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"
57 #include "raw_bitset.h"
58 #include "unionfind.h"
60 #include "hungarian.h"
63 #include "bechordal_t.h"
72 #include "bespillutil.h"
76 #define USE_FACTOR 1.0f
77 #define DEF_FACTOR 1.0f
78 #define NEIGHBOR_FACTOR 0.2f
79 #define AFF_SHOULD_BE_SAME 0.5f
81 #define SPLIT_DELTA 1.0f
82 #define MAX_OPTIMISTIC_SPLIT_RECURSION 0
84 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
86 static struct obstack obst;
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 useful 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 /* the common reason to hit this panic is when 1 of your nodes is not
874 * register pressure faithful */
875 panic("No register left for %+F\n", node);
878 reg = arch_register_for_index(cls, r);
879 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
884 * Add an permutation in front of a node and change the assignments
885 * due to this permutation.
887 * To understand this imagine a permutation like this:
897 * First we count how many destinations a single value has. At the same time
898 * we can be sure that each destination register has at most 1 source register
899 * (it can have 0 which means we don't care what value is in it).
900 * We ignore all fullfilled permuations (like 7->7)
901 * In a first pass we create as much copy instructions as possible as they
902 * are generally cheaper than exchanges. We do this by counting into how many
903 * destinations a register has to be copied (in the example it's 2 for register
904 * 3, or 1 for the registers 1,2,4 and 7).
905 * We can then create a copy into every destination register when the usecount
906 * of that register is 0 (= noone else needs the value in the register).
908 * After this step we should have cycles left. We implement a cyclic permutation
909 * of n registers with n-1 transpositions.
911 * @param live_nodes the set of live nodes, updated due to live range split
912 * @param before the node before we add the permutation
913 * @param permutation the permutation array indices are the destination
914 * registers, the values in the array are the source
917 static void permute_values(ir_nodeset_t *live_nodes, ir_node *before,
918 unsigned *permutation)
920 unsigned *n_used = ALLOCANZ(unsigned, n_regs);
924 /* determine how often each source register needs to be read */
925 for (r = 0; r < n_regs; ++r) {
926 unsigned old_reg = permutation[r];
929 value = assignments[old_reg];
931 /* nothing to do here, reg is not live. Mark it as fixpoint
932 * so we ignore it in the next steps */
940 block = get_nodes_block(before);
942 /* step1: create copies where immediately possible */
943 for (r = 0; r < n_regs; /* empty */) {
946 const arch_register_t *reg;
947 unsigned old_r = permutation[r];
949 /* - no need to do anything for fixed points.
950 - we can't copy if the value in the dest reg is still needed */
951 if (old_r == r || n_used[r] > 0) {
957 src = assignments[old_r];
958 copy = be_new_Copy(cls, block, src);
959 sched_add_before(before, copy);
960 reg = arch_register_for_index(cls, r);
961 DB((dbg, LEVEL_2, "Copy %+F (from %+F, before %+F) -> %s\n",
962 copy, src, before, reg->name));
963 mark_as_copy_of(copy, src);
966 if (live_nodes != NULL) {
967 ir_nodeset_insert(live_nodes, copy);
970 /* old register has 1 user less, permutation is resolved */
971 assert(arch_register_get_index(arch_get_irn_register(src)) == old_r);
974 assert(n_used[old_r] > 0);
976 if (n_used[old_r] == 0) {
977 if (live_nodes != NULL) {
978 ir_nodeset_remove(live_nodes, src);
980 free_reg_of_value(src);
983 /* advance or jump back (if this copy enabled another copy) */
984 if (old_r < r && n_used[old_r] == 0) {
991 /* at this point we only have "cycles" left which we have to resolve with
993 * TODO: if we have free registers left, then we should really use copy
994 * instructions for any cycle longer than 2 registers...
