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 ifg before, after or after each cloud", &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_irn_n(op, 0);
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)
603 unsigned int ret = 0;
605 for(n=get_irn_arity(irn)-1; n>=0; --n) {
606 const ir_node *arg = get_irn_n(irn, n);
608 if(has_reg_class(si, arg)) ++ret;
615 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
620 if( has_reg_class(si, op)
621 && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1)
622 && (opt_remats != REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0)
624 remat = get_remat_from_op(si, op, op);
626 add_remat(si, remat);
630 if(opt_remats == REMATS_ALL) {
631 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
633 for (n = get_irn_arity(op)-1; n>=0; --n) {
634 ir_node *arg = get_irn_n(op, n);
636 if(has_reg_class(si, arg)) {
637 /* try to get an inverse remat */
638 remat = get_remat_from_op(si, arg, op);
640 add_remat(si, remat);
648 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
651 ir_node *def_block = get_nodes_block(val);
657 /* if pos is at end of a basic block */
659 ret = (pos == def_block || block_dominates(def_block, pos));
660 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
664 /* else if this is a normal operation */
665 block = get_nodes_block(pos);
666 if(block == def_block) {
667 if(!sched_is_scheduled(val)) return 1;
669 ret = sched_comes_after(val, pos);
670 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
674 ret = block_dominates(def_block, block);
675 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
679 static INLINE ir_node *
680 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
682 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
686 * Returns first non-Phi node of block @p bb
688 static INLINE ir_node *
689 sched_block_first_nonphi(const ir_node * bb)
691 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
695 sched_skip_proj_predicator(const ir_node * irn, void * data)
697 return (is_Proj(irn));
700 static INLINE ir_node *
701 sched_next_nonproj(const ir_node * irn, int forward)
703 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
707 * Returns next operation node (non-Proj) after @p irn
708 * or the basic block of this node
710 static INLINE ir_node *
711 sched_next_op(const ir_node * irn)
713 ir_node *next = sched_next(irn);
718 return sched_next_nonproj(next, 1);
722 * Returns previous operation node (non-Proj) before @p irn
723 * or the basic block of this node
725 static INLINE ir_node *
726 sched_prev_op(const ir_node * irn)
728 ir_node *prev = sched_prev(irn);
733 return sched_next_nonproj(prev, 0);
737 sched_put_after(ir_node * insert, ir_node * irn)
739 if(is_Block(insert)) {
740 insert = sched_block_first_nonphi(insert);
742 insert = sched_next_op(insert);
744 sched_add_before(insert, irn);
748 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
750 if(is_Block(insert)) {
751 insert = sched_block_last_noncf(si, insert);
753 insert = sched_next_nonproj(insert, 0);
754 insert = sched_prev(insert);
756 sched_add_after(insert, irn);
760 * Tells you whether a @p remat can be placed before the irn @p pos
763 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
765 const ir_node *op = remat->op;
771 prev = sched_block_last_noncf(si, pos);
772 prev = sched_next_nonproj(prev, 0);
774 prev = sched_prev_op(pos);
776 /* do not remat if the rematted value is defined immediately before this op */
777 if(prev == remat->op) {
782 /* this should be just fine, the following OP will be using this value, right? */
784 /* only remat AFTER the real definition of a value (?) */
785 if(!value_is_defined_before(si, pos, remat->value)) {
786 // ir_fprintf(stderr, "error(not defined)");
791 for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
792 const ir_node *arg = get_irn_n(op, n);
794 if(opt_no_enlarge_liveness) {
795 if(has_reg_class(si, arg) && live) {
796 res &= pset_find_ptr((pset*)live, arg)?1:0;
798 res &= value_is_defined_before(si, pos, arg);
801 res &= value_is_defined_before(si, pos, arg);
809 * Tells you whether a @p remat can be placed after the irn @p pos
812 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
815 pos = sched_block_first_nonphi(pos);
817 pos = sched_next_op(pos);
820 /* only remat AFTER the real definition of a value (?) */
821 if(!value_is_defined_before(si, pos, remat->value)) {
825 return can_remat_before(si, remat, pos, live);
829 * Collect potetially rematerializable OPs
832 walker_remat_collector(ir_node * irn, void * data)
834 spill_ilp_t *si = data;
836 if(!is_Block(irn) && !is_Phi(irn)) {
837 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
838 get_remats_from_op(si, irn);
843 * Inserts a copy of @p irn before @p pos
846 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
851 bb = is_Block(pos)?pos:get_nodes_block(pos);
852 copy = exact_copy(irn);
854 _set_phi_class(copy, NULL);
855 set_nodes_block(copy, bb);
856 sched_put_before(si, pos, copy);
862 * Inserts a copy of @p irn after @p pos
865 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
870 bb = is_Block(pos)?pos:get_nodes_block(pos);
871 copy = exact_copy(irn);
873 _set_phi_class(copy, NULL);
874 set_nodes_block(copy, bb);
875 sched_put_after(pos, copy);
881 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
885 if(can_remat_after(si, remat, pos, live)) {
890 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
892 copy = insert_copy_after(si, remat->op, pos);
894 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
895 op = obstack_alloc(si->obst, sizeof(*op));
897 op->attr.remat.remat = remat;
898 op->attr.remat.pre = 0;
899 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
901 set_irn_link(copy, op);
902 pset_insert_ptr(si->all_possible_remats, copy);
904 proj_copy = insert_copy_after(si, remat->proj, copy);
905 set_irn_n(proj_copy, 0, copy);
906 set_irn_link(proj_copy, op);
907 pset_insert_ptr(si->all_possible_remats, proj_copy);
919 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
923 if(can_remat_before(si, remat, pos, live)) {
928 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
930 copy = insert_copy_before(si, remat->op, pos);
932 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
933 op = obstack_alloc(si->obst, sizeof(*op));
935 op->attr.remat.remat = remat;
936 op->attr.remat.pre = 1;
937 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
939 set_irn_link(copy, op);
940 pset_insert_ptr(si->all_possible_remats, copy);
942 proj_copy = insert_copy_after(si, remat->proj, copy);
943 set_irn_n(proj_copy, 0, copy);
944 set_irn_link(proj_copy, op);
945 pset_insert_ptr(si->all_possible_remats, proj_copy);
957 get_block_n_succs(const ir_node *block) {
958 const ir_edge_t *edge;
960 assert(edges_activated(current_ir_graph));
962 edge = get_block_succ_first(block);
966 edge = get_block_succ_next(block, edge);
971 is_merge_edge(const ir_node * bb)
974 return get_block_n_succs(bb) == 1;
980 is_diverge_edge(const ir_node * bb)
983 return get_Block_n_cfgpreds(bb) == 1;
989 walker_regclass_copy_insertor(ir_node * irn, void * data)
991 spill_ilp_t *si = data;
993 if(is_Phi(irn) && has_reg_class(si, irn)) {
996 for(n=get_irn_arity(irn)-1; n>=0; --n) {
997 ir_node *phi_arg = get_irn_n(irn, n);
998 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
1000 if(!has_reg_class(si, phi_arg)) {
1001 ir_node *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
1002 ir_node *pos = sched_block_last_noncf(si, bb);
1003 op_t *op = obstack_alloc(si->obst, sizeof(*op));
1005 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
1006 sched_add_after(pos, copy);
1007 set_irn_n(irn, n, copy);
1010 op->attr.live_range.args.reloads = NULL;
1011 op->attr.live_range.ilp = ILP_UNDEF;
1012 set_irn_link(copy, op);
1020 * Insert (so far unused) remats into the irg to
1021 * recompute the potential liveness of all values
1024 walker_remat_insertor(ir_node * bb, void * data)
1026 spill_ilp_t *si = data;
1027 spill_bb_t *spill_bb;
1030 pset *live = pset_new_ptr_default();
1032 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1034 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1035 ir_node *value = be_lv_get_irn(si->lv, bb, i);
1037 /* add remats at end of block */
1038 if (has_reg_class(si, value)) {
1039 pset_insert_ptr(live, value);
1043 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1044 set_irn_link(bb, spill_bb);
1046 irn = sched_last(bb);
1047 while(!sched_is_end(irn)) {
1054 next = sched_prev(irn);
1056 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
1058 if(is_Phi(irn) || is_Proj(irn)) {
1061 if(has_reg_class(si, irn)) {
1062 pset_remove_ptr(live, irn);
1065 op = obstack_alloc(si->obst, sizeof(*op));
1067 op->attr.live_range.args.reloads = NULL;
1068 op->attr.live_range.ilp = ILP_UNDEF;
1069 set_irn_link(irn, op);
1075 op = obstack_alloc(si->obst, sizeof(*op));
1077 op->attr.live_range.ilp = ILP_UNDEF;
1078 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1079 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1080 set_irn_link(irn, op);
1082 args = pset_new_ptr_default();
1084 /* collect arguments of op */
1085 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1086 ir_node *arg = get_irn_n(irn, n);
1088 pset_insert_ptr(args, arg);
1091 /* set args of op already live in epilog */
1092 pset_foreach(args, arg) {
1093 if(has_reg_class(si, arg)) {
1094 pset_insert_ptr(live, arg);
1097 /* delete defined value from live set */
1098 if(has_reg_class(si, irn)) {
1099 pset_remove_ptr(live, irn);
1103 remat_args = pset_new_ptr_default();
1105 /* insert all possible remats before irn */
1106 pset_foreach(args, arg) {
1107 remat_info_t *remat_info,
1111 /* continue if the operand has the wrong reg class
1113 if(!has_reg_class(si, arg))
1117 query.remats = NULL;
1118 query.remats_by_operand = NULL;
1119 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1125 if(remat_info->remats) {
1126 pset_foreach(remat_info->remats, remat) {
1127 ir_node *remat_irn = NULL;
1129 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
1130 if(opt_remat_while_live) {
1131 if(pset_find_ptr(live, remat->value)) {
1132 remat_irn = insert_remat_before(si, remat, irn, live);
1135 remat_irn = insert_remat_before(si, remat, irn, live);
1138 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1139 ir_node *remat_arg = get_irn_n(remat_irn, n);
1141 if(!has_reg_class(si, remat_arg)) continue;
1143 pset_insert_ptr(remat_args, remat_arg);
1150 /* now we add remat args to op's args because they could also die at this op */
1151 pset_foreach(args,arg) {
1152 if(pset_find_ptr(remat_args, arg)) {
1153 pset_remove_ptr(remat_args, arg);
1156 pset_foreach(remat_args,arg) {
1157 pset_insert_ptr(args, arg);
1160 /* insert all possible remats after irn */
1161 pset_foreach(args, arg) {
1162 remat_info_t *remat_info,
1166 /* continue if the operand has the wrong reg class */
1167 if(!has_reg_class(si, arg))
1171 query.remats = NULL;
1172 query.remats_by_operand = NULL;
1173 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1179 /* do not place post remats after jumps */
1180 if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
1182 if(remat_info->remats_by_operand) {
1183 pset_foreach(remat_info->remats_by_operand, remat) {
1184 /* do not insert remats producing the same value as one of the operands */
1185 if(!pset_find_ptr(args, remat->value)) {
1186 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
1187 if(opt_remat_while_live) {
