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;
87 static be_irg_t *birg;
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; /**< 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;
189 default_cls_req = req;
191 return default_cls_req;
195 * Link the allocation info of a node to a copy.
196 * Afterwards, both nodes uses the same allocation info.
197 * Copy must not have an allocation info assigned yet.
199 * @param copy the node that gets the allocation info assigned
200 * @param value the original node
202 static void mark_as_copy_of(ir_node *copy, ir_node *value)
205 allocation_info_t *info = get_allocation_info(value);
206 allocation_info_t *copy_info = get_allocation_info(copy);
208 /* find original value */
209 original = info->original_value;
210 if (original != value) {
211 info = get_allocation_info(original);
214 assert(info->original_value == original);
215 info->current_value = copy;
217 /* the copy should not be linked to something else yet */
218 assert(copy_info->original_value == copy);
219 copy_info->original_value = original;
221 /* copy over allocation preferences */
222 memcpy(copy_info->prefs, info->prefs, n_regs * sizeof(copy_info->prefs[0]));
226 * Calculate the penalties for every register on a node and its live neighbors.
228 * @param live_nodes the set of live nodes at the current position, may be NULL
229 * @param penalty the penalty to subtract from
230 * @param limited a raw bitset containing the limited set for the node
231 * @param node the node
233 static void give_penalties_for_limits(const ir_nodeset_t *live_nodes,
234 float penalty, const unsigned* limited,
237 ir_nodeset_iterator_t iter;
240 allocation_info_t *info = get_allocation_info(node);
243 /* give penalty for all forbidden regs */
244 for (r = 0; r < n_regs; ++r) {
245 if (rbitset_is_set(limited, r))
248 info->prefs[r] -= penalty;
251 /* all other live values should get a penalty for allowed regs */
252 if (live_nodes == NULL)
255 penalty *= NEIGHBOR_FACTOR;
256 n_allowed = rbitset_popcount(limited, n_regs);
258 /* only create a very weak penalty if multiple regs are allowed */
259 penalty = (penalty * 0.8f) / n_allowed;
261 foreach_ir_nodeset(live_nodes, neighbor, iter) {
262 allocation_info_t *neighbor_info;
264 /* TODO: if op is used on multiple inputs we might not do a
266 if (neighbor == node)
269 neighbor_info = get_allocation_info(neighbor);
270 for (r = 0; r < n_regs; ++r) {
271 if (!rbitset_is_set(limited, r))
274 neighbor_info->prefs[r] -= penalty;
280 * Calculate the preferences of a definition for the current register class.
281 * If the definition uses a limited set of registers, reduce the preferences
282 * for the limited register on the node and its neighbors.
284 * @param live_nodes the set of live nodes at the current node
285 * @param weight the weight
286 * @param node the current node
288 static void check_defs(const ir_nodeset_t *live_nodes, float weight,
291 const arch_register_req_t *req;
293 if (get_irn_mode(node) == mode_T) {
294 const ir_edge_t *edge;
295 foreach_out_edge(node, edge) {
296 ir_node *proj = get_edge_src_irn(edge);
297 check_defs(live_nodes, weight, proj);
302 if (!arch_irn_consider_in_reg_alloc(cls, node))
305 req = arch_get_register_req_out(node);
306 if (req->type & arch_register_req_type_limited) {
307 const unsigned *limited = req->limited;
308 float penalty = weight * DEF_FACTOR;
309 give_penalties_for_limits(live_nodes, penalty, limited, node);
312 if (req->type & arch_register_req_type_should_be_same) {
313 ir_node *insn = skip_Proj(node);
314 allocation_info_t *info = get_allocation_info(node);
315 int arity = get_irn_arity(insn);
318 float factor = 1.0f / rbitset_popcount(&req->other_same, arity);
319 for (i = 0; i < arity; ++i) {
322 allocation_info_t *op_info;
324 if (!rbitset_is_set(&req->other_same, i))
327 op = get_irn_n(insn, i);
329 /* if we the value at the should_be_same input doesn't die at the
330 * node, then it is no use to propagate the constraints (since a
331 * copy will emerge anyway) */
332 if (ir_nodeset_contains(live_nodes, op))
335 op_info = get_allocation_info(op);
336 for (r = 0; r < n_regs; ++r) {
337 op_info->prefs[r] += info->prefs[r] * factor;
344 * Walker: Runs an a block calculates the preferences for any
345 * node and every register from the considered register class.
347 static void analyze_block(ir_node *block, void *data)
349 float weight = (float)get_block_execfreq(execfreqs, block);
350 ir_nodeset_t live_nodes;
354 ir_nodeset_init(&live_nodes);
355 be_liveness_end_of_block(lv, cls, block, &live_nodes);
357 sched_foreach_reverse(block, node) {
358 allocation_info_t *info;
365 if (create_preferences)
366 check_defs(&live_nodes, weight, node);
369 arity = get_irn_arity(node);
371 /* the allocation info node currently only uses 1 unsigned value
372 to mark last used inputs. So we will fail for a node with more than
374 if (arity >= (int) sizeof(unsigned) * 8) {
375 panic("Node with more than %d inputs not supported yet",
376 (int) sizeof(unsigned) * 8);
379 info = get_allocation_info(node);
380 for (i = 0; i < arity; ++i) {
381 ir_node *op = get_irn_n(node, i);
382 if (!arch_irn_consider_in_reg_alloc(cls, op))
385 /* last usage of a value? */
386 if (!ir_nodeset_contains(&live_nodes, op)) {
387 rbitset_set(&info->last_uses, i);
391 be_liveness_transfer(cls, node, &live_nodes);
393 if (create_preferences) {
394 /* update weights based on usage constraints */
395 for (i = 0; i < arity; ++i) {
396 const arch_register_req_t *req;
397 const unsigned *limited;
398 ir_node *op = get_irn_n(node, i);
400 if (!arch_irn_consider_in_reg_alloc(cls, op))
403 req = arch_get_register_req(node, i);
404 if (!(req->type & arch_register_req_type_limited))
407 limited = req->limited;
408 give_penalties_for_limits(&live_nodes, weight * USE_FACTOR,
414 ir_nodeset_destroy(&live_nodes);
417 static void congruence_def(ir_nodeset_t *live_nodes, ir_node *node)
