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 permutate_values code
43 * - We have to pessimistically construct Phi_0s when not all predecessors
44 * of a block are known.
45 * - Phi color assignment should give bonus points towards registers already
46 * assigned at predecessors.
47 * - think about a smarter sequence of visiting the blocks. Sorted by
48 * execfreq might be good, or looptree from inner to outermost loops going
49 * over blocks in a reverse postorder
50 * - propagate preferences through Phis
61 #include "iredges_t.h"
62 #include "irgraph_t.h"
68 #include "bechordal_t.h"
77 #include "bespillutil.h"
80 #include "bipartite.h"
81 #include "hungarian.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 1.0f
89 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
91 static struct obstack obst;
92 static be_irg_t *birg;
94 static const arch_register_class_t *cls;
95 static const arch_register_req_t *default_cls_req;
97 static const ir_exec_freq *execfreqs;
98 static unsigned n_regs;
99 static bitset_t *ignore_regs;
101 /** info about the current assignment for a register */
102 struct assignment_t {
103 ir_node *value; /**< currently assigned value */
105 typedef struct assignment_t assignment_t;
107 /** currently active assignments (while processing a basic block) */
108 static assignment_t *assignments;
111 * allocation information: last_uses, register preferences
112 * the information is per firm-node.
114 struct allocation_info_t {
115 unsigned last_uses; /**< bitset indicating last uses (input pos) */
116 ir_node *current_value; /**< copy of the value that should be used */
117 ir_node *original_value; /**< for copies point to original value */
118 float prefs[0]; /**< register preferences */
120 typedef struct allocation_info_t allocation_info_t;
122 /** helper datastructure used when sorting register preferences */
127 typedef struct reg_pref_t reg_pref_t;
129 /** per basic-block information */
130 struct block_info_t {
131 bool processed; /**< indicate wether block is processed */
132 assignment_t assignments[0]; /**< register assignments at end of block */
134 typedef struct block_info_t block_info_t;
137 * Get the allocation info for a node.
138 * The info is allocated on the first visit of a node.
140 static allocation_info_t *get_allocation_info(ir_node *node)
142 allocation_info_t *info;
143 if (!irn_visited_else_mark(node)) {
144 size_t size = sizeof(info[0]) + n_regs * sizeof(info->prefs[0]);
145 info = obstack_alloc(&obst, size);
146 memset(info, 0, size);
147 info->current_value = node;
148 info->original_value = node;
149 set_irn_link(node, info);
151 info = get_irn_link(node);
158 * Get allocation information for a basic block
160 static block_info_t *get_block_info(ir_node *block)
164 assert(is_Block(block));
165 if (!irn_visited_else_mark(block)) {
166 size_t size = sizeof(info[0]) + n_regs * sizeof(info->assignments[0]);
167 info = obstack_alloc(&obst, size);
168 memset(info, 0, size);
169 set_irn_link(block, info);
171 info = get_irn_link(block);
178 * Get default register requirement for the current register class
180 static const arch_register_req_t *get_default_req_current_cls(void)
182 if (default_cls_req == NULL) {
183 struct obstack *obst = get_irg_obstack(irg);
184 arch_register_req_t *req = obstack_alloc(obst, sizeof(*req));
185 memset(req, 0, sizeof(*req));
187 req->type = arch_register_req_type_normal;
190 default_cls_req = req;
192 return default_cls_req;
196 * Link the allocation info of a node to a copy.
197 * Afterwards, both nodes uses the same allocation info.
198 * Copy must not have an allocation info assigned yet.
200 * @param copy the node that gets the allocation info assigned
201 * @param value the original node
203 static void mark_as_copy_of(ir_node *copy, ir_node *value)
206 allocation_info_t *info = get_allocation_info(value);
207 allocation_info_t *copy_info = get_allocation_info(copy);
209 /* find original value */
210 original = info->original_value;
211 if (original != value) {
212 info = get_allocation_info(original);
215 assert(info->original_value == original);
216 info->current_value = copy;
218 /* the copy should not be linked to something else yet */
219 assert(copy_info->original_value == copy);
220 /* copy over allocation preferences */
221 memcpy(copy_info->prefs, info->prefs, n_regs * sizeof(copy_info->prefs[0]));
222 copy_info->original_value = original;
226 * Calculate the penalties for every register on a node and its live neighbors.
