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 Preference Guided Register Assignment
23 * @author Matthias Braun
27 * The idea is to allocate registers in 2 passes:
28 * 1. A first pass to determine "preferred" registers for live-ranges. This
29 * calculates for each register and each live-range a value indicating
30 * the usefulness. (You can roughly think of the value as the negative
31 * costs needed for copies when the value is in the specific registers...)
33 * 2. Walk blocks and assigns registers in a greedy fashion. Preferring
34 * registers with high preferences. When register constraints are not met,
35 * add copies and split live-ranges.
38 * - make use of free registers in the permute_values code
50 #include "iredges_t.h"
51 #include "irgraph_t.h"
56 #include "raw_bitset.h"
57 #include "unionfind.h"
59 #include "hungarian.h"
62 #include "bechordal_t.h"
71 #include "bespillutil.h"
75 #define USE_FACTOR 1.0f
76 #define DEF_FACTOR 1.0f
77 #define NEIGHBOR_FACTOR 0.2f
78 #define AFF_SHOULD_BE_SAME 0.5f
80 #define SPLIT_DELTA 1.0f
81 #define MAX_OPTIMISTIC_SPLIT_RECURSION 0
83 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
85 static struct obstack obst;
86 static be_irg_t *birg;
88 static const arch_register_class_t *cls;
89 static const arch_register_req_t *default_cls_req;
91 static const ir_exec_freq *execfreqs;
92 static unsigned n_regs;
93 static unsigned *normal_regs;
94 static int *congruence_classes;
95 static ir_node **block_order;
96 static int n_block_order;
97 static int create_preferences = true;
98 static int create_congruence_classes = true;
99 static int propagate_phi_registers = true;
101 static const lc_opt_table_entry_t options[] = {
102 LC_OPT_ENT_BOOL("prefs", "use preference based coloring", &create_preferences),
103 LC_OPT_ENT_BOOL("congruences", "create congruence classes", &create_congruence_classes),
104 LC_OPT_ENT_BOOL("prop_phi", "propagate phi registers", &propagate_phi_registers),
108 /** currently active assignments (while processing a basic block)
109 * maps registers to values(their current copies) */
110 static ir_node **assignments;
113 * allocation information: last_uses, register preferences
114 * the information is per firm-node.
116 struct allocation_info_t {
117 unsigned last_uses; /**< bitset indicating last uses (input pos) */
118 ir_node *current_value; /**< copy of the value that should be used */
119 ir_node *original_value; /**< for copies point to original value */
120 float prefs[0]; /**< register preferences */
122 typedef struct allocation_info_t allocation_info_t;
124 /** helper datastructure used when sorting register preferences */
129 typedef struct reg_pref_t reg_pref_t;
131 /** per basic-block information */
132 struct block_info_t {
133 bool processed; /**< indicate whether block is processed */
134 ir_node *assignments[0]; /**< register assignments at end of block */
136 typedef struct block_info_t block_info_t;
139 * Get the allocation info for a node.
140 * The info is allocated on the first visit of a node.
142 static allocation_info_t *get_allocation_info(ir_node *node)
144 allocation_info_t *info = get_irn_link(node);
146 info = OALLOCFZ(&obst, allocation_info_t, prefs, n_regs);
147 info->current_value = node;
148 info->original_value = node;
149 set_irn_link(node, info);
155 static allocation_info_t *try_get_allocation_info(const ir_node *node)
157 return (allocation_info_t*) get_irn_link(node);
161 * Get allocation information for a basic block
163 static block_info_t *get_block_info(ir_node *block)
165 block_info_t *info = get_irn_link(block);
167 assert(is_Block(block));
169 info = OALLOCFZ(&obst, block_info_t, assignments, n_regs);
170 set_irn_link(block, info);
177 * Get default register requirement for the current register class
179 static const arch_register_req_t *get_default_req_current_cls(void)
181 if (default_cls_req == NULL) {
182 struct obstack *obst = get_irg_obstack(irg);
183 arch_register_req_t *req = OALLOCZ(obst, arch_register_req_t);
185 req->type = arch_register_req_type_normal;
188 default_cls_req = req;
190 return default_cls_req;
194 * Link the allocation info of a node to a copy.
195 * Afterwards, both nodes uses the same allocation info.
196 * Copy must not have an allocation info assigned yet.
198 * @param copy the node that gets the allocation info assigned
199 * @param value the original node
201 static void mark_as_copy_of(ir_node *copy, ir_node *value)
204 allocation_info_t *info = get_allocation_info(value);
205 allocation_info_t *copy_info = get_allocation_info(copy);
207 /* find original value */
208 original = info->original_value;
209 if (original != value) {
210 info = get_allocation_info(original);
213 assert(info->original_value == original);
214 info->current_value = copy;
216 /* the copy should not be linked to something else yet */
217 assert(copy_info->original_value == copy);
218 copy_info->original_value = original;
220 /* copy over allocation preferences */
221 memcpy(copy_info->prefs, info->prefs, n_regs * sizeof(copy_info->prefs[0]));
225 * Calculate the penalties for every register on a node and its live neighbors.
227 * @param live_nodes the set of live nodes at the current position, may be NULL
228 * @param penalty the penalty to subtract from
229 * @param limited a raw bitset containing the limited set for the node
230 * @param node the node
232 static void give_penalties_for_limits(const ir_nodeset_t *live_nodes,
233 float penalty, const unsigned* limited,
236 ir_nodeset_iterator_t iter;
239 allocation_info_t *info = get_allocation_info(node);
242 /* give penalty for all forbidden regs */
243 for (r = 0; r < n_regs; ++r) {
244 if (rbitset_is_set(limited, r))
247 info->prefs[r] -= penalty;
250 /* all other live values should get a penalty for allowed regs */
251 if (live_nodes == NULL)
254 penalty *= NEIGHBOR_FACTOR;
255 n_allowed = rbitset_popcount(limited, n_regs);
257 /* only create a very weak penalty if multiple regs are allowed */
258 penalty = (penalty * 0.8f) / n_allowed;
260 foreach_ir_nodeset(live_nodes, neighbor, iter) {
261 allocation_info_t *neighbor_info;
263 /* TODO: if op is used on multiple inputs we might not do a
265 if (neighbor == node)
268 neighbor_info = get_allocation_info(neighbor);
269 for (r = 0; r < n_regs; ++r) {
270 if (!rbitset_is_set(limited, r))
273 neighbor_info->prefs[r] -= penalty;
279 * Calculate the preferences of a definition for the current register class.
280 * If the definition uses a limited set of registers, reduce the preferences
281 * for the limited register on the node and its neighbors.