419 const arch_register_req_t *req;
421 if (get_irn_mode(node) == mode_T) {
422 const ir_edge_t *edge;
423 foreach_out_edge(node, edge) {
424 ir_node *def = get_edge_src_irn(edge);
425 congruence_def(live_nodes, def);
430 if (!arch_irn_consider_in_reg_alloc(cls, node))
433 /* should be same constraint? */
434 req = arch_get_register_req_out(node);
435 if (req->type & arch_register_req_type_should_be_same) {
436 ir_node *insn = skip_Proj(node);
437 int arity = get_irn_arity(insn);
439 unsigned node_idx = get_irn_idx(node);
440 node_idx = uf_find(congruence_classes, node_idx);
442 for (i = 0; i < arity; ++i) {
446 ir_nodeset_iterator_t iter;
447 bool interferes = false;
449 if (!rbitset_is_set(&req->other_same, i))
452 op = get_irn_n(insn, i);
453 op_idx = get_irn_idx(op);
454 op_idx = uf_find(congruence_classes, op_idx);
456 /* do we interfere with the value */
457 foreach_ir_nodeset(live_nodes, live, iter) {
458 int lv_idx = get_irn_idx(live);
459 lv_idx = uf_find(congruence_classes, lv_idx);
460 if (lv_idx == op_idx) {
465 /* don't put in same affinity class if we interfere */
469 node_idx = uf_union(congruence_classes, node_idx, op_idx);
470 DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
472 /* one should_be_same is enough... */
478 static void create_congruence_class(ir_node *block, void *data)
480 ir_nodeset_t live_nodes;
484 ir_nodeset_init(&live_nodes);
485 be_liveness_end_of_block(lv, cls, block, &live_nodes);
487 /* check should be same constraints */
488 sched_foreach_reverse(block, node) {
492 congruence_def(&live_nodes, node);
493 be_liveness_transfer(cls, node, &live_nodes);
496 /* check phi congruence classes */
497 sched_foreach_reverse_from(node, node) {
501 assert(is_Phi(node));
503 if (!arch_irn_consider_in_reg_alloc(cls, node))
506 node_idx = get_irn_idx(node);
507 node_idx = uf_find(congruence_classes, node_idx);
509 arity = get_irn_arity(node);
510 for (i = 0; i < arity; ++i) {
511 bool interferes = false;
512 ir_nodeset_iterator_t iter;
517 allocation_info_t *head_info;
518 allocation_info_t *other_info;
519 ir_node *op = get_Phi_pred(node, i);
520 int op_idx = get_irn_idx(op);
521 op_idx = uf_find(congruence_classes, op_idx);
523 /* do we interfere with the value */
524 foreach_ir_nodeset(&live_nodes, live, iter) {
525 int lv_idx = get_irn_idx(live);
526 lv_idx = uf_find(congruence_classes, lv_idx);
527 if (lv_idx == op_idx) {
532 /* don't put in same affinity class if we interfere */
535 /* any other phi has the same input? */
536 sched_foreach(block, phi) {
541 if (!arch_irn_consider_in_reg_alloc(cls, phi))
543 oop = get_Phi_pred(phi, i);
546 oop_idx = get_irn_idx(oop);
547 oop_idx = uf_find(congruence_classes, oop_idx);
548 if (oop_idx == op_idx) {
556 /* merge the 2 congruence classes and sum up their preferences */
557 old_node_idx = node_idx;
558 node_idx = uf_union(congruence_classes, node_idx, op_idx);
559 DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
562 old_node_idx = node_idx == old_node_idx ? op_idx : old_node_idx;
563 head_info = get_allocation_info(get_idx_irn(irg, node_idx));
564 other_info = get_allocation_info(get_idx_irn(irg, old_node_idx));
565 for (r = 0; r < n_regs; ++r) {
566 head_info->prefs[r] += other_info->prefs[r];
572 static void set_congruence_prefs(ir_node *node, void *data)
574 allocation_info_t *info;
575 allocation_info_t *head_info;
576 unsigned node_idx = get_irn_idx(node);
577 unsigned node_set = uf_find(congruence_classes, node_idx);
581 /* head of congruence class or not in any class */
582 if (node_set == node_idx)
585 if (!arch_irn_consider_in_reg_alloc(cls, node))
588 head_info = get_allocation_info(get_idx_irn(irg, node_set));
589 info = get_allocation_info(node);
591 memcpy(info->prefs, head_info->prefs, n_regs * sizeof(info->prefs[0]));
594 static void combine_congruence_classes(void)
596 size_t n = get_irg_last_idx(irg);
597 congruence_classes = XMALLOCN(int, n);
598 uf_init(congruence_classes, n);
600 /* create congruence classes */
601 irg_block_walk_graph(irg, create_congruence_class, NULL, NULL);
602 /* merge preferences */
603 irg_walk_graph(irg, set_congruence_prefs, NULL, NULL);
604 free(congruence_classes);
612 * Assign register reg to the given node.
614 * @param node the node
615 * @param reg the register
617 static void use_reg(ir_node *node, const arch_register_t *reg)
619 unsigned r = arch_register_get_index(reg);
620 assignments[r] = node;
621 arch_set_irn_register(node, reg);
624 static void free_reg_of_value(ir_node *node)
626 const arch_register_t *reg;
629 if (!arch_irn_consider_in_reg_alloc(cls, node))
632 reg = arch_get_irn_register(node);
633 r = arch_register_get_index(reg);
634 /* assignment->value may be NULL if a value is used at 2 inputs
635 so it gets freed twice. */
636 assert(assignments[r] == node || assignments[r] == NULL);
637 assignments[r] = NULL;
641 * Compare two register preferences in decreasing order.
643 static int compare_reg_pref(const void *e1, const void *e2)
645 const reg_pref_t *rp1 = (const reg_pref_t*) e1;
646 const reg_pref_t *rp2 = (const reg_pref_t*) e2;
647 if (rp1->pref < rp2->pref)
649 if (rp1->pref > rp2->pref)
654 static void fill_sort_candidates(reg_pref_t *regprefs,
655 const allocation_info_t *info)
659 for (r = 0; r < n_regs; ++r) {
660 float pref = info->prefs[r];
662 regprefs[r].pref = pref;
664 /* TODO: use a stable sort here to avoid unnecessary register jumping */
665 qsort(regprefs, n_regs, sizeof(regprefs[0]), compare_reg_pref);
668 static bool try_optimistic_split(ir_node *to_split, ir_node *before,
669 float pref, float pref_delta,
670 unsigned *forbidden_regs, int recursion)
672 const arch_register_t *from_reg;
673 const arch_register_t *reg;
674 ir_node *original_insn;
680 allocation_info_t *info = get_allocation_info(to_split);
683 float split_threshold;
687 /* stupid hack: don't optimisticallt split don't spill nodes...
