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
37 #include <lpp/lpp_net.h>
38 #include <lpp/lpp_cplex.h>
39 //#include <lc_pset.h>
40 #include <libcore/lc_bitset.h>
44 #include "besched_t.h"
49 #include "bespillremat.h"
51 #include "bepressurestat.h"
53 #include "bechordal_t.h"
59 //#define KEEPALIVE /* keep alive all inserted remats and dump graph with remats */
60 #define COLLECT_REMATS /* enable rematerialization */
61 #define COLLECT_INVERSE_REMATS /* enable placement of inverse remats */
62 #define REMAT_WHILE_LIVE /* only remat values that are live */
63 //#define NO_ENLARGE_L1V3N355 /* do not remat after the death of some operand */
64 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
65 #define MAY_DIE_AT_REMAT /* allow values to die after a pre remat */
66 #define NO_SINGLE_USE_REMATS /* do not repair schedule */
67 //#define KEEPALIVE_SPILLS
68 //#define KEEPALIVE_RELOADS
69 #define GOODWIN_REDUCTION
74 #define LPP_SERVER "i44pc52"
75 #define LPP_SOLVER "cplex"
81 #define ILP_TIMEOUT 120
85 typedef struct _spill_ilp_t {
86 const arch_register_class_t *cls;
88 const be_chordal_env_t *chordal_env;
92 pset *all_possible_remats;
97 set *values; /**< for collecting all definitions of values before running ssa-construction */
102 DEBUG_ONLY(firm_dbg_module_t * dbg);
105 typedef int ilp_var_t;
106 typedef int ilp_cst_t;
108 typedef struct _spill_bb_t {
113 typedef struct _remat_t {
114 const ir_node *op; /**< for copy_irn */
115 const ir_node *value; /**< the value which is being recomputed by this remat */
116 ir_node *proj; /**< not NULL if the above op produces a tuple */
117 int cost; /**< cost of this remat */
118 int inverse; /**< nonzero if this is an inverse remat */
122 * Data to be attached to each IR node. For remats this contains the ilp_var
123 * for this remat and for normal ops this contains the ilp_vars for
124 * reloading each operand
126 typedef struct _op_t {
131 remat_t *remat; /** the remat this op belongs to */
132 int pre; /** 1, if this is a pressure-increasing remat */
136 ir_node *op; /** the operation this live range belongs to */
145 typedef struct _defs_t {
147 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
148 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
151 typedef struct _remat_info_t {
152 const ir_node *irn; /**< the irn to which these remats belong */
153 pset *remats; /**< possible remats for this value */
154 pset *remats_by_operand; /**< remats with this value as operand */
157 typedef struct _keyval_t {
162 typedef struct _spill_t {
172 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
174 return chordal_has_class(si->chordal_env, irn);
179 cmp_remat(const void *a, const void *b)
181 const keyval_t *p = a;
182 const keyval_t *q = b;
183 const remat_t *r = p->val;
184 const remat_t *s = q->val;
188 return !(r == s || r->op == s->op);
192 cmp_remat(const void *a, const void *b)
194 const remat_t *r = a;
195 const remat_t *s = a;
197 return !(r == s || r->op == s->op);
201 cmp_spill(const void *a, const void *b, size_t size)
203 const spill_t *p = a;
204 const spill_t *q = b;
206 // return !(p->irn == q->irn && p->bb == q->bb);
207 return !(p->irn == q->irn);
211 set_find_keyval(set * set, void * key)
216 return set_find(set, &query, sizeof(query), HASH_PTR(key));
220 set_insert_keyval(set * set, void * key, void * val)
226 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
230 set_find_def(set * set, ir_node * value)
235 return set_find(set, &query, sizeof(query), HASH_PTR(value));
239 set_insert_def(set * set, ir_node * value)
246 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
250 set_find_spill(set * set, ir_node * value)
255 return set_find(set, &query, sizeof(query), HASH_PTR(value));
258 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
259 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
260 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
261 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
262 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
265 cmp_remat_info(const void *a, const void *b, size_t size)
267 const remat_info_t *p = a;
268 const remat_info_t *q = b;
270 return !(p->irn == q->irn);
274 cmp_defs(const void *a, const void *b, size_t size)
279 return !(p->value == q->value);
283 cmp_keyval(const void *a, const void *b, size_t size)
285 const keyval_t *p = a;
286 const keyval_t *q = b;
288 return !(p->key == q->key);
292 execution_frequency(const spill_ilp_t * si, const ir_node * irn)
297 return get_block_execfreq(si->execfreqs, irn) + FUDGE;
299 return get_block_execfreq(si->execfreqs, get_nodes_block(irn)) + FUDGE;
303 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
305 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
310 get_cost(const spill_ilp_t * si, const ir_node * irn)
312 if(be_is_Spill(irn)) {
314 } else if(be_is_Reload(irn)){
317 return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
323 * Checks, whether node and its operands have suitable reg classes
326 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
329 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
330 int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
334 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
337 for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
338 ir_node *op = get_irn_n(irn, n);
339 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
342 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
349 * Try to create a remat from @p op with destination value @p dest_value
351 static INLINE remat_t *
352 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
354 remat_t *remat = NULL;
356 // if(!mode_is_datab(get_irn_mode(dest_value)))
359 if(dest_value == op) {
360 const ir_node *proj = NULL;
362 if(is_Proj(dest_value)) {
363 op = get_irn_n(op, 0);
367 if(!is_rematerializable(si, op))
370 remat = obstack_alloc(si->obst, sizeof(*remat));
372 remat->cost = get_cost(si, op);
373 remat->value = dest_value;
377 arch_inverse_t inverse;
380 /* get the index of the operand we want to retrieve by the inverse op */
381 for (n = get_irn_arity(op)-1; n>=0; --n) {
382 ir_node *arg = get_irn_n(op, n);
384 if(arg == dest_value) break;
388 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", n, op));
390 /* else ask the backend to give an inverse op */
391 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
394 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
396 assert(inverse.n > 0 && "inverse op should have at least one node");
398 for(i=inverse.n-1; i>=0; --i) {
399 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
403 remat = obstack_alloc(si->obst, sizeof(*remat));
404 remat->op = inverse.nodes[0];
405 remat->cost = inverse.costs;
406 remat->value = dest_value;
407 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
410 assert(is_Proj(remat->proj));
412 assert(0 && "I can not handle remats with more than 2 nodes");
419 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
421 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
429 add_remat(const spill_ilp_t * si, const remat_t * remat)
431 remat_info_t *remat_info,
436 assert(remat->value);
438 query.irn = remat->value;
440 query.remats_by_operand = NULL;
441 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
443 if(remat_info->remats == NULL) {
444 remat_info->remats = new_pset(cmp_remat, 4096);
446 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
448 /* insert the remat into the remats_be_operand set of each argument of the remat op */
449 for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
450 ir_node *arg = get_irn_n(remat->op, n);
454 query.remats_by_operand = NULL;
455 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
457 if(remat_info->remats_by_operand == NULL) {
458 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
460 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
465 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
467 const ir_edge_t *edge = get_irn_out_edge_first(irn);
471 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
474 edge = get_irn_out_edge_next(irn, edge);
481 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
486 #ifdef NO_SINGLE_USE_REMATS
487 if(has_reg_class(si, op) && (get_irn_n_nonremat_edges(si, op) > 1)) {
489 if(has_reg_class(si, op)) {
491 remat = get_remat_from_op(si, op, op);
493 add_remat(si, remat);
497 #ifdef COLLECT_INVERSE_REMATS
498 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
500 for (n = get_irn_arity(op)-1; n>=0; --n) {
501 ir_node *arg = get_irn_n(op, n);
503 if(has_reg_class(si, arg)) {
504 /* try to get an inverse remat */
505 remat = get_remat_from_op(si, arg, op);
507 add_remat(si, remat);
516 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
519 ir_node *def_block = get_nodes_block(val);
525 /* if pos is at end of a basic block */
527 ret = (pos == def_block || block_dominates(def_block, pos));
528 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
532 /* else if this is a normal operation */
533 block = get_nodes_block(pos);
534 if(block == def_block) {
535 if(!sched_is_scheduled(val)) return 1;
537 ret = sched_comes_after(val, pos);
538 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
542 ret = block_dominates(def_block, block);
543 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
547 static INLINE ir_node *
548 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
550 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
554 * Returns first non-Phi node of block @p bb
556 static INLINE ir_node *
557 sched_block_first_nonphi(const ir_node * bb)
559 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
563 sched_skip_proj_predicator(const ir_node * irn, void * data)
565 return (is_Proj(irn));
568 static INLINE ir_node *
569 sched_next_nonproj(const ir_node * irn, int forward)
571 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
575 * Returns next operation node (non-Proj) after @p irn
576 * or the basic block of this node
578 static INLINE ir_node *
579 sched_next_op(const ir_node * irn)
581 ir_node *next = sched_next(irn);
586 return sched_next_nonproj(next, 1);
590 * Returns previous operation node (non-Proj) before @p irn
591 * or the basic block of this node
593 static INLINE ir_node *
594 sched_prev_op(const ir_node * irn)
596 ir_node *prev = sched_prev(irn);
601 return sched_next_nonproj(prev, 0);
605 sched_put_after(ir_node * insert, ir_node * irn)
607 if(is_Block(insert)) {
608 insert = sched_block_first_nonphi(insert);
610 insert = sched_next_op(insert);
612 sched_add_before(insert, irn);
616 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
618 if(is_Block(insert)) {
619 insert = sched_block_last_noncf(si, insert);
621 insert = sched_next_nonproj(insert, 0);
622 insert = sched_prev(insert);
624 sched_add_after(insert, irn);
628 * Tells you whether a @p remat can be placed before the irn @p pos
631 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
633 const ir_node *op = remat->op;
639 prev = sched_block_last_noncf(si, pos);
640 prev = sched_next_nonproj(prev, 0);
642 prev = sched_prev_op(pos);
644 /* do not remat if the rematted value is defined immediately before this op */
645 if(prev == remat->op) {
650 /* this should be just fine, the following OP will be using this value, right? */
652 /* only remat AFTER the real definition of a value (?) */
653 if(!value_is_defined_before(si, pos, remat->value)) {
654 // ir_fprintf(stderr, "error(not defined)");
659 for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
660 const ir_node *arg = get_irn_n(op, n);
662 #ifdef NO_ENLARGE_L1V3N355
663 if(has_reg_class(si, arg) && live) {
664 res &= pset_find_ptr(live, arg)?1:0;
666 res &= value_is_defined_before(si, pos, arg);
669 res &= value_is_defined_before(si, pos, arg);
677 * Tells you whether a @p remat can be placed after the irn @p pos
680 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
683 pos = sched_block_first_nonphi(pos);
685 pos = sched_next_op(pos);
688 /* only remat AFTER the real definition of a value (?) */
