2008-08-24 Jeff Law <law@redhat.com>
[official-gcc.git] / gcc / ira.c
blob40345c07c56d961f7937a5ab0e075ac94191e92d
1 /* Integrated Register Allocator (IRA) entry point.
2 Copyright (C) 2006, 2007, 2008
3 Free Software Foundation, Inc.
4 Contributed by Vladimir Makarov <vmakarov@redhat.com>.
6 This file is part of GCC.
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.
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.
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/>. */
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.
32 Major IRA notions are:
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.
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.
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:
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.
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.
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.
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.
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.
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.
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.
134 IRA major passes are:
136 o Building IRA internal representation which consists of the
137 following subpasses:
139 * First, IRA builds regions and creates allocnos (file
140 ira-build.c) and initializes most of their attributes.
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).
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).
151 * IRA removes low register pressure loops from the regions
152 mostly to speed IRA up (file ira-build.c).
154 * IRA propagates accumulated allocno info from lower region
155 allocnos to corresponding upper region allocnos (file
156 ira-build.c).
158 * IRA creates all caps (file ira-build.c).
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.
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:
171 * Optional aggressive coalescing of allocnos in the region.
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.
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.
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.
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.
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.
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.
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.
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
257 * sharing stack slots for the spilled pseudos based on IRA info
258 about pseudo-register conflicts.
260 * reassigning hard-registers to all spilled pseudos at the end
261 of each reload iteration.
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.
267 IRA uses a lot of data representing the target processors. These
268 data are initilized in file ira.c.
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).
277 Literature is worth to read for better understanding the code:
279 o Preston Briggs, Keith D. Cooper, Linda Torczon. Improvements to
280 Graph Coloring Register Allocation.
282 o David Callahan, Brian Koblenz. Register allocation via
283 hierarchical graph coloring.
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.
289 o Guei-Yuan Lueh, Thomas Gross, and Ali-Reza Adl-Tabatabai. Global
290 Register Allocation Based on Graph Fusion.
292 o Vladimir Makarov. The Integrated Register Allocator for GCC.
294 o Vladimir Makarov. The top-down register allocator for irregular
295 register file architectures.
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 "reload.h"
320 #include "errors.h"
321 #include "integrate.h"
322 #include "df.h"
323 #include "ggc.h"
324 #include "ira-int.h"
327 /* A modified value of flag `-fira-verbose' used internally. */
328 int internal_flag_ira_verbose;
330 /* Dump file of the allocator if it is not NULL. */
331 FILE *ira_dump_file;
333 /* Pools for allocnos, copies, allocno live ranges. */
334 alloc_pool allocno_pool, copy_pool, allocno_live_range_pool;
336 /* The number of elements in the following array. */
337 int ira_spilled_reg_stack_slots_num;
339 /* The following array contains info about spilled pseudo-registers
340 stack slots used in current function so far. */
341 struct ira_spilled_reg_stack_slot *ira_spilled_reg_stack_slots;
343 /* Correspondingly overall cost of the allocation, cost of the
344 allocnos assigned to hard-registers, cost of the allocnos assigned
345 to memory, cost of loads, stores and register move insns generated
346 for pseudo-register live range splitting (see ira-emit.c). */
347 int ira_overall_cost;
348 int ira_reg_cost, ira_mem_cost;
349 int ira_load_cost, ira_store_cost, ira_shuffle_cost;
350 int ira_move_loops_num, ira_additional_jumps_num;
352 /* Map: hard regs X modes -> set of hard registers for storing value
353 of given mode starting with given hard register. */
354 HARD_REG_SET ira_reg_mode_hard_regset[FIRST_PSEUDO_REGISTER][NUM_MACHINE_MODES];
356 /* The following two variables are array analogs of the macros
357 MEMORY_MOVE_COST and REGISTER_MOVE_COST. */
358 short int ira_memory_move_cost[MAX_MACHINE_MODE][N_REG_CLASSES][2];
359 move_table *ira_register_move_cost[MAX_MACHINE_MODE];
361 /* Similar to may_move_in_cost but it is calculated in IRA instead of
362 regclass. Another difference is that we take only available hard
363 registers into account to figure out that one register class is a
364 subset of the another one. */
365 move_table *ira_may_move_in_cost[MAX_MACHINE_MODE];
367 /* Similar to may_move_out_cost but it is calculated in IRA instead of
368 regclass. Another difference is that we take only available hard
369 registers into account to figure out that one register class is a
370 subset of the another one. */
371 move_table *ira_may_move_out_cost[MAX_MACHINE_MODE];
373 /* Register class subset relation: TRUE if the first class is a subset
374 of the second one considering only hard registers available for the
375 allocation. */
376 int ira_class_subset_p[N_REG_CLASSES][N_REG_CLASSES];
378 /* Temporary hard reg set used for a different calculation. */
379 static HARD_REG_SET temp_hard_regset;
383 /* The function sets up the map IRA_REG_MODE_HARD_REGSET. */
384 static void
385 setup_reg_mode_hard_regset (void)
387 int i, m, hard_regno;
389 for (m = 0; m < NUM_MACHINE_MODES; m++)
390 for (hard_regno = 0; hard_regno < FIRST_PSEUDO_REGISTER; hard_regno++)
392 CLEAR_HARD_REG_SET (ira_reg_mode_hard_regset[hard_regno][m]);
393 for (i = hard_regno_nregs[hard_regno][m] - 1; i >= 0; i--)
394 if (hard_regno + i < FIRST_PSEUDO_REGISTER)
395 SET_HARD_REG_BIT (ira_reg_mode_hard_regset[hard_regno][m],
396 hard_regno + i);
402 /* Hard registers that can not be used for the register allocator for
403 all functions of the current compilation unit. */
404 static HARD_REG_SET no_unit_alloc_regs;
406 /* Array of the number of hard registers of given class which are
407 available for allocation. The order is defined by the
