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"
77 #define USE_FACTOR 1.0f
78 #define DEF_FACTOR 1.0f
79 #define NEIGHBOR_FACTOR 0.2f
80 #define AFF_SHOULD_BE_SAME 0.5f
82 #define SPLIT_DELTA 1.0f
83 #define MAX_OPTIMISTIC_SPLIT_RECURSION 0
85 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
87 static struct obstack obst;
89 static const arch_register_class_t *cls;
90 static const arch_register_req_t *default_cls_req;
92 static const ir_exec_freq *execfreqs;
93 static unsigned n_regs;
94 static unsigned *normal_regs;
95 static int *congruence_classes;
96 static ir_node **block_order;
97 static int n_block_order;
98 static int create_preferences = true;
99 static int create_congruence_classes = true;
100 static int propagate_phi_registers = true;
102 static const lc_opt_table_entry_t options[] = {
103 LC_OPT_ENT_BOOL("prefs", "use preference based coloring", &create_preferences),
104 LC_OPT_ENT_BOOL("congruences", "create congruence classes", &create_congruence_classes),
105 LC_OPT_ENT_BOOL("prop_phi", "propagate phi registers", &propagate_phi_registers),
109 /** currently active assignments (while processing a basic block)
110 * maps registers to values(their current copies) */
111 static ir_node **assignments;
114 * allocation information: last_uses, register preferences
115 * the information is per firm-node.
117 struct allocation_info_t {
118 unsigned last_uses[2]; /**< bitset indicating last uses (input pos) */
119 ir_node *current_value; /**< copy of the value that should be used */
120 ir_node *original_value; /**< for copies point to original value */
121 float prefs[0]; /**< register preferences */
123 typedef struct allocation_info_t allocation_info_t;
125 /** helper datastructure used when sorting register preferences */
130 typedef struct reg_pref_t reg_pref_t;
132 /** per basic-block information */
133 struct block_info_t {
134 bool processed; /**< indicate whether block is processed */
135 ir_node *assignments[0]; /**< register assignments at end of block */
137 typedef struct block_info_t block_info_t;
140 * Get the allocation info for a node.
141 * The info is allocated on the first visit of a node.
143 static allocation_info_t *get_allocation_info(ir_node *node)
145 allocation_info_t *info = get_irn_link(node);
147 info = OALLOCFZ(&obst, allocation_info_t, prefs, n_regs);
148 info->current_value = node;
149 info->original_value = node;
150 set_irn_link(node, info);
156 static allocation_info_t *try_get_allocation_info(const ir_node *node)
158 return (allocation_info_t*) get_irn_link(node);
162 * Get allocation information for a basic block
164 static block_info_t *get_block_info(ir_node *block)
166 block_info_t *info = get_irn_link(block);
168 assert(is_Block(block));
170 info = OALLOCFZ(&obst, block_info_t, assignments, n_regs);
171 set_irn_link(block, info);
178 * Get default register requirement for the current register class
180 static const arch_register_req_t *get_default_req_current_cls(void)
182 if (default_cls_req == NULL) {
183 struct obstack *obst = get_irg_obstack(irg);
184 arch_register_req_t *req = OALLOCZ(obst, arch_register_req_t);
186 req->type = arch_register_req_type_normal;
190 default_cls_req = req;
192 return default_cls_req;
196 * Link the allocation info of a node to a copy.
197 * Afterwards, both nodes uses the same allocation info.
198 * Copy must not have an allocation info assigned yet.
200 * @param copy the node that gets the allocation info assigned
201 * @param value the original node
203 static void mark_as_copy_of(ir_node *copy, ir_node *value)
206 allocation_info_t *info = get_allocation_info(value);
207 allocation_info_t *copy_info = get_allocation_info(copy);
209 /* find original value */
210 original = info->original_value;
211 if (original != value) {
212 info = get_allocation_info(original);
215 assert(info->original_value == original);
216 info->current_value = copy;
218 /* the copy should not be linked to something else yet */
219 assert(copy_info->original_value == copy);
220 copy_info->original_value = original;
222 /* copy over allocation preferences */
223 memcpy(copy_info->prefs, info->prefs, n_regs * sizeof(copy_info->prefs[0]));
227 * Calculate the penalties for every register on a node and its live neighbors.
229 * @param live_nodes the set of live nodes at the current position, may be NULL
230 * @param penalty the penalty to subtract from
231 * @param limited a raw bitset containing the limited set for the node
232 * @param node the node
234 static void give_penalties_for_limits(const ir_nodeset_t *live_nodes,
235 float penalty, const unsigned* limited,
238 ir_nodeset_iterator_t iter;
241 allocation_info_t *info = get_allocation_info(node);
244 /* give penalty for all forbidden regs */
245 for (r = 0; r < n_regs; ++r) {
246 if (rbitset_is_set(limited, r))
249 info->prefs[r] -= penalty;
252 /* all other live values should get a penalty for allowed regs */
253 if (live_nodes == NULL)
256 penalty *= NEIGHBOR_FACTOR;
257 n_allowed = rbitset_popcount(limited, n_regs);
259 /* only create a very weak penalty if multiple regs are allowed */
260 penalty = (penalty * 0.8f) / n_allowed;
262 foreach_ir_nodeset(live_nodes, neighbor, iter) {
263 allocation_info_t *neighbor_info;
265 /* TODO: if op is used on multiple inputs we might not do a
267 if (neighbor == node)
270 neighbor_info = get_allocation_info(neighbor);
271 for (r = 0; r < n_regs; ++r) {
272 if (!rbitset_is_set(limited, r))
275 neighbor_info->prefs[r] -= penalty;
281 * Calculate the preferences of a definition for the current register class.
282 * If the definition uses a limited set of registers, reduce the preferences
283 * for the limited register on the node and its neighbors.
285 * @param live_nodes the set of live nodes at the current node
286 * @param weight the weight
287 * @param node the current node
289 static void check_defs(const ir_nodeset_t *live_nodes, float weight,
292 const arch_register_req_t *req = arch_get_register_req_out(node);
293 if (req->type & arch_register_req_type_limited) {
294 const unsigned *limited = req->limited;
295 float penalty = weight * DEF_FACTOR;
296 give_penalties_for_limits(live_nodes, penalty, limited, node);
299 if (req->type & arch_register_req_type_should_be_same) {
300 ir_node *insn = skip_Proj(node);
301 allocation_info_t *info = get_allocation_info(node);
302 int arity = get_irn_arity(insn);
305 float factor = 1.0f / rbitset_popcount(&req->other_same, arity);
306 for (i = 0; i < arity; ++i) {
309 allocation_info_t *op_info;
311 if (!rbitset_is_set(&req->other_same, i))
314 op = get_irn_n(insn, i);
316 /* if we the value at the should_be_same input doesn't die at the
317 * node, then it is no use to propagate the constraints (since a
318 * copy will emerge anyway) */
319 if (ir_nodeset_contains(live_nodes, op))
322 op_info = get_allocation_info(op);
323 for (r = 0; r < n_regs; ++r) {
324 op_info->prefs[r] += info->prefs[r] * factor;
331 * Walker: Runs an a block calculates the preferences for any
332 * node and every register from the considered register class.
