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
36 #include "irbackedge_t.h"
40 #include <lpp/lpp_net.h>
41 #include <lpp/lpp_cplex.h>
42 //#include <lc_pset.h>
43 //#include <libcore/lc_bitset.h>
47 #include "besched_t.h"
53 #include "bespillremat.h"
55 #include "bepressurestat.h"
56 #include "beprofile.h"
57 #include "bespilloptions.h"
59 #include "bechordal_t.h"
62 #include <libcore/lc_opts.h>
63 #include <libcore/lc_opts_enum.h>
64 #endif /* WITH_LIBCORE */
66 #define DUMP_PROBLEM 1
68 #define DUMP_SOLUTION 4
70 #define DUMP_PRESSURE 16
72 #define KEEPALIVE_REMATS 1
73 #define KEEPALIVE_SPILLS 2
74 #define KEEPALIVE_RELOADS 4
76 #define VERIFY_MEMINTERF 1
77 #define VERIFY_DOMINANCE 2
80 #define REMATS_BRIGGS 1
81 #define REMATS_NOINVERSE 2
84 static int opt_dump_flags = 0;
85 static int opt_log = 0;
86 static int opt_keep_alive = 0;
87 static int opt_goodwin = 1;
88 static int opt_memcopies = 1;
89 static int opt_memoperands = 1;
90 static int opt_verify = VERIFY_MEMINTERF;
91 static int opt_remats = REMATS_ALL;
92 static int opt_repair_schedule = 0;
93 static int opt_no_enlarge_liveness = 0;
94 static int opt_remat_while_live = 1;
95 static int opt_timeout = 300;
96 static double opt_cost_reload = 8.0;
97 static double opt_cost_memoperand = 7.0;
98 static double opt_cost_spill = 15.0;
99 static double opt_cost_remat = 1.0;
103 static const lc_opt_enum_mask_items_t dump_items[] = {
104 { "problem", DUMP_PROBLEM },
106 { "solution", DUMP_SOLUTION },
107 { "stats", DUMP_STATS },
108 { "pressure", DUMP_PRESSURE },
112 static lc_opt_enum_mask_var_t dump_var = {
113 &opt_dump_flags, dump_items
116 static const lc_opt_enum_mask_items_t keepalive_items[] = {
117 { "remats", KEEPALIVE_REMATS },
118 { "spills", KEEPALIVE_SPILLS },
119 { "reloads", KEEPALIVE_RELOADS },
123 static lc_opt_enum_mask_var_t keep_alive_var = {
124 &opt_keep_alive, keepalive_items
127 static const lc_opt_enum_mask_items_t remats_items[] = {
128 { "none", REMATS_NONE },
129 { "briggs", REMATS_BRIGGS },
130 { "noinverse", REMATS_NOINVERSE },
131 { "all", REMATS_ALL },
135 static lc_opt_enum_mask_var_t remats_var = {
136 &opt_remats, remats_items
139 static const lc_opt_table_entry_t options[] = {
140 LC_OPT_ENT_ENUM_MASK("keepalive", "keep alive inserted nodes", &keep_alive_var),
142 LC_OPT_ENT_BOOL ("goodwin", "activate goodwin reduction", &opt_goodwin),
143 LC_OPT_ENT_BOOL ("memcopies", "activate memcopy handling", &opt_memcopies),
144 LC_OPT_ENT_BOOL ("memoperands", "activate memoperands", &opt_memoperands),
145 LC_OPT_ENT_ENUM_INT ("remats", "type of remats to insert", &remats_var),
146 LC_OPT_ENT_BOOL ("repair_schedule", "repair the schedule by rematting once used nodes",&opt_repair_schedule),
147 LC_OPT_ENT_BOOL ("no_enlage_liveness", "do not enlarge liveness of operands of remats",&opt_no_enlarge_liveness),
148 LC_OPT_ENT_BOOL ("remat_while_live", "only remat where rematted value was live", &opt_remat_while_live),
150 LC_OPT_ENT_ENUM_MASK("dump", "dump problem, solution or statistical data", &dump_var),
151 LC_OPT_ENT_BOOL ("log", "activate the lpp log", &opt_log),
152 LC_OPT_ENT_INT ("timeout", "ILP solver timeout", &opt_timeout),
154 LC_OPT_ENT_DBL ("cost_reload", "cost of a reload", &opt_cost_reload),
155 LC_OPT_ENT_DBL ("cost_memoperand", "cost of a memory operand", &opt_cost_memoperand),
156 LC_OPT_ENT_DBL ("cost_spill", "cost of a spill instruction", &opt_cost_spill),
157 LC_OPT_ENT_DBL ("cost_remat", "cost of a rematerialization", &opt_cost_remat),
164 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
165 //#define SCHEDULE_PHIM /* insert phim nodes into schedule */
168 //#define SOLVE_LOCAL
169 #define LPP_SERVER "i44pc52"
170 #define LPP_SOLVER "cplex"
173 #define MAX_PATHS INT_MAX
176 typedef struct _spill_ilp_t {
177 const arch_register_class_t *cls;
182 struct obstack *obst;
184 pset *all_possible_remats;
187 set *values; /**< for collecting all definitions of values before running ssa-construction */
193 #ifndef SCHEDULE_PHIM
196 DEBUG_ONLY(firm_dbg_module_t * dbg);
199 typedef int ilp_var_t;
200 typedef int ilp_cst_t;
202 typedef struct _spill_bb_t {
207 typedef struct _remat_t {
208 const ir_node *op; /**< for copy_irn */
209 const ir_node *value; /**< the value which is being recomputed by this remat */
210 const ir_node *proj; /**< not NULL if the above op produces a tuple */
211 int cost; /**< cost of this remat */
212 int inverse; /**< nonzero if this is an inverse remat */
216 * Data to be attached to each IR node. For remats this contains the ilp_var
217 * for this remat and for normal ops this contains the ilp_vars for
218 * reloading each operand
220 typedef struct _op_t {
225 const remat_t *remat; /** the remat this op belongs to */
226 int pre; /** 1, if this is a pressure-increasing remat */
230 ir_node *op; /** the operation this live range belongs to */
239 typedef struct _defs_t {
240 const ir_node *value;
241 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
242 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
245 typedef struct _remat_info_t {
246 const ir_node *irn; /**< the irn to which these remats belong */
247 pset *remats; /**< possible remats for this value */
248 pset *remats_by_operand; /**< remats with this value as operand */
251 typedef struct _keyval_t {
256 typedef struct _spill_t {
265 typedef struct _memoperand_t {
266 ir_node *irn; /**< the irn */
267 unsigned int pos; /**< the position of the argument */
268 ilp_var_t ilp; /**< the ilp var for this memory operand */
272 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
274 return arch_irn_consider_in_reg_alloc(si->birg->main_env->arch_env,
280 cmp_remat(const void *a, const void *b)
282 const keyval_t *p = a;
283 const keyval_t *q = b;
284 const remat_t *r = p->val;
285 const remat_t *s = q->val;
289 return !(r == s || r->op == s->op);
293 cmp_remat(const void *a, const void *b)
295 const remat_t *r = a;
296 const remat_t *s = a;
298 return !(r == s || r->op == s->op);
302 cmp_spill(const void *a, const void *b, size_t size)
304 const spill_t *p = a;
305 const spill_t *q = b;
307 // return !(p->irn == q->irn && p->bb == q->bb);
308 return !(p->irn == q->irn);
312 cmp_memoperands(const void *a, const void *b, size_t size)
314 const memoperand_t *p = a;
315 const memoperand_t *q = b;
317 return !(p->irn == q->irn && p->pos == q->pos);
321 set_find_keyval(set * set, const void * key)
326 return set_find(set, &query, sizeof(query), HASH_PTR(key));
330 set_insert_keyval(set * set, void * key, void * val)
336 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
340 set_find_def(set * set, const ir_node * value)
345 return set_find(set, &query, sizeof(query), HASH_PTR(value));
349 set_insert_def(set * set, const ir_node * value)
356 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
359 static memoperand_t *
360 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
367 return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
370 static memoperand_t *
371 set_find_memoperand(set * set, const ir_node * irn, unsigned int pos)
375 query.irn = (ir_node*)irn;
377 return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
382 set_find_spill(set * set, const ir_node * value)
386 query.irn = (ir_node*)value;
387 return set_find(set, &query, sizeof(query), HASH_PTR(value));
390 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
391 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
392 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
393 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
394 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
397 cmp_remat_info(const void *a, const void *b, size_t size)
399 const remat_info_t *p = a;
400 const remat_info_t *q = b;
402 return !(p->irn == q->irn);
406 cmp_defs(const void *a, const void *b, size_t size)
411 return !(p->value == q->value);
415 cmp_keyval(const void *a, const void *b, size_t size)
417 const keyval_t *p = a;
418 const keyval_t *q = b;
420 return !(p->key == q->key);
424 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
427 if(be_profile_has_data())
428 return ((double)be_profile_get_block_execcount(get_block(irn))) + FUDGE;
430 #ifndef EXECFREQ_LOOPDEPH
431 return get_block_execfreq(si->birg->exec_freq, get_block(irn)) + FUDGE;
434 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
436 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
441 get_cost(const spill_ilp_t * si, const ir_node * irn)
443 if(be_is_Spill(irn)) {
444 return opt_cost_spill;
445 } else if(be_is_Reload(irn)){
446 return opt_cost_reload;
448 return arch_get_op_estimated_cost(si->birg->main_env->arch_env, irn);
453 * Checks, whether node and its operands have suitable reg classes
456 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
459 const arch_env_t *arch_env = si->birg->main_env->arch_env;
460 int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
464 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
467 for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
468 ir_node *op = get_irn_n(irn, n);
469 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
472 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
479 * Try to create a remat from @p op with destination value @p dest_value
481 static INLINE remat_t *
482 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
484 remat_t *remat = NULL;
486 // if(!mode_is_datab(get_irn_mode(dest_value)))
489 if(dest_value == op) {
490 const ir_node *proj = NULL;
492 if(is_Proj(dest_value)) {
493 op = get_Proj_pred(op);
497 if(!is_rematerializable(si, op))
500 remat = obstack_alloc(si->obst, sizeof(*remat));
502 remat->cost = (int)get_cost(si, op);
503 remat->value = dest_value;
507 arch_inverse_t inverse;
510 /* get the index of the operand we want to retrieve by the inverse op */
511 for (n = get_irn_arity(op)-1; n>=0; --n) {
512 ir_node *arg = get_irn_n(op, n);
514 if(arg == dest_value) break;
518 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", n, op));
520 /* else ask the backend to give an inverse op */
521 if(arch_get_inverse(si->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
524 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
526 assert(inverse.n > 0 && "inverse op should have at least one node");
528 for(i=inverse.n-1; i>=0; --i) {
529 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
533 remat = obstack_alloc(si->obst, sizeof(*remat));
534 remat->op = inverse.nodes[0];
535 remat->cost = inverse.costs;
536 remat->value = dest_value;
537 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
540 // Matze: commented this out, this doesn't seem to be true if
541 // the inverse is a simple operation with only 1 result...
