2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief New approach to allocation and copy coalescing
23 * @author Matthias Braun
27 * ... WE NEED A NAME FOR THIS ...
29 * Only a proof of concept at this moment...
31 * The idea is to allocate registers in 2 passes:
32 * 1. A first pass to determine "preferred" registers for live-ranges. This
33 * calculates for each register and each live-range a value indicating
34 * the usefulness. (You can roughly think of the value as the negative
35 * costs needed for copies when the value is in the specific registers...)
37 * 2. Walk blocks and assigns registers in a greedy fashion. Preferring
38 * registers with high preferences. When register constraints are not met,
39 * add copies and split live-ranges.
42 * - make use of free registers in the permute_values code
43 * - think about a smarter sequence of visiting the blocks. Sorted by
44 * execfreq might be good, or looptree from inner to outermost loops going
45 * over blocks in a reverse postorder
46 * - propagate preferences through Phis
58 #include "iredges_t.h"
59 #include "irgraph_t.h"
64 #include "raw_bitset.h"
65 #include "unionfind.h"
67 #include "hungarian.h"
70 #include "bechordal_t.h"
79 #include "bespillutil.h"
83 #define USE_FACTOR 1.0f
84 #define DEF_FACTOR 1.0f
85 #define NEIGHBOR_FACTOR 0.2f
86 #define AFF_SHOULD_BE_SAME 0.5f
88 #define SPLIT_DELTA 1.0f
89 #define MAX_OPTIMISTIC_SPLIT_RECURSION 0
91 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
93 static struct obstack obst;
94 static be_irg_t *birg;
96 static const arch_register_class_t *cls;
97 static const arch_register_req_t *default_cls_req;
99 static const ir_exec_freq *execfreqs;
100 static unsigned n_regs;
101 static unsigned *normal_regs;
102 static int *congruence_classes;
103 static ir_node **block_order;
104 static int n_block_order;
105 static int create_preferences = true;
106 static int create_congruence_classes = true;
107 static int propagate_phi_registers = true;
109 static const lc_opt_table_entry_t options[] = {
110 LC_OPT_ENT_BOOL("prefs", "use preference based coloring", &create_preferences),
111 LC_OPT_ENT_BOOL("congruences", "create congruence classes", &create_congruence_classes),
112 LC_OPT_ENT_BOOL("prop_phi", "propagate phi registers", &propagate_phi_registers),
116 /** currently active assignments (while processing a basic block)
117 * maps registers to values(their current copies) */
118 static ir_node **assignments;
121 * allocation information: last_uses, register preferences
122 * the information is per firm-node.
124 struct allocation_info_t {
125 unsigned last_uses; /**< bitset indicating last uses (input pos) */
126 ir_node *current_value; /**< copy of the value that should be used */
127 ir_node *original_value; /**< for copies point to original value */
128 float prefs[0]; /**< register preferences */
130 typedef struct allocation_info_t allocation_info_t;
132 /** helper datastructure used when sorting register preferences */
137 typedef struct reg_pref_t reg_pref_t;
139 /** per basic-block information */
140 struct block_info_t {
141 bool processed; /**< indicate wether block is processed */
142 ir_node *assignments[0]; /**< register assignments at end of block */
144 typedef struct block_info_t block_info_t;
147 * Get the allocation info for a node.
148 * The info is allocated on the first visit of a node.
150 static allocation_info_t *get_allocation_info(ir_node *node)
152 allocation_info_t *info = get_irn_link(node);
154 info = OALLOCFZ(&obst, allocation_info_t, prefs, n_regs);
155 info->current_value = node;
156 info->original_value = node;
157 set_irn_link(node, info);
164 * Get allocation information for a basic block
166 static block_info_t *get_block_info(ir_node *block)
168 block_info_t *info = get_irn_link(block);
170 assert(is_Block(block));
172 info = OALLOCFZ(&obst, block_info_t, assignments, n_regs);
173 set_irn_link(block, info);
180 * Get default register requirement for the current register class
182 static const arch_register_req_t *get_default_req_current_cls(void)
184 if (default_cls_req == NULL) {
185 struct obstack *obst = get_irg_obstack(irg);
186 arch_register_req_t *req = OALLOCZ(obst, arch_register_req_t);
188 req->type = arch_register_req_type_normal;
191 default_cls_req = req;
193 return default_cls_req;
197 * Link the allocation info of a node to a copy.
198 * Afterwards, both nodes uses the same allocation info.
199 * Copy must not have an allocation info assigned yet.
201 * @param copy the node that gets the allocation info assigned
202 * @param value the original node
204 static void mark_as_copy_of(ir_node *copy, ir_node *value)
207 allocation_info_t *info = get_allocation_info(value);
208 allocation_info_t *copy_info = get_allocation_info(copy);
210 /* find original value */
211 original = info->original_value;
212 if (original != value) {
213 info = get_allocation_info(original);
216 assert(info->original_value == original);
217 info->current_value = copy;
219 /* the copy should not be linked to something else yet */
220 assert(copy_info->original_value == copy);
221 copy_info->original_value = original;
223 /* copy over allocation preferences */
224 memcpy(copy_info->prefs, info->prefs, n_regs * sizeof(copy_info->prefs[0]));
228 * Calculate the penalties for every register on a node and its live neighbors.
230 * @param live_nodes the set of live nodes at the current position, may be NULL
231 * @param penalty the penalty to subtract from
232 * @param limited a raw bitset containing the limited set for the node
233 * @param node the node
235 static void give_penalties_for_limits(const ir_nodeset_t *live_nodes,
236 float penalty, const unsigned* limited,
239 ir_nodeset_iterator_t iter;
242 allocation_info_t *info = get_allocation_info(node);
245 /* give penalty for all forbidden regs */
246 for (r = 0; r < n_regs; ++r) {
247 if (rbitset_is_set(limited, r))
250 info->prefs[r] -= penalty;
253 /* all other live values should get a penalty for allowed regs */
254 if (live_nodes == NULL)
257 penalty *= NEIGHBOR_FACTOR;
258 n_allowed = rbitset_popcnt(limited, n_regs);
260 /* only create a very weak penalty if multiple regs are allowed */
261 penalty = (penalty * 0.8) / n_allowed;
263 foreach_ir_nodeset(live_nodes, neighbor, iter) {
264 allocation_info_t *neighbor_info;
266 /* TODO: if op is used on multiple inputs we might not do a
268 if (neighbor == node)
271 neighbor_info = get_allocation_info(neighbor);
272 for (r = 0; r < n_regs; ++r) {
273 if (!rbitset_is_set(limited, r))
276 neighbor_info->prefs[r] -= penalty;
282 * Calculate the preferences of a definition for the current register class.
283 * If the definition uses a limited set of registers, reduce the preferences
284 * for the limited register on the node and its neighbors.