995 * (this is probably architecture dependent, there might be archs where
996 * copies are preferable even for 2-cycles) */
998 /* create perms with the rest */
999 for (r = 0; r < n_regs; /* empty */) {
1000 const arch_register_t *reg;
1001 unsigned old_r = permutation[r];
1013 /* we shouldn't have copies from 1 value to multiple destinations left*/
1014 assert(n_used[old_r] == 1);
1016 /* exchange old_r and r2; after that old_r is a fixed point */
1017 r2 = permutation[old_r];
1019 in[0] = assignments[r2];
1020 in[1] = assignments[old_r];
1021 perm = be_new_Perm(cls, block, 2, in);
1022 sched_add_before(before, perm);
1023 DB((dbg, LEVEL_2, "Perm %+F (perm %+F,%+F, before %+F)\n",
1024 perm, in[0], in[1], before));
1026 proj0 = new_r_Proj(perm, get_irn_mode(in[0]), 0);
1027 mark_as_copy_of(proj0, in[0]);
1028 reg = arch_register_for_index(cls, old_r);
1029 use_reg(proj0, reg);
1031 proj1 = new_r_Proj(perm, get_irn_mode(in[1]), 1);
1032 mark_as_copy_of(proj1, in[1]);
1033 reg = arch_register_for_index(cls, r2);
1034 use_reg(proj1, reg);
1036 /* 1 value is now in the correct register */
1037 permutation[old_r] = old_r;
1038 /* the source of r changed to r2 */
1039 permutation[r] = r2;
1041 /* if we have reached a fixpoint update data structures */
1042 if (live_nodes != NULL) {
1043 ir_nodeset_remove(live_nodes, in[0]);
1044 ir_nodeset_remove(live_nodes, in[1]);
1045 ir_nodeset_remove(live_nodes, proj0);
1046 ir_nodeset_insert(live_nodes, proj1);
1050 #ifdef DEBUG_libfirm
1051 /* now we should only have fixpoints left */
1052 for (r = 0; r < n_regs; ++r) {
1053 assert(permutation[r] == r);
1059 * Free regs for values last used.
1061 * @param live_nodes set of live nodes, will be updated
1062 * @param node the node to consider
1064 static void free_last_uses(ir_nodeset_t *live_nodes, ir_node *node)
1066 allocation_info_t *info = get_allocation_info(node);
1067 const unsigned *last_uses = &info->last_uses;
1068 int arity = get_irn_arity(node);
1071 for (i = 0; i < arity; ++i) {
1074 /* check if one operand is the last use */
1075 if (!rbitset_is_set(last_uses, i))
1078 op = get_irn_n(node, i);
1079 free_reg_of_value(op);
1080 ir_nodeset_remove(live_nodes, op);
1085 * change inputs of a node to the current value (copies/perms)
1087 static void rewire_inputs(ir_node *node)
1090 int arity = get_irn_arity(node);
1092 for (i = 0; i < arity; ++i) {
1093 ir_node *op = get_irn_n(node, i);
1094 allocation_info_t *info = try_get_allocation_info(op);
1099 info = get_allocation_info(info->original_value);
1100 if (info->current_value != op) {
1101 set_irn_n(node, i, info->current_value);
1107 * Create a bitset of registers occupied with value living through an
1110 static void determine_live_through_regs(unsigned *bitset, ir_node *node)
1112 const allocation_info_t *info = get_allocation_info(node);
1117 /* mark all used registers as potentially live-through */
1118 for (r = 0; r < n_regs; ++r) {
1119 if (assignments[r] == NULL)
1121 if (!rbitset_is_set(normal_regs, r))
1124 rbitset_set(bitset, r);
1127 /* remove registers of value dying at the instruction */
1128 arity = get_irn_arity(node);
1129 for (i = 0; i < arity; ++i) {
1131 const arch_register_t *reg;
1133 if (!rbitset_is_set(&info->last_uses, i))
1136 op = get_irn_n(node, i);
1137 reg = arch_get_irn_register(op);
1138 rbitset_clear(bitset, arch_register_get_index(reg));
1143 * Enforce constraints at a node by live range splits.
