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 ILP based spilling & rematerialization
23 * @author Adam M. Szalkowski
49 #include "irnodeset.h"
51 #include "iredges_t.h"
54 #include "irbackedge_t.h"
55 #include "irprofile.h"
59 #include <lpp/lpp_net.h>
60 #include <lpp/lpp_cplex.h>
65 #include "besched_t.h"
66 #include "bessaconstr.h"
68 #include "beintlive_t.h"
72 #include "bespillremat.h"
74 #include "bepressurestat.h"
75 #include "bespilloptions.h"
76 #include "bechordal_t.h"
80 #include "lc_opts_enum.h"
82 #define DUMP_PROBLEM 1
84 #define DUMP_SOLUTION 4
86 #define DUMP_PRESSURE 16
88 #define KEEPALIVE_REMATS 1
89 #define KEEPALIVE_SPILLS 2
90 #define KEEPALIVE_RELOADS 4
92 #define VERIFY_MEMINTERF 1
93 #define VERIFY_DOMINANCE 2
96 #define REMATS_BRIGGS 1
97 #define REMATS_NOINVERSE 2
100 static unsigned opt_dump_flags = 0;
101 static int opt_log = 0;
102 static unsigned opt_keep_alive = 0;
103 static int opt_goodwin = 1;
104 static int opt_memcopies = 1;
105 static int opt_memoperands = 1;
106 static int opt_verify = VERIFY_MEMINTERF;
107 static unsigned opt_remats = REMATS_ALL;
108 static int opt_repair_schedule = 0;
109 static int opt_no_enlarge_liveness = 0;
110 static int opt_remat_while_live = 1;
111 static int opt_timeout = 300;
112 static double opt_cost_reload = 8.0;
113 static double opt_cost_memoperand = 7.0;
114 static double opt_cost_spill = 15.0;
115 static double opt_cost_remat = 1.0;
118 static const lc_opt_enum_mask_items_t dump_items[] = {
119 { "problem", DUMP_PROBLEM },
121 { "solution", DUMP_SOLUTION },
122 { "stats", DUMP_STATS },
123 { "pressure", DUMP_PRESSURE },
127 static lc_opt_enum_mask_var_t dump_var = {
128 &opt_dump_flags, dump_items
131 static const lc_opt_enum_mask_items_t keepalive_items[] = {
132 { "remats", KEEPALIVE_REMATS },
133 { "spills", KEEPALIVE_SPILLS },
134 { "reloads", KEEPALIVE_RELOADS },
138 static lc_opt_enum_mask_var_t keep_alive_var = {
139 &opt_keep_alive, keepalive_items
142 static const lc_opt_enum_mask_items_t remats_items[] = {
143 { "none", REMATS_NONE },
144 { "briggs", REMATS_BRIGGS },
145 { "noinverse", REMATS_NOINVERSE },
146 { "all", REMATS_ALL },
150 static lc_opt_enum_mask_var_t remats_var = {
151 &opt_remats, remats_items
154 static const lc_opt_table_entry_t options[] = {
155 LC_OPT_ENT_ENUM_MASK("keepalive", "keep alive inserted nodes", &keep_alive_var),
157 LC_OPT_ENT_BOOL ("goodwin", "activate goodwin reduction", &opt_goodwin),
158 LC_OPT_ENT_BOOL ("memcopies", "activate memcopy handling", &opt_memcopies),
159 LC_OPT_ENT_BOOL ("memoperands", "activate memoperands", &opt_memoperands),
160 LC_OPT_ENT_ENUM_INT ("remats", "type of remats to insert", &remats_var),
161 LC_OPT_ENT_BOOL ("repair_schedule", "repair the schedule by rematting once used nodes",&opt_repair_schedule),
162 LC_OPT_ENT_BOOL ("no_enlage_liveness", "do not enlarge liveness of operands of remats",&opt_no_enlarge_liveness),
163 LC_OPT_ENT_BOOL ("remat_while_live", "only remat where rematted value was live", &opt_remat_while_live),
165 LC_OPT_ENT_ENUM_MASK("dump", "dump problem, solution or statistical data", &dump_var),
166 LC_OPT_ENT_BOOL ("log", "activate the lpp log", &opt_log),
167 LC_OPT_ENT_INT ("timeout", "ILP solver timeout", &opt_timeout),
169 LC_OPT_ENT_DBL ("cost_reload", "cost of a reload", &opt_cost_reload),
170 LC_OPT_ENT_DBL ("cost_memoperand", "cost of a memory operand", &opt_cost_memoperand),
171 LC_OPT_ENT_DBL ("cost_spill", "cost of a spill instruction", &opt_cost_spill),
172 LC_OPT_ENT_DBL ("cost_remat", "cost of a rematerialization", &opt_cost_remat),
176 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
177 //#define SCHEDULE_PHIM /* insert phim nodes into schedule */
180 //#define SOLVE_LOCAL
181 #define LPP_SERVER "i44pc52"
182 #define LPP_SOLVER "cplex"
185 #define MAX_PATHS INT_MAX
188 typedef struct _spill_ilp_t {
189 const arch_register_class_t *cls;
194 struct obstack *obst;
196 pset *all_possible_remats;
199 set *values; /**< for collecting all definitions of values before running ssa-construction */
205 #ifndef SCHEDULE_PHIM
208 DEBUG_ONLY(firm_dbg_module_t * dbg);
211 typedef int ilp_var_t;
212 typedef int ilp_cst_t;
214 typedef struct _spill_bb_t {
219 typedef struct _remat_t {
220 const ir_node *op; /**< for copy_irn */
221 const ir_node *value; /**< the value which is being recomputed by this remat */
222 const ir_node *proj; /**< not NULL if the above op produces a tuple */
223 int cost; /**< cost of this remat */
224 int inverse; /**< nonzero if this is an inverse remat */
228 * Data to be attached to each IR node. For remats this contains the ilp_var
229 * for this remat and for normal ops this contains the ilp_vars for
230 * reloading each operand
232 typedef struct _op_t {
237 const remat_t *remat; /** the remat this op belongs to */
238 int pre; /** 1, if this is a pressure-increasing remat */
242 ir_node *op; /** the operation this live range belongs to */
251 typedef struct _defs_t {
252 const ir_node *value;
253 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
254 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
257 typedef struct _remat_info_t {
258 const ir_node *irn; /**< the irn to which these remats belong */
259 pset *remats; /**< possible remats for this value */
260 pset *remats_by_operand; /**< remats with this value as operand */
263 typedef struct _keyval_t {
268 typedef struct _spill_t {
277 typedef struct _memoperand_t {
278 ir_node *irn; /**< the irn */
279 unsigned int pos; /**< the position of the argument */
280 ilp_var_t ilp; /**< the ilp var for this memory operand */
284 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
286 return arch_irn_consider_in_reg_alloc(si->cls, irn);
291 cmp_remat(const void *a, const void *b)
293 const keyval_t *p = a;
294 const keyval_t *q = b;
295 const remat_t *r = p->val;
296 const remat_t *s = q->val;
300 return !(r == s || r->op == s->op);
304 cmp_remat(const void *a, const void *b)
306 const remat_t *r = a;
307 const remat_t *s = b;
309 return !(r == s || r->op == s->op);
313 cmp_spill(const void *a, const void *b, size_t size)
315 const spill_t *p = a;
316 const spill_t *q = b;
319 // return !(p->irn == q->irn && p->bb == q->bb);
320 return !(p->irn == q->irn);
324 cmp_memoperands(const void *a, const void *b, size_t size)
326 const memoperand_t *p = a;
327 const memoperand_t *q = b;
330 return !(p->irn == q->irn && p->pos == q->pos);
334 set_find_keyval(set * set, const void * key)
339 return set_find(set, &query, sizeof(query), HASH_PTR(key));
343 set_insert_keyval(set * set, void * key, void * val)
349 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
353 set_find_def(set * set, const ir_node * value)
358 return set_find(set, &query, sizeof(query), HASH_PTR(value));
362 set_insert_def(set * set, const ir_node * value)
369 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
372 static memoperand_t *
373 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
380 return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
383 static memoperand_t *
384 set_find_memoperand(set * set, const ir_node * irn, unsigned int pos)
388 query.irn = (ir_node*)irn;
390 return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
395 set_find_spill(set * set, const ir_node * value)
399 query.irn = (ir_node*)value;
400 return set_find(set, &query, sizeof(query), HASH_PTR(value));
403 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
404 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
405 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
406 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
407 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
410 cmp_remat_info(const void *a, const void *b, size_t size)
412 const remat_info_t *p = a;
413 const remat_info_t *q = b;
416 return !(p->irn == q->irn);
420 cmp_defs(const void *a, const void *b, size_t size)
426 return !(p->value == q->value);
430 cmp_keyval(const void *a, const void *b, size_t size)
432 const keyval_t *p = a;
433 const keyval_t *q = b;
436 return !(p->key == q->key);
440 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
443 if(ir_profile_has_data())
444 return ((double)ir_profile_get_block_execcount(get_block_const(irn))) + FUDGE;
446 #ifndef EXECFREQ_LOOPDEPH
447 return get_block_execfreq(si->birg->exec_freq, get_block_const(irn)) + FUDGE;
450 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
452 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
456 static double get_cost(const ir_node *irn)
458 if(be_is_Spill(irn)) {
459 return opt_cost_spill;
460 } else if(be_is_Reload(irn)){
461 return opt_cost_reload;
463 return arch_get_op_estimated_cost(irn);
468 * Checks, whether node and its operands have suitable reg classes
471 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
474 int remat = (arch_irn_get_flags(irn) & arch_irn_flags_rematerializable) != 0;
478 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
481 for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
482 ir_node *op = get_irn_n(irn, n);
483 remat &= has_reg_class(si, op) || arch_irn_get_flags(op) & arch_irn_flags_ignore || is_NoMem(op);
486 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
493 * Try to create a remat from @p op with destination value @p dest_value
495 static INLINE remat_t *
496 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
498 remat_t *remat = NULL;
500 // if(!mode_is_datab(get_irn_mode(dest_value)))
503 if(dest_value == op) {
504 const ir_node *proj = NULL;
506 if(is_Proj(dest_value)) {
507 op = get_Proj_pred(op);
511 if(!is_rematerializable(si, op))
514 remat = obstack_alloc(si->obst, sizeof(*remat));
516 remat->cost = (int)get_cost(op);
517 remat->value = dest_value;
521 arch_inverse_t inverse;
524 /* get the index of the operand we want to retrieve by the inverse op */
525 for (n = get_irn_arity(op)-1; n>=0; --n) {
526 ir_node *arg = get_irn_n(op, n);
528 if(arg == dest_value) break;
532 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", n, op));
534 /* else ask the backend to give an inverse op */
535 if(arch_get_inverse(op, n, &inverse, si->obst)) {
538 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
540 assert(inverse.n > 0 && "inverse op should have at least one node");
542 for(i=inverse.n-1; i>=0; --i) {
543 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
547 remat = obstack_alloc(si->obst, sizeof(*remat));
548 remat->op = inverse.nodes[0];
549 remat->cost = inverse.costs;
550 remat->value = dest_value;
551 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
554 // Matze: commented this out, this doesn't seem to be true if
555 // the inverse is a simple operation with only 1 result...