286 * @param live_nodes the set of live nodes at the current node
287 * @param weight the weight
288 * @param node the current node
290 static void check_defs(const ir_nodeset_t *live_nodes, float weight,
293 const arch_register_req_t *req;
295 if (get_irn_mode(node) == mode_T) {
296 const ir_edge_t *edge;
297 foreach_out_edge(node, edge) {
298 ir_node *proj = get_edge_src_irn(edge);
299 check_defs(live_nodes, weight, proj);
304 if (!arch_irn_consider_in_reg_alloc(cls, node))
307 req = arch_get_register_req_out(node);
308 if (req->type & arch_register_req_type_limited) {
309 const unsigned *limited = req->limited;
310 float penalty = weight * DEF_FACTOR;
311 give_penalties_for_limits(live_nodes, penalty, limited, node);
314 if (req->type & arch_register_req_type_should_be_same) {
315 ir_node *insn = skip_Proj(node);
316 allocation_info_t *info = get_allocation_info(node);
317 int arity = get_irn_arity(insn);
320 float factor = 1.0f / rbitset_popcnt(&req->other_same, arity);
321 for (i = 0; i < arity; ++i) {
324 allocation_info_t *op_info;
326 if (!rbitset_is_set(&req->other_same, i))
329 op = get_irn_n(insn, i);
331 /* if we the value at the should_be_same input doesn't die at the
332 * node, then it is no use to propagate the constraints (since a
333 * copy will emerge anyway) */
334 if (ir_nodeset_contains(live_nodes, op))
337 op_info = get_allocation_info(op);
338 for (r = 0; r < n_regs; ++r) {
339 op_info->prefs[r] += info->prefs[r] * factor;
346 * Walker: Runs an a block calculates the preferences for any
347 * node and every register from the considered register class.
349 static void analyze_block(ir_node *block, void *data)
351 float weight = get_block_execfreq(execfreqs, block);
352 ir_nodeset_t live_nodes;
356 ir_nodeset_init(&live_nodes);
357 be_liveness_end_of_block(lv, cls, block, &live_nodes);
359 sched_foreach_reverse(block, node) {
360 allocation_info_t *info;
367 if (create_preferences)
368 check_defs(&live_nodes, weight, node);
371 arity = get_irn_arity(node);
373 /* the allocation info node currently only uses 1 unsigned value
374 to mark last used inputs. So we will fail for a node with more than
376 if (arity >= (int) sizeof(unsigned) * 8) {
377 panic("Node with more than %d inputs not supported yet",
378 (int) sizeof(unsigned) * 8);
381 info = get_allocation_info(node);
382 for (i = 0; i < arity; ++i) {
383 ir_node *op = get_irn_n(node, i);
384 if (!arch_irn_consider_in_reg_alloc(cls, op))
387 /* last usage of a value? */
388 if (!ir_nodeset_contains(&live_nodes, op)) {
389 rbitset_set(&info->last_uses, i);
393 be_liveness_transfer(cls, node, &live_nodes);
395 if (create_preferences) {
396 /* update weights based on usage constraints */
397 for (i = 0; i < arity; ++i) {
398 const arch_register_req_t *req;
399 const unsigned *limited;
400 ir_node *op = get_irn_n(node, i);
402 if (!arch_irn_consider_in_reg_alloc(cls, op))
405 req = arch_get_register_req(node, i);
406 if (!(req->type & arch_register_req_type_limited))
409 limited = req->limited;
410 give_penalties_for_limits(&live_nodes, weight * USE_FACTOR, limited,
416 ir_nodeset_destroy(&live_nodes);
419 static void congruence_def(ir_nodeset_t *live_nodes, ir_node *node)
421 const arch_register_req_t *req;
423 if (get_irn_mode(node) == mode_T) {
424 const ir_edge_t *edge;
425 foreach_out_edge(node, edge) {
426 ir_node *def = get_edge_src_irn(edge);
427 congruence_def(live_nodes, def);
432 if (!arch_irn_consider_in_reg_alloc(cls, node))
435 /* should be same constraint? */
436 req = arch_get_register_req_out(node);
437 if (req->type & arch_register_req_type_should_be_same) {
438 ir_node *insn = skip_Proj(node);
439 int arity = get_irn_arity(insn);
441 unsigned node_idx = get_irn_idx(node);
442 node_idx = uf_find(congruence_classes, node_idx);
444 for (i = 0; i < arity; ++i) {
448 ir_nodeset_iterator_t iter;
449 bool interferes = false;
451 if (!rbitset_is_set(&req->other_same, i))
454 op = get_irn_n(insn, i);
455 op_idx = get_irn_idx(op);
456 op_idx = uf_find(congruence_classes, op_idx);
458 /* do we interfere with the value */
459 foreach_ir_nodeset(live_nodes, live, iter) {
460 int lv_idx = get_irn_idx(live);
461 lv_idx = uf_find(congruence_classes, lv_idx);
462 if (lv_idx == op_idx) {
467 /* don't put in same affinity class if we interfere */
471 node_idx = uf_union(congruence_classes, node_idx, op_idx);
472 DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
474 /* one should_be_same is enough... */
480 static void create_congurence_class(ir_node *block, void *data)
482 ir_nodeset_t live_nodes;
486 ir_nodeset_init(&live_nodes);
487 be_liveness_end_of_block(lv, cls, block, &live_nodes);
489 /* check should be same constraints */
490 sched_foreach_reverse(block, node) {
494 congruence_def(&live_nodes, node);
495 be_liveness_transfer(cls, node, &live_nodes);
498 /* check phi congruence classes */
499 sched_foreach_reverse_from(node, node) {
503 assert(is_Phi(node));
505 if (!arch_irn_consider_in_reg_alloc(cls, node))
508 node_idx = get_irn_idx(node);
509 node_idx = uf_find(congruence_classes, node_idx);
511 arity = get_irn_arity(node);
512 for (i = 0; i < arity; ++i) {
513 bool interferes = false;
514 ir_nodeset_iterator_t iter;
517 ir_node *op = get_Phi_pred(node, i);
518 int op_idx = get_irn_idx(op);
519 op_idx = uf_find(congruence_classes, op_idx);
521 /* do we interfere with the value */
522 foreach_ir_nodeset(&live_nodes, live, iter) {
523 int lv_idx = get_irn_idx(live);
524 lv_idx = uf_find(congruence_classes, lv_idx);
525 if (lv_idx == op_idx) {
530 /* don't put in same affinity class if we interfere */
533 /* any other phi has the same input? */
534 sched_foreach(block, phi) {
539 if (!arch_irn_consider_in_reg_alloc(cls, phi))
541 oop = get_Phi_pred(phi, i);
544 oop_idx = get_irn_idx(oop);
545 oop_idx = uf_find(congruence_classes, oop_idx);
546 if (oop_idx == op_idx) {
554 node_idx = uf_union(congruence_classes, node_idx, op_idx);
555 DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
561 static void merge_congruence_prefs(ir_node *node, void *data)
563 allocation_info_t *info;
564 allocation_info_t *head_info;
565 unsigned node_idx = get_irn_idx(node);
566 unsigned node_set = uf_find(congruence_classes, node_idx);
571 /* head of congruence class or not in any class */
572 if (node_set == node_idx)
575 if (!arch_irn_consider_in_reg_alloc(cls, node))
578 head_info = get_allocation_info(get_idx_irn(irg, node_set));
579 info = get_allocation_info(node);
581 for (r = 0; r < n_regs; ++r) {
582 head_info->prefs[r] += info->prefs[r];
586 static void set_congruence_prefs(ir_node *node, void *data)
588 allocation_info_t *info;
589 allocation_info_t *head_info;
590 unsigned node_idx = get_irn_idx(node);
591 unsigned node_set = uf_find(congruence_classes, node_idx);
595 /* head of congruence class or not in any class */
596 if (node_set == node_idx)
599 if (!arch_irn_consider_in_reg_alloc(cls, node))
602 head_info = get_allocation_info(get_idx_irn(irg, node_set));
603 info = get_allocation_info(node);
605 memcpy(info->prefs, head_info->prefs, n_regs * sizeof(info->prefs[0]));
608 static void combine_congruence_classes(void)
610 size_t n = get_irg_last_idx(irg);
611 congruence_classes = XMALLOCN(int, n);
612 uf_init(congruence_classes, n);
614 /* create congruence classes */
615 irg_block_walk_graph(irg, create_congurence_class, NULL, NULL);
616 /* merge preferences */
617 irg_walk_graph(irg, merge_congruence_prefs, NULL, NULL);
618 irg_walk_graph(irg, set_congruence_prefs, NULL, NULL);
619 free(congruence_classes);
627 * Assign register reg to the given node.
