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 2
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 bool create_preferences = true;
106 static bool create_congruence_classes = true;
107 static bool 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 check_defs(&live_nodes, weight, node);
370 arity = get_irn_arity(node);
372 /* the allocation info node currently only uses 1 unsigned value
373 to mark last used inputs. So we will fail for a node with more than
375 if (arity >= (int) sizeof(unsigned) * 8) {
376 panic("Node with more than %d inputs not supported yet",
377 (int) sizeof(unsigned) * 8);
380 info = get_allocation_info(node);
381 for (i = 0; i < arity; ++i) {
382 ir_node *op = get_irn_n(node, i);
383 if (!arch_irn_consider_in_reg_alloc(cls, op))
386 /* last usage of a value? */
387 if (!ir_nodeset_contains(&live_nodes, op)) {
388 rbitset_set(&info->last_uses, i);
392 be_liveness_transfer(cls, node, &live_nodes);
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, limited,
413 ir_nodeset_destroy(&live_nodes);
416 static void congruence_def(ir_nodeset_t *live_nodes, ir_node *node)
418 const arch_register_req_t *req;
420 if (get_irn_mode(node) == mode_T) {
421 const ir_edge_t *edge;
422 foreach_out_edge(node, edge) {
423 ir_node *def = get_edge_src_irn(edge);
424 congruence_def(live_nodes, def);
429 if (!arch_irn_consider_in_reg_alloc(cls, node))
432 /* should be same constraint? */
433 req = arch_get_register_req_out(node);
434 if (req->type & arch_register_req_type_should_be_same) {
435 ir_node *insn = skip_Proj(node);
436 int arity = get_irn_arity(insn);
438 unsigned node_idx = get_irn_idx(node);
439 node_idx = uf_find(congruence_classes, node_idx);
441 for (i = 0; i < arity; ++i) {
445 ir_nodeset_iterator_t iter;
446 bool interferes = false;
448 if (!rbitset_is_set(&req->other_same, i))
451 op = get_irn_n(insn, i);
452 op_idx = get_irn_idx(op);
453 op_idx = uf_find(congruence_classes, op_idx);
455 /* do we interfere with the value */
456 foreach_ir_nodeset(live_nodes, live, iter) {
457 int lv_idx = get_irn_idx(live);
458 lv_idx = uf_find(congruence_classes, lv_idx);
459 if (lv_idx == op_idx) {
464 /* don't put in same affinity class if we interfere */
468 node_idx = uf_union(congruence_classes, node_idx, op_idx);
469 DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
471 /* one should_be_same is enough... */
477 static void create_congurence_class(ir_node *block, void *data)
479 ir_nodeset_t live_nodes;
483 ir_nodeset_init(&live_nodes);
484 be_liveness_end_of_block(lv, cls, block, &live_nodes);
486 /* check should be same constraints */
487 sched_foreach_reverse(block, node) {
491 congruence_def(&live_nodes, node);
492 be_liveness_transfer(cls, node, &live_nodes);
495 /* check phi congruence classes */
496 sched_foreach_reverse_from(node, node) {
500 assert(is_Phi(node));
502 if (!arch_irn_consider_in_reg_alloc(cls, node))
505 node_idx = get_irn_idx(node);
506 node_idx = uf_find(congruence_classes, node_idx);
508 arity = get_irn_arity(node);
509 for (i = 0; i < arity; ++i) {
510 bool interferes = false;
511 ir_nodeset_iterator_t iter;
514 ir_node *op = get_Phi_pred(node, i);
515 int op_idx = get_irn_idx(op);
516 op_idx = uf_find(congruence_classes, op_idx);
518 /* do we interfere with the value */
519 foreach_ir_nodeset(&live_nodes, live, iter) {
520 int lv_idx = get_irn_idx(live);
521 lv_idx = uf_find(congruence_classes, lv_idx);
522 if (lv_idx == op_idx) {
527 /* don't put in same affinity class if we interfere */
530 /* any other phi has the same input? */
531 sched_foreach(block, phi) {
536 if (!arch_irn_consider_in_reg_alloc(cls, phi))
538 oop = get_Phi_pred(phi, i);
541 oop_idx = get_irn_idx(oop);
542 oop_idx = uf_find(congruence_classes, oop_idx);
543 if (oop_idx == op_idx) {
551 node_idx = uf_union(congruence_classes, node_idx, op_idx);
552 DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
558 static void merge_congruence_prefs(ir_node *node, void *data)
560 allocation_info_t *info;
561 allocation_info_t *head_info;
562 unsigned node_idx = get_irn_idx(node);
563 unsigned node_set = uf_find(congruence_classes, node_idx);
568 /* head of congruence class or not in any class */
569 if (node_set == node_idx)
572 if (!arch_irn_consider_in_reg_alloc(cls, node))
575 head_info = get_allocation_info(get_idx_irn(irg, node_set));
576 info = get_allocation_info(node);
578 for (r = 0; r < n_regs; ++r) {
579 head_info->prefs[r] += info->prefs[r];
583 static void set_congruence_prefs(ir_node *node, void *data)
585 allocation_info_t *info;
586 allocation_info_t *head_info;
587 unsigned node_idx = get_irn_idx(node);
588 unsigned node_set = uf_find(congruence_classes, node_idx);
592 /* head of congruence class or not in any class */
593 if (node_set == node_idx)
596 if (!arch_irn_consider_in_reg_alloc(cls, node))
599 head_info = get_allocation_info(get_idx_irn(irg, node_set));
600 info = get_allocation_info(node);
602 memcpy(info->prefs, head_info->prefs, n_regs * sizeof(info->prefs[0]));
605 static void combine_congruence_classes(void)
607 size_t n = get_irg_last_idx(irg);
608 congruence_classes = XMALLOCN(int, n);
609 uf_init(congruence_classes, n);
611 /* create congruence classes */
612 irg_block_walk_graph(irg, create_congurence_class, NULL, NULL);
613 /* merge preferences */
614 irg_walk_graph(irg, merge_congruence_prefs, NULL, NULL);
615 irg_walk_graph(irg, set_congruence_prefs, NULL, NULL);
616 free(congruence_classes);
624 * Assign register reg to the given node.