408 allocation order. */
409 short ira_class_hard_regs[N_REG_CLASSES][FIRST_PSEUDO_REGISTER];
411 /* The number of elements of the above array for given register
412 class. */
413 int ira_class_hard_regs_num[N_REG_CLASSES];
415 /* Index (in ira_class_hard_regs) for given register class and hard
416 register (in general case a hard register can belong to several
417 register classes). The index is negative for hard registers
418 unavailable for the allocation. */
419 short ira_class_hard_reg_index[N_REG_CLASSES][FIRST_PSEUDO_REGISTER];
421 /* The function sets up the three arrays declared above. */
422 static void
423 setup_class_hard_regs (void)
425 int cl, i, hard_regno, n;
426 HARD_REG_SET processed_hard_reg_set;
428 ira_assert (SHRT_MAX >= FIRST_PSEUDO_REGISTER);
429 /* We could call ORDER_REGS_FOR_LOCAL_ALLOC here (it is usually
430 putting hard callee-used hard registers first). But our
431 heuristics work better. */
432 for (cl = (int) N_REG_CLASSES - 1; cl >= 0; cl--)
434 COPY_HARD_REG_SET (temp_hard_regset, reg_class_contents[cl]);
435 AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
436 CLEAR_HARD_REG_SET (processed_hard_reg_set);
437 for (n = 0, i = 0; i < FIRST_PSEUDO_REGISTER; i++)
439 #ifdef REG_ALLOC_ORDER
440 hard_regno = reg_alloc_order[i];
441 #else
442 hard_regno = i;
443 #endif
444 if (TEST_HARD_REG_BIT (processed_hard_reg_set, hard_regno))
445 continue;
446 SET_HARD_REG_BIT (processed_hard_reg_set, hard_regno);
447 if (! TEST_HARD_REG_BIT (temp_hard_regset, hard_regno))
448 ira_class_hard_reg_index[cl][hard_regno] = -1;
449 else
451 ira_class_hard_reg_index[cl][hard_regno] = n;
452 ira_class_hard_regs[cl][n++] = hard_regno;
455 ira_class_hard_regs_num[cl] = n;
459 /* Number of given class hard registers available for the register
460 allocation for given classes. */
461 int ira_available_class_regs[N_REG_CLASSES];
463 /* Set up IRA_AVAILABLE_CLASS_REGS. */
464 static void
465 setup_available_class_regs (void)
467 int i, j;
469 memset (ira_available_class_regs, 0, sizeof (ira_available_class_regs));
470 for (i = 0; i < N_REG_CLASSES; i++)
472 COPY_HARD_REG_SET (temp_hard_regset, reg_class_contents[i]);
473 AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
474 for (j = 0; j < FIRST_PSEUDO_REGISTER; j++)
475 if (TEST_HARD_REG_BIT (temp_hard_regset, j))
476 ira_available_class_regs[i]++;
480 /* Set up global variables defining info about hard registers for the
481 allocation. These depend on USE_HARD_FRAME_P whose TRUE value means
482 that we can use the hard frame pointer for the allocation. */
483 static void
484 setup_alloc_regs (bool use_hard_frame_p)
486 COPY_HARD_REG_SET (no_unit_alloc_regs, fixed_reg_set);
487 if (! use_hard_frame_p)
488 SET_HARD_REG_BIT (no_unit_alloc_regs, HARD_FRAME_POINTER_REGNUM);
489 setup_class_hard_regs ();
490 setup_available_class_regs ();
495 /* Set up IRA_MEMORY_MOVE_COST, IRA_REGISTER_MOVE_COST. */
496 static void
497 setup_class_subset_and_memory_move_costs (void)
499 int cl, cl2;
500 enum machine_mode mode;
501 HARD_REG_SET temp_hard_regset2;
503 for (mode = 0; mode < MAX_MACHINE_MODE; mode++)
504 ira_memory_move_cost[mode][NO_REGS][0]
505 = ira_memory_move_cost[mode][NO_REGS][1] = SHRT_MAX;
506 for (cl = (int) N_REG_CLASSES - 1; cl >= 0; cl--)
508 if (cl != (int) NO_REGS)
509 for (mode = 0; mode < MAX_MACHINE_MODE; mode++)
511 ira_memory_move_cost[mode][cl][0] = MEMORY_MOVE_COST (mode, cl, 0);
512 ira_memory_move_cost[mode][cl][1] = MEMORY_MOVE_COST (mode, cl, 1);
513 /* Costs for NO_REGS are used in cost calculation on the
514 1st pass when the preferred register classes are not
515 known yet. In this case we take the best scenario. */
516 if (ira_memory_move_cost[mode][NO_REGS][0]
517 > ira_memory_move_cost[mode][cl][0])
518 ira_memory_move_cost[mode][NO_REGS][0]
519 = ira_memory_move_cost[mode][cl][0];
520 if (ira_memory_move_cost[mode][NO_REGS][1]
521 > ira_memory_move_cost[mode][cl][1])
522 ira_memory_move_cost[mode][NO_REGS][1]
523 = ira_memory_move_cost[mode][cl][1];
525 for (cl2 = (int) N_REG_CLASSES - 1; cl2 >= 0; cl2--)
527 COPY_HARD_REG_SET (temp_hard_regset, reg_class_contents[cl]);
528 AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
529 COPY_HARD_REG_SET (temp_hard_regset2, reg_class_contents[cl2]);
530 AND_COMPL_HARD_REG_SET (temp_hard_regset2, no_unit_alloc_regs);
531 ira_class_subset_p[cl][cl2]
532 = hard_reg_set_subset_p (temp_hard_regset, temp_hard_regset2);
539 /* Define the following macro if allocation through malloc if
540 preferable. */
541 #define IRA_NO_OBSTACK
543 #ifndef IRA_NO_OBSTACK
544 /* Obstack used for storing all dynamic data (except bitmaps) of the
545 IRA. */
546 static struct obstack ira_obstack;
547 #endif
549 /* Obstack used for storing all bitmaps of the IRA. */
550 static struct bitmap_obstack ira_bitmap_obstack;
552 /* Allocate memory of size LEN for IRA data. */
553 void *
554 ira_allocate (size_t len)
556 void *res;
558 #ifndef IRA_NO_OBSTACK
559 res = obstack_alloc (&ira_obstack, len);
560 #else
561 res = xmalloc (len);
562 #endif
563 return res;
566 /* Reallocate memory PTR of size LEN for IRA data. */
567 void *
568 ira_reallocate (void *ptr, size_t len)
570 void *res;
572 #ifndef IRA_NO_OBSTACK
573 res = obstack_alloc (&ira_obstack, len);
574 #else
575 res = xrealloc (ptr, len);
576 #endif
577 return res;
580 /* Free memory ADDR allocated for IRA data. */
581 void
582 ira_free (void *addr ATTRIBUTE_UNUSED)
584 #ifndef IRA_NO_OBSTACK
585 /* do nothing */
586 #else
587 free (addr);
588 #endif
592 /* Allocate and returns bitmap for IRA. */
593 bitmap
594 ira_allocate_bitmap (void)
596 return BITMAP_ALLOC (&ira_bitmap_obstack);
599 /* Free bitmap B allocated for IRA. */
600 void
601 ira_free_bitmap (bitmap b ATTRIBUTE_UNUSED)
603 /* do nothing */
608 /* Output information about allocation of all allocnos (except for
609 caps) into file F. */
610 void
611 ira_print_disposition (FILE *f)
613 int i, n, max_regno;
614 ira_allocno_t a;
615 basic_block bb;
617 fprintf (f, "Disposition:");
618 max_regno = max_reg_num ();
619 for (n = 0, i = FIRST_PSEUDO_REGISTER; i < max_regno; i++)
620 for (a = ira_regno_allocno_map[i];
621 a != NULL;
622 a = ALLOCNO_NEXT_REGNO_ALLOCNO (a))
624 if (n % 4 == 0)
625 fprintf (f, "\n");
626 n++;
627 fprintf (f, " %4d:r%-4d", ALLOCNO_NUM (a), ALLOCNO_REGNO (a));
628 if ((bb = ALLOCNO_LOOP_TREE_NODE (a)->bb) != NULL)
629 fprintf (f, "b%-3d", bb->index);
630 else
631 fprintf (f, "l%-3d", ALLOCNO_LOOP_TREE_NODE (a)->loop->num);
632 if (ALLOCNO_HARD_REGNO (a) >= 0)
633 fprintf (f, " %3d", ALLOCNO_HARD_REGNO (a));
634 else
635 fprintf (f, " mem");
637 fprintf (f, "\n");
640 /* Outputs information about allocation of all allocnos into
641 stderr. */
642 void
643 ira_debug_disposition (void)
645 ira_print_disposition (stderr);
650 /* For each reg class, table listing all the classes contained in it
