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 * - output constraints are not ensured. The algorithm fails to copy values
44 * away, so the registers for constrained outputs are free.
45 * - must_be_different constraint is not respected
46 * - We have to pessimistically construct Phi_0s when not all predecessors
47 * of a block are known.
48 * - Phi color assignment should give bonus points towards registers already
49 * assigned at predecessors.
50 * - think about a smarter sequence of visiting the blocks. Sorted by
51 * execfreq might be good, or looptree from inner to outermost loops going
52 * over blocks in a reverse postorder
60 #include "irgraph_t.h"
61 #include "iredges_t.h"
70 #include "bechordal_t.h"
75 #include "bespillutil.h"
78 #include "bipartite.h"
79 #include "hungarian.h"
81 #define USE_FACTOR 1.0f
82 #define DEF_FACTOR 1.0f
83 #define NEIGHBOR_FACTOR 0.2f
84 #define SHOULD_BE_SAME 1.0f
86 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
88 static struct obstack obst;
89 static be_irg_t *birg;
91 static const arch_register_class_t *cls;
93 static const ir_exec_freq *execfreqs;
94 static unsigned n_regs;
95 static bitset_t *ignore_regs;
97 /** info about the current assignment for a register */
99 ir_node *value; /**< currently assigned value */
101 typedef struct assignment_t assignment_t;
103 /** currently active assignments (while processing a basic block) */
104 static assignment_t *assignments;
107 * allocation information: last_uses, register preferences
108 * the information is per firm-node.
110 struct allocation_info_t {
111 unsigned last_uses; /**< bitset indicating last uses (input pos) */
112 assignment_t *current_assignment;
113 float prefs[0]; /**< register preferences */
115 typedef struct allocation_info_t allocation_info_t;
117 /** helper datastructure used when sorting register preferences */
122 typedef struct reg_pref_t reg_pref_t;
124 /** per basic-block information */
125 struct block_info_t {
126 int processed; /**< indicate wether block is processed */
127 assignment_t assignments[0]; /**< register assignments at end of block */
129 typedef struct block_info_t block_info_t;
132 * Get the allocation info for a node.
133 * The info is allocated on the first visit of a node.
135 static allocation_info_t *get_allocation_info(ir_node *node)
137 allocation_info_t *info;
138 if (!irn_visited(node)) {
139 size_t size = sizeof(info[0]) + n_regs * sizeof(info->prefs[0]);
140 info = obstack_alloc(&obst, size);
141 memset(info, 0, size);
142 set_irn_link(node, info);
143 mark_irn_visited(node);
145 info = get_irn_link(node);
152 * Get allocation information for a basic block
154 static block_info_t *get_block_info(ir_node *block)
158 assert(is_Block(block));
159 if (!irn_visited(block)) {
160 size_t size = sizeof(info[0]) + n_regs * sizeof(info->assignments[0]);
161 info = obstack_alloc(&obst, size);
162 memset(info, 0, size);
163 set_irn_link(block, info);
164 mark_irn_visited(block);
166 info = get_irn_link(block);
173 * Link the allocation info of a node to a copy.
174 * Afterwards, both nodes uses the same allocation info.
175 * Copy must not have an allocation info assigned yet.
177 * @param copy the node that gets the allocation info assigned
178 * @param value the original node
180 static void link_to(ir_node *copy, ir_node *value)
182 allocation_info_t *info = get_allocation_info(value);
183 assert(!irn_visited(copy));
184 set_irn_link(copy, info);
185 mark_irn_visited(copy);
189 * Calculate the penalties for every register on a node and its live neighbors.
191 * @param live_nodes the set of live nodes at the current position, may be NULL
192 * @param penalty the penalty to subtract from
193 * @param limited a raw bitset containing the limited set for the node
194 * @param node the node
196 static void give_penalties_for_limits(const ir_nodeset_t *live_nodes,
197 float penalty, const unsigned* limited,
200 ir_nodeset_iterator_t iter;
202 allocation_info_t *info = get_allocation_info(node);
205 /* give penalty for all forbidden regs */
206 for (r = 0; r < n_regs; ++r) {
207 if (rbitset_is_set(limited, r))
210 info->prefs[r] -= penalty;
213 /* all other live values should get a penalty for allowed regs */
214 if (live_nodes == NULL)
217 /* TODO: reduce penalty if there are multiple allowed registers... */
218 penalty *= NEIGHBOR_FACTOR;
219 foreach_ir_nodeset(live_nodes, neighbor, iter) {
220 allocation_info_t *neighbor_info;
222 /* TODO: if op is used on multiple inputs we might not do a
224 if (neighbor == node)
227 neighbor_info = get_allocation_info(neighbor);
228 for (r = 0; r < n_regs; ++r) {
229 if (!rbitset_is_set(limited, r))
232 neighbor_info->prefs[r] -= penalty;
238 * Calculate the preferences of a definition for the current register class.
