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"
52 #include "bespillremat.h"
54 #include "bepressurestat.h"
56 #include "bechordal_t.h"
59 #include <libcore/lc_opts.h>
60 #include <libcore/lc_opts_enum.h>
61 #endif /* WITH_LIBCORE */
63 #define DUMP_PROBLEM 1
65 #define DUMP_SOLUTION 4
67 #define KEEPALIVE_REMATS 1
68 #define KEEPALIVE_SPILLS 2
69 #define KEEPALIVE_RELOADS 4
71 #define VERIFY_MEMINTERF 1
72 #define VERIFY_DOMINANCE 2
75 #define REMATS_BRIGGS 1
76 #define REMATS_NOINVERSE 2
79 static int opt_dump_flags = 0;
80 static int opt_log = 0;
81 static int opt_keep_alive = 0;
82 static int opt_goodwin = 1;
83 static int opt_memcopies = 1;
84 static int opt_memoperands = 1;
85 static int opt_verify = VERIFY_MEMINTERF;
86 static int opt_remats = REMATS_ALL;
87 static int opt_repair_schedule = 0;
88 static int opt_no_enlarge_liveness = 0;
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),
141 LC_OPT_ENT_ENUM_MASK("dump", "dump problem, mps or solution", &dump_var),
142 LC_OPT_ENT_BOOL ("log", "activate the lpp log", &opt_log),
143 LC_OPT_ENT_INT ("timeout", "ILP solver timeout", &opt_timeout),
145 LC_OPT_ENT_DBL ("cost_reload", "cost of a reload", &opt_cost_reload),
146 LC_OPT_ENT_DBL ("cost_memoperand", "cost of a memory operand", &opt_cost_memoperand),
147 LC_OPT_ENT_DBL ("cost_spill", "cost of a spill instruction", &opt_cost_spill),
148 LC_OPT_ENT_DBL ("cost_remat", "cost of a rematerialization", &opt_cost_remat),
152 void be_spill_remat_register_options(lc_opt_entry_t *grp)
154 lc_opt_entry_t *my_grp = lc_opt_get_grp(grp, "remat");
155 lc_opt_add_table(my_grp, options);
160 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
161 //#define SCHEDULE_PHIM /* insert phim nodes into schedule */
164 //#define SOLVE_LOCAL
165 #define LPP_SERVER "i44pc52"
166 #define LPP_SOLVER "cplex"
169 #define MAX_PATHS INT_MAX
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_Proj_pred(op);
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)
605 irn = get_Proj_pred(irn);
607 for(n=get_irn_arity(irn)-1; n>=0; --n) {
608 const ir_node *arg = get_irn_n(irn, n);
610 if(has_reg_class(si, arg)) ++ret;
617 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
622 if( has_reg_class(si, op)
623 && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1)
624 && (opt_remats != REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0)
626 remat = get_remat_from_op(si, op, op);
628 add_remat(si, remat);
632 if(opt_remats == REMATS_ALL) {
633 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
635 for (n = get_irn_arity(op)-1; n>=0; --n) {
636 ir_node *arg = get_irn_n(op, n);
638 if(has_reg_class(si, arg)) {
639 /* try to get an inverse remat */
640 remat = get_remat_from_op(si, arg, op);
642 add_remat(si, remat);
650 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
653 ir_node *def_block = get_nodes_block(val);
659 /* if pos is at end of a basic block */
661 ret = (pos == def_block || block_dominates(def_block, pos));
662 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
666 /* else if this is a normal operation */
667 block = get_nodes_block(pos);
668 if(block == def_block) {
669 if(!sched_is_scheduled(val)) return 1;
671 ret = sched_comes_after(val, pos);
672 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
676 ret = block_dominates(def_block, block);
677 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
681 static INLINE ir_node *
682 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
684 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
688 * Returns first non-Phi node of block @p bb
690 static INLINE ir_node *
691 sched_block_first_nonphi(const ir_node * bb)
693 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
697 sched_skip_proj_predicator(const ir_node * irn, void * data)
699 return (is_Proj(irn));
702 static INLINE ir_node *
703 sched_next_nonproj(const ir_node * irn, int forward)
705 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
709 * Returns next operation node (non-Proj) after @p irn
710 * or the basic block of this node
712 static INLINE ir_node *
713 sched_next_op(const ir_node * irn)
715 ir_node *next = sched_next(irn);
720 return sched_next_nonproj(next, 1);
724 * Returns previous operation node (non-Proj) before @p irn
725 * or the basic block of this node
727 static INLINE ir_node *
728 sched_prev_op(const ir_node * irn)
730 ir_node *prev = sched_prev(irn);
735 return sched_next_nonproj(prev, 0);
739 sched_put_after(ir_node * insert, ir_node * irn)
741 if(is_Block(insert)) {
742 insert = sched_block_first_nonphi(insert);
744 insert = sched_next_op(insert);
746 sched_add_before(insert, irn);
750 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
752 if(is_Block(insert)) {
753 insert = sched_block_last_noncf(si, insert);
755 insert = sched_next_nonproj(insert, 0);
756 insert = sched_prev(insert);
758 sched_add_after(insert, irn);
762 * Tells you whether a @p remat can be placed before the irn @p pos
765 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
767 const ir_node *op = remat->op;
773 prev = sched_block_last_noncf(si, pos);
774 prev = sched_next_nonproj(prev, 0);
776 prev = sched_prev_op(pos);
778 /* do not remat if the rematted value is defined immediately before this op */
779 if(prev == remat->op) {
784 /* this should be just fine, the following OP will be using this value, right? */
786 /* only remat AFTER the real definition of a value (?) */
787 if(!value_is_defined_before(si, pos, remat->value)) {
788 // ir_fprintf(stderr, "error(not defined)");
793 for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
794 const ir_node *arg = get_irn_n(op, n);
796 if(opt_no_enlarge_liveness) {
797 if(has_reg_class(si, arg) && live) {
798 res &= pset_find_ptr((pset*)live, arg)?1:0;
800 res &= value_is_defined_before(si, pos, arg);
803 res &= value_is_defined_before(si, pos, arg);
811 * Tells you whether a @p remat can be placed after the irn @p pos
814 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
817 pos = sched_block_first_nonphi(pos);
819 pos = sched_next_op(pos);
822 /* only remat AFTER the real definition of a value (?) */
823 if(!value_is_defined_before(si, pos, remat->value)) {
827 return can_remat_before(si, remat, pos, live);
831 * Collect potetially rematerializable OPs
834 walker_remat_collector(ir_node * irn, void * data)
836 spill_ilp_t *si = data;
838 if(!is_Block(irn) && !is_Phi(irn)) {
839 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
840 get_remats_from_op(si, irn);
845 * Inserts a copy of @p irn before @p pos
848 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
853 bb = is_Block(pos)?pos:get_nodes_block(pos);
854 copy = exact_copy(irn);
856 _set_phi_class(copy, NULL);
857 set_nodes_block(copy, bb);
858 sched_put_before(si, pos, copy);
864 * Inserts a copy of @p irn after @p pos
867 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
872 bb = is_Block(pos)?pos:get_nodes_block(pos);
873 copy = exact_copy(irn);
875 _set_phi_class(copy, NULL);
876 set_nodes_block(copy, bb);
877 sched_put_after(pos, copy);
883 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
887 if(can_remat_after(si, remat, pos, live)) {
892 DBG((si->dbg, LEVEL_3, "\t >inserting remat2 %+F\n", remat->op));
894 copy = insert_copy_after(si, remat->op, pos);
896 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
897 op = obstack_alloc(si->obst, sizeof(*op));
899 op->attr.remat.remat = remat;
900 op->attr.remat.pre = 0;
901 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
903 set_irn_link(copy, op);
904 pset_insert_ptr(si->all_possible_remats, copy);
906 proj_copy = insert_copy_after(si, remat->proj, copy);
907 set_irn_n(proj_copy, 0, copy);
908 set_irn_link(proj_copy, op);
909 pset_insert_ptr(si->all_possible_remats, proj_copy);
921 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
925 if(can_remat_before(si, remat, pos, live)) {
930 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
932 copy = insert_copy_before(si, remat->op, pos);
934 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
935 op = obstack_alloc(si->obst, sizeof(*op));
937 op->attr.remat.remat = remat;
938 op->attr.remat.pre = 1;
939 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
941 set_irn_link(copy, op);
942 pset_insert_ptr(si->all_possible_remats, copy);
944 proj_copy = insert_copy_after(si, remat->proj, copy);
945 set_irn_n(proj_copy, 0, copy);
946 set_irn_link(proj_copy, op);
947 pset_insert_ptr(si->all_possible_remats, proj_copy);
959 get_block_n_succs(const ir_node *block) {
960 const ir_edge_t *edge;
962 assert(edges_activated(current_ir_graph));
964 edge = get_block_succ_first(block);
968 edge = get_block_succ_next(block, edge);
973 is_start_block(const ir_node * bb)
975 return get_irg_start_block(get_irn_irg(bb)) == bb;
979 is_merge_edge(const ir_node * bb)
981 if(is_start_block(bb))
985 return get_block_n_succs(bb) == 1;
991 is_diverge_edge(const ir_node * bb)
993 if(is_start_block(bb))
997 return get_Block_n_cfgpreds(bb) == 1;
1003 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1008 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1009 irn = be_lv_get_irn(si->lv, bb, i);
1011 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1012 pset_insert_ptr(live, irn);
1016 irn = sched_last(bb);
1018 /* all values eaten by control flow operations are also live until the end of the block */
1019 sched_foreach_reverse(bb, irn) {
1022 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1024 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1025 ir_node *arg = get_irn_n(irn,i);
1027 if(has_reg_class(si, arg)) {
1028 pset_insert_ptr(live, arg);
1035 walker_regclass_copy_insertor(ir_node * irn, void * data)
1037 spill_ilp_t *si = data;
1039 if(is_Phi(irn) && has_reg_class(si, irn)) {
1042 for(n=get_irn_arity(irn)-1; n>=0; --n) {
1043 ir_node *phi_arg = get_irn_n(irn, n);
1044 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
1046 if(!has_reg_class(si, phi_arg)) {
1047 ir_node *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
1048 ir_node *pos = sched_block_last_noncf(si, bb);
1049 op_t *op = obstack_alloc(si->obst, sizeof(*op));
1051 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
1052 sched_add_after(pos, copy);
1053 set_irn_n(irn, n, copy);
1056 op->attr.live_range.args.reloads = NULL;
1057 op->attr.live_range.ilp = ILP_UNDEF;
1058 set_irn_link(copy, op);
1065 * Insert (so far unused) remats into the irg to
1066 * recompute the potential liveness of all values
1069 walker_remat_insertor(ir_node * bb, void * data)
1071 spill_ilp_t *si = data;
1078 /* skip start block, no remats to do there */
1079 if(is_start_block(bb)) return;
1081 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1083 live = pset_new_ptr_default();
1084 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1085 ir_node *value = be_lv_get_irn(si->lv, bb, i);
1087 /* add remats at end of block */
1088 if (has_reg_class(si, value)) {
1089 pset_insert_ptr(live, value);
1093 irn = sched_last(bb);
1094 while(!sched_is_end(irn)) {
1100 next = sched_prev(irn);
1102 /* delete defined value from live set */
1103 if(has_reg_class(si, irn)) {
1104 pset_remove_ptr(live, irn);
1107 if(is_Phi(irn) || is_Proj(irn)) {
1112 args = pset_new_ptr_default();
1113 used = pset_new_ptr_default();
1115 /* collect arguments of op and set args of op already live in epilog */
1116 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1117 ir_node *arg = get_irn_n(irn, n);
1119 pset_insert_ptr(args, arg);
1120 if(has_reg_class(si, arg)) {
1121 pset_insert_ptr(live, arg);
1122 pset_insert_ptr(used, arg);
1126 /* insert all possible remats before irn */
1127 pset_foreach(args, arg) {
1128 remat_info_t *remat_info,
1131 /* continue if the operand has the wrong reg class */
1132 if(!has_reg_class(si, arg))
1136 query.remats = NULL;
1137 query.remats_by_operand = NULL;
