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 */
164 //#define SOLVE_LOCAL
165 #define LPP_SERVER "i44pc52"
166 #define LPP_SOLVER "cplex"
172 typedef struct _spill_ilp_t {
173 const arch_register_class_t *cls;
175 const be_chordal_env_t *chordal_env;
178 struct obstack *obst;
180 pset *all_possible_remats;
183 set *values; /**< for collecting all definitions of values before running ssa-construction */
188 DEBUG_ONLY(firm_dbg_module_t * dbg);
191 typedef int ilp_var_t;
192 typedef int ilp_cst_t;
194 typedef struct _spill_bb_t {
199 typedef struct _remat_t {
200 const ir_node *op; /**< for copy_irn */
201 const ir_node *value; /**< the value which is being recomputed by this remat */
202 const ir_node *proj; /**< not NULL if the above op produces a tuple */
203 int cost; /**< cost of this remat */
204 int inverse; /**< nonzero if this is an inverse remat */
208 * Data to be attached to each IR node. For remats this contains the ilp_var
209 * for this remat and for normal ops this contains the ilp_vars for
210 * reloading each operand
212 typedef struct _op_t {
217 const remat_t *remat; /** the remat this op belongs to */
218 int pre; /** 1, if this is a pressure-increasing remat */
222 ir_node *op; /** the operation this live range belongs to */
231 typedef struct _defs_t {
232 const ir_node *value;
233 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
234 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
237 typedef struct _remat_info_t {
238 const ir_node *irn; /**< the irn to which these remats belong */
239 pset *remats; /**< possible remats for this value */
240 pset *remats_by_operand; /**< remats with this value as operand */
243 typedef struct _keyval_t {
248 typedef struct _spill_t {
257 typedef struct _memoperand_t {
258 ir_node *irn; /**< the irn */
259 unsigned int pos; /**< the position of the argument */
260 ilp_var_t ilp; /**< the ilp var for this memory operand */
264 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
266 return chordal_has_class(si->chordal_env, irn);
271 cmp_remat(const void *a, const void *b)
273 const keyval_t *p = a;
274 const keyval_t *q = b;
275 const remat_t *r = p->val;
276 const remat_t *s = q->val;
280 return !(r == s || r->op == s->op);
284 cmp_remat(const void *a, const void *b)
286 const remat_t *r = a;
287 const remat_t *s = a;
289 return !(r == s || r->op == s->op);
293 cmp_spill(const void *a, const void *b, size_t size)
295 const spill_t *p = a;
296 const spill_t *q = b;
298 // return !(p->irn == q->irn && p->bb == q->bb);
299 return !(p->irn == q->irn);
303 cmp_memoperands(const void *a, const void *b, size_t size)
305 const memoperand_t *p = a;
306 const memoperand_t *q = b;
308 return !(p->irn == q->irn && p->pos == q->pos);
312 set_find_keyval(set * set, const void * key)
317 return set_find(set, &query, sizeof(query), HASH_PTR(key));
321 set_insert_keyval(set * set, void * key, void * val)
327 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
331 set_find_def(set * set, const ir_node * value)
336 return set_find(set, &query, sizeof(query), HASH_PTR(value));
340 set_insert_def(set * set, const ir_node * value)
347 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
350 static memoperand_t *
351 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
358 return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
361 static memoperand_t *
362 set_find_memoperand(set * set, const ir_node * irn, unsigned int pos)
366 query.irn = (ir_node*)irn;
368 return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
373 set_find_spill(set * set, const ir_node * value)
377 query.irn = (ir_node*)value;
378 return set_find(set, &query, sizeof(query), HASH_PTR(value));
381 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
382 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
383 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
384 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
385 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
388 cmp_remat_info(const void *a, const void *b, size_t size)
390 const remat_info_t *p = a;
391 const remat_info_t *q = b;
393 return !(p->irn == q->irn);
397 cmp_defs(const void *a, const void *b, size_t size)
402 return !(p->value == q->value);
406 cmp_keyval(const void *a, const void *b, size_t size)
408 const keyval_t *p = a;
409 const keyval_t *q = b;
411 return !(p->key == q->key);
415 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
418 #ifndef EXECFREQ_LOOPDEPH
419 return get_block_execfreq(si->chordal_env->exec_freq, get_block(irn)) + FUDGE;
422 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
424 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
429 get_cost(const spill_ilp_t * si, const ir_node * irn)
431 if(be_is_Spill(irn)) {
432 return opt_cost_spill;
433 } else if(be_is_Reload(irn)){
434 return opt_cost_reload;
436 return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
441 * Checks, whether node and its operands have suitable reg classes
444 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
447 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
448 int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
452 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
455 for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
456 ir_node *op = get_irn_n(irn, n);
457 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
460 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
467 * Try to create a remat from @p op with destination value @p dest_value
469 static INLINE remat_t *
470 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
472 remat_t *remat = NULL;
474 // if(!mode_is_datab(get_irn_mode(dest_value)))
477 if(dest_value == op) {
478 const ir_node *proj = NULL;
480 if(is_Proj(dest_value)) {
481 op = get_irn_n(op, 0);
485 if(!is_rematerializable(si, op))
488 remat = obstack_alloc(si->obst, sizeof(*remat));
490 remat->cost = get_cost(si, op);
491 remat->value = dest_value;
495 arch_inverse_t inverse;
498 /* get the index of the operand we want to retrieve by the inverse op */
499 for (n = get_irn_arity(op)-1; n>=0; --n) {
500 ir_node *arg = get_irn_n(op, n);
502 if(arg == dest_value) break;
506 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", n, op));
508 /* else ask the backend to give an inverse op */
509 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
512 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
514 assert(inverse.n > 0 && "inverse op should have at least one node");
516 for(i=inverse.n-1; i>=0; --i) {
517 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
521 remat = obstack_alloc(si->obst, sizeof(*remat));
522 remat->op = inverse.nodes[0];
523 remat->cost = inverse.costs;
524 remat->value = dest_value;
525 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
528 assert(is_Proj(remat->proj));
530 assert(0 && "I can not handle remats with more than 2 nodes");
537 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
539 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
547 add_remat(const spill_ilp_t * si, const remat_t * remat)
549 remat_info_t *remat_info,
554 assert(remat->value);
556 query.irn = remat->value;
558 query.remats_by_operand = NULL;
559 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
561 if(remat_info->remats == NULL) {
562 remat_info->remats = new_pset(cmp_remat, 4096);
564 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
566 /* insert the remat into the remats_be_operand set of each argument of the remat op */
567 for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
568 ir_node *arg = get_irn_n(remat->op, n);
572 query.remats_by_operand = NULL;
573 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
575 if(remat_info->remats_by_operand == NULL) {
576 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
578 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
583 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
585 const ir_edge_t *edge = get_irn_out_edge_first(irn);
589 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
592 edge = get_irn_out_edge_next(irn, edge);
599 get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
602 unsigned int ret = 0;
604 for(n=get_irn_arity(irn)-1; n>=0; --n) {
605 if(has_reg_class(si, irn)) ++ret;
612 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
617 if( has_reg_class(si, op)
618 && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1)
619 && (opt_remats != REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0)
621 remat = get_remat_from_op(si, op, op);
623 add_remat(si, remat);
627 if(opt_remats == REMATS_ALL) {
628 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
630 for (n = get_irn_arity(op)-1; n>=0; --n) {
631 ir_node *arg = get_irn_n(op, n);
633 if(has_reg_class(si, arg)) {
634 /* try to get an inverse remat */
635 remat = get_remat_from_op(si, arg, op);
637 add_remat(si, remat);
645 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
648 ir_node *def_block = get_nodes_block(val);
654 /* if pos is at end of a basic block */
656 ret = (pos == def_block || block_dominates(def_block, pos));
657 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
661 /* else if this is a normal operation */
662 block = get_nodes_block(pos);
663 if(block == def_block) {
664 if(!sched_is_scheduled(val)) return 1;
666 ret = sched_comes_after(val, pos);
667 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
671 ret = block_dominates(def_block, block);
672 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
676 static INLINE ir_node *
677 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
679 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
683 * Returns first non-Phi node of block @p bb
685 static INLINE ir_node *
686 sched_block_first_nonphi(const ir_node * bb)
688 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
692 sched_skip_proj_predicator(const ir_node * irn, void * data)
694 return (is_Proj(irn));
697 static INLINE ir_node *
698 sched_next_nonproj(const ir_node * irn, int forward)
700 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
704 * Returns next operation node (non-Proj) after @p irn
705 * or the basic block of this node
707 static INLINE ir_node *
708 sched_next_op(const ir_node * irn)
710 ir_node *next = sched_next(irn);
715 return sched_next_nonproj(next, 1);
719 * Returns previous operation node (non-Proj) before @p irn
720 * or the basic block of this node
722 static INLINE ir_node *
723 sched_prev_op(const ir_node * irn)
725 ir_node *prev = sched_prev(irn);
730 return sched_next_nonproj(prev, 0);
734 sched_put_after(ir_node * insert, ir_node * irn)
736 if(is_Block(insert)) {
737 insert = sched_block_first_nonphi(insert);
739 insert = sched_next_op(insert);
741 sched_add_before(insert, irn);
745 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
747 if(is_Block(insert)) {
748 insert = sched_block_last_noncf(si, insert);
750 insert = sched_next_nonproj(insert, 0);
751 insert = sched_prev(insert);
753 sched_add_after(insert, irn);
757 * Tells you whether a @p remat can be placed before the irn @p pos
760 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
762 const ir_node *op = remat->op;
768 prev = sched_block_last_noncf(si, pos);
769 prev = sched_next_nonproj(prev, 0);
771 prev = sched_prev_op(pos);
773 /* do not remat if the rematted value is defined immediately before this op */
774 if(prev == remat->op) {
779 /* this should be just fine, the following OP will be using this value, right? */
781 /* only remat AFTER the real definition of a value (?) */
782 if(!value_is_defined_before(si, pos, remat->value)) {
783 // ir_fprintf(stderr, "error(not defined)");
788 for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
789 const ir_node *arg = get_irn_n(op, n);
791 if(opt_no_enlarge_liveness) {
792 if(has_reg_class(si, arg) && live) {
793 res &= pset_find_ptr((pset*)live, arg)?1:0;
795 res &= value_is_defined_before(si, pos, arg);
798 res &= value_is_defined_before(si, pos, arg);
806 * Tells you whether a @p remat can be placed after the irn @p pos
809 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
812 pos = sched_block_first_nonphi(pos);
814 pos = sched_next_op(pos);