651 (excluding the class itself. Non-allocatable registers are
652 excluded from the consideration). */
653 static enum reg_class alloc_reg_class_subclasses[N_REG_CLASSES][N_REG_CLASSES];
655 /* Initialize the table of subclasses of each reg class. */
656 static void
657 setup_reg_subclasses (void)
659 int i, j;
660 HARD_REG_SET temp_hard_regset2;
662 for (i = 0; i < N_REG_CLASSES; i++)
663 for (j = 0; j < N_REG_CLASSES; j++)
664 alloc_reg_class_subclasses[i][j] = LIM_REG_CLASSES;
666 for (i = 0; i < N_REG_CLASSES; i++)
668 if (i == (int) NO_REGS)
669 continue;
671 COPY_HARD_REG_SET (temp_hard_regset, reg_class_contents[i]);
672 AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
673 if (hard_reg_set_equal_p (temp_hard_regset, ira_zero_hard_reg_set))
674 continue;
675 for (j = 0; j < N_REG_CLASSES; j++)
676 if (i != j)
678 enum reg_class *p;
680 COPY_HARD_REG_SET (temp_hard_regset2, reg_class_contents[j]);
681 AND_COMPL_HARD_REG_SET (temp_hard_regset2, no_unit_alloc_regs);
682 if (! hard_reg_set_subset_p (temp_hard_regset,
683 temp_hard_regset2))
684 continue;
685 p = &alloc_reg_class_subclasses[j][0];
686 while (*p != LIM_REG_CLASSES) p++;
687 *p = (enum reg_class) i;
694 /* Number of cover classes. Cover classes is non-intersected register
695 classes containing all hard-registers available for the
696 allocation. */
697 int ira_reg_class_cover_size;
699 /* The array containing cover classes (see also comments for macro
700 IRA_COVER_CLASSES). Only first IRA_REG_CLASS_COVER_SIZE elements are
701 used for this. */
702 enum reg_class ira_reg_class_cover[N_REG_CLASSES];
704 /* The number of elements in the subsequent array. */
705 int ira_important_classes_num;
707 /* The array containing non-empty classes (including non-empty cover
708 classes) which are subclasses of cover classes. Such classes is
709 important for calculation of the hard register usage costs. */
710 enum reg_class ira_important_classes[N_REG_CLASSES];
712 /* The array containing indexes of important classes in the previous
713 array. The array elements are defined only for important
714 classes. */
715 int ira_important_class_nums[N_REG_CLASSES];
717 #ifdef IRA_COVER_CLASSES
719 /* Check IRA_COVER_CLASSES and sets the four global variables defined
720 above. */
721 static void
722 setup_cover_and_important_classes (void)
724 int i, j;
725 enum reg_class cl;
726 static enum reg_class classes[] = IRA_COVER_CLASSES;
727 HARD_REG_SET temp_hard_regset2;
729 ira_reg_class_cover_size = 0;
730 for (i = 0; (cl = classes[i]) != LIM_REG_CLASSES; i++)
732 for (j = 0; j < i; j++)
733 if (reg_classes_intersect_p (cl, classes[j]))
734 gcc_unreachable ();
735 COPY_HARD_REG_SET (temp_hard_regset, reg_class_contents[cl]);
736 AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
737 if (! hard_reg_set_equal_p (temp_hard_regset, ira_zero_hard_reg_set))
738 ira_reg_class_cover[ira_reg_class_cover_size++] = cl;
740 ira_important_classes_num = 0;
741 for (cl = 0; cl < N_REG_CLASSES; cl++)
743 COPY_HARD_REG_SET (temp_hard_regset, reg_class_contents[cl]);
744 AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
745 if (! hard_reg_set_equal_p (temp_hard_regset, ira_zero_hard_reg_set))
746 for (j = 0; j < ira_reg_class_cover_size; j++)
748 COPY_HARD_REG_SET (temp_hard_regset, reg_class_contents[cl]);
749 AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
750 COPY_HARD_REG_SET (temp_hard_regset2,
751 reg_class_contents[ira_reg_class_cover[j]]);
752 AND_COMPL_HARD_REG_SET (temp_hard_regset2, no_unit_alloc_regs);
753 if (cl == ira_reg_class_cover[j]
754 || (hard_reg_set_subset_p (temp_hard_regset, temp_hard_regset2)
755 && ! hard_reg_set_equal_p (temp_hard_regset,
756 temp_hard_regset2)))
758 ira_important_class_nums[cl] = ira_important_classes_num;
759 ira_important_classes[ira_important_classes_num++] = cl;
764 #endif
766 /* Map of all register classes to corresponding cover class containing
767 the given class. If given class is not a subset of a cover class,
768 we translate it into the cheapest cover class. */
769 enum reg_class ira_class_translate[N_REG_CLASSES];
771 #ifdef IRA_COVER_CLASSES
773 /* Set up array IRA_CLASS_TRANSLATE. */
774 static void
775 setup_class_translate (void)
777 enum reg_class cl, cover_class, best_class, *cl_ptr;
778 enum machine_mode mode;
779 int i, cost, min_cost, best_cost;
781 for (cl = 0; cl < N_REG_CLASSES; cl++)
782 ira_class_translate[cl] = NO_REGS;
783 for (i = 0; i < ira_reg_class_cover_size; i++)
785 cover_class = ira_reg_class_cover[i];
786 for (cl_ptr = &alloc_reg_class_subclasses[cover_class][0];
787 (cl = *cl_ptr) != LIM_REG_CLASSES;
788 cl_ptr++)
790 if (ira_class_translate[cl] == NO_REGS)
791 ira_class_translate[cl] = cover_class;
792 #ifdef ENABLE_IRA_CHECKING
793 else
795 COPY_HARD_REG_SET (temp_hard_regset, reg_class_contents[cl]);
796 AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
797 if (! hard_reg_set_subset_p (temp_hard_regset,
798 ira_zero_hard_reg_set))
799 gcc_unreachable ();
801 #endif
803 ira_class_translate[cover_class] = cover_class;
805 /* For classes which are not fully covered by a cover class (in
806 other words covered by more one cover class), use the cheapest
807 cover class. */
808 for (cl = 0; cl < N_REG_CLASSES; cl++)
810 if (cl == NO_REGS || ira_class_translate[cl] != NO_REGS)
811 continue;
812 best_class = NO_REGS;
813 best_cost = INT_MAX;
814 for (i = 0; i < ira_reg_class_cover_size; i++)
816 cover_class = ira_reg_class_cover[i];
817 COPY_HARD_REG_SET (temp_hard_regset,
818 reg_class_contents[cover_class]);
819 AND_HARD_REG_SET (temp_hard_regset, reg_class_contents[cl]);
820 AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
821 if (! hard_reg_set_equal_p (temp_hard_regset, ira_zero_hard_reg_set))
823 min_cost = INT_MAX;
824 for (mode = 0; mode < MAX_MACHINE_MODE; mode++)
826 cost = (ira_memory_move_cost[mode][cl][0]
827 + ira_memory_move_cost[mode][cl][1]);
828 if (min_cost > cost)
829 min_cost = cost;
831 if (best_class == NO_REGS || best_cost > min_cost)
833 best_class = cover_class;
834 best_cost = min_cost;
838 ira_class_translate[cl] = best_class;
841 #endif
843 /* The biggest important reg_class inside of intersection of the two
844 reg_classes (that is calculated taking only hard registers
845 available for allocation into account). If the both reg_classes
846 contain no hard registers available for allocation, the value is
847 calculated by taking all hard-registers including fixed ones into
848 account. */
849 enum reg_class ira_reg_class_intersect[N_REG_CLASSES][N_REG_CLASSES];
851 /* The biggest important reg_class inside of union of the two
852 reg_classes (that is calculated taking only hard registers
853 available for allocation into account). If the both reg_classes
854 contain no hard registers available for allocation, the value is