689 if(!value_is_defined_before(si, pos, remat->value)) {
693 return can_remat_before(si, remat, pos, live);
697 * Collect potetially rematerializable OPs
700 walker_remat_collector(ir_node * irn, void * data)
702 spill_ilp_t *si = data;
704 if(!is_Block(irn) && !is_Phi(irn)) {
705 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
706 get_remats_from_op(si, irn);
711 * Inserts a copy of @p irn before @p pos
714 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
719 bb = is_Block(pos)?pos:get_nodes_block(pos);
720 copy = exact_copy(irn);
721 set_nodes_block(copy, bb);
722 sched_put_before(si, pos, copy);
728 * Inserts a copy of @p irn after @p pos
731 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
736 bb = is_Block(pos)?pos:get_nodes_block(pos);
737 copy = exact_copy(irn);
738 set_nodes_block(copy, bb);
739 sched_put_after(pos, copy);
745 insert_remat_after(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
749 if(can_remat_after(si, remat, pos, live)) {
754 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
756 copy = insert_copy_after(si, remat->op, pos);
758 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
759 op = obstack_alloc(si->obst, sizeof(*op));
761 op->attr.remat.remat = remat;
762 op->attr.remat.pre = 0;
763 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
765 set_irn_link(copy, op);
766 pset_insert_ptr(si->all_possible_remats, copy);
768 proj_copy = insert_copy_after(si, remat->proj, copy);
769 set_irn_n(proj_copy, 0, copy);
770 set_irn_link(proj_copy, op);
771 pset_insert_ptr(si->all_possible_remats, proj_copy);
783 insert_remat_before(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
787 if(can_remat_before(si, remat, pos, live)) {
792 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
794 copy = insert_copy_before(si, remat->op, pos);
796 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
797 op = obstack_alloc(si->obst, sizeof(*op));
799 op->attr.remat.remat = remat;
800 op->attr.remat.pre = 1;
801 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
803 set_irn_link(copy, op);
804 pset_insert_ptr(si->all_possible_remats, copy);
806 proj_copy = insert_copy_after(si, remat->proj, copy);
807 set_irn_n(proj_copy, 0, copy);
808 set_irn_link(proj_copy, op);
809 pset_insert_ptr(si->all_possible_remats, proj_copy);
821 get_block_n_succs(const ir_node *block) {
822 const ir_edge_t *edge;
824 assert(edges_activated(current_ir_graph));
826 edge = get_block_succ_first(block);
830 edge = get_block_succ_next(block, edge);
835 is_merge_edge(const ir_node * bb)
837 #ifdef GOODWIN_REDUCTION
838 return get_block_n_succs(bb) == 1;
845 is_diverge_edge(const ir_node * bb)
847 #ifdef GOODWIN_REDUCTION
848 return get_Block_n_cfgpreds(bb) == 1;
855 walker_regclass_copy_insertor(ir_node * irn, void * data)
857 spill_ilp_t *si = data;
859 if(is_Phi(irn) && has_reg_class(si, irn)) {
862 for(n=get_irn_arity(irn)-1; n>=0; --n) {
863 ir_node *phi_arg = get_irn_n(irn, n);
864 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
866 if(!has_reg_class(si, phi_arg)) {
867 ir_node *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
868 ir_node *pos = sched_block_last_noncf(si, bb);
869 op_t *op = obstack_alloc(si->obst, sizeof(*op));
871 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
872 sched_add_after(pos, copy);
873 set_irn_n(irn, n, copy);
876 op->attr.live_range.args.reloads = NULL;
877 op->attr.live_range.ilp = ILP_UNDEF;
878 set_irn_link(copy, op);
886 * Insert (so far unused) remats into the irg to
887 * recompute the potential liveness of all values
890 walker_remat_insertor(ir_node * bb, void * data)
892 spill_ilp_t *si = data;
893 spill_bb_t *spill_bb;
897 pset *live = pset_new_ptr_default();
899 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
901 live_foreach(bb, li) {
902 ir_node *value = (ir_node *) li->irn;
904 /* add remats at end of block */
905 if (live_is_end(li) && has_reg_class(si, value)) {
906 pset_insert_ptr(live, value);
910 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
911 set_irn_link(bb, spill_bb);
913 irn = sched_last(bb);
914 while(!sched_is_end(irn)) {
921 next = sched_prev(irn);
923 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
925 if(is_Phi(irn) || is_Proj(irn)) {
928 if(has_reg_class(si, irn)) {
929 pset_remove_ptr(live, irn);
932 op = obstack_alloc(si->obst, sizeof(*op));
934 op->attr.live_range.args.reloads = NULL;
935 op->attr.live_range.ilp = ILP_UNDEF;
936 set_irn_link(irn, op);
942 op = obstack_alloc(si->obst, sizeof(*op));
944 op->attr.live_range.ilp = ILP_UNDEF;
945 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
946 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
947 set_irn_link(irn, op);
949 args = pset_new_ptr_default();
951 /* collect arguments of op */
952 for (n = get_irn_arity(irn)-1; n>=0; --n) {
953 ir_node *arg = get_irn_n(irn, n);
955 pset_insert_ptr(args, arg);
958 /* set args of op already live in epilog */
959 pset_foreach(args, arg) {
960 if(has_reg_class(si, arg)) {
961 pset_insert_ptr(live, arg);
964 /* delete defined value from live set */
965 if(has_reg_class(si, irn)) {
966 pset_remove_ptr(live, irn);
970 remat_args = pset_new_ptr_default();
972 /* insert all possible remats before irn */
973 pset_foreach(args, arg) {
974 remat_info_t *remat_info,
978 /* continue if the operand has the wrong reg class
980 if(!has_reg_class(si, arg))
985 query.remats_by_operand = NULL;
986 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
992 if(remat_info->remats) {
993 pset_foreach(remat_info->remats, remat) {
994 ir_node *remat_irn = NULL;
996 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
997 #ifdef REMAT_WHILE_LIVE
998 if(pset_find_ptr(live, remat->value)) {
999 remat_irn = insert_remat_before(si, remat, irn, live);
1002 remat_irn = insert_remat_before(si, remat, irn, live);
1005 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1006 ir_node *remat_arg = get_irn_n(remat_irn, n);
1008 if(!has_reg_class(si, remat_arg)) continue;
1010 pset_insert_ptr(remat_args, remat_arg);
1017 /* now we add remat args to op's args because they could also die at this op */
1018 pset_foreach(args,arg) {
1019 if(pset_find_ptr(remat_args, arg)) {
1020 pset_remove_ptr(remat_args, arg);
1023 pset_foreach(remat_args,arg) {
1024 pset_insert_ptr(args, arg);
1027 /* insert all possible remats after irn */
1028 pset_foreach(args, arg) {
1029 remat_info_t *remat_info,
1033 /* continue if the operand has the wrong reg class */
1034 if(!has_reg_class(si, arg))
1038 query.remats = NULL;
1039 query.remats_by_operand = NULL;
1040 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1046 /* do not place post remats after jumps */
1047 if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
1049 if(remat_info->remats_by_operand) {
1050 pset_foreach(remat_info->remats_by_operand, remat) {
1051 /* do not insert remats producing the same value as one of the operands */
1052 if(!pset_find_ptr(args, remat->value)) {
1053 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
1054 #ifdef REMAT_WHILE_LIVE
1055 if(pset_find_ptr(live, remat->value)) {
1056 insert_remat_after(si, remat, irn, live);
1059 insert_remat_after(si, remat, irn, live);
1066 del_pset(remat_args);
1071 live_foreach(bb, li) {
1072 ir_node *value = (ir_node *) li->irn;
1074 /* add remats at end if successor has multiple predecessors */
1075 if(is_merge_edge(bb)) {
1076 /* add remats at end of block */
1077 if (live_is_end(li) && has_reg_class(si, value)) {
1078 remat_info_t *remat_info,
1083 query.remats = NULL;
1084 query.remats_by_operand = NULL;
1085 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1087 if(remat_info && remat_info->remats) {
1088 pset_foreach(remat_info->remats, remat) {
1089 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1091 insert_remat_before(si, remat, bb, NULL);
1096 if(is_diverge_edge(bb)) {
1097 /* add remat2s at beginning of block */
1098 if ((live_is_in(li) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) {
1099 remat_info_t *remat_info,
1104 query.remats = NULL;
1105 query.remats_by_operand = NULL;
1106 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1108 if(remat_info && remat_info->remats) {
1109 pset_foreach(remat_info->remats, remat) {
1110 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at beginning of block %+F\n", remat->op, bb));
1112 /* put the remat here if all its args are available */
1113 insert_remat_after(si, remat, bb, NULL);
1123 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1126 luke_endwalker(ir_node * bb, void * data)
1128 spill_ilp_t *si = (spill_ilp_t*)data;
1135 spill_bb_t *spill_bb = get_irn_link(bb);
1138 live = pset_new_ptr_default();
1139 use_end = pset_new_ptr_default();
1141 live_foreach(bb, li) {
1142 irn = (ir_node *) li->irn;
1143 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1146 pset_insert_ptr(live, irn);
1147 op = get_irn_link(irn);
1148 assert(!op->is_remat);
1152 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1153 /* their reg_out must always be set */
1154 sched_foreach_reverse(bb, irn) {
1157 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1159 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1160 ir_node *irn_arg = get_irn_n(irn, n);
1162 if(has_reg_class(si, irn_arg)) {
1163 pset_insert_ptr(use_end, irn_arg);
1168 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1169 //cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1170 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1172 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1174 /* if this is a merge edge we can reload at the end of this block */
1175 if(is_merge_edge(bb)) {
1176 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1177 } else if(pset_count(use_end)){
1178 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1180 spill_bb->reloads = NULL;
1183 pset_foreach(live,irn) {
1189 /* handle values used by control flow nodes later separately */
1190 if(pset_find_ptr(use_end, irn)) continue;
1193 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1195 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1197 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1198 spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1199 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1201 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1202 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1204 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1205 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1207 if(is_merge_edge(bb)) {
1211 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1212 reload = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1213 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1215 /* reload <= mem_out */
1216 rel_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1217 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1218 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1221 spill->reg_in = ILP_UNDEF;
1222 spill->mem_in = ILP_UNDEF;
1225 pset_foreach(use_end,irn) {
1229 ilp_cst_t end_use_req,
1234 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1236 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1238 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1239 spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1240 /* if irn is used at the end of the block, then it is live anyway */
1241 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1243 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1244 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1246 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1247 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1249 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1250 reload = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1251 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1253 /* reload <= mem_out */
1254 rel_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1255 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1256 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1258 spill->reg_in = ILP_UNDEF;
