2 * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Preference Guided Register Assignment
23 * @author Matthias Braun
26 * The idea is to allocate registers in 2 passes:
27 * 1. A first pass to determine "preferred" registers for live-ranges. This
28 * calculates for each register and each live-range a value indicating
29 * the usefulness. (You can roughly think of the value as the negative
30 * costs needed for copies when the value is in the specific registers...)
32 * 2. Walk blocks and assigns registers in a greedy fashion. Preferring
33 * registers with high preferences. When register constraints are not met,
34 * add copies and split live-ranges.
37 * - make use of free registers in the permute_values code
50 #include "iredges_t.h"
51 #include "irgraph_t.h"
59 #include "raw_bitset.h"
60 #include "unionfind.h"
62 #include "hungarian.h"
65 #include "bechordal_t.h"
74 #include "bespillutil.h"
79 #define USE_FACTOR 1.0f
80 #define DEF_FACTOR 1.0f
81 #define NEIGHBOR_FACTOR 0.2f
82 #define AFF_SHOULD_BE_SAME 0.5f
84 #define SPLIT_DELTA 1.0f
85 #define MAX_OPTIMISTIC_SPLIT_RECURSION 0
87 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
89 static struct obstack obst;
91 static const arch_register_class_t *cls;
93 static unsigned n_regs;
94 static unsigned *normal_regs;
95 static int *congruence_classes;
96 static ir_node **block_order;
97 static size_t n_block_order;
99 /** currently active assignments (while processing a basic block)
100 * maps registers to values(their current copies) */
101 static ir_node **assignments;
104 * allocation information: last_uses, register preferences
105 * the information is per firm-node.
107 struct allocation_info_t {
108 unsigned last_uses[2]; /**< bitset indicating last uses (input pos) */
109 ir_node *current_value; /**< copy of the value that should be used */
110 ir_node *original_value; /**< for copies point to original value */
111 float prefs[]; /**< register preferences */
113 typedef struct allocation_info_t allocation_info_t;
115 /** helper datastructure used when sorting register preferences */
120 typedef struct reg_pref_t reg_pref_t;
122 /** per basic-block information */
123 struct block_info_t {
124 bool processed; /**< indicate whether block is processed */
125 ir_node *assignments[]; /**< register assignments at end of block */
127 typedef struct block_info_t block_info_t;
130 * Get the allocation info for a node.
131 * The info is allocated on the first visit of a node.
133 static allocation_info_t *get_allocation_info(ir_node *node)
135 allocation_info_t *info = (allocation_info_t*)get_irn_link(node);
137 info = OALLOCFZ(&obst, allocation_info_t, prefs, n_regs);
138 info->current_value = node;
139 info->original_value = node;
140 set_irn_link(node, info);
146 static allocation_info_t *try_get_allocation_info(const ir_node *node)
148 return (allocation_info_t*) get_irn_link(node);
152 * Get allocation information for a basic block
154 static block_info_t *get_block_info(ir_node *block)
156 block_info_t *info = (block_info_t*)get_irn_link(block);
158 assert(is_Block(block));
160 info = OALLOCFZ(&obst, block_info_t, assignments, n_regs);
161 set_irn_link(block, info);
168 * Link the allocation info of a node to a copy.
169 * Afterwards, both nodes uses the same allocation info.
170 * Copy must not have an allocation info assigned yet.
172 * @param copy the node that gets the allocation info assigned
173 * @param value the original node
175 static void mark_as_copy_of(ir_node *copy, ir_node *value)
177 allocation_info_t *info = get_allocation_info(value);
178 allocation_info_t *copy_info = get_allocation_info(copy);
180 /* find original value */
181 ir_node *original = info->original_value;
182 if (original != value) {
183 info = get_allocation_info(original);
186 assert(info->original_value == original);
187 info->current_value = copy;
189 /* the copy should not be linked to something else yet */
190 assert(copy_info->original_value == copy);
191 copy_info->original_value = original;
193 /* copy over allocation preferences */
194 memcpy(copy_info->prefs, info->prefs, n_regs * sizeof(copy_info->prefs[0]));
198 * Calculate the penalties for every register on a node and its live neighbors.
200 * @param live_nodes the set of live nodes at the current position, may be NULL
201 * @param penalty the penalty to subtract from
202 * @param limited a raw bitset containing the limited set for the node
203 * @param node the node
205 static void give_penalties_for_limits(const ir_nodeset_t *live_nodes,
206 float penalty, const unsigned* limited,
209 allocation_info_t *info = get_allocation_info(node);
211 /* give penalty for all forbidden regs */
212 for (unsigned r = 0; r < n_regs; ++r) {
213 if (rbitset_is_set(limited, r))
216 info->prefs[r] -= penalty;
219 /* all other live values should get a penalty for allowed regs */
220 if (live_nodes == NULL)
223 penalty *= NEIGHBOR_FACTOR;
224 size_t n_allowed = rbitset_popcount(limited, n_regs);
226 /* only create a very weak penalty if multiple regs are allowed */
227 penalty = (penalty * 0.8f) / n_allowed;
229 foreach_ir_nodeset(live_nodes, neighbor, iter) {
230 allocation_info_t *neighbor_info;
232 /* TODO: if op is used on multiple inputs we might not do a
234 if (neighbor == node)
237 neighbor_info = get_allocation_info(neighbor);
238 for (unsigned r = 0; r < n_regs; ++r) {
239 if (!rbitset_is_set(limited, r))
242 neighbor_info->prefs[r] -= penalty;
248 * Calculate the preferences of a definition for the current register class.
249 * If the definition uses a limited set of registers, reduce the preferences
250 * for the limited register on the node and its neighbors.
252 * @param live_nodes the set of live nodes at the current node
253 * @param weight the weight
254 * @param node the current node
256 static void check_defs(ir_nodeset_t const *const live_nodes, float const weight, ir_node *const node, arch_register_req_t const *const req)
258 if (arch_register_req_is(req, limited)) {
259 const unsigned *limited = req->limited;
260 float penalty = weight * DEF_FACTOR;
261 give_penalties_for_limits(live_nodes, penalty, limited, node);
264 if (arch_register_req_is(req, should_be_same)) {
265 ir_node *insn = skip_Proj(node);
266 allocation_info_t *info = get_allocation_info(node);
267 int arity = get_irn_arity(insn);
269 float factor = 1.0f / rbitset_popcount(&req->other_same, arity);
270 for (int i = 0; i < arity; ++i) {
271 if (!rbitset_is_set(&req->other_same, i))
274 ir_node *op = get_irn_n(insn, i);
276 /* if we the value at the should_be_same input doesn't die at the
277 * node, then it is no use to propagate the constraints (since a
278 * copy will emerge anyway) */
279 if (ir_nodeset_contains(live_nodes, op))
282 allocation_info_t *op_info = get_allocation_info(op);
283 for (unsigned r = 0; r < n_regs; ++r) {
284 op_info->prefs[r] += info->prefs[r] * factor;
291 * Walker: Runs an a block calculates the preferences for any
292 * node and every register from the considered register class.
