2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief New approach to allocation and copy coalescing
23 * @author Matthias Braun
27 * ... WE NEED A NAME FOR THIS ...
29 * Only a proof of concept at this moment...
31 * The idea is to allocate registers in 2 passes:
32 * 1. A first pass to determine "preferred" registers for live-ranges. This
33 * calculates for each register and each live-range a value indicating
34 * the usefulness. (You can roughly think of the value as the negative
35 * costs needed for copies when the value is in the specific registers...)
37 * 2. Walk blocks and assigns registers in a greedy fashion. Preferring registers
38 * with high preferences. When register constraints are not met, add copies
39 * and split live-ranges.
42 * - output constraints are not ensured. The algorithm fails to copy values
43 * away, so the registers for constrained outputs are free.
44 * - must_be_different constraint is not respected
45 * - No parallel copies at basic block borders are created, no additional phis
46 * created after copies have been inserted.
47 * - Phi color assignment should give bonus points towards registers already
48 * assigned at predecessors.
49 * - think about a smarter sequence of visiting the blocks. Sorted by
50 * execfreq might be good, or looptree from inner to outermost loops going
51 * over blocks in a reverse postorder
59 #include "irgraph_t.h"
60 #include "iredges_t.h"
69 #include "bechordal_t.h"
70 #include "besched_t.h"
73 #include "bespilloptions.h"
76 #include "bipartite.h"
77 #include "hungarian.h"
79 #define USE_FACTOR 1.0f
80 #define DEF_FACTOR 1.0f
81 #define NEIGHBOR_FACTOR 0.2f
82 #define SHOULD_BE_SAME 1.0f
84 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
86 static struct obstack obst;
87 static be_irg_t *birg;
89 static const arch_register_class_t *cls;
91 static const ir_exec_freq *execfreqs;
92 static unsigned n_regs;
93 static bitset_t *ignore_regs;
95 typedef struct assignment_t assignment_t;
97 ir_node *value; /**< currently assigned value */
100 static assignment_t *assignments;
102 typedef struct allocation_info_t allocation_info_t;
103 struct allocation_info_t {
104 unsigned last_uses; /**< bitset indicating last uses (input pos) */
105 assignment_t *current_assignment;
106 float prefs[0]; /**< register preferences */
109 typedef struct reg_pref_t reg_pref_t;
115 static allocation_info_t *get_allocation_info(ir_node *node)
117 allocation_info_t *info;
118 if (!irn_visited(node)) {
119 size_t size = sizeof(info[0]) + n_regs * sizeof(float);
120 info = obstack_alloc(&obst, size);
121 memset(info, 0, size);
122 set_irn_link(node, info);
123 mark_irn_visited(node);
125 info = get_irn_link(node);
131 static void link_to(ir_node *copy, ir_node *value)
133 allocation_info_t *info = get_allocation_info(value);
134 assert(!irn_visited(copy));
135 set_irn_link(copy, info);
136 mark_irn_visited(copy);
139 static void give_penalties_for_limits(const ir_nodeset_t *live_nodes,
140 float penalty, const unsigned* limited,
143 ir_nodeset_iterator_t iter;
145 allocation_info_t *info = get_allocation_info(node);
148 /* give penalty for all forbidden regs */
149 for (r = 0; r < n_regs; ++r) {
150 if (rbitset_is_set(limited, r))
153 info->prefs[r] -= penalty;
156 /* all other live values should get a penalty for allowed regs */
157 if (live_nodes == NULL)
160 /* TODO: reduce penalty if there are multiple allowed registers... */
161 penalty *= NEIGHBOR_FACTOR;
162 foreach_ir_nodeset(live_nodes, neighbor, iter) {
163 allocation_info_t *neighbor_info;
165 /* TODO: if op is used on multiple inputs we might not do a