1188 if(pset_find_ptr(live, remat->value)) {
1189 insert_remat_after(si, remat, irn, live);
1192 insert_remat_after(si, remat, irn, live);
1199 del_pset(remat_args);
1204 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_in, i) {
1205 ir_node *value = be_lv_get_irn(si->lv, bb, i);
1207 /* add remats at end if successor has multiple predecessors */
1208 if(is_merge_edge(bb)) {
1209 /* add remats at end of block */
1210 if (be_is_live_end(si->lv, bb, value) && has_reg_class(si, value)) {
1211 remat_info_t *remat_info,
1216 query.remats = NULL;
1217 query.remats_by_operand = NULL;
1218 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1220 if(remat_info && remat_info->remats) {
1221 pset_foreach(remat_info->remats, remat) {
1222 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1224 insert_remat_before(si, remat, bb, NULL);
1229 if(is_diverge_edge(bb)) {
1230 /* add remat2s at beginning of block */
1231 if ((be_is_live_in(si->lv, bb, value) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) {
1232 remat_info_t *remat_info,
1237 query.remats = NULL;
1238 query.remats_by_operand = NULL;
1239 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1241 if(remat_info && remat_info->remats) {
1242 pset_foreach(remat_info->remats, remat) {
1243 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at beginning of block %+F\n", remat->op, bb));
1245 /* put the remat here if all its args are available */
1246 insert_remat_after(si, remat, bb, NULL);
1256 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1259 luke_endwalker(ir_node * bb, void * data)
1261 spill_ilp_t *si = (spill_ilp_t*)data;
1267 spill_bb_t *spill_bb = get_irn_link(bb);
1271 live = pset_new_ptr_default();
1272 use_end = pset_new_ptr_default();
1274 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1275 irn = be_lv_get_irn(si->lv, bb, i);
1276 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1279 pset_insert_ptr(live, irn);
1280 op = get_irn_link(irn);
1281 assert(!op->is_remat);
1285 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1286 /* their reg_out must always be set */
1287 sched_foreach_reverse(bb, irn) {
1290 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1292 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1293 ir_node *irn_arg = get_irn_n(irn, n);
1295 if(has_reg_class(si, irn_arg)) {
1296 pset_insert_ptr(use_end, irn_arg);
1301 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1302 //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1303 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1305 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1307 /* if this is a merge edge we can reload at the end of this block */
1308 if(is_merge_edge(bb)) {
1309 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1310 } else if(pset_count(use_end)){
1311 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1313 spill_bb->reloads = NULL;
1316 pset_foreach(live,irn) {
1320 int default_spilled;
1323 /* handle values used by control flow nodes later separately */
1324 if(pset_find_ptr(use_end, irn)) continue;
1327 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1329 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1331 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1332 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1333 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1335 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1336 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1338 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1339 /* by default spill value right after definition */
1340 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1341 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1343 if(is_merge_edge(bb)) {
1347 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1348 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1349 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1351 /* reload <= mem_out */
1352 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1353 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1354 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1357 spill->reg_in = ILP_UNDEF;
1358 spill->mem_in = ILP_UNDEF;
1361 pset_foreach(use_end,irn) {
1365 ilp_cst_t end_use_req,
1368 int default_spilled;
1371 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1373 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1375 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1376 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1377 /* if irn is used at the end of the block, then it is live anyway */
1378 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1380 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1381 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1383 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1384 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1385 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1387 /* reload for use be control flow op */
1388 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1389 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1390 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1392 /* reload <= mem_out */
1393 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1394 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1395 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1397 spill->reg_in = ILP_UNDEF;
1398 spill->mem_in = ILP_UNDEF;
1400 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1401 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1402 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1410 next_post_remat(const ir_node * irn)
1416 next = sched_block_first_nonphi(irn);
1418 next = sched_next_op(irn);
1421 if(sched_is_end(next))
1424 op = get_irn_link(next);
1425 if(op->is_remat && !op->attr.remat.pre) {
1434 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1440 ret = sched_block_last_noncf(si, irn);
1441 ret = sched_next(ret);
1442 ret = sched_prev_op(ret);
1444 ret = sched_prev_op(irn);
1447 if(sched_is_end(ret) || is_Phi(ret))
1450 op = (op_t*)get_irn_link(ret);
1451 if(op->is_remat && op->attr.remat.pre) {
1459 * Find a remat of value @p value in the epilog of @p pos
1462 find_post_remat(const ir_node * value, const ir_node * pos)
1464 while((pos = next_post_remat(pos)) != NULL) {
1467 op = get_irn_link(pos);
1468 assert(op->is_remat && !op->attr.remat.pre);
1470 if(op->attr.remat.remat->value == value)
1471 return (ir_node*)pos;
1474 const ir_edge_t *edge;
1475 foreach_out_edge(pos, edge) {
1476 ir_node *proj = get_edge_src_irn(edge);
1477 assert(is_Proj(proj));
1487 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1489 spill_bb_t *spill_bb = get_irn_link(bb);
1493 int default_spilled;
1496 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1498 double spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1500 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1502 spill->reg_out = ILP_UNDEF;
1503 spill->reg_in = ILP_UNDEF;
1504 spill->mem_in = ILP_UNDEF;
1506 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1507 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1509 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1510 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1511 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1518 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1523 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1524 irn = be_lv_get_irn(si->lv, bb, i);
1526 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1527 pset_insert_ptr(live, irn);
1531 irn = sched_last(bb);
1533 /* all values eaten by control flow operations are also live until the end of the block */
1534 sched_foreach_reverse(bb, irn) {
1537 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1539 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1540 ir_node *arg = get_irn_n(irn,i);
1542 if(has_reg_class(si, arg)) {
1543 pset_insert_ptr(live, arg);
1550 * Inserts ILP-constraints and variables for memory copying before the given position
1553 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1555 const ir_node *succ;
1556 const ir_edge_t *edge;
1557 spill_bb_t *spill_bb = get_irn_link(block);
1566 assert(edges_activated(current_ir_graph));
1568 edge = get_block_succ_first(block);
1574 edge = get_block_succ_next(block, edge);
1575 /* next block can only contain phis, if this is a merge edge */
1578 ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1579 copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1581 ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1582 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1584 pset_foreach(live, tmp) {
1587 op_t *op = get_irn_link(irn);
1588 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1590 spill = set_find_spill(spill_bb->ilp, tmp);
1593 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1595 lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1597 sched_foreach(succ, phi) {
1598 const ir_node *to_copy;
1600 spill_t *to_copy_spill;
1601 op_t *phi_op = get_irn_link(phi);
1602 ilp_var_t reload = ILP_UNDEF;
1605 if(!is_Phi(phi)) break;
1606 if(!has_reg_class(si, phi)) continue;
1608 to_copy = get_irn_n(phi, pos);
1610 to_copy_op = get_irn_link(to_copy);
1612 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1613 assert(to_copy_spill);
1615 if(spill_bb->reloads) {
1616 keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1619 reload = PTR_TO_INT(keyval->val);
1623 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1624 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1626 /* copy - reg_out - reload - remat - live_range <= 0 */
1627 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1628 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1629 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1630 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1631 foreach_pre_remat(si, block, tmp) {
1632 op_t *remat_op = get_irn_link(tmp);
1633 if(remat_op->attr.remat.remat->value == to_copy) {
1634 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1638 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1639 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1641 /* copy - reg_out - copyreg <= 0 */
1642 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1643 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1644 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1650 * Walk all irg blocks and emit this ILP
1653 luke_blockwalker(ir_node * bb, void * data)
1655 spill_ilp_t *si = (spill_ilp_t*)data;
1660 spill_bb_t *spill_bb = get_irn_link(bb);
1663 pset *defs = pset_new_ptr_default();
1664 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
1667 live = pset_new_ptr_default();
1669 /****************************************
1670 * B A S I C B L O C K E N D
1671 ***************************************/
1674 /* init live values at end of block */
1675 get_live_end(si, bb, live);
1677 pset_foreach(live, irn) {
1679 ilp_var_t reload = ILP_UNDEF;
1681 spill = set_find_spill(spill_bb->ilp, irn);
1684 if(spill_bb->reloads) {
1685 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1688 reload = PTR_TO_INT(keyval->val);
1692 op = get_irn_link(irn);
1693 assert(!op->is_remat);
1695 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1696 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1697 op->attr.live_range.op = bb;
1699 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1700 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1702 /* reg_out - reload - remat - live_range <= 0 */