542 //assert(is_Proj(remat->proj));
544 assert(0 && "I can not handle remats with more than 2 nodes");
551 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
553 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
561 add_remat(const spill_ilp_t * si, const remat_t * remat)
563 remat_info_t *remat_info,
568 assert(remat->value);
570 query.irn = remat->value;
572 query.remats_by_operand = NULL;
573 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
575 if(remat_info->remats == NULL) {
576 remat_info->remats = new_pset(cmp_remat, 4096);
578 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
580 /* insert the remat into the remats_be_operand set of each argument of the remat op */
581 for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
582 ir_node *arg = get_irn_n(remat->op, n);
586 query.remats_by_operand = NULL;
587 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
589 if(remat_info->remats_by_operand == NULL) {
590 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
592 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
597 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
599 const ir_edge_t *edge = get_irn_out_edge_first(irn);
603 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
606 edge = get_irn_out_edge_next(irn, edge);
613 get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
619 irn = get_Proj_pred(irn);
621 for(n=get_irn_arity(irn)-1; n>=0; --n) {
622 const ir_node *arg = get_irn_n(irn, n);
624 if(has_reg_class(si, arg)) ++ret;
631 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
636 if( has_reg_class(si, op)
637 && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1)
638 && (opt_remats != REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0)
640 remat = get_remat_from_op(si, op, op);
642 add_remat(si, remat);
646 if(opt_remats == REMATS_ALL) {
647 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
649 for (n = get_irn_arity(op)-1; n>=0; --n) {
650 ir_node *arg = get_irn_n(op, n);
652 if(has_reg_class(si, arg)) {
653 /* try to get an inverse remat */
654 remat = get_remat_from_op(si, arg, op);
656 add_remat(si, remat);
664 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
667 ir_node *def_block = get_nodes_block(val);
673 /* if pos is at end of a basic block */
675 ret = (pos == def_block || block_dominates(def_block, pos));
676 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
680 /* else if this is a normal operation */
681 block = get_nodes_block(pos);
682 if(block == def_block) {
683 if(!sched_is_scheduled(val)) return 1;
685 ret = sched_comes_after(val, pos);
686 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
690 ret = block_dominates(def_block, block);
691 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
695 static INLINE ir_node *
696 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
698 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->birg->main_env->arch_env);
702 * Returns first non-Phi node of block @p bb
704 static INLINE ir_node *
705 sched_block_first_nonphi(const ir_node * bb)
707 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
711 sched_skip_proj_predicator(const ir_node * irn, void * data)
713 return (is_Proj(irn));
716 static INLINE ir_node *
717 sched_next_nonproj(const ir_node * irn, int forward)
719 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
723 * Returns next operation node (non-Proj) after @p irn
724 * or the basic block of this node
726 static INLINE ir_node *
727 sched_next_op(const ir_node * irn)
729 ir_node *next = sched_next(irn);
734 return sched_next_nonproj(next, 1);
738 * Returns previous operation node (non-Proj) before @p irn
739 * or the basic block of this node
741 static INLINE ir_node *
742 sched_prev_op(const ir_node * irn)
744 ir_node *prev = sched_prev(irn);
749 return sched_next_nonproj(prev, 0);
753 sched_put_after(ir_node * insert, ir_node * irn)
755 if(is_Block(insert)) {
756 insert = sched_block_first_nonphi(insert);
758 insert = sched_next_op(insert);
761 sched_add_before(insert, irn);
765 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
767 if(is_Block(insert)) {
768 insert = sched_block_last_noncf(si, insert);
770 insert = sched_next_nonproj(insert, 0);
771 insert = sched_prev(insert);
774 sched_add_after(insert, irn);
778 next_post_remat(const ir_node * irn)
784 next = sched_block_first_nonphi(irn);
786 next = sched_next_op(irn);
789 if(sched_is_end(next))
792 op = get_irn_link(next);
793 if(op->is_remat && !op->attr.remat.pre) {
802 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
808 ret = sched_block_last_noncf(si, irn);
809 ret = sched_next(ret);
810 ret = sched_prev_op(ret);
812 ret = sched_prev_op(irn);
815 if(sched_is_end(ret) || is_Phi(ret))
818 op = (op_t*)get_irn_link(ret);
819 if(op->is_remat && op->attr.remat.pre) {
827 * Tells you whether a @p remat can be placed before the irn @p pos
830 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
832 const ir_node *op = remat->op;
838 prev = sched_block_last_noncf(si, pos);
839 prev = sched_next_nonproj(prev, 0);
841 prev = sched_prev_op(pos);
843 /* do not remat if the rematted value is defined immediately before this op */
844 if(prev == remat->op) {
849 /* this should be just fine, the following OP will be using this value, right? */
851 /* only remat AFTER the real definition of a value (?) */
852 if(!value_is_defined_before(si, pos, remat->value)) {
853 // ir_fprintf(stderr, "error(not defined)");
858 for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
859 const ir_node *arg = get_irn_n(op, n);
861 if(opt_no_enlarge_liveness) {
862 if(has_reg_class(si, arg) && live) {
863 res &= pset_find_ptr((pset*)live, arg)?1:0;
865 res &= value_is_defined_before(si, pos, arg);
868 res &= value_is_defined_before(si, pos, arg);
876 * Tells you whether a @p remat can be placed after the irn @p pos
879 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
882 pos = sched_block_first_nonphi(pos);
884 pos = sched_next_op(pos);
887 /* only remat AFTER the real definition of a value (?) */
888 if(!value_is_defined_before(si, pos, remat->value)) {
892 return can_remat_before(si, remat, pos, live);
896 * Collect potetially rematerializable OPs
899 walker_remat_collector(ir_node * irn, void * data)
901 spill_ilp_t *si = data;
903 if(!is_Block(irn) && !is_Phi(irn)) {
904 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
905 get_remats_from_op(si, irn);
910 * Inserts a copy of @p irn before @p pos
913 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
918 bb = is_Block(pos)?pos:get_nodes_block(pos);
919 copy = exact_copy(irn);
921 set_phi_class(si->pc, copy, NULL);
922 set_nodes_block(copy, bb);
923 sched_put_before(si, pos, copy);
929 * Inserts a copy of @p irn after @p pos
932 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
937 bb = is_Block(pos)?pos:get_nodes_block(pos);
938 copy = exact_copy(irn);
940 set_phi_class(si->pc, copy, NULL);
941 set_nodes_block(copy, bb);
942 sched_put_after(pos, copy);
948 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
952 if(can_remat_after(si, remat, pos, live)) {
957 DBG((si->dbg, LEVEL_3, "\t >inserting remat2 %+F\n", remat->op));
959 copy = insert_copy_after(si, remat->op, pos);
961 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
962 op = obstack_alloc(si->obst, sizeof(*op));
964 op->attr.remat.remat = remat;
965 op->attr.remat.pre = 0;
966 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
968 set_irn_link(copy, op);
969 pset_insert_ptr(si->all_possible_remats, copy);
971 proj_copy = insert_copy_after(si, remat->proj, copy);
972 set_irn_n(proj_copy, 0, copy);
973 set_irn_link(proj_copy, op);
974 pset_insert_ptr(si->all_possible_remats, proj_copy);
986 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
990 if(can_remat_before(si, remat, pos, live)) {
995 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
997 copy = insert_copy_before(si, remat->op, pos);
999 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
1000 op = obstack_alloc(si->obst, sizeof(*op));
1002 op->attr.remat.remat = remat;
1003 op->attr.remat.pre = 1;
1004 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
1006 set_irn_link(copy, op);
1007 pset_insert_ptr(si->all_possible_remats, copy);
1009 proj_copy = insert_copy_after(si, remat->proj, copy);
1010 set_irn_n(proj_copy, 0, copy);
1011 set_irn_link(proj_copy, op);
1012 pset_insert_ptr(si->all_possible_remats, proj_copy);
1024 get_block_n_succs(const ir_node *block) {
1025 const ir_edge_t *edge;
1027 assert(edges_activated(current_ir_graph));
1029 edge = get_block_succ_first(block);
1033 edge = get_block_succ_next(block, edge);
1034 return edge ? 2 : 1;
1038 is_start_block(const ir_node * bb)
1040 return get_irg_start_block(get_irn_irg(bb)) == bb;
1044 is_merge_edge(const ir_node * bb)
1046 if(is_start_block(bb))
1050 return get_block_n_succs(bb) == 1;
1056 is_diverge_edge(const ir_node * bb)
1058 if(is_start_block(bb))
1062 return get_Block_n_cfgpreds(bb) == 1;
1068 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1073 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1074 irn = be_lv_get_irn(si->lv, bb, i);
1076 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1077 pset_insert_ptr(live, irn);
1081 irn = sched_last(bb);
1083 /* all values eaten by control flow operations are also live until the end of the block */
1084 sched_foreach_reverse(bb, irn) {
1087 if(!sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) break;
1089 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1090 ir_node *arg = get_irn_n(irn,i);
1092 if(has_reg_class(si, arg)) {
1093 pset_insert_ptr(live, arg);
1098 * find values that are used by remats at end of block
1099 * and insert them into live set
1101 foreach_pre_remat(si, bb, irn) {
1104 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1105 ir_node *remat_arg = get_irn_n(irn, n);
1107 if(!has_reg_class(si, remat_arg)) continue;
1109 /* if value is becoming live through use by remat */
1110 if(!pset_find_ptr(live, remat_arg)) {
1111 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, irn));
1113 pset_insert_ptr(live, remat_arg);
1120 walker_regclass_copy_insertor(ir_node * irn, void * data)
1122 spill_ilp_t *si = data;
1124 if(is_Phi(irn) && has_reg_class(si, irn)) {
1127 for(n=get_irn_arity(irn)-1; n>=0; --n) {
1128 ir_node *phi_arg = get_irn_n(irn, n);
1129 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
1131 if(!has_reg_class(si, phi_arg)) {
1132 ir_node *copy = be_new_Copy(si->cls, si->birg->irg, bb, phi_arg);
1133 ir_node *pos = sched_block_last_noncf(si, bb);
1134 op_t *op = obstack_alloc(si->obst, sizeof(*op));
1136 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
1137 sched_add_after(pos, copy);
1138 set_irn_n(irn, n, copy);
1141 op->attr.live_range.args.reloads = NULL;
1142 op->attr.live_range.ilp = ILP_UNDEF;
1143 set_irn_link(copy, op);
1150 * Insert (so far unused) remats into the irg to
1151 * recompute the potential liveness of all values
1154 walker_remat_insertor(ir_node * bb, void * data)
1156 spill_ilp_t *si = data;
1163 /* skip start block, no remats to do there */
1164 if(is_start_block(bb)) return;
1166 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1168 live = pset_new_ptr_default();
1169 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1170 ir_node *value = be_lv_get_irn(si->lv, bb, i);
1172 /* add remats at end of block */
1173 if (has_reg_class(si, value)) {
1174 pset_insert_ptr(live, value);
1178 irn = sched_last(bb);
1179 while(!sched_is_end(irn)) {
1185 next = sched_prev(irn);
1187 /* delete defined value from live set */
1188 if(has_reg_class(si, irn)) {
1189 pset_remove_ptr(live, irn);
1192 if(is_Phi(irn) || is_Proj(irn)) {
1197 args = pset_new_ptr_default();
1198 used = pset_new_ptr_default();
1200 /* collect arguments of op and set args of op already live in epilog */
1201 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1202 ir_node *arg = get_irn_n(irn, n);
1204 pset_insert_ptr(args, arg);
1205 if(has_reg_class(si, arg)) {
1206 pset_insert_ptr(live, arg);
1207 pset_insert_ptr(used, arg);
1211 /* insert all possible remats before irn */
1212 pset_foreach(args, arg) {
1213 remat_info_t *remat_info,
1216 /* continue if the operand has the wrong reg class */
1217 if(!has_reg_class(si, arg))
1221 query.remats = NULL;
1222 query.remats_by_operand = NULL;
1223 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1229 if(remat_info->remats) {
1230 pset_foreach(remat_info->remats, remat) {
1231 ir_node *remat_irn = NULL;
1233 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
1234 remat_irn = insert_remat_before(si, remat, irn, live);
1237 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1238 ir_node *remat_arg = get_irn_n(remat_irn, n);
1240 /* collect args of remats which are not args of op */
1241 if(has_reg_class(si, remat_arg) && !pset_find_ptr(args, remat_arg)) {
1242 pset_insert_ptr(used, remat_arg);
1250 /* do not place post remats after jumps */
1251 if(sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) {
1257 /* insert all possible remats after irn */
1258 post_remats = pset_new_ptr_default();
1259 pset_foreach(used, arg) {
1260 remat_info_t *remat_info,
1263 /* continue if the operand has the wrong reg class */
1264 if(!has_reg_class(si, arg))
1268 query.remats = NULL;
1269 query.remats_by_operand = NULL;
1270 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1276 if(remat_info->remats_by_operand) {
1277 pset_foreach(remat_info->remats_by_operand, remat) {
1278 /* do not insert remats producing the same value as one of the operands */
1279 if(!pset_find_ptr(args, remat->value)) {
1280 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
1282 /* only remat values that can be used by real ops */
1283 if(!opt_remat_while_live || pset_find_ptr(live, remat->value)) {
1284 pset_insert_ptr(post_remats, remat);
1290 pset_foreach(post_remats, remat) {
1291 insert_remat_after(si, remat, irn, live);
1293 del_pset(post_remats);
1300 /* add remats at end if successor has multiple predecessors */
1301 if(is_merge_edge(bb)) {
1302 pset *live_out = pset_new_ptr_default();
1305 get_live_end(si, bb, live_out);
1307 /* add remats at end of block */
1308 pset_foreach(live_out, value) {
1309 remat_info_t *remat_info,
1313 query.remats = NULL;
1314 query.remats_by_operand = NULL;
1315 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1317 if(remat_info && remat_info->remats) {
1318 pset_foreach(remat_info->remats, remat) {
1319 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1321 insert_remat_before(si, remat, bb, live_out);
1328 if(is_diverge_edge(bb)) {
1329 pset *live_in = pset_new_ptr_default();
1332 be_lv_foreach(si->lv, bb, be_lv_state_in, i) {
1333 value = be_lv_get_irn(si->lv, bb, i);