855 calculated by taking all hard-registers including fixed ones into
856 account. In other words, the value is the corresponding
857 reg_class_subunion value. */
858 enum reg_class ira_reg_class_union[N_REG_CLASSES][N_REG_CLASSES];
860 #ifdef IRA_COVER_CLASSES
862 /* Set up IRA_REG_CLASS_INTERSECT and IRA_REG_CLASS_UNION. */
863 static void
864 setup_reg_class_intersect_union (void)
866 int i, cl1, cl2, cl3;
867 HARD_REG_SET intersection_set, union_set, temp_set2;
869 for (cl1 = 0; cl1 < N_REG_CLASSES; cl1++)
871 for (cl2 = 0; cl2 < N_REG_CLASSES; cl2++)
873 ira_reg_class_intersect[cl1][cl2] = NO_REGS;
874 COPY_HARD_REG_SET (temp_hard_regset, reg_class_contents[cl1]);
875 AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
876 COPY_HARD_REG_SET (temp_set2, reg_class_contents[cl2]);
877 AND_COMPL_HARD_REG_SET (temp_set2, no_unit_alloc_regs);
878 if (hard_reg_set_equal_p (temp_hard_regset, ira_zero_hard_reg_set)
879 && hard_reg_set_equal_p (temp_set2, ira_zero_hard_reg_set))
881 for (i = 0;; i++)
883 cl3 = reg_class_subclasses[cl1][i];
884 if (cl3 == LIM_REG_CLASSES)
885 break;
886 if (reg_class_subset_p (ira_reg_class_intersect[cl1][cl2],
887 cl3))
888 ira_reg_class_intersect[cl1][cl2] = cl3;
890 ira_reg_class_union[cl1][cl2] = reg_class_subunion[cl1][cl2];
891 continue;
893 ira_reg_class_union[cl1][cl2] = NO_REGS;
894 COPY_HARD_REG_SET (intersection_set, reg_class_contents[cl1]);
895 AND_HARD_REG_SET (intersection_set, reg_class_contents[cl2]);
896 AND_COMPL_HARD_REG_SET (intersection_set, no_unit_alloc_regs);
897 COPY_HARD_REG_SET (union_set, reg_class_contents[cl1]);
898 IOR_HARD_REG_SET (union_set, reg_class_contents[cl2]);
899 AND_COMPL_HARD_REG_SET (union_set, no_unit_alloc_regs);
900 for (i = 0; i < ira_important_classes_num; i++)
902 cl3 = ira_important_classes[i];
903 COPY_HARD_REG_SET (temp_hard_regset, reg_class_contents[cl3]);
904 AND_COMPL_HARD_REG_SET (temp_hard_regset, no_unit_alloc_regs);
905 if (hard_reg_set_subset_p (temp_hard_regset, intersection_set))
907 COPY_HARD_REG_SET
908 (temp_set2,
909 reg_class_contents[(int)
910 ira_reg_class_intersect[cl1][cl2]]);
911 AND_COMPL_HARD_REG_SET (temp_set2, no_unit_alloc_regs);
912 if (! hard_reg_set_subset_p (temp_hard_regset, temp_set2))
913 ira_reg_class_intersect[cl1][cl2] = (enum reg_class) cl3;
915 if (hard_reg_set_subset_p (temp_hard_regset, union_set))
917 COPY_HARD_REG_SET
918 (temp_set2,
919 reg_class_contents[(int) ira_reg_class_union[cl1][cl2]]);
920 AND_COMPL_HARD_REG_SET (temp_set2, no_unit_alloc_regs);
921 if (hard_reg_set_subset_p (temp_set2, temp_hard_regset))
922 ira_reg_class_union[cl1][cl2] = (enum reg_class) cl3;
929 #endif
931 /* Output all cover classes and the translation map into file F. */
932 static void
933 print_class_cover (FILE *f)
935 static const char *const reg_class_names[] = REG_CLASS_NAMES;
936 int i;
938 fprintf (f, "Class cover:\n");
939 for (i = 0; i < ira_reg_class_cover_size; i++)
940 fprintf (f, " %s", reg_class_names[ira_reg_class_cover[i]]);
941 fprintf (f, "\nClass translation:\n");
942 for (i = 0; i < N_REG_CLASSES; i++)
943 fprintf (f, " %s -> %s\n", reg_class_names[i],
944 reg_class_names[ira_class_translate[i]]);
947 /* Output all cover classes and the translation map into
948 stderr. */
949 void
950 ira_debug_class_cover (void)
952 print_class_cover (stderr);
955 /* Set up different arrays concerning class subsets, cover and
956 important classes. */
957 static void
958 find_reg_class_closure (void)
960 setup_reg_subclasses ();
961 #ifdef IRA_COVER_CLASSES
962 setup_cover_and_important_classes ();
963 setup_class_translate ();
964 setup_reg_class_intersect_union ();
965 #endif
970 /* Map: register class x machine mode -> number of hard registers of
971 given class needed to store value of given mode. If the number is
972 different, the size will be negative. */
973 int ira_reg_class_nregs[N_REG_CLASSES][MAX_MACHINE_MODE];
975 /* Maximal value of the previous array elements. */
976 int ira_max_nregs;
978 /* Form IRA_REG_CLASS_NREGS map. */
979 static void
980 setup_reg_class_nregs (void)
982 int m;
983 enum reg_class cl;
985 ira_max_nregs = -1;
986 for (cl = 0; cl < N_REG_CLASSES; cl++)
987 for (m = 0; m < MAX_MACHINE_MODE; m++)
989 ira_reg_class_nregs[cl][m] = CLASS_MAX_NREGS (cl, m);
990 if (ira_max_nregs < ira_reg_class_nregs[cl][m])
991 ira_max_nregs = ira_reg_class_nregs[cl][m];
997 /* Array whose values are hard regset of hard registers available for
998 the allocation of given register class whose HARD_REGNO_MODE_OK
999 values for given mode are zero. */
1000 HARD_REG_SET prohibited_class_mode_regs[N_REG_CLASSES][NUM_MACHINE_MODES];
1002 /* Set up PROHIBITED_CLASS_MODE_REGS. */
1003 static void
1004 setup_prohibited_class_mode_regs (void)
1006 int i, j, k, hard_regno;
1007 enum reg_class cl;
1009 for (i = 0; i < ira_reg_class_cover_size; i++)
1011 cl = ira_reg_class_cover[i];
1012 for (j = 0; j < NUM_MACHINE_MODES; j++)
1014 CLEAR_HARD_REG_SET (prohibited_class_mode_regs[cl][j]);
1015 for (k = ira_class_hard_regs_num[cl] - 1; k >= 0; k--)
1017 hard_regno = ira_class_hard_regs[cl][k];
1018 if (! HARD_REGNO_MODE_OK (hard_regno, j))
1019 SET_HARD_REG_BIT (prohibited_class_mode_regs[cl][j],
1020 hard_regno);
1028 /* Allocate and initialize IRA_REGISTER_MOVE_COST,
1029 IRA_MAY_MOVE_IN_COST, and IRA_MAY_MOVE_OUT_COST for MODE if it is
1030 not done yet. */
1031 void
1032 ira_init_register_move_cost (enum machine_mode mode)
1034 int cl1, cl2;
1036 ira_assert (ira_register_move_cost[mode] == NULL
1037 && ira_may_move_in_cost[mode] == NULL
1038 && ira_may_move_out_cost[mode] == NULL);
1039 if (move_cost[mode] == NULL)
1040 init_move_cost (mode);
1041 ira_register_move_cost[mode] = move_cost[mode];
1042 /* Don't use ira_allocate because the tables exist out of scope of a
1043 IRA call. */
1044 ira_may_move_in_cost[mode]
1045 = (move_table *) xmalloc (sizeof (move_table) * N_REG_CLASSES);
1046 memcpy (ira_may_move_in_cost[mode], may_move_in_cost[mode],
1047 sizeof (move_table) * N_REG_CLASSES);
1048 ira_may_move_out_cost[mode]
1049 = (move_table *) xmalloc (sizeof (move_table) * N_REG_CLASSES);
1050 memcpy (ira_may_move_out_cost[mode], may_move_out_cost[mode],
1051 sizeof (move_table) * N_REG_CLASSES);
1052 for (cl1 = 0; cl1 < N_REG_CLASSES; cl1++)
1054 for (cl2 = 0; cl2 < N_REG_CLASSES; cl2++)
1056 if (ira_class_subset_p[cl1][cl2])
1057 ira_may_move_in_cost[mode][cl1][cl2] = 0;
1058 if (ira_class_subset_p[cl2][cl1])
1059 ira_may_move_out_cost[mode][cl1][cl2] = 0;
1066 /* Hard regsets whose all bits are correspondingly zero or one. */
1067 HARD_REG_SET ira_zero_hard_reg_set;
1068 HARD_REG_SET ira_one_hard_reg_set;
1070 /* This is called once during compiler work. It sets up
1071 different arrays whose values don't depend on the compiled
1072 function. */
1073 void
1074 ira_init_once (void)
1076 enum machine_mode mode;
1078 CLEAR_HARD_REG_SET (ira_zero_hard_reg_set);