1703 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1704 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1705 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1706 foreach_pre_remat(si, bb, tmp) {
1707 op_t *remat_op = get_irn_link(tmp);
1708 if(remat_op->attr.remat.remat->value == irn) {
1709 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1712 /* maybe we should also assure that reg_out >= live_range etc. */
1716 insert_mem_copy_position(si, live, bb);
1719 * start new live ranges for values used by remats at end of block
1720 * and assure the remat args are available
1722 foreach_pre_remat(si, bb, tmp) {
1723 op_t *remat_op = get_irn_link(tmp);
1726 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1727 ir_node *remat_arg = get_irn_n(tmp, n);
1728 op_t *arg_op = get_irn_link(remat_arg);
1731 if(!has_reg_class(si, remat_arg)) continue;
1733 /* if value is becoming live through use by remat */
1734 if(!pset_find_ptr(live, remat_arg)) {
1735 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1736 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1738 arg_op->attr.live_range.ilp = prev_lr;
1739 arg_op->attr.live_range.op = bb;
1741 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1743 pset_insert_ptr(live, remat_arg);
1744 add_to_spill_bb(si, bb, remat_arg);
1747 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1748 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1749 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1751 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1752 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1754 /* use reload placed for this argument */
1755 if(spill_bb->reloads) {
1756 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1759 ilp_var_t reload = PTR_TO_INT(keyval->val);
1761 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1766 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1771 /**************************************
1772 * B A S I C B L O C K B O D Y
1773 **************************************/
1775 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1781 ilp_cst_t check_pre,
1787 ilp_cst_t one_memoperand;
1789 /* iterate only until first phi */
1793 op = get_irn_link(irn);
1795 if(op->is_remat) continue;
1796 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1798 /* collect defined values */
1799 if(has_reg_class(si, irn)) {
1800 pset_insert_ptr(defs, irn);
1804 if(is_Proj(irn)) continue;
1807 * init set of irn's arguments
1808 * and all possibly used values around this op
1809 * and values defined by post remats
1811 args = new_set(cmp_keyval, get_irn_arity(irn));
1812 used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1813 remat_defs = pset_new_ptr(pset_count(live));
1815 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1816 ir_node *irn_arg = get_irn_n(irn, n);
1817 if(has_reg_class(si, irn_arg)) {
1818 set_insert_keyval(args, irn_arg, (void*)n);
1819 pset_insert_ptr(used, irn_arg);
1822 foreach_post_remat(irn, tmp) {
1823 op_t *remat_op = get_irn_link(tmp);
1825 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1827 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1828 ir_node *remat_arg = get_irn_n(tmp, n);
1829 if(has_reg_class(si, remat_arg)) {
1830 pset_insert_ptr(used, remat_arg);
1834 foreach_pre_remat(si, irn, tmp) {
1835 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1836 ir_node *remat_arg = get_irn_n(tmp, n);
1837 if(has_reg_class(si, remat_arg)) {
1838 pset_insert_ptr(used, remat_arg);
1843 /**********************************
1844 * I N E P I L O G O F irn
1845 **********************************/
1847 /* ensure each dying value is used by only one post remat */
1848 pset_foreach(used, tmp) {
1849 ir_node *value = tmp;
1850 op_t *value_op = get_irn_link(value);
1855 foreach_post_remat(irn, remat) {
1856 op_t *remat_op = get_irn_link(remat);
1858 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1859 ir_node *remat_arg = get_irn_n(remat, n);
1861 /* if value is used by this remat add it to constraint */
1862 if(remat_arg == value) {
1864 /* sum remat2s <= 1 + n_remats*live_range */
1865 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1866 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
1870 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1876 // value_op->attr.live_range.ilp != ILP_UNDEF
1877 if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
1878 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1882 /* ensure at least one value dies at post remat */
1883 foreach_post_remat(irn, tmp) {
1884 op_t *remat_op = get_irn_link(tmp);
1885 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
1888 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
1889 remat_arg = get_irn_n(tmp, n);
1891 if(has_reg_class(si, remat_arg)) {
1893 /* does arg always die at this op? */
1894 if(!pset_find_ptr(live, remat_arg))
1895 goto skip_one_must_die;
1897 pset_insert_ptr(remat_args, remat_arg);
1901 /* remat + \sum live_range(remat_arg) <= |args| */
1902 ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
1903 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
1904 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1906 pset_foreach(remat_args, remat_arg) {
1907 op_t *arg_op = get_irn_link(remat_arg);
1909 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1913 del_pset(remat_args);
1916 /* new live ranges for values from L\U defined by post remats */
1917 pset_foreach(live, tmp) {
1918 ir_node *value = tmp;
1919 op_t *value_op = get_irn_link(value);
1921 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
1922 ilp_var_t prev_lr = ILP_UNDEF;
1925 if(pset_find_ptr(remat_defs, value)) {
1927 /* next_live_range <= prev_live_range + sum remat2s */
1928 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1929 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1931 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1932 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1934 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1935 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1937 foreach_post_remat(irn, remat) {
1938 op_t *remat_op = get_irn_link(remat);
1940 /* if value is being rematerialized by this remat */
1941 if(value == remat_op->attr.remat.remat->value) {
1942 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1946 value_op->attr.live_range.ilp = prev_lr;
1947 value_op->attr.live_range.op = irn;
1952 /* requirements for post remats and start live ranges from L/U' for values dying here */
1953 foreach_post_remat(irn, tmp) {
1954 op_t *remat_op = get_irn_link(tmp);
1957 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1958 ir_node *remat_arg = get_irn_n(tmp, n);
1959 op_t *arg_op = get_irn_link(remat_arg);
1961 if(!has_reg_class(si, remat_arg)) continue;
1963 /* only for values in L\U (TODO and D?), the others are handled with post_use */
1964 if(!pset_find_ptr(used, remat_arg)) {
1965 /* remat <= live_range(remat_arg) */
1966 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
1967 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1969 /* if value is becoming live through use by remat2 */
1970 if(!pset_find_ptr(live, remat_arg)) {
1973 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1974 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1976 arg_op->attr.live_range.ilp = lr;
1977 arg_op->attr.live_range.op = irn;
1979 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
1981 pset_insert_ptr(live, remat_arg);
1982 add_to_spill_bb(si, bb, remat_arg);
1985 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1986 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1991 d = pset_count(defs);
1992 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
1994 /* count how many regs irn needs for arguments */
1995 u = set_count(args);
1998 /* check the register pressure in the epilog */
1999 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
2000 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
2001 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
2003 /* add L\U' to check_post */
2004 pset_foreach(live, tmp) {
2005 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
2006 /* if a live value is not used by irn */
2007 tmp_op = get_irn_link(tmp);
2008 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
2012 /***********************************************************
2013 * 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
2014 **********************************************************/
2017 pset_foreach(used, tmp) {
2023 op_t *arg_op = get_irn_link(arg);
2026 spill = add_to_spill_bb(si, bb, arg);
2028 /* new live range for each used value */
2029 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2030 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2032 /* the epilog stuff - including post_use, check_post, check_post_remat */
2033 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2034 post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2036 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2038 /* arg is live throughout epilog if the next live_range is in a register */
2039 if(pset_find_ptr(live, arg)) {
2040 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
2042 /* post_use >= next_lr + remat */
2043 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2044 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2045 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2046 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2050 /* if value is not an arg of op and not possibly defined by post remat
2051 * then it may only die and not become live
2053 if(!set_find_keyval(args, arg)) {
2054 /* post_use <= prev_lr */
2055 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2056 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2057 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2058 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2060 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2061 /* next_lr <= prev_lr */
2062 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2063 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2064 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2065 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2070 /* forall post remat which use arg add a similar cst */
2071 foreach_post_remat(irn, remat) {
2074 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2075 ir_node *remat_arg = get_irn_n(remat, n);
2076 op_t *remat_op = get_irn_link(remat);
2078 if(remat_arg == arg) {
2079 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
2081 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2082 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2083 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2084 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2089 if(opt_memoperands) {
2090 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2091 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2092 ilp_var_t memoperand;
2094 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2095 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0);
2096 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2098 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2099 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2101 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2102 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2103 // if(arg_op->attr.live_range.ilp != ILP_UNDEF)
2104 // lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2109 /* new live range begins for each used value */