1335 if(has_reg_class(si, value)) {
1336 pset_insert_ptr(live_in, value);
1339 /* add phis to live_in */
1340 sched_foreach(bb, value) {
1341 if(!is_Phi(value)) break;
1343 if(has_reg_class(si, value)) {
1344 pset_insert_ptr(live_in, value);
1348 /* add remat2s at beginning of block */
1349 post_remats = pset_new_ptr_default();
1350 pset_foreach(live_in, value) {
1351 remat_info_t *remat_info,
1355 query.remats = NULL;
1356 query.remats_by_operand = NULL;
1357 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1359 if(remat_info && remat_info->remats_by_operand) {
1360 pset_foreach(remat_info->remats_by_operand, remat) {
1361 DBG((si->dbg, LEVEL_4, "\t considering remat2 %+F at beginning of block %+F\n", remat->op, bb));
1363 /* put the remat here if all its args are available and result is still live */
1364 if(!opt_remat_while_live || pset_find_ptr(live_in, remat->value)) {
1365 pset_insert_ptr(post_remats, remat);
1370 pset_foreach(post_remats, remat) {
1371 insert_remat_after(si, remat, bb, live_in);
1373 del_pset(post_remats);
1379 can_be_copied(const ir_node * bb, const ir_node * irn)
1381 const ir_edge_t *edge = get_block_succ_first(bb);
1382 const ir_node *next_bb = edge->src;
1383 int pos = edge->pos;
1386 assert(is_merge_edge(bb));
1388 sched_foreach(next_bb, phi) {
1389 const ir_node *phi_arg;
1391 if(!is_Phi(phi)) break;
1393 phi_arg = get_irn_n(phi, pos);
1395 if(phi_arg == irn) {
1403 * Initialize additional node info
1406 luke_initializer(ir_node * bb, void * data)
1408 spill_ilp_t *si = (spill_ilp_t*)data;
1409 spill_bb_t *spill_bb;
1412 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1413 set_irn_link(bb, spill_bb);
1415 sched_foreach(bb, irn) {
1418 op = obstack_alloc(si->obst, sizeof(*op));
1420 op->attr.live_range.ilp = ILP_UNDEF;
1423 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1424 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1426 } else if(!is_Proj(irn)) {
1427 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1428 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1430 op->attr.live_range.args.reloads = NULL;
1432 set_irn_link(irn, op);
1438 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1441 luke_endwalker(ir_node * bb, void * data)
1443 spill_ilp_t *si = (spill_ilp_t*)data;
1449 spill_bb_t *spill_bb = get_irn_link(bb);
1452 live = pset_new_ptr_default();
1453 use_end = pset_new_ptr_default();
1455 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1456 irn = be_lv_get_irn(si->lv, bb, i);
1457 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1458 pset_insert_ptr(live, irn);
1462 * find values that are used by remats at end of block
1463 * and insert them into live set
1465 foreach_pre_remat(si, bb, irn) {
1468 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1469 ir_node *remat_arg = get_irn_n(irn, n);
1471 if(has_reg_class(si, remat_arg)) {
1472 pset_insert_ptr(live, remat_arg);
1477 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1478 /* their reg_out must always be set */
1479 sched_foreach_reverse(bb, irn) {
1482 if(!sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) break;
1484 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1485 ir_node *irn_arg = get_irn_n(irn, n);
1487 if(has_reg_class(si, irn_arg)) {
1488 pset_insert_ptr(use_end, irn_arg);
1493 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1494 //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1495 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1497 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1499 /* if this is a merge edge we can reload at the end of this block */
1500 if(is_merge_edge(bb)) {
1501 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1502 } else if(pset_count(use_end)){
1503 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1505 spill_bb->reloads = NULL;
1508 pset_foreach(live,irn) {
1512 int default_spilled;
1515 /* handle values used by control flow nodes later separately */
1516 if(pset_find_ptr(use_end, irn)) continue;
1519 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1521 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1523 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1524 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1525 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1527 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1528 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1530 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1531 /* by default spill value right after definition */
1532 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1533 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1535 if(is_merge_edge(bb)) {
1539 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1540 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), can_be_copied(bb, irn));
1541 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1543 /* reload <= mem_out */
1544 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1545 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1546 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1549 spill->reg_in = ILP_UNDEF;
1550 spill->mem_in = ILP_UNDEF;
1553 pset_foreach(use_end,irn) {
1557 ilp_cst_t end_use_req,
1560 int default_spilled;
1563 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1565 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1567 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1568 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1570 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1571 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1573 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1574 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1575 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1577 /* reload for use be control flow op */
1578 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1579 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1580 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1582 /* reload <= mem_out */
1583 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1584 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1585 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1587 spill->reg_in = ILP_UNDEF;
1588 spill->mem_in = ILP_UNDEF;
1590 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1591 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1592 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1600 * Find a remat of value @p value in the epilog of @p pos
1603 find_post_remat(const ir_node * value, const ir_node * pos)
1605 while((pos = next_post_remat(pos)) != NULL) {
1608 op = get_irn_link(pos);
1609 assert(op->is_remat && !op->attr.remat.pre);
1611 if(op->attr.remat.remat->value == value)
1612 return (ir_node*)pos;
1615 const ir_edge_t *edge;
1616 foreach_out_edge(pos, edge) {
1617 ir_node *proj = get_edge_src_irn(edge);
1618 assert(is_Proj(proj));
1628 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1630 spill_bb_t *spill_bb = get_irn_link(bb);
1634 int default_spilled;
1637 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1639 double spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1641 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1643 spill->reg_out = ILP_UNDEF;
1644 spill->reg_in = ILP_UNDEF;
1645 spill->mem_in = ILP_UNDEF;
1647 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1648 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1650 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1651 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1652 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1659 * Inserts ILP-constraints and variables for memory copying before the given position
1662 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1664 const ir_node *succ;
1665 const ir_edge_t *edge;
1666 spill_bb_t *spill_bb = get_irn_link(block);
1675 assert(edges_activated(current_ir_graph));
1677 edge = get_block_succ_first(block);
1683 edge = get_block_succ_next(block, edge);
1684 /* next block can only contain phis, if this is a merge edge */
1687 ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1688 copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1690 ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1691 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1693 pset_foreach(live, tmp) {
1696 op_t *op = get_irn_link(irn);
1697 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1699 spill = set_find_spill(spill_bb->ilp, tmp);
1702 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1704 lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1706 sched_foreach(succ, phi) {
1707 const ir_node *to_copy;
1709 spill_t *to_copy_spill;
1710 op_t *phi_op = get_irn_link(phi);
1711 ilp_var_t reload = ILP_UNDEF;
1714 if(!is_Phi(phi)) break;
1715 if(!has_reg_class(si, phi)) continue;
1717 to_copy = get_irn_n(phi, pos);
1718 to_copy_op = get_irn_link(to_copy);
1720 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1721 assert(to_copy_spill);
1723 if(spill_bb->reloads) {
1724 keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1727 reload = PTR_TO_INT(keyval->val);
1731 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1732 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1734 /* copy - reg_out - reload - remat - live_range <= 0 */
1735 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1736 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1737 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1738 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1739 foreach_pre_remat(si, block, tmp) {
1740 op_t *remat_op = get_irn_link(tmp);
1741 if(remat_op->attr.remat.remat->value == to_copy) {
1742 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1746 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1747 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1749 /* copy - reg_out - copyreg <= 0 */
1750 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1751 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1752 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1758 * Walk all irg blocks and emit this ILP
1761 luke_blockwalker(ir_node * bb, void * data)
1763 spill_ilp_t *si = (spill_ilp_t*)data;
1768 spill_bb_t *spill_bb = get_irn_link(bb);
1771 pset *defs = pset_new_ptr_default();
1772 const arch_env_t *arch_env = si->birg->main_env->arch_env;
1774 live = pset_new_ptr_default();
1776 /****************************************
1777 * B A S I C B L O C K E N D
1778 ***************************************/
1781 /* init live values at end of block */
1782 get_live_end(si, bb, live);
1784 pset_foreach(live, irn) {
1786 ilp_var_t reload = ILP_UNDEF;
1788 spill = set_find_spill(spill_bb->ilp, irn);
1791 if(spill_bb->reloads) {
1792 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1795 reload = PTR_TO_INT(keyval->val);
1799 op = get_irn_link(irn);
1800 assert(!op->is_remat);
1802 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1803 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1804 op->attr.live_range.op = bb;
1806 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1807 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1809 /* reg_out - reload - remat - live_range <= 0 */
1810 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1811 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1812 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1813 foreach_pre_remat(si, bb, tmp) {
1814 op_t *remat_op = get_irn_link(tmp);
1815 if(remat_op->attr.remat.remat->value == irn) {
1816 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1819 ir_snprintf(buf, sizeof(buf), "reg_out2_%N_%N", bb, irn);
1820 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
1822 /* value may only die at bb end if it is used for a mem copy */
1823 /* reg_out + \sum copy - reload - remat - live_range >= 0 */
1824 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1825 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1826 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1827 foreach_pre_remat(si, bb, tmp) {
1828 op_t *remat_op = get_irn_link(tmp);
1829 if(remat_op->attr.remat.remat->value == irn) {
1830 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1833 if(is_merge_edge(bb)) {
1834 const ir_edge_t *edge = get_block_succ_first(bb);
1835 const ir_node *next_bb = edge->src;
1836 int pos = edge->pos;
1839 sched_foreach(next_bb, phi) {
1840 const ir_node *phi_arg;
1842 if(!is_Phi(phi)) break;
1844 phi_arg = get_irn_n(phi, pos);
1846 if(phi_arg == irn) {
1847 op_t *phi_op = get_irn_link(phi);
1848 ilp_var_t copy = phi_op->attr.live_range.args.copies[pos];
1850 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
1857 insert_mem_copy_position(si, live, bb);
1860 * assure the remat args are available
1862 foreach_pre_remat(si, bb, tmp) {
1863 op_t *remat_op = get_irn_link(tmp);
1866 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1867 ir_node *remat_arg = get_irn_n(tmp, n);
1868 op_t *arg_op = get_irn_link(remat_arg);
1870 if(!has_reg_class(si, remat_arg)) continue;
1872 spill = set_find_spill(spill_bb->ilp, remat_arg);
1875 /* arguments of remats have to be live until the very end of the block
1876 * remat = reg_out(remat_arg) and (reload(remat_arg) or live_range(remat_arg)),
1877 * no remats, they could be in wrong order
1880 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1881 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1883 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 3.0);
1884 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, -2.0);
1885 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1887 /* use reload placed for this argument */
1888 if(spill_bb->reloads) {
1889 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1892 ilp_var_t reload = PTR_TO_INT(keyval->val);
1894 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1899 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1904 /**************************************
1905 * B A S I C B L O C K B O D Y
1906 **************************************/
1908 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1914 ilp_cst_t check_pre,
1920 ilp_cst_t one_memoperand;
1922 /* iterate only until first phi */
1926 op = get_irn_link(irn);
1928 if(op->is_remat) continue;
1930 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1932 /* collect defined values */
1933 if(has_reg_class(si, irn)) {
1934 pset_insert_ptr(defs, irn);
1938 if(is_Proj(irn)) continue;
1941 * init set of irn's arguments
1942 * and all possibly used values around this op
1943 * and values defined by post remats
1945 args = new_set(cmp_keyval, get_irn_arity(irn));
1946 used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1947 remat_defs = pset_new_ptr(pset_count(live));