688 * (so we don't split away the values produced because of
689 * must_be_different constraints) */
690 original_insn = skip_Proj(info->original_value);
691 if (arch_irn_get_flags(original_insn) & arch_irn_flags_dont_spill)
694 from_reg = arch_get_irn_register(to_split);
695 from_r = arch_register_get_index(from_reg);
696 block = get_nodes_block(before);
697 split_threshold = (float)get_block_execfreq(execfreqs, block) * SPLIT_DELTA;
699 if (pref_delta < split_threshold*0.5)
702 /* find the best free position where we could move to */
703 prefs = ALLOCAN(reg_pref_t, n_regs);
704 fill_sort_candidates(prefs, info);
705 for (i = 0; i < n_regs; ++i) {
709 bool old_source_state;
711 /* we need a normal register which is not an output register
712 an different from the current register of to_split */
714 if (!rbitset_is_set(normal_regs, r))
716 if (rbitset_is_set(forbidden_regs, r))
721 /* is the split worth it? */
722 delta = pref_delta + prefs[i].pref;
723 if (delta < split_threshold) {
724 DB((dbg, LEVEL_3, "Not doing optimistical split of %+F (depth %d), win %f too low\n",
725 to_split, recursion, delta));
729 /* if the register is free then we can do the split */
730 if (assignments[r] == NULL)
733 /* otherwise we might try recursively calling optimistic_split */
734 if (recursion+1 > MAX_OPTIMISTIC_SPLIT_RECURSION)
737 apref = prefs[i].pref;
738 apref_delta = i+1 < n_regs ? apref - prefs[i+1].pref : 0;
739 apref_delta += pref_delta - split_threshold;
741 /* our source register isn't a useful destination for recursive
743 old_source_state = rbitset_is_set(forbidden_regs, from_r);
744 rbitset_set(forbidden_regs, from_r);
745 /* try recursive split */
746 res = try_optimistic_split(assignments[r], before, apref,
747 apref_delta, forbidden_regs, recursion+1);
748 /* restore our destination */
749 if (old_source_state) {
750 rbitset_set(forbidden_regs, from_r);
752 rbitset_clear(forbidden_regs, from_r);
761 reg = arch_register_for_index(cls, r);
762 copy = be_new_Copy(cls, block, to_split);
763 mark_as_copy_of(copy, to_split);
764 /* hacky, but correct here */
765 if (assignments[arch_register_get_index(from_reg)] == to_split)
766 free_reg_of_value(to_split);
768 sched_add_before(before, copy);
771 "Optimistic live-range split %+F move %+F(%s) -> %s before %+F (win %f, depth %d)\n",
772 copy, to_split, from_reg->name, reg->name, before, delta, recursion));
777 * Determine and assign a register for node @p node
779 static void assign_reg(const ir_node *block, ir_node *node,
780 unsigned *forbidden_regs)
782 const arch_register_t *reg;
783 allocation_info_t *info;
784 const arch_register_req_t *req;
785 reg_pref_t *reg_prefs;
788 const unsigned *allowed_regs;
791 assert(!is_Phi(node));
792 assert(arch_irn_consider_in_reg_alloc(cls, node));
794 /* preassigned register? */
795 reg = arch_get_irn_register(node);
797 DB((dbg, LEVEL_2, "Preassignment %+F -> %s\n", node, reg->name));
802 /* give should_be_same boni */
803 info = get_allocation_info(node);
804 req = arch_get_register_req_out(node);
806 in_node = skip_Proj(node);
807 if (req->type & arch_register_req_type_should_be_same) {
808 float weight = (float)get_block_execfreq(execfreqs, block);
809 int arity = get_irn_arity(in_node);
812 assert(arity <= (int) sizeof(req->other_same) * 8);
813 for (i = 0; i < arity; ++i) {
815 const arch_register_t *reg;
817 if (!rbitset_is_set(&req->other_same, i))
820 in = get_irn_n(in_node, i);
821 reg = arch_get_irn_register(in);
823 r = arch_register_get_index(reg);
825 /* if the value didn't die here then we should not propagate the
826 * should_be_same info */
827 if (assignments[r] == in)
830 info->prefs[r] += weight * AFF_SHOULD_BE_SAME;
834 /* create list of register candidates and sort by their preference */
835 DB((dbg, LEVEL_2, "Candidates for %+F:", node));
836 reg_prefs = alloca(n_regs * sizeof(reg_prefs[0]));
837 fill_sort_candidates(reg_prefs, info);
838 for (i = 0; i < n_regs; ++i) {
839 unsigned num = reg_prefs[i].num;
840 const arch_register_t *reg;
842 if (!rbitset_is_set(normal_regs, num))
845 reg = arch_register_for_index(cls, num);
846 DB((dbg, LEVEL_2, " %s(%f)", reg->name, reg_prefs[i].pref));
848 DB((dbg, LEVEL_2, "\n"));
850 allowed_regs = normal_regs;
851 if (req->type & arch_register_req_type_limited) {
852 allowed_regs = req->limited;
855 for (i = 0; i < n_regs; ++i) {
860 r = reg_prefs[i].num;
861 if (!rbitset_is_set(allowed_regs, r))
863 if (assignments[r] == NULL)
865 pref = reg_prefs[i].pref;
866 delta = i+1 < n_regs ? pref - reg_prefs[i+1].pref : 0;
867 before = skip_Proj(node);
868 res = try_optimistic_split(assignments[r], before,
869 pref, delta, forbidden_regs, 0);
874 panic("No register left for %+F\n", node);
877 reg = arch_register_for_index(cls, r);
878 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
883 * Add an permutation in front of a node and change the assignments
884 * due to this permutation.
886 * To understand this imagine a permutation like this:
896 * First we count how many destinations a single value has. At the same time
897 * we can be sure that each destination register has at most 1 source register
898 * (it can have 0 which means we don't care what value is in it).
899 * We ignore all fullfilled permuations (like 7->7)
900 * In a first pass we create as much copy instructions as possible as they
901 * are generally cheaper than exchanges. We do this by counting into how many
902 * destinations a register has to be copied (in the example it's 2 for register
903 * 3, or 1 for the registers 1,2,4 and 7).
904 * We can then create a copy into every destination register when the usecount
905 * of that register is 0 (= noone else needs the value in the register).
907 * After this step we should have cycles left. We implement a cyclic permutation
908 * of n registers with n-1 transpositions.