334 static void analyze_block(ir_node *block, void *data)
336 float weight = (float)get_block_execfreq(execfreqs, block);
337 ir_nodeset_t live_nodes;
341 ir_nodeset_init(&live_nodes);
342 be_liveness_end_of_block(lv, cls, block, &live_nodes);
344 sched_foreach_reverse(block, node) {
345 allocation_info_t *info;
352 if (create_preferences) {
354 be_foreach_definition(node, cls, value,
355 check_defs(&live_nodes, weight, value);
360 arity = get_irn_arity(node);
362 /* the allocation info node currently only uses 1 unsigned value
363 to mark last used inputs. So we will fail for a node with more than
365 if (arity >= (int) sizeof(info->last_uses) * 8) {
366 panic("Node with more than %d inputs not supported yet",
367 (int) sizeof(info->last_uses) * 8);
370 info = get_allocation_info(node);
371 for (i = 0; i < arity; ++i) {
372 ir_node *op = get_irn_n(node, i);
373 const arch_register_req_t *req = arch_get_register_req_out(op);
377 /* last usage of a value? */
378 if (!ir_nodeset_contains(&live_nodes, op)) {
379 rbitset_set(info->last_uses, i);
383 be_liveness_transfer(cls, node, &live_nodes);
385 if (create_preferences) {
386 /* update weights based on usage constraints */
387 for (i = 0; i < arity; ++i) {
388 const arch_register_req_t *req;
389 const unsigned *limited;
390 ir_node *op = get_irn_n(node, i);
392 if (!arch_irn_consider_in_reg_alloc(cls, op))
395 req = arch_get_register_req(node, i);
396 if (!(req->type & arch_register_req_type_limited))
399 limited = req->limited;
400 give_penalties_for_limits(&live_nodes, weight * USE_FACTOR,
406 ir_nodeset_destroy(&live_nodes);
409 static void congruence_def(ir_nodeset_t *live_nodes, const ir_node *node)
411 const arch_register_req_t *req = arch_get_register_req_out(node);
413 /* should be same constraint? */
414 if (req->type & arch_register_req_type_should_be_same) {
415 const ir_node *insn = skip_Proj_const(node);
416 int arity = get_irn_arity(insn);
418 unsigned node_idx = get_irn_idx(node);
419 node_idx = uf_find(congruence_classes, node_idx);
421 for (i = 0; i < arity; ++i) {
425 ir_nodeset_iterator_t iter;
426 bool interferes = false;
428 if (!rbitset_is_set(&req->other_same, i))
431 op = get_irn_n(insn, i);
432 op_idx = get_irn_idx(op);
433 op_idx = uf_find(congruence_classes, op_idx);
435 /* do we interfere with the value */
436 foreach_ir_nodeset(live_nodes, live, iter) {
437 int lv_idx = get_irn_idx(live);
438 lv_idx = uf_find(congruence_classes, lv_idx);
439 if (lv_idx == op_idx) {
444 /* don't put in same affinity class if we interfere */
448 node_idx = uf_union(congruence_classes, node_idx, op_idx);
449 DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
451 /* one should_be_same is enough... */
457 static void create_congruence_class(ir_node *block, void *data)
459 ir_nodeset_t live_nodes;
463 ir_nodeset_init(&live_nodes);
464 be_liveness_end_of_block(lv, cls, block, &live_nodes);
466 /* check should be same constraints */
467 sched_foreach_reverse(block, node) {
472 be_foreach_definition(node, cls, value,
473 congruence_def(&live_nodes, value);
475 be_liveness_transfer(cls, node, &live_nodes);
478 /* check phi congruence classes */
479 sched_foreach_reverse_from(node, node) {
483 assert(is_Phi(node));
485 if (!arch_irn_consider_in_reg_alloc(cls, node))
488 node_idx = get_irn_idx(node);
489 node_idx = uf_find(congruence_classes, node_idx);
491 arity = get_irn_arity(node);
492 for (i = 0; i < arity; ++i) {
493 bool interferes = false;
494 ir_nodeset_iterator_t iter;
499 allocation_info_t *head_info;
500 allocation_info_t *other_info;
501 ir_node *op = get_Phi_pred(node, i);
502 int op_idx = get_irn_idx(op);
503 op_idx = uf_find(congruence_classes, op_idx);
505 /* do we interfere with the value */
506 foreach_ir_nodeset(&live_nodes, live, iter) {
507 int lv_idx = get_irn_idx(live);
508 lv_idx = uf_find(congruence_classes, lv_idx);
509 if (lv_idx == op_idx) {
514 /* don't put in same affinity class if we interfere */
517 /* any other phi has the same input? */
518 sched_foreach(block, phi) {
523 if (!arch_irn_consider_in_reg_alloc(cls, phi))
525 oop = get_Phi_pred(phi, i);
528 oop_idx = get_irn_idx(oop);
529 oop_idx = uf_find(congruence_classes, oop_idx);
530 if (oop_idx == op_idx) {
538 /* merge the 2 congruence classes and sum up their preferences */
539 old_node_idx = node_idx;
540 node_idx = uf_union(congruence_classes, node_idx, op_idx);
541 DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
544 old_node_idx = node_idx == old_node_idx ? op_idx : old_node_idx;
545 head_info = get_allocation_info(get_idx_irn(irg, node_idx));
546 other_info = get_allocation_info(get_idx_irn(irg, old_node_idx));
547 for (r = 0; r < n_regs; ++r) {
548 head_info->prefs[r] += other_info->prefs[r];
554 static void set_congruence_prefs(ir_node *node, void *data)
556 allocation_info_t *info;
557 allocation_info_t *head_info;
558 unsigned node_idx = get_irn_idx(node);
559 unsigned node_set = uf_find(congruence_classes, node_idx);
563 /* head of congruence class or not in any class */
564 if (node_set == node_idx)
567 if (!arch_irn_consider_in_reg_alloc(cls, node))
570 head_info = get_allocation_info(get_idx_irn(irg, node_set));
571 info = get_allocation_info(node);
573 memcpy(info->prefs, head_info->prefs, n_regs * sizeof(info->prefs[0]));
576 static void combine_congruence_classes(void)
578 size_t n = get_irg_last_idx(irg);
579 congruence_classes = XMALLOCN(int, n);
580 uf_init(congruence_classes, n);
582 /* create congruence classes */
583 irg_block_walk_graph(irg, create_congruence_class, NULL, NULL);
584 /* merge preferences */
585 irg_walk_graph(irg, set_congruence_prefs, NULL, NULL);
586 free(congruence_classes);
592 * Assign register reg to the given node.