629 * @param node the node
630 * @param reg the register
632 static void use_reg(ir_node *node, const arch_register_t *reg)
634 unsigned r = arch_register_get_index(reg);
635 assignments[r] = node;
636 arch_set_irn_register(node, reg);
639 static void free_reg_of_value(ir_node *node)
641 const arch_register_t *reg;
644 if (!arch_irn_consider_in_reg_alloc(cls, node))
647 reg = arch_get_irn_register(node);
648 r = arch_register_get_index(reg);
649 /* assignment->value may be NULL if a value is used at 2 inputs
650 so it gets freed twice. */
651 assert(assignments[r] == node || assignments[r] == NULL);
652 assignments[r] = NULL;
656 * Compare two register preferences in decreasing order.
658 static int compare_reg_pref(const void *e1, const void *e2)
660 const reg_pref_t *rp1 = (const reg_pref_t*) e1;
661 const reg_pref_t *rp2 = (const reg_pref_t*) e2;
662 if (rp1->pref < rp2->pref)
664 if (rp1->pref > rp2->pref)
669 static void fill_sort_candidates(reg_pref_t *regprefs,
670 const allocation_info_t *info)
674 for (r = 0; r < n_regs; ++r) {
675 float pref = info->prefs[r];
677 regprefs[r].pref = pref;
679 /* TODO: use a stable sort here to avoid unnecessary register jumping */
680 qsort(regprefs, n_regs, sizeof(regprefs[0]), compare_reg_pref);
683 static bool try_optimistic_split(ir_node *to_split, ir_node *before,
684 float pref, float pref_delta,
685 unsigned *forbidden_regs, int recursion)
687 const arch_register_t *from_reg;
688 const arch_register_t *reg;
689 ir_node *original_insn;
695 allocation_info_t *info = get_allocation_info(to_split);
698 float split_threshold;
702 /* stupid hack: don't optimisticallt split don't spill nodes...
703 * (so we don't split away the values produced because of
704 * must_be_different constraints) */
705 original_insn = skip_Proj(info->original_value);
706 if (arch_irn_get_flags(original_insn) & arch_irn_flags_dont_spill)
709 from_reg = arch_get_irn_register(to_split);
710 from_r = arch_register_get_index(from_reg);
711 block = get_nodes_block(before);
712 split_threshold = get_block_execfreq(execfreqs, block) * SPLIT_DELTA;
714 if (pref_delta < split_threshold*0.5)
717 /* find the best free position where we could move to */
718 prefs = ALLOCAN(reg_pref_t, n_regs);
719 fill_sort_candidates(prefs, info);
720 for (i = 0; i < n_regs; ++i) {
724 bool old_source_state;
726 /* we need a normal register which is not an output register
727 an different from the current register of to_split */
729 if (!rbitset_is_set(normal_regs, r))
731 if (rbitset_is_set(forbidden_regs, r))
736 /* is the split worth it? */
737 delta = pref_delta + prefs[i].pref;
738 if (delta < split_threshold) {
739 DB((dbg, LEVEL_3, "Not doing optimistical split of %+F (depth %d), win %f too low\n",
740 to_split, recursion, delta));
744 /* if the register is free then we can do the split */
745 if (assignments[r] == NULL)
748 /* otherwise we might try recursively calling optimistic_split */
749 if (recursion+1 > MAX_OPTIMISTIC_SPLIT_RECURSION)
752 apref = prefs[i].pref;
753 apref_delta = i+1 < n_regs ? apref - prefs[i+1].pref : 0;
754 apref_delta += pref_delta - split_threshold;
756 /* our source register isn't a usefull destination for recursive
758 old_source_state = rbitset_is_set(forbidden_regs, from_r);
759 rbitset_set(forbidden_regs, from_r);
760 /* try recursive split */
761 res = try_optimistic_split(assignments[r], before, apref,
762 apref_delta, forbidden_regs, recursion+1);
763 /* restore our destination */
764 if (old_source_state) {
765 rbitset_set(forbidden_regs, from_r);
767 rbitset_clear(forbidden_regs, from_r);
776 reg = arch_register_for_index(cls, r);
777 copy = be_new_Copy(cls, block, to_split);
778 mark_as_copy_of(copy, to_split);
779 /* hacky, but correct here */
780 if (assignments[arch_register_get_index(from_reg)] == to_split)
781 free_reg_of_value(to_split);
783 sched_add_before(before, copy);
786 "Optimistic live-range split %+F move %+F(%s) -> %s before %+F (win %f, depth %d)\n",
787 copy, to_split, from_reg->name, reg->name, before, delta, recursion));
792 * Determine and assign a register for node @p node
794 static void assign_reg(const ir_node *block, ir_node *node,
795 unsigned *forbidden_regs)
797 const arch_register_t *reg;
798 allocation_info_t *info;
799 const arch_register_req_t *req;
800 reg_pref_t *reg_prefs;
803 const unsigned *allowed_regs;
806 assert(!is_Phi(node));
807 assert(arch_irn_consider_in_reg_alloc(cls, node));
809 /* preassigned register? */
810 reg = arch_get_irn_register(node);
812 DB((dbg, LEVEL_2, "Preassignment %+F -> %s\n", node, reg->name));
817 /* give should_be_same boni */
818 info = get_allocation_info(node);
819 req = arch_get_register_req_out(node);
821 in_node = skip_Proj(node);
822 if (req->type & arch_register_req_type_should_be_same) {
823 float weight = get_block_execfreq(execfreqs, block);
824 int arity = get_irn_arity(in_node);
827 assert(arity <= (int) sizeof(req->other_same) * 8);
828 for (i = 0; i < arity; ++i) {
830 const arch_register_t *reg;
832 if (!rbitset_is_set(&req->other_same, i))
835 in = get_irn_n(in_node, i);
836 reg = arch_get_irn_register(in);
838 r = arch_register_get_index(reg);
840 /* if the value didn't die here then we should not propagate the
841 * should_be_same info */
842 if (assignments[r] == in)
845 info->prefs[r] += weight * AFF_SHOULD_BE_SAME;
849 /* create list of register candidates and sort by their preference */
850 DB((dbg, LEVEL_2, "Candidates for %+F:", node));
851 reg_prefs = alloca(n_regs * sizeof(reg_prefs[0]));
852 fill_sort_candidates(reg_prefs, info);
853 for (i = 0; i < n_regs; ++i) {
854 unsigned num = reg_prefs[i].num;
855 const arch_register_t *reg;
857 if (!rbitset_is_set(normal_regs, num))
860 reg = arch_register_for_index(cls, num);
861 DB((dbg, LEVEL_2, " %s(%f)", reg->name, reg_prefs[i].pref));
863 DB((dbg, LEVEL_2, "\n"));
865 allowed_regs = normal_regs;
866 if (req->type & arch_register_req_type_limited) {
867 allowed_regs = req->limited;
870 for (i = 0; i < n_regs; ++i) {
871 r = reg_prefs[i].num;
872 if (!rbitset_is_set(allowed_regs, r))
874 if (assignments[r] == NULL)
876 float pref = reg_prefs[i].pref;
877 float delta = i+1 < n_regs ? pref - reg_prefs[i+1].pref : 0;
878 ir_node *before = skip_Proj(node);
879 bool res = try_optimistic_split(assignments[r], before,
880 pref, delta, forbidden_regs, 0);
885 panic("No register left for %+F\n", node);
888 reg = arch_register_for_index(cls, r);
889 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
894 * Add an permutation in front of a node and change the assignments
895 * due to this permutation.