1138 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1144 if(remat_info->remats) {
1145 pset_foreach(remat_info->remats, remat) {
1146 ir_node *remat_irn = NULL;
1148 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
1149 remat_irn = insert_remat_before(si, remat, irn, live);
1152 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1153 ir_node *remat_arg = get_irn_n(remat_irn, n);
1155 /* collect args of remats which are not args of op */
1156 if(has_reg_class(si, remat_arg) && !pset_find_ptr(args, remat_arg)) {
1157 pset_insert_ptr(used, remat_arg);
1165 /* do not place post remats after jumps */
1166 if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) {
1172 /* insert all possible remats after irn */
1173 post_remats = pset_new_ptr_default();
1174 pset_foreach(used, arg) {
1175 remat_info_t *remat_info,
1178 /* continue if the operand has the wrong reg class */
1179 if(!has_reg_class(si, arg))
1183 query.remats = NULL;
1184 query.remats_by_operand = NULL;
1185 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1191 if(remat_info->remats_by_operand) {
1192 pset_foreach(remat_info->remats_by_operand, remat) {
1193 /* do not insert remats producing the same value as one of the operands */
1194 if(!pset_find_ptr(args, remat->value)) {
1195 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
1197 /* only remat values that can be used be real ops */
1198 if(pset_find_ptr(live, remat->value)) {
1199 pset_insert_ptr(post_remats, remat);
1205 pset_foreach(post_remats, remat) {
1206 insert_remat_after(si, remat, irn, live);
1208 del_pset(post_remats);
1215 /* add remats at end if successor has multiple predecessors */
1216 if(is_merge_edge(bb)) {
1217 pset *live_out = pset_new_ptr_default();
1220 get_live_end(si, bb, live_out);
1222 /* add remats at end of block */
1223 pset_foreach(live_out, value) {
1224 remat_info_t *remat_info,
1228 query.remats = NULL;
1229 query.remats_by_operand = NULL;
1230 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1232 if(remat_info && remat_info->remats) {
1233 pset_foreach(remat_info->remats, remat) {
1234 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1236 insert_remat_before(si, remat, bb, live_out);
1243 if(is_diverge_edge(bb)) {
1244 pset *live_in = pset_new_ptr_default();
1247 be_lv_foreach(si->lv, bb, be_lv_state_in, i) {
1248 value = be_lv_get_irn(si->lv, bb, i);
1250 if(has_reg_class(si, value)) {
1251 pset_insert_ptr(live_in, value);
1254 /* add phis to live_in */
1255 sched_foreach(bb, value) {
1256 if(!is_Phi(value)) break;
1258 if(has_reg_class(si, value)) {
1259 pset_insert_ptr(live_in, value);
1263 /* add remat2s at beginning of block */
1264 post_remats = pset_new_ptr_default();
1265 pset_foreach(live_in, value) {
1266 remat_info_t *remat_info,
1270 query.remats = NULL;
1271 query.remats_by_operand = NULL;
1272 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1274 if(remat_info && remat_info->remats_by_operand) {
1275 pset_foreach(remat_info->remats_by_operand, remat) {
1276 DBG((si->dbg, LEVEL_4, "\t considering remat2 %+F at beginning of block %+F\n", remat->op, bb));
1278 /* put the remat here if all its args are available and result is still live */
1279 if(pset_find_ptr(live_in, remat->value)) {
1280 pset_insert_ptr(post_remats, remat);
1285 pset_foreach(post_remats, remat) {
1286 insert_remat_after(si, remat, bb, live_in);
1288 del_pset(post_remats);
1294 can_be_copied(const ir_node * bb, const ir_node * irn)
1296 const ir_edge_t *edge = get_block_succ_first(bb);
1297 const ir_node *next_bb = edge->src;
1298 int pos = edge->pos;
1301 assert(is_merge_edge(bb));
1303 sched_foreach(next_bb, phi) {
1304 const ir_node *phi_arg;
1306 if(!is_Phi(phi)) break;
1308 phi_arg = get_irn_n(phi, pos);
1310 if(phi_arg == irn) {
1318 * Initialize additional node info
1321 luke_initializer(ir_node * bb, void * data)
1323 spill_ilp_t *si = (spill_ilp_t*)data;
1324 spill_bb_t *spill_bb;
1327 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1328 set_irn_link(bb, spill_bb);
1330 sched_foreach(bb, irn) {
1333 op = obstack_alloc(si->obst, sizeof(*op));
1335 op->attr.live_range.ilp = ILP_UNDEF;
1338 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1339 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1341 } else if(!is_Proj(irn)) {
1342 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1343 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1345 op->attr.live_range.args.reloads = NULL;
1347 set_irn_link(irn, op);
1353 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1356 luke_endwalker(ir_node * bb, void * data)
1358 spill_ilp_t *si = (spill_ilp_t*)data;
1364 spill_bb_t *spill_bb = get_irn_link(bb);
1367 live = pset_new_ptr_default();
1368 use_end = pset_new_ptr_default();
1370 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1371 irn = be_lv_get_irn(si->lv, bb, i);
1372 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1373 pset_insert_ptr(live, irn);
1377 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1378 /* their reg_out must always be set */
1379 sched_foreach_reverse(bb, irn) {
1382 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1384 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1385 ir_node *irn_arg = get_irn_n(irn, n);
1387 if(has_reg_class(si, irn_arg)) {
1388 pset_insert_ptr(use_end, irn_arg);
1393 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1394 //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1395 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1397 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1399 /* if this is a merge edge we can reload at the end of this block */
1400 if(is_merge_edge(bb)) {
1401 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1402 } else if(pset_count(use_end)){
1403 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1405 spill_bb->reloads = NULL;
1408 pset_foreach(live,irn) {
1412 int default_spilled;
1415 /* handle values used by control flow nodes later separately */
1416 if(pset_find_ptr(use_end, irn)) continue;
1419 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1421 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1423 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1424 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1425 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1427 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1428 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1430 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1431 /* by default spill value right after definition */
1432 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1433 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1435 if(is_merge_edge(bb)) {
1439 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1440 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), can_be_copied(bb, irn));
1441 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1443 /* reload <= mem_out */
1444 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1445 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1446 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1449 spill->reg_in = ILP_UNDEF;
1450 spill->mem_in = ILP_UNDEF;
1453 pset_foreach(use_end,irn) {
1457 ilp_cst_t end_use_req,
1460 int default_spilled;
1463 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1465 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1467 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1468 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1469 /* if irn is used at the end of the block, then it is live anyway */
1470 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1472 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1473 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1475 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1476 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1477 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1479 /* reload for use be control flow op */
1480 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1481 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1482 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1484 /* reload <= mem_out */
1485 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1486 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1487 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1489 spill->reg_in = ILP_UNDEF;
1490 spill->mem_in = ILP_UNDEF;
1492 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1493 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1494 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1502 next_post_remat(const ir_node * irn)
1508 next = sched_block_first_nonphi(irn);
1510 next = sched_next_op(irn);
1513 if(sched_is_end(next))
1516 op = get_irn_link(next);
1517 if(op->is_remat && !op->attr.remat.pre) {
1526 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1532 ret = sched_block_last_noncf(si, irn);
1533 ret = sched_next(ret);
1534 ret = sched_prev_op(ret);
1536 ret = sched_prev_op(irn);
1539 if(sched_is_end(ret) || is_Phi(ret))
1542 op = (op_t*)get_irn_link(ret);
1543 if(op->is_remat && op->attr.remat.pre) {
1551 * Find a remat of value @p value in the epilog of @p pos
1554 find_post_remat(const ir_node * value, const ir_node * pos)
1556 while((pos = next_post_remat(pos)) != NULL) {
1559 op = get_irn_link(pos);
1560 assert(op->is_remat && !op->attr.remat.pre);
1562 if(op->attr.remat.remat->value == value)
1563 return (ir_node*)pos;
1566 const ir_edge_t *edge;
1567 foreach_out_edge(pos, edge) {
1568 ir_node *proj = get_edge_src_irn(edge);
1569 assert(is_Proj(proj));
1579 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1581 spill_bb_t *spill_bb = get_irn_link(bb);
1585 int default_spilled;
1588 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1590 double spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1592 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1594 spill->reg_out = ILP_UNDEF;
1595 spill->reg_in = ILP_UNDEF;
1596 spill->mem_in = ILP_UNDEF;
1598 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1599 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1601 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1602 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1603 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1610 * Inserts ILP-constraints and variables for memory copying before the given position
1613 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1615 const ir_node *succ;
1616 const ir_edge_t *edge;
1617 spill_bb_t *spill_bb = get_irn_link(block);
1626 assert(edges_activated(current_ir_graph));
1628 edge = get_block_succ_first(block);
1634 edge = get_block_succ_next(block, edge);
1635 /* next block can only contain phis, if this is a merge edge */
1638 ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1639 copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1641 ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1642 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1644 pset_foreach(live, tmp) {
1647 op_t *op = get_irn_link(irn);
1648 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1650 spill = set_find_spill(spill_bb->ilp, tmp);
1653 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1655 lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1657 sched_foreach(succ, phi) {
1658 const ir_node *to_copy;
1660 spill_t *to_copy_spill;
1661 op_t *phi_op = get_irn_link(phi);
1662 ilp_var_t reload = ILP_UNDEF;
1665 if(!is_Phi(phi)) break;
1666 if(!has_reg_class(si, phi)) continue;
1668 to_copy = get_irn_n(phi, pos);
1669 to_copy_op = get_irn_link(to_copy);
1671 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1672 assert(to_copy_spill);
1674 if(spill_bb->reloads) {
1675 keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1678 reload = PTR_TO_INT(keyval->val);
