1 /* Change pseudos by memory.
2    Copyright (C) 2010-2014 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 "tm.h"
62 #include "rtl.h"
63 #include "tm_p.h"
64 #include "insn-config.h"
65 #include "recog.h"
66 #include "output.h"
67 #include "regs.h"
68 #include "hard-reg-set.h"
69 #include "flags.h"
70 #include "function.h"
71 #include "expr.h"
72 #include "basic-block.h"
73 #include "except.h"
74 #include "timevar.h"
75 #include "target.h"
76 #include "lra-int.h"
77 #include "ira.h"
78 #include "df.h"
79 
80 
81 /* Max regno at the start of the pass.	*/
82 static int regs_num;
83 
84 /* Map spilled regno -> hard regno used instead of memory for
85    spilling.  */
86 static rtx *spill_hard_reg;
87 
88 /* The structure describes stack slot of a spilled pseudo.  */
89 struct pseudo_slot
90 {
91   /* Number (0, 1, ...) of the stack slot to which given pseudo
92      belongs.  */
93   int slot_num;
94   /* First or next slot with the same slot number.  */
95   struct pseudo_slot *next, *first;
96   /* Memory representing the spilled pseudo.  */
97   rtx mem;
98 };
99 
100 /* The stack slots for each spilled pseudo.  Indexed by regnos.	 */
101 static struct pseudo_slot *pseudo_slots;
102 
103 /* The structure describes a register or a stack slot which can be
104    used for several spilled pseudos.  */
105 struct slot
106 {
107   /* First pseudo with given stack slot.  */
108   int regno;
109   /* Hard reg into which the slot pseudos are spilled.	The value is
110      negative for pseudos spilled into memory.	*/
111   int hard_regno;
112   /* Memory representing the all stack slot.  It can be different from
113      memory representing a pseudo belonging to give stack slot because
114      pseudo can be placed in a part of the corresponding stack slot.
115      The value is NULL for pseudos spilled into a hard reg.  */
116   rtx mem;
117   /* Combined live ranges of all pseudos belonging to given slot.  It
118      is used to figure out that a new spilled pseudo can use given
119      stack slot.  */
120   lra_live_range_t live_ranges;
121 };
122 
123 /* Array containing info about the stack slots.	 The array element is
124    indexed by the stack slot number in the range [0..slots_num).  */
125 static struct slot *slots;
126 /* The number of the stack slots currently existing.  */
127 static int slots_num;
128 
129 /* Set up memory of the spilled pseudo I.  The function can allocate
130    the corresponding stack slot if it is not done yet.	*/
131 static void
assign_mem_slot(int i)132 assign_mem_slot (int i)
133 {
134   rtx x = NULL_RTX;
135   enum machine_mode mode = GET_MODE (regno_reg_rtx[i]);
136   unsigned int inherent_size = PSEUDO_REGNO_BYTES (i);
137   unsigned int inherent_align = GET_MODE_ALIGNMENT (mode);
138   unsigned int max_ref_width = GET_MODE_SIZE (lra_reg_info[i].biggest_mode);
139   unsigned int total_size = MAX (inherent_size, max_ref_width);
140   unsigned int min_align = max_ref_width * BITS_PER_UNIT;
141   int adjust = 0;
142 
143   lra_assert (regno_reg_rtx[i] != NULL_RTX && REG_P (regno_reg_rtx[i])
144 	      && lra_reg_info[i].nrefs != 0 && reg_renumber[i] < 0);
145 
146   x = slots[pseudo_slots[i].slot_num].mem;
147 
148   /* We can use a slot already allocated because it is guaranteed the
149      slot provides both enough inherent space and enough total
150      space.  */
151   if (x)
152     ;
153   /* Each pseudo has an inherent size which comes from its own mode,
154      and a total size which provides room for paradoxical subregs
155      which refer to the pseudo reg in wider modes.  We allocate a new
156      slot, making sure that it has enough inherent space and total
157      space.  */
158   else
159     {
160       rtx stack_slot;
161 
162       /* No known place to spill from => no slot to reuse.  */
163       x = assign_stack_local (mode, total_size,
164 			      min_align > inherent_align
165 			      || total_size > inherent_size ? -1 : 0);
166       stack_slot = x;
167       /* Cancel the big-endian correction done in assign_stack_local.
