1 /** vim: set sw=4 ts=4:
4 * @author Adam M. Szalkowski & Sebastian Hack
6 * ILP based spilling & rematerialization
8 * Copyright (C) 2006 Universitaet Karlsruhe
9 * Released under the GPL
32 #include "phiclass_t.h"
39 #include <lpp/lpp_net.h>
40 #include <lpp/lpp_cplex.h>
41 //#include <lc_pset.h>
42 #include <libcore/lc_bitset.h>
46 #include "besched_t.h"
51 #include "bespillremat.h"
53 #include "bepressurestat.h"
55 #include "bechordal_t.h"
58 #include <libcore/lc_opts.h>
59 #include <libcore/lc_opts_enum.h>
60 #endif /* WITH_LIBCORE */
62 #define DUMP_PROBLEM 1
64 #define DUMP_SOLUTION 4
66 #define KEEPALIVE_REMATS 1
67 #define KEEPALIVE_SPILLS 2
68 #define KEEPALIVE_RELOADS 4
70 #define VERIFY_MEMINTERF 1
71 #define VERIFY_DOMINANCE 2
74 #define REMATS_BRIGGS 1
75 #define REMATS_NOINVERSE 2
78 static int opt_dump_flags = 0;
79 static int opt_log = 0;
80 static int opt_keep_alive = 0;
81 static int opt_goodwin = 1;
82 static int opt_memcopies = 1;
83 static int opt_memoperands = 1;
84 static int opt_verify = VERIFY_MEMINTERF;
85 static int opt_remats = REMATS_ALL;
86 static int opt_repair_schedule = 0;
87 static int opt_no_enlarge_liveness = 0;
88 static int opt_remat_while_live = 1;
89 static int opt_timeout = 300;
90 static double opt_cost_reload = 8.0;
91 static double opt_cost_memoperand = 7.0;
92 static double opt_cost_spill = 50.0;
93 static double opt_cost_remat = 1.0;
97 static const lc_opt_enum_mask_items_t dump_items[] = {
98 { "problem", DUMP_PROBLEM },
100 { "solution", DUMP_SOLUTION },
104 static lc_opt_enum_mask_var_t dump_var = {
105 &opt_dump_flags, dump_items
108 static const lc_opt_enum_mask_items_t keepalive_items[] = {
109 { "remats", KEEPALIVE_REMATS },
110 { "spills", KEEPALIVE_SPILLS },
111 { "reloads", KEEPALIVE_RELOADS },
115 static lc_opt_enum_mask_var_t keep_alive_var = {
116 &opt_keep_alive, keepalive_items
119 static const lc_opt_enum_mask_items_t remats_items[] = {
120 { "none", REMATS_NONE },
121 { "briggs", REMATS_BRIGGS },
122 { "noinverse", REMATS_NOINVERSE },
123 { "all", REMATS_ALL },
127 static lc_opt_enum_mask_var_t remats_var = {
128 &opt_remats, remats_items
131 static const lc_opt_table_entry_t options[] = {
132 LC_OPT_ENT_ENUM_MASK("keepalive", "keep alive remats, spills or reloads", &keep_alive_var),
134 LC_OPT_ENT_BOOL ("goodwin", "activate goodwin reduction", &opt_goodwin),
135 LC_OPT_ENT_BOOL ("memcopies", "activate memcopy handling", &opt_memcopies),
136 LC_OPT_ENT_BOOL ("memoperands", "activate memoperands", &opt_memoperands),
137 LC_OPT_ENT_ENUM_INT ("remats", "type of remats to insert (none, briggs, noinverse or all)",&remats_var),
138 LC_OPT_ENT_BOOL ("repair_schedule", "repair the schedule by rematting once used nodes",&opt_repair_schedule),
139 LC_OPT_ENT_BOOL ("no_enlage_liveness", "do not enlarge liveness of operands of remats",&opt_no_enlarge_liveness),
140 LC_OPT_ENT_BOOL ("remat_while_live", "remat only values that can be used by real ops", &opt_remat_while_live),
142 LC_OPT_ENT_ENUM_MASK("dump", "dump problem, mps or solution", &dump_var),
143 LC_OPT_ENT_BOOL ("log", "activate the lpp log", &opt_log),
144 LC_OPT_ENT_INT ("timeout", "ILP solver timeout", &opt_timeout),
146 LC_OPT_ENT_DBL ("cost_reload", "cost of a reload", &opt_cost_reload),
147 LC_OPT_ENT_DBL ("cost_memoperand", "cost of a memory operand", &opt_cost_memoperand),
148 LC_OPT_ENT_DBL ("cost_spill", "cost of a spill instruction", &opt_cost_spill),
149 LC_OPT_ENT_DBL ("cost_remat", "cost of a rematerialization", &opt_cost_remat),
153 void be_spill_remat_register_options(lc_opt_entry_t *grp)
155 lc_opt_entry_t *my_grp = lc_opt_get_grp(grp, "remat");
156 lc_opt_add_table(my_grp, options);
161 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
162 //#define SCHEDULE_PHIM /* insert phim nodes into schedule */
165 //#define SOLVE_LOCAL
166 #define LPP_SERVER "i44pc52"
167 #define LPP_SOLVER "cplex"
170 #define MAX_PATHS INT_MAX
173 typedef struct _spill_ilp_t {
174 const arch_register_class_t *cls;
176 const be_chordal_env_t *chordal_env;
179 struct obstack *obst;
181 pset *all_possible_remats;
184 set *values; /**< for collecting all definitions of values before running ssa-construction */
189 DEBUG_ONLY(firm_dbg_module_t * dbg);
192 typedef int ilp_var_t;
193 typedef int ilp_cst_t;
195 typedef struct _spill_bb_t {
200 typedef struct _remat_t {
201 const ir_node *op; /**< for copy_irn */
202 const ir_node *value; /**< the value which is being recomputed by this remat */
203 const ir_node *proj; /**< not NULL if the above op produces a tuple */
204 int cost; /**< cost of this remat */
205 int inverse; /**< nonzero if this is an inverse remat */
209 * Data to be attached to each IR node. For remats this contains the ilp_var
210 * for this remat and for normal ops this contains the ilp_vars for
211 * reloading each operand
213 typedef struct _op_t {
218 const remat_t *remat; /** the remat this op belongs to */
219 int pre; /** 1, if this is a pressure-increasing remat */
223 ir_node *op; /** the operation this live range belongs to */
232 typedef struct _defs_t {
233 const ir_node *value;
234 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
235 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
238 typedef struct _remat_info_t {
239 const ir_node *irn; /**< the irn to which these remats belong */
240 pset *remats; /**< possible remats for this value */
241 pset *remats_by_operand; /**< remats with this value as operand */
244 typedef struct _keyval_t {
249 typedef struct _spill_t {
258 typedef struct _memoperand_t {
259 ir_node *irn; /**< the irn */
260 unsigned int pos; /**< the position of the argument */
261 ilp_var_t ilp; /**< the ilp var for this memory operand */
265 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
267 return chordal_has_class(si->chordal_env, irn);
272 cmp_remat(const void *a, const void *b)
274 const keyval_t *p = a;
275 const keyval_t *q = b;
276 const remat_t *r = p->val;
277 const remat_t *s = q->val;
281 return !(r == s || r->op == s->op);
285 cmp_remat(const void *a, const void *b)
287 const remat_t *r = a;
288 const remat_t *s = a;
290 return !(r == s || r->op == s->op);
294 cmp_spill(const void *a, const void *b, size_t size)
296 const spill_t *p = a;
297 const spill_t *q = b;
299 // return !(p->irn == q->irn && p->bb == q->bb);
300 return !(p->irn == q->irn);
304 cmp_memoperands(const void *a, const void *b, size_t size)
306 const memoperand_t *p = a;
307 const memoperand_t *q = b;
309 return !(p->irn == q->irn && p->pos == q->pos);
313 set_find_keyval(set * set, const void * key)
318 return set_find(set, &query, sizeof(query), HASH_PTR(key));
322 set_insert_keyval(set * set, void * key, void * val)
328 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
332 set_find_def(set * set, const ir_node * value)
337 return set_find(set, &query, sizeof(query), HASH_PTR(value));
341 set_insert_def(set * set, const ir_node * value)
348 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
351 static memoperand_t *
352 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
359 return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
362 static memoperand_t *
363 set_find_memoperand(set * set, const ir_node * irn, unsigned int pos)
367 query.irn = (ir_node*)irn;
369 return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
374 set_find_spill(set * set, const ir_node * value)
378 query.irn = (ir_node*)value;
379 return set_find(set, &query, sizeof(query), HASH_PTR(value));
382 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
383 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
384 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
385 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
386 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
389 cmp_remat_info(const void *a, const void *b, size_t size)
391 const remat_info_t *p = a;
392 const remat_info_t *q = b;
394 return !(p->irn == q->irn);
398 cmp_defs(const void *a, const void *b, size_t size)
403 return !(p->value == q->value);
407 cmp_keyval(const void *a, const void *b, size_t size)
409 const keyval_t *p = a;
410 const keyval_t *q = b;
412 return !(p->key == q->key);
416 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
419 #ifndef EXECFREQ_LOOPDEPH
420 return get_block_execfreq(si->chordal_env->exec_freq, get_block(irn)) + FUDGE;
423 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
425 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
430 get_cost(const spill_ilp_t * si, const ir_node * irn)
432 if(be_is_Spill(irn)) {
433 return opt_cost_spill;
434 } else if(be_is_Reload(irn)){
435 return opt_cost_reload;
437 return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
442 * Checks, whether node and its operands have suitable reg classes
445 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
448 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
449 int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
453 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
456 for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
457 ir_node *op = get_irn_n(irn, n);
458 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
461 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
468 * Try to create a remat from @p op with destination value @p dest_value
470 static INLINE remat_t *
471 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
473 remat_t *remat = NULL;
475 // if(!mode_is_datab(get_irn_mode(dest_value)))
478 if(dest_value == op) {
479 const ir_node *proj = NULL;
481 if(is_Proj(dest_value)) {
482 op = get_Proj_pred(op);
486 if(!is_rematerializable(si, op))
489 remat = obstack_alloc(si->obst, sizeof(*remat));
491 remat->cost = get_cost(si, op);
492 remat->value = dest_value;
496 arch_inverse_t inverse;
499 /* get the index of the operand we want to retrieve by the inverse op */
500 for (n = get_irn_arity(op)-1; n>=0; --n) {
501 ir_node *arg = get_irn_n(op, n);
503 if(arg == dest_value) break;
507 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", n, op));
509 /* else ask the backend to give an inverse op */
510 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
513 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
515 assert(inverse.n > 0 && "inverse op should have at least one node");
517 for(i=inverse.n-1; i>=0; --i) {
518 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
522 remat = obstack_alloc(si->obst, sizeof(*remat));
523 remat->op = inverse.nodes[0];
524 remat->cost = inverse.costs;
525 remat->value = dest_value;
526 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
529 assert(is_Proj(remat->proj));
531 assert(0 && "I can not handle remats with more than 2 nodes");
538 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
540 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
548 add_remat(const spill_ilp_t * si, const remat_t * remat)
550 remat_info_t *remat_info,
555 assert(remat->value);
557 query.irn = remat->value;
559 query.remats_by_operand = NULL;
560 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
562 if(remat_info->remats == NULL) {
563 remat_info->remats = new_pset(cmp_remat, 4096);
565 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
567 /* insert the remat into the remats_be_operand set of each argument of the remat op */
568 for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
569 ir_node *arg = get_irn_n(remat->op, n);
573 query.remats_by_operand = NULL;
574 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
576 if(remat_info->remats_by_operand == NULL) {
577 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
579 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
584 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
586 const ir_edge_t *edge = get_irn_out_edge_first(irn);
590 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
593 edge = get_irn_out_edge_next(irn, edge);
600 get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
606 irn = get_Proj_pred(irn);
608 for(n=get_irn_arity(irn)-1; n>=0; --n) {
609 const ir_node *arg = get_irn_n(irn, n);
611 if(has_reg_class(si, arg)) ++ret;
618 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
623 if( has_reg_class(si, op)
624 && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1)
625 && (opt_remats != REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0)
627 remat = get_remat_from_op(si, op, op);
629 add_remat(si, remat);
633 if(opt_remats == REMATS_ALL) {
634 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
636 for (n = get_irn_arity(op)-1; n>=0; --n) {
637 ir_node *arg = get_irn_n(op, n);
639 if(has_reg_class(si, arg)) {
640 /* try to get an inverse remat */
641 remat = get_remat_from_op(si, arg, op);
643 add_remat(si, remat);
651 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
654 ir_node *def_block = get_nodes_block(val);
660 /* if pos is at end of a basic block */
662 ret = (pos == def_block || block_dominates(def_block, pos));