1145 * @param live_nodes the set of live nodes, might be changed
1146 * @param node the current node
1148 static void enforce_constraints(ir_nodeset_t *live_nodes, ir_node *node,
1149 unsigned *forbidden_regs)
1151 int arity = get_irn_arity(node);
1153 hungarian_problem_t *bp;
1155 unsigned *assignment;
1157 /* construct a list of register occupied by live-through values */
1158 unsigned *live_through_regs = NULL;
1160 /* see if any use constraints are not met */
1162 for (i = 0; i < arity; ++i) {
1163 ir_node *op = get_irn_n(node, i);
1164 const arch_register_t *reg;
1165 const arch_register_req_t *req;
1166 const unsigned *limited;
1169 if (!arch_irn_consider_in_reg_alloc(cls, op))
1172 /* are there any limitations for the i'th operand? */
1173 req = arch_get_register_req(node, i);
1174 if (!(req->type & arch_register_req_type_limited))
1177 limited = req->limited;
1178 reg = arch_get_irn_register(op);
1179 r = arch_register_get_index(reg);
1180 if (!rbitset_is_set(limited, r)) {
1181 /* found an assignment outside the limited set */
1187 /* is any of the live-throughs using a constrained output register? */
1188 if (get_irn_mode(node) == mode_T) {
1189 const ir_edge_t *edge;
1191 foreach_out_edge(node, edge) {
1192 ir_node *proj = get_edge_src_irn(edge);
1193 const arch_register_req_t *req;
1195 if (!arch_irn_consider_in_reg_alloc(cls, proj))
1198 req = arch_get_register_req_out(proj);
1199 if (!(req->type & arch_register_req_type_limited))
1202 if (live_through_regs == NULL) {
1203 rbitset_alloca(live_through_regs, n_regs);
1204 determine_live_through_regs(live_through_regs, node);
1207 rbitset_or(forbidden_regs, req->limited, n_regs);
1208 if (rbitsets_have_common(req->limited, live_through_regs, n_regs)) {
1213 if (arch_irn_consider_in_reg_alloc(cls, node)) {
1214 const arch_register_req_t *req = arch_get_register_req_out(node);
1215 if (req->type & arch_register_req_type_limited) {
1216 rbitset_alloca(live_through_regs, n_regs);
1217 determine_live_through_regs(live_through_regs, node);
1218 if (rbitsets_have_common(req->limited, live_through_regs, n_regs)) {
1220 rbitset_or(forbidden_regs, req->limited, n_regs);
1229 /* create these arrays if we haven't yet */
1230 if (live_through_regs == NULL) {
1231 rbitset_alloca(live_through_regs, n_regs);
1234 /* at this point we have to construct a bipartite matching problem to see
1235 * which values should go to which registers
1236 * Note: We're building the matrix in "reverse" - source registers are
1237 * right, destinations left because this will produce the solution
1238 * in the format required for permute_values.
1240 bp = hungarian_new(n_regs, n_regs, HUNGARIAN_MATCH_PERFECT);
1242 /* add all combinations, then remove not allowed ones */
1243 for (l = 0; l < n_regs; ++l) {
1244 if (!rbitset_is_set(normal_regs, l)) {
1245 hungarian_add(bp, l, l, 1);
1249 for (r = 0; r < n_regs; ++r) {
1250 if (!rbitset_is_set(normal_regs, r))
1252 /* livethrough values may not use constrainted output registers */
1253 if (rbitset_is_set(live_through_regs, l)
1254 && rbitset_is_set(forbidden_regs, r))
1257 hungarian_add(bp, r, l, l == r ? 9 : 8);
1261 for (i = 0; i < arity; ++i) {
1262 ir_node *op = get_irn_n(node, i);
1263 const arch_register_t *reg;
1264 const arch_register_req_t *req;
1265 const unsigned *limited;
1266 unsigned current_reg;
1268 if (!arch_irn_consider_in_reg_alloc(cls, op))
1271 req = arch_get_register_req(node, i);
1272 if (!(req->type & arch_register_req_type_limited))
1275 limited = req->limited;
1276 reg = arch_get_irn_register(op);
1277 current_reg = arch_register_get_index(reg);
1278 for (r = 0; r < n_regs; ++r) {
1279 if (rbitset_is_set(limited, r))
1281 hungarian_remove(bp, r, current_reg);
1285 //hungarian_print_cost_matrix(bp, 1);
1286 hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1288 assignment = ALLOCAN(unsigned, n_regs);
1289 res = hungarian_solve(bp, assignment, NULL, 0);
1293 fprintf(stderr, "Swap result:");