1259 spill->mem_in = ILP_UNDEF;
1261 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1262 end_use_req = lpp_add_cst(si->lpp, buf, lpp_equal, 1);
1263 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1271 next_post_remat(const ir_node * irn)
1276 irn = sched_block_first_nonphi(irn);
1278 irn = sched_next_op(irn);
1281 if(sched_is_end(irn))
1284 op = (op_t*)get_irn_link(irn);
1285 if(op->is_remat && !op->attr.remat.pre) {
1294 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1300 ret = sched_block_last_noncf(si, irn);
1301 ret = sched_next(ret);
1302 ret = sched_prev_op(ret);
1304 ret = sched_prev_op(irn);
1307 if(sched_is_end(ret) || is_Phi(ret))
1310 op = (op_t*)get_irn_link(ret);
1311 if(op->is_remat && op->attr.remat.pre) {
1319 * Find a remat of value @p value in the epilog of @p pos
1322 find_post_remat(const ir_node * value, const ir_node * pos)
1324 while((pos = next_post_remat(pos)) != NULL) {
1327 op = get_irn_link(pos);
1328 assert(op->is_remat && !op->attr.remat.pre);
1330 if(op->attr.remat.remat->value == value)
1331 return (ir_node*)pos;
1334 const ir_edge_t *edge;
1335 foreach_out_edge(pos, edge) {
1336 ir_node *proj = get_edge_src_irn(edge);
1337 assert(is_Proj(proj));
1347 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1349 spill_bb_t *spill_bb = get_irn_link(bb);
1355 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1357 double spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1359 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1361 spill->reg_out = ILP_UNDEF;
1362 spill->reg_in = ILP_UNDEF;
1363 spill->mem_in = ILP_UNDEF;
1365 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1366 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1368 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1369 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1376 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1381 live_foreach(bb, li) {
1382 irn = (ir_node *) li->irn;
1384 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1385 pset_insert_ptr(live, irn);
1389 irn = sched_last(bb);
1391 /* all values eaten by control flow operations are also live until the end of the block */
1392 sched_foreach_reverse(bb, irn) {
1395 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1397 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1398 ir_node *arg = get_irn_n(irn,i);
1400 if(has_reg_class(si, arg)) {
1401 pset_insert_ptr(live, arg);
1408 * Inserts ILP-constraints and variables for memory copying before the given position
1411 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1413 const ir_node *succ;
1414 const ir_edge_t *edge;
1415 spill_bb_t *spill_bb = get_irn_link(block);
1424 assert(edges_activated(current_ir_graph));
1426 edge = get_block_succ_first(block);
1432 edge = get_block_succ_next(block, edge);
1433 /* next block can only contain phis, if this is a merge edge */
1436 ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1437 copyreg = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1439 ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1440 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1442 pset_foreach(live, tmp) {
1445 op_t *op = get_irn_link(irn);
1446 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1448 spill = set_find_spill(spill_bb->ilp, tmp);
1451 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1453 lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1455 sched_foreach(succ, phi) {
1456 const ir_node *to_copy;
1458 spill_t *to_copy_spill;
1459 op_t *phi_op = get_irn_link(phi);
1460 ilp_var_t reload = ILP_UNDEF;
1463 if(!is_Phi(phi)) break;
1464 if(!has_reg_class(si, phi)) continue;
1466 to_copy = get_irn_n(phi, pos);
1468 to_copy_op = get_irn_link(to_copy);
1470 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1471 assert(to_copy_spill);
1473 if(spill_bb->reloads) {
1474 keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1477 reload = PTR_TO_INT(keyval->val);
1481 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N", block, to_copy);
1482 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1484 /* copy - reg_out - reload - remat - live_range <= 0 */
1485 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1486 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1487 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1488 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1489 foreach_pre_remat(si, block, tmp) {
1490 op_t *remat_op = get_irn_link(tmp);
1491 if(remat_op->attr.remat.remat->value == to_copy) {
1492 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1496 ir_snprintf(buf, sizeof(buf), "copyreq_%N_%N", block, to_copy);
1497 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1499 /* copy - reg_out - copyreg <= 0 */
1500 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1501 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1502 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1508 * Walk all irg blocks and emit this ILP
1511 luke_blockwalker(ir_node * bb, void * data)
1513 spill_ilp_t *si = (spill_ilp_t*)data;
1518 spill_bb_t *spill_bb = get_irn_link(bb);
1521 pset *defs = pset_new_ptr_default();
1524 live = pset_new_ptr_default();
1526 /****************************************
1527 * B A S I C B L O C K E N D
1528 ***************************************/
1531 /* init live values at end of block */
1532 get_live_end(si, bb, live);
1534 pset_foreach(live, irn) {
1536 ilp_var_t reload = ILP_UNDEF;
1538 spill = set_find_spill(spill_bb->ilp, irn);
1541 if(spill_bb->reloads) {
1542 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1545 reload = PTR_TO_INT(keyval->val);
1549 op = get_irn_link(irn);
1550 assert(!op->is_remat);
1552 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1553 op->attr.live_range.ilp = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1554 op->attr.live_range.op = bb;
1556 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1557 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1559 /* reg_out - reload - remat - live_range <= 0 */
1560 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1561 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1562 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1563 foreach_pre_remat(si, bb, tmp) {
1564 op_t *remat_op = get_irn_link(tmp);
1565 if(remat_op->attr.remat.remat->value == irn) {
1566 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1569 /* maybe we should also assure that reg_out >= live_range etc. */
1572 #ifndef NO_MEMCOPIES
1573 insert_mem_copy_position(si, live, bb);
1577 * start new live ranges for values used by remats at end of block
1578 * and assure the remat args are available
1580 foreach_pre_remat(si, bb, tmp) {
1581 op_t *remat_op = get_irn_link(tmp);
1584 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1585 ir_node *remat_arg = get_irn_n(tmp, n);
1586 op_t *arg_op = get_irn_link(remat_arg);
1589 if(!has_reg_class(si, remat_arg)) continue;
1591 /* if value is becoming live through use by remat */
1592 if(!pset_find_ptr(live, remat_arg)) {
1593 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1594 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1596 arg_op->attr.live_range.ilp = prev_lr;
1597 arg_op->attr.live_range.op = bb;
1599 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1601 pset_insert_ptr(live, remat_arg);
1602 add_to_spill_bb(si, bb, remat_arg);
1605 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1606 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1607 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1609 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1610 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1612 /* use reload placed for this argument */
1613 if(spill_bb->reloads) {
1614 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1617 ilp_var_t reload = PTR_TO_INT(keyval->val);
1619 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1624 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1629 /**************************************
1630 * B A S I C B L O C K B O D Y
1631 **************************************/
1633 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1639 ilp_cst_t check_pre,
1646 /* iterate only until first phi */
1650 op = get_irn_link(irn);
1652 if(op->is_remat) continue;
1653 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1655 /* collect defined values */
1656 if(has_reg_class(si, irn)) {
1657 pset_insert_ptr(defs, irn);
1661 if(is_Proj(irn)) continue;
1664 * init set of irn's arguments
1665 * and all possibly used values around this op
1666 * and values defined by post remats
1668 args = new_set(cmp_keyval, get_irn_arity(irn));
1669 used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1670 remat_defs = pset_new_ptr(pset_count(live));
1672 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1673 ir_node *irn_arg = get_irn_n(irn, n);
1674 if(has_reg_class(si, irn_arg)) {
1675 set_insert_keyval(args, irn_arg, (void*)n);
1676 pset_insert_ptr(used, irn_arg);
1679 foreach_post_remat(irn, tmp) {
1680 op_t *remat_op = get_irn_link(tmp);
1682 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1684 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1685 ir_node *remat_arg = get_irn_n(tmp, n);
1686 if(has_reg_class(si, remat_arg)) {
1687 pset_insert_ptr(used, remat_arg);
1691 foreach_pre_remat(si, irn, tmp) {
1692 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1693 ir_node *remat_arg = get_irn_n(tmp, n);
1694 if(has_reg_class(si, remat_arg)) {
1695 pset_insert_ptr(used, remat_arg);
1700 /**********************************
1701 * I N E P I L O G O F irn
1702 **********************************/
1704 /* ensure each dying value is used by only one post remat */
1705 pset_foreach(live, tmp) {
1706 ir_node *value = tmp;
1707 op_t *value_op = get_irn_link(value);
1712 foreach_post_remat(irn, remat) {
1713 op_t *remat_op = get_irn_link(remat);
1715 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1716 ir_node *remat_arg = get_irn_n(remat, n);
1718 /* if value is used by this remat add it to constraint */
1719 if(remat_arg == value) {
1721 /* sum remat2s <= 1 + n_remats*live_range */
1722 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1723 cst = lpp_add_cst(si->lpp, buf, lpp_less, 1.0);
1727 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1733 if(value_op->attr.live_range.ilp != ILP_UNDEF && cst != ILP_UNDEF) {
1734 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1740 /* new live ranges for values from L\U defined by post remats */
1741 pset_foreach(live, tmp) {
1742 ir_node *value = tmp;
1743 op_t *value_op = get_irn_link(value);
1745 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
1746 ilp_var_t prev_lr = ILP_UNDEF;
1749 if(pset_find_ptr(remat_defs, value)) {
1751 /* next_live_range <= prev_live_range + sum remat2s */
1752 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1753 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1755 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1756 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1758 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1759 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1761 foreach_post_remat(irn, remat) {
1762 op_t *remat_op = get_irn_link(remat);
1764 /* if value is being rematerialized by this remat */
1765 if(value == remat_op->attr.remat.remat->value) {
1766 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1770 value_op->attr.live_range.ilp = prev_lr;
1771 value_op->attr.live_range.op = irn;
1776 /* requirements for post remats and start live ranges from L/U' for values dying here */
1777 foreach_post_remat(irn, tmp) {
1778 op_t *remat_op = get_irn_link(tmp);
1781 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1782 ir_node *remat_arg = get_irn_n(tmp, n);
1783 op_t *arg_op = get_irn_link(remat_arg);
1785 if(!has_reg_class(si, remat_arg)) continue;
1787 /* only for values in L\U (TODO and D?), the others are handled with post_use */
1788 if(!pset_find_ptr(used, remat_arg)) {