663 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
667 /* else if this is a normal operation */
668 block = get_nodes_block(pos);
669 if(block == def_block) {
670 if(!sched_is_scheduled(val)) return 1;
672 ret = sched_comes_after(val, pos);
673 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
677 ret = block_dominates(def_block, block);
678 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
682 static INLINE ir_node *
683 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
685 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
689 * Returns first non-Phi node of block @p bb
691 static INLINE ir_node *
692 sched_block_first_nonphi(const ir_node * bb)
694 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
698 sched_skip_proj_predicator(const ir_node * irn, void * data)
700 return (is_Proj(irn));
703 static INLINE ir_node *
704 sched_next_nonproj(const ir_node * irn, int forward)
706 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
710 * Returns next operation node (non-Proj) after @p irn
711 * or the basic block of this node
713 static INLINE ir_node *
714 sched_next_op(const ir_node * irn)
716 ir_node *next = sched_next(irn);
721 return sched_next_nonproj(next, 1);
725 * Returns previous operation node (non-Proj) before @p irn
726 * or the basic block of this node
728 static INLINE ir_node *
729 sched_prev_op(const ir_node * irn)
731 ir_node *prev = sched_prev(irn);
736 return sched_next_nonproj(prev, 0);
740 sched_put_after(ir_node * insert, ir_node * irn)
742 if(is_Block(insert)) {
743 insert = sched_block_first_nonphi(insert);
745 insert = sched_next_op(insert);
747 sched_add_before(insert, irn);
751 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
753 if(is_Block(insert)) {
754 insert = sched_block_last_noncf(si, insert);
756 insert = sched_next_nonproj(insert, 0);
757 insert = sched_prev(insert);
759 sched_add_after(insert, irn);
763 * Tells you whether a @p remat can be placed before the irn @p pos
766 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
768 const ir_node *op = remat->op;
774 prev = sched_block_last_noncf(si, pos);
775 prev = sched_next_nonproj(prev, 0);
777 prev = sched_prev_op(pos);
779 /* do not remat if the rematted value is defined immediately before this op */
780 if(prev == remat->op) {
785 /* this should be just fine, the following OP will be using this value, right? */
787 /* only remat AFTER the real definition of a value (?) */
788 if(!value_is_defined_before(si, pos, remat->value)) {
789 // ir_fprintf(stderr, "error(not defined)");
794 for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
795 const ir_node *arg = get_irn_n(op, n);
797 if(opt_no_enlarge_liveness) {
798 if(has_reg_class(si, arg) && live) {
799 res &= pset_find_ptr((pset*)live, arg)?1:0;
801 res &= value_is_defined_before(si, pos, arg);
804 res &= value_is_defined_before(si, pos, arg);
812 * Tells you whether a @p remat can be placed after the irn @p pos
815 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
818 pos = sched_block_first_nonphi(pos);
820 pos = sched_next_op(pos);
823 /* only remat AFTER the real definition of a value (?) */
824 if(!value_is_defined_before(si, pos, remat->value)) {
828 return can_remat_before(si, remat, pos, live);
832 * Collect potetially rematerializable OPs
835 walker_remat_collector(ir_node * irn, void * data)
837 spill_ilp_t *si = data;
839 if(!is_Block(irn) && !is_Phi(irn)) {
840 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
841 get_remats_from_op(si, irn);
846 * Inserts a copy of @p irn before @p pos
849 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
854 bb = is_Block(pos)?pos:get_nodes_block(pos);
855 copy = exact_copy(irn);
857 _set_phi_class(copy, NULL);
858 set_nodes_block(copy, bb);
859 sched_put_before(si, pos, copy);
865 * Inserts a copy of @p irn after @p pos
868 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
873 bb = is_Block(pos)?pos:get_nodes_block(pos);
874 copy = exact_copy(irn);
876 _set_phi_class(copy, NULL);
877 set_nodes_block(copy, bb);
878 sched_put_after(pos, copy);
884 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
888 if(can_remat_after(si, remat, pos, live)) {
893 DBG((si->dbg, LEVEL_3, "\t >inserting remat2 %+F\n", remat->op));
895 copy = insert_copy_after(si, remat->op, pos);
897 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
898 op = obstack_alloc(si->obst, sizeof(*op));
900 op->attr.remat.remat = remat;
901 op->attr.remat.pre = 0;
902 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
904 set_irn_link(copy, op);
905 pset_insert_ptr(si->all_possible_remats, copy);
907 proj_copy = insert_copy_after(si, remat->proj, copy);
908 set_irn_n(proj_copy, 0, copy);
909 set_irn_link(proj_copy, op);
910 pset_insert_ptr(si->all_possible_remats, proj_copy);
922 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
926 if(can_remat_before(si, remat, pos, live)) {
931 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
933 copy = insert_copy_before(si, remat->op, pos);
935 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
936 op = obstack_alloc(si->obst, sizeof(*op));
938 op->attr.remat.remat = remat;
939 op->attr.remat.pre = 1;
940 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
942 set_irn_link(copy, op);
943 pset_insert_ptr(si->all_possible_remats, copy);
945 proj_copy = insert_copy_after(si, remat->proj, copy);
946 set_irn_n(proj_copy, 0, copy);
947 set_irn_link(proj_copy, op);
948 pset_insert_ptr(si->all_possible_remats, proj_copy);
960 get_block_n_succs(const ir_node *block) {
961 const ir_edge_t *edge;
963 assert(edges_activated(current_ir_graph));
965 edge = get_block_succ_first(block);
969 edge = get_block_succ_next(block, edge);
974 is_start_block(const ir_node * bb)
976 return get_irg_start_block(get_irn_irg(bb)) == bb;
980 is_before_frame(const ir_node * bb, const ir_node * irn)
982 const ir_node *frame = get_irg_frame(get_irn_irg(bb));
984 if(is_start_block(bb) && sched_get_time_step(frame) >= sched_get_time_step(irn))
991 is_merge_edge(const ir_node * bb)
993 if(is_start_block(bb))
997 return get_block_n_succs(bb) == 1;
1003 is_diverge_edge(const ir_node * bb)
1005 if(is_start_block(bb))
1009 return get_Block_n_cfgpreds(bb) == 1;
1015 walker_regclass_copy_insertor(ir_node * irn, void * data)
1017 spill_ilp_t *si = data;
1019 if(is_Phi(irn) && has_reg_class(si, irn)) {
1022 for(n=get_irn_arity(irn)-1; n>=0; --n) {
1023 ir_node *phi_arg = get_irn_n(irn, n);
1024 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
1026 if(!has_reg_class(si, phi_arg)) {
1027 ir_node *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
1028 ir_node *pos = sched_block_last_noncf(si, bb);
1029 op_t *op = obstack_alloc(si->obst, sizeof(*op));
1031 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
1032 sched_add_after(pos, copy);
1033 set_irn_n(irn, n, copy);
1036 op->attr.live_range.args.reloads = NULL;
1037 op->attr.live_range.ilp = ILP_UNDEF;
1038 set_irn_link(copy, op);
1046 * Insert (so far unused) remats into the irg to
1047 * recompute the potential liveness of all values
1050 walker_remat_insertor(ir_node * bb, void * data)
1052 spill_ilp_t *si = data;
1053 spill_bb_t *spill_bb;
1056 pset *live = pset_new_ptr_default();
1058 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1060 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1061 ir_node *value = be_lv_get_irn(si->lv, bb, i);
1063 /* add remats at end of block */
1064 if (has_reg_class(si, value)) {
1065 pset_insert_ptr(live, value);
1069 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1070 set_irn_link(bb, spill_bb);
1072 irn = sched_last(bb);
1073 while(!sched_is_end(irn)) {
1080 next = sched_prev(irn);
1082 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
1084 if(is_Phi(irn) || is_Proj(irn)) {
1087 if(has_reg_class(si, irn)) {
1088 pset_remove_ptr(live, irn);
1091 op = obstack_alloc(si->obst, sizeof(*op));
1093 op->attr.live_range.args.reloads = NULL;
1094 op->attr.live_range.ilp = ILP_UNDEF;
1095 set_irn_link(irn, op);
1101 op = obstack_alloc(si->obst, sizeof(*op));
1103 op->attr.live_range.ilp = ILP_UNDEF;
1104 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1105 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1106 set_irn_link(irn, op);
1108 args = pset_new_ptr_default();
1110 /* collect arguments of op */
1111 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1112 ir_node *arg = get_irn_n(irn, n);
1114 pset_insert_ptr(args, arg);
1117 /* set args of op already live in epilog */
1118 pset_foreach(args, arg) {
1119 if(has_reg_class(si, arg)) {
1120 pset_insert_ptr(live, arg);
1123 /* delete defined value from live set */
1124 if(has_reg_class(si, irn)) {
1125 pset_remove_ptr(live, irn);
1129 remat_args = pset_new_ptr_default();
1131 /* insert all possible remats before irn */
1132 pset_foreach(args, arg) {
1133 remat_info_t *remat_info,
1137 /* continue if the operand has the wrong reg class
1139 if(!has_reg_class(si, arg))
1143 query.remats = NULL;
1144 query.remats_by_operand = NULL;
1145 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1151 if(remat_info->remats) {
1152 pset_foreach(remat_info->remats, remat) {
1153 ir_node *remat_irn = NULL;
1155 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
1156 if(opt_remat_while_live) {
1157 if(pset_find_ptr(live, remat->value)) {
1158 remat_irn = insert_remat_before(si, remat, irn, live);
1161 remat_irn = insert_remat_before(si, remat, irn, live);
1164 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1165 ir_node *remat_arg = get_irn_n(remat_irn, n);
1167 if(!has_reg_class(si, remat_arg)) continue;
1169 pset_insert_ptr(remat_args, remat_arg);
1176 /* now we add remat args to op's args because they could also die at this op */
1177 pset_foreach(args,arg) {
1178 if(pset_find_ptr(remat_args, arg)) {
1179 pset_remove_ptr(remat_args, arg);
1182 pset_foreach(remat_args,arg) {
1183 pset_insert_ptr(args, arg);
1186 /* insert all possible remats after irn */
1187 pset_foreach(args, arg) {
1188 remat_info_t *remat_info,
1192 /* continue if the operand has the wrong reg class */
1193 if(!has_reg_class(si, arg))
1197 query.remats = NULL;
1198 query.remats_by_operand = NULL;
1199 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1205 /* do not place post remats after jumps */
1206 if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
1208 if(remat_info->remats_by_operand) {
1209 pset_foreach(remat_info->remats_by_operand, remat) {
1210 /* do not insert remats producing the same value as one of the operands */
1211 if(!pset_find_ptr(args, remat->value)) {
1212 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
1213 if(opt_remat_while_live) {
1214 if(pset_find_ptr(live, remat->value)) {
1215 insert_remat_after(si, remat, irn, live);
1218 insert_remat_after(si, remat, irn, live);
1225 del_pset(remat_args);
1230 /* add remats at end if successor has multiple predecessors */
1231 if(is_merge_edge(bb)) {
1232 pset *live_out = pset_new_ptr_default();
1235 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1236 value = be_lv_get_irn(si->lv, bb, i);
1238 if (has_reg_class(si, value)) {
1239 pset_insert_ptr(live_out, value);
1243 /* add remats at end of block */
1244 pset_foreach(live_out, value) {
1245 remat_info_t *remat_info,
1250 query.remats = NULL;
1251 query.remats_by_operand = NULL;
1252 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1254 if(remat_info && remat_info->remats) {
1255 pset_foreach(remat_info->remats, remat) {
1256 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1258 insert_remat_before(si, remat, bb, live_out);
1265 if(is_diverge_edge(bb)) {
1266 pset *live_in = pset_new_ptr_default();
1269 be_lv_foreach(si->lv, bb, be_lv_state_in, i) {
1270 value = be_lv_get_irn(si->lv, bb, i);
1272 if(has_reg_class(si, value)) {
1273 pset_insert_ptr(live_in, value);
1276 sched_foreach(bb, value) {
1277 if(!is_Phi(value)) break;
1279 if(has_reg_class(si, value)) {
1280 pset_insert_ptr(live_in, value);
1284 /* add remat2s at beginning of block */
1285 pset_foreach(live_in, value) {
1286 remat_info_t *remat_info,
1291 query.remats = NULL;
1292 query.remats_by_operand = NULL;
1293 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1295 if(remat_info && remat_info->remats_by_operand) {
1296 pset_foreach(remat_info->remats_by_operand, remat) {
1297 DBG((si->dbg, LEVEL_4, "\t considering remat2 %+F at beginning of block %+F\n", remat->op, bb));
1299 /* put the remat here if all its args are available */
1300 insert_remat_after(si, remat, bb, live_in);
1310 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1313 luke_endwalker(ir_node * bb, void * data)
1315 spill_ilp_t *si = (spill_ilp_t*)data;
1321 spill_bb_t *spill_bb = get_irn_link(bb);
1325 live = pset_new_ptr_default();
1326 use_end = pset_new_ptr_default();
1328 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1329 irn = be_lv_get_irn(si->lv, bb, i);