556 //assert(is_Proj(remat->proj));
558 assert(0 && "I can not handle remats with more than 2 nodes");
565 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
567 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
575 add_remat(const spill_ilp_t * si, const remat_t * remat)
577 remat_info_t *remat_info,
582 assert(remat->value);
584 query.irn = remat->value;
586 query.remats_by_operand = NULL;
587 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
589 if(remat_info->remats == NULL) {
590 remat_info->remats = new_pset(cmp_remat, 4096);
592 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
594 /* insert the remat into the remats_be_operand set of each argument of the remat op */
595 for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
596 ir_node *arg = get_irn_n(remat->op, n);
600 query.remats_by_operand = NULL;
601 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
603 if(remat_info->remats_by_operand == NULL) {
604 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
606 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
611 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
613 const ir_edge_t *edge = get_irn_out_edge_first(irn);
617 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
620 edge = get_irn_out_edge_next(irn, edge);
627 get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
633 irn = get_Proj_pred(irn);
635 for(n=get_irn_arity(irn)-1; n>=0; --n) {
636 const ir_node *arg = get_irn_n(irn, n);
638 if(has_reg_class(si, arg)) ++ret;
645 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
650 if( has_reg_class(si, op)
651 && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1)
652 && (opt_remats != REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0)
654 remat = get_remat_from_op(si, op, op);
656 add_remat(si, remat);
660 if(opt_remats == REMATS_ALL) {
661 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
663 for (n = get_irn_arity(op)-1; n>=0; --n) {
664 ir_node *arg = get_irn_n(op, n);
666 if(has_reg_class(si, arg)) {
667 /* try to get an inverse remat */
668 remat = get_remat_from_op(si, arg, op);
670 add_remat(si, remat);
678 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
681 ir_node *def_block = get_nodes_block(val);
688 /* if pos is at end of a basic block */
690 ret = (pos == def_block || block_dominates(def_block, pos));
691 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
695 /* else if this is a normal operation */
696 block = get_nodes_block(pos);
697 if(block == def_block) {
698 if(!sched_is_scheduled(val)) return 1;
700 ret = sched_comes_after(val, pos);
701 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
705 ret = block_dominates(def_block, block);
706 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
710 static INLINE ir_node *
711 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
713 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->birg->main_env->arch_env);
717 * Returns first non-Phi node of block @p bb
719 static INLINE ir_node *
720 sched_block_first_nonphi(const ir_node * bb)
722 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
726 sched_skip_proj_predicator(const ir_node * irn, void * data)
729 return (is_Proj(irn));
732 static INLINE ir_node *
733 sched_next_nonproj(const ir_node * irn, int forward)
735 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
739 * Returns next operation node (non-Proj) after @p irn
740 * or the basic block of this node
742 static INLINE ir_node *
743 sched_next_op(const ir_node * irn)
745 ir_node *next = sched_next(irn);
750 return sched_next_nonproj(next, 1);
754 * Returns previous operation node (non-Proj) before @p irn
755 * or the basic block of this node
757 static INLINE ir_node *
758 sched_prev_op(const ir_node * irn)
760 ir_node *prev = sched_prev(irn);
765 return sched_next_nonproj(prev, 0);
769 sched_put_after(ir_node * insert, ir_node * irn)
771 if(is_Block(insert)) {
772 insert = sched_block_first_nonphi(insert);
774 insert = sched_next_op(insert);
777 sched_add_before(insert, irn);
781 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
783 if(is_Block(insert)) {
784 insert = sched_block_last_noncf(si, insert);
786 insert = sched_next_nonproj(insert, 0);
787 insert = sched_prev(insert);
790 sched_add_after(insert, irn);
794 next_post_remat(const ir_node * irn)
800 next = sched_block_first_nonphi(irn);
802 next = sched_next_op(irn);
805 if(sched_is_end(next))
808 op = get_irn_link(next);
809 if(op->is_remat && !op->attr.remat.pre) {
818 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
824 ret = sched_block_last_noncf(si, irn);
825 ret = sched_next(ret);
826 ret = sched_prev_op(ret);
828 ret = sched_prev_op(irn);
831 if(sched_is_end(ret) || is_Phi(ret))
834 op = (op_t*)get_irn_link(ret);
835 if(op->is_remat && op->attr.remat.pre) {
843 * Tells you whether a @p remat can be placed before the irn @p pos
846 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
848 const ir_node *op = remat->op;
854 prev = sched_block_last_noncf(si, pos);
855 prev = sched_next_nonproj(prev, 0);
857 prev = sched_prev_op(pos);
859 /* do not remat if the rematted value is defined immediately before this op */
860 if(prev == remat->op) {
865 /* this should be just fine, the following OP will be using this value, right? */
867 /* only remat AFTER the real definition of a value (?) */
868 if(!value_is_defined_before(si, pos, remat->value)) {
869 // ir_fprintf(stderr, "error(not defined)");
874 for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
875 const ir_node *arg = get_irn_n(op, n);
877 if(opt_no_enlarge_liveness) {
878 if(has_reg_class(si, arg) && live) {
879 res &= pset_find_ptr((pset*)live, arg)?1:0;
881 res &= value_is_defined_before(si, pos, arg);
884 res &= value_is_defined_before(si, pos, arg);
892 * Tells you whether a @p remat can be placed after the irn @p pos
895 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
898 pos = sched_block_first_nonphi(pos);
900 pos = sched_next_op(pos);
903 /* only remat AFTER the real definition of a value (?) */
904 if(!value_is_defined_before(si, pos, remat->value)) {
908 return can_remat_before(si, remat, pos, live);
912 * Collect potetially rematerializable OPs
915 walker_remat_collector(ir_node * irn, void * data)
917 spill_ilp_t *si = data;
919 if(!is_Block(irn) && !is_Phi(irn)) {
920 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
921 get_remats_from_op(si, irn);
926 * Inserts a copy of @p irn before @p pos
929 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
934 bb = is_Block(pos)?pos:get_nodes_block(pos);
935 copy = exact_copy(irn);
937 set_phi_class(si->pc, copy, NULL);
938 set_nodes_block(copy, bb);
939 sched_put_before(si, pos, copy);
945 * Inserts a copy of @p irn after @p pos
948 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
953 bb = is_Block(pos)?pos:get_nodes_block(pos);
954 copy = exact_copy(irn);
956 set_phi_class(si->pc, copy, NULL);
957 set_nodes_block(copy, bb);
958 sched_put_after(pos, copy);
964 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
968 if(can_remat_after(si, remat, pos, live)) {
973 DBG((si->dbg, LEVEL_3, "\t >inserting remat2 %+F\n", remat->op));
975 copy = insert_copy_after(si, remat->op, pos);
977 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
978 op = obstack_alloc(si->obst, sizeof(*op));
980 op->attr.remat.remat = remat;
981 op->attr.remat.pre = 0;
982 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
984 set_irn_link(copy, op);
985 pset_insert_ptr(si->all_possible_remats, copy);
987 proj_copy = insert_copy_after(si, remat->proj, copy);
988 set_irn_n(proj_copy, 0, copy);
989 set_irn_link(proj_copy, op);
990 pset_insert_ptr(si->all_possible_remats, proj_copy);
1002 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
1006 if(can_remat_before(si, remat, pos, live)) {
1011 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
1013 copy = insert_copy_before(si, remat->op, pos);
1015 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
1016 op = obstack_alloc(si->obst, sizeof(*op));
1018 op->attr.remat.remat = remat;
1019 op->attr.remat.pre = 1;
1020 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
1022 set_irn_link(copy, op);
1023 pset_insert_ptr(si->all_possible_remats, copy);
1025 proj_copy = insert_copy_after(si, remat->proj, copy);
1026 set_irn_n(proj_copy, 0, copy);
1027 set_irn_link(proj_copy, op);
1028 pset_insert_ptr(si->all_possible_remats, proj_copy);
1040 get_block_n_succs(const ir_node *block) {
1041 const ir_edge_t *edge;
1043 assert(edges_activated(current_ir_graph));
1045 edge = get_block_succ_first(block);
1049 edge = get_block_succ_next(block, edge);
1050 return edge ? 2 : 1;
1054 is_start_block(const ir_node * bb)
1056 return get_irg_start_block(get_irn_irg(bb)) == bb;
1060 is_merge_edge(const ir_node * bb)
1062 if(is_start_block(bb))
1066 return get_block_n_succs(bb) == 1;
1072 is_diverge_edge(const ir_node * bb)
1074 if(is_start_block(bb))
1078 return get_Block_n_cfgpreds(bb) == 1;
1084 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1089 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1090 irn = be_lv_get_irn(si->lv, bb, i);
1092 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1093 pset_insert_ptr(live, irn);
1097 irn = sched_last(bb);
1099 /* all values eaten by control flow operations are also live until the end of the block */
1100 sched_foreach_reverse(bb, irn) {
1103 if(!sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) break;
1105 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1106 ir_node *arg = get_irn_n(irn,i);
1108 if(has_reg_class(si, arg)) {
1109 pset_insert_ptr(live, arg);
1114 * find values that are used by remats at end of block
1115 * and insert them into live set
1117 foreach_pre_remat(si, bb, irn) {
1120 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1121 ir_node *remat_arg = get_irn_n(irn, n);
1123 if(!has_reg_class(si, remat_arg)) continue;
1125 /* if value is becoming live through use by remat */
1126 if(!pset_find_ptr(live, remat_arg)) {
1127 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, irn));
1129 pset_insert_ptr(live, remat_arg);
1136 walker_regclass_copy_insertor(ir_node * irn, void * data)
1138 spill_ilp_t *si = data;
1140 if(is_Phi(irn) && has_reg_class(si, irn)) {
1143 for(n=get_irn_arity(irn)-1; n>=0; --n) {
1144 ir_node *phi_arg = get_irn_n(irn, n);
1145 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
1147 if(!has_reg_class(si, phi_arg)) {
1148 ir_node *copy = be_new_Copy(si->cls, si->birg->irg, bb, phi_arg);
1149 ir_node *pos = sched_block_last_noncf(si, bb);
1150 op_t *op = obstack_alloc(si->obst, sizeof(*op));
1152 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
1153 sched_add_after(pos, copy);
1154 set_irn_n(irn, n, copy);
1157 op->attr.live_range.args.reloads = NULL;
1158 op->attr.live_range.ilp = ILP_UNDEF;
1159 set_irn_link(copy, op);
1166 * Insert (so far unused) remats into the irg to
1167 * recompute the potential liveness of all values
1170 walker_remat_insertor(ir_node * bb, void * data)
1172 spill_ilp_t *si = data;
1179 /* skip start block, no remats to do there */
1180 if(is_start_block(bb)) return;
1182 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1184 live = pset_new_ptr_default();
1185 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1186 ir_node *value = be_lv_get_irn(si->lv, bb, i);
1188 /* add remats at end of block */
1189 if (has_reg_class(si, value)) {
1190 pset_insert_ptr(live, value);
1194 irn = sched_last(bb);
1195 while(!sched_is_end(irn)) {
1201 next = sched_prev(irn);
1203 /* delete defined value from live set */
1204 if(has_reg_class(si, irn)) {
1205 pset_remove_ptr(live, irn);
1208 if(is_Phi(irn) || is_Proj(irn)) {
1213 args = pset_new_ptr_default();
1214 used = pset_new_ptr_default();
1216 /* collect arguments of op and set args of op already live in epilog */
1217 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1218 ir_node *arg = get_irn_n(irn, n);
1220 pset_insert_ptr(args, arg);
1221 if(has_reg_class(si, arg)) {
1222 pset_insert_ptr(live, arg);
1223 pset_insert_ptr(used, arg);
1227 /* insert all possible remats before irn */
1228 pset_foreach(args, arg) {
1229 remat_info_t *remat_info,
1232 /* continue if the operand has the wrong reg class */
1233 if(!has_reg_class(si, arg))
1237 query.remats = NULL;
1238 query.remats_by_operand = NULL;
1239 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1245 if(remat_info->remats) {
1246 pset_foreach(remat_info->remats, remat) {
1247 ir_node *remat_irn = NULL;
1249 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
1250 remat_irn = insert_remat_before(si, remat, irn, live);
1253 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1254 ir_node *remat_arg = get_irn_n(remat_irn, n);
1256 /* collect args of remats which are not args of op */
1257 if(has_reg_class(si, remat_arg) && !pset_find_ptr(args, remat_arg)) {
1258 pset_insert_ptr(used, remat_arg);
1266 /* do not place post remats after jumps */
1267 if(sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) {
1273 /* insert all possible remats after irn */
1274 post_remats = pset_new_ptr_default();
1275 pset_foreach(used, arg) {
1276 remat_info_t *remat_info,
1279 /* continue if the operand has the wrong reg class */
1280 if(!has_reg_class(si, arg))
1284 query.remats = NULL;
1285 query.remats_by_operand = NULL;
1286 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1292 if(remat_info->remats_by_operand) {
1293 pset_foreach(remat_info->remats_by_operand, remat) {
1294 /* do not insert remats producing the same value as one of the operands */
1295 if(!pset_find_ptr(args, remat->value)) {
1296 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
1298 /* only remat values that can be used by real ops */
1299 if(!opt_remat_while_live || pset_find_ptr(live, remat->value)) {
1300 pset_insert_ptr(post_remats, remat);
1306 pset_foreach(post_remats, remat) {
1307 insert_remat_after(si, remat, irn, live);
1309 del_pset(post_remats);
1316 /* add remats at end if successor has multiple predecessors */
1317 if(is_merge_edge(bb)) {
1318 pset *live_out = pset_new_ptr_default();
1321 get_live_end(si, bb, live_out);
1323 /* add remats at end of block */
1324 pset_foreach(live_out, value) {
1325 remat_info_t *remat_info,
1329 query.remats = NULL;
1330 query.remats_by_operand = NULL;
1331 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1333 if(remat_info && remat_info->remats) {
1334 pset_foreach(remat_info->remats, remat) {
1335 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1337 insert_remat_before(si, remat, bb, live_out);
1344 if(is_diverge_edge(bb)) {
1345 pset *live_in = pset_new_ptr_default();
1348 be_lv_foreach(si->lv, bb, be_lv_state_in, i) {
1349 value = be_lv_get_irn(si->lv, bb, i);