239 * If the definition uses a limited set of registers, reduce the preferences
240 * for the limited register on the node and its neighbors.
242 * @param live_nodes the set of live nodes at the current node
243 * @param weight the weight
244 * @param node the current node
246 static void check_defs(const ir_nodeset_t *live_nodes, float weight,
249 const arch_register_req_t *req;
251 if (get_irn_mode(node) == mode_T) {
252 const ir_edge_t *edge;
253 foreach_out_edge(node, edge) {
254 ir_node *proj = get_edge_src_irn(edge);
255 check_defs(live_nodes, weight, proj);
260 if (!arch_irn_consider_in_reg_alloc(cls, node))
263 req = arch_get_register_req_out(node);
264 if (req->type & arch_register_req_type_limited) {
265 const unsigned *limited = req->limited;
266 float penalty = weight * DEF_FACTOR;
267 give_penalties_for_limits(live_nodes, penalty, limited, node);
270 if (req->type & arch_register_req_type_should_be_same) {
271 ir_node *insn = skip_Proj(node);
272 allocation_info_t *info = get_allocation_info(node);
273 int arity = get_irn_arity(insn);
276 float factor = 1.0f / rbitset_popcnt(&req->other_same, arity);
277 for (i = 0; i < arity; ++i) {
280 allocation_info_t *op_info;
282 if (!rbitset_is_set(&req->other_same, i))
285 op = get_irn_n(insn, i);
286 op_info = get_allocation_info(op);
287 for (r = 0; r < n_regs; ++r) {
288 if (bitset_is_set(ignore_regs, r))
290 op_info->prefs[r] += info->prefs[r] * factor;
297 * Walker: Runs an a block calculates the preferences for any
298 * node and every register from the considered register class.
300 static void analyze_block(ir_node *block, void *data)
302 float weight = get_block_execfreq(execfreqs, block);
303 ir_nodeset_t live_nodes;
307 ir_nodeset_init(&live_nodes);
308 be_liveness_end_of_block(lv, cls, block, &live_nodes);
310 sched_foreach_reverse(block, node) {
311 allocation_info_t *info;
315 /* TODO: handle constrained phi-nodes */
319 /* TODO give/take penalties for should_be_same/different) */
320 check_defs(&live_nodes, weight, node);
323 arity = get_irn_arity(node);
324 /* I was lazy, and only allocated 1 unsigned
325 => maximum of 32 uses per node (rewrite if necessary) */
326 assert(arity <= (int) sizeof(unsigned) * 8);
328 info = get_allocation_info(node);
329 for (i = 0; i < arity; ++i) {
330 ir_node *op = get_irn_n(node, i);
331 if (!arch_irn_consider_in_reg_alloc(cls, op))
334 /* last usage of a value? */
335 if (!ir_nodeset_contains(&live_nodes, op)) {
336 rbitset_set(&info->last_uses, i);
340 be_liveness_transfer(cls, node, &live_nodes);
342 /* update weights based on usage constraints */
343 for (i = 0; i < arity; ++i) {
344 const arch_register_req_t *req;
345 const unsigned *limited;
346 ir_node *op = get_irn_n(node, i);
348 if (!arch_irn_consider_in_reg_alloc(cls, op))
351 req = arch_get_register_req(node, i);
352 if ((req->type & arch_register_req_type_limited) == 0)
355 /* TODO: give penalties to neighbors for precolored nodes! */
357 limited = req->limited;
358 give_penalties_for_limits(&live_nodes, weight * USE_FACTOR, limited,
363 ir_nodeset_destroy(&live_nodes);
367 * Assign register reg to the given node.