594 * @param node the node
595 * @param reg the register
597 static void use_reg(ir_node *node, const arch_register_t *reg)
599 unsigned r = arch_register_get_index(reg);
600 assignments[r] = node;
601 arch_set_irn_register(node, reg);
604 static void free_reg_of_value(ir_node *node)
606 const arch_register_t *reg;
609 if (!arch_irn_consider_in_reg_alloc(cls, node))
612 reg = arch_get_irn_register(node);
613 r = arch_register_get_index(reg);
614 /* assignment->value may be NULL if a value is used at 2 inputs
615 so it gets freed twice. */
616 assert(assignments[r] == node || assignments[r] == NULL);
617 assignments[r] = NULL;
621 * Compare two register preferences in decreasing order.
623 static int compare_reg_pref(const void *e1, const void *e2)
625 const reg_pref_t *rp1 = (const reg_pref_t*) e1;
626 const reg_pref_t *rp2 = (const reg_pref_t*) e2;
627 if (rp1->pref < rp2->pref)
629 if (rp1->pref > rp2->pref)
634 static void fill_sort_candidates(reg_pref_t *regprefs,
635 const allocation_info_t *info)
639 for (r = 0; r < n_regs; ++r) {
640 float pref = info->prefs[r];
642 regprefs[r].pref = pref;
644 /* TODO: use a stable sort here to avoid unnecessary register jumping */
645 qsort(regprefs, n_regs, sizeof(regprefs[0]), compare_reg_pref);
648 static bool try_optimistic_split(ir_node *to_split, ir_node *before,
649 float pref, float pref_delta,
650 unsigned *forbidden_regs, int recursion)
652 const arch_register_t *from_reg;
653 const arch_register_t *reg;
654 ir_node *original_insn;
660 allocation_info_t *info = get_allocation_info(to_split);
663 float split_threshold;
667 /* stupid hack: don't optimisticallt split don't spill nodes...
668 * (so we don't split away the values produced because of
669 * must_be_different constraints) */
670 original_insn = skip_Proj(info->original_value);
671 if (arch_irn_get_flags(original_insn) & arch_irn_flags_dont_spill)
674 from_reg = arch_get_irn_register(to_split);
675 from_r = arch_register_get_index(from_reg);
676 block = get_nodes_block(before);
677 split_threshold = (float)get_block_execfreq(execfreqs, block) * SPLIT_DELTA;
679 if (pref_delta < split_threshold*0.5)
682 /* find the best free position where we could move to */
683 prefs = ALLOCAN(reg_pref_t, n_regs);
684 fill_sort_candidates(prefs, info);
685 for (i = 0; i < n_regs; ++i) {
689 bool old_source_state;
691 /* we need a normal register which is not an output register
692 an different from the current register of to_split */
694 if (!rbitset_is_set(normal_regs, r))
696 if (rbitset_is_set(forbidden_regs, r))
701 /* is the split worth it? */
702 delta = pref_delta + prefs[i].pref;
703 if (delta < split_threshold) {
704 DB((dbg, LEVEL_3, "Not doing optimistical split of %+F (depth %d), win %f too low\n",
705 to_split, recursion, delta));
709 /* if the register is free then we can do the split */
710 if (assignments[r] == NULL)
713 /* otherwise we might try recursively calling optimistic_split */
714 if (recursion+1 > MAX_OPTIMISTIC_SPLIT_RECURSION)
717 apref = prefs[i].pref;
718 apref_delta = i+1 < n_regs ? apref - prefs[i+1].pref : 0;
719 apref_delta += pref_delta - split_threshold;
721 /* our source register isn't a useful destination for recursive
723 old_source_state = rbitset_is_set(forbidden_regs, from_r);
724 rbitset_set(forbidden_regs, from_r);
725 /* try recursive split */
726 res = try_optimistic_split(assignments[r], before, apref,
727 apref_delta, forbidden_regs, recursion+1);
728 /* restore our destination */
729 if (old_source_state) {
730 rbitset_set(forbidden_regs, from_r);
732 rbitset_clear(forbidden_regs, from_r);
741 reg = arch_register_for_index(cls, r);
742 copy = be_new_Copy(cls, block, to_split);
743 mark_as_copy_of(copy, to_split);
744 /* hacky, but correct here */
745 if (assignments[arch_register_get_index(from_reg)] == to_split)
746 free_reg_of_value(to_split);
748 sched_add_before(before, copy);
751 "Optimistic live-range split %+F move %+F(%s) -> %s before %+F (win %f, depth %d)\n",
752 copy, to_split, from_reg->name, reg->name, before, delta, recursion));
757 * Determine and assign a register for node @p node
759 static void assign_reg(const ir_node *block, ir_node *node,
760 unsigned *forbidden_regs)
762 const arch_register_t *reg;
763 allocation_info_t *info;
764 const arch_register_req_t *req;
765 reg_pref_t *reg_prefs;
768 const unsigned *allowed_regs;
771 assert(!is_Phi(node));
772 /* preassigned register? */
773 reg = arch_get_irn_register(node);
775 DB((dbg, LEVEL_2, "Preassignment %+F -> %s\n", node, reg->name));
780 req = arch_get_register_req_out(node);
781 /* ignore reqs must be preassigned */
782 assert (! (req->type & arch_register_req_type_ignore));
784 /* give should_be_same boni */
785 info = get_allocation_info(node);
786 in_node = skip_Proj(node);
787 if (req->type & arch_register_req_type_should_be_same) {
788 float weight = (float)get_block_execfreq(execfreqs, block);
789 int arity = get_irn_arity(in_node);
792 assert(arity <= (int) sizeof(req->other_same) * 8);
793 for (i = 0; i < arity; ++i) {
795 const arch_register_t *reg;
797 if (!rbitset_is_set(&req->other_same, i))
800 in = get_irn_n(in_node, i);
801 reg = arch_get_irn_register(in);
803 r = arch_register_get_index(reg);
805 /* if the value didn't die here then we should not propagate the
806 * should_be_same info */
807 if (assignments[r] == in)
810 info->prefs[r] += weight * AFF_SHOULD_BE_SAME;
814 /* create list of register candidates and sort by their preference */
815 DB((dbg, LEVEL_2, "Candidates for %+F:", node));
816 reg_prefs = alloca(n_regs * sizeof(reg_prefs[0]));
817 fill_sort_candidates(reg_prefs, info);
818 for (i = 0; i < n_regs; ++i) {
819 unsigned num = reg_prefs[i].num;
820 const arch_register_t *reg;
822 if (!rbitset_is_set(normal_regs, num))
825 reg = arch_register_for_index(cls, num);
826 DB((dbg, LEVEL_2, " %s(%f)", reg->name, reg_prefs[i].pref));
828 DB((dbg, LEVEL_2, "\n"));
830 allowed_regs = normal_regs;
831 if (req->type & arch_register_req_type_limited) {
832 allowed_regs = req->limited;
835 for (i = 0; i < n_regs; ++i) {
840 r = reg_prefs[i].num;
841 if (!rbitset_is_set(allowed_regs, r))
843 if (assignments[r] == NULL)
845 pref = reg_prefs[i].pref;
846 delta = i+1 < n_regs ? pref - reg_prefs[i+1].pref : 0;
847 before = skip_Proj(node);
848 res = try_optimistic_split(assignments[r], before,
849 pref, delta, forbidden_regs, 0);
854 /* the common reason to hit this panic is when 1 of your nodes is not
855 * register pressure faithful */
856 panic("No register left for %+F\n", node);
859 reg = arch_register_for_index(cls, r);
860 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
865 * Add an permutation in front of a node and change the assignments
866 * due to this permutation.