294 static void analyze_block(ir_node *block, void *data)
296 float weight = (float)get_block_execfreq(block);
297 ir_nodeset_t live_nodes;
300 ir_nodeset_init(&live_nodes);
301 be_liveness_end_of_block(lv, cls, block, &live_nodes);
303 sched_foreach_reverse(block, node) {
307 be_foreach_definition(node, cls, value, req,
308 check_defs(&live_nodes, weight, value, req);
312 int arity = get_irn_arity(node);
314 /* the allocation info node currently only uses 1 unsigned value
315 to mark last used inputs. So we will fail for a node with more than
317 allocation_info_t *info = get_allocation_info(node);
318 if (arity >= (int) sizeof(info->last_uses) * 8) {
319 panic("Node with more than %d inputs not supported yet",
320 (int) sizeof(info->last_uses) * 8);
323 for (int i = 0; i < arity; ++i) {
324 ir_node *op = get_irn_n(node, i);
325 const arch_register_req_t *req = arch_get_irn_register_req(op);
329 /* last usage of a value? */
330 if (!ir_nodeset_contains(&live_nodes, op)) {
331 rbitset_set(info->last_uses, i);
335 be_liveness_transfer(cls, node, &live_nodes);
337 /* update weights based on usage constraints */
338 be_foreach_use(node, cls, req, op, op_req,
339 if (!arch_register_req_is(req, limited))
342 give_penalties_for_limits(&live_nodes, weight * USE_FACTOR, req->limited, op);
346 ir_nodeset_destroy(&live_nodes);
349 static void congruence_def(ir_nodeset_t *const live_nodes, ir_node const *const node, arch_register_req_t const *const req)
351 /* should be same constraint? */
352 if (arch_register_req_is(req, should_be_same)) {
353 const ir_node *insn = skip_Proj_const(node);
354 int arity = get_irn_arity(insn);
355 unsigned node_idx = get_irn_idx(node);
356 node_idx = uf_find(congruence_classes, node_idx);
358 for (int i = 0; i < arity; ++i) {
359 if (!rbitset_is_set(&req->other_same, i))
362 ir_node *op = get_irn_n(insn, i);
363 int op_idx = get_irn_idx(op);
364 op_idx = uf_find(congruence_classes, op_idx);
366 /* do we interfere with the value */
367 bool interferes = false;
368 foreach_ir_nodeset(live_nodes, live, iter) {
369 int lv_idx = get_irn_idx(live);
370 lv_idx = uf_find(congruence_classes, lv_idx);
371 if (lv_idx == op_idx) {
376 /* don't put in same affinity class if we interfere */
380 uf_union(congruence_classes, node_idx, op_idx);
381 DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
383 /* one should_be_same is enough... */
389 static void create_congruence_class(ir_node *block, void *data)
391 ir_nodeset_t live_nodes;
394 ir_nodeset_init(&live_nodes);
395 be_liveness_end_of_block(lv, cls, block, &live_nodes);
397 /* check should be same constraints */
398 ir_node *last_phi = NULL;
399 sched_foreach_reverse(block, node) {
405 be_foreach_definition(node, cls, value, req,
406 congruence_def(&live_nodes, value, req);
408 be_liveness_transfer(cls, node, &live_nodes);
411 ir_nodeset_destroy(&live_nodes);
415 /* check phi congruence classes */
416 sched_foreach_reverse_from(last_phi, phi) {
419 if (!arch_irn_consider_in_reg_alloc(cls, phi))
422 int node_idx = get_irn_idx(phi);
423 node_idx = uf_find(congruence_classes, node_idx);
425 int arity = get_irn_arity(phi);
426 for (int i = 0; i < arity; ++i) {
427 ir_node *op = get_Phi_pred(phi, i);
428 int op_idx = get_irn_idx(op);
429 op_idx = uf_find(congruence_classes, op_idx);
431 /* do we interfere with the value */
432 bool interferes = false;
433 foreach_ir_nodeset(&live_nodes, live, iter) {
434 int lv_idx = get_irn_idx(live);
435 lv_idx = uf_find(congruence_classes, lv_idx);
436 if (lv_idx == op_idx) {
441 /* don't put in same affinity class if we interfere */
444 /* any other phi has the same input? */
445 sched_foreach(block, phi) {
450 if (!arch_irn_consider_in_reg_alloc(cls, phi))
452 oop = get_Phi_pred(phi, i);
455 oop_idx = get_irn_idx(oop);
456 oop_idx = uf_find(congruence_classes, oop_idx);
457 if (oop_idx == op_idx) {
465 /* merge the 2 congruence classes and sum up their preferences */
466 int old_node_idx = node_idx;
467 node_idx = uf_union(congruence_classes, node_idx, op_idx);
468 DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
471 old_node_idx = node_idx == old_node_idx ? op_idx : old_node_idx;
472 allocation_info_t *head_info
473 = get_allocation_info(get_idx_irn(irg, node_idx));
474 allocation_info_t *other_info
475 = get_allocation_info(get_idx_irn(irg, old_node_idx));
476 for (unsigned r = 0; r < n_regs; ++r) {
477 head_info->prefs[r] += other_info->prefs[r];
481 ir_nodeset_destroy(&live_nodes);
484 static void set_congruence_prefs(ir_node *node, void *data)
487 unsigned node_idx = get_irn_idx(node);
488 unsigned node_set = uf_find(congruence_classes, node_idx);
490 /* head of congruence class or not in any class */
491 if (node_set == node_idx)
494 if (!arch_irn_consider_in_reg_alloc(cls, node))
497 ir_node *head = get_idx_irn(irg, node_set);
498 allocation_info_t *head_info = get_allocation_info(head);
499 allocation_info_t *info = get_allocation_info(node);
501 memcpy(info->prefs, head_info->prefs, n_regs * sizeof(info->prefs[0]));
504 static void combine_congruence_classes(void)
506 size_t n = get_irg_last_idx(irg);
507 congruence_classes = XMALLOCN(int, n);
508 uf_init(congruence_classes, n);
510 /* create congruence classes */
511 irg_block_walk_graph(irg, create_congruence_class, NULL, NULL);
512 /* merge preferences */
513 irg_walk_graph(irg, set_congruence_prefs, NULL, NULL);
514 free(congruence_classes);
520 * Assign register reg to the given node.
522 * @param node the node
523 * @param reg the register
525 static void use_reg(ir_node *node, const arch_register_t *reg, unsigned width)
527 unsigned r = reg->index;
528 for (unsigned r0 = r; r0 < r + width; ++r0)
529 assignments[r0] = node;
530 arch_set_irn_register(node, reg);
533 static void free_reg_of_value(ir_node *node)
535 if (!arch_irn_consider_in_reg_alloc(cls, node))
538 const arch_register_t *reg = arch_get_irn_register(node);
539 const arch_register_req_t *req = arch_get_irn_register_req(node);
540 unsigned r = reg->index;
541 /* assignment->value may be NULL if a value is used at 2 inputs
542 * so it gets freed twice. */
543 for (unsigned r0 = r; r0 < r + req->width; ++r0) {
544 assert(assignments[r0] == node || assignments[r0] == NULL);
545 assignments[r0] = NULL;
550 * Compare two register preferences in decreasing order.