369 * @param node the node
370 * @param reg the register
372 static void use_reg(ir_node *node, const arch_register_t *reg)
374 unsigned r = arch_register_get_index(reg);
375 assignment_t *assignment = &assignments[r];
376 allocation_info_t *info;
378 assert(assignment->value == NULL);
379 assignment->value = node;
381 info = get_allocation_info(node);
382 info->current_assignment = assignment;
384 arch_set_irn_register(node, reg);
388 * Compare two register preferences in decreasing order.
390 static int compare_reg_pref(const void *e1, const void *e2)
392 const reg_pref_t *rp1 = (const reg_pref_t*) e1;
393 const reg_pref_t *rp2 = (const reg_pref_t*) e2;
394 if (rp1->pref < rp2->pref)
396 if (rp1->pref > rp2->pref)
401 static void fill_sort_candidates(reg_pref_t *regprefs,
402 const allocation_info_t *info)
406 for (r = 0; r < n_regs; ++r) {
407 float pref = info->prefs[r];
408 if (bitset_is_set(ignore_regs, r)) {
412 regprefs[r].pref = pref;
414 /* TODO: use a stable sort here to avoid unnecessary register jumping */
415 qsort(regprefs, n_regs, sizeof(regprefs[0]), compare_reg_pref);
419 * Determine and assign a register for node @p node
421 static void assign_reg(const ir_node *block, ir_node *node)
423 const arch_register_t *reg;
424 allocation_info_t *info;
425 const arch_register_req_t *req;
426 reg_pref_t *reg_prefs;
430 assert(arch_irn_consider_in_reg_alloc(cls, node));
432 /* preassigned register? */
433 reg = arch_get_irn_register(node);
435 DB((dbg, LEVEL_2, "Preassignment %+F -> %s\n", node, reg->name));
440 /* give should_be_same boni */
441 info = get_allocation_info(node);
442 req = arch_get_register_req_out(node);
444 in_node = skip_Proj(node);
445 if (req->type & arch_register_req_type_should_be_same) {
446 float weight = get_block_execfreq(execfreqs, block);
447 int arity = get_irn_arity(in_node);
450 assert(arity <= (int) sizeof(req->other_same) * 8);
451 for (i = 0; i < arity; ++i) {
453 const arch_register_t *reg;
455 if (!rbitset_is_set(&req->other_same, i))
458 in = get_irn_n(in_node, i);
459 reg = arch_get_irn_register(in);
461 r = arch_register_get_index(reg);
462 if (bitset_is_set(ignore_regs, r))
464 info->prefs[r] += weight * SHOULD_BE_SAME;
468 /* TODO: handle must_be_different */
471 DB((dbg, LEVEL_2, "Candidates for %+F:", node));
472 reg_prefs = alloca(n_regs * sizeof(reg_prefs[0]));
473 fill_sort_candidates(reg_prefs, info);
474 for (i = 0; i < n_regs; ++i) {
475 unsigned num = reg_prefs[i].num;
476 const arch_register_t *reg = arch_register_for_index(cls, num);
477 DB((dbg, LEVEL_2, " %s(%f)", reg->name, reg_prefs[i].pref));
479 DB((dbg, LEVEL_2, "\n"));
481 for (i = 0; i < n_regs; ++i) {
482 unsigned r = reg_prefs[i].num;
483 /* ignores should be last and we should have a non-ignore left */
484 assert(!bitset_is_set(ignore_regs, r));
486 TODO: It might be better to copy the value occupying the register around here, find out when... */
487 if (assignments[r].value != NULL)
489 reg = arch_register_for_index(cls, r);
490 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
496 static void free_reg_of_value(ir_node *node)
498 allocation_info_t *info;
499 assignment_t *assignment;
502 if (!arch_irn_consider_in_reg_alloc(cls, node))
505 info = get_allocation_info(node);
506 assignment = info->current_assignment;
508 assert(assignment != NULL);
510 r = assignment - assignments;
511 DB((dbg, LEVEL_2, "Value %+F ended, freeing %s\n",
512 node, arch_register_for_index(cls, r)->name));
513 assignment->value = NULL;
514 info->current_assignment = NULL;
518 * Return the index of the currently assigned register of a node.
520 static unsigned get_current_reg(ir_node *node)
522 allocation_info_t *info = get_allocation_info(node);
523 assignment_t *assignment = info->current_assignment;
524 return assignment - assignments;
528 * Return the currently assigned assignment of a node.