1351 if(has_reg_class(si, value)) {
1352 pset_insert_ptr(live_in, value);
1355 /* add phis to live_in */
1356 sched_foreach(bb, value) {
1357 if(!is_Phi(value)) break;
1359 if(has_reg_class(si, value)) {
1360 pset_insert_ptr(live_in, value);
1364 /* add remat2s at beginning of block */
1365 post_remats = pset_new_ptr_default();
1366 pset_foreach(live_in, value) {
1367 remat_info_t *remat_info,
1371 query.remats = NULL;
1372 query.remats_by_operand = NULL;
1373 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1375 if(remat_info && remat_info->remats_by_operand) {
1376 pset_foreach(remat_info->remats_by_operand, remat) {
1377 DBG((si->dbg, LEVEL_4, "\t considering remat2 %+F at beginning of block %+F\n", remat->op, bb));
1379 /* put the remat here if all its args are available and result is still live */
1380 if(!opt_remat_while_live || pset_find_ptr(live_in, remat->value)) {
1381 pset_insert_ptr(post_remats, remat);
1386 pset_foreach(post_remats, remat) {
1387 insert_remat_after(si, remat, bb, live_in);
1389 del_pset(post_remats);
1395 can_be_copied(const ir_node * bb, const ir_node * irn)
1397 const ir_edge_t *edge = get_block_succ_first(bb);
1398 const ir_node *next_bb = edge->src;
1399 int pos = edge->pos;
1402 assert(is_merge_edge(bb));
1404 sched_foreach(next_bb, phi) {
1405 const ir_node *phi_arg;
1407 if(!is_Phi(phi)) break;
1409 phi_arg = get_irn_n(phi, pos);
1411 if(phi_arg == irn) {
1419 * Initialize additional node info
1422 luke_initializer(ir_node * bb, void * data)
1424 spill_ilp_t *si = (spill_ilp_t*)data;
1425 spill_bb_t *spill_bb;
1428 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1429 set_irn_link(bb, spill_bb);
1431 sched_foreach(bb, irn) {
1434 op = obstack_alloc(si->obst, sizeof(*op));
1436 op->attr.live_range.ilp = ILP_UNDEF;
1439 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1440 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1442 } else if(!is_Proj(irn)) {
1443 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1444 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1446 op->attr.live_range.args.reloads = NULL;
1448 set_irn_link(irn, op);
1454 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1457 luke_endwalker(ir_node * bb, void * data)
1459 spill_ilp_t *si = (spill_ilp_t*)data;
1465 spill_bb_t *spill_bb = get_irn_link(bb);
1468 live = pset_new_ptr_default();
1469 use_end = pset_new_ptr_default();
1471 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1472 irn = be_lv_get_irn(si->lv, bb, i);
1473 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1474 pset_insert_ptr(live, irn);
1478 * find values that are used by remats at end of block
1479 * and insert them into live set
1481 foreach_pre_remat(si, bb, irn) {
1484 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1485 ir_node *remat_arg = get_irn_n(irn, n);
1487 if(has_reg_class(si, remat_arg)) {
1488 pset_insert_ptr(live, remat_arg);
1493 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1494 /* their reg_out must always be set */
1495 sched_foreach_reverse(bb, irn) {
1498 if(!sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) break;
1500 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1501 ir_node *irn_arg = get_irn_n(irn, n);
1503 if(has_reg_class(si, irn_arg)) {
1504 pset_insert_ptr(use_end, irn_arg);
1509 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1510 //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1511 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1513 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1515 /* if this is a merge edge we can reload at the end of this block */
1516 if(is_merge_edge(bb)) {
1517 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1518 } else if(pset_count(use_end)){
1519 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1521 spill_bb->reloads = NULL;
1524 pset_foreach(live,irn) {
1528 int default_spilled;
1531 /* handle values used by control flow nodes later separately */
1532 if(pset_find_ptr(use_end, irn)) continue;
1535 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1537 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1539 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1540 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1541 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1543 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1544 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1546 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1547 /* by default spill value right after definition */
1548 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1549 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1551 if(is_merge_edge(bb)) {
1555 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1556 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), can_be_copied(bb, irn));
1557 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1559 /* reload <= mem_out */
1560 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1561 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1562 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1565 spill->reg_in = ILP_UNDEF;
1566 spill->mem_in = ILP_UNDEF;
1569 pset_foreach(use_end,irn) {
1573 ilp_cst_t end_use_req,
1576 int default_spilled;
1579 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1581 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1583 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1584 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1586 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1587 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1589 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1590 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1591 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1593 /* reload for use be control flow op */
1594 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1595 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1596 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1598 /* reload <= mem_out */
1599 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1600 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1601 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1603 spill->reg_in = ILP_UNDEF;
1604 spill->mem_in = ILP_UNDEF;
1606 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1607 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1608 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1617 * Find a remat of value @p value in the epilog of @p pos
1620 find_post_remat(const ir_node * value, const ir_node * pos)
1622 while((pos = next_post_remat(pos)) != NULL) {
1625 op = get_irn_link(pos);
1626 assert(op->is_remat && !op->attr.remat.pre);
1628 if(op->attr.remat.remat->value == value)
1629 return (ir_node*)pos;
1632 const ir_edge_t *edge;
1633 foreach_out_edge(pos, edge) {
1634 ir_node *proj = get_edge_src_irn(edge);
1635 assert(is_Proj(proj));
1646 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1648 spill_bb_t *spill_bb = get_irn_link(bb);
1652 int default_spilled;
1655 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1657 double spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1659 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1661 spill->reg_out = ILP_UNDEF;
1662 spill->reg_in = ILP_UNDEF;
1663 spill->mem_in = ILP_UNDEF;
1665 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1666 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1668 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1669 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1670 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1677 * Inserts ILP-constraints and variables for memory copying before the given position
1680 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1682 const ir_node *succ;
1683 const ir_edge_t *edge;
1684 spill_bb_t *spill_bb = get_irn_link(block);
1693 assert(edges_activated(current_ir_graph));
1695 edge = get_block_succ_first(block);
1701 edge = get_block_succ_next(block, edge);
1702 /* next block can only contain phis, if this is a merge edge */
1705 ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1706 copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1708 ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1709 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1711 pset_foreach(live, tmp) {
1714 op_t *op = get_irn_link(irn);
1715 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1717 spill = set_find_spill(spill_bb->ilp, tmp);
1720 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1722 lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1724 sched_foreach(succ, phi) {
1725 const ir_node *to_copy;
1727 spill_t *to_copy_spill;
1728 op_t *phi_op = get_irn_link(phi);
1729 ilp_var_t reload = ILP_UNDEF;
1732 if(!is_Phi(phi)) break;
1733 if(!has_reg_class(si, phi)) continue;
1735 to_copy = get_irn_n(phi, pos);
1736 to_copy_op = get_irn_link(to_copy);
1738 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1739 assert(to_copy_spill);
1741 if(spill_bb->reloads) {
1742 keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1745 reload = PTR_TO_INT(keyval->val);
1749 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1750 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1752 /* copy - reg_out - reload - remat - live_range <= 0 */
1753 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1754 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1755 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1756 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1757 foreach_pre_remat(si, block, tmp) {
1758 op_t *remat_op = get_irn_link(tmp);
1759 if(remat_op->attr.remat.remat->value == to_copy) {
1760 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1764 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1765 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1767 /* copy - reg_out - copyreg <= 0 */
1768 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1769 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1770 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1776 * Walk all irg blocks and emit this ILP
1779 luke_blockwalker(ir_node * bb, void * data)
1781 spill_ilp_t *si = (spill_ilp_t*)data;
1786 spill_bb_t *spill_bb = get_irn_link(bb);
1789 pset *defs = pset_new_ptr_default();
1791 live = pset_new_ptr_default();
1793 /****************************************
1794 * B A S I C B L O C K E N D
1795 ***************************************/
1798 /* init live values at end of block */
1799 get_live_end(si, bb, live);
1801 pset_foreach(live, irn) {
1803 ilp_var_t reload = ILP_UNDEF;
1805 spill = set_find_spill(spill_bb->ilp, irn);
1808 if(spill_bb->reloads) {
1809 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1812 reload = PTR_TO_INT(keyval->val);
1816 op = get_irn_link(irn);
1817 assert(!op->is_remat);
1819 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1820 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1821 op->attr.live_range.op = bb;
1823 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1824 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1826 /* reg_out - reload - remat - live_range <= 0 */
1827 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1828 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1829 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1830 foreach_pre_remat(si, bb, tmp) {
1831 op_t *remat_op = get_irn_link(tmp);
1832 if(remat_op->attr.remat.remat->value == irn) {
1833 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1836 ir_snprintf(buf, sizeof(buf), "reg_out2_%N_%N", bb, irn);
1837 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
1839 /* value may only die at bb end if it is used for a mem copy */
1840 /* reg_out + \sum copy - reload - remat - live_range >= 0 */
1841 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1842 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1843 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1844 foreach_pre_remat(si, bb, tmp) {
1845 op_t *remat_op = get_irn_link(tmp);
1846 if(remat_op->attr.remat.remat->value == irn) {
1847 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1850 if(is_merge_edge(bb)) {
1851 const ir_edge_t *edge = get_block_succ_first(bb);
1852 const ir_node *next_bb = edge->src;
1853 int pos = edge->pos;
1856 sched_foreach(next_bb, phi) {
1857 const ir_node *phi_arg;
1859 if(!is_Phi(phi)) break;
1861 phi_arg = get_irn_n(phi, pos);
1863 if(phi_arg == irn) {
1864 op_t *phi_op = get_irn_link(phi);
1865 ilp_var_t copy = phi_op->attr.live_range.args.copies[pos];
1867 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
1874 insert_mem_copy_position(si, live, bb);
1877 * assure the remat args are available
1879 foreach_pre_remat(si, bb, tmp) {
1880 op_t *remat_op = get_irn_link(tmp);
1883 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1884 ir_node *remat_arg = get_irn_n(tmp, n);
1885 op_t *arg_op = get_irn_link(remat_arg);
1887 if(!has_reg_class(si, remat_arg)) continue;
1889 spill = set_find_spill(spill_bb->ilp, remat_arg);
1892 /* arguments of remats have to be live until the very end of the block
1893 * remat = reg_out(remat_arg) and (reload(remat_arg) or live_range(remat_arg)),
1894 * no remats, they could be in wrong order
1897 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1898 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1900 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 3.0);
1901 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, -2.0);
1902 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1904 /* use reload placed for this argument */
1905 if(spill_bb->reloads) {
1906 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1909 ilp_var_t reload = PTR_TO_INT(keyval->val);
1911 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1916 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1921 /**************************************
1922 * B A S I C B L O C K B O D Y
1923 **************************************/
1925 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1931 ilp_cst_t check_pre,
1937 ilp_cst_t one_memoperand = -1;
1939 /* iterate only until first phi */
1943 op = get_irn_link(irn);
1945 if(op->is_remat) continue;
1947 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1949 /* collect defined values */
1950 if(has_reg_class(si, irn)) {
1951 pset_insert_ptr(defs, irn);
1955 if(is_Proj(irn)) continue;
1958 * init set of irn's arguments
1959 * and all possibly used values around this op
1960 * and values defined by post remats
1962 args = new_set(cmp_keyval, get_irn_arity(irn));
1963 used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1964 remat_defs = pset_new_ptr(pset_count(live));
1966 if(!is_start_block(bb) || !be_is_Barrier(irn)) {
1967 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1968 ir_node *irn_arg = get_irn_n(irn, n);