868 * To understand this imagine a permutation like this:
878 * First we count how many destinations a single value has. At the same time
879 * we can be sure that each destination register has at most 1 source register
880 * (it can have 0 which means we don't care what value is in it).
881 * We ignore all fullfilled permuations (like 7->7)
882 * In a first pass we create as much copy instructions as possible as they
883 * are generally cheaper than exchanges. We do this by counting into how many
884 * destinations a register has to be copied (in the example it's 2 for register
885 * 3, or 1 for the registers 1,2,4 and 7).
886 * We can then create a copy into every destination register when the usecount
887 * of that register is 0 (= noone else needs the value in the register).
889 * After this step we should have cycles left. We implement a cyclic permutation
890 * of n registers with n-1 transpositions.
892 * @param live_nodes the set of live nodes, updated due to live range split
893 * @param before the node before we add the permutation
894 * @param permutation the permutation array indices are the destination
895 * registers, the values in the array are the source
898 static void permute_values(ir_nodeset_t *live_nodes, ir_node *before,
899 unsigned *permutation)
901 unsigned *n_used = ALLOCANZ(unsigned, n_regs);
905 /* determine how often each source register needs to be read */
906 for (r = 0; r < n_regs; ++r) {
907 unsigned old_reg = permutation[r];
910 value = assignments[old_reg];
912 /* nothing to do here, reg is not live. Mark it as fixpoint
913 * so we ignore it in the next steps */
921 block = get_nodes_block(before);
923 /* step1: create copies where immediately possible */
924 for (r = 0; r < n_regs; /* empty */) {
927 const arch_register_t *reg;
928 unsigned old_r = permutation[r];
930 /* - no need to do anything for fixed points.
931 - we can't copy if the value in the dest reg is still needed */
932 if (old_r == r || n_used[r] > 0) {
938 src = assignments[old_r];
939 copy = be_new_Copy(cls, block, src);
940 sched_add_before(before, copy);
941 reg = arch_register_for_index(cls, r);
942 DB((dbg, LEVEL_2, "Copy %+F (from %+F, before %+F) -> %s\n",
943 copy, src, before, reg->name));
944 mark_as_copy_of(copy, src);
947 if (live_nodes != NULL) {
948 ir_nodeset_insert(live_nodes, copy);
951 /* old register has 1 user less, permutation is resolved */
952 assert(arch_register_get_index(arch_get_irn_register(src)) == old_r);
955 assert(n_used[old_r] > 0);
957 if (n_used[old_r] == 0) {
958 if (live_nodes != NULL) {
959 ir_nodeset_remove(live_nodes, src);
961 free_reg_of_value(src);
964 /* advance or jump back (if this copy enabled another copy) */
965 if (old_r < r && n_used[old_r] == 0) {
972 /* at this point we only have "cycles" left which we have to resolve with
974 * TODO: if we have free registers left, then we should really use copy
975 * instructions for any cycle longer than 2 registers...
976 * (this is probably architecture dependent, there might be archs where
977 * copies are preferable even for 2-cycles) */
979 /* create perms with the rest */
980 for (r = 0; r < n_regs; /* empty */) {
981 const arch_register_t *reg;
982 unsigned old_r = permutation[r];
994 /* we shouldn't have copies from 1 value to multiple destinations left*/
995 assert(n_used[old_r] == 1);
997 /* exchange old_r and r2; after that old_r is a fixed point */
998 r2 = permutation[old_r];
1000 in[0] = assignments[r2];
1001 in[1] = assignments[old_r];
1002 perm = be_new_Perm(cls, block, 2, in);
1003 sched_add_before(before, perm);
1004 DB((dbg, LEVEL_2, "Perm %+F (perm %+F,%+F, before %+F)\n",
1005 perm, in[0], in[1], before));
1007 proj0 = new_r_Proj(perm, get_irn_mode(in[0]), 0);
1008 mark_as_copy_of(proj0, in[0]);
1009 reg = arch_register_for_index(cls, old_r);
1010 use_reg(proj0, reg);
1012 proj1 = new_r_Proj(perm, get_irn_mode(in[1]), 1);
1013 mark_as_copy_of(proj1, in[1]);
1014 reg = arch_register_for_index(cls, r2);
1015 use_reg(proj1, reg);
1017 /* 1 value is now in the correct register */
1018 permutation[old_r] = old_r;
1019 /* the source of r changed to r2 */
1020 permutation[r] = r2;
1022 /* if we have reached a fixpoint update data structures */
1023 if (live_nodes != NULL) {
1024 ir_nodeset_remove(live_nodes, in[0]);
1025 ir_nodeset_remove(live_nodes, in[1]);
1026 ir_nodeset_remove(live_nodes, proj0);
1027 ir_nodeset_insert(live_nodes, proj1);
1031 #ifdef DEBUG_libfirm
1032 /* now we should only have fixpoints left */
1033 for (r = 0; r < n_regs; ++r) {
1034 assert(permutation[r] == r);
1040 * Free regs for values last used.