626 * @param node the node
627 * @param reg the register
629 static void use_reg(ir_node *node, const arch_register_t *reg)
631 unsigned r = arch_register_get_index(reg);
632 assignments[r] = node;
633 arch_set_irn_register(node, reg);
636 static void free_reg_of_value(ir_node *node)
638 const arch_register_t *reg;
641 if (!arch_irn_consider_in_reg_alloc(cls, node))
644 reg = arch_get_irn_register(node);
645 r = arch_register_get_index(reg);
646 /* assignment->value may be NULL if a value is used at 2 inputs
647 so it gets freed twice. */
648 assert(assignments[r] == node || assignments[r] == NULL);
649 assignments[r] = NULL;
653 * Compare two register preferences in decreasing order.
655 static int compare_reg_pref(const void *e1, const void *e2)
657 const reg_pref_t *rp1 = (const reg_pref_t*) e1;
658 const reg_pref_t *rp2 = (const reg_pref_t*) e2;
659 if (rp1->pref < rp2->pref)
661 if (rp1->pref > rp2->pref)
666 static void fill_sort_candidates(reg_pref_t *regprefs,
667 const allocation_info_t *info)
671 for (r = 0; r < n_regs; ++r) {
672 float pref = info->prefs[r];
674 regprefs[r].pref = pref;
676 /* TODO: use a stable sort here to avoid unnecessary register jumping */
677 qsort(regprefs, n_regs, sizeof(regprefs[0]), compare_reg_pref);
680 static bool try_optimistic_split(ir_node *to_split, ir_node *before,
681 float pref, float pref_delta,
682 unsigned *forbidden_regs, int recursion)
684 const arch_register_t *from_reg;
685 const arch_register_t *reg;
686 ir_node *original_insn;
692 allocation_info_t *info = get_allocation_info(to_split);
695 float split_threshold;
699 /* stupid hack: don't optimisticallt split don't spill nodes...
700 * (so we don't split away the values produced because of
701 * must_be_different constraints) */
702 original_insn = skip_Proj(info->original_value);
703 if (arch_irn_get_flags(original_insn) & arch_irn_flags_dont_spill)
706 from_reg = arch_get_irn_register(to_split);
707 from_r = arch_register_get_index(from_reg);
708 block = get_nodes_block(before);
709 split_threshold = get_block_execfreq(execfreqs, block) * SPLIT_DELTA;
711 if (pref_delta < split_threshold*0.5)
714 /* find the best free position where we could move to */
715 prefs = ALLOCAN(reg_pref_t, n_regs);
716 fill_sort_candidates(prefs, info);
717 for (i = 0; i < n_regs; ++i) {
721 bool old_source_state;
723 /* we need a normal register which is not an output register
724 an different from the current register of to_split */
726 if (!rbitset_is_set(normal_regs, r))
728 if (rbitset_is_set(forbidden_regs, r))
733 /* is the split worth it? */
734 delta = pref_delta + prefs[i].pref;
735 if (delta < split_threshold) {
736 DB((dbg, LEVEL_3, "Not doing optimistical split of %+F (depth %d), win %f too low\n",
737 to_split, recursion, delta));
741 /* if the register is free then we can do the split */
742 if (assignments[r] == NULL)
745 /* otherwise we might try recursively calling optimistic_split */
746 if (recursion+1 > MAX_OPTIMISTIC_SPLIT_RECURSION)
749 apref = prefs[i].pref;
750 apref_delta = i+1 < n_regs ? apref - prefs[i+1].pref : 0;
751 apref_delta += pref_delta - split_threshold;
753 /* our source register isn't a usefull destination for recursive
755 old_source_state = rbitset_is_set(forbidden_regs, from_r);
756 rbitset_set(forbidden_regs, from_r);
757 /* try recursive split */
758 res = try_optimistic_split(assignments[r], before, apref,
759 apref_delta, forbidden_regs, recursion+1);
760 /* restore our destination */
761 if (old_source_state) {
762 rbitset_set(forbidden_regs, from_r);
764 rbitset_clear(forbidden_regs, from_r);
773 reg = arch_register_for_index(cls, r);
774 copy = be_new_Copy(cls, block, to_split);
775 mark_as_copy_of(copy, to_split);
776 /* hacky, but correct here */
777 if (assignments[arch_register_get_index(from_reg)] == to_split)
778 free_reg_of_value(to_split);
780 sched_add_before(before, copy);
783 "Optimistic live-range split %+F move %+F(%s) -> %s before %+F (win %f, depth %d)\n",
784 copy, to_split, from_reg->name, reg->name, before, delta, recursion));
789 * Determine and assign a register for node @p node
791 static void assign_reg(const ir_node *block, ir_node *node,
792 unsigned *forbidden_regs)
794 const arch_register_t *reg;
795 allocation_info_t *info;
796 const arch_register_req_t *req;
797 reg_pref_t *reg_prefs;
800 const unsigned *allowed_regs;
803 assert(!is_Phi(node));
804 assert(arch_irn_consider_in_reg_alloc(cls, node));
806 /* preassigned register? */
807 reg = arch_get_irn_register(node);
809 DB((dbg, LEVEL_2, "Preassignment %+F -> %s\n", node, reg->name));
814 /* give should_be_same boni */
815 info = get_allocation_info(node);
816 req = arch_get_register_req_out(node);
818 in_node = skip_Proj(node);
819 if (req->type & arch_register_req_type_should_be_same) {
820 float weight = get_block_execfreq(execfreqs, block);
821 int arity = get_irn_arity(in_node);
824 assert(arity <= (int) sizeof(req->other_same) * 8);
825 for (i = 0; i < arity; ++i) {
827 const arch_register_t *reg;
829 if (!rbitset_is_set(&req->other_same, i))
832 in = get_irn_n(in_node, i);
833 reg = arch_get_irn_register(in);
835 r = arch_register_get_index(reg);
837 /* if the value didn't die here then we should not propagate the
838 * should_be_same info */
839 if (assignments[r] == in)
842 info->prefs[r] += weight * AFF_SHOULD_BE_SAME;
846 /* create list of register candidates and sort by their preference */
847 DB((dbg, LEVEL_2, "Candidates for %+F:", node));
848 reg_prefs = alloca(n_regs * sizeof(reg_prefs[0]));
849 fill_sort_candidates(reg_prefs, info);
850 for (i = 0; i < n_regs; ++i) {