2110 arg_op->attr.live_range.ilp = prev_lr;
2111 arg_op->attr.live_range.op = irn;
2113 /*if(!pset_find_ptr(live, arg)) {
2114 pset_insert_ptr(live, arg);
2115 add_to_spill_bb(si, bb, arg);
2117 pset_insert_ptr(live, arg);
2121 /* just to be sure */
2122 check_post = ILP_UNDEF;
2131 /* check the register pressure in the prolog */
2132 /* sum_{L\U} lr <= k - |U| */
2133 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2134 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2136 /* for the prolog remove defined values from the live set */
2137 pset_foreach(defs, tmp) {
2138 pset_remove_ptr(live, tmp);
2141 if(opt_memoperands) {
2142 ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2143 one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2146 /***********************************************************
2147 * 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
2148 **********************************************************/
2151 set_foreach(args, keyval) {
2153 const ir_node *arg = keyval->key;
2154 int i = PTR_TO_INT(keyval->val);
2155 op_t *arg_op = get_irn_link(arg);
2156 ilp_cst_t requirements;
2159 spill = set_find_spill(spill_bb->ilp, arg);
2162 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2163 op->attr.live_range.args.reloads[i] = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
2165 /* reload <= mem_out */
2166 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2167 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2168 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2169 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2171 /* requirement: arg must be in register for use */
2172 /* reload + remat + live_range == 1 */
2173 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2174 requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2176 lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2177 lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2178 foreach_pre_remat(si, irn, tmp) {
2179 op_t *remat_op = get_irn_link(tmp);
2180 if(remat_op->attr.remat.remat->value == arg) {
2181 lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2185 if(opt_memoperands) {
2187 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2188 if(get_irn_n(irn, n) == arg) {
2192 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2193 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2194 memoperand_t *memoperand;
2195 memoperand = set_find_memoperand(si->memoperands, irn, n);
2197 /* memoperand <= mem_out */
2198 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2199 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2200 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2201 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2203 /* the memoperand is only sufficient if it is used once by the op */
2204 if(n_memoperands == 1)
2205 lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2207 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2209 /* we have one more free register if we use a memory operand */
2210 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2216 /* iterate over L\U */
2217 pset_foreach(live, tmp) {
2218 if(!set_find_keyval(args, tmp)) {
2219 /* if a live value is not used by irn */
2220 tmp_op = get_irn_link(tmp);
2221 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2226 /* requirements for remats */
2227 /* start new live ranges for values used by remats */
2228 foreach_pre_remat(si, irn, tmp) {
2229 op_t *remat_op = get_irn_link(tmp);
2232 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2233 ir_node *remat_arg = get_irn_n(tmp, n);
2234 op_t *arg_op = get_irn_link(remat_arg);
2236 if(!has_reg_class(si, remat_arg)) continue;
2238 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2239 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2240 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2242 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2243 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2245 /* if remat arg is also used by current op then we can use reload placed for this argument */
2246 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2247 int index = (int)keyval->val;
2249 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2257 /*************************
2258 * D O N E W I T H O P
2259 *************************/
2261 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2263 pset_foreach(live, tmp) {
2264 assert(has_reg_class(si, tmp));
2267 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2268 ir_node *arg = get_irn_n(irn, n);
2270 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2273 del_pset(remat_defs);
2277 defs = pset_new_ptr_default();
2282 /***************************************
2283 * B E G I N N I N G O F B L O C K
2284 ***************************************/
2287 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2288 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2290 pset_foreach(live, irn) {
2291 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2294 /* construct mem_outs for all values */
2296 set_foreach(spill_bb->ilp, spill) {
2297 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2298 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2300 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2301 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2303 if(pset_find_ptr(live, spill->irn)) {
2304 int default_spilled;
2305 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2307 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2308 default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2309 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2310 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2312 if(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2314 op_t *op = get_irn_link(spill->irn);
2316 /* do we have to copy a phi argument? */
2317 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2318 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2320 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2321 const ir_node *arg = get_irn_n(spill->irn, n);
2327 /* argument already done? */
2328 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2330 /* get sum of execution frequencies of blocks with the same phi argument */
2331 for(m=n; m>=0; --m) {
2332 const ir_node *arg2 = get_irn_n(spill->irn, m);
2335 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2339 /* copies are not for free */
2340 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2341 var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
2343 for(m=n; m>=0; --m) {
2344 const ir_node *arg2 = get_irn_n(spill->irn, m);
2347 op->attr.live_range.args.copies[m] = var;
2352 /* copy <= mem_in */
2353 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2354 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2355 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2356 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2364 /* L\U is empty at bb start */
2365 /* arg is live throughout epilog if it is reg_in into this block */
2367 /* check the register pressure at the beginning of the block
2370 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2371 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2373 pset_foreach(live, irn) {
2376 spill = set_find_spill(spill_bb->ilp, irn);
2379 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2380 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2382 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2384 /* spill + mem_in <= 1 */
2385 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2386 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2388 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2389 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2392 foreach_post_remat(bb, irn) {
2393 op_t *remat_op = get_irn_link(irn);
2395 DBG((si->dbg, LEVEL_4, "\t next post remat: %+F\n", irn));
2396 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
2398 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2401 /* forall post remats add requirements */
2402 foreach_post_remat(bb, tmp) {
2405 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2406 ir_node *remat_arg = get_irn_n(tmp, n);
2407 op_t *remat_op = get_irn_link(tmp);
2409 if(!has_reg_class(si, remat_arg)) continue;
2411 spill = set_find_spill(spill_bb->ilp, remat_arg);
2414 /* remat <= reg_in_argument */
2415 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2416 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2417 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2418 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2422 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2423 pset_foreach(live, irn) {
2427 spill = set_find_spill(spill_bb->ilp, irn);
2428 assert(spill && spill->irn == irn);
2430 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2431 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2434 ir_node *phi_arg = get_Phi_pred(irn, n);
2435 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2436 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2438 op_t *op = get_irn_link(irn);
2440 /* although the phi is in the right regclass one or more of
2441 * its arguments can be in a different one or at least to
2444 if(has_reg_class(si, phi_arg)) {
2445 /* mem_in < mem_out_arg + copy */
2446 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2447 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2449 /* reg_in < reg_out_arg */
2450 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2451 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2453 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2454 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2456 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2459 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2461 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2463 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2467 /* else assure the value arrives on all paths in the same resource */
2469 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2472 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2473 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2476 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2477 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2478 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2479 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2481 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2482 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2484 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2487 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2488 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2493 /* first live ranges from reg_ins */
2494 pset_foreach(live, irn) {
2495 op_t *op = get_irn_link(irn);
2497 spill = set_find_spill(spill_bb->ilp, irn);
2498 assert(spill && spill->irn == irn);
2500 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2501 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2502 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2503 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2505 foreach_post_remat(bb, tmp) {
2506 op_t *remat_op = get_irn_link(tmp);
2508 if(remat_op->attr.remat.remat->value == irn) {
2509 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2514 /* walk forward now and compute constraints for placing spills */
2515 /* this must only be done for values that are not defined in this block */