167 if (neighbor == node)
170 neighbor_info = get_allocation_info(neighbor);
171 for (r = 0; r < n_regs; ++r) {
172 if (!rbitset_is_set(limited, r))
175 neighbor_info->prefs[r] -= penalty;
180 static void check_defs(const ir_nodeset_t *live_nodes, float weight,
183 const arch_register_req_t *req;
185 if (get_irn_mode(node) == mode_T) {
186 const ir_edge_t *edge;
187 foreach_out_edge(node, edge) {
188 ir_node *proj = get_edge_src_irn(edge);
189 check_defs(live_nodes, weight, proj);
194 if (!arch_irn_consider_in_reg_alloc(cls, node))
197 req = arch_get_register_req_out(node);
198 if (req->type & arch_register_req_type_limited) {
199 const unsigned *limited = req->limited;
200 float penalty = weight * DEF_FACTOR;
201 give_penalties_for_limits(live_nodes, penalty, limited, node);
204 if (req->type & arch_register_req_type_should_be_same) {
205 ir_node *insn = skip_Proj(node);
206 allocation_info_t *info = get_allocation_info(node);
207 int arity = get_irn_arity(insn);
210 float factor = 1.0f / rbitset_popcnt(&req->other_same, arity);
211 for (i = 0; i < arity; ++i) {
214 allocation_info_t *op_info;
216 if (!rbitset_is_set(&req->other_same, i))
219 op = get_irn_n(insn, i);
220 op_info = get_allocation_info(op);
221 for (r = 0; r < n_regs; ++r) {
222 if (bitset_is_set(ignore_regs, r))
224 op_info->prefs[r] += info->prefs[r] * factor;
230 static void analyze_block(ir_node *block, void *data)
232 float weight = get_block_execfreq(execfreqs, block);
233 ir_nodeset_t live_nodes;
237 ir_nodeset_init(&live_nodes);
238 be_liveness_end_of_block(lv, cls, block, &live_nodes);
240 sched_foreach_reverse(block, node) {
241 allocation_info_t *info;
245 /* TODO: handle constrained phi-nodes */
249 /* TODO give/take penalties for should_be_same/different) */
250 check_defs(&live_nodes, weight, node);
253 arity = get_irn_arity(node);
254 /* I was lazy, and only allocated 1 unsigned
255 => maximum of 32 uses per node (rewrite if necessary) */
256 assert(arity <= (int) sizeof(unsigned) * 8);
258 info = get_allocation_info(node);
259 for (i = 0; i < arity; ++i) {
260 ir_node *op = get_irn_n(node, i);
261 if (!arch_irn_consider_in_reg_alloc(cls, op))
264 /* last usage of a value? */
265 if (!ir_nodeset_contains(&live_nodes, op)) {
266 rbitset_set(&info->last_uses, i);
270 be_liveness_transfer(cls, node, &live_nodes);
272 /* update weights based on usage constraints */
273 for (i = 0; i < arity; ++i) {
274 const arch_register_req_t *req;
275 const unsigned *limited;
276 ir_node *op = get_irn_n(node, i);
278 if (!arch_irn_consider_in_reg_alloc(cls, op))
281 req = arch_get_register_req(node, i);
282 if ((req->type & arch_register_req_type_limited) == 0)
285 /* TODO: give penalties to neighbors for precolored nodes! */
287 limited = req->limited;
288 give_penalties_for_limits(&live_nodes, weight * USE_FACTOR, limited,
293 ir_nodeset_destroy(&live_nodes);
296 static void use_reg(ir_node *node, const arch_register_t *reg)
298 unsigned r = arch_register_get_index(reg);
299 assignment_t *assignment = &assignments[r];
300 allocation_info_t *info;
302 assert(assignment->value == NULL);
303 assignment->value = node;
305 info = get_allocation_info(node);
306 info->current_assignment = assignment;
308 arch_set_irn_register(node, reg);
311 static int compare_reg_pref(const void *e1, const void *e2)
313 const reg_pref_t *rp1 = (const reg_pref_t*) e1;
314 const reg_pref_t *rp2 = (const reg_pref_t*) e2;
315 if (rp1->pref < rp2->pref)
317 if (rp1->pref > rp2->pref)
322 static void fill_sort_candidates(reg_pref_t *regprefs,
323 const allocation_info_t *info)
327 for (r = 0; r < n_regs; ++r) {