1969 if(has_reg_class(si, irn_arg)) {
1970 set_insert_keyval(args, irn_arg, (void*)n);
1971 pset_insert_ptr(used, irn_arg);
1974 foreach_post_remat(irn, tmp) {
1975 op_t *remat_op = get_irn_link(tmp);
1977 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1979 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1980 ir_node *remat_arg = get_irn_n(tmp, n);
1981 if(has_reg_class(si, remat_arg)) {
1982 pset_insert_ptr(used, remat_arg);
1986 foreach_pre_remat(si, irn, tmp) {
1987 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1988 ir_node *remat_arg = get_irn_n(tmp, n);
1989 if(has_reg_class(si, remat_arg)) {
1990 pset_insert_ptr(used, remat_arg);
1996 /**********************************
1997 * I N E P I L O G O F irn
1998 **********************************/
2000 /* ensure each dying value is used by only one post remat */
2001 pset_foreach(used, tmp) {
2002 ir_node *value = tmp;
2003 op_t *value_op = get_irn_link(value);
2008 foreach_post_remat(irn, remat) {
2009 op_t *remat_op = get_irn_link(remat);
2011 for(n=get_irn_arity(remat)-1; n>=0; --n) {
2012 ir_node *remat_arg = get_irn_n(remat, n);
2014 /* if value is used by this remat add it to constraint */
2015 if(remat_arg == value) {
2017 /* sum remat2s <= 1 + n_remats*live_range */
2018 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
2019 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2023 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2029 if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
2030 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
2034 /* ensure at least one value dies at post remat */
2035 foreach_post_remat(irn, tmp) {
2036 op_t *remat_op = get_irn_link(tmp);
2037 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
2040 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2041 remat_arg = get_irn_n(tmp, n);
2043 if(has_reg_class(si, remat_arg)) {
2045 /* does arg always die at this op? */
2046 if(!pset_find_ptr(live, remat_arg))
2047 goto skip_one_must_die;
2049 pset_insert_ptr(remat_args, remat_arg);
2053 /* remat + \sum live_range(remat_arg) <= |args| */
2054 ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
2055 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2056 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2058 pset_foreach(remat_args, remat_arg) {
2059 op_t *arg_op = get_irn_link(remat_arg);
2061 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2065 del_pset(remat_args);
2068 /* new live ranges for values from L\U defined by post remats */
2069 pset_foreach(live, tmp) {
2070 ir_node *value = tmp;
2071 op_t *value_op = get_irn_link(value);
2073 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
2074 ilp_var_t prev_lr = ILP_UNDEF;
2077 if(pset_find_ptr(remat_defs, value)) {
2079 /* next_live_range <= prev_live_range + sum remat2s */
2080 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
2081 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2083 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
2084 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2086 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
2087 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2089 foreach_post_remat(irn, remat) {
2090 op_t *remat_op = get_irn_link(remat);
2092 /* if value is being rematerialized by this remat */
2093 if(value == remat_op->attr.remat.remat->value) {
2094 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2098 value_op->attr.live_range.ilp = prev_lr;
2099 value_op->attr.live_range.op = irn;
2104 /* requirements for post remats and start live ranges from L/U' for values dying here */
2105 foreach_post_remat(irn, tmp) {
2106 op_t *remat_op = get_irn_link(tmp);
2109 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2110 ir_node *remat_arg = get_irn_n(tmp, n);
2111 op_t *arg_op = get_irn_link(remat_arg);
2113 if(!has_reg_class(si, remat_arg)) continue;
2115 /* only for values in L\U (TODO and D?), the others are handled with post_use */
2116 if(!pset_find_ptr(used, remat_arg)) {
2117 /* remat <= live_range(remat_arg) */
2118 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
2119 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2121 /* if value is becoming live through use by remat2 */
2122 if(!pset_find_ptr(live, remat_arg)) {
2125 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
2126 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2128 arg_op->attr.live_range.ilp = lr;
2129 arg_op->attr.live_range.op = irn;
2131 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
2133 pset_insert_ptr(live, remat_arg);
2134 add_to_spill_bb(si, bb, remat_arg);
2137 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2138 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2143 d = pset_count(defs);
2144 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
2146 /* count how many regs irn needs for arguments */
2147 u = set_count(args);
2150 /* check the register pressure in the epilog */
2151 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
2152 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
2153 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
2155 /* add L\U' to check_post */
2156 pset_foreach(live, tmp) {
2157 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
2158 /* if a live value is not used by irn */
2159 tmp_op = get_irn_link(tmp);
2160 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
2164 /***********************************************************
2165 * I T E R A T I O N O V E R U S E S F O R E P I L O G
2166 **********************************************************/
2169 pset_foreach(used, tmp) {
2175 op_t *arg_op = get_irn_link(arg);
2178 spill = add_to_spill_bb(si, bb, arg);
2180 /* new live range for each used value */
2181 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2182 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2184 /* the epilog stuff - including post_use, check_post, check_post_remat */
2185 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2186 post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2188 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2190 /* arg is live throughout epilog if the next live_range is in a register */
2191 if(pset_find_ptr(live, arg)) {
2192 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
2194 /* post_use >= next_lr + remat */
2195 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2196 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2197 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2198 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2201 /* forall post remat which use arg add a similar cst */
2202 foreach_post_remat(irn, remat) {
2205 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2206 ir_node *remat_arg = get_irn_n(remat, n);
2207 op_t *remat_op = get_irn_link(remat);
2209 if(remat_arg == arg) {
2210 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
2212 /* post_use >= remat */
2213 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2214 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2215 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2216 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2221 /* if value is not an arg of op and not possibly defined by post remat
2222 * then it may only die and not become live
2224 if(!set_find_keyval(args, arg)) {
2225 /* post_use <= prev_lr */
2226 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2227 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2228 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2229 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2231 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2232 /* next_lr <= prev_lr */
2233 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2234 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2235 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2236 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2240 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2241 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2242 if (get_irn_n(irn, n) == arg &&
2243 arch_possible_memory_operand(irn, n)) {
2244 ilp_var_t memoperand;
2246 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2247 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0);
2248 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2250 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2251 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2253 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2254 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2259 /* new live range begins for each used value */
2260 arg_op->attr.live_range.ilp = prev_lr;
2261 arg_op->attr.live_range.op = irn;
2263 pset_insert_ptr(live, arg);
2266 /* just to be sure */
2267 check_post = ILP_UNDEF;
2269 /* allow original defintions to be removed */
2270 if(opt_repair_schedule) {
2271 pset_foreach(defs, tmp) {
2272 op_t *tmp_op = get_irn_link(tmp);
2273 spill_t *spill = set_find_spill(spill_bb->ilp, tmp);
2278 ir_snprintf(buf, sizeof(buf), "delete_%N", tmp);
2279 delete = lpp_add_var_default(si->lpp, buf, lpp_binary, -1.0 * get_cost(irn) * execution_frequency(si, bb), 0.0);
2281 /* op may not be killed if its first live_range is 1 */
2282 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2283 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2284 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2285 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2287 /* op may not be killed if it is spilled after the definition */
2288 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2289 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2290 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2291 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2296 ir_snprintf(buf, sizeof(buf), "keep_%N", tmp);
2297 keep = lpp_add_var_default(si->lpp, buf, lpp_binary, get_cost(irn) * execution_frequency(si, bb), 1.0);
2299 /* op may not be killed if its first live_range is 1 */
2300 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2301 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2302 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2303 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, -1.0);
2305 /* op may not be killed if it is spilled after the definition */
2306 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2307 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2308 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2309 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2314 pset_foreach(defs, tmp) {
2315 op_t *tmp_op = get_irn_link(tmp);
2316 spill_t *spill = set_find_spill(spill_bb->ilp, tmp);
2319 /* live_range or spill should be 1
2320 TODO: lr should be live until first use */
2321 ir_snprintf(buf, sizeof(buf), "nokillorig_%N", tmp);
2322 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 1.0);
2323 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2324 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2334 /* check the register pressure in the prolog */
2335 /* sum_{L\U} lr <= k - |U| */
2336 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2337 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2339 /* for the prolog remove defined values from the live set */
2340 pset_foreach(defs, tmp) {
2341 pset_remove_ptr(live, tmp);
2344 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2345 ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2346 one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2349 /***********************************************************
2350 * I T E R A T I O N O V E R A R G S F O R P R O L O G
2351 **********************************************************/
2354 set_foreach(args, keyval) {
2356 const ir_node *arg = keyval->key;
2357 int i = PTR_TO_INT(keyval->val);
2358 op_t *arg_op = get_irn_link(arg);
2359 ilp_cst_t requirements;
2362 spill = set_find_spill(spill_bb->ilp, arg);
2365 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2366 op->attr.live_range.args.reloads[i] = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
2368 /* reload <= mem_out */
2369 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2370 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2371 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2372 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2374 /* requirement: arg must be in register for use */
2375 /* reload + remat + live_range == 1 */
2376 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2377 requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2379 lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2380 lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2381 foreach_pre_remat(si, irn, tmp) {
2382 op_t *remat_op = get_irn_link(tmp);
2383 if(remat_op->attr.remat.remat->value == arg) {
2384 lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2388 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2390 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2391 if(get_irn_n(irn, n) == arg) {
2395 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2396 if (get_irn_n(irn, n) == arg &&
2397 arch_possible_memory_operand(irn, n)) {
2398 memoperand_t *memoperand;
2399 memoperand = set_find_memoperand(si->memoperands, irn, n);
2401 /* memoperand <= mem_out */
2402 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2403 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2404 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2405 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2407 /* the memoperand is only sufficient if it is used once by the op */
2408 if(n_memoperands == 1)
2409 lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2411 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2413 /* we have one more free register if we use a memory operand */
2414 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2420 /* iterate over L\U */
2421 pset_foreach(live, tmp) {
2422 if(!set_find_keyval(args, tmp)) {
2423 /* if a live value is not used by irn */
2424 tmp_op = get_irn_link(tmp);
2425 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2429 /* requirements for remats */
2430 foreach_pre_remat(si, irn, tmp) {
2431 op_t *remat_op = get_irn_link(tmp);
2434 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2435 ir_node *remat_arg = get_irn_n(tmp, n);
2436 op_t *arg_op = get_irn_link(remat_arg);
2438 if(!has_reg_class(si, remat_arg)) continue;
2440 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2441 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2442 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2444 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2445 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2447 /* if remat arg is also used by current op then we can use reload placed for this argument */