1294 for (i = 0; i < (int) n_regs; ++i) {
1295 fprintf(stderr, " %d", assignment[i]);
1297 fprintf(stderr, "\n");
1302 permute_values(live_nodes, node, assignment);
1305 /** test wether a node @p n is a copy of the value of node @p of */
1306 static bool is_copy_of(ir_node *value, ir_node *test_value)
1308 allocation_info_t *test_info;
1309 allocation_info_t *info;
1311 if (value == test_value)
1314 info = get_allocation_info(value);
1315 test_info = get_allocation_info(test_value);
1316 return test_info->original_value == info->original_value;
1320 * find a value in the end-assignment of a basic block
1321 * @returns the index into the assignment array if found
1324 static int find_value_in_block_info(block_info_t *info, ir_node *value)
1327 ir_node **assignments = info->assignments;
1328 for (r = 0; r < n_regs; ++r) {
1329 ir_node *a_value = assignments[r];
1331 if (a_value == NULL)
1333 if (is_copy_of(a_value, value))
1341 * Create the necessary permutations at the end of a basic block to fullfill
1342 * the register assignment for phi-nodes in the next block
1344 static void add_phi_permutations(ir_node *block, int p)
1347 unsigned *permutation;
1348 ir_node **old_assignments;
1349 bool need_permutation;
1351 ir_node *pred = get_Block_cfgpred_block(block, p);
1353 block_info_t *pred_info = get_block_info(pred);
1355 /* predecessor not processed yet? nothing to do */
1356 if (!pred_info->processed)
1359 permutation = ALLOCAN(unsigned, n_regs);
1360 for (r = 0; r < n_regs; ++r) {
1364 /* check phi nodes */
1365 need_permutation = false;
1366 node = sched_first(block);
1367 for ( ; is_Phi(node); node = sched_next(node)) {
1368 const arch_register_t *reg;
1373 if (!arch_irn_consider_in_reg_alloc(cls, node))
1376 op = get_Phi_pred(node, p);
1377 if (!arch_irn_consider_in_reg_alloc(cls, op))
1380 a = find_value_in_block_info(pred_info, op);
1383 reg = arch_get_irn_register(node);
1384 regn = arch_register_get_index(reg);
1386 permutation[regn] = a;
1387 need_permutation = true;
1391 if (need_permutation) {
1392 /* permute values at end of predecessor */
1393 old_assignments = assignments;
1394 assignments = pred_info->assignments;
1395 permute_values(NULL, be_get_end_of_block_insertion_point(pred),
1397 assignments = old_assignments;
1400 /* change phi nodes to use the copied values */
1401 node = sched_first(block);
1402 for ( ; is_Phi(node); node = sched_next(node)) {
1406 if (!arch_irn_consider_in_reg_alloc(cls, node))
1409 op = get_Phi_pred(node, p);
1410 /* no need to do anything for Unknown inputs */
1411 if (!arch_irn_consider_in_reg_alloc(cls, op))
1414 /* we have permuted all values into the correct registers so we can
1415 simply query which value occupies the phis register in the
1417 a = arch_register_get_index(arch_get_irn_register(node));
1418 op = pred_info->assignments[a];
1419 set_Phi_pred(node, p, op);
1424 * Set preferences for a phis register based on the registers used on the
1427 static void adapt_phi_prefs(ir_node *phi)
1430 int arity = get_irn_arity(phi);
1431 ir_node *block = get_nodes_block(phi);
1432 allocation_info_t *info = get_allocation_info(phi);
1434 for (i = 0; i < arity; ++i) {
1435 ir_node *op = get_irn_n(phi, i);
1436 const arch_register_t *reg = arch_get_irn_register(op);
1437 ir_node *pred_block;
1438 block_info_t *pred_block_info;
1444 /* we only give the bonus if the predecessor already has registers
1445 * assigned, otherwise we only see a dummy value
1446 * and any conclusions about its register are useless */
1447 pred_block = get_Block_cfgpred_block(block, i);
1448 pred_block_info = get_block_info(pred_block);
1449 if (!pred_block_info->processed)
1452 /* give bonus for already assigned register */
1453 weight = (float)get_block_execfreq(execfreqs, pred_block);
1454 r = arch_register_get_index(reg);
1455 info->prefs[r] += weight * AFF_PHI;
1460 * After a phi has been assigned a register propagate preference inputs
1461 * to the phi inputs.
