111
|
1 /* Change pseudos by memory.
|
|
2 Copyright (C) 2010-2017 Free Software Foundation, Inc.
|
|
3 Contributed by Vladimir Makarov <vmakarov@redhat.com>.
|
|
4
|
|
5 This file is part of GCC.
|
|
6
|
|
7 GCC is free software; you can redistribute it and/or modify it under
|
|
8 the terms of the GNU General Public License as published by the Free
|
|
9 Software Foundation; either version 3, or (at your option) any later
|
|
10 version.
|
|
11
|
|
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
|
|
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
|
|
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
|
|
15 for more details.
|
|
16
|
|
17 You should have received a copy of the GNU General Public License
|
|
18 along with GCC; see the file COPYING3. If not see
|
|
19 <http://www.gnu.org/licenses/>. */
|
|
20
|
|
21
|
|
22 /* This file contains code for a pass to change spilled pseudos into
|
|
23 memory.
|
|
24
|
|
25 The pass creates necessary stack slots and assigns spilled pseudos
|
|
26 to the stack slots in following way:
|
|
27
|
|
28 for all spilled pseudos P most frequently used first do
|
|
29 for all stack slots S do
|
|
30 if P doesn't conflict with pseudos assigned to S then
|
|
31 assign S to P and goto to the next pseudo process
|
|
32 end
|
|
33 end
|
|
34 create new stack slot S and assign P to S
|
|
35 end
|
|
36
|
|
37 The actual algorithm is bit more complicated because of different
|
|
38 pseudo sizes.
|
|
39
|
|
40 After that the code changes spilled pseudos (except ones created
|
|
41 from scratches) by corresponding stack slot memory in RTL.
|
|
42
|
|
43 If at least one stack slot was created, we need to run more passes
|
|
44 because we have new addresses which should be checked and because
|
|
45 the old address displacements might change and address constraints
|
|
46 (or insn memory constraints) might not be satisfied any more.
|
|
47
|
|
48 For some targets, the pass can spill some pseudos into hard
|
|
49 registers of different class (usually into vector registers)
|
|
50 instead of spilling them into memory if it is possible and
|
|
51 profitable. Spilling GENERAL_REGS pseudo into SSE registers for
|
|
52 Intel Corei7 is an example of such optimization. And this is
|
|
53 actually recommended by Intel optimization guide.
|
|
54
|
|
55 The file also contains code for final change of pseudos on hard
|
|
56 regs correspondingly assigned to them. */
|
|
57
|
|
58 #include "config.h"
|
|
59 #include "system.h"
|
|
60 #include "coretypes.h"
|
|
61 #include "backend.h"
|
|
62 #include "target.h"
|
|
63 #include "rtl.h"
|
|
64 #include "df.h"
|
|
65 #include "insn-config.h"
|
|
66 #include "regs.h"
|
|
67 #include "memmodel.h"
|
|
68 #include "ira.h"
|
|
69 #include "recog.h"
|
|
70 #include "output.h"
|
|
71 #include "cfgrtl.h"
|
|
72 #include "lra.h"
|
|
73 #include "lra-int.h"
|
|
74
|
|
75
|
|
76 /* Max regno at the start of the pass. */
|
|
77 static int regs_num;
|
|
78
|
|
79 /* Map spilled regno -> hard regno used instead of memory for
|
|
80 spilling. */
|
|
81 static rtx *spill_hard_reg;
|
|
82
|
|
83 /* The structure describes stack slot of a spilled pseudo. */
|
|
84 struct pseudo_slot
|
|
85 {
|
|
86 /* Number (0, 1, ...) of the stack slot to which given pseudo
|
|
87 belongs. */
|
|
88 int slot_num;
|
|
89 /* First or next slot with the same slot number. */
|
|
90 struct pseudo_slot *next, *first;
|
|
91 /* Memory representing the spilled pseudo. */
|
|
92 rtx mem;
|
|
93 };
|
|
94
|
|
95 /* The stack slots for each spilled pseudo. Indexed by regnos. */
|
|
96 static struct pseudo_slot *pseudo_slots;
|
|
97
|
|
98 /* The structure describes a register or a stack slot which can be
|
|
99 used for several spilled pseudos. */
|
|
100 struct slot
|
|
101 {
|
|
102 /* First pseudo with given stack slot. */
|
|
103 int regno;
|
|
104 /* Hard reg into which the slot pseudos are spilled. The value is
|
|
105 negative for pseudos spilled into memory. */
|
|
106 int hard_regno;
|
|
107 /* Maximum alignment required by all users of the slot. */
|
|
108 unsigned int align;
|
|
109 /* Maximum size required by all users of the slot. */
|
|
110 HOST_WIDE_INT size;
|
|
111 /* Memory representing the all stack slot. It can be different from
|
|
112 memory representing a pseudo belonging to give stack slot because
|
|
113 pseudo can be placed in a part of the corresponding stack slot.
