | max = max_reg_num (); |
1364 | ||
1365 | /* First see if there are any multi-word pseudo-registers. If there | |
1366 | aren't, there is nothing we can do. This should speed up this | |
1367 | pass in the normal case, since it should be faster than scanning | |
1368 | all the insns. */ | |
1369 | { | |
1370 | unsigned int i; | |
af4ba423 | 1371 | bool useful_modes_seen = false; |
e53a16e7 ILT |
1372 | |
1373 | for (i = FIRST_PSEUDO_REGISTER; i < max; ++i) | |
af4ba423 KZ |
1374 | if (regno_reg_rtx[i] != NULL) |
1375 | { | |
1376 | enum machine_mode mode = GET_MODE (regno_reg_rtx[i]); | |
1377 | if (choices[false].move_modes_to_split[(int) mode] | |
1378 | || choices[true].move_modes_to_split[(int) mode]) | |
1379 | { | |
1380 | useful_modes_seen = true; | |
1381 | break; | |
1382 | } | |
1383 | } | |
1384 | ||
1385 | if (!useful_modes_seen) | |
e53a16e7 | 1386 | { |
af4ba423 KZ |
1387 | if (dump_file) |
1388 | fprintf (dump_file, "Nothing to lower in this function.\n"); | |
1389 | return; | |
e53a16e7 | 1390 | } |
e53a16e7 ILT |
1391 | } |
1392 | ||
8d074192 | 1393 | if (df) |
af4ba423 KZ |
1394 | { |
1395 | df_set_flags (DF_DEFER_INSN_RESCAN); | |
1396 | run_word_dce (); | |
1397 | } | |
8d074192 | 1398 | |
af4ba423 KZ |
1399 | /* FIXME: It may be possible to change this code to look for each |
1400 | multi-word pseudo-register and to find each insn which sets or | |
1401 | uses that register. That should be faster than scanning all the | |
1402 | insns. */ | |
e53a16e7 ILT |
1403 | |
1404 | decomposable_context = BITMAP_ALLOC (NULL); | |
1405 | non_decomposable_context = BITMAP_ALLOC (NULL); | |
402464a0 | 1406 | subreg_context = BITMAP_ALLOC (NULL); |
e53a16e7 ILT |
1407 | |
1408 | reg_copy_graph = VEC_alloc (bitmap, heap, max); | |
1409 | VEC_safe_grow (bitmap, heap, reg_copy_graph, max); | |
1410 | memset (VEC_address (bitmap, reg_copy_graph), 0, sizeof (bitmap) * max); | |
1411 | ||
af4ba423 | 1412 | speed_p = optimize_function_for_speed_p (cfun); |
e53a16e7 ILT |
1413 | FOR_EACH_BB (bb) |
1414 | { | |
1415 | rtx insn; | |
1416 | ||
1417 | FOR_BB_INSNS (bb, insn) | |
1418 | { | |
1419 | rtx set; | |
1420 | enum classify_move_insn cmi; | |
1421 | int i, n; | |
1422 | ||
1423 | if (!INSN_P (insn) | |
1424 | || GET_CODE (PATTERN (insn)) == CLOBBER | |
1425 | || GET_CODE (PATTERN (insn)) == USE) | |
1426 | continue; | |
1427 | ||
d5785e76 JJ |
1428 | recog_memoized (insn); |
1429 | ||
af4ba423 | 1430 | if (find_decomposable_shift_zext (insn, speed_p)) |
e0892570 AK |
1431 | continue; |
1432 | ||
e53a16e7 ILT |
1433 | extract_insn (insn); |
1434 | ||
af4ba423 | 1435 | set = simple_move (insn, speed_p); |
e53a16e7 ILT |
1436 | |
1437 | if (!set) | |
1438 | cmi = NOT_SIMPLE_MOVE; | |
1439 | else | |
1440 | { | |
c2c47e8f UW |
1441 | /* We mark pseudo-to-pseudo copies as decomposable during the |
1442 | second pass only. The first pass is so early that there is | |
1443 | good chance such moves will be optimized away completely by | |