530 static assignment_t *get_current_assignment(ir_node *node)
532 allocation_info_t *info = get_allocation_info(node);
533 return info->current_assignment;
537 * Add an permutation in front of a node and change the assignments
538 * due to this permutation.
540 * To understand this imagine a permutation like this:
550 * First we count how many destinations a single value has. At the same time
551 * we can be sure that each destination register has at most 1 source register
552 * (it can have 0 which means we don't care what value is in it).
553 * We ignore all fullfilled permuations (like 7->7)
554 * In a first pass we create as much copy instructions as possible as they
555 * are generally cheaper than exchanges. We do this by counting into how many
556 * destinations a register has to be copied (in the example it's 2 for register
557 * 3, or 1 for the registers 1,2,4 and 7).
558 * We can then create a copy into every destination register when the usecount
559 * of that register is 0 (= noone else needs the value in the register).
561 * After this step we should have cycles left. We implement a cyclic permutation
562 * of n registers with n-1 transpositions.
564 * @param live_nodes the set of live nodes, updated due to live range split
565 * @param before the node before we add the permutation
566 * @param permutation the permutation array indices are the destination
567 * registers, the values in the array are the source
570 static void permutate_values(ir_nodeset_t *live_nodes, ir_node *before,
571 unsigned *permutation)
574 ir_node **ins = ALLOCANZ(ir_node*, n_regs);
575 unsigned *n_used = ALLOCANZ(unsigned, n_regs);
578 /* create a list of permutations. Leave out fix points. */
579 for (r = 0; r < n_regs; ++r) {
580 unsigned old_reg = permutation[r];
581 assignment_t *assignment;
584 /* no need to do anything for a fixpoint */
588 assignment = &assignments[old_reg];
589 value = assignment->value;
591 /* nothing to do here, reg is not live. Mark it as fixpoint
592 * so we ignore it in the next steps */
597 ins[old_reg] = value;
600 /* free occupation infos, we'll add the values back later */
601 if (live_nodes != NULL) {
602 free_reg_of_value(value);
603 ir_nodeset_remove(live_nodes, value);
607 block = get_nodes_block(before);
609 /* step1: create copies where immediately possible */
610 for (r = 0; r < n_regs; /* empty */) {
613 const arch_register_t *reg;
614 unsigned old_r = permutation[r];
616 /* - no need to do anything for fixed points.
617 - we can't copy if the value in the dest reg is still needed */
618 if (old_r == r || n_used[r] > 0) {
625 copy = be_new_Copy(cls, block, src);
626 reg = arch_register_for_index(cls, r);
627 DB((dbg, LEVEL_2, "Copy %+F (from %+F) -> %s\n", copy, src, reg->name));
630 sched_add_before(before, copy);
632 /* old register has 1 user less, permutation is resolved */
633 assert(arch_register_get_index(arch_get_irn_register(src)) == old_r);
634 assert(n_used[old_r] > 0);
638 /* advance or jump back (this copy could have enabled another copy) */
639 if (old_r < r && n_used[old_r] == 0) {
646 /* at this point we only have "cycles" left which we have to resolve with
648 * TODO: if we have free registers left, then we should really use copy
649 * instructions for any cycle longer than 2 registers...
650 * (this is probably architecture dependent, there might be archs where
651 * copies are preferable even for 2 cycles)
654 /* create perms with the rest */
655 for (r = 0; r < n_regs; /* empty */) {
656 const arch_register_t *reg;
657 unsigned old_r = permutation[r];
669 /* we shouldn't have copies from 1 value to multiple destinations left*/
670 assert(n_used[old_r] == 1);
672 /* exchange old_r and r2; after that old_r is a fixed point */
673 r2 = permutation[old_r];
677 perm = be_new_Perm(cls, block, 2, in);
679 proj0 = new_r_Proj(block, perm, get_irn_mode(in[0]), 0);
680 link_to(proj0, in[0]);
681 reg = arch_register_for_index(cls, old_r);
684 proj1 = new_r_Proj(block, perm, get_irn_mode(in[1]), 1);
686 /* 1 value is now in the correct register */
687 permutation[old_r] = old_r;
688 /* the source of r changed to r2 */
691 reg = arch_register_for_index(cls, r2);
693 /* if we have reached a fixpoint update data structures */
694 link_to(proj1, in[1]);
697 arch_set_irn_register(proj1, reg);
702 /* now we should only have fixpoints left */
703 for (r = 0; r < n_regs; ++r) {
704 assert(permutation[r] == r);
710 * Free regs for values last used.