1949 if(!is_start_block(bb) || !be_is_Barrier(irn)) {
1950 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1951 ir_node *irn_arg = get_irn_n(irn, n);
1952 if(has_reg_class(si, irn_arg)) {
1953 set_insert_keyval(args, irn_arg, (void*)n);
1954 pset_insert_ptr(used, irn_arg);
1957 foreach_post_remat(irn, tmp) {
1958 op_t *remat_op = get_irn_link(tmp);
1960 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1962 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1963 ir_node *remat_arg = get_irn_n(tmp, n);
1964 if(has_reg_class(si, remat_arg)) {
1965 pset_insert_ptr(used, remat_arg);
1969 foreach_pre_remat(si, irn, tmp) {
1970 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1971 ir_node *remat_arg = get_irn_n(tmp, n);
1972 if(has_reg_class(si, remat_arg)) {
1973 pset_insert_ptr(used, remat_arg);
1979 /**********************************
1980 * I N E P I L O G O F irn
1981 **********************************/
1983 /* ensure each dying value is used by only one post remat */
1984 pset_foreach(used, tmp) {
1985 ir_node *value = tmp;
1986 op_t *value_op = get_irn_link(value);
1991 foreach_post_remat(irn, remat) {
1992 op_t *remat_op = get_irn_link(remat);
1994 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1995 ir_node *remat_arg = get_irn_n(remat, n);
1997 /* if value is used by this remat add it to constraint */
1998 if(remat_arg == value) {
2000 /* sum remat2s <= 1 + n_remats*live_range */
2001 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
2002 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2006 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2012 if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
2013 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
2017 /* ensure at least one value dies at post remat */
2018 foreach_post_remat(irn, tmp) {
2019 op_t *remat_op = get_irn_link(tmp);
2020 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
2023 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2024 remat_arg = get_irn_n(tmp, n);
2026 if(has_reg_class(si, remat_arg)) {
2028 /* does arg always die at this op? */
2029 if(!pset_find_ptr(live, remat_arg))
2030 goto skip_one_must_die;
2032 pset_insert_ptr(remat_args, remat_arg);
2036 /* remat + \sum live_range(remat_arg) <= |args| */
2037 ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
2038 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2039 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2041 pset_foreach(remat_args, remat_arg) {
2042 op_t *arg_op = get_irn_link(remat_arg);
2044 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2048 del_pset(remat_args);
2051 /* new live ranges for values from L\U defined by post remats */
2052 pset_foreach(live, tmp) {
2053 ir_node *value = tmp;
2054 op_t *value_op = get_irn_link(value);
2056 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
2057 ilp_var_t prev_lr = ILP_UNDEF;
2060 if(pset_find_ptr(remat_defs, value)) {
2062 /* next_live_range <= prev_live_range + sum remat2s */
2063 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
2064 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2066 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
2067 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2069 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
2070 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2072 foreach_post_remat(irn, remat) {
2073 op_t *remat_op = get_irn_link(remat);
2075 /* if value is being rematerialized by this remat */
2076 if(value == remat_op->attr.remat.remat->value) {
2077 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2081 value_op->attr.live_range.ilp = prev_lr;
2082 value_op->attr.live_range.op = irn;
2087 /* requirements for post remats and start live ranges from L/U' for values dying here */
2088 foreach_post_remat(irn, tmp) {
2089 op_t *remat_op = get_irn_link(tmp);
2092 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2093 ir_node *remat_arg = get_irn_n(tmp, n);
2094 op_t *arg_op = get_irn_link(remat_arg);
2096 if(!has_reg_class(si, remat_arg)) continue;
2098 /* only for values in L\U (TODO and D?), the others are handled with post_use */
2099 if(!pset_find_ptr(used, remat_arg)) {
2100 /* remat <= live_range(remat_arg) */
2101 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
2102 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2104 /* if value is becoming live through use by remat2 */
2105 if(!pset_find_ptr(live, remat_arg)) {
2108 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
2109 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2111 arg_op->attr.live_range.ilp = lr;
2112 arg_op->attr.live_range.op = irn;
2114 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
2116 pset_insert_ptr(live, remat_arg);
2117 add_to_spill_bb(si, bb, remat_arg);
2120 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2121 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2126 d = pset_count(defs);
2127 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
2129 /* count how many regs irn needs for arguments */
2130 u = set_count(args);
2133 /* check the register pressure in the epilog */
2134 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
2135 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
2136 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
2138 /* add L\U' to check_post */
2139 pset_foreach(live, tmp) {
2140 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
2141 /* if a live value is not used by irn */
2142 tmp_op = get_irn_link(tmp);
2143 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
2147 /***********************************************************
2148 * 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
2149 **********************************************************/
2152 pset_foreach(used, tmp) {
2158 op_t *arg_op = get_irn_link(arg);
2161 spill = add_to_spill_bb(si, bb, arg);
2163 /* new live range for each used value */
2164 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2165 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2167 /* the epilog stuff - including post_use, check_post, check_post_remat */
2168 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2169 post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2171 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2173 /* arg is live throughout epilog if the next live_range is in a register */
2174 if(pset_find_ptr(live, arg)) {
2175 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
2177 /* post_use >= next_lr + remat */
2178 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2179 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2180 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2181 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2184 /* forall post remat which use arg add a similar cst */
2185 foreach_post_remat(irn, remat) {
2188 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2189 ir_node *remat_arg = get_irn_n(remat, n);
2190 op_t *remat_op = get_irn_link(remat);
2192 if(remat_arg == arg) {
2193 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
2195 /* post_use >= remat */
2196 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2197 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2198 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2199 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2204 /* if value is not an arg of op and not possibly defined by post remat
2205 * then it may only die and not become live
2207 if(!set_find_keyval(args, arg)) {
2208 /* post_use <= prev_lr */
2209 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2210 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2211 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2212 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2214 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2215 /* next_lr <= prev_lr */
2216 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2217 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2218 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2219 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2223 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2224 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2225 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2226 ilp_var_t memoperand;
2228 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2229 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0);
2230 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2232 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2233 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2235 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2236 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2241 /* new live range begins for each used value */
2242 arg_op->attr.live_range.ilp = prev_lr;
2243 arg_op->attr.live_range.op = irn;
2245 pset_insert_ptr(live, arg);
2248 /* just to be sure */
2249 check_post = ILP_UNDEF;
2251 /* allow original defintions to be removed */
2252 if(opt_repair_schedule) {
2253 pset_foreach(defs, tmp) {
2254 op_t *tmp_op = get_irn_link(tmp);
2255 spill_t *spill = set_find_spill(spill_bb->ilp, tmp);
2260 ir_snprintf(buf, sizeof(buf), "delete_%N", tmp);
2261 delete = lpp_add_var_default(si->lpp, buf, lpp_binary, -1.0*get_cost(si, irn)*execution_frequency(si, bb), 0.0);
2263 /* op may not be killed if its first live_range is 1 */
2264 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2265 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2266 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2267 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2269 /* op may not be killed if it is spilled after the definition */
2270 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2271 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2272 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2273 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2278 ir_snprintf(buf, sizeof(buf), "keep_%N", tmp);
2279 keep = lpp_add_var_default(si->lpp, buf, lpp_binary, get_cost(si, irn)*execution_frequency(si, bb), 1.0);
2281 /* op may not be killed if its first live_range is 1 */
2282 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2283 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2284 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2285 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, -1.0);
2287 /* op may not be killed if it is spilled after the definition */
2288 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2289 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2290 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2291 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2296 pset_foreach(defs, tmp) {
2297 op_t *tmp_op = get_irn_link(tmp);
2298 spill_t *spill = set_find_spill(spill_bb->ilp, tmp);
2301 /* live_range or spill should be 1
2302 TODO: lr should be live until first use */
2303 ir_snprintf(buf, sizeof(buf), "nokillorig_%N", tmp);
2304 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 1.0);
2305 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2306 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2316 /* check the register pressure in the prolog */
2317 /* sum_{L\U} lr <= k - |U| */
2318 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2319 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2321 /* for the prolog remove defined values from the live set */
2322 pset_foreach(defs, tmp) {
2323 pset_remove_ptr(live, tmp);
2326 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2327 ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2328 one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2331 /***********************************************************
2332 * 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
2333 **********************************************************/
2336 set_foreach(args, keyval) {
2338 const ir_node *arg = keyval->key;
2339 int i = PTR_TO_INT(keyval->val);
2340 op_t *arg_op = get_irn_link(arg);
2341 ilp_cst_t requirements;
2344 spill = set_find_spill(spill_bb->ilp, arg);
2347 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2348 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);
2350 /* reload <= mem_out */
2351 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2352 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2353 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2354 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2356 /* requirement: arg must be in register for use */
2357 /* reload + remat + live_range == 1 */
2358 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2359 requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2361 lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2362 lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2363 foreach_pre_remat(si, irn, tmp) {
2364 op_t *remat_op = get_irn_link(tmp);
2365 if(remat_op->attr.remat.remat->value == arg) {
2366 lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2370 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2372 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2373 if(get_irn_n(irn, n) == arg) {
2377 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2378 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2379 memoperand_t *memoperand;
2380 memoperand = set_find_memoperand(si->memoperands, irn, n);
2382 /* memoperand <= mem_out */
2383 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2384 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2385 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2386 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2388 /* the memoperand is only sufficient if it is used once by the op */
2389 if(n_memoperands == 1)
2390 lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2392 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2394 /* we have one more free register if we use a memory operand */
2395 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2401 /* iterate over L\U */
2402 pset_foreach(live, tmp) {
2403 if(!set_find_keyval(args, tmp)) {
2404 /* if a live value is not used by irn */
2405 tmp_op = get_irn_link(tmp);
2406 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2410 /* requirements for remats */
2411 foreach_pre_remat(si, irn, tmp) {
2412 op_t *remat_op = get_irn_link(tmp);
2415 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2416 ir_node *remat_arg = get_irn_n(tmp, n);
2417 op_t *arg_op = get_irn_link(remat_arg);
2419 if(!has_reg_class(si, remat_arg)) continue;
2421 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2422 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2423 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2425 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2426 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2428 /* if remat arg is also used by current op then we can use reload placed for this argument */