1682 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1683 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1685 /* copy - reg_out - reload - remat - live_range <= 0 */
1686 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1687 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1688 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1689 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1690 foreach_pre_remat(si, block, tmp) {
1691 op_t *remat_op = get_irn_link(tmp);
1692 if(remat_op->attr.remat.remat->value == to_copy) {
1693 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1697 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1698 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1700 /* copy - reg_out - copyreg <= 0 */
1701 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1702 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1703 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1709 * Walk all irg blocks and emit this ILP
1712 luke_blockwalker(ir_node * bb, void * data)
1714 spill_ilp_t *si = (spill_ilp_t*)data;
1719 spill_bb_t *spill_bb = get_irn_link(bb);
1722 pset *defs = pset_new_ptr_default();
1723 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
1725 live = pset_new_ptr_default();
1727 /****************************************
1728 * B A S I C B L O C K E N D
1729 ***************************************/
1732 /* init live values at end of block */
1733 get_live_end(si, bb, live);
1735 pset_foreach(live, irn) {
1737 ilp_var_t reload = ILP_UNDEF;
1739 spill = set_find_spill(spill_bb->ilp, irn);
1742 if(spill_bb->reloads) {
1743 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1746 reload = PTR_TO_INT(keyval->val);
1750 op = get_irn_link(irn);
1751 assert(!op->is_remat);
1753 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1754 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1755 op->attr.live_range.op = bb;
1757 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1758 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1760 /* reg_out - reload - remat - live_range <= 0 */
1761 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1762 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1763 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1764 foreach_pre_remat(si, bb, tmp) {
1765 op_t *remat_op = get_irn_link(tmp);
1766 if(remat_op->attr.remat.remat->value == irn) {
1767 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1770 ir_snprintf(buf, sizeof(buf), "reg_out2_%N_%N", bb, irn);
1771 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
1773 /* value may only die at bb end if it is used for a mem copy */
1774 /* reg_out + \sum copy - reload - remat - live_range >= 0 */
1775 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1776 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1777 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1778 foreach_pre_remat(si, bb, tmp) {
1779 op_t *remat_op = get_irn_link(tmp);
1780 if(remat_op->attr.remat.remat->value == irn) {
1781 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1784 if(is_merge_edge(bb)) {
1785 const ir_edge_t *edge = get_block_succ_first(bb);
1786 const ir_node *next_bb = edge->src;
1787 int pos = edge->pos;
1790 sched_foreach(next_bb, phi) {
1791 const ir_node *phi_arg;
1793 if(!is_Phi(phi)) break;
1795 phi_arg = get_irn_n(phi, pos);
1797 if(phi_arg == irn) {
1798 op_t *phi_op = get_irn_link(phi);
1799 ilp_var_t copy = phi_op->attr.live_range.args.copies[pos];
1801 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
1808 insert_mem_copy_position(si, live, bb);
1811 * start new live ranges for values used by remats at end of block
1812 * and assure the remat args are available
1814 foreach_pre_remat(si, bb, tmp) {
1815 op_t *remat_op = get_irn_link(tmp);
1818 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1819 ir_node *remat_arg = get_irn_n(tmp, n);
1820 op_t *arg_op = get_irn_link(remat_arg);
1823 if(!has_reg_class(si, remat_arg)) continue;
1825 /* if value is becoming live through use by remat */
1826 if(!pset_find_ptr(live, remat_arg)) {
1827 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1828 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1830 arg_op->attr.live_range.ilp = prev_lr;
1831 arg_op->attr.live_range.op = bb;
1833 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1835 pset_insert_ptr(live, remat_arg);
1836 add_to_spill_bb(si, bb, remat_arg);
1839 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1840 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1841 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1843 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1844 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1846 /* use reload placed for this argument */
1847 if(spill_bb->reloads) {
1848 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1851 ilp_var_t reload = PTR_TO_INT(keyval->val);
1853 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1858 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1863 /**************************************
1864 * B A S I C B L O C K B O D Y
1865 **************************************/
1867 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1873 ilp_cst_t check_pre,
1879 ilp_cst_t one_memoperand;
1881 /* iterate only until first phi */
1885 op = get_irn_link(irn);
1887 if(op->is_remat) continue;
1889 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1891 /* collect defined values */
1892 if(has_reg_class(si, irn)) {
1893 pset_insert_ptr(defs, irn);
1897 if(is_Proj(irn)) continue;
1900 * init set of irn's arguments
1901 * and all possibly used values around this op
1902 * and values defined by post remats
1904 args = new_set(cmp_keyval, get_irn_arity(irn));
1905 used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1906 remat_defs = pset_new_ptr(pset_count(live));
1908 if(!is_start_block(bb) || !be_is_Barrier(irn)) {
1909 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1910 ir_node *irn_arg = get_irn_n(irn, n);
1911 if(has_reg_class(si, irn_arg)) {
1912 set_insert_keyval(args, irn_arg, (void*)n);
1913 pset_insert_ptr(used, irn_arg);
1916 foreach_post_remat(irn, tmp) {
1917 op_t *remat_op = get_irn_link(tmp);
1919 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1921 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1922 ir_node *remat_arg = get_irn_n(tmp, n);
1923 if(has_reg_class(si, remat_arg)) {
1924 pset_insert_ptr(used, remat_arg);
1928 foreach_pre_remat(si, irn, tmp) {
1929 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1930 ir_node *remat_arg = get_irn_n(tmp, n);
1931 if(has_reg_class(si, remat_arg)) {
1932 pset_insert_ptr(used, remat_arg);
1938 /**********************************
1939 * I N E P I L O G O F irn
1940 **********************************/
1942 /* ensure each dying value is used by only one post remat */
1943 pset_foreach(used, tmp) {
1944 ir_node *value = tmp;
1945 op_t *value_op = get_irn_link(value);
1950 foreach_post_remat(irn, remat) {
1951 op_t *remat_op = get_irn_link(remat);
1953 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1954 ir_node *remat_arg = get_irn_n(remat, n);
1956 /* if value is used by this remat add it to constraint */
1957 if(remat_arg == value) {
1959 /* sum remat2s <= 1 + n_remats*live_range */
1960 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1961 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
1965 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1971 if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
1972 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1976 /* ensure at least one value dies at post remat */
1977 foreach_post_remat(irn, tmp) {
1978 op_t *remat_op = get_irn_link(tmp);
1979 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
1982 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
1983 remat_arg = get_irn_n(tmp, n);
1985 if(has_reg_class(si, remat_arg)) {
1987 /* does arg always die at this op? */
1988 if(!pset_find_ptr(live, remat_arg))
1989 goto skip_one_must_die;
1991 pset_insert_ptr(remat_args, remat_arg);
1995 /* remat + \sum live_range(remat_arg) <= |args| */
1996 ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
1997 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
1998 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2000 pset_foreach(remat_args, remat_arg) {
2001 op_t *arg_op = get_irn_link(remat_arg);
2003 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2007 del_pset(remat_args);
2010 /* new live ranges for values from L\U defined by post remats */
2011 pset_foreach(live, tmp) {
2012 ir_node *value = tmp;
2013 op_t *value_op = get_irn_link(value);
2015 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
2016 ilp_var_t prev_lr = ILP_UNDEF;
2019 if(pset_find_ptr(remat_defs, value)) {
2021 /* next_live_range <= prev_live_range + sum remat2s */
2022 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
2023 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2025 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
2026 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2028 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
2029 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2031 foreach_post_remat(irn, remat) {
2032 op_t *remat_op = get_irn_link(remat);
2034 /* if value is being rematerialized by this remat */
2035 if(value == remat_op->attr.remat.remat->value) {
2036 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2040 value_op->attr.live_range.ilp = prev_lr;
2041 value_op->attr.live_range.op = irn;
2046 /* requirements for post remats and start live ranges from L/U' for values dying here */
2047 foreach_post_remat(irn, tmp) {
2048 op_t *remat_op = get_irn_link(tmp);
2051 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2052 ir_node *remat_arg = get_irn_n(tmp, n);
2053 op_t *arg_op = get_irn_link(remat_arg);
2055 if(!has_reg_class(si, remat_arg)) continue;
2057 /* only for values in L\U (TODO and D?), the others are handled with post_use */
2058 if(!pset_find_ptr(used, remat_arg)) {
2059 /* remat <= live_range(remat_arg) */
2060 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
2061 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2063 /* if value is becoming live through use by remat2 */
2064 if(!pset_find_ptr(live, remat_arg)) {
2067 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
2068 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2070 arg_op->attr.live_range.ilp = lr;
2071 arg_op->attr.live_range.op = irn;
2073 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
2075 pset_insert_ptr(live, remat_arg);
2076 add_to_spill_bb(si, bb, remat_arg);
2079 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2080 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2085 d = pset_count(defs);
2086 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
2088 /* count how many regs irn needs for arguments */
2089 u = set_count(args);
2092 /* check the register pressure in the epilog */
2093 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
2094 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
2095 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
2097 /* add L\U' to check_post */
2098 pset_foreach(live, tmp) {
2099 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
2100 /* if a live value is not used by irn */
2101 tmp_op = get_irn_link(tmp);
2102 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
2106 /***********************************************************
2107 * 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
2108 **********************************************************/
2111 pset_foreach(used, tmp) {
2117 op_t *arg_op = get_irn_link(arg);
2120 spill = add_to_spill_bb(si, bb, arg);
2122 /* new live range for each used value */
2123 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2124 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2126 /* the epilog stuff - including post_use, check_post, check_post_remat */