283 * @param live_nodes the set of live nodes at the current node
284 * @param weight the weight
285 * @param node the current node
287 static void check_defs(const ir_nodeset_t *live_nodes, float weight,
290 const arch_register_req_t *req;
292 if (get_irn_mode(node) == mode_T) {
293 const ir_edge_t *edge;
294 foreach_out_edge(node, edge) {
295 ir_node *proj = get_edge_src_irn(edge);
296 check_defs(live_nodes, weight, proj);
301 if (!arch_irn_consider_in_reg_alloc(cls, node))
304 req = arch_get_register_req_out(node);
305 if (req->type & arch_register_req_type_limited) {
306 const unsigned *limited = req->limited;
307 float penalty = weight * DEF_FACTOR;
308 give_penalties_for_limits(live_nodes, penalty, limited, node);
311 if (req->type & arch_register_req_type_should_be_same) {
312 ir_node *insn = skip_Proj(node);
313 allocation_info_t *info = get_allocation_info(node);
314 int arity = get_irn_arity(insn);
317 float factor = 1.0f / rbitset_popcount(&req->other_same, arity);
318 for (i = 0; i < arity; ++i) {
321 allocation_info_t *op_info;
323 if (!rbitset_is_set(&req->other_same, i))
326 op = get_irn_n(insn, i);
328 /* if we the value at the should_be_same input doesn't die at the
329 * node, then it is no use to propagate the constraints (since a
330 * copy will emerge anyway) */
331 if (ir_nodeset_contains(live_nodes, op))
334 op_info = get_allocation_info(op);
335 for (r = 0; r < n_regs; ++r) {
336 op_info->prefs[r] += info->prefs[r] * factor;
343 * Walker: Runs an a block calculates the preferences for any
344 * node and every register from the considered register class.
346 static void analyze_block(ir_node *block, void *data)
348 float weight = (float)get_block_execfreq(execfreqs, block);
349 ir_nodeset_t live_nodes;
353 ir_nodeset_init(&live_nodes);
354 be_liveness_end_of_block(lv, cls, block, &live_nodes);
356 sched_foreach_reverse(block, node) {
357 allocation_info_t *info;
364 if (create_preferences)
365 check_defs(&live_nodes, weight, node);
368 arity = get_irn_arity(node);
370 /* the allocation info node currently only uses 1 unsigned value
371 to mark last used inputs. So we will fail for a node with more than
373 if (arity >= (int) sizeof(unsigned) * 8) {
374 panic("Node with more than %d inputs not supported yet",
375 (int) sizeof(unsigned) * 8);
378 info = get_allocation_info(node);
379 for (i = 0; i < arity; ++i) {
380 ir_node *op = get_irn_n(node, i);
381 if (!arch_irn_consider_in_reg_alloc(cls, op))
384 /* last usage of a value? */
385 if (!ir_nodeset_contains(&live_nodes, op)) {
386 rbitset_set(&info->last_uses, i);
390 be_liveness_transfer(cls, node, &live_nodes);
392 if (create_preferences) {
393 /* update weights based on usage constraints */
394 for (i = 0; i < arity; ++i) {
395 const arch_register_req_t *req;
396 const unsigned *limited;
397 ir_node *op = get_irn_n(node, i);
399 if (!arch_irn_consider_in_reg_alloc(cls, op))
402 req = arch_get_register_req(node, i);
403 if (!(req->type & arch_register_req_type_limited))
406 limited = req->limited;
407 give_penalties_for_limits(&live_nodes, weight * USE_FACTOR,
413 ir_nodeset_destroy(&live_nodes);
416 static void congruence_def(ir_nodeset_t *live_nodes, ir_node *node)
418 const arch_register_req_t *req;
420 if (get_irn_mode(node) == mode_T) {
421 const ir_edge_t *edge;
422 foreach_out_edge(node, edge) {
423 ir_node *def = get_edge_src_irn(edge);
424 congruence_def(live_nodes, def);
429 if (!arch_irn_consider_in_reg_alloc(cls, node))
432 /* should be same constraint? */
433 req = arch_get_register_req_out(node);
434 if (req->type & arch_register_req_type_should_be_same) {
435 ir_node *insn = skip_Proj(node);
436 int arity = get_irn_arity(insn);
438 unsigned node_idx = get_irn_idx(node);
439 node_idx = uf_find(congruence_classes, node_idx);
441 for (i = 0; i < arity; ++i) {
445 ir_nodeset_iterator_t iter;
446 bool interferes = false;
448 if (!rbitset_is_set(&req->other_same, i))
451 op = get_irn_n(insn, i);
452 op_idx = get_irn_idx(op);
453 op_idx = uf_find(congruence_classes, op_idx);
455 /* do we interfere with the value */
456 foreach_ir_nodeset(live_nodes, live, iter) {
457 int lv_idx = get_irn_idx(live);
458 lv_idx = uf_find(congruence_classes, lv_idx);
459 if (lv_idx == op_idx) {
464 /* don't put in same affinity class if we interfere */
468 node_idx = uf_union(congruence_classes, node_idx, op_idx);
469 DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
471 /* one should_be_same is enough... */
477 static void create_congruence_class(ir_node *block, void *data)
479 ir_nodeset_t live_nodes;
483 ir_nodeset_init(&live_nodes);
484 be_liveness_end_of_block(lv, cls, block, &live_nodes);
486 /* check should be same constraints */
487 sched_foreach_reverse(block, node) {
491 congruence_def(&live_nodes, node);
492 be_liveness_transfer(cls, node, &live_nodes);
495 /* check phi congruence classes */
496 sched_foreach_reverse_from(node, node) {
500 assert(is_Phi(node));
502 if (!arch_irn_consider_in_reg_alloc(cls, node))
505 node_idx = get_irn_idx(node);
506 node_idx = uf_find(congruence_classes, node_idx);
508 arity = get_irn_arity(node);
509 for (i = 0; i < arity; ++i) {
510 bool interferes = false;
511 ir_nodeset_iterator_t iter;
516 allocation_info_t *head_info;
517 allocation_info_t *other_info;
518 ir_node *op = get_Phi_pred(node, i);
519 int op_idx = get_irn_idx(op);
520 op_idx = uf_find(congruence_classes, op_idx);
522 /* do we interfere with the value */
523 foreach_ir_nodeset(&live_nodes, live, iter) {
524 int lv_idx = get_irn_idx(live);
525 lv_idx = uf_find(congruence_classes, lv_idx);
526 if (lv_idx == op_idx) {
531 /* don't put in same affinity class if we interfere */
534 /* any other phi has the same input? */
535 sched_foreach(block, phi) {
540 if (!arch_irn_consider_in_reg_alloc(cls, phi))
542 oop = get_Phi_pred(phi, i);
545 oop_idx = get_irn_idx(oop);
546 oop_idx = uf_find(congruence_classes, oop_idx);
547 if (oop_idx == op_idx) {
555 /* merge the 2 congruence classes and sum up their preferences */
556 old_node_idx = node_idx;
557 node_idx = uf_union(congruence_classes, node_idx, op_idx);
558 DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
561 old_node_idx = node_idx == old_node_idx ? op_idx : old_node_idx;
562 head_info = get_allocation_info(get_idx_irn(irg, node_idx));
563 other_info = get_allocation_info(get_idx_irn(irg, old_node_idx));
564 for (r = 0; r < n_regs; ++r) {
565 head_info->prefs[r] += other_info->prefs[r];
571 static void set_congruence_prefs(ir_node *node, void *data)
573 allocation_info_t *info;
574 allocation_info_t *head_info;
575 unsigned node_idx = get_irn_idx(node);
576 unsigned node_set = uf_find(congruence_classes, node_idx);
580 /* head of congruence class or not in any class */
581 if (node_set == node_idx)
584 if (!arch_irn_consider_in_reg_alloc(cls, node))
587 head_info = get_allocation_info(get_idx_irn(irg, node_set));
588 info = get_allocation_info(node);
590 memcpy(info->prefs, head_info->prefs, n_regs * sizeof(info->prefs[0]));
593 static void combine_congruence_classes(void)
595 size_t n = get_irg_last_idx(irg);
596 congruence_classes = XMALLOCN(int, n);
597 uf_init(congruence_classes, n);
599 /* create congruence classes */
600 irg_block_walk_graph(irg, create_congruence_class, NULL, NULL);
601 /* merge preferences */
602 irg_walk_graph(irg, set_congruence_prefs, NULL, NULL);
603 free(congruence_classes);
611 * Assign register reg to the given node.