2429 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2430 int index = (int)keyval->val;
2432 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2440 /*************************
2441 * D O N E W I T H O P
2442 *************************/
2444 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2446 pset_foreach(live, tmp) {
2447 assert(has_reg_class(si, tmp));
2450 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2451 ir_node *arg = get_irn_n(irn, n);
2453 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2456 del_pset(remat_defs);
2460 defs = pset_new_ptr_default();
2462 /* skip everything above barrier in start block */
2463 if(is_start_block(bb) && be_is_Barrier(irn)) {
2464 assert(pset_count(live) == 0);
2473 /***************************************
2474 * B E G I N N I N G O F B L O C K
2475 ***************************************/
2478 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2479 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2481 pset_foreach(live, irn) {
2482 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2485 /* construct mem_outs for all values */
2486 set_foreach(spill_bb->ilp, spill) {
2487 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2488 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2490 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2491 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2493 if(pset_find_ptr(live, spill->irn)) {
2494 int default_spilled;
2495 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2497 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2498 default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2499 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2500 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2502 if(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2504 op_t *op = get_irn_link(spill->irn);
2506 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2507 const ir_node *arg = get_irn_n(spill->irn, n);
2513 /* argument already done? */
2514 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2516 /* get sum of execution frequencies of blocks with the same phi argument */
2517 for(m=n; m>=0; --m) {
2518 const ir_node *arg2 = get_irn_n(spill->irn, m);
2521 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2525 /* copies are not for free */
2526 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2527 var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
2529 for(m=n; m>=0; --m) {
2530 const ir_node *arg2 = get_irn_n(spill->irn, m);
2533 op->attr.live_range.args.copies[m] = var;
2538 /* copy <= mem_in */
2539 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2540 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2541 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2542 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2549 foreach_post_remat(bb, tmp) {
2551 op_t *remat_op = get_irn_link(tmp);
2552 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
2555 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2556 remat_arg = get_irn_n(tmp, n);
2558 if(has_reg_class(si, remat_arg)) {
2559 pset_insert_ptr(remat_args, remat_arg);
2563 /* remat + \sum live_range(remat_arg) <= |args| */
2564 ir_snprintf(buf, sizeof(buf), "one_must_die_%N", tmp);
2565 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2566 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2568 pset_foreach(remat_args, remat_arg) {
2569 if(pset_find_ptr(live, remat_arg)) {
2570 op_t *remat_arg_op = get_irn_link(remat_arg);
2571 lpp_set_factor_fast(si->lpp, cst, remat_arg_op->attr.live_range.ilp, 1.0);
2574 del_pset(remat_args);
2577 foreach_post_remat(bb, tmp) {
2580 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2581 ir_node *remat_arg = get_irn_n(tmp, n);
2583 /* if value is becoming live through use by remat2 */
2584 if(has_reg_class(si, remat_arg) && !pset_find_ptr(live, remat_arg)) {
2585 op_t *remat_arg_op = get_irn_link(remat_arg);
2588 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 at bb start %+F\n", remat_arg, tmp));
2590 pset_insert_ptr(live, remat_arg);
2591 spill = add_to_spill_bb(si, bb, remat_arg);
2592 remat_arg_op->attr.live_range.ilp = ILP_UNDEF;
2594 /* we need reg_in and mem_in for this value; they will be referenced later */
2595 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", remat_arg, bb);
2596 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2597 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", remat_arg, bb);
2598 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2601 /* optimization: all memory stuff should be 0, for we do not want to insert reloads for remats */
2602 ir_snprintf(buf, sizeof(buf), "nomem_%N_%N", remat_arg, bb);
2603 nomem = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0.0);
2604 lpp_set_factor_fast(si->lpp, nomem, spill->spill, 1.0);
2609 /* L\U is empty at bb start */
2610 /* arg is live throughout epilog if it is reg_in into this block */
2612 /* check the register pressure at the beginning of the block
2615 /* reg_in entspricht post_use */
2617 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2618 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2620 pset_foreach(live, irn) {
2623 spill = set_find_spill(spill_bb->ilp, irn);
2626 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2627 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2629 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2631 /* spill + mem_in <= 1 */
2632 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2633 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2635 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2636 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2638 } /* post_remats are NOT included in register pressure check because
2639 they do not increase regpressure */
2641 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2642 pset_foreach(live, irn) {
2646 spill = set_find_spill(spill_bb->ilp, irn);
2647 assert(spill && spill->irn == irn);
2649 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2650 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2653 ir_node *phi_arg = get_Phi_pred(irn, n);
2654 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2655 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2657 op_t *op = get_irn_link(irn);
2659 /* although the phi is in the right regclass one or more of
2660 * its arguments can be in a different one or at least to
2663 if(has_reg_class(si, phi_arg)) {
2664 /* mem_in < mem_out_arg + copy */
2665 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2666 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2668 /* reg_in < reg_out_arg */
2669 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2670 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2672 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2673 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2675 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2678 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2680 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2682 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2686 /* else assure the value arrives on all paths in the same resource */
2688 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2691 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2692 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2695 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2696 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2697 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2698 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2700 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2701 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2703 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2706 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2707 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2712 foreach_post_remat(bb, tmp) {
2715 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2716 ir_node *remat_arg = get_irn_n(tmp, n);
2717 op_t *remat_op = get_irn_link(tmp);
2719 if(!has_reg_class(si, remat_arg)) continue;
2721 spill = set_find_spill(spill_bb->ilp, remat_arg);
2724 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2725 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2726 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2727 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2731 pset_foreach(live, irn) {
2732 const op_t *op = get_irn_link(irn);
2733 const ir_node *remat;
2738 foreach_post_remat(bb, remat) {
2741 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2742 const ir_node *arg = get_irn_n(remat, n);
2745 const op_t *remat_op = get_irn_link(remat);
2747 if(cst == ILP_UNDEF) {
2748 /* sum remat2s <= 1 + n_remats*live_range */
2749 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", irn, bb);
2750 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2752 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2758 if(cst != ILP_UNDEF && op->attr.live_range.ilp != ILP_UNDEF) {
2759 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -n_remats);
2763 /* first live ranges from reg_ins */
2764 pset_foreach(live, irn) {
2765 op_t *op = get_irn_link(irn);
2767 if(op->attr.live_range.ilp != ILP_UNDEF) {
2769 spill = set_find_spill(spill_bb->ilp, irn);
2770 assert(spill && spill->irn == irn);
2772 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2773 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2774 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2775 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2777 foreach_post_remat(bb, tmp) {
2778 op_t *remat_op = get_irn_link(tmp);
2780 if(remat_op->attr.remat.remat->value == irn) {
2781 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2787 /* walk forward now and compute constraints for placing spills */
2788 /* this must only be done for values that are not defined in this block */
2789 pset_foreach(live, irn) {
2791 * if value is defined in this block we can anways place the spill directly after the def
2792 * -> no constraint necessary
2794 if(!is_Phi(irn) && get_nodes_block(irn) == bb) {
2799 spill = set_find_spill(spill_bb->ilp, irn);
2802 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2803 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2805 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2806 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2809 sched_foreach_op(bb, tmp) {
2810 op_t *op = get_irn_link(tmp);
2812 if(is_Phi(tmp)) continue;
2813 assert(!is_Proj(tmp));
2816 const ir_node *value = op->attr.remat.remat->value;
2819 /* only collect remats up to the first real use of a value */
2820 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2825 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2826 ir_node *arg = get_irn_n(tmp, n);
2829 /* if a value is used stop collecting remats */
2842 typedef struct _irnlist_t {
2843 struct list_head list;
2847 typedef struct _interference_t {
2848 struct list_head blocklist;
2854 cmp_interference(const void *a, const void *b, size_t size)
2856 const interference_t *p = a;
2857 const interference_t *q = b;
2859 return !(p->a == q->a && p->b == q->b);
2862 static interference_t *
2863 set_find_interference(set * set, ir_node * a, ir_node * b)
2865 interference_t query;
2867 query.a = (a>b)?a:b;
2868 query.b = (a>b)?b:a;
2870 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2873 static interference_t *
2874 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2876 interference_t query,
2878 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2882 result = set_find_interference(set, a, b);
2885 list_add(&list->list, &result->blocklist);
2889 query.a = (a>b)?a:b;
2890 query.b = (a>b)?b:a;
2892 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2894 INIT_LIST_HEAD(&result->blocklist);
2895 list_add(&list->list, &result->blocklist);
2901 values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
2903 const ir_edge_t *edge;
2905 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2906 /* both values are live in, so they interfere */
2910 /* ensure a dominates b */
2911 if(value_dominates(b,a)) {
2917 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2920 /* the following code is stolen from bera.c */
2921 if(be_is_live_end(si->lv, bb, a))
2924 foreach_out_edge(a, edge) {
2925 const ir_node *user = edge->src;
2926 if(get_nodes_block(user) == bb
2929 && !pset_find_ptr(si->inverse_ops, user)
2930 && value_dominates(b, user))
2938 * Walk all irg blocks and collect interfering values inside of phi classes
2941 luke_interferencewalker(ir_node * bb, void * data)
2943 spill_ilp_t *si = (spill_ilp_t*)data;
2946 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2947 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
2948 op_t *a_op = get_irn_link(a);
2951 /* a is only interesting if it is in my register class and if it is inside a phi class */
2952 if (has_reg_class(si, a) && get_phi_class(si->pc, a)) {
2953 if (a_op->is_remat || pset_find_ptr(si->inverse_ops, a))
2956 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)) {
2957 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
2958 op_t *b_op = get_irn_link(b);
2960 /* a and b are only interesting if they are in the same phi class */
2961 if (has_reg_class(si, b) && get_phi_class(si->pc, a) == get_phi_class(si->pc, b)) {
2962 if (b_op->is_remat || pset_find_ptr(si->inverse_ops, b))
2965 if (values_interfere_in_block(si, bb, a, b)) {
2966 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2967 set_insert_interference(si, si->interferences, a, b, bb);
2975 static unsigned int copy_path_id = 0;
2978 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2985 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2986 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2988 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2990 pset_foreach(copies, ptr) {
2991 copy = PTR_TO_INT(ptr);
2992 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2997 * @parameter copies contains a path of copies which lead us to irn
2998 * @parameter visited contains a set of nodes already visited on this path
3001 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
3003 const ir_edge_t *edge;
3004 op_t *op = get_irn_link(irn);
3005 pset *visited_users = pset_new_ptr_default();
3008 if(op->is_remat) return 0;
3010 pset_insert_ptr(visited, irn);
3014 pset *visited_operands = pset_new_ptr(get_irn_arity(irn));
3016 /* visit all operands */
3017 for(n=get_irn_arity(irn)-1; n>=0; --n) {