1789 /* remat <= live_rang(remat_arg) */
1790 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
1791 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1793 /* if value is becoming live through use by remat2 */
1794 if(!pset_find_ptr(live, remat_arg)) {
1797 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1798 lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1800 arg_op->attr.live_range.ilp = lr;
1801 arg_op->attr.live_range.op = irn;
1803 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
1805 pset_insert_ptr(live, remat_arg);
1806 add_to_spill_bb(si, bb, remat_arg);
1809 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1810 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1815 d = pset_count(defs);
1816 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
1818 /* count how many regs irn needs for arguments */
1819 u = set_count(args);
1822 /* check the register pressure in the epilog */
1823 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
1824 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
1825 check_post = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - d);
1827 /* add L\U' to check_post */
1828 pset_foreach(live, tmp) {
1829 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
1830 /* if a live value is not used by irn */
1831 tmp_op = get_irn_link(tmp);
1832 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
1836 /***********************************************************
1837 * 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
1838 **********************************************************/
1841 pset_foreach(used, tmp) {
1847 op_t *arg_op = get_irn_link(arg);
1850 spill = add_to_spill_bb(si, bb, arg);
1852 /* new live range for each used value */
1853 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
1854 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1856 /* the epilog stuff - including post_use, check_post, check_post_remat */
1857 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
1858 post_use = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1860 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
1862 /* arg is live throughout epilog if the next live_range is in a register */
1863 if(pset_find_ptr(live, arg)) {
1864 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
1866 /* post_use >= next_lr + remat */
1867 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1868 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1869 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
1870 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1874 /* if value is not an arg of op and not possibly defined by post remat
1875 * then it may only die and not become live
1877 if(!set_find_keyval(args, arg)) {
1878 /* post_use <= prev_lr */
1879 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
1880 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1881 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1882 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1884 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
1885 /* next_lr <= prev_lr */
1886 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
1887 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1888 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1889 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1895 /* forall post remat which use arg add a similar cst */
1896 foreach_post_remat(irn, remat) {
1899 for (n=get_irn_arity(remat)-1; n>=0; --n) {
1900 ir_node *remat_arg = get_irn_n(remat, n);
1901 op_t *remat_op = get_irn_link(remat);
1903 if(remat_arg == arg) {
1904 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
1906 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1907 cst = lpp_add_cst(si->lpp, buf, lpp_greater, 0.0);
1908 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1909 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1914 /* new live range begins for each used value */
1915 arg_op->attr.live_range.ilp = prev_lr;
1916 arg_op->attr.live_range.op = irn;
1918 /*if(!pset_find_ptr(live, arg)) {
1919 pset_insert_ptr(live, arg);
1920 add_to_spill_bb(si, bb, arg);
1922 pset_insert_ptr(live, arg);
1926 /* just to be sure */
1927 check_post = ILP_UNDEF;
1936 /* check the register pressure in the prolog */
1937 /* sum_{L\U} lr <= k - |U| */
1938 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
1939 check_pre = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - u);
1941 /* for the prolog remove defined values from the live set */
1942 pset_foreach(defs, tmp) {
1943 pset_remove_ptr(live, tmp);
1946 /***********************************************************
1947 * 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
1948 **********************************************************/
1951 set_foreach(args, keyval) {
1953 ir_node *arg = keyval->key;
1954 int i = PTR_TO_INT(keyval->val);
1955 op_t *arg_op = get_irn_link(arg);
1957 spill = set_find_spill(spill_bb->ilp, arg);
1960 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
1961 op->attr.live_range.args.reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1963 /* reload <= mem_out */
1964 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
1965 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1966 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
1967 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
1969 /* requirement: arg must be in register for use */
1970 /* reload + remat + live_range == 1 */
1971 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
1972 cst = lpp_add_cst(si->lpp, buf, lpp_equal, 1.0);
1974 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1975 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
1976 foreach_pre_remat(si, irn, tmp) {
1977 op_t *remat_op = get_irn_link(tmp);
1978 if(remat_op->attr.remat.remat->value == arg) {
1979 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1984 /* iterate over L\U */
1985 pset_foreach(live, tmp) {
1986 if(!set_find_keyval(args, tmp)) {
1987 /* if a live value is not used by irn */
1988 tmp_op = get_irn_link(tmp);
1989 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
1994 /* requirements for remats */
1995 /* start new live ranges for values used by remats */
1996 foreach_pre_remat(si, irn, tmp) {
1997 op_t *remat_op = get_irn_link(tmp);
2000 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2001 ir_node *remat_arg = get_irn_n(tmp, n);
2002 op_t *arg_op = get_irn_link(remat_arg);
2005 if(!has_reg_class(si, remat_arg)) continue;
2007 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2008 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2009 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2011 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2012 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2014 /* if remat arg is also used by current op then we can use reload placed for this argument */
2015 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2016 int index = (int)keyval->val;
2018 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2026 /*************************
2027 * D O N E W I T H O P
2028 *************************/
2030 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2032 pset_foreach(live, tmp) {
2033 assert(has_reg_class(si, tmp));
2036 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2037 ir_node *arg = get_irn_n(irn, n);
2039 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2042 del_pset(remat_defs);
2046 defs = pset_new_ptr_default();
2051 /***************************************
2052 * B E G I N N I N G O F B L O C K
2053 ***************************************/
2056 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2057 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2059 pset_foreach(live, irn) {
2060 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2063 /* construct mem_outs for all values */
2065 set_foreach(spill_bb->ilp, spill) {
2066 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2067 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2069 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2070 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2072 if(pset_find_ptr(live, spill->irn)) {
2073 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2075 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2076 spill->mem_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2077 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2079 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2081 op_t *op = get_irn_link(spill->irn);
2083 /* do we have to copy a phi argument? */
2084 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2085 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2087 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2088 const ir_node *arg = get_irn_n(spill->irn, n);
2094 /* argument already done? */
2095 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2097 /* get sum of execution frequencies of blocks with the same phi argument */
2098 for(m=n; m>=0; --m) {
2099 const ir_node *arg2 = get_irn_n(spill->irn, m);
2102 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2106 /* copies are not for free */
2107 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2108 var = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE * freq);
2110 for(m=n; m>=0; --m) {
2111 const ir_node *arg2 = get_irn_n(spill->irn, m);
2114 op->attr.live_range.args.copies[m] = var;
2118 /* copy <= mem_in */
2119 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2120 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2121 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2122 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2129 /* L\U is empty at bb start */
2130 /* arg is live throughout epilog if it is reg_in into this block */
2132 /* check the register pressure at the beginning of the block
2135 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2136 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
2138 pset_foreach(live, irn) {
2141 spill = set_find_spill(spill_bb->ilp, irn);
2144 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2145 spill->reg_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2147 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2149 /* spill + mem_in <= 1 */
2150 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2151 nospill = lpp_add_cst(si->lpp, buf, lpp_less, 1);
2153 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2154 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2157 foreach_post_remat(bb, irn) {
2158 op_t *remat_op = get_irn_link(irn);
2160 DBG((si->dbg, LEVEL_4, "\t next post remat: %+F\n", irn));
2161 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
2163 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2166 /* forall post remats add requirements */
2167 foreach_post_remat(bb, tmp) {
2170 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2171 ir_node *remat_arg = get_irn_n(tmp, n);
2172 op_t *remat_op = get_irn_link(tmp);
2174 if(!has_reg_class(si, remat_arg)) continue;
2176 spill = set_find_spill(spill_bb->ilp, remat_arg);
2179 /* remat <= reg_in_argument */
2180 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2181 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2182 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2183 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2187 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2188 pset_foreach(live, irn) {
2192 spill = set_find_spill(spill_bb->ilp, irn);
2193 assert(spill && spill->irn == irn);
2195 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2196 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2199 ir_node *phi_arg = get_Phi_pred(irn, n);
2200 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2201 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2204 /* although the phi is in the right regclass one or more of
2205 * its arguments can be in a different one or at least to
2208 if(has_reg_class(si, phi_arg)) {