2127 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2128 post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2130 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2132 /* arg is live throughout epilog if the next live_range is in a register */
2133 if(pset_find_ptr(live, arg)) {
2134 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
2136 /* post_use >= next_lr + remat */
2137 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2138 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2139 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2140 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2143 /* forall post remat which use arg add a similar cst */
2144 foreach_post_remat(irn, remat) {
2147 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2148 ir_node *remat_arg = get_irn_n(remat, n);
2149 op_t *remat_op = get_irn_link(remat);
2151 if(remat_arg == arg) {
2152 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
2154 /* post_use >= remat */
2155 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2156 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2157 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2158 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2163 /* if value is not an arg of op and not possibly defined by post remat
2164 * then it may only die and not become live
2166 if(!set_find_keyval(args, arg)) {
2167 /* post_use <= prev_lr */
2168 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2169 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2170 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2171 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2173 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2174 /* next_lr <= prev_lr */
2175 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2176 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2177 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2178 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2182 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2183 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2184 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2185 ilp_var_t memoperand;
2187 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2188 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0);
2189 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2191 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2192 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2194 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2195 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2200 /* new live range begins for each used value */
2201 arg_op->attr.live_range.ilp = prev_lr;
2202 arg_op->attr.live_range.op = irn;
2204 pset_insert_ptr(live, arg);
2207 /* just to be sure */
2208 check_post = ILP_UNDEF;
2215 /* check the register pressure in the prolog */
2216 /* sum_{L\U} lr <= k - |U| */
2217 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2218 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2220 /* for the prolog remove defined values from the live set */
2221 pset_foreach(defs, tmp) {
2222 pset_remove_ptr(live, tmp);
2225 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2226 ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2227 one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2230 /***********************************************************
2231 * 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
2232 **********************************************************/
2235 set_foreach(args, keyval) {
2237 const ir_node *arg = keyval->key;
2238 int i = PTR_TO_INT(keyval->val);
2239 op_t *arg_op = get_irn_link(arg);
2240 ilp_cst_t requirements;
2243 spill = set_find_spill(spill_bb->ilp, arg);
2246 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2247 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);
2249 /* reload <= mem_out */
2250 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2251 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2252 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2253 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2255 /* requirement: arg must be in register for use */
2256 /* reload + remat + live_range == 1 */
2257 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2258 requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2260 lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2261 lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2262 foreach_pre_remat(si, irn, tmp) {
2263 op_t *remat_op = get_irn_link(tmp);
2264 if(remat_op->attr.remat.remat->value == arg) {
2265 lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2269 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2271 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2272 if(get_irn_n(irn, n) == arg) {
2276 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2277 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2278 memoperand_t *memoperand;
2279 memoperand = set_find_memoperand(si->memoperands, irn, n);
2281 /* memoperand <= mem_out */
2282 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2283 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2284 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2285 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2287 /* the memoperand is only sufficient if it is used once by the op */
2288 if(n_memoperands == 1)
2289 lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2291 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2293 /* we have one more free register if we use a memory operand */
2294 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2300 /* iterate over L\U */
2301 pset_foreach(live, tmp) {
2302 if(!set_find_keyval(args, tmp)) {
2303 /* if a live value is not used by irn */
2304 tmp_op = get_irn_link(tmp);
2305 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2309 /* requirements for remats */
2310 foreach_pre_remat(si, irn, tmp) {
2311 op_t *remat_op = get_irn_link(tmp);
2314 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2315 ir_node *remat_arg = get_irn_n(tmp, n);
2316 op_t *arg_op = get_irn_link(remat_arg);
2318 if(!has_reg_class(si, remat_arg)) continue;
2320 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2321 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2322 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2324 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2325 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2327 /* if remat arg is also used by current op then we can use reload placed for this argument */
2328 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2329 int index = (int)keyval->val;
2331 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2339 /*************************
2340 * D O N E W I T H O P
2341 *************************/
2343 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2345 pset_foreach(live, tmp) {
2346 assert(has_reg_class(si, tmp));
2349 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2350 ir_node *arg = get_irn_n(irn, n);
2352 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2355 del_pset(remat_defs);
2359 defs = pset_new_ptr_default();
2361 /* skip everything above barrier in start block */
2362 if(is_start_block(bb) && be_is_Barrier(irn)) {
2363 assert(pset_count(live) == 0);
2372 /***************************************
2373 * B E G I N N I N G O F B L O C K
2374 ***************************************/
2377 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2378 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2380 pset_foreach(live, irn) {
2381 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2384 /* construct mem_outs for all values */
2385 set_foreach(spill_bb->ilp, spill) {
2386 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2387 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2389 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2390 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2392 if(pset_find_ptr(live, spill->irn)) {
2393 int default_spilled;
2394 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2396 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2397 default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2398 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2399 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2401 if(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2403 op_t *op = get_irn_link(spill->irn);
2405 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2406 const ir_node *arg = get_irn_n(spill->irn, n);
2412 /* argument already done? */
2413 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2415 /* get sum of execution frequencies of blocks with the same phi argument */
2416 for(m=n; m>=0; --m) {
2417 const ir_node *arg2 = get_irn_n(spill->irn, m);
2420 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2424 /* copies are not for free */
2425 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2426 var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
2428 for(m=n; m>=0; --m) {
2429 const ir_node *arg2 = get_irn_n(spill->irn, m);
2432 op->attr.live_range.args.copies[m] = var;
2437 /* copy <= mem_in */
2438 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2439 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2440 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2441 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2448 foreach_post_remat(bb, tmp) {
2450 op_t *remat_op = get_irn_link(tmp);
2451 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
2454 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2455 remat_arg = get_irn_n(tmp, n);
2457 if(has_reg_class(si, remat_arg)) {
2458 pset_insert_ptr(remat_args, remat_arg);
2462 /* remat + \sum live_range(remat_arg) <= |args| */
2463 ir_snprintf(buf, sizeof(buf), "one_must_die_%N", tmp);
2464 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2465 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2467 pset_foreach(remat_args, remat_arg) {
2468 if(pset_find_ptr(live, remat_arg)) {
2469 op_t *remat_arg_op = get_irn_link(remat_arg);
2470 lpp_set_factor_fast(si->lpp, cst, remat_arg_op->attr.live_range.ilp, 1.0);
2473 del_pset(remat_args);
2476 foreach_post_remat(bb, tmp) {
2479 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2480 ir_node *remat_arg = get_irn_n(tmp, n);
2482 /* if value is becoming live through use by remat2 */
2483 if(has_reg_class(si, remat_arg) && !pset_find_ptr(live, remat_arg)) {
2484 op_t *remat_arg_op = get_irn_link(remat_arg);
2487 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 at bb start %+F\n", remat_arg, tmp));
2489 pset_insert_ptr(live, remat_arg);
2490 spill = add_to_spill_bb(si, bb, remat_arg);
2491 remat_arg_op->attr.live_range.ilp = ILP_UNDEF;
2493 /* we need reg_in and mem_in for this value; they will be referenced later */
2494 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", remat_arg, bb);
2495 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2496 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", remat_arg, bb);
2497 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2500 /* optimization: all memory stuff should be 0, for we do not want to insert reloads for remats */
2501 ir_snprintf(buf, sizeof(buf), "nomem_%N_%N", remat_arg, bb);
2502 nomem = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0.0);
2503 lpp_set_factor_fast(si->lpp, nomem, spill->spill, 1.0);
2508 /* L\U is empty at bb start */
2509 /* arg is live throughout epilog if it is reg_in into this block */
2511 /* check the register pressure at the beginning of the block
2514 /* reg_in entspricht post_use */
2516 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2517 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2519 pset_foreach(live, irn) {