1079 SET_HARD_REG_SET (ira_one_hard_reg_set);
1080 for (mode = 0; mode < MAX_MACHINE_MODE; mode++)
1082 ira_register_move_cost[mode] = NULL;
1083 ira_may_move_in_cost[mode] = NULL;
1084 ira_may_move_out_cost[mode] = NULL;
1086 ira_init_costs_once ();
1089 /* Free ira_register_move_cost, ira_may_move_in_cost, and
1090 ira_may_move_out_cost for each mode. */
1091 static void
1092 free_register_move_costs (void)
1094 enum machine_mode mode;
1096 for (mode = 0; mode < MAX_MACHINE_MODE; mode++)
1098 if (ira_may_move_in_cost[mode] != NULL)
1099 free (ira_may_move_in_cost[mode]);
1100 if (ira_may_move_out_cost[mode] != NULL)
1101 free (ira_may_move_out_cost[mode]);
1102 ira_register_move_cost[mode] = NULL;
1103 ira_may_move_in_cost[mode] = NULL;
1104 ira_may_move_out_cost[mode] = NULL;
1108 /* This is called every time when register related information is
1109 changed. */
1110 void
1111 ira_init (void)
1113 free_register_move_costs ();
1114 setup_reg_mode_hard_regset ();
1115 setup_alloc_regs (flag_omit_frame_pointer != 0);
1116 setup_class_subset_and_memory_move_costs ();
1117 find_reg_class_closure ();
1118 setup_reg_class_nregs ();
1119 setup_prohibited_class_mode_regs ();
1120 ira_init_costs ();
1123 /* Function called once at the end of compiler work. */
1124 void
1125 ira_finish_once (void)
1127 ira_finish_costs_once ();
1128 free_register_move_costs ();
1133 /* Array whose values are hard regset of hard registers for which
1134 move of the hard register in given mode into itself is
1135 prohibited. */
1136 HARD_REG_SET ira_prohibited_mode_move_regs[NUM_MACHINE_MODES];
1138 /* Flag of that the above array has been initialized. */
1139 static bool ira_prohibited_mode_move_regs_initialized_p = false;
1141 /* Set up IRA_PROHIBITED_MODE_MOVE_REGS. */
1142 static void
1143 setup_prohibited_mode_move_regs (void)
1145 int i, j;
1146 rtx test_reg1, test_reg2, move_pat, move_insn;
1148 if (ira_prohibited_mode_move_regs_initialized_p)
1149 return;
1150 ira_prohibited_mode_move_regs_initialized_p = true;
1151 test_reg1 = gen_rtx_REG (VOIDmode, 0);
1152 test_reg2 = gen_rtx_REG (VOIDmode, 0);
1153 move_pat = gen_rtx_SET (VOIDmode, test_reg1, test_reg2);
1154 move_insn = gen_rtx_INSN (VOIDmode, 0, 0, 0, 0, 0, move_pat, -1, 0);
1155 for (i = 0; i < NUM_MACHINE_MODES; i++)
1157 SET_HARD_REG_SET (ira_prohibited_mode_move_regs[i]);
1158 for (j = 0; j < FIRST_PSEUDO_REGISTER; j++)
1160 if (! HARD_REGNO_MODE_OK (j, i))
1161 continue;
1162 SET_REGNO (test_reg1, j);
1163 PUT_MODE (test_reg1, i);
1164 SET_REGNO (test_reg2, j);
1165 PUT_MODE (test_reg2, i);
1166 INSN_CODE (move_insn) = -1;
1167 recog_memoized (move_insn);
1168 if (INSN_CODE (move_insn) < 0)
1169 continue;
1170 extract_insn (move_insn);
1171 if (! constrain_operands (1))
1172 continue;
1173 CLEAR_HARD_REG_BIT (ira_prohibited_mode_move_regs[i], j);
1180 /* Function specific hard registers that can not be used for the
1181 register allocation. */
1182 HARD_REG_SET ira_no_alloc_regs;
1184 /* Return TRUE if *LOC contains an asm. */
1185 static int
1186 insn_contains_asm_1 (rtx *loc, void *data ATTRIBUTE_UNUSED)
1188 if ( !*loc)
1189 return FALSE;
1190 if (GET_CODE (*loc) == ASM_OPERANDS)
1191 return TRUE;
1192 return FALSE;
1196 /* Return TRUE if INSN contains an ASM. */
1197 static bool
1198 insn_contains_asm (rtx insn)
1200 return for_each_rtx (&insn, insn_contains_asm_1, NULL);
1203 /* Set up regs_asm_clobbered. */
1204 static void
1205 compute_regs_asm_clobbered (char *regs_asm_clobbered)
1207 basic_block bb;
1209 memset (regs_asm_clobbered, 0, sizeof (char) * FIRST_PSEUDO_REGISTER);
1211 FOR_EACH_BB (bb)
1213 rtx insn;
1214 FOR_BB_INSNS_REVERSE (bb, insn)
1216 struct df_ref **def_rec;
1218 if (insn_contains_asm (insn))
1219 for (def_rec = DF_INSN_DEFS (insn); *def_rec; def_rec++)
1221 struct df_ref *def = *def_rec;
1222 unsigned int dregno = DF_REF_REGNO (def);
1223 if (dregno < FIRST_PSEUDO_REGISTER)
1225 unsigned int i;
1226 enum machine_mode mode = GET_MODE (DF_REF_REAL_REG (def));
1227 unsigned int end = dregno
1228 + hard_regno_nregs[dregno][mode] - 1;
1230 for (i = dregno; i <= end; ++i)
1231 regs_asm_clobbered[i] = 1;
1239 /* Set up ELIMINABLE_REGSET, IRA_NO_ALLOC_REGS, and REGS_EVER_LIVE. */
1240 static void
1241 setup_eliminable_regset (void)
1243 int i;
1244 /* Like regs_ever_live, but 1 if a reg is set or clobbered from an
1245 asm. Unlike regs_ever_live, elements of this array corresponding
1246 to eliminable regs (like the frame pointer) are set if an asm
1247 sets them. */
1248 char *regs_asm_clobbered
1249 = (char *) alloca (FIRST_PSEUDO_REGISTER * sizeof (char));
1250 #ifdef ELIMINABLE_REGS
1251 static const struct {const int from, to; } eliminables[] = ELIMINABLE_REGS;
1252 #endif
1253 /* FIXME: If EXIT_IGNORE_STACK is set, we will not save and restore
1254 sp for alloca. So we can't eliminate the frame pointer in that
1255 case. At some point, we should improve this by emitting the
1256 sp-adjusting insns for this case. */
1257 int need_fp
1258 = (! flag_omit_frame_pointer
1259 || (cfun->calls_alloca && EXIT_IGNORE_STACK)
1260 || crtl->accesses_prior_frames
1261 || crtl->stack_realign_needed
1262 || FRAME_POINTER_REQUIRED);
1264 frame_pointer_needed = need_fp;
1266 COPY_HARD_REG_SET (ira_no_alloc_regs, no_unit_alloc_regs);
1267 CLEAR_HARD_REG_SET (eliminable_regset);
1269 compute_regs_asm_clobbered (regs_asm_clobbered);
1270 /* Build the regset of all eliminable registers and show we can't
1271 use those that we already know won't be eliminated. */
1272 #ifdef ELIMINABLE_REGS
1273 for (i = 0; i < (int) ARRAY_SIZE (eliminables); i++)
1275 bool cannot_elim
1276 = (! CAN_ELIMINATE (eliminables[i].from, eliminables[i].to)
1277 || (eliminables[i].to == STACK_POINTER_REGNUM && need_fp));
1279 if (! regs_asm_clobbered[eliminables[i].from])
1281 SET_HARD_REG_BIT (eliminable_regset, eliminables[i].from);
1283 if (cannot_elim)
1284 SET_HARD_REG_BIT (ira_no_alloc_regs, eliminables[i].from);
1286 else if (cannot_elim)
1287 error ("%s cannot be used in asm here",
1288 reg_names[eliminables[i].from]);
1289 else
1290 df_set_regs_ever_live (eliminables[i].from, true);
1292 #if FRAME_POINTER_REGNUM != HARD_FRAME_POINTER_REGNUM
1293 if (! regs_asm_clobbered[HARD_FRAME_POINTER_REGNUM])
1295 SET_HARD_REG_BIT (eliminable_regset, HARD_FRAME_POINTER_REGNUM);
1296 if (need_fp)
1297 SET_HARD_REG_BIT (ira_no_alloc_regs, HARD_FRAME_POINTER_REGNUM);
1299 else if (need_fp)
1300 error ("%s cannot be used in asm here",
1301 reg_names[HARD_FRAME_POINTER_REGNUM]);
1302 else
1303 df_set_regs_ever_live (HARD_FRAME_POINTER_REGNUM, true);
1304 #endif
1306 #else
1307 if (! regs_asm_clobbered[FRAME_POINTER_REGNUM])
1309 SET_HARD_REG_BIT (eliminable_regset, FRAME_POINTER_REGNUM);
1310 if (need_fp)
1311 SET_HARD_REG_BIT (ira_no_alloc_regs, FRAME_POINTER_REGNUM);
1313 else if (need_fp)