897 * To understand this imagine a permutation like this:
907 * First we count how many destinations a single value has. At the same time
908 * we can be sure that each destination register has at most 1 source register
909 * (it can have 0 which means we don't care what value is in it).
910 * We ignore all fullfilled permuations (like 7->7)
911 * In a first pass we create as much copy instructions as possible as they
912 * are generally cheaper than exchanges. We do this by counting into how many
913 * destinations a register has to be copied (in the example it's 2 for register
914 * 3, or 1 for the registers 1,2,4 and 7).
915 * We can then create a copy into every destination register when the usecount
916 * of that register is 0 (= noone else needs the value in the register).
918 * After this step we should have cycles left. We implement a cyclic permutation
919 * of n registers with n-1 transpositions.
921 * @param live_nodes the set of live nodes, updated due to live range split
922 * @param before the node before we add the permutation
923 * @param permutation the permutation array indices are the destination
924 * registers, the values in the array are the source
927 static void permute_values(ir_nodeset_t *live_nodes, ir_node *before,
928 unsigned *permutation)
930 unsigned *n_used = ALLOCANZ(unsigned, n_regs);
934 /* determine how often each source register needs to be read */
935 for (r = 0; r < n_regs; ++r) {
936 unsigned old_reg = permutation[r];
939 value = assignments[old_reg];
941 /* nothing to do here, reg is not live. Mark it as fixpoint
942 * so we ignore it in the next steps */
950 block = get_nodes_block(before);
952 /* step1: create copies where immediately possible */
953 for (r = 0; r < n_regs; /* empty */) {
956 const arch_register_t *reg;
957 unsigned old_r = permutation[r];
959 /* - no need to do anything for fixed points.
960 - we can't copy if the value in the dest reg is still needed */
961 if (old_r == r || n_used[r] > 0) {
967 src = assignments[old_r];
968 copy = be_new_Copy(cls, block, src);
969 sched_add_before(before, copy);
970 reg = arch_register_for_index(cls, r);
971 DB((dbg, LEVEL_2, "Copy %+F (from %+F, before %+F) -> %s\n",
972 copy, src, before, reg->name));
973 mark_as_copy_of(copy, src);
976 if (live_nodes != NULL) {
977 ir_nodeset_insert(live_nodes, copy);
980 /* old register has 1 user less, permutation is resolved */
981 assert(arch_register_get_index(arch_get_irn_register(src)) == old_r);
984 assert(n_used[old_r] > 0);
986 if (n_used[old_r] == 0) {
987 if (live_nodes != NULL) {
988 ir_nodeset_remove(live_nodes, src);
990 free_reg_of_value(src);
993 /* advance or jump back (if this copy enabled another copy) */
994 if (old_r < r && n_used[old_r] == 0) {
1001 /* at this point we only have "cycles" left which we have to resolve with
1003 * TODO: if we have free registers left, then we should really use copy
1004 * instructions for any cycle longer than 2 registers...
1005 * (this is probably architecture dependent, there might be archs where
1006 * copies are preferable even for 2-cycles) */
1008 /* create perms with the rest */
1009 for (r = 0; r < n_regs; /* empty */) {
1010 const arch_register_t *reg;
1011 unsigned old_r = permutation[r];
1023 /* we shouldn't have copies from 1 value to multiple destinations left*/
1024 assert(n_used[old_r] == 1);
1026 /* exchange old_r and r2; after that old_r is a fixed point */
1027 r2 = permutation[old_r];
1029 in[0] = assignments[r2];
1030 in[1] = assignments[old_r];
1031 perm = be_new_Perm(cls, block, 2, in);
1032 sched_add_before(before, perm);
1033 DB((dbg, LEVEL_2, "Perm %+F (perm %+F,%+F, before %+F)\n",
1034 perm, in[0], in[1], before));
1036 proj0 = new_r_Proj(block, perm, get_irn_mode(in[0]), 0);
1037 mark_as_copy_of(proj0, in[0]);
1038 reg = arch_register_for_index(cls, old_r);
1039 use_reg(proj0, reg);
1041 proj1 = new_r_Proj(block, perm, get_irn_mode(in[1]), 1);
1042 mark_as_copy_of(proj1, in[1]);
1043 reg = arch_register_for_index(cls, r2);
1044 use_reg(proj1, reg);
1046 /* 1 value is now in the correct register */
1047 permutation[old_r] = old_r;
1048 /* the source of r changed to r2 */
1049 permutation[r] = r2;
1051 /* if we have reached a fixpoint update data structures */
1052 if (live_nodes != NULL) {
1053 ir_nodeset_remove(live_nodes, in[0]);
1054 ir_nodeset_remove(live_nodes, in[1]);
1055 ir_nodeset_remove(live_nodes, proj0);
1056 ir_nodeset_insert(live_nodes, proj1);
1060 #ifdef DEBUG_libfirm
1061 /* now we should only have fixpoints left */
1062 for (r = 0; r < n_regs; ++r) {
1063 assert(permutation[r] == r);
1069 * Free regs for values last used.