1330 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1333 pset_insert_ptr(live, irn);
1334 op = get_irn_link(irn);
1335 assert(!op->is_remat);
1339 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1340 /* their reg_out must always be set */
1341 sched_foreach_reverse(bb, irn) {
1344 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1346 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1347 ir_node *irn_arg = get_irn_n(irn, n);
1349 if(has_reg_class(si, irn_arg)) {
1350 pset_insert_ptr(use_end, irn_arg);
1355 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1356 //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1357 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1359 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1361 /* if this is a merge edge we can reload at the end of this block */
1362 if(is_merge_edge(bb)) {
1363 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1364 } else if(pset_count(use_end)){
1365 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1367 spill_bb->reloads = NULL;
1370 pset_foreach(live,irn) {
1374 int default_spilled;
1377 /* handle values used by control flow nodes later separately */
1378 if(pset_find_ptr(use_end, irn)) continue;
1381 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1383 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1385 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1386 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1387 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1389 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1390 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1392 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1393 /* by default spill value right after definition */
1394 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1395 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1397 if(is_merge_edge(bb)) {
1401 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1402 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1403 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1405 /* reload <= mem_out */
1406 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1407 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1408 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1411 spill->reg_in = ILP_UNDEF;
1412 spill->mem_in = ILP_UNDEF;
1415 pset_foreach(use_end,irn) {
1419 ilp_cst_t end_use_req,
1422 int default_spilled;
1425 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1427 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1429 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1430 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1431 /* if irn is used at the end of the block, then it is live anyway */
1432 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1434 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1435 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1437 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1438 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1439 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1441 /* reload for use be control flow op */
1442 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1443 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1444 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1446 /* reload <= mem_out */
1447 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1448 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1449 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1451 spill->reg_in = ILP_UNDEF;
1452 spill->mem_in = ILP_UNDEF;
1454 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1455 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1456 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1464 next_post_remat(const ir_node * irn)
1470 next = sched_block_first_nonphi(irn);
1472 next = sched_next_op(irn);
1475 if(sched_is_end(next))
1478 op = get_irn_link(next);
1479 if(op->is_remat && !op->attr.remat.pre) {
1488 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1494 ret = sched_block_last_noncf(si, irn);
1495 ret = sched_next(ret);
1496 ret = sched_prev_op(ret);
1498 ret = sched_prev_op(irn);
1501 if(sched_is_end(ret) || is_Phi(ret))
1504 op = (op_t*)get_irn_link(ret);
1505 if(op->is_remat && op->attr.remat.pre) {
1513 * Find a remat of value @p value in the epilog of @p pos
1516 find_post_remat(const ir_node * value, const ir_node * pos)
1518 while((pos = next_post_remat(pos)) != NULL) {
1521 op = get_irn_link(pos);
1522 assert(op->is_remat && !op->attr.remat.pre);
1524 if(op->attr.remat.remat->value == value)
1525 return (ir_node*)pos;
1528 const ir_edge_t *edge;
1529 foreach_out_edge(pos, edge) {
1530 ir_node *proj = get_edge_src_irn(edge);
1531 assert(is_Proj(proj));
1541 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1543 spill_bb_t *spill_bb = get_irn_link(bb);
1547 int default_spilled;
1550 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1552 double spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1554 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1556 spill->reg_out = ILP_UNDEF;
1557 spill->reg_in = ILP_UNDEF;
1558 spill->mem_in = ILP_UNDEF;
1560 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1561 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1563 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1564 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1565 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1572 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1577 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1578 irn = be_lv_get_irn(si->lv, bb, i);
1580 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1581 pset_insert_ptr(live, irn);
1585 irn = sched_last(bb);
1587 /* all values eaten by control flow operations are also live until the end of the block */
1588 sched_foreach_reverse(bb, irn) {
1591 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1593 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1594 ir_node *arg = get_irn_n(irn,i);
1596 if(has_reg_class(si, arg)) {
1597 pset_insert_ptr(live, arg);
1604 * Inserts ILP-constraints and variables for memory copying before the given position
1607 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1609 const ir_node *succ;
1610 const ir_edge_t *edge;
1611 spill_bb_t *spill_bb = get_irn_link(block);
1620 assert(edges_activated(current_ir_graph));
1622 edge = get_block_succ_first(block);
1628 edge = get_block_succ_next(block, edge);
1629 /* next block can only contain phis, if this is a merge edge */
1632 ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1633 copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1635 ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1636 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1638 pset_foreach(live, tmp) {
1641 op_t *op = get_irn_link(irn);
1642 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1644 spill = set_find_spill(spill_bb->ilp, tmp);
1647 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1649 lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1651 sched_foreach(succ, phi) {
1652 const ir_node *to_copy;
1654 spill_t *to_copy_spill;
1655 op_t *phi_op = get_irn_link(phi);
1656 ilp_var_t reload = ILP_UNDEF;
1659 if(!is_Phi(phi)) break;
1660 if(!has_reg_class(si, phi)) continue;
1662 to_copy = get_irn_n(phi, pos);
1664 to_copy_op = get_irn_link(to_copy);
1666 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1667 assert(to_copy_spill);
1669 if(spill_bb->reloads) {
1670 keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1673 reload = PTR_TO_INT(keyval->val);
1677 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1678 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1680 /* copy - reg_out - reload - remat - live_range <= 0 */
1681 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1682 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1683 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1684 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1685 foreach_pre_remat(si, block, tmp) {
1686 op_t *remat_op = get_irn_link(tmp);
1687 if(remat_op->attr.remat.remat->value == to_copy) {
1688 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1692 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1693 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1695 /* copy - reg_out - copyreg <= 0 */
1696 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1697 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1698 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1704 * Walk all irg blocks and emit this ILP
1707 luke_blockwalker(ir_node * bb, void * data)
1709 spill_ilp_t *si = (spill_ilp_t*)data;
1714 spill_bb_t *spill_bb = get_irn_link(bb);
1717 pset *defs = pset_new_ptr_default();
1718 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
1721 live = pset_new_ptr_default();
1723 /****************************************
1724 * B A S I C B L O C K E N D
1725 ***************************************/
1728 /* init live values at end of block */
1729 get_live_end(si, bb, live);
1731 pset_foreach(live, irn) {
1733 ilp_var_t reload = ILP_UNDEF;
1735 spill = set_find_spill(spill_bb->ilp, irn);
1738 if(spill_bb->reloads) {
1739 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1742 reload = PTR_TO_INT(keyval->val);
1746 op = get_irn_link(irn);
1747 assert(!op->is_remat);
1749 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1750 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1751 op->attr.live_range.op = bb;
1753 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1754 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1756 /* reg_out - reload - remat - live_range <= 0 */
1757 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1758 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1759 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1760 foreach_pre_remat(si, bb, tmp) {
1761 op_t *remat_op = get_irn_link(tmp);
1762 if(remat_op->attr.remat.remat->value == irn) {
1763 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1766 /* maybe we should also assure that reg_out >= live_range etc. */
1770 insert_mem_copy_position(si, live, bb);
1772 /* allow only one argument to die at pre remat. If two value die check_pre does
1773 * not ensure a correct register pressure FIXME (verify this is really necessary!) */
1774 foreach_pre_remat(si, bb, tmp) {
1775 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
1778 op_t *remat_op = get_irn_link(tmp);
1780 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
1781 remat_arg = get_irn_n(tmp, n);
1783 if(has_reg_class(si, remat_arg)) {
1784 pset_insert_ptr(remat_args, remat_arg);
1788 if(pset_count(remat_args)) {
1789 /* \sum_args reg_out >= #args * remat - 1 */
1790 ir_snprintf(buf, sizeof(buf), "one_may_die_%N", tmp);
1791 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, -1.0);
1792 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -pset_count(remat_args));
1794 pset_foreach(remat_args, remat_arg) {
1795 spill = set_find_spill(spill_bb->ilp, remat_arg);
1798 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1803 del_pset(remat_args);
1807 * start new live ranges for values used by remats at end of block
1808 * and assure the remat args are available
1810 foreach_pre_remat(si, bb, tmp) {
1811 op_t *remat_op = get_irn_link(tmp);
1814 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1815 ir_node *remat_arg = get_irn_n(tmp, n);
1816 op_t *arg_op = get_irn_link(remat_arg);
1819 if(!has_reg_class(si, remat_arg)) continue;
1821 /* if value is becoming live through use by remat */
1822 if(!pset_find_ptr(live, remat_arg)) {
1823 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1824 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1826 arg_op->attr.live_range.ilp = prev_lr;
1827 arg_op->attr.live_range.op = bb;
1829 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1831 pset_insert_ptr(live, remat_arg);
1832 add_to_spill_bb(si, bb, remat_arg);
1835 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1836 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1837 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1839 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1840 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1842 /* use reload placed for this argument */
1843 if(spill_bb->reloads) {
1844 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1847 ilp_var_t reload = PTR_TO_INT(keyval->val);
1849 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1854 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1859 /**************************************
1860 * B A S I C B L O C K B O D Y
1861 **************************************/
1863 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1869 ilp_cst_t check_pre,
1875 ilp_cst_t one_memoperand;
1877 /* iterate only until first phi */
1881 op = get_irn_link(irn);