2209 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2210 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2211 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2212 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2214 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2215 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2217 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2220 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2221 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2225 /* else assure the value arrives on all paths in the same resource */
2227 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2230 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2231 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2234 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2235 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2236 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2237 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2239 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2240 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2242 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2245 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2246 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2251 /* first live ranges from reg_ins */
2252 pset_foreach(live, irn) {
2253 op_t *op = get_irn_link(irn);
2255 spill = set_find_spill(spill_bb->ilp, irn);
2256 assert(spill && spill->irn == irn);
2258 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2259 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2260 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2261 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2263 foreach_post_remat(bb, tmp) {
2264 op_t *remat_op = get_irn_link(tmp);
2266 if(remat_op->attr.remat.remat->value == irn) {
2267 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2272 /* walk forward now and compute constraints for placing spills */
2273 /* this must only be done for values that are not defined in this block */
2274 /* TODO are these values at start of block? if yes, just check whether this is a diverge edge and skip the loop */
2275 pset_foreach(live, irn) {
2277 * if value is defined in this block we can anways place the spill directly after the def
2278 * -> no constraint necessary
2280 if(!is_Phi(irn) && get_nodes_block(irn) == bb) continue;
2283 spill = set_find_spill(spill_bb->ilp, irn);
2286 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2287 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2289 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2290 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2293 sched_foreach_op(bb, tmp) {
2294 op_t *op = get_irn_link(tmp);
2296 if(is_Phi(tmp)) continue;
2297 assert(!is_Proj(tmp));
2300 ir_node *value = op->attr.remat.remat->value;
2303 /* only collect remats up to the first use of a value */
2304 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2309 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2310 ir_node *arg = get_irn_n(tmp, n);
2313 /* if a value is used stop collecting remats */
2319 if(cst == ILP_UNDEF) break;
2327 typedef struct _irnlist_t {
2328 struct list_head list;
2332 typedef struct _interference_t {
2333 struct list_head blocklist;
2339 cmp_interference(const void *a, const void *b, size_t size)
2341 const interference_t *p = a;
2342 const interference_t *q = b;
2344 return !(p->a == q->a && p->b == q->b);
2347 static interference_t *
2348 set_find_interference(set * set, ir_node * a, ir_node * b)
2350 interference_t query;
2352 query.a = (a>b)?a:b;
2353 query.b = (a>b)?b:a;
2355 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2358 static interference_t *
2359 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2361 interference_t query,
2363 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2367 result = set_find_interference(set, a, b);
2370 list_add(&list->list, &result->blocklist);
2374 query.a = (a>b)?a:b;
2375 query.b = (a>b)?b:a;
2377 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2379 INIT_LIST_HEAD(&result->blocklist);
2380 list_add(&list->list, &result->blocklist);
2386 values_interfere_in_block(ir_node * bb, ir_node * a, ir_node * b)
2388 const ir_edge_t *edge;
2390 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2391 /* both values are live in, so they interfere */
2395 /* ensure a dominates b */
2396 if(value_dominates(b,a)) {
2402 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2405 /* the following code is stolen from bera.c */
2406 if(is_live_end(bb, a))
2409 foreach_out_edge(a, edge) {
2410 const ir_node *user = edge->src;
2411 if(get_nodes_block(user) == bb
2414 && value_dominates(b, user))
2422 * Walk all irg blocks and collect interfering values inside of phi classes
2425 luke_interferencewalker(ir_node * bb, void * data)
2427 spill_ilp_t *si = (spill_ilp_t*)data;
2431 live_foreach(bb, li1) {
2432 ir_node *a = (ir_node *) li1->irn;
2433 op_t *a_op = get_irn_link(a);
2435 if(a_op->is_remat) continue;
2437 /* a is only interesting if it is in my register class and if it is inside a phi class */
2438 if (has_reg_class(si, a) && get_phi_class(a)) {
2439 for(li2=li1->next; li2; li2 = li2->next) {
2440 ir_node *b = (ir_node *) li2->irn;
2441 op_t *b_op = get_irn_link(b);
2443 if(b_op->is_remat) continue;
2445 /* a and b are only interesting if they are in the same phi class */
2446 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2447 if(values_interfere_in_block(bb, a, b)) {
2448 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2449 set_insert_interference(si, si->interferences, a, b, bb);
2457 static unsigned int copy_path_id = 0;
2460 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2467 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2468 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2470 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2472 pset_foreach(copies, ptr) {
2473 copy = PTR_TO_INT(ptr);
2474 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2479 * @parameter copies contains a path of copies which lead us to irn
2480 * @parameter visited contains a set of nodes already visited on this path
2483 find_copy_path(spill_ilp_t * si, ir_node * irn, ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2486 op_t *op = get_irn_link(irn);
2488 if(op->is_remat) return;
2490 pset_insert_ptr(visited, irn);
2495 /* visit all operands */
2496 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2497 ir_node *arg = get_irn_n(irn, n);
2498 ilp_var_t copy = op->attr.live_range.args.copies[n];
2500 if(!has_reg_class(si, arg)) continue;
2503 pset_insert(copies, INT_TO_PTR(copy), copy);
2504 write_copy_path_cst(si, copies, any_interfere);
2505 pset_remove(copies, INT_TO_PTR(copy), copy);
2507 if(!pset_find_ptr(visited, arg)) {
2508 pset_insert(copies, INT_TO_PTR(copy), copy);
2509 find_copy_path(si, arg, target, any_interfere, copies, visited);
2510 pset_remove(copies, INT_TO_PTR(copy), copy);
2516 /* visit all uses which are phis */
2517 foreach_out_edge(irn, edge) {
2518 ir_node *user = edge->src;
2519 int pos = edge->pos;
2520 op_t *op = get_irn_link(user);
2523 if(!is_Phi(user)) continue;
2524 if(!has_reg_class(si, user)) continue;
2526 copy = op->attr.live_range.args.copies[pos];
2528 if(user == target) {
2529 pset_insert(copies, INT_TO_PTR(copy), copy);
2530 write_copy_path_cst(si, copies, any_interfere);
2531 pset_remove(copies, INT_TO_PTR(copy), copy);
2533 if(!pset_find_ptr(visited, user)) {
2534 pset_insert(copies, INT_TO_PTR(copy), copy);
2535 find_copy_path(si, user, target, any_interfere, copies, visited);
2536 pset_remove(copies, INT_TO_PTR(copy), copy);
2541 pset_remove_ptr(visited, irn);
2545 gen_copy_constraints(spill_ilp_t * si, ir_node * a, ir_node * b, ilp_var_t any_interfere)
2547 pset * copies = pset_new_ptr_default();
2548 pset * visited = pset_new_ptr_default();
2550 find_copy_path(si, a, b, any_interfere, copies, visited);
2558 memcopyhandler(spill_ilp_t * si)
2560 interference_t *interference;
2562 /* teste Speicherwerte auf Interferenz */
2564 /* analyze phi classes */
2565 phi_class_compute(si->chordal_env->irg);
2567 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
2568 irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
2570 // phi_class_free(si->chordal_env->irg);
2572 /* now lets emit the ILP unequations for the crap */
2573 set_foreach(si->interferences, interference) {
2575 ilp_var_t interfere,
2577 ilp_cst_t any_interfere_cst,
2579 const ir_node *a = interference->a;
2580 const ir_node *b = interference->b;
2582 /* any_interf <= \sum interf */
2583 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
2584 any_interfere_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2585 any_interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2587 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
2589 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
2590 const ir_node *bb = irnlist->irn;
2591 spill_bb_t *spill_bb = get_irn_link(bb);
2598 spilla = set_find_spill(spill_bb->ilp, a);
2602 spillb = set_find_spill(spill_bb->ilp, b);
2605 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
2606 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
2607 /* 2: - mem_in_a - spill_a + interfere <= 0 */
2608 /* 3: - mem_in_b - spill_b + interfere <= 0 */
2609 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
2610 interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2612 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
2613 cst = lpp_add_cst(si->lpp, buf, lpp_less, 1);
2615 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
2616 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
2617 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
2618 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
2619 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
2621 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
2622 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2624 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2625 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
2626 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
2628 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
2629 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2631 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2632 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
2633 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
2636 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
2638 /* any_interfere >= interf */
2639 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
2640 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2642 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2643 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2646 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
2647 gen_copy_constraints(si,a,b,any_interfere);
2654 memcopyinsertor(spill_ilp_t * si)
2656 /* weise Spillkontexte zu. Sorge bei Phis dafuer, dass gleiche
2657 * Kontexte zusammenfliessen (Operanden und Ergebnis hat gleichen
2673 return fabs(x) < 0.00001;
2677 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
2679 spill_ilp_t *si = get_irg_link(current_ir_graph);
2681 if(pset_find_ptr(si->all_possible_remats, n)) {
2682 op_t *op = (op_t*)get_irn_link(n);
2683 assert(op && op->is_remat);
2685 if(!op->attr.remat.remat->inverse) {
2686 if(op->attr.remat.pre) {
2687 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
2689 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
2694 op_t *op = (op_t*)get_irn_link(n);
2695 assert(op && op->is_remat);
2697 if(op->attr.remat.pre) {
2698 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
2700 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
2711 dump_graph_with_remats(ir_graph * irg, const char * suffix)
2713 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
2714 be_dump(irg, suffix, dump_ir_block_graph_sched);
2715 set_dump_node_vcgattr_hook(NULL);
2720 * Edge hook to dump the schedule edges with annotated register pressure.
