xref: /netbsd-src/external/gpl3/gcc.old/dist/gcc/ira.c (revision 07ece4eabb6d327c320416d49d51617a7c0fb3be)
1 /* Integrated Register Allocator (IRA) entry point.
2    Copyright (C) 2006, 2007, 2008, 2009, 2010
3    Free Software Foundation, Inc.
4    Contributed by Vladimir Makarov <vmakarov@redhat.com>.
5 
6 This file is part of GCC.
7 
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
11 version.
12 
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
16 for more details.
17 
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3.  If not see
20 <http://www.gnu.org/licenses/>.  */
21 
22 /* The integrated register allocator (IRA) is a
23    regional register allocator performing graph coloring on a top-down
24    traversal of nested regions.  Graph coloring in a region is based
25    on Chaitin-Briggs algorithm.  It is called integrated because
26    register coalescing, register live range splitting, and choosing a
27    better hard register are done on-the-fly during coloring.  Register
28    coalescing and choosing a cheaper hard register is done by hard
29    register preferencing during hard register assigning.  The live
30    range splitting is a byproduct of the regional register allocation.
31 
32    Major IRA notions are:
33 
34      o *Region* is a part of CFG where graph coloring based on
35        Chaitin-Briggs algorithm is done.  IRA can work on any set of
36        nested CFG regions forming a tree.  Currently the regions are
37        the entire function for the root region and natural loops for
38        the other regions.  Therefore data structure representing a
39        region is called loop_tree_node.
40 
41      o *Cover class* is a register class belonging to a set of
42        non-intersecting register classes containing all of the
43        hard-registers available for register allocation.  The set of
44        all cover classes for a target is defined in the corresponding
45        machine-description file according some criteria.  Such notion
46        is needed because Chaitin-Briggs algorithm works on
47        non-intersected register classes.
48 
49      o *Allocno* represents the live range of a pseudo-register in a
50        region.  Besides the obvious attributes like the corresponding
51        pseudo-register number, cover class, conflicting allocnos and
52        conflicting hard-registers, there are a few allocno attributes
53        which are important for understanding the allocation algorithm:
54 
55        - *Live ranges*.  This is a list of ranges of *program
56          points* where the allocno lives.  Program points represent
57          places where a pseudo can be born or become dead (there are
58          approximately two times more program points than the insns)
59          and they are represented by integers starting with 0.  The
60          live ranges are used to find conflicts between allocnos of
61          different cover classes.  They also play very important role
62          for the transformation of the IRA internal representation of
63          several regions into a one region representation.  The later is
64          used during the reload pass work because each allocno
65          represents all of the corresponding pseudo-registers.
66 
67        - *Hard-register costs*.  This is a vector of size equal to the
68          number of available hard-registers of the allocno's cover
69          class.  The cost of a callee-clobbered hard-register for an
70          allocno is increased by the cost of save/restore code around
71          the calls through the given allocno's life.  If the allocno
72          is a move instruction operand and another operand is a
73          hard-register of the allocno's cover class, the cost of the
74          hard-register is decreased by the move cost.
75 
76          When an allocno is assigned, the hard-register with minimal
77          full cost is used.  Initially, a hard-register's full cost is
78          the corresponding value from the hard-register's cost vector.
79          If the allocno is connected by a *copy* (see below) to
80          another allocno which has just received a hard-register, the
81          cost of the hard-register is decreased.  Before choosing a
82          hard-register for an allocno, the allocno's current costs of
83          the hard-registers are modified by the conflict hard-register
84          costs of all of the conflicting allocnos which are not
85          assigned yet.
86 
87        - *Conflict hard-register costs*.  This is a vector of the same
88          size as the hard-register costs vector.  To permit an
89          unassigned allocno to get a better hard-register, IRA uses
90          this vector to calculate the final full cost of the
91          available hard-registers.  Conflict hard-register costs of an
92          unassigned allocno are also changed with a change of the
93          hard-register cost of the allocno when a copy involving the
94          allocno is processed as described above.  This is done to
95          show other unassigned allocnos that a given allocno prefers
96          some hard-registers in order to remove the move instruction
97          corresponding to the copy.
98 
99      o *Cap*.  If a pseudo-register does not live in a region but
100        lives in a nested region, IRA creates a special allocno called
101        a cap in the outer region.  A region cap is also created for a
102        subregion cap.
103 
104      o *Copy*.  Allocnos can be connected by copies.  Copies are used
105        to modify hard-register costs for allocnos during coloring.
106        Such modifications reflects a preference to use the same
107        hard-register for the allocnos connected by copies.  Usually
108        copies are created for move insns (in this case it results in
109        register coalescing).  But IRA also creates copies for operands
110        of an insn which should be assigned to the same hard-register
111        due to constraints in the machine description (it usually
112        results in removing a move generated in reload to satisfy
113        the constraints) and copies referring to the allocno which is
114        the output operand of an instruction and the allocno which is
115        an input operand dying in the instruction (creation of such
116        copies results in less register shuffling).  IRA *does not*
117        create copies between the same register allocnos from different
118        regions because we use another technique for propagating
119        hard-register preference on the borders of regions.
120 
121    Allocnos (including caps) for the upper region in the region tree
122    *accumulate* information important for coloring from allocnos with
123    the same pseudo-register from nested regions.  This includes
124    hard-register and memory costs, conflicts with hard-registers,
125    allocno conflicts, allocno copies and more.  *Thus, attributes for
126    allocnos in a region have the same values as if the region had no
127    subregions*.  It means that attributes for allocnos in the
128    outermost region corresponding to the function have the same values
129    as though the allocation used only one region which is the entire
130    function.  It also means that we can look at IRA work as if the
131    first IRA did allocation for all function then it improved the
132    allocation for loops then their subloops and so on.
133 
134    IRA major passes are:
135 
136      o Building IRA internal representation which consists of the
137        following subpasses:
138 
139        * First, IRA builds regions and creates allocnos (file
140          ira-build.c) and initializes most of their attributes.
141 
142        * Then IRA finds a cover class for each allocno and calculates
143          its initial (non-accumulated) cost of memory and each
144          hard-register of its cover class (file ira-cost.c).
145 
146        * IRA creates live ranges of each allocno, calulates register
147          pressure for each cover class in each region, sets up
148          conflict hard registers for each allocno and info about calls
149          the allocno lives through (file ira-lives.c).
150 
151        * IRA removes low register pressure loops from the regions
152          mostly to speed IRA up (file ira-build.c).
153 
154        * IRA propagates accumulated allocno info from lower region
155          allocnos to corresponding upper region allocnos (file
156          ira-build.c).
157 
158        * IRA creates all caps (file ira-build.c).
159 
160        * Having live-ranges of allocnos and their cover classes, IRA
161          creates conflicting allocnos of the same cover class for each
162          allocno.  Conflicting allocnos are stored as a bit vector or
163          array of pointers to the conflicting allocnos whatever is
164          more profitable (file ira-conflicts.c).  At this point IRA
165          creates allocno copies.
166 
167      o Coloring.  Now IRA has all necessary info to start graph coloring
168        process.  It is done in each region on top-down traverse of the
169        region tree (file ira-color.c).  There are following subpasses:
170 
171        * Optional aggressive coalescing of allocnos in the region.
172 
173        * Putting allocnos onto the coloring stack.  IRA uses Briggs
174          optimistic coloring which is a major improvement over
175          Chaitin's coloring.  Therefore IRA does not spill allocnos at
176          this point.  There is some freedom in the order of putting
177          allocnos on the stack which can affect the final result of
178          the allocation.  IRA uses some heuristics to improve the order.
179 
180        * Popping the allocnos from the stack and assigning them hard
181          registers.  If IRA can not assign a hard register to an
182          allocno and the allocno is coalesced, IRA undoes the
183          coalescing and puts the uncoalesced allocnos onto the stack in
184          the hope that some such allocnos will get a hard register
185          separately.  If IRA fails to assign hard register or memory
186          is more profitable for it, IRA spills the allocno.  IRA
187          assigns the allocno the hard-register with minimal full
188          allocation cost which reflects the cost of usage of the
189          hard-register for the allocno and cost of usage of the
190          hard-register for allocnos conflicting with given allocno.
191 
192        * After allono assigning in the region, IRA modifies the hard
193          register and memory costs for the corresponding allocnos in
194          the subregions to reflect the cost of possible loads, stores,
195          or moves on the border of the region and its subregions.
196          When default regional allocation algorithm is used
197          (-fira-algorithm=mixed), IRA just propagates the assignment
198          for allocnos if the register pressure in the region for the
199          corresponding cover class is less than number of available
200          hard registers for given cover class.
201 
202      o Spill/restore code moving.  When IRA performs an allocation
203        by traversing regions in top-down order, it does not know what
204        happens below in the region tree.  Therefore, sometimes IRA
205        misses opportunities to perform a better allocation.  A simple
206        optimization tries to improve allocation in a region having
207        subregions and containing in another region.  If the
208        corresponding allocnos in the subregion are spilled, it spills
209        the region allocno if it is profitable.  The optimization
210        implements a simple iterative algorithm performing profitable
211        transformations while they are still possible.  It is fast in
212        practice, so there is no real need for a better time complexity
213        algorithm.
214 
215      o Code change.  After coloring, two allocnos representing the same
216        pseudo-register outside and inside a region respectively may be
217        assigned to different locations (hard-registers or memory).  In
218        this case IRA creates and uses a new pseudo-register inside the
219        region and adds code to move allocno values on the region's
220        borders.  This is done during top-down traversal of the regions
221        (file ira-emit.c).  In some complicated cases IRA can create a
222        new allocno to move allocno values (e.g. when a swap of values
223        stored in two hard-registers is needed).  At this stage, the
224        new allocno is marked as spilled.  IRA still creates the
225        pseudo-register and the moves on the region borders even when
226        both allocnos were assigned to the same hard-register.  If the
227        reload pass spills a pseudo-register for some reason, the
228        effect will be smaller because another allocno will still be in
229        the hard-register.  In most cases, this is better then spilling
230        both allocnos.  If reload does not change the allocation
231        for the two pseudo-registers, the trivial move will be removed
232        by post-reload optimizations.  IRA does not generate moves for
233        allocnos assigned to the same hard register when the default
234        regional allocation algorithm is used and the register pressure
235        in the region for the corresponding allocno cover class is less
236        than number of available hard registers for given cover class.
237        IRA also does some optimizations to remove redundant stores and
238        to reduce code duplication on the region borders.
239 
240      o Flattening internal representation.  After changing code, IRA
241        transforms its internal representation for several regions into
242        one region representation (file ira-build.c).  This process is
243        called IR flattening.  Such process is more complicated than IR
244        rebuilding would be, but is much faster.
245 
246      o After IR flattening, IRA tries to assign hard registers to all
247        spilled allocnos.  This is impelemented by a simple and fast
248        priority coloring algorithm (see function
249        ira_reassign_conflict_allocnos::ira-color.c).  Here new allocnos
250        created during the code change pass can be assigned to hard
251        registers.
252 
253      o At the end IRA calls the reload pass.  The reload pass
254        communicates with IRA through several functions in file
255        ira-color.c to improve its decisions in
256 
257        * sharing stack slots for the spilled pseudos based on IRA info
258          about pseudo-register conflicts.
259 
260        * reassigning hard-registers to all spilled pseudos at the end
261          of each reload iteration.
262 
263        * choosing a better hard-register to spill based on IRA info
264          about pseudo-register live ranges and the register pressure
265          in places where the pseudo-register lives.
266 
267    IRA uses a lot of data representing the target processors.  These
268    data are initilized in file ira.c.
269 
270    If function has no loops (or the loops are ignored when
271    -fira-algorithm=CB is used), we have classic Chaitin-Briggs
272    coloring (only instead of separate pass of coalescing, we use hard
273    register preferencing).  In such case, IRA works much faster
274    because many things are not made (like IR flattening, the
275    spill/restore optimization, and the code change).
276 
277    Literature is worth to read for better understanding the code:
278 
279    o Preston Briggs, Keith D. Cooper, Linda Torczon.  Improvements to
280      Graph Coloring Register Allocation.
281 
282    o David Callahan, Brian Koblenz.  Register allocation via
283      hierarchical graph coloring.
284 
285    o Keith Cooper, Anshuman Dasgupta, Jason Eckhardt. Revisiting Graph
286      Coloring Register Allocation: A Study of the Chaitin-Briggs and
287      Callahan-Koblenz Algorithms.
288 
289    o Guei-Yuan Lueh, Thomas Gross, and Ali-Reza Adl-Tabatabai. Global
290      Register Allocation Based on Graph Fusion.
291 
292    o Vladimir Makarov. The Integrated Register Allocator for GCC.
293 
294    o Vladimir Makarov.  The top-down register allocator for irregular
295      register file architectures.