712 * @param live_nodes set of live nodes, will be updated
713 * @param node the node to consider
715 static void free_last_uses(ir_nodeset_t *live_nodes, ir_node *node)
717 allocation_info_t *info = get_allocation_info(node);
718 int arity = get_irn_arity(node);
720 for (i = 0; i < arity; ++i) {
723 /* check if one operand is the last use */
724 if (!rbitset_is_set(&info->last_uses, i))
727 op = get_irn_n(node, i);
728 free_reg_of_value(op);
729 ir_nodeset_remove(live_nodes, op);
734 * Create a bitset of registers occupied with value living through an
737 static void determine_live_through_regs(unsigned *bitset, ir_node *node)
739 const allocation_info_t *info = get_allocation_info(node);
744 /* mark all used registers as potentially live-through */
745 for (r = 0; r < n_regs; ++r) {
746 const assignment_t *assignment = &assignments[r];
747 if (assignment->value == NULL)
750 rbitset_set(bitset, r);
753 /* remove registers of value dying at the instruction */
754 arity = get_irn_arity(node);
755 for (i = 0; i < arity; ++i) {
757 const arch_register_t *reg;
759 if (!rbitset_is_set(&info->last_uses, i))
762 op = get_irn_n(node, i);
763 reg = arch_get_irn_register(op);
764 rbitset_clear(bitset, arch_register_get_index(reg));
769 * Enforce constraints at a node by live range splits.
771 * @param live_nodes the set of live nodes, might be changed
772 * @param node the current node
774 static void enforce_constraints(ir_nodeset_t *live_nodes, ir_node *node)
776 int arity = get_irn_arity(node);
778 hungarian_problem_t *bp;
780 unsigned *assignment;
782 /* see if any use constraints are not met */
784 for (i = 0; i < arity; ++i) {
785 ir_node *op = get_irn_n(node, i);
786 const arch_register_req_t *req;
787 const unsigned *limited;
790 if (!arch_irn_consider_in_reg_alloc(cls, op))
793 /* are there any limitations for the i'th operand? */
794 req = arch_get_register_req(node, i);
795 if ((req->type & arch_register_req_type_limited) == 0)
798 limited = req->limited;
799 r = get_current_reg(op);
800 if (!rbitset_is_set(limited, r)) {
801 /* found an assignement outside the limited set */
807 /* construct a list of register occupied by live-through values */
808 unsigned *live_through_regs = NULL;
809 unsigned *output_regs = NULL;
811 /* is any of the live-throughs using a constrainted output register? */
812 if (get_irn_mode(node) == mode_T) {
813 const ir_edge_t *edge;
815 foreach_out_edge(node, edge) {
816 ir_node *proj = get_edge_src_irn(edge);
817 const arch_register_req_t *req;
819 if (!arch_irn_consider_in_reg_alloc(cls, proj))
822 req = arch_get_register_req_out(proj);
823 if (! (req->type & arch_register_req_type_limited))
826 if (live_through_regs == NULL) {
827 rbitset_alloca(live_through_regs, n_regs);
828 determine_live_through_regs(live_through_regs, node);
830 rbitset_alloca(output_regs, n_regs);
833 rbitset_or(output_regs, req->limited, n_regs);
834 if (rbitsets_have_common(req->limited, live_through_regs, n_regs)) {
840 if (arch_irn_consider_in_reg_alloc(cls, node)) {
841 const arch_register_req_t *req = arch_get_register_req_out(node);
842 if (req->type & arch_register_req_type_limited) {
843 rbitset_alloca(live_through_regs, n_regs);
844 determine_live_through_regs(live_through_regs, node);
845 if (rbitsets_have_common(req->limited, live_through_regs, n_regs)) {
848 rbitset_alloca(output_regs, n_regs);
849 rbitset_or(output_regs, req->limited, n_regs);
858 if (live_through_regs == NULL) {
859 rbitset_alloca(live_through_regs, n_regs);
860 rbitset_alloca(output_regs, n_regs);
863 /* swap values around */
