1 /** vim: set sw=4 ts=4:
4 * @author Adam M. Szalkowski & Sebastian Hack
6 * ILP based spilling & rematerialization
8 * Copyright (C) 2006 Universitaet Karlsruhe
9 * Released under the GPL
32 #include "phiclass_t.h"
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 */
192 #ifndef SCHEDULE_PHIM
195 DEBUG_ONLY(firm_dbg_module_t * dbg);
198 typedef int ilp_var_t;
199 typedef int ilp_cst_t;
201 typedef struct _spill_bb_t {
206 typedef struct _remat_t {
207 const ir_node *op; /**< for copy_irn */
208 const ir_node *value; /**< the value which is being recomputed by this remat */
209 const ir_node *proj; /**< not NULL if the above op produces a tuple */
210 int cost; /**< cost of this remat */
211 int inverse; /**< nonzero if this is an inverse remat */
215 * Data to be attached to each IR node. For remats this contains the ilp_var
216 * for this remat and for normal ops this contains the ilp_vars for
217 * reloading each operand
219 typedef struct _op_t {
224 const remat_t *remat; /** the remat this op belongs to */
225 int pre; /** 1, if this is a pressure-increasing remat */
229 ir_node *op; /** the operation this live range belongs to */
238 typedef struct _defs_t {
239 const ir_node *value;
240 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
241 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
244 typedef struct _remat_info_t {
245 const ir_node *irn; /**< the irn to which these remats belong */
246 pset *remats; /**< possible remats for this value */
247 pset *remats_by_operand; /**< remats with this value as operand */
250 typedef struct _keyval_t {
255 typedef struct _spill_t {
264 typedef struct _memoperand_t {
265 ir_node *irn; /**< the irn */
266 unsigned int pos; /**< the position of the argument */
267 ilp_var_t ilp; /**< the ilp var for this memory operand */
271 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
273 return arch_irn_consider_in_reg_alloc(si->birg->main_env->arch_env,
279 cmp_remat(const void *a, const void *b)
281 const keyval_t *p = a;
282 const keyval_t *q = b;
283 const remat_t *r = p->val;
284 const remat_t *s = q->val;
288 return !(r == s || r->op == s->op);
292 cmp_remat(const void *a, const void *b)
294 const remat_t *r = a;
295 const remat_t *s = a;
297 return !(r == s || r->op == s->op);
301 cmp_spill(const void *a, const void *b, size_t size)
303 const spill_t *p = a;
304 const spill_t *q = b;
306 // return !(p->irn == q->irn && p->bb == q->bb);
307 return !(p->irn == q->irn);
311 cmp_memoperands(const void *a, const void *b, size_t size)
313 const memoperand_t *p = a;
314 const memoperand_t *q = b;
316 return !(p->irn == q->irn && p->pos == q->pos);
320 set_find_keyval(set * set, const void * key)
325 return set_find(set, &query, sizeof(query), HASH_PTR(key));
329 set_insert_keyval(set * set, void * key, void * val)
335 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
339 set_find_def(set * set, const ir_node * value)
344 return set_find(set, &query, sizeof(query), HASH_PTR(value));
348 set_insert_def(set * set, const ir_node * value)
355 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
358 static memoperand_t *
359 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
366 return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
369 static memoperand_t *
370 set_find_memoperand(set * set, const ir_node * irn, unsigned int pos)
374 query.irn = (ir_node*)irn;
376 return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
381 set_find_spill(set * set, const ir_node * value)
385 query.irn = (ir_node*)value;
386 return set_find(set, &query, sizeof(query), HASH_PTR(value));
389 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
390 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
391 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
392 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
393 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
396 cmp_remat_info(const void *a, const void *b, size_t size)
398 const remat_info_t *p = a;
399 const remat_info_t *q = b;
401 return !(p->irn == q->irn);
405 cmp_defs(const void *a, const void *b, size_t size)
410 return !(p->value == q->value);
414 cmp_keyval(const void *a, const void *b, size_t size)
416 const keyval_t *p = a;
417 const keyval_t *q = b;
419 return !(p->key == q->key);
423 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
426 if(be_profile_has_data())
427 return ((double)be_profile_get_block_execcount(get_block(irn))) + FUDGE;
429 #ifndef EXECFREQ_LOOPDEPH
430 return get_block_execfreq(si->birg->exec_freq, get_block(irn)) + FUDGE;
433 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
435 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
440 get_cost(const spill_ilp_t * si, const ir_node * irn)
442 if(be_is_Spill(irn)) {
443 return opt_cost_spill;
444 } else if(be_is_Reload(irn)){
445 return opt_cost_reload;
447 return arch_get_op_estimated_cost(si->birg->main_env->arch_env, irn);
452 * Checks, whether node and its operands have suitable reg classes
455 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
458 const arch_env_t *arch_env = si->birg->main_env->arch_env;
459 int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
463 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
466 for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
467 ir_node *op = get_irn_n(irn, n);
468 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
471 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
478 * Try to create a remat from @p op with destination value @p dest_value
480 static INLINE remat_t *
481 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
483 remat_t *remat = NULL;
485 // if(!mode_is_datab(get_irn_mode(dest_value)))
488 if(dest_value == op) {
489 const ir_node *proj = NULL;
491 if(is_Proj(dest_value)) {
492 op = get_Proj_pred(op);
496 if(!is_rematerializable(si, op))
499 remat = obstack_alloc(si->obst, sizeof(*remat));
501 remat->cost = (int)get_cost(si, op);
502 remat->value = dest_value;
506 arch_inverse_t inverse;
509 /* get the index of the operand we want to retrieve by the inverse op */
510 for (n = get_irn_arity(op)-1; n>=0; --n) {
511 ir_node *arg = get_irn_n(op, n);
513 if(arg == dest_value) break;
517 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", n, op));
519 /* else ask the backend to give an inverse op */
520 if(arch_get_inverse(si->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
523 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
525 assert(inverse.n > 0 && "inverse op should have at least one node");
527 for(i=inverse.n-1; i>=0; --i) {
528 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
532 remat = obstack_alloc(si->obst, sizeof(*remat));
533 remat->op = inverse.nodes[0];
534 remat->cost = inverse.costs;
535 remat->value = dest_value;
536 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
539 // Matze: commented this out, this doesn't seem to be true if
540 // the inverse is a simple operation with only 1 result...
541 //assert(is_Proj(remat->proj));
543 assert(0 && "I can not handle remats with more than 2 nodes");
550 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
552 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
560 add_remat(const spill_ilp_t * si, const remat_t * remat)
562 remat_info_t *remat_info,
567 assert(remat->value);
569 query.irn = remat->value;
571 query.remats_by_operand = NULL;
572 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
574 if(remat_info->remats == NULL) {
575 remat_info->remats = new_pset(cmp_remat, 4096);
577 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
579 /* insert the remat into the remats_be_operand set of each argument of the remat op */
580 for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
581 ir_node *arg = get_irn_n(remat->op, n);
585 query.remats_by_operand = NULL;
586 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
588 if(remat_info->remats_by_operand == NULL) {
589 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
591 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
596 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
598 const ir_edge_t *edge = get_irn_out_edge_first(irn);
602 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
605 edge = get_irn_out_edge_next(irn, edge);
612 get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
618 irn = get_Proj_pred(irn);
620 for(n=get_irn_arity(irn)-1; n>=0; --n) {
621 const ir_node *arg = get_irn_n(irn, n);
623 if(has_reg_class(si, arg)) ++ret;
630 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
635 if( has_reg_class(si, op)
636 && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1)
637 && (opt_remats != REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0)
639 remat = get_remat_from_op(si, op, op);
641 add_remat(si, remat);
645 if(opt_remats == REMATS_ALL) {
646 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
648 for (n = get_irn_arity(op)-1; n>=0; --n) {
649 ir_node *arg = get_irn_n(op, n);
651 if(has_reg_class(si, arg)) {
652 /* try to get an inverse remat */
653 remat = get_remat_from_op(si, arg, op);
655 add_remat(si, remat);
663 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
666 ir_node *def_block = get_nodes_block(val);
672 /* if pos is at end of a basic block */
674 ret = (pos == def_block || block_dominates(def_block, pos));
675 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
679 /* else if this is a normal operation */
680 block = get_nodes_block(pos);
681 if(block == def_block) {
682 if(!sched_is_scheduled(val)) return 1;
684 ret = sched_comes_after(val, pos);
685 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
689 ret = block_dominates(def_block, block);
690 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
694 static INLINE ir_node *
695 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
697 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->birg->main_env->arch_env);
701 * Returns first non-Phi node of block @p bb
703 static INLINE ir_node *
704 sched_block_first_nonphi(const ir_node * bb)
706 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
710 sched_skip_proj_predicator(const ir_node * irn, void * data)
712 return (is_Proj(irn));
715 static INLINE ir_node *
716 sched_next_nonproj(const ir_node * irn, int forward)
718 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
722 * Returns next operation node (non-Proj) after @p irn
723 * or the basic block of this node
725 static INLINE ir_node *
726 sched_next_op(const ir_node * irn)
728 ir_node *next = sched_next(irn);
733 return sched_next_nonproj(next, 1);
737 * Returns previous operation node (non-Proj) before @p irn
738 * or the basic block of this node
740 static INLINE ir_node *
741 sched_prev_op(const ir_node * irn)
743 ir_node *prev = sched_prev(irn);
748 return sched_next_nonproj(prev, 0);
752 sched_put_after(ir_node * insert, ir_node * irn)
754 if(is_Block(insert)) {
755 insert = sched_block_first_nonphi(insert);
757 insert = sched_next_op(insert);
760 sched_add_before(insert, irn);
764 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
766 if(is_Block(insert)) {
767 insert = sched_block_last_noncf(si, insert);
769 insert = sched_next_nonproj(insert, 0);
770 insert = sched_prev(insert);
773 sched_add_after(insert, irn);
777 next_post_remat(const ir_node * irn)
783 next = sched_block_first_nonphi(irn);
785 next = sched_next_op(irn);
788 if(sched_is_end(next))
791 op = get_irn_link(next);
792 if(op->is_remat && !op->attr.remat.pre) {
801 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
807 ret = sched_block_last_noncf(si, irn);
808 ret = sched_next(ret);
809 ret = sched_prev_op(ret);
811 ret = sched_prev_op(irn);
814 if(sched_is_end(ret) || is_Phi(ret))
817 op = (op_t*)get_irn_link(ret);
818 if(op->is_remat && op->attr.remat.pre) {
826 * Tells you whether a @p remat can be placed before the irn @p pos
829 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
831 const ir_node *op = remat->op;
837 prev = sched_block_last_noncf(si, pos);
838 prev = sched_next_nonproj(prev, 0);
840 prev = sched_prev_op(pos);
842 /* do not remat if the rematted value is defined immediately before this op */
843 if(prev == remat->op) {
848 /* this should be just fine, the following OP will be using this value, right? */
850 /* only remat AFTER the real definition of a value (?) */
851 if(!value_is_defined_before(si, pos, remat->value)) {
852 // ir_fprintf(stderr, "error(not defined)");
857 for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
858 const ir_node *arg = get_irn_n(op, n);
860 if(opt_no_enlarge_liveness) {
861 if(has_reg_class(si, arg) && live) {
862 res &= pset_find_ptr((pset*)live, arg)?1:0;
864 res &= value_is_defined_before(si, pos, arg);
867 res &= value_is_defined_before(si, pos, arg);
875 * Tells you whether a @p remat can be placed after the irn @p pos
878 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
881 pos = sched_block_first_nonphi(pos);
883 pos = sched_next_op(pos);
886 /* only remat AFTER the real definition of a value (?) */
887 if(!value_is_defined_before(si, pos, remat->value)) {
891 return can_remat_before(si, remat, pos, live);
895 * Collect potetially rematerializable OPs
898 walker_remat_collector(ir_node * irn, void * data)
900 spill_ilp_t *si = data;
902 if(!is_Block(irn) && !is_Phi(irn)) {
903 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
904 get_remats_from_op(si, irn);
909 * Inserts a copy of @p irn before @p pos
912 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
917 bb = is_Block(pos)?pos:get_nodes_block(pos);
918 copy = exact_copy(irn);
920 _set_phi_class(copy, NULL);
921 set_nodes_block(copy, bb);
922 sched_put_before(si, pos, copy);