613 * @param node the node
614 * @param reg the register
616 static void use_reg(ir_node *node, const arch_register_t *reg)
618 unsigned r = arch_register_get_index(reg);
619 assignments[r] = node;
620 arch_set_irn_register(node, reg);
623 static void free_reg_of_value(ir_node *node)
625 const arch_register_t *reg;
628 if (!arch_irn_consider_in_reg_alloc(cls, node))
631 reg = arch_get_irn_register(node);
632 r = arch_register_get_index(reg);
633 /* assignment->value may be NULL if a value is used at 2 inputs
634 so it gets freed twice. */
635 assert(assignments[r] == node || assignments[r] == NULL);
636 assignments[r] = NULL;
640 * Compare two register preferences in decreasing order.
642 static int compare_reg_pref(const void *e1, const void *e2)
644 const reg_pref_t *rp1 = (const reg_pref_t*) e1;
645 const reg_pref_t *rp2 = (const reg_pref_t*) e2;
646 if (rp1->pref < rp2->pref)
648 if (rp1->pref > rp2->pref)
653 static void fill_sort_candidates(reg_pref_t *regprefs,
654 const allocation_info_t *info)
658 for (r = 0; r < n_regs; ++r) {
659 float pref = info->prefs[r];
661 regprefs[r].pref = pref;
663 /* TODO: use a stable sort here to avoid unnecessary register jumping */
664 qsort(regprefs, n_regs, sizeof(regprefs[0]), compare_reg_pref);
667 static bool try_optimistic_split(ir_node *to_split, ir_node *before,
668 float pref, float pref_delta,
669 unsigned *forbidden_regs, int recursion)
671 const arch_register_t *from_reg;
672 const arch_register_t *reg;
673 ir_node *original_insn;
679 allocation_info_t *info = get_allocation_info(to_split);
682 float split_threshold;
686 /* stupid hack: don't optimisticallt split don't spill nodes...
687 * (so we don't split away the values produced because of
688 * must_be_different constraints) */
689 original_insn = skip_Proj(info->original_value);
690 if (arch_irn_get_flags(original_insn) & arch_irn_flags_dont_spill)
693 from_reg = arch_get_irn_register(to_split);
694 from_r = arch_register_get_index(from_reg);
695 block = get_nodes_block(before);
696 split_threshold = (float)get_block_execfreq(execfreqs, block) * SPLIT_DELTA;
698 if (pref_delta < split_threshold*0.5)
701 /* find the best free position where we could move to */
702 prefs = ALLOCAN(reg_pref_t, n_regs);
703 fill_sort_candidates(prefs, info);
704 for (i = 0; i < n_regs; ++i) {
708 bool old_source_state;
710 /* we need a normal register which is not an output register
711 an different from the current register of to_split */
713 if (!rbitset_is_set(normal_regs, r))
715 if (rbitset_is_set(forbidden_regs, r))
720 /* is the split worth it? */
721 delta = pref_delta + prefs[i].pref;
722 if (delta < split_threshold) {
723 DB((dbg, LEVEL_3, "Not doing optimistical split of %+F (depth %d), win %f too low\n",
724 to_split, recursion, delta));
728 /* if the register is free then we can do the split */
729 if (assignments[r] == NULL)
732 /* otherwise we might try recursively calling optimistic_split */
733 if (recursion+1 > MAX_OPTIMISTIC_SPLIT_RECURSION)
736 apref = prefs[i].pref;
737 apref_delta = i+1 < n_regs ? apref - prefs[i+1].pref : 0;
738 apref_delta += pref_delta - split_threshold;
740 /* our source register isn't a usefull destination for recursive
742 old_source_state = rbitset_is_set(forbidden_regs, from_r);
743 rbitset_set(forbidden_regs, from_r);
744 /* try recursive split */
745 res = try_optimistic_split(assignments[r], before, apref,
746 apref_delta, forbidden_regs, recursion+1);
747 /* restore our destination */
748 if (old_source_state) {
749 rbitset_set(forbidden_regs, from_r);
751 rbitset_clear(forbidden_regs, from_r);
760 reg = arch_register_for_index(cls, r);
761 copy = be_new_Copy(cls, block, to_split);
762 mark_as_copy_of(copy, to_split);
763 /* hacky, but correct here */
764 if (assignments[arch_register_get_index(from_reg)] == to_split)
765 free_reg_of_value(to_split);
767 sched_add_before(before, copy);
770 "Optimistic live-range split %+F move %+F(%s) -> %s before %+F (win %f, depth %d)\n",
771 copy, to_split, from_reg->name, reg->name, before, delta, recursion));
776 * Determine and assign a register for node @p node
778 static void assign_reg(const ir_node *block, ir_node *node,
779 unsigned *forbidden_regs)
781 const arch_register_t *reg;
782 allocation_info_t *info;
783 const arch_register_req_t *req;
784 reg_pref_t *reg_prefs;
787 const unsigned *allowed_regs;
790 assert(!is_Phi(node));
791 assert(arch_irn_consider_in_reg_alloc(cls, node));
793 /* preassigned register? */
794 reg = arch_get_irn_register(node);
796 DB((dbg, LEVEL_2, "Preassignment %+F -> %s\n", node, reg->name));
801 /* give should_be_same boni */
802 info = get_allocation_info(node);
803 req = arch_get_register_req_out(node);
805 in_node = skip_Proj(node);
806 if (req->type & arch_register_req_type_should_be_same) {
807 float weight = (float)get_block_execfreq(execfreqs, block);
808 int arity = get_irn_arity(in_node);
811 assert(arity <= (int) sizeof(req->other_same) * 8);
812 for (i = 0; i < arity; ++i) {
814 const arch_register_t *reg;
816 if (!rbitset_is_set(&req->other_same, i))
819 in = get_irn_n(in_node, i);
820 reg = arch_get_irn_register(in);
822 r = arch_register_get_index(reg);
824 /* if the value didn't die here then we should not propagate the
825 * should_be_same info */
826 if (assignments[r] == in)
829 info->prefs[r] += weight * AFF_SHOULD_BE_SAME;
833 /* create list of register candidates and sort by their preference */
834 DB((dbg, LEVEL_2, "Candidates for %+F:", node));
835 reg_prefs = alloca(n_regs * sizeof(reg_prefs[0]));
836 fill_sort_candidates(reg_prefs, info);
837 for (i = 0; i < n_regs; ++i) {
838 unsigned num = reg_prefs[i].num;
839 const arch_register_t *reg;
841 if (!rbitset_is_set(normal_regs, num))
844 reg = arch_register_for_index(cls, num);
845 DB((dbg, LEVEL_2, " %s(%f)", reg->name, reg_prefs[i].pref));
847 DB((dbg, LEVEL_2, "\n"));
849 allowed_regs = normal_regs;
850 if (req->type & arch_register_req_type_limited) {
851 allowed_regs = req->limited;
854 for (i = 0; i < n_regs; ++i) {
859 r = reg_prefs[i].num;
860 if (!rbitset_is_set(allowed_regs, r))
862 if (assignments[r] == NULL)
864 pref = reg_prefs[i].pref;
865 delta = i+1 < n_regs ? pref - reg_prefs[i+1].pref : 0;
866 before = skip_Proj(node);
867 res = try_optimistic_split(assignments[r], before,
868 pref, delta, forbidden_regs, 0);
873 panic("No register left for %+F\n", node);
876 reg = arch_register_for_index(cls, r);
877 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
882 * Add an permutation in front of a node and change the assignments
883 * due to this permutation.