1071 * @param live_nodes set of live nodes, will be updated
1072 * @param node the node to consider
1074 static void free_last_uses(ir_nodeset_t *live_nodes, ir_node *node)
1076 allocation_info_t *info = get_allocation_info(node);
1077 const unsigned *last_uses = &info->last_uses;
1078 int arity = get_irn_arity(node);
1081 for (i = 0; i < arity; ++i) {
1084 /* check if one operand is the last use */
1085 if (!rbitset_is_set(last_uses, i))
1088 op = get_irn_n(node, i);
1089 free_reg_of_value(op);
1090 ir_nodeset_remove(live_nodes, op);
1095 * change inputs of a node to the current value (copies/perms)
1097 static void rewire_inputs(ir_node *node)
1100 int arity = get_irn_arity(node);
1102 for (i = 0; i < arity; ++i) {
1103 ir_node *op = get_irn_n(node, i);
1104 allocation_info_t *info;
1106 if (!arch_irn_consider_in_reg_alloc(cls, op))
1109 info = get_allocation_info(op);
1110 info = get_allocation_info(info->original_value);
1111 if (info->current_value != op) {
1112 set_irn_n(node, i, info->current_value);
1118 * Create a bitset of registers occupied with value living through an
1121 static void determine_live_through_regs(unsigned *bitset, ir_node *node)
1123 const allocation_info_t *info = get_allocation_info(node);
1128 /* mark all used registers as potentially live-through */
1129 for (r = 0; r < n_regs; ++r) {
1130 if (assignments[r] == NULL)
1132 if (!rbitset_is_set(normal_regs, r))
1135 rbitset_set(bitset, r);
1138 /* remove registers of value dying at the instruction */
1139 arity = get_irn_arity(node);
1140 for (i = 0; i < arity; ++i) {
1142 const arch_register_t *reg;
1144 if (!rbitset_is_set(&info->last_uses, i))
1147 op = get_irn_n(node, i);
1148 reg = arch_get_irn_register(op);
1149 rbitset_clear(bitset, arch_register_get_index(reg));
1154 * Enforce constraints at a node by live range splits.
1156 * @param live_nodes the set of live nodes, might be changed
1157 * @param node the current node
1159 static void enforce_constraints(ir_nodeset_t *live_nodes, ir_node *node,
1160 unsigned *forbidden_regs)
1162 int arity = get_irn_arity(node);
1164 hungarian_problem_t *bp;
1166 unsigned *assignment;
1168 /* construct a list of register occupied by live-through values */
1169 unsigned *live_through_regs = NULL;
1171 /* see if any use constraints are not met */
1173 for (i = 0; i < arity; ++i) {
1174 ir_node *op = get_irn_n(node, i);
1175 const arch_register_t *reg;
1176 const arch_register_req_t *req;
1177 const unsigned *limited;
1180 if (!arch_irn_consider_in_reg_alloc(cls, op))
1183 /* are there any limitations for the i'th operand? */
1184 req = arch_get_register_req(node, i);
1185 if (!(req->type & arch_register_req_type_limited))
1188 limited = req->limited;
1189 reg = arch_get_irn_register(op);
1190 r = arch_register_get_index(reg);
1191 if (!rbitset_is_set(limited, r)) {
1192 /* found an assignment outside the limited set */
1198 /* is any of the live-throughs using a constrained output register? */
1199 if (get_irn_mode(node) == mode_T) {
1200 const ir_edge_t *edge;
1202 foreach_out_edge(node, edge) {
1203 ir_node *proj = get_edge_src_irn(edge);
1204 const arch_register_req_t *req;
1206 if (!arch_irn_consider_in_reg_alloc(cls, proj))
1209 req = arch_get_register_req_out(proj);
1210 if (!(req->type & arch_register_req_type_limited))
1213 if (live_through_regs == NULL) {
1214 rbitset_alloca(live_through_regs, n_regs);
1215 determine_live_through_regs(live_through_regs, node);
1218 rbitset_or(forbidden_regs, req->limited, n_regs);
1219 if (rbitsets_have_common(req->limited, live_through_regs, n_regs)) {
1224 if (arch_irn_consider_in_reg_alloc(cls, node)) {
1225 const arch_register_req_t *req = arch_get_register_req_out(node);
1226 if (req->type & arch_register_req_type_limited) {
1227 rbitset_alloca(live_through_regs, n_regs);
1228 determine_live_through_regs(live_through_regs, node);
1229 if (rbitsets_have_common(req->limited, live_through_regs, n_regs)) {
1231 rbitset_or(forbidden_regs, req->limited, n_regs);
1240 /* create these arrays if we haven't yet */
1241 if (live_through_regs == NULL) {
1242 rbitset_alloca(live_through_regs, n_regs);
1245 /* at this point we have to construct a bipartite matching problem to see
1246 * which values should go to which registers
1247 * Note: We're building the matrix in "reverse" - source registers are
1248 * right, destinations left because this will produce the solution
1249 * in the format required for permute_values.
1251 bp = hungarian_new(n_regs, n_regs, HUNGARIAN_MATCH_PERFECT);
1253 /* add all combinations, then remove not allowed ones */
1254 for (l = 0; l < n_regs; ++l) {
1255 if (!rbitset_is_set(normal_regs, l)) {
1256 hungarian_add(bp, l, l, 1);
1260 for (r = 0; r < n_regs; ++r) {
1261 if (!rbitset_is_set(normal_regs, r))
1263 /* livethrough values may not use constrainted output registers */
1264 if (rbitset_is_set(live_through_regs, l)
1265 && rbitset_is_set(forbidden_regs, r))
1268 hungarian_add(bp, r, l, l == r ? 9 : 8);
1272 for (i = 0; i < arity; ++i) {
1273 ir_node *op = get_irn_n(node, i);
1274 const arch_register_t *reg;
1275 const arch_register_req_t *req;
1276 const unsigned *limited;
1277 unsigned current_reg;
1279 if (!arch_irn_consider_in_reg_alloc(cls, op))
1282 req = arch_get_register_req(node, i);
1283 if (!(req->type & arch_register_req_type_limited))
1286 limited = req->limited;
1287 reg = arch_get_irn_register(op);
1288 current_reg = arch_register_get_index(reg);
1289 for (r = 0; r < n_regs; ++r) {
1290 if (rbitset_is_set(limited, r))
1292 hungarian_remv(bp, r, current_reg);
1296 //hungarian_print_cost_matrix(bp, 1);
1297 hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1299 assignment = ALLOCAN(unsigned, n_regs);
1300 res = hungarian_solve(bp, (int*) assignment, NULL, 0);
1304 fprintf(stderr, "Swap result:");
1305 for (i = 0; i < (int) n_regs; ++i) {
1306 fprintf(stderr, " %d", assignment[i]);
1308 fprintf(stderr, "\n");
1313 permute_values(live_nodes, node, assignment);
1316 /** test wether a node @p n is a copy of the value of node @p of */