864 bp = hungarian_new(n_regs, n_regs, HUNGARIAN_MATCH_PERFECT);
866 /* add all combinations, then remove not allowed ones */
867 for (l = 0; l < n_regs; ++l) {
868 if (bitset_is_set(ignore_regs, l)) {
869 hungarian_add(bp, l, l, 90);
873 for (r = 0; r < n_regs; ++r) {
874 if (bitset_is_set(ignore_regs, r))
876 /* livethrough values may not use constrainted output registers */
877 if (rbitset_is_set(live_through_regs, l)
878 && rbitset_is_set(output_regs, r))
881 hungarian_add(bp, l, r, l == r ? 90 : 89);
885 for (i = 0; i < arity; ++i) {
886 ir_node *op = get_irn_n(node, i);
887 const arch_register_req_t *req;
888 const unsigned *limited;
889 unsigned current_reg;
891 if (!arch_irn_consider_in_reg_alloc(cls, op))
894 req = arch_get_register_req(node, i);
895 if ((req->type & arch_register_req_type_limited) == 0)
898 limited = req->limited;
899 current_reg = get_current_reg(op);
900 for (r = 0; r < n_regs; ++r) {
901 if (rbitset_is_set(limited, r))
903 hungarian_remv(bp, current_reg, r);
907 hungarian_print_costmatrix(bp, 1);
908 hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
910 assignment = ALLOCAN(unsigned, n_regs);
911 res = hungarian_solve(bp, (int*) assignment, &dummy, 0);
914 printf("Swap result:");
915 for (p = 0; p < n_regs; ++p) {
916 printf(" %d", assignment[p]);
922 permutate_values(live_nodes, node, assignment);
925 /** test wether a node @p n is a copy of the value of node @p of */
926 static int is_copy_of(ir_node *n, ir_node *of)
928 allocation_info_t *of_info;
936 of_info = get_allocation_info(of);
940 return of_info == get_irn_link(n);
943 /** find a value in the end-assignment of a basic block
944 * @returns the index into the assignment array if found
947 static int find_value_in_block_info(block_info_t *info, ir_node *value)
950 assignment_t *assignments = info->assignments;
951 for (r = 0; r < n_regs; ++r) {
952 const assignment_t *assignment = &assignments[r];
953 if (is_copy_of(assignment->value, value))
961 * Create the necessary permutations at the end of a basic block to fullfill
962 * the register assignment for phi-nodes in the next block
964 static void add_phi_permutations(ir_node *block, int p)
967 unsigned *permutation;
968 assignment_t *old_assignments;
969 int need_permutation;
971 ir_node *pred = get_Block_cfgpred_block(block, p);
973 block_info_t *pred_info = get_block_info(pred);
975 /* predecessor not processed yet? nothing to do */
976 if (!pred_info->processed)
979 permutation = ALLOCAN(unsigned, n_regs);
980 for (r = 0; r < n_regs; ++r) {
984 /* check phi nodes */
985 need_permutation = 0;
986 node = sched_first(block);
987 for ( ; is_Phi(node); node = sched_next(node)) {
988 const arch_register_t *reg;
993 if (!arch_irn_consider_in_reg_alloc(cls, node))
996 op = get_Phi_pred(node, p);
997 a = find_value_in_block_info(pred_info, op);
1000 reg = arch_get_irn_register(node);
1001 regn = arch_register_get_index(reg);
1003 permutation[regn] = a;
1004 need_permutation = 1;
1008 old_assignments = assignments;
1009 assignments = pred_info->assignments;
1010 permutate_values(NULL, be_get_end_of_block_insertion_point(pred),
1012 assignments = old_assignments;
1014 node = sched_first(block);
1015 for ( ; is_Phi(node); node = sched_next(node)) {
1019 if (!arch_irn_consider_in_reg_alloc(cls, node))
1022 op = get_Phi_pred(node, p);
1023 /* TODO: optimize */
1024 a = find_value_in_block_info(pred_info, op);
1027 op = pred_info->assignments[a].value;
1028 set_Phi_pred(node, p, op);
1033 * Walker: assign registers to all nodes of a block that
1034 * needs registers from the currently considered register class.