|
|
114 The value is NULL for pseudos spilled into a hard reg. */
|
|
115 rtx mem;
|
|
116 /* Combined live ranges of all pseudos belonging to given slot. It
|
|
117 is used to figure out that a new spilled pseudo can use given
|
|
118 stack slot. */
|
|
119 lra_live_range_t live_ranges;
|
|
120 };
|
|
121
|
|
122 /* Array containing info about the stack slots. The array element is
|
|
123 indexed by the stack slot number in the range [0..slots_num). */
|
|
124 static struct slot *slots;
|
|
125 /* The number of the stack slots currently existing. */
|
|
126 static int slots_num;
|
|
127
|
|
128 /* Set up memory of the spilled pseudo I. The function can allocate
|
|
129 the corresponding stack slot if it is not done yet. */
|
|
130 static void
|
|
131 assign_mem_slot (int i)
|
|
132 {
|
|
133 rtx x = NULL_RTX;
|
|
134 machine_mode mode = GET_MODE (regno_reg_rtx[i]);
|
|
135 HOST_WIDE_INT inherent_size = PSEUDO_REGNO_BYTES (i);
|
|
136 machine_mode wider_mode
|
|
137 = wider_subreg_mode (mode, lra_reg_info[i].biggest_mode);
|
|
138 HOST_WIDE_INT total_size = GET_MODE_SIZE (wider_mode);
|
|
139 HOST_WIDE_INT adjust = 0;
|
|
140
|
|
141 lra_assert (regno_reg_rtx[i] != NULL_RTX && REG_P (regno_reg_rtx[i])
|
|
142 && lra_reg_info[i].nrefs != 0 && reg_renumber[i] < 0);
|
|
143
|
|
144 unsigned int slot_num = pseudo_slots[i].slot_num;
|
|
145 x = slots[slot_num].mem;
|
|
146 if (!x)
|
|
147 {
|
|
148 x = assign_stack_local (BLKmode, slots[slot_num].size,
|
|
149 slots[slot_num].align);
|
|
150 slots[slot_num].mem = x;
|
|
151 }
|
|
152
|
|
153 /* On a big endian machine, the "address" of the slot is the address
|
|
154 of the low part that fits its inherent mode. */
|
|
155 adjust += subreg_size_lowpart_offset (inherent_size, total_size);
|
|
156 x = adjust_address_nv (x, GET_MODE (regno_reg_rtx[i]), adjust);
|
|
157
|
|
158 /* Set all of the memory attributes as appropriate for a spill. */
|
|
159 set_mem_attrs_for_spill (x);
|
|
160 pseudo_slots[i].mem = x;
|
|
161 }
|
|
162
|
|
163 /* Sort pseudos according their usage frequencies. */
|
|
164 static int
|
|
165 regno_freq_compare (const void *v1p, const void *v2p)
|
|
166 {
|
|
167 const int regno1 = *(const int *) v1p;
|
|
168 const int regno2 = *(const int *) v2p;
|
|
169 int diff;
|
|
170
|
|
171 if ((diff = lra_reg_info[regno2].freq - lra_reg_info[regno1].freq) != 0)
|
|
172 return diff;
|
|
173 return regno1 - regno2;
|
|
174 }
|
|
175
|
|
176 /* Sort pseudos according to their slots, putting the slots in the order
|
|
177 that they should be allocated. Slots with lower numbers have the highest
|
|
178 priority and should get the smallest displacement from the stack or
|
|
179 frame pointer (whichever is being used).
|
|
180
|
|
181 The first allocated slot is always closest to the frame pointer,
|
|
182 so prefer lower slot numbers when frame_pointer_needed. If the stack
|
|
183 and frame grow in the same direction, then the first allocated slot is
|
|
184 always closest to the initial stack pointer and furthest away from the
|
|
185 final stack pointer, so allocate higher numbers first when using the
|
|
186 stack pointer in that case. The reverse is true if the stack and
|
|
187 frame grow in opposite directions. */
|
|
188 static int
|
|
189 pseudo_reg_slot_compare (const void *v1p, const void *v2p)
|
|
190 {
|
|
191 const int regno1 = *(const int *) v1p;
|
|
192 const int regno2 = *(const int *) v2p;
|
|
193 int diff, slot_num1, slot_num2;
|
|
194 int total_size1, total_size2;
|
|
195
|
|
196 slot_num1 = pseudo_slots[regno1].slot_num;
|
|
197 slot_num2 = pseudo_slots[regno2].slot_num;
|
|
198 if ((diff = slot_num1 - slot_num2) != 0)
|
|
199 return (frame_pointer_needed
|
|
200 || (!FRAME_GROWS_DOWNWARD) == STACK_GROWS_DOWNWARD ? diff : -diff);
|
|
201 total_size1 = GET_MODE_SIZE (lra_reg_info[regno1].biggest_mode);
|
|
202 total_size2 = GET_MODE_SIZE (lra_reg_info[regno2].biggest_mode);
|
|
203 if ((diff = total_size2 - total_size1) != 0)
|
|
204 return diff;
|
|
205 return regno1 - regno2;
|
|
206 }
|
|
207
|
|
208 /* Assign spill hard registers to N pseudos in PSEUDO_REGNOS which is
|
|
209 sorted in order of highest frequency first. Put the pseudos which
|
|