2516 /* TODO are these values at start of block? if yes, just check whether this is a diverge edge and skip the loop */
2517 pset_foreach(live, irn) {
2519 * if value is defined in this block we can anways place the spill directly after the def
2520 * -> no constraint necessary
2522 if(!is_Phi(irn) && get_nodes_block(irn) == bb) continue;
2525 spill = set_find_spill(spill_bb->ilp, irn);
2528 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2529 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2531 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2532 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2535 sched_foreach_op(bb, tmp) {
2536 op_t *op = get_irn_link(tmp);
2538 if(is_Phi(tmp)) continue;
2539 assert(!is_Proj(tmp));
2542 const ir_node *value = op->attr.remat.remat->value;
2545 /* only collect remats up to the first real use of a value */
2546 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2551 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2552 ir_node *arg = get_irn_n(tmp, n);
2555 /* if a value is used stop collecting remats */
2568 typedef struct _irnlist_t {
2569 struct list_head list;
2573 typedef struct _interference_t {
2574 struct list_head blocklist;
2580 cmp_interference(const void *a, const void *b, size_t size)
2582 const interference_t *p = a;
2583 const interference_t *q = b;
2585 return !(p->a == q->a && p->b == q->b);
2588 static interference_t *
2589 set_find_interference(set * set, ir_node * a, ir_node * b)
2591 interference_t query;
2593 query.a = (a>b)?a:b;
2594 query.b = (a>b)?b:a;
2596 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2599 static interference_t *
2600 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2602 interference_t query,
2604 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2608 result = set_find_interference(set, a, b);
2611 list_add(&list->list, &result->blocklist);
2615 query.a = (a>b)?a:b;
2616 query.b = (a>b)?b:a;
2618 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2620 INIT_LIST_HEAD(&result->blocklist);
2621 list_add(&list->list, &result->blocklist);
2627 values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
2629 const ir_edge_t *edge;
2631 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2632 /* both values are live in, so they interfere */
2636 /* ensure a dominates b */
2637 if(value_dominates(b,a)) {
2643 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2646 /* the following code is stolen from bera.c */
2647 if(be_is_live_end(si->lv, bb, a))
2650 foreach_out_edge(a, edge) {
2651 const ir_node *user = edge->src;
2652 if(get_nodes_block(user) == bb
2655 && !pset_find_ptr(si->inverse_ops, user)
2656 && value_dominates(b, user))
2664 * Walk all irg blocks and collect interfering values inside of phi classes
2667 luke_interferencewalker(ir_node * bb, void * data)
2669 spill_ilp_t *si = (spill_ilp_t*)data;
2672 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2673 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
2674 op_t *a_op = get_irn_link(a);
2677 /* a is only interesting if it is in my register class and if it is inside a phi class */
2678 if (has_reg_class(si, a) && get_phi_class(a)) {
2679 if(a_op->is_remat || pset_find_ptr(si->inverse_ops, a))
2682 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)) {
2683 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
2684 op_t *b_op = get_irn_link(b);
2687 /* a and b are only interesting if they are in the same phi class */
2688 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2689 if(b_op->is_remat || pset_find_ptr(si->inverse_ops, b))
2692 if(values_interfere_in_block(si, bb, a, b)) {
2693 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2694 set_insert_interference(si, si->interferences, a, b, bb);
2702 static unsigned int copy_path_id = 0;
2705 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2712 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2713 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2715 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2717 pset_foreach(copies, ptr) {
2718 copy = PTR_TO_INT(ptr);
2719 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2724 * @parameter copies contains a path of copies which lead us to irn
2725 * @parameter visited contains a set of nodes already visited on this path
2728 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2730 const ir_edge_t *edge;
2731 op_t *op = get_irn_link(irn);
2732 pset *visited_users = pset_new_ptr_default();
2735 if(op->is_remat) return 0;
2737 pset_insert_ptr(visited, irn);
2741 pset *visited_operands = pset_new_ptr(get_irn_arity(irn));
2743 /* visit all operands */
2744 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2745 ir_node *arg = get_irn_n(irn, n);
2746 ilp_var_t copy = op->attr.live_range.args.copies[n];
2748 if(!has_reg_class(si, arg)) continue;
2749 if(pset_find_ptr(visited_operands, arg)) continue;
2750 pset_insert_ptr(visited_operands, arg);
2753 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2754 del_pset(visited_operands);
2755 del_pset(visited_users);
2756 pset_remove_ptr(visited, irn);
2759 pset_insert(copies, INT_TO_PTR(copy), copy);
2760 write_copy_path_cst(si, copies, any_interfere);
2761 pset_remove(copies, INT_TO_PTR(copy), copy);
2762 } else if(!pset_find_ptr(visited, arg)) {
2763 pset_insert(copies, INT_TO_PTR(copy), copy);
2764 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
2765 pset_remove(copies, INT_TO_PTR(copy), copy);
2767 if(paths > MAX_PATHS) {
2768 if(pset_count(copies) == 0) {
2772 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2773 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2774 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2775 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2776 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2780 del_pset(visited_operands);
2781 del_pset(visited_users);
2782 pset_remove_ptr(visited, irn);
2785 } else if(pset_count(copies) == 0) {
2791 del_pset(visited_operands);
2794 /* visit all uses which are phis */
2795 foreach_out_edge(irn, edge) {
2796 ir_node *user = edge->src;
2797 int pos = edge->pos;
2798 op_t *op = get_irn_link(user);
2801 if(!is_Phi(user)) continue;
2802 if(!has_reg_class(si, user)) continue;
2803 if(pset_find_ptr(visited_users, user)) continue;
2804 pset_insert_ptr(visited_users, user);
2806 copy = op->attr.live_range.args.copies[pos];
2808 if(user == target) {
2809 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2810 del_pset(visited_users);
2811 pset_remove_ptr(visited, irn);
2814 pset_insert(copies, INT_TO_PTR(copy), copy);
2815 write_copy_path_cst(si, copies, any_interfere);
2816 pset_remove(copies, INT_TO_PTR(copy), copy);
2817 } else if(!pset_find_ptr(visited, user)) {
2818 pset_insert(copies, INT_TO_PTR(copy), copy);
2819 paths += find_copy_path(si, user, target, any_interfere, copies, visited);
2820 pset_remove(copies, INT_TO_PTR(copy), copy);
2822 if(paths > MAX_PATHS) {
2823 if(pset_count(copies) == 0) {
2827 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2828 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2829 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2830 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2831 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2835 del_pset(visited_users);
2836 pset_remove_ptr(visited, irn);
2839 } else if(pset_count(copies) == 0) {
2845 del_pset(visited_users);
2846 pset_remove_ptr(visited, irn);
2851 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
2853 pset * copies = pset_new_ptr_default();
2854 pset * visited = pset_new_ptr_default();
2856 find_copy_path(si, a, b, any_interfere, copies, visited);
2864 memcopyhandler(spill_ilp_t * si)
2866 interference_t *interference;
2868 /* teste Speicherwerte auf Interferenz */
2870 /* analyze phi classes */
2871 phi_class_compute(si->chordal_env->irg);
2873 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
2874 irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
2876 /* now lets emit the ILP unequations for the crap */
2877 set_foreach(si->interferences, interference) {
2879 ilp_var_t interfere,
2881 ilp_cst_t any_interfere_cst,
2883 const ir_node *a = interference->a;
2884 const ir_node *b = interference->b;
2886 /* any_interf <= \sum interf */
2887 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
2888 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2889 any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2891 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
2893 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
2894 const ir_node *bb = irnlist->irn;
2895 spill_bb_t *spill_bb = get_irn_link(bb);
2900 spilla = set_find_spill(spill_bb->ilp, a);
2903 spillb = set_find_spill(spill_bb->ilp, b);
2906 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
2907 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
2908 /* 2: - mem_in_a - spill_a + interfere <= 0 */
2909 /* 3: - mem_in_b - spill_b + interfere <= 0 */
2910 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
2911 interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2913 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
2914 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2916 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
2917 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
2918 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
2919 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
2920 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
2922 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
2923 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2925 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2926 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
2927 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
2929 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
2930 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2932 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2933 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
2934 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
2937 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
2939 /* any_interfere >= interf */
2940 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
2941 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2943 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2944 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2947 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
2948 gen_copy_constraints(si,a,b,any_interfere);
2956 return fabs(x) < 0.00001;
2959 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
2961 spill_ilp_t *si = get_irg_link(current_ir_graph);
2963 if(pset_find_ptr(si->all_possible_remats, n)) {
2964 op_t *op = (op_t*)get_irn_link(n);
2965 assert(op && op->is_remat);
2967 if(!op->attr.remat.remat->inverse) {
2968 if(op->attr.remat.pre) {
2969 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
2971 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
2976 op_t *op = (op_t*)get_irn_link(n);
2977 assert(op && op->is_remat);
2979 if(op->attr.remat.pre) {
2980 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
2982 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
2993 dump_graph_with_remats(ir_graph * irg, const char * suffix)
2995 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
2996 be_dump(irg, suffix, dump_ir_block_graph_sched);
2997 set_dump_node_vcgattr_hook(NULL);
3001 * Edge hook to dump the schedule edges with annotated register pressure.
