1 /** vim: set sw=4 ts=4:
4 * @author Adam M. Szalkowski & Sebastian Hack
6 * ILP based spilling & rematerialization
8 * Copyright (C) 2006 Universitaet Karlsruhe
9 * Released under the GPL
32 #include "phiclass_t.h"
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
70 //#define NO_MEMCOPIES
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);
722 _set_phi_class(copy, NULL);
723 set_nodes_block(copy, bb);
724 sched_put_before(si, pos, copy);
730 * Inserts a copy of @p irn after @p pos
733 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
738 bb = is_Block(pos)?pos:get_nodes_block(pos);
739 copy = exact_copy(irn);
741 _set_phi_class(copy, NULL);
742 set_nodes_block(copy, bb);
743 sched_put_after(pos, copy);
749 insert_remat_after(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
753 if(can_remat_after(si, remat, pos, live)) {
758 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
760 copy = insert_copy_after(si, remat->op, pos);
762 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
763 op = obstack_alloc(si->obst, sizeof(*op));
765 op->attr.remat.remat = remat;
766 op->attr.remat.pre = 0;
767 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
769 set_irn_link(copy, op);
770 pset_insert_ptr(si->all_possible_remats, copy);
772 proj_copy = insert_copy_after(si, remat->proj, copy);
773 set_irn_n(proj_copy, 0, copy);
774 set_irn_link(proj_copy, op);
775 pset_insert_ptr(si->all_possible_remats, proj_copy);
787 insert_remat_before(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
791 if(can_remat_before(si, remat, pos, live)) {
796 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
798 copy = insert_copy_before(si, remat->op, pos);
800 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
801 op = obstack_alloc(si->obst, sizeof(*op));
803 op->attr.remat.remat = remat;
804 op->attr.remat.pre = 1;
805 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
807 set_irn_link(copy, op);
808 pset_insert_ptr(si->all_possible_remats, copy);
810 proj_copy = insert_copy_after(si, remat->proj, copy);
811 set_irn_n(proj_copy, 0, copy);
812 set_irn_link(proj_copy, op);
813 pset_insert_ptr(si->all_possible_remats, proj_copy);
825 get_block_n_succs(const ir_node *block) {
826 const ir_edge_t *edge;
828 assert(edges_activated(current_ir_graph));
830 edge = get_block_succ_first(block);
834 edge = get_block_succ_next(block, edge);
839 is_merge_edge(const ir_node * bb)
841 #ifdef GOODWIN_REDUCTION
842 return get_block_n_succs(bb) == 1;
849 is_diverge_edge(const ir_node * bb)
851 #ifdef GOODWIN_REDUCTION
852 return get_Block_n_cfgpreds(bb) == 1;
859 walker_regclass_copy_insertor(ir_node * irn, void * data)
861 spill_ilp_t *si = data;
863 if(is_Phi(irn) && has_reg_class(si, irn)) {
866 for(n=get_irn_arity(irn)-1; n>=0; --n) {
867 ir_node *phi_arg = get_irn_n(irn, n);
868 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
870 if(!has_reg_class(si, phi_arg)) {
871 ir_node *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
872 ir_node *pos = sched_block_last_noncf(si, bb);
873 op_t *op = obstack_alloc(si->obst, sizeof(*op));
875 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
876 sched_add_after(pos, copy);
877 set_irn_n(irn, n, copy);
880 op->attr.live_range.args.reloads = NULL;
881 op->attr.live_range.ilp = ILP_UNDEF;
882 set_irn_link(copy, op);
890 * Insert (so far unused) remats into the irg to
891 * recompute the potential liveness of all values
894 walker_remat_insertor(ir_node * bb, void * data)
896 spill_ilp_t *si = data;
897 spill_bb_t *spill_bb;
901 pset *live = pset_new_ptr_default();
903 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
905 live_foreach(bb, li) {
906 ir_node *value = (ir_node *) li->irn;
908 /* add remats at end of block */
909 if (live_is_end(li) && has_reg_class(si, value)) {
910 pset_insert_ptr(live, value);
914 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
915 set_irn_link(bb, spill_bb);
917 irn = sched_last(bb);
918 while(!sched_is_end(irn)) {
925 next = sched_prev(irn);
927 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
929 if(is_Phi(irn) || is_Proj(irn)) {
932 if(has_reg_class(si, irn)) {
933 pset_remove_ptr(live, irn);
936 op = obstack_alloc(si->obst, sizeof(*op));
938 op->attr.live_range.args.reloads = NULL;
939 op->attr.live_range.ilp = ILP_UNDEF;
940 set_irn_link(irn, op);
946 op = obstack_alloc(si->obst, sizeof(*op));
948 op->attr.live_range.ilp = ILP_UNDEF;
949 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
950 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
951 set_irn_link(irn, op);
953 args = pset_new_ptr_default();
955 /* collect arguments of op */
956 for (n = get_irn_arity(irn)-1; n>=0; --n) {
957 ir_node *arg = get_irn_n(irn, n);
959 pset_insert_ptr(args, arg);
962 /* set args of op already live in epilog */
963 pset_foreach(args, arg) {
964 if(has_reg_class(si, arg)) {
965 pset_insert_ptr(live, arg);
968 /* delete defined value from live set */
969 if(has_reg_class(si, irn)) {
970 pset_remove_ptr(live, irn);
974 remat_args = pset_new_ptr_default();
976 /* insert all possible remats before irn */
977 pset_foreach(args, arg) {
978 remat_info_t *remat_info,
982 /* continue if the operand has the wrong reg class
984 if(!has_reg_class(si, arg))
989 query.remats_by_operand = NULL;
990 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
996 if(remat_info->remats) {
997 pset_foreach(remat_info->remats, remat) {
998 ir_node *remat_irn = NULL;
1000 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
1001 #ifdef REMAT_WHILE_LIVE
1002 if(pset_find_ptr(live, remat->value)) {
1003 remat_irn = insert_remat_before(si, remat, irn, live);
1006 remat_irn = insert_remat_before(si, remat, irn, live);
1009 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1010 ir_node *remat_arg = get_irn_n(remat_irn, n);
1012 if(!has_reg_class(si, remat_arg)) continue;
1014 pset_insert_ptr(remat_args, remat_arg);