851 unsigned num = reg_prefs[i].num;
852 const arch_register_t *reg;
854 if (!rbitset_is_set(normal_regs, num))
857 reg = arch_register_for_index(cls, num);
858 DB((dbg, LEVEL_2, " %s(%f)", reg->name, reg_prefs[i].pref));
860 DB((dbg, LEVEL_2, "\n"));
862 allowed_regs = normal_regs;
863 if (req->type & arch_register_req_type_limited) {
864 allowed_regs = req->limited;
867 for (i = 0; i < n_regs; ++i) {
868 r = reg_prefs[i].num;
869 if (!rbitset_is_set(allowed_regs, r))
871 if (assignments[r] == NULL)
873 float pref = reg_prefs[i].pref;
874 float delta = i+1 < n_regs ? pref - reg_prefs[i+1].pref : 0;
875 ir_node *before = skip_Proj(node);
876 bool res = try_optimistic_split(assignments[r], before,
877 pref, delta, forbidden_regs, 0);
882 panic("No register left for %+F\n", node);
885 reg = arch_register_for_index(cls, r);
886 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
891 * Add an permutation in front of a node and change the assignments
892 * due to this permutation.
894 * To understand this imagine a permutation like this:
904 * First we count how many destinations a single value has. At the same time
905 * we can be sure that each destination register has at most 1 source register
906 * (it can have 0 which means we don't care what value is in it).
907 * We ignore all fullfilled permuations (like 7->7)
908 * In a first pass we create as much copy instructions as possible as they
909 * are generally cheaper than exchanges. We do this by counting into how many
910 * destinations a register has to be copied (in the example it's 2 for register
911 * 3, or 1 for the registers 1,2,4 and 7).
912 * We can then create a copy into every destination register when the usecount
913 * of that register is 0 (= noone else needs the value in the register).
915 * After this step we should have cycles left. We implement a cyclic permutation
916 * of n registers with n-1 transpositions.
918 * @param live_nodes the set of live nodes, updated due to live range split
919 * @param before the node before we add the permutation
920 * @param permutation the permutation array indices are the destination
921 * registers, the values in the array are the source
924 static void permute_values(ir_nodeset_t *live_nodes, ir_node *before,
925 unsigned *permutation)
927 unsigned *n_used = ALLOCANZ(unsigned, n_regs);
931 /* determine how often each source register needs to be read */
932 for (r = 0; r < n_regs; ++r) {
933 unsigned old_reg = permutation[r];
936 value = assignments[old_reg];
938 /* nothing to do here, reg is not live. Mark it as fixpoint
939 * so we ignore it in the next steps */
947 block = get_nodes_block(before);
949 /* step1: create copies where immediately possible */
950 for (r = 0; r < n_regs; /* empty */) {
953 const arch_register_t *reg;
954 unsigned old_r = permutation[r];
956 /* - no need to do anything for fixed points.
957 - we can't copy if the value in the dest reg is still needed */
958 if (old_r == r || n_used[r] > 0) {
964 src = assignments[old_r];
965 copy = be_new_Copy(cls, block, src);
966 sched_add_before(before, copy);
967 reg = arch_register_for_index(cls, r);
968 DB((dbg, LEVEL_2, "Copy %+F (from %+F, before %+F) -> %s\n",
969 copy, src, before, reg->name));
970 mark_as_copy_of(copy, src);
973 if (live_nodes != NULL) {
974 ir_nodeset_insert(live_nodes, copy);
977 /* old register has 1 user less, permutation is resolved */
978 assert(arch_register_get_index(arch_get_irn_register(src)) == old_r);
981 assert(n_used[old_r] > 0);
983 if (n_used[old_r] == 0) {
984 if (live_nodes != NULL) {
985 ir_nodeset_remove(live_nodes, src);
987 free_reg_of_value(src);
990 /* advance or jump back (if this copy enabled another copy) */
991 if (old_r < r && n_used[old_r] == 0) {
998 /* at this point we only have "cycles" left which we have to resolve with
1000 * TODO: if we have free registers left, then we should really use copy
1001 * instructions for any cycle longer than 2 registers...
1002 * (this is probably architecture dependent, there might be archs where
1003 * copies are preferable even for 2-cycles) */
1005 /* create perms with the rest */
1006 for (r = 0; r < n_regs; /* empty */) {
1007 const arch_register_t *reg;
1008 unsigned old_r = permutation[r];
1020 /* we shouldn't have copies from 1 value to multiple destinations left*/
1021 assert(n_used[old_r] == 1);
1023 /* exchange old_r and r2; after that old_r is a fixed point */
1024 r2 = permutation[old_r];
1026 in[0] = assignments[r2];
1027 in[1] = assignments[old_r];
1028 perm = be_new_Perm(cls, block, 2, in);
1029 sched_add_before(before, perm);
1030 DB((dbg, LEVEL_2, "Perm %+F (perm %+F,%+F, before %+F)\n",
1031 perm, in[0], in[1], before));
1033 proj0 = new_r_Proj(block, perm, get_irn_mode(in[0]), 0);
1034 mark_as_copy_of(proj0, in[0]);
1035 reg = arch_register_for_index(cls, old_r);
1036 use_reg(proj0, reg);
1038 proj1 = new_r_Proj(block, perm, get_irn_mode(in[1]), 1);
1039 mark_as_copy_of(proj1, in[1]);
1040 reg = arch_register_for_index(cls, r2);
1041 use_reg(proj1, reg);
1043 /* 1 value is now in the correct register */
1044 permutation[old_r] = old_r;
1045 /* the source of r changed to r2 */
1046 permutation[r] = r2;
1048 /* if we have reached a fixpoint update data structures */
1049 if (live_nodes != NULL) {
1050 ir_nodeset_remove(live_nodes, in[0]);
1051 ir_nodeset_remove(live_nodes, in[1]);
1052 ir_nodeset_remove(live_nodes, proj0);
1053 ir_nodeset_insert(live_nodes, proj1);
1057 #ifdef DEBUG_libfirm
1058 /* now we should only have fixpoints left */
1059 for (r = 0; r < n_regs; ++r) {
1060 assert(permutation[r] == r);
1066 * Free regs for values last used.