3004 sched_pressure_edge_hook(FILE *F, ir_node *irn)
3006 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
3007 ir_node *prev = sched_prev(irn);
3008 fprintf(F, "edge:{sourcename:\"");
3010 fprintf(F, "\" targetname:\"");
3012 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
3013 fprintf(F, "\" color:magenta}\n");
3019 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
3021 DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3023 dump_consts_local(0);
3024 set_dump_node_edge_hook(sched_pressure_edge_hook);
3025 dump_ir_block_graph(irg, suffix);
3026 set_dump_node_edge_hook(old_edge_hook);
3030 walker_pressure_annotator(ir_node * bb, void * data)
3032 spill_ilp_t *si = data;
3035 pset *live = pset_new_ptr_default();
3038 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3039 irn = be_lv_get_irn(si->lv, bb, i);
3041 if (has_reg_class(si, irn)) {
3042 pset_insert_ptr(live, irn);
3046 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3048 sched_foreach_reverse(bb, irn) {
3050 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3054 if(has_reg_class(si, irn)) {
3055 pset_remove_ptr(live, irn);
3056 if(is_Proj(irn)) ++projs;
3059 if(!is_Proj(irn)) projs = 0;
3061 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3062 ir_node *arg = get_irn_n(irn, n);
3064 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3066 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3073 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3075 be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
3079 connect_all_remats_with_keep(spill_ilp_t * si)
3087 n_remats = pset_count(si->all_possible_remats);
3089 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3092 pset_foreach(si->all_possible_remats, irn) {
3097 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
3099 obstack_free(si->obst, ins);
3104 connect_all_spills_with_keep(spill_ilp_t * si)
3113 n_spills = pset_count(si->spills);
3115 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3118 pset_foreach(si->spills, irn) {
3123 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
3125 obstack_free(si->obst, ins);
3129 /** insert a spill at an arbitrary position */
3130 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3132 ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
3133 ir_graph *irg = get_irn_irg(bl);
3134 ir_node *frame = get_irg_frame(irg);
3138 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
3139 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3141 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3144 * search the right insertion point. a spill of a phi cannot be put
3145 * directly after the phi, if there are some phis behind the one which
3146 * is spilled. Also, a spill of a Proj must be after all Projs of the
3149 * Here's one special case:
3150 * If the spill is in the start block, the spill must be after the frame
3151 * pointer is set up. This is done by setting insert to the end of the block
3152 * which is its default initialization (see above).
3155 if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3158 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3161 sched_add_after(insert, spill);
3166 delete_remat(spill_ilp_t * si, ir_node * remat) {
3168 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3170 sched_remove(remat);
3172 /* kill links to operands */
3173 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3174 set_irn_n(remat, n, bad);
3179 clean_remat_info(spill_ilp_t * si)
3183 remat_info_t *remat_info;
3184 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3186 set_foreach(si->remat_info, remat_info) {
3187 if(!remat_info->remats) continue;
3189 pset_foreach(remat_info->remats, remat)
3191 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3192 set_irn_n((ir_node*)remat->proj, -1, bad);
3193 set_irn_n((ir_node*)remat->proj, 0, bad);
3196 if(get_irn_n_edges(remat->op) == 0) {
3197 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3198 set_irn_n((ir_node*)remat->op, n, bad);
3203 if(remat_info->remats) del_pset(remat_info->remats);
3204 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3209 delete_unnecessary_remats(spill_ilp_t * si)
3211 if(opt_keep_alive & KEEPALIVE_REMATS) {
3213 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3216 // ir_node *end = get_irg_end(si->chordal_env->irg);
3219 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3220 ir_node *keep_arg = get_irn_n(si->keep, n);
3221 op_t *arg_op = get_irn_link(keep_arg);
3224 assert(arg_op->is_remat);
3226 name = si->lpp->vars[arg_op->attr.remat.ilp];
3228 if(is_zero(name->value)) {
3229 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
3230 /* TODO check whether reload is preferred over remat (could be bug) */
3231 delete_remat(si, keep_arg);
3233 if(!arg_op->attr.remat.remat->inverse) {
3234 if(arg_op->attr.remat.pre) {
3235 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3237 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3240 if(arg_op->attr.remat.pre) {
3241 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3243 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3248 set_irn_n(si->keep, n, bad);
3251 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
3252 ir_node *end_arg = get_End_keepalive(end, i);
3254 if(end_arg != si->keep) {
3255 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
3258 keeps = obstack_finish(si->obst);
3259 set_End_keepalives(end, n-1, keeps);
3260 obstack_free(si->obst, keeps);
3263 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
3268 pset_foreach(si->all_possible_remats, remat) {
3269 op_t *remat_op = get_irn_link(remat);
3270 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
3272 if(is_zero(name->value)) {
3273 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
3274 /* TODO check whether reload is preferred over remat (could be bug) */
3275 delete_remat(si, remat);
3277 if(!remat_op->attr.remat.remat->inverse) {
3278 if(remat_op->attr.remat.pre) {
3279 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3281 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3284 if(remat_op->attr.remat.pre) {
3285 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3287 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3296 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3298 pset *spills = pset_new_ptr_default();
3300 const ir_node *next;
3303 defs = set_find_def(si->values, value);
3305 if(defs && defs->spills) {
3306 for(next = defs->spills; next; next = get_irn_link(next)) {
3307 pset_insert_ptr(spills, next);
3315 * @param before The node after which the spill will be placed in the schedule
3318 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3322 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3324 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3326 spill = be_spill2(arch_env, irn, before);
3328 defs = set_insert_def(si->values, value);
3331 /* enter into the linked list */
3332 set_irn_link(spill, defs->spills);
3333 defs->spills = spill;
3335 if(opt_keep_alive & KEEPALIVE_SPILLS)
3336 pset_insert_ptr(si->spills, spill);
3342 * @param before The Phi node which has to be spilled
3345 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3352 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3354 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3355 ins[n] = si->m_unknown;
3358 mem_phi = new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3360 defs = set_insert_def(si->values, phi);
3363 /* enter into the linked list */
3364 set_irn_link(mem_phi, defs->spills);
3365 defs->spills = mem_phi;
3367 #ifdef SCHEDULE_PHIM
3368 sched_add_after(phi, mem_phi);
3371 if(opt_keep_alive & KEEPALIVE_SPILLS)
3372 pset_insert_ptr(si->spills, mem_phi);
3379 * Add remat to list of defs, destroys link field!