817 /* only remat AFTER the real definition of a value (?) */
818 if(!value_is_defined_before(si, pos, remat->value)) {
822 return can_remat_before(si, remat, pos, live);
826 * Collect potetially rematerializable OPs
829 walker_remat_collector(ir_node * irn, void * data)
831 spill_ilp_t *si = data;
833 if(!is_Block(irn) && !is_Phi(irn)) {
834 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
835 get_remats_from_op(si, irn);
840 * Inserts a copy of @p irn before @p pos
843 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
848 bb = is_Block(pos)?pos:get_nodes_block(pos);
849 copy = exact_copy(irn);
851 _set_phi_class(copy, NULL);
852 set_nodes_block(copy, bb);
853 sched_put_before(si, pos, copy);
859 * Inserts a copy of @p irn after @p pos
862 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
867 bb = is_Block(pos)?pos:get_nodes_block(pos);
868 copy = exact_copy(irn);
870 _set_phi_class(copy, NULL);
871 set_nodes_block(copy, bb);
872 sched_put_after(pos, copy);
878 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
882 if(can_remat_after(si, remat, pos, live)) {
887 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
889 copy = insert_copy_after(si, remat->op, pos);
891 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
892 op = obstack_alloc(si->obst, sizeof(*op));
894 op->attr.remat.remat = remat;
895 op->attr.remat.pre = 0;
896 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
898 set_irn_link(copy, op);
899 pset_insert_ptr(si->all_possible_remats, copy);
901 proj_copy = insert_copy_after(si, remat->proj, copy);
902 set_irn_n(proj_copy, 0, copy);
903 set_irn_link(proj_copy, op);
904 pset_insert_ptr(si->all_possible_remats, proj_copy);
916 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
920 if(can_remat_before(si, remat, pos, live)) {
925 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
927 copy = insert_copy_before(si, remat->op, pos);
929 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
930 op = obstack_alloc(si->obst, sizeof(*op));
932 op->attr.remat.remat = remat;
933 op->attr.remat.pre = 1;
934 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
936 set_irn_link(copy, op);
937 pset_insert_ptr(si->all_possible_remats, copy);
939 proj_copy = insert_copy_after(si, remat->proj, copy);
940 set_irn_n(proj_copy, 0, copy);
941 set_irn_link(proj_copy, op);
942 pset_insert_ptr(si->all_possible_remats, proj_copy);
954 get_block_n_succs(const ir_node *block) {
955 const ir_edge_t *edge;
957 assert(edges_activated(current_ir_graph));
959 edge = get_block_succ_first(block);
963 edge = get_block_succ_next(block, edge);
968 is_merge_edge(const ir_node * bb)
971 return get_block_n_succs(bb) == 1;
977 is_diverge_edge(const ir_node * bb)
980 return get_Block_n_cfgpreds(bb) == 1;
986 walker_regclass_copy_insertor(ir_node * irn, void * data)
988 spill_ilp_t *si = data;
990 if(is_Phi(irn) && has_reg_class(si, irn)) {
993 for(n=get_irn_arity(irn)-1; n>=0; --n) {
994 ir_node *phi_arg = get_irn_n(irn, n);
995 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
997 if(!has_reg_class(si, phi_arg)) {
998 ir_node *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
999 ir_node *pos = sched_block_last_noncf(si, bb);
1000 op_t *op = obstack_alloc(si->obst, sizeof(*op));
1002 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
1003 sched_add_after(pos, copy);
1004 set_irn_n(irn, n, copy);
1007 op->attr.live_range.args.reloads = NULL;
1008 op->attr.live_range.ilp = ILP_UNDEF;
1009 set_irn_link(copy, op);
1017 * Insert (so far unused) remats into the irg to
1018 * recompute the potential liveness of all values
1021 walker_remat_insertor(ir_node * bb, void * data)
1023 spill_ilp_t *si = data;
1024 spill_bb_t *spill_bb;
1027 pset *live = pset_new_ptr_default();
1029 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1031 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1032 ir_node *value = be_lv_get_irn(si->lv, bb, i);
1034 /* add remats at end of block */
1035 if (has_reg_class(si, value)) {
1036 pset_insert_ptr(live, value);
1040 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1041 set_irn_link(bb, spill_bb);
1043 irn = sched_last(bb);
1044 while(!sched_is_end(irn)) {
1051 next = sched_prev(irn);
1053 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
1055 if(is_Phi(irn) || is_Proj(irn)) {
1058 if(has_reg_class(si, irn)) {
1059 pset_remove_ptr(live, irn);
1062 op = obstack_alloc(si->obst, sizeof(*op));
1064 op->attr.live_range.args.reloads = NULL;
1065 op->attr.live_range.ilp = ILP_UNDEF;
1066 set_irn_link(irn, op);
1072 op = obstack_alloc(si->obst, sizeof(*op));
1074 op->attr.live_range.ilp = ILP_UNDEF;
1075 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1076 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1077 set_irn_link(irn, op);
1079 args = pset_new_ptr_default();
1081 /* collect arguments of op */
1082 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1083 ir_node *arg = get_irn_n(irn, n);
1085 pset_insert_ptr(args, arg);
1088 /* set args of op already live in epilog */
1089 pset_foreach(args, arg) {
1090 if(has_reg_class(si, arg)) {
1091 pset_insert_ptr(live, arg);
1094 /* delete defined value from live set */
1095 if(has_reg_class(si, irn)) {
1096 pset_remove_ptr(live, irn);
1100 remat_args = pset_new_ptr_default();
1102 /* insert all possible remats before irn */
1103 pset_foreach(args, arg) {
1104 remat_info_t *remat_info,
1108 /* continue if the operand has the wrong reg class
1110 if(!has_reg_class(si, arg))
1114 query.remats = NULL;
1115 query.remats_by_operand = NULL;
1116 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1122 if(remat_info->remats) {
1123 pset_foreach(remat_info->remats, remat) {
1124 ir_node *remat_irn = NULL;
1126 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
1127 if(opt_remat_while_live) {
1128 if(pset_find_ptr(live, remat->value)) {
1129 remat_irn = insert_remat_before(si, remat, irn, live);
1132 remat_irn = insert_remat_before(si, remat, irn, live);
1135 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1136 ir_node *remat_arg = get_irn_n(remat_irn, n);
1138 if(!has_reg_class(si, remat_arg)) continue;
1140 pset_insert_ptr(remat_args, remat_arg);
1147 /* now we add remat args to op's args because they could also die at this op */
1148 pset_foreach(args,arg) {
1149 if(pset_find_ptr(remat_args, arg)) {
1150 pset_remove_ptr(remat_args, arg);
1153 pset_foreach(remat_args,arg) {
1154 pset_insert_ptr(args, arg);
1157 /* insert all possible remats after irn */
1158 pset_foreach(args, arg) {
1159 remat_info_t *remat_info,
1163 /* continue if the operand has the wrong reg class */
1164 if(!has_reg_class(si, arg))
1168 query.remats = NULL;
1169 query.remats_by_operand = NULL;
1170 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1176 /* do not place post remats after jumps */
1177 if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
1179 if(remat_info->remats_by_operand) {
1180 pset_foreach(remat_info->remats_by_operand, remat) {
1181 /* do not insert remats producing the same value as one of the operands */
1182 if(!pset_find_ptr(args, remat->value)) {
1183 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
1184 if(opt_remat_while_live) {
1185 if(pset_find_ptr(live, remat->value)) {
1186 insert_remat_after(si, remat, irn, live);
1189 insert_remat_after(si, remat, irn, live);
1196 del_pset(remat_args);
1201 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_in, i) {
1202 ir_node *value = be_lv_get_irn(si->lv, bb, i);
1204 /* add remats at end if successor has multiple predecessors */
1205 if(is_merge_edge(bb)) {
1206 /* add remats at end of block */
1207 if (be_is_live_end(si->lv, bb, value) && has_reg_class(si, value)) {
1208 remat_info_t *remat_info,
1213 query.remats = NULL;
1214 query.remats_by_operand = NULL;
1215 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1217 if(remat_info && remat_info->remats) {
1218 pset_foreach(remat_info->remats, remat) {
1219 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1221 insert_remat_before(si, remat, bb, NULL);
1226 if(is_diverge_edge(bb)) {
1227 /* add remat2s at beginning of block */
1228 if ((be_is_live_in(si->lv, bb, value) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) {
1229 remat_info_t *remat_info,
1234 query.remats = NULL;
1235 query.remats_by_operand = NULL;
1236 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1238 if(remat_info && remat_info->remats) {
1239 pset_foreach(remat_info->remats, remat) {
1240 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at beginning of block %+F\n", remat->op, bb));
1242 /* put the remat here if all its args are available */
1243 insert_remat_after(si, remat, bb, NULL);
1253 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1256 luke_endwalker(ir_node * bb, void * data)
1258 spill_ilp_t *si = (spill_ilp_t*)data;
1264 spill_bb_t *spill_bb = get_irn_link(bb);
1268 live = pset_new_ptr_default();
1269 use_end = pset_new_ptr_default();
1271 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1272 irn = be_lv_get_irn(si->lv, bb, i);
1273 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1276 pset_insert_ptr(live, irn);
1277 op = get_irn_link(irn);
1278 assert(!op->is_remat);
1282 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1283 /* their reg_out must always be set */
1284 sched_foreach_reverse(bb, irn) {
1287 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1289 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1290 ir_node *irn_arg = get_irn_n(irn, n);
1292 if(has_reg_class(si, irn_arg)) {
1293 pset_insert_ptr(use_end, irn_arg);
1298 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1299 //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1300 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1302 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1304 /* if this is a merge edge we can reload at the end of this block */
1305 if(is_merge_edge(bb)) {
1306 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1307 } else if(pset_count(use_end)){
1308 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1310 spill_bb->reloads = NULL;
1313 pset_foreach(live,irn) {
1317 int default_spilled;
1320 /* handle values used by control flow nodes later separately */
1321 if(pset_find_ptr(use_end, irn)) continue;
1324 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1326 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1328 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1329 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1330 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1332 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1333 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1335 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1336 /* by default spill value right after definition */
1337 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1338 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1340 if(is_merge_edge(bb)) {
1344 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1345 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1346 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1348 /* reload <= mem_out */
1349 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1350 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1351 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1354 spill->reg_in = ILP_UNDEF;
1355 spill->mem_in = ILP_UNDEF;
1358 pset_foreach(use_end,irn) {
1362 ilp_cst_t end_use_req,
1365 int default_spilled;
1368 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1370 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1372 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1373 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1374 /* if irn is used at the end of the block, then it is live anyway */
1375 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1377 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1378 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1380 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1381 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1382 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1384 /* reload for use be control flow op */