168 	 Get the address of the beginning of the slot.	This is so we
169 	 can do a big-endian correction unconditionally below.	*/
170       if (BYTES_BIG_ENDIAN)
171 	{
172 	  adjust = inherent_size - total_size;
173 	  if (adjust)
174 	    stack_slot
175 	      = adjust_address_nv (x,
176 				   mode_for_size (total_size * BITS_PER_UNIT,
177 						  MODE_INT, 1),
178 				   adjust);
179 	}
180       slots[pseudo_slots[i].slot_num].mem = stack_slot;
181     }
182 
183   /* On a big endian machine, the "address" of the slot is the address
184      of the low part that fits its inherent mode.  */
185   if (BYTES_BIG_ENDIAN && inherent_size < total_size)
186     adjust += (total_size - inherent_size);
187 
188   x = adjust_address_nv (x, GET_MODE (regno_reg_rtx[i]), adjust);
189 
190   /* Set all of the memory attributes as appropriate for a spill.  */
191   set_mem_attrs_for_spill (x);
192   pseudo_slots[i].mem = x;
193 }
194 
195 /* Sort pseudos according their usage frequencies.  */
196 static int
regno_freq_compare(const void * v1p,const void * v2p)197 regno_freq_compare (const void *v1p, const void *v2p)
198 {
199   const int regno1 = *(const int *) v1p;
200   const int regno2 = *(const int *) v2p;
201   int diff;
202 
203   if ((diff = lra_reg_info[regno2].freq - lra_reg_info[regno1].freq) != 0)
204     return diff;
205   return regno1 - regno2;
206 }
207 
208 /* Redefine STACK_GROWS_DOWNWARD in terms of 0 or 1.  */
209 #ifdef STACK_GROWS_DOWNWARD
210 # undef STACK_GROWS_DOWNWARD
211 # define STACK_GROWS_DOWNWARD 1
212 #else
213 # define STACK_GROWS_DOWNWARD 0
214 #endif
215 
216 /* Sort pseudos according to their slots, putting the slots in the order
217    that they should be allocated.  Slots with lower numbers have the highest
218    priority and should get the smallest displacement from the stack or
219    frame pointer (whichever is being used).
220 
221    The first allocated slot is always closest to the frame pointer,
222    so prefer lower slot numbers when frame_pointer_needed.  If the stack
223    and frame grow in the same direction, then the first allocated slot is
224    always closest to the initial stack pointer and furthest away from the
225    final stack pointer, so allocate higher numbers first when using the
226    stack pointer in that case.  The reverse is true if the stack and
227    frame grow in opposite directions.  */
228 static int
pseudo_reg_slot_compare(const void * v1p,const void * v2p)229 pseudo_reg_slot_compare (const void *v1p, const void *v2p)
230 {
231   const int regno1 = *(const int *) v1p;
232   const int regno2 = *(const int *) v2p;
233   int diff, slot_num1, slot_num2;
234   int total_size1, total_size2;
235 
236   slot_num1 = pseudo_slots[regno1].slot_num;
237   slot_num2 = pseudo_slots[regno2].slot_num;
238   if ((diff = slot_num1 - slot_num2) != 0)
239     return (frame_pointer_needed
240 	    || !FRAME_GROWS_DOWNWARD == STACK_GROWS_DOWNWARD ? diff : -diff);
241   total_size1 = GET_MODE_SIZE (lra_reg_info[regno1].biggest_mode);
242   total_size2 = GET_MODE_SIZE (lra_reg_info[regno2].biggest_mode);
243   if ((diff = total_size2 - total_size1) != 0)
244     return diff;
245   return regno1 - regno2;
246 }
247 
248 /* Assign spill hard registers to N pseudos in PSEUDO_REGNOS which is
249    sorted in order of highest frequency first.  Put the pseudos which
250    did not get a spill hard register at the beginning of array
251    PSEUDO_REGNOS.  Return the number of such pseudos.  */