3382 insert_remat(spill_ilp_t * si, ir_node * remat)
3385 op_t *remat_op = get_irn_link(remat);
3387 assert(remat_op->is_remat);
3389 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3392 /* enter into the linked list */
3393 set_irn_link(remat, defs->remats);
3394 defs->remats = remat;
3399 * Add reload before operation and add to list of defs
3402 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3407 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3409 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3411 defs = set_find_def(si->values, value);
3413 spill = defs->spills;
3414 assert(spill && "no spill placed before reload");
3416 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3418 /* enter into the linked list */
3419 set_irn_link(reload, defs->remats);
3420 defs->remats = reload;
3425 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3429 ir_node *value = get_irn_n(memoperand->irn, memoperand->pos);
3431 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3433 DBG((si->dbg, LEVEL_2, "\t inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3435 defs = set_find_def(si->values, value);
3437 spill = defs->spills;
3438 assert(spill && "no spill placed before reload");
3440 reload = be_reload(arch_env, si->cls, memoperand->irn, get_irn_mode(value), spill);
3442 arch_perform_memory_operand(arch_env, memoperand->irn, reload, memoperand->pos);
3443 sched_remove(reload);
3446 void insert_memoperands(spill_ilp_t * si)
3448 memoperand_t *memoperand;
3451 set_foreach(si->memoperands, memoperand) {
3452 name = si->lpp->vars[memoperand->ilp];
3453 if(!is_zero(name->value)) {
3454 perform_memory_operand(si, memoperand);
3460 walker_spill_placer(ir_node * bb, void * data) {
3461 spill_ilp_t *si = (spill_ilp_t*)data;
3463 spill_bb_t *spill_bb = get_irn_link(bb);
3464 pset *spills_to_do = pset_new_ptr_default();
3467 set_foreach(spill_bb->ilp, spill) {
3470 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3471 name = si->lpp->vars[spill->mem_in];
3472 if(!is_zero(name->value)) {
3475 mem_phi = insert_mem_phi(si, spill->irn);
3477 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3481 name = si->lpp->vars[spill->spill];
3482 if(!is_zero(name->value)) {
3483 /* place spill directly after definition */
3484 if(get_nodes_block(spill->irn) == bb) {
3485 insert_spill(si, spill->irn, spill->irn, spill->irn);
3489 /* place spill at bb start */
3490 if(spill->reg_in > 0) {
3491 name = si->lpp->vars[spill->reg_in];
3492 if(!is_zero(name->value)) {
3493 insert_spill(si, spill->irn, spill->irn, bb);
3497 /* place spill after a remat */
3498 pset_insert_ptr(spills_to_do, spill->irn);
3501 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3504 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3505 op_t *op = get_irn_link(irn);
3507 if(be_is_Spill(irn)) continue;
3510 /* TODO fix this if we want to support remats with more than two nodes */
3511 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3512 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3514 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3517 if(pset_find_ptr(spills_to_do, irn)) {
3518 pset_remove_ptr(spills_to_do, irn);
3520 insert_spill(si, irn, irn, irn);
3526 assert(pset_count(spills_to_do) == 0);
3528 /* afterwards free data in block */
3529 del_pset(spills_to_do);
3533 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3535 ir_node *insert_pos = bb;
3537 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3539 /* find last definition of arg value in block */
3544 defs = set_find_def(si->values, value);
3546 if(defs && defs->remats) {
3547 for(next = defs->remats; next; next = get_irn_link(next)) {
3548 if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3549 last = sched_get_time_step(next);
3555 if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3556 last = sched_get_time_step(value);
3560 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
3562 spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3568 phim_fixer(spill_ilp_t *si) {
3571 set_foreach(si->values, defs) {
3572 const ir_node *phi = defs->value;
3573 op_t *op = get_irn_link(phi);
3574 ir_node *phi_m = NULL;
3575 ir_node *next = defs->spills;
3578 if(!is_Phi(phi)) continue;
3581 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3585 next = get_irn_link(next);
3588 if(!phi_m) continue;
3590 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3591 ir_node *value = get_irn_n(phi, n);
3592 defs_t *val_defs = set_find_def(si->values, value);
3594 /* a spill of this value */
3599 ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3600 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3602 if(!is_zero(name->value)) {
3603 spill = insert_mem_copy(si, pred, value);
3605 spill = val_defs->spills;
3608 spill = val_defs->spills;
3611 assert(spill && "no spill placed before PhiM");
3612 set_irn_n(phi_m, n, spill);
3618 walker_reload_placer(ir_node * bb, void * data) {
3619 spill_ilp_t *si = (spill_ilp_t*)data;
3621 spill_bb_t *spill_bb = get_irn_link(bb);
3623 /* reloads at end of block */
3624 if(spill_bb->reloads) {
3627 set_foreach(spill_bb->reloads, keyval) {
3628 ir_node *irn = (ir_node*)keyval->key;
3629 ilp_var_t reload = PTR_TO_INT(keyval->val);
3632 name = si->lpp->vars[reload];
3633 if(!is_zero(name->value)) {
3635 ir_node *insert_pos = bb;
3636 ir_node *prev = sched_block_last_noncf(si, bb);
3637 op_t *prev_op = get_irn_link(prev);
3639 while(be_is_Spill(prev)) {
3640 prev = sched_prev(prev);
3643 prev_op = get_irn_link(prev);
3645 /* insert reload before pre-remats */
3646 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3647 && prev_op->is_remat && prev_op->attr.remat.pre) {
3651 prev = sched_prev(prev);
3652 } while(be_is_Spill(prev));
3654 prev_op = get_irn_link(prev);
3658 reload = insert_reload(si, irn, insert_pos);
3660 if(opt_keep_alive & KEEPALIVE_RELOADS)
3661 pset_insert_ptr(si->spills, reload);
3666 /* walk and insert more reloads and collect remats */
3667 sched_foreach_reverse(bb, irn) {
3668 op_t *op = get_irn_link(irn);
3670 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3671 if(is_Phi(irn)) break;
3674 if(get_irn_mode(irn) != mode_T) {
3675 insert_remat(si, irn);
3680 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3681 ir_node *arg = get_irn_n(irn, n);
3683 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3686 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3687 if(!is_zero(name->value)) {
3689 ir_node *insert_pos = irn;
3690 ir_node *prev = sched_prev(insert_pos);
3693 while(be_is_Spill(prev)) {
3694 prev = sched_prev(prev);
3697 prev_op = get_irn_link(prev);
3699 /* insert reload before pre-remats */
3700 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3701 && prev_op->is_remat && prev_op->attr.remat.pre) {
3705 prev = sched_prev(prev);
3706 } while(be_is_Spill(prev));
3708 prev_op = get_irn_link(prev);
3712 reload = insert_reload(si, arg, insert_pos);
3714 set_irn_n(irn, n, reload);
3716 if(opt_keep_alive & KEEPALIVE_RELOADS)
3717 pset_insert_ptr(si->spills, reload);
3724 del_set(spill_bb->ilp);
3725 if(spill_bb->reloads) del_set(spill_bb->reloads);
3729 walker_collect_used(ir_node * irn, void * data)
3731 lc_bitset_t *used = data;
3733 lc_bitset_set(used, get_irn_idx(irn));
3736 struct kill_helper {
3742 walker_kill_unused(ir_node * bb, void * data)
3744 struct kill_helper *kh = data;
3745 ir_node *bad = get_irg_bad(get_irn_irg(bb));
3749 for(irn=sched_first(bb); !sched_is_end(irn);) {
3750 ir_node *next = sched_next(irn);
3753 if(!lc_bitset_is_set(kh->used, get_irn_idx(irn))) {
3754 if(be_is_Spill(irn) || be_is_Reload(irn)) {
3755 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)));
3757 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3763 set_nodes_block(irn, bad);
3764 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3765 set_irn_n(irn, n, bad);
3773 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3775 struct kill_helper kh;
3777 kh.used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3780 irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3781 irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3783 lc_bitset_free(kh.used);
3787 print_irn_pset(pset * p)
3791 pset_foreach(p, irn) {
3792 ir_printf("%+F\n", irn);
3797 dump_phi_class(spill_ilp_t * si, pset * phiclass, const char * file)
3799 FILE *f = fopen(file, "w");
3801 interference_t *interference;
3803 pset_break(phiclass);
3804 set_break(si->interferences);
3806 ir_fprintf(f, "digraph phiclass {\n");
3808 pset_foreach(phiclass, irn) {
3810 ir_fprintf(f, " %F%N [shape=box]\n",irn,irn);
3813 pset_foreach(phiclass, irn) {
3816 if(!is_Phi(irn)) continue;
3818 for(n=get_irn_arity(irn)-1; n>=0; --n) {
3819 ir_node *arg = get_irn_n(irn, n);
3821 ir_fprintf(f, " %F%N -> %F%N\n",irn,irn,arg,arg);
3825 set_foreach(si->interferences, interference) {
3826 const ir_node *a = interference->a;
3827 const ir_node *b = interference->b;
3828 if(get_phi_class(a) == phiclass) {
3829 ir_fprintf(f, " %F%N -> %F%N [color=red,dir=none,style=bold]\n",a,a,b,b);
3838 rewire_uses(spill_ilp_t * si)
3840 dom_front_info_t *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
3842 pset *ignore = pset_new_ptr(1);
3844 pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
3846 /* then fix uses of spills */
3847 set_foreach(si->values, defs) {