2522 spill = set_find_spill(spill_bb->ilp, irn);
2525 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2526 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2528 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2530 /* spill + mem_in <= 1 */
2531 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2532 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2534 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2535 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2537 } /* post_remats are NOT included in register pressure check because
2538 they do not increase regpressure */
2540 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2541 pset_foreach(live, irn) {
2545 spill = set_find_spill(spill_bb->ilp, irn);
2546 assert(spill && spill->irn == irn);
2548 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2549 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2552 ir_node *phi_arg = get_Phi_pred(irn, n);
2553 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2554 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2556 op_t *op = get_irn_link(irn);
2558 /* although the phi is in the right regclass one or more of
2559 * its arguments can be in a different one or at least to
2562 if(has_reg_class(si, phi_arg)) {
2563 /* mem_in < mem_out_arg + copy */
2564 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2565 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2567 /* reg_in < reg_out_arg */
2568 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2569 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2571 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2572 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2574 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2577 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2579 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2581 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2585 /* else assure the value arrives on all paths in the same resource */
2587 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2590 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2591 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2594 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2595 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2596 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2597 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2599 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2600 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2602 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2605 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2606 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2611 foreach_post_remat(bb, tmp) {
2614 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2615 ir_node *remat_arg = get_irn_n(tmp, n);
2616 op_t *remat_op = get_irn_link(tmp);
2618 if(!has_reg_class(si, remat_arg)) continue;
2620 spill = set_find_spill(spill_bb->ilp, remat_arg);
2623 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2624 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2625 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2626 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2630 pset_foreach(live, irn) {
2631 const op_t *op = get_irn_link(irn);
2632 const ir_node *remat;
2637 foreach_post_remat(bb, remat) {
2640 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2641 const ir_node *arg = get_irn_n(remat, n);
2644 const op_t *remat_op = get_irn_link(remat);
2646 if(cst == ILP_UNDEF) {
2647 /* sum remat2s <= 1 + n_remats*live_range */
2648 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", irn, bb);
2649 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2651 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2657 if(cst != ILP_UNDEF && op->attr.live_range.ilp != ILP_UNDEF) {
2658 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -n_remats);
2662 /* first live ranges from reg_ins */
2663 pset_foreach(live, irn) {
2664 op_t *op = get_irn_link(irn);
2666 if(op->attr.live_range.ilp != ILP_UNDEF) {
2668 spill = set_find_spill(spill_bb->ilp, irn);
2669 assert(spill && spill->irn == irn);
2671 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2672 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2673 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2674 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2676 foreach_post_remat(bb, tmp) {
2677 op_t *remat_op = get_irn_link(tmp);
2679 if(remat_op->attr.remat.remat->value == irn) {
2680 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2686 /* walk forward now and compute constraints for placing spills */
2687 /* this must only be done for values that are not defined in this block */
2688 pset_foreach(live, irn) {
2690 * if value is defined in this block we can anways place the spill directly after the def
2691 * -> no constraint necessary
2693 if(!is_Phi(irn) && get_nodes_block(irn) == bb) {
2698 spill = set_find_spill(spill_bb->ilp, irn);
2701 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2702 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2704 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2705 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2708 sched_foreach_op(bb, tmp) {
2709 op_t *op = get_irn_link(tmp);
2711 if(is_Phi(tmp)) continue;
2712 assert(!is_Proj(tmp));
2715 const ir_node *value = op->attr.remat.remat->value;
2718 /* only collect remats up to the first real use of a value */
2719 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2724 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2725 ir_node *arg = get_irn_n(tmp, n);
2728 /* if a value is used stop collecting remats */
2741 typedef struct _irnlist_t {
2742 struct list_head list;
2746 typedef struct _interference_t {
2747 struct list_head blocklist;
2753 cmp_interference(const void *a, const void *b, size_t size)
2755 const interference_t *p = a;
2756 const interference_t *q = b;
2758 return !(p->a == q->a && p->b == q->b);
2761 static interference_t *
2762 set_find_interference(set * set, ir_node * a, ir_node * b)
2764 interference_t query;
2766 query.a = (a>b)?a:b;
2767 query.b = (a>b)?b:a;
2769 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2772 static interference_t *
2773 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2775 interference_t query,
2777 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2781 result = set_find_interference(set, a, b);
2784 list_add(&list->list, &result->blocklist);
2788 query.a = (a>b)?a:b;
2789 query.b = (a>b)?b:a;
2791 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2793 INIT_LIST_HEAD(&result->blocklist);
2794 list_add(&list->list, &result->blocklist);
2800 values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
2802 const ir_edge_t *edge;
2804 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2805 /* both values are live in, so they interfere */
2809 /* ensure a dominates b */
2810 if(value_dominates(b,a)) {
2816 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2819 /* the following code is stolen from bera.c */
2820 if(be_is_live_end(si->lv, bb, a))
2823 foreach_out_edge(a, edge) {
2824 const ir_node *user = edge->src;
2825 if(get_nodes_block(user) == bb
2828 && !pset_find_ptr(si->inverse_ops, user)
2829 && value_dominates(b, user))
2837 * Walk all irg blocks and collect interfering values inside of phi classes
2840 luke_interferencewalker(ir_node * bb, void * data)
2842 spill_ilp_t *si = (spill_ilp_t*)data;
2845 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2846 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
2847 op_t *a_op = get_irn_link(a);
2850 /* a is only interesting if it is in my register class and if it is inside a phi class */
2851 if (has_reg_class(si, a) && get_phi_class(a)) {
2852 if(a_op->is_remat || pset_find_ptr(si->inverse_ops, a))
2855 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)) {
2856 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
2857 op_t *b_op = get_irn_link(b);
2860 /* a and b are only interesting if they are in the same phi class */
2861 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2862 if(b_op->is_remat || pset_find_ptr(si->inverse_ops, b))
2865 if(values_interfere_in_block(si, bb, a, b)) {
2866 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2867 set_insert_interference(si, si->interferences, a, b, bb);
2875 static unsigned int copy_path_id = 0;
2878 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2885 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2886 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2888 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2890 pset_foreach(copies, ptr) {
2891 copy = PTR_TO_INT(ptr);
2892 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2897 * @parameter copies contains a path of copies which lead us to irn
2898 * @parameter visited contains a set of nodes already visited on this path
2901 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2903 const ir_edge_t *edge;
2904 op_t *op = get_irn_link(irn);
2905 pset *visited_users = pset_new_ptr_default();
2908 if(op->is_remat) return 0;
2910 pset_insert_ptr(visited, irn);
2914 pset *visited_operands = pset_new_ptr(get_irn_arity(irn));
2916 /* visit all operands */
2917 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2918 ir_node *arg = get_irn_n(irn, n);
2919 ilp_var_t copy = op->attr.live_range.args.copies[n];
2921 if(!has_reg_class(si, arg)) continue;
2922 if(pset_find_ptr(visited_operands, arg)) continue;
2923 pset_insert_ptr(visited_operands, arg);
2926 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2927 del_pset(visited_operands);
2928 del_pset(visited_users);
2929 pset_remove_ptr(visited, irn);
2932 pset_insert(copies, INT_TO_PTR(copy), copy);
2933 write_copy_path_cst(si, copies, any_interfere);
2934 pset_remove(copies, INT_TO_PTR(copy), copy);
2935 } else if(!pset_find_ptr(visited, arg)) {
2936 pset_insert(copies, INT_TO_PTR(copy), copy);
2937 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
2938 pset_remove(copies, INT_TO_PTR(copy), copy);
2940 if(paths > MAX_PATHS) {
2941 if(pset_count(copies) == 0) {
2945 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2946 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2947 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2948 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2949 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2953 del_pset(visited_operands);
2954 del_pset(visited_users);
2955 pset_remove_ptr(visited, irn);
2958 } else if(pset_count(copies) == 0) {
2964 del_pset(visited_operands);
2967 /* visit all uses which are phis */
2968 foreach_out_edge(irn, edge) {
2969 ir_node *user = edge->src;
2970 int pos = edge->pos;
2971 op_t *op = get_irn_link(user);
2974 if(!is_Phi(user)) continue;
2975 if(!has_reg_class(si, user)) continue;
2976 if(pset_find_ptr(visited_users, user)) continue;
2977 pset_insert_ptr(visited_users, user);
2979 copy = op->attr.live_range.args.copies[pos];
2981 if(user == target) {
2982 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2983 del_pset(visited_users);
2984 pset_remove_ptr(visited, irn);
2987 pset_insert(copies, INT_TO_PTR(copy), copy);
2988 write_copy_path_cst(si, copies, any_interfere);
2989 pset_remove(copies, INT_TO_PTR(copy), copy);
2990 } else if(!pset_find_ptr(visited, user)) {
2991 pset_insert(copies, INT_TO_PTR(copy), copy);
2992 paths += find_copy_path(si, user, target, any_interfere, copies, visited);
2993 pset_remove(copies, INT_TO_PTR(copy), copy);
2995 if(paths > MAX_PATHS) {
2996 if(pset_count(copies) == 0) {
3000 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3001 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3002 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3003 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3004 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3008 del_pset(visited_users);