885 * To understand this imagine a permutation like this:
895 * First we count how many destinations a single value has. At the same time
896 * we can be sure that each destination register has at most 1 source register
897 * (it can have 0 which means we don't care what value is in it).
898 * We ignore all fullfilled permuations (like 7->7)
899 * In a first pass we create as much copy instructions as possible as they
900 * are generally cheaper than exchanges. We do this by counting into how many
901 * destinations a register has to be copied (in the example it's 2 for register
902 * 3, or 1 for the registers 1,2,4 and 7).
903 * We can then create a copy into every destination register when the usecount
904 * of that register is 0 (= noone else needs the value in the register).
906 * After this step we should have cycles left. We implement a cyclic permutation
907 * of n registers with n-1 transpositions.
909 * @param live_nodes the set of live nodes, updated due to live range split
910 * @param before the node before we add the permutation
911 * @param permutation the permutation array indices are the destination
912 * registers, the values in the array are the source
915 static void permute_values(ir_nodeset_t *live_nodes, ir_node *before,
916 unsigned *permutation)
918 unsigned *n_used = ALLOCANZ(unsigned, n_regs);
922 /* determine how often each source register needs to be read */
923 for (r = 0; r < n_regs; ++r) {
924 unsigned old_reg = permutation[r];
927 value = assignments[old_reg];
929 /* nothing to do here, reg is not live. Mark it as fixpoint
930 * so we ignore it in the next steps */
938 block = get_nodes_block(before);
940 /* step1: create copies where immediately possible */
941 for (r = 0; r < n_regs; /* empty */) {
944 const arch_register_t *reg;
945 unsigned old_r = permutation[r];
947 /* - no need to do anything for fixed points.
948 - we can't copy if the value in the dest reg is still needed */
949 if (old_r == r || n_used[r] > 0) {
955 src = assignments[old_r];
956 copy = be_new_Copy(cls, block, src);
957 sched_add_before(before, copy);
958 reg = arch_register_for_index(cls, r);
959 DB((dbg, LEVEL_2, "Copy %+F (from %+F, before %+F) -> %s\n",
960 copy, src, before, reg->name));
961 mark_as_copy_of(copy, src);
964 if (live_nodes != NULL) {
965 ir_nodeset_insert(live_nodes, copy);
968 /* old register has 1 user less, permutation is resolved */
969 assert(arch_register_get_index(arch_get_irn_register(src)) == old_r);
972 assert(n_used[old_r] > 0);
974 if (n_used[old_r] == 0) {
975 if (live_nodes != NULL) {
976 ir_nodeset_remove(live_nodes, src);
978 free_reg_of_value(src);
981 /* advance or jump back (if this copy enabled another copy) */
982 if (old_r < r && n_used[old_r] == 0) {
989 /* at this point we only have "cycles" left which we have to resolve with
991 * TODO: if we have free registers left, then we should really use copy
992 * instructions for any cycle longer than 2 registers...
993 * (this is probably architecture dependent, there might be archs where
994 * copies are preferable even for 2-cycles) */
996 /* create perms with the rest */
997 for (r = 0; r < n_regs; /* empty */) {
998 const arch_register_t *reg;
999 unsigned old_r = permutation[r];
1011 /* we shouldn't have copies from 1 value to multiple destinations left*/
1012 assert(n_used[old_r] == 1);
1014 /* exchange old_r and r2; after that old_r is a fixed point */
1015 r2 = permutation[old_r];
1017 in[0] = assignments[r2];
1018 in[1] = assignments[old_r];
1019 perm = be_new_Perm(cls, block, 2, in);
1020 sched_add_before(before, perm);
1021 DB((dbg, LEVEL_2, "Perm %+F (perm %+F,%+F, before %+F)\n",
1022 perm, in[0], in[1], before));
1024 proj0 = new_r_Proj(perm, get_irn_mode(in[0]), 0);
1025 mark_as_copy_of(proj0, in[0]);
1026 reg = arch_register_for_index(cls, old_r);
1027 use_reg(proj0, reg);
1029 proj1 = new_r_Proj(perm, get_irn_mode(in[1]), 1);
1030 mark_as_copy_of(proj1, in[1]);
1031 reg = arch_register_for_index(cls, r2);
1032 use_reg(proj1, reg);
1034 /* 1 value is now in the correct register */
1035 permutation[old_r] = old_r;
1036 /* the source of r changed to r2 */
1037 permutation[r] = r2;
1039 /* if we have reached a fixpoint update data structures */
1040 if (live_nodes != NULL) {
1041 ir_nodeset_remove(live_nodes, in[0]);
1042 ir_nodeset_remove(live_nodes, in[1]);
1043 ir_nodeset_remove(live_nodes, proj0);
1044 ir_nodeset_insert(live_nodes, proj1);
1048 #ifdef DEBUG_libfirm
1049 /* now we should only have fixpoints left */
1050 for (r = 0; r < n_regs; ++r) {
1051 assert(permutation[r] == r);
1057 * Free regs for values last used.