296 
297 */
298 
299 
300 #include "config.h"
301 #include "system.h"
302 #include "coretypes.h"
303 #include "tm.h"
304 #include "regs.h"
305 #include "rtl.h"
306 #include "tm_p.h"
307 #include "target.h"
308 #include "flags.h"
309 #include "obstack.h"
310 #include "bitmap.h"
311 #include "hard-reg-set.h"
312 #include "basic-block.h"
313 #include "expr.h"
314 #include "recog.h"
315 #include "params.h"
316 #include "timevar.h"
317 #include "tree-pass.h"
318 #include "output.h"
319 #include "except.h"
320 #include "reload.h"
321 #include "errors.h"
322 #include "integrate.h"
323 #include "df.h"
324 #include "ggc.h"
325 #include "ira-int.h"
326 
327 
328 /* A modified value of flag `-fira-verbose' used internally.  */
329 int internal_flag_ira_verbose;
330 
331 /* Dump file of the allocator if it is not NULL.  */
332 FILE *ira_dump_file;
333 
334 /* Pools for allocnos, copies, allocno live ranges.  */
335 alloc_pool allocno_pool, copy_pool, allocno_live_range_pool;
336 
337 /* The number of elements in the following array.  */
338 int ira_spilled_reg_stack_slots_num;
339 
340 /* The following array contains info about spilled pseudo-registers
341    stack slots used in current function so far.  */
342 struct ira_spilled_reg_stack_slot *ira_spilled_reg_stack_slots;
343 
344 /* Correspondingly overall cost of the allocation, cost of the
345    allocnos assigned to hard-registers, cost of the allocnos assigned
346    to memory, cost of loads, stores and register move insns generated
347    for pseudo-register live range splitting (see ira-emit.c).  */
348 int ira_overall_cost;
349 int ira_reg_cost, ira_mem_cost;
350 int ira_load_cost, ira_store_cost, ira_shuffle_cost;
351 int ira_move_loops_num, ira_additional_jumps_num;
352 
353 /* All registers that can be eliminated.  */
354 
355 HARD_REG_SET eliminable_regset;
356 
357 /* Map: hard regs X modes -> set of hard registers for storing value
358    of given mode starting with given hard register.  */
359 HARD_REG_SET ira_reg_mode_hard_regset[FIRST_PSEUDO_REGISTER][NUM_MACHINE_MODES];
360 
361 /* The following two variables are array analogs of the macros
362    MEMORY_MOVE_COST and REGISTER_MOVE_COST.  */
363 short int ira_memory_move_cost[MAX_MACHINE_MODE][N_REG_CLASSES][2];
364 move_table *ira_register_move_cost[MAX_MACHINE_MODE];
365 
366 /* Similar to may_move_in_cost but it is calculated in IRA instead of
367    regclass.  Another difference is that we take only available hard
368    registers into account to figure out that one register class is a
369    subset of the another one.  */
370 move_table *ira_may_move_in_cost[MAX_MACHINE_MODE];
371 
372 /* Similar to may_move_out_cost but it is calculated in IRA instead of
373    regclass.  Another difference is that we take only available hard
374    registers into account to figure out that one register class is a
375    subset of the another one.  */
376 move_table *ira_may_move_out_cost[MAX_MACHINE_MODE];
377 
378 /* Register class subset relation: TRUE if the first class is a subset
379    of the second one considering only hard registers available for the
380    allocation.  */
381 int ira_class_subset_p[N_REG_CLASSES][N_REG_CLASSES];
382 
383 /* Temporary hard reg set used for a different calculation.  */
384 static HARD_REG_SET temp_hard_regset;
385 
386 
387 
388 /* The function sets up the map IRA_REG_MODE_HARD_REGSET.  */
389 static void
390 setup_reg_mode_hard_regset (void)
391 {
392   int i, m, hard_regno;
393 
394   for (m = 0; m < NUM_MACHINE_MODES; m++)
395     for (hard_regno = 0; hard_regno < FIRST_PSEUDO_REGISTER; hard_regno++)
396       {
397 	CLEAR_HARD_REG_SET (ira_reg_mode_hard_regset[hard_regno][m]);
398 	for (i = hard_regno_nregs[hard_regno][m] - 1; i >= 0; i--)
399 	  if (hard_regno + i < FIRST_PSEUDO_REGISTER)
400 	    SET_HARD_REG_BIT (ira_reg_mode_hard_regset[hard_regno][m],
401 			      hard_regno + i);
402       }
403 }
404 
405 
406 
407 /* Hard registers that can not be used for the register allocator for
408    all functions of the current compilation unit.  */
409 static HARD_REG_SET no_unit_alloc_regs;
410 
411 /* Array of the number of hard registers of given class which are
412    available for allocation.  The order is defined by the
413    allocation order.  */
414 short ira_class_hard_regs[N_REG_CLASSES][FIRST_PSEUDO_REGISTER];
415 
416 /* The number of elements of the above array for given register
417    class.  */
418 int ira_class_hard_regs_num[N_REG_CLASSES];
419 
420 /* Index (in ira_class_hard_regs) for given register class and hard
421    register (in general case a hard register can belong to several
422    register classes).  The index is negative for hard registers
423    unavailable for the allocation. */
424 short ira_class_hard_reg_index[N_REG_CLASSES][FIRST_PSEUDO_REGISTER];
425 
426 /* The function sets up the three arrays declared above.  */
427 static void
428 setup_class_hard_regs (void)
429 {
430   int cl, i, hard_regno, n;
431   HARD_REG_SET processed_hard_reg_set;
432 
433   ira_assert (SHRT_MAX >= FIRST_PSEUDO_REGISTER);
434   /* We could call ORDER_REGS_FOR_LOCAL_ALLOC here (it is usually
435      putting hard callee-used hard registers first).  But our
436      heuristics work better.  */
437   for (cl = (int) N_REG_CLASSES - 1; cl >= 0; cl--)
438     {
439       COPY_HARD_REG_SET (temp_hard_regset, reg_class_contents[cl]);
440       AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
441       CLEAR_HARD_REG_SET (processed_hard_reg_set);
442       for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
443 	ira_class_hard_reg_index[cl][0] = -1;
444       for (n = 0, i = 0; i < FIRST_PSEUDO_REGISTER; i++)
445 	{
446 #ifdef REG_ALLOC_ORDER
447 	  hard_regno = reg_alloc_order[i];
448 #else
449 	  hard_regno = i;
450 #endif
451 	  if (TEST_HARD_REG_BIT (processed_hard_reg_set, hard_regno))
452 	    continue;
453 	  SET_HARD_REG_BIT (processed_hard_reg_set, hard_regno);
454       	  if (! TEST_HARD_REG_BIT (temp_hard_regset, hard_regno))
455 	    ira_class_hard_reg_index[cl][hard_regno] = -1;
456 	  else
457 	    {
458 	      ira_class_hard_reg_index[cl][hard_regno] = n;
459 	      ira_class_hard_regs[cl][n++] = hard_regno;
460 	    }
461 	}
462       ira_class_hard_regs_num[cl] = n;
463     }
464 }
465 
466 /* Number of given class hard registers available for the register
467    allocation for given classes.  */
468 int ira_available_class_regs[N_REG_CLASSES];
469 
470 /* Set up IRA_AVAILABLE_CLASS_REGS.  */
471 static void
472 setup_available_class_regs (void)
473 {
474   int i, j;
475 
476   memset (ira_available_class_regs, 0, sizeof (ira_available_class_regs));
477   for (i = 0; i < N_REG_CLASSES; i++)
478     {
479       COPY_HARD_REG_SET (temp_hard_regset, reg_class_contents[i]);
480       AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
481       for (j = 0; j < FIRST_PSEUDO_REGISTER; j++)
482 	if (TEST_HARD_REG_BIT (temp_hard_regset, j))
483 	  ira_available_class_regs[i]++;
484     }
485 }
486 
487 /* Set up global variables defining info about hard registers for the
488    allocation.  These depend on USE_HARD_FRAME_P whose TRUE value means
489    that we can use the hard frame pointer for the allocation.  */
490 static void
491 setup_alloc_regs (bool use_hard_frame_p)
492 {
493   COPY_HARD_REG_SET (no_unit_alloc_regs, fixed_reg_set);
494   if (! use_hard_frame_p)
495     SET_HARD_REG_BIT (no_unit_alloc_regs, HARD_FRAME_POINTER_REGNUM);
496   setup_class_hard_regs ();
497   setup_available_class_regs ();
498 }
499 
500 
501 
502 /* Set up IRA_MEMORY_MOVE_COST, IRA_REGISTER_MOVE_COST.  */
503 static void
504 setup_class_subset_and_memory_move_costs (void)
505 {
506   int cl, cl2, mode;
507   HARD_REG_SET temp_hard_regset2;
508 
509   for (mode = 0; mode < MAX_MACHINE_MODE; mode++)
510     ira_memory_move_cost[mode][NO_REGS][0]
511       = ira_memory_move_cost[mode][NO_REGS][1] = SHRT_MAX;
512   for (cl = (int) N_REG_CLASSES - 1; cl >= 0; cl--)
513     {
514       if (cl != (int) NO_REGS)
515 	for (mode = 0; mode < MAX_MACHINE_MODE; mode++)
516 	  {
517 	    ira_memory_move_cost[mode][cl][0] =
518 	      MEMORY_MOVE_COST ((enum machine_mode) mode,
519 				(enum reg_class) cl, 0);
520 	    ira_memory_move_cost[mode][cl][1] =
521 	      MEMORY_MOVE_COST ((enum machine_mode) mode,
522 				(enum reg_class) cl, 1);
523 	    /* Costs for NO_REGS are used in cost calculation on the
524 	       1st pass when the preferred register classes are not
525 	       known yet.  In this case we take the best scenario.  */
526 	    if (ira_memory_move_cost[mode][NO_REGS][0]
527 		> ira_memory_move_cost[mode][cl][0])
528 	      ira_memory_move_cost[mode][NO_REGS][0]
529 		= ira_memory_move_cost[mode][cl][0];
530 	    if (ira_memory_move_cost[mode][NO_REGS][1]
531 		> ira_memory_move_cost[mode][cl][1])
532 	      ira_memory_move_cost[mode][NO_REGS][1]
533 		= ira_memory_move_cost[mode][cl][1];
534 	  }
535       for (cl2 = (int) N_REG_CLASSES - 1; cl2 >= 0; cl2--)
536 	{
537 	  COPY_HARD_REG_SET (temp_hard_regset, reg_class_contents[cl]);
538 	  AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
539 	  COPY_HARD_REG_SET (temp_hard_regset2, reg_class_contents[cl2]);
540 	  AND_COMPL_HARD_REG_SET (temp_hard_regset2, no_unit_alloc_regs);
541 	  ira_class_subset_p[cl][cl2]
542 	    = hard_reg_set_subset_p (temp_hard_regset, temp_hard_regset2);
543 	}
544     }
545 }
546 
547 
548 
549 /* Define the following macro if allocation through malloc if
550    preferable.  */
551 #define IRA_NO_OBSTACK
552 
553 #ifndef IRA_NO_OBSTACK
554 /* Obstack used for storing all dynamic data (except bitmaps) of the
555    IRA.  */
556 static struct obstack ira_obstack;
557 #endif
558 
559 /* Obstack used for storing all bitmaps of the IRA.  */
560 static struct bitmap_obstack ira_bitmap_obstack;
561 
562 /* Allocate memory of size LEN for IRA data.  */
563 void *
564 ira_allocate (size_t len)
565 {
566   void *res;
567 
568 #ifndef IRA_NO_OBSTACK
569   res = obstack_alloc (&ira_obstack, len);
570 #else
571   res = xmalloc (len);
572 #endif
573   return res;
574 }
575 
576 /* Reallocate memory PTR of size LEN for IRA data.  */
577 void *
578 ira_reallocate (void *ptr, size_t len)
579 {
580   void *res;
581 
582 #ifndef IRA_NO_OBSTACK
583   res = obstack_alloc (&ira_obstack, len);
584 #else
585   res = xrealloc (ptr, len);
586 #endif
587   return res;
588 }
589 
590 /* Free memory ADDR allocated for IRA data.  */
591 void
592 ira_free (void *addr ATTRIBUTE_UNUSED)
593 {
594 #ifndef IRA_NO_OBSTACK
595   /* do nothing */
596 #else
597   free (addr);
598 #endif
599 }
600 
601 
602 /* Allocate and returns bitmap for IRA.  */
603 bitmap
604 ira_allocate_bitmap (void)
605 {
606   return BITMAP_ALLOC (&ira_bitmap_obstack);
607 }
608 
609 /* Free bitmap B allocated for IRA.  */
610 void
611 ira_free_bitmap (bitmap b ATTRIBUTE_UNUSED)
612 {
613   /* do nothing */
614 }
615 
616 
617 
618 /* Output information about allocation of all allocnos (except for
619    caps) into file F.  */
620 void
621 ira_print_disposition (FILE *f)
622 {
623   int i, n, max_regno;
624   ira_allocno_t a;
625   basic_block bb;
626 
627   fprintf (f, "Disposition:");
628   max_regno = max_reg_num ();
629   for (n = 0, i = FIRST_PSEUDO_REGISTER; i < max_regno; i++)
630     for (a = ira_regno_allocno_map[i];
631 	 a != NULL;
632 	 a = ALLOCNO_NEXT_REGNO_ALLOCNO (a))
633       {
634 	if (n % 4 == 0)
635 	  fprintf (f, "\n");
636 	n++;
637 	fprintf (f, " %4d:r%-4d", ALLOCNO_NUM (a), ALLOCNO_REGNO (a));
638 	if ((bb = ALLOCNO_LOOP_TREE_NODE (a)->bb) != NULL)
639 	  fprintf (f, "b%-3d", bb->index);
640 	else
641 	  fprintf (f, "l%-3d", ALLOCNO_LOOP_TREE_NODE (a)->loop->num);
642 	if (ALLOCNO_HARD_REGNO (a) >= 0)
643 	  fprintf (f, " %3d", ALLOCNO_HARD_REGNO (a));
644 	else
645 	  fprintf (f, " mem");
646       }
647   fprintf (f, "\n");
648 }
649 
650 /* Outputs information about allocation of all allocnos into
651    stderr.  */
652 void
653 ira_debug_disposition (void)
654 {
655   ira_print_disposition (stderr);
656 }
657 
658 
659 
660 /* For each reg class, table listing all the classes contained in it
661    (excluding the class itself.  Non-allocatable registers are
662    excluded from the consideration).  */
663 static enum reg_class alloc_reg_class_subclasses[N_REG_CLASSES][N_REG_CLASSES];
664 
665 /* Initialize the table of subclasses of each reg class.  */
666 static void
667 setup_reg_subclasses (void)
668 {
669   int i, j;
670   HARD_REG_SET temp_hard_regset2;
671 
672   for (i = 0; i < N_REG_CLASSES; i++)
673     for (j = 0; j < N_REG_CLASSES; j++)
674       alloc_reg_class_subclasses[i][j] = LIM_REG_CLASSES;
675 
676   for (i = 0; i < N_REG_CLASSES; i++)
677     {
678       if (i == (int) NO_REGS)
679 	continue;
680 
681       COPY_HARD_REG_SET (temp_hard_regset, reg_class_contents[i]);
682       AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
683       if (hard_reg_set_empty_p (temp_hard_regset))
684 	continue;
685       for (j = 0; j < N_REG_CLASSES; j++)
686 	if (i != j)
687 	  {
688 	    enum reg_class *p;
689 
690 	    COPY_HARD_REG_SET (temp_hard_regset2, reg_class_contents[j]);
691 	    AND_COMPL_HARD_REG_SET (temp_hard_regset2, no_unit_alloc_regs);
692 	    if (! hard_reg_set_subset_p (temp_hard_regset,
693 					 temp_hard_regset2))
694 	      continue;
695 	    p = &alloc_reg_class_subclasses[j][0];
696 	    while (*p != LIM_REG_CLASSES) p++;
697 	    *p = (enum reg_class) i;
698 	  }
699     }
700 }
701 
702 
703 
704 /* Number of cover classes.  Cover classes is non-intersected register
705    classes containing all hard-registers available for the
706    allocation.  */
707 int ira_reg_class_cover_size;
708 
709 /* The array containing cover classes (see also comments for macro
710    IRA_COVER_CLASSES).  Only first IRA_REG_CLASS_COVER_SIZE elements are
711    used for this.  */
712 enum reg_class ira_reg_class_cover[N_REG_CLASSES];
713 
714 /* The number of elements in the subsequent array.  */
715 int ira_important_classes_num;
716 
717 /* The array containing non-empty classes (including non-empty cover
718    classes) which are subclasses of cover classes.  Such classes is
719    important for calculation of the hard register usage costs.  */
720 enum reg_class ira_important_classes[N_REG_CLASSES];
721 
722 /* The array containing indexes of important classes in the previous
723    array.  The array elements are defined only for important
724    classes.  */
725 int ira_important_class_nums[N_REG_CLASSES];
726 
727 /* Set the four global variables defined above.  */
728 static void
729 setup_cover_and_important_classes (void)
730 {
731   int i, j, n, cl;
732   bool set_p;
733   const enum reg_class *cover_classes;
734   HARD_REG_SET temp_hard_regset2;
735   static enum reg_class classes[LIM_REG_CLASSES + 1];
736 
737   if (targetm.ira_cover_classes == NULL)
738     cover_classes = NULL;
739   else
740     cover_classes = targetm.ira_cover_classes ();
741   if (cover_classes == NULL)