210 did not get a spill hard register at the beginning of array
|
|
211 PSEUDO_REGNOS. Return the number of such pseudos. */
|
|
212 static int
|
|
213 assign_spill_hard_regs (int *pseudo_regnos, int n)
|
|
214 {
|
|
215 int i, k, p, regno, res, spill_class_size, hard_regno, nr;
|
|
216 enum reg_class rclass, spill_class;
|
|
217 machine_mode mode;
|
|
218 lra_live_range_t r;
|
|
219 rtx_insn *insn;
|
|
220 rtx set;
|
|
221 basic_block bb;
|
|
222 HARD_REG_SET conflict_hard_regs;
|
|
223 bitmap setjump_crosses = regstat_get_setjmp_crosses ();
|
|
224 /* Hard registers which can not be used for any purpose at given
|
|
225 program point because they are unallocatable or already allocated
|
|
226 for other pseudos. */
|
|
227 HARD_REG_SET *reserved_hard_regs;
|
|
228
|
|
229 if (! lra_reg_spill_p)
|
|
230 return n;
|
|
231 /* Set up reserved hard regs for every program point. */
|
|
232 reserved_hard_regs = XNEWVEC (HARD_REG_SET, lra_live_max_point);
|
|
233 for (p = 0; p < lra_live_max_point; p++)
|
|
234 COPY_HARD_REG_SET (reserved_hard_regs[p], lra_no_alloc_regs);
|
|
235 for (i = FIRST_PSEUDO_REGISTER; i < regs_num; i++)
|
|
236 if (lra_reg_info[i].nrefs != 0
|
|
237 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
|
|
238 for (r = lra_reg_info[i].live_ranges; r != NULL; r = r->next)
|
|
239 for (p = r->start; p <= r->finish; p++)
|
|
240 add_to_hard_reg_set (&reserved_hard_regs[p],
|
|
241 lra_reg_info[i].biggest_mode, hard_regno);
|
|
242 auto_bitmap ok_insn_bitmap (®_obstack);
|
|
243 FOR_EACH_BB_FN (bb, cfun)
|
|
244 FOR_BB_INSNS (bb, insn)
|
|
245 if (DEBUG_INSN_P (insn)
|
|
246 || ((set = single_set (insn)) != NULL_RTX
|
|
247 && REG_P (SET_SRC (set)) && REG_P (SET_DEST (set))))
|
|
248 bitmap_set_bit (ok_insn_bitmap, INSN_UID (insn));
|
|
249 for (res = i = 0; i < n; i++)
|
|
250 {
|
|
251 regno = pseudo_regnos[i];
|
|
252 rclass = lra_get_allocno_class (regno);
|
|
253 if (bitmap_bit_p (setjump_crosses, regno)
|
|
254 || (spill_class
|
|
255 = ((enum reg_class)
|
|
256 targetm.spill_class ((reg_class_t) rclass,
|
|
257 PSEUDO_REGNO_MODE (regno)))) == NO_REGS
|
|
258 || bitmap_intersect_compl_p (&lra_reg_info[regno].insn_bitmap,
|
|
259 ok_insn_bitmap))
|
|
260 {
|
|
261 pseudo_regnos[res++] = regno;
|
|
262 continue;
|
|
263 }
|
|
264 lra_assert (spill_class != NO_REGS);
|
|
265 COPY_HARD_REG_SET (conflict_hard_regs,
|
|
266 lra_reg_info[regno].conflict_hard_regs);
|
|
267 for (r = lra_reg_info[regno].live_ranges; r != NULL; r = r->next)
|
|
268 for (p = r->start; p <= r->finish; p++)
|
|
269 IOR_HARD_REG_SET (conflict_hard_regs, reserved_hard_regs[p]);
|
|
270 spill_class_size = ira_class_hard_regs_num[spill_class];
|
|
271 mode = lra_reg_info[regno].biggest_mode;
|
|
272 for (k = 0; k < spill_class_size; k++)
|
|
273 {
|
|
274 hard_regno = ira_class_hard_regs[spill_class][k];
|
|
275 if (! overlaps_hard_reg_set_p (conflict_hard_regs, mode, hard_regno))
|
|
276 break;
|
|
277 }
|
|
278 if (k >= spill_class_size)
|
|
279 {
|
|
280 /* There is no available regs -- assign memory later. */
|
|
281 pseudo_regnos[res++] = regno;
|
|
282 continue;
|
|
283 }
|
|
284 if (lra_dump_file != NULL)
|
|
285 fprintf (lra_dump_file, " Spill r%d into hr%d\n", regno, hard_regno);
|
|
286 /* Update reserved_hard_regs. */
|
|
287 for (r = lra_reg_info[regno].live_ranges; r != NULL; r = r->next)
|
|
288 for (p = r->start; p <= r->finish; p++)
|
|
289 add_to_hard_reg_set (&reserved_hard_regs[p],
|
|
290 lra_reg_info[regno].biggest_mode, hard_regno);
|
|
291 spill_hard_reg[regno]
|
|
292 = gen_raw_REG (PSEUDO_REGNO_MODE (regno), hard_regno);
|
|
293 for (nr = 0;
|
|
294 nr < hard_regno_nregs (hard_regno,
|
|
295 lra_reg_info[regno].biggest_mode);
|
|
296 nr++)
|
|
297 /* Just loop. */
|
|
298 df_set_regs_ever_live (hard_regno + nr, true);
|
|
299 }
|
|
300 free (reserved_hard_regs);
|
|
301 return res;
|
|
302 }
|
|
303
|
|
304 /* Add pseudo REGNO to slot SLOT_NUM. */
|
|
305 static void
|
|
306 add_pseudo_to_slot (int regno, int slot_num)
|
|
307 {
|
|
308 struct pseudo_slot *first;
|
|
309
|
|
310 /* Each pseudo has an inherent size which comes from its own mode,
|
|
311 and a total size which provides room for paradoxical subregs.