1036 static void allocate_coalesce_block(ir_node *block, void *data)
1040 ir_nodeset_t live_nodes;
1041 ir_nodeset_iterator_t iter;
1042 ir_node *node, *start;
1044 block_info_t *block_info;
1045 block_info_t **pred_block_infos;
1048 DB((dbg, LEVEL_2, "Allocating in block %+F\n",
1051 /* clear assignments */
1052 block_info = get_block_info(block);
1053 assignments = block_info->assignments;
1055 for (r = 0; r < n_regs; ++r) {
1056 assignment_t *assignment = &assignments[r];
1057 ir_node *value = assignment->value;
1058 allocation_info_t *info;
1063 info = get_allocation_info(value);
1064 info->current_assignment = assignment;
1067 ir_nodeset_init(&live_nodes);
1069 /* gather regalloc infos of predecessor blocks */
1070 n_preds = get_Block_n_cfgpreds(block);
1071 pred_block_infos = ALLOCAN(block_info_t*, n_preds);
1072 for (i = 0; i < n_preds; ++i) {
1073 ir_node *pred = get_Block_cfgpred_block(block, i);
1074 pred_block_infos[i] = get_block_info(pred);
1077 /* collect live-in nodes and preassigned values */
1078 be_lv_foreach(lv, block, be_lv_state_in, i) {
1079 const arch_register_t *reg;
1081 node = be_lv_get_irn(lv, block, i);
1082 if (!arch_irn_consider_in_reg_alloc(cls, node))
1085 /* remember that this node is live at the beginning of the block */
1086 ir_nodeset_insert(&live_nodes, node);
1088 /* if the node already has a register assigned use it */
1089 reg = arch_get_irn_register(node);
1091 /* TODO: consult pred-block infos here. The value could be copied
1092 away in some/all predecessor blocks. We need to construct
1093 phi-nodes in this case.
1094 We even need to construct some Phi_0 like constructs in cases
1095 where the predecessor allocation is not determined yet. */
1100 /* handle phis... */
1101 node = sched_first(block);
1102 for ( ; is_Phi(node); node = sched_next(node)) {
1103 const arch_register_t *reg;
1105 if (!arch_irn_consider_in_reg_alloc(cls, node))
1108 /* fill in regs already assigned */
1109 reg = arch_get_irn_register(node);
1113 /* TODO: give boni for registers already assigned at the
1115 assign_reg(block, node);
1120 /* assign regs for live-in values */
1121 foreach_ir_nodeset(&live_nodes, node, iter) {
1122 const arch_register_t *reg = arch_get_irn_register(node);
1126 assign_reg(block, node);
1129 /* permutate values at end of predecessor blocks in case of phi-nodes */
1132 for (p = 0; p < n_preds; ++p) {
1133 add_phi_permutations(block, p);
1137 /* assign instructions in the block */
1138 for (node = start; !sched_is_end(node); node = sched_next(node)) {
1139 int arity = get_irn_arity(node);
1142 /* enforce use constraints */
1143 enforce_constraints(&live_nodes, node);
1145 /* exchange values to copied values where needed */
1146 for (i = 0; i < arity; ++i) {
1147 ir_node *op = get_irn_n(node, i);
1148 assignment_t *assignment;
1150 if (!arch_irn_consider_in_reg_alloc(cls, op))
1152 assignment = get_current_assignment(op);
1153 assert(assignment != NULL);
1154 if (op != assignment->value) {
1155 set_irn_n(node, i, assignment->value);
1159 /* free registers of values last used at this instruction */
1160 free_last_uses(&live_nodes, node);
1162 /* assign output registers */
1163 /* TODO: 2 phases: first: pre-assigned ones, 2nd real regs */
1164 if (get_irn_mode(node) == mode_T) {
1165 const ir_edge_t *edge;
1166 foreach_out_edge(node, edge) {
1167 ir_node *proj = get_edge_src_irn(edge);
1168 if (!arch_irn_consider_in_reg_alloc(cls, proj))
1170 assign_reg(block, proj);
1172 } else if (arch_irn_consider_in_reg_alloc(cls, node)) {
1173 assign_reg(block, node);
1177 ir_nodeset_destroy(&live_nodes);
1180 block_info->processed = 1;
1182 /* if we have exactly 1 successor then we might be able to produce phi
1184 if (get_irn_n_edges_kind(block, EDGE_KIND_BLOCK) == 1) {
1185 const ir_edge_t *edge
1186 = get_irn_out_edge_first_kind(block, EDGE_KIND_BLOCK);
1187 ir_node *succ = get_edge_src_irn(edge);
1188 int p = get_edge_src_pos(edge);
1189 block_info_t *succ_info = get_block_info(succ);
1191 if (succ_info->processed) {
1192 add_phi_permutations(succ, p);
1198 * Run the register allocator for the current register class.