1385 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1386 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1387 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1389 /* reload <= mem_out */
1390 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1391 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1392 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1394 spill->reg_in = ILP_UNDEF;
1395 spill->mem_in = ILP_UNDEF;
1397 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1398 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1399 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1407 next_post_remat(const ir_node * irn)
1413 next = sched_block_first_nonphi(irn);
1415 next = sched_next_op(irn);
1418 if(sched_is_end(next))
1421 op = get_irn_link(next);
1422 if(op->is_remat && !op->attr.remat.pre) {
1431 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1437 ret = sched_block_last_noncf(si, irn);
1438 ret = sched_next(ret);
1439 ret = sched_prev_op(ret);
1441 ret = sched_prev_op(irn);
1444 if(sched_is_end(ret) || is_Phi(ret))
1447 op = (op_t*)get_irn_link(ret);
1448 if(op->is_remat && op->attr.remat.pre) {
1456 * Find a remat of value @p value in the epilog of @p pos
1459 find_post_remat(const ir_node * value, const ir_node * pos)
1461 while((pos = next_post_remat(pos)) != NULL) {
1464 op = get_irn_link(pos);
1465 assert(op->is_remat && !op->attr.remat.pre);
1467 if(op->attr.remat.remat->value == value)
1468 return (ir_node*)pos;
1471 const ir_edge_t *edge;
1472 foreach_out_edge(pos, edge) {
1473 ir_node *proj = get_edge_src_irn(edge);
1474 assert(is_Proj(proj));
1484 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1486 spill_bb_t *spill_bb = get_irn_link(bb);
1490 int default_spilled;
1493 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1495 double spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1497 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1499 spill->reg_out = ILP_UNDEF;
1500 spill->reg_in = ILP_UNDEF;
1501 spill->mem_in = ILP_UNDEF;
1503 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1504 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1506 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1507 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1508 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1515 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1520 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1521 irn = be_lv_get_irn(si->lv, bb, i);
1523 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1524 pset_insert_ptr(live, irn);
1528 irn = sched_last(bb);
1530 /* all values eaten by control flow operations are also live until the end of the block */
1531 sched_foreach_reverse(bb, irn) {
1534 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1536 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1537 ir_node *arg = get_irn_n(irn,i);
1539 if(has_reg_class(si, arg)) {
1540 pset_insert_ptr(live, arg);
1547 * Inserts ILP-constraints and variables for memory copying before the given position
1550 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1552 const ir_node *succ;
1553 const ir_edge_t *edge;
1554 spill_bb_t *spill_bb = get_irn_link(block);
1563 assert(edges_activated(current_ir_graph));
1565 edge = get_block_succ_first(block);
1571 edge = get_block_succ_next(block, edge);
1572 /* next block can only contain phis, if this is a merge edge */
1575 ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1576 copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1578 ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1579 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1581 pset_foreach(live, tmp) {
1584 op_t *op = get_irn_link(irn);
1585 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1587 spill = set_find_spill(spill_bb->ilp, tmp);
1590 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1592 lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1594 sched_foreach(succ, phi) {
1595 const ir_node *to_copy;
1597 spill_t *to_copy_spill;
1598 op_t *phi_op = get_irn_link(phi);
1599 ilp_var_t reload = ILP_UNDEF;
1602 if(!is_Phi(phi)) break;
1603 if(!has_reg_class(si, phi)) continue;
1605 to_copy = get_irn_n(phi, pos);
1607 to_copy_op = get_irn_link(to_copy);
1609 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1610 assert(to_copy_spill);
1612 if(spill_bb->reloads) {
1613 keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1616 reload = PTR_TO_INT(keyval->val);
1620 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1621 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1623 /* copy - reg_out - reload - remat - live_range <= 0 */
1624 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1625 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1626 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1627 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1628 foreach_pre_remat(si, block, tmp) {
1629 op_t *remat_op = get_irn_link(tmp);
1630 if(remat_op->attr.remat.remat->value == to_copy) {
1631 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1635 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1636 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1638 /* copy - reg_out - copyreg <= 0 */
1639 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1640 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1641 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1647 * Walk all irg blocks and emit this ILP
1650 luke_blockwalker(ir_node * bb, void * data)
1652 spill_ilp_t *si = (spill_ilp_t*)data;
1657 spill_bb_t *spill_bb = get_irn_link(bb);
1660 pset *defs = pset_new_ptr_default();
1661 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
1664 live = pset_new_ptr_default();
1666 /****************************************
1667 * B A S I C B L O C K E N D
1668 ***************************************/
1671 /* init live values at end of block */
1672 get_live_end(si, bb, live);
1674 pset_foreach(live, irn) {
1676 ilp_var_t reload = ILP_UNDEF;
1678 spill = set_find_spill(spill_bb->ilp, irn);
1681 if(spill_bb->reloads) {
1682 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1685 reload = PTR_TO_INT(keyval->val);
1689 op = get_irn_link(irn);
1690 assert(!op->is_remat);
1692 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1693 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1694 op->attr.live_range.op = bb;
1696 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1697 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1699 /* reg_out - reload - remat - live_range <= 0 */
1700 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1701 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1702 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1703 foreach_pre_remat(si, bb, tmp) {
1704 op_t *remat_op = get_irn_link(tmp);
1705 if(remat_op->attr.remat.remat->value == irn) {
1706 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1709 /* maybe we should also assure that reg_out >= live_range etc. */
1713 insert_mem_copy_position(si, live, bb);
1716 * start new live ranges for values used by remats at end of block
1717 * and assure the remat args are available
1719 foreach_pre_remat(si, bb, tmp) {
1720 op_t *remat_op = get_irn_link(tmp);
1723 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1724 ir_node *remat_arg = get_irn_n(tmp, n);
1725 op_t *arg_op = get_irn_link(remat_arg);
1728 if(!has_reg_class(si, remat_arg)) continue;
1730 /* if value is becoming live through use by remat */
1731 if(!pset_find_ptr(live, remat_arg)) {
1732 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1733 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1735 arg_op->attr.live_range.ilp = prev_lr;
1736 arg_op->attr.live_range.op = bb;
1738 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1740 pset_insert_ptr(live, remat_arg);
1741 add_to_spill_bb(si, bb, remat_arg);
1744 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1745 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1746 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1748 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1749 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1751 /* use reload placed for this argument */
1752 if(spill_bb->reloads) {
1753 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1756 ilp_var_t reload = PTR_TO_INT(keyval->val);
1758 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1763 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1768 /**************************************
1769 * B A S I C B L O C K B O D Y
1770 **************************************/
1772 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1778 ilp_cst_t check_pre,
1784 ilp_cst_t one_memoperand;
1786 /* iterate only until first phi */
1790 op = get_irn_link(irn);
1792 if(op->is_remat) continue;
1793 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1795 /* collect defined values */
1796 if(has_reg_class(si, irn)) {
1797 pset_insert_ptr(defs, irn);
1801 if(is_Proj(irn)) continue;
1804 * init set of irn's arguments
1805 * and all possibly used values around this op
1806 * and values defined by post remats
1808 args = new_set(cmp_keyval, get_irn_arity(irn));
1809 used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1810 remat_defs = pset_new_ptr(pset_count(live));
1812 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1813 ir_node *irn_arg = get_irn_n(irn, n);
1814 if(has_reg_class(si, irn_arg)) {
1815 set_insert_keyval(args, irn_arg, (void*)n);
1816 pset_insert_ptr(used, irn_arg);
1819 foreach_post_remat(irn, tmp) {
1820 op_t *remat_op = get_irn_link(tmp);
1822 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1824 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1825 ir_node *remat_arg = get_irn_n(tmp, n);
1826 if(has_reg_class(si, remat_arg)) {
1827 pset_insert_ptr(used, remat_arg);
1831 foreach_pre_remat(si, irn, tmp) {
1832 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1833 ir_node *remat_arg = get_irn_n(tmp, n);
1834 if(has_reg_class(si, remat_arg)) {
1835 pset_insert_ptr(used, remat_arg);
1840 /**********************************
1841 * I N E P I L O G O F irn
1842 **********************************/
1844 /* ensure each dying value is used by only one post remat */
1845 pset_foreach(used, tmp) {
1846 ir_node *value = tmp;
1847 op_t *value_op = get_irn_link(value);
1852 foreach_post_remat(irn, remat) {
1853 op_t *remat_op = get_irn_link(remat);
1855 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1856 ir_node *remat_arg = get_irn_n(remat, n);
1858 /* if value is used by this remat add it to constraint */
1859 if(remat_arg == value) {
1861 /* sum remat2s <= 1 + n_remats*live_range */
1862 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1863 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
1867 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1873 // value_op->attr.live_range.ilp != ILP_UNDEF
1874 if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
1875 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1879 /* ensure at least one value dies at post remat */
1880 foreach_post_remat(irn, tmp) {
1881 op_t *remat_op = get_irn_link(tmp);
1882 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
1885 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
1886 remat_arg = get_irn_n(tmp, n);
1888 if(has_reg_class(si, remat_arg)) {
1890 /* does arg always die at this op? */
1891 if(!pset_find_ptr(live, remat_arg))
1892 goto skip_one_must_die;
1894 pset_insert_ptr(remat_args, remat_arg);
1898 /* remat + \sum live_range(remat_arg) <= |args| */
1899 ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
1900 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
1901 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1903 pset_foreach(remat_args, remat_arg) {
1904 op_t *arg_op = get_irn_link(remat_arg);
1906 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1910 del_pset(remat_args);
1913 /* new live ranges for values from L\U defined by post remats */
1914 pset_foreach(live, tmp) {
1915 ir_node *value = tmp;
1916 op_t *value_op = get_irn_link(value);
1918 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
1919 ilp_var_t prev_lr = ILP_UNDEF;
1922 if(pset_find_ptr(remat_defs, value)) {