1883 if(op->is_remat) continue;
1884 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1886 /* collect defined values */
1887 if(has_reg_class(si, irn)) {
1888 pset_insert_ptr(defs, irn);
1892 if(is_Proj(irn)) continue;
1895 * init set of irn's arguments
1896 * and all possibly used values around this op
1897 * and values defined by post remats
1899 args = new_set(cmp_keyval, get_irn_arity(irn));
1900 used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1901 remat_defs = pset_new_ptr(pset_count(live));
1903 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1904 ir_node *irn_arg = get_irn_n(irn, n);
1905 if(has_reg_class(si, irn_arg)) {
1906 set_insert_keyval(args, irn_arg, (void*)n);
1907 pset_insert_ptr(used, irn_arg);
1910 foreach_post_remat(irn, tmp) {
1911 op_t *remat_op = get_irn_link(tmp);
1913 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1915 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1916 ir_node *remat_arg = get_irn_n(tmp, n);
1917 if(has_reg_class(si, remat_arg)) {
1918 pset_insert_ptr(used, remat_arg);
1922 foreach_pre_remat(si, irn, tmp) {
1923 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1924 ir_node *remat_arg = get_irn_n(tmp, n);
1925 if(has_reg_class(si, remat_arg)) {
1926 pset_insert_ptr(used, remat_arg);
1931 /**********************************
1932 * I N E P I L O G O F irn
1933 **********************************/
1935 /* ensure each dying value is used by only one post remat */
1936 pset_foreach(used, tmp) {
1937 ir_node *value = tmp;
1938 op_t *value_op = get_irn_link(value);
1943 foreach_post_remat(irn, remat) {
1944 op_t *remat_op = get_irn_link(remat);
1946 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1947 ir_node *remat_arg = get_irn_n(remat, n);
1949 /* if value is used by this remat add it to constraint */
1950 if(remat_arg == value) {
1952 /* sum remat2s <= 1 + n_remats*live_range */
1953 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1954 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
1958 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1964 if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
1965 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1969 /* ensure at least one value dies at post remat */
1970 foreach_post_remat(irn, tmp) {
1971 op_t *remat_op = get_irn_link(tmp);
1972 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
1975 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
1976 remat_arg = get_irn_n(tmp, n);
1978 if(has_reg_class(si, remat_arg)) {
1980 /* does arg always die at this op? */
1981 if(!pset_find_ptr(live, remat_arg))
1982 goto skip_one_must_die;
1984 pset_insert_ptr(remat_args, remat_arg);
1988 /* remat + \sum live_range(remat_arg) <= |args| */
1989 ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
1990 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
1991 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1993 pset_foreach(remat_args, remat_arg) {
1994 op_t *arg_op = get_irn_link(remat_arg);
1996 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2000 del_pset(remat_args);
2003 /* new live ranges for values from L\U defined by post remats */
2004 pset_foreach(live, tmp) {
2005 ir_node *value = tmp;
2006 op_t *value_op = get_irn_link(value);
2008 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
2009 ilp_var_t prev_lr = ILP_UNDEF;
2012 if(pset_find_ptr(remat_defs, value)) {
2014 /* next_live_range <= prev_live_range + sum remat2s */
2015 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
2016 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2018 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
2019 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2021 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
2022 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2024 foreach_post_remat(irn, remat) {
2025 op_t *remat_op = get_irn_link(remat);
2027 /* if value is being rematerialized by this remat */
2028 if(value == remat_op->attr.remat.remat->value) {
2029 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2033 value_op->attr.live_range.ilp = prev_lr;
2034 value_op->attr.live_range.op = irn;
2039 /* requirements for post remats and start live ranges from L/U' for values dying here */
2040 foreach_post_remat(irn, tmp) {
2041 op_t *remat_op = get_irn_link(tmp);
2044 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2045 ir_node *remat_arg = get_irn_n(tmp, n);
2046 op_t *arg_op = get_irn_link(remat_arg);
2048 if(!has_reg_class(si, remat_arg)) continue;
2050 /* only for values in L\U (TODO and D?), the others are handled with post_use */
2051 if(!pset_find_ptr(used, remat_arg)) {
2052 /* remat <= live_range(remat_arg) */
2053 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
2054 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2056 /* if value is becoming live through use by remat2 */
2057 if(!pset_find_ptr(live, remat_arg)) {
2060 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
2061 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2063 arg_op->attr.live_range.ilp = lr;
2064 arg_op->attr.live_range.op = irn;
2066 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
2068 pset_insert_ptr(live, remat_arg);
2069 add_to_spill_bb(si, bb, remat_arg);
2072 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2073 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2078 d = pset_count(defs);
2079 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
2081 /* count how many regs irn needs for arguments */
2082 u = set_count(args);
2085 /* check the register pressure in the epilog */
2086 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
2087 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
2088 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
2090 /* add L\U' to check_post */
2091 pset_foreach(live, tmp) {
2092 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
2093 /* if a live value is not used by irn */
2094 tmp_op = get_irn_link(tmp);
2095 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
2099 /* allow only one argument to die at pre remat. If two value die check_pre does
2100 * not ensure a correct register pressure FIXME (verify this is really necessary!) */
2101 foreach_pre_remat(si, irn, tmp) {
2102 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
2105 op_t *remat_op = get_irn_link(tmp);
2107 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2108 remat_arg = get_irn_n(tmp, n);
2110 if(has_reg_class(si, remat_arg)) {
2111 pset_insert_ptr(remat_args, remat_arg);
2115 if(pset_count(remat_args)) {
2116 /* \sum_args next(lr) >= #args * remat - 1 */
2117 ir_snprintf(buf, sizeof(buf), "one_may_die_%N", tmp);
2118 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, -1.0);
2119 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -pset_count(remat_args));
2121 pset_foreach(remat_args, remat_arg) {
2122 op_t *arg_op = get_irn_link(remat_arg);
2124 if(arg_op->attr.live_range.ilp != ILP_UNDEF) {
2125 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2130 del_pset(remat_args);
2133 /***********************************************************
2134 * 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
2135 **********************************************************/
2138 pset_foreach(used, tmp) {
2144 op_t *arg_op = get_irn_link(arg);
2147 spill = add_to_spill_bb(si, bb, arg);
2149 /* new live range for each used value */
2150 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2151 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, is_before_frame(bb, irn)?1.0:0.0);
2153 /* the epilog stuff - including post_use, check_post, check_post_remat */
2154 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2155 post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2157 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2159 /* arg is live throughout epilog if the next live_range is in a register */
2160 if(pset_find_ptr(live, arg)) {
2161 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
2163 /* post_use >= next_lr + remat */
2164 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2165 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2166 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2167 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2171 /* if value is not an arg of op and not possibly defined by post remat
2172 * then it may only die and not become live
2174 if(!set_find_keyval(args, arg)) {
2175 /* post_use <= prev_lr */
2176 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2177 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2178 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2179 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2181 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2182 /* next_lr <= prev_lr */
2183 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2184 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2185 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2186 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2191 /* forall post remat which use arg add a similar cst */
2192 foreach_post_remat(irn, remat) {
2195 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2196 ir_node *remat_arg = get_irn_n(remat, n);
2197 op_t *remat_op = get_irn_link(remat);
2199 if(remat_arg == arg) {
2200 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
2202 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2203 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2204 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2205 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2210 if(opt_memoperands) {
2211 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2212 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2213 ilp_var_t memoperand;
2215 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2216 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0);
2217 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2219 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2220 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2222 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2223 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2228 /* new live range begins for each used value */
2229 arg_op->attr.live_range.ilp = prev_lr;
2230 arg_op->attr.live_range.op = irn;
2232 pset_insert_ptr(live, arg);
2235 /* just to be sure */
2236 check_post = ILP_UNDEF;
2245 /* check the register pressure in the prolog */
2246 /* sum_{L\U} lr <= k - |U| */
2247 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2248 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2250 /* for the prolog remove defined values from the live set */
2251 pset_foreach(defs, tmp) {
2252 pset_remove_ptr(live, tmp);
2255 if(opt_memoperands) {
2256 ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2257 one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2260 /***********************************************************
2261 * 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
2262 **********************************************************/
2265 set_foreach(args, keyval) {
2267 const ir_node *arg = keyval->key;
2268 int i = PTR_TO_INT(keyval->val);
2269 op_t *arg_op = get_irn_link(arg);
2270 ilp_cst_t requirements;
2273 spill = set_find_spill(spill_bb->ilp, arg);
2276 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2277 op->attr.live_range.args.reloads[i] = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), is_before_frame(bb, irn)?0.0:1.0);
2279 /* reload <= mem_out */
2280 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2281 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2282 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2283 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2285 /* requirement: arg must be in register for use */
2286 /* reload + remat + live_range == 1 */
2287 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2288 requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2290 lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2291 lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2292 foreach_pre_remat(si, irn, tmp) {
2293 op_t *remat_op = get_irn_link(tmp);
2294 if(remat_op->attr.remat.remat->value == arg) {
2295 lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2299 if(opt_memoperands) {
2301 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2302 if(get_irn_n(irn, n) == arg) {
2306 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2307 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2308 memoperand_t *memoperand;
2309 memoperand = set_find_memoperand(si->memoperands, irn, n);
2311 /* memoperand <= mem_out */
2312 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2313 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2314 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2315 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2317 /* the memoperand is only sufficient if it is used once by the op */