3850 const ir_node *next = defs->remats;
3853 reloads = pset_new_ptr_default();
3856 if(be_is_Reload(next)) {
3857 pset_insert_ptr(reloads, next);
3861 next = get_irn_link(next);
3864 spills = get_spills_for_value(si, defs->value);
3865 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));
3866 if(pset_count(spills) > 1) {
3867 //assert(pset_count(reloads) > 0);
3868 // print_irn_pset(spills);
3869 // print_irn_pset(reloads);
3871 be_ssa_constr_set_ignore(dfi, si->lv, spills, ignore);
3878 /* first fix uses of remats and reloads */
3879 set_foreach(si->values, defs) {
3881 const ir_node *next = defs->remats;
3884 nodes = pset_new_ptr_default();
3885 pset_insert_ptr(nodes, defs->value);
3888 pset_insert_ptr(nodes, next);
3889 next = get_irn_link(next);
3892 if(pset_count(nodes) > 1) {
3893 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
3894 be_ssa_constr_set(dfi, si->lv, nodes);
3901 // remove_unused_defs(si);
3903 be_free_dominance_frontiers(dfi);
3908 writeback_results(spill_ilp_t * si)
3910 /* walk through the graph and collect all spills, reloads and remats for a value */
3912 si->values = new_set(cmp_defs, 4096);
3914 DBG((si->dbg, LEVEL_1, "Applying results\n"));
3915 delete_unnecessary_remats(si);
3916 si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
3917 irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
3918 irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
3920 insert_memoperands(si);
3923 /* clean the remat info! there are still back-edges leading there! */
3924 clean_remat_info(si);
3928 connect_all_spills_with_keep(si);
3930 del_set(si->values);
3934 get_n_regs(spill_ilp_t * si)
3936 int arch_n_regs = arch_register_class_n_regs(si->cls);
3940 for(i=0; i<arch_n_regs; i++) {
3941 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
3946 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
3951 walker_reload_mover(ir_node * bb, void * data)
3953 spill_ilp_t *si = data;
3956 sched_foreach(bb, tmp) {
3957 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
3958 ir_node *reload = tmp;
3961 /* move reload upwards */
3963 int pressure = (int)get_irn_link(reload);
3964 if(pressure < si->n_regs) {
3965 irn = sched_prev(reload);
3966 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
3967 sched_remove(reload);
3968 pressure = (int)get_irn_link(irn);
3970 while(pressure < si->n_regs) {
3971 if( sched_is_end(irn) ||
3972 (be_is_Reload(irn) && has_reg_class(si, irn)) ||
3973 /* do not move reload before its spill */
3974 (irn == be_get_Reload_mem(reload)) ||
3975 /* do not move before phi */
3978 set_irn_link(irn, INT_TO_PTR(pressure+1));
3979 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
3980 irn = sched_prev(irn);
3982 pressure = (int)get_irn_link(irn);
3985 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
3986 sched_put_after(irn, reload);
3993 move_reloads_upward(spill_ilp_t * si)
3995 irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
4000 * Walk all irg blocks and check for interfering spills inside of phi classes
4003 luke_meminterferencechecker(ir_node * bb, void * data)
4005 spill_ilp_t *si = (spill_ilp_t*)data;
4008 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
4009 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
4011 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
4013 /* a is only interesting if it is in my register class and if it is inside a phi class */
4014 if (has_reg_class(si, a) && get_phi_class(a)) {
4015 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)) {
4016 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
4018 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
4020 /* a and b are only interesting if they are in the same phi class */
4021 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
4022 if(values_interfere_in_block(si, bb, a, b)) {
4023 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4032 verify_phiclasses(spill_ilp_t * si)
4034 /* analyze phi classes */
4035 phi_class_compute(si->chordal_env->irg);
4037 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4038 irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
4042 walker_spillslotassigner(ir_node * irn, void * data)
4046 if(!be_is_Spill(irn)) return;
4048 /* set spill context to phi class if it has one ;) */
4051 // Matze: not needed anymore
4052 cls = get_phi_class(irn);
4054 be_set_Spill_context(irn, cls);
4056 be_set_Spill_context(irn, irn);
4062 assign_spillslots(spill_ilp_t * si)
4064 DBG((si->dbg, LEVEL_2, "\t calling spill slot assigner\n"));
4065 irg_walk_graph(si->chordal_env->irg, walker_spillslotassigner, NULL, si);
4069 be_spill_remat(const be_chordal_env_t * chordal_env)
4072 char problem_name[256];
4073 char dump_suffix[256];
4074 char dump_suffix2[256];
4075 struct obstack obst;
4078 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
4079 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
4080 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
4082 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4083 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4085 if(opt_verify & VERIFY_DOMINANCE)
4086 be_check_dominance(chordal_env->irg);
4088 obstack_init(&obst);
4089 si.chordal_env = chordal_env;
4091 si.cls = chordal_env->cls;
4092 si.lpp = new_lpp(problem_name, lpp_minimize);
4093 si.remat_info = new_set(cmp_remat_info, 4096);
4094 si.interferences = new_set(cmp_interference, 32);
4095 si.memoperands = new_set(cmp_memoperands, 128);
4096 si.all_possible_remats = pset_new_ptr_default();
4097 si.spills = pset_new_ptr_default();
4098 si.inverse_ops = pset_new_ptr_default();
4099 si.lv = chordal_env->lv;
4101 si.n_regs = get_n_regs(&si);
4103 set_irg_link(chordal_env->irg, &si);
4104 compute_doms(chordal_env->irg);
4106 /* compute phi classes */
4107 // phi_class_compute(chordal_env->irg);
4109 be_analyze_regpressure(chordal_env, "-pre");
4112 /* collect remats */
4113 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4114 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
4117 /* insert possible remats */
4118 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4119 irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
4120 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4122 if(opt_keep_alive & KEEPALIVE_REMATS) {
4123 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4124 connect_all_remats_with_keep(&si);
4125 /* dump graph with inserted remats */
4126 dump_graph_with_remats(chordal_env->irg, dump_suffix);
4129 /* insert copies for phi arguments not in my regclass */
4130 irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
4132 /* recompute liveness */
4133 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4134 be_liveness_recompute(si.lv);
4138 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4139 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4140 irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
4142 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4143 irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
4146 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4147 memcopyhandler(&si);
4150 if(opt_dump_flags & DUMP_PROBLEM) {
4152 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4153 if ((f = fopen(buf, "wt")) != NULL) {
4154 lpp_dump_plain(si.lpp, f);
4159 if(opt_dump_flags & DUMP_MPS) {
4162 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4163 if((f = fopen(buf, "wt")) != NULL) {
4164 mps_write_mps(si.lpp, s_mps_fixed, f);
4168 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4169 if((f = fopen(buf, "wt")) != NULL) {
4170 mps_write_mst(si.lpp, s_mps_fixed, f);
4175 lpp_check_startvals(si.lpp);
4178 DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4179 lpp_set_time_limit(si.lpp, opt_timeout);
4182 lpp_set_log(si.lpp, stdout);
4185 lpp_solve_cplex(si.lpp);
4187 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4189 assert(lpp_is_sol_valid(si.lpp)
4190 && "solution of ILP must be valid");
4192 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));
4194 if(opt_dump_flags & DUMP_SOLUTION) {
4198 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4199 if ((f = fopen(buf, "wt")) != NULL) {
4201 for (i = 0; i < si.lpp->var_next; ++i) {
4202 lpp_name_t *name = si.lpp->vars[i];
4203 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4209 writeback_results(&si);
4213 kill_all_unused_values_in_schedule(&si);
4215 if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
4216 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
4218 // move reloads upwards
4219 be_liveness_recompute(si.lv);
4220 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4221 move_reloads_upward(&si);
4224 verify_phiclasses(&si);
4225 assign_spillslots(&si);
4228 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4230 dump_pressure_graph(&si, dump_suffix2);
4232 be_analyze_regpressure(chordal_env, "-post");
4234 if(opt_verify & VERIFY_DOMINANCE)
4235 be_check_dominance(chordal_env->irg);
4237 free_dom(chordal_env->irg);
4238 del_set(si.interferences);
4239 del_pset(si.inverse_ops);
4240 del_pset(si.all_possible_remats);
4241 del_set(si.memoperands);
4242 del_pset(si.spills);
4244 obstack_free(&obst, NULL);
4245 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4248 #else /* WITH_ILP */
4251 only_that_you_can_compile_without_WITH_ILP_defined(void)
4255 #endif /* WITH_ILP */