1068 * @param live_nodes set of live nodes, will be updated
1069 * @param node the node to consider
1071 static void free_last_uses(ir_nodeset_t *live_nodes, ir_node *node)
1073 allocation_info_t *info = get_allocation_info(node);
1074 const unsigned *last_uses = &info->last_uses;
1075 int arity = get_irn_arity(node);
1078 for (i = 0; i < arity; ++i) {
1081 /* check if one operand is the last use */
1082 if (!rbitset_is_set(last_uses, i))
1085 op = get_irn_n(node, i);
1086 free_reg_of_value(op);
1087 ir_nodeset_remove(live_nodes, op);
1092 * change inputs of a node to the current value (copies/perms)
1094 static void rewire_inputs(ir_node *node)
1097 int arity = get_irn_arity(node);
1099 for (i = 0; i < arity; ++i) {
1100 ir_node *op = get_irn_n(node, i);
1101 allocation_info_t *info;
1103 if (!arch_irn_consider_in_reg_alloc(cls, op))
1106 info = get_allocation_info(op);
1107 info = get_allocation_info(info->original_value);
1108 if (info->current_value != op) {
1109 set_irn_n(node, i, info->current_value);
1115 * Create a bitset of registers occupied with value living through an
1118 static void determine_live_through_regs(unsigned *bitset, ir_node *node)
1120 const allocation_info_t *info = get_allocation_info(node);
1125 /* mark all used registers as potentially live-through */
1126 for (r = 0; r < n_regs; ++r) {
1127 if (assignments[r] == NULL)
1129 if (!rbitset_is_set(normal_regs, r))
1132 rbitset_set(bitset, r);
1135 /* remove registers of value dying at the instruction */
1136 arity = get_irn_arity(node);
1137 for (i = 0; i < arity; ++i) {
1139 const arch_register_t *reg;
1141 if (!rbitset_is_set(&info->last_uses, i))
1144 op = get_irn_n(node, i);
1145 reg = arch_get_irn_register(op);
1146 rbitset_clear(bitset, arch_register_get_index(reg));
1151 * Enforce constraints at a node by live range splits.
1153 * @param live_nodes the set of live nodes, might be changed
1154 * @param node the current node
1156 static void enforce_constraints(ir_nodeset_t *live_nodes, ir_node *node,
1157 unsigned *forbidden_regs)
1159 int arity = get_irn_arity(node);
1161 hungarian_problem_t *bp;
1163 unsigned *assignment;
1165 /* construct a list of register occupied by live-through values */
1166 unsigned *live_through_regs = NULL;
1168 /* see if any use constraints are not met */
1170 for (i = 0; i < arity; ++i) {
1171 ir_node *op = get_irn_n(node, i);
1172 const arch_register_t *reg;
1173 const arch_register_req_t *req;
1174 const unsigned *limited;
1177 if (!arch_irn_consider_in_reg_alloc(cls, op))
1180 /* are there any limitations for the i'th operand? */
1181 req = arch_get_register_req(node, i);
1182 if (!(req->type & arch_register_req_type_limited))
1185 limited = req->limited;
1186 reg = arch_get_irn_register(op);
1187 r = arch_register_get_index(reg);
1188 if (!rbitset_is_set(limited, r)) {
1189 /* found an assignment outside the limited set */
1195 /* is any of the live-throughs using a constrained output register? */
1196 if (get_irn_mode(node) == mode_T) {
1197 const ir_edge_t *edge;
1199 foreach_out_edge(node, edge) {
1200 ir_node *proj = get_edge_src_irn(edge);
1201 const arch_register_req_t *req;
1203 if (!arch_irn_consider_in_reg_alloc(cls, proj))
1206 req = arch_get_register_req_out(proj);
1207 if (!(req->type & arch_register_req_type_limited))
1210 if (live_through_regs == NULL) {
1211 rbitset_alloca(live_through_regs, n_regs);
1212 determine_live_through_regs(live_through_regs, node);
1215 rbitset_or(forbidden_regs, req->limited, n_regs);
1216 if (rbitsets_have_common(req->limited, live_through_regs, n_regs)) {
1221 if (arch_irn_consider_in_reg_alloc(cls, node)) {
1222 const arch_register_req_t *req = arch_get_register_req_out(node);
1223 if (req->type & arch_register_req_type_limited) {
1224 rbitset_alloca(live_through_regs, n_regs);
1225 determine_live_through_regs(live_through_regs, node);
1226 if (rbitsets_have_common(req->limited, live_through_regs, n_regs)) {
1228 rbitset_or(forbidden_regs, req->limited, n_regs);
1237 /* create these arrays if we haven't yet */
1238 if (live_through_regs == NULL) {
1239 rbitset_alloca(live_through_regs, n_regs);
1242 /* at this point we have to construct a bipartite matching problem to see
1243 * which values should go to which registers
1244 * Note: We're building the matrix in "reverse" - source registers are
1245 * right, destinations left because this will produce the solution
1246 * in the format required for permute_values.