2723 sched_pressure_edge_hook(FILE *F, ir_node *irn)
2725 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
2726 ir_node *prev = sched_prev(irn);
2727 fprintf(F, "edge:{sourcename:\"");
2729 fprintf(F, "\" targetname:\"");
2731 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
2732 fprintf(F, "\" color:magenta}\n");
2738 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
2740 DUMP_NODE_EDGE_FUNC old = get_dump_node_edge_hook();
2742 dump_consts_local(0);
2743 set_dump_node_edge_hook(sched_pressure_edge_hook);
2744 dump_ir_block_graph(irg, suffix);
2745 set_dump_node_edge_hook(old);
2749 walker_pressure_annotator(ir_node * bb, void * data)
2751 spill_ilp_t *si = data;
2755 pset *live = pset_new_ptr_default();
2758 live_foreach(bb, li) {
2759 irn = (ir_node *) li->irn;
2761 if (live_is_end(li) && has_reg_class(si, irn)) {
2762 pset_insert_ptr(live, irn);
2766 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
2768 sched_foreach_reverse(bb, irn) {
2770 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
2774 if(has_reg_class(si, irn)) {
2775 pset_remove_ptr(live, irn);
2776 if(is_Proj(irn)) ++projs;
2779 if(!is_Proj(irn)) projs = 0;
2781 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2782 ir_node *arg = get_irn_n(irn, n);
2784 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
2786 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
2793 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
2795 be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
2800 connect_all_remats_with_keep(spill_ilp_t * si)
2808 n_remats = pset_count(si->all_possible_remats);
2810 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
2813 pset_foreach(si->all_possible_remats, irn) {
2818 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
2820 obstack_free(si->obst, ins);
2826 connect_all_spills_with_keep(spill_ilp_t * si)
2835 n_spills = pset_count(si->spills);
2837 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
2840 pset_foreach(si->spills, irn) {
2845 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
2847 obstack_free(si->obst, ins);
2851 /** insert a spill at an arbitrary position */
2852 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert, ir_node *ctx)
2854 ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
2855 ir_graph *irg = get_irn_irg(bl);
2856 ir_node *frame = get_irg_frame(irg);
2860 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
2861 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
2863 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn, ctx);
2866 * search the right insertion point. a spill of a phi cannot be put
2867 * directly after the phi, if there are some phis behind the one which
2868 * is spilled. Also, a spill of a Proj must be after all Projs of the
2871 * Here's one special case:
2872 * If the spill is in the start block, the spill must be after the frame
2873 * pointer is set up. This is done by setting insert to the end of the block
2874 * which is its default initialization (see above).
2877 if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
2880 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
2883 sched_add_after(insert, spill);
2888 delete_remat(spill_ilp_t * si, ir_node * remat) {
2890 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2892 sched_remove(remat);
2894 /* kill links to operands */
2895 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
2896 set_irn_n(remat, n, bad);
2901 clean_remat_info(spill_ilp_t * si)
2905 remat_info_t *remat_info;
2906 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2908 set_foreach(si->remat_info, remat_info) {
2909 if(!remat_info->remats) continue;
2911 pset_foreach(remat_info->remats, remat)
2913 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
2914 set_irn_n(remat->proj, -1, bad);
2915 set_irn_n(remat->proj, 0, bad);
2918 if(get_irn_n_edges(remat->op) == 0) {
2919 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
2920 set_irn_n(remat->op, n, bad);
2925 if(remat_info->remats) del_pset(remat_info->remats);
2926 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
2931 delete_unnecessary_remats(spill_ilp_t * si)
2935 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2938 ir_node *end = get_irg_end(si->chordal_env->irg);
2941 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
2942 ir_node *keep_arg = get_irn_n(si->keep, n);
2943 op_t *arg_op = get_irn_link(keep_arg);
2946 assert(arg_op->is_remat);
2948 name = si->lpp->vars[arg_op->attr.remat.ilp];
2950 if(is_zero(name->value)) {
2951 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
2952 /* TODO check whether reload is preferred over remat (could be bug) */
2953 delete_remat(si, keep_arg);
2955 if(!arg_op->attr.remat.remat->inverse) {
2956 if(arg_op->attr.remat.pre) {
2957 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
2959 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
2962 if(arg_op->attr.remat.pre) {
2963 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
2965 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
2970 set_irn_n(si->keep, n, bad);
2973 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
2974 ir_node *end_arg = get_End_keepalive(end, i);
2976 if(end_arg != si->keep) {
2977 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
2980 keeps = obstack_finish(si->obst);
2981 set_End_keepalives(end, n-1, keeps);
2982 obstack_free(si->obst, keeps);
2985 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
2990 pset_foreach(si->all_possible_remats, remat) {
2991 op_t *remat_op = get_irn_link(remat);
2992 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
2994 if(is_zero(name->value)) {
2995 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
2996 /* TODO check whether reload is preferred over remat (could be bug) */
2997 delete_remat(si, remat);
2999 if(!remat_op->attr.remat.remat->inverse) {
3000 if(remat_op->attr.remat.pre) {
3001 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3003 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3006 if(remat_op->attr.remat.pre) {
3007 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3009 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3018 get_spills_for_value(spill_ilp_t * si, ir_node * value)
3020 pset *spills = pset_new_ptr_default();
3021 // pset *visited = pset_new_ptr_default();
3023 // collect_spills(si, value, spills, visited);
3024 // del_pset(visited);
3028 defs = set_find_def(si->values, value);
3030 if(defs && defs->spills) {
3031 for(next = defs->spills; next; next = get_irn_link(next)) {
3032 pset_insert_ptr(spills, next);
3040 get_remats_for_value(spill_ilp_t * si, ir_node * value)
3042 pset *remats = pset_new_ptr_default();
3047 defs = set_find_def(si->values, value);
3049 if(defs && defs->remats) {
3050 for(next = defs->remats; next; next = get_irn_link(next)) {
3051 pset_insert_ptr(remats, next);
3060 * @param before The node after which the spill will be placed in the schedule
3062 /* TODO set context properly */
3064 insert_spill(spill_ilp_t * si, ir_node * irn, ir_node * value, ir_node * before)
3068 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3070 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3072 spill = be_spill2(arch_env, irn, before, irn);
3074 defs = set_insert_def(si->values, value);
3077 /* enter into the linked list */
3078 set_irn_link(spill, defs->spills);
3079 defs->spills = spill;
3081 #ifdef KEEPALIVE_SPILLS
3082 pset_insert_ptr(si->spills, spill);
3089 insert_mem_copy(spill_ilp_t * si, const ir_node * bb, const ir_node * arg)
3091 ir_node *prev = sched_block_last_noncf(si, bb);
3092 ir_node *insert_pos = sched_next(prev);
3093 op_t *prev_op = get_irn_link(prev);
3094 pset *remats = get_remats_for_value(si, arg);
3096 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3098 /* start from end of block and search a position for memcopy (spill) until block's last op begins */
3099 while(be_is_Spill(prev)) {
3100 prev = sched_prev(prev);
3103 prev_op = get_irn_link(prev);
3105 while(!sched_is_end(prev) && !is_Phi(prev)
3106 && prev_op->is_remat && prev_op->attr.remat.pre) {
3110 if(pset_find_ptr(remats, prev)) {
3111 insert_pos = sched_next(insert_pos);
3116 prev = sched_prev(prev);
3117 } while(be_is_Spill(prev));
3119 prev_op = get_irn_link(prev);
3121 insert_pos = sched_prev(insert_pos);
3123 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", arg, insert_pos));
3125 spill = be_spill2(arch_env, arg, insert_pos, arg);
3133 * @param before The Phi node which has to be spilled
3136 insert_mem_phi(spill_ilp_t * si, const ir_node * phi)
3142 op_t *op = get_irn_link(phi);
3144 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3146 #ifndef NO_MEMCOPIES
3147 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3148 ir_node *arg = get_irn_n(phi, n);
3149 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(phi), n);
3150 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3152 if(!is_zero(name->value)) {
3153 ins[n] = insert_mem_copy(si, bb, arg);
3155 ins[n] = si->m_unknown;
3159 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3160 ins[n] = si->m_unknown;
3164 mem_phi = new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3166 defs = set_insert_def(si->values, phi);
3169 /* enter into the linked list */
3170 set_irn_link(mem_phi, defs->spills);
3171 defs->spills = mem_phi;
3173 sched_add_after(phi, mem_phi);
3175 #ifdef KEEPALIVE_SPILLS
3176 pset_insert_ptr(si->spills, mem_phi);
3184 * Add remat to list of defs, destroys link field!