928 * Inserts a copy of @p irn after @p pos
931 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
936 bb = is_Block(pos)?pos:get_nodes_block(pos);
937 copy = exact_copy(irn);
939 _set_phi_class(copy, NULL);
940 set_nodes_block(copy, bb);
941 sched_put_after(pos, copy);
947 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
951 if(can_remat_after(si, remat, pos, live)) {
956 DBG((si->dbg, LEVEL_3, "\t >inserting remat2 %+F\n", remat->op));
958 copy = insert_copy_after(si, remat->op, pos);
960 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
961 op = obstack_alloc(si->obst, sizeof(*op));
963 op->attr.remat.remat = remat;
964 op->attr.remat.pre = 0;
965 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
967 set_irn_link(copy, op);
968 pset_insert_ptr(si->all_possible_remats, copy);
970 proj_copy = insert_copy_after(si, remat->proj, copy);
971 set_irn_n(proj_copy, 0, copy);
972 set_irn_link(proj_copy, op);
973 pset_insert_ptr(si->all_possible_remats, proj_copy);
985 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
989 if(can_remat_before(si, remat, pos, live)) {
994 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
996 copy = insert_copy_before(si, remat->op, pos);
998 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
999 op = obstack_alloc(si->obst, sizeof(*op));
1001 op->attr.remat.remat = remat;
1002 op->attr.remat.pre = 1;
1003 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
1005 set_irn_link(copy, op);
1006 pset_insert_ptr(si->all_possible_remats, copy);
1008 proj_copy = insert_copy_after(si, remat->proj, copy);
1009 set_irn_n(proj_copy, 0, copy);
1010 set_irn_link(proj_copy, op);
1011 pset_insert_ptr(si->all_possible_remats, proj_copy);
1023 get_block_n_succs(const ir_node *block) {
1024 const ir_edge_t *edge;
1026 assert(edges_activated(current_ir_graph));
1028 edge = get_block_succ_first(block);
1032 edge = get_block_succ_next(block, edge);
1033 return edge ? 2 : 1;
1037 is_start_block(const ir_node * bb)
1039 return get_irg_start_block(get_irn_irg(bb)) == bb;
1043 is_merge_edge(const ir_node * bb)
1045 if(is_start_block(bb))
1049 return get_block_n_succs(bb) == 1;
1055 is_diverge_edge(const ir_node * bb)
1057 if(is_start_block(bb))
1061 return get_Block_n_cfgpreds(bb) == 1;
1067 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1072 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1073 irn = be_lv_get_irn(si->lv, bb, i);
1075 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1076 pset_insert_ptr(live, irn);
1080 irn = sched_last(bb);
1082 /* all values eaten by control flow operations are also live until the end of the block */
1083 sched_foreach_reverse(bb, irn) {
1086 if(!sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) break;
1088 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1089 ir_node *arg = get_irn_n(irn,i);
1091 if(has_reg_class(si, arg)) {
1092 pset_insert_ptr(live, arg);
1097 * find values that are used by remats at end of block
1098 * and insert them into live set
1100 foreach_pre_remat(si, bb, irn) {
1103 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1104 ir_node *remat_arg = get_irn_n(irn, n);
1106 if(!has_reg_class(si, remat_arg)) continue;
1108 /* if value is becoming live through use by remat */
1109 if(!pset_find_ptr(live, remat_arg)) {
1110 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, irn));
1112 pset_insert_ptr(live, remat_arg);
1119 walker_regclass_copy_insertor(ir_node * irn, void * data)
1121 spill_ilp_t *si = data;
1123 if(is_Phi(irn) && has_reg_class(si, irn)) {
1126 for(n=get_irn_arity(irn)-1; n>=0; --n) {
1127 ir_node *phi_arg = get_irn_n(irn, n);
1128 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
1130 if(!has_reg_class(si, phi_arg)) {
1131 ir_node *copy = be_new_Copy(si->cls, si->birg->irg, bb, phi_arg);
1132 ir_node *pos = sched_block_last_noncf(si, bb);
1133 op_t *op = obstack_alloc(si->obst, sizeof(*op));
1135 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
1136 sched_add_after(pos, copy);
1137 set_irn_n(irn, n, copy);
1140 op->attr.live_range.args.reloads = NULL;
1141 op->attr.live_range.ilp = ILP_UNDEF;
1142 set_irn_link(copy, op);
1149 * Insert (so far unused) remats into the irg to
1150 * recompute the potential liveness of all values
1153 walker_remat_insertor(ir_node * bb, void * data)
1155 spill_ilp_t *si = data;
1162 /* skip start block, no remats to do there */
1163 if(is_start_block(bb)) return;
1165 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1167 live = pset_new_ptr_default();
1168 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1169 ir_node *value = be_lv_get_irn(si->lv, bb, i);
1171 /* add remats at end of block */
1172 if (has_reg_class(si, value)) {
1173 pset_insert_ptr(live, value);
1177 irn = sched_last(bb);
1178 while(!sched_is_end(irn)) {
1184 next = sched_prev(irn);
1186 /* delete defined value from live set */
1187 if(has_reg_class(si, irn)) {
1188 pset_remove_ptr(live, irn);
1191 if(is_Phi(irn) || is_Proj(irn)) {
1196 args = pset_new_ptr_default();
1197 used = pset_new_ptr_default();
1199 /* collect arguments of op and set args of op already live in epilog */
1200 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1201 ir_node *arg = get_irn_n(irn, n);
1203 pset_insert_ptr(args, arg);
1204 if(has_reg_class(si, arg)) {
1205 pset_insert_ptr(live, arg);
1206 pset_insert_ptr(used, arg);
1210 /* insert all possible remats before irn */
1211 pset_foreach(args, arg) {
1212 remat_info_t *remat_info,
1215 /* continue if the operand has the wrong reg class */
1216 if(!has_reg_class(si, arg))
1220 query.remats = NULL;
1221 query.remats_by_operand = NULL;
1222 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1228 if(remat_info->remats) {
1229 pset_foreach(remat_info->remats, remat) {
1230 ir_node *remat_irn = NULL;
1232 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
1233 remat_irn = insert_remat_before(si, remat, irn, live);
1236 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1237 ir_node *remat_arg = get_irn_n(remat_irn, n);
1239 /* collect args of remats which are not args of op */
1240 if(has_reg_class(si, remat_arg) && !pset_find_ptr(args, remat_arg)) {
1241 pset_insert_ptr(used, remat_arg);
1249 /* do not place post remats after jumps */
1250 if(sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) {
1256 /* insert all possible remats after irn */
1257 post_remats = pset_new_ptr_default();
1258 pset_foreach(used, arg) {
1259 remat_info_t *remat_info,
1262 /* continue if the operand has the wrong reg class */
1263 if(!has_reg_class(si, arg))
1267 query.remats = NULL;
1268 query.remats_by_operand = NULL;
1269 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1275 if(remat_info->remats_by_operand) {
1276 pset_foreach(remat_info->remats_by_operand, remat) {
1277 /* do not insert remats producing the same value as one of the operands */
1278 if(!pset_find_ptr(args, remat->value)) {
1279 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
1281 /* only remat values that can be used by real ops */
1282 if(!opt_remat_while_live || pset_find_ptr(live, remat->value)) {
1283 pset_insert_ptr(post_remats, remat);
1289 pset_foreach(post_remats, remat) {
1290 insert_remat_after(si, remat, irn, live);
1292 del_pset(post_remats);
1299 /* add remats at end if successor has multiple predecessors */
1300 if(is_merge_edge(bb)) {
1301 pset *live_out = pset_new_ptr_default();
1304 get_live_end(si, bb, live_out);
1306 /* add remats at end of block */
1307 pset_foreach(live_out, value) {
1308 remat_info_t *remat_info,
1312 query.remats = NULL;
1313 query.remats_by_operand = NULL;
1314 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1316 if(remat_info && remat_info->remats) {
1317 pset_foreach(remat_info->remats, remat) {
1318 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1320 insert_remat_before(si, remat, bb, live_out);
1327 if(is_diverge_edge(bb)) {
1328 pset *live_in = pset_new_ptr_default();
1331 be_lv_foreach(si->lv, bb, be_lv_state_in, i) {
1332 value = be_lv_get_irn(si->lv, bb, i);
1334 if(has_reg_class(si, value)) {
1335 pset_insert_ptr(live_in, value);
1338 /* add phis to live_in */
1339 sched_foreach(bb, value) {
1340 if(!is_Phi(value)) break;
1342 if(has_reg_class(si, value)) {
1343 pset_insert_ptr(live_in, value);
1347 /* add remat2s at beginning of block */
1348 post_remats = pset_new_ptr_default();
1349 pset_foreach(live_in, value) {
1350 remat_info_t *remat_info,
1354 query.remats = NULL;
1355 query.remats_by_operand = NULL;
1356 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1358 if(remat_info && remat_info->remats_by_operand) {
1359 pset_foreach(remat_info->remats_by_operand, remat) {
1360 DBG((si->dbg, LEVEL_4, "\t considering remat2 %+F at beginning of block %+F\n", remat->op, bb));
1362 /* put the remat here if all its args are available and result is still live */
1363 if(!opt_remat_while_live || pset_find_ptr(live_in, remat->value)) {
1364 pset_insert_ptr(post_remats, remat);
1369 pset_foreach(post_remats, remat) {
1370 insert_remat_after(si, remat, bb, live_in);
1372 del_pset(post_remats);
1378 can_be_copied(const ir_node * bb, const ir_node * irn)
1380 const ir_edge_t *edge = get_block_succ_first(bb);
1381 const ir_node *next_bb = edge->src;
1382 int pos = edge->pos;
1385 assert(is_merge_edge(bb));
1387 sched_foreach(next_bb, phi) {
1388 const ir_node *phi_arg;
1390 if(!is_Phi(phi)) break;
1392 phi_arg = get_irn_n(phi, pos);
1394 if(phi_arg == irn) {
1402 * Initialize additional node info
1405 luke_initializer(ir_node * bb, void * data)
1407 spill_ilp_t *si = (spill_ilp_t*)data;
1408 spill_bb_t *spill_bb;
1411 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1412 set_irn_link(bb, spill_bb);
1414 sched_foreach(bb, irn) {
1417 op = obstack_alloc(si->obst, sizeof(*op));
1419 op->attr.live_range.ilp = ILP_UNDEF;
1422 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1423 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1425 } else if(!is_Proj(irn)) {
1426 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1427 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1429 op->attr.live_range.args.reloads = NULL;
1431 set_irn_link(irn, op);
1437 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1440 luke_endwalker(ir_node * bb, void * data)
1442 spill_ilp_t *si = (spill_ilp_t*)data;
1448 spill_bb_t *spill_bb = get_irn_link(bb);
1451 live = pset_new_ptr_default();
1452 use_end = pset_new_ptr_default();
1454 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1455 irn = be_lv_get_irn(si->lv, bb, i);
1456 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1457 pset_insert_ptr(live, irn);
1461 * find values that are used by remats at end of block
1462 * and insert them into live set
1464 foreach_pre_remat(si, bb, irn) {
1467 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1468 ir_node *remat_arg = get_irn_n(irn, n);
1470 if(has_reg_class(si, remat_arg)) {
1471 pset_insert_ptr(live, remat_arg);
1476 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1477 /* their reg_out must always be set */
1478 sched_foreach_reverse(bb, irn) {
1481 if(!sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) break;
1483 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1484 ir_node *irn_arg = get_irn_n(irn, n);
1486 if(has_reg_class(si, irn_arg)) {
1487 pset_insert_ptr(use_end, irn_arg);
1492 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1493 //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1494 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1496 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1498 /* if this is a merge edge we can reload at the end of this block */
1499 if(is_merge_edge(bb)) {
1500 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1501 } else if(pset_count(use_end)){
1502 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1504 spill_bb->reloads = NULL;
1507 pset_foreach(live,irn) {
1511 int default_spilled;
1514 /* handle values used by control flow nodes later separately */
1515 if(pset_find_ptr(use_end, irn)) continue;
1518 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1520 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1522 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1523 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1524 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1526 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1527 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1529 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1530 /* by default spill value right after definition */
1531 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1532 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1534 if(is_merge_edge(bb)) {
1538 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1539 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), can_be_copied(bb, irn));
1540 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1542 /* reload <= mem_out */
1543 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1544 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1545 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1548 spill->reg_in = ILP_UNDEF;
1549 spill->mem_in = ILP_UNDEF;
1552 pset_foreach(use_end,irn) {
1556 ilp_cst_t end_use_req,
1559 int default_spilled;
1562 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1564 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1566 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1567 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1569 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1570 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1572 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1573 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1574 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1576 /* reload for use be control flow op */