1248 bp = hungarian_new(n_regs, n_regs, HUNGARIAN_MATCH_PERFECT);
1250 /* add all combinations, then remove not allowed ones */
1251 for (l = 0; l < n_regs; ++l) {
1252 if (!rbitset_is_set(normal_regs, l)) {
1253 hungarian_add(bp, l, l, 1);
1257 for (r = 0; r < n_regs; ++r) {
1258 if (!rbitset_is_set(normal_regs, r))
1260 /* livethrough values may not use constrainted output registers */
1261 if (rbitset_is_set(live_through_regs, l)
1262 && rbitset_is_set(forbidden_regs, r))
1265 hungarian_add(bp, r, l, l == r ? 9 : 8);
1269 for (i = 0; i < arity; ++i) {
1270 ir_node *op = get_irn_n(node, i);
1271 const arch_register_t *reg;
1272 const arch_register_req_t *req;
1273 const unsigned *limited;
1274 unsigned current_reg;
1276 if (!arch_irn_consider_in_reg_alloc(cls, op))
1279 req = arch_get_register_req(node, i);
1280 if (!(req->type & arch_register_req_type_limited))
1283 limited = req->limited;
1284 reg = arch_get_irn_register(op);
1285 current_reg = arch_register_get_index(reg);
1286 for (r = 0; r < n_regs; ++r) {
1287 if (rbitset_is_set(limited, r))
1289 hungarian_remv(bp, r, current_reg);
1293 //hungarian_print_cost_matrix(bp, 1);
1294 hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1296 assignment = ALLOCAN(unsigned, n_regs);
1297 res = hungarian_solve(bp, (int*) assignment, NULL, 0);
1301 fprintf(stderr, "Swap result:");
1302 for (i = 0; i < (int) n_regs; ++i) {
1303 fprintf(stderr, " %d", assignment[i]);
1305 fprintf(stderr, "\n");
1310 permute_values(live_nodes, node, assignment);
1313 /** test wether a node @p n is a copy of the value of node @p of */
1314 static bool is_copy_of(ir_node *value, ir_node *test_value)
1316 allocation_info_t *test_info;
1317 allocation_info_t *info;
1319 if (value == test_value)
1322 info = get_allocation_info(value);
1323 test_info = get_allocation_info(test_value);
1324 return test_info->original_value == info->original_value;
1328 * find a value in the end-assignment of a basic block
1329 * @returns the index into the assignment array if found
1332 static int find_value_in_block_info(block_info_t *info, ir_node *value)
1335 ir_node **assignments = info->assignments;
1336 for (r = 0; r < n_regs; ++r) {
1337 ir_node *a_value = assignments[r];
1339 if (a_value == NULL)
1341 if (is_copy_of(a_value, value))
1349 * Create the necessary permutations at the end of a basic block to fullfill
1350 * the register assignment for phi-nodes in the next block
1352 static void add_phi_permutations(ir_node *block, int p)
1355 unsigned *permutation;
1356 ir_node **old_assignments;
1357 bool need_permutation;
1359 ir_node *pred = get_Block_cfgpred_block(block, p);
1361 block_info_t *pred_info = get_block_info(pred);
1363 /* predecessor not processed yet? nothing to do */
1364 if (!pred_info->processed)
1367 permutation = ALLOCAN(unsigned, n_regs);
1368 for (r = 0; r < n_regs; ++r) {
1372 /* check phi nodes */
1373 need_permutation = false;
1374 node = sched_first(block);
1375 for ( ; is_Phi(node); node = sched_next(node)) {
1376 const arch_register_t *reg;
1381 if (!arch_irn_consider_in_reg_alloc(cls, node))
1384 op = get_Phi_pred(node, p);
1385 if (!arch_irn_consider_in_reg_alloc(cls, op))
1388 a = find_value_in_block_info(pred_info, op);
1391 reg = arch_get_irn_register(node);
1392 regn = arch_register_get_index(reg);
1394 permutation[regn] = a;
1395 need_permutation = true;
1399 if (need_permutation) {
1400 /* permute values at end of predecessor */
1401 old_assignments = assignments;
1402 assignments = pred_info->assignments;
1403 permute_values(NULL, be_get_end_of_block_insertion_point(pred),
1405 assignments = old_assignments;
1408 /* change phi nodes to use the copied values */
1409 node = sched_first(block);
1410 for ( ; is_Phi(node); node = sched_next(node)) {
1414 if (!arch_irn_consider_in_reg_alloc(cls, node))
1417 op = get_Phi_pred(node, p);
1418 /* no need to do anything for Unknown inputs */
1419 if (!arch_irn_consider_in_reg_alloc(cls, op))
1422 /* we have permuted all values into the correct registers so we can
1423 simply query which value occupies the phis register in the
1425 a = arch_register_get_index(arch_get_irn_register(node));
1426 op = pred_info->assignments[a];
1427 set_Phi_pred(node, p, op);
1432 * Set preferences for a phis register based on the registers used on the
1435 static void adapt_phi_prefs(ir_node *phi)
1438 int arity = get_irn_arity(phi);
1439 ir_node *block = get_nodes_block(phi);
1440 allocation_info_t *info = get_allocation_info(phi);
1442 for (i = 0; i < arity; ++i) {
1443 ir_node *op = get_irn_n(phi, i);
1444 const arch_register_t *reg = arch_get_irn_register(op);
1445 ir_node *pred_block;
1446 block_info_t *pred_block_info;
1452 /* we only give the bonus if the predecessor already has registers
1453 * assigned, otherwise we only see a dummy value
1454 * and any conclusions about its register are useless */
1455 pred_block = get_Block_cfgpred_block(block, i);
1456 pred_block_info = get_block_info(pred_block);
1457 if (!pred_block_info->processed)
1460 /* give bonus for already assigned register */
1461 weight = get_block_execfreq(execfreqs, pred_block);
1462 r = arch_register_get_index(reg);
1463 info->prefs[r] += weight * AFF_PHI;
1468 * After a phi has been assigned a register propagate preference inputs
1469 * to the phi inputs.