1042 * @param live_nodes set of live nodes, will be updated
1043 * @param node the node to consider
1045 static void free_last_uses(ir_nodeset_t *live_nodes, ir_node *node)
1047 allocation_info_t *info = get_allocation_info(node);
1048 const unsigned *last_uses = info->last_uses;
1049 int arity = get_irn_arity(node);
1052 for (i = 0; i < arity; ++i) {
1055 /* check if one operand is the last use */
1056 if (!rbitset_is_set(last_uses, i))
1059 op = get_irn_n(node, i);
1060 free_reg_of_value(op);
1061 ir_nodeset_remove(live_nodes, op);
1066 * change inputs of a node to the current value (copies/perms)
1068 static void rewire_inputs(ir_node *node)
1071 int arity = get_irn_arity(node);
1073 for (i = 0; i < arity; ++i) {
1074 ir_node *op = get_irn_n(node, i);
1075 allocation_info_t *info = try_get_allocation_info(op);
1080 info = get_allocation_info(info->original_value);
1081 if (info->current_value != op) {
1082 set_irn_n(node, i, info->current_value);
1088 * Create a bitset of registers occupied with value living through an
1091 static void determine_live_through_regs(unsigned *bitset, ir_node *node)
1093 const allocation_info_t *info = get_allocation_info(node);
1098 /* mark all used registers as potentially live-through */
1099 for (r = 0; r < n_regs; ++r) {
1100 if (assignments[r] == NULL)
1102 if (!rbitset_is_set(normal_regs, r))
1105 rbitset_set(bitset, r);
1108 /* remove registers of value dying at the instruction */
1109 arity = get_irn_arity(node);
1110 for (i = 0; i < arity; ++i) {
1112 const arch_register_t *reg;
1114 if (!rbitset_is_set(info->last_uses, i))
1117 op = get_irn_n(node, i);
1118 reg = arch_get_irn_register(op);
1119 rbitset_clear(bitset, arch_register_get_index(reg));
1124 * Enforce constraints at a node by live range splits.
1126 * @param live_nodes the set of live nodes, might be changed
1127 * @param node the current node
1129 static void enforce_constraints(ir_nodeset_t *live_nodes, ir_node *node,
1130 unsigned *forbidden_regs)
1132 int arity = get_irn_arity(node);
1134 hungarian_problem_t *bp;
1136 unsigned *assignment;
1139 /* construct a list of register occupied by live-through values */
1140 unsigned *live_through_regs = NULL;
1142 /* see if any use constraints are not met */
1144 for (i = 0; i < arity; ++i) {
1145 ir_node *op = get_irn_n(node, i);
1146 const arch_register_t *reg;
1147 const arch_register_req_t *req;
1148 const unsigned *limited;
1151 if (!arch_irn_consider_in_reg_alloc(cls, op))
1154 /* are there any limitations for the i'th operand? */
1155 req = arch_get_register_req(node, i);
1156 if (!(req->type & arch_register_req_type_limited))
1159 limited = req->limited;
1160 reg = arch_get_irn_register(op);
1161 r = arch_register_get_index(reg);
1162 if (!rbitset_is_set(limited, r)) {
1163 /* found an assignment outside the limited set */
1169 /* is any of the live-throughs using a constrained output register? */
1170 be_foreach_definition(node, cls, value,
1171 if (! (req_->type & arch_register_req_type_limited))
1173 if (live_through_regs == NULL) {
1174 rbitset_alloca(live_through_regs, n_regs);
1175 determine_live_through_regs(live_through_regs, node);
1177 rbitset_or(forbidden_regs, req_->limited, n_regs);
1178 if (rbitsets_have_common(req_->limited, live_through_regs, n_regs))
1185 /* create these arrays if we haven't yet */
1186 if (live_through_regs == NULL) {
1187 rbitset_alloca(live_through_regs, n_regs);
1190 /* at this point we have to construct a bipartite matching problem to see
1191 * which values should go to which registers
1192 * Note: We're building the matrix in "reverse" - source registers are
1193 * right, destinations left because this will produce the solution
1194 * in the format required for permute_values.
1196 bp = hungarian_new(n_regs, n_regs, HUNGARIAN_MATCH_PERFECT);
1198 /* add all combinations, then remove not allowed ones */
1199 for (l = 0; l < n_regs; ++l) {
1200 if (!rbitset_is_set(normal_regs, l)) {
1201 hungarian_add(bp, l, l, 1);
1205 for (r = 0; r < n_regs; ++r) {
1206 if (!rbitset_is_set(normal_regs, r))
1208 /* livethrough values may not use constrainted output registers */
1209 if (rbitset_is_set(live_through_regs, l)
1210 && rbitset_is_set(forbidden_regs, r))
1213 hungarian_add(bp, r, l, l == r ? 9 : 8);
1217 for (i = 0; i < arity; ++i) {
1218 ir_node *op = get_irn_n(node, i);
1219 const arch_register_t *reg;
1220 const arch_register_req_t *req;
1221 const unsigned *limited;
1222 unsigned current_reg;
1224 if (!arch_irn_consider_in_reg_alloc(cls, op))
1227 req = arch_get_register_req(node, i);
1228 if (!(req->type & arch_register_req_type_limited))
1231 limited = req->limited;
1232 reg = arch_get_irn_register(op);
1233 current_reg = arch_register_get_index(reg);
1234 for (r = 0; r < n_regs; ++r) {
1235 if (rbitset_is_set(limited, r))
1237 hungarian_remove(bp, r, current_reg);
1241 //hungarian_print_cost_matrix(bp, 1);
1242 hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1244 assignment = ALLOCAN(unsigned, n_regs);
1245 res = hungarian_solve(bp, assignment, NULL, 0);
1249 fprintf(stderr, "Swap result:");
1250 for (i = 0; i < (int) n_regs; ++i) {
1251 fprintf(stderr, " %d", assignment[i]);
1253 fprintf(stderr, "\n");
1258 permute_values(live_nodes, node, assignment);
1261 /** test wether a node @p n is a copy of the value of node @p of */
1262 static bool is_copy_of(ir_node *value, ir_node *test_value)
1264 allocation_info_t *test_info;
1265 allocation_info_t *info;
1267 if (value == test_value)
1270 info = get_allocation_info(value);
1271 test_info = get_allocation_info(test_value);
1272 return test_info->original_value == info->original_value;
1276 * find a value in the end-assignment of a basic block
1277 * @returns the index into the assignment array if found
1280 static int find_value_in_block_info(block_info_t *info, ir_node *value)