228 * @param live_nodes the set of live nodes at the current position, may be NULL
229 * @param penalty the penalty to subtract from
230 * @param limited a raw bitset containing the limited set for the node
231 * @param node the node
233 static void give_penalties_for_limits(const ir_nodeset_t *live_nodes,
234 float penalty, const unsigned* limited,
237 ir_nodeset_iterator_t iter;
239 allocation_info_t *info = get_allocation_info(node);
242 /* give penalty for all forbidden regs */
243 for (r = 0; r < n_regs; ++r) {
244 if (rbitset_is_set(limited, r))
247 info->prefs[r] -= penalty;
250 /* all other live values should get a penalty for allowed regs */
251 if (live_nodes == NULL)
254 /* TODO: reduce penalty if there are multiple allowed registers... */
255 penalty *= NEIGHBOR_FACTOR;
256 foreach_ir_nodeset(live_nodes, neighbor, iter) {
257 allocation_info_t *neighbor_info;
259 /* TODO: if op is used on multiple inputs we might not do a
261 if (neighbor == node)
264 neighbor_info = get_allocation_info(neighbor);
265 for (r = 0; r < n_regs; ++r) {
266 if (!rbitset_is_set(limited, r))
269 neighbor_info->prefs[r] -= penalty;
275 * Calculate the preferences of a definition for the current register class.
276 * If the definition uses a limited set of registers, reduce the preferences
277 * for the limited register on the node and its neighbors.
279 * @param live_nodes the set of live nodes at the current node
280 * @param weight the weight
281 * @param node the current node
283 static void check_defs(const ir_nodeset_t *live_nodes, float weight,
286 const arch_register_req_t *req;
288 if (get_irn_mode(node) == mode_T) {
289 const ir_edge_t *edge;
290 foreach_out_edge(node, edge) {
291 ir_node *proj = get_edge_src_irn(edge);
292 check_defs(live_nodes, weight, proj);
297 if (!arch_irn_consider_in_reg_alloc(cls, node))
300 req = arch_get_register_req_out(node);
301 if (req->type & arch_register_req_type_limited) {
302 const unsigned *limited = req->limited;
303 float penalty = weight * DEF_FACTOR;
304 give_penalties_for_limits(live_nodes, penalty, limited, node);
307 if (req->type & arch_register_req_type_should_be_same) {
308 ir_node *insn = skip_Proj(node);
309 allocation_info_t *info = get_allocation_info(node);
310 int arity = get_irn_arity(insn);
313 float factor = 1.0f / rbitset_popcnt(&req->other_same, arity);
314 for (i = 0; i < arity; ++i) {
317 allocation_info_t *op_info;
319 if (!rbitset_is_set(&req->other_same, i))
322 op = get_irn_n(insn, i);
323 op_info = get_allocation_info(op);
324 for (r = 0; r < n_regs; ++r) {
325 if (bitset_is_set(ignore_regs, r))
327 op_info->prefs[r] += info->prefs[r] * factor;
334 * Walker: Runs an a block calculates the preferences for any
335 * node and every register from the considered register class.
337 static void analyze_block(ir_node *block, void *data)
339 float weight = get_block_execfreq(execfreqs, block);
340 ir_nodeset_t live_nodes;
344 ir_nodeset_init(&live_nodes);
345 be_liveness_end_of_block(lv, cls, block, &live_nodes);
347 sched_foreach_reverse(block, node) {
348 allocation_info_t *info;
355 /* TODO give/take penalties for should_be_same/different) */
356 check_defs(&live_nodes, weight, node);
359 arity = get_irn_arity(node);
361 /* the allocation info node currently only uses 1 unsigned value
362 to mark last used inputs. So we will fail for a node with more than
364 if (arity >= (int) sizeof(unsigned) * 8) {
365 panic("Node with more than %d inputs not supported yet",
366 (int) sizeof(unsigned) * 8);
369 info = get_allocation_info(node);
370 for (i = 0; i < arity; ++i) {
371 ir_node *op = get_irn_n(node, i);
372 if (!arch_irn_consider_in_reg_alloc(cls, op))
375 /* last usage of a value? */
376 if (!ir_nodeset_contains(&live_nodes, op)) {
377 rbitset_set(&info->last_uses, i);
381 be_liveness_transfer(cls, node, &live_nodes);
383 /* update weights based on usage constraints */
384 for (i = 0; i < arity; ++i) {
385 const arch_register_req_t *req;
386 const unsigned *limited;
387 ir_node *op = get_irn_n(node, i);
389 if (!arch_irn_consider_in_reg_alloc(cls, op))
392 req = arch_get_register_req(node, i);
393 if (!(req->type & arch_register_req_type_limited))
396 /* TODO: give penalties to neighbors for precolored nodes! */
398 limited = req->limited;
399 give_penalties_for_limits(&live_nodes, weight * USE_FACTOR, limited,
404 ir_nodeset_destroy(&live_nodes);
408 * Assign register reg to the given node.
410 * @param node the node
411 * @param reg the register
413 static void use_reg(ir_node *node, const arch_register_t *reg)
415 unsigned r = arch_register_get_index(reg);
416 assignment_t *assignment = &assignments[r];
418 //assert(assignment->value == NULL);
419 assignment->value = node;
421 arch_set_irn_register(node, reg);
425 * Compare two register preferences in decreasing order.