2448 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2449 int index = (int)keyval->val;
2451 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2459 /*************************
2460 * D O N E W I T H O P
2461 *************************/
2463 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2465 pset_foreach(live, tmp) {
2466 assert(has_reg_class(si, tmp));
2470 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2471 ir_node *arg = get_irn_n(irn, n);
2473 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2477 del_pset(remat_defs);
2481 defs = pset_new_ptr_default();
2483 /* skip everything above barrier in start block */
2484 if(is_start_block(bb) && be_is_Barrier(irn)) {
2485 assert(pset_count(live) == 0);
2494 /***************************************
2495 * B E G I N N I N G O F B L O C K
2496 ***************************************/
2499 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2500 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2502 pset_foreach(live, irn) {
2503 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2506 /* construct mem_outs for all values */
2507 set_foreach(spill_bb->ilp, spill) {
2508 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2509 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2511 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2512 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2514 if(pset_find_ptr(live, spill->irn)) {
2515 int default_spilled;
2516 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2518 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2519 default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2520 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2521 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2523 if(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2525 op_t *op = get_irn_link(spill->irn);
2527 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2528 const ir_node *arg = get_irn_n(spill->irn, n);
2534 /* argument already done? */
2535 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2537 /* get sum of execution frequencies of blocks with the same phi argument */
2538 for(m=n; m>=0; --m) {
2539 const ir_node *arg2 = get_irn_n(spill->irn, m);
2542 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2546 /* copies are not for free */
2547 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2548 var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
2550 for(m=n; m>=0; --m) {
2551 const ir_node *arg2 = get_irn_n(spill->irn, m);
2554 op->attr.live_range.args.copies[m] = var;
2559 /* copy <= mem_in */
2560 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2561 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2562 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2563 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2570 foreach_post_remat(bb, tmp) {
2572 op_t *remat_op = get_irn_link(tmp);
2573 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
2576 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2577 remat_arg = get_irn_n(tmp, n);
2579 if(has_reg_class(si, remat_arg)) {
2580 pset_insert_ptr(remat_args, remat_arg);
2584 /* remat + \sum live_range(remat_arg) <= |args| */
2585 ir_snprintf(buf, sizeof(buf), "one_must_die_%N", tmp);
2586 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2587 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2589 pset_foreach(remat_args, remat_arg) {
2590 if(pset_find_ptr(live, remat_arg)) {
2591 op_t *remat_arg_op = get_irn_link(remat_arg);
2592 lpp_set_factor_fast(si->lpp, cst, remat_arg_op->attr.live_range.ilp, 1.0);
2595 del_pset(remat_args);
2598 foreach_post_remat(bb, tmp) {
2601 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2602 ir_node *remat_arg = get_irn_n(tmp, n);
2604 /* if value is becoming live through use by remat2 */
2605 if(has_reg_class(si, remat_arg) && !pset_find_ptr(live, remat_arg)) {
2606 op_t *remat_arg_op = get_irn_link(remat_arg);
2609 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 at bb start %+F\n", remat_arg, tmp));
2611 pset_insert_ptr(live, remat_arg);
2612 spill = add_to_spill_bb(si, bb, remat_arg);
2613 remat_arg_op->attr.live_range.ilp = ILP_UNDEF;
2615 /* we need reg_in and mem_in for this value; they will be referenced later */
2616 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", remat_arg, bb);
2617 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2618 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", remat_arg, bb);
2619 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2622 /* optimization: all memory stuff should be 0, for we do not want to insert reloads for remats */
2623 ir_snprintf(buf, sizeof(buf), "nomem_%N_%N", remat_arg, bb);
2624 nomem = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0.0);
2625 lpp_set_factor_fast(si->lpp, nomem, spill->spill, 1.0);
2630 /* L\U is empty at bb start */
2631 /* arg is live throughout epilog if it is reg_in into this block */
2633 /* check the register pressure at the beginning of the block
2636 /* reg_in entspricht post_use */
2638 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2639 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2641 pset_foreach(live, irn) {
2644 spill = set_find_spill(spill_bb->ilp, irn);
2647 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2648 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2650 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2652 /* spill + mem_in <= 1 */
2653 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2654 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2656 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2657 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2659 } /* post_remats are NOT included in register pressure check because
2660 they do not increase regpressure */
2662 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2663 pset_foreach(live, irn) {
2667 spill = set_find_spill(spill_bb->ilp, irn);
2668 assert(spill && spill->irn == irn);
2670 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2671 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2674 ir_node *phi_arg = get_Phi_pred(irn, n);
2675 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2676 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2678 op_t *op = get_irn_link(irn);
2680 /* although the phi is in the right regclass one or more of
2681 * its arguments can be in a different one or at least to
2684 if(has_reg_class(si, phi_arg)) {
2685 /* mem_in < mem_out_arg + copy */
2686 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2687 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2689 /* reg_in < reg_out_arg */
2690 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2691 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2693 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2694 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2696 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2699 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2701 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2703 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2707 /* else assure the value arrives on all paths in the same resource */
2709 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2712 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2713 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2716 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2717 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2718 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2719 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2721 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2722 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2724 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2727 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2728 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2733 foreach_post_remat(bb, tmp) {
2736 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2737 ir_node *remat_arg = get_irn_n(tmp, n);
2738 op_t *remat_op = get_irn_link(tmp);
2740 if(!has_reg_class(si, remat_arg)) continue;
2742 spill = set_find_spill(spill_bb->ilp, remat_arg);
2745 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2746 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2747 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2748 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2752 pset_foreach(live, irn) {
2753 const op_t *op = get_irn_link(irn);
2754 const ir_node *remat;
2759 foreach_post_remat(bb, remat) {
2762 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2763 const ir_node *arg = get_irn_n(remat, n);
2766 const op_t *remat_op = get_irn_link(remat);
2768 if(cst == ILP_UNDEF) {
2769 /* sum remat2s <= 1 + n_remats*live_range */
2770 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", irn, bb);
2771 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2773 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2779 if(cst != ILP_UNDEF && op->attr.live_range.ilp != ILP_UNDEF) {
2780 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -n_remats);
2784 /* first live ranges from reg_ins */
2785 pset_foreach(live, irn) {
2786 op_t *op = get_irn_link(irn);
2788 if(op->attr.live_range.ilp != ILP_UNDEF) {
2790 spill = set_find_spill(spill_bb->ilp, irn);
2791 assert(spill && spill->irn == irn);
2793 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2794 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2795 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2796 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2798 foreach_post_remat(bb, tmp) {
2799 op_t *remat_op = get_irn_link(tmp);
2801 if(remat_op->attr.remat.remat->value == irn) {
2802 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2808 /* walk forward now and compute constraints for placing spills */
2809 /* this must only be done for values that are not defined in this block */
2810 pset_foreach(live, irn) {
2812 * if value is defined in this block we can anways place the spill directly after the def
2813 * -> no constraint necessary
2815 if(!is_Phi(irn) && get_nodes_block(irn) == bb) {
2820 spill = set_find_spill(spill_bb->ilp, irn);
2823 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2824 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2826 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2827 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2830 sched_foreach_op(bb, tmp) {
2831 op_t *op = get_irn_link(tmp);
2833 if(is_Phi(tmp)) continue;
2834 assert(!is_Proj(tmp));
2837 const ir_node *value = op->attr.remat.remat->value;
2840 /* only collect remats up to the first real use of a value */
2841 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2846 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2847 ir_node *arg = get_irn_n(tmp, n);
2850 /* if a value is used stop collecting remats */
2863 typedef struct _irnlist_t {
2864 struct list_head list;
2868 typedef struct _interference_t {
2869 struct list_head blocklist;
2875 cmp_interference(const void *a, const void *b, size_t size)
2877 const interference_t *p = a;
2878 const interference_t *q = b;
2881 return !(p->a == q->a && p->b == q->b);
2884 static interference_t *
2885 set_find_interference(set * set, ir_node * a, ir_node * b)
2887 interference_t query;
2889 query.a = (a>b)?a:b;
2890 query.b = (a>b)?b:a;
2892 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2895 static interference_t *
2896 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2898 interference_t query,
2900 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2904 result = set_find_interference(set, a, b);
2907 list_add(&list->list, &result->blocklist);
2911 query.a = (a>b)?a:b;
2912 query.b = (a>b)?b:a;
2914 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2916 INIT_LIST_HEAD(&result->blocklist);
2917 list_add(&list->list, &result->blocklist);
2923 int values_interfere_in_block(const spill_ilp_t *si, const ir_node *bb, const ir_node *a, const ir_node *b)
2925 const ir_edge_t *edge;
2927 if (get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2928 /* both values are live in, so they interfere */
2932 /* ensure a dominates b */
2933 if (value_dominates(b, a)) {
2939 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2942 /* the following code is stolen from bera.c */
2943 if (be_is_live_end(si->lv, bb, a))
2946 foreach_out_edge(a, edge) {
2947 const ir_node *user = edge->src;
2948 if (get_nodes_block(user) == bb
2951 && ! pset_find_ptr(si->inverse_ops, user)
2952 && value_dominates(b, user))
2960 * Walk all irg blocks and collect interfering values inside of phi classes
2963 luke_interferencewalker(ir_node * bb, void * data)
2965 spill_ilp_t *si = (spill_ilp_t*)data;
2968 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2969 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
2970 op_t *a_op = get_irn_link(a);
2973 /* a is only interesting if it is in my register class and if it is inside a phi class */
2974 if (has_reg_class(si, a) && get_phi_class(si->pc, a)) {
2975 if (a_op->is_remat || pset_find_ptr(si->inverse_ops, a))
2978 for (l2 = _be_lv_next_irn(si->lv, bb, 0xff, l1 + 1); l2 >= 0; l2 = _be_lv_next_irn(si->lv, bb, 0xff, l2 + 1)) {
2979 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
2980 op_t *b_op = get_irn_link(b);
2982 /* a and b are only interesting if they are in the same phi class */
2983 if (has_reg_class(si, b) && get_phi_class(si->pc, a) == get_phi_class(si->pc, b)) {
2984 if (b_op->is_remat || pset_find_ptr(si->inverse_ops, b))
2987 if (values_interfere_in_block(si, bb, a, b)) {
2988 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2989 set_insert_interference(si, si->interferences, a, b, bb);
2997 static unsigned int copy_path_id = 0;
3000 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
3007 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
3008 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3010 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
3012 pset_foreach(copies, ptr) {
3013 copy = PTR_TO_INT(ptr);
3014 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
3019 * @parameter copies contains a path of copies which lead us to irn
3020 * @parameter visited contains a set of nodes already visited on this path
3023 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
3025 const ir_edge_t *edge;
3026 op_t *op = get_irn_link(irn);
3027 pset *visited_users = pset_new_ptr_default();
3030 if(op->is_remat) return 0;
3032 pset_insert_ptr(visited, irn);
3036 pset *visited_operands = pset_new_ptr(get_irn_arity(irn));
3038 /* visit all operands */
3039 for(n=get_irn_arity(irn)-1; n>=0; --n) {