252 static int
assign_spill_hard_regs(int * pseudo_regnos,int n)253 assign_spill_hard_regs (int *pseudo_regnos, int n)
254 {
255   int i, k, p, regno, res, spill_class_size, hard_regno, nr;
256   enum reg_class rclass, spill_class;
257   enum machine_mode mode;
258   lra_live_range_t r;
259   rtx insn, set;
260   basic_block bb;
261   HARD_REG_SET conflict_hard_regs;
262   bitmap_head ok_insn_bitmap;
263   bitmap setjump_crosses = regstat_get_setjmp_crosses ();
264   /* Hard registers which can not be used for any purpose at given
265      program point because they are unallocatable or already allocated
266      for other pseudos.	 */
267   HARD_REG_SET *reserved_hard_regs;
268 
269   if (! lra_reg_spill_p)
270     return n;
271   /* Set up reserved hard regs for every program point.	 */
272   reserved_hard_regs = XNEWVEC (HARD_REG_SET, lra_live_max_point);
273   for (p = 0; p < lra_live_max_point; p++)
274     COPY_HARD_REG_SET (reserved_hard_regs[p], lra_no_alloc_regs);
275   for (i = FIRST_PSEUDO_REGISTER; i < regs_num; i++)
276     if (lra_reg_info[i].nrefs != 0
277 	&& (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
278       for (r = lra_reg_info[i].live_ranges; r != NULL; r = r->next)
279 	for (p = r->start; p <= r->finish; p++)
280 	  add_to_hard_reg_set (&reserved_hard_regs[p],
281 			       lra_reg_info[i].biggest_mode, hard_regno);
282   bitmap_initialize (&ok_insn_bitmap, &reg_obstack);
283   FOR_EACH_BB_FN (bb, cfun)
284     FOR_BB_INSNS (bb, insn)
285       if (DEBUG_INSN_P (insn)
286 	  || ((set = single_set (insn)) != NULL_RTX
287 	      && REG_P (SET_SRC (set)) && REG_P (SET_DEST (set))))
288 	bitmap_set_bit (&ok_insn_bitmap, INSN_UID (insn));
289   for (res = i = 0; i < n; i++)
290     {
291       regno = pseudo_regnos[i];
292       rclass = lra_get_allocno_class (regno);
293       if (bitmap_bit_p (setjump_crosses, regno)
294 	  || (spill_class
295 	      = ((enum reg_class)
296 		 targetm.spill_class ((reg_class_t) rclass,
297 				      PSEUDO_REGNO_MODE (regno)))) == NO_REGS
298 	  || bitmap_intersect_compl_p (&lra_reg_info[regno].insn_bitmap,
299 				       &ok_insn_bitmap))
300 	{
301 	  pseudo_regnos[res++] = regno;
302 	  continue;
303 	}
304       lra_assert (spill_class != NO_REGS);
305       COPY_HARD_REG_SET (conflict_hard_regs,
306 			 lra_reg_info[regno].conflict_hard_regs);
307       for (r = lra_reg_info[regno].live_ranges; r != NULL; r = r->next)
308 	for (p = r->start; p <= r->finish; p++)
309 	  IOR_HARD_REG_SET (conflict_hard_regs, reserved_hard_regs[p]);
310       spill_class_size = ira_class_hard_regs_num[spill_class];
311       mode = lra_reg_info[regno].biggest_mode;
312       for (k = 0; k < spill_class_size; k++)
313 	{
314 	  hard_regno = ira_class_hard_regs[spill_class][k];
315 	  if (! overlaps_hard_reg_set_p (conflict_hard_regs, mode, hard_regno))
316 	    break;
317 	}
318       if (k >= spill_class_size)
319 	{
320 	   /* There is no available regs -- assign memory later.  */
321 	  pseudo_regnos[res++] = regno;
322 	  continue;
323 	}
324       if (lra_dump_file != NULL)
325 	fprintf (lra_dump_file, "  Spill r%d into hr%d\n", regno, hard_regno);
326       /* Update reserved_hard_regs.  */
327       for (r = lra_reg_info[regno].live_ranges; r != NULL; r = r->next)
328 	for (p = r->start; p <= r->finish; p++)
329 	  add_to_hard_reg_set (&reserved_hard_regs[p],
330 			       lra_reg_info[regno].biggest_mode, hard_regno);