427 static int compare_reg_pref(const void *e1, const void *e2)
429 const reg_pref_t *rp1 = (const reg_pref_t*) e1;
430 const reg_pref_t *rp2 = (const reg_pref_t*) e2;
431 if (rp1->pref < rp2->pref)
433 if (rp1->pref > rp2->pref)
438 static void fill_sort_candidates(reg_pref_t *regprefs,
439 const allocation_info_t *info)
443 for (r = 0; r < n_regs; ++r) {
444 float pref = info->prefs[r];
445 if (bitset_is_set(ignore_regs, r)) {
449 regprefs[r].pref = pref;
451 /* TODO: use a stable sort here to avoid unnecessary register jumping */
452 qsort(regprefs, n_regs, sizeof(regprefs[0]), compare_reg_pref);
456 * Determine and assign a register for node @p node
458 static void assign_reg(const ir_node *block, ir_node *node)
460 const arch_register_t *reg;
461 allocation_info_t *info;
462 const arch_register_req_t *req;
463 reg_pref_t *reg_prefs;
467 assert(arch_irn_consider_in_reg_alloc(cls, node));
469 /* preassigned register? */
470 reg = arch_get_irn_register(node);
472 DB((dbg, LEVEL_2, "Preassignment %+F -> %s\n", node, reg->name));
477 /* give should_be_same boni */
478 info = get_allocation_info(node);
479 req = arch_get_register_req_out(node);
481 in_node = skip_Proj(node);
482 if (req->type & arch_register_req_type_should_be_same) {
483 float weight = get_block_execfreq(execfreqs, block);
484 int arity = get_irn_arity(in_node);
487 assert(arity <= (int) sizeof(req->other_same) * 8);
488 for (i = 0; i < arity; ++i) {
490 const arch_register_t *reg;
492 if (!rbitset_is_set(&req->other_same, i))
495 in = get_irn_n(in_node, i);
496 reg = arch_get_irn_register(in);
498 r = arch_register_get_index(reg);
499 if (bitset_is_set(ignore_regs, r))
501 info->prefs[r] += weight * AFF_SHOULD_BE_SAME;
505 DB((dbg, LEVEL_2, "Candidates for %+F:", node));
506 reg_prefs = alloca(n_regs * sizeof(reg_prefs[0]));
507 fill_sort_candidates(reg_prefs, info);
508 for (i = 0; i < n_regs; ++i) {
509 unsigned num = reg_prefs[i].num;
510 const arch_register_t *reg = arch_register_for_index(cls, num);
511 DB((dbg, LEVEL_2, " %s(%f)", reg->name, reg_prefs[i].pref));
513 DB((dbg, LEVEL_2, "\n"));
515 for (i = 0; i < n_regs; ++i) {
516 unsigned r = reg_prefs[i].num;
517 /* ignores are last and we should have at least 1 non-ignore left */
518 assert(!bitset_is_set(ignore_regs, r));
520 TODO: It might be better to copy the value occupying the register
521 around here instead of trying the next one, find out when... */
522 if (assignments[r].value != NULL)
524 reg = arch_register_for_index(cls, r);
525 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
531 static void free_reg_of_value(ir_node *node)
533 assignment_t *assignment;
534 const arch_register_t *reg;
537 if (!arch_irn_consider_in_reg_alloc(cls, node))
540 reg = arch_get_irn_register(node);
541 r = arch_register_get_index(reg);
542 assignment = &assignments[r];
543 assert(assignment->value == node);
544 assignment->value = NULL;
548 * Add an permutation in front of a node and change the assignments
549 * due to this permutation.
551 * To understand this imagine a permutation like this:
561 * First we count how many destinations a single value has. At the same time
562 * we can be sure that each destination register has at most 1 source register
563 * (it can have 0 which means we don't care what value is in it).
564 * We ignore all fullfilled permuations (like 7->7)
565 * In a first pass we create as much copy instructions as possible as they
566 * are generally cheaper than exchanges. We do this by counting into how many
567 * destinations a register has to be copied (in the example it's 2 for register
568 * 3, or 1 for the registers 1,2,4 and 7).
569 * We can then create a copy into every destination register when the usecount
570 * of that register is 0 (= noone else needs the value in the register).
572 * After this step we should have cycles left. We implement a cyclic permutation
573 * of n registers with n-1 transpositions.