1924 /* next_live_range <= prev_live_range + sum remat2s */
1925 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1926 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1928 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1929 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1931 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1932 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1934 foreach_post_remat(irn, remat) {
1935 op_t *remat_op = get_irn_link(remat);
1937 /* if value is being rematerialized by this remat */
1938 if(value == remat_op->attr.remat.remat->value) {
1939 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1943 value_op->attr.live_range.ilp = prev_lr;
1944 value_op->attr.live_range.op = irn;
1949 /* requirements for post remats and start live ranges from L/U' for values dying here */
1950 foreach_post_remat(irn, tmp) {
1951 op_t *remat_op = get_irn_link(tmp);
1954 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1955 ir_node *remat_arg = get_irn_n(tmp, n);
1956 op_t *arg_op = get_irn_link(remat_arg);
1958 if(!has_reg_class(si, remat_arg)) continue;
1960 /* only for values in L\U (TODO and D?), the others are handled with post_use */
1961 if(!pset_find_ptr(used, remat_arg)) {
1962 /* remat <= live_range(remat_arg) */
1963 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
1964 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1966 /* if value is becoming live through use by remat2 */
1967 if(!pset_find_ptr(live, remat_arg)) {
1970 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1971 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1973 arg_op->attr.live_range.ilp = lr;
1974 arg_op->attr.live_range.op = irn;
1976 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
1978 pset_insert_ptr(live, remat_arg);
1979 add_to_spill_bb(si, bb, remat_arg);
1982 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1983 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1988 d = pset_count(defs);
1989 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
1991 /* count how many regs irn needs for arguments */
1992 u = set_count(args);
1995 /* check the register pressure in the epilog */
1996 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
1997 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
1998 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
2000 /* add L\U' to check_post */
2001 pset_foreach(live, tmp) {
2002 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
2003 /* if a live value is not used by irn */
2004 tmp_op = get_irn_link(tmp);
2005 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
2009 /***********************************************************
2010 * 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
2011 **********************************************************/
2014 pset_foreach(used, tmp) {
2020 op_t *arg_op = get_irn_link(arg);
2023 spill = add_to_spill_bb(si, bb, arg);
2025 /* new live range for each used value */
2026 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2027 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2029 /* the epilog stuff - including post_use, check_post, check_post_remat */
2030 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2031 post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2033 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2035 /* arg is live throughout epilog if the next live_range is in a register */
2036 if(pset_find_ptr(live, arg)) {
2037 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
2039 /* post_use >= next_lr + remat */
2040 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2041 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2042 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2043 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2047 /* if value is not an arg of op and not possibly defined by post remat
2048 * then it may only die and not become live
2050 if(!set_find_keyval(args, arg)) {
2051 /* post_use <= prev_lr */
2052 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2053 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2054 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2055 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2057 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2058 /* next_lr <= prev_lr */
2059 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2060 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2061 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2062 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2067 /* forall post remat which use arg add a similar cst */
2068 foreach_post_remat(irn, remat) {
2071 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2072 ir_node *remat_arg = get_irn_n(remat, n);
2073 op_t *remat_op = get_irn_link(remat);
2075 if(remat_arg == arg) {
2076 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
2078 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2079 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2080 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2081 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2086 if(opt_memoperands) {
2087 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2088 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2089 ilp_var_t memoperand;
2091 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2092 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0);
2093 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2095 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2096 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2098 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2099 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2100 // if(arg_op->attr.live_range.ilp != ILP_UNDEF)
2101 // lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2106 /* new live range begins for each used value */
2107 arg_op->attr.live_range.ilp = prev_lr;
2108 arg_op->attr.live_range.op = irn;
2110 /*if(!pset_find_ptr(live, arg)) {
2111 pset_insert_ptr(live, arg);
2112 add_to_spill_bb(si, bb, arg);
2114 pset_insert_ptr(live, arg);
2118 /* just to be sure */
2119 check_post = ILP_UNDEF;
2128 /* check the register pressure in the prolog */
2129 /* sum_{L\U} lr <= k - |U| */
2130 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2131 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2133 /* for the prolog remove defined values from the live set */
2134 pset_foreach(defs, tmp) {
2135 pset_remove_ptr(live, tmp);
2138 if(opt_memoperands) {
2139 ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2140 one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2143 /***********************************************************
2144 * 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
2145 **********************************************************/
2148 set_foreach(args, keyval) {
2150 const ir_node *arg = keyval->key;
2151 int i = PTR_TO_INT(keyval->val);
2152 op_t *arg_op = get_irn_link(arg);
2153 ilp_cst_t requirements;
2156 spill = set_find_spill(spill_bb->ilp, arg);
2159 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2160 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);
2162 /* reload <= mem_out */
2163 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2164 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2165 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2166 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2168 /* requirement: arg must be in register for use */
2169 /* reload + remat + live_range == 1 */
2170 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2171 requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2173 lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2174 lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2175 foreach_pre_remat(si, irn, tmp) {
2176 op_t *remat_op = get_irn_link(tmp);
2177 if(remat_op->attr.remat.remat->value == arg) {
2178 lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2182 if(opt_memoperands) {
2184 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2185 if(get_irn_n(irn, n) == arg) {
2189 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2190 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2191 memoperand_t *memoperand;
2192 memoperand = set_find_memoperand(si->memoperands, irn, n);
2194 /* memoperand <= mem_out */
2195 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2196 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2197 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2198 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2200 /* the memoperand is only sufficient if it is used once by the op */
2201 if(n_memoperands == 1)
2202 lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2204 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2206 /* we have one more free register if we use a memory operand */
2207 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2213 /* iterate over L\U */
2214 pset_foreach(live, tmp) {
2215 if(!set_find_keyval(args, tmp)) {
2216 /* if a live value is not used by irn */
2217 tmp_op = get_irn_link(tmp);
2218 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2223 /* requirements for remats */
2224 /* start new live ranges for values used by remats */
2225 foreach_pre_remat(si, irn, tmp) {
2226 op_t *remat_op = get_irn_link(tmp);
2229 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2230 ir_node *remat_arg = get_irn_n(tmp, n);
2231 op_t *arg_op = get_irn_link(remat_arg);
2233 if(!has_reg_class(si, remat_arg)) continue;
2235 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2236 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2237 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2239 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2240 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2242 /* if remat arg is also used by current op then we can use reload placed for this argument */
2243 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2244 int index = (int)keyval->val;
2246 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2254 /*************************
2255 * D O N E W I T H O P
2256 *************************/
2258 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2260 pset_foreach(live, tmp) {
2261 assert(has_reg_class(si, tmp));
2264 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2265 ir_node *arg = get_irn_n(irn, n);
2267 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2270 del_pset(remat_defs);
2274 defs = pset_new_ptr_default();
2279 /***************************************
2280 * B E G I N N I N G O F B L O C K
2281 ***************************************/
2284 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2285 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2287 pset_foreach(live, irn) {
2288 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2291 /* construct mem_outs for all values */
2293 set_foreach(spill_bb->ilp, spill) {
2294 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2295 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2297 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2298 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2300 if(pset_find_ptr(live, spill->irn)) {
2301 int default_spilled;
2302 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2304 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2305 default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2306 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2307 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2309 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2311 op_t *op = get_irn_link(spill->irn);
2313 /* do we have to copy a phi argument? */
2314 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2315 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2317 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2318 const ir_node *arg = get_irn_n(spill->irn, n);
2324 /* argument already done? */
2325 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2327 /* get sum of execution frequencies of blocks with the same phi argument */
2328 for(m=n; m>=0; --m) {
2329 const ir_node *arg2 = get_irn_n(spill->irn, m);
2332 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2336 /* copies are not for free */