3187 insert_remat(spill_ilp_t * si, ir_node * remat)
3190 op_t *remat_op = get_irn_link(remat);
3192 assert(remat_op->is_remat);
3194 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3197 /* enter into the linked list */
3198 set_irn_link(remat, defs->remats);
3199 defs->remats = remat;
3204 collect_spills(spill_ilp_t * si, ir_node * value, pset * spills, pset * visited)
3209 defs = set_find_def(si->values, value);
3211 if(defs && defs->spills) {
3212 for(next = defs->spills; next; next = get_irn_link(next)) {
3213 pset_insert_ptr(spills, next);
3215 } else if (is_Phi(value)) {
3217 if(!pset_find_ptr(visited, value)) {
3221 pset_insert_ptr(visited, value);
3222 for(i=0, n=get_irn_arity(value); i<n; ++i) {
3223 ir_node *arg = get_irn_n(value, i);
3225 collect_spills(si, arg, spills, visited);
3229 // assert(0 && "Phi operand not spilled");
3235 * Add reload before operation and add to list of defs
3238 insert_reload(spill_ilp_t * si, const ir_node * value, const ir_node * after)
3243 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3245 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3247 defs = set_find_def(si->values, value);
3248 /* get a spill of this value */
3250 if((!defs || !defs->spills) && is_Phi(value)) {
3253 spills = get_spills_for_value(si, value);
3255 spill = pset_first(spills);
3259 defs = set_insert_def(si->values, value);
3261 defs->spills = spill;
3262 set_irn_link(spill, NULL);
3264 spill = defs->spills;
3267 spill = defs->spills;
3268 assert(spill && "no spill placed before reload");
3270 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3272 /* enter into the linked list */
3273 set_irn_link(reload, defs->remats);
3274 defs->remats = reload;
3280 walker_spill_placer(ir_node * bb, void * data) {
3281 spill_ilp_t *si = (spill_ilp_t*)data;
3283 spill_bb_t *spill_bb = get_irn_link(bb);
3284 pset *spills_to_do = pset_new_ptr_default();
3287 set_foreach(spill_bb->ilp, spill) {
3290 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3291 name = si->lpp->vars[spill->mem_in];
3292 if(!is_zero(name->value)) {
3295 mem_phi = insert_mem_phi(si, spill->irn);
3297 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3301 name = si->lpp->vars[spill->spill];
3302 if(!is_zero(name->value)) {
3303 /* place spill directly after definition */
3304 if(get_nodes_block(spill->irn) == bb) {
3305 insert_spill(si, spill->irn, spill->irn, spill->irn);
3309 /* place spill at bb start */
3310 if(spill->reg_in > 0) {
3311 name = si->lpp->vars[spill->reg_in];
3312 if(!is_zero(name->value)) {
3313 insert_spill(si, spill->irn, spill->irn, bb);
3317 /* place spill after a remat */
3318 pset_insert_ptr(spills_to_do, spill->irn);
3321 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3324 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3325 op_t *op = get_irn_link(irn);
3327 if(be_is_Spill(irn)) continue;
3330 /* TODO fix this if we want to support remats with more than two nodes */
3331 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3332 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3334 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3337 if(pset_find_ptr(spills_to_do, irn)) {
3338 pset_remove_ptr(spills_to_do, irn);
3340 insert_spill(si, irn, irn, irn);
3346 assert(pset_count(spills_to_do) == 0);
3348 /* afterwards free data in block */
3349 del_pset(spills_to_do);
3354 phim_fixer(spill_ilp_t *si) {
3357 set_foreach(si->values, defs) {
3358 const ir_node *phi = defs->value;
3359 ir_node *phi_m = NULL;
3360 ir_node *next = defs->spills;
3363 if(!is_Phi(phi)) continue;
3366 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3370 next = get_irn_link(next);
3373 if(!phi_m) continue;
3375 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3376 const ir_node *value = get_irn_n(phi, n);
3377 defs_t *val_defs = set_find_def(si->values, value);
3378 ir_node *arg = get_irn_n(phi_m, n);
3380 /* get a spill of this value */
3381 ir_node *spill = val_defs->spills;
3383 assert(spill && "no spill placed before PhiM");
3386 set_irn_n(phi_m, n, spill);
3392 walker_reload_placer(ir_node * bb, void * data) {
3393 spill_ilp_t *si = (spill_ilp_t*)data;
3395 spill_bb_t *spill_bb = get_irn_link(bb);
3399 /* reloads at end of block */
3400 if(spill_bb->reloads) {
3403 set_foreach(spill_bb->reloads, keyval) {
3404 ir_node *irn = (ir_node*)keyval->key;
3405 ilp_var_t reload = PTR_TO_INT(keyval->val);
3408 name = si->lpp->vars[reload];
3409 if(!is_zero(name->value)) {
3411 ir_node *insert_pos = bb;
3412 ir_node *prev = sched_block_last_noncf(si, bb);
3413 op_t *prev_op = get_irn_link(prev);
3415 while(be_is_Spill(prev)) {
3416 prev = sched_prev(prev);
3419 prev_op = get_irn_link(prev);
3421 /* insert reload before pre-remats */
3422 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3423 && prev_op->is_remat && prev_op->attr.remat.pre) {
3427 prev = sched_prev(prev);
3428 } while(be_is_Spill(prev));
3430 prev_op = get_irn_link(prev);
3433 // /* insert reload before pre-remats */
3434 // while(!sched_is_end(prev) && !be_is_Reload(prev) //FIXME && !be_is_Spill(prev)
3435 // && !is_Phi(prev) && prev_op->is_remat && prev_op->attr.remat.pre) {
3436 // insert_pos = prev;
3438 // prev = sched_prev(insert_pos);
3439 // prev_op = get_irn_link(prev);
3442 reload = insert_reload(si, irn, insert_pos);
3444 #ifdef KEEPALIVE_RELOADS
3445 pset_insert_ptr(si->spills, reload);
3451 /* walk and insert more reloads and collect remats */
3452 sched_foreach_reverse(bb, irn) {
3453 op_t *op = get_irn_link(irn);
3455 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3456 if(is_Phi(irn)) break;
3459 if(get_irn_mode(irn) != mode_T) {
3460 insert_remat(si, irn);
3465 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3466 ir_node *arg = get_irn_n(irn, n);
3468 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3471 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3472 if(!is_zero(name->value)) {
3474 ir_node *insert_pos = irn;
3475 ir_node *prev = sched_prev(insert_pos);
3478 while(be_is_Spill(prev)) {
3479 prev = sched_prev(prev);
3482 prev_op = get_irn_link(prev);
3484 /* insert reload before pre-remats */
3485 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3486 && prev_op->is_remat && prev_op->attr.remat.pre) {
3490 prev = sched_prev(prev);
3491 } while(be_is_Spill(prev));
3493 prev_op = get_irn_link(prev);
3497 reload = insert_reload(si, arg, insert_pos);
3499 set_irn_n(irn, n, reload);
3501 #ifdef KEEPALIVE_RELOADS
3502 pset_insert_ptr(si->spills, reload);
3510 del_set(spill_bb->ilp);
3511 if(spill_bb->reloads) del_set(spill_bb->reloads);
3515 walker_collect_used(ir_node * irn, void * data)
3517 lc_bitset_t *used = data;
3519 lc_bitset_set(used, get_irn_idx(irn));
3522 struct kill_helper {
3528 walker_kill_unused(ir_node * bb, void * data)
3530 struct kill_helper *kh = data;
3531 const ir_node *bad = get_irg_bad(get_irn_irg(bb));
3535 for(irn=sched_first(bb); !sched_is_end(irn);) {
3536 ir_node *next = sched_next(irn);
3539 if(!lc_bitset_is_set(kh->used, get_irn_idx(irn))) {
3540 if(be_is_Spill(irn) || be_is_Reload(irn)) {
3541 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)));
3543 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3549 set_nodes_block(irn, bad);
3550 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3551 set_irn_n(irn, n, bad);
3559 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3561 struct kill_helper kh;
3563 kh.used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3566 irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3567 irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3569 lc_bitset_free(kh.used);
3573 print_irn_pset(pset * p)
3577 pset_foreach(p, irn) {
3578 ir_printf("%+F\n", irn);
3583 rewire_uses(spill_ilp_t * si)
3585 dom_front_info_t *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
3587 pset *ignore = pset_new_ptr(1);
3589 pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
3591 /* then fix uses of spills */
3592 set_foreach(si->values, defs) {
3595 ir_node *next = defs->remats;
3598 reloads = pset_new_ptr_default();
3601 if(be_is_Reload(next)) {
3602 pset_insert_ptr(reloads, next);
3606 next = get_irn_link(next);
3609 spills = get_spills_for_value(si, defs->value);
3610 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));
3611 if(pset_count(spills) > 1) {
3612 //assert(pset_count(reloads) > 0);