575 * @param live_nodes the set of live nodes, updated due to live range split
576 * @param before the node before we add the permutation
577 * @param permutation the permutation array indices are the destination
578 * registers, the values in the array are the source
581 static void permutate_values(ir_nodeset_t *live_nodes, ir_node *before,
582 unsigned *permutation)
584 ir_node **ins = ALLOCANZ(ir_node*, n_regs);
585 unsigned *n_used = ALLOCANZ(unsigned, n_regs);
589 /* create a list of permutations. Leave out fix points. */
590 for (r = 0; r < n_regs; ++r) {
591 unsigned old_reg = permutation[r];
592 assignment_t *assignment;
595 /* no need to do anything for a fixpoint */
599 assignment = &assignments[old_reg];
600 value = assignment->value;
602 /* nothing to do here, reg is not live. Mark it as fixpoint
603 * so we ignore it in the next steps */
608 ins[old_reg] = value;
611 /* free occupation infos, we'll add the values back later */
612 if (live_nodes != NULL) {
613 ir_nodeset_remove(live_nodes, value);
617 block = get_nodes_block(before);
619 /* step1: create copies where immediately possible */
620 for (r = 0; r < n_regs; /* empty */) {
623 const arch_register_t *reg;
624 unsigned old_r = permutation[r];
626 /* - no need to do anything for fixed points.
627 - we can't copy if the value in the dest reg is still needed */
628 if (old_r == r || n_used[r] > 0) {
635 copy = be_new_Copy(cls, block, src);
636 sched_add_before(before, copy);
637 reg = arch_register_for_index(cls, r);
638 DB((dbg, LEVEL_2, "Copy %+F (from %+F, before %+F) -> %s\n",
639 copy, src, before, reg->name));
640 mark_as_copy_of(copy, src);
643 if (live_nodes != NULL) {
644 ir_nodeset_insert(live_nodes, copy);
647 /* old register has 1 user less, permutation is resolved */
648 assert(arch_register_get_index(arch_get_irn_register(src)) == old_r);
651 assert(n_used[old_r] > 0);
653 if (n_used[old_r] == 0) {
654 free_reg_of_value(src);
657 /* advance or jump back (if this copy enabled another copy) */
658 if (old_r < r && n_used[old_r] == 0) {
665 /* at this point we only have "cycles" left which we have to resolve with
667 * TODO: if we have free registers left, then we should really use copy
668 * instructions for any cycle longer than 2 registers...
669 * (this is probably architecture dependent, there might be archs where
670 * copies are preferable even for 2-cycles) */
672 /* create perms with the rest */
673 for (r = 0; r < n_regs; /* empty */) {
674 const arch_register_t *reg;
675 unsigned old_r = permutation[r];
687 /* we shouldn't have copies from 1 value to multiple destinations left*/
688 assert(n_used[old_r] == 1);
690 /* exchange old_r and r2; after that old_r is a fixed point */
691 r2 = permutation[old_r];
695 perm = be_new_Perm(cls, block, 2, in);
696 sched_add_before(before, perm);
697 DB((dbg, LEVEL_2, "Perm %+F (perm %+F,%+F, before %+F)\n",
698 perm, in[0], in[1], before));
700 proj0 = new_r_Proj(block, perm, get_irn_mode(in[0]), 0);
701 mark_as_copy_of(proj0, in[0]);
702 reg = arch_register_for_index(cls, old_r);
704 if (live_nodes != NULL) {
705 ir_nodeset_insert(live_nodes, proj0);
708 proj1 = new_r_Proj(block, perm, get_irn_mode(in[1]), 1);
710 /* 1 value is now in the correct register */
711 permutation[old_r] = old_r;
712 /* the source of r changed to r2 */
715 reg = arch_register_for_index(cls, r2);
717 /* if we have reached a fixpoint update data structures */
718 mark_as_copy_of(proj1, in[1]);
720 if (live_nodes != NULL) {
721 ir_nodeset_insert(live_nodes, proj1);
724 arch_set_irn_register(proj1, reg);
729 /* now we should only have fixpoints left */
730 for (r = 0; r < n_regs; ++r) {
731 assert(permutation[r] == r);
737 * Free regs for values last used.
739 * @param live_nodes set of live nodes, will be updated
740 * @param node the node to consider
742 static void free_last_uses(ir_nodeset_t *live_nodes, ir_node *node)
744 allocation_info_t *info = get_allocation_info(node);
745 const unsigned *last_uses = &info->last_uses;
746 int arity = get_irn_arity(node);
748 for (i = 0; i < arity; ++i) {
751 /* check if one operand is the last use */
752 if (!rbitset_is_set(last_uses, i))
755 op = get_irn_n(node, i);
756 free_reg_of_value(op);
757 ir_nodeset_remove(live_nodes, op);
762 * Create a bitset of registers occupied with value living through an
765 static void determine_live_through_regs(unsigned *bitset, ir_node *node)
767 const allocation_info_t *info = get_allocation_info(node);
772 /* mark all used registers as potentially live-through */
773 for (r = 0; r < n_regs; ++r) {
774 const assignment_t *assignment = &assignments[r];
775 if (assignment->value == NULL)
778 rbitset_set(bitset, r);
781 /* remove registers of value dying at the instruction */
782 arity = get_irn_arity(node);
783 for (i = 0; i < arity; ++i) {
785 const arch_register_t *reg;
787 if (!rbitset_is_set(&info->last_uses, i))
790 op = get_irn_n(node, i);
791 reg = arch_get_irn_register(op);
792 rbitset_clear(bitset, arch_register_get_index(reg));
797 * Enforce constraints at a node by live range splits.