1471 static void propagate_phi_register(ir_node *phi, unsigned assigned_r)
1474 ir_node *block = get_nodes_block(phi);
1475 int arity = get_irn_arity(phi);
1477 for (i = 0; i < arity; ++i) {
1478 ir_node *op = get_Phi_pred(phi, i);
1479 allocation_info_t *info = get_allocation_info(op);
1480 ir_node *pred_block = get_Block_cfgpred_block(block, i);
1483 = get_block_execfreq(execfreqs, pred_block) * AFF_PHI;
1485 if (info->prefs[assigned_r] >= weight)
1488 /* promote the prefered register */
1489 for (r = 0; r < n_regs; ++r) {
1490 if (r == assigned_r) {
1491 info->prefs[r] = AFF_PHI * weight;
1493 info->prefs[r] -= AFF_PHI * weight;
1498 propagate_phi_register(op, assigned_r);
1502 static void assign_phi_registers(ir_node *block)
1509 hungarian_problem_t *bp;
1511 /* count phi nodes */
1512 sched_foreach(block, node) {
1515 if (!arch_irn_consider_in_reg_alloc(cls, node))
1523 /* build a bipartite matching problem for all phi nodes */
1524 bp = hungarian_new(n_phis, n_regs, HUNGARIAN_MATCH_PERFECT);
1526 sched_foreach(block, node) {
1529 allocation_info_t *info;
1532 if (!arch_irn_consider_in_reg_alloc(cls, node))
1535 /* give boni for predecessor colorings */
1536 adapt_phi_prefs(node);
1537 /* add stuff to bipartite problem */
1538 info = get_allocation_info(node);
1539 DB((dbg, LEVEL_3, "Prefs for %+F: ", node));
1540 for (r = 0; r < n_regs; ++r) {
1543 if (!rbitset_is_set(normal_regs, r))
1546 costs = info->prefs[r];
1547 costs = costs < 0 ? -logf(-costs+1) : logf(costs+1);
1550 hungarian_add(bp, n, r, costs);
1551 DB((dbg, LEVEL_3, " %s(%f)", arch_register_for_index(cls, r)->name,
1554 DB((dbg, LEVEL_3, "\n"));
1558 //hungarian_print_cost_matrix(bp, 7);
1559 hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1561 assignment = ALLOCAN(int, n_regs);
1562 res = hungarian_solve(bp, assignment, NULL, 0);
1567 sched_foreach(block, node) {
1569 const arch_register_t *reg;
1573 if (!arch_irn_consider_in_reg_alloc(cls, node))
1576 r = assignment[n++];
1577 assert(rbitset_is_set(normal_regs, r));
1578 reg = arch_register_for_index(cls, r);
1579 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
1582 /* adapt preferences for phi inputs */
1583 if (propagate_phi_registers)
1584 propagate_phi_register(node, r);
1589 * Walker: assign registers to all nodes of a block that
1590 * need registers from the currently considered register class.
1592 static void allocate_coalesce_block(ir_node *block, void *data)
1595 ir_nodeset_t live_nodes;
1596 ir_nodeset_iterator_t iter;
1599 block_info_t *block_info;
1600 block_info_t **pred_block_infos;
1602 unsigned *forbidden_regs; /**< collects registers which must
1603 not be used for optimistic splits */
1606 DB((dbg, LEVEL_2, "* Block %+F\n", block));
1608 /* clear assignments */
1609 block_info = get_block_info(block);
1610 assignments = block_info->assignments;
1612 ir_nodeset_init(&live_nodes);
1614 /* gather regalloc infos of predecessor blocks */
1615 n_preds = get_Block_n_cfgpreds(block);
1616 pred_block_infos = ALLOCAN(block_info_t*, n_preds);
1617 for (i = 0; i < n_preds; ++i) {
1618 ir_node *pred = get_Block_cfgpred_block(block, i);
1619 block_info_t *pred_info = get_block_info(pred);
1620 pred_block_infos[i] = pred_info;
1623 phi_ins = ALLOCAN(ir_node*, n_preds);
1625 /* collect live-in nodes and preassigned values */
1626 be_lv_foreach(lv, block, be_lv_state_in, i) {
1627 const arch_register_t *reg;
1629 bool need_phi = false;
1631 node = be_lv_get_irn(lv, block, i);
1632 if (!arch_irn_consider_in_reg_alloc(cls, node))
1635 /* check all predecessors for this value, if it is not everywhere the
1636 same or unknown then we have to construct a phi
1637 (we collect the potential phi inputs here) */
1638 for (p = 0; p < n_preds; ++p) {
1639 block_info_t *pred_info = pred_block_infos[p];
1641 if (!pred_info->processed) {
1642 /* use node for now, it will get fixed later */
1646 int a = find_value_in_block_info(pred_info, node);
1648 /* must live out of predecessor */
1650 phi_ins[p] = pred_info->assignments[a];
1651 /* different value from last time? then we need a phi */
1652 if (p > 0 && phi_ins[p-1] != phi_ins[p]) {
1659 ir_mode *mode = get_irn_mode(node);
1660 const arch_register_req_t *req = get_default_req_current_cls();
1664 phi = new_r_Phi(block, n_preds, phi_ins, mode);
1665 be_set_phi_reg_req(phi, req);
1667 DB((dbg, LEVEL_3, "Create Phi %+F (for %+F) -", phi, node));
1668 #ifdef DEBUG_libfirm
1669 for (i = 0; i < n_preds; ++i) {
1670 DB((dbg, LEVEL_3, " %+F", phi_ins[i]));
1672 DB((dbg, LEVEL_3, "\n"));
1674 mark_as_copy_of(phi, node);
1675 sched_add_after(block, phi);
1679 allocation_info_t *info = get_allocation_info(node);
1680 info->current_value = phi_ins[0];
1682 /* Grab 1 of the inputs we constructed (might not be the same as
1683 * "node" as we could see the same copy of the value in all
1688 /* if the node already has a register assigned use it */
1689 reg = arch_get_irn_register(node);
1694 /* remember that this node is live at the beginning of the block */
1695 ir_nodeset_insert(&live_nodes, node);
1698 rbitset_alloca(forbidden_regs, n_regs);
1700 /* handle phis... */
1701 assign_phi_registers(block);
1703 /* all live-ins must have a register */
1704 #ifdef DEBUG_libfirm
1705 foreach_ir_nodeset(&live_nodes, node, iter) {