1317 static bool is_copy_of(ir_node *value, ir_node *test_value)
1319 allocation_info_t *test_info;
1320 allocation_info_t *info;
1322 if (value == test_value)
1325 info = get_allocation_info(value);
1326 test_info = get_allocation_info(test_value);
1327 return test_info->original_value == info->original_value;
1331 * find a value in the end-assignment of a basic block
1332 * @returns the index into the assignment array if found
1335 static int find_value_in_block_info(block_info_t *info, ir_node *value)
1338 ir_node **assignments = info->assignments;
1339 for (r = 0; r < n_regs; ++r) {
1340 ir_node *a_value = assignments[r];
1342 if (a_value == NULL)
1344 if (is_copy_of(a_value, value))
1352 * Create the necessary permutations at the end of a basic block to fullfill
1353 * the register assignment for phi-nodes in the next block
1355 static void add_phi_permutations(ir_node *block, int p)
1358 unsigned *permutation;
1359 ir_node **old_assignments;
1360 bool need_permutation;
1362 ir_node *pred = get_Block_cfgpred_block(block, p);
1364 block_info_t *pred_info = get_block_info(pred);
1366 /* predecessor not processed yet? nothing to do */
1367 if (!pred_info->processed)
1370 permutation = ALLOCAN(unsigned, n_regs);
1371 for (r = 0; r < n_regs; ++r) {
1375 /* check phi nodes */
1376 need_permutation = false;
1377 node = sched_first(block);
1378 for ( ; is_Phi(node); node = sched_next(node)) {
1379 const arch_register_t *reg;
1384 if (!arch_irn_consider_in_reg_alloc(cls, node))
1387 op = get_Phi_pred(node, p);
1388 if (!arch_irn_consider_in_reg_alloc(cls, op))
1391 a = find_value_in_block_info(pred_info, op);
1394 reg = arch_get_irn_register(node);
1395 regn = arch_register_get_index(reg);
1397 permutation[regn] = a;
1398 need_permutation = true;
1402 if (need_permutation) {
1403 /* permute values at end of predecessor */
1404 old_assignments = assignments;
1405 assignments = pred_info->assignments;
1406 permute_values(NULL, be_get_end_of_block_insertion_point(pred),
1408 assignments = old_assignments;
1411 /* change phi nodes to use the copied values */
1412 node = sched_first(block);
1413 for ( ; is_Phi(node); node = sched_next(node)) {
1417 if (!arch_irn_consider_in_reg_alloc(cls, node))
1420 op = get_Phi_pred(node, p);
1421 /* no need to do anything for Unknown inputs */
1422 if (!arch_irn_consider_in_reg_alloc(cls, op))
1425 /* we have permuted all values into the correct registers so we can
1426 simply query which value occupies the phis register in the
1428 a = arch_register_get_index(arch_get_irn_register(node));
1429 op = pred_info->assignments[a];
1430 set_Phi_pred(node, p, op);
1435 * Set preferences for a phis register based on the registers used on the
1438 static void adapt_phi_prefs(ir_node *phi)
1441 int arity = get_irn_arity(phi);
1442 ir_node *block = get_nodes_block(phi);
1443 allocation_info_t *info = get_allocation_info(phi);
1445 for (i = 0; i < arity; ++i) {
1446 ir_node *op = get_irn_n(phi, i);
1447 const arch_register_t *reg = arch_get_irn_register(op);
1448 ir_node *pred_block;
1449 block_info_t *pred_block_info;
1455 /* we only give the bonus if the predecessor already has registers
1456 * assigned, otherwise we only see a dummy value
1457 * and any conclusions about its register are useless */
1458 pred_block = get_Block_cfgpred_block(block, i);
1459 pred_block_info = get_block_info(pred_block);
1460 if (!pred_block_info->processed)
1463 /* give bonus for already assigned register */
1464 weight = get_block_execfreq(execfreqs, pred_block);
1465 r = arch_register_get_index(reg);
1466 info->prefs[r] += weight * AFF_PHI;
1471 * After a phi has been assigned a register propagate preference inputs
1472 * to the phi inputs.
1474 static void propagate_phi_register(ir_node *phi, unsigned assigned_r)
1477 ir_node *block = get_nodes_block(phi);
1478 int arity = get_irn_arity(phi);
1480 for (i = 0; i < arity; ++i) {
1481 ir_node *op = get_Phi_pred(phi, i);
1482 allocation_info_t *info = get_allocation_info(op);
1483 ir_node *pred_block = get_Block_cfgpred_block(block, i);
1486 = get_block_execfreq(execfreqs, pred_block) * AFF_PHI;
1488 if (info->prefs[assigned_r] >= weight)
1491 /* promote the prefered register */
1492 for (r = 0; r < n_regs; ++r) {
1493 if (info->prefs[r] > -weight) {
1494 info->prefs[r] = -weight;
1497 info->prefs[assigned_r] = weight;
1500 propagate_phi_register(op, assigned_r);
1504 static void assign_phi_registers(ir_node *block)
1511 hungarian_problem_t *bp;
1513 /* count phi nodes */
1514 sched_foreach(block, node) {
1517 if (!arch_irn_consider_in_reg_alloc(cls, node))
1525 /* build a bipartite matching problem for all phi nodes */
1526 bp = hungarian_new(n_phis, n_regs, HUNGARIAN_MATCH_PERFECT);
1528 sched_foreach(block, node) {
1531 allocation_info_t *info;
1534 if (!arch_irn_consider_in_reg_alloc(cls, node))
1537 /* give boni for predecessor colorings */
1538 adapt_phi_prefs(node);
1539 /* add stuff to bipartite problem */
1540 info = get_allocation_info(node);
1541 DB((dbg, LEVEL_3, "Prefs for %+F: ", node));
1542 for (r = 0; r < n_regs; ++r) {
1545 if (!rbitset_is_set(normal_regs, r))
1548 costs = info->prefs[r];
1549 costs = costs < 0 ? -logf(-costs+1) : logf(costs+1);
1552 hungarian_add(bp, n, r, costs);
1553 DB((dbg, LEVEL_3, " %s(%f)", arch_register_for_index(cls, r)->name,
1556 DB((dbg, LEVEL_3, "\n"));
1560 //hungarian_print_cost_matrix(bp, 7);
1561 hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1563 assignment = ALLOCAN(int, n_regs);
1564 res = hungarian_solve(bp, assignment, NULL, 0);
1569 sched_foreach(block, node) {
1571 const arch_register_t *reg;
1575 if (!arch_irn_consider_in_reg_alloc(cls, node))
1578 r = assignment[n++];
1579 assert(rbitset_is_set(normal_regs, r));
1580 reg = arch_register_for_index(cls, r);
1581 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
1584 /* adapt preferences for phi inputs */
1585 if (propagate_phi_registers)
1586 propagate_phi_register(node, r);
1591 * Walker: assign registers to all nodes of a block that
1592 * need registers from the currently considered register class.