2318 if(n_memoperands == 1)
2319 lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2321 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2323 /* we have one more free register if we use a memory operand */
2324 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2330 /* iterate over L\U */
2331 pset_foreach(live, tmp) {
2332 if(!set_find_keyval(args, tmp)) {
2333 /* if a live value is not used by irn */
2334 tmp_op = get_irn_link(tmp);
2335 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2340 /* requirements for remats */
2341 foreach_pre_remat(si, irn, tmp) {
2342 op_t *remat_op = get_irn_link(tmp);
2345 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2346 ir_node *remat_arg = get_irn_n(tmp, n);
2347 op_t *arg_op = get_irn_link(remat_arg);
2349 if(!has_reg_class(si, remat_arg)) continue;
2351 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2352 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2353 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2355 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2356 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2358 /* if remat arg is also used by current op then we can use reload placed for this argument */
2359 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2360 int index = (int)keyval->val;
2362 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2370 /*************************
2371 * D O N E W I T H O P
2372 *************************/
2374 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2376 pset_foreach(live, tmp) {
2377 assert(has_reg_class(si, tmp));
2380 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2381 ir_node *arg = get_irn_n(irn, n);
2383 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2386 del_pset(remat_defs);
2390 defs = pset_new_ptr_default();
2395 /***************************************
2396 * B E G I N N I N G O F B L O C K
2397 ***************************************/
2400 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2401 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2403 pset_foreach(live, irn) {
2404 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2407 /* construct mem_outs for all values */
2409 set_foreach(spill_bb->ilp, spill) {
2410 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2411 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2413 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2414 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2416 if(pset_find_ptr(live, spill->irn)) {
2417 int default_spilled;
2418 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2420 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2421 default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2422 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2423 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2425 if(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2427 op_t *op = get_irn_link(spill->irn);
2429 /* do we have to copy a phi argument? */
2430 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2431 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2433 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2434 const ir_node *arg = get_irn_n(spill->irn, n);
2440 /* argument already done? */
2441 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2443 /* get sum of execution frequencies of blocks with the same phi argument */
2444 for(m=n; m>=0; --m) {
2445 const ir_node *arg2 = get_irn_n(spill->irn, m);
2448 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2452 /* copies are not for free */
2453 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2454 var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
2456 for(m=n; m>=0; --m) {
2457 const ir_node *arg2 = get_irn_n(spill->irn, m);
2460 op->attr.live_range.args.copies[m] = var;
2465 /* copy <= mem_in */
2466 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2467 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2468 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2469 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2476 foreach_post_remat(bb, tmp) {
2478 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
2479 op_t *remat_op = get_irn_link(tmp);
2482 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2483 remat_arg = get_irn_n(tmp, n);
2484 if(has_reg_class(si, remat_arg)) {
2485 pset_insert_ptr(remat_args, remat_arg);
2488 assert(pset_count(remat_args) > 0 && "post remats should have at least one arg");
2490 /* remat + \sum live_range(remat_arg) <= |args| */
2491 ir_snprintf(buf, sizeof(buf), "one_must_die_%N", tmp);
2492 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2493 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2495 pset_foreach(remat_args, remat_arg) {
2496 /* if value is becoming live through use by remat2 */
2497 if(!pset_find_ptr(live, remat_arg)) {
2498 op_t *remat_arg_op = get_irn_link(remat_arg);
2501 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 at bb start %+F\n", remat_arg, tmp));
2503 pset_insert_ptr(live, remat_arg);
2504 spill = add_to_spill_bb(si, bb, remat_arg);
2505 remat_arg_op->attr.live_range.ilp = ILP_UNDEF;
2507 /* we need reg_in and mem_in for this value; they will be referenced later */
2508 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", remat_arg, bb);
2509 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2510 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", remat_arg, bb);
2511 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2514 /* optimization: all memory stuff should be 0, for we do not want to insert reloads for remats */
2515 ir_snprintf(buf, sizeof(buf), "nomem_%N_%N", remat_arg, bb);
2516 nomem = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0.0);
2518 lpp_set_factor_fast(si->lpp, nomem, spill->spill, 1.0);
2519 if(spill_bb->reloads) {
2520 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
2523 ilp_var_t reload = PTR_TO_INT(keyval->val);
2524 lpp_set_factor_fast(si->lpp, nomem, reload, 1.0);
2528 op_t *remat_arg_op = get_irn_link(remat_arg);
2529 lpp_set_factor_fast(si->lpp, cst, remat_arg_op->attr.live_range.ilp, 1.0);
2532 del_pset(remat_args);
2535 /* L\U is empty at bb start */
2536 /* arg is live throughout epilog if it is reg_in into this block */
2538 /* check the register pressure at the beginning of the block
2541 /* reg_in entspricht post_use */
2543 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2544 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2546 pset_foreach(live, irn) {
2549 spill = set_find_spill(spill_bb->ilp, irn);
2552 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2553 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2555 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2557 /* spill + mem_in <= 1 */
2558 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2559 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2561 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2562 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2564 } /* post_remats are NOT included in register pressure check because
2565 they do not increase regpressure */
2567 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2568 pset_foreach(live, irn) {
2572 spill = set_find_spill(spill_bb->ilp, irn);
2573 assert(spill && spill->irn == irn);
2575 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2576 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2579 ir_node *phi_arg = get_Phi_pred(irn, n);
2580 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2581 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2583 op_t *op = get_irn_link(irn);
2585 /* although the phi is in the right regclass one or more of
2586 * its arguments can be in a different one or at least to
2589 if(has_reg_class(si, phi_arg)) {
2590 /* mem_in < mem_out_arg + copy */
2591 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2592 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2594 /* reg_in < reg_out_arg */
2595 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2596 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2598 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2599 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2601 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2604 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2606 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2608 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2612 /* else assure the value arrives on all paths in the same resource */
2614 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2617 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2618 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2621 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2622 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2623 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2624 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2626 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2627 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2629 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2632 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2633 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2638 foreach_post_remat(bb, tmp) {
2641 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2642 ir_node *remat_arg = get_irn_n(tmp, n);
2643 op_t *remat_op = get_irn_link(tmp);
2645 if(!has_reg_class(si, remat_arg)) continue;
2647 spill = set_find_spill(spill_bb->ilp, remat_arg);
2650 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2651 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2652 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2653 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2657 pset_foreach(live, irn) {
2658 const op_t *op = get_irn_link(irn);
2659 const ir_node *remat;
2664 foreach_post_remat(bb, remat) {
2667 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2668 const ir_node *arg = get_irn_n(remat, n);
2671 const op_t *remat_op = get_irn_link(remat);
2673 if(cst == ILP_UNDEF) {
2674 /* sum remat2s <= 1 + n_remats*live_range */
2675 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", irn, bb);
2676 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2678 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2684 if(cst != ILP_UNDEF && op->attr.live_range.ilp != ILP_UNDEF) {
2685 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -n_remats);
2689 /* first live ranges from reg_ins */
2690 pset_foreach(live, irn) {
2691 op_t *op = get_irn_link(irn);
2693 if(op->attr.live_range.ilp != ILP_UNDEF) {
2695 spill = set_find_spill(spill_bb->ilp, irn);
2696 assert(spill && spill->irn == irn);
2698 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2699 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2700 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2701 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2703 foreach_post_remat(bb, tmp) {
2704 op_t *remat_op = get_irn_link(tmp);
2706 if(remat_op->attr.remat.remat->value == irn) {
2707 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2713 /* walk forward now and compute constraints for placing spills */
2714 /* this must only be done for values that are not defined in this block */
2715 /* TODO are these values at start of block? if yes, just check whether this is a diverge edge and skip the loop */
2716 pset_foreach(live, irn) {
2718 * if value is defined in this block we can anways place the spill directly after the def
2719 * -> no constraint necessary
2721 if(!is_Phi(irn) && get_nodes_block(irn) == bb) continue;
2724 spill = set_find_spill(spill_bb->ilp, irn);
2727 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2728 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2730 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2731 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2734 sched_foreach_op(bb, tmp) {
2735 op_t *op = get_irn_link(tmp);
2737 if(is_Phi(tmp)) continue;
2738 assert(!is_Proj(tmp));
2741 const ir_node *value = op->attr.remat.remat->value;
2744 /* only collect remats up to the first real use of a value */
2745 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2750 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2751 ir_node *arg = get_irn_n(tmp, n);
2754 /* if a value is used stop collecting remats */
2767 typedef struct _irnlist_t {
2768 struct list_head list;
2772 typedef struct _interference_t {
2773 struct list_head blocklist;
2779 cmp_interference(const void *a, const void *b, size_t size)
2781 const interference_t *p = a;
2782 const interference_t *q = b;
2784 return !(p->a == q->a && p->b == q->b);
2787 static interference_t *
2788 set_find_interference(set * set, ir_node * a, ir_node * b)
2790 interference_t query;
2792 query.a = (a>b)?a:b;
2793 query.b = (a>b)?b:a;
2795 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2798 static interference_t *