1314 error ("%s cannot be used in asm here", reg_names[FRAME_POINTER_REGNUM]);
1315 else
1316 df_set_regs_ever_live (FRAME_POINTER_REGNUM, true);
1317 #endif
1322 /* The length of the following two arrays. */
1323 int ira_reg_equiv_len;
1325 /* The element value is TRUE if the corresponding regno value is
1326 invariant. */
1327 bool *ira_reg_equiv_invariant_p;
1329 /* The element value is equiv constant of given pseudo-register or
1330 NULL_RTX. */
1331 rtx *ira_reg_equiv_const;
1333 /* Set up the two arrays declared above. */
1334 static void
1335 find_reg_equiv_invariant_const (void)
1337 int i;
1338 bool invariant_p;
1339 rtx list, insn, note, constant, x;
1341 for (i = FIRST_PSEUDO_REGISTER; i < reg_equiv_init_size; i++)
1343 constant = NULL_RTX;
1344 invariant_p = false;
1345 for (list = reg_equiv_init[i]; list != NULL_RTX; list = XEXP (list, 1))
1347 insn = XEXP (list, 0);
1348 note = find_reg_note (insn, REG_EQUIV, NULL_RTX);
1350 if (note == NULL_RTX)
1351 continue;
1353 x = XEXP (note, 0);
1355 if (! function_invariant_p (x)
1356 || ! flag_pic
1357 /* A function invariant is often CONSTANT_P but may
1358 include a register. We promise to only pass CONSTANT_P
1359 objects to LEGITIMATE_PIC_OPERAND_P. */
1360 || (CONSTANT_P (x) && LEGITIMATE_PIC_OPERAND_P (x)))
1362 /* It can happen that a REG_EQUIV note contains a MEM
1363 that is not a legitimate memory operand. As later
1364 stages of the reload assume that all addresses found
1365 in the reg_equiv_* arrays were originally legitimate,
1366 we ignore such REG_EQUIV notes. */
1367 if (memory_operand (x, VOIDmode))
1368 invariant_p = MEM_READONLY_P (x);
1369 else if (function_invariant_p (x))
1371 if (GET_CODE (x) == PLUS
1372 || x == frame_pointer_rtx || x == arg_pointer_rtx)
1373 invariant_p = true;
1374 else
1375 constant = x;
1379 ira_reg_equiv_invariant_p[i] = invariant_p;
1380 ira_reg_equiv_const[i] = constant;
1386 /* Set up REG_RENUMBER and CALLER_SAVE_NEEDED (used by reload) from
1387 the allocation found by IRA. */
1388 static void
1389 setup_reg_renumber (void)
1391 int regno, hard_regno;
1392 ira_allocno_t a;
1393 ira_allocno_iterator ai;
1395 caller_save_needed = 0;
1396 FOR_EACH_ALLOCNO (a, ai)
1398 /* There are no caps at this point. */
1399 ira_assert (ALLOCNO_CAP_MEMBER (a) == NULL);
1400 if (! ALLOCNO_ASSIGNED_P (a))
1401 /* It can happen if A is not referenced but partially anticipated
1402 somewhere in a region. */
1403 ALLOCNO_ASSIGNED_P (a) = true;
1404 ira_free_allocno_updated_costs (a);
1405 hard_regno = ALLOCNO_HARD_REGNO (a);
1406 regno = (int) REGNO (ALLOCNO_REG (a));
1407 reg_renumber[regno] = (hard_regno < 0 ? -1 : hard_regno);
1408 if (hard_regno >= 0 && ALLOCNO_CALLS_CROSSED_NUM (a) != 0
1409 && ! ira_hard_reg_not_in_set_p (hard_regno, ALLOCNO_MODE (a),
1410 call_used_reg_set))
1412 ira_assert (!optimize || flag_caller_saves
1413 || regno >= ira_reg_equiv_len
1414 || ira_reg_equiv_const[regno]
1415 || ira_reg_equiv_invariant_p[regno]);
1416 caller_save_needed = 1;
1421 /* Set up allocno assignment flags for further allocation
1422 improvements. */
1423 static void
1424 setup_allocno_assignment_flags (void)
1426 int hard_regno;
1427 ira_allocno_t a;
1428 ira_allocno_iterator ai;
1430 FOR_EACH_ALLOCNO (a, ai)
1432 if (! ALLOCNO_ASSIGNED_P (a))
1433 /* It can happen if A is not referenced but partially anticipated
1434 somewhere in a region. */
1435 ira_free_allocno_updated_costs (a);
1436 hard_regno = ALLOCNO_HARD_REGNO (a);
1437 /* Don't assign hard registers to allocnos which are destination
1438 of removed store at the end of loop. It has no sense to keep
1439 the same value in different hard registers. It is also
1440 impossible to assign hard registers correctly to such
1441 allocnos because the cost info and info about intersected
1442 calls are incorrect for them. */
1443 ALLOCNO_ASSIGNED_P (a) = (hard_regno >= 0
1444 || ALLOCNO_MEM_OPTIMIZED_DEST_P (a)
1445 || (ALLOCNO_MEMORY_COST (a)
1446 - ALLOCNO_COVER_CLASS_COST (a)) < 0);
1447 ira_assert (hard_regno < 0
1448 || ! ira_hard_reg_not_in_set_p (hard_regno, ALLOCNO_MODE (a),
1449 reg_class_contents
1450 [ALLOCNO_COVER_CLASS (a)]));
1454 /* Evaluate overall allocation cost and the costs for using hard
1455 registers and memory for allocnos. */
1456 static void
1457 calculate_allocation_cost (void)
1459 int hard_regno, cost;
1460 ira_allocno_t a;
1461 ira_allocno_iterator ai;
1463 ira_overall_cost = ira_reg_cost = ira_mem_cost = 0;
1464 FOR_EACH_ALLOCNO (a, ai)
1466 hard_regno = ALLOCNO_HARD_REGNO (a);
1467 ira_assert (hard_regno < 0
1468 || ! ira_hard_reg_not_in_set_p
1469 (hard_regno, ALLOCNO_MODE (a),
1470 reg_class_contents[ALLOCNO_COVER_CLASS (a)]));
1471 if (hard_regno < 0)
1473 cost = ALLOCNO_MEMORY_COST (a);
1474 ira_mem_cost += cost;
1476 else if (ALLOCNO_HARD_REG_COSTS (a) != NULL)
1478 cost = (ALLOCNO_HARD_REG_COSTS (a)
1479 [ira_class_hard_reg_index
1480 [ALLOCNO_COVER_CLASS (a)][hard_regno]]);
1481 ira_reg_cost += cost;
1483 else
1485 cost = ALLOCNO_COVER_CLASS_COST (a);
1486 ira_reg_cost += cost;
1488 ira_overall_cost += cost;
1491 if (internal_flag_ira_verbose > 0 && ira_dump_file != NULL)
1493 fprintf (ira_dump_file,
1494 "+++Costs: overall %d, reg %d, mem %d, ld %d, st %d, move %d\n",
1495 ira_overall_cost, ira_reg_cost, ira_mem_cost,
1496 ira_load_cost, ira_store_cost, ira_shuffle_cost);
1497 fprintf (ira_dump_file, "+++ move loops %d, new jumps %d\n",
1498 ira_move_loops_num, ira_additional_jumps_num);
1503 #ifdef ENABLE_IRA_CHECKING
1504 /* Check the correctness of the allocation. We do need this because
1505 of complicated code to transform more one region internal
1506 representation into one region representation. */
1507 static void
1508 check_allocation (void)
1510 ira_allocno_t a, conflict_a;
1511 int hard_regno, conflict_hard_regno, nregs, conflict_nregs;
1512 ira_allocno_conflict_iterator aci;
1513 ira_allocno_iterator ai;
1515 FOR_EACH_ALLOCNO (a, ai)
1517 if (ALLOCNO_CAP_MEMBER (a) != NULL
1518 || (hard_regno = ALLOCNO_HARD_REGNO (a)) < 0)
1519 continue;
1520 nregs = hard_regno_nregs[hard_regno][ALLOCNO_MODE (a)];
1521 FOR_EACH_ALLOCNO_CONFLICT (a, conflict_a, aci)
1522 if ((conflict_hard_regno = ALLOCNO_HARD_REGNO (conflict_a)) >= 0)
1524 conflict_nregs
1525 = (hard_regno_nregs
1526 [conflict_hard_regno][ALLOCNO_MODE (conflict_a)]);
1527 if ((conflict_hard_regno <= hard_regno
1528 && hard_regno < conflict_hard_regno + conflict_nregs)
1529 || (hard_regno <= conflict_hard_regno
1530 && conflict_hard_regno < hard_regno + nregs))
1532 fprintf (stderr, "bad allocation for %d and %d\n",
1533 ALLOCNO_REGNO (a), ALLOCNO_REGNO (conflict_a));
1534 gcc_unreachable ();
1539 #endif
1541 /* Fix values of array REG_EQUIV_INIT after live range splitting done
1542 by IRA. */
1543 static void
1544 fix_reg_equiv_init (void)
1546 int max_regno = max_reg_num ();