742     ira_assert (flag_ira_algorithm == IRA_ALGORITHM_PRIORITY);
743   else
744     {
745       for (i = 0; (cl = cover_classes[i]) != LIM_REG_CLASSES; i++)
746 	classes[i] = (enum reg_class) cl;
747       classes[i] = LIM_REG_CLASSES;
748     }
749 
750   if (flag_ira_algorithm == IRA_ALGORITHM_PRIORITY)
751     {
752       n = 0;
753       for (i = 0; i <= LIM_REG_CLASSES; i++)
754 	{
755 	  if (i == NO_REGS)
756 	    continue;
757 #ifdef CONSTRAINT_NUM_DEFINED_P
758 	  for (j = 0; j < CONSTRAINT__LIMIT; j++)
759 	    if ((int) REG_CLASS_FOR_CONSTRAINT ((enum constraint_num) j) == i)
760 	      break;
761 	  if (j < CONSTRAINT__LIMIT)
762 	    {
763 	      classes[n++] = (enum reg_class) i;
764 	      continue;
765 	    }
766 #endif
767 	  COPY_HARD_REG_SET (temp_hard_regset, reg_class_contents[i]);
768 	  AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
769 	  for (j = 0; j < LIM_REG_CLASSES; j++)
770 	    {
771 	      if (i == j)
772 		continue;
773 	      COPY_HARD_REG_SET (temp_hard_regset2, reg_class_contents[j]);
774 	      AND_COMPL_HARD_REG_SET (temp_hard_regset2,
775 				      no_unit_alloc_regs);
776 	      if (hard_reg_set_equal_p (temp_hard_regset,
777 					temp_hard_regset2))
778 		    break;
779 	    }
780 	  if (j >= i)
781 	    classes[n++] = (enum reg_class) i;
782 	}
783       classes[n] = LIM_REG_CLASSES;
784     }
785 
786   ira_reg_class_cover_size = 0;
787   for (i = 0; (cl = classes[i]) != LIM_REG_CLASSES; i++)
788     {
789       for (j = 0; j < i; j++)
790 	if (flag_ira_algorithm != IRA_ALGORITHM_PRIORITY
791 	    && reg_classes_intersect_p ((enum reg_class) cl, classes[j]))
792 	  gcc_unreachable ();
793       COPY_HARD_REG_SET (temp_hard_regset, reg_class_contents[cl]);
794       AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
795       if (! hard_reg_set_empty_p (temp_hard_regset))
796 	ira_reg_class_cover[ira_reg_class_cover_size++] = (enum reg_class) cl;
797     }
798   ira_important_classes_num = 0;
799   for (cl = 0; cl < N_REG_CLASSES; cl++)
800     {
801       COPY_HARD_REG_SET (temp_hard_regset, reg_class_contents[cl]);
802       AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
803       if (! hard_reg_set_empty_p (temp_hard_regset))
804 	{
805 	  set_p = false;
806 	  for (j = 0; j < ira_reg_class_cover_size; j++)
807 	    {
808 	      COPY_HARD_REG_SET (temp_hard_regset, reg_class_contents[cl]);
809 	      AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
810 	      COPY_HARD_REG_SET (temp_hard_regset2,
811 				 reg_class_contents[ira_reg_class_cover[j]]);
812 	      AND_COMPL_HARD_REG_SET (temp_hard_regset2, no_unit_alloc_regs);
813 	      if ((enum reg_class) cl == ira_reg_class_cover[j]
814 		  || hard_reg_set_equal_p (temp_hard_regset,
815 					   temp_hard_regset2))
816 		break;
817 	      else if (hard_reg_set_subset_p (temp_hard_regset,
818 					      temp_hard_regset2))
819 		set_p = true;
820 	    }
821 	  if (set_p && j >= ira_reg_class_cover_size)
822 	    ira_important_classes[ira_important_classes_num++]
823 	      = (enum reg_class) cl;
824 	}
825     }
826   for (j = 0; j < ira_reg_class_cover_size; j++)
827     ira_important_classes[ira_important_classes_num++]
828       = ira_reg_class_cover[j];
829 }
830 
831 /* Map of all register classes to corresponding cover class containing
832    the given class.  If given class is not a subset of a cover class,
833    we translate it into the cheapest cover class.  */
834 enum reg_class ira_class_translate[N_REG_CLASSES];
835 
836 /* Set up array IRA_CLASS_TRANSLATE.  */
837 static void
838 setup_class_translate (void)
839 {
840   int cl, mode;
841   enum reg_class cover_class, best_class, *cl_ptr;
842   int i, cost, min_cost, best_cost;
843 
844   for (cl = 0; cl < N_REG_CLASSES; cl++)
845     ira_class_translate[cl] = NO_REGS;
846 
847   if (flag_ira_algorithm == IRA_ALGORITHM_PRIORITY)
848     for (cl = 0; cl < LIM_REG_CLASSES; cl++)
849       {
850 	COPY_HARD_REG_SET (temp_hard_regset, reg_class_contents[cl]);
851 	AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
852 	for (i = 0; i < ira_reg_class_cover_size; i++)
853 	  {
854 	    HARD_REG_SET temp_hard_regset2;
855 
856 	    cover_class = ira_reg_class_cover[i];
857 	    COPY_HARD_REG_SET (temp_hard_regset2,
858 			       reg_class_contents[cover_class]);
859 	    AND_COMPL_HARD_REG_SET (temp_hard_regset2, no_unit_alloc_regs);
860 	    if (hard_reg_set_equal_p (temp_hard_regset, temp_hard_regset2))
861 	      ira_class_translate[cl] = cover_class;
862 	  }
863       }
864   for (i = 0; i < ira_reg_class_cover_size; i++)
865     {
866       cover_class = ira_reg_class_cover[i];
867       if (flag_ira_algorithm != IRA_ALGORITHM_PRIORITY)
868 	for (cl_ptr = &alloc_reg_class_subclasses[cover_class][0];
869 	     (cl = *cl_ptr) != LIM_REG_CLASSES;
870 	     cl_ptr++)
871 	  {
872 	    if (ira_class_translate[cl] == NO_REGS)
873 	      ira_class_translate[cl] = cover_class;
874 #ifdef ENABLE_IRA_CHECKING
875 	    else
876 	      {
877 		COPY_HARD_REG_SET (temp_hard_regset, reg_class_contents[cl]);
878 		AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
879 		if (! hard_reg_set_empty_p (temp_hard_regset))
880 		  gcc_unreachable ();
881 	      }
882 #endif
883 	  }
884       ira_class_translate[cover_class] = cover_class;
885     }
886   /* For classes which are not fully covered by a cover class (in
887      other words covered by more one cover class), use the cheapest
888      cover class.  */
889   for (cl = 0; cl < N_REG_CLASSES; cl++)
890     {
891       if (cl == NO_REGS || ira_class_translate[cl] != NO_REGS)
892 	continue;
893       best_class = NO_REGS;
894       best_cost = INT_MAX;
895       for (i = 0; i < ira_reg_class_cover_size; i++)
896 	{
897 	  cover_class = ira_reg_class_cover[i];
898 	  COPY_HARD_REG_SET (temp_hard_regset,
899 			     reg_class_contents[cover_class]);
900 	  AND_HARD_REG_SET (temp_hard_regset, reg_class_contents[cl]);
901 	  AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
902 	  if (! hard_reg_set_empty_p (temp_hard_regset))
903 	    {
904 	      min_cost = INT_MAX;
905 	      for (mode = 0; mode < MAX_MACHINE_MODE; mode++)
906 		{
907 		  cost = (ira_memory_move_cost[mode][cl][0]
908 			  + ira_memory_move_cost[mode][cl][1]);
909 		  if (min_cost > cost)
910 		    min_cost = cost;
911 		}
912 	      if (best_class == NO_REGS || best_cost > min_cost)
913 		{
914 		  best_class = cover_class;
915 		  best_cost = min_cost;
916 		}
917 	    }
918 	}
919       ira_class_translate[cl] = best_class;
920     }
921 }
922 
923 /* Order numbers of cover classes in original target cover class
924    array, -1 for non-cover classes.  */
925 static int cover_class_order[N_REG_CLASSES];
926 
927 /* The function used to sort the important classes.  */
928 static int
929 comp_reg_classes_func (const void *v1p, const void *v2p)
930 {
931   enum reg_class cl1 = *(const enum reg_class *) v1p;
932   enum reg_class cl2 = *(const enum reg_class *) v2p;
933   int diff;
934 
935   cl1 = ira_class_translate[cl1];
936   cl2 = ira_class_translate[cl2];
937   if (cl1 != NO_REGS && cl2 != NO_REGS
938       && (diff = cover_class_order[cl1] - cover_class_order[cl2]) != 0)
939     return diff;
940   return (int) cl1 - (int) cl2;
941 }
942 
943 /* Reorder important classes according to the order of their cover
944    classes.  Set up array ira_important_class_nums too.  */
945 static void
946 reorder_important_classes (void)
947 {
948   int i;
949 
950   for (i = 0; i < N_REG_CLASSES; i++)
951     cover_class_order[i] = -1;
952   for (i = 0; i < ira_reg_class_cover_size; i++)
953     cover_class_order[ira_reg_class_cover[i]] = i;
954   qsort (ira_important_classes, ira_important_classes_num,
955 	 sizeof (enum reg_class), comp_reg_classes_func);
956   for (i = 0; i < ira_important_classes_num; i++)
957     ira_important_class_nums[ira_important_classes[i]] = i;
958 }
959 
960 /* The biggest important reg_class inside of intersection of the two
961    reg_classes (that is calculated taking only hard registers
962    available for allocation into account).  If the both reg_classes
963    contain no hard registers available for allocation, the value is
964    calculated by taking all hard-registers including fixed ones into
965    account.  */
966 enum reg_class ira_reg_class_intersect[N_REG_CLASSES][N_REG_CLASSES];
967 
968 /* True if the two classes (that is calculated taking only hard
969    registers available for allocation into account) are
970    intersected.  */
971 bool ira_reg_classes_intersect_p[N_REG_CLASSES][N_REG_CLASSES];
972 
973 /* Important classes with end marker LIM_REG_CLASSES which are
974    supersets with given important class (the first index).  That
975    includes given class itself.  This is calculated taking only hard
976    registers available for allocation into account.  */
977 enum reg_class ira_reg_class_super_classes[N_REG_CLASSES][N_REG_CLASSES];
978 
979 /* The biggest important reg_class inside of union of the two
980    reg_classes (that is calculated taking only hard registers
981    available for allocation into account).  If the both reg_classes
982    contain no hard registers available for allocation, the value is
983    calculated by taking all hard-registers including fixed ones into
984    account.  In other words, the value is the corresponding
985    reg_class_subunion value.  */
986 enum reg_class ira_reg_class_union[N_REG_CLASSES][N_REG_CLASSES];
987 
988 /* Set up the above reg class relations.  */
989 static void
990 setup_reg_class_relations (void)
991 {
992   int i, cl1, cl2, cl3;
993   HARD_REG_SET intersection_set, union_set, temp_set2;
994   bool important_class_p[N_REG_CLASSES];
995 
996   memset (important_class_p, 0, sizeof (important_class_p));
997   for (i = 0; i < ira_important_classes_num; i++)
998     important_class_p[ira_important_classes[i]] = true;
999   for (cl1 = 0; cl1 < N_REG_CLASSES; cl1++)
1000     {
1001       ira_reg_class_super_classes[cl1][0] = LIM_REG_CLASSES;
1002       for (cl2 = 0; cl2 < N_REG_CLASSES; cl2++)
1003 	{
1004 	  ira_reg_classes_intersect_p[cl1][cl2] = false;
1005 	  ira_reg_class_intersect[cl1][cl2] = NO_REGS;
1006 	  COPY_HARD_REG_SET (temp_hard_regset, reg_class_contents[cl1]);
1007 	  AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
1008 	  COPY_HARD_REG_SET (temp_set2, reg_class_contents[cl2]);
1009 	  AND_COMPL_HARD_REG_SET (temp_set2, no_unit_alloc_regs);
1010 	  if (hard_reg_set_empty_p (temp_hard_regset)
1011 	      && hard_reg_set_empty_p (temp_set2))
1012 	    {
1013 	      for (i = 0;; i++)
1014 		{
1015 		  cl3 = reg_class_subclasses[cl1][i];
1016 		  if (cl3 == LIM_REG_CLASSES)
1017 		    break;
1018 		  if (reg_class_subset_p (ira_reg_class_intersect[cl1][cl2],
1019 					  (enum reg_class) cl3))
1020 		    ira_reg_class_intersect[cl1][cl2] = (enum reg_class) cl3;
1021 		}
1022 	      ira_reg_class_union[cl1][cl2] = reg_class_subunion[cl1][cl2];
1023 	      continue;
1024 	    }
1025 	  ira_reg_classes_intersect_p[cl1][cl2]
1026 	    = hard_reg_set_intersect_p (temp_hard_regset, temp_set2);
1027 	  if (important_class_p[cl1] && important_class_p[cl2]
1028 	      && hard_reg_set_subset_p (temp_hard_regset, temp_set2))
1029 	    {
1030 	      enum reg_class *p;
1031 
1032 	      p = &ira_reg_class_super_classes[cl1][0];
1033 	      while (*p != LIM_REG_CLASSES)
1034 		p++;
1035 	      *p++ = (enum reg_class) cl2;
1036 	      *p = LIM_REG_CLASSES;
1037 	    }
1038 	  ira_reg_class_union[cl1][cl2] = NO_REGS;
1039 	  COPY_HARD_REG_SET (intersection_set, reg_class_contents[cl1]);
1040 	  AND_HARD_REG_SET (intersection_set, reg_class_contents[cl2]);
1041 	  AND_COMPL_HARD_REG_SET (intersection_set, no_unit_alloc_regs);
1042 	  COPY_HARD_REG_SET (union_set, reg_class_contents[cl1]);
1043 	  IOR_HARD_REG_SET (union_set, reg_class_contents[cl2]);
1044 	  AND_COMPL_HARD_REG_SET (union_set, no_unit_alloc_regs);
1045 	  for (i = 0; i < ira_important_classes_num; i++)
1046 	    {
1047 	      cl3 = ira_important_classes[i];
1048 	      COPY_HARD_REG_SET (temp_hard_regset, reg_class_contents[cl3]);
1049 	      AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
1050 	      if (hard_reg_set_subset_p (temp_hard_regset, intersection_set))
1051 		{
1052 		  COPY_HARD_REG_SET
1053 		    (temp_set2,
1054 		     reg_class_contents[(int)
1055 					ira_reg_class_intersect[cl1][cl2]]);
1056 		  AND_COMPL_HARD_REG_SET (temp_set2, no_unit_alloc_regs);
1057 	 	  if (! hard_reg_set_subset_p (temp_hard_regset, temp_set2)
1058 		      /* Ignore unavailable hard registers and prefer
1059 			 smallest class for debugging purposes.  */
1060 		      || (hard_reg_set_equal_p (temp_hard_regset, temp_set2)
1061 			  && hard_reg_set_subset_p
1062 			     (reg_class_contents[cl3],
1063 			      reg_class_contents
1064 			      [(int) ira_reg_class_intersect[cl1][cl2]])))
1065 		    ira_reg_class_intersect[cl1][cl2] = (enum reg_class) cl3;
1066 		}
1067 	      if (hard_reg_set_subset_p (temp_hard_regset, union_set))
1068 		{
1069 		  COPY_HARD_REG_SET
1070 		    (temp_set2,
1071 		     reg_class_contents[(int) ira_reg_class_union[cl1][cl2]]);
1072 		  AND_COMPL_HARD_REG_SET (temp_set2, no_unit_alloc_regs);
1073 	 	  if (ira_reg_class_union[cl1][cl2] == NO_REGS
1074 		      || (hard_reg_set_subset_p (temp_set2, temp_hard_regset)
1075 
1076 			  && (! hard_reg_set_equal_p (temp_set2,
1077 						      temp_hard_regset)
1078 			      /* Ignore unavailable hard registers and
1079 				 prefer smallest class for debugging
1080 				 purposes.  */
1081 			      || hard_reg_set_subset_p
1082 			         (reg_class_contents[cl3],
1083 				  reg_class_contents
1084 				  [(int) ira_reg_class_union[cl1][cl2]]))))
1085 		    ira_reg_class_union[cl1][cl2] = (enum reg_class) cl3;
1086 		}
1087 	    }
1088 	}
1089     }
1090 }
1091 
1092 /* Output all cover classes and the translation map into file F.  */
1093 static void
1094 print_class_cover (FILE *f)
1095 {
1096   static const char *const reg_class_names[] = REG_CLASS_NAMES;
1097   int i;
1098 
1099   fprintf (f, "Class cover:\n");
1100   for (i = 0; i < ira_reg_class_cover_size; i++)
1101     fprintf (f, " %s", reg_class_names[ira_reg_class_cover[i]]);
1102   fprintf (f, "\nClass translation:\n");
1103   for (i = 0; i < N_REG_CLASSES; i++)
1104     fprintf (f, " %s -> %s\n", reg_class_names[i],
1105 	     reg_class_names[ira_class_translate[i]]);
1106 }
1107 
1108 /* Output all cover classes and the translation map into
1109    stderr.  */
1110 void
1111 ira_debug_class_cover (void)
1112 {
1113   print_class_cover (stderr);
1114 }
1115 
1116 /* Set up different arrays concerning class subsets, cover and
1117    important classes.  */
1118 static void
1119 find_reg_class_closure (void)
1120 {
1121   setup_reg_subclasses ();
1122   setup_cover_and_important_classes ();
1123   setup_class_translate ();
1124   reorder_important_classes ();
1125   setup_reg_class_relations ();
1126 }
1127 