3040 ir_node *arg = get_irn_n(irn, n);
3041 ilp_var_t copy = op->attr.live_range.args.copies[n];
3043 if(!has_reg_class(si, arg)) continue;
3044 if(pset_find_ptr(visited_operands, arg)) continue;
3045 pset_insert_ptr(visited_operands, arg);
3048 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3049 del_pset(visited_operands);
3050 del_pset(visited_users);
3051 pset_remove_ptr(visited, irn);
3054 pset_insert(copies, INT_TO_PTR(copy), copy);
3055 write_copy_path_cst(si, copies, any_interfere);
3056 pset_remove(copies, INT_TO_PTR(copy), copy);
3057 } else if(!pset_find_ptr(visited, arg)) {
3058 pset_insert(copies, INT_TO_PTR(copy), copy);
3059 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
3060 pset_remove(copies, INT_TO_PTR(copy), copy);
3062 if(paths > MAX_PATHS) {
3063 if(pset_count(copies) == 0) {
3067 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3068 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3069 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3070 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3071 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3075 del_pset(visited_operands);
3076 del_pset(visited_users);
3077 pset_remove_ptr(visited, irn);
3080 } else if(pset_count(copies) == 0) {
3086 del_pset(visited_operands);
3089 /* visit all uses which are phis */
3090 foreach_out_edge(irn, edge) {
3091 ir_node *user = edge->src;
3092 int pos = edge->pos;
3093 op_t *op = get_irn_link(user);
3096 if(!is_Phi(user)) continue;
3097 if(!has_reg_class(si, user)) continue;
3098 if(pset_find_ptr(visited_users, user)) continue;
3099 pset_insert_ptr(visited_users, user);
3101 copy = op->attr.live_range.args.copies[pos];
3103 if(user == target) {
3104 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3105 del_pset(visited_users);
3106 pset_remove_ptr(visited, irn);
3109 pset_insert(copies, INT_TO_PTR(copy), copy);
3110 write_copy_path_cst(si, copies, any_interfere);
3111 pset_remove(copies, INT_TO_PTR(copy), copy);
3112 } else if(!pset_find_ptr(visited, user)) {
3113 pset_insert(copies, INT_TO_PTR(copy), copy);
3114 paths += find_copy_path(si, user, target, any_interfere, copies, visited);
3115 pset_remove(copies, INT_TO_PTR(copy), copy);
3117 if(paths > MAX_PATHS) {
3118 if(pset_count(copies) == 0) {
3122 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3123 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3124 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3125 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3126 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3130 del_pset(visited_users);
3131 pset_remove_ptr(visited, irn);
3134 } else if(pset_count(copies) == 0) {
3140 del_pset(visited_users);
3141 pset_remove_ptr(visited, irn);
3146 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
3148 pset * copies = pset_new_ptr_default();
3149 pset * visited = pset_new_ptr_default();
3151 find_copy_path(si, a, b, any_interfere, copies, visited);
3159 memcopyhandler(spill_ilp_t * si)
3161 interference_t *interference;
3163 /* teste Speicherwerte auf Interferenz */
3165 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
3166 irg_block_walk_graph(si->birg->irg, luke_interferencewalker, NULL, si);
3168 /* now lets emit the ILP unequations for the crap */
3169 set_foreach(si->interferences, interference) {
3171 ilp_var_t interfere, any_interfere;
3172 ilp_cst_t any_interfere_cst, cst;
3173 const ir_node *a = interference->a;
3174 const ir_node *b = interference->b;
3176 /* any_interf <= \sum interf */
3177 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
3178 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3179 any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3181 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
3183 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
3184 const ir_node *bb = irnlist->irn;
3185 spill_bb_t *spill_bb = get_irn_link(bb);
3190 spilla = set_find_spill(spill_bb->ilp, a);
3193 spillb = set_find_spill(spill_bb->ilp, b);
3196 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
3197 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
3198 /* 2: - mem_in_a - spill_a + interfere <= 0 */
3199 /* 3: - mem_in_b - spill_b + interfere <= 0 */
3200 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
3201 interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3203 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
3204 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
3206 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
3207 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
3208 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
3209 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
3210 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
3212 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
3213 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3215 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3216 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
3217 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
3219 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
3220 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3222 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3223 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
3224 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
3227 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
3229 /* any_interfere >= interf */
3230 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
3231 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3233 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3234 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3237 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
3238 gen_copy_constraints(si,a,b,any_interfere);
3246 return fabs(x) < 0.00001;
3250 * node attribute hook for changing colors
3252 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
3254 spill_ilp_t *si = get_irg_link(current_ir_graph);
3257 if(pset_find_ptr(si->all_possible_remats, n)) {
3258 op_t *op = (op_t*)get_irn_link(n);
3259 assert(op && op->is_remat);
3261 if(!op->attr.remat.remat->inverse) {
3262 if(op->attr.remat.pre) {
3263 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
3265 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
3270 op_t *op = (op_t*)get_irn_link(n);
3271 assert(op && op->is_remat);
3273 if(op->attr.remat.pre) {
3274 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
3276 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
3287 dump_graph_with_remats(ir_graph * irg, const char * suffix)
3289 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
3290 be_dump(irg, suffix, dump_ir_block_graph_sched);
3291 set_dump_node_vcgattr_hook(NULL);
3295 * Edge hook to dump the schedule edges with annotated register pressure.
3298 sched_pressure_edge_hook(FILE *F, ir_node *irn)
3300 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
3301 ir_node *prev = sched_prev(irn);
3302 fprintf(F, "edge:{sourcename:\"");
3304 fprintf(F, "\" targetname:\"");
3306 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
3307 fprintf(F, "\" color:magenta}\n");
3313 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
3315 DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3317 dump_consts_local(0);
3318 set_dump_node_edge_hook(sched_pressure_edge_hook);
3319 dump_ir_block_graph(irg, suffix);
3320 set_dump_node_edge_hook(old_edge_hook);
3324 walker_pressure_annotator(ir_node * bb, void * data)
3326 spill_ilp_t *si = data;
3329 pset *live = pset_new_ptr_default();
3332 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3333 irn = be_lv_get_irn(si->lv, bb, i);
3335 if (has_reg_class(si, irn)) {
3336 pset_insert_ptr(live, irn);
3340 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3342 sched_foreach_reverse(bb, irn) {
3344 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3348 if(has_reg_class(si, irn)) {
3349 pset_remove_ptr(live, irn);
3350 if(is_Proj(irn)) ++projs;
3353 if(!is_Proj(irn)) projs = 0;
3355 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3356 ir_node *arg = get_irn_n(irn, n);
3358 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3360 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3367 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3369 be_dump(si->birg->irg, suffix, dump_ir_block_graph_sched_pressure);
3373 connect_all_remats_with_keep(spill_ilp_t * si)
3381 n_remats = pset_count(si->all_possible_remats);
3383 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3386 pset_foreach(si->all_possible_remats, irn) {
3391 si->keep = be_new_Keep(si->cls, si->birg->irg, get_irg_end_block(si->birg->irg), n_remats, ins);
3393 obstack_free(si->obst, ins);
3398 connect_all_spills_with_keep(spill_ilp_t * si)
3407 n_spills = pset_count(si->spills);
3409 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3412 pset_foreach(si->spills, irn) {
3417 keep = be_new_Keep(si->cls, si->birg->irg, get_irg_end_block(si->birg->irg), n_spills, ins);
3419 obstack_free(si->obst, ins);
3423 /** insert a spill at an arbitrary position */
3424 static ir_node *be_spill2(ir_node *irn, ir_node *insert)
3426 ir_node *bl = is_Block(insert) ? insert : get_nodes_block(insert);
3427 ir_graph *irg = get_irn_irg(bl);
3428 ir_node *frame = get_irg_frame(irg);
3431 const arch_register_class_t *cls = arch_get_irn_reg_class(irn, -1);
3432 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(frame, -1);
3434 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3437 * search the right insertion point. a spill of a phi cannot be put
3438 * directly after the phi, if there are some phis behind the one which
3439 * is spilled. Also, a spill of a Proj must be after all Projs of the
3442 * Here's one special case:
3443 * If the spill is in the start block, the spill must be after the frame
3444 * pointer is set up. This is done by setting insert to the end of the block
3445 * which is its default initialization (see above).
3448 if (bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3451 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3454 sched_add_after(insert, spill);
3459 delete_remat(spill_ilp_t * si, ir_node * remat) {
3461 ir_node *bad = get_irg_bad(si->birg->irg);
3463 sched_remove(remat);
3465 /* kill links to operands */
3466 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3467 set_irn_n(remat, n, bad);
3472 clean_remat_info(spill_ilp_t * si)
3476 remat_info_t *remat_info;
3477 ir_node *bad = get_irg_bad(si->birg->irg);
3479 set_foreach(si->remat_info, remat_info) {
3480 if(!remat_info->remats) continue;
3482 pset_foreach(remat_info->remats, remat)
3484 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3485 if(sched_is_scheduled(remat->proj)) {
3486 sched_remove((ir_node*)remat->proj);
3488 set_irn_n((ir_node*)remat->proj, -1, bad);
3489 set_irn_n((ir_node*)remat->proj, 0, bad);
3492 if(get_irn_n_edges(remat->op) == 0) {
3493 if(sched_is_scheduled(remat->op)) {
3494 sched_remove((ir_node*)remat->op);
3496 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3497 set_irn_n((ir_node*)remat->op, n, bad);
3502 if(remat_info->remats) del_pset(remat_info->remats);
3503 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3508 delete_unnecessary_remats(spill_ilp_t * si)
3510 if(opt_keep_alive & KEEPALIVE_REMATS) {
3512 ir_node *bad = get_irg_bad(si->birg->irg);
3515 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3516 ir_node *keep_arg = get_irn_n(si->keep, n);
3517 op_t *arg_op = get_irn_link(keep_arg);
3520 assert(arg_op->is_remat);
3522 name = si->lpp->vars[arg_op->attr.remat.ilp];
3524 if(is_zero(name->value)) {
3525 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
3526 /* TODO check whether reload is preferred over remat (could be bug) */
3527 delete_remat(si, keep_arg);
3529 if(!arg_op->attr.remat.remat->inverse) {
3530 if(arg_op->attr.remat.pre) {
3531 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3533 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3536 if(arg_op->attr.remat.pre) {
3537 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3539 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3544 set_irn_n(si->keep, n, bad);
3547 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
3552 pset_foreach(si->all_possible_remats, remat) {
3553 op_t *remat_op = get_irn_link(remat);
3554 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
3556 if(is_zero(name->value)) {
3557 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
3558 /* TODO check whether reload is preferred over remat (could be bug) */
3559 delete_remat(si, remat);
3561 if(!remat_op->attr.remat.remat->inverse) {
3562 if(remat_op->attr.remat.pre) {
3563 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3565 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3568 if(remat_op->attr.remat.pre) {
3569 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3571 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3580 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3582 pset *spills = pset_new_ptr_default();
3584 const ir_node *next;
3587 defs = set_find_def(si->values, value);
3589 if(defs && defs->spills) {
3590 for(next = defs->spills; next; next = get_irn_link(next)) {
3591 pset_insert_ptr(spills, next);
3599 new_r_PhiM_nokeep(ir_graph * irg, ir_node *block, int arity, ir_node **in)
3603 assert( get_irn_arity(block) == arity );
3605 res = new_ir_node(NULL, irg, block, op_Phi, mode_M, arity, in);
3606 res->attr.phi.u.backedge = new_backedge_arr(irg->obst, arity);
3612 * @param before The node after which the spill will be placed in the schedule
3615 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3620 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3622 spill = be_spill2(irn, before);
3624 defs = set_insert_def(si->values, value);
3627 /* enter into the linked list */
3628 set_irn_link(spill, defs->spills);
3629 defs->spills = spill;
3631 if(opt_keep_alive & KEEPALIVE_SPILLS)
3632 pset_insert_ptr(si->spills, spill);
3638 * @param before The Phi node which has to be spilled
3641 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3648 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3650 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3651 ins[n] = si->m_unknown;
3654 mem_phi = new_r_PhiM_nokeep(si->birg->irg, get_nodes_block(phi), get_irn_arity(phi), ins);
3656 defs = set_insert_def(si->values, phi);
3659 /* enter into the linked list */
3660 set_irn_link(mem_phi, defs->spills);
3661 defs->spills = mem_phi;
3663 #ifdef SCHEDULE_PHIM
3664 sched_add_after(phi, mem_phi);
3666 pset_insert_ptr(si->phims, mem_phi);
3669 if(opt_keep_alive & KEEPALIVE_SPILLS)
3670 pset_insert_ptr(si->spills, mem_phi);
3677 * Add remat to list of defs, destroys link field!