3018 ir_node *arg = get_irn_n(irn, n);
3019 ilp_var_t copy = op->attr.live_range.args.copies[n];
3021 if(!has_reg_class(si, arg)) continue;
3022 if(pset_find_ptr(visited_operands, arg)) continue;
3023 pset_insert_ptr(visited_operands, arg);
3026 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3027 del_pset(visited_operands);
3028 del_pset(visited_users);
3029 pset_remove_ptr(visited, irn);
3032 pset_insert(copies, INT_TO_PTR(copy), copy);
3033 write_copy_path_cst(si, copies, any_interfere);
3034 pset_remove(copies, INT_TO_PTR(copy), copy);
3035 } else if(!pset_find_ptr(visited, arg)) {
3036 pset_insert(copies, INT_TO_PTR(copy), copy);
3037 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
3038 pset_remove(copies, INT_TO_PTR(copy), copy);
3040 if(paths > MAX_PATHS) {
3041 if(pset_count(copies) == 0) {
3045 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3046 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3047 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3048 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3049 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3053 del_pset(visited_operands);
3054 del_pset(visited_users);
3055 pset_remove_ptr(visited, irn);
3058 } else if(pset_count(copies) == 0) {
3064 del_pset(visited_operands);
3067 /* visit all uses which are phis */
3068 foreach_out_edge(irn, edge) {
3069 ir_node *user = edge->src;
3070 int pos = edge->pos;
3071 op_t *op = get_irn_link(user);
3074 if(!is_Phi(user)) continue;
3075 if(!has_reg_class(si, user)) continue;
3076 if(pset_find_ptr(visited_users, user)) continue;
3077 pset_insert_ptr(visited_users, user);
3079 copy = op->attr.live_range.args.copies[pos];
3081 if(user == target) {
3082 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3083 del_pset(visited_users);
3084 pset_remove_ptr(visited, irn);
3087 pset_insert(copies, INT_TO_PTR(copy), copy);
3088 write_copy_path_cst(si, copies, any_interfere);
3089 pset_remove(copies, INT_TO_PTR(copy), copy);
3090 } else if(!pset_find_ptr(visited, user)) {
3091 pset_insert(copies, INT_TO_PTR(copy), copy);
3092 paths += find_copy_path(si, user, target, any_interfere, copies, visited);
3093 pset_remove(copies, INT_TO_PTR(copy), copy);
3095 if(paths > MAX_PATHS) {
3096 if(pset_count(copies) == 0) {
3100 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3101 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3102 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3103 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3104 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3108 del_pset(visited_users);
3109 pset_remove_ptr(visited, irn);
3112 } else if(pset_count(copies) == 0) {
3118 del_pset(visited_users);
3119 pset_remove_ptr(visited, irn);
3124 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
3126 pset * copies = pset_new_ptr_default();
3127 pset * visited = pset_new_ptr_default();
3129 find_copy_path(si, a, b, any_interfere, copies, visited);
3137 memcopyhandler(spill_ilp_t * si)
3139 interference_t *interference;
3141 /* teste Speicherwerte auf Interferenz */
3143 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
3144 irg_block_walk_graph(si->birg->irg, luke_interferencewalker, NULL, si);
3146 /* now lets emit the ILP unequations for the crap */
3147 set_foreach(si->interferences, interference) {
3149 ilp_var_t interfere, any_interfere;
3150 ilp_cst_t any_interfere_cst, cst;
3151 const ir_node *a = interference->a;
3152 const ir_node *b = interference->b;
3154 /* any_interf <= \sum interf */
3155 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
3156 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3157 any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3159 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
3161 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
3162 const ir_node *bb = irnlist->irn;
3163 spill_bb_t *spill_bb = get_irn_link(bb);
3168 spilla = set_find_spill(spill_bb->ilp, a);
3171 spillb = set_find_spill(spill_bb->ilp, b);
3174 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
3175 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
3176 /* 2: - mem_in_a - spill_a + interfere <= 0 */
3177 /* 3: - mem_in_b - spill_b + interfere <= 0 */
3178 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
3179 interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3181 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
3182 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
3184 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
3185 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
3186 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
3187 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
3188 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
3190 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
3191 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3193 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3194 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
3195 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
3197 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
3198 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3200 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3201 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
3202 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
3205 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
3207 /* any_interfere >= interf */
3208 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
3209 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3211 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3212 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3215 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
3216 gen_copy_constraints(si,a,b,any_interfere);
3224 return fabs(x) < 0.00001;
3227 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
3229 spill_ilp_t *si = get_irg_link(current_ir_graph);
3231 if(pset_find_ptr(si->all_possible_remats, n)) {
3232 op_t *op = (op_t*)get_irn_link(n);
3233 assert(op && op->is_remat);
3235 if(!op->attr.remat.remat->inverse) {
3236 if(op->attr.remat.pre) {
3237 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
3239 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
3244 op_t *op = (op_t*)get_irn_link(n);
3245 assert(op && op->is_remat);
3247 if(op->attr.remat.pre) {
3248 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
3250 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
3261 dump_graph_with_remats(ir_graph * irg, const char * suffix)
3263 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
3264 be_dump(irg, suffix, dump_ir_block_graph_sched);
3265 set_dump_node_vcgattr_hook(NULL);
3269 * Edge hook to dump the schedule edges with annotated register pressure.
3272 sched_pressure_edge_hook(FILE *F, ir_node *irn)
3274 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
3275 ir_node *prev = sched_prev(irn);
3276 fprintf(F, "edge:{sourcename:\"");
3278 fprintf(F, "\" targetname:\"");
3280 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
3281 fprintf(F, "\" color:magenta}\n");
3287 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
3289 DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3291 dump_consts_local(0);
3292 set_dump_node_edge_hook(sched_pressure_edge_hook);
3293 dump_ir_block_graph(irg, suffix);
3294 set_dump_node_edge_hook(old_edge_hook);
3298 walker_pressure_annotator(ir_node * bb, void * data)
3300 spill_ilp_t *si = data;
3303 pset *live = pset_new_ptr_default();
3306 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3307 irn = be_lv_get_irn(si->lv, bb, i);
3309 if (has_reg_class(si, irn)) {
3310 pset_insert_ptr(live, irn);
3314 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3316 sched_foreach_reverse(bb, irn) {
3318 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3322 if(has_reg_class(si, irn)) {
3323 pset_remove_ptr(live, irn);
3324 if(is_Proj(irn)) ++projs;
3327 if(!is_Proj(irn)) projs = 0;
3329 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3330 ir_node *arg = get_irn_n(irn, n);
3332 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3334 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3341 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3343 be_dump(si->birg->irg, suffix, dump_ir_block_graph_sched_pressure);
3347 connect_all_remats_with_keep(spill_ilp_t * si)
3355 n_remats = pset_count(si->all_possible_remats);
3357 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3360 pset_foreach(si->all_possible_remats, irn) {
3365 si->keep = be_new_Keep(si->cls, si->birg->irg, get_irg_end_block(si->birg->irg), n_remats, ins);
3367 obstack_free(si->obst, ins);
3372 connect_all_spills_with_keep(spill_ilp_t * si)
3381 n_spills = pset_count(si->spills);
3383 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3386 pset_foreach(si->spills, irn) {
3391 keep = be_new_Keep(si->cls, si->birg->irg, get_irg_end_block(si->birg->irg), n_spills, ins);
3393 obstack_free(si->obst, ins);
3397 /** insert a spill at an arbitrary position */
3398 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3400 ir_node *bl = is_Block(insert) ? insert : get_nodes_block(insert);
3401 ir_graph *irg = get_irn_irg(bl);
3402 ir_node *frame = get_irg_frame(irg);
3405 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
3406 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3408 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3411 * search the right insertion point. a spill of a phi cannot be put
3412 * directly after the phi, if there are some phis behind the one which
3413 * is spilled. Also, a spill of a Proj must be after all Projs of the
3416 * Here's one special case:
3417 * If the spill is in the start block, the spill must be after the frame
3418 * pointer is set up. This is done by setting insert to the end of the block
3419 * which is its default initialization (see above).
3422 if (bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3425 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3428 sched_add_after(insert, spill);
3433 delete_remat(spill_ilp_t * si, ir_node * remat) {
3435 ir_node *bad = get_irg_bad(si->birg->irg);
3437 sched_remove(remat);
3439 /* kill links to operands */
3440 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3441 set_irn_n(remat, n, bad);
3446 clean_remat_info(spill_ilp_t * si)
3450 remat_info_t *remat_info;
3451 ir_node *bad = get_irg_bad(si->birg->irg);
3453 set_foreach(si->remat_info, remat_info) {
3454 if(!remat_info->remats) continue;
3456 pset_foreach(remat_info->remats, remat)
3458 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3459 if(sched_is_scheduled(remat->proj)) {
3460 sched_remove((ir_node*)remat->proj);
3462 set_irn_n((ir_node*)remat->proj, -1, bad);
3463 set_irn_n((ir_node*)remat->proj, 0, bad);
3466 if(get_irn_n_edges(remat->op) == 0) {
3467 if(sched_is_scheduled(remat->op)) {
3468 sched_remove((ir_node*)remat->op);
3470 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3471 set_irn_n((ir_node*)remat->op, n, bad);
3476 if(remat_info->remats) del_pset(remat_info->remats);
3477 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3482 delete_unnecessary_remats(spill_ilp_t * si)
3484 if(opt_keep_alive & KEEPALIVE_REMATS) {
3486 ir_node *bad = get_irg_bad(si->birg->irg);
3489 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3490 ir_node *keep_arg = get_irn_n(si->keep, n);
3491 op_t *arg_op = get_irn_link(keep_arg);
3494 assert(arg_op->is_remat);
3496 name = si->lpp->vars[arg_op->attr.remat.ilp];
3498 if(is_zero(name->value)) {
3499 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
3500 /* TODO check whether reload is preferred over remat (could be bug) */
3501 delete_remat(si, keep_arg);
3503 if(!arg_op->attr.remat.remat->inverse) {
3504 if(arg_op->attr.remat.pre) {
3505 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3507 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3510 if(arg_op->attr.remat.pre) {
3511 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3513 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3518 set_irn_n(si->keep, n, bad);
3521 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
3526 pset_foreach(si->all_possible_remats, remat) {
3527 op_t *remat_op = get_irn_link(remat);
3528 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
3530 if(is_zero(name->value)) {
3531 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
3532 /* TODO check whether reload is preferred over remat (could be bug) */
3533 delete_remat(si, remat);
3535 if(!remat_op->attr.remat.remat->inverse) {
3536 if(remat_op->attr.remat.pre) {
3537 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3539 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3542 if(remat_op->attr.remat.pre) {
3543 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3545 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3554 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3556 pset *spills = pset_new_ptr_default();
3558 const ir_node *next;
3561 defs = set_find_def(si->values, value);
3563 if(defs && defs->spills) {
3564 for(next = defs->spills; next; next = get_irn_link(next)) {
3565 pset_insert_ptr(spills, next);
3573 new_r_PhiM_nokeep(ir_graph * irg, ir_node *block, int arity, ir_node **in)
3577 assert( get_irn_arity(block) == arity );
3579 res = new_ir_node(NULL, irg, block, op_Phi, mode_M, arity, in);
3580 res->attr.phi_backedge = new_backedge_arr(irg->obst, arity);
3586 * @param before The node after which the spill will be placed in the schedule
3589 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3593 const arch_env_t *arch_env = si->birg->main_env->arch_env;
3595 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3597 spill = be_spill2(arch_env, irn, before);
3599 defs = set_insert_def(si->values, value);
3602 /* enter into the linked list */
3603 set_irn_link(spill, defs->spills);
3604 defs->spills = spill;
3606 if(opt_keep_alive & KEEPALIVE_SPILLS)
3607 pset_insert_ptr(si->spills, spill);
3613 * @param before The Phi node which has to be spilled
3616 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3623 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3625 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3626 ins[n] = si->m_unknown;
3629 mem_phi = new_r_PhiM_nokeep(si->birg->irg, get_nodes_block(phi), get_irn_arity(phi), ins);
3631 defs = set_insert_def(si->values, phi);
3634 /* enter into the linked list */
3635 set_irn_link(mem_phi, defs->spills);
3636 defs->spills = mem_phi;
3638 #ifdef SCHEDULE_PHIM
3639 sched_add_after(phi, mem_phi);
3641 pset_insert_ptr(si->phims, mem_phi);
3644 if(opt_keep_alive & KEEPALIVE_SPILLS)
3645 pset_insert_ptr(si->spills, mem_phi);
3652 * Add remat to list of defs, destroys link field!