1463 static void propagate_phi_register(ir_node *phi, unsigned assigned_r)
1466 ir_node *block = get_nodes_block(phi);
1467 int arity = get_irn_arity(phi);
1469 for (i = 0; i < arity; ++i) {
1470 ir_node *op = get_Phi_pred(phi, i);
1471 allocation_info_t *info = get_allocation_info(op);
1472 ir_node *pred_block = get_Block_cfgpred_block(block, i);
1475 = (float)get_block_execfreq(execfreqs, pred_block) * AFF_PHI;
1477 if (info->prefs[assigned_r] >= weight)
1480 /* promote the prefered register */
1481 for (r = 0; r < n_regs; ++r) {
1482 if (info->prefs[r] > -weight) {
1483 info->prefs[r] = -weight;
1486 info->prefs[assigned_r] = weight;
1489 propagate_phi_register(op, assigned_r);
1493 static void assign_phi_registers(ir_node *block)
1498 unsigned *assignment;
1500 hungarian_problem_t *bp;
1502 /* count phi nodes */
1503 sched_foreach(block, node) {
1506 if (!arch_irn_consider_in_reg_alloc(cls, node))
1514 /* build a bipartite matching problem for all phi nodes */
1515 bp = hungarian_new(n_phis, n_regs, HUNGARIAN_MATCH_PERFECT);
1517 sched_foreach(block, node) {
1520 allocation_info_t *info;
1523 if (!arch_irn_consider_in_reg_alloc(cls, node))
1526 /* give boni for predecessor colorings */
1527 adapt_phi_prefs(node);
1528 /* add stuff to bipartite problem */
1529 info = get_allocation_info(node);
1530 DB((dbg, LEVEL_3, "Prefs for %+F: ", node));
1531 for (r = 0; r < n_regs; ++r) {
1534 if (!rbitset_is_set(normal_regs, r))
1537 costs = info->prefs[r];
1538 costs = costs < 0 ? -logf(-costs+1) : logf(costs+1);
1541 hungarian_add(bp, n, r, (int)costs);
1542 DB((dbg, LEVEL_3, " %s(%f)", arch_register_for_index(cls, r)->name,
1545 DB((dbg, LEVEL_3, "\n"));
1549 //hungarian_print_cost_matrix(bp, 7);
1550 hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1552 assignment = ALLOCAN(unsigned, n_regs);
1553 res = hungarian_solve(bp, assignment, NULL, 0);
1558 sched_foreach(block, node) {
1560 const arch_register_t *reg;
1564 if (!arch_irn_consider_in_reg_alloc(cls, node))
1567 r = assignment[n++];
1568 assert(rbitset_is_set(normal_regs, r));
1569 reg = arch_register_for_index(cls, r);
1570 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
1573 /* adapt preferences for phi inputs */
1574 if (propagate_phi_registers)
1575 propagate_phi_register(node, r);
1580 * Walker: assign registers to all nodes of a block that
1581 * need registers from the currently considered register class.