552 static int compare_reg_pref(const void *e1, const void *e2)
554 const reg_pref_t *rp1 = (const reg_pref_t*) e1;
555 const reg_pref_t *rp2 = (const reg_pref_t*) e2;
556 if (rp1->pref < rp2->pref)
558 if (rp1->pref > rp2->pref)
563 static void fill_sort_candidates(reg_pref_t *regprefs,
564 const allocation_info_t *info)
566 for (unsigned r = 0; r < n_regs; ++r) {
567 float pref = info->prefs[r];
569 regprefs[r].pref = pref;
571 /* TODO: use a stable sort here to avoid unnecessary register jumping */
572 qsort(regprefs, n_regs, sizeof(regprefs[0]), compare_reg_pref);
575 static bool try_optimistic_split(ir_node *to_split, ir_node *before,
576 float pref, float pref_delta,
577 unsigned *forbidden_regs, int recursion)
581 allocation_info_t *info = get_allocation_info(to_split);
584 /* stupid hack: don't optimistically split don't spill nodes...
585 * (so we don't split away the values produced because of
586 * must_be_different constraints) */
587 ir_node *original_insn = skip_Proj(info->original_value);
588 if (arch_get_irn_flags(original_insn) & arch_irn_flags_dont_spill)
591 const arch_register_t *from_reg = arch_get_irn_register(to_split);
592 unsigned from_r = from_reg->index;
593 ir_node *block = get_nodes_block(before);
594 float split_threshold = (float)get_block_execfreq(block) * SPLIT_DELTA;
596 if (pref_delta < split_threshold*0.5)
599 /* find the best free position where we could move to */
600 reg_pref_t *prefs = ALLOCAN(reg_pref_t, n_regs);
601 fill_sort_candidates(prefs, info);
603 for (i = 0; i < n_regs; ++i) {
604 /* we need a normal register which is not an output register
605 an different from the current register of to_split */
607 if (!rbitset_is_set(normal_regs, r))
609 if (rbitset_is_set(forbidden_regs, r))
614 /* is the split worth it? */
615 delta = pref_delta + prefs[i].pref;
616 if (delta < split_threshold) {
617 DB((dbg, LEVEL_3, "Not doing optimistical split of %+F (depth %d), win %f too low\n",
618 to_split, recursion, delta));
622 /* if the register is free then we can do the split */
623 if (assignments[r] == NULL)
626 /* otherwise we might try recursively calling optimistic_split */
627 if (recursion+1 > MAX_OPTIMISTIC_SPLIT_RECURSION)
630 float apref = prefs[i].pref;
631 float apref_delta = i+1 < n_regs ? apref - prefs[i+1].pref : 0;
632 apref_delta += pref_delta - split_threshold;
634 /* our source register isn't a useful destination for recursive
636 bool old_source_state = rbitset_is_set(forbidden_regs, from_r);
637 rbitset_set(forbidden_regs, from_r);
638 /* try recursive split */
639 bool res = try_optimistic_split(assignments[r], before, apref,
640 apref_delta, forbidden_regs, recursion+1);
641 /* restore our destination */
642 if (old_source_state) {
643 rbitset_set(forbidden_regs, from_r);
645 rbitset_clear(forbidden_regs, from_r);
654 const arch_register_t *reg = arch_register_for_index(cls, r);
655 ir_node *copy = be_new_Copy(block, to_split);
657 mark_as_copy_of(copy, to_split);
658 /* hacky, but correct here */
659 if (assignments[from_reg->index] == to_split)
660 free_reg_of_value(to_split);
661 use_reg(copy, reg, width);
662 sched_add_before(before, copy);
665 "Optimistic live-range split %+F move %+F(%s) -> %s before %+F (win %f, depth %d)\n",
666 copy, to_split, from_reg->name, reg->name, before, delta, recursion));
671 * Determine and assign a register for node @p node
673 static void assign_reg(ir_node const *const block, ir_node *const node, arch_register_req_t const *const req, unsigned *const forbidden_regs)
675 assert(!is_Phi(node));
676 /* preassigned register? */
677 arch_register_t const *final_reg = arch_get_irn_register(node);
678 unsigned const width = req->width;
679 if (final_reg != NULL) {
680 DB((dbg, LEVEL_2, "Preassignment %+F -> %s\n", node, final_reg->name));
681 use_reg(node, final_reg, width);
685 /* ignore reqs must be preassigned */
686 assert(!arch_register_req_is(req, ignore));
688 /* give should_be_same boni */
689 allocation_info_t *info = get_allocation_info(node);
690 ir_node *in_node = skip_Proj(node);
691 if (arch_register_req_is(req, should_be_same)) {
692 float weight = (float)get_block_execfreq(block);
693 int arity = get_irn_arity(in_node);
695 assert(arity <= (int) sizeof(req->other_same) * 8);
696 for (int i = 0; i < arity; ++i) {
697 if (!rbitset_is_set(&req->other_same, i))
700 ir_node *in = get_irn_n(in_node, i);
701 const arch_register_t *reg = arch_get_irn_register(in);
702 unsigned reg_index = reg->index;
704 /* if the value didn't die here then we should not propagate the
705 * should_be_same info */
706 if (assignments[reg_index] == in)
709 info->prefs[reg_index] += weight * AFF_SHOULD_BE_SAME;
713 /* create list of register candidates and sort by their preference */
714 DB((dbg, LEVEL_2, "Candidates for %+F:", node));
715 reg_pref_t *reg_prefs = ALLOCAN(reg_pref_t, n_regs);
716 fill_sort_candidates(reg_prefs, info);
717 for (unsigned r = 0; r < n_regs; ++r) {
718 unsigned num = reg_prefs[r].num;
719 if (!rbitset_is_set(normal_regs, num))
721 const arch_register_t *reg = arch_register_for_index(cls, num);
722 DB((dbg, LEVEL_2, " %s(%f)", reg->name, reg_prefs[r].pref));
724 DB((dbg, LEVEL_2, "\n"));
726 const unsigned *allowed_regs = normal_regs;
727 if (arch_register_req_is(req, limited)) {
728 allowed_regs = req->limited;
731 unsigned final_reg_index = 0;
733 for (r = 0; r < n_regs; ++r) {
734 final_reg_index = reg_prefs[r].num;
735 if (!rbitset_is_set(allowed_regs, final_reg_index))
737 /* alignment constraint? */
739 if (arch_register_req_is(req, aligned) && (final_reg_index % width) != 0)
742 for (unsigned r0 = r+1; r0 < r+width; ++r0) {
743 if (assignments[r0] != NULL)
746 /* TODO: attempt optimistic split here */
751 if (assignments[final_reg_index] == NULL)
753 float pref = reg_prefs[r].pref;
754 float delta = r+1 < n_regs ? pref - reg_prefs[r+1].pref : 0;
755 ir_node *before = skip_Proj(node);
757 = try_optimistic_split(assignments[final_reg_index], before, pref,
758 delta, forbidden_regs, 0);
763 /* the common reason to hit this panic is when 1 of your nodes is not
764 * register pressure faithful */
765 panic("No register left for %+F\n", node);
768 final_reg = arch_register_for_index(cls, final_reg_index);
769 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, final_reg->name));
770 use_reg(node, final_reg, width);
774 * Add an permutation in front of a node and change the assignments
775 * due to this permutation.