1594 static void allocate_coalesce_block(ir_node *block, void *data)
1597 ir_nodeset_t live_nodes;
1598 ir_nodeset_iterator_t iter;
1601 block_info_t *block_info;
1602 block_info_t **pred_block_infos;
1604 unsigned *forbidden_regs; /**< collects registers which must
1605 not be used for optimistic splits */
1608 DB((dbg, LEVEL_2, "* Block %+F\n", block));
1610 /* clear assignments */
1611 block_info = get_block_info(block);
1612 assignments = block_info->assignments;
1614 ir_nodeset_init(&live_nodes);
1616 /* gather regalloc infos of predecessor blocks */
1617 n_preds = get_Block_n_cfgpreds(block);
1618 pred_block_infos = ALLOCAN(block_info_t*, n_preds);
1619 for (i = 0; i < n_preds; ++i) {
1620 ir_node *pred = get_Block_cfgpred_block(block, i);
1621 block_info_t *pred_info = get_block_info(pred);
1622 pred_block_infos[i] = pred_info;
1625 phi_ins = ALLOCAN(ir_node*, n_preds);
1627 /* collect live-in nodes and preassigned values */
1628 be_lv_foreach(lv, block, be_lv_state_in, i) {
1629 const arch_register_t *reg;
1631 bool need_phi = false;
1633 node = be_lv_get_irn(lv, block, i);
1634 if (!arch_irn_consider_in_reg_alloc(cls, node))
1637 /* check all predecessors for this value, if it is not everywhere the
1638 same or unknown then we have to construct a phi
1639 (we collect the potential phi inputs here) */
1640 for (p = 0; p < n_preds; ++p) {
1641 block_info_t *pred_info = pred_block_infos[p];
1643 if (!pred_info->processed) {
1644 /* use node for now, it will get fixed later */
1648 int a = find_value_in_block_info(pred_info, node);
1650 /* must live out of predecessor */
1652 phi_ins[p] = pred_info->assignments[a];
1653 /* different value from last time? then we need a phi */
1654 if (p > 0 && phi_ins[p-1] != phi_ins[p]) {
1661 ir_mode *mode = get_irn_mode(node);
1662 const arch_register_req_t *req = get_default_req_current_cls();
1666 phi = new_r_Phi(block, n_preds, phi_ins, mode);
1667 be_set_phi_reg_req(phi, req);
1669 DB((dbg, LEVEL_3, "Create Phi %+F (for %+F) -", phi, node));
1670 #ifdef DEBUG_libfirm
1671 for (i = 0; i < n_preds; ++i) {
1672 DB((dbg, LEVEL_3, " %+F", phi_ins[i]));
1674 DB((dbg, LEVEL_3, "\n"));
1676 mark_as_copy_of(phi, node);
1677 sched_add_after(block, phi);
1681 allocation_info_t *info = get_allocation_info(node);
1682 info->current_value = phi_ins[0];
1684 /* Grab 1 of the inputs we constructed (might not be the same as
1685 * "node" as we could see the same copy of the value in all
1690 /* if the node already has a register assigned use it */
1691 reg = arch_get_irn_register(node);
1696 /* remember that this node is live at the beginning of the block */
1697 ir_nodeset_insert(&live_nodes, node);
1700 rbitset_alloca(forbidden_regs, n_regs);
1702 /* handle phis... */
1703 assign_phi_registers(block);
1705 /* all live-ins must have a register */
1706 #ifdef DEBUG_libfirm
1707 foreach_ir_nodeset(&live_nodes, node, iter) {
1708 const arch_register_t *reg = arch_get_irn_register(node);
1709 assert(reg != NULL);
1713 /* assign instructions in the block */
1714 sched_foreach(block, node) {
1718 /* phis are already assigned */
1722 rewire_inputs(node);
1724 /* enforce use constraints */
1725 rbitset_clear_all(forbidden_regs, n_regs);
1726 enforce_constraints(&live_nodes, node, forbidden_regs);
1728 rewire_inputs(node);
1730 /* we may not use registers used for inputs for optimistic splits */
1731 arity = get_irn_arity(node);
1732 for (i = 0; i < arity; ++i) {
1733 ir_node *op = get_irn_n(node, i);
1734 const arch_register_t *reg;
1735 if (!arch_irn_consider_in_reg_alloc(cls, op))
1738 reg = arch_get_irn_register(op);
1739 rbitset_set(forbidden_regs, arch_register_get_index(reg));
1742 /* free registers of values last used at this instruction */
1743 free_last_uses(&live_nodes, node);
1745 /* assign output registers */
1746 /* TODO: 2 phases: first: pre-assigned ones, 2nd real regs */
1747 if (get_irn_mode(node) == mode_T) {
1748 const ir_edge_t *edge;
1749 foreach_out_edge(node, edge) {
1750 ir_node *proj = get_edge_src_irn(edge);
1751 if (!arch_irn_consider_in_reg_alloc(cls, proj))
1753 assign_reg(block, proj, forbidden_regs);
1755 } else if (arch_irn_consider_in_reg_alloc(cls, node)) {
1756 assign_reg(block, node, forbidden_regs);
1760 ir_nodeset_destroy(&live_nodes);
1763 block_info->processed = true;
1765 /* permute values at end of predecessor blocks in case of phi-nodes */
1768 for (p = 0; p < n_preds; ++p) {
1769 add_phi_permutations(block, p);
1773 /* if we have exactly 1 successor then we might be able to produce phi
1775 if (get_irn_n_edges_kind(block, EDGE_KIND_BLOCK) == 1) {
1776 const ir_edge_t *edge
1777 = get_irn_out_edge_first_kind(block, EDGE_KIND_BLOCK);
1778 ir_node *succ = get_edge_src_irn(edge);
1779 int p = get_edge_src_pos(edge);
1780 block_info_t *succ_info = get_block_info(succ);
1782 if (succ_info->processed) {
1783 add_phi_permutations(succ, p);
1788 typedef struct block_costs_t block_costs_t;
1789 struct block_costs_t {
1790 float costs; /**< costs of the block */
1791 int dfs_num; /**< depth first search number (to detect backedges) */
1794 static int cmp_block_costs(const void *d1, const void *d2)
1796 const ir_node * const *block1 = d1;
1797 const ir_node * const *block2 = d2;
1798 const block_costs_t *info1 = get_irn_link(*block1);
1799 const block_costs_t *info2 = get_irn_link(*block2);
1800 return QSORT_CMP(info2->costs, info1->costs);
1803 static void determine_block_order(void)
1806 ir_node **blocklist = be_get_cfgpostorder(irg);
1807 int n_blocks = ARR_LEN(blocklist);
1809 pdeq *worklist = new_pdeq();
1810 ir_node **order = XMALLOCN(ir_node*, n_blocks);
1813 /* clear block links... */
1814 for (i = 0; i < n_blocks; ++i) {
1815 ir_node *block = blocklist[i];
1816 set_irn_link(block, NULL);
1819 /* walk blocks in reverse postorder, the costs for each block are the
1820 * sum of the costs of its predecessors (excluding the costs on backedges
1821 * which we can't determine) */
1822 for (i = n_blocks-1; i >= 0; --i) {
1823 block_costs_t *cost_info;
1824 ir_node *block = blocklist[i];