1128 
1129 
1130 /* Map: hard register number -> cover class it belongs to.  If the
1131    corresponding class is NO_REGS, the hard register is not available
1132    for allocation.  */
1133 enum reg_class ira_hard_regno_cover_class[FIRST_PSEUDO_REGISTER];
1134 
1135 /* Set up the array above.  */
1136 static void
1137 setup_hard_regno_cover_class (void)
1138 {
1139   int i, j;
1140   enum reg_class cl;
1141 
1142   for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
1143     {
1144       ira_hard_regno_cover_class[i] = NO_REGS;
1145       for (j = 0; j < ira_reg_class_cover_size; j++)
1146 	{
1147 	  cl = ira_reg_class_cover[j];
1148 	  if (ira_class_hard_reg_index[cl][i] >= 0)
1149 	    {
1150 	      ira_hard_regno_cover_class[i] = cl;
1151 	      break;
1152 	    }
1153 	}
1154 
1155     }
1156 }
1157 
1158 
1159 
1160 /* Map: register class x machine mode -> number of hard registers of
1161    given class needed to store value of given mode.  If the number is
1162    different, the size will be negative.  */
1163 int ira_reg_class_nregs[N_REG_CLASSES][MAX_MACHINE_MODE];
1164 
1165 /* Maximal value of the previous array elements.  */
1166 int ira_max_nregs;
1167 
1168 /* Form IRA_REG_CLASS_NREGS map.  */
1169 static void
1170 setup_reg_class_nregs (void)
1171 {
1172   int cl, m;
1173 
1174   ira_max_nregs = -1;
1175   for (cl = 0; cl < N_REG_CLASSES; cl++)
1176     for (m = 0; m < MAX_MACHINE_MODE; m++)
1177       {
1178 	ira_reg_class_nregs[cl][m] = CLASS_MAX_NREGS ((enum reg_class) cl,
1179 						      (enum machine_mode) m);
1180 	if (ira_max_nregs < ira_reg_class_nregs[cl][m])
1181 	  ira_max_nregs = ira_reg_class_nregs[cl][m];
1182       }
1183 }
1184 
1185 
1186 
1187 /* Array whose values are hard regset of hard registers available for
1188    the allocation of given register class whose HARD_REGNO_MODE_OK
1189    values for given mode are zero.  */
1190 HARD_REG_SET prohibited_class_mode_regs[N_REG_CLASSES][NUM_MACHINE_MODES];
1191 
1192 /* Set up PROHIBITED_CLASS_MODE_REGS.  */
1193 static void
1194 setup_prohibited_class_mode_regs (void)
1195 {
1196   int i, j, k, hard_regno;
1197   enum reg_class cl;
1198 
1199   for (i = 0; i < ira_reg_class_cover_size; i++)
1200     {
1201       cl = ira_reg_class_cover[i];
1202       for (j = 0; j < NUM_MACHINE_MODES; j++)
1203 	{
1204 	  CLEAR_HARD_REG_SET (prohibited_class_mode_regs[cl][j]);
1205 	  for (k = ira_class_hard_regs_num[cl] - 1; k >= 0; k--)
1206 	    {
1207 	      hard_regno = ira_class_hard_regs[cl][k];
1208 	      if (! HARD_REGNO_MODE_OK (hard_regno, (enum machine_mode) j))
1209 		SET_HARD_REG_BIT (prohibited_class_mode_regs[cl][j],
1210 				  hard_regno);
1211 	    }
1212 	}
1213     }
1214 }
1215 
1216 
1217 
1218 /* Allocate and initialize IRA_REGISTER_MOVE_COST,
1219    IRA_MAY_MOVE_IN_COST, and IRA_MAY_MOVE_OUT_COST for MODE if it is
1220    not done yet.  */
1221 void
1222 ira_init_register_move_cost (enum machine_mode mode)
1223 {
1224   int cl1, cl2;
1225 
1226   ira_assert (ira_register_move_cost[mode] == NULL
1227 	      && ira_may_move_in_cost[mode] == NULL
1228 	      && ira_may_move_out_cost[mode] == NULL);
1229   if (move_cost[mode] == NULL)
1230     init_move_cost (mode);
1231   ira_register_move_cost[mode] = move_cost[mode];
1232   /* Don't use ira_allocate because the tables exist out of scope of a
1233      IRA call.  */
1234   ira_may_move_in_cost[mode]
1235     = (move_table *) xmalloc (sizeof (move_table) * N_REG_CLASSES);
1236   memcpy (ira_may_move_in_cost[mode], may_move_in_cost[mode],
1237 	  sizeof (move_table) * N_REG_CLASSES);
1238   ira_may_move_out_cost[mode]
1239     = (move_table *) xmalloc (sizeof (move_table) * N_REG_CLASSES);
1240   memcpy (ira_may_move_out_cost[mode], may_move_out_cost[mode],
1241 	  sizeof (move_table) * N_REG_CLASSES);
1242   for (cl1 = 0; cl1 < N_REG_CLASSES; cl1++)
1243     {
1244       for (cl2 = 0; cl2 < N_REG_CLASSES; cl2++)
1245 	{
1246 	  if (ira_class_subset_p[cl1][cl2])
1247 	    ira_may_move_in_cost[mode][cl1][cl2] = 0;
1248 	  if (ira_class_subset_p[cl2][cl1])
1249 	    ira_may_move_out_cost[mode][cl1][cl2] = 0;
1250 	}
1251     }
1252 }
1253 
1254 
1255 
1256 /* This is called once during compiler work.  It sets up
1257    different arrays whose values don't depend on the compiled
1258    function.  */
1259 void
1260 ira_init_once (void)
1261 {
1262   int mode;
1263 
1264   for (mode = 0; mode < MAX_MACHINE_MODE; mode++)
1265     {
1266       ira_register_move_cost[mode] = NULL;
1267       ira_may_move_in_cost[mode] = NULL;
1268       ira_may_move_out_cost[mode] = NULL;
1269     }
1270   ira_init_costs_once ();
1271 }
1272 
1273 /* Free ira_register_move_cost, ira_may_move_in_cost, and
1274    ira_may_move_out_cost for each mode.  */
1275 static void
1276 free_register_move_costs (void)
1277 {
1278   int mode;
1279 
1280   for (mode = 0; mode < MAX_MACHINE_MODE; mode++)
1281     {
1282       if (ira_may_move_in_cost[mode] != NULL)
1283 	free (ira_may_move_in_cost[mode]);
1284       if (ira_may_move_out_cost[mode] != NULL)
1285 	free (ira_may_move_out_cost[mode]);
1286       ira_register_move_cost[mode] = NULL;
1287       ira_may_move_in_cost[mode] = NULL;
1288       ira_may_move_out_cost[mode] = NULL;
1289     }
1290 }
1291 
1292 /* This is called every time when register related information is
1293    changed.  */
1294 void
1295 ira_init (void)
1296 {
1297   free_register_move_costs ();
1298   setup_reg_mode_hard_regset ();
1299   setup_alloc_regs (flag_omit_frame_pointer != 0);
1300   setup_class_subset_and_memory_move_costs ();
1301   find_reg_class_closure ();
1302   setup_hard_regno_cover_class ();
1303   setup_reg_class_nregs ();
1304   setup_prohibited_class_mode_regs ();
1305   ira_init_costs ();
1306 }
1307 
1308 /* Function called once at the end of compiler work.  */
1309 void
1310 ira_finish_once (void)
1311 {
1312   ira_finish_costs_once ();
1313   free_register_move_costs ();
1314 }
1315 
1316 
1317 
1318 /* Array whose values are hard regset of hard registers for which
1319    move of the hard register in given mode into itself is
1320    prohibited.  */
1321 HARD_REG_SET ira_prohibited_mode_move_regs[NUM_MACHINE_MODES];
1322 
1323 /* Flag of that the above array has been initialized.  */
1324 static bool ira_prohibited_mode_move_regs_initialized_p = false;
1325 
1326 /* Set up IRA_PROHIBITED_MODE_MOVE_REGS.  */
1327 static void
1328 setup_prohibited_mode_move_regs (void)
1329 {
1330   int i, j;
1331   rtx test_reg1, test_reg2, move_pat, move_insn;
1332 
1333   if (ira_prohibited_mode_move_regs_initialized_p)
1334     return;
1335   ira_prohibited_mode_move_regs_initialized_p = true;
1336   test_reg1 = gen_rtx_REG (VOIDmode, 0);
1337   test_reg2 = gen_rtx_REG (VOIDmode, 0);
1338   move_pat = gen_rtx_SET (VOIDmode, test_reg1, test_reg2);
1339   move_insn = gen_rtx_INSN (VOIDmode, 0, 0, 0, 0, 0, move_pat, -1, 0);
1340   for (i = 0; i < NUM_MACHINE_MODES; i++)
1341     {
1342       SET_HARD_REG_SET (ira_prohibited_mode_move_regs[i]);
1343       for (j = 0; j < FIRST_PSEUDO_REGISTER; j++)
1344 	{
1345 	  if (! HARD_REGNO_MODE_OK (j, (enum machine_mode) i))
1346 	    continue;
1347 	  SET_REGNO (test_reg1, j);
1348 	  PUT_MODE (test_reg1, (enum machine_mode) i);
1349 	  SET_REGNO (test_reg2, j);
1350 	  PUT_MODE (test_reg2, (enum machine_mode) i);
1351 	  INSN_CODE (move_insn) = -1;
1352 	  recog_memoized (move_insn);
1353 	  if (INSN_CODE (move_insn) < 0)
1354 	    continue;
1355 	  extract_insn (move_insn);
1356 	  if (! constrain_operands (1))
1357 	    continue;
1358 	  CLEAR_HARD_REG_BIT (ira_prohibited_mode_move_regs[i], j);
1359 	}
1360     }
1361 }
1362 
1363 
1364 
1365 /* Function specific hard registers that can not be used for the
1366    register allocation.  */
1367 HARD_REG_SET ira_no_alloc_regs;
1368 
1369 /* Return TRUE if *LOC contains an asm.  */
1370 static int
1371 insn_contains_asm_1 (rtx *loc, void *data ATTRIBUTE_UNUSED)
1372 {
1373   if ( !*loc)
1374     return FALSE;
1375   if (GET_CODE (*loc) == ASM_OPERANDS)
1376     return TRUE;
1377   return FALSE;
1378 }
1379 
1380 
1381 /* Return TRUE if INSN contains an ASM.  */
1382 static bool
1383 insn_contains_asm (rtx insn)
1384 {
1385   return for_each_rtx (&insn, insn_contains_asm_1, NULL);
1386 }
1387 
1388 /* Set up regs_asm_clobbered.  */
1389 static void
1390 compute_regs_asm_clobbered (char *regs_asm_clobbered)
1391 {
1392   basic_block bb;
1393 
1394   memset (regs_asm_clobbered, 0, sizeof (char) * FIRST_PSEUDO_REGISTER);
1395 
1396   FOR_EACH_BB (bb)
1397     {
1398       rtx insn;
1399       FOR_BB_INSNS_REVERSE (bb, insn)
1400 	{
1401 	  df_ref *def_rec;
1402 
1403 	  if (insn_contains_asm (insn))
1404 	    for (def_rec = DF_INSN_DEFS (insn); *def_rec; def_rec++)
1405 	      {
1406 		df_ref def = *def_rec;
1407 		unsigned int dregno = DF_REF_REGNO (def);
1408 		if (dregno < FIRST_PSEUDO_REGISTER)
1409 		  {
1410 		    unsigned int i;
1411 		    enum machine_mode mode = GET_MODE (DF_REF_REAL_REG (def));
1412 		    unsigned int end = dregno
1413 		      + hard_regno_nregs[dregno][mode] - 1;
1414 
1415 		    for (i = dregno; i <= end; ++i)
1416 		      regs_asm_clobbered[i] = 1;
1417 		  }
1418 	      }
1419 	}
1420     }
1421 }
1422 
1423 
1424 /* Set up ELIMINABLE_REGSET, IRA_NO_ALLOC_REGS, and REGS_EVER_LIVE.  */
1425 void
1426 ira_setup_eliminable_regset (void)
1427 {
1428   /* Like regs_ever_live, but 1 if a reg is set or clobbered from an
1429      asm.  Unlike regs_ever_live, elements of this array corresponding
1430      to eliminable regs (like the frame pointer) are set if an asm
1431      sets them.  */
1432   char *regs_asm_clobbered
1433     = (char *) alloca (FIRST_PSEUDO_REGISTER * sizeof (char));
1434 #ifdef ELIMINABLE_REGS
1435   int i;
1436   static const struct {const int from, to; } eliminables[] = ELIMINABLE_REGS;
1437 #endif
1438   /* FIXME: If EXIT_IGNORE_STACK is set, we will not save and restore
1439      sp for alloca.  So we can't eliminate the frame pointer in that
1440      case.  At some point, we should improve this by emitting the
1441      sp-adjusting insns for this case.  */
1442   int need_fp
1443     = (! flag_omit_frame_pointer
1444        || (cfun->calls_alloca && EXIT_IGNORE_STACK)
1445        /* We need the frame pointer to catch stack overflow exceptions
1446 	  if the stack pointer is moving.  */
1447        || (flag_stack_check && STACK_CHECK_MOVING_SP)
1448        || crtl->accesses_prior_frames
1449        || crtl->stack_realign_needed
1450        || targetm.frame_pointer_required ());
1451 
1452   frame_pointer_needed = need_fp;
1453 
1454   COPY_HARD_REG_SET (ira_no_alloc_regs, no_unit_alloc_regs);
1455   CLEAR_HARD_REG_SET (eliminable_regset);
1456 
1457   compute_regs_asm_clobbered (regs_asm_clobbered);
1458   /* Build the regset of all eliminable registers and show we can't
1459      use those that we already know won't be eliminated.  */
1460 #ifdef ELIMINABLE_REGS
1461   for (i = 0; i < (int) ARRAY_SIZE (eliminables); i++)
1462     {
1463       bool cannot_elim
1464 	= (! targetm.can_eliminate (eliminables[i].from, eliminables[i].to)
1465 	   || (eliminables[i].to == STACK_POINTER_REGNUM && need_fp));
1466 
1467       if (! regs_asm_clobbered[eliminables[i].from])
1468 	{
1469 	    SET_HARD_REG_BIT (eliminable_regset, eliminables[i].from);
1470 
1471 	    if (cannot_elim)
1472 	      SET_HARD_REG_BIT (ira_no_alloc_regs, eliminables[i].from);
1473 	}
1474       else if (cannot_elim)
1475 	error ("%s cannot be used in asm here",
1476 	       reg_names[eliminables[i].from]);
1477       else
1478 	df_set_regs_ever_live (eliminables[i].from, true);
1479     }
1480 #if FRAME_POINTER_REGNUM != HARD_FRAME_POINTER_REGNUM
1481   if (! regs_asm_clobbered[HARD_FRAME_POINTER_REGNUM])
1482     {
1483       SET_HARD_REG_BIT (eliminable_regset, HARD_FRAME_POINTER_REGNUM);
1484       if (need_fp)
1485 	SET_HARD_REG_BIT (ira_no_alloc_regs, HARD_FRAME_POINTER_REGNUM);
1486     }
1487   else if (need_fp)
1488     error ("%s cannot be used in asm here",
1489 	   reg_names[HARD_FRAME_POINTER_REGNUM]);
1490   else
1491     df_set_regs_ever_live (HARD_FRAME_POINTER_REGNUM, true);
1492 #endif
1493 
1494 #else
1495   if (! regs_asm_clobbered[FRAME_POINTER_REGNUM])
1496     {
1497       SET_HARD_REG_BIT (eliminable_regset, FRAME_POINTER_REGNUM);
1498       if (need_fp)
1499 	SET_HARD_REG_BIT (ira_no_alloc_regs, FRAME_POINTER_REGNUM);
1500     }
1501   else if (need_fp)
1502     error ("%s cannot be used in asm here", reg_names[FRAME_POINTER_REGNUM]);
1503   else
1504     df_set_regs_ever_live (FRAME_POINTER_REGNUM, true);
1505 #endif
1506 }
1507 
1508 
1509 
1510 /* The length of the following two arrays.  */
1511 int ira_reg_equiv_len;
1512 
1513 /* The element value is TRUE if the corresponding regno value is
1514    invariant.  */
1515 bool *ira_reg_equiv_invariant_p;
1516 
1517 /* The element value is equiv constant of given pseudo-register or
1518    NULL_RTX.  */
1519 rtx *ira_reg_equiv_const;
1520 
1521 /* Set up the two arrays declared above.  */
1522 static void
1523 find_reg_equiv_invariant_const (void)
1524 {
1525   int i;
1526   bool invariant_p;
1527   rtx list, insn, note, constant, x;
1528 
1529   for (i = FIRST_PSEUDO_REGISTER; i < reg_equiv_init_size; i++)
1530     {
1531       constant = NULL_RTX;
1532       invariant_p = false;
1533       for (list = reg_equiv_init[i]; list != NULL_RTX; list = XEXP (list, 1))
1534 	{
1535 	  insn = XEXP (list, 0);
1536 	  note = find_reg_note (insn, REG_EQUIV, NULL_RTX);
1537 
1538 	  if (note == NULL_RTX)
1539 	    continue;
1540 
1541 	  x = XEXP (note, 0);
1542 
1543 	  if (! function_invariant_p (x)
1544 	      || ! flag_pic
1545 	      /* A function invariant is often CONSTANT_P but may
1546 		 include a register.  We promise to only pass CONSTANT_P
1547 		 objects to LEGITIMATE_PIC_OPERAND_P.  */
1548 	      || (CONSTANT_P (x) && LEGITIMATE_PIC_OPERAND_P (x)))
1549 	    {
1550 	      /* It can happen that a REG_EQUIV note contains a MEM
1551 		 that is not a legitimate memory operand.  As later
1552 		 stages of the reload assume that all addresses found
1553 		 in the reg_equiv_* arrays were originally legitimate,
1554 		 we ignore such REG_EQUIV notes.  */
1555 	      if (memory_operand (x, VOIDmode))
1556 		invariant_p = MEM_READONLY_P (x);
1557 	      else if (function_invariant_p (x))
1558 		{
1559 		  if (GET_CODE (x) == PLUS
1560 		      || x == frame_pointer_rtx || x == arg_pointer_rtx)
1561 		    invariant_p = true;
1562 		  else
1563 		    constant = x;
1564 		}
1565 	    }
1566 	}
1567       ira_reg_equiv_invariant_p[i] = invariant_p;
1568       ira_reg_equiv_const[i] = constant;
1569     }
1570 }
1571 
1572 
1573 
1574 /* Vector of substitutions of register numbers,
1575    used to map pseudo regs into hardware regs.
1576    This is set up as a result of register allocation.
1577    Element N is the hard reg assigned to pseudo reg N,
1578    or is -1 if no hard reg was assigned.