1577 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1578 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1579 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1581 /* reload <= mem_out */
1582 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1583 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1584 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1586 spill->reg_in = ILP_UNDEF;
1587 spill->mem_in = ILP_UNDEF;
1589 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1590 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1591 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1599 * Find a remat of value @p value in the epilog of @p pos
1602 find_post_remat(const ir_node * value, const ir_node * pos)
1604 while((pos = next_post_remat(pos)) != NULL) {
1607 op = get_irn_link(pos);
1608 assert(op->is_remat && !op->attr.remat.pre);
1610 if(op->attr.remat.remat->value == value)
1611 return (ir_node*)pos;
1614 const ir_edge_t *edge;
1615 foreach_out_edge(pos, edge) {
1616 ir_node *proj = get_edge_src_irn(edge);
1617 assert(is_Proj(proj));
1627 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1629 spill_bb_t *spill_bb = get_irn_link(bb);
1633 int default_spilled;
1636 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1638 double spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1640 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1642 spill->reg_out = ILP_UNDEF;
1643 spill->reg_in = ILP_UNDEF;
1644 spill->mem_in = ILP_UNDEF;
1646 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1647 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1649 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1650 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1651 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1658 * Inserts ILP-constraints and variables for memory copying before the given position
1661 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1663 const ir_node *succ;
1664 const ir_edge_t *edge;
1665 spill_bb_t *spill_bb = get_irn_link(block);
1674 assert(edges_activated(current_ir_graph));
1676 edge = get_block_succ_first(block);
1682 edge = get_block_succ_next(block, edge);
1683 /* next block can only contain phis, if this is a merge edge */
1686 ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1687 copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1689 ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1690 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1692 pset_foreach(live, tmp) {
1695 op_t *op = get_irn_link(irn);
1696 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1698 spill = set_find_spill(spill_bb->ilp, tmp);
1701 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1703 lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1705 sched_foreach(succ, phi) {
1706 const ir_node *to_copy;
1708 spill_t *to_copy_spill;
1709 op_t *phi_op = get_irn_link(phi);
1710 ilp_var_t reload = ILP_UNDEF;
1713 if(!is_Phi(phi)) break;
1714 if(!has_reg_class(si, phi)) continue;
1716 to_copy = get_irn_n(phi, pos);
1717 to_copy_op = get_irn_link(to_copy);
1719 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1720 assert(to_copy_spill);
1722 if(spill_bb->reloads) {
1723 keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1726 reload = PTR_TO_INT(keyval->val);
1730 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1731 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1733 /* copy - reg_out - reload - remat - live_range <= 0 */
1734 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1735 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1736 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1737 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1738 foreach_pre_remat(si, block, tmp) {
1739 op_t *remat_op = get_irn_link(tmp);
1740 if(remat_op->attr.remat.remat->value == to_copy) {
1741 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1745 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1746 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1748 /* copy - reg_out - copyreg <= 0 */
1749 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1750 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1751 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1757 * Walk all irg blocks and emit this ILP
1760 luke_blockwalker(ir_node * bb, void * data)
1762 spill_ilp_t *si = (spill_ilp_t*)data;
1767 spill_bb_t *spill_bb = get_irn_link(bb);
1770 pset *defs = pset_new_ptr_default();
1771 const arch_env_t *arch_env = si->birg->main_env->arch_env;
1773 live = pset_new_ptr_default();
1775 /****************************************
1776 * B A S I C B L O C K E N D
1777 ***************************************/
1780 /* init live values at end of block */
1781 get_live_end(si, bb, live);
1783 pset_foreach(live, irn) {
1785 ilp_var_t reload = ILP_UNDEF;
1787 spill = set_find_spill(spill_bb->ilp, irn);
1790 if(spill_bb->reloads) {
1791 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1794 reload = PTR_TO_INT(keyval->val);
1798 op = get_irn_link(irn);
1799 assert(!op->is_remat);
1801 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1802 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1803 op->attr.live_range.op = bb;
1805 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1806 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1808 /* reg_out - reload - remat - live_range <= 0 */
1809 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1810 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1811 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1812 foreach_pre_remat(si, bb, tmp) {
1813 op_t *remat_op = get_irn_link(tmp);
1814 if(remat_op->attr.remat.remat->value == irn) {
1815 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1818 ir_snprintf(buf, sizeof(buf), "reg_out2_%N_%N", bb, irn);
1819 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
1821 /* value may only die at bb end if it is used for a mem copy */
1822 /* reg_out + \sum copy - reload - remat - live_range >= 0 */
1823 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1824 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1825 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1826 foreach_pre_remat(si, bb, tmp) {
1827 op_t *remat_op = get_irn_link(tmp);
1828 if(remat_op->attr.remat.remat->value == irn) {
1829 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1832 if(is_merge_edge(bb)) {
1833 const ir_edge_t *edge = get_block_succ_first(bb);
1834 const ir_node *next_bb = edge->src;
1835 int pos = edge->pos;
1838 sched_foreach(next_bb, phi) {
1839 const ir_node *phi_arg;
1841 if(!is_Phi(phi)) break;
1843 phi_arg = get_irn_n(phi, pos);
1845 if(phi_arg == irn) {
1846 op_t *phi_op = get_irn_link(phi);
1847 ilp_var_t copy = phi_op->attr.live_range.args.copies[pos];
1849 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
1856 insert_mem_copy_position(si, live, bb);
1859 * assure the remat args are available
1861 foreach_pre_remat(si, bb, tmp) {
1862 op_t *remat_op = get_irn_link(tmp);
1865 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1866 ir_node *remat_arg = get_irn_n(tmp, n);
1867 op_t *arg_op = get_irn_link(remat_arg);
1869 if(!has_reg_class(si, remat_arg)) continue;
1871 spill = set_find_spill(spill_bb->ilp, remat_arg);
1874 /* arguments of remats have to be live until the very end of the block
1875 * remat = reg_out(remat_arg) and (reload(remat_arg) or live_range(remat_arg)),
1876 * no remats, they could be in wrong order
1879 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1880 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1882 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 3.0);
1883 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, -2.0);
1884 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1886 /* use reload placed for this argument */
1887 if(spill_bb->reloads) {
1888 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1891 ilp_var_t reload = PTR_TO_INT(keyval->val);
1893 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1898 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1903 /**************************************
1904 * B A S I C B L O C K B O D Y
1905 **************************************/
1907 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1913 ilp_cst_t check_pre,
1919 ilp_cst_t one_memoperand;
1921 /* iterate only until first phi */
1925 op = get_irn_link(irn);
1927 if(op->is_remat) continue;
1929 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1931 /* collect defined values */
1932 if(has_reg_class(si, irn)) {
1933 pset_insert_ptr(defs, irn);
1937 if(is_Proj(irn)) continue;
1940 * init set of irn's arguments
1941 * and all possibly used values around this op
1942 * and values defined by post remats
1944 args = new_set(cmp_keyval, get_irn_arity(irn));
1945 used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1946 remat_defs = pset_new_ptr(pset_count(live));
1948 if(!is_start_block(bb) || !be_is_Barrier(irn)) {
1949 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1950 ir_node *irn_arg = get_irn_n(irn, n);
1951 if(has_reg_class(si, irn_arg)) {
1952 set_insert_keyval(args, irn_arg, (void*)n);
1953 pset_insert_ptr(used, irn_arg);
1956 foreach_post_remat(irn, tmp) {
1957 op_t *remat_op = get_irn_link(tmp);
1959 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1961 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1962 ir_node *remat_arg = get_irn_n(tmp, n);
1963 if(has_reg_class(si, remat_arg)) {
1964 pset_insert_ptr(used, remat_arg);
1968 foreach_pre_remat(si, irn, tmp) {
1969 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1970 ir_node *remat_arg = get_irn_n(tmp, n);
1971 if(has_reg_class(si, remat_arg)) {
1972 pset_insert_ptr(used, remat_arg);
1978 /**********************************
1979 * I N E P I L O G O F irn
1980 **********************************/
1982 /* ensure each dying value is used by only one post remat */
1983 pset_foreach(used, tmp) {
1984 ir_node *value = tmp;
1985 op_t *value_op = get_irn_link(value);
1990 foreach_post_remat(irn, remat) {
1991 op_t *remat_op = get_irn_link(remat);
1993 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1994 ir_node *remat_arg = get_irn_n(remat, n);
1996 /* if value is used by this remat add it to constraint */
1997 if(remat_arg == value) {
1999 /* sum remat2s <= 1 + n_remats*live_range */
2000 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
2001 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2005 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2011 if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
2012 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
2016 /* ensure at least one value dies at post remat */
2017 foreach_post_remat(irn, tmp) {
2018 op_t *remat_op = get_irn_link(tmp);
2019 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
2022 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2023 remat_arg = get_irn_n(tmp, n);
2025 if(has_reg_class(si, remat_arg)) {
2027 /* does arg always die at this op? */
2028 if(!pset_find_ptr(live, remat_arg))
2029 goto skip_one_must_die;
2031 pset_insert_ptr(remat_args, remat_arg);
2035 /* remat + \sum live_range(remat_arg) <= |args| */
2036 ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
2037 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2038 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2040 pset_foreach(remat_args, remat_arg) {
2041 op_t *arg_op = get_irn_link(remat_arg);
2043 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2047 del_pset(remat_args);
2050 /* new live ranges for values from L\U defined by post remats */
2051 pset_foreach(live, tmp) {
2052 ir_node *value = tmp;
2053 op_t *value_op = get_irn_link(value);
2055 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
2056 ilp_var_t prev_lr = ILP_UNDEF;
2059 if(pset_find_ptr(remat_defs, value)) {
2061 /* next_live_range <= prev_live_range + sum remat2s */
2062 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
2063 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2065 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
2066 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2068 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
2069 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2071 foreach_post_remat(irn, remat) {
2072 op_t *remat_op = get_irn_link(remat);
2074 /* if value is being rematerialized by this remat */
2075 if(value == remat_op->attr.remat.remat->value) {
2076 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2080 value_op->attr.live_range.ilp = prev_lr;
2081 value_op->attr.live_range.op = irn;
2086 /* requirements for post remats and start live ranges from L/U' for values dying here */
2087 foreach_post_remat(irn, tmp) {
2088 op_t *remat_op = get_irn_link(tmp);
2091 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2092 ir_node *remat_arg = get_irn_n(tmp, n);
2093 op_t *arg_op = get_irn_link(remat_arg);
2095 if(!has_reg_class(si, remat_arg)) continue;
2097 /* only for values in L\U (TODO and D?), the others are handled with post_use */
2098 if(!pset_find_ptr(used, remat_arg)) {
2099 /* remat <= live_range(remat_arg) */
2100 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
2101 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2103 /* if value is becoming live through use by remat2 */
2104 if(!pset_find_ptr(live, remat_arg)) {
2107 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
2108 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2110 arg_op->attr.live_range.ilp = lr;