777 * To understand this imagine a permutation like this:
787 * First we count how many destinations a single value has. At the same time
788 * we can be sure that each destination register has at most 1 source register
789 * (it can have 0 which means we don't care what value is in it).
790 * We ignore all fulfilled permuations (like 7->7)
791 * In a first pass we create as much copy instructions as possible as they
792 * are generally cheaper than exchanges. We do this by counting into how many
793 * destinations a register has to be copied (in the example it's 2 for register
794 * 3, or 1 for the registers 1,2,4 and 7).
795 * We can then create a copy into every destination register when the usecount
796 * of that register is 0 (= noone else needs the value in the register).
798 * After this step we should only have cycles left. We implement a cyclic
799 * permutation of n registers with n-1 transpositions.
801 * @param live_nodes the set of live nodes, updated due to live range split
802 * @param before the node before we add the permutation
803 * @param permutation the permutation array indices are the destination
804 * registers, the values in the array are the source
807 static void permute_values(ir_nodeset_t *live_nodes, ir_node *before,
808 unsigned *permutation)
810 unsigned *n_used = ALLOCANZ(unsigned, n_regs);
812 /* determine how often each source register needs to be read */
813 for (unsigned r = 0; r < n_regs; ++r) {
814 unsigned old_reg = permutation[r];
817 value = assignments[old_reg];
819 /* nothing to do here, reg is not live. Mark it as fixpoint
820 * so we ignore it in the next steps */
828 ir_node *block = get_nodes_block(before);
830 /* step1: create copies where immediately possible */
831 for (unsigned r = 0; r < n_regs; /* empty */) {
832 unsigned old_r = permutation[r];
834 /* - no need to do anything for fixed points.
835 - we can't copy if the value in the dest reg is still needed */
836 if (old_r == r || n_used[r] > 0) {
842 ir_node *src = assignments[old_r];
843 ir_node *copy = be_new_Copy(block, src);
844 sched_add_before(before, copy);
845 const arch_register_t *reg = arch_register_for_index(cls, r);
846 DB((dbg, LEVEL_2, "Copy %+F (from %+F, before %+F) -> %s\n",
847 copy, src, before, reg->name));
848 mark_as_copy_of(copy, src);
849 unsigned width = 1; /* TODO */
850 use_reg(copy, reg, width);
852 if (live_nodes != NULL) {
853 ir_nodeset_insert(live_nodes, copy);
856 /* old register has 1 user less, permutation is resolved */
857 assert(arch_get_irn_register(src)->index == old_r);
860 assert(n_used[old_r] > 0);
862 if (n_used[old_r] == 0) {
863 if (live_nodes != NULL) {
864 ir_nodeset_remove(live_nodes, src);
866 free_reg_of_value(src);
869 /* advance or jump back (if this copy enabled another copy) */
870 if (old_r < r && n_used[old_r] == 0) {
877 /* at this point we only have "cycles" left which we have to resolve with
879 * TODO: if we have free registers left, then we should really use copy
880 * instructions for any cycle longer than 2 registers...
881 * (this is probably architecture dependent, there might be archs where
882 * copies are preferable even for 2-cycles) */
884 /* create perms with the rest */
885 for (unsigned r = 0; r < n_regs; /* empty */) {
886 unsigned old_r = permutation[r];
893 /* we shouldn't have copies from 1 value to multiple destinations left*/
894 assert(n_used[old_r] == 1);
896 /* exchange old_r and r2; after that old_r is a fixed point */
897 unsigned r2 = permutation[old_r];
899 ir_node *in[2] = { assignments[r2], assignments[old_r] };
900 ir_node *perm = be_new_Perm(cls, block, 2, in);
901 sched_add_before(before, perm);
902 DB((dbg, LEVEL_2, "Perm %+F (perm %+F,%+F, before %+F)\n",
903 perm, in[0], in[1], before));
905 unsigned width = 1; /* TODO */
907 ir_node *proj0 = new_r_Proj(perm, get_irn_mode(in[0]), 0);
908 mark_as_copy_of(proj0, in[0]);
909 const arch_register_t *reg0 = arch_register_for_index(cls, old_r);
910 use_reg(proj0, reg0, width);
912 ir_node *proj1 = new_r_Proj(perm, get_irn_mode(in[1]), 1);
913 mark_as_copy_of(proj1, in[1]);
914 const arch_register_t *reg1 = arch_register_for_index(cls, r2);
915 use_reg(proj1, reg1, width);
917 /* 1 value is now in the correct register */
918 permutation[old_r] = old_r;
919 /* the source of r changed to r2 */
922 /* if we have reached a fixpoint update data structures */
923 if (live_nodes != NULL) {
924 ir_nodeset_remove(live_nodes, in[0]);
925 ir_nodeset_remove(live_nodes, in[1]);
926 ir_nodeset_remove(live_nodes, proj0);
927 ir_nodeset_insert(live_nodes, proj1);
932 /* now we should only have fixpoints left */
933 for (unsigned r = 0; r < n_regs; ++r) {
934 assert(permutation[r] == r);
940 * Free regs for values last used.
942 * @param live_nodes set of live nodes, will be updated
943 * @param node the node to consider
945 static void free_last_uses(ir_nodeset_t *live_nodes, ir_node *node)
947 allocation_info_t *info = get_allocation_info(node);
948 const unsigned *last_uses = info->last_uses;
949 int arity = get_irn_arity(node);
951 for (int i = 0; i < arity; ++i) {
952 /* check if one operand is the last use */
953 if (!rbitset_is_set(last_uses, i))
956 ir_node *op = get_irn_n(node, i);
957 free_reg_of_value(op);
958 ir_nodeset_remove(live_nodes, op);
963 * change inputs of a node to the current value (copies/perms)
965 static void rewire_inputs(ir_node *node)
967 int arity = get_irn_arity(node);
968 for (int i = 0; i < arity; ++i) {
969 ir_node *op = get_irn_n(node, i);
970 allocation_info_t *info = try_get_allocation_info(op);
975 info = get_allocation_info(info->original_value);
976 if (info->current_value != op) {
977 set_irn_n(node, i, info->current_value);
983 * Create a bitset of registers occupied with value living through an
986 static void determine_live_through_regs(unsigned *bitset, ir_node *node)