1059 * @param live_nodes set of live nodes, will be updated
1060 * @param node the node to consider
1062 static void free_last_uses(ir_nodeset_t *live_nodes, ir_node *node)
1064 allocation_info_t *info = get_allocation_info(node);
1065 const unsigned *last_uses = &info->last_uses;
1066 int arity = get_irn_arity(node);
1069 for (i = 0; i < arity; ++i) {
1072 /* check if one operand is the last use */
1073 if (!rbitset_is_set(last_uses, i))
1076 op = get_irn_n(node, i);
1077 free_reg_of_value(op);
1078 ir_nodeset_remove(live_nodes, op);
1083 * change inputs of a node to the current value (copies/perms)
1085 static void rewire_inputs(ir_node *node)
1088 int arity = get_irn_arity(node);
1090 for (i = 0; i < arity; ++i) {
1091 ir_node *op = get_irn_n(node, i);
1092 allocation_info_t *info = try_get_allocation_info(op);
1097 info = get_allocation_info(info->original_value);
1098 if (info->current_value != op) {
1099 set_irn_n(node, i, info->current_value);
1105 * Create a bitset of registers occupied with value living through an
1108 static void determine_live_through_regs(unsigned *bitset, ir_node *node)
1110 const allocation_info_t *info = get_allocation_info(node);
1115 /* mark all used registers as potentially live-through */
1116 for (r = 0; r < n_regs; ++r) {
1117 if (assignments[r] == NULL)
1119 if (!rbitset_is_set(normal_regs, r))
1122 rbitset_set(bitset, r);
1125 /* remove registers of value dying at the instruction */
1126 arity = get_irn_arity(node);
1127 for (i = 0; i < arity; ++i) {
1129 const arch_register_t *reg;
1131 if (!rbitset_is_set(&info->last_uses, i))
1134 op = get_irn_n(node, i);
1135 reg = arch_get_irn_register(op);
1136 rbitset_clear(bitset, arch_register_get_index(reg));
1141 * Enforce constraints at a node by live range splits.
1143 * @param live_nodes the set of live nodes, might be changed
1144 * @param node the current node
1146 static void enforce_constraints(ir_nodeset_t *live_nodes, ir_node *node,
1147 unsigned *forbidden_regs)
1149 int arity = get_irn_arity(node);
1151 hungarian_problem_t *bp;
1153 unsigned *assignment;
1155 /* construct a list of register occupied by live-through values */
1156 unsigned *live_through_regs = NULL;
1158 /* see if any use constraints are not met */
1160 for (i = 0; i < arity; ++i) {
1161 ir_node *op = get_irn_n(node, i);
1162 const arch_register_t *reg;
1163 const arch_register_req_t *req;
1164 const unsigned *limited;
1167 if (!arch_irn_consider_in_reg_alloc(cls, op))
1170 /* are there any limitations for the i'th operand? */
1171 req = arch_get_register_req(node, i);
1172 if (!(req->type & arch_register_req_type_limited))
1175 limited = req->limited;
1176 reg = arch_get_irn_register(op);
1177 r = arch_register_get_index(reg);
1178 if (!rbitset_is_set(limited, r)) {
1179 /* found an assignment outside the limited set */
1185 /* is any of the live-throughs using a constrained output register? */
1186 if (get_irn_mode(node) == mode_T) {
1187 const ir_edge_t *edge;
1189 foreach_out_edge(node, edge) {
1190 ir_node *proj = get_edge_src_irn(edge);
1191 const arch_register_req_t *req;
1193 if (!arch_irn_consider_in_reg_alloc(cls, proj))
1196 req = arch_get_register_req_out(proj);
1197 if (!(req->type & arch_register_req_type_limited))
1200 if (live_through_regs == NULL) {
1201 rbitset_alloca(live_through_regs, n_regs);
1202 determine_live_through_regs(live_through_regs, node);
1205 rbitset_or(forbidden_regs, req->limited, n_regs);
1206 if (rbitsets_have_common(req->limited, live_through_regs, n_regs)) {
1211 if (arch_irn_consider_in_reg_alloc(cls, node)) {
1212 const arch_register_req_t *req = arch_get_register_req_out(node);
1213 if (req->type & arch_register_req_type_limited) {
1214 rbitset_alloca(live_through_regs, n_regs);
1215 determine_live_through_regs(live_through_regs, node);
1216 if (rbitsets_have_common(req->limited, live_through_regs, n_regs)) {
1218 rbitset_or(forbidden_regs, req->limited, n_regs);
1227 /* create these arrays if we haven't yet */
1228 if (live_through_regs == NULL) {
1229 rbitset_alloca(live_through_regs, n_regs);
1232 /* at this point we have to construct a bipartite matching problem to see
1233 * which values should go to which registers
1234 * Note: We're building the matrix in "reverse" - source registers are
1235 * right, destinations left because this will produce the solution
1236 * in the format required for permute_values.
1238 bp = hungarian_new(n_regs, n_regs, HUNGARIAN_MATCH_PERFECT);
1240 /* add all combinations, then remove not allowed ones */
1241 for (l = 0; l < n_regs; ++l) {
1242 if (!rbitset_is_set(normal_regs, l)) {
1243 hungarian_add(bp, l, l, 1);
1247 for (r = 0; r < n_regs; ++r) {
1248 if (!rbitset_is_set(normal_regs, r))
1250 /* livethrough values may not use constrainted output registers */
1251 if (rbitset_is_set(live_through_regs, l)
1252 && rbitset_is_set(forbidden_regs, r))
1255 hungarian_add(bp, r, l, l == r ? 9 : 8);
1259 for (i = 0; i < arity; ++i) {
1260 ir_node *op = get_irn_n(node, i);
1261 const arch_register_t *reg;
1262 const arch_register_req_t *req;
1263 const unsigned *limited;
1264 unsigned current_reg;
1266 if (!arch_irn_consider_in_reg_alloc(cls, op))
1269 req = arch_get_register_req(node, i);
1270 if (!(req->type & arch_register_req_type_limited))
1273 limited = req->limited;
1274 reg = arch_get_irn_register(op);
1275 current_reg = arch_register_get_index(reg);
1276 for (r = 0; r < n_regs; ++r) {
1277 if (rbitset_is_set(limited, r))
1279 hungarian_remv(bp, r, current_reg);
1283 //hungarian_print_cost_matrix(bp, 1);
1284 hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1286 assignment = ALLOCAN(unsigned, n_regs);
1287 res = hungarian_solve(bp, (int*) assignment, NULL, 0);
1291 fprintf(stderr, "Swap result:");
1292 for (i = 0; i < (int) n_regs; ++i) {
1293 fprintf(stderr, " %d", assignment[i]);
1295 fprintf(stderr, "\n");
1300 permute_values(live_nodes, node, assignment);
1303 /** test wether a node @p n is a copy of the value of node @p of */
1304 static bool is_copy_of(ir_node *value, ir_node *test_value)