331       spill_hard_reg[regno]
332 	= gen_raw_REG (PSEUDO_REGNO_MODE (regno), hard_regno);
333       for (nr = 0;
334 	   nr < hard_regno_nregs[hard_regno][lra_reg_info[regno].biggest_mode];
335 	   nr++)
336 	/* Just loop.  */
337 	df_set_regs_ever_live (hard_regno + nr, true);
338     }
339   bitmap_clear (&ok_insn_bitmap);
340   free (reserved_hard_regs);
341   return res;
342 }
343 
344 /* Add pseudo REGNO to slot SLOT_NUM.  */
345 static void
add_pseudo_to_slot(int regno,int slot_num)346 add_pseudo_to_slot (int regno, int slot_num)
347 {
348   struct pseudo_slot *first;
349 
350   if (slots[slot_num].regno < 0)
351     {
352       /* It is the first pseudo in the slot.  */
353       slots[slot_num].regno = regno;
354       pseudo_slots[regno].first = &pseudo_slots[regno];
355       pseudo_slots[regno].next = NULL;
356     }
357   else
358     {
359       first = pseudo_slots[regno].first = &pseudo_slots[slots[slot_num].regno];
360       pseudo_slots[regno].next = first->next;
361       first->next = &pseudo_slots[regno];
362     }
363   pseudo_slots[regno].mem = NULL_RTX;
364   pseudo_slots[regno].slot_num = slot_num;
365   slots[slot_num].live_ranges
366     = lra_merge_live_ranges (slots[slot_num].live_ranges,
367 			     lra_copy_live_range_list
368 			     (lra_reg_info[regno].live_ranges));
369 }
370 
371 /* Assign stack slot numbers to pseudos in array PSEUDO_REGNOS of
372    length N.  Sort pseudos in PSEUDO_REGNOS for subsequent assigning
373    memory stack slots.	*/
374 static void
assign_stack_slot_num_and_sort_pseudos(int * pseudo_regnos,int n)375 assign_stack_slot_num_and_sort_pseudos (int *pseudo_regnos, int n)
376 {
377   int i, j, regno;
378 
379   slots_num = 0;
380   /* Assign stack slot numbers to spilled pseudos, use smaller numbers
381      for most frequently used pseudos.	*/
382   for (i = 0; i < n; i++)
383     {
384       regno = pseudo_regnos[i];
385       if (! flag_ira_share_spill_slots)
386 	j = slots_num;
387       else
388 	{
389 	  for (j = 0; j < slots_num; j++)
390 	    if (slots[j].hard_regno < 0
391 		&& ! (lra_intersected_live_ranges_p
392 		      (slots[j].live_ranges,
393 		       lra_reg_info[regno].live_ranges)))
394 	      break;
395 	}
396       if (j >= slots_num)
397 	{
398 	  /* New slot.	*/
399 	  slots[j].live_ranges = NULL;
400 	  slots[j].regno = slots[j].hard_regno = -1;
401 	  slots[j].mem = NULL_RTX;
402 	  slots_num++;
403 	}
404       add_pseudo_to_slot (regno, j);
405     }
406   /* Sort regnos according to their slot numbers.  */
407   qsort (pseudo_regnos, n, sizeof (int), pseudo_reg_slot_compare);
408 }
409 
410 /* Recursively process LOC in INSN and change spilled pseudos to the
411    corresponding memory or spilled hard reg.  Ignore spilled pseudos
412    created from the scratches.	*/
413 static void
remove_pseudos(rtx * loc,rtx insn)414 remove_pseudos (rtx *loc, rtx insn)
415 {
416   int i;
417   rtx hard_reg;
418   const char *fmt;
419   enum rtx_code code;
420 
421   if (*loc == NULL_RTX)
422     return;
423   code = GET_CODE (*loc);
424   if (code == REG && (i = REGNO (*loc)) >= FIRST_PSEUDO_REGISTER
425       && lra_get_regno_hard_regno (i) < 0
426       /* We do not want to assign memory for former scratches because
427 	 it might result in an address reload for some targets.	 In
428 	 any case we transform such pseudos not getting hard registers
429 	 into scratches back.  */
430       && ! lra_former_scratch_p (i))
431     {