|
|
312 We need to make sure the size and alignment of the slot are
|
|
313 sufficient for both. */
|
|
314 machine_mode mode = wider_subreg_mode (PSEUDO_REGNO_MODE (regno),
|
|
315 lra_reg_info[regno].biggest_mode);
|
|
316 unsigned int align = spill_slot_alignment (mode);
|
|
317 slots[slot_num].align = MAX (slots[slot_num].align, align);
|
|
318 slots[slot_num].size = MAX (slots[slot_num].size, GET_MODE_SIZE (mode));
|
|
319
|
|
320 if (slots[slot_num].regno < 0)
|
|
321 {
|
|
322 /* It is the first pseudo in the slot. */
|
|
323 slots[slot_num].regno = regno;
|
|
324 pseudo_slots[regno].first = &pseudo_slots[regno];
|
|
325 pseudo_slots[regno].next = NULL;
|
|
326 }
|
|
327 else
|
|
328 {
|
|
329 first = pseudo_slots[regno].first = &pseudo_slots[slots[slot_num].regno];
|
|
330 pseudo_slots[regno].next = first->next;
|
|
331 first->next = &pseudo_slots[regno];
|
|
332 }
|
|
333 pseudo_slots[regno].mem = NULL_RTX;
|
|
334 pseudo_slots[regno].slot_num = slot_num;
|
|
335 slots[slot_num].live_ranges
|
|
336 = lra_merge_live_ranges (slots[slot_num].live_ranges,
|
|
337 lra_copy_live_range_list
|
|
338 (lra_reg_info[regno].live_ranges));
|
|
339 }
|
|
340
|
|
341 /* Assign stack slot numbers to pseudos in array PSEUDO_REGNOS of
|
|
342 length N. Sort pseudos in PSEUDO_REGNOS for subsequent assigning
|
|
343 memory stack slots. */
|
|
344 static void
|
|
345 assign_stack_slot_num_and_sort_pseudos (int *pseudo_regnos, int n)
|
|
346 {
|
|
347 int i, j, regno;
|
|
348
|
|
349 slots_num = 0;
|
|
350 /* Assign stack slot numbers to spilled pseudos, use smaller numbers
|
|
351 for most frequently used pseudos. */
|
|
352 for (i = 0; i < n; i++)
|
|
353 {
|
|
354 regno = pseudo_regnos[i];
|
|
355 if (! flag_ira_share_spill_slots)
|
|
356 j = slots_num;
|
|
357 else
|
|
358 {
|
|
359 for (j = 0; j < slots_num; j++)
|
|
360 if (slots[j].hard_regno < 0
|
|
361 && ! (lra_intersected_live_ranges_p
|
|
362 (slots[j].live_ranges,
|
|
363 lra_reg_info[regno].live_ranges)))
|
|
364 break;
|
|
365 }
|
|
366 if (j >= slots_num)
|
|
367 {
|
|
368 /* New slot. */
|
|
369 slots[j].live_ranges = NULL;
|
|
370 slots[j].size = 0;
|
|
371 slots[j].align = BITS_PER_UNIT;
|
|
372 slots[j].regno = slots[j].hard_regno = -1;
|
|
373 slots[j].mem = NULL_RTX;
|
|
374 slots_num++;
|
|
375 }
|
|
376 add_pseudo_to_slot (regno, j);
|
|
377 }
|
|
378 /* Sort regnos according to their slot numbers. */
|
|
379 qsort (pseudo_regnos, n, sizeof (int), pseudo_reg_slot_compare);
|
|
380 }
|
|
381
|
|
382 /* Recursively process LOC in INSN and change spilled pseudos to the
|
|
383 corresponding memory or spilled hard reg. Ignore spilled pseudos
|
|
384 created from the scratches. Return true if the pseudo nrefs equal
|
|
385 to 0 (don't change the pseudo in this case). Otherwise return false. */
|
|
386 static bool
|
|
387 remove_pseudos (rtx *loc, rtx_insn *insn)
|
|
388 {
|
|
389 int i;
|
|
390 rtx hard_reg;
|
|
391 const char *fmt;
|
|
392 enum rtx_code code;
|
|
393 bool res = false;
|
|
394
|
|
395 if (*loc == NULL_RTX)
|
|
396 return res;
|
|
397 code = GET_CODE (*loc);
|
|
398 if (code == REG && (i = REGNO (*loc)) >= FIRST_PSEUDO_REGISTER
|
|
399 && lra_get_regno_hard_regno (i) < 0
|
|
400 /* We do not want to assign memory for former scratches because
|
|
401 it might result in an address reload for some targets. In
|
|
402 any case we transform such pseudos not getting hard registers
|
|
403 into scratches back. */