910 * @param live_nodes the set of live nodes, updated due to live range split
911 * @param before the node before we add the permutation
912 * @param permutation the permutation array indices are the destination
913 * registers, the values in the array are the source
916 static void permute_values(ir_nodeset_t *live_nodes, ir_node *before,
917 unsigned *permutation)
919 unsigned *n_used = ALLOCANZ(unsigned, n_regs);
923 /* determine how often each source register needs to be read */
924 for (r = 0; r < n_regs; ++r) {
925 unsigned old_reg = permutation[r];
928 value = assignments[old_reg];
930 /* nothing to do here, reg is not live. Mark it as fixpoint
931 * so we ignore it in the next steps */
939 block = get_nodes_block(before);
941 /* step1: create copies where immediately possible */
942 for (r = 0; r < n_regs; /* empty */) {
945 const arch_register_t *reg;
946 unsigned old_r = permutation[r];
948 /* - no need to do anything for fixed points.
949 - we can't copy if the value in the dest reg is still needed */
950 if (old_r == r || n_used[r] > 0) {
956 src = assignments[old_r];
957 copy = be_new_Copy(cls, block, src);
958 sched_add_before(before, copy);
959 reg = arch_register_for_index(cls, r);
960 DB((dbg, LEVEL_2, "Copy %+F (from %+F, before %+F) -> %s\n",
961 copy, src, before, reg->name));
962 mark_as_copy_of(copy, src);
965 if (live_nodes != NULL) {
966 ir_nodeset_insert(live_nodes, copy);
969 /* old register has 1 user less, permutation is resolved */
970 assert(arch_register_get_index(arch_get_irn_register(src)) == old_r);
973 assert(n_used[old_r] > 0);
975 if (n_used[old_r] == 0) {
976 if (live_nodes != NULL) {
977 ir_nodeset_remove(live_nodes, src);
979 free_reg_of_value(src);
982 /* advance or jump back (if this copy enabled another copy) */
983 if (old_r < r && n_used[old_r] == 0) {
990 /* at this point we only have "cycles" left which we have to resolve with
992 * TODO: if we have free registers left, then we should really use copy
993 * instructions for any cycle longer than 2 registers...
994 * (this is probably architecture dependent, there might be archs where
995 * copies are preferable even for 2-cycles) */
997 /* create perms with the rest */
998 for (r = 0; r < n_regs; /* empty */) {
999 const arch_register_t *reg;
1000 unsigned old_r = permutation[r];
1012 /* we shouldn't have copies from 1 value to multiple destinations left*/
1013 assert(n_used[old_r] == 1);
1015 /* exchange old_r and r2; after that old_r is a fixed point */
1016 r2 = permutation[old_r];
1018 in[0] = assignments[r2];
1019 in[1] = assignments[old_r];
1020 perm = be_new_Perm(cls, block, 2, in);
1021 sched_add_before(before, perm);
1022 DB((dbg, LEVEL_2, "Perm %+F (perm %+F,%+F, before %+F)\n",
1023 perm, in[0], in[1], before));
1025 proj0 = new_r_Proj(perm, get_irn_mode(in[0]), 0);
1026 mark_as_copy_of(proj0, in[0]);
1027 reg = arch_register_for_index(cls, old_r);
1028 use_reg(proj0, reg);
1030 proj1 = new_r_Proj(perm, get_irn_mode(in[1]), 1);
1031 mark_as_copy_of(proj1, in[1]);
1032 reg = arch_register_for_index(cls, r2);
1033 use_reg(proj1, reg);
1035 /* 1 value is now in the correct register */
1036 permutation[old_r] = old_r;
1037 /* the source of r changed to r2 */
1038 permutation[r] = r2;
1040 /* if we have reached a fixpoint update data structures */
1041 if (live_nodes != NULL) {
1042 ir_nodeset_remove(live_nodes, in[0]);
1043 ir_nodeset_remove(live_nodes, in[1]);
1044 ir_nodeset_remove(live_nodes, proj0);
1045 ir_nodeset_insert(live_nodes, proj1);
1049 #ifdef DEBUG_libfirm
1050 /* now we should only have fixpoints left */
1051 for (r = 0; r < n_regs; ++r) {
1052 assert(permutation[r] == r);
1058 * Free regs for values last used.
1060 * @param live_nodes set of live nodes, will be updated
1061 * @param node the node to consider
1063 static void free_last_uses(ir_nodeset_t *live_nodes, ir_node *node)
1065 allocation_info_t *info = get_allocation_info(node);
1066 const unsigned *last_uses = &info->last_uses;
1067 int arity = get_irn_arity(node);
1070 for (i = 0; i < arity; ++i) {
1073 /* check if one operand is the last use */
1074 if (!rbitset_is_set(last_uses, i))
1077 op = get_irn_n(node, i);
1078 free_reg_of_value(op);
1079 ir_nodeset_remove(live_nodes, op);
1084 * change inputs of a node to the current value (copies/perms)
1086 static void rewire_inputs(ir_node *node)
1089 int arity = get_irn_arity(node);
1091 for (i = 0; i < arity; ++i) {
1092 ir_node *op = get_irn_n(node, i);
1093 allocation_info_t *info = try_get_allocation_info(op);
1098 info = get_allocation_info(info->original_value);
1099 if (info->current_value != op) {
1100 set_irn_n(node, i, info->current_value);
1106 * Create a bitset of registers occupied with value living through an
1109 static void determine_live_through_regs(unsigned *bitset, ir_node *node)
1111 const allocation_info_t *info = get_allocation_info(node);
1116 /* mark all used registers as potentially live-through */
1117 for (r = 0; r < n_regs; ++r) {
1118 if (assignments[r] == NULL)
1120 if (!rbitset_is_set(normal_regs, r))
1123 rbitset_set(bitset, r);
1126 /* remove registers of value dying at the instruction */
1127 arity = get_irn_arity(node);
1128 for (i = 0; i < arity; ++i) {
1130 const arch_register_t *reg;
1132 if (!rbitset_is_set(&info->last_uses, i))
1135 op = get_irn_n(node, i);
1136 reg = arch_get_irn_register(op);
1137 rbitset_clear(bitset, arch_register_get_index(reg));
1142 * Enforce constraints at a node by live range splits.