432       if ((hard_reg = spill_hard_reg[i]) != NULL_RTX)
433 	*loc = copy_rtx (hard_reg);
434       else
435 	{
436 	  rtx x = lra_eliminate_regs_1 (insn, pseudo_slots[i].mem,
437 					GET_MODE (pseudo_slots[i].mem),
438 					false, false, true);
439 	  *loc = x != pseudo_slots[i].mem ? x : copy_rtx (x);
440 	}
441       return;
442     }
443 
444   fmt = GET_RTX_FORMAT (code);
445   for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
446     {
447       if (fmt[i] == 'e')
448 	remove_pseudos (&XEXP (*loc, i), insn);
449       else if (fmt[i] == 'E')
450 	{
451 	  int j;
452 
453 	  for (j = XVECLEN (*loc, i) - 1; j >= 0; j--)
454 	    remove_pseudos (&XVECEXP (*loc, i, j), insn);
455 	}
456     }
457 }
458 
459 /* Convert spilled pseudos into their stack slots or spill hard regs,
460    put insns to process on the constraint stack (that is all insns in
461    which pseudos were changed to memory or spill hard regs).   */
462 static void
spill_pseudos(void)463 spill_pseudos (void)
464 {
465   basic_block bb;
466   rtx insn;
467   int i;
468   bitmap_head spilled_pseudos, changed_insns;
469 
470   bitmap_initialize (&spilled_pseudos, &reg_obstack);
471   bitmap_initialize (&changed_insns, &reg_obstack);
472   for (i = FIRST_PSEUDO_REGISTER; i < regs_num; i++)
473     {
474       if (lra_reg_info[i].nrefs != 0 && lra_get_regno_hard_regno (i) < 0
475 	  && ! lra_former_scratch_p (i))
476 	{
477 	  bitmap_set_bit (&spilled_pseudos, i);
478 	  bitmap_ior_into (&changed_insns, &lra_reg_info[i].insn_bitmap);
479 	}
480     }
481   FOR_EACH_BB_FN (bb, cfun)
482     {
483       FOR_BB_INSNS (bb, insn)
484 	if (bitmap_bit_p (&changed_insns, INSN_UID (insn)))
485 	  {
486 	    rtx *link_loc, link;
487 	    remove_pseudos (&PATTERN (insn), insn);
488 	    if (CALL_P (insn))
489 	      remove_pseudos (&CALL_INSN_FUNCTION_USAGE (insn), insn);
490 	    for (link_loc = &REG_NOTES (insn);
491 		 (link = *link_loc) != NULL_RTX;
492 		 link_loc = &XEXP (link, 1))
493 	      {
494 		switch (REG_NOTE_KIND (link))
495 		  {
496 		  case REG_FRAME_RELATED_EXPR:
497 		  case REG_CFA_DEF_CFA:
498 		  case REG_CFA_ADJUST_CFA:
499 		  case REG_CFA_OFFSET:
500 		  case REG_CFA_REGISTER:
501 		  case REG_CFA_EXPRESSION:
502 		  case REG_CFA_RESTORE:
503 		  case REG_CFA_SET_VDRAP:
504 		    remove_pseudos (&XEXP (link, 0), insn);
505 		    break;
506 		  default:
507 		    break;
508 		  }
509 	      }
510 	    if (lra_dump_file != NULL)
511 	      fprintf (lra_dump_file,
512 		       "Changing spilled pseudos to memory in insn #%u\n",
513 		       INSN_UID (insn));
514 	    lra_push_insn (insn);
515 	    if (lra_reg_spill_p || targetm.different_addr_displacement_p ())
516 	      lra_set_used_insn_alternative (insn, -1);
517 	  }
518 	else if (CALL_P (insn))
519 	  /* Presence of any pseudo in CALL_INSN_FUNCTION_USAGE does
520 	     not affect value of insn_bitmap of the corresponding
521 	     lra_reg_info.  That is because we don't need to reload
522 	     pseudos in CALL_INSN_FUNCTION_USAGEs.  So if we process
523 	     only insns in the insn_bitmap of given pseudo here, we
524 	     can miss the pseudo in some
525 	     CALL_INSN_FUNCTION_USAGEs.  */
526 	  remove_pseudos (&CALL_INSN_FUNCTION_USAGE (insn), insn);
527       bitmap_and_compl_into (df_get_live_in (bb), &spilled_pseudos);
528       bitmap_and_compl_into (df_get_live_out (bb), &spilled_pseudos);
529     }
530   bitmap_clear (&spilled_pseudos);
531   bitmap_clear (&changed_insns);
532 }