3680 insert_remat(spill_ilp_t * si, ir_node * remat)
3683 op_t *remat_op = get_irn_link(remat);
3685 assert(remat_op->is_remat);
3687 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3690 /* enter into the linked list */
3691 set_irn_link(remat, defs->remats);
3692 defs->remats = remat;
3697 * Add reload before operation and add to list of defs
3700 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3705 const arch_env_t *arch_env = si->birg->main_env->arch_env;
3707 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3709 defs = set_find_def(si->values, value);
3711 spill = defs->spills;
3712 assert(spill && "no spill placed before reload");
3714 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3716 /* enter into the linked list */
3717 set_irn_link(reload, defs->remats);
3718 defs->remats = reload;
3723 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3726 ir_node *value = get_irn_n(memoperand->irn, memoperand->pos);
3729 DBG((si->dbg, LEVEL_2, "\t inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3731 defs = set_find_def(si->values, value);
3733 spill = defs->spills;
3734 assert(spill && "no spill placed before reload");
3736 arch_perform_memory_operand(memoperand->irn, spill, memoperand->pos);
3739 void insert_memoperands(spill_ilp_t * si)
3741 memoperand_t *memoperand;
3744 set_foreach(si->memoperands, memoperand) {
3745 name = si->lpp->vars[memoperand->ilp];
3746 if(!is_zero(name->value)) {
3747 perform_memory_operand(si, memoperand);
3753 walker_spill_placer(ir_node * bb, void * data) {
3754 spill_ilp_t *si = (spill_ilp_t*)data;
3756 spill_bb_t *spill_bb = get_irn_link(bb);
3757 pset *spills_to_do = pset_new_ptr_default();
3760 set_foreach(spill_bb->ilp, spill) {
3763 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3764 name = si->lpp->vars[spill->mem_in];
3765 if(!is_zero(name->value)) {
3768 mem_phi = insert_mem_phi(si, spill->irn);
3770 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3774 name = si->lpp->vars[spill->spill];
3775 if(!is_zero(name->value)) {
3776 /* place spill directly after definition */
3777 if(get_nodes_block(spill->irn) == bb) {
3778 insert_spill(si, spill->irn, spill->irn, spill->irn);
3782 /* place spill at bb start */
3783 if(spill->reg_in > 0) {
3784 name = si->lpp->vars[spill->reg_in];
3785 if(!is_zero(name->value)) {
3786 insert_spill(si, spill->irn, spill->irn, bb);
3790 /* place spill after a remat */
3791 pset_insert_ptr(spills_to_do, spill->irn);
3794 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3797 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3798 op_t *op = get_irn_link(irn);
3800 if(be_is_Spill(irn)) continue;
3803 /* TODO fix this if we want to support remats with more than two nodes */
3804 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3805 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3807 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3810 if(pset_find_ptr(spills_to_do, irn)) {
3811 pset_remove_ptr(spills_to_do, irn);
3813 insert_spill(si, irn, irn, irn);
3819 assert(pset_count(spills_to_do) == 0);
3821 /* afterwards free data in block */
3822 del_pset(spills_to_do);
3826 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3828 ir_node *insert_pos = bb;
3831 /* find last definition of arg value in block */
3836 defs = set_find_def(si->values, value);
3838 if(defs && defs->remats) {
3839 for(next = defs->remats; next; next = get_irn_link(next)) {
3840 if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3841 last = sched_get_time_step(next);
3847 if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3848 last = sched_get_time_step(value);
3852 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
3854 spill = be_spill2(is_Block(insert_pos)?value:insert_pos, insert_pos);
3860 phim_fixer(spill_ilp_t *si) {
3863 set_foreach(si->values, defs) {
3864 const ir_node *phi = defs->value;
3865 op_t *op = get_irn_link(phi);
3866 ir_node *phi_m = NULL;
3867 ir_node *next = defs->spills;
3870 if(!is_Phi(phi)) continue;
3873 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3877 next = get_irn_link(next);
3880 if(!phi_m) continue;
3882 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3883 ir_node *value = get_irn_n(phi, n);
3884 defs_t *val_defs = set_find_def(si->values, value);
3886 /* a spill of this value */
3891 ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3892 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3894 if(!is_zero(name->value)) {
3895 spill = insert_mem_copy(si, pred, value);
3897 spill = val_defs->spills;
3900 spill = val_defs->spills;
3903 assert(spill && "no spill placed before PhiM");
3904 set_irn_n(phi_m, n, spill);
3910 walker_reload_placer(ir_node * bb, void * data) {
3911 spill_ilp_t *si = (spill_ilp_t*)data;
3913 spill_bb_t *spill_bb = get_irn_link(bb);
3915 /* reloads at end of block */
3916 if(spill_bb->reloads) {
3919 set_foreach(spill_bb->reloads, keyval) {
3920 ir_node *irn = (ir_node*)keyval->key;
3921 ilp_var_t reload = PTR_TO_INT(keyval->val);
3924 name = si->lpp->vars[reload];
3925 if(!is_zero(name->value)) {
3927 ir_node *insert_pos = bb;
3928 ir_node *prev = sched_block_last_noncf(si, bb);
3929 op_t *prev_op = get_irn_link(prev);
3931 while(be_is_Spill(prev)) {
3932 prev = sched_prev(prev);
3935 prev_op = get_irn_link(prev);
3937 /* insert reload before pre-remats */
3938 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3939 && prev_op->is_remat && prev_op->attr.remat.pre) {
3943 prev = sched_prev(prev);
3944 } while(be_is_Spill(prev));
3946 prev_op = get_irn_link(prev);
3950 reload = insert_reload(si, irn, insert_pos);
3952 if(opt_keep_alive & KEEPALIVE_RELOADS)
3953 pset_insert_ptr(si->spills, reload);
3958 /* walk and insert more reloads and collect remats */
3959 sched_foreach_reverse(bb, irn) {
3960 op_t *op = get_irn_link(irn);
3962 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3963 if(is_Phi(irn)) break;
3966 if(get_irn_mode(irn) != mode_T) {
3967 insert_remat(si, irn);
3972 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3973 ir_node *arg = get_irn_n(irn, n);
3975 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3978 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3979 if(!is_zero(name->value)) {
3981 ir_node *insert_pos = irn;
3982 ir_node *prev = sched_prev(insert_pos);
3985 while(be_is_Spill(prev)) {
3986 prev = sched_prev(prev);
3989 prev_op = get_irn_link(prev);
3991 /* insert reload before pre-remats */
3992 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3993 && prev_op->is_remat && prev_op->attr.remat.pre) {
3997 prev = sched_prev(prev);
3998 } while(be_is_Spill(prev));
4000 prev_op = get_irn_link(prev);
4004 reload = insert_reload(si, arg, insert_pos);
4006 assert(reload && "no reload returned");
4007 set_irn_n(irn, n, reload);
4009 if(opt_keep_alive & KEEPALIVE_RELOADS)
4010 pset_insert_ptr(si->spills, reload);
4017 del_set(spill_bb->ilp);
4018 if(spill_bb->reloads) del_set(spill_bb->reloads);
4022 walker_collect_used(ir_node * irn, void * data)
4024 bitset_t *used = data;
4026 bitset_set(used, get_irn_idx(irn));
4029 struct kill_helper {
4035 walker_kill_unused(ir_node * bb, void * data)
4037 struct kill_helper *kh = data;
4038 ir_node *bad = get_irg_bad(get_irn_irg(bb));
4042 for(irn=sched_first(bb); !sched_is_end(irn);) {
4043 ir_node *next = sched_next(irn);
4046 if(!bitset_is_set(kh->used, get_irn_idx(irn))) {
4047 if(be_is_Spill(irn) || be_is_Reload(irn)) {
4048 DBG((kh->si->dbg, LEVEL_1, "\t SUBOPTIMAL! %+F IS UNUSED (cost: %g)\n", irn, get_cost(irn) * execution_frequency(kh->si, bb)));
4050 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
4056 set_nodes_block(irn, bad);
4057 for (n=get_irn_arity(irn)-1; n>=0; --n) {
4058 set_irn_n(irn, n, bad);
4065 #ifndef SCHEDULE_PHIM
4067 kill_unused_phims(spill_ilp_t * si, struct kill_helper * kh)
4070 ir_node *bad = get_irg_bad(si->birg->irg);
4073 pset_foreach(si->phims, phi) {
4074 if(!bitset_is_set(kh->used, get_irn_idx(phi))) {
4076 set_nodes_block(phi, bad);
4077 for (n=get_irn_arity(phi)-1; n>=0; --n) {
4078 set_irn_n(phi, n, bad);
4086 kill_all_unused_values_in_schedule(spill_ilp_t * si)
4088 struct kill_helper kh;
4090 kh.used = bitset_malloc(get_irg_last_idx(si->birg->irg));
4093 irg_walk_graph(si->birg->irg, walker_collect_used, NULL, kh.used);
4094 #ifndef SCHEDULE_PHIM
4095 kill_unused_phims(si, &kh);
4097 irg_block_walk_graph(si->birg->irg, walker_kill_unused, NULL, &kh);
4099 bitset_free(kh.used);
4103 print_irn_pset(pset * p)
4107 pset_foreach(p, irn) {
4108 ir_printf("%+F\n", irn);
4113 dump_phi_class(spill_ilp_t *si, ir_node **phiclass, const char * file)
4115 FILE *f = fopen(file, "w");
4117 interference_t *interference;
4120 set_break(si->interferences);
4122 ir_fprintf(f, "digraph phiclass {\n");
4124 for (i = ARR_LEN(phiclass) - 1; i >= 0; --i) {
4127 ir_fprintf(f, " %F%N [shape=box]\n", irn, irn);
4130 for (i = ARR_LEN(phiclass) - 1; i >= 0; --i) {
4137 for (n = get_irn_arity(irn) - 1; n >= 0; --n) {
4138 ir_node *arg = get_irn_n(irn, n);
4140 ir_fprintf(f, " %F%N -> %F%N\n", irn, irn, arg, arg);
4144 set_foreach(si->interferences, interference) {
4145 const ir_node *a = interference->a;
4146 const ir_node *b = interference->b;
4147 if (get_phi_class(si->pc, (ir_node *)a) == phiclass) {
4148 ir_fprintf(f, " %F%N -> %F%N [color=red,dir=none,style=bold]\n", a, a, b, b);
4157 rewire_uses(spill_ilp_t * si)
4160 ir_nodeset_t ignore;
4162 ir_nodeset_init(&ignore);
4163 ir_nodeset_insert(&ignore, get_irg_end(si->birg->irg));
4165 /* then fix uses of spills */
4166 set_foreach(si->values, defs) {
4169 const ir_node *next = defs->remats;
4172 reloads = pset_new_ptr_default();
4175 if(be_is_Reload(next)) {
4176 pset_insert_ptr(reloads, next);
4180 next = get_irn_link(next);
4183 spills = get_spills_for_value(si, defs->value);
4184 DBG((si->dbg, LEVEL_2, "\t %d remats, %d reloads, and %d spills for value %+F\n", remats, pset_count(reloads), pset_count(spills), defs->value));
4185 if(pset_count(spills) > 1) {
4186 be_ssa_construction_env_t senv;
4188 //assert(pset_count(reloads) > 0);
4189 // print_irn_pset(spills);
4190 // print_irn_pset(reloads);
4192 be_ssa_construction_init(&senv, si->birg);
4193 be_ssa_construction_set_ignore_uses(&senv, &ignore);
4194 pset_foreach(spills, node) {
4195 be_ssa_construction_add_copy(&senv, node);
4197 pset_foreach(spills, node) {
4198 be_ssa_construction_fix_users(&senv, node);
4200 be_ssa_construction_update_liveness_phis(&senv, si->lv);
4201 pset_foreach(spills, node) {
4202 be_liveness_update(si->lv, node);