1444 | subsequent optimizations anyway. | |
1445 | ||
1446 | However, we call find_pseudo_copy even during the first pass | |
1447 | so as to properly set up the reg_copy_graph. */ | |
4a8cae83 | 1448 | if (find_pseudo_copy (set)) |
c2c47e8f | 1449 | cmi = decompose_copies? DECOMPOSABLE_SIMPLE_MOVE : SIMPLE_MOVE; |
e53a16e7 ILT |
1450 | else |
1451 | cmi = SIMPLE_MOVE; | |
1452 | } | |
1453 | ||
1454 | n = recog_data.n_operands; | |
1455 | for (i = 0; i < n; ++i) | |
1456 | { | |
1457 | for_each_rtx (&recog_data.operand[i], | |
1458 | find_decomposable_subregs, | |
1459 | &cmi); | |
1460 | ||
1461 | /* We handle ASM_OPERANDS as a special case to support | |
1462 | things like x86 rdtsc which returns a DImode value. | |
1463 | We can decompose the output, which will certainly be | |
1464 | operand 0, but not the inputs. */ | |
1465 | ||
1466 | if (cmi == SIMPLE_MOVE | |
1467 | && GET_CODE (SET_SRC (set)) == ASM_OPERANDS) | |
1468 | { | |
1469 | gcc_assert (i == 0); | |
1470 | cmi = NOT_SIMPLE_MOVE; | |
1471 | } | |
1472 | } | |
1473 | } | |
1474 | } | |
1475 | ||
1476 | bitmap_and_compl_into (decomposable_context, non_decomposable_context); | |
1477 | if (!bitmap_empty_p (decomposable_context)) | |
1478 | { | |
73663bb7 | 1479 | sbitmap sub_blocks; |
7984c787 SB |
1480 | unsigned int i; |
1481 | sbitmap_iterator sbi; | |
e53a16e7 ILT |
1482 | bitmap_iterator iter; |
1483 | unsigned int regno; | |
1484 | ||
1485 | propagate_pseudo_copies (); | |
1486 | ||
73663bb7 | 1487 | sub_blocks = sbitmap_alloc (last_basic_block); |
f61e445a | 1488 | bitmap_clear (sub_blocks); |
e53a16e7 ILT |
1489 | |
1490 | EXECUTE_IF_SET_IN_BITMAP (decomposable_context, 0, regno, iter) | |
1491 | decompose_register (regno); | |
1492 | ||
1493 | FOR_EACH_BB (bb) | |
1494 | { | |
ba4807a0 | 1495 | rtx insn; |
e53a16e7 | 1496 | |
ba4807a0 | 1497 | FOR_BB_INSNS (bb, insn) |
e53a16e7 | 1498 | { |
11895e28 | 1499 | rtx pat; |
e53a16e7 ILT |
1500 | |
1501 | if (!INSN_P (insn)) | |
1502 | continue; | |
1503 | ||
e53a16e7 ILT |
1504 | pat = PATTERN (insn); |
1505 | if (GET_CODE (pat) == CLOBBER) | |
d4fd3465 | 1506 | resolve_clobber (pat, insn); |
e53a16e7 | 1507 | else if (GET_CODE (pat) == USE) |
d4fd3465 | 1508 | resolve_use (pat, insn); |
b5b8b0ac AO |
1509 | else if (DEBUG_INSN_P (insn)) |
1510 | resolve_debug (insn); | |
e53a16e7 ILT |
1511 | else |
1512 | { | |
1513 | rtx set; | |
1514 | int i; | |
1515 | ||
1516 | recog_memoized (insn); | |
1517 | extract_insn (insn); | |
1518 | ||
af4ba423 | 1519 | set = simple_move (insn, speed_p); |
e53a16e7 ILT |
1520 | if (set) |
1521 | { | |
1522 | rtx orig_insn = insn; | |
73663bb7 | 1523 | bool cfi = control_flow_insn_p (insn); |
e53a16e7 | 1524 | |
7984c787 SB |
1525 | /* We can end up splitting loads to multi-word pseudos |
1526 | into separate loads to machine word size pseudos. | |
1527 | When this happens, we first had one load that can | |