1144 * @param live_nodes the set of live nodes, might be changed
1145 * @param node the current node
1147 static void enforce_constraints(ir_nodeset_t *live_nodes, ir_node *node,
1148 unsigned *forbidden_regs)
1150 int arity = get_irn_arity(node);
1152 hungarian_problem_t *bp;
1154 unsigned *assignment;
1156 /* construct a list of register occupied by live-through values */
1157 unsigned *live_through_regs = NULL;
1159 /* see if any use constraints are not met */
1161 for (i = 0; i < arity; ++i) {
1162 ir_node *op = get_irn_n(node, i);
1163 const arch_register_t *reg;
1164 const arch_register_req_t *req;
1165 const unsigned *limited;
1168 if (!arch_irn_consider_in_reg_alloc(cls, op))
1171 /* are there any limitations for the i'th operand? */
1172 req = arch_get_register_req(node, i);
1173 if (!(req->type & arch_register_req_type_limited))
1176 limited = req->limited;
1177 reg = arch_get_irn_register(op);
1178 r = arch_register_get_index(reg);
1179 if (!rbitset_is_set(limited, r)) {
1180 /* found an assignment outside the limited set */
1186 /* is any of the live-throughs using a constrained output register? */
1187 if (get_irn_mode(node) == mode_T) {
1188 const ir_edge_t *edge;
1190 foreach_out_edge(node, edge) {
1191 ir_node *proj = get_edge_src_irn(edge);
1192 const arch_register_req_t *req;
1194 if (!arch_irn_consider_in_reg_alloc(cls, proj))
1197 req = arch_get_register_req_out(proj);
1198 if (!(req->type & arch_register_req_type_limited))
1201 if (live_through_regs == NULL) {
1202 rbitset_alloca(live_through_regs, n_regs);
1203 determine_live_through_regs(live_through_regs, node);
1206 rbitset_or(forbidden_regs, req->limited, n_regs);
1207 if (rbitsets_have_common(req->limited, live_through_regs, n_regs)) {
1212 if (arch_irn_consider_in_reg_alloc(cls, node)) {
1213 const arch_register_req_t *req = arch_get_register_req_out(node);
1214 if (req->type & arch_register_req_type_limited) {
1215 rbitset_alloca(live_through_regs, n_regs);
1216 determine_live_through_regs(live_through_regs, node);
1217 if (rbitsets_have_common(req->limited, live_through_regs, n_regs)) {
1219 rbitset_or(forbidden_regs, req->limited, n_regs);
1228 /* create these arrays if we haven't yet */
1229 if (live_through_regs == NULL) {
1230 rbitset_alloca(live_through_regs, n_regs);
1233 /* at this point we have to construct a bipartite matching problem to see
1234 * which values should go to which registers
1235 * Note: We're building the matrix in "reverse" - source registers are
1236 * right, destinations left because this will produce the solution
1237 * in the format required for permute_values.
1239 bp = hungarian_new(n_regs, n_regs, HUNGARIAN_MATCH_PERFECT);
1241 /* add all combinations, then remove not allowed ones */
1242 for (l = 0; l < n_regs; ++l) {
1243 if (!rbitset_is_set(normal_regs, l)) {
1244 hungarian_add(bp, l, l, 1);
1248 for (r = 0; r < n_regs; ++r) {
1249 if (!rbitset_is_set(normal_regs, r))
1251 /* livethrough values may not use constrainted output registers */
1252 if (rbitset_is_set(live_through_regs, l)
1253 && rbitset_is_set(forbidden_regs, r))
1256 hungarian_add(bp, r, l, l == r ? 9 : 8);
1260 for (i = 0; i < arity; ++i) {
1261 ir_node *op = get_irn_n(node, i);
1262 const arch_register_t *reg;
1263 const arch_register_req_t *req;
1264 const unsigned *limited;
1265 unsigned current_reg;
1267 if (!arch_irn_consider_in_reg_alloc(cls, op))
1270 req = arch_get_register_req(node, i);
1271 if (!(req->type & arch_register_req_type_limited))
1274 limited = req->limited;
1275 reg = arch_get_irn_register(op);
1276 current_reg = arch_register_get_index(reg);
1277 for (r = 0; r < n_regs; ++r) {
1278 if (rbitset_is_set(limited, r))
1280 hungarian_remv(bp, r, current_reg);
1284 //hungarian_print_cost_matrix(bp, 1);
1285 hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1287 assignment = ALLOCAN(unsigned, n_regs);
1288 res = hungarian_solve(bp, (int*) assignment, NULL, 0);
1292 fprintf(stderr, "Swap result:");
1293 for (i = 0; i < (int) n_regs; ++i) {
1294 fprintf(stderr, " %d", assignment[i]);
1296 fprintf(stderr, "\n");
1301 permute_values(live_nodes, node, assignment);
1304 /** test wether a node @p n is a copy of the value of node @p of */
1305 static bool is_copy_of(ir_node *value, ir_node *test_value)
1307 allocation_info_t *test_info;
1308 allocation_info_t *info;
1310 if (value == test_value)
1313 info = get_allocation_info(value);
1314 test_info = get_allocation_info(test_value);
1315 return test_info->original_value == info->original_value;
1319 * find a value in the end-assignment of a basic block
1320 * @returns the index into the assignment array if found
1323 static int find_value_in_block_info(block_info_t *info, ir_node *value)
1326 ir_node **assignments = info->assignments;
1327 for (r = 0; r < n_regs; ++r) {
1328 ir_node *a_value = assignments[r];
1330 if (a_value == NULL)
1332 if (is_copy_of(a_value, value))
1340 * Create the necessary permutations at the end of a basic block to fullfill
1341 * the register assignment for phi-nodes in the next block
1343 static void add_phi_permutations(ir_node *block, int p)
1346 unsigned *permutation;
1347 ir_node **old_assignments;