1579    If N is a hard reg number, element N is N.  */
1580 short *reg_renumber;
1581 
1582 /* Set up REG_RENUMBER and CALLER_SAVE_NEEDED (used by reload) from
1583    the allocation found by IRA.  */
1584 static void
1585 setup_reg_renumber (void)
1586 {
1587   int regno, hard_regno;
1588   ira_allocno_t a;
1589   ira_allocno_iterator ai;
1590 
1591   caller_save_needed = 0;
1592   FOR_EACH_ALLOCNO (a, ai)
1593     {
1594       /* There are no caps at this point.  */
1595       ira_assert (ALLOCNO_CAP_MEMBER (a) == NULL);
1596       if (! ALLOCNO_ASSIGNED_P (a))
1597 	/* It can happen if A is not referenced but partially anticipated
1598 	   somewhere in a region.  */
1599 	ALLOCNO_ASSIGNED_P (a) = true;
1600       ira_free_allocno_updated_costs (a);
1601       hard_regno = ALLOCNO_HARD_REGNO (a);
1602       regno = (int) REGNO (ALLOCNO_REG (a));
1603       reg_renumber[regno] = (hard_regno < 0 ? -1 : hard_regno);
1604       if (hard_regno >= 0 && ALLOCNO_CALLS_CROSSED_NUM (a) != 0
1605 	  && ! ira_hard_reg_not_in_set_p (hard_regno, ALLOCNO_MODE (a),
1606 					  call_used_reg_set))
1607 	{
1608 	  ira_assert (!optimize || flag_caller_saves
1609 		      || regno >= ira_reg_equiv_len
1610 		      || ira_reg_equiv_const[regno]
1611 		      || ira_reg_equiv_invariant_p[regno]);
1612 	  caller_save_needed = 1;
1613 	}
1614     }
1615 }
1616 
1617 /* Set up allocno assignment flags for further allocation
1618    improvements.  */
1619 static void
1620 setup_allocno_assignment_flags (void)
1621 {
1622   int hard_regno;
1623   ira_allocno_t a;
1624   ira_allocno_iterator ai;
1625 
1626   FOR_EACH_ALLOCNO (a, ai)
1627     {
1628       if (! ALLOCNO_ASSIGNED_P (a))
1629 	/* It can happen if A is not referenced but partially anticipated
1630 	   somewhere in a region.  */
1631 	ira_free_allocno_updated_costs (a);
1632       hard_regno = ALLOCNO_HARD_REGNO (a);
1633       /* Don't assign hard registers to allocnos which are destination
1634 	 of removed store at the end of loop.  It has no sense to keep
1635 	 the same value in different hard registers.  It is also
1636 	 impossible to assign hard registers correctly to such
1637 	 allocnos because the cost info and info about intersected
1638 	 calls are incorrect for them.  */
1639       ALLOCNO_ASSIGNED_P (a) = (hard_regno >= 0
1640 				|| ALLOCNO_MEM_OPTIMIZED_DEST_P (a)
1641 				|| (ALLOCNO_MEMORY_COST (a)
1642 				    - ALLOCNO_COVER_CLASS_COST (a)) < 0);
1643       ira_assert (hard_regno < 0
1644 		  || ! ira_hard_reg_not_in_set_p (hard_regno, ALLOCNO_MODE (a),
1645 						  reg_class_contents
1646 						  [ALLOCNO_COVER_CLASS (a)]));
1647     }
1648 }
1649 
1650 /* Evaluate overall allocation cost and the costs for using hard
1651    registers and memory for allocnos.  */
1652 static void
1653 calculate_allocation_cost (void)
1654 {
1655   int hard_regno, cost;
1656   ira_allocno_t a;
1657   ira_allocno_iterator ai;
1658 
1659   ira_overall_cost = ira_reg_cost = ira_mem_cost = 0;
1660   FOR_EACH_ALLOCNO (a, ai)
1661     {
1662       hard_regno = ALLOCNO_HARD_REGNO (a);
1663       ira_assert (hard_regno < 0
1664 		  || ! ira_hard_reg_not_in_set_p
1665 		       (hard_regno, ALLOCNO_MODE (a),
1666 			reg_class_contents[ALLOCNO_COVER_CLASS (a)]));
1667       if (hard_regno < 0)
1668 	{
1669 	  cost = ALLOCNO_MEMORY_COST (a);
1670 	  ira_mem_cost += cost;
1671 	}
1672       else if (ALLOCNO_HARD_REG_COSTS (a) != NULL)
1673 	{
1674 	  cost = (ALLOCNO_HARD_REG_COSTS (a)
1675 		  [ira_class_hard_reg_index
1676 		   [ALLOCNO_COVER_CLASS (a)][hard_regno]]);
1677 	  ira_reg_cost += cost;
1678 	}
1679       else
1680 	{
1681 	  cost = ALLOCNO_COVER_CLASS_COST (a);
1682 	  ira_reg_cost += cost;
1683 	}
1684       ira_overall_cost += cost;
1685     }
1686 
1687   if (internal_flag_ira_verbose > 0 && ira_dump_file != NULL)
1688     {
1689       fprintf (ira_dump_file,
1690 	       "+++Costs: overall %d, reg %d, mem %d, ld %d, st %d, move %d\n",
1691 	       ira_overall_cost, ira_reg_cost, ira_mem_cost,
1692 	       ira_load_cost, ira_store_cost, ira_shuffle_cost);
1693       fprintf (ira_dump_file, "+++       move loops %d, new jumps %d\n",
1694 	       ira_move_loops_num, ira_additional_jumps_num);
1695     }
1696 
1697 }
1698 
1699 #ifdef ENABLE_IRA_CHECKING
1700 /* Check the correctness of the allocation.  We do need this because
1701    of complicated code to transform more one region internal
1702    representation into one region representation.  */
1703 static void
1704 check_allocation (void)
1705 {
1706   ira_allocno_t a, conflict_a;
1707   int hard_regno, conflict_hard_regno, nregs, conflict_nregs;
1708   ira_allocno_conflict_iterator aci;
1709   ira_allocno_iterator ai;
1710 
1711   FOR_EACH_ALLOCNO (a, ai)
1712     {
1713       if (ALLOCNO_CAP_MEMBER (a) != NULL
1714 	  || (hard_regno = ALLOCNO_HARD_REGNO (a)) < 0)
1715 	continue;
1716       nregs = hard_regno_nregs[hard_regno][ALLOCNO_MODE (a)];
1717       FOR_EACH_ALLOCNO_CONFLICT (a, conflict_a, aci)
1718 	if ((conflict_hard_regno = ALLOCNO_HARD_REGNO (conflict_a)) >= 0)
1719 	  {
1720 	    conflict_nregs
1721 	      = (hard_regno_nregs
1722 		 [conflict_hard_regno][ALLOCNO_MODE (conflict_a)]);
1723 	    if ((conflict_hard_regno <= hard_regno
1724 		 && hard_regno < conflict_hard_regno + conflict_nregs)
1725 		|| (hard_regno <= conflict_hard_regno
1726 		    && conflict_hard_regno < hard_regno + nregs))
1727 	      {
1728 		fprintf (stderr, "bad allocation for %d and %d\n",
1729 			 ALLOCNO_REGNO (a), ALLOCNO_REGNO (conflict_a));
1730 		gcc_unreachable ();
1731 	      }
1732 	  }
1733     }
1734 }
1735 #endif
1736 
1737 /* Fix values of array REG_EQUIV_INIT after live range splitting done
1738    by IRA.  */
1739 static void
1740 fix_reg_equiv_init (void)
1741 {
1742   int max_regno = max_reg_num ();
1743   int i, new_regno;
1744   rtx x, prev, next, insn, set;
1745 
1746   if (reg_equiv_init_size < max_regno)
1747     {
1748       reg_equiv_init
1749 	= (rtx *) ggc_realloc (reg_equiv_init, max_regno * sizeof (rtx));
1750       while (reg_equiv_init_size < max_regno)
1751 	reg_equiv_init[reg_equiv_init_size++] = NULL_RTX;
1752       for (i = FIRST_PSEUDO_REGISTER; i < reg_equiv_init_size; i++)
1753 	for (prev = NULL_RTX, x = reg_equiv_init[i]; x != NULL_RTX; x = next)
1754 	  {
1755 	    next = XEXP (x, 1);
1756 	    insn = XEXP (x, 0);
1757 	    set = single_set (insn);
1758 	    ira_assert (set != NULL_RTX
1759 			&& (REG_P (SET_DEST (set)) || REG_P (SET_SRC (set))));
1760 	    if (REG_P (SET_DEST (set))
1761 		&& ((int) REGNO (SET_DEST (set)) == i
1762 		    || (int) ORIGINAL_REGNO (SET_DEST (set)) == i))
1763 	      new_regno = REGNO (SET_DEST (set));
1764 	    else if (REG_P (SET_SRC (set))
1765 		     && ((int) REGNO (SET_SRC (set)) == i
1766 			 || (int) ORIGINAL_REGNO (SET_SRC (set)) == i))
1767 	      new_regno = REGNO (SET_SRC (set));
1768 	    else
1769  	      gcc_unreachable ();
1770 	    if (new_regno == i)
1771 	      prev = x;
1772 	    else
1773 	      {
1774 		if (prev == NULL_RTX)
1775 		  reg_equiv_init[i] = next;
1776 		else
1777 		  XEXP (prev, 1) = next;
1778 		XEXP (x, 1) = reg_equiv_init[new_regno];
1779 		reg_equiv_init[new_regno] = x;
1780 	      }
1781 	  }
1782     }
1783 }
1784 
1785 #ifdef ENABLE_IRA_CHECKING
1786 /* Print redundant memory-memory copies.  */
1787 static void
1788 print_redundant_copies (void)
1789 {
1790   int hard_regno;
1791   ira_allocno_t a;
1792   ira_copy_t cp, next_cp;
1793   ira_allocno_iterator ai;
1794 
1795   FOR_EACH_ALLOCNO (a, ai)
1796     {
1797       if (ALLOCNO_CAP_MEMBER (a) != NULL)
1798 	/* It is a cap. */
1799 	continue;
1800       hard_regno = ALLOCNO_HARD_REGNO (a);
1801       if (hard_regno >= 0)
1802 	continue;
1803       for (cp = ALLOCNO_COPIES (a); cp != NULL; cp = next_cp)
1804 	if (cp->first == a)
1805 	  next_cp = cp->next_first_allocno_copy;
1806 	else
1807 	  {
1808 	    next_cp = cp->next_second_allocno_copy;
1809 	    if (internal_flag_ira_verbose > 4 && ira_dump_file != NULL
1810 		&& cp->insn != NULL_RTX
1811 		&& ALLOCNO_HARD_REGNO (cp->first) == hard_regno)
1812 	      fprintf (ira_dump_file,
1813 		       "        Redundant move from %d(freq %d):%d\n",
1814 		       INSN_UID (cp->insn), cp->freq, hard_regno);
1815 	  }
1816     }
1817 }
1818 #endif
1819 
1820 /* Setup preferred and alternative classes for new pseudo-registers
1821    created by IRA starting with START.  */
1822 static void
1823 setup_preferred_alternate_classes_for_new_pseudos (int start)
1824 {
1825   int i, old_regno;
1826   int max_regno = max_reg_num ();
1827 
1828   for (i = start; i < max_regno; i++)
1829     {
1830       old_regno = ORIGINAL_REGNO (regno_reg_rtx[i]);
1831       ira_assert (i != old_regno);
1832       setup_reg_classes (i, reg_preferred_class (old_regno),
1833 			 reg_alternate_class (old_regno),
1834 			 reg_cover_class (old_regno));
1835       if (internal_flag_ira_verbose > 2 && ira_dump_file != NULL)
1836 	fprintf (ira_dump_file,
1837 		 "    New r%d: setting preferred %s, alternative %s\n",
1838 		 i, reg_class_names[reg_preferred_class (old_regno)],
1839 		 reg_class_names[reg_alternate_class (old_regno)]);
1840     }
1841 }
1842 
1843 
1844 
1845 /* Regional allocation can create new pseudo-registers.  This function
1846    expands some arrays for pseudo-registers.  */
1847 static void
1848 expand_reg_info (int old_size)
1849 {
1850   int i;
1851   int size = max_reg_num ();
1852 
1853   resize_reg_info ();
1854   for (i = old_size; i < size; i++)
1855     setup_reg_classes (i, GENERAL_REGS, ALL_REGS, GENERAL_REGS);
1856 }
1857 
1858 /* Return TRUE if there is too high register pressure in the function.
1859    It is used to decide when stack slot sharing is worth to do.  */
1860 static bool
1861 too_high_register_pressure_p (void)
1862 {
1863   int i;
1864   enum reg_class cover_class;
1865 
1866   for (i = 0; i < ira_reg_class_cover_size; i++)
1867     {
1868       cover_class = ira_reg_class_cover[i];
1869       if (ira_loop_tree_root->reg_pressure[cover_class] > 10000)
1870 	return true;
1871     }
1872   return false;
1873 }
1874 
1875 
1876 
1877 /* Indicate that hard register number FROM was eliminated and replaced with
1878    an offset from hard register number TO.  The status of hard registers live
1879    at the start of a basic block is updated by replacing a use of FROM with
1880    a use of TO.  */
1881 
1882 void
1883 mark_elimination (int from, int to)
1884 {
1885   basic_block bb;
1886 
1887   FOR_EACH_BB (bb)
1888     {
1889       /* We don't use LIVE info in IRA.  */
1890       regset r = DF_LR_IN (bb);
1891 
1892       if (REGNO_REG_SET_P (r, from))
1893 	{
1894 	  CLEAR_REGNO_REG_SET (r, from);
1895 	  SET_REGNO_REG_SET (r, to);
1896 	}
1897     }
1898 }
1899 
1900 
1901 
1902 struct equivalence
1903 {
1904   /* Set when a REG_EQUIV note is found or created.  Use to
1905      keep track of what memory accesses might be created later,
1906      e.g. by reload.  */
1907   rtx replacement;
1908   rtx *src_p;
1909   /* The list of each instruction which initializes this register.  */
1910   rtx init_insns;
1911   /* Loop depth is used to recognize equivalences which appear
1912      to be present within the same loop (or in an inner loop).  */
1913   int loop_depth;
1914   /* Nonzero if this had a preexisting REG_EQUIV note.  */
1915   int is_arg_equivalence;
1916   /* Set when an attempt should be made to replace a register
1917      with the associated src_p entry.  */
1918   char replace;
1919 };
1920 
1921 /* reg_equiv[N] (where N is a pseudo reg number) is the equivalence
1922    structure for that register.  */
1923 static struct equivalence *reg_equiv;
1924 
1925 /* Used for communication between the following two functions: contains
1926    a MEM that we wish to ensure remains unchanged.  */
1927 static rtx equiv_mem;
1928 
1929 /* Set nonzero if EQUIV_MEM is modified.  */
1930 static int equiv_mem_modified;
1931 
1932 /* If EQUIV_MEM is modified by modifying DEST, indicate that it is modified.
1933    Called via note_stores.  */
1934 static void
1935 validate_equiv_mem_from_store (rtx dest, const_rtx set ATTRIBUTE_UNUSED,
1936 			       void *data ATTRIBUTE_UNUSED)
1937 {
1938   if ((REG_P (dest)
1939        && reg_overlap_mentioned_p (dest, equiv_mem))
1940       || (MEM_P (dest)
1941 	  && true_dependence (dest, VOIDmode, equiv_mem, rtx_varies_p)))
1942     equiv_mem_modified = 1;
1943 }
1944 
1945 /* Verify that no store between START and the death of REG invalidates
1946    MEMREF.  MEMREF is invalidated by modifying a register used in MEMREF,
1947    by storing into an overlapping memory location, or with a non-const
1948    CALL_INSN.
1949 
1950    Return 1 if MEMREF remains valid.  */
1951 static int
1952 validate_equiv_mem (rtx start, rtx reg, rtx memref)
1953 {
1954   rtx insn;
1955   rtx note;
1956 
1957   equiv_mem = memref;
1958   equiv_mem_modified = 0;
1959 
1960   /* If the memory reference has side effects or is volatile, it isn't a
1961      valid equivalence.  */
1962   if (side_effects_p (memref))
1963     return 0;
1964 
1965   for (insn = start; insn && ! equiv_mem_modified; insn = NEXT_INSN (insn))
1966     {
1967       if (! INSN_P (insn))
1968 	continue;
1969 
1970       if (find_reg_note (insn, REG_DEAD, reg))
1971 	return 1;
1972 
1973       if (CALL_P (insn) && ! MEM_READONLY_P (memref)
1974 	  && ! RTL_CONST_OR_PURE_CALL_P (insn))
1975 	return 0;
1976 
1977       note_stores (PATTERN (insn), validate_equiv_mem_from_store, NULL);
1978 
1979       /* If a register mentioned in MEMREF is modified via an
1980 	 auto-increment, we lose the equivalence.  Do the same if one
1981 	 dies; although we could extend the life, it doesn't seem worth
1982 	 the trouble.  */
1983 
1984       for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
1985 	if ((REG_NOTE_KIND (note) == REG_INC
1986 	     || REG_NOTE_KIND (note) == REG_DEAD)
1987 	    && REG_P (XEXP (note, 0))
1988 	    && reg_overlap_mentioned_p (XEXP (note, 0), memref))
1989 	  return 0;
1990     }
1991 
1992   return 0;
1993 }
1994 
1995 /* Returns zero if X is known to be invariant.  */
1996 static int
1997 equiv_init_varies_p (rtx x)
1998 {
1999   RTX_CODE code = GET_CODE (x);
2000   int i;
2001   const char *fmt;
2002 
2003   switch (code)
2004     {
2005     case MEM:
2006       return !MEM_READONLY_P (x) || equiv_init_varies_p (XEXP (x, 0));
2007 
2008     case CONST:
2009     case CONST_INT:
2010     case CONST_DOUBLE:
2011     case CONST_FIXED:
2012     case CONST_VECTOR:
2013     case SYMBOL_REF:
2014     case LABEL_REF:
2015       return 0;
2016 
2017     case REG:
2018       return reg_equiv[REGNO (x)].replace == 0 && rtx_varies_p (x, 0);
2019 
2020     case ASM_OPERANDS:
2021       if (MEM_VOLATILE_P (x))
2022 	return 1;
2023 
2024       /* Fall through.  */
2025 
2026     default:
2027       break;
2028     }
2029 
2030   fmt = GET_RTX_FORMAT (code);
2031   for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2032     if (fmt[i] == 'e')
2033       {
2034 	if (equiv_init_varies_p (XEXP (x, i)))
2035 	  return 1;
2036       }
2037     else if (fmt[i] == 'E')
2038       {
2039 	int j;
2040 	for (j = 0; j < XVECLEN (x, i); j++)
2041 	  if (equiv_init_varies_p (XVECEXP (x, i, j)))
2042 	    return 1;
2043       }
2044 
2045   return 0;
2046 }
2047 
2048 /* Returns nonzero if X (used to initialize register REGNO) is movable.