2111 arg_op->attr.live_range.op = irn;
2113 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
2115 pset_insert_ptr(live, remat_arg);
2116 add_to_spill_bb(si, bb, remat_arg);
2119 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2120 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2125 d = pset_count(defs);
2126 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
2128 /* count how many regs irn needs for arguments */
2129 u = set_count(args);
2132 /* check the register pressure in the epilog */
2133 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
2134 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
2135 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
2137 /* add L\U' to check_post */
2138 pset_foreach(live, tmp) {
2139 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
2140 /* if a live value is not used by irn */
2141 tmp_op = get_irn_link(tmp);
2142 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
2146 /***********************************************************
2147 * 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
2148 **********************************************************/
2151 pset_foreach(used, tmp) {
2157 op_t *arg_op = get_irn_link(arg);
2160 spill = add_to_spill_bb(si, bb, arg);
2162 /* new live range for each used value */
2163 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2164 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2166 /* the epilog stuff - including post_use, check_post, check_post_remat */
2167 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2168 post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2170 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2172 /* arg is live throughout epilog if the next live_range is in a register */
2173 if(pset_find_ptr(live, arg)) {
2174 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
2176 /* post_use >= next_lr + remat */
2177 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2178 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2179 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2180 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2183 /* forall post remat which use arg add a similar cst */
2184 foreach_post_remat(irn, remat) {
2187 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2188 ir_node *remat_arg = get_irn_n(remat, n);
2189 op_t *remat_op = get_irn_link(remat);
2191 if(remat_arg == arg) {
2192 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
2194 /* post_use >= remat */
2195 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2196 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2197 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2198 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2203 /* if value is not an arg of op and not possibly defined by post remat
2204 * then it may only die and not become live
2206 if(!set_find_keyval(args, arg)) {
2207 /* post_use <= prev_lr */
2208 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2209 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2210 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2211 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2213 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2214 /* next_lr <= prev_lr */
2215 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2216 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2217 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2218 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2222 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2223 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2224 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2225 ilp_var_t memoperand;
2227 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2228 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0);
2229 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2231 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2232 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2234 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2235 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2240 /* new live range begins for each used value */
2241 arg_op->attr.live_range.ilp = prev_lr;
2242 arg_op->attr.live_range.op = irn;
2244 pset_insert_ptr(live, arg);
2247 /* just to be sure */
2248 check_post = ILP_UNDEF;
2250 /* allow original defintions to be removed */
2251 if(opt_repair_schedule) {
2252 pset_foreach(defs, tmp) {
2253 op_t *tmp_op = get_irn_link(tmp);
2254 spill_t *spill = set_find_spill(spill_bb->ilp, tmp);
2259 ir_snprintf(buf, sizeof(buf), "delete_%N", tmp);
2260 delete = lpp_add_var_default(si->lpp, buf, lpp_binary, -1.0*get_cost(si, irn)*execution_frequency(si, bb), 0.0);
2262 /* op may not be killed if its first live_range is 1 */
2263 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2264 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2265 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2266 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2268 /* op may not be killed if it is spilled after the definition */
2269 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2270 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2271 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2272 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2277 ir_snprintf(buf, sizeof(buf), "keep_%N", tmp);
2278 keep = lpp_add_var_default(si->lpp, buf, lpp_binary, get_cost(si, irn)*execution_frequency(si, bb), 1.0);
2280 /* op may not be killed if its first live_range is 1 */
2281 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2282 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2283 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2284 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, -1.0);
2286 /* op may not be killed if it is spilled after the definition */
2287 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2288 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2289 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2290 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2295 pset_foreach(defs, tmp) {
2296 op_t *tmp_op = get_irn_link(tmp);
2297 spill_t *spill = set_find_spill(spill_bb->ilp, tmp);
2300 /* live_range or spill should be 1
2301 TODO: lr should be live until first use */
2302 ir_snprintf(buf, sizeof(buf), "nokillorig_%N", tmp);
2303 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 1.0);
2304 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2305 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2315 /* check the register pressure in the prolog */
2316 /* sum_{L\U} lr <= k - |U| */
2317 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2318 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2320 /* for the prolog remove defined values from the live set */
2321 pset_foreach(defs, tmp) {
2322 pset_remove_ptr(live, tmp);
2325 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2326 ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2327 one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2330 /***********************************************************
2331 * 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
2332 **********************************************************/
2335 set_foreach(args, keyval) {
2337 const ir_node *arg = keyval->key;
2338 int i = PTR_TO_INT(keyval->val);
2339 op_t *arg_op = get_irn_link(arg);
2340 ilp_cst_t requirements;
2343 spill = set_find_spill(spill_bb->ilp, arg);
2346 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2347 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);
2349 /* reload <= mem_out */
2350 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2351 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2352 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2353 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2355 /* requirement: arg must be in register for use */
2356 /* reload + remat + live_range == 1 */
2357 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2358 requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2360 lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2361 lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2362 foreach_pre_remat(si, irn, tmp) {
2363 op_t *remat_op = get_irn_link(tmp);
2364 if(remat_op->attr.remat.remat->value == arg) {
2365 lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2369 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2371 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2372 if(get_irn_n(irn, n) == arg) {
2376 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2377 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2378 memoperand_t *memoperand;
2379 memoperand = set_find_memoperand(si->memoperands, irn, n);
2381 /* memoperand <= mem_out */
2382 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2383 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2384 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2385 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2387 /* the memoperand is only sufficient if it is used once by the op */
2388 if(n_memoperands == 1)
2389 lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2391 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2393 /* we have one more free register if we use a memory operand */
2394 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2400 /* iterate over L\U */
2401 pset_foreach(live, tmp) {
2402 if(!set_find_keyval(args, tmp)) {
2403 /* if a live value is not used by irn */
2404 tmp_op = get_irn_link(tmp);
2405 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2409 /* requirements for remats */
2410 foreach_pre_remat(si, irn, tmp) {
2411 op_t *remat_op = get_irn_link(tmp);
2414 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2415 ir_node *remat_arg = get_irn_n(tmp, n);
2416 op_t *arg_op = get_irn_link(remat_arg);
2418 if(!has_reg_class(si, remat_arg)) continue;
2420 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2421 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2422 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2424 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2425 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2427 /* if remat arg is also used by current op then we can use reload placed for this argument */
2428 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2429 int index = (int)keyval->val;
2431 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2439 /*************************
2440 * D O N E W I T H O P
2441 *************************/
2443 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2445 pset_foreach(live, tmp) {
2446 assert(has_reg_class(si, tmp));
2449 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2450 ir_node *arg = get_irn_n(irn, n);
2452 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2455 del_pset(remat_defs);
2459 defs = pset_new_ptr_default();
2461 /* skip everything above barrier in start block */
2462 if(is_start_block(bb) && be_is_Barrier(irn)) {
2463 assert(pset_count(live) == 0);
2472 /***************************************
2473 * B E G I N N I N G O F B L O C K
2474 ***************************************/
2477 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2478 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2480 pset_foreach(live, irn) {
2481 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2484 /* construct mem_outs for all values */
2485 set_foreach(spill_bb->ilp, spill) {
2486 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2487 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2489 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2490 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2492 if(pset_find_ptr(live, spill->irn)) {
2493 int default_spilled;
2494 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2496 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2497 default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2498 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2499 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2501 if(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2503 op_t *op = get_irn_link(spill->irn);
2505 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2506 const ir_node *arg = get_irn_n(spill->irn, n);
2512 /* argument already done? */
2513 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2515 /* get sum of execution frequencies of blocks with the same phi argument */
2516 for(m=n; m>=0; --m) {
2517 const ir_node *arg2 = get_irn_n(spill->irn, m);
2520 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2524 /* copies are not for free */
2525 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2526 var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
2528 for(m=n; m>=0; --m) {
2529 const ir_node *arg2 = get_irn_n(spill->irn, m);
2532 op->attr.live_range.args.copies[m] = var;
2537 /* copy <= mem_in */
2538 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2539 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2540 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2541 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2548 foreach_post_remat(bb, tmp) {
2550 op_t *remat_op = get_irn_link(tmp);
2551 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
2554 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2555 remat_arg = get_irn_n(tmp, n);
2557 if(has_reg_class(si, remat_arg)) {
2558 pset_insert_ptr(remat_args, remat_arg);
2562 /* remat + \sum live_range(remat_arg) <= |args| */
2563 ir_snprintf(buf, sizeof(buf), "one_must_die_%N", tmp);
2564 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2565 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2567 pset_foreach(remat_args, remat_arg) {