988 const allocation_info_t *info = get_allocation_info(node);
990 /* mark all used registers as potentially live-through */
991 for (unsigned r = 0; r < n_regs; ++r) {
992 if (assignments[r] == NULL)
994 if (!rbitset_is_set(normal_regs, r))
997 rbitset_set(bitset, r);
1000 /* remove registers of value dying at the instruction */
1001 int arity = get_irn_arity(node);
1002 for (int i = 0; i < arity; ++i) {
1003 if (!rbitset_is_set(info->last_uses, i))
1006 ir_node *op = get_irn_n(node, i);
1007 const arch_register_t *reg = arch_get_irn_register(op);
1008 rbitset_clear(bitset, reg->index);
1012 static void solve_lpp(ir_nodeset_t *live_nodes, ir_node *node,
1013 unsigned *forbidden_regs, unsigned *live_through_regs)
1015 unsigned *forbidden_edges = rbitset_malloc(n_regs * n_regs);
1016 int *lpp_vars = XMALLOCNZ(int, n_regs*n_regs);
1018 lpp_t *lpp = lpp_new("prefalloc", lpp_minimize);
1019 //lpp_set_time_limit(lpp, 20);
1020 lpp_set_log(lpp, stdout);
1022 /** mark some edges as forbidden */
1023 be_foreach_use(node, cls, req, op, op_req,
1024 if (!arch_register_req_is(req, limited))
1027 const unsigned *limited = req->limited;
1028 const arch_register_t *reg = arch_get_irn_register(op);
1029 unsigned current_reg = reg->index;
1030 for (unsigned r = 0; r < n_regs; ++r) {
1031 if (rbitset_is_set(limited, r))
1034 rbitset_set(forbidden_edges, current_reg*n_regs + r);
1038 /* add all combinations, except for not allowed ones */
1039 for (unsigned l = 0; l < n_regs; ++l) {
1040 if (!rbitset_is_set(normal_regs, l)) {
1042 snprintf(name, sizeof(name), "%u_to_%u", l, l);
1043 lpp_vars[l*n_regs+l] = lpp_add_var(lpp, name, lpp_binary, 1);
1047 for (unsigned r = 0; r < n_regs; ++r) {
1048 if (!rbitset_is_set(normal_regs, r))
1050 if (rbitset_is_set(forbidden_edges, l*n_regs + r))
1052 /* livethrough values may not use constrained output registers */
1053 if (rbitset_is_set(live_through_regs, l)
1054 && rbitset_is_set(forbidden_regs, r))
1058 snprintf(name, sizeof(name), "%u_to_%u", l, r);
1060 double costs = l==r ? 9 : 8;
1061 lpp_vars[l*n_regs+r]
1062 = lpp_add_var(lpp, name, lpp_binary, costs);
1063 assert(lpp_vars[l*n_regs+r] > 0);
1066 /* add constraints */
1067 for (unsigned l = 0; l < n_regs; ++l) {
1068 /* only 1 destination per register */
1069 int constraint = -1;
1070 for (unsigned r = 0; r < n_regs; ++r) {
1071 int var = lpp_vars[l*n_regs+r];
1074 if (constraint < 0) {
1076 snprintf(name, sizeof(name), "%u_to_dest", l);
1077 constraint = lpp_add_cst(lpp, name, lpp_equal, 1);
1079 lpp_set_factor_fast(lpp, constraint, var, 1);
1081 /* each destination used by at most 1 value */
1083 for (unsigned r = 0; r < n_regs; ++r) {
1084 int var = lpp_vars[r*n_regs+l];
1087 if (constraint < 0) {
1089 snprintf(name, sizeof(name), "one_to_%u", l);
1090 constraint = lpp_add_cst(lpp, name, lpp_less_equal, 1);
1092 lpp_set_factor_fast(lpp, constraint, var, 1);
1096 lpp_dump_plain(lpp, fopen("lppdump.txt", "w"));
1099 lpp_solve(lpp, be_options.ilp_server, be_options.ilp_solver);
1100 if (!lpp_is_sol_valid(lpp))
1101 panic("ilp solution not valid!");
1103 unsigned *assignment = ALLOCAN(unsigned, n_regs);
1104 for (unsigned l = 0; l < n_regs; ++l) {
1105 unsigned dest_reg = (unsigned)-1;
1106 for (unsigned r = 0; r < n_regs; ++r) {
1107 int var = lpp_vars[l*n_regs+r];
1110 double val = lpp_get_var_sol(lpp, var);
1112 assert(dest_reg == (unsigned)-1);
1116 assert(dest_reg != (unsigned)-1);
1117 assignment[dest_reg] = l;
1120 fprintf(stderr, "Assignment: ");
1121 for (unsigned l = 0; l < n_regs; ++l) {
1122 fprintf(stderr, "%u ", assignment[l]);
1124 fprintf(stderr, "\n");
1126 permute_values(live_nodes, node, assignment);
1130 static bool is_aligned(unsigned num, unsigned alignment)
1132 unsigned mask = alignment-1;
1133 assert(is_po2(alignment));
1134 return (num&mask) == 0;
1138 * Enforce constraints at a node by live range splits.
1140 * @param live_nodes the set of live nodes, might be changed
1141 * @param node the current node
1143 static void enforce_constraints(ir_nodeset_t *live_nodes, ir_node *node,
1144 unsigned *forbidden_regs)
1146 /* see if any use constraints are not met and whether double-width
1147 * values are involved */
1148 bool double_width = false;
1150 be_foreach_use(node, cls, req, op, op_req,
1151 /* are there any limitations for the i'th operand? */
1153 double_width = true;
1154 const arch_register_t *reg = arch_get_irn_register(op);
1155 unsigned reg_index = reg->index;
1156 if (arch_register_req_is(req, aligned)) {
1157 if (!is_aligned(reg_index, req->width)) {
1162 if (!arch_register_req_is(req, limited))
1165 const unsigned *limited = req->limited;
1166 if (!rbitset_is_set(limited, reg_index)) {
1167 /* found an assignment outside the limited set */
1173 /* is any of the live-throughs using a constrained output register? */
1174 unsigned *live_through_regs = NULL;
1175 be_foreach_definition(node, cls, value, req,
1178 double_width = true;
1179 if (!arch_register_req_is(req, limited))
1181 if (live_through_regs == NULL) {
1182 live_through_regs = rbitset_alloca(n_regs);
1183 determine_live_through_regs(live_through_regs, node);
1185 rbitset_or(forbidden_regs, req->limited, n_regs);
1186 if (rbitsets_have_common(req->limited, live_through_regs, n_regs))
1193 /* create these arrays if we haven't yet */
1194 if (live_through_regs == NULL) {
1195 live_through_regs = rbitset_alloca(n_regs);
1199 /* only the ILP variant can solve this yet */
1200 solve_lpp(live_nodes, node, forbidden_regs, live_through_regs);
1204 /* at this point we have to construct a bipartite matching problem to see
1205 * which values should go to which registers
1206 * Note: We're building the matrix in "reverse" - source registers are
1207 * right, destinations left because this will produce the solution
1208 * in the format required for permute_values.