3613 // print_irn_pset(spills);
3614 // print_irn_pset(reloads);
3616 be_ssa_constr_set_ignore(dfi, spills, ignore);
3623 /* first fix uses of remats and reloads */
3624 set_foreach(si->values, defs) {
3626 ir_node *next = defs->remats;
3629 nodes = pset_new_ptr_default();
3630 pset_insert_ptr(nodes, defs->value);
3633 pset_insert_ptr(nodes, next);
3634 next = get_irn_link(next);
3637 if(pset_count(nodes) > 1) {
3638 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
3639 be_ssa_constr_set(dfi, nodes);
3646 // remove_unused_defs(si);
3648 be_free_dominance_frontiers(dfi);
3653 writeback_results(spill_ilp_t * si)
3655 /* walk through the graph and collect all spills, reloads and remats for a value */
3657 si->values = new_set(cmp_defs, 4096);
3659 DBG((si->dbg, LEVEL_1, "Applying results\n"));
3660 delete_unnecessary_remats(si);
3661 si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
3662 irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
3664 irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
3666 /* clean the remat info! there are still back-edges leading there! */
3667 clean_remat_info(si);
3671 connect_all_spills_with_keep(si);
3673 del_set(si->values);
3677 get_n_regs(spill_ilp_t * si)
3679 int arch_n_regs = arch_register_class_n_regs(si->cls);
3683 for(i=0; i<arch_n_regs; i++) {
3684 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
3689 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
3694 walker_reload_mover(ir_node * bb, void * data)
3696 spill_ilp_t *si = data;
3699 sched_foreach(bb, tmp) {
3700 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
3701 ir_node *reload = tmp;
3704 /* move reload upwards */
3706 int pressure = (int)get_irn_link(reload);
3707 if(pressure < si->n_regs) {
3708 irn = sched_prev(reload);
3709 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
3710 sched_remove(reload);
3711 pressure = (int)get_irn_link(irn);
3713 while(pressure < si->n_regs) {
3714 if(sched_is_end(irn) || (be_is_Reload(irn) && has_reg_class(si, irn))) break;
3716 set_irn_link(irn, INT_TO_PTR(pressure+1));
3717 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
3718 irn = sched_prev(irn);
3720 pressure = (int)get_irn_link(irn);
3723 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
3724 sched_put_after(irn, reload);
3731 move_reloads_upward(spill_ilp_t * si)
3733 irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
3738 * Walk all irg blocks and check for interfering spills inside of phi classes
3741 luke_meminterferencechecker(ir_node * bb, void * data)
3743 spill_ilp_t *si = (spill_ilp_t*)data;
3747 live_foreach(bb, li1) {
3748 ir_node *a = (ir_node *) li1->irn;
3750 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
3752 /* a is only interesting if it is inside a phi class */
3753 if (get_phi_class(a)) {
3754 for(li2=li1->next; li2; li2 = li2->next) {
3755 ir_node *b = (ir_node *) li2->irn;
3757 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
3759 /* a and b are only interesting if they are in the same phi class */
3760 if(get_phi_class(a) == get_phi_class(b)) {
3761 if(values_interfere_in_block(bb, a, b)) {
3762 ir_fprintf(stderr, "Spills interfere in %+F: %+F, %+F\n", bb, a, b);
3771 verify_phiclasses(spill_ilp_t * si)
3773 /* analyze phi classes */
3774 phi_class_compute(si->chordal_env->irg);
3776 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
3777 irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
3781 walker_spillslotassigner(ir_node * irn, void * data)
3783 spill_ilp_t *si = (spill_ilp_t*)data;
3786 if(!be_is_Spill(irn)) return;
3788 /* set spill context to phi class if it has one ;) */
3790 cls = get_phi_class(irn);
3792 be_set_Spill_context(irn, cls);
3794 be_set_Spill_context(irn, irn);
3799 assign_spillslots(spill_ilp_t * si)
3801 DBG((si->dbg, LEVEL_2, "\t calling spill slot assigner\n"));
3802 irg_walk_graph(si->chordal_env->irg, walker_spillslotassigner, NULL, si);
3806 be_spill_remat(const be_chordal_env_t * chordal_env)
3808 char problem_name[256];
3809 char dump_suffix[256];
3810 char dump_suffix2[256];
3811 char dump_suffix3[256];
3812 struct obstack obst;
3815 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
3816 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
3817 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
3819 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
3820 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
3822 obstack_init(&obst);
3823 si.chordal_env = chordal_env;
3825 si.cls = chordal_env->cls;
3826 si.lpp = new_lpp(problem_name, lpp_minimize);
3827 si.remat_info = new_set(cmp_remat_info, 4096);
3828 si.interferences = new_set(cmp_interference, 32);
3829 si.all_possible_remats = pset_new_ptr_default();
3830 si.spills = pset_new_ptr_default();
3831 si.inverse_ops = pset_new_ptr_default();
3832 #ifndef EXECFREQ_LOOPDEPH
3833 si.execfreqs = compute_execfreq(chordal_env->irg);
3835 si.execfreqs = NULL;
3840 si.n_regs = get_n_regs(&si);
3842 set_irg_link(chordal_env->irg, &si);
3843 compute_doms(chordal_env->irg);
3845 /* compute phi classes */
3846 // phi_class_compute(chordal_env->irg);
3848 be_analyze_regpressure(chordal_env, "-pre");
3850 #ifdef COLLECT_REMATS
3851 /* collect remats */
3852 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
3853 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
3856 /* insert possible remats */
3857 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
3858 irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
3859 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
3862 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
3863 connect_all_remats_with_keep(&si);
3864 /* dump graph with inserted remats */
3865 dump_graph_with_remats(chordal_env->irg, dump_suffix);
3868 /* insert copies for phi arguments not in my regclass */
3869 irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
3871 /* recompute liveness */
3872 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
3873 be_liveness(chordal_env->irg);
3877 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
3878 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
3879 irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
3881 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
3882 irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
3884 #ifndef NO_MEMCOPIES
3885 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
3886 memcopyhandler(&si);
3894 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
3895 if ((f = fopen(buf, "wt")) != NULL) {
3896 lpp_dump_plain(si.lpp, f);
3903 DBG((si.dbg, LEVEL_1, "\tSolving %F\n", chordal_env->irg));
3905 lpp_set_time_limit(si.lpp, ILP_TIMEOUT);
3909 lpp_solve_cplex(si.lpp);
3911 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
3913 assert(lpp_is_sol_valid(si.lpp)
3914 && "solution of ILP must be valid");
3916 DBG((si.dbg, LEVEL_1, "\t%s: iterations: %d, solution time: %g, objective function: %g\n", problem_name, si.lpp->iterations, si.lpp->sol_time, is_zero(si.lpp->objval)?0.0:si.lpp->objval));
3918 #ifdef DUMP_SOLUTION
3923 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
3924 if ((f = fopen(buf, "wt")) != NULL) {
3926 for (i = 0; i < si.lpp->var_next; ++i) {
3927 lpp_name_t *name = si.lpp->vars[i];
3928 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
3935 writeback_results(&si);
3939 kill_all_unused_values_in_schedule(&si);
3941 #if defined(KEEPALIVE_SPILLS) || defined(KEEPALIVE_RELOADS)
3942 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
3945 // move reloads upwards
3946 be_liveness(chordal_env->irg);
3947 //irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
3948 //move_reloads_upward(&si);
3950 #ifndef NO_MEMCOPIES
3951 verify_phiclasses(&si);
3952 assign_spillslots(&si);
3955 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
3957 dump_pressure_graph(&si, dump_suffix2);
3959 be_analyze_regpressure(chordal_env, "-post");
3961 free_dom(chordal_env->irg);
3962 del_set(si.interferences);
3963 del_pset(si.inverse_ops);
3964 del_pset(si.all_possible_remats);
3965 del_pset(si.spills);
3966 #ifndef EXECFREQ_LOOPDEPH
3967 free_execfreq(si.execfreqs);
3970 obstack_free(&obst, NULL);
3971 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
3974 #else /* WITH_ILP */
3977 only_that_you_can_compile_without_WITH_ILP_defined(void)
3981 #endif /* WITH_ILP */