1528 | throw, and after resolve_simple_move we'll have a | |
1529 | bunch of loads (at least two). All those loads may | |
1530 | trap if we can have non-call exceptions, so they | |
1531 | all will end the current basic block. We split the | |
1532 | block after the outer loop over all insns, but we | |
1533 | make sure here that we will be able to split the | |
1534 | basic block and still produce the correct control | |
1535 | flow graph for it. */ | |
1536 | gcc_assert (!cfi | |
8f4f502f | 1537 | || (cfun->can_throw_non_call_exceptions |
7984c787 SB |
1538 | && can_throw_internal (insn))); |
1539 | ||
e53a16e7 ILT |
1540 | insn = resolve_simple_move (set, insn); |
1541 | if (insn != orig_insn) | |
1542 | { | |
e53a16e7 ILT |
1543 | recog_memoized (insn); |
1544 | extract_insn (insn); | |
73663bb7 ILT |
1545 | |
1546 | if (cfi) | |
1547 | SET_BIT (sub_blocks, bb->index); | |
e53a16e7 ILT |
1548 | } |
1549 | } | |
e0892570 AK |
1550 | else |
1551 | { | |
1552 | rtx decomposed_shift; | |
1553 | ||
1554 | decomposed_shift = resolve_shift_zext (insn); | |
1555 | if (decomposed_shift != NULL_RTX) | |
1556 | { | |
e0892570 AK |
1557 | insn = decomposed_shift; |
1558 | recog_memoized (insn); | |
1559 | extract_insn (insn); | |
1560 | } | |
1561 | } | |
e53a16e7 ILT |
1562 | |
1563 | for (i = recog_data.n_operands - 1; i >= 0; --i) | |
1564 | for_each_rtx (recog_data.operand_loc[i], | |
1565 | resolve_subreg_use, | |
1566 | insn); | |
1567 | ||
1568 | resolve_reg_notes (insn); | |
1569 | ||
1570 | if (num_validated_changes () > 0) | |
1571 | { | |
1572 | for (i = recog_data.n_dups - 1; i >= 0; --i) | |
1573 | { | |
1574 | rtx *pl = recog_data.dup_loc[i]; | |
1575 | int dup_num = recog_data.dup_num[i]; | |
1576 | rtx *px = recog_data.operand_loc[dup_num]; | |
1577 | ||
1a309dfb | 1578 | validate_unshare_change (insn, pl, *px, 1); |
e53a16e7 ILT |
1579 | } |
1580 | ||
1581 | i = apply_change_group (); | |
1582 | gcc_assert (i); | |
e53a16e7 ILT |
1583 | } |
1584 | } | |
e53a16e7 ILT |
1585 | } |
1586 | } | |
1587 | ||
7984c787 SB |
1588 | /* If we had insns to split that caused control flow insns in the middle |
1589 | of a basic block, split those blocks now. Note that we only handle | |
1590 | the case where splitting a load has caused multiple possibly trapping | |
1591 | loads to appear. */ | |
1592 | EXECUTE_IF_SET_IN_SBITMAP (sub_blocks, 0, i, sbi) | |
1593 | { | |
1594 | rtx insn, end; | |
1595 | edge fallthru; | |
1596 | ||
1597 | bb = BASIC_BLOCK (i); | |
1598 | insn = BB_HEAD (bb); | |
1599 | end = BB_END (bb); | |
1600 | ||
1601 | while (insn != end) | |
1602 | { | |
1603 | if (control_flow_insn_p (insn)) | |
1604 | { | |
1605 | /* Split the block after insn. There will be a fallthru | |
1606 | edge, which is OK so we keep it. We have to create the | |
1607 | exception edges ourselves. */ | |
1608 | fallthru = split_block (bb, insn); | |
1609 | rtl_make_eh_edge (NULL, bb, BB_END (bb)); | |
1610 | bb = fallthru->dest; | |