799 * @param live_nodes the set of live nodes, might be changed
800 * @param node the current node
802 static void enforce_constraints(ir_nodeset_t *live_nodes, ir_node *node)
804 int arity = get_irn_arity(node);
806 hungarian_problem_t *bp;
808 unsigned *assignment;
810 /* see if any use constraints are not met */
812 for (i = 0; i < arity; ++i) {
813 ir_node *op = get_irn_n(node, i);
814 const arch_register_t *reg;
815 const arch_register_req_t *req;
816 const unsigned *limited;
819 if (!arch_irn_consider_in_reg_alloc(cls, op))
822 /* are there any limitations for the i'th operand? */
823 req = arch_get_register_req(node, i);
824 if (!(req->type & arch_register_req_type_limited))
827 limited = req->limited;
828 reg = arch_get_irn_register(op);
829 r = arch_register_get_index(reg);
830 if (!rbitset_is_set(limited, r)) {
831 /* found an assignement outside the limited set */
837 /* construct a list of register occupied by live-through values */
838 unsigned *live_through_regs = NULL;
839 unsigned *output_regs = NULL;
841 /* is any of the live-throughs using a constrainted output register? */
842 if (get_irn_mode(node) == mode_T) {
843 const ir_edge_t *edge;
845 foreach_out_edge(node, edge) {
846 ir_node *proj = get_edge_src_irn(edge);
847 const arch_register_req_t *req;
849 if (!arch_irn_consider_in_reg_alloc(cls, proj))
852 req = arch_get_register_req_out(proj);
853 if (!(req->type & arch_register_req_type_limited))
856 if (live_through_regs == NULL) {
857 rbitset_alloca(live_through_regs, n_regs);
858 determine_live_through_regs(live_through_regs, node);
860 rbitset_alloca(output_regs, n_regs);
863 rbitset_or(output_regs, req->limited, n_regs);
864 if (rbitsets_have_common(req->limited, live_through_regs, n_regs)) {
869 if (arch_irn_consider_in_reg_alloc(cls, node)) {
870 const arch_register_req_t *req = arch_get_register_req_out(node);
871 if (req->type & arch_register_req_type_limited) {
872 rbitset_alloca(live_through_regs, n_regs);
873 determine_live_through_regs(live_through_regs, node);
874 if (rbitsets_have_common(req->limited, live_through_regs, n_regs)) {
877 rbitset_alloca(output_regs, n_regs);
878 rbitset_or(output_regs, req->limited, n_regs);
887 /* create these arrays if we haven't yet */
888 if (output_regs == NULL) {
889 if (live_through_regs == NULL) {
890 rbitset_alloca(live_through_regs, n_regs);
892 rbitset_alloca(output_regs, n_regs);
895 /* at this point we have to construct a bipartite matching problem to see
896 which values should go to which registers */
897 bp = hungarian_new(n_regs, n_regs, HUNGARIAN_MATCH_PERFECT);
899 /* add all combinations, then remove not allowed ones */
900 for (l = 0; l < n_regs; ++l) {
901 if (bitset_is_set(ignore_regs, l)) {
902 hungarian_add(bp, l, l, 1);
906 for (r = 0; r < n_regs; ++r) {
907 if (bitset_is_set(ignore_regs, r))
909 /* livethrough values may not use constrainted output registers */
910 if (rbitset_is_set(live_through_regs, l)
911 && rbitset_is_set(output_regs, r))
914 hungarian_add(bp, r, l, l == r ? 9 : 8);
918 for (i = 0; i < arity; ++i) {
919 ir_node *op = get_irn_n(node, i);
920 const arch_register_t *reg;
921 const arch_register_req_t *req;
922 const unsigned *limited;
923 unsigned current_reg;
925 if (!arch_irn_consider_in_reg_alloc(cls, op))
928 req = arch_get_register_req(node, i);
929 if (!(req->type & arch_register_req_type_limited))
932 limited = req->limited;
933 reg = arch_get_irn_register(op);
934 current_reg = arch_register_get_index(reg);
935 for (r = 0; r < n_regs; ++r) {
936 if (rbitset_is_set(limited, r))
938 hungarian_remv(bp, r, current_reg);
942 hungarian_print_costmatrix(bp, 1);
943 hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
945 assignment = ALLOCAN(unsigned, n_regs);
946 res = hungarian_solve(bp, (int*) assignment, &dummy, 0);