|
|
404 && ! lra_former_scratch_p (i))
|
|
405 {
|
|
406 if (lra_reg_info[i].nrefs == 0
|
|
407 && pseudo_slots[i].mem == NULL && spill_hard_reg[i] == NULL)
|
|
408 return true;
|
|
409 if ((hard_reg = spill_hard_reg[i]) != NULL_RTX)
|
|
410 *loc = copy_rtx (hard_reg);
|
|
411 else
|
|
412 {
|
|
413 rtx x = lra_eliminate_regs_1 (insn, pseudo_slots[i].mem,
|
|
414 GET_MODE (pseudo_slots[i].mem),
|
|
415 false, false, 0, true);
|
|
416 *loc = x != pseudo_slots[i].mem ? x : copy_rtx (x);
|
|
417 }
|
|
418 return res;
|
|
419 }
|
|
420
|
|
421 fmt = GET_RTX_FORMAT (code);
|
|
422 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
|
|
423 {
|
|
424 if (fmt[i] == 'e')
|
|
425 res = remove_pseudos (&XEXP (*loc, i), insn) || res;
|
|
426 else if (fmt[i] == 'E')
|
|
427 {
|
|
428 int j;
|
|
429
|
|
430 for (j = XVECLEN (*loc, i) - 1; j >= 0; j--)
|
|
431 res = remove_pseudos (&XVECEXP (*loc, i, j), insn) || res;
|
|
432 }
|
|
433 }
|
|
434 return res;
|
|
435 }
|
|
436
|
|
437 /* Convert spilled pseudos into their stack slots or spill hard regs,
|
|
438 put insns to process on the constraint stack (that is all insns in
|
|
439 which pseudos were changed to memory or spill hard regs). */
|
|
440 static void
|
|
441 spill_pseudos (void)
|
|
442 {
|
|
443 basic_block bb;
|
|
444 rtx_insn *insn, *curr;
|
|
445 int i;
|
|
446
|
|
447 auto_bitmap spilled_pseudos (®_obstack);
|
|
448 auto_bitmap changed_insns (®_obstack);
|
|
449 for (i = FIRST_PSEUDO_REGISTER; i < regs_num; i++)
|
|
450 {
|
|
451 if (lra_reg_info[i].nrefs != 0 && lra_get_regno_hard_regno (i) < 0
|
|
452 && ! lra_former_scratch_p (i))
|
|
453 {
|
|
454 bitmap_set_bit (spilled_pseudos, i);
|
|
455 bitmap_ior_into (changed_insns, &lra_reg_info[i].insn_bitmap);
|
|
456 }
|
|
457 }
|
|
458 FOR_EACH_BB_FN (bb, cfun)
|
|
459 {
|
|
460 FOR_BB_INSNS_SAFE (bb, insn, curr)
|
|
461 {
|
|
462 bool removed_pseudo_p = false;
|
|
463
|
|
464 if (bitmap_bit_p (changed_insns, INSN_UID (insn)))
|
|
465 {
|
|
466 rtx *link_loc, link;
|
|
467
|
|
468 removed_pseudo_p = remove_pseudos (&PATTERN (insn), insn);
|
|
469 if (CALL_P (insn)
|
|
470 && remove_pseudos (&CALL_INSN_FUNCTION_USAGE (insn), insn))
|
|
471 removed_pseudo_p = true;
|
|
472 for (link_loc = ®_NOTES (insn);
|
|
473 (link = *link_loc) != NULL_RTX;
|
|
474 link_loc = &XEXP (link, 1))
|
|
475 {
|
|
476 switch (REG_NOTE_KIND (link))
|
|
477 {
|
|
478 case REG_FRAME_RELATED_EXPR:
|
|
479 case REG_CFA_DEF_CFA:
|
|
480 case REG_CFA_ADJUST_CFA:
|
|
481 case REG_CFA_OFFSET:
|
|
482 case REG_CFA_REGISTER:
|
|
483 case REG_CFA_EXPRESSION:
|
|
484 case REG_CFA_RESTORE:
|
|
485 case REG_CFA_SET_VDRAP:
|
|
486 if (remove_pseudos (&XEXP (link, 0), insn))
|
|
487 removed_pseudo_p = true;
|
|
488 break;
|
|
489 default:
|
|
490 break;
|
|
491 }
|
|
492 }
|
|
493 if (lra_dump_file != NULL)
|
|
494 fprintf (lra_dump_file,
|
|
495 "Changing spilled pseudos to memory in insn #%u\n",
|
|
496 INSN_UID (insn));
|
|
497 lra_push_insn (insn);
|
|
498 if (lra_reg_spill_p || targetm.different_addr_displacement_p ())
|
|
499 lra_set_used_insn_alternative (insn, -1);
|
|
500 }
|
|
501 else if (CALL_P (insn)
|
|
502 /* Presence of any pseudo in CALL_INSN_FUNCTION_USAGE
|
|
503 does not affect value of insn_bitmap of the
|
|
504 corresponding lra_reg_info. That is because we
|
|
505 don't need to reload pseudos in
|
|
506 CALL_INSN_FUNCTION_USAGEs. So if we process only
|
|
507 insns in the insn_bitmap of given pseudo here, we