3655 insert_remat(spill_ilp_t * si, ir_node * remat)
3658 op_t *remat_op = get_irn_link(remat);
3660 assert(remat_op->is_remat);
3662 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3665 /* enter into the linked list */
3666 set_irn_link(remat, defs->remats);
3667 defs->remats = remat;
3672 * Add reload before operation and add to list of defs
3675 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3680 const arch_env_t *arch_env = si->birg->main_env->arch_env;
3682 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3684 defs = set_find_def(si->values, value);
3686 spill = defs->spills;
3687 assert(spill && "no spill placed before reload");
3689 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3691 /* enter into the linked list */
3692 set_irn_link(reload, defs->remats);
3693 defs->remats = reload;
3698 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3701 ir_node *value = get_irn_n(memoperand->irn, memoperand->pos);
3703 const arch_env_t *arch_env = si->birg->main_env->arch_env;
3705 DBG((si->dbg, LEVEL_2, "\t inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3707 defs = set_find_def(si->values, value);
3709 spill = defs->spills;
3710 assert(spill && "no spill placed before reload");
3712 arch_perform_memory_operand(arch_env, memoperand->irn, spill, memoperand->pos);
3715 void insert_memoperands(spill_ilp_t * si)
3717 memoperand_t *memoperand;
3720 set_foreach(si->memoperands, memoperand) {
3721 name = si->lpp->vars[memoperand->ilp];
3722 if(!is_zero(name->value)) {
3723 perform_memory_operand(si, memoperand);
3729 walker_spill_placer(ir_node * bb, void * data) {
3730 spill_ilp_t *si = (spill_ilp_t*)data;
3732 spill_bb_t *spill_bb = get_irn_link(bb);
3733 pset *spills_to_do = pset_new_ptr_default();
3736 set_foreach(spill_bb->ilp, spill) {
3739 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3740 name = si->lpp->vars[spill->mem_in];
3741 if(!is_zero(name->value)) {
3744 mem_phi = insert_mem_phi(si, spill->irn);
3746 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3750 name = si->lpp->vars[spill->spill];
3751 if(!is_zero(name->value)) {
3752 /* place spill directly after definition */
3753 if(get_nodes_block(spill->irn) == bb) {
3754 insert_spill(si, spill->irn, spill->irn, spill->irn);
3758 /* place spill at bb start */
3759 if(spill->reg_in > 0) {
3760 name = si->lpp->vars[spill->reg_in];
3761 if(!is_zero(name->value)) {
3762 insert_spill(si, spill->irn, spill->irn, bb);
3766 /* place spill after a remat */
3767 pset_insert_ptr(spills_to_do, spill->irn);
3770 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3773 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3774 op_t *op = get_irn_link(irn);
3776 if(be_is_Spill(irn)) continue;
3779 /* TODO fix this if we want to support remats with more than two nodes */
3780 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3781 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3783 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3786 if(pset_find_ptr(spills_to_do, irn)) {
3787 pset_remove_ptr(spills_to_do, irn);
3789 insert_spill(si, irn, irn, irn);
3795 assert(pset_count(spills_to_do) == 0);
3797 /* afterwards free data in block */
3798 del_pset(spills_to_do);
3802 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3804 ir_node *insert_pos = bb;
3806 const arch_env_t *arch_env = si->birg->main_env->arch_env;
3808 /* find last definition of arg value in block */
3813 defs = set_find_def(si->values, value);
3815 if(defs && defs->remats) {
3816 for(next = defs->remats; next; next = get_irn_link(next)) {
3817 if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3818 last = sched_get_time_step(next);
3824 if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3825 last = sched_get_time_step(value);
3829 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
3831 spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3837 phim_fixer(spill_ilp_t *si) {
3840 set_foreach(si->values, defs) {
3841 const ir_node *phi = defs->value;
3842 op_t *op = get_irn_link(phi);
3843 ir_node *phi_m = NULL;
3844 ir_node *next = defs->spills;
3847 if(!is_Phi(phi)) continue;
3850 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3854 next = get_irn_link(next);
3857 if(!phi_m) continue;
3859 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3860 ir_node *value = get_irn_n(phi, n);
3861 defs_t *val_defs = set_find_def(si->values, value);
3863 /* a spill of this value */
3868 ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3869 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3871 if(!is_zero(name->value)) {
3872 spill = insert_mem_copy(si, pred, value);
3874 spill = val_defs->spills;
3877 spill = val_defs->spills;
3880 assert(spill && "no spill placed before PhiM");
3881 set_irn_n(phi_m, n, spill);
3887 walker_reload_placer(ir_node * bb, void * data) {
3888 spill_ilp_t *si = (spill_ilp_t*)data;
3890 spill_bb_t *spill_bb = get_irn_link(bb);
3892 /* reloads at end of block */
3893 if(spill_bb->reloads) {
3896 set_foreach(spill_bb->reloads, keyval) {
3897 ir_node *irn = (ir_node*)keyval->key;
3898 ilp_var_t reload = PTR_TO_INT(keyval->val);
3901 name = si->lpp->vars[reload];
3902 if(!is_zero(name->value)) {
3904 ir_node *insert_pos = bb;
3905 ir_node *prev = sched_block_last_noncf(si, bb);
3906 op_t *prev_op = get_irn_link(prev);
3908 while(be_is_Spill(prev)) {
3909 prev = sched_prev(prev);
3912 prev_op = get_irn_link(prev);
3914 /* insert reload before pre-remats */
3915 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3916 && prev_op->is_remat && prev_op->attr.remat.pre) {
3920 prev = sched_prev(prev);
3921 } while(be_is_Spill(prev));
3923 prev_op = get_irn_link(prev);
3927 reload = insert_reload(si, irn, insert_pos);
3929 if(opt_keep_alive & KEEPALIVE_RELOADS)
3930 pset_insert_ptr(si->spills, reload);
3935 /* walk and insert more reloads and collect remats */
3936 sched_foreach_reverse(bb, irn) {
3937 op_t *op = get_irn_link(irn);
3939 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3940 if(is_Phi(irn)) break;
3943 if(get_irn_mode(irn) != mode_T) {
3944 insert_remat(si, irn);
3949 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3950 ir_node *arg = get_irn_n(irn, n);
3952 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3955 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3956 if(!is_zero(name->value)) {
3958 ir_node *insert_pos = irn;
3959 ir_node *prev = sched_prev(insert_pos);
3962 while(be_is_Spill(prev)) {
3963 prev = sched_prev(prev);
3966 prev_op = get_irn_link(prev);
3968 /* insert reload before pre-remats */
3969 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3970 && prev_op->is_remat && prev_op->attr.remat.pre) {
3974 prev = sched_prev(prev);
3975 } while(be_is_Spill(prev));
3977 prev_op = get_irn_link(prev);
3981 reload = insert_reload(si, arg, insert_pos);
3983 assert(reload && "no reload returned");
3984 set_irn_n(irn, n, reload);
3986 if(opt_keep_alive & KEEPALIVE_RELOADS)
3987 pset_insert_ptr(si->spills, reload);
3994 del_set(spill_bb->ilp);
3995 if(spill_bb->reloads) del_set(spill_bb->reloads);
3999 walker_collect_used(ir_node * irn, void * data)
4001 bitset_t *used = data;
4003 bitset_set(used, get_irn_idx(irn));
4006 struct kill_helper {
4012 walker_kill_unused(ir_node * bb, void * data)
4014 struct kill_helper *kh = data;
4015 ir_node *bad = get_irg_bad(get_irn_irg(bb));
4019 for(irn=sched_first(bb); !sched_is_end(irn);) {
4020 ir_node *next = sched_next(irn);
4023 if(!bitset_is_set(kh->used, get_irn_idx(irn))) {
4024 if(be_is_Spill(irn) || be_is_Reload(irn)) {
4025 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)));
4027 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
4033 set_nodes_block(irn, bad);
4034 for (n=get_irn_arity(irn)-1; n>=0; --n) {
4035 set_irn_n(irn, n, bad);
4042 #ifndef SCHEDULE_PHIM
4044 kill_unused_phims(spill_ilp_t * si, struct kill_helper * kh)
4047 ir_node *bad = get_irg_bad(si->birg->irg);
4050 pset_foreach(si->phims, phi) {
4051 if(!bitset_is_set(kh->used, get_irn_idx(phi))) {
4053 set_nodes_block(phi, bad);
4054 for (n=get_irn_arity(phi)-1; n>=0; --n) {
4055 set_irn_n(phi, n, bad);
4063 kill_all_unused_values_in_schedule(spill_ilp_t * si)
4065 struct kill_helper kh;
4067 kh.used = bitset_malloc(get_irg_last_idx(si->birg->irg));
4070 irg_walk_graph(si->birg->irg, walker_collect_used, NULL, kh.used);
4071 #ifndef SCHEDULE_PHIM
4072 kill_unused_phims(si, &kh);
4074 irg_block_walk_graph(si->birg->irg, walker_kill_unused, NULL, &kh);
4076 bitset_free(kh.used);
4080 print_irn_pset(pset * p)
4084 pset_foreach(p, irn) {
4085 ir_printf("%+F\n", irn);
4090 dump_phi_class(spill_ilp_t *si, ir_node **phiclass, const char * file)
4092 FILE *f = fopen(file, "w");
4094 interference_t *interference;
4097 set_break(si->interferences);
4099 ir_fprintf(f, "digraph phiclass {\n");
4101 for (i = ARR_LEN(phiclass) - 1; i >= 0; --i) {
4104 ir_fprintf(f, " %F%N [shape=box]\n", irn, irn);
4107 for (i = ARR_LEN(phiclass) - 1; i >= 0; --i) {
4114 for (n = get_irn_arity(irn) - 1; n >= 0; --n) {
4115 ir_node *arg = get_irn_n(irn, n);
4117 ir_fprintf(f, " %F%N -> %F%N\n", irn, irn, arg, arg);
4121 set_foreach(si->interferences, interference) {
4122 const ir_node *a = interference->a;
4123 const ir_node *b = interference->b;
4124 if (get_phi_class(si->pc, a) == phiclass) {
4125 ir_fprintf(f, " %F%N -> %F%N [color=red,dir=none,style=bold]\n", a, a, b, b);
4134 rewire_uses(spill_ilp_t * si)
4137 pset *ignore = pset_new_ptr(1);
4138 be_dom_front_info_t *dom_front = si->birg->dom_front;
4140 pset_insert_ptr(ignore, get_irg_end(si->birg->irg));
4142 /* then fix uses of spills */
4143 set_foreach(si->values, defs) {
4146 const ir_node *next = defs->remats;
4149 reloads = pset_new_ptr_default();
4152 if(be_is_Reload(next)) {
4153 pset_insert_ptr(reloads, next);
4157 next = get_irn_link(next);
4160 spills = get_spills_for_value(si, defs->value);
4161 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));
4162 if(pset_count(spills) > 1) {
4163 //assert(pset_count(reloads) > 0);
4164 // print_irn_pset(spills);