1348 bool need_permutation;
1350 ir_node *pred = get_Block_cfgpred_block(block, p);
1352 block_info_t *pred_info = get_block_info(pred);
1354 /* predecessor not processed yet? nothing to do */
1355 if (!pred_info->processed)
1358 permutation = ALLOCAN(unsigned, n_regs);
1359 for (r = 0; r < n_regs; ++r) {
1363 /* check phi nodes */
1364 need_permutation = false;
1365 node = sched_first(block);
1366 for ( ; is_Phi(node); node = sched_next(node)) {
1367 const arch_register_t *reg;
1372 if (!arch_irn_consider_in_reg_alloc(cls, node))
1375 op = get_Phi_pred(node, p);
1376 if (!arch_irn_consider_in_reg_alloc(cls, op))
1379 a = find_value_in_block_info(pred_info, op);
1382 reg = arch_get_irn_register(node);
1383 regn = arch_register_get_index(reg);
1385 permutation[regn] = a;
1386 need_permutation = true;
1390 if (need_permutation) {
1391 /* permute values at end of predecessor */
1392 old_assignments = assignments;
1393 assignments = pred_info->assignments;
1394 permute_values(NULL, be_get_end_of_block_insertion_point(pred),
1396 assignments = old_assignments;
1399 /* change phi nodes to use the copied values */
1400 node = sched_first(block);
1401 for ( ; is_Phi(node); node = sched_next(node)) {
1405 if (!arch_irn_consider_in_reg_alloc(cls, node))
1408 op = get_Phi_pred(node, p);
1409 /* no need to do anything for Unknown inputs */
1410 if (!arch_irn_consider_in_reg_alloc(cls, op))
1413 /* we have permuted all values into the correct registers so we can
1414 simply query which value occupies the phis register in the
1416 a = arch_register_get_index(arch_get_irn_register(node));
1417 op = pred_info->assignments[a];
1418 set_Phi_pred(node, p, op);
1423 * Set preferences for a phis register based on the registers used on the
1426 static void adapt_phi_prefs(ir_node *phi)
1429 int arity = get_irn_arity(phi);
1430 ir_node *block = get_nodes_block(phi);
1431 allocation_info_t *info = get_allocation_info(phi);
1433 for (i = 0; i < arity; ++i) {
1434 ir_node *op = get_irn_n(phi, i);
1435 const arch_register_t *reg = arch_get_irn_register(op);
1436 ir_node *pred_block;
1437 block_info_t *pred_block_info;
1443 /* we only give the bonus if the predecessor already has registers
1444 * assigned, otherwise we only see a dummy value
1445 * and any conclusions about its register are useless */
1446 pred_block = get_Block_cfgpred_block(block, i);
1447 pred_block_info = get_block_info(pred_block);
1448 if (!pred_block_info->processed)
1451 /* give bonus for already assigned register */
1452 weight = (float)get_block_execfreq(execfreqs, pred_block);
1453 r = arch_register_get_index(reg);
1454 info->prefs[r] += weight * AFF_PHI;
1459 * After a phi has been assigned a register propagate preference inputs
1460 * to the phi inputs.
1462 static void propagate_phi_register(ir_node *phi, unsigned assigned_r)
1465 ir_node *block = get_nodes_block(phi);
1466 int arity = get_irn_arity(phi);
1468 for (i = 0; i < arity; ++i) {
1469 ir_node *op = get_Phi_pred(phi, i);
1470 allocation_info_t *info = get_allocation_info(op);
1471 ir_node *pred_block = get_Block_cfgpred_block(block, i);
1474 = (float)get_block_execfreq(execfreqs, pred_block) * AFF_PHI;
1476 if (info->prefs[assigned_r] >= weight)
1479 /* promote the prefered register */
1480 for (r = 0; r < n_regs; ++r) {
1481 if (info->prefs[r] > -weight) {
1482 info->prefs[r] = -weight;
1485 info->prefs[assigned_r] = weight;
1488 propagate_phi_register(op, assigned_r);
1492 static void assign_phi_registers(ir_node *block)
1499 hungarian_problem_t *bp;
1501 /* count phi nodes */
1502 sched_foreach(block, node) {
1505 if (!arch_irn_consider_in_reg_alloc(cls, node))
1513 /* build a bipartite matching problem for all phi nodes */
1514 bp = hungarian_new(n_phis, n_regs, HUNGARIAN_MATCH_PERFECT);
1516 sched_foreach(block, node) {
1519 allocation_info_t *info;
1522 if (!arch_irn_consider_in_reg_alloc(cls, node))
1525 /* give boni for predecessor colorings */
1526 adapt_phi_prefs(node);
1527 /* add stuff to bipartite problem */
1528 info = get_allocation_info(node);
1529 DB((dbg, LEVEL_3, "Prefs for %+F: ", node));
1530 for (r = 0; r < n_regs; ++r) {
1533 if (!rbitset_is_set(normal_regs, r))
1536 costs = info->prefs[r];
1537 costs = costs < 0 ? -logf(-costs+1) : logf(costs+1);
1540 hungarian_add(bp, n, r, (int)costs);
1541 DB((dbg, LEVEL_3, " %s(%f)", arch_register_for_index(cls, r)->name,
1544 DB((dbg, LEVEL_3, "\n"));
1548 //hungarian_print_cost_matrix(bp, 7);
1549 hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1551 assignment = ALLOCAN(int, n_regs);
1552 res = hungarian_solve(bp, assignment, NULL, 0);
1557 sched_foreach(block, node) {
1559 const arch_register_t *reg;
1563 if (!arch_irn_consider_in_reg_alloc(cls, node))
1566 r = assignment[n++];
1567 assert(rbitset_is_set(normal_regs, r));
1568 reg = arch_register_for_index(cls, r);
1569 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
1572 /* adapt preferences for phi inputs */
1573 if (propagate_phi_registers)
1574 propagate_phi_register(node, r);
1579 * Walker: assign registers to all nodes of a block that
1580 * need registers from the currently considered register class.