|
|
508 can miss the pseudo in some
|
|
509 CALL_INSN_FUNCTION_USAGEs. */
|
|
510 && remove_pseudos (&CALL_INSN_FUNCTION_USAGE (insn), insn))
|
|
511 removed_pseudo_p = true;
|
|
512 if (removed_pseudo_p)
|
|
513 {
|
|
514 lra_assert (DEBUG_INSN_P (insn));
|
|
515 lra_invalidate_insn_data (insn);
|
|
516 INSN_VAR_LOCATION_LOC (insn) = gen_rtx_UNKNOWN_VAR_LOC ();
|
|
517 if (lra_dump_file != NULL)
|
|
518 fprintf (lra_dump_file,
|
|
519 "Debug insn #%u is reset because it referenced "
|
|
520 "removed pseudo\n", INSN_UID (insn));
|
|
521 }
|
|
522 bitmap_and_compl_into (df_get_live_in (bb), spilled_pseudos);
|
|
523 bitmap_and_compl_into (df_get_live_out (bb), spilled_pseudos);
|
|
524 }
|
|
525 }
|
|
526 }
|
|
527
|
|
528 /* Return true if we need to change some pseudos into memory. */
|
|
529 bool
|
|
530 lra_need_for_spills_p (void)
|
|
531 {
|
|
532 int i; max_regno = max_reg_num ();
|
|
533
|
|
534 for (i = FIRST_PSEUDO_REGISTER; i < max_regno; i++)
|
|
535 if (lra_reg_info[i].nrefs != 0 && lra_get_regno_hard_regno (i) < 0
|
|
536 && ! lra_former_scratch_p (i))
|
|
537 return true;
|
|
538 return false;
|
|
539 }
|
|
540
|
|
541 /* Change spilled pseudos into memory or spill hard regs. Put changed
|
|
542 insns on the constraint stack (these insns will be considered on
|
|
543 the next constraint pass). The changed insns are all insns in
|
|
544 which pseudos were changed. */
|
|
545 void
|
|
546 lra_spill (void)
|
|
547 {
|
|
548 int i, n, curr_regno;
|
|
549 int *pseudo_regnos;
|
|
550
|
|
551 regs_num = max_reg_num ();
|
|
552 spill_hard_reg = XNEWVEC (rtx, regs_num);
|
|
553 pseudo_regnos = XNEWVEC (int, regs_num);
|
|
554 for (n = 0, i = FIRST_PSEUDO_REGISTER; i < regs_num; i++)
|
|
555 if (lra_reg_info[i].nrefs != 0 && lra_get_regno_hard_regno (i) < 0
|
|
556 /* We do not want to assign memory for former scratches. */
|
|
557 && ! lra_former_scratch_p (i))
|
|
558 pseudo_regnos[n++] = i;
|
|
559 lra_assert (n > 0);
|
|
560 pseudo_slots = XNEWVEC (struct pseudo_slot, regs_num);
|
|
561 for (i = FIRST_PSEUDO_REGISTER; i < regs_num; i++)
|
|
562 {
|
|
563 spill_hard_reg[i] = NULL_RTX;
|
|
564 pseudo_slots[i].mem = NULL_RTX;
|
|
565 }
|
|
566 slots = XNEWVEC (struct slot, regs_num);
|
|
567 /* Sort regnos according their usage frequencies. */
|
|
568 qsort (pseudo_regnos, n, sizeof (int), regno_freq_compare);
|
|
569 n = assign_spill_hard_regs (pseudo_regnos, n);
|
|
570 assign_stack_slot_num_and_sort_pseudos (pseudo_regnos, n);
|
|
571 for (i = 0; i < n; i++)
|
|
572 if (pseudo_slots[pseudo_regnos[i]].mem == NULL_RTX)
|
|
573 assign_mem_slot (pseudo_regnos[i]);
|
|
574 if (n > 0 && crtl->stack_alignment_needed)
|
|
575 /* If we have a stack frame, we must align it now. The stack size
|
|
576 may be a part of the offset computation for register
|
|
577 elimination. */
|
|
578 assign_stack_local (BLKmode, 0, crtl->stack_alignment_needed);
|
|
579 if (lra_dump_file != NULL)
|
|
580 {
|
|
581 for (i = 0; i < slots_num; i++)
|
|
582 {
|
|
583 fprintf (lra_dump_file, " Slot %d regnos (width = %d):", i,
|
|
584 GET_MODE_SIZE (GET_MODE (slots[i].mem)));
|
|
585 for (curr_regno = slots[i].regno;;
|
|
586 curr_regno = pseudo_slots[curr_regno].next - pseudo_slots)
|
|
587 {
|
|
588 fprintf (lra_dump_file, " %d", curr_regno);
|
|
589 if (pseudo_slots[curr_regno].next == NULL)
|
|
590 break;
|
|
591 }
|
|
592 fprintf (lra_dump_file, "\n");
|
|
593 }
|
|
594 }
|
|