4204 be_ssa_construction_destroy(&senv);
4211 /* first fix uses of remats and reloads */
4212 set_foreach(si->values, defs) {
4213 const ir_node *next = defs->remats;
4217 be_ssa_construction_env_t senv;
4219 be_ssa_construction_init(&senv, si->birg);
4221 if(sched_is_scheduled(defs->value)) {
4222 be_ssa_construction_add_copy(&senv, (ir_node*) defs->value);
4226 next = defs->remats;
4228 be_ssa_construction_add_copy(&senv, (ir_node*) next);
4229 next = get_irn_link(next);
4232 if(sched_is_scheduled(defs->value)) {
4233 be_ssa_construction_fix_users(&senv, (ir_node*) defs->value);
4236 next = defs->remats;
4238 be_ssa_construction_fix_users(&senv, (ir_node*) next);
4239 next = get_irn_link(next);
4242 be_ssa_construction_update_liveness_phis(&senv, si->lv);
4243 if(sched_is_scheduled(defs->value)) {
4244 be_liveness_update(si->lv, (ir_node*) defs->value);
4247 next = defs->remats;
4249 be_liveness_update(si->lv, (ir_node*) next);
4250 next = get_irn_link(next);
4253 be_ssa_construction_destroy(&senv);
4257 ir_nodeset_destroy(&ignore);
4258 // remove_unused_defs(si);
4263 writeback_results(spill_ilp_t * si)
4265 /* walk through the graph and collect all spills, reloads and remats for a value */
4267 si->values = new_set(cmp_defs, 4096);
4269 DBG((si->dbg, LEVEL_1, "Applying results\n"));
4270 delete_unnecessary_remats(si);
4271 si->m_unknown = new_r_Unknown(si->birg->irg, mode_M);
4272 irg_block_walk_graph(si->birg->irg, walker_spill_placer, NULL, si);
4273 irg_block_walk_graph(si->birg->irg, walker_reload_placer, NULL, si);
4275 insert_memoperands(si);
4278 /* clean the remat info! there are still back-edges leading there! */
4279 clean_remat_info(si);
4283 connect_all_spills_with_keep(si);
4285 del_set(si->values);
4289 get_n_regs(spill_ilp_t * si)
4291 int arch_n_regs = arch_register_class_n_regs(si->cls);
4293 bitset_t *arch_regs = bitset_malloc(arch_n_regs);
4294 bitset_t *abi_regs = bitset_malloc(arch_n_regs);
4296 arch_put_non_ignore_regs(si->cls, arch_regs);
4297 be_abi_put_ignore_regs(si->birg->abi, si->cls, abi_regs);
4299 bitset_andnot(arch_regs, abi_regs);
4300 arch_n_regs = bitset_popcnt(arch_regs);
4302 bitset_free(arch_regs);
4303 bitset_free(abi_regs);
4305 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", arch_n_regs, si->cls->name));
4310 walker_reload_mover(ir_node * bb, void * data)
4312 spill_ilp_t *si = data;
4315 sched_foreach(bb, tmp) {
4316 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
4317 ir_node *reload = tmp;
4320 /* move reload upwards */
4322 int pressure = (int)get_irn_link(reload);
4323 if(pressure < si->n_regs) {
4324 irn = sched_prev(reload);
4325 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
4326 sched_remove(reload);
4327 pressure = (int)get_irn_link(irn);
4329 while(pressure < si->n_regs) {
4330 if( sched_is_end(irn) ||
4331 (be_is_Reload(irn) && has_reg_class(si, irn)) ||
4332 /* do not move reload before its spill */
4333 (irn == be_get_Reload_mem(reload)) ||
4334 /* do not move before phi */
4337 set_irn_link(irn, INT_TO_PTR(pressure+1));
4338 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
4339 irn = sched_prev(irn);
4341 pressure = (int)get_irn_link(irn);
4344 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
4345 sched_put_after(irn, reload);
4352 move_reloads_upward(spill_ilp_t * si)
4354 irg_block_walk_graph(si->birg->irg, walker_reload_mover, NULL, si);
4359 * Walk all irg blocks and check for interfering spills inside of phi classes
4362 luke_meminterferencechecker(ir_node * bb, void * data)
4364 spill_ilp_t *si = (spill_ilp_t*)data;
4367 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
4368 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
4370 if (! be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T))
4373 /* a is only interesting if it is in my register class and if it is inside a phi class */
4374 if (has_reg_class(si, a) && get_phi_class(si->pc, a)) {
4375 for (l2 = _be_lv_next_irn(si->lv, bb, 0xff, l1 + 1); l2 >= 0; l2 = _be_lv_next_irn(si->lv, bb, 0xff, l2 + 1)) {
4376 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
4378 if (! be_is_Spill(b) && (! is_Phi(b) || get_irn_mode(b) != mode_T))
4381 /* a and b are only interesting if they are in the same phi class */
4382 if (has_reg_class(si, b) && get_phi_class(si->pc, a) == get_phi_class(si->pc, b)) {
4383 if (values_interfere_in_block(si, bb, a, b)) {
4384 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4393 verify_phiclasses(spill_ilp_t * si)
4395 /* analyze phi classes */
4396 phi_class_free(si->pc);
4397 si->pc = phi_class_new_from_irg(si->birg->irg, 0);
4399 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4400 irg_block_walk_graph(si->birg->irg, luke_meminterferencechecker, NULL, si);
4404 be_spill_remat(be_irg_t *birg, const arch_register_class_t *cls)
4407 char problem_name[256];
4408 char dump_suffix[256];
4409 char dump_suffix2[256];
4410 struct obstack obst;
4412 ir_graph *irg = be_get_birg_irg(birg);
4414 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", irg, cls->name);
4415 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", cls->name);
4416 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", cls->name);
4418 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4419 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4421 if(opt_verify & VERIFY_DOMINANCE)
4422 be_check_dominance(irg);
4424 be_assure_liveness(birg);
4426 obstack_init(&obst);
4430 si.lpp = new_lpp(problem_name, lpp_minimize);
4431 si.remat_info = new_set(cmp_remat_info, 4096);
4432 si.interferences = new_set(cmp_interference, 32);
4433 si.memoperands = new_set(cmp_memoperands, 128);
4434 si.all_possible_remats = pset_new_ptr_default();
4435 si.spills = pset_new_ptr_default();
4436 si.inverse_ops = pset_new_ptr_default();
4439 si.n_regs = get_n_regs(&si);
4441 set_irg_link(irg, &si);
4444 /* compute phi classes */
4445 // phi_class_compute(irg);
4447 if(opt_dump_flags & DUMP_STATS)
4448 be_analyze_regpressure(birg, cls, "-pre");
4450 DBG((si.dbg, LEVEL_2, "\t initializing\n"));
4451 irg_block_walk_graph(irg, luke_initializer, NULL, &si);
4454 /* collect remats */
4455 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4456 irg_walk_graph(irg, walker_remat_collector, NULL, &si);
4459 /* insert possible remats */
4460 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4461 irg_block_walk_graph(irg, walker_remat_insertor, NULL, &si);
4462 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4464 if(opt_keep_alive & KEEPALIVE_REMATS) {
4465 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4466 connect_all_remats_with_keep(&si);
4467 /* dump graph with inserted remats */
4468 dump_graph_with_remats(irg, dump_suffix);
4471 /* insert copies for phi arguments not in my regclass */
4472 irg_walk_graph(irg, walker_regclass_copy_insertor, NULL, &si);
4474 /* recompute liveness */
4475 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4476 be_liveness_recompute(si.lv);
4479 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4480 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4481 irg_block_walk_graph(irg, luke_endwalker, NULL, &si);
4483 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4484 irg_block_walk_graph(irg, luke_blockwalker, NULL, &si);
4486 si.pc = phi_class_new_from_irg(birg->irg, 0);
4487 if (opt_memcopies) {
4488 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4489 memcopyhandler(&si);
4492 if (opt_dump_flags & DUMP_PROBLEM) {
4494 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4495 if ((f = fopen(buf, "wt")) != NULL) {
4496 lpp_dump_plain(si.lpp, f);
4501 if (opt_dump_flags & DUMP_MPS) {
4504 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4505 if ((f = fopen(buf, "wt")) != NULL) {
4506 mps_write_mps(si.lpp, s_mps_fixed, f);
4510 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4511 if ((f = fopen(buf, "wt")) != NULL) {
4512 mps_write_mst(si.lpp, s_mps_fixed, f);
4517 lpp_check_startvals(si.lpp);
4520 DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4521 lpp_set_time_limit(si.lpp, opt_timeout);
4524 lpp_set_log(si.lpp, stdout);
4527 lpp_solve_cplex(si.lpp);
4529 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4531 assert(lpp_is_sol_valid(si.lpp)
4532 && "solution of ILP must be valid");
4534 DBG((si.dbg, LEVEL_1, "\t%s: iterations: %d, solution time: %g, objective function: %g, best bound: %g\n", problem_name, si.lpp->iterations, si.lpp->sol_time, is_zero(si.lpp->objval)?0.0:si.lpp->objval, is_zero(si.lpp->best_bound)?0.0:si.lpp->best_bound));
4536 if(opt_dump_flags & DUMP_SOLUTION) {
4540 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4541 if ((f = fopen(buf, "wt")) != NULL) {
4543 for (i = 0; i < si.lpp->var_next; ++i) {
4544 lpp_name_t *name = si.lpp->vars[i];
4545 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4551 #ifndef SCHEDULE_PHIM
4552 si.phims = pset_new_ptr_default();
4554 writeback_results(&si);
4559 kill_all_unused_values_in_schedule(&si);
4561 #if !defined(SCHEDULE_PHIM) && defined(SOLVE)
4565 if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
4566 be_dump(irg, "-spills-placed", dump_ir_block_graph);
4568 // move reloads upwards
4569 be_liveness_recompute(si.lv);
4570 irg_block_walk_graph(irg, walker_pressure_annotator, NULL, &si);
4571 move_reloads_upward(&si);
4574 verify_phiclasses(&si);
4577 irg_block_walk_graph(irg, walker_pressure_annotator, NULL, &si);
4579 if(opt_dump_flags & DUMP_PRESSURE)
4580 dump_pressure_graph(&si, dump_suffix2);
4582 if(opt_dump_flags & DUMP_STATS)
4583 be_analyze_regpressure(birg, cls, "-post");
4585 if(opt_verify & VERIFY_DOMINANCE)
4586 be_check_dominance(irg);
4589 del_set(si.interferences);
4590 del_pset(si.inverse_ops);
4591 del_pset(si.all_possible_remats);
4592 del_set(si.memoperands);
4593 del_pset(si.spills);
4595 phi_class_free(si.pc);
4596 obstack_free(&obst, NULL);
4597 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4600 void be_init_spillremat(void)
4602 static be_spiller_t remat_spiller = {
4605 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
4606 lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
4607 lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
4608 lc_opt_entry_t *remat_grp = lc_opt_get_grp(chordal_grp, "remat");
4610 be_register_spiller("remat", &remat_spiller);
4611 lc_opt_add_table(remat_grp, options);
4614 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillremat);
4616 #else /* WITH_ILP */
4618 static __attribute__((unused))
4619 void only_that_you_can_compile_without_WITH_ILP_defined(void)
4623 #endif /* WITH_ILP */