2049    X is only movable if the registers it uses have equivalent initializations
2050    which appear to be within the same loop (or in an inner loop) and movable
2051    or if they are not candidates for local_alloc and don't vary.  */
2052 static int
2053 equiv_init_movable_p (rtx x, int regno)
2054 {
2055   int i, j;
2056   const char *fmt;
2057   enum rtx_code code = GET_CODE (x);
2058 
2059   switch (code)
2060     {
2061     case SET:
2062       return equiv_init_movable_p (SET_SRC (x), regno);
2063 
2064     case CC0:
2065     case CLOBBER:
2066       return 0;
2067 
2068     case PRE_INC:
2069     case PRE_DEC:
2070     case POST_INC:
2071     case POST_DEC:
2072     case PRE_MODIFY:
2073     case POST_MODIFY:
2074       return 0;
2075 
2076     case REG:
2077       return (reg_equiv[REGNO (x)].loop_depth >= reg_equiv[regno].loop_depth
2078 	      && reg_equiv[REGNO (x)].replace)
2079 	     || (REG_BASIC_BLOCK (REGNO (x)) < NUM_FIXED_BLOCKS && ! rtx_varies_p (x, 0));
2080 
2081     case UNSPEC_VOLATILE:
2082       return 0;
2083 
2084     case ASM_OPERANDS:
2085       if (MEM_VOLATILE_P (x))
2086 	return 0;
2087 
2088       /* Fall through.  */
2089 
2090     default:
2091       break;
2092     }
2093 
2094   fmt = GET_RTX_FORMAT (code);
2095   for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2096     switch (fmt[i])
2097       {
2098       case 'e':
2099 	if (! equiv_init_movable_p (XEXP (x, i), regno))
2100 	  return 0;
2101 	break;
2102       case 'E':
2103 	for (j = XVECLEN (x, i) - 1; j >= 0; j--)
2104 	  if (! equiv_init_movable_p (XVECEXP (x, i, j), regno))
2105 	    return 0;
2106 	break;
2107       }
2108 
2109   return 1;
2110 }
2111 
2112 /* TRUE if X uses any registers for which reg_equiv[REGNO].replace is true.  */
2113 static int
2114 contains_replace_regs (rtx x)
2115 {
2116   int i, j;
2117   const char *fmt;
2118   enum rtx_code code = GET_CODE (x);
2119 
2120   switch (code)
2121     {
2122     case CONST_INT:
2123     case CONST:
2124     case LABEL_REF:
2125     case SYMBOL_REF:
2126     case CONST_DOUBLE:
2127     case CONST_FIXED:
2128     case CONST_VECTOR:
2129     case PC:
2130     case CC0:
2131     case HIGH:
2132       return 0;
2133 
2134     case REG:
2135       return reg_equiv[REGNO (x)].replace;
2136 
2137     default:
2138       break;
2139     }
2140 
2141   fmt = GET_RTX_FORMAT (code);
2142   for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2143     switch (fmt[i])
2144       {
2145       case 'e':
2146 	if (contains_replace_regs (XEXP (x, i)))
2147 	  return 1;
2148 	break;
2149       case 'E':
2150 	for (j = XVECLEN (x, i) - 1; j >= 0; j--)
2151 	  if (contains_replace_regs (XVECEXP (x, i, j)))
2152 	    return 1;
2153 	break;
2154       }
2155 
2156   return 0;
2157 }
2158 
2159 /* TRUE if X references a memory location that would be affected by a store
2160    to MEMREF.  */
2161 static int
2162 memref_referenced_p (rtx memref, rtx x)
2163 {
2164   int i, j;
2165   const char *fmt;
2166   enum rtx_code code = GET_CODE (x);
2167 
2168   switch (code)
2169     {
2170     case CONST_INT:
2171     case CONST:
2172     case LABEL_REF:
2173     case SYMBOL_REF:
2174     case CONST_DOUBLE:
2175     case CONST_FIXED:
2176     case CONST_VECTOR:
2177     case PC:
2178     case CC0:
2179     case HIGH:
2180     case LO_SUM:
2181       return 0;
2182 
2183     case REG:
2184       return (reg_equiv[REGNO (x)].replacement
2185 	      && memref_referenced_p (memref,
2186 				      reg_equiv[REGNO (x)].replacement));
2187 
2188     case MEM:
2189       if (true_dependence (memref, VOIDmode, x, rtx_varies_p))
2190 	return 1;
2191       break;
2192 
2193     case SET:
2194       /* If we are setting a MEM, it doesn't count (its address does), but any
2195 	 other SET_DEST that has a MEM in it is referencing the MEM.  */
2196       if (MEM_P (SET_DEST (x)))
2197 	{
2198 	  if (memref_referenced_p (memref, XEXP (SET_DEST (x), 0)))
2199 	    return 1;
2200 	}
2201       else if (memref_referenced_p (memref, SET_DEST (x)))
2202 	return 1;
2203 
2204       return memref_referenced_p (memref, SET_SRC (x));
2205 
2206     default:
2207       break;
2208     }
2209 
2210   fmt = GET_RTX_FORMAT (code);
2211   for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2212     switch (fmt[i])
2213       {
2214       case 'e':
2215 	if (memref_referenced_p (memref, XEXP (x, i)))
2216 	  return 1;
2217 	break;
2218       case 'E':
2219 	for (j = XVECLEN (x, i) - 1; j >= 0; j--)
2220 	  if (memref_referenced_p (memref, XVECEXP (x, i, j)))
2221 	    return 1;
2222 	break;
2223       }
2224 
2225   return 0;
2226 }
2227 
2228 /* TRUE if some insn in the range (START, END] references a memory location
2229    that would be affected by a store to MEMREF.  */
2230 static int
2231 memref_used_between_p (rtx memref, rtx start, rtx end)
2232 {
2233   rtx insn;
2234 
2235   for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2236        insn = NEXT_INSN (insn))
2237     {
2238       if (!NONDEBUG_INSN_P (insn))
2239 	continue;
2240 
2241       if (memref_referenced_p (memref, PATTERN (insn)))
2242 	return 1;
2243 
2244       /* Nonconst functions may access memory.  */
2245       if (CALL_P (insn) && (! RTL_CONST_CALL_P (insn)))
2246 	return 1;
2247     }
2248 
2249   return 0;
2250 }
2251 
2252 /* Mark REG as having no known equivalence.
2253    Some instructions might have been processed before and furnished
2254    with REG_EQUIV notes for this register; these notes will have to be
2255    removed.
2256    STORE is the piece of RTL that does the non-constant / conflicting
2257    assignment - a SET, CLOBBER or REG_INC note.  It is currently not used,
2258    but needs to be there because this function is called from note_stores.  */
2259 static void
2260 no_equiv (rtx reg, const_rtx store ATTRIBUTE_UNUSED, void *data ATTRIBUTE_UNUSED)
2261 {
2262   int regno;
2263   rtx list;
2264 
2265   if (!REG_P (reg))
2266     return;
2267   regno = REGNO (reg);
2268   list = reg_equiv[regno].init_insns;
2269   if (list == const0_rtx)
2270     return;
2271   reg_equiv[regno].init_insns = const0_rtx;
2272   reg_equiv[regno].replacement = NULL_RTX;
2273   /* This doesn't matter for equivalences made for argument registers, we
2274      should keep their initialization insns.  */
2275   if (reg_equiv[regno].is_arg_equivalence)
2276     return;
2277   reg_equiv_init[regno] = NULL_RTX;
2278   for (; list; list =  XEXP (list, 1))
2279     {
2280       rtx insn = XEXP (list, 0);
2281       remove_note (insn, find_reg_note (insn, REG_EQUIV, NULL_RTX));
2282     }
2283 }
2284 
2285 /* In DEBUG_INSN location adjust REGs from CLEARED_REGS bitmap to the
2286    equivalent replacement.  */
2287 
2288 static rtx
2289 adjust_cleared_regs (rtx loc, const_rtx old_rtx ATTRIBUTE_UNUSED, void *data)
2290 {
2291   if (REG_P (loc))
2292     {
2293       bitmap cleared_regs = (bitmap) data;
2294       if (bitmap_bit_p (cleared_regs, REGNO (loc)))
2295 	return simplify_replace_fn_rtx (*reg_equiv[REGNO (loc)].src_p,
2296 					NULL_RTX, adjust_cleared_regs, data);
2297     }
2298   return NULL_RTX;
2299 }
2300 
2301 /* Nonzero if we recorded an equivalence for a LABEL_REF.  */
2302 static int recorded_label_ref;
2303 
2304 /* Find registers that are equivalent to a single value throughout the
2305    compilation (either because they can be referenced in memory or are set once
2306    from a single constant).  Lower their priority for a register.
2307 
2308    If such a register is only referenced once, try substituting its value
2309    into the using insn.  If it succeeds, we can eliminate the register
2310    completely.
2311 
2312    Initialize the REG_EQUIV_INIT array of initializing insns.
2313 
2314    Return non-zero if jump label rebuilding should be done.  */
2315 static int
2316 update_equiv_regs (void)
2317 {
2318   rtx insn;
2319   basic_block bb;
2320   int loop_depth;
2321   bitmap cleared_regs;
2322 
2323   /* We need to keep track of whether or not we recorded a LABEL_REF so
2324      that we know if the jump optimizer needs to be rerun.  */
2325   recorded_label_ref = 0;
2326 
2327   reg_equiv = XCNEWVEC (struct equivalence, max_regno);
2328   reg_equiv_init = GGC_CNEWVEC (rtx, max_regno);
2329   reg_equiv_init_size = max_regno;
2330 
2331   init_alias_analysis ();
2332 
2333   /* Scan the insns and find which registers have equivalences.  Do this
2334      in a separate scan of the insns because (due to -fcse-follow-jumps)
2335      a register can be set below its use.  */
2336   FOR_EACH_BB (bb)
2337     {
2338       loop_depth = bb->loop_depth;
2339 
2340       for (insn = BB_HEAD (bb);
2341 	   insn != NEXT_INSN (BB_END (bb));
2342 	   insn = NEXT_INSN (insn))
2343 	{
2344 	  rtx note;
2345 	  rtx set;
2346 	  rtx dest, src;
2347 	  int regno;
2348 
2349 	  if (! INSN_P (insn))
2350 	    continue;
2351 
2352 	  for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
2353 	    if (REG_NOTE_KIND (note) == REG_INC)
2354 	      no_equiv (XEXP (note, 0), note, NULL);
2355 
2356 	  set = single_set (insn);
2357 
2358 	  /* If this insn contains more (or less) than a single SET,
2359 	     only mark all destinations as having no known equivalence.  */
2360 	  if (set == 0)
2361 	    {
2362 	      note_stores (PATTERN (insn), no_equiv, NULL);
2363 	      continue;
2364 	    }
2365 	  else if (GET_CODE (PATTERN (insn)) == PARALLEL)
2366 	    {
2367 	      int i;
2368 
2369 	      for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
2370 		{
2371 		  rtx part = XVECEXP (PATTERN (insn), 0, i);
2372 		  if (part != set)
2373 		    note_stores (part, no_equiv, NULL);
2374 		}
2375 	    }
2376 
2377 	  dest = SET_DEST (set);
2378 	  src = SET_SRC (set);
2379 
2380 	  /* See if this is setting up the equivalence between an argument
2381 	     register and its stack slot.  */
2382 	  note = find_reg_note (insn, REG_EQUIV, NULL_RTX);
2383 	  if (note)
2384 	    {
2385 	      gcc_assert (REG_P (dest));
2386 	      regno = REGNO (dest);
2387 
2388 	      /* Note that we don't want to clear reg_equiv_init even if there
2389 		 are multiple sets of this register.  */
2390 	      reg_equiv[regno].is_arg_equivalence = 1;
2391 
2392 	      /* Record for reload that this is an equivalencing insn.  */
2393 	      if (rtx_equal_p (src, XEXP (note, 0)))
2394 		reg_equiv_init[regno]
2395 		  = gen_rtx_INSN_LIST (VOIDmode, insn, reg_equiv_init[regno]);
2396 
2397 	      /* Continue normally in case this is a candidate for
2398 		 replacements.  */
2399 	    }
2400 
2401 	  if (!optimize)
2402 	    continue;
2403 
2404 	  /* We only handle the case of a pseudo register being set
2405 	     once, or always to the same value.  */
2406 	  /* ??? The mn10200 port breaks if we add equivalences for
2407 	     values that need an ADDRESS_REGS register and set them equivalent
2408 	     to a MEM of a pseudo.  The actual problem is in the over-conservative
2409 	     handling of INPADDR_ADDRESS / INPUT_ADDRESS / INPUT triples in
2410 	     calculate_needs, but we traditionally work around this problem
2411 	     here by rejecting equivalences when the destination is in a register
2412 	     that's likely spilled.  This is fragile, of course, since the
2413 	     preferred class of a pseudo depends on all instructions that set
2414 	     or use it.  */
2415 
2416 	  if (!REG_P (dest)
2417 	      || (regno = REGNO (dest)) < FIRST_PSEUDO_REGISTER
2418 	      || reg_equiv[regno].init_insns == const0_rtx
2419 	      || (CLASS_LIKELY_SPILLED_P (reg_preferred_class (regno))
2420 		  && MEM_P (src) && ! reg_equiv[regno].is_arg_equivalence))
2421 	    {
2422 	      /* This might be setting a SUBREG of a pseudo, a pseudo that is
2423 		 also set somewhere else to a constant.  */
2424 	      note_stores (set, no_equiv, NULL);
2425 	      continue;
2426 	    }
2427 
2428 	  note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
2429 
2430 	  /* cse sometimes generates function invariants, but doesn't put a
2431 	     REG_EQUAL note on the insn.  Since this note would be redundant,
2432 	     there's no point creating it earlier than here.  */
2433 	  if (! note && ! rtx_varies_p (src, 0))
2434 	    note = set_unique_reg_note (insn, REG_EQUAL, copy_rtx (src));
2435 
2436 	  /* Don't bother considering a REG_EQUAL note containing an EXPR_LIST
2437 	     since it represents a function call */
2438 	  if (note && GET_CODE (XEXP (note, 0)) == EXPR_LIST)
2439 	    note = NULL_RTX;
2440 
2441 	  if (DF_REG_DEF_COUNT (regno) != 1
2442 	      && (! note
2443 		  || rtx_varies_p (XEXP (note, 0), 0)
2444 		  || (reg_equiv[regno].replacement
2445 		      && ! rtx_equal_p (XEXP (note, 0),
2446 					reg_equiv[regno].replacement))))
2447 	    {
2448 	      no_equiv (dest, set, NULL);
2449 	      continue;
2450 	    }
2451 	  /* Record this insn as initializing this register.  */
2452 	  reg_equiv[regno].init_insns
2453 	    = gen_rtx_INSN_LIST (VOIDmode, insn, reg_equiv[regno].init_insns);
2454 
2455 	  /* If this register is known to be equal to a constant, record that
2456 	     it is always equivalent to the constant.  */
2457 	  if (DF_REG_DEF_COUNT (regno) == 1
2458 	      && note && ! rtx_varies_p (XEXP (note, 0), 0))
2459 	    {
2460 	      rtx note_value = XEXP (note, 0);
2461 	      remove_note (insn, note);
2462 	      set_unique_reg_note (insn, REG_EQUIV, note_value);
2463 	    }
2464 
2465 	  /* If this insn introduces a "constant" register, decrease the priority
2466 	     of that register.  Record this insn if the register is only used once
2467 	     more and the equivalence value is the same as our source.
2468 
2469 	     The latter condition is checked for two reasons:  First, it is an
2470 	     indication that it may be more efficient to actually emit the insn
2471 	     as written (if no registers are available, reload will substitute
2472 	     the equivalence).  Secondly, it avoids problems with any registers
2473 	     dying in this insn whose death notes would be missed.
2474 
2475 	     If we don't have a REG_EQUIV note, see if this insn is loading
2476 	     a register used only in one basic block from a MEM.  If so, and the
2477 	     MEM remains unchanged for the life of the register, add a REG_EQUIV
2478 	     note.  */
2479 
2480 	  note = find_reg_note (insn, REG_EQUIV, NULL_RTX);
2481 
2482 	  if (note == 0 && REG_BASIC_BLOCK (regno) >= NUM_FIXED_BLOCKS
2483 	      && MEM_P (SET_SRC (set))
2484 	      && validate_equiv_mem (insn, dest, SET_SRC (set)))
2485 	    note = set_unique_reg_note (insn, REG_EQUIV, copy_rtx (SET_SRC (set)));
2486 
2487 	  if (note)
2488 	    {
2489 	      int regno = REGNO (dest);
2490 	      rtx x = XEXP (note, 0);
2491 
2492 	      /* If we haven't done so, record for reload that this is an
2493 		 equivalencing insn.  */
2494 	      if (!reg_equiv[regno].is_arg_equivalence)
2495 		reg_equiv_init[regno]
2496 		  = gen_rtx_INSN_LIST (VOIDmode, insn, reg_equiv_init[regno]);
2497 
2498 	      /* Record whether or not we created a REG_EQUIV note for a LABEL_REF.
2499 		 We might end up substituting the LABEL_REF for uses of the
2500 		 pseudo here or later.  That kind of transformation may turn an
2501 		 indirect jump into a direct jump, in which case we must rerun the
2502 		 jump optimizer to ensure that the JUMP_LABEL fields are valid.  */
2503 	      if (GET_CODE (x) == LABEL_REF
2504 		  || (GET_CODE (x) == CONST
2505 		      && GET_CODE (XEXP (x, 0)) == PLUS
2506 		      && (GET_CODE (XEXP (XEXP (x, 0), 0)) == LABEL_REF)))
2507 		recorded_label_ref = 1;
2508 
2509 	      reg_equiv[regno].replacement = x;
2510 	      reg_equiv[regno].src_p = &SET_SRC (set);
2511 	      reg_equiv[regno].loop_depth = loop_depth;
2512 
2513 	      /* Don't mess with things live during setjmp.  */
2514 	      if (REG_LIVE_LENGTH (regno) >= 0 && optimize)
2515 		{
2516 		  /* Note that the statement below does not affect the priority
2517 		     in local-alloc!  */
2518 		  REG_LIVE_LENGTH (regno) *= 2;
2519 
2520 		  /* If the register is referenced exactly twice, meaning it is
2521 		     set once and used once, indicate that the reference may be
2522 		     replaced by the equivalence we computed above.  Do this
2523 		     even if the register is only used in one block so that
2524 		     dependencies can be handled where the last register is
2525 		     used in a different block (i.e. HIGH / LO_SUM sequences)
2526 		     and to reduce the number of registers alive across
2527 		     calls.  */
2528 
2529 		  if (REG_N_REFS (regno) == 2
2530 		      && (rtx_equal_p (x, src)
2531 			  || ! equiv_init_varies_p (src))
2532 		      && NONJUMP_INSN_P (insn)
2533 		      && equiv_init_movable_p (PATTERN (insn), regno))
2534 		    reg_equiv[regno].replace = 1;
2535 		}
2536 	    }
2537 	}
2538     }
2539 
2540   if (!optimize)
2541     goto out;
2542 
2543   /* A second pass, to gather additional equivalences with memory.  This needs
2544      to be done after we know which registers we are going to replace.  */
2545 
2546   for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
2547     {
2548       rtx set, src, dest;
2549       unsigned regno;
2550 
2551       if (! INSN_P (insn))
2552 	continue;
2553 
2554       set = single_set (insn);
2555       if (! set)
2556 	continue;
2557 
2558       dest = SET_DEST (set);
2559       src = SET_SRC (set);
2560 
2561       /* If this sets a MEM to the contents of a REG that is only used
2562 	 in a single basic block, see if the register is always equivalent
2563 	 to that memory location and if moving the store from INSN to the
2564 	 insn that set REG is safe.  If so, put a REG_EQUIV note on the
2565 	 initializing insn.