2337 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2338 var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
2340 for(m=n; m>=0; --m) {
2341 const ir_node *arg2 = get_irn_n(spill->irn, m);
2344 op->attr.live_range.args.copies[m] = var;
2349 /* copy <= mem_in */
2350 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2351 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2352 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2353 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2361 /* L\U is empty at bb start */
2362 /* arg is live throughout epilog if it is reg_in into this block */
2364 /* check the register pressure at the beginning of the block
2367 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2368 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2370 pset_foreach(live, irn) {
2373 spill = set_find_spill(spill_bb->ilp, irn);
2376 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2377 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2379 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2381 /* spill + mem_in <= 1 */
2382 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2383 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2385 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2386 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2389 foreach_post_remat(bb, irn) {
2390 op_t *remat_op = get_irn_link(irn);
2392 DBG((si->dbg, LEVEL_4, "\t next post remat: %+F\n", irn));
2393 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
2395 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2398 /* forall post remats add requirements */
2399 foreach_post_remat(bb, tmp) {
2402 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2403 ir_node *remat_arg = get_irn_n(tmp, n);
2404 op_t *remat_op = get_irn_link(tmp);
2406 if(!has_reg_class(si, remat_arg)) continue;
2408 spill = set_find_spill(spill_bb->ilp, remat_arg);
2411 /* remat <= reg_in_argument */
2412 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2413 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2414 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2415 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2419 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2420 pset_foreach(live, irn) {
2424 spill = set_find_spill(spill_bb->ilp, irn);
2425 assert(spill && spill->irn == irn);
2427 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2428 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2431 ir_node *phi_arg = get_Phi_pred(irn, n);
2432 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2433 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2435 op_t *op = get_irn_link(irn);
2437 /* although the phi is in the right regclass one or more of
2438 * its arguments can be in a different one or at least to
2441 if(has_reg_class(si, phi_arg)) {
2442 /* mem_in < mem_out_arg + copy */
2443 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2444 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2446 /* reg_in < reg_out_arg */
2447 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2448 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2450 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2451 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2453 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2456 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2457 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2458 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2462 /* else assure the value arrives on all paths in the same resource */
2464 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2467 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2468 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2471 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2472 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2473 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2474 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2476 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2477 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2479 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2482 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2483 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2488 /* first live ranges from reg_ins */
2489 pset_foreach(live, irn) {
2490 op_t *op = get_irn_link(irn);
2492 spill = set_find_spill(spill_bb->ilp, irn);
2493 assert(spill && spill->irn == irn);
2495 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2496 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2497 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2498 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2500 foreach_post_remat(bb, tmp) {
2501 op_t *remat_op = get_irn_link(tmp);
2503 if(remat_op->attr.remat.remat->value == irn) {
2504 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2509 /* walk forward now and compute constraints for placing spills */
2510 /* this must only be done for values that are not defined in this block */
2511 /* TODO are these values at start of block? if yes, just check whether this is a diverge edge and skip the loop */
2512 pset_foreach(live, irn) {
2514 * if value is defined in this block we can anways place the spill directly after the def
2515 * -> no constraint necessary
2517 if(!is_Phi(irn) && get_nodes_block(irn) == bb) continue;
2520 spill = set_find_spill(spill_bb->ilp, irn);
2523 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2524 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2526 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2527 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2530 sched_foreach_op(bb, tmp) {
2531 op_t *op = get_irn_link(tmp);
2533 if(is_Phi(tmp)) continue;
2534 assert(!is_Proj(tmp));
2537 const ir_node *value = op->attr.remat.remat->value;
2540 /* only collect remats up to the first real use of a value */
2541 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2546 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2547 ir_node *arg = get_irn_n(tmp, n);
2550 /* if a value is used stop collecting remats */
2563 typedef struct _irnlist_t {
2564 struct list_head list;
2568 typedef struct _interference_t {
2569 struct list_head blocklist;
2575 cmp_interference(const void *a, const void *b, size_t size)
2577 const interference_t *p = a;
2578 const interference_t *q = b;
2580 return !(p->a == q->a && p->b == q->b);
2583 static interference_t *
2584 set_find_interference(set * set, ir_node * a, ir_node * b)
2586 interference_t query;
2588 query.a = (a>b)?a:b;
2589 query.b = (a>b)?b:a;
2591 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2594 static interference_t *
2595 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2597 interference_t query,
2599 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2603 result = set_find_interference(set, a, b);
2606 list_add(&list->list, &result->blocklist);
2610 query.a = (a>b)?a:b;
2611 query.b = (a>b)?b:a;
2613 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2615 INIT_LIST_HEAD(&result->blocklist);
2616 list_add(&list->list, &result->blocklist);
2622 values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
2624 const ir_edge_t *edge;
2626 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2627 /* both values are live in, so they interfere */
2631 /* ensure a dominates b */
2632 if(value_dominates(b,a)) {
2638 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2641 /* the following code is stolen from bera.c */
2642 if(be_is_live_end(si->lv, bb, a))
2645 foreach_out_edge(a, edge) {
2646 const ir_node *user = edge->src;
2647 if(get_nodes_block(user) == bb
2650 && !pset_find_ptr(si->inverse_ops, user)
2651 && value_dominates(b, user))
2659 * Walk all irg blocks and collect interfering values inside of phi classes
2662 luke_interferencewalker(ir_node * bb, void * data)
2664 spill_ilp_t *si = (spill_ilp_t*)data;
2667 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2668 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
2669 op_t *a_op = get_irn_link(a);
2672 /* a is only interesting if it is in my register class and if it is inside a phi class */
2673 if (has_reg_class(si, a) && get_phi_class(a)) {
2674 if(a_op->is_remat || pset_find_ptr(si->inverse_ops, a))
2677 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)) {
2678 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
2679 op_t *b_op = get_irn_link(b);
2682 /* a and b are only interesting if they are in the same phi class */
2683 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2684 if(b_op->is_remat || pset_find_ptr(si->inverse_ops, b))
2687 if(values_interfere_in_block(si, bb, a, b)) {
2688 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2689 set_insert_interference(si, si->interferences, a, b, bb);
2697 static unsigned int copy_path_id = 0;
2700 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2707 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2708 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2710 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2712 pset_foreach(copies, ptr) {
2713 copy = PTR_TO_INT(ptr);
2714 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2719 * @parameter copies contains a path of copies which lead us to irn
2720 * @parameter visited contains a set of nodes already visited on this path
2723 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2725 const ir_edge_t *edge;
2726 op_t *op = get_irn_link(irn);
2727 pset *visited_users = pset_new_ptr_default();
2730 if(op->is_remat) return 0;
2732 pset_insert_ptr(visited, irn);
2736 pset *visited_operands = pset_new_ptr(get_irn_arity(irn));
2738 /* visit all operands */
2739 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2740 ir_node *arg = get_irn_n(irn, n);
2741 ilp_var_t copy = op->attr.live_range.args.copies[n];
2743 if(!has_reg_class(si, arg)) continue;
2744 if(pset_find_ptr(visited_operands, arg)) continue;
2745 pset_insert_ptr(visited_operands, arg);
2748 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2749 del_pset(visited_operands);
2750 del_pset(visited_users);
2751 pset_remove_ptr(visited, irn);
2754 pset_insert(copies, INT_TO_PTR(copy), copy);
2755 write_copy_path_cst(si, copies, any_interfere);
2756 pset_remove(copies, INT_TO_PTR(copy), copy);
2757 } else if(!pset_find_ptr(visited, arg)) {
2758 pset_insert(copies, INT_TO_PTR(copy), copy);
2759 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
2760 pset_remove(copies, INT_TO_PTR(copy), copy);
2762 /*if(paths > MAX_PATHS) {
2763 if(pset_count(copies) == 0) {
2767 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2768 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2769 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2770 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2771 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2775 del_pset(visited_operands);
2776 del_pset(visited_users);
2777 pset_remove_ptr(visited, irn);
2780 } else if(pset_count(copies) == 0) {
2786 del_pset(visited_operands);
2789 /* visit all uses which are phis */
2790 foreach_out_edge(irn, edge) {
2791 ir_node *user = edge->src;
2792 int pos = edge->pos;
2793 op_t *op = get_irn_link(user);
2796 if(!is_Phi(user)) continue;
2797 if(!has_reg_class(si, user)) continue;
2798 if(pset_find_ptr(visited_users, user)) continue;
2799 pset_insert_ptr(visited_users, user);
2801 copy = op->attr.live_range.args.copies[pos];
2803 if(user == target) {
2804 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2805 del_pset(visited_users);
2806 pset_remove_ptr(visited, irn);
2809 pset_insert(copies, INT_TO_PTR(copy), copy);
2810 write_copy_path_cst(si, copies, any_interfere);
2811 pset_remove(copies, INT_TO_PTR(copy), copy);
2812 } else if(!pset_find_ptr(visited, user)) {
2813 pset_insert(copies, INT_TO_PTR(copy), copy);
2814 paths += find_copy_path(si, user, target, any_interfere, copies, visited);
2815 pset_remove(copies, INT_TO_PTR(copy), copy);
2817 /*if(paths > MAX_PATHS) {
2818 if(pset_count(copies) == 0) {