2568 if(pset_find_ptr(live, remat_arg)) {
2569 op_t *remat_arg_op = get_irn_link(remat_arg);
2570 lpp_set_factor_fast(si->lpp, cst, remat_arg_op->attr.live_range.ilp, 1.0);
2573 del_pset(remat_args);
2576 foreach_post_remat(bb, tmp) {
2579 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2580 ir_node *remat_arg = get_irn_n(tmp, n);
2582 /* if value is becoming live through use by remat2 */
2583 if(has_reg_class(si, remat_arg) && !pset_find_ptr(live, remat_arg)) {
2584 op_t *remat_arg_op = get_irn_link(remat_arg);
2587 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 at bb start %+F\n", remat_arg, tmp));
2589 pset_insert_ptr(live, remat_arg);
2590 spill = add_to_spill_bb(si, bb, remat_arg);
2591 remat_arg_op->attr.live_range.ilp = ILP_UNDEF;
2593 /* we need reg_in and mem_in for this value; they will be referenced later */
2594 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", remat_arg, bb);
2595 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2596 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", remat_arg, bb);
2597 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2600 /* optimization: all memory stuff should be 0, for we do not want to insert reloads for remats */
2601 ir_snprintf(buf, sizeof(buf), "nomem_%N_%N", remat_arg, bb);
2602 nomem = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0.0);
2603 lpp_set_factor_fast(si->lpp, nomem, spill->spill, 1.0);
2608 /* L\U is empty at bb start */
2609 /* arg is live throughout epilog if it is reg_in into this block */
2611 /* check the register pressure at the beginning of the block
2614 /* reg_in entspricht post_use */
2616 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2617 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2619 pset_foreach(live, irn) {
2622 spill = set_find_spill(spill_bb->ilp, irn);
2625 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2626 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2628 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2630 /* spill + mem_in <= 1 */
2631 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2632 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2634 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2635 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2637 } /* post_remats are NOT included in register pressure check because
2638 they do not increase regpressure */
2640 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2641 pset_foreach(live, irn) {
2645 spill = set_find_spill(spill_bb->ilp, irn);
2646 assert(spill && spill->irn == irn);
2648 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2649 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2652 ir_node *phi_arg = get_Phi_pred(irn, n);
2653 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2654 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2656 op_t *op = get_irn_link(irn);
2658 /* although the phi is in the right regclass one or more of
2659 * its arguments can be in a different one or at least to
2662 if(has_reg_class(si, phi_arg)) {
2663 /* mem_in < mem_out_arg + copy */
2664 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2665 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2667 /* reg_in < reg_out_arg */
2668 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2669 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2671 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2672 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2674 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2677 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2679 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2681 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2685 /* else assure the value arrives on all paths in the same resource */
2687 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2690 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2691 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2694 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2695 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2696 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2697 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2699 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2700 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2702 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2705 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2706 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2711 foreach_post_remat(bb, tmp) {
2714 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2715 ir_node *remat_arg = get_irn_n(tmp, n);
2716 op_t *remat_op = get_irn_link(tmp);
2718 if(!has_reg_class(si, remat_arg)) continue;
2720 spill = set_find_spill(spill_bb->ilp, remat_arg);
2723 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2724 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2725 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2726 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2730 pset_foreach(live, irn) {
2731 const op_t *op = get_irn_link(irn);
2732 const ir_node *remat;
2737 foreach_post_remat(bb, remat) {
2740 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2741 const ir_node *arg = get_irn_n(remat, n);
2744 const op_t *remat_op = get_irn_link(remat);
2746 if(cst == ILP_UNDEF) {
2747 /* sum remat2s <= 1 + n_remats*live_range */
2748 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", irn, bb);
2749 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2751 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2757 if(cst != ILP_UNDEF && op->attr.live_range.ilp != ILP_UNDEF) {
2758 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -n_remats);
2762 /* first live ranges from reg_ins */
2763 pset_foreach(live, irn) {
2764 op_t *op = get_irn_link(irn);
2766 if(op->attr.live_range.ilp != ILP_UNDEF) {
2768 spill = set_find_spill(spill_bb->ilp, irn);
2769 assert(spill && spill->irn == irn);
2771 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2772 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2773 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2774 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2776 foreach_post_remat(bb, tmp) {
2777 op_t *remat_op = get_irn_link(tmp);
2779 if(remat_op->attr.remat.remat->value == irn) {
2780 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2786 /* walk forward now and compute constraints for placing spills */
2787 /* this must only be done for values that are not defined in this block */
2788 pset_foreach(live, irn) {
2790 * if value is defined in this block we can anways place the spill directly after the def
2791 * -> no constraint necessary
2793 if(!is_Phi(irn) && get_nodes_block(irn) == bb) {
2798 spill = set_find_spill(spill_bb->ilp, irn);
2801 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2802 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2804 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2805 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2808 sched_foreach_op(bb, tmp) {
2809 op_t *op = get_irn_link(tmp);
2811 if(is_Phi(tmp)) continue;
2812 assert(!is_Proj(tmp));
2815 const ir_node *value = op->attr.remat.remat->value;
2818 /* only collect remats up to the first real use of a value */
2819 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2824 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2825 ir_node *arg = get_irn_n(tmp, n);
2828 /* if a value is used stop collecting remats */
2841 typedef struct _irnlist_t {
2842 struct list_head list;
2846 typedef struct _interference_t {
2847 struct list_head blocklist;
2853 cmp_interference(const void *a, const void *b, size_t size)
2855 const interference_t *p = a;
2856 const interference_t *q = b;
2858 return !(p->a == q->a && p->b == q->b);
2861 static interference_t *
2862 set_find_interference(set * set, ir_node * a, ir_node * b)
2864 interference_t query;
2866 query.a = (a>b)?a:b;
2867 query.b = (a>b)?b:a;
2869 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2872 static interference_t *
2873 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2875 interference_t query,
2877 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2881 result = set_find_interference(set, a, b);
2884 list_add(&list->list, &result->blocklist);
2888 query.a = (a>b)?a:b;
2889 query.b = (a>b)?b:a;
2891 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2893 INIT_LIST_HEAD(&result->blocklist);
2894 list_add(&list->list, &result->blocklist);
2900 values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
2902 const ir_edge_t *edge;
2904 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2905 /* both values are live in, so they interfere */
2909 /* ensure a dominates b */
2910 if(value_dominates(b,a)) {
2916 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2919 /* the following code is stolen from bera.c */
2920 if(be_is_live_end(si->lv, bb, a))
2923 foreach_out_edge(a, edge) {
2924 const ir_node *user = edge->src;
2925 if(get_nodes_block(user) == bb
2928 && !pset_find_ptr(si->inverse_ops, user)
2929 && value_dominates(b, user))
2937 * Walk all irg blocks and collect interfering values inside of phi classes
2940 luke_interferencewalker(ir_node * bb, void * data)
2942 spill_ilp_t *si = (spill_ilp_t*)data;
2945 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2946 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
2947 op_t *a_op = get_irn_link(a);
2950 /* a is only interesting if it is in my register class and if it is inside a phi class */
2951 if (has_reg_class(si, a) && get_phi_class(a)) {
2952 if(a_op->is_remat || pset_find_ptr(si->inverse_ops, a))
2955 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)) {
2956 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
2957 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(a) == get_phi_class(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 /* analyze phi classes */
3144 phi_class_compute(si->birg->irg);
3146 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
3147 irg_block_walk_graph(si->birg->irg, luke_interferencewalker, NULL, si);
3149 /* now lets emit the ILP unequations for the crap */
3150 set_foreach(si->interferences, interference) {
3152 ilp_var_t interfere,
3154 ilp_cst_t any_interfere_cst,
3156 const ir_node *a = interference->a;
3157 const ir_node *b = interference->b;
3159 /* any_interf <= \sum interf */
3160 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
3161 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3162 any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3164 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
3166 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
3167 const ir_node *bb = irnlist->irn;
3168 spill_bb_t *spill_bb = get_irn_link(bb);
3173 spilla = set_find_spill(spill_bb->ilp, a);
3176 spillb = set_find_spill(spill_bb->ilp, b);
3179 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
3180 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
3181 /* 2: - mem_in_a - spill_a + interfere <= 0 */
3182 /* 3: - mem_in_b - spill_b + interfere <= 0 */
3183 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
3184 interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3186 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
3187 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
3189 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
3190 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
3191 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
3192 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
3193 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
3195 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
3196 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3198 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3199 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
3200 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
3202 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
3203 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3205 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3206 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
3207 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
3210 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
3212 /* any_interfere >= interf */
3213 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
3214 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3216 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3217 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3220 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
3221 gen_copy_constraints(si,a,b,any_interfere);
3229 return fabs(x) < 0.00001;
3232 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
3234 spill_ilp_t *si = get_irg_link(current_ir_graph);
3236 if(pset_find_ptr(si->all_possible_remats, n)) {
3237 op_t *op = (op_t*)get_irn_link(n);
3238 assert(op && op->is_remat);
3240 if(!op->attr.remat.remat->inverse) {
3241 if(op->attr.remat.pre) {
3242 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
3244 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
3249 op_t *op = (op_t*)get_irn_link(n);
3250 assert(op && op->is_remat);
3252 if(op->attr.remat.pre) {
3253 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
3255 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
3266 dump_graph_with_remats(ir_graph * irg, const char * suffix)
3268 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
3269 be_dump(irg, suffix, dump_ir_block_graph_sched);
3270 set_dump_node_vcgattr_hook(NULL);
3274 * Edge hook to dump the schedule edges with annotated register pressure.