2566 
2567 	 Don't add a REG_EQUIV note if the insn already has one.  The existing
2568 	 REG_EQUIV is likely more useful than the one we are adding.
2569 
2570 	 If one of the regs in the address has reg_equiv[REGNO].replace set,
2571 	 then we can't add this REG_EQUIV note.  The reg_equiv[REGNO].replace
2572 	 optimization may move the set of this register immediately before
2573 	 insn, which puts it after reg_equiv[REGNO].init_insns, and hence
2574 	 the mention in the REG_EQUIV note would be to an uninitialized
2575 	 pseudo.  */
2576 
2577       if (MEM_P (dest) && REG_P (src)
2578 	  && (regno = REGNO (src)) >= FIRST_PSEUDO_REGISTER
2579 	  && REG_BASIC_BLOCK (regno) >= NUM_FIXED_BLOCKS
2580 	  && DF_REG_DEF_COUNT (regno) == 1
2581 	  && reg_equiv[regno].init_insns != 0
2582 	  && reg_equiv[regno].init_insns != const0_rtx
2583 	  && ! find_reg_note (XEXP (reg_equiv[regno].init_insns, 0),
2584 			      REG_EQUIV, NULL_RTX)
2585 	  && ! contains_replace_regs (XEXP (dest, 0)))
2586 	{
2587 	  rtx init_insn = XEXP (reg_equiv[regno].init_insns, 0);
2588 	  if (validate_equiv_mem (init_insn, src, dest)
2589 	      && ! memref_used_between_p (dest, init_insn, insn)
2590 	      /* Attaching a REG_EQUIV note will fail if INIT_INSN has
2591 		 multiple sets.  */
2592 	      && set_unique_reg_note (init_insn, REG_EQUIV, copy_rtx (dest)))
2593 	    {
2594 	      /* This insn makes the equivalence, not the one initializing
2595 		 the register.  */
2596 	      reg_equiv_init[regno]
2597 		= gen_rtx_INSN_LIST (VOIDmode, insn, NULL_RTX);
2598 	      df_notes_rescan (init_insn);
2599 	    }
2600 	}
2601     }
2602 
2603   cleared_regs = BITMAP_ALLOC (NULL);
2604   /* Now scan all regs killed in an insn to see if any of them are
2605      registers only used that once.  If so, see if we can replace the
2606      reference with the equivalent form.  If we can, delete the
2607      initializing reference and this register will go away.  If we
2608      can't replace the reference, and the initializing reference is
2609      within the same loop (or in an inner loop), then move the register
2610      initialization just before the use, so that they are in the same
2611      basic block.  */
2612   FOR_EACH_BB_REVERSE (bb)
2613     {
2614       loop_depth = bb->loop_depth;
2615       for (insn = BB_END (bb);
2616 	   insn != PREV_INSN (BB_HEAD (bb));
2617 	   insn = PREV_INSN (insn))
2618 	{
2619 	  rtx link;
2620 
2621 	  if (! INSN_P (insn))
2622 	    continue;
2623 
2624 	  /* Don't substitute into a non-local goto, this confuses CFG.  */
2625 	  if (JUMP_P (insn)
2626 	      && find_reg_note (insn, REG_NON_LOCAL_GOTO, NULL_RTX))
2627 	    continue;
2628 
2629 	  for (link = REG_NOTES (insn); link; link = XEXP (link, 1))
2630 	    {
2631 	      if (REG_NOTE_KIND (link) == REG_DEAD
2632 		  /* Make sure this insn still refers to the register.  */
2633 		  && reg_mentioned_p (XEXP (link, 0), PATTERN (insn)))
2634 		{
2635 		  int regno = REGNO (XEXP (link, 0));
2636 		  rtx equiv_insn;
2637 
2638 		  if (! reg_equiv[regno].replace
2639 		      || reg_equiv[regno].loop_depth < loop_depth)
2640 		    continue;
2641 
2642 		  /* reg_equiv[REGNO].replace gets set only when
2643 		     REG_N_REFS[REGNO] is 2, i.e. the register is set
2644 		     once and used once.  (If it were only set, but not used,
2645 		     flow would have deleted the setting insns.)  Hence
2646 		     there can only be one insn in reg_equiv[REGNO].init_insns.  */
2647 		  gcc_assert (reg_equiv[regno].init_insns
2648 			      && !XEXP (reg_equiv[regno].init_insns, 1));
2649 		  equiv_insn = XEXP (reg_equiv[regno].init_insns, 0);
2650 
2651 		  /* We may not move instructions that can throw, since
2652 		     that changes basic block boundaries and we are not
2653 		     prepared to adjust the CFG to match.  */
2654 		  if (can_throw_internal (equiv_insn))
2655 		    continue;
2656 
2657 		  if (asm_noperands (PATTERN (equiv_insn)) < 0
2658 		      && validate_replace_rtx (regno_reg_rtx[regno],
2659 					       *(reg_equiv[regno].src_p), insn))
2660 		    {
2661 		      rtx equiv_link;
2662 		      rtx last_link;
2663 		      rtx note;
2664 
2665 		      /* Find the last note.  */
2666 		      for (last_link = link; XEXP (last_link, 1);
2667 			   last_link = XEXP (last_link, 1))
2668 			;
2669 
2670 		      /* Append the REG_DEAD notes from equiv_insn.  */
2671 		      equiv_link = REG_NOTES (equiv_insn);
2672 		      while (equiv_link)
2673 			{
2674 			  note = equiv_link;
2675 			  equiv_link = XEXP (equiv_link, 1);
2676 			  if (REG_NOTE_KIND (note) == REG_DEAD)
2677 			    {
2678 			      remove_note (equiv_insn, note);
2679 			      XEXP (last_link, 1) = note;
2680 			      XEXP (note, 1) = NULL_RTX;
2681 			      last_link = note;
2682 			    }
2683 			}
2684 
2685 		      remove_death (regno, insn);
2686 		      SET_REG_N_REFS (regno, 0);
2687 		      REG_FREQ (regno) = 0;
2688 		      delete_insn (equiv_insn);
2689 
2690 		      reg_equiv[regno].init_insns
2691 			= XEXP (reg_equiv[regno].init_insns, 1);
2692 
2693 		      reg_equiv_init[regno] = NULL_RTX;
2694 		      bitmap_set_bit (cleared_regs, regno);
2695 		    }
2696 		  /* Move the initialization of the register to just before
2697 		     INSN.  Update the flow information.  */
2698 		  else if (prev_nondebug_insn (insn) != equiv_insn)
2699 		    {
2700 		      rtx new_insn;
2701 
2702 		      new_insn = emit_insn_before (PATTERN (equiv_insn), insn);
2703 		      REG_NOTES (new_insn) = REG_NOTES (equiv_insn);
2704 		      REG_NOTES (equiv_insn) = 0;
2705 		      /* Rescan it to process the notes.  */
2706 		      df_insn_rescan (new_insn);
2707 
2708 		      /* Make sure this insn is recognized before
2709 			 reload begins, otherwise
2710 			 eliminate_regs_in_insn will die.  */
2711 		      INSN_CODE (new_insn) = INSN_CODE (equiv_insn);
2712 
2713 		      delete_insn (equiv_insn);
2714 
2715 		      XEXP (reg_equiv[regno].init_insns, 0) = new_insn;
2716 
2717 		      REG_BASIC_BLOCK (regno) = bb->index;
2718 		      REG_N_CALLS_CROSSED (regno) = 0;
2719 		      REG_FREQ_CALLS_CROSSED (regno) = 0;
2720 		      REG_N_THROWING_CALLS_CROSSED (regno) = 0;
2721 		      REG_LIVE_LENGTH (regno) = 2;
2722 
2723 		      if (insn == BB_HEAD (bb))
2724 			BB_HEAD (bb) = PREV_INSN (insn);
2725 
2726 		      reg_equiv_init[regno]
2727 			= gen_rtx_INSN_LIST (VOIDmode, new_insn, NULL_RTX);
2728 		      bitmap_set_bit (cleared_regs, regno);
2729 		    }
2730 		}
2731 	    }
2732 	}
2733     }
2734 
2735   if (!bitmap_empty_p (cleared_regs))
2736     {
2737       FOR_EACH_BB (bb)
2738 	{
2739 	  bitmap_and_compl_into (DF_LIVE_IN (bb), cleared_regs);
2740 	  bitmap_and_compl_into (DF_LIVE_OUT (bb), cleared_regs);
2741 	  bitmap_and_compl_into (DF_LR_IN (bb), cleared_regs);
2742 	  bitmap_and_compl_into (DF_LR_OUT (bb), cleared_regs);
2743 	}
2744 
2745       /* Last pass - adjust debug insns referencing cleared regs.  */
2746       if (MAY_HAVE_DEBUG_INSNS)
2747 	for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
2748 	  if (DEBUG_INSN_P (insn))
2749 	    {
2750 	      rtx old_loc = INSN_VAR_LOCATION_LOC (insn);
2751 	      INSN_VAR_LOCATION_LOC (insn)
2752 		= simplify_replace_fn_rtx (old_loc, NULL_RTX,
2753 					   adjust_cleared_regs,
2754 					   (void *) cleared_regs);
2755 	      if (old_loc != INSN_VAR_LOCATION_LOC (insn))
2756 		df_insn_rescan (insn);
2757 	    }
2758     }
2759 
2760   BITMAP_FREE (cleared_regs);
2761 
2762   out:
2763   /* Clean up.  */
2764 
2765   end_alias_analysis ();
2766   free (reg_equiv);
2767   return recorded_label_ref;
2768 }
2769 
2770 
2771 
2772 /* Print chain C to FILE.  */
2773 static void
2774 print_insn_chain (FILE *file, struct insn_chain *c)
2775 {
2776   fprintf (file, "insn=%d, ", INSN_UID(c->insn));
2777   bitmap_print (file, &c->live_throughout, "live_throughout: ", ", ");
2778   bitmap_print (file, &c->dead_or_set, "dead_or_set: ", "\n");
2779 }
2780 
2781 
2782 /* Print all reload_insn_chains to FILE.  */
2783 static void
2784 print_insn_chains (FILE *file)
2785 {
2786   struct insn_chain *c;
2787   for (c = reload_insn_chain; c ; c = c->next)
2788     print_insn_chain (file, c);
2789 }
2790 
2791 /* Return true if pseudo REGNO should be added to set live_throughout
2792    or dead_or_set of the insn chains for reload consideration.  */
2793 static bool
2794 pseudo_for_reload_consideration_p (int regno)
2795 {
2796   /* Consider spilled pseudos too for IRA because they still have a
2797      chance to get hard-registers in the reload when IRA is used.  */
2798   return (reg_renumber[regno] >= 0 || ira_conflicts_p);
2799 }
2800 
2801 /* Init LIVE_SUBREGS[ALLOCNUM] and LIVE_SUBREGS_USED[ALLOCNUM] using
2802    REG to the number of nregs, and INIT_VALUE to get the
2803    initialization.  ALLOCNUM need not be the regno of REG.  */
2804 static void
2805 init_live_subregs (bool init_value, sbitmap *live_subregs,
2806 		   int *live_subregs_used, int allocnum, rtx reg)
2807 {
2808   unsigned int regno = REGNO (SUBREG_REG (reg));
2809   int size = GET_MODE_SIZE (GET_MODE (regno_reg_rtx[regno]));
2810 
2811   gcc_assert (size > 0);
2812 
2813   /* Been there, done that.  */
2814   if (live_subregs_used[allocnum])
2815     return;
2816 
2817   /* Create a new one with zeros.  */
2818   if (live_subregs[allocnum] == NULL)
2819     live_subregs[allocnum] = sbitmap_alloc (size);
2820 
2821   /* If the entire reg was live before blasting into subregs, we need
2822      to init all of the subregs to ones else init to 0.  */
2823   if (init_value)
2824     sbitmap_ones (live_subregs[allocnum]);
2825   else
2826     sbitmap_zero (live_subregs[allocnum]);
2827 
2828   /* Set the number of bits that we really want.  */
2829   live_subregs_used[allocnum] = size;
2830 }
2831 
2832 /* Walk the insns of the current function and build reload_insn_chain,
2833    and record register life information.  */
2834 static void
2835 build_insn_chain (void)
2836 {
2837   unsigned int i;
2838   struct insn_chain **p = &reload_insn_chain;
2839   basic_block bb;
2840   struct insn_chain *c = NULL;
2841   struct insn_chain *next = NULL;
2842   bitmap live_relevant_regs = BITMAP_ALLOC (NULL);
2843   bitmap elim_regset = BITMAP_ALLOC (NULL);
2844   /* live_subregs is a vector used to keep accurate information about
2845      which hardregs are live in multiword pseudos.  live_subregs and
2846      live_subregs_used are indexed by pseudo number.  The live_subreg
2847      entry for a particular pseudo is only used if the corresponding
2848      element is non zero in live_subregs_used.  The value in
2849      live_subregs_used is number of bytes that the pseudo can
2850      occupy.  */
2851   sbitmap *live_subregs = XCNEWVEC (sbitmap, max_regno);
2852   int *live_subregs_used = XNEWVEC (int, max_regno);
2853 
2854   for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2855     if (TEST_HARD_REG_BIT (eliminable_regset, i))
2856       bitmap_set_bit (elim_regset, i);
2857   FOR_EACH_BB_REVERSE (bb)
2858     {
2859       bitmap_iterator bi;
2860       rtx insn;
2861 
2862       CLEAR_REG_SET (live_relevant_regs);
2863       memset (live_subregs_used, 0, max_regno * sizeof (int));
2864 
2865       EXECUTE_IF_SET_IN_BITMAP (DF_LR_OUT (bb), 0, i, bi)
2866 	{
2867 	  if (i >= FIRST_PSEUDO_REGISTER)
2868 	    break;
2869 	  bitmap_set_bit (live_relevant_regs, i);
2870 	}
2871 
2872       EXECUTE_IF_SET_IN_BITMAP (DF_LR_OUT (bb),
2873 				FIRST_PSEUDO_REGISTER, i, bi)
2874 	{
2875 	  if (pseudo_for_reload_consideration_p (i))
2876 	    bitmap_set_bit (live_relevant_regs, i);
2877 	}
2878 
2879       FOR_BB_INSNS_REVERSE (bb, insn)
2880 	{
2881 	  if (!NOTE_P (insn) && !BARRIER_P (insn))
2882 	    {
2883 	      unsigned int uid = INSN_UID (insn);
2884 	      df_ref *def_rec;
2885 	      df_ref *use_rec;
2886 
2887 	      c = new_insn_chain ();
2888 	      c->next = next;
2889 	      next = c;
2890 	      *p = c;
2891 	      p = &c->prev;
2892 
2893 	      c->insn = insn;
2894 	      c->block = bb->index;
2895 
2896 	      if (INSN_P (insn))
2897 		for (def_rec = DF_INSN_UID_DEFS (uid); *def_rec; def_rec++)
2898 		  {
2899 		    df_ref def = *def_rec;
2900 		    unsigned int regno = DF_REF_REGNO (def);
2901 
2902 		    /* Ignore may clobbers because these are generated
2903 		       from calls. However, every other kind of def is
2904 		       added to dead_or_set.  */
2905 		    if (!DF_REF_FLAGS_IS_SET (def, DF_REF_MAY_CLOBBER))
2906 		      {
2907 			if (regno < FIRST_PSEUDO_REGISTER)
2908 			  {
2909 			    if (!fixed_regs[regno])
2910 			      bitmap_set_bit (&c->dead_or_set, regno);
2911 			  }
2912 			else if (pseudo_for_reload_consideration_p (regno))
2913 			  bitmap_set_bit (&c->dead_or_set, regno);
2914 		      }
2915 
2916 		    if ((regno < FIRST_PSEUDO_REGISTER
2917 			 || reg_renumber[regno] >= 0
2918 			 || ira_conflicts_p)
2919 			&& (!DF_REF_FLAGS_IS_SET (def, DF_REF_CONDITIONAL)))
2920 		      {
2921 			rtx reg = DF_REF_REG (def);
2922 
2923 			/* We can model subregs, but not if they are
2924 			   wrapped in ZERO_EXTRACTS.  */
2925 			if (GET_CODE (reg) == SUBREG
2926 			    && !DF_REF_FLAGS_IS_SET (def, DF_REF_ZERO_EXTRACT))
2927 			  {
2928 			    unsigned int start = SUBREG_BYTE (reg);
2929 			    unsigned int last = start
2930 			      + GET_MODE_SIZE (GET_MODE (reg));
2931 
2932 			    init_live_subregs
2933 			      (bitmap_bit_p (live_relevant_regs, regno),
2934 			       live_subregs, live_subregs_used, regno, reg);
2935 
2936 			    if (!DF_REF_FLAGS_IS_SET
2937 				(def, DF_REF_STRICT_LOW_PART))
2938 			      {
2939 				/* Expand the range to cover entire words.