1200 static void be_straight_alloc_cls(void)
1202 n_regs = arch_register_class_n_regs(cls);
1203 lv = be_assure_liveness(birg);
1204 be_liveness_assure_sets(lv);
1205 be_liveness_assure_chk(lv);
1209 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_IRN_VISITED);
1210 inc_irg_visited(irg);
1212 DB((dbg, LEVEL_2, "=== Allocating registers of %s ===\n", cls->name));
1214 irg_block_walk_graph(irg, NULL, analyze_block, NULL);
1215 irg_block_walk_graph(irg, NULL, allocate_coalesce_block, NULL);
1217 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_IRN_VISITED);
1221 * Run the spiller on the current graph.
1223 static void spill(void)
1225 /* make sure all nodes show their real register pressure */
1226 BE_TIMER_PUSH(t_ra_constr);
1227 be_pre_spill_prepare_constr(birg, cls);
1228 BE_TIMER_POP(t_ra_constr);
1231 BE_TIMER_PUSH(t_ra_spill);
1232 be_do_spill(birg, cls);
1233 BE_TIMER_POP(t_ra_spill);
1235 BE_TIMER_PUSH(t_ra_spill_apply);
1236 check_for_memory_operands(irg);
1237 BE_TIMER_POP(t_ra_spill_apply);
1241 * The straight register allocator for a whole procedure.
1243 static void be_straight_alloc(be_irg_t *new_birg)
1245 const arch_env_t *arch_env = new_birg->main_env->arch_env;
1246 int n_cls = arch_env_get_n_reg_class(arch_env);
1249 obstack_init(&obst);
1252 irg = be_get_birg_irg(birg);
1253 execfreqs = birg->exec_freq;
1255 /* TODO: extract some of the stuff from bechordal allocator, like
1256 * statistics, time measurements, etc. and use them here too */
1258 for (c = 0; c < n_cls; ++c) {
1259 cls = arch_env_get_reg_class(arch_env, c);
1260 if (arch_register_class_flags(cls) & arch_register_class_flag_manual_ra)
1263 stat_ev_ctx_push_str("bestraight_cls", cls->name);
1265 n_regs = cls->n_regs;
1266 ignore_regs = bitset_malloc(n_regs);
1267 be_put_ignore_regs(birg, cls, ignore_regs);
1271 /* verify schedule and register pressure */
1272 BE_TIMER_PUSH(t_verify);
1273 if (birg->main_env->options->vrfy_option == BE_CH_VRFY_WARN) {
1274 be_verify_schedule(birg);
1275 be_verify_register_pressure(birg, cls, irg);
1276 } else if (birg->main_env->options->vrfy_option == BE_CH_VRFY_ASSERT) {
1277 assert(be_verify_schedule(birg) && "Schedule verification failed");
1278 assert(be_verify_register_pressure(birg, cls, irg)
1279 && "Register pressure verification failed");
1281 BE_TIMER_POP(t_verify);
1283 BE_TIMER_PUSH(t_ra_color);
1284 be_straight_alloc_cls();
1285 BE_TIMER_POP(t_ra_color);
1287 bitset_free(ignore_regs);
1289 stat_ev_ctx_pop("bestraight_cls");
1292 BE_TIMER_PUSH(t_verify);
1293 if (birg->main_env->options->vrfy_option == BE_CH_VRFY_WARN) {
1294 be_verify_register_allocation(birg);
1295 } else if(birg->main_env->options->vrfy_option == BE_CH_VRFY_ASSERT) {
1296 assert(be_verify_register_allocation(birg)
1297 && "Register allocation invalid");
1299 BE_TIMER_POP(t_verify);
1301 obstack_free(&obst, NULL);
1305 * Initializes this module.
1307 void be_init_straight_alloc(void)
1309 static be_ra_t be_ra_straight = {
1313 FIRM_DBG_REGISTER(dbg, "firm.be.straightalloc");
1315 be_register_allocator("straight", &be_ra_straight);
1318 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_straight_alloc);