1706 const arch_register_t *reg = arch_get_irn_register(node);
1707 assert(reg != NULL);
1711 /* assign instructions in the block */
1712 sched_foreach(block, node) {
1716 /* phis are already assigned */
1720 rewire_inputs(node);
1722 /* enforce use constraints */
1723 rbitset_clear_all(forbidden_regs, n_regs);
1724 enforce_constraints(&live_nodes, node, forbidden_regs);
1726 rewire_inputs(node);
1728 /* we may not use registers used for inputs for optimistic splits */
1729 arity = get_irn_arity(node);
1730 for (i = 0; i < arity; ++i) {
1731 ir_node *op = get_irn_n(node, i);
1732 const arch_register_t *reg;
1733 if (!arch_irn_consider_in_reg_alloc(cls, op))
1736 reg = arch_get_irn_register(op);
1737 rbitset_set(forbidden_regs, arch_register_get_index(reg));
1740 /* free registers of values last used at this instruction */
1741 free_last_uses(&live_nodes, node);
1743 /* assign output registers */
1744 /* TODO: 2 phases: first: pre-assigned ones, 2nd real regs */
1745 if (get_irn_mode(node) == mode_T) {
1746 const ir_edge_t *edge;
1747 foreach_out_edge(node, edge) {
1748 ir_node *proj = get_edge_src_irn(edge);
1749 if (!arch_irn_consider_in_reg_alloc(cls, proj))
1751 assign_reg(block, proj, forbidden_regs);
1753 } else if (arch_irn_consider_in_reg_alloc(cls, node)) {
1754 assign_reg(block, node, forbidden_regs);
1758 ir_nodeset_destroy(&live_nodes);
1761 block_info->processed = true;
1763 /* permute values at end of predecessor blocks in case of phi-nodes */
1766 for (p = 0; p < n_preds; ++p) {
1767 add_phi_permutations(block, p);
1771 /* if we have exactly 1 successor then we might be able to produce phi
1773 if (get_irn_n_edges_kind(block, EDGE_KIND_BLOCK) == 1) {
1774 const ir_edge_t *edge
1775 = get_irn_out_edge_first_kind(block, EDGE_KIND_BLOCK);
1776 ir_node *succ = get_edge_src_irn(edge);
1777 int p = get_edge_src_pos(edge);
1778 block_info_t *succ_info = get_block_info(succ);
1780 if (succ_info->processed) {
1781 add_phi_permutations(succ, p);
1786 typedef struct block_costs_t block_costs_t;
1787 struct block_costs_t {
1788 float costs; /**< costs of the block */
1789 int dfs_num; /**< depth first search number (to detect backedges) */
1792 static int cmp_block_costs(const void *d1, const void *d2)
1794 const ir_node * const *block1 = d1;
1795 const ir_node * const *block2 = d2;
1796 const block_costs_t *info1 = get_irn_link(*block1);
1797 const block_costs_t *info2 = get_irn_link(*block2);
1798 return QSORT_CMP(info2->costs, info1->costs);
1801 static void determine_block_order(void)
1804 ir_node **blocklist = be_get_cfgpostorder(irg);
1805 int n_blocks = ARR_LEN(blocklist);
1807 pdeq *worklist = new_pdeq();
1808 ir_node **order = XMALLOCN(ir_node*, n_blocks);
1811 /* clear block links... */
1812 for (i = 0; i < n_blocks; ++i) {
1813 ir_node *block = blocklist[i];
1814 set_irn_link(block, NULL);
1817 /* walk blocks in reverse postorder, the costs for each block are the
1818 * sum of the costs of its predecessors (excluding the costs on backedges
1819 * which we can't determine) */
1820 for (i = n_blocks-1; i >= 0; --i) {
1821 block_costs_t *cost_info;
1822 ir_node *block = blocklist[i];
1824 float execfreq = get_block_execfreq(execfreqs, block);
1825 float costs = execfreq;
1826 int n_cfgpreds = get_Block_n_cfgpreds(block);
1828 for (p = 0; p < n_cfgpreds; ++p) {
1829 ir_node *pred_block = get_Block_cfgpred_block(block, p);
1830 block_costs_t *pred_costs = get_irn_link(pred_block);
1831 /* we don't have any info for backedges */
1832 if (pred_costs == NULL)
1834 costs += pred_costs->costs;
1837 cost_info = OALLOCZ(&obst, block_costs_t);
1838 cost_info->costs = costs;
1839 cost_info->dfs_num = dfs_num++;
1840 set_irn_link(block, cost_info);
1843 /* sort array by block costs */
1844 qsort(blocklist, n_blocks, sizeof(blocklist[0]), cmp_block_costs);
1846 ir_reserve_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1847 inc_irg_block_visited(irg);
1849 for (i = 0; i < n_blocks; ++i) {
1850 ir_node *block = blocklist[i];
1851 if (Block_block_visited(block))
1854 /* continually add predecessors with highest costs to worklist
1855 * (without using backedges) */
1857 block_costs_t *info = get_irn_link(block);
1858 ir_node *best_pred = NULL;
1859 float best_costs = -1;
1860 int n_cfgpred = get_Block_n_cfgpreds(block);
1863 pdeq_putr(worklist, block);
1864 mark_Block_block_visited(block);
1865 for (i = 0; i < n_cfgpred; ++i) {
1866 ir_node *pred_block = get_Block_cfgpred_block(block, i);
1867 block_costs_t *pred_info = get_irn_link(pred_block);
1869 /* ignore backedges */
1870 if (pred_info->dfs_num > info->dfs_num)
1873 if (info->costs > best_costs) {
1874 best_costs = info->costs;
1875 best_pred = pred_block;
1879 } while(block != NULL && !Block_block_visited(block));
1881 /* now put all nodes in the worklist in our final order */
1882 while (!pdeq_empty(worklist)) {
1883 ir_node *pblock = pdeq_getr(worklist);
1884 assert(order_p < n_blocks);
1885 order[order_p++] = pblock;
1888 assert(order_p == n_blocks);
1891 ir_free_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1893 DEL_ARR_F(blocklist);
1895 obstack_free(&obst, NULL);
1896 obstack_init(&obst);
1898 block_order = order;
1899 n_block_order = n_blocks;
1903 * Run the register allocator for the current register class.