950 printf("Swap result:");
951 for (i = 0; i < (int) n_regs; ++i) {
952 printf(" %d", assignment[i]);
959 permutate_values(live_nodes, node, assignment);
962 /** test wether a node @p n is a copy of the value of node @p of */
963 static bool is_copy_of(ir_node *value, ir_node *test_value)
965 allocation_info_t *test_info;
966 allocation_info_t *info;
968 if (value == test_value)
971 info = get_allocation_info(value);
972 test_info = get_allocation_info(test_value);
973 return test_info->original_value == info->original_value;
977 * find a value in the end-assignment of a basic block
978 * @returns the index into the assignment array if found
981 static int find_value_in_block_info(block_info_t *info, ir_node *value)
984 assignment_t *assignments = info->assignments;
985 for (r = 0; r < n_regs; ++r) {
986 const assignment_t *assignment = &assignments[r];
987 ir_node *a_value = assignment->value;
991 if (is_copy_of(a_value, value))
999 * Create the necessary permutations at the end of a basic block to fullfill
1000 * the register assignment for phi-nodes in the next block
1002 static void add_phi_permutations(ir_node *block, int p)
1005 unsigned *permutation;
1006 assignment_t *old_assignments;
1007 bool need_permutation;
1009 ir_node *pred = get_Block_cfgpred_block(block, p);
1011 block_info_t *pred_info = get_block_info(pred);
1013 /* predecessor not processed yet? nothing to do */
1014 if (!pred_info->processed)
1017 permutation = ALLOCAN(unsigned, n_regs);
1018 for (r = 0; r < n_regs; ++r) {
1022 /* check phi nodes */
1023 need_permutation = false;
1024 node = sched_first(block);
1025 for ( ; is_Phi(node); node = sched_next(node)) {
1026 const arch_register_t *reg;
1031 if (!arch_irn_consider_in_reg_alloc(cls, node))
1034 op = get_Phi_pred(node, p);
1035 a = find_value_in_block_info(pred_info, op);
1038 reg = arch_get_irn_register(node);
1039 regn = arch_register_get_index(reg);
1041 permutation[regn] = a;
1042 need_permutation = true;
1046 if (need_permutation) {
1047 /* permutate values at end of predecessor */
1048 old_assignments = assignments;
1049 assignments = pred_info->assignments;
1050 permutate_values(NULL, be_get_end_of_block_insertion_point(pred),
1052 assignments = old_assignments;
1055 /* change phi nodes to use the copied values */
1056 node = sched_first(block);
1057 for ( ; is_Phi(node); node = sched_next(node)) {
1061 if (!arch_irn_consider_in_reg_alloc(cls, node))
1064 /* we have permutated all values into the correct registers so we can
1065 simply query which value occupies the phis register in the
1067 a = arch_register_get_index(arch_get_irn_register(node));
1068 op = pred_info->assignments[a].value;
1069 set_Phi_pred(node, p, op);
1073 static void handle_phi_prefs(ir_node *phi)
1076 int arity = get_irn_arity(phi);
1077 ir_node *block = get_nodes_block(phi);
1078 allocation_info_t *info = get_allocation_info(phi);
1080 for (i = 0; i < arity; ++i) {
1081 ir_node *op = get_irn_n(phi, i);
1082 const arch_register_t *reg = arch_get_irn_register(op);
1090 /* give bonus for already assigned register */
1091 pred = get_Block_cfgpred_block(block, i);
1092 weight = get_block_execfreq(execfreqs, pred);
1093 r = arch_register_get_index(reg);
1094 info->prefs[r] += weight * AFF_PHI;
1099 * change inputs of a node to the current value (copies/perms)
1101 static void rewire_inputs(ir_node *node)
1104 int arity = get_irn_arity(node);
1106 for (i = 0; i < arity; ++i) {
1107 ir_node *op = get_irn_n(node, i);
1108 allocation_info_t *info;
1110 if (!arch_irn_consider_in_reg_alloc(cls, op))
1113 info = get_allocation_info(op);
1114 if (info->current_value != op) {
1115 set_irn_n(node, i, info->current_value);
1121 * Walker: assign registers to all nodes of a block that
1122 * need registers from the currently considered register class.