1583 static void allocate_coalesce_block(ir_node *block, void *data)
1586 ir_nodeset_t live_nodes;
1589 block_info_t *block_info;
1590 block_info_t **pred_block_infos;
1592 unsigned *forbidden_regs; /**< collects registers which must
1593 not be used for optimistic splits */
1596 DB((dbg, LEVEL_2, "* Block %+F\n", block));
1598 /* clear assignments */
1599 block_info = get_block_info(block);
1600 assignments = block_info->assignments;
1602 ir_nodeset_init(&live_nodes);
1604 /* gather regalloc infos of predecessor blocks */
1605 n_preds = get_Block_n_cfgpreds(block);
1606 pred_block_infos = ALLOCAN(block_info_t*, n_preds);
1607 for (i = 0; i < n_preds; ++i) {
1608 ir_node *pred = get_Block_cfgpred_block(block, i);
1609 block_info_t *pred_info = get_block_info(pred);
1610 pred_block_infos[i] = pred_info;
1613 phi_ins = ALLOCAN(ir_node*, n_preds);
1615 /* collect live-in nodes and preassigned values */
1616 be_lv_foreach(lv, block, be_lv_state_in, i) {
1617 const arch_register_t *reg;
1619 bool need_phi = false;
1621 node = be_lv_get_irn(lv, block, i);
1622 if (!arch_irn_consider_in_reg_alloc(cls, node))
1625 /* check all predecessors for this value, if it is not everywhere the
1626 same or unknown then we have to construct a phi
1627 (we collect the potential phi inputs here) */
1628 for (p = 0; p < n_preds; ++p) {
1629 block_info_t *pred_info = pred_block_infos[p];
1631 if (!pred_info->processed) {
1632 /* use node for now, it will get fixed later */
1636 int a = find_value_in_block_info(pred_info, node);
1638 /* must live out of predecessor */
1640 phi_ins[p] = pred_info->assignments[a];
1641 /* different value from last time? then we need a phi */
1642 if (p > 0 && phi_ins[p-1] != phi_ins[p]) {
1649 ir_mode *mode = get_irn_mode(node);
1650 const arch_register_req_t *req = get_default_req_current_cls();
1653 phi = new_r_Phi(block, n_preds, phi_ins, mode);
1654 be_set_phi_reg_req(phi, req);
1656 DB((dbg, LEVEL_3, "Create Phi %+F (for %+F) -", phi, node));
1657 #ifdef DEBUG_libfirm
1660 for (i = 0; i < n_preds; ++i) {
1661 DB((dbg, LEVEL_3, " %+F", phi_ins[i]));
1663 DB((dbg, LEVEL_3, "\n"));
1666 mark_as_copy_of(phi, node);
1667 sched_add_after(block, phi);
1671 allocation_info_t *info = get_allocation_info(node);
1672 info->current_value = phi_ins[0];
1674 /* Grab 1 of the inputs we constructed (might not be the same as
1675 * "node" as we could see the same copy of the value in all
1680 /* if the node already has a register assigned use it */
1681 reg = arch_get_irn_register(node);
1686 /* remember that this node is live at the beginning of the block */
1687 ir_nodeset_insert(&live_nodes, node);
1690 rbitset_alloca(forbidden_regs, n_regs);
1692 /* handle phis... */
1693 assign_phi_registers(block);
1695 /* all live-ins must have a register */
1696 #ifdef DEBUG_libfirm
1698 ir_nodeset_iterator_t iter;
1699 foreach_ir_nodeset(&live_nodes, node, iter) {
1700 const arch_register_t *reg = arch_get_irn_register(node);
1701 assert(reg != NULL);
1706 /* assign instructions in the block */
1707 sched_foreach(block, node) {
1711 /* phis are already assigned */
1715 rewire_inputs(node);
1717 /* enforce use constraints */
1718 rbitset_clear_all(forbidden_regs, n_regs);
1719 enforce_constraints(&live_nodes, node, forbidden_regs);
1721 rewire_inputs(node);
1723 /* we may not use registers used for inputs for optimistic splits */
1724 arity = get_irn_arity(node);
1725 for (i = 0; i < arity; ++i) {
1726 ir_node *op = get_irn_n(node, i);
1727 const arch_register_t *reg;
1728 if (!arch_irn_consider_in_reg_alloc(cls, op))
1731 reg = arch_get_irn_register(op);
1732 rbitset_set(forbidden_regs, arch_register_get_index(reg));
1735 /* free registers of values last used at this instruction */
1736 free_last_uses(&live_nodes, node);
1738 /* assign output registers */
1739 /* TODO: 2 phases: first: pre-assigned ones, 2nd real regs */
1740 if (get_irn_mode(node) == mode_T) {
1741 const ir_edge_t *edge;
1742 foreach_out_edge(node, edge) {
1743 ir_node *proj = get_edge_src_irn(edge);
1744 if (!arch_irn_consider_in_reg_alloc(cls, proj))