1283 ir_node **assignments = info->assignments;
1284 for (r = 0; r < n_regs; ++r) {
1285 ir_node *a_value = assignments[r];
1287 if (a_value == NULL)
1289 if (is_copy_of(a_value, value))
1297 * Create the necessary permutations at the end of a basic block to fullfill
1298 * the register assignment for phi-nodes in the next block
1300 static void add_phi_permutations(ir_node *block, int p)
1303 unsigned *permutation;
1304 ir_node **old_assignments;
1305 bool need_permutation;
1307 ir_node *pred = get_Block_cfgpred_block(block, p);
1309 block_info_t *pred_info = get_block_info(pred);
1311 /* predecessor not processed yet? nothing to do */
1312 if (!pred_info->processed)
1315 permutation = ALLOCAN(unsigned, n_regs);
1316 for (r = 0; r < n_regs; ++r) {
1320 /* check phi nodes */
1321 need_permutation = false;
1322 phi = sched_first(block);
1323 for ( ; is_Phi(phi); phi = sched_next(phi)) {
1324 const arch_register_t *reg;
1325 const arch_register_t *op_reg;
1330 if (!arch_irn_consider_in_reg_alloc(cls, phi))
1333 op = get_Phi_pred(phi, p);
1334 a = find_value_in_block_info(pred_info, op);
1337 reg = arch_get_irn_register(phi);
1338 regn = arch_register_get_index(reg);
1339 /* same register? nothing to do */
1343 op = pred_info->assignments[a];
1344 op_reg = arch_get_irn_register(op);
1345 /* virtual or joker registers are ok too */
1346 if ((op_reg->type & arch_register_type_joker)
1347 || (op_reg->type & arch_register_type_virtual))
1350 permutation[regn] = a;
1351 need_permutation = true;
1354 if (need_permutation) {
1355 /* permute values at end of predecessor */
1356 old_assignments = assignments;
1357 assignments = pred_info->assignments;
1358 permute_values(NULL, be_get_end_of_block_insertion_point(pred),
1360 assignments = old_assignments;
1363 /* change phi nodes to use the copied values */
1364 phi = sched_first(block);
1365 for ( ; is_Phi(phi); phi = sched_next(phi)) {
1369 if (!arch_irn_consider_in_reg_alloc(cls, phi))
1372 op = get_Phi_pred(phi, p);
1374 /* we have permuted all values into the correct registers so we can
1375 simply query which value occupies the phis register in the
1377 a = arch_register_get_index(arch_get_irn_register(phi));
1378 op = pred_info->assignments[a];
1379 set_Phi_pred(phi, p, op);
1384 * Set preferences for a phis register based on the registers used on the
1387 static void adapt_phi_prefs(ir_node *phi)
1390 int arity = get_irn_arity(phi);
1391 ir_node *block = get_nodes_block(phi);
1392 allocation_info_t *info = get_allocation_info(phi);
1394 for (i = 0; i < arity; ++i) {
1395 ir_node *op = get_irn_n(phi, i);
1396 const arch_register_t *reg = arch_get_irn_register(op);
1397 ir_node *pred_block;
1398 block_info_t *pred_block_info;
1404 /* we only give the bonus if the predecessor already has registers
1405 * assigned, otherwise we only see a dummy value
1406 * and any conclusions about its register are useless */
1407 pred_block = get_Block_cfgpred_block(block, i);
1408 pred_block_info = get_block_info(pred_block);
1409 if (!pred_block_info->processed)
1412 /* give bonus for already assigned register */
1413 weight = (float)get_block_execfreq(execfreqs, pred_block);
1414 r = arch_register_get_index(reg);
1415 info->prefs[r] += weight * AFF_PHI;
1420 * After a phi has been assigned a register propagate preference inputs
1421 * to the phi inputs.
1423 static void propagate_phi_register(ir_node *phi, unsigned assigned_r)
1426 ir_node *block = get_nodes_block(phi);
1427 int arity = get_irn_arity(phi);
1429 for (i = 0; i < arity; ++i) {
1430 ir_node *op = get_Phi_pred(phi, i);
1431 allocation_info_t *info = get_allocation_info(op);
1432 ir_node *pred_block = get_Block_cfgpred_block(block, i);
1435 = (float)get_block_execfreq(execfreqs, pred_block) * AFF_PHI;
1437 if (info->prefs[assigned_r] >= weight)
1440 /* promote the prefered register */
1441 for (r = 0; r < n_regs; ++r) {
1442 if (info->prefs[r] > -weight) {
1443 info->prefs[r] = -weight;
1446 info->prefs[assigned_r] = weight;
1449 propagate_phi_register(op, assigned_r);
1453 static void assign_phi_registers(ir_node *block)
1458 unsigned *assignment;
1460 hungarian_problem_t *bp;
1462 /* count phi nodes */
1463 sched_foreach(block, node) {
1466 if (!arch_irn_consider_in_reg_alloc(cls, node))
1474 /* build a bipartite matching problem for all phi nodes */
1475 bp = hungarian_new(n_phis, n_regs, HUNGARIAN_MATCH_PERFECT);
1477 sched_foreach(block, node) {
1480 allocation_info_t *info;
1483 if (!arch_irn_consider_in_reg_alloc(cls, node))
1486 /* give boni for predecessor colorings */
1487 adapt_phi_prefs(node);
1488 /* add stuff to bipartite problem */
1489 info = get_allocation_info(node);
1490 DB((dbg, LEVEL_3, "Prefs for %+F: ", node));
1491 for (r = 0; r < n_regs; ++r) {
1494 if (!rbitset_is_set(normal_regs, r))
1497 costs = info->prefs[r];
1498 costs = costs < 0 ? -logf(-costs+1) : logf(costs+1);
1501 hungarian_add(bp, n, r, (int)costs);
1502 DB((dbg, LEVEL_3, " %s(%f)", arch_register_for_index(cls, r)->name,
1505 DB((dbg, LEVEL_3, "\n"));
1509 //hungarian_print_cost_matrix(bp, 7);
1510 hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1512 assignment = ALLOCAN(unsigned, n_regs);
1513 res = hungarian_solve(bp, assignment, NULL, 0);
1518 sched_foreach(block, node) {
1520 const arch_register_t *reg;
1524 if (!arch_irn_consider_in_reg_alloc(cls, node))
1527 r = assignment[n++];
1528 assert(rbitset_is_set(normal_regs, r));
1529 reg = arch_register_for_index(cls, r);
1530 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
1533 /* adapt preferences for phi inputs */
1534 if (propagate_phi_registers)
1535 propagate_phi_register(node, r);
1540 * Walker: assign registers to all nodes of a block that
1541 * need registers from the currently considered register class.