1210 hungarian_problem_t *bp
1211 = hungarian_new(n_regs, n_regs, HUNGARIAN_MATCH_PERFECT);
1213 /* add all combinations, then remove not allowed ones */
1214 for (unsigned l = 0; l < n_regs; ++l) {
1215 if (!rbitset_is_set(normal_regs, l)) {
1216 hungarian_add(bp, l, l, 1);
1220 for (unsigned r = 0; r < n_regs; ++r) {
1221 if (!rbitset_is_set(normal_regs, r))
1223 /* livethrough values may not use constrainted output registers */
1224 if (rbitset_is_set(live_through_regs, l)
1225 && rbitset_is_set(forbidden_regs, r))
1228 hungarian_add(bp, r, l, l == r ? 9 : 8);
1232 be_foreach_use(node, cls, req, op, op_req,
1233 if (!arch_register_req_is(req, limited))
1236 const unsigned *limited = req->limited;
1237 const arch_register_t *reg = arch_get_irn_register(op);
1238 unsigned current_reg = reg->index;
1239 for (unsigned r = 0; r < n_regs; ++r) {
1240 if (rbitset_is_set(limited, r))
1242 hungarian_remove(bp, r, current_reg);
1246 //hungarian_print_cost_matrix(bp, 1);
1247 hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1249 unsigned *assignment = ALLOCAN(unsigned, n_regs);
1250 int res = hungarian_solve(bp, assignment, NULL, 0);
1254 fprintf(stderr, "Swap result:");
1255 for (i = 0; i < (int) n_regs; ++i) {
1256 fprintf(stderr, " %d", assignment[i]);
1258 fprintf(stderr, "\n");
1263 permute_values(live_nodes, node, assignment);
1266 /** test whether a node @p n is a copy of the value of node @p of */
1267 static bool is_copy_of(ir_node *value, ir_node *test_value)
1269 if (value == test_value)
1272 allocation_info_t *info = get_allocation_info(value);
1273 allocation_info_t *test_info = get_allocation_info(test_value);
1274 return test_info->original_value == info->original_value;
1278 * find a value in the end-assignment of a basic block
1279 * @returns the index into the assignment array if found
1282 static int find_value_in_block_info(block_info_t *info, ir_node *value)
1284 ir_node **end_assignments = info->assignments;
1285 for (unsigned r = 0; r < n_regs; ++r) {
1286 ir_node *a_value = end_assignments[r];
1288 if (a_value == NULL)
1290 if (is_copy_of(a_value, value))
1298 * Create the necessary permutations at the end of a basic block to fullfill
1299 * the register assignment for phi-nodes in the next block
1301 static void add_phi_permutations(ir_node *block, int p)
1303 ir_node *pred = get_Block_cfgpred_block(block, p);
1304 block_info_t *pred_info = get_block_info(pred);
1306 /* predecessor not processed yet? nothing to do */
1307 if (!pred_info->processed)
1310 unsigned *permutation = ALLOCAN(unsigned, n_regs);
1311 for (unsigned r = 0; r < n_regs; ++r) {
1315 /* check phi nodes */
1316 bool need_permutation = false;
1317 ir_node *phi = sched_first(block);
1318 for ( ; is_Phi(phi); phi = sched_next(phi)) {
1319 if (!arch_irn_consider_in_reg_alloc(cls, phi))
1322 ir_node *phi_pred = get_Phi_pred(phi, p);
1323 int a = find_value_in_block_info(pred_info, phi_pred);
1326 const arch_register_t *reg = arch_get_irn_register(phi);
1327 int regn = reg->index;
1328 /* same register? nothing to do */
1332 ir_node *op = pred_info->assignments[a];
1333 const arch_register_t *op_reg = arch_get_irn_register(op);
1334 /* Virtual registers are ok, too. */
1335 if (op_reg->type & arch_register_type_virtual)
1338 permutation[regn] = a;
1339 need_permutation = true;
1342 if (need_permutation) {
1343 /* permute values at end of predecessor */
1344 ir_node **old_assignments = assignments;
1345 assignments = pred_info->assignments;
1346 permute_values(NULL, be_get_end_of_block_insertion_point(pred),
1348 assignments = old_assignments;
1351 /* change phi nodes to use the copied values */
1352 phi = sched_first(block);
1353 for ( ; is_Phi(phi); phi = sched_next(phi)) {
1354 if (!arch_irn_consider_in_reg_alloc(cls, phi))
1357 /* we have permuted all values into the correct registers so we can
1358 simply query which value occupies the phis register in the
1360 int a = arch_get_irn_register(phi)->index;
1361 ir_node *op = pred_info->assignments[a];
1362 set_Phi_pred(phi, p, op);
1367 * Set preferences for a phis register based on the registers used on the
1370 static void adapt_phi_prefs(ir_node *phi)
1372 ir_node *block = get_nodes_block(phi);
1373 allocation_info_t *info = get_allocation_info(phi);
1375 int arity = get_irn_arity(phi);
1376 for (int i = 0; i < arity; ++i) {
1377 ir_node *op = get_irn_n(phi, i);
1378 const arch_register_t *reg = arch_get_irn_register(op);
1382 /* we only give the bonus if the predecessor already has registers
1383 * assigned, otherwise we only see a dummy value
1384 * and any conclusions about its register are useless */
1385 ir_node *pred_block = get_Block_cfgpred_block(block, i);
1386 block_info_t *pred_block_info = get_block_info(pred_block);
1387 if (!pred_block_info->processed)
1390 /* give bonus for already assigned register */
1391 float weight = (float)get_block_execfreq(pred_block);
1392 info->prefs[reg->index] += weight * AFF_PHI;
1397 * After a phi has been assigned a register propagate preference inputs
1398 * to the phi inputs.
1400 static void propagate_phi_register(ir_node *phi, unsigned assigned_r)
1402 ir_node *block = get_nodes_block(phi);
1404 int arity = get_irn_arity(phi);
1405 for (int i = 0; i < arity; ++i) {
1406 ir_node *op = get_Phi_pred(phi, i);
1407 allocation_info_t *info = get_allocation_info(op);
1408 ir_node *pred_block = get_Block_cfgpred_block(block, i);
1410 = (float)get_block_execfreq(pred_block) * AFF_PHI;
1412 if (info->prefs[assigned_r] >= weight)
1415 /* promote the prefered register */
1416 for (unsigned r = 0; r < n_regs; ++r) {
1417 if (info->prefs[r] > -weight) {
1418 info->prefs[r] = -weight;
1421 info->prefs[assigned_r] = weight;
1424 propagate_phi_register(op, assigned_r);
1428 static void assign_phi_registers(ir_node *block)
1430 /* count phi nodes */
1432 sched_foreach(block, node) {
1435 if (!arch_irn_consider_in_reg_alloc(cls, node))
1443 /* build a bipartite matching problem for all phi nodes */
1444 hungarian_problem_t *bp
1445 = hungarian_new(n_phis, n_regs, HUNGARIAN_MATCH_PERFECT);
1447 sched_foreach(block, node) {
1450 if (!arch_irn_consider_in_reg_alloc(cls, node))
1453 /* give boni for predecessor colorings */
1454 adapt_phi_prefs(node);
1455 /* add stuff to bipartite problem */
1456 allocation_info_t *info = get_allocation_info(node);
1457 DB((dbg, LEVEL_3, "Prefs for %+F: ", node));
1458 for (unsigned r = 0; r < n_regs; ++r) {
1459 if (!rbitset_is_set(normal_regs, r))
1462 float costs = info->prefs[r];
1463 costs = costs < 0 ? -logf(-costs+1) : logf(costs+1);
1466 hungarian_add(bp, n, r, (int)costs);
1467 DB((dbg, LEVEL_3, " %s(%f)", arch_register_for_index(cls, r)->name,
1470 DB((dbg, LEVEL_3, "\n"));
1474 //hungarian_print_cost_matrix(bp, 7);
1475 hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1477 unsigned *assignment = ALLOCAN(unsigned, n_regs);
1478 int res = hungarian_solve(bp, assignment, NULL, 0);
1483 sched_foreach(block, node) {
1486 if (!arch_irn_consider_in_reg_alloc(cls, node))
1488 const arch_register_req_t *req
1489 = arch_get_irn_register_req(node);
1491 unsigned r = assignment[n++];
1492 assert(rbitset_is_set(normal_regs, r));
1493 const arch_register_t *reg = arch_register_for_index(cls, r);
1494 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
1495 use_reg(node, reg, req->width);
1497 /* adapt preferences for phi inputs */
1498 propagate_phi_register(node, r);
1502 static arch_register_req_t *allocate_reg_req(ir_graph *irg)
1504 struct obstack *obst = be_get_be_obst(irg);
1505 arch_register_req_t *req = OALLOCZ(obst, arch_register_req_t);
1510 * Walker: assign registers to all nodes of a block that
1511 * need registers from the currently considered register class.