1905 static void be_straight_alloc_cls(void)
1909 lv = be_assure_liveness(birg);
1910 be_liveness_assure_sets(lv);
1911 be_liveness_assure_chk(lv);
1913 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
1915 DB((dbg, LEVEL_2, "=== Allocating registers of %s ===\n", cls->name));
1917 be_clear_links(irg);
1919 if (create_preferences)
1920 irg_block_walk_graph(irg, NULL, analyze_block, NULL);
1921 if (create_congruence_classes)
1922 combine_congruence_classes();
1924 for (i = 0; i < n_block_order; ++i) {
1925 ir_node *block = block_order[i];
1926 allocate_coalesce_block(block, NULL);
1929 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
1932 static void dump(int mask, ir_graph *irg, const char *suffix,
1933 void (*dumper)(ir_graph *, const char *))
1935 if(birg->main_env->options->dump_flags & mask)
1936 be_dump(irg, suffix, dumper);
1940 * Run the spiller on the current graph.
1942 static void spill(void)
1944 /* make sure all nodes show their real register pressure */
1945 BE_TIMER_PUSH(t_ra_constr);
1946 be_pre_spill_prepare_constr(birg, cls);
1947 BE_TIMER_POP(t_ra_constr);
1949 dump(DUMP_RA, irg, "-spillprepare", dump_ir_block_graph_sched);
1952 BE_TIMER_PUSH(t_ra_spill);
1953 be_do_spill(birg, cls);
1954 BE_TIMER_POP(t_ra_spill);
1956 BE_TIMER_PUSH(t_ra_spill_apply);
1957 check_for_memory_operands(irg);
1958 BE_TIMER_POP(t_ra_spill_apply);
1960 dump(DUMP_RA, irg, "-spill", dump_ir_block_graph_sched);
1964 * The straight register allocator for a whole procedure.
1966 static void be_straight_alloc(be_irg_t *new_birg)
1968 const arch_env_t *arch_env = new_birg->main_env->arch_env;
1969 int n_cls = arch_env_get_n_reg_class(arch_env);
1972 obstack_init(&obst);
1975 irg = be_get_birg_irg(birg);
1976 execfreqs = birg->exec_freq;
1978 /* determine a good coloring order */
1979 determine_block_order();
1981 for (c = 0; c < n_cls; ++c) {
1982 cls = arch_env_get_reg_class(arch_env, c);
1983 default_cls_req = NULL;
1984 if (arch_register_class_flags(cls) & arch_register_class_flag_manual_ra)
1987 stat_ev_ctx_push_str("regcls", cls->name);
1989 n_regs = arch_register_class_n_regs(cls);
1990 normal_regs = rbitset_malloc(n_regs);
1991 be_abi_set_non_ignore_regs(birg->abi, cls, normal_regs);
1995 /* verify schedule and register pressure */
1996 BE_TIMER_PUSH(t_verify);
1997 if (birg->main_env->options->vrfy_option == BE_VRFY_WARN) {
1998 be_verify_schedule(birg);
1999 be_verify_register_pressure(birg, cls, irg);
2000 } else if (birg->main_env->options->vrfy_option == BE_VRFY_ASSERT) {
2001 assert(be_verify_schedule(birg) && "Schedule verification failed");
2002 assert(be_verify_register_pressure(birg, cls, irg)
2003 && "Register pressure verification failed");
2005 BE_TIMER_POP(t_verify);
2007 BE_TIMER_PUSH(t_ra_color);
2008 be_straight_alloc_cls();
2009 BE_TIMER_POP(t_ra_color);
2011 /* we most probably constructed new Phis so liveness info is invalid
2013 /* TODO: test liveness_introduce */
2014 be_liveness_invalidate(lv);
2017 stat_ev_ctx_pop("regcls");
2020 BE_TIMER_PUSH(t_ra_spill_apply);
2021 be_abi_fix_stack_nodes(birg->abi);
2022 BE_TIMER_POP(t_ra_spill_apply);
2024 BE_TIMER_PUSH(t_verify);
2025 if (birg->main_env->options->vrfy_option == BE_VRFY_WARN) {
2026 be_verify_register_allocation(birg);
2027 } else if (birg->main_env->options->vrfy_option == BE_VRFY_ASSERT) {
2028 assert(be_verify_register_allocation(birg)
2029 && "Register allocation invalid");
2031 BE_TIMER_POP(t_verify);
2033 obstack_free(&obst, NULL);
2037 * Initializes this module.
2039 void be_init_straight_alloc(void)
2041 static be_ra_t be_ra_straight = {
2044 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
2045 lc_opt_entry_t *straightalloc_group = lc_opt_get_grp(be_grp, "straightalloc");
2046 lc_opt_add_table(straightalloc_group, options);
2048 be_register_allocator("straight", &be_ra_straight);
2049 FIRM_DBG_REGISTER(dbg, "firm.be.straightalloc");
2052 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_straight_alloc);