2822 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2823 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2824 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2825 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2826 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2830 del_pset(visited_users);
2831 pset_remove_ptr(visited, irn);
2834 } else if(pset_count(copies) == 0) {
2840 del_pset(visited_users);
2841 pset_remove_ptr(visited, irn);
2846 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
2848 pset * copies = pset_new_ptr_default();
2849 pset * visited = pset_new_ptr_default();
2851 find_copy_path(si, a, b, any_interfere, copies, visited);
2859 memcopyhandler(spill_ilp_t * si)
2861 interference_t *interference;
2863 /* teste Speicherwerte auf Interferenz */
2865 /* analyze phi classes */
2866 phi_class_compute(si->chordal_env->irg);
2868 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
2869 irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
2871 /* now lets emit the ILP unequations for the crap */
2872 set_foreach(si->interferences, interference) {
2874 ilp_var_t interfere,
2876 ilp_cst_t any_interfere_cst,
2878 const ir_node *a = interference->a;
2879 const ir_node *b = interference->b;
2881 /* any_interf <= \sum interf */
2882 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
2883 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2884 any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2886 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
2888 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
2889 const ir_node *bb = irnlist->irn;
2890 spill_bb_t *spill_bb = get_irn_link(bb);
2895 spilla = set_find_spill(spill_bb->ilp, a);
2898 spillb = set_find_spill(spill_bb->ilp, b);
2901 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
2902 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
2903 /* 2: - mem_in_a - spill_a + interfere <= 0 */
2904 /* 3: - mem_in_b - spill_b + interfere <= 0 */
2905 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
2906 interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2908 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
2909 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2911 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
2912 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
2913 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
2914 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
2915 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
2917 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
2918 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2920 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2921 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
2922 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
2924 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
2925 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2927 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2928 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
2929 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
2932 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
2934 /* any_interfere >= interf */
2935 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
2936 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2938 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2939 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2942 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
2943 gen_copy_constraints(si,a,b,any_interfere);
2951 return fabs(x) < 0.00001;
2954 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
2956 spill_ilp_t *si = get_irg_link(current_ir_graph);
2958 if(pset_find_ptr(si->all_possible_remats, n)) {
2959 op_t *op = (op_t*)get_irn_link(n);
2960 assert(op && op->is_remat);
2962 if(!op->attr.remat.remat->inverse) {
2963 if(op->attr.remat.pre) {
2964 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
2966 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
2971 op_t *op = (op_t*)get_irn_link(n);
2972 assert(op && op->is_remat);
2974 if(op->attr.remat.pre) {
2975 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
2977 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
2988 dump_graph_with_remats(ir_graph * irg, const char * suffix)
2990 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
2991 be_dump(irg, suffix, dump_ir_block_graph_sched);
2992 set_dump_node_vcgattr_hook(NULL);
2996 * Edge hook to dump the schedule edges with annotated register pressure.
2999 sched_pressure_edge_hook(FILE *F, ir_node *irn)
3001 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
3002 ir_node *prev = sched_prev(irn);
3003 fprintf(F, "edge:{sourcename:\"");
3005 fprintf(F, "\" targetname:\"");
3007 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
3008 fprintf(F, "\" color:magenta}\n");
3014 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
3016 DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3018 dump_consts_local(0);
3019 set_dump_node_edge_hook(sched_pressure_edge_hook);
3020 dump_ir_block_graph(irg, suffix);
3021 set_dump_node_edge_hook(old_edge_hook);
3025 walker_pressure_annotator(ir_node * bb, void * data)
3027 spill_ilp_t *si = data;
3030 pset *live = pset_new_ptr_default();
3033 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3034 irn = be_lv_get_irn(si->lv, bb, i);
3036 if (has_reg_class(si, irn)) {
3037 pset_insert_ptr(live, irn);
3041 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3043 sched_foreach_reverse(bb, irn) {
3045 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3049 if(has_reg_class(si, irn)) {
3050 pset_remove_ptr(live, irn);
3051 if(is_Proj(irn)) ++projs;
3054 if(!is_Proj(irn)) projs = 0;
3056 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3057 ir_node *arg = get_irn_n(irn, n);
3059 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3061 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3068 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3070 be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
3074 connect_all_remats_with_keep(spill_ilp_t * si)
3082 n_remats = pset_count(si->all_possible_remats);
3084 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3087 pset_foreach(si->all_possible_remats, irn) {
3092 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
3094 obstack_free(si->obst, ins);
3099 connect_all_spills_with_keep(spill_ilp_t * si)
3108 n_spills = pset_count(si->spills);
3110 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3113 pset_foreach(si->spills, irn) {
3118 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
3120 obstack_free(si->obst, ins);
3124 /** insert a spill at an arbitrary position */
3125 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3127 ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
3128 ir_graph *irg = get_irn_irg(bl);
3129 ir_node *frame = get_irg_frame(irg);
3133 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
3134 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3136 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3139 * search the right insertion point. a spill of a phi cannot be put
3140 * directly after the phi, if there are some phis behind the one which
3141 * is spilled. Also, a spill of a Proj must be after all Projs of the
3144 * Here's one special case:
3145 * If the spill is in the start block, the spill must be after the frame
3146 * pointer is set up. This is done by setting insert to the end of the block
3147 * which is its default initialization (see above).
3150 if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3153 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3156 sched_add_after(insert, spill);
3161 delete_remat(spill_ilp_t * si, ir_node * remat) {
3163 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3165 sched_remove(remat);
3167 /* kill links to operands */
3168 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3169 set_irn_n(remat, n, bad);
3174 clean_remat_info(spill_ilp_t * si)
3178 remat_info_t *remat_info;
3179 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3181 set_foreach(si->remat_info, remat_info) {
3182 if(!remat_info->remats) continue;
3184 pset_foreach(remat_info->remats, remat)
3186 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3187 set_irn_n((ir_node*)remat->proj, -1, bad);
3188 set_irn_n((ir_node*)remat->proj, 0, bad);
3191 if(get_irn_n_edges(remat->op) == 0) {
3192 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3193 set_irn_n((ir_node*)remat->op, n, bad);
3198 if(remat_info->remats) del_pset(remat_info->remats);
3199 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3204 delete_unnecessary_remats(spill_ilp_t * si)
3206 if(opt_keep_alive & KEEPALIVE_REMATS) {
3208 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3211 // ir_node *end = get_irg_end(si->chordal_env->irg);
3214 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3215 ir_node *keep_arg = get_irn_n(si->keep, n);
3216 op_t *arg_op = get_irn_link(keep_arg);
3219 assert(arg_op->is_remat);
3221 name = si->lpp->vars[arg_op->attr.remat.ilp];
3223 if(is_zero(name->value)) {
3224 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
3225 /* TODO check whether reload is preferred over remat (could be bug) */
3226 delete_remat(si, keep_arg);
3228 if(!arg_op->attr.remat.remat->inverse) {
3229 if(arg_op->attr.remat.pre) {
3230 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3232 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3235 if(arg_op->attr.remat.pre) {
3236 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3238 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3243 set_irn_n(si->keep, n, bad);
3246 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
3247 ir_node *end_arg = get_End_keepalive(end, i);
3249 if(end_arg != si->keep) {
3250 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
3253 keeps = obstack_finish(si->obst);
3254 set_End_keepalives(end, n-1, keeps);
3255 obstack_free(si->obst, keeps);
3258 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
3263 pset_foreach(si->all_possible_remats, remat) {
3264 op_t *remat_op = get_irn_link(remat);
3265 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
3267 if(is_zero(name->value)) {
3268 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
3269 /* TODO check whether reload is preferred over remat (could be bug) */
3270 delete_remat(si, remat);
3272 if(!remat_op->attr.remat.remat->inverse) {
3273 if(remat_op->attr.remat.pre) {
3274 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3276 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3279 if(remat_op->attr.remat.pre) {
3280 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3282 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3291 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3293 pset *spills = pset_new_ptr_default();
3295 const ir_node *next;
3298 defs = set_find_def(si->values, value);
3300 if(defs && defs->spills) {
3301 for(next = defs->spills; next; next = get_irn_link(next)) {
3302 pset_insert_ptr(spills, next);
3310 * @param before The node after which the spill will be placed in the schedule
3312 /* TODO set context properly */
3314 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3318 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3320 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3322 spill = be_spill2(arch_env, irn, before);
3324 defs = set_insert_def(si->values, value);
3327 /* enter into the linked list */
3328 set_irn_link(spill, defs->spills);
3329 defs->spills = spill;
3331 if(opt_keep_alive & KEEPALIVE_SPILLS)
3332 pset_insert_ptr(si->spills, spill);
3338 * @param before The Phi node which has to be spilled
3341 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3348 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3350 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3351 ins[n] = si->m_unknown;
3354 mem_phi = new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3356 defs = set_insert_def(si->values, phi);
3359 /* enter into the linked list */
3360 set_irn_link(mem_phi, defs->spills);
3361 defs->spills = mem_phi;
3363 sched_add_after(phi, mem_phi);
3365 if(opt_keep_alive & KEEPALIVE_SPILLS)
3366 pset_insert_ptr(si->spills, mem_phi);
3373 * Add remat to list of defs, destroys link field!