1124 static void allocate_coalesce_block(ir_node *block, void *data)
1127 ir_nodeset_t live_nodes;
1128 ir_nodeset_iterator_t iter;
1129 ir_node *node, *start;
1131 block_info_t *block_info;
1132 block_info_t **pred_block_infos;
1136 DB((dbg, LEVEL_2, "* Block %+F\n", block));
1138 /* clear assignments */
1139 block_info = get_block_info(block);
1140 assignments = block_info->assignments;
1142 ir_nodeset_init(&live_nodes);
1144 /* gather regalloc infos of predecessor blocks */
1145 n_preds = get_Block_n_cfgpreds(block);
1146 pred_block_infos = ALLOCAN(block_info_t*, n_preds);
1147 for (i = 0; i < n_preds; ++i) {
1148 ir_node *pred = get_Block_cfgpred_block(block, i);
1149 block_info_t *pred_info = get_block_info(pred);
1150 pred_block_infos[i] = pred_info;
1153 phi_ins = ALLOCAN(ir_node*, n_preds);
1155 /* collect live-in nodes and preassigned values */
1156 be_lv_foreach(lv, block, be_lv_state_in, i) {
1157 const arch_register_t *reg;
1160 node = be_lv_get_irn(lv, block, i);
1161 if (!arch_irn_consider_in_reg_alloc(cls, node))
1164 /* check all predecessors for this value, if it is not everywhere the
1165 same or unknown then we have to construct a phi
1166 (we collect the potential phi inputs here) */
1167 bool need_phi = false;
1168 for (p = 0; p < n_preds; ++p) {
1169 block_info_t *pred_info = pred_block_infos[p];
1171 if (!pred_info->processed) {
1172 /* use node for now, it will get fixed later */
1176 int a = find_value_in_block_info(pred_info, node);
1178 /* must live out of predecessor */
1180 phi_ins[p] = pred_info->assignments[a].value;
1181 /* different value from last time? then we need a phi */
1182 if (p > 0 && phi_ins[p-1] != phi_ins[p]) {
1189 ir_mode *mode = get_irn_mode(node);
1190 const arch_register_req_t *req = get_default_req_current_cls();
1193 phi = new_r_Phi(block, n_preds, phi_ins, mode);
1194 be_set_phi_reg_req(phi, req);
1196 DB((dbg, LEVEL_3, "Create Phi %+F (for %+F)\n", phi, node));
1198 mark_as_copy_of(phi, node);
1199 sched_add_after(block, phi);
1203 allocation_info_t *info = get_allocation_info(node);
1204 info->current_value = phi_ins[0];
1206 /* Grab 1 of the inputs we constructed (might not be the same as
1207 * "node" as we could see the same copy of the value in all
1212 /* if the node already has a register assigned use it */
1213 reg = arch_get_irn_register(node);
1215 /* TODO: consult pred-block infos here. The value could be copied
1216 away in some/all predecessor blocks. We need to construct
1217 phi-nodes in this case.
1218 We even need to construct some Phi_0 like constructs in cases
1219 where the predecessor allocation is not determined yet. */
1223 /* remember that this node is live at the beginning of the block */
1224 ir_nodeset_insert(&live_nodes, node);
1227 /* handle phis... */
1228 node = sched_first(block);
1229 for ( ; is_Phi(node); node = sched_next(node)) {
1230 const arch_register_t *reg;
1232 if (!arch_irn_consider_in_reg_alloc(cls, node))
1235 /* fill in regs already assigned */
1236 reg = arch_get_irn_register(node);
1240 /* TODO: give boni for registers already assigned at the
1242 handle_phi_prefs(node);
1243 assign_reg(block, node);
1248 /* assign regs for live-in values */
1249 foreach_ir_nodeset(&live_nodes, node, iter) {
1250 const arch_register_t *reg = arch_get_irn_register(node);
1254 assign_reg(block, node);
1257 /* assign instructions in the block */
1258 for (node = start; !sched_is_end(node); node = sched_next(node)) {
1259 /* enforce use constraints */
1260 enforce_constraints(&live_nodes, node);
1262 rewire_inputs(node);
1264 /* free registers of values last used at this instruction */
1265 free_last_uses(&live_nodes, node);
1267 /* assign output registers */
1268 /* TODO: 2 phases: first: pre-assigned ones, 2nd real regs */
1269 if (get_irn_mode(node) == mode_T) {
1270 const ir_edge_t *edge;
1271 foreach_out_edge(node, edge) {
1272 ir_node *proj = get_edge_src_irn(edge);
1273 if (!arch_irn_consider_in_reg_alloc(cls, proj))
1275 assign_reg(block, proj);
1277 } else if (arch_irn_consider_in_reg_alloc(cls, node)) {
1278 assign_reg(block, node);
1282 ir_nodeset_destroy(&live_nodes);
1285 block_info->processed = true;
1287 /* permutate values at end of predecessor blocks in case of phi-nodes */
1290 for (p = 0; p < n_preds; ++p) {
1291 add_phi_permutations(block, p);
1295 /* if we have exactly 1 successor then we might be able to produce phi
1297 if (get_irn_n_edges_kind(block, EDGE_KIND_BLOCK) == 1) {
1298 const ir_edge_t *edge
1299 = get_irn_out_edge_first_kind(block, EDGE_KIND_BLOCK);
1300 ir_node *succ = get_edge_src_irn(edge);
1301 int p = get_edge_src_pos(edge);
1302 block_info_t *succ_info = get_block_info(succ);
1304 if (succ_info->processed) {
1305 add_phi_permutations(succ, p);
1311 * Run the register allocator for the current register class.