3277 sched_pressure_edge_hook(FILE *F, ir_node *irn)
3279 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
3280 ir_node *prev = sched_prev(irn);
3281 fprintf(F, "edge:{sourcename:\"");
3283 fprintf(F, "\" targetname:\"");
3285 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
3286 fprintf(F, "\" color:magenta}\n");
3292 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
3294 DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3296 dump_consts_local(0);
3297 set_dump_node_edge_hook(sched_pressure_edge_hook);
3298 dump_ir_block_graph(irg, suffix);
3299 set_dump_node_edge_hook(old_edge_hook);
3303 walker_pressure_annotator(ir_node * bb, void * data)
3305 spill_ilp_t *si = data;
3308 pset *live = pset_new_ptr_default();
3311 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3312 irn = be_lv_get_irn(si->lv, bb, i);
3314 if (has_reg_class(si, irn)) {
3315 pset_insert_ptr(live, irn);
3319 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3321 sched_foreach_reverse(bb, irn) {
3323 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3327 if(has_reg_class(si, irn)) {
3328 pset_remove_ptr(live, irn);
3329 if(is_Proj(irn)) ++projs;
3332 if(!is_Proj(irn)) projs = 0;
3334 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3335 ir_node *arg = get_irn_n(irn, n);
3337 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3339 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3346 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3348 be_dump(si->birg->irg, suffix, dump_ir_block_graph_sched_pressure);
3352 connect_all_remats_with_keep(spill_ilp_t * si)
3360 n_remats = pset_count(si->all_possible_remats);
3362 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3365 pset_foreach(si->all_possible_remats, irn) {
3370 si->keep = be_new_Keep(si->cls, si->birg->irg, get_irg_end_block(si->birg->irg), n_remats, ins);
3372 obstack_free(si->obst, ins);
3377 connect_all_spills_with_keep(spill_ilp_t * si)
3386 n_spills = pset_count(si->spills);
3388 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3391 pset_foreach(si->spills, irn) {
3396 keep = be_new_Keep(si->cls, si->birg->irg, get_irg_end_block(si->birg->irg), n_spills, ins);
3398 obstack_free(si->obst, ins);
3402 /** insert a spill at an arbitrary position */
3403 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3405 ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
3406 ir_graph *irg = get_irn_irg(bl);
3407 ir_node *frame = get_irg_frame(irg);
3410 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
3412 spill = be_new_Spill(cls, irg, bl, irn);
3415 * search the right insertion point. a spill of a phi cannot be put
3416 * directly after the phi, if there are some phis behind the one which
3417 * is spilled. Also, a spill of a Proj must be after all Projs of the
3420 * Here's one special case:
3421 * If the spill is in the start block, the spill must be after the frame
3422 * pointer is set up. This is done by setting insert to the end of the block
3423 * which is its default initialization (see above).
3426 if (bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3429 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3432 sched_add_after(insert, spill);
3437 delete_remat(spill_ilp_t * si, ir_node * remat) {
3439 ir_node *bad = get_irg_bad(si->birg->irg);
3441 sched_remove(remat);
3443 /* kill links to operands */
3444 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3445 set_irn_n(remat, n, bad);
3450 clean_remat_info(spill_ilp_t * si)
3454 remat_info_t *remat_info;
3455 ir_node *bad = get_irg_bad(si->birg->irg);
3457 set_foreach(si->remat_info, remat_info) {
3458 if(!remat_info->remats) continue;
3460 pset_foreach(remat_info->remats, remat)
3462 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3463 if(sched_is_scheduled(remat->proj)) {
3464 sched_remove((ir_node*)remat->proj);
3466 set_irn_n((ir_node*)remat->proj, -1, bad);
3467 set_irn_n((ir_node*)remat->proj, 0, bad);
3470 if(get_irn_n_edges(remat->op) == 0) {
3471 if(sched_is_scheduled(remat->op)) {
3472 sched_remove((ir_node*)remat->op);
3474 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3475 set_irn_n((ir_node*)remat->op, n, bad);
3480 if(remat_info->remats) del_pset(remat_info->remats);
3481 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3486 delete_unnecessary_remats(spill_ilp_t * si)
3488 if(opt_keep_alive & KEEPALIVE_REMATS) {
3490 ir_node *bad = get_irg_bad(si->birg->irg);
3493 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3494 ir_node *keep_arg = get_irn_n(si->keep, n);
3495 op_t *arg_op = get_irn_link(keep_arg);
3498 assert(arg_op->is_remat);
3500 name = si->lpp->vars[arg_op->attr.remat.ilp];
3502 if(is_zero(name->value)) {
3503 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
3504 /* TODO check whether reload is preferred over remat (could be bug) */
3505 delete_remat(si, keep_arg);
3507 if(!arg_op->attr.remat.remat->inverse) {
3508 if(arg_op->attr.remat.pre) {
3509 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3511 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3514 if(arg_op->attr.remat.pre) {
3515 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3517 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3522 set_irn_n(si->keep, n, bad);
3525 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
3530 pset_foreach(si->all_possible_remats, remat) {
3531 op_t *remat_op = get_irn_link(remat);
3532 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
3534 if(is_zero(name->value)) {
3535 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
3536 /* TODO check whether reload is preferred over remat (could be bug) */
3537 delete_remat(si, remat);
3539 if(!remat_op->attr.remat.remat->inverse) {
3540 if(remat_op->attr.remat.pre) {
3541 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3543 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3546 if(remat_op->attr.remat.pre) {
3547 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3549 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3558 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3560 pset *spills = pset_new_ptr_default();
3562 const ir_node *next;
3565 defs = set_find_def(si->values, value);
3567 if(defs && defs->spills) {
3568 for(next = defs->spills; next; next = get_irn_link(next)) {
3569 pset_insert_ptr(spills, next);
3577 new_r_PhiM_nokeep(ir_graph * irg, ir_node *block, int arity, ir_node **in)
3581 assert( get_irn_arity(block) == arity );
3583 res = new_ir_node(NULL, irg, block, op_Phi, mode_M, arity, in);
3584 res->attr.phi_backedge = new_backedge_arr(irg->obst, arity);
3590 * @param before The node after which the spill will be placed in the schedule
3593 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3597 const arch_env_t *arch_env = si->birg->main_env->arch_env;
3599 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3601 spill = be_spill2(arch_env, irn, before);
3603 defs = set_insert_def(si->values, value);
3606 /* enter into the linked list */
3607 set_irn_link(spill, defs->spills);
3608 defs->spills = spill;
3610 if(opt_keep_alive & KEEPALIVE_SPILLS)
3611 pset_insert_ptr(si->spills, spill);
3617 * @param before The Phi node which has to be spilled
3620 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3627 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3629 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3630 ins[n] = si->m_unknown;
3633 mem_phi = new_r_PhiM_nokeep(si->birg->irg, get_nodes_block(phi), get_irn_arity(phi), ins);
3635 defs = set_insert_def(si->values, phi);
3638 /* enter into the linked list */
3639 set_irn_link(mem_phi, defs->spills);
3640 defs->spills = mem_phi;
3642 #ifdef SCHEDULE_PHIM
3643 sched_add_after(phi, mem_phi);
3645 pset_insert_ptr(si->phims, mem_phi);
3648 if(opt_keep_alive & KEEPALIVE_SPILLS)
3649 pset_insert_ptr(si->spills, mem_phi);
3656 * Add remat to list of defs, destroys link field!