595 spill_pseudos ();
|
|
596 free (slots);
|
|
597 free (pseudo_slots);
|
|
598 free (pseudo_regnos);
|
|
599 free (spill_hard_reg);
|
|
600 }
|
|
601
|
|
602 /* Apply alter_subreg for subregs of regs in *LOC. Use FINAL_P for
|
|
603 alter_subreg calls. Return true if any subreg of reg is
|
|
604 processed. */
|
|
605 static bool
|
|
606 alter_subregs (rtx *loc, bool final_p)
|
|
607 {
|
|
608 int i;
|
|
609 rtx x = *loc;
|
|
610 bool res;
|
|
611 const char *fmt;
|
|
612 enum rtx_code code;
|
|
613
|
|
614 if (x == NULL_RTX)
|
|
615 return false;
|
|
616 code = GET_CODE (x);
|
|
617 if (code == SUBREG && REG_P (SUBREG_REG (x)))
|
|
618 {
|
|
619 lra_assert (REGNO (SUBREG_REG (x)) < FIRST_PSEUDO_REGISTER);
|
|
620 alter_subreg (loc, final_p);
|
|
621 return true;
|
|
622 }
|
|
623 fmt = GET_RTX_FORMAT (code);
|
|
624 res = false;
|
|
625 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
|
|
626 {
|
|
627 if (fmt[i] == 'e')
|
|
628 {
|
|
629 if (alter_subregs (&XEXP (x, i), final_p))
|
|
630 res = true;
|
|
631 }
|
|
632 else if (fmt[i] == 'E')
|
|
633 {
|
|
634 int j;
|
|
635
|
|
636 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
|
|
637 if (alter_subregs (&XVECEXP (x, i, j), final_p))
|
|
638 res = true;
|
|
639 }
|
|
640 }
|
|
641 return res;
|
|
642 }
|
|
643
|
|
644 /* Return true if REGNO is used for return in the current
|
|
645 function. */
|
|
646 static bool
|
|
647 return_regno_p (unsigned int regno)
|
|
648 {
|
|
649 rtx outgoing = crtl->return_rtx;
|
|
650
|
|
651 if (! outgoing)
|
|
652 return false;
|
|
653
|
|
654 if (REG_P (outgoing))
|
|
655 return REGNO (outgoing) == regno;
|
|
656 else if (GET_CODE (outgoing) == PARALLEL)
|
|
657 {
|
|
658 int i;
|
|
659
|
|
660 for (i = 0; i < XVECLEN (outgoing, 0); i++)
|
|
661 {
|
|
662 rtx x = XEXP (XVECEXP (outgoing, 0, i), 0);
|
|
663
|
|
664 if (REG_P (x) && REGNO (x) == regno)
|
|
665 return true;
|
|
666 }
|
|
667 }
|
|
668 return false;
|
|
669 }
|
|
670
|
|
671 /* Return true if REGNO is in one of subsequent USE after INSN in the
|
|
672 same BB. */
|
|
673 static bool
|
|
674 regno_in_use_p (rtx_insn *insn, unsigned int regno)
|
|
675 {
|
|
676 static lra_insn_recog_data_t id;
|
|
677 static struct lra_static_insn_data *static_id;
|
|
678 struct lra_insn_reg *reg;
|
|
679 int i, arg_regno;
|
|
680 basic_block bb = BLOCK_FOR_INSN (insn);
|
|
681
|
|
682 while ((insn = next_nondebug_insn (insn)) != NULL_RTX)
|
|
683 {
|
|
684 if (BARRIER_P (insn) || bb != BLOCK_FOR_INSN (insn))
|
|
685 return false;
|
|
686 if (! INSN_P (insn))
|
|
687 continue;
|
|
688 if (GET_CODE (PATTERN (insn)) == USE
|
|
689 && REG_P (XEXP (PATTERN (insn), 0))
|
|
690 && regno == REGNO (XEXP (PATTERN (insn), 0)))
|
|
691 return true;
|
|
692 /* Check that the regno is not modified. */
|
|
693 id = lra_get_insn_recog_data (insn);
|
|
694 for (reg = id->regs; reg != NULL; reg = reg->next)
|
|
695 if (reg->type != OP_IN && reg->regno == (int) regno)
|
|
696 return false;
|
|
697 static_id = id->insn_static_data;
|
|
698 for (reg = static_id->hard_regs; reg != NULL; reg = reg->next)
|
|
699 if (reg->type != OP_IN && reg->regno == (int) regno)
|
|
700 return false;
|
|
701 if (id->arg_hard_regs != NULL)
|
|
702 for (i = 0; (arg_regno = id->arg_hard_regs[i]) >= 0; i++)
|
|
703 if ((int) regno == (arg_regno >= FIRST_PSEUDO_REGISTER
|
|
704 ? arg_regno : arg_regno - FIRST_PSEUDO_REGISTER))
|
|
705 return false;
|
|
706 }
|
|
707 return false;
|
|