1547 int i, new_regno;
1548 rtx x, prev, next, insn, set;
1550 if (reg_equiv_init_size < max_regno)
1552 reg_equiv_init
1553 = (rtx *) ggc_realloc (reg_equiv_init, max_regno * sizeof (rtx));
1554 while (reg_equiv_init_size < max_regno)
1555 reg_equiv_init[reg_equiv_init_size++] = NULL_RTX;
1556 for (i = FIRST_PSEUDO_REGISTER; i < reg_equiv_init_size; i++)
1557 for (prev = NULL_RTX, x = reg_equiv_init[i]; x != NULL_RTX; x = next)
1559 next = XEXP (x, 1);
1560 insn = XEXP (x, 0);
1561 set = single_set (insn);
1562 ira_assert (set != NULL_RTX
1563 && (REG_P (SET_DEST (set)) || REG_P (SET_SRC (set))));
1564 if (REG_P (SET_DEST (set))
1565 && ((int) REGNO (SET_DEST (set)) == i
1566 || (int) ORIGINAL_REGNO (SET_DEST (set)) == i))
1567 new_regno = REGNO (SET_DEST (set));
1568 else if (REG_P (SET_SRC (set))
1569 && ((int) REGNO (SET_SRC (set)) == i
1570 || (int) ORIGINAL_REGNO (SET_SRC (set)) == i))
1571 new_regno = REGNO (SET_SRC (set));
1572 else
1573 gcc_unreachable ();
1574 if (new_regno == i)
1575 prev = x;
1576 else
1578 if (prev == NULL_RTX)
1579 reg_equiv_init[i] = next;
1580 else
1581 XEXP (prev, 1) = next;
1582 XEXP (x, 1) = reg_equiv_init[new_regno];
1583 reg_equiv_init[new_regno] = x;
1589 #ifdef ENABLE_IRA_CHECKING
1590 /* Print redundant memory-memory copies. */
1591 static void
1592 print_redundant_copies (void)
1594 int hard_regno;
1595 ira_allocno_t a;
1596 ira_copy_t cp, next_cp;
1597 ira_allocno_iterator ai;
1599 FOR_EACH_ALLOCNO (a, ai)
1601 if (ALLOCNO_CAP_MEMBER (a) != NULL)
1602 /* It is a cap. */
1603 continue;
1604 hard_regno = ALLOCNO_HARD_REGNO (a);
1605 if (hard_regno >= 0)
1606 continue;
1607 for (cp = ALLOCNO_COPIES (a); cp != NULL; cp = next_cp)
1608 if (cp->first == a)
1609 next_cp = cp->next_first_allocno_copy;
1610 else
1612 next_cp = cp->next_second_allocno_copy;
1613 if (internal_flag_ira_verbose > 4 && ira_dump_file != NULL
1614 && cp->insn != NULL_RTX
1615 && ALLOCNO_HARD_REGNO (cp->first) == hard_regno)
1616 fprintf (ira_dump_file,
1617 " Redundant move from %d(freq %d):%d\n",
1618 INSN_UID (cp->insn), cp->freq, hard_regno);
1622 #endif
1624 /* Setup preferred and alternative classes for new pseudo-registers
1625 created by IRA starting with START. */
1626 static void
1627 setup_preferred_alternate_classes_for_new_pseudos (int start)
1629 int i, old_regno;
1630 int max_regno = max_reg_num ();
1632 for (i = start; i < max_regno; i++)
1634 old_regno = ORIGINAL_REGNO (regno_reg_rtx[i]);
1635 ira_assert (i != old_regno);
1636 setup_reg_classes (i, reg_preferred_class (old_regno),
1637 reg_alternate_class (old_regno));
1638 if (internal_flag_ira_verbose > 2 && ira_dump_file != NULL)
1639 fprintf (ira_dump_file,
1640 " New r%d: setting preferred %s, alternative %s\n",
1641 i, reg_class_names[reg_preferred_class (old_regno)],
1642 reg_class_names[reg_alternate_class (old_regno)]);
1648 /* Regional allocation can create new pseudo-registers. This function
1649 expands some arrays for pseudo-registers. */
1650 static void
1651 expand_reg_info (int old_size)
1653 int i;
1654 int size = max_reg_num ();
1656 resize_reg_info ();
1657 for (i = old_size; i < size; i++)
1659 reg_renumber[i] = -1;
1660 setup_reg_classes (i, GENERAL_REGS, ALL_REGS);
1666 /* This page contains code for sorting the insn chain used by reload.
1667 In the old register allocator, the insn chain order corresponds to
1668 the order of insns in RTL. By putting insns with higher execution
1669 frequency BBs first, reload has a better chance to generate less
1670 expensive operand reloads for such insns. */
1672 /* Map bb index -> order number in the BB chain in RTL code. */
1673 static int *basic_block_order_nums;
1675 /* Map chain insn uid -> order number in the insn chain before sorting
1676 the insn chain. */
1677 static int *chain_insn_order;
1679 /* The function is used to sort insn chain according insn execution
1680 frequencies. */
1681 static int
1682 chain_freq_compare (const void *v1p, const void *v2p)
1684 const struct insn_chain *c1 = *(struct insn_chain * const *)v1p;
1685 const struct insn_chain *c2 = *(struct insn_chain * const *)v2p;
1686 int diff;
1688 diff = (BASIC_BLOCK (c2->block)->frequency
1689 - BASIC_BLOCK (c1->block)->frequency);
1690 if (diff)
1691 return diff;
1692 /* Keep the same order in BB scope. */
1693 return (chain_insn_order[INSN_UID(c1->insn)]
1694 - chain_insn_order[INSN_UID(c2->insn)]);
1697 /* Sort the insn chain according insn original order. */
1698 static int
1699 chain_bb_compare (const void *v1p, const void *v2p)
1701 const struct insn_chain *c1 = *(struct insn_chain * const *)v1p;
1702 const struct insn_chain *c2 = *(struct insn_chain * const *)v2p;
1703 int diff;
1705 diff = (basic_block_order_nums[c1->block]
1706 - basic_block_order_nums[c2->block]);
1707 if (diff)
1708 return diff;
1709 /* Keep the same order in BB scope. */
1710 return (chain_insn_order[INSN_UID(c1->insn)]
1711 - chain_insn_order[INSN_UID(c2->insn)]);
1714 /* Sort the insn chain according to insn frequencies if
1715 FREQ_P or according to insn original order otherwise. */
1716 void
1717 ira_sort_insn_chain (bool freq_p)
1719 struct insn_chain *chain, **chain_arr;
1720 basic_block bb;
1721 int i, n;
1723 chain_insn_order = (int *) ira_allocate (get_max_uid () * sizeof (int));
1724 for (n = 0, chain = reload_insn_chain; chain != 0; chain = chain->next)
1726 chain_insn_order[INSN_UID (chain->insn)] = n;
1727 n++;
1729 if (n <= 1)
1730 return;
1731 chain_arr
1732 = (struct insn_chain **) ira_allocate (n * sizeof (struct insn_chain *));
1733 basic_block_order_nums
1734 = (int *) ira_allocate (sizeof (int) * last_basic_block);
1735 n = 0;
1736 FOR_EACH_BB (bb)
1738 basic_block_order_nums[bb->index] = n++;
1740 for (n = 0, chain = reload_insn_chain; chain != 0; chain = chain->next)
1741 chain_arr[n++] = chain;
1742 qsort (chain_arr, n, sizeof (struct insn_chain *),
1743 freq_p ? chain_freq_compare : chain_bb_compare);
1744 ira_free (chain_insn_order);
1745 for (i = 1; i < n - 1; i++)
1747 chain_arr[i]->next = chain_arr[i + 1];
1748 chain_arr[i]->prev = chain_arr[i - 1];
1750 chain_arr[i]->next = NULL;
1751 chain_arr[i]->prev = chain_arr[i - 1];
1752 reload_insn_chain = chain_arr[0];
1753 reload_insn_chain->prev = NULL;
1754 reload_insn_chain->next = chain_arr[1];
1755 ira_free (basic_block_order_nums);
1756 ira_free (chain_arr);
1761 /* All natural loops. */
1762 struct loops ira_loops;
1764 /* This is the main entry of IRA. */
1765 static void
1766 ira (FILE *f)
1768 int overall_cost_before, allocated_reg_info_size;
1769 bool loops_p;
1770 int max_regno_before_ira, ira_max_point_before_emit;
1771 int rebuild_p;
1772 int saved_flag_ira_algorithm;
1773 basic_block bb;
1775 timevar_push (TV_IRA);
1777 if (flag_ira_verbose < 10)
1779 internal_flag_ira_verbose = flag_ira_verbose;
1780 ira_dump_file = f;
1782 else
1784 internal_flag_ira_verbose = flag_ira_verbose - 10;