3659 insert_remat(spill_ilp_t * si, ir_node * remat)
3662 op_t *remat_op = get_irn_link(remat);
3664 assert(remat_op->is_remat);
3666 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3669 /* enter into the linked list */
3670 set_irn_link(remat, defs->remats);
3671 defs->remats = remat;
3676 * Add reload before operation and add to list of defs
3679 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3684 const arch_env_t *arch_env = si->birg->main_env->arch_env;
3686 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3688 defs = set_find_def(si->values, value);
3690 spill = defs->spills;
3691 assert(spill && "no spill placed before reload");
3693 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3695 /* enter into the linked list */
3696 set_irn_link(reload, defs->remats);
3697 defs->remats = reload;
3702 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3705 ir_node *value = get_irn_n(memoperand->irn, memoperand->pos);
3707 const arch_env_t *arch_env = si->birg->main_env->arch_env;
3709 DBG((si->dbg, LEVEL_2, "\t inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3711 defs = set_find_def(si->values, value);
3713 spill = defs->spills;
3714 assert(spill && "no spill placed before reload");
3716 arch_perform_memory_operand(arch_env, memoperand->irn, spill, memoperand->pos);
3719 void insert_memoperands(spill_ilp_t * si)
3721 memoperand_t *memoperand;
3724 set_foreach(si->memoperands, memoperand) {
3725 name = si->lpp->vars[memoperand->ilp];
3726 if(!is_zero(name->value)) {
3727 perform_memory_operand(si, memoperand);
3733 walker_spill_placer(ir_node * bb, void * data) {
3734 spill_ilp_t *si = (spill_ilp_t*)data;
3736 spill_bb_t *spill_bb = get_irn_link(bb);
3737 pset *spills_to_do = pset_new_ptr_default();
3740 set_foreach(spill_bb->ilp, spill) {
3743 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3744 name = si->lpp->vars[spill->mem_in];
3745 if(!is_zero(name->value)) {
3748 mem_phi = insert_mem_phi(si, spill->irn);
3750 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3754 name = si->lpp->vars[spill->spill];
3755 if(!is_zero(name->value)) {
3756 /* place spill directly after definition */
3757 if(get_nodes_block(spill->irn) == bb) {
3758 insert_spill(si, spill->irn, spill->irn, spill->irn);
3762 /* place spill at bb start */
3763 if(spill->reg_in > 0) {
3764 name = si->lpp->vars[spill->reg_in];
3765 if(!is_zero(name->value)) {
3766 insert_spill(si, spill->irn, spill->irn, bb);
3770 /* place spill after a remat */
3771 pset_insert_ptr(spills_to_do, spill->irn);
3774 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3777 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3778 op_t *op = get_irn_link(irn);
3780 if(be_is_Spill(irn)) continue;
3783 /* TODO fix this if we want to support remats with more than two nodes */
3784 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3785 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3787 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3790 if(pset_find_ptr(spills_to_do, irn)) {
3791 pset_remove_ptr(spills_to_do, irn);
3793 insert_spill(si, irn, irn, irn);
3799 assert(pset_count(spills_to_do) == 0);
3801 /* afterwards free data in block */
3802 del_pset(spills_to_do);
3806 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3808 ir_node *insert_pos = bb;
3810 const arch_env_t *arch_env = si->birg->main_env->arch_env;
3812 /* find last definition of arg value in block */
3817 defs = set_find_def(si->values, value);
3819 if(defs && defs->remats) {
3820 for(next = defs->remats; next; next = get_irn_link(next)) {
3821 if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3822 last = sched_get_time_step(next);
3828 if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3829 last = sched_get_time_step(value);
3833 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
3835 spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3841 phim_fixer(spill_ilp_t *si) {
3844 set_foreach(si->values, defs) {
3845 const ir_node *phi = defs->value;
3846 op_t *op = get_irn_link(phi);
3847 ir_node *phi_m = NULL;
3848 ir_node *next = defs->spills;
3851 if(!is_Phi(phi)) continue;
3854 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3858 next = get_irn_link(next);
3861 if(!phi_m) continue;
3863 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3864 ir_node *value = get_irn_n(phi, n);
3865 defs_t *val_defs = set_find_def(si->values, value);
3867 /* a spill of this value */
3872 ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3873 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3875 if(!is_zero(name->value)) {
3876 spill = insert_mem_copy(si, pred, value);
3878 spill = val_defs->spills;
3881 spill = val_defs->spills;
3884 assert(spill && "no spill placed before PhiM");
3885 set_irn_n(phi_m, n, spill);
3891 walker_reload_placer(ir_node * bb, void * data) {
3892 spill_ilp_t *si = (spill_ilp_t*)data;
3894 spill_bb_t *spill_bb = get_irn_link(bb);
3896 /* reloads at end of block */
3897 if(spill_bb->reloads) {
3900 set_foreach(spill_bb->reloads, keyval) {
3901 ir_node *irn = (ir_node*)keyval->key;
3902 ilp_var_t reload = PTR_TO_INT(keyval->val);
3905 name = si->lpp->vars[reload];
3906 if(!is_zero(name->value)) {
3908 ir_node *insert_pos = bb;
3909 ir_node *prev = sched_block_last_noncf(si, bb);
3910 op_t *prev_op = get_irn_link(prev);
3912 while(be_is_Spill(prev)) {
3913 prev = sched_prev(prev);
3916 prev_op = get_irn_link(prev);
3918 /* insert reload before pre-remats */
3919 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3920 && prev_op->is_remat && prev_op->attr.remat.pre) {
3924 prev = sched_prev(prev);
3925 } while(be_is_Spill(prev));
3927 prev_op = get_irn_link(prev);
3931 reload = insert_reload(si, irn, insert_pos);
3933 if(opt_keep_alive & KEEPALIVE_RELOADS)
3934 pset_insert_ptr(si->spills, reload);
3939 /* walk and insert more reloads and collect remats */
3940 sched_foreach_reverse(bb, irn) {
3941 op_t *op = get_irn_link(irn);
3943 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3944 if(is_Phi(irn)) break;
3947 if(get_irn_mode(irn) != mode_T) {
3948 insert_remat(si, irn);
3953 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3954 ir_node *arg = get_irn_n(irn, n);
3956 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3959 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3960 if(!is_zero(name->value)) {
3962 ir_node *insert_pos = irn;
3963 ir_node *prev = sched_prev(insert_pos);
3966 while(be_is_Spill(prev)) {
3967 prev = sched_prev(prev);
3970 prev_op = get_irn_link(prev);
3972 /* insert reload before pre-remats */
3973 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3974 && prev_op->is_remat && prev_op->attr.remat.pre) {
3978 prev = sched_prev(prev);
3979 } while(be_is_Spill(prev));
3981 prev_op = get_irn_link(prev);
3985 reload = insert_reload(si, arg, insert_pos);
3987 assert(reload && "no reload returned");
3988 set_irn_n(irn, n, reload);
3990 if(opt_keep_alive & KEEPALIVE_RELOADS)
3991 pset_insert_ptr(si->spills, reload);
3998 del_set(spill_bb->ilp);
3999 if(spill_bb->reloads) del_set(spill_bb->reloads);
4003 walker_collect_used(ir_node * irn, void * data)
4005 bitset_t *used = data;
4007 bitset_set(used, get_irn_idx(irn));
4010 struct kill_helper {
4016 walker_kill_unused(ir_node * bb, void * data)
4018 struct kill_helper *kh = data;
4019 ir_node *bad = get_irg_bad(get_irn_irg(bb));
4023 for(irn=sched_first(bb); !sched_is_end(irn);) {
4024 ir_node *next = sched_next(irn);
4027 if(!bitset_is_set(kh->used, get_irn_idx(irn))) {
4028 if(be_is_Spill(irn) || be_is_Reload(irn)) {
4029 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)));
4031 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
4037 set_nodes_block(irn, bad);
4038 for (n=get_irn_arity(irn)-1; n>=0; --n) {
4039 set_irn_n(irn, n, bad);
4046 #ifndef SCHEDULE_PHIM
4048 kill_unused_phims(spill_ilp_t * si, struct kill_helper * kh)
4051 ir_node *bad = get_irg_bad(si->birg->irg);
4054 pset_foreach(si->phims, phi) {
4055 if(!bitset_is_set(kh->used, get_irn_idx(phi))) {
4057 set_nodes_block(phi, bad);
4058 for (n=get_irn_arity(phi)-1; n>=0; --n) {
4059 set_irn_n(phi, n, bad);
4067 kill_all_unused_values_in_schedule(spill_ilp_t * si)
4069 struct kill_helper kh;
4071 kh.used = bitset_malloc(get_irg_last_idx(si->birg->irg));
4074 irg_walk_graph(si->birg->irg, walker_collect_used, NULL, kh.used);
4075 #ifndef SCHEDULE_PHIM
4076 kill_unused_phims(si, &kh);
4078 irg_block_walk_graph(si->birg->irg, walker_kill_unused, NULL, &kh);
4080 bitset_free(kh.used);
4084 print_irn_pset(pset * p)
4088 pset_foreach(p, irn) {
4089 ir_printf("%+F\n", irn);
4094 dump_phi_class(spill_ilp_t * si, pset * phiclass, const char * file)
4096 FILE *f = fopen(file, "w");
4098 interference_t *interference;
4100 pset_break(phiclass);
4101 set_break(si->interferences);
4103 ir_fprintf(f, "digraph phiclass {\n");
4105 pset_foreach(phiclass, irn) {
4107 ir_fprintf(f, " %F%N [shape=box]\n",irn,irn);
4110 pset_foreach(phiclass, irn) {
4113 if(!is_Phi(irn)) continue;
4115 for(n=get_irn_arity(irn)-1; n>=0; --n) {
4116 ir_node *arg = get_irn_n(irn, n);
4118 ir_fprintf(f, " %F%N -> %F%N\n",irn,irn,arg,arg);
4122 set_foreach(si->interferences, interference) {
4123 const ir_node *a = interference->a;
4124 const ir_node *b = interference->b;
4125 if(get_phi_class(a) == phiclass) {
4126 ir_fprintf(f, " %F%N -> %F%N [color=red,dir=none,style=bold]\n",a,a,b,b);
4135 rewire_uses(spill_ilp_t * si)
4138 pset *ignore = pset_new_ptr(1);
4139 be_dom_front_info_t *dom_front = si->birg->dom_front;
4141 pset_insert_ptr(ignore, get_irg_end(si->birg->irg));
4143 /* then fix uses of spills */
4144 set_foreach(si->values, defs) {
4147 const ir_node *next = defs->remats;
4150 reloads = pset_new_ptr_default();
4153 if(be_is_Reload(next)) {
4154 pset_insert_ptr(reloads, next);