1513 static void allocate_coalesce_block(ir_node *block, void *data)
1516 DB((dbg, LEVEL_2, "* Block %+F\n", block));
1518 /* clear assignments */
1519 block_info_t *block_info = get_block_info(block);
1520 assignments = block_info->assignments;
1522 ir_nodeset_t live_nodes;
1523 ir_nodeset_init(&live_nodes);
1525 /* gather regalloc infos of predecessor blocks */
1526 int n_preds = get_Block_n_cfgpreds(block);
1527 block_info_t **pred_block_infos = ALLOCAN(block_info_t*, n_preds);
1528 for (int i = 0; i < n_preds; ++i) {
1529 ir_node *pred = get_Block_cfgpred_block(block, i);
1530 block_info_t *pred_info = get_block_info(pred);
1531 pred_block_infos[i] = pred_info;
1534 ir_node **phi_ins = ALLOCAN(ir_node*, n_preds);
1536 /* collect live-in nodes and preassigned values */
1537 be_lv_foreach(lv, block, be_lv_state_in, node) {
1538 const arch_register_req_t *req = arch_get_irn_register_req(node);
1539 if (req->cls != cls)
1542 if (arch_register_req_is(req, ignore)) {
1543 allocation_info_t *info = get_allocation_info(node);
1544 info->current_value = node;
1546 const arch_register_t *reg = arch_get_irn_register(node);
1547 assert(reg != NULL); /* ignore values must be preassigned */
1548 use_reg(node, reg, req->width);
1552 /* check all predecessors for this value, if it is not everywhere the
1553 same or unknown then we have to construct a phi
1554 (we collect the potential phi inputs here) */
1555 bool need_phi = false;
1556 for (int p = 0; p < n_preds; ++p) {
1557 block_info_t *pred_info = pred_block_infos[p];
1559 if (!pred_info->processed) {
1560 /* use node for now, it will get fixed later */
1564 int a = find_value_in_block_info(pred_info, node);
1566 /* must live out of predecessor */
1568 phi_ins[p] = pred_info->assignments[a];
1569 /* different value from last time? then we need a phi */
1570 if (p > 0 && phi_ins[p-1] != phi_ins[p]) {
1577 ir_mode *mode = get_irn_mode(node);
1578 const arch_register_req_t *phi_req = cls->class_req;
1579 if (req->width > 1) {
1580 arch_register_req_t *new_req = allocate_reg_req(irg);
1582 new_req->type = req->type & arch_register_req_type_aligned;
1583 new_req->width = req->width;
1586 ir_node *phi = be_new_Phi(block, n_preds, phi_ins, mode,
1589 DB((dbg, LEVEL_3, "Create Phi %+F (for %+F) -", phi, node));
1590 #ifdef DEBUG_libfirm
1591 for (int pi = 0; pi < n_preds; ++pi) {
1592 DB((dbg, LEVEL_3, " %+F", phi_ins[pi]));
1594 DB((dbg, LEVEL_3, "\n"));
1596 mark_as_copy_of(phi, node);
1597 sched_add_after(block, phi);
1601 allocation_info_t *info = get_allocation_info(node);
1602 info->current_value = phi_ins[0];
1604 /* Grab 1 of the inputs we constructed (might not be the same as
1605 * "node" as we could see the same copy of the value in all
1610 /* if the node already has a register assigned use it */
1611 const arch_register_t *reg = arch_get_irn_register(node);
1613 use_reg(node, reg, req->width);
1616 /* remember that this node is live at the beginning of the block */
1617 ir_nodeset_insert(&live_nodes, node);
1620 /** Collects registers which must not be used for optimistic splits. */
1621 unsigned *const forbidden_regs = rbitset_alloca(n_regs);
1623 /* handle phis... */
1624 assign_phi_registers(block);
1626 /* all live-ins must have a register */
1628 foreach_ir_nodeset(&live_nodes, node, iter) {
1629 const arch_register_t *reg = arch_get_irn_register(node);
1630 assert(reg != NULL);
1634 /* assign instructions in the block */
1635 sched_foreach(block, node) {
1636 /* phis are already assigned */
1640 rewire_inputs(node);
1642 /* enforce use constraints */
1643 rbitset_clear_all(forbidden_regs, n_regs);
1644 enforce_constraints(&live_nodes, node, forbidden_regs);
1646 rewire_inputs(node);
1648 /* we may not use registers used for inputs for optimistic splits */
1649 be_foreach_use(node, cls, in_req, op, op_req,
1650 const arch_register_t *reg = arch_get_irn_register(op);
1651 rbitset_set(forbidden_regs, reg->index);
1654 /* free registers of values last used at this instruction */
1655 free_last_uses(&live_nodes, node);
1657 /* assign output registers */
1658 be_foreach_definition_(node, cls, value, req,
1659 assign_reg(block, value, req, forbidden_regs);
1663 ir_nodeset_destroy(&live_nodes);
1666 block_info->processed = true;
1668 /* permute values at end of predecessor blocks in case of phi-nodes */
1670 for (int p = 0; p < n_preds; ++p) {
1671 add_phi_permutations(block, p);
1675 /* if we have exactly 1 successor then we might be able to produce phi
1677 if (get_irn_n_edges_kind(block, EDGE_KIND_BLOCK) == 1) {
1678 const ir_edge_t *edge
1679 = get_irn_out_edge_first_kind(block, EDGE_KIND_BLOCK);
1680 ir_node *succ = get_edge_src_irn(edge);
1681 int p = get_edge_src_pos(edge);
1682 block_info_t *succ_info = get_block_info(succ);
1684 if (succ_info->processed) {
1685 add_phi_permutations(succ, p);
1690 typedef struct block_costs_t block_costs_t;
1691 struct block_costs_t {
1692 float costs; /**< costs of the block */
1693 int dfs_num; /**< depth first search number (to detect backedges) */
1696 static int cmp_block_costs(const void *d1, const void *d2)
1698 const ir_node * const *block1 = (const ir_node**)d1;
1699 const ir_node * const *block2 = (const ir_node**)d2;
1700 const block_costs_t *info1 = (const block_costs_t*)get_irn_link(*block1);
1701 const block_costs_t *info2 = (const block_costs_t*)get_irn_link(*block2);
1702 return QSORT_CMP(info2->costs, info1->costs);
1705 static void determine_block_order(void)
1707 ir_node **blocklist = be_get_cfgpostorder(irg);
1708 size_t n_blocks = ARR_LEN(blocklist);
1710 pdeq *worklist = new_pdeq();
1711 ir_node **order = XMALLOCN(ir_node*, n_blocks);
1714 /* clear block links... */
1715 for (size_t p = 0; p < n_blocks; ++p) {
1716 ir_node *block = blocklist[p];
1717 set_irn_link(block, NULL);