533 
534 /* Return true if we need to change some pseudos into memory.  */
535 bool
lra_need_for_spills_p(void)536 lra_need_for_spills_p (void)
537 {
538   int i; max_regno = max_reg_num ();
539 
540   for (i = FIRST_PSEUDO_REGISTER; i < max_regno; i++)
541     if (lra_reg_info[i].nrefs != 0 && lra_get_regno_hard_regno (i) < 0
542 	&& ! lra_former_scratch_p (i))
543       return true;
544   return false;
545 }
546 
547 /* Change spilled pseudos into memory or spill hard regs.  Put changed
548    insns on the constraint stack (these insns will be considered on
549    the next constraint pass).  The changed insns are all insns in
550    which pseudos were changed.  */
551 void
lra_spill(void)552 lra_spill (void)
553 {
554   int i, n, curr_regno;
555   int *pseudo_regnos;
556 
557   regs_num = max_reg_num ();
558   spill_hard_reg = XNEWVEC (rtx, regs_num);
559   pseudo_regnos = XNEWVEC (int, regs_num);
560   for (n = 0, i = FIRST_PSEUDO_REGISTER; i < regs_num; i++)
561     if (lra_reg_info[i].nrefs != 0 && lra_get_regno_hard_regno (i) < 0
562 	/* We do not want to assign memory for former scratches.  */
563 	&& ! lra_former_scratch_p (i))
564       {
565 	spill_hard_reg[i] = NULL_RTX;
566 	pseudo_regnos[n++] = i;
567       }
568   lra_assert (n > 0);
569   pseudo_slots = XNEWVEC (struct pseudo_slot, regs_num);
570   slots = XNEWVEC (struct slot, regs_num);
571   /* Sort regnos according their usage frequencies.  */
572   qsort (pseudo_regnos, n, sizeof (int), regno_freq_compare);
573   n = assign_spill_hard_regs (pseudo_regnos, n);
574   assign_stack_slot_num_and_sort_pseudos (pseudo_regnos, n);
575   for (i = 0; i < n; i++)
576     if (pseudo_slots[pseudo_regnos[i]].mem == NULL_RTX)
577       assign_mem_slot (pseudo_regnos[i]);
578   if (n > 0 && crtl->stack_alignment_needed)
579     /* If we have a stack frame, we must align it now.  The stack size
580        may be a part of the offset computation for register
581        elimination.  */
582     assign_stack_local (BLKmode, 0, crtl->stack_alignment_needed);
583   if (lra_dump_file != NULL)
584     {
585       for (i = 0; i < slots_num; i++)
586 	{
587 	  fprintf (lra_dump_file, "  Slot %d regnos (width = %d):", i,
588 		   GET_MODE_SIZE (GET_MODE (slots[i].mem)));
589 	  for (curr_regno = slots[i].regno;;
590 	       curr_regno = pseudo_slots[curr_regno].next - pseudo_slots)
591 	    {
592 	      fprintf (lra_dump_file, "	 %d", curr_regno);
593 	      if (pseudo_slots[curr_regno].next == NULL)
594 		break;
595 	    }
596 	  fprintf (lra_dump_file, "\n");
597 	}
598     }
599   spill_pseudos ();
600   free (slots);
601   free (pseudo_slots);
602   free (pseudo_regnos);
603   free (spill_hard_reg);
604 }
605 
606 /* Apply alter_subreg for subregs of regs in *LOC.  Use FINAL_P for
607    alter_subreg calls. Return true if any subreg of reg is
608    processed.  */
609 static bool
alter_subregs(rtx * loc,bool final_p)610 alter_subregs (rtx *loc, bool final_p)
611 {
612   int i;
613   rtx x = *loc;
614   bool res;
615   const char *fmt;
616   enum rtx_code code;
617 
618   if (x == NULL_RTX)
619     return false;
620   code = GET_CODE (x);
621   if (code == SUBREG && REG_P (SUBREG_REG (x)))
622     {
623       lra_assert (REGNO (SUBREG_REG (x)) < FIRST_PSEUDO_REGISTER);
624       alter_subreg (loc, final_p);
625       return true;
626     }
627   fmt = GET_RTX_FORMAT (code);
628   res = false;
629   for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
630     {