2940 				   Bytes added here are "don't care".  */
2941 				start
2942 				  = start / UNITS_PER_WORD * UNITS_PER_WORD;
2943 				last = ((last + UNITS_PER_WORD - 1)
2944 					/ UNITS_PER_WORD * UNITS_PER_WORD);
2945 			      }
2946 
2947 			    /* Ignore the paradoxical bits.  */
2948 			    if ((int)last > live_subregs_used[regno])
2949 			      last = live_subregs_used[regno];
2950 
2951 			    while (start < last)
2952 			      {
2953 				RESET_BIT (live_subregs[regno], start);
2954 				start++;
2955 			      }
2956 
2957 			    if (sbitmap_empty_p (live_subregs[regno]))
2958 			      {
2959 				live_subregs_used[regno] = 0;
2960 				bitmap_clear_bit (live_relevant_regs, regno);
2961 			      }
2962 			    else
2963 			      /* Set live_relevant_regs here because
2964 				 that bit has to be true to get us to
2965 				 look at the live_subregs fields.  */
2966 			      bitmap_set_bit (live_relevant_regs, regno);
2967 			  }
2968 			else
2969 			  {
2970 			    /* DF_REF_PARTIAL is generated for
2971 			       subregs, STRICT_LOW_PART, and
2972 			       ZERO_EXTRACT.  We handle the subreg
2973 			       case above so here we have to keep from
2974 			       modeling the def as a killing def.  */
2975 			    if (!DF_REF_FLAGS_IS_SET (def, DF_REF_PARTIAL))
2976 			      {
2977 				bitmap_clear_bit (live_relevant_regs, regno);
2978 				live_subregs_used[regno] = 0;
2979 			      }
2980 			  }
2981 		      }
2982 		  }
2983 
2984 	      bitmap_and_compl_into (live_relevant_regs, elim_regset);
2985 	      bitmap_copy (&c->live_throughout, live_relevant_regs);
2986 
2987 	      if (INSN_P (insn))
2988 		for (use_rec = DF_INSN_UID_USES (uid); *use_rec; use_rec++)
2989 		  {
2990 		    df_ref use = *use_rec;
2991 		    unsigned int regno = DF_REF_REGNO (use);
2992 		    rtx reg = DF_REF_REG (use);
2993 
2994 		    /* DF_REF_READ_WRITE on a use means that this use
2995 		       is fabricated from a def that is a partial set
2996 		       to a multiword reg.  Here, we only model the
2997 		       subreg case that is not wrapped in ZERO_EXTRACT
2998 		       precisely so we do not need to look at the
2999 		       fabricated use. */
3000 		    if (DF_REF_FLAGS_IS_SET (use, DF_REF_READ_WRITE)
3001 			&& !DF_REF_FLAGS_IS_SET (use, DF_REF_ZERO_EXTRACT)
3002 			&& DF_REF_FLAGS_IS_SET (use, DF_REF_SUBREG))
3003 		      continue;
3004 
3005 		    /* Add the last use of each var to dead_or_set.  */
3006 		    if (!bitmap_bit_p (live_relevant_regs, regno))
3007 		      {
3008 			if (regno < FIRST_PSEUDO_REGISTER)
3009 			  {
3010 			    if (!fixed_regs[regno])
3011 			      bitmap_set_bit (&c->dead_or_set, regno);
3012 			  }
3013 			else if (pseudo_for_reload_consideration_p (regno))
3014 			  bitmap_set_bit (&c->dead_or_set, regno);
3015 		      }
3016 
3017 		    if (regno < FIRST_PSEUDO_REGISTER
3018 			|| pseudo_for_reload_consideration_p (regno))
3019 		      {
3020 			if (GET_CODE (reg) == SUBREG
3021 			    && !DF_REF_FLAGS_IS_SET (use,
3022 						     DF_REF_SIGN_EXTRACT
3023 						     | DF_REF_ZERO_EXTRACT))
3024 			  {
3025 			    unsigned int start = SUBREG_BYTE (reg);
3026 			    unsigned int last = start
3027 			      + GET_MODE_SIZE (GET_MODE (reg));
3028 
3029 			    init_live_subregs
3030 			      (bitmap_bit_p (live_relevant_regs, regno),
3031 			       live_subregs, live_subregs_used, regno, reg);
3032 
3033 			    /* Ignore the paradoxical bits.  */
3034 			    if ((int)last > live_subregs_used[regno])
3035 			      last = live_subregs_used[regno];
3036 
3037 			    while (start < last)
3038 			      {
3039 				SET_BIT (live_subregs[regno], start);
3040 				start++;
3041 			      }
3042 			  }
3043 			else
3044 			  /* Resetting the live_subregs_used is
3045 			     effectively saying do not use the subregs
3046 			     because we are reading the whole
3047 			     pseudo.  */
3048 			  live_subregs_used[regno] = 0;
3049 			bitmap_set_bit (live_relevant_regs, regno);
3050 		      }
3051 		  }
3052 	    }
3053 	}
3054 
3055       /* FIXME!! The following code is a disaster.  Reload needs to see the
3056 	 labels and jump tables that are just hanging out in between
3057 	 the basic blocks.  See pr33676.  */
3058       insn = BB_HEAD (bb);
3059 
3060       /* Skip over the barriers and cruft.  */
3061       while (insn && (BARRIER_P (insn) || NOTE_P (insn)
3062 		      || BLOCK_FOR_INSN (insn) == bb))
3063 	insn = PREV_INSN (insn);
3064 
3065       /* While we add anything except barriers and notes, the focus is
3066 	 to get the labels and jump tables into the
3067 	 reload_insn_chain.  */
3068       while (insn)
3069 	{
3070 	  if (!NOTE_P (insn) && !BARRIER_P (insn))
3071 	    {
3072 	      if (BLOCK_FOR_INSN (insn))
3073 		break;
3074 
3075 	      c = new_insn_chain ();
3076 	      c->next = next;
3077 	      next = c;
3078 	      *p = c;
3079 	      p = &c->prev;
3080 
3081 	      /* The block makes no sense here, but it is what the old
3082 		 code did.  */
3083 	      c->block = bb->index;
3084 	      c->insn = insn;
3085 	      bitmap_copy (&c->live_throughout, live_relevant_regs);
3086 	    }
3087 	  insn = PREV_INSN (insn);
3088 	}
3089     }
3090 
3091   for (i = 0; i < (unsigned int) max_regno; i++)
3092     if (live_subregs[i])
3093       free (live_subregs[i]);
3094 
3095   reload_insn_chain = c;
3096   *p = NULL;
3097 
3098   free (live_subregs);
3099   free (live_subregs_used);
3100   BITMAP_FREE (live_relevant_regs);
3101   BITMAP_FREE (elim_regset);
3102 
3103   if (dump_file)
3104     print_insn_chains (dump_file);
3105 }
3106 
3107 
3108 
3109 /* All natural loops.  */
3110 struct loops ira_loops;
3111 
3112 /* True if we have allocno conflicts.  It is false for non-optimized
3113    mode or when the conflict table is too big.  */
3114 bool ira_conflicts_p;
3115 
3116 /* This is the main entry of IRA.  */
3117 static void
3118 ira (FILE *f)
3119 {
3120   int overall_cost_before, allocated_reg_info_size;
3121   bool loops_p;
3122   int max_regno_before_ira, ira_max_point_before_emit;
3123   int rebuild_p;
3124   int saved_flag_ira_share_spill_slots;
3125   basic_block bb;
3126 
3127   timevar_push (TV_IRA);
3128 
3129   if (flag_caller_saves)
3130     init_caller_save ();
3131 
3132   if (flag_ira_verbose < 10)
3133     {
3134       internal_flag_ira_verbose = flag_ira_verbose;
3135       ira_dump_file = f;
3136     }
3137   else
3138     {
3139       internal_flag_ira_verbose = flag_ira_verbose - 10;
3140       ira_dump_file = stderr;
3141     }
3142 
3143   ira_conflicts_p = optimize > 0;
3144   setup_prohibited_mode_move_regs ();
3145 
3146   df_note_add_problem ();
3147 
3148   if (optimize == 1)
3149     {
3150       df_live_add_problem ();
3151       df_live_set_all_dirty ();
3152     }
3153 #ifdef ENABLE_CHECKING
3154   df->changeable_flags |= DF_VERIFY_SCHEDULED;
3155 #endif
3156   df_analyze ();
3157   df_clear_flags (DF_NO_INSN_RESCAN);
3158   regstat_init_n_sets_and_refs ();
3159   regstat_compute_ri ();
3160 
3161   /* If we are not optimizing, then this is the only place before
3162      register allocation where dataflow is done.  And that is needed
3163      to generate these warnings.  */
3164   if (warn_clobbered)
3165     generate_setjmp_warnings ();
3166 
3167   /* Determine if the current function is a leaf before running IRA
3168      since this can impact optimizations done by the prologue and
3169      epilogue thus changing register elimination offsets.  */
3170   current_function_is_leaf = leaf_function_p ();
3171 
3172   if (resize_reg_info () && flag_ira_loop_pressure)
3173     ira_set_pseudo_classes (ira_dump_file);
3174 
3175   rebuild_p = update_equiv_regs ();
3176 
3177 #ifndef IRA_NO_OBSTACK
3178   gcc_obstack_init (&ira_obstack);
3179 #endif
3180   bitmap_obstack_initialize (&ira_bitmap_obstack);
3181   if (optimize)
3182     {
3183       max_regno = max_reg_num ();
3184       ira_reg_equiv_len = max_regno;
3185       ira_reg_equiv_invariant_p
3186 	= (bool *) ira_allocate (max_regno * sizeof (bool));
3187       memset (ira_reg_equiv_invariant_p, 0, max_regno * sizeof (bool));
3188       ira_reg_equiv_const = (rtx *) ira_allocate (max_regno * sizeof (rtx));
3189       memset (ira_reg_equiv_const, 0, max_regno * sizeof (rtx));
3190       find_reg_equiv_invariant_const ();
3191       if (rebuild_p)
3192 	{
3193 	  timevar_push (TV_JUMP);
3194 	  rebuild_jump_labels (get_insns ());
3195 	  purge_all_dead_edges ();
3196 	  timevar_pop (TV_JUMP);
3197 	}
3198     }
3199 
3200   max_regno_before_ira = allocated_reg_info_size = max_reg_num ();
3201   ira_setup_eliminable_regset ();
3202 
3203   ira_overall_cost = ira_reg_cost = ira_mem_cost = 0;
3204   ira_load_cost = ira_store_cost = ira_shuffle_cost = 0;
3205   ira_move_loops_num = ira_additional_jumps_num = 0;
3206 
3207   ira_assert (current_loops == NULL);
3208   flow_loops_find (&ira_loops);
3209   record_loop_exits ();
3210   current_loops = &ira_loops;
3211 
3212   if (internal_flag_ira_verbose > 0 && ira_dump_file != NULL)
3213     fprintf (ira_dump_file, "Building IRA IR\n");
3214   loops_p = ira_build (optimize
3215 		       && (flag_ira_region == IRA_REGION_ALL
3216 			   || flag_ira_region == IRA_REGION_MIXED));
3217 
3218   ira_assert (ira_conflicts_p || !loops_p);
3219 
3220   saved_flag_ira_share_spill_slots = flag_ira_share_spill_slots;
3221   if (too_high_register_pressure_p () || cfun->calls_setjmp)
3222     /* It is just wasting compiler's time to pack spilled pseudos into
3223        stack slots in this case -- prohibit it.  We also do this if
3224        there is setjmp call because a variable not modified between
3225        setjmp and longjmp the compiler is required to preserve its
3226        value and sharing slots does not guarantee it.  */
3227     flag_ira_share_spill_slots = FALSE;
3228 
3229   ira_color ();
3230 
3231   ira_max_point_before_emit = ira_max_point;
3232 
3233   ira_emit (loops_p);
3234 
3235   if (ira_conflicts_p)
3236     {
3237       max_regno = max_reg_num ();
3238 
3239       if (! loops_p)
3240 	ira_initiate_assign ();
3241       else
3242 	{
3243 	  expand_reg_info (allocated_reg_info_size);
3244 	  setup_preferred_alternate_classes_for_new_pseudos
3245 	    (allocated_reg_info_size);
3246 	  allocated_reg_info_size = max_regno;
3247 
3248 	  if (internal_flag_ira_verbose > 0 && ira_dump_file != NULL)
3249 	    fprintf (ira_dump_file, "Flattening IR\n");
3250 	  ira_flattening (max_regno_before_ira, ira_max_point_before_emit);
3251 	  /* New insns were generated: add notes and recalculate live
3252 	     info.  */
3253 	  df_analyze ();
3254 
3255 	  flow_loops_find (&ira_loops);
3256 	  record_loop_exits ();
3257 	  current_loops = &ira_loops;
3258 
3259 	  setup_allocno_assignment_flags ();
3260 	  ira_initiate_assign ();
3261 	  ira_reassign_conflict_allocnos (max_regno);
3262 	}
3263     }
3264 
3265   setup_reg_renumber ();
3266 
3267   calculate_allocation_cost ();
3268 
3269 #ifdef ENABLE_IRA_CHECKING
3270   if (ira_conflicts_p)
3271     check_allocation ();
3272 #endif
3273 
3274   delete_trivially_dead_insns (get_insns (), max_reg_num ());
3275   max_regno = max_reg_num ();
3276 
3277   /* And the reg_equiv_memory_loc array.  */
3278   VEC_safe_grow (rtx, gc, reg_equiv_memory_loc_vec, max_regno);
3279   memset (VEC_address (rtx, reg_equiv_memory_loc_vec), 0,
3280 	  sizeof (rtx) * max_regno);
3281   reg_equiv_memory_loc = VEC_address (rtx, reg_equiv_memory_loc_vec);
3282 
3283   if (max_regno != max_regno_before_ira)
3284     {
3285       regstat_free_n_sets_and_refs ();
3286       regstat_free_ri ();
3287       regstat_init_n_sets_and_refs ();
3288       regstat_compute_ri ();
3289     }
3290 
3291   allocate_initial_values (reg_equiv_memory_loc);
3292 
3293   overall_cost_before = ira_overall_cost;
3294   if (ira_conflicts_p)
3295     {
3296       fix_reg_equiv_init ();
3297 
3298 #ifdef ENABLE_IRA_CHECKING
3299       print_redundant_copies ();
3300 #endif
3301 
3302       ira_spilled_reg_stack_slots_num = 0;
3303       ira_spilled_reg_stack_slots
3304 	= ((struct ira_spilled_reg_stack_slot *)
3305 	   ira_allocate (max_regno
3306 			 * sizeof (struct ira_spilled_reg_stack_slot)));
3307       memset (ira_spilled_reg_stack_slots, 0,
3308 	      max_regno * sizeof (struct ira_spilled_reg_stack_slot));
3309     }
3310 
3311   timevar_pop (TV_IRA);
3312 
3313   timevar_push (TV_RELOAD);
3314   df_set_flags (DF_NO_INSN_RESCAN);
3315   build_insn_chain ();
3316 
3317   reload_completed = !reload (get_insns (), ira_conflicts_p);
3318 
3319   finish_subregs_of_mode ();
3320 
3321   timevar_pop (TV_RELOAD);
3322 
3323   timevar_push (TV_IRA);
3324 
3325   if (ira_conflicts_p)
3326     {
3327       ira_free (ira_spilled_reg_stack_slots);
3328 
3329       ira_finish_assign ();
3330 
3331     }
3332   if (internal_flag_ira_verbose > 0 && ira_dump_file != NULL
3333       && overall_cost_before != ira_overall_cost)
3334     fprintf (ira_dump_file, "+++Overall after reload %d\n", ira_overall_cost);
3335   ira_destroy ();
3336 
3337   flag_ira_share_spill_slots = saved_flag_ira_share_spill_slots;
3338 
3339   flow_loops_free (&ira_loops);
3340   free_dominance_info (CDI_DOMINATORS);
3341   FOR_ALL_BB (bb)
3342     bb->loop_father = NULL;
3343   current_loops = NULL;
3344 
3345   regstat_free_ri ();
3346   regstat_free_n_sets_and_refs ();
3347 
3348   if (optimize)
3349     {
3350       cleanup_cfg (CLEANUP_EXPENSIVE);
3351 
3352       ira_free (ira_reg_equiv_invariant_p);
3353       ira_free (ira_reg_equiv_const);
3354     }
3355 
3356   bitmap_obstack_release (&ira_bitmap_obstack);
3357 #ifndef IRA_NO_OBSTACK
3358   obstack_free (&ira_obstack, NULL);
3359 #endif
3360 
3361   /* The code after the reload has changed so much that at this point
3362      we might as well just rescan everything.  Not that
3363      df_rescan_all_insns is not going to help here because it does not
3364      touch the artificial uses and defs.  */
3365   df_finish_pass (true);
3366   if (optimize > 1)
3367     df_live_add_problem ();
3368   df_scan_alloc (NULL);
3369   df_scan_blocks ();
3370 
3371   if (optimize)
3372     df_analyze ();
3373 
3374   timevar_pop (TV_IRA);
3375 }
3376 
3377 
3378 
3379 static bool
3380 gate_ira (void)
3381 {
3382   return true;
3383 }
3384 
3385 /* Run the integrated register allocator.  */
3386 static unsigned int
3387 rest_of_handle_ira (void)
3388 {
3389   ira (dump_file);
3390   return 0;
3391 }
3392 
3393 struct rtl_opt_pass pass_ira =
3394 {
3395  {
3396   RTL_PASS,
3397   "ira",                                /* name */
3398   gate_ira,                             /* gate */
3399   rest_of_handle_ira,		        /* execute */
3400   NULL,                                 /* sub */
3401   NULL,                                 /* next */
3402   0,                                    /* static_pass_number */
3403   TV_NONE,	                        /* tv_id */
3404   0,                                    /* properties_required */
3405   0,                                    /* properties_provided */
3406   0,                                    /* properties_destroyed */
3407   0,                                    /* todo_flags_start */
3408   TODO_dump_func |
3409   TODO_ggc_collect                      /* todo_flags_finish */
3410  }
3411 };
3412