1313 static void be_straight_alloc_cls(void)
1315 lv = be_assure_liveness(birg);
1316 be_liveness_assure_sets(lv);
1317 be_liveness_assure_chk(lv);
1319 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_IRN_VISITED);
1320 inc_irg_visited(irg);
1322 DB((dbg, LEVEL_2, "=== Allocating registers of %s ===\n", cls->name));
1324 irg_block_walk_graph(irg, NULL, analyze_block, NULL);
1325 /* we need some dominance pre-order walk to ensure we see all
1326 * definitions/create copies before we encounter their users */
1327 dom_tree_walk_irg(irg, allocate_coalesce_block, NULL, NULL);
1329 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_IRN_VISITED);
1332 static void dump(int mask, ir_graph *irg, const char *suffix,
1333 void (*dumper)(ir_graph *, const char *))
1335 if(birg->main_env->options->dump_flags & mask)
1336 be_dump(irg, suffix, dumper);
1340 * Run the spiller on the current graph.
1342 static void spill(void)
1344 /* make sure all nodes show their real register pressure */
1345 BE_TIMER_PUSH(t_ra_constr);
1346 be_pre_spill_prepare_constr(birg, cls);
1347 BE_TIMER_POP(t_ra_constr);
1349 dump(DUMP_RA, irg, "-spillprepare", dump_ir_block_graph_sched);
1352 BE_TIMER_PUSH(t_ra_spill);
1353 be_do_spill(birg, cls);
1354 BE_TIMER_POP(t_ra_spill);
1356 BE_TIMER_PUSH(t_ra_spill_apply);
1357 check_for_memory_operands(irg);
1358 BE_TIMER_POP(t_ra_spill_apply);
1360 dump(DUMP_RA, irg, "-spill", dump_ir_block_graph_sched);
1364 * The straight register allocator for a whole procedure.
1366 static void be_straight_alloc(be_irg_t *new_birg)
1368 const arch_env_t *arch_env = new_birg->main_env->arch_env;
1369 int n_cls = arch_env_get_n_reg_class(arch_env);
1372 obstack_init(&obst);
1375 irg = be_get_birg_irg(birg);
1376 execfreqs = birg->exec_freq;
1378 /* TODO: extract some of the stuff from bechordal allocator, like
1379 * statistics, time measurements, etc. and use them here too */
1381 for (c = 0; c < n_cls; ++c) {
1382 cls = arch_env_get_reg_class(arch_env, c);
1383 default_cls_req = NULL;
1384 if (arch_register_class_flags(cls) & arch_register_class_flag_manual_ra)
1387 stat_ev_ctx_push_str("regcls", cls->name);
1389 n_regs = arch_register_class_n_regs(cls);
1390 ignore_regs = bitset_malloc(n_regs);
1391 be_put_ignore_regs(birg, cls, ignore_regs);
1395 /* verify schedule and register pressure */
1396 BE_TIMER_PUSH(t_verify);
1397 if (birg->main_env->options->vrfy_option == BE_VRFY_WARN) {
1398 be_verify_schedule(birg);
1399 be_verify_register_pressure(birg, cls, irg);
1400 } else if (birg->main_env->options->vrfy_option == BE_VRFY_ASSERT) {
1401 assert(be_verify_schedule(birg) && "Schedule verification failed");
1402 assert(be_verify_register_pressure(birg, cls, irg)
1403 && "Register pressure verification failed");
1405 BE_TIMER_POP(t_verify);
1407 BE_TIMER_PUSH(t_ra_color);
1408 be_straight_alloc_cls();
1409 BE_TIMER_POP(t_ra_color);
1411 /* we most probably constructed new Phis so liveness info is invalid
1413 /* TODO: test liveness_introduce */
1414 be_liveness_invalidate(lv);
1416 bitset_free(ignore_regs);
1418 stat_ev_ctx_pop("regcls");
1421 BE_TIMER_PUSH(t_ra_spill_apply);
1422 be_abi_fix_stack_nodes(birg->abi);
1423 BE_TIMER_POP(t_ra_spill_apply);
1425 BE_TIMER_PUSH(t_verify);
1426 if (birg->main_env->options->vrfy_option == BE_VRFY_WARN) {
1427 be_verify_register_allocation(birg);
1428 } else if (birg->main_env->options->vrfy_option == BE_VRFY_ASSERT) {
1429 assert(be_verify_register_allocation(birg)
1430 && "Register allocation invalid");
1432 BE_TIMER_POP(t_verify);
1434 obstack_free(&obst, NULL);
1438 * Initializes this module.
1440 void be_init_straight_alloc(void)
1442 static be_ra_t be_ra_straight = {
1446 FIRM_DBG_REGISTER(dbg, "firm.be.straightalloc");
1448 be_register_allocator("straight", &be_ra_straight);
1451 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_straight_alloc);