1611 | insn = BB_HEAD (bb); | |
1612 | } | |
1613 | else | |
1614 | insn = NEXT_INSN (insn); | |
1615 | } | |
1616 | } | |
73663bb7 | 1617 | |
73663bb7 | 1618 | sbitmap_free (sub_blocks); |
e53a16e7 ILT |
1619 | } |
1620 | ||
1621 | { | |
1622 | unsigned int i; | |
1623 | bitmap b; | |
1624 | ||
ac47786e | 1625 | FOR_EACH_VEC_ELT (bitmap, reg_copy_graph, i, b) |
e53a16e7 ILT |
1626 | if (b) |
1627 | BITMAP_FREE (b); | |
1628 | } | |
1629 | ||
b8698a0f | 1630 | VEC_free (bitmap, heap, reg_copy_graph); |
e53a16e7 ILT |
1631 | |
1632 | BITMAP_FREE (decomposable_context); | |
1633 | BITMAP_FREE (non_decomposable_context); | |
402464a0 | 1634 | BITMAP_FREE (subreg_context); |
e53a16e7 ILT |
1635 | } |
1636 | \f | |
1637 | /* Gate function for lower subreg pass. */ | |
1638 | ||
1639 | static bool | |
1640 | gate_handle_lower_subreg (void) | |
1641 | { | |
1642 | return flag_split_wide_types != 0; | |
1643 | } | |
1644 | ||
1645 | /* Implement first lower subreg pass. */ | |
1646 | ||
1647 | static unsigned int | |
1648 | rest_of_handle_lower_subreg (void) | |
1649 | { | |
c2c47e8f | 1650 | decompose_multiword_subregs (false); |
e53a16e7 ILT |
1651 | return 0; |
1652 | } | |
1653 | ||
1654 | /* Implement second lower subreg pass. */ | |
1655 | ||
1656 | static unsigned int | |
1657 | rest_of_handle_lower_subreg2 (void) | |
1658 | { | |
c2c47e8f | 1659 | decompose_multiword_subregs (true); |
e53a16e7 ILT |
1660 | return 0; |
1661 | } | |
1662 | ||
8ddbbcae | 1663 | struct rtl_opt_pass pass_lower_subreg = |
e53a16e7 | 1664 | { |
8ddbbcae JH |
1665 | { |
1666 | RTL_PASS, | |
00b251a0 | 1667 | "subreg1", /* name */ |
e53a16e7 ILT |
1668 | gate_handle_lower_subreg, /* gate */ |
1669 | rest_of_handle_lower_subreg, /* execute */ | |
1670 | NULL, /* sub */ | |
1671 | NULL, /* next */ | |
1672 | 0, /* static_pass_number */ | |
1673 | TV_LOWER_SUBREG, /* tv_id */ | |
1674 | 0, /* properties_required */ | |
1675 | 0, /* properties_provided */ | |
1676 | 0, /* properties_destroyed */ | |
1677 | 0, /* todo_flags_start */ | |
73663bb7 | 1678 | TODO_ggc_collect | |
8ddbbcae JH |
1679 | TODO_verify_flow /* todo_flags_finish */ |
1680 | } | |
e53a16e7 ILT |
1681 | }; |
1682 | ||
8ddbbcae | 1683 | struct rtl_opt_pass pass_lower_subreg2 = |
e53a16e7 | 1684 | { |
8ddbbcae JH |
1685 | { |
1686 | RTL_PASS, | |
e53a16e7 ILT |
1687 | "subreg2", /* name */ |
1688 | gate_handle_lower_subreg, /* gate */ | |
1689 | rest_of_handle_lower_subreg2, /* execute */ | |
1690 | NULL, /* sub */ | |
1691 | NULL, /* next */ | |
1692 | 0, /* static_pass_number */ | |
1693 | TV_LOWER_SUBREG, /* tv_id */ | |
1694 | 0, /* properties_required */ | |
1695 | 0, /* properties_provided */ | |
1696 | 0, /* properties_destroyed */ | |
1697 | 0, /* todo_flags_start */ | |
a36b8a1e | 1698 | TODO_df_finish | TODO_verify_rtl_sharing | |
73663bb7 | 1699 | TODO_ggc_collect | |
8ddbbcae JH |
1700 | TODO_verify_flow /* todo_flags_finish */ |
1701 | } | |
e53a16e7 | 1702 | }; |