1720 /* walk blocks in reverse postorder, the costs for each block are the
1721 * sum of the costs of its predecessors (excluding the costs on backedges
1722 * which we can't determine) */
1723 for (size_t p = n_blocks; p > 0;) {
1724 block_costs_t *cost_info;
1725 ir_node *block = blocklist[--p];
1727 float execfreq = (float)get_block_execfreq(block);
1728 float costs = execfreq;
1729 int n_cfgpreds = get_Block_n_cfgpreds(block);
1730 for (int p2 = 0; p2 < n_cfgpreds; ++p2) {
1731 ir_node *pred_block = get_Block_cfgpred_block(block, p2);
1732 block_costs_t *pred_costs = (block_costs_t*)get_irn_link(pred_block);
1733 /* we don't have any info for backedges */
1734 if (pred_costs == NULL)
1736 costs += pred_costs->costs;
1739 cost_info = OALLOCZ(&obst, block_costs_t);
1740 cost_info->costs = costs;
1741 cost_info->dfs_num = dfs_num++;
1742 set_irn_link(block, cost_info);
1745 /* sort array by block costs */
1746 qsort(blocklist, n_blocks, sizeof(blocklist[0]), cmp_block_costs);
1748 ir_reserve_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1749 inc_irg_block_visited(irg);
1751 for (size_t p = 0; p < n_blocks; ++p) {
1752 ir_node *block = blocklist[p];
1753 if (Block_block_visited(block))
1756 /* continually add predecessors with highest costs to worklist
1757 * (without using backedges) */
1759 block_costs_t *info = (block_costs_t*)get_irn_link(block);
1760 ir_node *best_pred = NULL;
1761 float best_costs = -1;
1762 int n_cfgpred = get_Block_n_cfgpreds(block);
1764 pdeq_putr(worklist, block);
1765 mark_Block_block_visited(block);
1766 for (int i = 0; i < n_cfgpred; ++i) {
1767 ir_node *pred_block = get_Block_cfgpred_block(block, i);
1768 block_costs_t *pred_info = (block_costs_t*)get_irn_link(pred_block);
1770 /* ignore backedges */
1771 if (pred_info->dfs_num > info->dfs_num)
1774 if (info->costs > best_costs) {
1775 best_costs = info->costs;
1776 best_pred = pred_block;
1780 } while (block != NULL && !Block_block_visited(block));
1782 /* now put all nodes in the worklist in our final order */
1783 while (!pdeq_empty(worklist)) {
1784 ir_node *pblock = (ir_node*)pdeq_getr(worklist);
1785 assert(order_p < n_blocks);
1786 order[order_p++] = pblock;
1789 assert(order_p == n_blocks);
1792 ir_free_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1794 DEL_ARR_F(blocklist);
1796 obstack_free(&obst, NULL);
1797 obstack_init(&obst);
1799 block_order = order;
1800 n_block_order = n_blocks;
1803 static void free_block_order(void)
1809 * Run the register allocator for the current register class.
1811 static void be_pref_alloc_cls(void)
1813 be_assure_live_sets(irg);
1814 lv = be_get_irg_liveness(irg);
1816 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
1818 DB((dbg, LEVEL_2, "=== Allocating registers of %s ===\n", cls->name));
1820 be_clear_links(irg);
1822 irg_block_walk_graph(irg, NULL, analyze_block, NULL);
1823 combine_congruence_classes();
1825 for (size_t i = 0; i < n_block_order; ++i) {
1826 ir_node *block = block_order[i];
1827 allocate_coalesce_block(block, NULL);
1830 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
1833 static void dump(int mask, ir_graph *irg, const char *suffix)
1835 if (be_options.dump_flags & mask)
1836 dump_ir_graph(irg, suffix);
1840 * Run the spiller on the current graph.
1842 static void spill(void)
1844 /* make sure all nodes show their real register pressure */
1845 be_timer_push(T_RA_CONSTR);
1846 be_pre_spill_prepare_constr(irg, cls);
1847 be_timer_pop(T_RA_CONSTR);
1849 dump(DUMP_RA, irg, "spillprepare");
1852 be_timer_push(T_RA_SPILL);
1853 be_do_spill(irg, cls);
1854 be_timer_pop(T_RA_SPILL);
1856 be_timer_push(T_RA_SPILL_APPLY);
1857 check_for_memory_operands(irg);
1858 be_timer_pop(T_RA_SPILL_APPLY);
1860 dump(DUMP_RA, irg, "spill");
1864 * The pref register allocator for a whole procedure.
1866 static void be_pref_alloc(ir_graph *new_irg)
1868 obstack_init(&obst);
1872 /* determine a good coloring order */
1873 determine_block_order();
1875 const arch_env_t *arch_env = be_get_irg_arch_env(new_irg);
1876 int n_cls = arch_env->n_register_classes;
1877 for (int c = 0; c < n_cls; ++c) {
1878 cls = &arch_env->register_classes[c];
1879 if (arch_register_class_flags(cls) & arch_register_class_flag_manual_ra)
1882 stat_ev_ctx_push_str("regcls", cls->name);
1884 n_regs = arch_register_class_n_regs(cls);
1885 normal_regs = rbitset_malloc(n_regs);
1886 be_set_allocatable_regs(irg, cls, normal_regs);
1890 /* verify schedule and register pressure */
1891 be_timer_push(T_VERIFY);
1892 if (be_options.verify_option == BE_VERIFY_WARN) {
1893 be_verify_schedule(irg);
1894 be_verify_register_pressure(irg, cls);
1895 } else if (be_options.verify_option == BE_VERIFY_ASSERT) {
1896 assert(be_verify_schedule(irg) && "Schedule verification failed");
1897 assert(be_verify_register_pressure(irg, cls)
1898 && "Register pressure verification failed");
1900 be_timer_pop(T_VERIFY);
1902 be_timer_push(T_RA_COLOR);
1903 be_pref_alloc_cls();
1904 be_timer_pop(T_RA_COLOR);
1906 /* we most probably constructed new Phis so liveness info is invalid
1908 be_invalidate_live_sets(irg);
1911 stat_ev_ctx_pop("regcls");
1916 be_timer_push(T_RA_SPILL_APPLY);
1917 be_abi_fix_stack_nodes(irg);
1918 be_timer_pop(T_RA_SPILL_APPLY);
1920 be_timer_push(T_VERIFY);
1921 if (be_options.verify_option == BE_VERIFY_WARN) {
1922 be_verify_register_allocation(irg);
1923 } else if (be_options.verify_option == BE_VERIFY_ASSERT) {
1924 assert(be_verify_register_allocation(irg)
1925 && "Register allocation invalid");
1927 be_timer_pop(T_VERIFY);
1929 obstack_free(&obst, NULL);
1932 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_pref_alloc)
1933 void be_init_pref_alloc(void)
1935 static be_ra_t be_ra_pref = { be_pref_alloc };
1936 be_register_allocator("pref", &be_ra_pref);
1937 FIRM_DBG_REGISTER(dbg, "firm.be.prefalloc");