631       if (fmt[i] == 'e')
632 	{
633 	  if (alter_subregs (&XEXP (x, i), final_p))
634 	    res = true;
635 	}
636       else if (fmt[i] == 'E')
637 	{
638 	  int j;
639 
640 	  for (j = XVECLEN (x, i) - 1; j >= 0; j--)
641 	    if (alter_subregs (&XVECEXP (x, i, j), final_p))
642 	      res = true;
643 	}
644     }
645   return res;
646 }
647 
648 /* Return true if REGNO is used for return in the current
649    function.  */
650 static bool
return_regno_p(unsigned int regno)651 return_regno_p (unsigned int regno)
652 {
653   rtx outgoing = crtl->return_rtx;
654 
655   if (! outgoing)
656     return false;
657 
658   if (REG_P (outgoing))
659     return REGNO (outgoing) == regno;
660   else if (GET_CODE (outgoing) == PARALLEL)
661     {
662       int i;
663 
664       for (i = 0; i < XVECLEN (outgoing, 0); i++)
665 	{
666 	  rtx x = XEXP (XVECEXP (outgoing, 0, i), 0);
667 
668 	  if (REG_P (x) && REGNO (x) == regno)
669 	    return true;
670 	}
671     }
672   return false;
673 }
674 
675 /* Final change of pseudos got hard registers into the corresponding
676    hard registers and removing temporary clobbers.  */
677 void
lra_final_code_change(void)678 lra_final_code_change (void)
679 {
680   int i, hard_regno;
681   basic_block bb;
682   rtx insn, curr;
683   int max_regno = max_reg_num ();
684 
685   for (i = FIRST_PSEUDO_REGISTER; i < max_regno; i++)
686     if (lra_reg_info[i].nrefs != 0
687 	&& (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
688       SET_REGNO (regno_reg_rtx[i], hard_regno);
689   FOR_EACH_BB_FN (bb, cfun)
690     FOR_BB_INSNS_SAFE (bb, insn, curr)
691       if (INSN_P (insn))
692 	{
693 	  rtx pat = PATTERN (insn);
694 
695 	  if (GET_CODE (pat) == CLOBBER && LRA_TEMP_CLOBBER_P (pat))
696 	    {
697 	      /* Remove clobbers temporarily created in LRA.  We don't
698 		 need them anymore and don't want to waste compiler
699 		 time processing them in a few subsequent passes.  */
700 	      lra_invalidate_insn_data (insn);
701 	      delete_insn (insn);
702 	      continue;
703 	    }
704 
705 	  /* IRA can generate move insns involving pseudos.  It is
706 	     better remove them earlier to speed up compiler a bit.
707 	     It is also better to do it here as they might not pass
708 	     final RTL check in LRA, (e.g. insn moving a control
709 	     register into itself).  So remove an useless move insn
710 	     unless next insn is USE marking the return reg (we should
711 	     save this as some subsequent optimizations assume that
712 	     such original insns are saved).  */
713 	  if (NONJUMP_INSN_P (insn) && GET_CODE (pat) == SET
714 	      && REG_P (SET_SRC (pat)) && REG_P (SET_DEST (pat))
715 	      && REGNO (SET_SRC (pat)) == REGNO (SET_DEST (pat))
716 	      && ! return_regno_p (REGNO (SET_SRC (pat))))
717 	    {
718 	      lra_invalidate_insn_data (insn);
719 	      delete_insn (insn);
720 	      continue;
721 	    }
722 
723 	  lra_insn_recog_data_t id = lra_get_insn_recog_data (insn);
724 	  struct lra_static_insn_data *static_id = id->insn_static_data;
725 	  bool insn_change_p = false;
726 
727 	  for (i = id->insn_static_data->n_operands - 1; i >= 0; i--)
728 	    if ((DEBUG_INSN_P (insn) || ! static_id->operand[i].is_operator)
729 		&& alter_subregs (id->operand_loc[i], ! DEBUG_INSN_P (insn)))
730 	      {
731 		lra_update_dup (id, i);
732 		insn_change_p = true;
733 	      }
734 	  if (insn_change_p)
735 	    lra_update_operator_dups (id);
736 	}
737 }
738