2799 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2801 interference_t query,
2803 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2807 result = set_find_interference(set, a, b);
2810 list_add(&list->list, &result->blocklist);
2814 query.a = (a>b)?a:b;
2815 query.b = (a>b)?b:a;
2817 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2819 INIT_LIST_HEAD(&result->blocklist);
2820 list_add(&list->list, &result->blocklist);
2826 values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
2828 const ir_edge_t *edge;
2830 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2831 /* both values are live in, so they interfere */
2835 /* ensure a dominates b */
2836 if(value_dominates(b,a)) {
2842 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2845 /* the following code is stolen from bera.c */
2846 if(be_is_live_end(si->lv, bb, a))
2849 foreach_out_edge(a, edge) {
2850 const ir_node *user = edge->src;
2851 if(get_nodes_block(user) == bb
2854 && !pset_find_ptr(si->inverse_ops, user)
2855 && value_dominates(b, user))
2863 * Walk all irg blocks and collect interfering values inside of phi classes
2866 luke_interferencewalker(ir_node * bb, void * data)
2868 spill_ilp_t *si = (spill_ilp_t*)data;
2871 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2872 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
2873 op_t *a_op = get_irn_link(a);
2876 /* a is only interesting if it is in my register class and if it is inside a phi class */
2877 if (has_reg_class(si, a) && get_phi_class(a)) {
2878 if(a_op->is_remat || pset_find_ptr(si->inverse_ops, a))
2881 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)) {
2882 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
2883 op_t *b_op = get_irn_link(b);
2886 /* a and b are only interesting if they are in the same phi class */
2887 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2888 if(b_op->is_remat || pset_find_ptr(si->inverse_ops, b))
2891 if(values_interfere_in_block(si, bb, a, b)) {
2892 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2893 set_insert_interference(si, si->interferences, a, b, bb);
2901 static unsigned int copy_path_id = 0;
2904 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2911 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2912 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2914 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2916 pset_foreach(copies, ptr) {
2917 copy = PTR_TO_INT(ptr);
2918 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2923 * @parameter copies contains a path of copies which lead us to irn
2924 * @parameter visited contains a set of nodes already visited on this path
2927 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2929 const ir_edge_t *edge;
2930 op_t *op = get_irn_link(irn);
2931 pset *visited_users = pset_new_ptr_default();
2934 if(op->is_remat) return 0;
2936 pset_insert_ptr(visited, irn);
2940 pset *visited_operands = pset_new_ptr(get_irn_arity(irn));
2942 /* visit all operands */
2943 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2944 ir_node *arg = get_irn_n(irn, n);
2945 ilp_var_t copy = op->attr.live_range.args.copies[n];
2947 if(!has_reg_class(si, arg)) continue;
2948 if(pset_find_ptr(visited_operands, arg)) continue;
2949 pset_insert_ptr(visited_operands, arg);
2952 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2953 del_pset(visited_operands);
2954 del_pset(visited_users);
2955 pset_remove_ptr(visited, irn);
2958 pset_insert(copies, INT_TO_PTR(copy), copy);
2959 write_copy_path_cst(si, copies, any_interfere);
2960 pset_remove(copies, INT_TO_PTR(copy), copy);
2961 } else if(!pset_find_ptr(visited, arg)) {
2962 pset_insert(copies, INT_TO_PTR(copy), copy);
2963 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
2964 pset_remove(copies, INT_TO_PTR(copy), copy);
2966 if(paths > MAX_PATHS) {
2967 if(pset_count(copies) == 0) {
2971 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2972 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2973 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2974 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2975 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2979 del_pset(visited_operands);
2980 del_pset(visited_users);
2981 pset_remove_ptr(visited, irn);
2984 } else if(pset_count(copies) == 0) {
2990 del_pset(visited_operands);
2993 /* visit all uses which are phis */
2994 foreach_out_edge(irn, edge) {
2995 ir_node *user = edge->src;
2996 int pos = edge->pos;
2997 op_t *op = get_irn_link(user);
3000 if(!is_Phi(user)) continue;
3001 if(!has_reg_class(si, user)) continue;
3002 if(pset_find_ptr(visited_users, user)) continue;
3003 pset_insert_ptr(visited_users, user);
3005 copy = op->attr.live_range.args.copies[pos];
3007 if(user == target) {
3008 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3009 del_pset(visited_users);
3010 pset_remove_ptr(visited, irn);
3013 pset_insert(copies, INT_TO_PTR(copy), copy);
3014 write_copy_path_cst(si, copies, any_interfere);
3015 pset_remove(copies, INT_TO_PTR(copy), copy);
3016 } else if(!pset_find_ptr(visited, user)) {
3017 pset_insert(copies, INT_TO_PTR(copy), copy);
3018 paths += find_copy_path(si, user, target, any_interfere, copies, visited);
3019 pset_remove(copies, INT_TO_PTR(copy), copy);
3021 if(paths > MAX_PATHS) {
3022 if(pset_count(copies) == 0) {
3026 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3027 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3028 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3029 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3030 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3034 del_pset(visited_users);
3035 pset_remove_ptr(visited, irn);
3038 } else if(pset_count(copies) == 0) {
3044 del_pset(visited_users);
3045 pset_remove_ptr(visited, irn);
3050 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
3052 pset * copies = pset_new_ptr_default();
3053 pset * visited = pset_new_ptr_default();
3055 find_copy_path(si, a, b, any_interfere, copies, visited);
3063 memcopyhandler(spill_ilp_t * si)
3065 interference_t *interference;
3067 /* teste Speicherwerte auf Interferenz */
3069 /* analyze phi classes */
3070 phi_class_compute(si->chordal_env->irg);
3072 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
3073 irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
3075 /* now lets emit the ILP unequations for the crap */
3076 set_foreach(si->interferences, interference) {
3078 ilp_var_t interfere,
3080 ilp_cst_t any_interfere_cst,
3082 const ir_node *a = interference->a;
3083 const ir_node *b = interference->b;
3085 /* any_interf <= \sum interf */
3086 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
3087 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3088 any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3090 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
3092 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
3093 const ir_node *bb = irnlist->irn;
3094 spill_bb_t *spill_bb = get_irn_link(bb);
3099 spilla = set_find_spill(spill_bb->ilp, a);
3102 spillb = set_find_spill(spill_bb->ilp, b);
3105 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
3106 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
3107 /* 2: - mem_in_a - spill_a + interfere <= 0 */
3108 /* 3: - mem_in_b - spill_b + interfere <= 0 */
3109 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
3110 interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3112 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
3113 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
3115 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
3116 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
3117 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
3118 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
3119 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
3121 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
3122 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3124 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3125 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
3126 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
3128 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
3129 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3131 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3132 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
3133 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
3136 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
3138 /* any_interfere >= interf */
3139 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
3140 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3142 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3143 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3146 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
3147 gen_copy_constraints(si,a,b,any_interfere);
3155 return fabs(x) < 0.00001;
3158 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
3160 spill_ilp_t *si = get_irg_link(current_ir_graph);
3162 if(pset_find_ptr(si->all_possible_remats, n)) {
3163 op_t *op = (op_t*)get_irn_link(n);
3164 assert(op && op->is_remat);
3166 if(!op->attr.remat.remat->inverse) {
3167 if(op->attr.remat.pre) {
3168 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
3170 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
3175 op_t *op = (op_t*)get_irn_link(n);
3176 assert(op && op->is_remat);
3178 if(op->attr.remat.pre) {
3179 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
3181 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
3192 dump_graph_with_remats(ir_graph * irg, const char * suffix)
3194 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
3195 be_dump(irg, suffix, dump_ir_block_graph_sched);
3196 set_dump_node_vcgattr_hook(NULL);
3200 * Edge hook to dump the schedule edges with annotated register pressure.
3203 sched_pressure_edge_hook(FILE *F, ir_node *irn)
3205 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
3206 ir_node *prev = sched_prev(irn);
3207 fprintf(F, "edge:{sourcename:\"");
3209 fprintf(F, "\" targetname:\"");
3211 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
3212 fprintf(F, "\" color:magenta}\n");
3218 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
3220 DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3222 dump_consts_local(0);
3223 set_dump_node_edge_hook(sched_pressure_edge_hook);
3224 dump_ir_block_graph(irg, suffix);
3225 set_dump_node_edge_hook(old_edge_hook);
3229 walker_pressure_annotator(ir_node * bb, void * data)
3231 spill_ilp_t *si = data;
3234 pset *live = pset_new_ptr_default();
3237 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3238 irn = be_lv_get_irn(si->lv, bb, i);
3240 if (has_reg_class(si, irn)) {
3241 pset_insert_ptr(live, irn);
3245 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3247 sched_foreach_reverse(bb, irn) {
3249 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3253 if(has_reg_class(si, irn)) {
3254 pset_remove_ptr(live, irn);
3255 if(is_Proj(irn)) ++projs;
3258 if(!is_Proj(irn)) projs = 0;
3260 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3261 ir_node *arg = get_irn_n(irn, n);
3263 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3265 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3272 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3274 be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
3278 connect_all_remats_with_keep(spill_ilp_t * si)
3286 n_remats = pset_count(si->all_possible_remats);
3288 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3291 pset_foreach(si->all_possible_remats, irn) {
3296 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
3298 obstack_free(si->obst, ins);
3303 connect_all_spills_with_keep(spill_ilp_t * si)
3312 n_spills = pset_count(si->spills);
3314 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3317 pset_foreach(si->spills, irn) {
3322 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
3324 obstack_free(si->obst, ins);
3328 /** insert a spill at an arbitrary position */
3329 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3331 ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
3332 ir_graph *irg = get_irn_irg(bl);
3333 ir_node *frame = get_irg_frame(irg);
3337 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
3338 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3340 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3343 * search the right insertion point. a spill of a phi cannot be put
3344 * directly after the phi, if there are some phis behind the one which
3345 * is spilled. Also, a spill of a Proj must be after all Projs of the
3348 * Here's one special case:
3349 * If the spill is in the start block, the spill must be after the frame
3350 * pointer is set up. This is done by setting insert to the end of the block
3351 * which is its default initialization (see above).