1582 static void allocate_coalesce_block(ir_node *block, void *data)
1585 ir_nodeset_t live_nodes;
1588 block_info_t *block_info;
1589 block_info_t **pred_block_infos;
1591 unsigned *forbidden_regs; /**< collects registers which must
1592 not be used for optimistic splits */
1595 DB((dbg, LEVEL_2, "* Block %+F\n", block));
1597 /* clear assignments */
1598 block_info = get_block_info(block);
1599 assignments = block_info->assignments;
1601 ir_nodeset_init(&live_nodes);
1603 /* gather regalloc infos of predecessor blocks */
1604 n_preds = get_Block_n_cfgpreds(block);
1605 pred_block_infos = ALLOCAN(block_info_t*, n_preds);
1606 for (i = 0; i < n_preds; ++i) {
1607 ir_node *pred = get_Block_cfgpred_block(block, i);
1608 block_info_t *pred_info = get_block_info(pred);
1609 pred_block_infos[i] = pred_info;
1612 phi_ins = ALLOCAN(ir_node*, n_preds);
1614 /* collect live-in nodes and preassigned values */
1615 be_lv_foreach(lv, block, be_lv_state_in, i) {
1616 const arch_register_t *reg;
1618 bool need_phi = false;
1620 node = be_lv_get_irn(lv, block, i);
1621 if (!arch_irn_consider_in_reg_alloc(cls, node))
1624 /* check all predecessors for this value, if it is not everywhere the
1625 same or unknown then we have to construct a phi
1626 (we collect the potential phi inputs here) */
1627 for (p = 0; p < n_preds; ++p) {
1628 block_info_t *pred_info = pred_block_infos[p];
1630 if (!pred_info->processed) {
1631 /* use node for now, it will get fixed later */
1635 int a = find_value_in_block_info(pred_info, node);
1637 /* must live out of predecessor */
1639 phi_ins[p] = pred_info->assignments[a];
1640 /* different value from last time? then we need a phi */
1641 if (p > 0 && phi_ins[p-1] != phi_ins[p]) {
1648 ir_mode *mode = get_irn_mode(node);
1649 const arch_register_req_t *req = get_default_req_current_cls();
1652 phi = new_r_Phi(block, n_preds, phi_ins, mode);
1653 be_set_phi_reg_req(phi, req);
1655 DB((dbg, LEVEL_3, "Create Phi %+F (for %+F) -", phi, node));
1656 #ifdef DEBUG_libfirm
1659 for (i = 0; i < n_preds; ++i) {
1660 DB((dbg, LEVEL_3, " %+F", phi_ins[i]));
1662 DB((dbg, LEVEL_3, "\n"));
1665 mark_as_copy_of(phi, node);
1666 sched_add_after(block, phi);
1670 allocation_info_t *info = get_allocation_info(node);
1671 info->current_value = phi_ins[0];
1673 /* Grab 1 of the inputs we constructed (might not be the same as
1674 * "node" as we could see the same copy of the value in all
1679 /* if the node already has a register assigned use it */
1680 reg = arch_get_irn_register(node);
1685 /* remember that this node is live at the beginning of the block */
1686 ir_nodeset_insert(&live_nodes, node);
1689 rbitset_alloca(forbidden_regs, n_regs);
1691 /* handle phis... */
1692 assign_phi_registers(block);
1694 /* all live-ins must have a register */
1695 #ifdef DEBUG_libfirm
1697 ir_nodeset_iterator_t iter;
1698 foreach_ir_nodeset(&live_nodes, node, iter) {
1699 const arch_register_t *reg = arch_get_irn_register(node);
1700 assert(reg != NULL);
1705 /* assign instructions in the block */
1706 sched_foreach(block, node) {
1710 /* phis are already assigned */
1714 rewire_inputs(node);
1716 /* enforce use constraints */
1717 rbitset_clear_all(forbidden_regs, n_regs);
1718 enforce_constraints(&live_nodes, node, forbidden_regs);
1720 rewire_inputs(node);
1722 /* we may not use registers used for inputs for optimistic splits */
1723 arity = get_irn_arity(node);
1724 for (i = 0; i < arity; ++i) {
1725 ir_node *op = get_irn_n(node, i);
1726 const arch_register_t *reg;
1727 if (!arch_irn_consider_in_reg_alloc(cls, op))
1730 reg = arch_get_irn_register(op);
1731 rbitset_set(forbidden_regs, arch_register_get_index(reg));
1734 /* free registers of values last used at this instruction */
1735 free_last_uses(&live_nodes, node);
1737 /* assign output registers */
1738 /* TODO: 2 phases: first: pre-assigned ones, 2nd real regs */
1739 if (get_irn_mode(node) == mode_T) {
1740 const ir_edge_t *edge;
1741 foreach_out_edge(node, edge) {
1742 ir_node *proj = get_edge_src_irn(edge);
1743 if (!arch_irn_consider_in_reg_alloc(cls, proj))
1745 assign_reg(block, proj, forbidden_regs);
1747 } else if (arch_irn_consider_in_reg_alloc(cls, node)) {
1748 assign_reg(block, node, forbidden_regs);
1752 ir_nodeset_destroy(&live_nodes);
1755 block_info->processed = true;
1757 /* permute values at end of predecessor blocks in case of phi-nodes */
1760 for (p = 0; p < n_preds; ++p) {
1761 add_phi_permutations(block, p);
1765 /* if we have exactly 1 successor then we might be able to produce phi
1767 if (get_irn_n_edges_kind(block, EDGE_KIND_BLOCK) == 1) {
1768 const ir_edge_t *edge
1769 = get_irn_out_edge_first_kind(block, EDGE_KIND_BLOCK);
1770 ir_node *succ = get_edge_src_irn(edge);
1771 int p = get_edge_src_pos(edge);
1772 block_info_t *succ_info = get_block_info(succ);
1774 if (succ_info->processed) {
1775 add_phi_permutations(succ, p);
1780 typedef struct block_costs_t block_costs_t;
1781 struct block_costs_t {
1782 float costs; /**< costs of the block */
1783 int dfs_num; /**< depth first search number (to detect backedges) */
1786 static int cmp_block_costs(const void *d1, const void *d2)
1788 const ir_node * const *block1 = d1;
1789 const ir_node * const *block2 = d2;
1790 const block_costs_t *info1 = get_irn_link(*block1);
1791 const block_costs_t *info2 = get_irn_link(*block2);
1792 return QSORT_CMP(info2->costs, info1->costs);
1795 static void determine_block_order(void)
1798 ir_node **blocklist = be_get_cfgpostorder(irg);
1799 int n_blocks = ARR_LEN(blocklist);
1801 pdeq *worklist = new_pdeq();
1802 ir_node **order = XMALLOCN(ir_node*, n_blocks);
1805 /* clear block links... */
1806 for (i = 0; i < n_blocks; ++i) {
1807 ir_node *block = blocklist[i];
1808 set_irn_link(block, NULL);
1811 /* walk blocks in reverse postorder, the costs for each block are the
1812 * sum of the costs of its predecessors (excluding the costs on backedges
1813 * which we can't determine) */
1814 for (i = n_blocks-1; i >= 0; --i) {
1815 block_costs_t *cost_info;
1816 ir_node *block = blocklist[i];
1818 float execfreq = (float)get_block_execfreq(execfreqs, block);
1819 float costs = execfreq;
1820 int n_cfgpreds = get_Block_n_cfgpreds(block);
1822 for (p = 0; p < n_cfgpreds; ++p) {
1823 ir_node *pred_block = get_Block_cfgpred_block(block, p);
1824 block_costs_t *pred_costs = get_irn_link(pred_block);
1825 /* we don't have any info for backedges */
1826 if (pred_costs == NULL)
1828 costs += pred_costs->costs;
1831 cost_info = OALLOCZ(&obst, block_costs_t);
1832 cost_info->costs = costs;
1833 cost_info->dfs_num = dfs_num++;
1834 set_irn_link(block, cost_info);
1837 /* sort array by block costs */
1838 qsort(blocklist, n_blocks, sizeof(blocklist[0]), cmp_block_costs);
1840 ir_reserve_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1841 inc_irg_block_visited(irg);
1843 for (i = 0; i < n_blocks; ++i) {
1844 ir_node *block = blocklist[i];
1845 if (Block_block_visited(block))
1848 /* continually add predecessors with highest costs to worklist
1849 * (without using backedges) */
1851 block_costs_t *info = get_irn_link(block);
1852 ir_node *best_pred = NULL;
1853 float best_costs = -1;
1854 int n_cfgpred = get_Block_n_cfgpreds(block);
1857 pdeq_putr(worklist, block);
1858 mark_Block_block_visited(block);
1859 for (i = 0; i < n_cfgpred; ++i) {
1860 ir_node *pred_block = get_Block_cfgpred_block(block, i);
1861 block_costs_t *pred_info = get_irn_link(pred_block);
1863 /* ignore backedges */
1864 if (pred_info->dfs_num > info->dfs_num)
1867 if (info->costs > best_costs) {
1868 best_costs = info->costs;
1869 best_pred = pred_block;
1873 } while (block != NULL && !Block_block_visited(block));
1875 /* now put all nodes in the worklist in our final order */
1876 while (!pdeq_empty(worklist)) {
1877 ir_node *pblock = pdeq_getr(worklist);
1878 assert(order_p < n_blocks);
1879 order[order_p++] = pblock;
1882 assert(order_p == n_blocks);
1885 ir_free_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1887 DEL_ARR_F(blocklist);
1889 obstack_free(&obst, NULL);
1890 obstack_init(&obst);
1892 block_order = order;
1893 n_block_order = n_blocks;
1897 * Run the register allocator for the current register class.