3376 insert_remat(spill_ilp_t * si, ir_node * remat)
3379 op_t *remat_op = get_irn_link(remat);
3381 assert(remat_op->is_remat);
3383 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3386 /* enter into the linked list */
3387 set_irn_link(remat, defs->remats);
3388 defs->remats = remat;
3393 * Add reload before operation and add to list of defs
3396 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3401 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3403 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3405 defs = set_find_def(si->values, value);
3407 spill = defs->spills;
3408 assert(spill && "no spill placed before reload");
3410 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3412 /* enter into the linked list */
3413 set_irn_link(reload, defs->remats);
3414 defs->remats = reload;
3419 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3423 ir_node *value = get_irn_n(memoperand->irn, memoperand->pos);
3425 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3427 DBG((si->dbg, LEVEL_2, "\t inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3429 defs = set_find_def(si->values, value);
3431 spill = defs->spills;
3432 assert(spill && "no spill placed before reload");
3434 reload = be_reload(arch_env, si->cls, memoperand->irn, get_irn_mode(value), spill);
3436 arch_perform_memory_operand(arch_env, memoperand->irn, reload, memoperand->pos);
3437 sched_remove(reload);
3440 void insert_memoperands(spill_ilp_t * si)
3442 memoperand_t *memoperand;
3445 set_foreach(si->memoperands, memoperand) {
3446 name = si->lpp->vars[memoperand->ilp];
3447 if(!is_zero(name->value)) {
3448 perform_memory_operand(si, memoperand);
3454 walker_spill_placer(ir_node * bb, void * data) {
3455 spill_ilp_t *si = (spill_ilp_t*)data;
3457 spill_bb_t *spill_bb = get_irn_link(bb);
3458 pset *spills_to_do = pset_new_ptr_default();
3461 set_foreach(spill_bb->ilp, spill) {
3464 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3465 name = si->lpp->vars[spill->mem_in];
3466 if(!is_zero(name->value)) {
3469 mem_phi = insert_mem_phi(si, spill->irn);
3471 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3475 name = si->lpp->vars[spill->spill];
3476 if(!is_zero(name->value)) {
3477 /* place spill directly after definition */
3478 if(get_nodes_block(spill->irn) == bb) {
3479 insert_spill(si, spill->irn, spill->irn, spill->irn);
3483 /* place spill at bb start */
3484 if(spill->reg_in > 0) {
3485 name = si->lpp->vars[spill->reg_in];
3486 if(!is_zero(name->value)) {
3487 insert_spill(si, spill->irn, spill->irn, bb);
3491 /* place spill after a remat */
3492 pset_insert_ptr(spills_to_do, spill->irn);
3495 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3498 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3499 op_t *op = get_irn_link(irn);
3501 if(be_is_Spill(irn)) continue;
3504 /* TODO fix this if we want to support remats with more than two nodes */
3505 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3506 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3508 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3511 if(pset_find_ptr(spills_to_do, irn)) {
3512 pset_remove_ptr(spills_to_do, irn);
3514 insert_spill(si, irn, irn, irn);
3520 assert(pset_count(spills_to_do) == 0);
3522 /* afterwards free data in block */
3523 del_pset(spills_to_do);
3527 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3529 ir_node *insert_pos = bb;
3531 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3533 /* find last definition of arg value in block */
3538 defs = set_find_def(si->values, value);
3540 if(defs && defs->remats) {
3541 for(next = defs->remats; next; next = get_irn_link(next)) {
3542 if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3543 last = sched_get_time_step(next);
3549 if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3550 last = sched_get_time_step(value);
3554 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
3556 spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3562 phim_fixer(spill_ilp_t *si) {
3565 set_foreach(si->values, defs) {
3566 const ir_node *phi = defs->value;
3567 op_t *op = get_irn_link(phi);
3568 ir_node *phi_m = NULL;
3569 ir_node *next = defs->spills;
3572 if(!is_Phi(phi)) continue;
3575 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3579 next = get_irn_link(next);
3582 if(!phi_m) continue;
3584 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3585 ir_node *value = get_irn_n(phi, n);
3586 defs_t *val_defs = set_find_def(si->values, value);
3588 /* a spill of this value */
3593 ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3594 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3596 if(!is_zero(name->value)) {
3597 spill = insert_mem_copy(si, pred, value);
3599 spill = val_defs->spills;
3602 spill = val_defs->spills;
3605 assert(spill && "no spill placed before PhiM");
3606 set_irn_n(phi_m, n, spill);
3612 walker_reload_placer(ir_node * bb, void * data) {
3613 spill_ilp_t *si = (spill_ilp_t*)data;
3615 spill_bb_t *spill_bb = get_irn_link(bb);
3617 /* reloads at end of block */
3618 if(spill_bb->reloads) {
3621 set_foreach(spill_bb->reloads, keyval) {
3622 ir_node *irn = (ir_node*)keyval->key;
3623 ilp_var_t reload = PTR_TO_INT(keyval->val);
3626 name = si->lpp->vars[reload];
3627 if(!is_zero(name->value)) {
3629 ir_node *insert_pos = bb;
3630 ir_node *prev = sched_block_last_noncf(si, bb);
3631 op_t *prev_op = get_irn_link(prev);
3633 while(be_is_Spill(prev)) {
3634 prev = sched_prev(prev);
3637 prev_op = get_irn_link(prev);
3639 /* insert reload before pre-remats */
3640 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3641 && prev_op->is_remat && prev_op->attr.remat.pre) {
3645 prev = sched_prev(prev);
3646 } while(be_is_Spill(prev));
3648 prev_op = get_irn_link(prev);
3652 reload = insert_reload(si, irn, insert_pos);
3654 if(opt_keep_alive & KEEPALIVE_RELOADS)
3655 pset_insert_ptr(si->spills, reload);
3660 /* walk and insert more reloads and collect remats */
3661 sched_foreach_reverse(bb, irn) {
3662 op_t *op = get_irn_link(irn);
3664 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3665 if(is_Phi(irn)) break;
3668 if(get_irn_mode(irn) != mode_T) {
3669 insert_remat(si, irn);
3674 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3675 ir_node *arg = get_irn_n(irn, n);
3677 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3680 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3681 if(!is_zero(name->value)) {
3683 ir_node *insert_pos = irn;
3684 ir_node *prev = sched_prev(insert_pos);
3687 while(be_is_Spill(prev)) {
3688 prev = sched_prev(prev);
3691 prev_op = get_irn_link(prev);
3693 /* insert reload before pre-remats */
3694 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3695 && prev_op->is_remat && prev_op->attr.remat.pre) {
3699 prev = sched_prev(prev);
3700 } while(be_is_Spill(prev));
3702 prev_op = get_irn_link(prev);
3706 reload = insert_reload(si, arg, insert_pos);
3708 set_irn_n(irn, n, reload);
3710 if(opt_keep_alive & KEEPALIVE_RELOADS)
3711 pset_insert_ptr(si->spills, reload);
3718 del_set(spill_bb->ilp);
3719 if(spill_bb->reloads) del_set(spill_bb->reloads);
3723 walker_collect_used(ir_node * irn, void * data)
3725 lc_bitset_t *used = data;
3727 lc_bitset_set(used, get_irn_idx(irn));
3730 struct kill_helper {
3736 walker_kill_unused(ir_node * bb, void * data)
3738 struct kill_helper *kh = data;
3739 ir_node *bad = get_irg_bad(get_irn_irg(bb));
3743 for(irn=sched_first(bb); !sched_is_end(irn);) {
3744 ir_node *next = sched_next(irn);
3747 if(!lc_bitset_is_set(kh->used, get_irn_idx(irn))) {
3748 if(be_is_Spill(irn) || be_is_Reload(irn)) {
3749 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)));
3751 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3757 set_nodes_block(irn, bad);
3758 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3759 set_irn_n(irn, n, bad);
3767 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3769 struct kill_helper kh;
3771 kh.used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3774 irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3775 irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3777 lc_bitset_free(kh.used);
3781 print_irn_pset(pset * p)
3785 pset_foreach(p, irn) {
3786 ir_printf("%+F\n", irn);
3791 dump_phi_class(spill_ilp_t * si, pset * phiclass, const char * file)
3793 FILE *f = fopen(file, "w");
3795 interference_t *interference;
3797 pset_break(phiclass);
3798 set_break(si->interferences);
3800 ir_fprintf(f, "digraph phiclass {\n");
3802 pset_foreach(phiclass, irn) {
3804 ir_fprintf(f, " %F%N [shape=box]\n",irn,irn);
3807 pset_foreach(phiclass, irn) {
3810 if(!is_Phi(irn)) continue;
3812 for(n=get_irn_arity(irn)-1; n>=0; --n) {
3813 ir_node *arg = get_irn_n(irn, n);
3815 ir_fprintf(f, " %F%N -> %F%N\n",irn,irn,arg,arg);
3819 set_foreach(si->interferences, interference) {
3820 const ir_node *a = interference->a;
3821 const ir_node *b = interference->b;
3822 if(get_phi_class(a) == phiclass) {
3823 ir_fprintf(f, " %F%N -> %F%N [color=red,dir=none,style=bold]\n",a,a,b,b);
3832 rewire_uses(spill_ilp_t * si)
3834 dom_front_info_t *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
3836 pset *ignore = pset_new_ptr(1);
3838 pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
3840 /* then fix uses of spills */
3841 set_foreach(si->values, defs) {