3354 if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3357 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3360 sched_add_after(insert, spill);
3365 delete_remat(spill_ilp_t * si, ir_node * remat) {
3367 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3369 sched_remove(remat);
3371 /* kill links to operands */
3372 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3373 set_irn_n(remat, n, bad);
3378 clean_remat_info(spill_ilp_t * si)
3382 remat_info_t *remat_info;
3383 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3385 set_foreach(si->remat_info, remat_info) {
3386 if(!remat_info->remats) continue;
3388 pset_foreach(remat_info->remats, remat)
3390 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3391 set_irn_n((ir_node*)remat->proj, -1, bad);
3392 set_irn_n((ir_node*)remat->proj, 0, bad);
3395 if(get_irn_n_edges(remat->op) == 0) {
3396 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3397 set_irn_n((ir_node*)remat->op, n, bad);
3402 if(remat_info->remats) del_pset(remat_info->remats);
3403 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3408 delete_unnecessary_remats(spill_ilp_t * si)
3410 if(opt_keep_alive & KEEPALIVE_REMATS) {
3412 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3415 // ir_node *end = get_irg_end(si->chordal_env->irg);
3418 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3419 ir_node *keep_arg = get_irn_n(si->keep, n);
3420 op_t *arg_op = get_irn_link(keep_arg);
3423 assert(arg_op->is_remat);
3425 name = si->lpp->vars[arg_op->attr.remat.ilp];
3427 if(is_zero(name->value)) {
3428 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
3429 /* TODO check whether reload is preferred over remat (could be bug) */
3430 delete_remat(si, keep_arg);
3432 if(!arg_op->attr.remat.remat->inverse) {
3433 if(arg_op->attr.remat.pre) {
3434 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3436 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3439 if(arg_op->attr.remat.pre) {
3440 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3442 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3447 set_irn_n(si->keep, n, bad);
3450 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
3451 ir_node *end_arg = get_End_keepalive(end, i);
3453 if(end_arg != si->keep) {
3454 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
3457 keeps = obstack_finish(si->obst);
3458 set_End_keepalives(end, n-1, keeps);
3459 obstack_free(si->obst, keeps);
3462 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
3467 pset_foreach(si->all_possible_remats, remat) {
3468 op_t *remat_op = get_irn_link(remat);
3469 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
3471 if(is_zero(name->value)) {
3472 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
3473 /* TODO check whether reload is preferred over remat (could be bug) */
3474 delete_remat(si, remat);
3476 if(!remat_op->attr.remat.remat->inverse) {
3477 if(remat_op->attr.remat.pre) {
3478 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3480 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3483 if(remat_op->attr.remat.pre) {
3484 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3486 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3495 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3497 pset *spills = pset_new_ptr_default();
3499 const ir_node *next;
3502 defs = set_find_def(si->values, value);
3504 if(defs && defs->spills) {
3505 for(next = defs->spills; next; next = get_irn_link(next)) {
3506 pset_insert_ptr(spills, next);
3514 * @param before The node after which the spill will be placed in the schedule
3517 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3521 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3523 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3525 spill = be_spill2(arch_env, irn, before);
3527 defs = set_insert_def(si->values, value);
3530 /* enter into the linked list */
3531 set_irn_link(spill, defs->spills);
3532 defs->spills = spill;
3534 if(opt_keep_alive & KEEPALIVE_SPILLS)
3535 pset_insert_ptr(si->spills, spill);
3541 * @param before The Phi node which has to be spilled
3544 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3551 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3553 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3554 ins[n] = si->m_unknown;
3557 mem_phi = new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3559 defs = set_insert_def(si->values, phi);
3562 /* enter into the linked list */
3563 set_irn_link(mem_phi, defs->spills);
3564 defs->spills = mem_phi;
3566 #ifdef SCHEDULE_PHIM
3567 sched_add_after(phi, mem_phi);
3570 if(opt_keep_alive & KEEPALIVE_SPILLS)
3571 pset_insert_ptr(si->spills, mem_phi);
3578 * Add remat to list of defs, destroys link field!
3581 insert_remat(spill_ilp_t * si, ir_node * remat)
3584 op_t *remat_op = get_irn_link(remat);
3586 assert(remat_op->is_remat);
3588 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3591 /* enter into the linked list */
3592 set_irn_link(remat, defs->remats);
3593 defs->remats = remat;
3598 * Add reload before operation and add to list of defs
3601 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3606 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3608 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3610 defs = set_find_def(si->values, value);
3612 spill = defs->spills;
3613 assert(spill && "no spill placed before reload");
3615 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3617 /* enter into the linked list */
3618 set_irn_link(reload, defs->remats);
3619 defs->remats = reload;
3624 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3627 ir_node *value = get_irn_n(memoperand->irn, memoperand->pos);
3629 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3631 DBG((si->dbg, LEVEL_2, "\t inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3633 defs = set_find_def(si->values, value);
3635 spill = defs->spills;
3636 assert(spill && "no spill placed before reload");
3638 arch_perform_memory_operand(arch_env, memoperand->irn, spill, memoperand->pos);
3641 void insert_memoperands(spill_ilp_t * si)
3643 memoperand_t *memoperand;
3646 set_foreach(si->memoperands, memoperand) {
3647 name = si->lpp->vars[memoperand->ilp];
3648 if(!is_zero(name->value)) {
3649 perform_memory_operand(si, memoperand);
3655 walker_spill_placer(ir_node * bb, void * data) {
3656 spill_ilp_t *si = (spill_ilp_t*)data;
3658 spill_bb_t *spill_bb = get_irn_link(bb);
3659 pset *spills_to_do = pset_new_ptr_default();
3662 set_foreach(spill_bb->ilp, spill) {
3665 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3666 name = si->lpp->vars[spill->mem_in];
3667 if(!is_zero(name->value)) {
3670 mem_phi = insert_mem_phi(si, spill->irn);
3672 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3676 name = si->lpp->vars[spill->spill];
3677 if(!is_zero(name->value)) {
3678 /* place spill directly after definition */
3679 if(get_nodes_block(spill->irn) == bb) {
3680 insert_spill(si, spill->irn, spill->irn, spill->irn);
3684 /* place spill at bb start */
3685 if(spill->reg_in > 0) {
3686 name = si->lpp->vars[spill->reg_in];
3687 if(!is_zero(name->value)) {
3688 insert_spill(si, spill->irn, spill->irn, bb);
3692 /* place spill after a remat */
3693 pset_insert_ptr(spills_to_do, spill->irn);
3696 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3699 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3700 op_t *op = get_irn_link(irn);
3702 if(be_is_Spill(irn)) continue;
3705 /* TODO fix this if we want to support remats with more than two nodes */
3706 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3707 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3709 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3712 if(pset_find_ptr(spills_to_do, irn)) {
3713 pset_remove_ptr(spills_to_do, irn);
3715 insert_spill(si, irn, irn, irn);
3721 assert(pset_count(spills_to_do) == 0);
3723 /* afterwards free data in block */
3724 del_pset(spills_to_do);
3728 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3730 ir_node *insert_pos = bb;
3732 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3734 /* find last definition of arg value in block */
3739 defs = set_find_def(si->values, value);
3741 if(defs && defs->remats) {
3742 for(next = defs->remats; next; next = get_irn_link(next)) {
3743 if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3744 last = sched_get_time_step(next);
3750 if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3751 last = sched_get_time_step(value);
3755 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
3757 spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3763 phim_fixer(spill_ilp_t *si) {
3766 set_foreach(si->values, defs) {
3767 const ir_node *phi = defs->value;
3768 op_t *op = get_irn_link(phi);
3769 ir_node *phi_m = NULL;
3770 ir_node *next = defs->spills;
3773 if(!is_Phi(phi)) continue;
3776 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3780 next = get_irn_link(next);
3783 if(!phi_m) continue;
3785 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3786 ir_node *value = get_irn_n(phi, n);
3787 defs_t *val_defs = set_find_def(si->values, value);
3789 /* a spill of this value */
3794 ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3795 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3797 if(!is_zero(name->value)) {
3798 spill = insert_mem_copy(si, pred, value);
3800 spill = val_defs->spills;
3803 spill = val_defs->spills;
3806 assert(spill && "no spill placed before PhiM");
3807 set_irn_n(phi_m, n, spill);
3813 walker_reload_placer(ir_node * bb, void * data) {
3814 spill_ilp_t *si = (spill_ilp_t*)data;
3816 spill_bb_t *spill_bb = get_irn_link(bb);
3818 /* reloads at end of block */
3819 if(spill_bb->reloads) {
3822 set_foreach(spill_bb->reloads, keyval) {
3823 ir_node *irn = (ir_node*)keyval->key;
3824 ilp_var_t reload = PTR_TO_INT(keyval->val);
3827 name = si->lpp->vars[reload];
3828 if(!is_zero(name->value)) {
3830 ir_node *insert_pos = bb;
3831 ir_node *prev = sched_block_last_noncf(si, bb);
3832 op_t *prev_op = get_irn_link(prev);
3834 while(be_is_Spill(prev)) {
3835 prev = sched_prev(prev);
3838 prev_op = get_irn_link(prev);
3840 /* insert reload before pre-remats */
3841 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3842 && prev_op->is_remat && prev_op->attr.remat.pre) {
3846 prev = sched_prev(prev);
3847 } while(be_is_Spill(prev));
3849 prev_op = get_irn_link(prev);
3853 reload = insert_reload(si, irn, insert_pos);
3855 if(opt_keep_alive & KEEPALIVE_RELOADS)
3856 pset_insert_ptr(si->spills, reload);
3861 /* walk and insert more reloads and collect remats */
3862 sched_foreach_reverse(bb, irn) {
3863 op_t *op = get_irn_link(irn);
3865 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3866 if(is_Phi(irn)) break;
3869 if(get_irn_mode(irn) != mode_T) {
3870 insert_remat(si, irn);
3875 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3876 ir_node *arg = get_irn_n(irn, n);
3878 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3881 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3882 if(!is_zero(name->value)) {
3884 ir_node *insert_pos = irn;
3885 ir_node *prev = sched_prev(insert_pos);
3888 while(be_is_Spill(prev)) {
3889 prev = sched_prev(prev);
3892 prev_op = get_irn_link(prev);
3894 /* insert reload before pre-remats */
3895 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3896 && prev_op->is_remat && prev_op->attr.remat.pre) {
3900 prev = sched_prev(prev);
3901 } while(be_is_Spill(prev));
3903 prev_op = get_irn_link(prev);