1826 float execfreq = get_block_execfreq(execfreqs, block);
1827 float costs = execfreq;
1828 int n_cfgpreds = get_Block_n_cfgpreds(block);
1830 for (p = 0; p < n_cfgpreds; ++p) {
1831 ir_node *pred_block = get_Block_cfgpred_block(block, p);
1832 block_costs_t *pred_costs = get_irn_link(pred_block);
1833 /* we don't have any info for backedges */
1834 if (pred_costs == NULL)
1836 costs += pred_costs->costs;
1839 cost_info = OALLOCZ(&obst, block_costs_t);
1840 cost_info->costs = costs;
1841 cost_info->dfs_num = dfs_num++;
1842 set_irn_link(block, cost_info);
1845 /* sort array by block costs */
1846 qsort(blocklist, n_blocks, sizeof(blocklist[0]), cmp_block_costs);
1848 ir_reserve_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1849 inc_irg_block_visited(irg);
1851 for (i = 0; i < n_blocks; ++i) {
1852 ir_node *block = blocklist[i];
1853 if (Block_block_visited(block))
1856 /* continually add predecessors with highest costs to worklist
1857 * (without using backedges) */
1859 block_costs_t *info = get_irn_link(block);
1860 ir_node *best_pred = NULL;
1861 float best_costs = -1;
1862 int n_cfgpred = get_Block_n_cfgpreds(block);
1865 pdeq_putr(worklist, block);
1866 mark_Block_block_visited(block);
1867 for (i = 0; i < n_cfgpred; ++i) {
1868 ir_node *pred_block = get_Block_cfgpred_block(block, i);
1869 block_costs_t *pred_info = get_irn_link(pred_block);
1871 /* ignore backedges */
1872 if (pred_info->dfs_num > info->dfs_num)
1875 if (info->costs > best_costs) {
1876 best_costs = info->costs;
1877 best_pred = pred_block;
1881 } while(block != NULL && !Block_block_visited(block));
1883 /* now put all nodes in the worklist in our final order */
1884 while (!pdeq_empty(worklist)) {
1885 ir_node *pblock = pdeq_getr(worklist);
1886 assert(order_p < n_blocks);
1887 order[order_p++] = pblock;
1890 assert(order_p == n_blocks);
1893 ir_free_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1895 DEL_ARR_F(blocklist);
1897 obstack_free(&obst, NULL);
1898 obstack_init(&obst);
1900 block_order = order;
1901 n_block_order = n_blocks;
1905 * Run the register allocator for the current register class.
1907 static void be_straight_alloc_cls(void)
1911 lv = be_assure_liveness(birg);
1912 be_liveness_assure_sets(lv);
1913 be_liveness_assure_chk(lv);
1915 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
1917 DB((dbg, LEVEL_2, "=== Allocating registers of %s ===\n", cls->name));
1919 be_clear_links(irg);
1921 irg_block_walk_graph(irg, NULL, analyze_block, NULL);
1922 if (create_congruence_classes)
1923 combine_congruence_classes();
1925 for (i = 0; i < n_block_order; ++i) {
1926 ir_node *block = block_order[i];
1927 allocate_coalesce_block(block, NULL);
1930 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
1933 static void dump(int mask, ir_graph *irg, const char *suffix,
1934 void (*dumper)(ir_graph *, const char *))
1936 if(birg->main_env->options->dump_flags & mask)
1937 be_dump(irg, suffix, dumper);
1941 * Run the spiller on the current graph.
1943 static void spill(void)
1945 /* make sure all nodes show their real register pressure */
1946 BE_TIMER_PUSH(t_ra_constr);
1947 be_pre_spill_prepare_constr(birg, cls);
1948 BE_TIMER_POP(t_ra_constr);
1950 dump(DUMP_RA, irg, "-spillprepare", dump_ir_block_graph_sched);
1953 BE_TIMER_PUSH(t_ra_spill);
1954 be_do_spill(birg, cls);
1955 BE_TIMER_POP(t_ra_spill);
1957 BE_TIMER_PUSH(t_ra_spill_apply);
1958 check_for_memory_operands(irg);
1959 BE_TIMER_POP(t_ra_spill_apply);
1961 dump(DUMP_RA, irg, "-spill", dump_ir_block_graph_sched);
1965 * The straight register allocator for a whole procedure.
1967 static void be_straight_alloc(be_irg_t *new_birg)
1969 const arch_env_t *arch_env = new_birg->main_env->arch_env;
1970 int n_cls = arch_env_get_n_reg_class(arch_env);
1973 obstack_init(&obst);
1976 irg = be_get_birg_irg(birg);
1977 execfreqs = birg->exec_freq;
1979 /* determine a good coloring order */
1980 determine_block_order();
1982 for (c = 0; c < n_cls; ++c) {
1983 cls = arch_env_get_reg_class(arch_env, c);
1984 default_cls_req = NULL;
1985 if (arch_register_class_flags(cls) & arch_register_class_flag_manual_ra)
1988 stat_ev_ctx_push_str("regcls", cls->name);
1990 n_regs = arch_register_class_n_regs(cls);
1991 normal_regs = rbitset_malloc(n_regs);
1992 be_abi_set_non_ignore_regs(birg->abi, cls, normal_regs);
1996 /* verify schedule and register pressure */
1997 BE_TIMER_PUSH(t_verify);
1998 if (birg->main_env->options->vrfy_option == BE_VRFY_WARN) {
1999 be_verify_schedule(birg);
2000 be_verify_register_pressure(birg, cls, irg);
2001 } else if (birg->main_env->options->vrfy_option == BE_VRFY_ASSERT) {
2002 assert(be_verify_schedule(birg) && "Schedule verification failed");
2003 assert(be_verify_register_pressure(birg, cls, irg)
2004 && "Register pressure verification failed");
2006 BE_TIMER_POP(t_verify);
2008 BE_TIMER_PUSH(t_ra_color);
2009 be_straight_alloc_cls();
2010 BE_TIMER_POP(t_ra_color);
2012 /* we most probably constructed new Phis so liveness info is invalid
2014 /* TODO: test liveness_introduce */
2015 be_liveness_invalidate(lv);
2018 stat_ev_ctx_pop("regcls");
2021 BE_TIMER_PUSH(t_ra_spill_apply);
2022 be_abi_fix_stack_nodes(birg->abi);
2023 BE_TIMER_POP(t_ra_spill_apply);
2025 BE_TIMER_PUSH(t_verify);
2026 if (birg->main_env->options->vrfy_option == BE_VRFY_WARN) {
2027 be_verify_register_allocation(birg);
2028 } else if (birg->main_env->options->vrfy_option == BE_VRFY_ASSERT) {
2029 assert(be_verify_register_allocation(birg)
2030 && "Register allocation invalid");
2032 BE_TIMER_POP(t_verify);
2034 obstack_free(&obst, NULL);
2038 * Initializes this module.
2040 void be_init_straight_alloc(void)
2042 static be_ra_t be_ra_straight = {
2045 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
2046 lc_opt_entry_t *straightalloc_group = lc_opt_get_grp(be_grp, "straightalloc");
2047 lc_opt_add_table(straightalloc_group, options);
2049 be_register_allocator("straight", &be_ra_straight);
2050 FIRM_DBG_REGISTER(dbg, "firm.be.straightalloc");
2053 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_straight_alloc);