3844 const ir_node *next = defs->remats;
3847 reloads = pset_new_ptr_default();
3850 if(be_is_Reload(next)) {
3851 pset_insert_ptr(reloads, next);
3855 next = get_irn_link(next);
3858 spills = get_spills_for_value(si, defs->value);
3859 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));
3860 if(pset_count(spills) > 1) {
3861 //assert(pset_count(reloads) > 0);
3862 // print_irn_pset(spills);
3863 // print_irn_pset(reloads);
3865 be_ssa_constr_set_ignore(dfi, si->lv, spills, ignore);
3872 /* first fix uses of remats and reloads */
3873 set_foreach(si->values, defs) {
3875 const ir_node *next = defs->remats;
3878 nodes = pset_new_ptr_default();
3879 pset_insert_ptr(nodes, defs->value);
3882 pset_insert_ptr(nodes, next);
3883 next = get_irn_link(next);
3886 if(pset_count(nodes) > 1) {
3887 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
3888 be_ssa_constr_set(dfi, si->lv, nodes);
3895 // remove_unused_defs(si);
3897 be_free_dominance_frontiers(dfi);
3902 writeback_results(spill_ilp_t * si)
3904 /* walk through the graph and collect all spills, reloads and remats for a value */
3906 si->values = new_set(cmp_defs, 4096);
3908 DBG((si->dbg, LEVEL_1, "Applying results\n"));
3909 delete_unnecessary_remats(si);
3910 si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
3911 irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
3912 irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
3914 insert_memoperands(si);
3917 /* clean the remat info! there are still back-edges leading there! */
3918 clean_remat_info(si);
3922 connect_all_spills_with_keep(si);
3924 del_set(si->values);
3928 get_n_regs(spill_ilp_t * si)
3930 int arch_n_regs = arch_register_class_n_regs(si->cls);
3934 for(i=0; i<arch_n_regs; i++) {
3935 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
3940 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
3945 walker_reload_mover(ir_node * bb, void * data)
3947 spill_ilp_t *si = data;
3950 sched_foreach(bb, tmp) {
3951 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
3952 ir_node *reload = tmp;
3955 /* move reload upwards */
3957 int pressure = (int)get_irn_link(reload);
3958 if(pressure < si->n_regs) {
3959 irn = sched_prev(reload);
3960 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
3961 sched_remove(reload);
3962 pressure = (int)get_irn_link(irn);
3964 while(pressure < si->n_regs) {
3965 if( sched_is_end(irn) ||
3966 (be_is_Reload(irn) && has_reg_class(si, irn)) ||
3967 /* do not move reload before its spill */
3968 (irn == be_get_Reload_mem(reload)) ||
3969 /* do not move before phi */
3972 set_irn_link(irn, INT_TO_PTR(pressure+1));
3973 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
3974 irn = sched_prev(irn);
3976 pressure = (int)get_irn_link(irn);
3979 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
3980 sched_put_after(irn, reload);
3987 move_reloads_upward(spill_ilp_t * si)
3989 irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
3994 * Walk all irg blocks and check for interfering spills inside of phi classes
3997 luke_meminterferencechecker(ir_node * bb, void * data)
3999 spill_ilp_t *si = (spill_ilp_t*)data;
4002 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
4003 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
4005 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
4007 /* a is only interesting if it is in my register class and if it is inside a phi class */
4008 if (has_reg_class(si, a) && get_phi_class(a)) {
4009 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)) {
4010 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
4012 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
4014 /* a and b are only interesting if they are in the same phi class */
4015 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
4016 if(values_interfere_in_block(si, bb, a, b)) {
4017 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4026 verify_phiclasses(spill_ilp_t * si)
4028 /* analyze phi classes */
4029 phi_class_compute(si->chordal_env->irg);
4031 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4032 irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
4036 walker_spillslotassigner(ir_node * irn, void * data)
4040 if(!be_is_Spill(irn)) return;
4042 /* set spill context to phi class if it has one ;) */
4045 // Matze: not needed anymore
4046 cls = get_phi_class(irn);
4048 be_set_Spill_context(irn, cls);
4050 be_set_Spill_context(irn, irn);
4056 assign_spillslots(spill_ilp_t * si)
4058 DBG((si->dbg, LEVEL_2, "\t calling spill slot assigner\n"));
4059 irg_walk_graph(si->chordal_env->irg, walker_spillslotassigner, NULL, si);
4063 be_spill_remat(const be_chordal_env_t * chordal_env)
4066 char problem_name[256];
4067 char dump_suffix[256];
4068 char dump_suffix2[256];
4069 struct obstack obst;
4072 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
4073 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
4074 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
4076 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4077 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4079 if(opt_verify & VERIFY_DOMINANCE)
4080 be_check_dominance(chordal_env->irg);
4082 obstack_init(&obst);
4083 si.chordal_env = chordal_env;
4085 si.cls = chordal_env->cls;
4086 si.lpp = new_lpp(problem_name, lpp_minimize);
4087 si.remat_info = new_set(cmp_remat_info, 4096);
4088 si.interferences = new_set(cmp_interference, 32);
4089 si.memoperands = new_set(cmp_memoperands, 128);
4090 si.all_possible_remats = pset_new_ptr_default();
4091 si.spills = pset_new_ptr_default();
4092 si.inverse_ops = pset_new_ptr_default();
4093 si.lv = chordal_env->lv;
4095 si.n_regs = get_n_regs(&si);
4097 set_irg_link(chordal_env->irg, &si);
4098 compute_doms(chordal_env->irg);
4100 /* compute phi classes */
4101 // phi_class_compute(chordal_env->irg);
4103 be_analyze_regpressure(chordal_env, "-pre");
4106 /* collect remats */
4107 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4108 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
4111 /* insert possible remats */
4112 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4113 irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
4114 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4116 if(opt_keep_alive & KEEPALIVE_REMATS) {
4117 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4118 connect_all_remats_with_keep(&si);
4119 /* dump graph with inserted remats */
4120 dump_graph_with_remats(chordal_env->irg, dump_suffix);
4123 /* insert copies for phi arguments not in my regclass */
4124 irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
4126 /* recompute liveness */
4127 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4128 be_liveness_recompute(si.lv);
4132 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4133 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4134 irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
4136 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4137 irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
4140 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4141 memcopyhandler(&si);
4144 if(opt_dump_flags & DUMP_PROBLEM) {
4146 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4147 if ((f = fopen(buf, "wt")) != NULL) {
4148 lpp_dump_plain(si.lpp, f);
4153 if(opt_dump_flags & DUMP_MPS) {
4156 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4157 if((f = fopen(buf, "wt")) != NULL) {
4158 mps_write_mps(si.lpp, s_mps_fixed, f);
4162 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4163 if((f = fopen(buf, "wt")) != NULL) {
4164 mps_write_mst(si.lpp, s_mps_fixed, f);
4169 lpp_check_startvals(si.lpp);
4172 DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4173 lpp_set_time_limit(si.lpp, opt_timeout);
4176 lpp_set_log(si.lpp, stdout);
4179 lpp_solve_cplex(si.lpp);
4181 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4183 assert(lpp_is_sol_valid(si.lpp)
4184 && "solution of ILP must be valid");
4186 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));
4188 if(opt_dump_flags & DUMP_SOLUTION) {
4192 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4193 if ((f = fopen(buf, "wt")) != NULL) {
4195 for (i = 0; i < si.lpp->var_next; ++i) {
4196 lpp_name_t *name = si.lpp->vars[i];
4197 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4203 writeback_results(&si);
4207 kill_all_unused_values_in_schedule(&si);
4209 if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
4210 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
4212 // move reloads upwards
4213 be_liveness_recompute(si.lv);
4214 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4215 move_reloads_upward(&si);
4218 verify_phiclasses(&si);
4219 assign_spillslots(&si);
4222 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4224 dump_pressure_graph(&si, dump_suffix2);
4226 be_analyze_regpressure(chordal_env, "-post");
4228 if(opt_verify & VERIFY_DOMINANCE)
4229 be_check_dominance(chordal_env->irg);
4231 free_dom(chordal_env->irg);
4232 del_set(si.interferences);
4233 del_pset(si.inverse_ops);
4234 del_pset(si.all_possible_remats);
4235 del_set(si.memoperands);
4236 del_pset(si.spills);
4238 obstack_free(&obst, NULL);
4239 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4242 #else /* WITH_ILP */
4245 only_that_you_can_compile_without_WITH_ILP_defined(void)
4249 #endif /* WITH_ILP */