3907 reload = insert_reload(si, arg, insert_pos);
3909 set_irn_n(irn, n, reload);
3911 if(opt_keep_alive & KEEPALIVE_RELOADS)
3912 pset_insert_ptr(si->spills, reload);
3919 del_set(spill_bb->ilp);
3920 if(spill_bb->reloads) del_set(spill_bb->reloads);
3924 walker_collect_used(ir_node * irn, void * data)
3926 lc_bitset_t *used = data;
3928 lc_bitset_set(used, get_irn_idx(irn));
3931 struct kill_helper {
3937 walker_kill_unused(ir_node * bb, void * data)
3939 struct kill_helper *kh = data;
3940 ir_node *bad = get_irg_bad(get_irn_irg(bb));
3944 for(irn=sched_first(bb); !sched_is_end(irn);) {
3945 ir_node *next = sched_next(irn);
3948 if(!lc_bitset_is_set(kh->used, get_irn_idx(irn))) {
3949 if(be_is_Spill(irn) || be_is_Reload(irn)) {
3950 DBG((kh->si->dbg, LEVEL_1, "\t SUBOPTIMAL! %+F IS UNUSED (cost: %g)\n", irn, get_cost(kh->si, irn)*execution_frequency(kh->si, bb)));
3952 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3958 set_nodes_block(irn, bad);
3959 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3960 set_irn_n(irn, n, bad);
3968 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3970 struct kill_helper kh;
3972 kh.used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3975 irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3976 irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3978 lc_bitset_free(kh.used);
3982 print_irn_pset(pset * p)
3986 pset_foreach(p, irn) {
3987 ir_printf("%+F\n", irn);
3992 dump_phi_class(spill_ilp_t * si, pset * phiclass, const char * file)
3994 FILE *f = fopen(file, "w");
3996 interference_t *interference;
3998 pset_break(phiclass);
3999 set_break(si->interferences);
4001 ir_fprintf(f, "digraph phiclass {\n");
4003 pset_foreach(phiclass, irn) {
4005 ir_fprintf(f, " %F%N [shape=box]\n",irn,irn);
4008 pset_foreach(phiclass, irn) {
4011 if(!is_Phi(irn)) continue;
4013 for(n=get_irn_arity(irn)-1; n>=0; --n) {
4014 ir_node *arg = get_irn_n(irn, n);
4016 ir_fprintf(f, " %F%N -> %F%N\n",irn,irn,arg,arg);
4020 set_foreach(si->interferences, interference) {
4021 const ir_node *a = interference->a;
4022 const ir_node *b = interference->b;
4023 if(get_phi_class(a) == phiclass) {
4024 ir_fprintf(f, " %F%N -> %F%N [color=red,dir=none,style=bold]\n",a,a,b,b);
4033 rewire_uses(spill_ilp_t * si)
4035 dom_front_info_t *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
4037 pset *ignore = pset_new_ptr(1);
4039 pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
4041 /* then fix uses of spills */
4042 set_foreach(si->values, defs) {
4045 const ir_node *next = defs->remats;
4048 reloads = pset_new_ptr_default();
4051 if(be_is_Reload(next)) {
4052 pset_insert_ptr(reloads, next);
4056 next = get_irn_link(next);
4059 spills = get_spills_for_value(si, defs->value);
4060 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));
4061 if(pset_count(spills) > 1) {
4062 //assert(pset_count(reloads) > 0);
4063 // print_irn_pset(spills);
4064 // print_irn_pset(reloads);
4066 be_ssa_constr_set_ignore(dfi, si->lv, spills, ignore);
4073 /* first fix uses of remats and reloads */
4074 set_foreach(si->values, defs) {
4076 const ir_node *next = defs->remats;
4079 nodes = pset_new_ptr_default();
4080 pset_insert_ptr(nodes, defs->value);
4083 pset_insert_ptr(nodes, next);
4084 next = get_irn_link(next);
4087 if(pset_count(nodes) > 1) {
4088 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
4089 be_ssa_constr_set(dfi, si->lv, nodes);
4096 // remove_unused_defs(si);
4098 be_free_dominance_frontiers(dfi);
4103 writeback_results(spill_ilp_t * si)
4105 /* walk through the graph and collect all spills, reloads and remats for a value */
4107 si->values = new_set(cmp_defs, 4096);
4109 DBG((si->dbg, LEVEL_1, "Applying results\n"));
4110 delete_unnecessary_remats(si);
4111 si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
4112 irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
4113 irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
4115 insert_memoperands(si);
4118 /* clean the remat info! there are still back-edges leading there! */
4119 clean_remat_info(si);
4123 connect_all_spills_with_keep(si);
4125 del_set(si->values);
4129 get_n_regs(spill_ilp_t * si)
4131 int arch_n_regs = arch_register_class_n_regs(si->cls);
4135 for(i=0; i<arch_n_regs; i++) {
4136 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
4141 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
4146 walker_reload_mover(ir_node * bb, void * data)
4148 spill_ilp_t *si = data;
4151 sched_foreach(bb, tmp) {
4152 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
4153 ir_node *reload = tmp;
4156 /* move reload upwards */
4158 int pressure = (int)get_irn_link(reload);
4159 if(pressure < si->n_regs) {
4160 irn = sched_prev(reload);
4161 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
4162 sched_remove(reload);
4163 pressure = (int)get_irn_link(irn);
4165 while(pressure < si->n_regs) {
4166 if( sched_is_end(irn) ||
4167 (be_is_Reload(irn) && has_reg_class(si, irn)) ||
4168 /* do not move reload before its spill */
4169 (irn == be_get_Reload_mem(reload)) ||
4170 /* do not move before phi */
4173 set_irn_link(irn, INT_TO_PTR(pressure+1));
4174 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
4175 irn = sched_prev(irn);
4177 pressure = (int)get_irn_link(irn);
4180 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
4181 sched_put_after(irn, reload);
4188 move_reloads_upward(spill_ilp_t * si)
4190 irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
4195 * Walk all irg blocks and check for interfering spills inside of phi classes
4198 luke_meminterferencechecker(ir_node * bb, void * data)
4200 spill_ilp_t *si = (spill_ilp_t*)data;
4203 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
4204 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
4206 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
4208 /* a is only interesting if it is in my register class and if it is inside a phi class */
4209 if (has_reg_class(si, a) && get_phi_class(a)) {
4210 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)) {
4211 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
4213 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
4215 /* a and b are only interesting if they are in the same phi class */
4216 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
4217 if(values_interfere_in_block(si, bb, a, b)) {
4218 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4227 verify_phiclasses(spill_ilp_t * si)
4229 /* analyze phi classes */
4230 phi_class_compute(si->chordal_env->irg);
4232 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4233 irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
4237 be_spill_remat(const be_chordal_env_t * chordal_env)
4240 char problem_name[256];
4241 char dump_suffix[256];
4242 char dump_suffix2[256];
4243 struct obstack obst;
4246 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
4247 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
4248 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
4250 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4251 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4253 if(opt_verify & VERIFY_DOMINANCE)
4254 be_check_dominance(chordal_env->irg);
4256 obstack_init(&obst);
4257 si.chordal_env = chordal_env;
4259 si.cls = chordal_env->cls;
4260 si.lpp = new_lpp(problem_name, lpp_minimize);
4261 si.remat_info = new_set(cmp_remat_info, 4096);
4262 si.interferences = new_set(cmp_interference, 32);
4263 si.memoperands = new_set(cmp_memoperands, 128);
4264 si.all_possible_remats = pset_new_ptr_default();
4265 si.spills = pset_new_ptr_default();
4266 si.inverse_ops = pset_new_ptr_default();
4267 si.lv = chordal_env->lv;
4269 si.n_regs = get_n_regs(&si);
4271 set_irg_link(chordal_env->irg, &si);
4272 compute_doms(chordal_env->irg);
4274 /* compute phi classes */
4275 // phi_class_compute(chordal_env->irg);
4277 be_analyze_regpressure(chordal_env, "-pre");
4280 /* collect remats */
4281 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4282 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
4285 /* insert possible remats */
4286 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4287 irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
4288 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4290 if(opt_keep_alive & KEEPALIVE_REMATS) {
4291 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4292 connect_all_remats_with_keep(&si);
4293 /* dump graph with inserted remats */
4294 dump_graph_with_remats(chordal_env->irg, dump_suffix);
4297 /* insert copies for phi arguments not in my regclass */
4298 irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
4300 /* recompute liveness */
4301 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4302 be_liveness_recompute(si.lv);
4306 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4307 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4308 irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
4310 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4311 irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
4314 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4315 memcopyhandler(&si);
4318 if(opt_dump_flags & DUMP_PROBLEM) {
4320 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4321 if ((f = fopen(buf, "wt")) != NULL) {
4322 lpp_dump_plain(si.lpp, f);
4327 if(opt_dump_flags & DUMP_MPS) {
4330 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4331 if((f = fopen(buf, "wt")) != NULL) {
4332 mps_write_mps(si.lpp, s_mps_fixed, f);
4336 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4337 if((f = fopen(buf, "wt")) != NULL) {
4338 mps_write_mst(si.lpp, s_mps_fixed, f);
4343 lpp_check_startvals(si.lpp);
4346 DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4347 lpp_set_time_limit(si.lpp, opt_timeout);
4350 lpp_set_log(si.lpp, stdout);
4353 lpp_solve_cplex(si.lpp);
4355 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4357 assert(lpp_is_sol_valid(si.lpp)
4358 && "solution of ILP must be valid");
4360 DBG((si.dbg, LEVEL_1, "\t%s: iterations: %d, solution time: %g, objective function: %g\n", problem_name, si.lpp->iterations, si.lpp->sol_time, is_zero(si.lpp->objval)?0.0:si.lpp->objval));
4362 if(opt_dump_flags & DUMP_SOLUTION) {
4366 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4367 if ((f = fopen(buf, "wt")) != NULL) {
4369 for (i = 0; i < si.lpp->var_next; ++i) {
4370 lpp_name_t *name = si.lpp->vars[i];
4371 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4377 writeback_results(&si);
4381 kill_all_unused_values_in_schedule(&si);
4383 if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
4384 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
4386 // move reloads upwards
4387 be_liveness_recompute(si.lv);
4388 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4389 move_reloads_upward(&si);
4392 verify_phiclasses(&si);
4395 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4397 dump_pressure_graph(&si, dump_suffix2);
4399 be_analyze_regpressure(chordal_env, "-post");
4401 if(opt_verify & VERIFY_DOMINANCE)
4402 be_check_dominance(chordal_env->irg);
4404 free_dom(chordal_env->irg);
4405 del_set(si.interferences);
4406 del_pset(si.inverse_ops);
4407 del_pset(si.all_possible_remats);
4408 del_set(si.memoperands);
4409 del_pset(si.spills);
4411 obstack_free(&obst, NULL);
4412 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4415 #else /* WITH_ILP */
4418 only_that_you_can_compile_without_WITH_ILP_defined(void)
4422 #endif /* WITH_ILP */