1785 ira_dump_file = stderr;
1788 setup_prohibited_mode_move_regs ();
1790 df_note_add_problem ();
1792 if (optimize == 1)
1794 df_live_add_problem ();
1795 df_live_set_all_dirty ();
1797 #ifdef ENABLE_CHECKING
1798 df->changeable_flags |= DF_VERIFY_SCHEDULED;
1799 #endif
1800 df_analyze ();
1801 df_clear_flags (DF_NO_INSN_RESCAN);
1802 regstat_init_n_sets_and_refs ();
1803 regstat_compute_ri ();
1805 /* If we are not optimizing, then this is the only place before
1806 register allocation where dataflow is done. And that is needed
1807 to generate these warnings. */
1808 if (warn_clobbered)
1809 generate_setjmp_warnings ();
1811 rebuild_p = update_equiv_regs ();
1813 #ifndef IRA_NO_OBSTACK
1814 gcc_obstack_init (&ira_obstack);
1815 #endif
1816 bitmap_obstack_initialize (&ira_bitmap_obstack);
1817 if (optimize)
1819 max_regno = max_reg_num ();
1820 ira_reg_equiv_len = max_regno;
1821 ira_reg_equiv_invariant_p
1822 = (bool *) ira_allocate (max_regno * sizeof (bool));
1823 memset (ira_reg_equiv_invariant_p, 0, max_regno * sizeof (bool));
1824 ira_reg_equiv_const = (rtx *) ira_allocate (max_regno * sizeof (rtx));
1825 memset (ira_reg_equiv_const, 0, max_regno * sizeof (rtx));
1826 find_reg_equiv_invariant_const ();
1827 if (rebuild_p)
1829 timevar_push (TV_JUMP);
1830 rebuild_jump_labels (get_insns ());
1831 purge_all_dead_edges ();
1832 timevar_pop (TV_JUMP);
1836 max_regno_before_ira = allocated_reg_info_size = max_reg_num ();
1837 allocate_reg_info ();
1838 setup_eliminable_regset ();
1840 ira_overall_cost = ira_reg_cost = ira_mem_cost = 0;
1841 ira_load_cost = ira_store_cost = ira_shuffle_cost = 0;
1842 ira_move_loops_num = ira_additional_jumps_num = 0;
1844 ira_assert (current_loops == NULL);
1845 flow_loops_find (&ira_loops);
1846 current_loops = &ira_loops;
1847 saved_flag_ira_algorithm = flag_ira_algorithm;
1848 if (optimize && number_of_loops () > (unsigned) IRA_MAX_LOOPS_NUM)
1849 flag_ira_algorithm = IRA_ALGORITHM_CB;
1851 if (internal_flag_ira_verbose > 0 && ira_dump_file != NULL)
1852 fprintf (ira_dump_file, "Building IRA IR\n");
1853 loops_p = ira_build (optimize
1854 && (flag_ira_algorithm == IRA_ALGORITHM_REGIONAL
1855 || flag_ira_algorithm == IRA_ALGORITHM_MIXED));
1856 if (optimize)
1857 ira_color ();
1858 else
1859 ira_fast_allocation ();
1861 ira_max_point_before_emit = ira_max_point;
1863 ira_emit (loops_p);
1865 if (optimize)
1867 max_regno = max_reg_num ();
1869 if (! loops_p)
1870 ira_initiate_assign ();
1871 else
1873 expand_reg_info (allocated_reg_info_size);
1874 setup_preferred_alternate_classes_for_new_pseudos
1875 (allocated_reg_info_size);
1876 allocated_reg_info_size = max_regno;
1878 if (internal_flag_ira_verbose > 0 && ira_dump_file != NULL)
1879 fprintf (ira_dump_file, "Flattening IR\n");
1880 ira_flattening (max_regno_before_ira, ira_max_point_before_emit);
1881 /* New insns were generated: add notes and recalculate live
1882 info. */
1883 df_analyze ();
1885 flow_loops_find (&ira_loops);
1886 current_loops = &ira_loops;
1888 setup_allocno_assignment_flags ();
1889 ira_initiate_assign ();
1890 ira_reassign_conflict_allocnos (max_regno);
1894 setup_reg_renumber ();
1896 calculate_allocation_cost ();
1898 #ifdef ENABLE_IRA_CHECKING
1899 if (optimize)
1900 check_allocation ();
1901 #endif
1903 delete_trivially_dead_insns (get_insns (), max_reg_num ());
1904 max_regno = max_reg_num ();
1906 /* Determine if the current function is a leaf before running IRA
1907 since this can impact optimizations done by the prologue and
1908 epilogue thus changing register elimination offsets. */
1909 current_function_is_leaf = leaf_function_p ();
1911 /* And the reg_equiv_memory_loc array. */
1912 VEC_safe_grow (rtx, gc, reg_equiv_memory_loc_vec, max_regno);
1913 memset (VEC_address (rtx, reg_equiv_memory_loc_vec), 0,
1914 sizeof (rtx) * max_regno);
1915 reg_equiv_memory_loc = VEC_address (rtx, reg_equiv_memory_loc_vec);
1917 if (max_regno != max_regno_before_ira)
1919 regstat_free_n_sets_and_refs ();
1920 regstat_free_ri ();
1921 regstat_init_n_sets_and_refs ();
1922 regstat_compute_ri ();
1925 allocate_initial_values (reg_equiv_memory_loc);
1927 overall_cost_before = ira_overall_cost;
1928 if (optimize)
1930 fix_reg_equiv_init ();
1932 #ifdef ENABLE_IRA_CHECKING
1933 print_redundant_copies ();
1934 #endif
1936 ira_spilled_reg_stack_slots_num = 0;
1937 ira_spilled_reg_stack_slots
1938 = ((struct ira_spilled_reg_stack_slot *)
1939 ira_allocate (max_regno
1940 * sizeof (struct ira_spilled_reg_stack_slot)));
1941 memset (ira_spilled_reg_stack_slots, 0,
1942 max_regno * sizeof (struct ira_spilled_reg_stack_slot));
1945 timevar_pop (TV_IRA);
1947 timevar_push (TV_RELOAD);
1948 df_set_flags (DF_NO_INSN_RESCAN);
1949 build_insn_chain ();
1951 if (optimize)
1952 ira_sort_insn_chain (true);
1954 reload_completed = !reload (get_insns (), optimize > 0);
1956 timevar_pop (TV_RELOAD);
1958 timevar_push (TV_IRA);
1960 if (optimize)
1962 ira_free (ira_spilled_reg_stack_slots);
1964 ira_finish_assign ();
1967 if (internal_flag_ira_verbose > 0 && ira_dump_file != NULL
1968 && overall_cost_before != ira_overall_cost)
1969 fprintf (ira_dump_file, "+++Overall after reload %d\n", ira_overall_cost);
1970 ira_destroy ();
1972 flow_loops_free (&ira_loops);
1973 free_dominance_info (CDI_DOMINATORS);
1974 FOR_ALL_BB (bb)
1975 bb->loop_father = NULL;
1976 current_loops = NULL;
1978 flag_ira_algorithm = saved_flag_ira_algorithm;
1980 regstat_free_ri ();
1981 regstat_free_n_sets_and_refs ();
1983 if (optimize)
1985 cleanup_cfg (CLEANUP_EXPENSIVE);
1987 ira_free (ira_reg_equiv_invariant_p);
1988 ira_free (ira_reg_equiv_const);
1991 bitmap_obstack_release (&ira_bitmap_obstack);
1992 #ifndef IRA_NO_OBSTACK
1993 obstack_free (&ira_obstack, NULL);
1994 #endif
1996 /* The code after the reload has changed so much that at this point
1997 we might as well just rescan everything. Not that
1998 df_rescan_all_insns is not going to help here because it does not
1999 touch the artificial uses and defs. */
2000 df_finish_pass (true);
2001 if (optimize > 1)
2002 df_live_add_problem ();
2003 df_scan_alloc (NULL);
2004 df_scan_blocks ();
2006 if (optimize)
2007 df_analyze ();
2009 timevar_pop (TV_IRA);
2014 static bool
2015 gate_ira (void)
2017 return flag_ira != 0;
2020 /* Run the integrated register allocator. */
2021 static unsigned int
2022 rest_of_handle_ira (void)
2024 ira (dump_file);
2025 return 0;
2028 struct rtl_opt_pass pass_ira =
2031 RTL_PASS,
2032 "ira", /* name */
2033 gate_ira, /* gate */
2034 rest_of_handle_ira, /* execute */
2035 NULL, /* sub */
2036 NULL, /* next */
2037 0, /* static_pass_number */
2038 0, /* tv_id */
2039 0, /* properties_required */
2040 0, /* properties_provided */
2041 0, /* properties_destroyed */
2042 0, /* todo_flags_start */
2043 TODO_dump_func |
2044 TODO_ggc_collect /* todo_flags_finish */