708 }
|
|
709
|
|
710 /* Final change of pseudos got hard registers into the corresponding
|
|
711 hard registers and removing temporary clobbers. */
|
|
712 void
|
|
713 lra_final_code_change (void)
|
|
714 {
|
|
715 int i, hard_regno;
|
|
716 basic_block bb;
|
|
717 rtx_insn *insn, *curr;
|
|
718 int max_regno = max_reg_num ();
|
|
719
|
|
720 for (i = FIRST_PSEUDO_REGISTER; i < max_regno; i++)
|
|
721 if (lra_reg_info[i].nrefs != 0
|
|
722 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
|
|
723 SET_REGNO (regno_reg_rtx[i], hard_regno);
|
|
724 FOR_EACH_BB_FN (bb, cfun)
|
|
725 FOR_BB_INSNS_SAFE (bb, insn, curr)
|
|
726 if (INSN_P (insn))
|
|
727 {
|
|
728 rtx pat = PATTERN (insn);
|
|
729
|
|
730 if (GET_CODE (pat) == CLOBBER && LRA_TEMP_CLOBBER_P (pat))
|
|
731 {
|
|
732 /* Remove clobbers temporarily created in LRA. We don't
|
|
733 need them anymore and don't want to waste compiler
|
|
734 time processing them in a few subsequent passes. */
|
|
735 lra_invalidate_insn_data (insn);
|
|
736 delete_insn (insn);
|
|
737 continue;
|
|
738 }
|
|
739
|
|
740 /* IRA can generate move insns involving pseudos. It is
|
|
741 better remove them earlier to speed up compiler a bit.
|
|
742 It is also better to do it here as they might not pass
|
|
743 final RTL check in LRA, (e.g. insn moving a control
|
|
744 register into itself). So remove an useless move insn
|
|
745 unless next insn is USE marking the return reg (we should
|
|
746 save this as some subsequent optimizations assume that
|
|
747 such original insns are saved). */
|
|
748 if (NONJUMP_INSN_P (insn) && GET_CODE (pat) == SET
|
|
749 && REG_P (SET_SRC (pat)) && REG_P (SET_DEST (pat))
|
|
750 && REGNO (SET_SRC (pat)) == REGNO (SET_DEST (pat))
|
|
751 && (! return_regno_p (REGNO (SET_SRC (pat)))
|
|
752 || ! regno_in_use_p (insn, REGNO (SET_SRC (pat)))))
|
|
753 {
|
|
754 lra_invalidate_insn_data (insn);
|
|
755 delete_insn (insn);
|
|
756 continue;
|
|
757 }
|
|
758
|
|
759 lra_insn_recog_data_t id = lra_get_insn_recog_data (insn);
|
|
760 struct lra_insn_reg *reg;
|
|
761
|
|
762 for (reg = id->regs; reg != NULL; reg = reg->next)
|
|
763 if (reg->regno >= FIRST_PSEUDO_REGISTER
|
|
764 && lra_reg_info [reg->regno].nrefs == 0)
|
|
765 break;
|
|
766
|
|
767 if (reg != NULL)
|
|
768 {
|
|
769 /* Pseudos still can be in debug insns in some very rare
|
|
770 and complicated cases, e.g. the pseudo was removed by
|
|
771 inheritance and the debug insn is not EBBs where the
|
|
772 inheritance happened. It is difficult and time
|
|
773 consuming to find what hard register corresponds the
|
|
774 pseudo -- so just remove the debug insn. Another
|
|
775 solution could be assigning hard reg/memory but it
|
|
776 would be a misleading info. It is better not to have
|
|
777 info than have it wrong. */
|
|
778 lra_assert (DEBUG_INSN_P (insn));
|
|
779 lra_invalidate_insn_data (insn);
|
|
780 delete_insn (insn);
|
|
781 continue;
|
|
782 }
|
|
783
|
|
784 struct lra_static_insn_data *static_id = id->insn_static_data;
|
|
785 bool insn_change_p = false;
|
|
786
|
|
787 for (i = id->insn_static_data->n_operands - 1; i >= 0; i--)
|
|
788 if ((DEBUG_INSN_P (insn) || ! static_id->operand[i].is_operator)
|
|
789 && alter_subregs (id->operand_loc[i], ! DEBUG_INSN_P (insn)))
|
|
790 {
|
|
791 lra_update_dup (id, i);
|
|
792 insn_change_p = true;
|
|
793 }
|
|
794 if (insn_change_p)
|
|
795 lra_update_operator_dups (id);
|
|
796 }
|
|
797 }
|