1306 allocation_info_t *test_info;
1307 allocation_info_t *info;
1309 if (value == test_value)
1312 info = get_allocation_info(value);
1313 test_info = get_allocation_info(test_value);
1314 return test_info->original_value == info->original_value;
1318 * find a value in the end-assignment of a basic block
1319 * @returns the index into the assignment array if found
1322 static int find_value_in_block_info(block_info_t *info, ir_node *value)
1325 ir_node **assignments = info->assignments;
1326 for (r = 0; r < n_regs; ++r) {
1327 ir_node *a_value = assignments[r];
1329 if (a_value == NULL)
1331 if (is_copy_of(a_value, value))
1339 * Create the necessary permutations at the end of a basic block to fullfill
1340 * the register assignment for phi-nodes in the next block
1342 static void add_phi_permutations(ir_node *block, int p)
1345 unsigned *permutation;
1346 ir_node **old_assignments;
1347 bool need_permutation;
1349 ir_node *pred = get_Block_cfgpred_block(block, p);
1351 block_info_t *pred_info = get_block_info(pred);
1353 /* predecessor not processed yet? nothing to do */
1354 if (!pred_info->processed)
1357 permutation = ALLOCAN(unsigned, n_regs);
1358 for (r = 0; r < n_regs; ++r) {
1362 /* check phi nodes */
1363 need_permutation = false;
1364 node = sched_first(block);
1365 for ( ; is_Phi(node); node = sched_next(node)) {
1366 const arch_register_t *reg;
1371 if (!arch_irn_consider_in_reg_alloc(cls, node))
1374 op = get_Phi_pred(node, p);
1375 if (!arch_irn_consider_in_reg_alloc(cls, op))
1378 a = find_value_in_block_info(pred_info, op);
1381 reg = arch_get_irn_register(node);
1382 regn = arch_register_get_index(reg);
1384 permutation[regn] = a;
1385 need_permutation = true;
1389 if (need_permutation) {
1390 /* permute values at end of predecessor */
1391 old_assignments = assignments;
1392 assignments = pred_info->assignments;
1393 permute_values(NULL, be_get_end_of_block_insertion_point(pred),
1395 assignments = old_assignments;
1398 /* change phi nodes to use the copied values */
1399 node = sched_first(block);
1400 for ( ; is_Phi(node); node = sched_next(node)) {
1404 if (!arch_irn_consider_in_reg_alloc(cls, node))
1407 op = get_Phi_pred(node, p);
1408 /* no need to do anything for Unknown inputs */
1409 if (!arch_irn_consider_in_reg_alloc(cls, op))
1412 /* we have permuted all values into the correct registers so we can
1413 simply query which value occupies the phis register in the
1415 a = arch_register_get_index(arch_get_irn_register(node));
1416 op = pred_info->assignments[a];
1417 set_Phi_pred(node, p, op);
1422 * Set preferences for a phis register based on the registers used on the
1425 static void adapt_phi_prefs(ir_node *phi)
1428 int arity = get_irn_arity(phi);
1429 ir_node *block = get_nodes_block(phi);
1430 allocation_info_t *info = get_allocation_info(phi);
1432 for (i = 0; i < arity; ++i) {
1433 ir_node *op = get_irn_n(phi, i);
1434 const arch_register_t *reg = arch_get_irn_register(op);
1435 ir_node *pred_block;
1436 block_info_t *pred_block_info;
1442 /* we only give the bonus if the predecessor already has registers
1443 * assigned, otherwise we only see a dummy value
1444 * and any conclusions about its register are useless */
1445 pred_block = get_Block_cfgpred_block(block, i);
1446 pred_block_info = get_block_info(pred_block);
1447 if (!pred_block_info->processed)
1450 /* give bonus for already assigned register */
1451 weight = (float)get_block_execfreq(execfreqs, pred_block);
1452 r = arch_register_get_index(reg);
1453 info->prefs[r] += weight * AFF_PHI;
1458 * After a phi has been assigned a register propagate preference inputs
1459 * to the phi inputs.
1461 static void propagate_phi_register(ir_node *phi, unsigned assigned_r)
1464 ir_node *block = get_nodes_block(phi);
1465 int arity = get_irn_arity(phi);
1467 for (i = 0; i < arity; ++i) {
1468 ir_node *op = get_Phi_pred(phi, i);
1469 allocation_info_t *info = get_allocation_info(op);
1470 ir_node *pred_block = get_Block_cfgpred_block(block, i);
1473 = (float)get_block_execfreq(execfreqs, pred_block) * AFF_PHI;
1475 if (info->prefs[assigned_r] >= weight)
1478 /* promote the prefered register */
1479 for (r = 0; r < n_regs; ++r) {
1480 if (info->prefs[r] > -weight) {
1481 info->prefs[r] = -weight;
1484 info->prefs[assigned_r] = weight;
1487 propagate_phi_register(op, assigned_r);
1491 static void assign_phi_registers(ir_node *block)
1498 hungarian_problem_t *bp;
1500 /* count phi nodes */
1501 sched_foreach(block, node) {
1504 if (!arch_irn_consider_in_reg_alloc(cls, node))
1512 /* build a bipartite matching problem for all phi nodes */
1513 bp = hungarian_new(n_phis, n_regs, HUNGARIAN_MATCH_PERFECT);
1515 sched_foreach(block, node) {
1518 allocation_info_t *info;
1521 if (!arch_irn_consider_in_reg_alloc(cls, node))
1524 /* give boni for predecessor colorings */
1525 adapt_phi_prefs(node);
1526 /* add stuff to bipartite problem */
1527 info = get_allocation_info(node);
1528 DB((dbg, LEVEL_3, "Prefs for %+F: ", node));
1529 for (r = 0; r < n_regs; ++r) {
1532 if (!rbitset_is_set(normal_regs, r))
1535 costs = info->prefs[r];
1536 costs = costs < 0 ? -logf(-costs+1) : logf(costs+1);
1539 hungarian_add(bp, n, r, (int)costs);
1540 DB((dbg, LEVEL_3, " %s(%f)", arch_register_for_index(cls, r)->name,
1543 DB((dbg, LEVEL_3, "\n"));
1547 //hungarian_print_cost_matrix(bp, 7);
1548 hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1550 assignment = ALLOCAN(int, n_regs);
1551 res = hungarian_solve(bp, assignment, NULL, 0);
1556 sched_foreach(block, node) {
1558 const arch_register_t *reg;
1562 if (!arch_irn_consider_in_reg_alloc(cls, node))
1565 r = assignment[n++];
1566 assert(rbitset_is_set(normal_regs, r));
1567 reg = arch_register_for_index(cls, r);
1568 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
1571 /* adapt preferences for phi inputs */
1572 if (propagate_phi_registers)
1573 propagate_phi_register(node, r);
1578 * Walker: assign registers to all nodes of a block that
1579 * need registers from the currently considered register class.