1899 static void be_pref_alloc_cls(void)
1903 lv = be_assure_liveness(irg);
1904 be_liveness_assure_sets(lv);
1906 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
1908 DB((dbg, LEVEL_2, "=== Allocating registers of %s ===\n", cls->name));
1910 be_clear_links(irg);
1912 irg_block_walk_graph(irg, NULL, analyze_block, NULL);
1913 if (create_congruence_classes)
1914 combine_congruence_classes();
1916 for (i = 0; i < n_block_order; ++i) {
1917 ir_node *block = block_order[i];
1918 allocate_coalesce_block(block, NULL);
1921 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
1924 static void dump(int mask, ir_graph *irg, const char *suffix)
1926 if (birg->main_env->options->dump_flags & mask)
1927 dump_ir_graph(irg, suffix);
1931 * Run the spiller on the current graph.
1933 static void spill(void)
1935 /* make sure all nodes show their real register pressure */
1936 be_timer_push(T_RA_CONSTR);
1937 be_pre_spill_prepare_constr(birg, cls);
1938 be_timer_pop(T_RA_CONSTR);
1940 dump(DUMP_RA, irg, "-spillprepare");
1943 be_timer_push(T_RA_SPILL);
1944 be_do_spill(birg, cls);
1945 be_timer_pop(T_RA_SPILL);
1947 be_timer_push(T_RA_SPILL_APPLY);
1948 check_for_memory_operands(irg);
1949 be_timer_pop(T_RA_SPILL_APPLY);
1951 dump(DUMP_RA, irg, "-spill");
1955 * The pref register allocator for a whole procedure.
1957 static void be_pref_alloc(be_irg_t *new_birg)
1959 const arch_env_t *arch_env = new_birg->main_env->arch_env;
1960 int n_cls = arch_env_get_n_reg_class(arch_env);
1963 obstack_init(&obst);
1966 irg = be_get_birg_irg(birg);
1967 execfreqs = birg->exec_freq;
1969 /* determine a good coloring order */
1970 determine_block_order();
1972 for (c = 0; c < n_cls; ++c) {
1973 cls = arch_env_get_reg_class(arch_env, c);
1974 default_cls_req = NULL;
1975 if (arch_register_class_flags(cls) & arch_register_class_flag_manual_ra)
1978 stat_ev_ctx_push_str("regcls", cls->name);
1980 n_regs = arch_register_class_n_regs(cls);
1981 normal_regs = rbitset_malloc(n_regs);
1982 be_abi_set_non_ignore_regs(birg->abi, cls, normal_regs);
1986 /* verify schedule and register pressure */
1987 be_timer_push(T_VERIFY);
1988 if (birg->main_env->options->vrfy_option == BE_VRFY_WARN) {
1989 be_verify_schedule(birg);
1990 be_verify_register_pressure(birg, cls, irg);
1991 } else if (birg->main_env->options->vrfy_option == BE_VRFY_ASSERT) {
1992 assert(be_verify_schedule(birg) && "Schedule verification failed");
1993 assert(be_verify_register_pressure(birg, cls, irg)
1994 && "Register pressure verification failed");
1996 be_timer_pop(T_VERIFY);
1998 be_timer_push(T_RA_COLOR);
1999 be_pref_alloc_cls();
2000 be_timer_pop(T_RA_COLOR);
2002 /* we most probably constructed new Phis so liveness info is invalid
2004 /* TODO: test liveness_introduce */
2005 be_liveness_invalidate(lv);
2008 stat_ev_ctx_pop("regcls");
2011 be_timer_push(T_RA_SPILL_APPLY);
2012 be_abi_fix_stack_nodes(birg->abi);
2013 be_timer_pop(T_RA_SPILL_APPLY);
2015 be_timer_push(T_VERIFY);
2016 if (birg->main_env->options->vrfy_option == BE_VRFY_WARN) {
2017 be_verify_register_allocation(birg);
2018 } else if (birg->main_env->options->vrfy_option == BE_VRFY_ASSERT) {
2019 assert(be_verify_register_allocation(birg)
2020 && "Register allocation invalid");
2022 be_timer_pop(T_VERIFY);
2024 obstack_free(&obst, NULL);
2027 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_pref_alloc);
2028 void be_init_pref_alloc(void)
2030 static be_ra_t be_ra_pref = {
2033 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
2034 lc_opt_entry_t *prefalloc_group = lc_opt_get_grp(be_grp, "prefalloc");
2035 lc_opt_add_table(prefalloc_group, options);
2037 be_register_allocator("pref", &be_ra_pref);
2038 FIRM_DBG_REGISTER(dbg, "firm.be.prefalloc");