3009 pset_remove_ptr(visited, irn);
3012 } else if(pset_count(copies) == 0) {
3018 del_pset(visited_users);
3019 pset_remove_ptr(visited, irn);
3024 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
3026 pset * copies = pset_new_ptr_default();
3027 pset * visited = pset_new_ptr_default();
3029 find_copy_path(si, a, b, any_interfere, copies, visited);
3037 memcopyhandler(spill_ilp_t * si)
3039 interference_t *interference;
3041 /* teste Speicherwerte auf Interferenz */
3043 /* analyze phi classes */
3044 phi_class_compute(si->chordal_env->irg);
3046 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
3047 irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
3049 /* now lets emit the ILP unequations for the crap */
3050 set_foreach(si->interferences, interference) {
3052 ilp_var_t interfere,
3054 ilp_cst_t any_interfere_cst,
3056 const ir_node *a = interference->a;
3057 const ir_node *b = interference->b;
3059 /* any_interf <= \sum interf */
3060 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
3061 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3062 any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3064 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
3066 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
3067 const ir_node *bb = irnlist->irn;
3068 spill_bb_t *spill_bb = get_irn_link(bb);
3073 spilla = set_find_spill(spill_bb->ilp, a);
3076 spillb = set_find_spill(spill_bb->ilp, b);
3079 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
3080 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
3081 /* 2: - mem_in_a - spill_a + interfere <= 0 */
3082 /* 3: - mem_in_b - spill_b + interfere <= 0 */
3083 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
3084 interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3086 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
3087 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
3089 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
3090 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
3091 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
3092 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
3093 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
3095 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
3096 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3098 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3099 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
3100 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
3102 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
3103 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3105 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3106 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
3107 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
3110 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
3112 /* any_interfere >= interf */
3113 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
3114 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3116 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3117 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3120 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
3121 gen_copy_constraints(si,a,b,any_interfere);
3129 return fabs(x) < 0.00001;
3132 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
3134 spill_ilp_t *si = get_irg_link(current_ir_graph);
3136 if(pset_find_ptr(si->all_possible_remats, n)) {
3137 op_t *op = (op_t*)get_irn_link(n);
3138 assert(op && op->is_remat);
3140 if(!op->attr.remat.remat->inverse) {
3141 if(op->attr.remat.pre) {
3142 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
3144 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
3149 op_t *op = (op_t*)get_irn_link(n);
3150 assert(op && op->is_remat);
3152 if(op->attr.remat.pre) {
3153 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
3155 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
3166 dump_graph_with_remats(ir_graph * irg, const char * suffix)
3168 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
3169 be_dump(irg, suffix, dump_ir_block_graph_sched);
3170 set_dump_node_vcgattr_hook(NULL);
3174 * Edge hook to dump the schedule edges with annotated register pressure.
3177 sched_pressure_edge_hook(FILE *F, ir_node *irn)
3179 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
3180 ir_node *prev = sched_prev(irn);
3181 fprintf(F, "edge:{sourcename:\"");
3183 fprintf(F, "\" targetname:\"");
3185 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
3186 fprintf(F, "\" color:magenta}\n");
3192 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
3194 DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3196 dump_consts_local(0);
3197 set_dump_node_edge_hook(sched_pressure_edge_hook);
3198 dump_ir_block_graph(irg, suffix);
3199 set_dump_node_edge_hook(old_edge_hook);
3203 walker_pressure_annotator(ir_node * bb, void * data)
3205 spill_ilp_t *si = data;
3208 pset *live = pset_new_ptr_default();
3211 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3212 irn = be_lv_get_irn(si->lv, bb, i);
3214 if (has_reg_class(si, irn)) {
3215 pset_insert_ptr(live, irn);
3219 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3221 sched_foreach_reverse(bb, irn) {
3223 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3227 if(has_reg_class(si, irn)) {
3228 pset_remove_ptr(live, irn);
3229 if(is_Proj(irn)) ++projs;
3232 if(!is_Proj(irn)) projs = 0;
3234 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3235 ir_node *arg = get_irn_n(irn, n);
3237 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3239 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3246 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3248 be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
3252 connect_all_remats_with_keep(spill_ilp_t * si)
3260 n_remats = pset_count(si->all_possible_remats);
3262 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3265 pset_foreach(si->all_possible_remats, irn) {
3270 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
3272 obstack_free(si->obst, ins);
3277 connect_all_spills_with_keep(spill_ilp_t * si)
3286 n_spills = pset_count(si->spills);
3288 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3291 pset_foreach(si->spills, irn) {
3296 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
3298 obstack_free(si->obst, ins);
3302 /** insert a spill at an arbitrary position */
3303 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3305 ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
3306 ir_graph *irg = get_irn_irg(bl);
3307 ir_node *frame = get_irg_frame(irg);
3311 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
3312 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3314 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3317 * search the right insertion point. a spill of a phi cannot be put
3318 * directly after the phi, if there are some phis behind the one which
3319 * is spilled. Also, a spill of a Proj must be after all Projs of the
3322 * Here's one special case:
3323 * If the spill is in the start block, the spill must be after the frame
3324 * pointer is set up. This is done by setting insert to the end of the block
3325 * which is its default initialization (see above).
3328 if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3331 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3334 sched_add_after(insert, spill);
3339 delete_remat(spill_ilp_t * si, ir_node * remat) {
3341 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3343 sched_remove(remat);
3345 /* kill links to operands */
3346 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3347 set_irn_n(remat, n, bad);
3352 clean_remat_info(spill_ilp_t * si)
3356 remat_info_t *remat_info;
3357 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3359 set_foreach(si->remat_info, remat_info) {
3360 if(!remat_info->remats) continue;
3362 pset_foreach(remat_info->remats, remat)
3364 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3365 if(sched_is_scheduled(remat->proj)) {
3366 sched_remove((ir_node*)remat->proj);
3368 set_irn_n((ir_node*)remat->proj, -1, bad);
3369 set_irn_n((ir_node*)remat->proj, 0, bad);
3372 if(get_irn_n_edges(remat->op) == 0) {
3373 if(sched_is_scheduled(remat->op)) {
3374 sched_remove((ir_node*)remat->op);
3376 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3377 set_irn_n((ir_node*)remat->op, n, bad);
3382 if(remat_info->remats) del_pset(remat_info->remats);
3383 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3388 delete_unnecessary_remats(spill_ilp_t * si)
3390 if(opt_keep_alive & KEEPALIVE_REMATS) {
3392 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3395 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3396 ir_node *keep_arg = get_irn_n(si->keep, n);
3397 op_t *arg_op = get_irn_link(keep_arg);
3400 assert(arg_op->is_remat);
3402 name = si->lpp->vars[arg_op->attr.remat.ilp];
3404 if(is_zero(name->value)) {
3405 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
3406 /* TODO check whether reload is preferred over remat (could be bug) */
3407 delete_remat(si, keep_arg);
3409 if(!arg_op->attr.remat.remat->inverse) {
3410 if(arg_op->attr.remat.pre) {
3411 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3413 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3416 if(arg_op->attr.remat.pre) {
3417 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3419 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3424 set_irn_n(si->keep, n, bad);
3427 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
3432 pset_foreach(si->all_possible_remats, remat) {
3433 op_t *remat_op = get_irn_link(remat);
3434 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
3436 if(is_zero(name->value)) {
3437 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
3438 /* TODO check whether reload is preferred over remat (could be bug) */
3439 delete_remat(si, remat);
3441 if(!remat_op->attr.remat.remat->inverse) {
3442 if(remat_op->attr.remat.pre) {
3443 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3445 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3448 if(remat_op->attr.remat.pre) {
3449 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3451 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3460 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3462 pset *spills = pset_new_ptr_default();
3464 const ir_node *next;
3467 defs = set_find_def(si->values, value);
3469 if(defs && defs->spills) {
3470 for(next = defs->spills; next; next = get_irn_link(next)) {
3471 pset_insert_ptr(spills, next);
3479 * @param before The node after which the spill will be placed in the schedule
3482 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3486 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3488 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3490 spill = be_spill2(arch_env, irn, before);
3492 defs = set_insert_def(si->values, value);
3495 /* enter into the linked list */
3496 set_irn_link(spill, defs->spills);
3497 defs->spills = spill;
3499 if(opt_keep_alive & KEEPALIVE_SPILLS)
3500 pset_insert_ptr(si->spills, spill);
3506 * @param before The Phi node which has to be spilled
3509 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3516 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3518 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3519 ins[n] = si->m_unknown;
3522 mem_phi = new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3524 defs = set_insert_def(si->values, phi);
3527 /* enter into the linked list */
3528 set_irn_link(mem_phi, defs->spills);
3529 defs->spills = mem_phi;
3531 #ifdef SCHEDULE_PHIM
3532 sched_add_after(phi, mem_phi);
3535 if(opt_keep_alive & KEEPALIVE_SPILLS)
3536 pset_insert_ptr(si->spills, mem_phi);
3543 * Add remat to list of defs, destroys link field!