328 float pref = info->prefs[r];
329 if (bitset_is_set(ignore_regs, r)) {
333 regprefs[r].pref = pref;
335 /* TODO: use a stable sort here to avoid unnecessary register jumping */
336 qsort(regprefs, n_regs, sizeof(regprefs[0]), compare_reg_pref);
339 static void assign_reg(const ir_node *block, ir_node *node)
341 const arch_register_t *reg;
342 allocation_info_t *info;
343 const arch_register_req_t *req;
344 reg_pref_t *reg_prefs;
348 assert(arch_irn_consider_in_reg_alloc(cls, node));
350 /* preassigned register? */
351 reg = arch_get_irn_register(node);
353 DB((dbg, LEVEL_2, "Preassignment %+F -> %s\n", node, reg->name));
358 /* give should_be_same boni */
359 info = get_allocation_info(node);
360 req = arch_get_register_req_out(node);
362 in_node = skip_Proj(node);
363 if (req->type & arch_register_req_type_should_be_same) {
364 float weight = get_block_execfreq(execfreqs, block);
365 int arity = get_irn_arity(in_node);
368 assert(arity <= (int) sizeof(req->other_same) * 8);
369 for (i = 0; i < arity; ++i) {
371 const arch_register_t *reg;
373 if (!rbitset_is_set(&req->other_same, i))
376 in = get_irn_n(in_node, i);
377 reg = arch_get_irn_register(in);
379 r = arch_register_get_index(reg);
380 if (bitset_is_set(ignore_regs, r))
382 info->prefs[r] += weight * SHOULD_BE_SAME;
386 /* TODO: handle must_be_different */
389 DB((dbg, LEVEL_2, "Candidates for %+F:", node));
390 reg_prefs = alloca(n_regs * sizeof(reg_prefs[0]));
391 fill_sort_candidates(reg_prefs, info);
392 for (i = 0; i < n_regs; ++i) {
393 unsigned num = reg_prefs[i].num;
394 const arch_register_t *reg = arch_register_for_index(cls, num);
395 DB((dbg, LEVEL_2, " %s(%f)", reg->name, reg_prefs[i].pref));
397 DB((dbg, LEVEL_2, "\n"));
399 for (i = 0; i < n_regs; ++i) {
400 unsigned r = reg_prefs[i].num;
401 /* ignores should be last and we should have a non-ignore left */
402 assert(!bitset_is_set(ignore_regs, r));
403 /* already used? TODO: It might be better to copy the value occupying the register around here, find out when... */
404 if (assignments[r].value != NULL)
406 use_reg(node, arch_register_for_index(cls, r));
411 static void free_reg_of_value(ir_node *node)
413 allocation_info_t *info;
414 assignment_t *assignment;
417 if (!arch_irn_consider_in_reg_alloc(cls, node))
420 info = get_allocation_info(node);
421 assignment = info->current_assignment;
423 assert(assignment != NULL);
425 r = assignment - assignments;
426 DB((dbg, LEVEL_2, "Value %+F ended, freeing %s\n",
427 node, arch_register_for_index(cls, r)->name));
428 assignment->value = NULL;
429 info->current_assignment = NULL;
432 static unsigned get_current_reg(ir_node *node)
434 allocation_info_t *info = get_allocation_info(node);
435 assignment_t *assignment = info->current_assignment;
436 return assignment - assignments;
439 static assignment_t *get_current_assignment(ir_node *node)
441 allocation_info_t *info = get_allocation_info(node);
442 return info->current_assignment;
445 static void permutate_values(ir_nodeset_t *live_nodes, ir_node *before,
446 unsigned *permutation)
448 ir_node *block, *perm;
449 ir_node **in = ALLOCAN(ir_node*, n_regs);
453 for (r = 0; r < n_regs; ++r) {
454 unsigned new_reg = permutation[r];
455 assignment_t *assignment;
461 assignment = &assignments[r];
462 value = assignment->value;
464 /* nothing to do here, reg is not live */
471 free_reg_of_value(value);
472 ir_nodeset_remove(live_nodes, value);
475 block = get_nodes_block(before);
476 perm = be_new_Perm(cls, irg, block, i, in);
478 sched_add_before(before, perm);
481 for (r = 0; r < n_regs; ++r) {
482 unsigned new_reg = permutation[r];
486 const arch_register_t *reg;