1746 assign_reg(block, proj, forbidden_regs);
1748 } else if (arch_irn_consider_in_reg_alloc(cls, node)) {
1749 assign_reg(block, node, forbidden_regs);
1753 ir_nodeset_destroy(&live_nodes);
1756 block_info->processed = true;
1758 /* permute values at end of predecessor blocks in case of phi-nodes */
1761 for (p = 0; p < n_preds; ++p) {
1762 add_phi_permutations(block, p);
1766 /* if we have exactly 1 successor then we might be able to produce phi
1768 if (get_irn_n_edges_kind(block, EDGE_KIND_BLOCK) == 1) {
1769 const ir_edge_t *edge
1770 = get_irn_out_edge_first_kind(block, EDGE_KIND_BLOCK);
1771 ir_node *succ = get_edge_src_irn(edge);
1772 int p = get_edge_src_pos(edge);
1773 block_info_t *succ_info = get_block_info(succ);
1775 if (succ_info->processed) {
1776 add_phi_permutations(succ, p);
1781 typedef struct block_costs_t block_costs_t;
1782 struct block_costs_t {
1783 float costs; /**< costs of the block */
1784 int dfs_num; /**< depth first search number (to detect backedges) */
1787 static int cmp_block_costs(const void *d1, const void *d2)
1789 const ir_node * const *block1 = d1;
1790 const ir_node * const *block2 = d2;
1791 const block_costs_t *info1 = get_irn_link(*block1);
1792 const block_costs_t *info2 = get_irn_link(*block2);
1793 return QSORT_CMP(info2->costs, info1->costs);
1796 static void determine_block_order(void)
1799 ir_node **blocklist = be_get_cfgpostorder(irg);
1800 int n_blocks = ARR_LEN(blocklist);
1802 pdeq *worklist = new_pdeq();
1803 ir_node **order = XMALLOCN(ir_node*, n_blocks);
1806 /* clear block links... */
1807 for (i = 0; i < n_blocks; ++i) {
1808 ir_node *block = blocklist[i];
1809 set_irn_link(block, NULL);
1812 /* walk blocks in reverse postorder, the costs for each block are the
1813 * sum of the costs of its predecessors (excluding the costs on backedges
1814 * which we can't determine) */
1815 for (i = n_blocks-1; i >= 0; --i) {
1816 block_costs_t *cost_info;
1817 ir_node *block = blocklist[i];
1819 float execfreq = (float)get_block_execfreq(execfreqs, block);
1820 float costs = execfreq;
1821 int n_cfgpreds = get_Block_n_cfgpreds(block);
1823 for (p = 0; p < n_cfgpreds; ++p) {
1824 ir_node *pred_block = get_Block_cfgpred_block(block, p);
1825 block_costs_t *pred_costs = get_irn_link(pred_block);
1826 /* we don't have any info for backedges */
1827 if (pred_costs == NULL)
1829 costs += pred_costs->costs;
1832 cost_info = OALLOCZ(&obst, block_costs_t);
1833 cost_info->costs = costs;
1834 cost_info->dfs_num = dfs_num++;
1835 set_irn_link(block, cost_info);
1838 /* sort array by block costs */
1839 qsort(blocklist, n_blocks, sizeof(blocklist[0]), cmp_block_costs);
1841 ir_reserve_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1842 inc_irg_block_visited(irg);
1844 for (i = 0; i < n_blocks; ++i) {
1845 ir_node *block = blocklist[i];
1846 if (Block_block_visited(block))
1849 /* continually add predecessors with highest costs to worklist
1850 * (without using backedges) */
1852 block_costs_t *info = get_irn_link(block);
1853 ir_node *best_pred = NULL;
1854 float best_costs = -1;
1855 int n_cfgpred = get_Block_n_cfgpreds(block);
1858 pdeq_putr(worklist, block);
1859 mark_Block_block_visited(block);
1860 for (i = 0; i < n_cfgpred; ++i) {
1861 ir_node *pred_block = get_Block_cfgpred_block(block, i);
1862 block_costs_t *pred_info = get_irn_link(pred_block);
1864 /* ignore backedges */
1865 if (pred_info->dfs_num > info->dfs_num)
1868 if (info->costs > best_costs) {
1869 best_costs = info->costs;
1870 best_pred = pred_block;
1874 } while (block != NULL && !Block_block_visited(block));
1876 /* now put all nodes in the worklist in our final order */
1877 while (!pdeq_empty(worklist)) {
1878 ir_node *pblock = pdeq_getr(worklist);
1879 assert(order_p < n_blocks);
1880 order[order_p++] = pblock;
1883 assert(order_p == n_blocks);
1886 ir_free_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1888 DEL_ARR_F(blocklist);
1890 obstack_free(&obst, NULL);
1891 obstack_init(&obst);
1893 block_order = order;
1894 n_block_order = n_blocks;
1898 * Run the register allocator for the current register class.