3546 insert_remat(spill_ilp_t * si, ir_node * remat)
3549 op_t *remat_op = get_irn_link(remat);
3551 assert(remat_op->is_remat);
3553 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3556 /* enter into the linked list */
3557 set_irn_link(remat, defs->remats);
3558 defs->remats = remat;
3563 * Add reload before operation and add to list of defs
3566 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3571 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3573 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3575 defs = set_find_def(si->values, value);
3577 spill = defs->spills;
3578 assert(spill && "no spill placed before reload");
3580 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3582 /* enter into the linked list */
3583 set_irn_link(reload, defs->remats);
3584 defs->remats = reload;
3589 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3592 ir_node *value = get_irn_n(memoperand->irn, memoperand->pos);
3594 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3596 DBG((si->dbg, LEVEL_2, "\t inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3598 defs = set_find_def(si->values, value);
3600 spill = defs->spills;
3601 assert(spill && "no spill placed before reload");
3603 arch_perform_memory_operand(arch_env, memoperand->irn, spill, memoperand->pos);
3606 void insert_memoperands(spill_ilp_t * si)
3608 memoperand_t *memoperand;
3611 set_foreach(si->memoperands, memoperand) {
3612 name = si->lpp->vars[memoperand->ilp];
3613 if(!is_zero(name->value)) {
3614 perform_memory_operand(si, memoperand);
3620 walker_spill_placer(ir_node * bb, void * data) {
3621 spill_ilp_t *si = (spill_ilp_t*)data;
3623 spill_bb_t *spill_bb = get_irn_link(bb);
3624 pset *spills_to_do = pset_new_ptr_default();
3627 set_foreach(spill_bb->ilp, spill) {
3630 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3631 name = si->lpp->vars[spill->mem_in];
3632 if(!is_zero(name->value)) {
3635 mem_phi = insert_mem_phi(si, spill->irn);
3637 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3641 name = si->lpp->vars[spill->spill];
3642 if(!is_zero(name->value)) {
3643 /* place spill directly after definition */
3644 if(get_nodes_block(spill->irn) == bb) {
3645 insert_spill(si, spill->irn, spill->irn, spill->irn);
3649 /* place spill at bb start */
3650 if(spill->reg_in > 0) {
3651 name = si->lpp->vars[spill->reg_in];
3652 if(!is_zero(name->value)) {
3653 insert_spill(si, spill->irn, spill->irn, bb);
3657 /* place spill after a remat */
3658 pset_insert_ptr(spills_to_do, spill->irn);
3661 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3664 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3665 op_t *op = get_irn_link(irn);
3667 if(be_is_Spill(irn)) continue;
3670 /* TODO fix this if we want to support remats with more than two nodes */
3671 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3672 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3674 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3677 if(pset_find_ptr(spills_to_do, irn)) {
3678 pset_remove_ptr(spills_to_do, irn);
3680 insert_spill(si, irn, irn, irn);
3686 assert(pset_count(spills_to_do) == 0);
3688 /* afterwards free data in block */
3689 del_pset(spills_to_do);
3693 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3695 ir_node *insert_pos = bb;
3697 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3699 /* find last definition of arg value in block */
3704 defs = set_find_def(si->values, value);
3706 if(defs && defs->remats) {
3707 for(next = defs->remats; next; next = get_irn_link(next)) {
3708 if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3709 last = sched_get_time_step(next);
3715 if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3716 last = sched_get_time_step(value);
3720 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
3722 spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3728 phim_fixer(spill_ilp_t *si) {
3731 set_foreach(si->values, defs) {
3732 const ir_node *phi = defs->value;
3733 op_t *op = get_irn_link(phi);
3734 ir_node *phi_m = NULL;
3735 ir_node *next = defs->spills;
3738 if(!is_Phi(phi)) continue;
3741 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3745 next = get_irn_link(next);
3748 if(!phi_m) continue;
3750 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3751 ir_node *value = get_irn_n(phi, n);
3752 defs_t *val_defs = set_find_def(si->values, value);
3754 /* a spill of this value */
3759 ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3760 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3762 if(!is_zero(name->value)) {
3763 spill = insert_mem_copy(si, pred, value);
3765 spill = val_defs->spills;
3768 spill = val_defs->spills;
3771 assert(spill && "no spill placed before PhiM");
3772 set_irn_n(phi_m, n, spill);
3778 walker_reload_placer(ir_node * bb, void * data) {
3779 spill_ilp_t *si = (spill_ilp_t*)data;
3781 spill_bb_t *spill_bb = get_irn_link(bb);
3783 /* reloads at end of block */
3784 if(spill_bb->reloads) {
3787 set_foreach(spill_bb->reloads, keyval) {
3788 ir_node *irn = (ir_node*)keyval->key;
3789 ilp_var_t reload = PTR_TO_INT(keyval->val);
3792 name = si->lpp->vars[reload];
3793 if(!is_zero(name->value)) {
3795 ir_node *insert_pos = bb;
3796 ir_node *prev = sched_block_last_noncf(si, bb);
3797 op_t *prev_op = get_irn_link(prev);
3799 while(be_is_Spill(prev)) {
3800 prev = sched_prev(prev);
3803 prev_op = get_irn_link(prev);
3805 /* insert reload before pre-remats */
3806 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3807 && prev_op->is_remat && prev_op->attr.remat.pre) {
3811 prev = sched_prev(prev);
3812 } while(be_is_Spill(prev));
3814 prev_op = get_irn_link(prev);
3818 reload = insert_reload(si, irn, insert_pos);
3820 if(opt_keep_alive & KEEPALIVE_RELOADS)
3821 pset_insert_ptr(si->spills, reload);
3826 /* walk and insert more reloads and collect remats */
3827 sched_foreach_reverse(bb, irn) {
3828 op_t *op = get_irn_link(irn);
3830 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3831 if(is_Phi(irn)) break;
3834 if(get_irn_mode(irn) != mode_T) {
3835 insert_remat(si, irn);
3840 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3841 ir_node *arg = get_irn_n(irn, n);
3843 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3846 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3847 if(!is_zero(name->value)) {
3849 ir_node *insert_pos = irn;
3850 ir_node *prev = sched_prev(insert_pos);
3853 while(be_is_Spill(prev)) {
3854 prev = sched_prev(prev);
3857 prev_op = get_irn_link(prev);
3859 /* insert reload before pre-remats */
3860 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3861 && prev_op->is_remat && prev_op->attr.remat.pre) {
3865 prev = sched_prev(prev);
3866 } while(be_is_Spill(prev));
3868 prev_op = get_irn_link(prev);
3872 reload = insert_reload(si, arg, insert_pos);
3874 set_irn_n(irn, n, reload);
3876 if(opt_keep_alive & KEEPALIVE_RELOADS)
3877 pset_insert_ptr(si->spills, reload);
3884 del_set(spill_bb->ilp);
3885 if(spill_bb->reloads) del_set(spill_bb->reloads);
3889 walker_collect_used(ir_node * irn, void * data)
3891 bitset_t *used = data;
3893 bitset_set(used, get_irn_idx(irn));
3896 struct kill_helper {
3902 walker_kill_unused(ir_node * bb, void * data)
3904 struct kill_helper *kh = data;
3905 ir_node *bad = get_irg_bad(get_irn_irg(bb));
3909 for(irn=sched_first(bb); !sched_is_end(irn);) {
3910 ir_node *next = sched_next(irn);
3913 if(!bitset_is_set(kh->used, get_irn_idx(irn))) {
3914 if(be_is_Spill(irn) || be_is_Reload(irn)) {
3915 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)));
3917 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3923 set_nodes_block(irn, bad);
3924 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3925 set_irn_n(irn, n, bad);
3933 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3935 struct kill_helper kh;
3937 kh.used = bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3940 irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3941 irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3943 bitset_free(kh.used);
3947 print_irn_pset(pset * p)
3951 pset_foreach(p, irn) {
3952 ir_printf("%+F\n", irn);
3957 dump_phi_class(spill_ilp_t * si, pset * phiclass, const char * file)
3959 FILE *f = fopen(file, "w");
3961 interference_t *interference;
3963 pset_break(phiclass);
3964 set_break(si->interferences);
3966 ir_fprintf(f, "digraph phiclass {\n");
3968 pset_foreach(phiclass, irn) {
3970 ir_fprintf(f, " %F%N [shape=box]\n",irn,irn);
3973 pset_foreach(phiclass, irn) {
3976 if(!is_Phi(irn)) continue;
3978 for(n=get_irn_arity(irn)-1; n>=0; --n) {
3979 ir_node *arg = get_irn_n(irn, n);
3981 ir_fprintf(f, " %F%N -> %F%N\n",irn,irn,arg,arg);
3985 set_foreach(si->interferences, interference) {
3986 const ir_node *a = interference->a;
3987 const ir_node *b = interference->b;
3988 if(get_phi_class(a) == phiclass) {
3989 ir_fprintf(f, " %F%N -> %F%N [color=red,dir=none,style=bold]\n",a,a,b,b);
3998 rewire_uses(spill_ilp_t * si)
4000 dom_front_info_t *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
4002 pset *ignore = pset_new_ptr(1);
4004 pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
4006 /* then fix uses of spills */
4007 set_foreach(si->values, defs) {
4010 const ir_node *next = defs->remats;
4013 reloads = pset_new_ptr_default();