492 mode = get_irn_mode(value);
493 proj = new_r_Proj(irg, block, perm, mode, i);
495 reg = arch_register_for_index(cls, new_reg);
497 link_to(proj, value);
499 ir_nodeset_insert(live_nodes, proj);
505 /* free regs for values last used */
506 static void free_last_uses(ir_nodeset_t *live_nodes, ir_node *node)
508 allocation_info_t *info = get_allocation_info(node);
509 int arity = get_irn_arity(node);
511 for (i = 0; i < arity; ++i) {
514 if (!rbitset_is_set(&info->last_uses, i))
517 op = get_irn_n(node, i);
518 free_reg_of_value(op);
519 ir_nodeset_remove(live_nodes, op);
523 static void enforce_constraints(ir_nodeset_t *live_nodes, ir_node *node)
525 int arity = get_irn_arity(node);
527 hungarian_problem_t *bp;
529 unsigned *assignment;
531 /* see if any use constraints are not met */
533 for (i = 0; i < arity; ++i) {
534 ir_node *op = get_irn_n(node, i);
535 const arch_register_req_t *req;
536 const unsigned *limited;
539 if (!arch_irn_consider_in_reg_alloc(cls, op))
542 req = arch_get_register_req(node, i);
543 if ((req->type & arch_register_req_type_limited) == 0)
546 limited = req->limited;
547 r = get_current_reg(op);
548 if (!rbitset_is_set(limited, r)) {
557 /* swap values around */
558 bp = hungarian_new(n_regs, n_regs, HUNGARIAN_MATCH_PERFECT);
560 /* add all combinations, then remove not allowed ones */
561 for (l = 0; l < n_regs; ++l) {
562 if (bitset_is_set(ignore_regs, l)) {
563 hungarian_add(bp, l, l, 90);
567 for (r = 0; r < n_regs; ++r) {
568 if (bitset_is_set(ignore_regs, r))
571 hungarian_add(bp, l, r, l == r ? 90 : 89);
575 for (i = 0; i < arity; ++i) {
576 ir_node *op = get_irn_n(node, i);
577 const arch_register_req_t *req;
578 const unsigned *limited;
579 unsigned current_reg;
581 if (!arch_irn_consider_in_reg_alloc(cls, op))
584 req = arch_get_register_req(node, i);
585 if ((req->type & arch_register_req_type_limited) == 0)
588 limited = req->limited;
589 current_reg = get_current_reg(op);
590 for (r = 0; r < n_regs; ++r) {
591 if (rbitset_is_set(limited, r))
593 hungarian_remv(bp, current_reg, r);
597 hungarian_print_costmatrix(bp, 1);
598 hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
600 assignment = ALLOCAN(unsigned, n_regs);
601 res = hungarian_solve(bp, (int*) assignment, &dummy, 0);
604 printf("Swap result:");
605 for (p = 0; p < n_regs; ++p) {
606 printf(" %d", assignment[p]);
612 permutate_values(live_nodes, node, assignment);
615 static void allocate_coalesce_block(ir_node *block, void *data)
618 ir_nodeset_t live_nodes;
619 ir_nodeset_iterator_t iter;
620 ir_node *node, *start;
622 /* clear assignments */
623 memset(assignments, 0, n_regs * sizeof(assignments[0]));
627 /* collect live-in nodes and preassigned values */
628 ir_nodeset_init(&live_nodes);
629 be_lv_foreach(lv, block, be_lv_state_in, i) {
630 const arch_register_t *reg;
632 node = be_lv_get_irn(lv, block, i);
633 if (!arch_irn_consider_in_reg_alloc(cls, node))
636 ir_nodeset_insert(&live_nodes, node);
638 /* fill in regs already assigned */
639 reg = arch_get_irn_register(node);
646 node = sched_first(block);
647 for ( ; is_Phi(node); node = sched_next(node)) {
648 const arch_register_t *reg;
650 if (!arch_irn_consider_in_reg_alloc(cls, node))
653 /* fill in regs already assigned */
654 reg = arch_get_irn_register(node);
658 /* TODO: give boni for registers already assigned at the
664 /* assign regs for live-in values */
665 foreach_ir_nodeset(&live_nodes, node, iter) {
666 assign_reg(block, node);
669 /* assign instructions in the block */
670 for (node = start; !sched_is_end(node); node = sched_next(node)) {
671 int arity = get_irn_arity(node);