4158 next = get_irn_link(next);
4161 spills = get_spills_for_value(si, defs->value);
4162 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));
4163 if(pset_count(spills) > 1) {
4164 //assert(pset_count(reloads) > 0);
4165 // print_irn_pset(spills);
4166 // print_irn_pset(reloads);
4168 be_ssa_constr_set_ignore(dom_front, si->lv, spills, ignore);
4175 /* first fix uses of remats and reloads */
4176 set_foreach(si->values, defs) {
4178 const ir_node *next = defs->remats;
4182 nodes = pset_new_ptr_default();
4183 if(sched_is_scheduled(defs->value)) {
4184 pset_insert_ptr(nodes, defs->value);
4189 pset_insert_ptr(nodes, next);
4190 next = get_irn_link(next);
4193 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-orig_kept, defs->value));
4194 be_ssa_constr_set(dom_front, si->lv, nodes);
4200 // remove_unused_defs(si);
4205 writeback_results(spill_ilp_t * si)
4207 /* walk through the graph and collect all spills, reloads and remats for a value */
4209 si->values = new_set(cmp_defs, 4096);
4211 DBG((si->dbg, LEVEL_1, "Applying results\n"));
4212 delete_unnecessary_remats(si);
4213 si->m_unknown = new_r_Unknown(si->birg->irg, mode_M);
4214 irg_block_walk_graph(si->birg->irg, walker_spill_placer, NULL, si);
4215 irg_block_walk_graph(si->birg->irg, walker_reload_placer, NULL, si);
4217 insert_memoperands(si);
4220 /* clean the remat info! there are still back-edges leading there! */
4221 clean_remat_info(si);
4225 connect_all_spills_with_keep(si);
4227 del_set(si->values);
4231 get_n_regs(spill_ilp_t * si)
4233 int arch_n_regs = arch_register_class_n_regs(si->cls);
4235 bitset_t *arch_regs = bitset_malloc(arch_n_regs);
4236 bitset_t *abi_regs = bitset_malloc(arch_n_regs);
4238 arch_put_non_ignore_regs(si->birg->main_env->arch_env, si->cls, arch_regs);
4239 be_abi_put_ignore_regs(si->birg->abi, si->cls, abi_regs);
4241 bitset_andnot(arch_regs, abi_regs);
4242 arch_n_regs = bitset_popcnt(arch_regs);
4244 bitset_free(arch_regs);
4245 bitset_free(abi_regs);
4247 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", arch_n_regs, si->cls->name));
4252 walker_reload_mover(ir_node * bb, void * data)
4254 spill_ilp_t *si = data;
4257 sched_foreach(bb, tmp) {
4258 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
4259 ir_node *reload = tmp;
4262 /* move reload upwards */
4264 int pressure = (int)get_irn_link(reload);
4265 if(pressure < si->n_regs) {
4266 irn = sched_prev(reload);
4267 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
4268 sched_remove(reload);
4269 pressure = (int)get_irn_link(irn);
4271 while(pressure < si->n_regs) {
4272 if( sched_is_end(irn) ||
4273 (be_is_Reload(irn) && has_reg_class(si, irn)) ||
4274 /* do not move reload before its spill */
4275 (irn == be_get_Reload_mem(reload)) ||
4276 /* do not move before phi */
4279 set_irn_link(irn, INT_TO_PTR(pressure+1));
4280 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
4281 irn = sched_prev(irn);
4283 pressure = (int)get_irn_link(irn);
4286 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
4287 sched_put_after(irn, reload);
4294 move_reloads_upward(spill_ilp_t * si)
4296 irg_block_walk_graph(si->birg->irg, walker_reload_mover, NULL, si);
4301 * Walk all irg blocks and check for interfering spills inside of phi classes
4304 luke_meminterferencechecker(ir_node * bb, void * data)
4306 spill_ilp_t *si = (spill_ilp_t*)data;
4309 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
4310 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
4312 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
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(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)) continue;
4321 /* a and b are only interesting if they are in the same phi class */
4322 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
4323 if(values_interfere_in_block(si, bb, a, b)) {
4324 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4333 verify_phiclasses(spill_ilp_t * si)
4335 /* analyze phi classes */
4336 phi_class_compute(si->birg->irg);
4338 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4339 irg_block_walk_graph(si->birg->irg, luke_meminterferencechecker, NULL, si);
4343 be_spill_remat(be_irg_t *birg, const arch_register_class_t *cls)
4346 char problem_name[256];
4347 char dump_suffix[256];
4348 char dump_suffix2[256];
4349 struct obstack obst;
4351 ir_graph *irg = be_get_birg_irg(birg);
4353 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", irg, cls->name);
4354 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", cls->name);
4355 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", cls->name);
4357 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4358 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4360 if(opt_verify & VERIFY_DOMINANCE)
4361 be_check_dominance(irg);
4363 be_assure_dom_front(birg);
4364 be_assure_liveness(birg);
4366 obstack_init(&obst);
4370 si.lpp = new_lpp(problem_name, lpp_minimize);
4371 si.remat_info = new_set(cmp_remat_info, 4096);
4372 si.interferences = new_set(cmp_interference, 32);
4373 si.memoperands = new_set(cmp_memoperands, 128);
4374 si.all_possible_remats = pset_new_ptr_default();
4375 si.spills = pset_new_ptr_default();
4376 si.inverse_ops = pset_new_ptr_default();
4379 si.n_regs = get_n_regs(&si);
4381 set_irg_link(irg, &si);
4384 /* compute phi classes */
4385 // phi_class_compute(irg);
4387 if(opt_dump_flags & DUMP_STATS)
4388 be_analyze_regpressure(birg, cls, "-pre");
4390 DBG((si.dbg, LEVEL_2, "\t initializing\n"));
4391 irg_block_walk_graph(irg, luke_initializer, NULL, &si);
4394 /* collect remats */
4395 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4396 irg_walk_graph(irg, walker_remat_collector, NULL, &si);
4399 /* insert possible remats */
4400 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4401 irg_block_walk_graph(irg, walker_remat_insertor, NULL, &si);
4402 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4404 if(opt_keep_alive & KEEPALIVE_REMATS) {
4405 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4406 connect_all_remats_with_keep(&si);
4407 /* dump graph with inserted remats */
4408 dump_graph_with_remats(irg, dump_suffix);
4411 /* insert copies for phi arguments not in my regclass */
4412 irg_walk_graph(irg, walker_regclass_copy_insertor, NULL, &si);
4414 /* recompute liveness */
4415 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4416 be_liveness_recompute(si.lv);
4419 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4420 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4421 irg_block_walk_graph(irg, luke_endwalker, NULL, &si);
4423 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4424 irg_block_walk_graph(irg, luke_blockwalker, NULL, &si);
4427 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4428 memcopyhandler(&si);
4431 if(opt_dump_flags & DUMP_PROBLEM) {
4433 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4434 if ((f = fopen(buf, "wt")) != NULL) {
4435 lpp_dump_plain(si.lpp, f);
4440 if(opt_dump_flags & DUMP_MPS) {
4443 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4444 if((f = fopen(buf, "wt")) != NULL) {
4445 mps_write_mps(si.lpp, s_mps_fixed, f);
4449 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4450 if((f = fopen(buf, "wt")) != NULL) {
4451 mps_write_mst(si.lpp, s_mps_fixed, f);
4456 lpp_check_startvals(si.lpp);
4459 DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4460 lpp_set_time_limit(si.lpp, opt_timeout);
4463 lpp_set_log(si.lpp, stdout);
4466 lpp_solve_cplex(si.lpp);
4468 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4470 assert(lpp_is_sol_valid(si.lpp)
4471 && "solution of ILP must be valid");
4473 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));
4475 if(opt_dump_flags & DUMP_SOLUTION) {
4479 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4480 if ((f = fopen(buf, "wt")) != NULL) {
4482 for (i = 0; i < si.lpp->var_next; ++i) {
4483 lpp_name_t *name = si.lpp->vars[i];
4484 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4490 #ifndef SCHEDULE_PHIM
4491 si.phims = pset_new_ptr_default();
4493 writeback_results(&si);
4498 kill_all_unused_values_in_schedule(&si);
4500 #if !defined(SCHEDULE_PHIM) && defined(SOLVE)
4504 if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
4505 be_dump(irg, "-spills-placed", dump_ir_block_graph);
4507 // move reloads upwards
4508 be_liveness_recompute(si.lv);
4509 irg_block_walk_graph(irg, walker_pressure_annotator, NULL, &si);
4510 move_reloads_upward(&si);
4513 verify_phiclasses(&si);
4516 irg_block_walk_graph(irg, walker_pressure_annotator, NULL, &si);
4518 if(opt_dump_flags & DUMP_PRESSURE)
4519 dump_pressure_graph(&si, dump_suffix2);
4521 if(opt_dump_flags & DUMP_STATS)
4522 be_analyze_regpressure(birg, cls, "-post");
4524 if(opt_verify & VERIFY_DOMINANCE)
4525 be_check_dominance(irg);
4528 del_set(si.interferences);
4529 del_pset(si.inverse_ops);
4530 del_pset(si.all_possible_remats);
4531 del_set(si.memoperands);
4532 del_pset(si.spills);
4534 obstack_free(&obst, NULL);
4535 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4538 static void be_spill_remat_oldinterface(const be_chordal_env_t *cenv)
4540 return be_spill_remat(cenv->birg, cenv->cls);
4543 void be_init_spillremat(void)
4545 static be_spiller_t remat_spiller = {
4546 be_spill_remat_oldinterface
4548 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
4549 lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
4550 lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
4551 lc_opt_entry_t *remat_grp = lc_opt_get_grp(chordal_grp, "remat");
4553 be_register_spiller("remat", &remat_spiller);
4554 lc_opt_add_table(remat_grp, options);
4557 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillremat);
4559 #else /* WITH_ILP */
4562 only_that_you_can_compile_without_WITH_ILP_defined(void)
4566 #endif /* WITH_ILP */