4016 if(be_is_Reload(next)) {
4017 pset_insert_ptr(reloads, next);
4021 next = get_irn_link(next);
4024 spills = get_spills_for_value(si, defs->value);
4025 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));
4026 if(pset_count(spills) > 1) {
4027 //assert(pset_count(reloads) > 0);
4028 // print_irn_pset(spills);
4029 // print_irn_pset(reloads);
4031 be_ssa_constr_set_ignore(dfi, si->lv, spills, ignore);
4038 /* first fix uses of remats and reloads */
4039 set_foreach(si->values, defs) {
4041 const ir_node *next = defs->remats;
4045 nodes = pset_new_ptr_default();
4046 if(sched_is_scheduled(defs->value)) {
4047 pset_insert_ptr(nodes, defs->value);
4052 pset_insert_ptr(nodes, next);
4053 next = get_irn_link(next);
4056 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-orig_kept, defs->value));
4057 be_ssa_constr_set(dfi, si->lv, nodes);
4063 // remove_unused_defs(si);
4065 be_free_dominance_frontiers(dfi);
4070 writeback_results(spill_ilp_t * si)
4072 /* walk through the graph and collect all spills, reloads and remats for a value */
4074 si->values = new_set(cmp_defs, 4096);
4076 DBG((si->dbg, LEVEL_1, "Applying results\n"));
4077 delete_unnecessary_remats(si);
4078 si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
4079 irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
4080 irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
4082 insert_memoperands(si);
4085 /* clean the remat info! there are still back-edges leading there! */
4086 clean_remat_info(si);
4090 connect_all_spills_with_keep(si);
4092 del_set(si->values);
4096 get_n_regs(spill_ilp_t * si)
4098 int arch_n_regs = arch_register_class_n_regs(si->cls);
4100 bitset_t *arch_regs = bitset_malloc(arch_n_regs);
4101 bitset_t *abi_regs = bitset_malloc(arch_n_regs);
4103 arch_put_non_ignore_regs(si->chordal_env->birg->main_env->arch_env, si->cls, arch_regs);
4104 be_abi_put_ignore_regs(si->chordal_env->birg->abi, si->cls, abi_regs);
4106 bitset_andnot(arch_regs, abi_regs);
4107 arch_n_regs = bitset_popcnt(arch_regs);
4109 bitset_free(arch_regs);
4110 bitset_free(abi_regs);
4112 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", arch_n_regs, si->cls->name));
4117 walker_reload_mover(ir_node * bb, void * data)
4119 spill_ilp_t *si = data;
4122 sched_foreach(bb, tmp) {
4123 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
4124 ir_node *reload = tmp;
4127 /* move reload upwards */
4129 int pressure = (int)get_irn_link(reload);
4130 if(pressure < si->n_regs) {
4131 irn = sched_prev(reload);
4132 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
4133 sched_remove(reload);
4134 pressure = (int)get_irn_link(irn);
4136 while(pressure < si->n_regs) {
4137 if( sched_is_end(irn) ||
4138 (be_is_Reload(irn) && has_reg_class(si, irn)) ||
4139 /* do not move reload before its spill */
4140 (irn == be_get_Reload_mem(reload)) ||
4141 /* do not move before phi */
4144 set_irn_link(irn, INT_TO_PTR(pressure+1));
4145 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
4146 irn = sched_prev(irn);
4148 pressure = (int)get_irn_link(irn);
4151 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
4152 sched_put_after(irn, reload);
4159 move_reloads_upward(spill_ilp_t * si)
4161 irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
4166 * Walk all irg blocks and check for interfering spills inside of phi classes
4169 luke_meminterferencechecker(ir_node * bb, void * data)
4171 spill_ilp_t *si = (spill_ilp_t*)data;
4174 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
4175 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
4177 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
4179 /* a is only interesting if it is in my register class and if it is inside a phi class */
4180 if (has_reg_class(si, a) && get_phi_class(a)) {
4181 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)) {
4182 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
4184 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
4186 /* a and b are only interesting if they are in the same phi class */
4187 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
4188 if(values_interfere_in_block(si, bb, a, b)) {
4189 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4198 verify_phiclasses(spill_ilp_t * si)
4200 /* analyze phi classes */
4201 phi_class_compute(si->chordal_env->irg);
4203 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4204 irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
4208 be_spill_remat(const be_chordal_env_t * chordal_env)
4211 char problem_name[256];
4212 char dump_suffix[256];
4213 char dump_suffix2[256];
4214 struct obstack obst;
4217 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
4218 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
4219 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
4221 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4222 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4224 if(opt_verify & VERIFY_DOMINANCE)
4225 be_check_dominance(chordal_env->irg);
4227 obstack_init(&obst);
4228 si.chordal_env = chordal_env;
4230 si.cls = chordal_env->cls;
4231 si.lpp = new_lpp(problem_name, lpp_minimize);
4232 si.remat_info = new_set(cmp_remat_info, 4096);
4233 si.interferences = new_set(cmp_interference, 32);
4234 si.memoperands = new_set(cmp_memoperands, 128);
4235 si.all_possible_remats = pset_new_ptr_default();
4236 si.spills = pset_new_ptr_default();
4237 si.inverse_ops = pset_new_ptr_default();
4238 si.lv = chordal_env->lv;
4240 si.n_regs = get_n_regs(&si);
4242 set_irg_link(chordal_env->irg, &si);
4243 compute_doms(chordal_env->irg);
4245 /* compute phi classes */
4246 // phi_class_compute(chordal_env->irg);
4248 be_analyze_regpressure(chordal_env, "-pre");
4250 DBG((si.dbg, LEVEL_2, "\t initializing\n"));
4251 irg_block_walk_graph(chordal_env->irg, luke_initializer, NULL, &si);
4254 /* collect remats */
4255 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4256 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
4259 /* insert possible remats */
4260 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4261 irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
4262 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4264 if(opt_keep_alive & KEEPALIVE_REMATS) {
4265 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4266 connect_all_remats_with_keep(&si);
4267 /* dump graph with inserted remats */
4268 dump_graph_with_remats(chordal_env->irg, dump_suffix);
4271 /* insert copies for phi arguments not in my regclass */
4272 irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
4274 /* recompute liveness */
4275 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4276 be_liveness_recompute(si.lv);
4279 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4280 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4281 irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
4283 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4284 irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
4287 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4288 memcopyhandler(&si);
4291 if(opt_dump_flags & DUMP_PROBLEM) {
4293 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4294 if ((f = fopen(buf, "wt")) != NULL) {
4295 lpp_dump_plain(si.lpp, f);
4300 if(opt_dump_flags & DUMP_MPS) {
4303 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4304 if((f = fopen(buf, "wt")) != NULL) {
4305 mps_write_mps(si.lpp, s_mps_fixed, f);
4309 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4310 if((f = fopen(buf, "wt")) != NULL) {
4311 mps_write_mst(si.lpp, s_mps_fixed, f);
4316 lpp_check_startvals(si.lpp);
4319 DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4320 lpp_set_time_limit(si.lpp, opt_timeout);
4323 lpp_set_log(si.lpp, stdout);
4326 lpp_solve_cplex(si.lpp);
4328 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4330 assert(lpp_is_sol_valid(si.lpp)
4331 && "solution of ILP must be valid");
4333 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));
4335 if(opt_dump_flags & DUMP_SOLUTION) {
4339 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4340 if ((f = fopen(buf, "wt")) != NULL) {
4342 for (i = 0; i < si.lpp->var_next; ++i) {
4343 lpp_name_t *name = si.lpp->vars[i];
4344 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4350 writeback_results(&si);
4354 kill_all_unused_values_in_schedule(&si);
4356 if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
4357 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
4359 // move reloads upwards
4360 be_liveness_recompute(si.lv);
4361 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4362 move_reloads_upward(&si);
4365 verify_phiclasses(&si);
4368 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4370 dump_pressure_graph(&si, dump_suffix2);
4372 be_analyze_regpressure(chordal_env, "-post");
4374 if(opt_verify & VERIFY_DOMINANCE)
4375 be_check_dominance(chordal_env->irg);
4377 free_dom(chordal_env->irg);
4378 del_set(si.interferences);
4379 del_pset(si.inverse_ops);
4380 del_pset(si.all_possible_remats);
4381 del_set(si.memoperands);
4382 del_pset(si.spills);
4384 obstack_free(&obst, NULL);
4385 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4388 #else /* WITH_ILP */
4391 only_that_you_can_compile_without_WITH_ILP_defined(void)
4395 #endif /* WITH_ILP */