1543 static void allocate_coalesce_block(ir_node *block, void *data)
1546 ir_nodeset_t live_nodes;
1549 block_info_t *block_info;
1550 block_info_t **pred_block_infos;
1552 unsigned *forbidden_regs; /**< collects registers which must
1553 not be used for optimistic splits */
1556 DB((dbg, LEVEL_2, "* Block %+F\n", block));
1558 /* clear assignments */
1559 block_info = get_block_info(block);
1560 assignments = block_info->assignments;
1562 ir_nodeset_init(&live_nodes);
1564 /* gather regalloc infos of predecessor blocks */
1565 n_preds = get_Block_n_cfgpreds(block);
1566 pred_block_infos = ALLOCAN(block_info_t*, n_preds);
1567 for (i = 0; i < n_preds; ++i) {
1568 ir_node *pred = get_Block_cfgpred_block(block, i);
1569 block_info_t *pred_info = get_block_info(pred);
1570 pred_block_infos[i] = pred_info;
1573 phi_ins = ALLOCAN(ir_node*, n_preds);
1575 /* collect live-in nodes and preassigned values */
1576 be_lv_foreach(lv, block, be_lv_state_in, i) {
1577 bool need_phi = false;
1578 const arch_register_req_t *req;
1579 const arch_register_t *reg;
1582 node = be_lv_get_irn(lv, block, i);
1583 req = arch_get_register_req_out(node);
1584 if (req->cls != cls)
1587 if (req->type & arch_register_req_type_ignore) {
1588 allocation_info_t *info = get_allocation_info(node);
1589 info->current_value = node;
1591 reg = arch_get_irn_register(node);
1592 assert(reg != NULL); /* ignore values must be preassigned */
1597 /* check all predecessors for this value, if it is not everywhere the
1598 same or unknown then we have to construct a phi
1599 (we collect the potential phi inputs here) */
1600 for (p = 0; p < n_preds; ++p) {
1601 block_info_t *pred_info = pred_block_infos[p];
1603 if (!pred_info->processed) {
1604 /* use node for now, it will get fixed later */
1608 int a = find_value_in_block_info(pred_info, node);
1610 /* must live out of predecessor */
1612 phi_ins[p] = pred_info->assignments[a];
1613 /* different value from last time? then we need a phi */
1614 if (p > 0 && phi_ins[p-1] != phi_ins[p]) {
1621 ir_mode *mode = get_irn_mode(node);
1622 const arch_register_req_t *req = get_default_req_current_cls();
1625 phi = new_r_Phi(block, n_preds, phi_ins, mode);
1626 be_set_phi_reg_req(phi, req);
1628 DB((dbg, LEVEL_3, "Create Phi %+F (for %+F) -", phi, node));
1629 #ifdef DEBUG_libfirm
1632 for (i = 0; i < n_preds; ++i) {
1633 DB((dbg, LEVEL_3, " %+F", phi_ins[i]));
1635 DB((dbg, LEVEL_3, "\n"));
1638 mark_as_copy_of(phi, node);
1639 sched_add_after(block, phi);
1643 allocation_info_t *info = get_allocation_info(node);
1644 info->current_value = phi_ins[0];
1646 /* Grab 1 of the inputs we constructed (might not be the same as
1647 * "node" as we could see the same copy of the value in all
1652 /* if the node already has a register assigned use it */
1653 reg = arch_get_irn_register(node);
1658 /* remember that this node is live at the beginning of the block */
1659 ir_nodeset_insert(&live_nodes, node);
1662 rbitset_alloca(forbidden_regs, n_regs);
1664 /* handle phis... */
1665 assign_phi_registers(block);
1667 /* all live-ins must have a register */
1668 #ifdef DEBUG_libfirm
1670 ir_nodeset_iterator_t iter;
1671 foreach_ir_nodeset(&live_nodes, node, iter) {
1672 const arch_register_t *reg = arch_get_irn_register(node);
1673 assert(reg != NULL);
1678 /* assign instructions in the block */
1679 sched_foreach(block, node) {
1684 /* phis are already assigned */
1688 rewire_inputs(node);
1690 /* enforce use constraints */
1691 rbitset_clear_all(forbidden_regs, n_regs);
1692 enforce_constraints(&live_nodes, node, forbidden_regs);
1694 rewire_inputs(node);
1696 /* we may not use registers used for inputs for optimistic splits */
1697 arity = get_irn_arity(node);
1698 for (i = 0; i < arity; ++i) {
1699 ir_node *op = get_irn_n(node, i);
1700 const arch_register_t *reg;
1701 if (!arch_irn_consider_in_reg_alloc(cls, op))
1704 reg = arch_get_irn_register(op);
1705 rbitset_set(forbidden_regs, arch_register_get_index(reg));
1708 /* free registers of values last used at this instruction */
1709 free_last_uses(&live_nodes, node);
1711 /* assign output registers */
1712 be_foreach_definition_(node, cls, value,
1713 assign_reg(block, value, forbidden_regs);
1717 ir_nodeset_destroy(&live_nodes);
1720 block_info->processed = true;
1722 /* permute values at end of predecessor blocks in case of phi-nodes */
1725 for (p = 0; p < n_preds; ++p) {
1726 add_phi_permutations(block, p);
1730 /* if we have exactly 1 successor then we might be able to produce phi
1732 if (get_irn_n_edges_kind(block, EDGE_KIND_BLOCK) == 1) {
1733 const ir_edge_t *edge
1734 = get_irn_out_edge_first_kind(block, EDGE_KIND_BLOCK);
1735 ir_node *succ = get_edge_src_irn(edge);
1736 int p = get_edge_src_pos(edge);
1737 block_info_t *succ_info = get_block_info(succ);
1739 if (succ_info->processed) {
1740 add_phi_permutations(succ, p);
1745 typedef struct block_costs_t block_costs_t;
1746 struct block_costs_t {
1747 float costs; /**< costs of the block */
1748 int dfs_num; /**< depth first search number (to detect backedges) */
1751 static int cmp_block_costs(const void *d1, const void *d2)
1753 const ir_node * const *block1 = d1;
1754 const ir_node * const *block2 = d2;
1755 const block_costs_t *info1 = get_irn_link(*block1);
1756 const block_costs_t *info2 = get_irn_link(*block2);
1757 return QSORT_CMP(info2->costs, info1->costs);
1760 static void determine_block_order(void)
1763 ir_node **blocklist = be_get_cfgpostorder(irg);
1764 int n_blocks = ARR_LEN(blocklist);
1766 pdeq *worklist = new_pdeq();
1767 ir_node **order = XMALLOCN(ir_node*, n_blocks);
1770 /* clear block links... */
1771 for (i = 0; i < n_blocks; ++i) {
1772 ir_node *block = blocklist[i];
1773 set_irn_link(block, NULL);
1776 /* walk blocks in reverse postorder, the costs for each block are the
1777 * sum of the costs of its predecessors (excluding the costs on backedges
1778 * which we can't determine) */