1581 static void allocate_coalesce_block(ir_node *block, void *data)
1584 ir_nodeset_t live_nodes;
1587 block_info_t *block_info;
1588 block_info_t **pred_block_infos;
1590 unsigned *forbidden_regs; /**< collects registers which must
1591 not be used for optimistic splits */
1594 DB((dbg, LEVEL_2, "* Block %+F\n", block));
1596 /* clear assignments */
1597 block_info = get_block_info(block);
1598 assignments = block_info->assignments;
1600 ir_nodeset_init(&live_nodes);
1602 /* gather regalloc infos of predecessor blocks */
1603 n_preds = get_Block_n_cfgpreds(block);
1604 pred_block_infos = ALLOCAN(block_info_t*, n_preds);
1605 for (i = 0; i < n_preds; ++i) {
1606 ir_node *pred = get_Block_cfgpred_block(block, i);
1607 block_info_t *pred_info = get_block_info(pred);
1608 pred_block_infos[i] = pred_info;
1611 phi_ins = ALLOCAN(ir_node*, n_preds);
1613 /* collect live-in nodes and preassigned values */
1614 be_lv_foreach(lv, block, be_lv_state_in, i) {
1615 const arch_register_t *reg;
1617 bool need_phi = false;
1619 node = be_lv_get_irn(lv, block, i);
1620 if (!arch_irn_consider_in_reg_alloc(cls, node))
1623 /* check all predecessors for this value, if it is not everywhere the
1624 same or unknown then we have to construct a phi
1625 (we collect the potential phi inputs here) */
1626 for (p = 0; p < n_preds; ++p) {
1627 block_info_t *pred_info = pred_block_infos[p];
1629 if (!pred_info->processed) {
1630 /* use node for now, it will get fixed later */
1634 int a = find_value_in_block_info(pred_info, node);
1636 /* must live out of predecessor */
1638 phi_ins[p] = pred_info->assignments[a];
1639 /* different value from last time? then we need a phi */
1640 if (p > 0 && phi_ins[p-1] != phi_ins[p]) {
1647 ir_mode *mode = get_irn_mode(node);
1648 const arch_register_req_t *req = get_default_req_current_cls();
1651 phi = new_r_Phi(block, n_preds, phi_ins, mode);
1652 be_set_phi_reg_req(phi, req);
1654 DB((dbg, LEVEL_3, "Create Phi %+F (for %+F) -", phi, node));
1655 #ifdef DEBUG_libfirm
1658 for (i = 0; i < n_preds; ++i) {
1659 DB((dbg, LEVEL_3, " %+F", phi_ins[i]));
1661 DB((dbg, LEVEL_3, "\n"));
1664 mark_as_copy_of(phi, node);
1665 sched_add_after(block, phi);
1669 allocation_info_t *info = get_allocation_info(node);
1670 info->current_value = phi_ins[0];
1672 /* Grab 1 of the inputs we constructed (might not be the same as
1673 * "node" as we could see the same copy of the value in all
1678 /* if the node already has a register assigned use it */
1679 reg = arch_get_irn_register(node);
1684 /* remember that this node is live at the beginning of the block */
1685 ir_nodeset_insert(&live_nodes, node);
1688 rbitset_alloca(forbidden_regs, n_regs);
1690 /* handle phis... */
1691 assign_phi_registers(block);
1693 /* all live-ins must have a register */
1694 #ifdef DEBUG_libfirm
1696 ir_nodeset_iterator_t iter;
1697 foreach_ir_nodeset(&live_nodes, node, iter) {
1698 const arch_register_t *reg = arch_get_irn_register(node);
1699 assert(reg != NULL);
1704 /* assign instructions in the block */
1705 sched_foreach(block, node) {
1709 /* phis are already assigned */
1713 rewire_inputs(node);
1715 /* enforce use constraints */
1716 rbitset_clear_all(forbidden_regs, n_regs);
1717 enforce_constraints(&live_nodes, node, forbidden_regs);
1719 rewire_inputs(node);
1721 /* we may not use registers used for inputs for optimistic splits */
1722 arity = get_irn_arity(node);
1723 for (i = 0; i < arity; ++i) {
1724 ir_node *op = get_irn_n(node, i);
1725 const arch_register_t *reg;
1726 if (!arch_irn_consider_in_reg_alloc(cls, op))
1729 reg = arch_get_irn_register(op);
1730 rbitset_set(forbidden_regs, arch_register_get_index(reg));
1733 /* free registers of values last used at this instruction */
1734 free_last_uses(&live_nodes, node);
1736 /* assign output registers */
1737 /* TODO: 2 phases: first: pre-assigned ones, 2nd real regs */
1738 if (get_irn_mode(node) == mode_T) {
1739 const ir_edge_t *edge;
1740 foreach_out_edge(node, edge) {
1741 ir_node *proj = get_edge_src_irn(edge);
1742 if (!arch_irn_consider_in_reg_alloc(cls, proj))
1744 assign_reg(block, proj, forbidden_regs);
1746 } else if (arch_irn_consider_in_reg_alloc(cls, node)) {
1747 assign_reg(block, node, forbidden_regs);
1751 ir_nodeset_destroy(&live_nodes);
1754 block_info->processed = true;
1756 /* permute values at end of predecessor blocks in case of phi-nodes */
1759 for (p = 0; p < n_preds; ++p) {
1760 add_phi_permutations(block, p);
1764 /* if we have exactly 1 successor then we might be able to produce phi
1766 if (get_irn_n_edges_kind(block, EDGE_KIND_BLOCK) == 1) {
1767 const ir_edge_t *edge
1768 = get_irn_out_edge_first_kind(block, EDGE_KIND_BLOCK);
1769 ir_node *succ = get_edge_src_irn(edge);
1770 int p = get_edge_src_pos(edge);
1771 block_info_t *succ_info = get_block_info(succ);
1773 if (succ_info->processed) {
1774 add_phi_permutations(succ, p);
1779 typedef struct block_costs_t block_costs_t;
1780 struct block_costs_t {
1781 float costs; /**< costs of the block */
1782 int dfs_num; /**< depth first search number (to detect backedges) */
1785 static int cmp_block_costs(const void *d1, const void *d2)
1787 const ir_node * const *block1 = d1;
1788 const ir_node * const *block2 = d2;
1789 const block_costs_t *info1 = get_irn_link(*block1);
1790 const block_costs_t *info2 = get_irn_link(*block2);
1791 return QSORT_CMP(info2->costs, info1->costs);
1794 static void determine_block_order(void)
1797 ir_node **blocklist = be_get_cfgpostorder(irg);
1798 int n_blocks = ARR_LEN(blocklist);
1800 pdeq *worklist = new_pdeq();
1801 ir_node **order = XMALLOCN(ir_node*, n_blocks);
1804 /* clear block links... */
1805 for (i = 0; i < n_blocks; ++i) {
1806 ir_node *block = blocklist[i];
1807 set_irn_link(block, NULL);
1810 /* walk blocks in reverse postorder, the costs for each block are the
1811 * sum of the costs of its predecessors (excluding the costs on backedges
1812 * which we can't determine) */
1813 for (i = n_blocks-1; i >= 0; --i) {
1814 block_costs_t *cost_info;