1900 static void be_pref_alloc_cls(void)
1904 lv = be_assure_liveness(irg);
1905 be_liveness_assure_sets(lv);
1907 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
1909 DB((dbg, LEVEL_2, "=== Allocating registers of %s ===\n", cls->name));
1911 be_clear_links(irg);
1913 irg_block_walk_graph(irg, NULL, analyze_block, NULL);
1914 if (create_congruence_classes)
1915 combine_congruence_classes();
1917 for (i = 0; i < n_block_order; ++i) {
1918 ir_node *block = block_order[i];
1919 allocate_coalesce_block(block, NULL);
1922 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
1925 static void dump(int mask, ir_graph *irg, const char *suffix)
1927 if (be_get_irg_options(irg)->dump_flags & mask)
1928 dump_ir_graph(irg, suffix);
1932 * Run the spiller on the current graph.
1934 static void spill(void)
1936 /* make sure all nodes show their real register pressure */
1937 be_timer_push(T_RA_CONSTR);
1938 be_pre_spill_prepare_constr(irg, cls);
1939 be_timer_pop(T_RA_CONSTR);
1941 dump(DUMP_RA, irg, "-spillprepare");
1944 be_timer_push(T_RA_SPILL);
1945 be_do_spill(irg, cls);
1946 be_timer_pop(T_RA_SPILL);
1948 be_timer_push(T_RA_SPILL_APPLY);
1949 check_for_memory_operands(irg);
1950 be_timer_pop(T_RA_SPILL_APPLY);
1952 dump(DUMP_RA, irg, "-spill");
1956 * The pref register allocator for a whole procedure.
1958 static void be_pref_alloc(ir_graph *new_irg)
1960 const arch_env_t *arch_env = be_get_irg_arch_env(new_irg);
1961 int n_cls = arch_env_get_n_reg_class(arch_env);
1964 obstack_init(&obst);
1967 execfreqs = be_get_irg_exec_freq(irg);
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(be_get_irg_abi(irg), cls, normal_regs);
1986 /* verify schedule and register pressure */
1987 be_timer_push(T_VERIFY);
1988 if (be_get_irg_options(irg)->vrfy_option == BE_VRFY_WARN) {
1989 be_verify_schedule(irg);
1990 be_verify_register_pressure(irg, cls);
1991 } else if (be_get_irg_options(irg)->vrfy_option == BE_VRFY_ASSERT) {
1992 assert(be_verify_schedule(irg) && "Schedule verification failed");
1993 assert(be_verify_register_pressure(irg, cls)
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(irg);
2013 be_timer_pop(T_RA_SPILL_APPLY);
2015 be_timer_push(T_VERIFY);
2016 if (be_get_irg_options(irg)->vrfy_option == BE_VRFY_WARN) {
2017 be_verify_register_allocation(irg);
2018 } else if (be_get_irg_options(irg)->vrfy_option == BE_VRFY_ASSERT) {
2019 assert(be_verify_register_allocation(irg)
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");