674 /* enforce use constraints */
675 enforce_constraints(&live_nodes, node);
677 /* exchange values to copied values where needed */
678 for (i = 0; i < arity; ++i) {
679 ir_node *op = get_irn_n(node, i);
680 assignment_t *assignment;
682 if (!arch_irn_consider_in_reg_alloc(cls, op))
684 assignment = get_current_assignment(op);
685 assert(assignment != NULL);
686 if (op != assignment->value) {
687 set_irn_n(node, i, assignment->value);
691 free_last_uses(&live_nodes, node);
693 /* assign output registers */
694 /* TODO: 2 phases: first: pre-assigned ones, 2nd real regs */
695 if (get_irn_mode(node) == mode_T) {
696 const ir_edge_t *edge;
697 foreach_out_edge(node, edge) {
698 ir_node *proj = get_edge_src_irn(edge);
699 if (!arch_irn_consider_in_reg_alloc(cls, proj))
701 assign_reg(block, proj);
703 } else if (arch_irn_consider_in_reg_alloc(cls, node)) {
704 assign_reg(block, node);
708 foreach_ir_nodeset(&live_nodes, node, iter) {
709 free_reg_of_value(node);
712 ir_nodeset_destroy(&live_nodes);
715 void be_straight_alloc_cls(void)
717 n_regs = arch_register_class_n_regs(cls);
718 lv = be_assure_liveness(birg);
719 be_liveness_assure_sets(lv);
720 be_liveness_assure_chk(lv);
722 assignments = obstack_alloc(&obst, n_regs * sizeof(assignments[0]));
724 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_IRN_VISITED);
725 inc_irg_visited(irg);
727 irg_block_walk_graph(irg, analyze_block, NULL, NULL);
728 irg_block_walk_graph(irg, allocate_coalesce_block, NULL, NULL);
730 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_IRN_VISITED);
733 static void spill(void)
735 /* TODO: rewrite pre_spill_prepare to work without chordal_env... */
736 be_chordal_env_t cenv;
737 memset(&cenv, 0, sizeof(cenv));
742 cenv.ignore_colors = ignore_regs;
744 /* make sure all nodes show their real register pressure */
745 be_pre_spill_prepare_constr(&cenv);
748 be_do_spill(birg, cls);
750 check_for_memory_operands(irg);
753 static void be_straight_alloc(be_irg_t *new_birg)
755 const arch_env_t *arch_env = new_birg->main_env->arch_env;
756 int n_cls = arch_env_get_n_reg_class(arch_env);
762 irg = be_get_birg_irg(birg);
763 execfreqs = birg->exec_freq;
765 /* TODO: extract some of the stuff from bechordal allocator, like
766 * statistics, time measurements, etc. and use them here too */
768 for (c = 0; c < n_cls; ++c) {
769 cls = arch_env_get_reg_class(arch_env, c);
770 if (arch_register_class_flags(cls) & arch_register_class_flag_manual_ra)
773 stat_ev_ctx_push_str("bestraight_cls", cls->name);
775 n_regs = cls->n_regs;
776 ignore_regs = bitset_malloc(n_regs);
777 be_put_ignore_regs(birg, cls, ignore_regs);
781 /* verify schedule and register pressure */
782 BE_TIMER_PUSH(t_verify);
783 if (birg->main_env->options->vrfy_option == BE_CH_VRFY_WARN) {
784 be_verify_schedule(birg);
785 be_verify_register_pressure(birg, cls, irg);
786 } else if (birg->main_env->options->vrfy_option == BE_CH_VRFY_ASSERT) {
787 assert(be_verify_schedule(birg) && "Schedule verification failed");
788 assert(be_verify_register_pressure(birg, cls, irg)
789 && "Register pressure verification failed");
791 BE_TIMER_POP(t_verify);
793 BE_TIMER_PUSH(t_ra_color);
794 be_straight_alloc_cls();
795 BE_TIMER_POP(t_ra_color);
797 bitset_free(ignore_regs);
799 /* TODO: dump intermediate results */
801 stat_ev_ctx_pop("bestraight_cls");
804 obstack_free(&obst, NULL);
807 static be_ra_t be_ra_straight = {
811 void be_init_straight_alloc(void)
813 FIRM_DBG_REGISTER(dbg, "firm.be.straightalloc");
815 be_register_allocator("straight", &be_ra_straight);
818 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_straight_alloc);