1815 ir_node *block = blocklist[i];
1817 float execfreq = (float)get_block_execfreq(execfreqs, block);
1818 float costs = execfreq;
1819 int n_cfgpreds = get_Block_n_cfgpreds(block);
1821 for (p = 0; p < n_cfgpreds; ++p) {
1822 ir_node *pred_block = get_Block_cfgpred_block(block, p);
1823 block_costs_t *pred_costs = get_irn_link(pred_block);
1824 /* we don't have any info for backedges */
1825 if (pred_costs == NULL)
1827 costs += pred_costs->costs;
1830 cost_info = OALLOCZ(&obst, block_costs_t);
1831 cost_info->costs = costs;
1832 cost_info->dfs_num = dfs_num++;
1833 set_irn_link(block, cost_info);
1836 /* sort array by block costs */
1837 qsort(blocklist, n_blocks, sizeof(blocklist[0]), cmp_block_costs);
1839 ir_reserve_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1840 inc_irg_block_visited(irg);
1842 for (i = 0; i < n_blocks; ++i) {
1843 ir_node *block = blocklist[i];
1844 if (Block_block_visited(block))
1847 /* continually add predecessors with highest costs to worklist
1848 * (without using backedges) */
1850 block_costs_t *info = get_irn_link(block);
1851 ir_node *best_pred = NULL;
1852 float best_costs = -1;
1853 int n_cfgpred = get_Block_n_cfgpreds(block);
1856 pdeq_putr(worklist, block);
1857 mark_Block_block_visited(block);
1858 for (i = 0; i < n_cfgpred; ++i) {
1859 ir_node *pred_block = get_Block_cfgpred_block(block, i);
1860 block_costs_t *pred_info = get_irn_link(pred_block);
1862 /* ignore backedges */
1863 if (pred_info->dfs_num > info->dfs_num)
1866 if (info->costs > best_costs) {
1867 best_costs = info->costs;
1868 best_pred = pred_block;
1872 } while (block != NULL && !Block_block_visited(block));
1874 /* now put all nodes in the worklist in our final order */
1875 while (!pdeq_empty(worklist)) {
1876 ir_node *pblock = pdeq_getr(worklist);
1877 assert(order_p < n_blocks);
1878 order[order_p++] = pblock;
1881 assert(order_p == n_blocks);
1884 ir_free_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1886 DEL_ARR_F(blocklist);
1888 obstack_free(&obst, NULL);
1889 obstack_init(&obst);
1891 block_order = order;
1892 n_block_order = n_blocks;
1896 * Run the register allocator for the current register class.
1898 static void be_pref_alloc_cls(void)
1902 lv = be_assure_liveness(birg);
1903 be_liveness_assure_sets(lv);
1905 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
1907 DB((dbg, LEVEL_2, "=== Allocating registers of %s ===\n", cls->name));
1909 be_clear_links(irg);
1911 irg_block_walk_graph(irg, NULL, analyze_block, NULL);
1912 if (create_congruence_classes)
1913 combine_congruence_classes();
1915 for (i = 0; i < n_block_order; ++i) {
1916 ir_node *block = block_order[i];
1917 allocate_coalesce_block(block, NULL);
1920 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
1923 static void dump(int mask, ir_graph *irg, const char *suffix,
1924 void (*dumper)(ir_graph *, const char *))
1926 if (birg->main_env->options->dump_flags & mask)
1927 be_dump(irg, suffix, dumper);
1931 * Run the spiller on the current graph.
1933 static void spill(void)
1935 /* make sure all nodes show their real register pressure */
1936 be_timer_push(T_RA_CONSTR);
1937 be_pre_spill_prepare_constr(birg, cls);
1938 be_timer_pop(T_RA_CONSTR);
1940 dump(DUMP_RA, irg, "-spillprepare", dump_ir_block_graph_sched);
1943 be_timer_push(T_RA_SPILL);
1944 be_do_spill(birg, cls);
1945 be_timer_pop(T_RA_SPILL);
1947 be_timer_push(T_RA_SPILL_APPLY);
1948 check_for_memory_operands(irg);
1949 be_timer_pop(T_RA_SPILL_APPLY);
1951 dump(DUMP_RA, irg, "-spill", dump_ir_block_graph_sched);
1955 * The pref register allocator for a whole procedure.
1957 static void be_pref_alloc(be_irg_t *new_birg)
1959 const arch_env_t *arch_env = new_birg->main_env->arch_env;
1960 int n_cls = arch_env_get_n_reg_class(arch_env);
1963 obstack_init(&obst);
1966 irg = be_get_birg_irg(birg);
1967 execfreqs = birg->exec_freq;
1969 /* determine a good coloring order */
1970 determine_block_order();
1972 for (c = 0; c < n_cls; ++c) {
1973 cls = arch_env_get_reg_class(arch_env, c);
1974 default_cls_req = NULL;
1975 if (arch_register_class_flags(cls) & arch_register_class_flag_manual_ra)
1978 stat_ev_ctx_push_str("regcls", cls->name);
1980 n_regs = arch_register_class_n_regs(cls);
1981 normal_regs = rbitset_malloc(n_regs);
1982 be_abi_set_non_ignore_regs(birg->abi, cls, normal_regs);
1986 /* verify schedule and register pressure */
1987 be_timer_push(T_VERIFY);
1988 if (birg->main_env->options->vrfy_option == BE_VRFY_WARN) {
1989 be_verify_schedule(birg);
1990 be_verify_register_pressure(birg, cls, irg);
1991 } else if (birg->main_env->options->vrfy_option == BE_VRFY_ASSERT) {
1992 assert(be_verify_schedule(birg) && "Schedule verification failed");
1993 assert(be_verify_register_pressure(birg, cls, irg)
1994 && "Register pressure verification failed");
1996 be_timer_pop(T_VERIFY);
1998 be_timer_push(T_RA_COLOR);
1999 be_pref_alloc_cls();
2000 be_timer_pop(T_RA_COLOR);
2002 /* we most probably constructed new Phis so liveness info is invalid
2004 /* TODO: test liveness_introduce */
2005 be_liveness_invalidate(lv);
2008 stat_ev_ctx_pop("regcls");
2011 be_timer_push(T_RA_SPILL_APPLY);
2012 be_abi_fix_stack_nodes(birg->abi);
2013 be_timer_pop(T_RA_SPILL_APPLY);
2015 be_timer_push(T_VERIFY);
2016 if (birg->main_env->options->vrfy_option == BE_VRFY_WARN) {
2017 be_verify_register_allocation(birg);
2018 } else if (birg->main_env->options->vrfy_option == BE_VRFY_ASSERT) {
2019 assert(be_verify_register_allocation(birg)
2020 && "Register allocation invalid");
2022 be_timer_pop(T_VERIFY);
2024 obstack_free(&obst, NULL);
2027 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_pref_alloc);
2028 void be_init_pref_alloc(void)
2030 static be_ra_t be_ra_pref = {
2033 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
2034 lc_opt_entry_t *prefalloc_group = lc_opt_get_grp(be_grp, "prefalloc");
2035 lc_opt_add_table(prefalloc_group, options);
2037 be_register_allocator("pref", &be_ra_pref);
2038 FIRM_DBG_REGISTER(dbg, "firm.be.prefalloc");