4165 // print_irn_pset(reloads);
4167 be_ssa_constr_set_ignore(dom_front, si->lv, spills, ignore);
4174 /* first fix uses of remats and reloads */
4175 set_foreach(si->values, defs) {
4177 const ir_node *next = defs->remats;
4181 nodes = pset_new_ptr_default();
4182 if(sched_is_scheduled(defs->value)) {
4183 pset_insert_ptr(nodes, defs->value);
4188 pset_insert_ptr(nodes, next);
4189 next = get_irn_link(next);
4192 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-orig_kept, defs->value));
4193 be_ssa_constr_set(dom_front, si->lv, nodes);
4199 // remove_unused_defs(si);
4204 writeback_results(spill_ilp_t * si)
4206 /* walk through the graph and collect all spills, reloads and remats for a value */
4208 si->values = new_set(cmp_defs, 4096);
4210 DBG((si->dbg, LEVEL_1, "Applying results\n"));
4211 delete_unnecessary_remats(si);
4212 si->m_unknown = new_r_Unknown(si->birg->irg, mode_M);
4213 irg_block_walk_graph(si->birg->irg, walker_spill_placer, NULL, si);
4214 irg_block_walk_graph(si->birg->irg, walker_reload_placer, NULL, si);
4216 insert_memoperands(si);
4219 /* clean the remat info! there are still back-edges leading there! */
4220 clean_remat_info(si);
4224 connect_all_spills_with_keep(si);
4226 del_set(si->values);
4230 get_n_regs(spill_ilp_t * si)
4232 int arch_n_regs = arch_register_class_n_regs(si->cls);
4234 bitset_t *arch_regs = bitset_malloc(arch_n_regs);
4235 bitset_t *abi_regs = bitset_malloc(arch_n_regs);
4237 arch_put_non_ignore_regs(si->birg->main_env->arch_env, si->cls, arch_regs);
4238 be_abi_put_ignore_regs(si->birg->abi, si->cls, abi_regs);
4240 bitset_andnot(arch_regs, abi_regs);
4241 arch_n_regs = bitset_popcnt(arch_regs);
4243 bitset_free(arch_regs);
4244 bitset_free(abi_regs);
4246 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", arch_n_regs, si->cls->name));
4251 walker_reload_mover(ir_node * bb, void * data)
4253 spill_ilp_t *si = data;
4256 sched_foreach(bb, tmp) {
4257 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
4258 ir_node *reload = tmp;
4261 /* move reload upwards */
4263 int pressure = (int)get_irn_link(reload);
4264 if(pressure < si->n_regs) {
4265 irn = sched_prev(reload);
4266 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
4267 sched_remove(reload);
4268 pressure = (int)get_irn_link(irn);
4270 while(pressure < si->n_regs) {
4271 if( sched_is_end(irn) ||
4272 (be_is_Reload(irn) && has_reg_class(si, irn)) ||
4273 /* do not move reload before its spill */
4274 (irn == be_get_Reload_mem(reload)) ||
4275 /* do not move before phi */
4278 set_irn_link(irn, INT_TO_PTR(pressure+1));
4279 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
4280 irn = sched_prev(irn);
4282 pressure = (int)get_irn_link(irn);
4285 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
4286 sched_put_after(irn, reload);
4293 move_reloads_upward(spill_ilp_t * si)
4295 irg_block_walk_graph(si->birg->irg, walker_reload_mover, NULL, si);
4300 * Walk all irg blocks and check for interfering spills inside of phi classes
4303 luke_meminterferencechecker(ir_node * bb, void * data)
4305 spill_ilp_t *si = (spill_ilp_t*)data;
4308 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
4309 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
4311 if (! be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T))
4314 /* a is only interesting if it is in my register class and if it is inside a phi class */
4315 if (has_reg_class(si, a) && get_phi_class(si->pc, a)) {
4316 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)) {
4317 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
4319 if (! be_is_Spill(b) && (! is_Phi(b) || get_irn_mode(b) != mode_T))
4322 /* a and b are only interesting if they are in the same phi class */
4323 if (has_reg_class(si, b) && get_phi_class(si->pc, a) == get_phi_class(si->pc, b)) {
4324 if (values_interfere_in_block(si, bb, a, b)) {
4325 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4334 verify_phiclasses(spill_ilp_t * si)
4336 /* analyze phi classes */
4337 phi_class_free(si->pc);
4338 si->pc = phi_class_new_from_irg(si->birg->irg);
4340 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4341 irg_block_walk_graph(si->birg->irg, luke_meminterferencechecker, NULL, si);
4345 be_spill_remat(be_irg_t *birg, const arch_register_class_t *cls)
4348 char problem_name[256];
4349 char dump_suffix[256];
4350 char dump_suffix2[256];
4351 struct obstack obst;
4353 ir_graph *irg = be_get_birg_irg(birg);
4355 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", irg, cls->name);
4356 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", cls->name);
4357 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", cls->name);
4359 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4360 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4362 if(opt_verify & VERIFY_DOMINANCE)
4363 be_check_dominance(irg);
4365 be_assure_dom_front(birg);
4366 be_assure_liveness(birg);
4368 obstack_init(&obst);
4372 si.lpp = new_lpp(problem_name, lpp_minimize);
4373 si.remat_info = new_set(cmp_remat_info, 4096);
4374 si.interferences = new_set(cmp_interference, 32);
4375 si.memoperands = new_set(cmp_memoperands, 128);
4376 si.all_possible_remats = pset_new_ptr_default();
4377 si.spills = pset_new_ptr_default();
4378 si.inverse_ops = pset_new_ptr_default();
4381 si.n_regs = get_n_regs(&si);
4383 set_irg_link(irg, &si);
4386 /* compute phi classes */
4387 // phi_class_compute(irg);
4389 if(opt_dump_flags & DUMP_STATS)
4390 be_analyze_regpressure(birg, cls, "-pre");
4392 DBG((si.dbg, LEVEL_2, "\t initializing\n"));
4393 irg_block_walk_graph(irg, luke_initializer, NULL, &si);
4396 /* collect remats */
4397 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4398 irg_walk_graph(irg, walker_remat_collector, NULL, &si);
4401 /* insert possible remats */
4402 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4403 irg_block_walk_graph(irg, walker_remat_insertor, NULL, &si);
4404 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4406 if(opt_keep_alive & KEEPALIVE_REMATS) {
4407 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4408 connect_all_remats_with_keep(&si);
4409 /* dump graph with inserted remats */
4410 dump_graph_with_remats(irg, dump_suffix);
4413 /* insert copies for phi arguments not in my regclass */
4414 irg_walk_graph(irg, walker_regclass_copy_insertor, NULL, &si);
4416 /* recompute liveness */
4417 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4418 be_liveness_recompute(si.lv);
4421 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4422 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4423 irg_block_walk_graph(irg, luke_endwalker, NULL, &si);
4425 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4426 irg_block_walk_graph(irg, luke_blockwalker, NULL, &si);
4428 si.pc = phi_class_new_from_irg(birg->irg);
4429 if (opt_memcopies) {
4430 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4431 memcopyhandler(&si);
4434 if (opt_dump_flags & DUMP_PROBLEM) {
4436 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4437 if ((f = fopen(buf, "wt")) != NULL) {
4438 lpp_dump_plain(si.lpp, f);
4443 if (opt_dump_flags & DUMP_MPS) {
4446 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4447 if ((f = fopen(buf, "wt")) != NULL) {
4448 mps_write_mps(si.lpp, s_mps_fixed, f);
4452 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4453 if ((f = fopen(buf, "wt")) != NULL) {
4454 mps_write_mst(si.lpp, s_mps_fixed, f);
4459 lpp_check_startvals(si.lpp);
4462 DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4463 lpp_set_time_limit(si.lpp, opt_timeout);
4466 lpp_set_log(si.lpp, stdout);
4469 lpp_solve_cplex(si.lpp);
4471 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4473 assert(lpp_is_sol_valid(si.lpp)
4474 && "solution of ILP must be valid");
4476 DBG((si.dbg, LEVEL_1, "\t%s: iterations: %d, solution time: %g, objective function: %g, best bound: %g\n", problem_name, si.lpp->iterations, si.lpp->sol_time, is_zero(si.lpp->objval)?0.0:si.lpp->objval, is_zero(si.lpp->best_bound)?0.0:si.lpp->best_bound));
4478 if(opt_dump_flags & DUMP_SOLUTION) {
4482 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4483 if ((f = fopen(buf, "wt")) != NULL) {
4485 for (i = 0; i < si.lpp->var_next; ++i) {
4486 lpp_name_t *name = si.lpp->vars[i];
4487 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4493 #ifndef SCHEDULE_PHIM
4494 si.phims = pset_new_ptr_default();
4496 writeback_results(&si);
4501 kill_all_unused_values_in_schedule(&si);
4503 #if !defined(SCHEDULE_PHIM) && defined(SOLVE)
4507 if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
4508 be_dump(irg, "-spills-placed", dump_ir_block_graph);
4510 // move reloads upwards
4511 be_liveness_recompute(si.lv);
4512 irg_block_walk_graph(irg, walker_pressure_annotator, NULL, &si);
4513 move_reloads_upward(&si);
4516 verify_phiclasses(&si);
4519 irg_block_walk_graph(irg, walker_pressure_annotator, NULL, &si);
4521 if(opt_dump_flags & DUMP_PRESSURE)
4522 dump_pressure_graph(&si, dump_suffix2);
4524 if(opt_dump_flags & DUMP_STATS)
4525 be_analyze_regpressure(birg, cls, "-post");
4527 if(opt_verify & VERIFY_DOMINANCE)
4528 be_check_dominance(irg);
4531 del_set(si.interferences);
4532 del_pset(si.inverse_ops);
4533 del_pset(si.all_possible_remats);
4534 del_set(si.memoperands);
4535 del_pset(si.spills);
4537 phi_class_free(si.pc);
4538 obstack_free(&obst, NULL);
4539 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4542 void be_init_spillremat(void)
4544 static be_spiller_t remat_spiller = {
4547 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
4548 lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
4549 lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
4550 lc_opt_entry_t *remat_grp = lc_opt_get_grp(chordal_grp, "remat");
4552 be_register_spiller("remat", &remat_spiller);
4553 lc_opt_add_table(remat_grp, options);
4556 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillremat);
4558 #else /* WITH_ILP */
4561 only_that_you_can_compile_without_WITH_ILP_defined(void)
4565 #endif /* WITH_ILP */