1779 for (i = n_blocks-1; i >= 0; --i) {
1780 block_costs_t *cost_info;
1781 ir_node *block = blocklist[i];
1783 float execfreq = (float)get_block_execfreq(execfreqs, block);
1784 float costs = execfreq;
1785 int n_cfgpreds = get_Block_n_cfgpreds(block);
1787 for (p = 0; p < n_cfgpreds; ++p) {
1788 ir_node *pred_block = get_Block_cfgpred_block(block, p);
1789 block_costs_t *pred_costs = get_irn_link(pred_block);
1790 /* we don't have any info for backedges */
1791 if (pred_costs == NULL)
1793 costs += pred_costs->costs;
1796 cost_info = OALLOCZ(&obst, block_costs_t);
1797 cost_info->costs = costs;
1798 cost_info->dfs_num = dfs_num++;
1799 set_irn_link(block, cost_info);
1802 /* sort array by block costs */
1803 qsort(blocklist, n_blocks, sizeof(blocklist[0]), cmp_block_costs);
1805 ir_reserve_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1806 inc_irg_block_visited(irg);
1808 for (i = 0; i < n_blocks; ++i) {
1809 ir_node *block = blocklist[i];
1810 if (Block_block_visited(block))
1813 /* continually add predecessors with highest costs to worklist
1814 * (without using backedges) */
1816 block_costs_t *info = get_irn_link(block);
1817 ir_node *best_pred = NULL;
1818 float best_costs = -1;
1819 int n_cfgpred = get_Block_n_cfgpreds(block);
1822 pdeq_putr(worklist, block);
1823 mark_Block_block_visited(block);
1824 for (i = 0; i < n_cfgpred; ++i) {
1825 ir_node *pred_block = get_Block_cfgpred_block(block, i);
1826 block_costs_t *pred_info = get_irn_link(pred_block);
1828 /* ignore backedges */
1829 if (pred_info->dfs_num > info->dfs_num)
1832 if (info->costs > best_costs) {
1833 best_costs = info->costs;
1834 best_pred = pred_block;
1838 } while (block != NULL && !Block_block_visited(block));
1840 /* now put all nodes in the worklist in our final order */
1841 while (!pdeq_empty(worklist)) {
1842 ir_node *pblock = pdeq_getr(worklist);
1843 assert(order_p < n_blocks);
1844 order[order_p++] = pblock;
1847 assert(order_p == n_blocks);
1850 ir_free_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1852 DEL_ARR_F(blocklist);
1854 obstack_free(&obst, NULL);
1855 obstack_init(&obst);
1857 block_order = order;
1858 n_block_order = n_blocks;
1862 * Run the register allocator for the current register class.
1864 static void be_pref_alloc_cls(void)
1868 lv = be_assure_liveness(irg);
1869 be_liveness_assure_sets(lv);
1871 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
1873 DB((dbg, LEVEL_2, "=== Allocating registers of %s ===\n", cls->name));
1875 be_clear_links(irg);
1877 irg_block_walk_graph(irg, NULL, analyze_block, NULL);
1878 if (create_congruence_classes)
1879 combine_congruence_classes();
1881 for (i = 0; i < n_block_order; ++i) {
1882 ir_node *block = block_order[i];
1883 allocate_coalesce_block(block, NULL);
1886 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
1889 static void dump(int mask, ir_graph *irg, const char *suffix)
1891 if (be_get_irg_options(irg)->dump_flags & mask)
1892 dump_ir_graph(irg, suffix);
1896 * Run the spiller on the current graph.
1898 static void spill(void)
1900 /* make sure all nodes show their real register pressure */
1901 be_timer_push(T_RA_CONSTR);
1902 be_pre_spill_prepare_constr(irg, cls);
1903 be_timer_pop(T_RA_CONSTR);
1905 dump(DUMP_RA, irg, "-spillprepare");
1908 be_timer_push(T_RA_SPILL);
1909 be_do_spill(irg, cls);
1910 be_timer_pop(T_RA_SPILL);
1912 be_timer_push(T_RA_SPILL_APPLY);
1913 check_for_memory_operands(irg);
1914 be_timer_pop(T_RA_SPILL_APPLY);
1916 dump(DUMP_RA, irg, "-spill");
1920 * The pref register allocator for a whole procedure.
1922 static void be_pref_alloc(ir_graph *new_irg)
1924 const arch_env_t *arch_env = be_get_irg_arch_env(new_irg);
1925 int n_cls = arch_env_get_n_reg_class(arch_env);
1928 obstack_init(&obst);
1931 execfreqs = be_get_irg_exec_freq(irg);
1933 /* determine a good coloring order */
1934 determine_block_order();
1936 for (c = 0; c < n_cls; ++c) {
1937 cls = arch_env_get_reg_class(arch_env, c);
1938 default_cls_req = NULL;
1939 if (arch_register_class_flags(cls) & arch_register_class_flag_manual_ra)
1942 stat_ev_ctx_push_str("regcls", cls->name);
1944 n_regs = arch_register_class_n_regs(cls);
1945 normal_regs = rbitset_malloc(n_regs);
1946 be_abi_set_non_ignore_regs(be_get_irg_abi(irg), cls, normal_regs);
1950 /* verify schedule and register pressure */
1951 be_timer_push(T_VERIFY);
1952 if (be_get_irg_options(irg)->verify_option == BE_VERIFY_WARN) {
1953 be_verify_schedule(irg);
1954 be_verify_register_pressure(irg, cls);
1955 } else if (be_get_irg_options(irg)->verify_option == BE_VERIFY_ASSERT) {
1956 assert(be_verify_schedule(irg) && "Schedule verification failed");
1957 assert(be_verify_register_pressure(irg, cls)
1958 && "Register pressure verification failed");
1960 be_timer_pop(T_VERIFY);
1962 be_timer_push(T_RA_COLOR);
1963 be_pref_alloc_cls();
1964 be_timer_pop(T_RA_COLOR);
1966 /* we most probably constructed new Phis so liveness info is invalid
1968 /* TODO: test liveness_introduce */
1969 be_liveness_invalidate(lv);
1972 stat_ev_ctx_pop("regcls");
1975 be_timer_push(T_RA_SPILL_APPLY);
1976 be_abi_fix_stack_nodes(irg);
1977 be_timer_pop(T_RA_SPILL_APPLY);
1979 be_timer_push(T_VERIFY);
1980 if (be_get_irg_options(irg)->verify_option == BE_VERIFY_WARN) {
1981 be_verify_register_allocation(irg);
1982 } else if (be_get_irg_options(irg)->verify_option == BE_VERIFY_ASSERT) {
1983 assert(be_verify_register_allocation(irg)
1984 && "Register allocation invalid");
1986 be_timer_pop(T_VERIFY);
1988 obstack_free(&obst, NULL);
1991 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_pref_alloc);
1992 void be_init_pref_alloc(void)
1994 static be_ra_t be_ra_pref = {
1997 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
1998 lc_opt_entry_t *prefalloc_group = lc_opt_get_grp(be_grp, "prefalloc");
1999 lc_opt_add_table(prefalloc_group, options);
2001 be_register_allocator("pref", &be_ra_pref);
2002 FIRM_DBG_REGISTER(dbg, "firm.be.prefalloc");