1021 /* now we add remat args to op's args because they could also die at this op */
1022 pset_foreach(args,arg) {
1023 if(pset_find_ptr(remat_args, arg)) {
1024 pset_remove_ptr(remat_args, arg);
1027 pset_foreach(remat_args,arg) {
1028 pset_insert_ptr(args, arg);
1031 /* insert all possible remats after irn */
1032 pset_foreach(args, arg) {
1033 remat_info_t *remat_info,
1037 /* continue if the operand has the wrong reg class */
1038 if(!has_reg_class(si, arg))
1042 query.remats = NULL;
1043 query.remats_by_operand = NULL;
1044 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1050 /* do not place post remats after jumps */
1051 if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
1053 if(remat_info->remats_by_operand) {
1054 pset_foreach(remat_info->remats_by_operand, remat) {
1055 /* do not insert remats producing the same value as one of the operands */
1056 if(!pset_find_ptr(args, remat->value)) {
1057 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
1058 #ifdef REMAT_WHILE_LIVE
1059 if(pset_find_ptr(live, remat->value)) {
1060 insert_remat_after(si, remat, irn, live);
1063 insert_remat_after(si, remat, irn, live);
1070 del_pset(remat_args);
1075 live_foreach(bb, li) {
1076 ir_node *value = (ir_node *) li->irn;
1078 /* add remats at end if successor has multiple predecessors */
1079 if(is_merge_edge(bb)) {
1080 /* add remats at end of block */
1081 if (live_is_end(li) && has_reg_class(si, value)) {
1082 remat_info_t *remat_info,
1087 query.remats = NULL;
1088 query.remats_by_operand = NULL;
1089 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1091 if(remat_info && remat_info->remats) {
1092 pset_foreach(remat_info->remats, remat) {
1093 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1095 insert_remat_before(si, remat, bb, NULL);
1100 if(is_diverge_edge(bb)) {
1101 /* add remat2s at beginning of block */
1102 if ((live_is_in(li) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) {
1103 remat_info_t *remat_info,
1108 query.remats = NULL;
1109 query.remats_by_operand = NULL;
1110 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1112 if(remat_info && remat_info->remats) {
1113 pset_foreach(remat_info->remats, remat) {
1114 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at beginning of block %+F\n", remat->op, bb));
1116 /* put the remat here if all its args are available */
1117 insert_remat_after(si, remat, bb, NULL);
1127 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1130 luke_endwalker(ir_node * bb, void * data)
1132 spill_ilp_t *si = (spill_ilp_t*)data;
1139 spill_bb_t *spill_bb = get_irn_link(bb);
1142 live = pset_new_ptr_default();
1143 use_end = pset_new_ptr_default();
1145 live_foreach(bb, li) {
1146 irn = (ir_node *) li->irn;
1147 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1150 pset_insert_ptr(live, irn);
1151 op = get_irn_link(irn);
1152 assert(!op->is_remat);
1156 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1157 /* their reg_out must always be set */
1158 sched_foreach_reverse(bb, irn) {
1161 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1163 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1164 ir_node *irn_arg = get_irn_n(irn, n);
1166 if(has_reg_class(si, irn_arg)) {
1167 pset_insert_ptr(use_end, irn_arg);
1172 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1173 //cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1174 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1176 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1178 /* if this is a merge edge we can reload at the end of this block */
1179 if(is_merge_edge(bb)) {
1180 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1181 } else if(pset_count(use_end)){
1182 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1184 spill_bb->reloads = NULL;
1187 pset_foreach(live,irn) {
1193 /* handle values used by control flow nodes later separately */
1194 if(pset_find_ptr(use_end, irn)) continue;
1197 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1199 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1201 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1202 spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1203 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1205 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1206 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1208 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1209 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1211 if(is_merge_edge(bb)) {
1215 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1216 reload = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1217 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1219 /* reload <= mem_out */
1220 rel_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1221 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1222 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1225 spill->reg_in = ILP_UNDEF;
1226 spill->mem_in = ILP_UNDEF;
1229 pset_foreach(use_end,irn) {
1233 ilp_cst_t end_use_req,
1238 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1240 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1242 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1243 spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1244 /* if irn is used at the end of the block, then it is live anyway */
1245 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1247 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1248 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1250 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1251 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1253 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1254 reload = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1255 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1257 /* reload <= mem_out */
1258 rel_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1259 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1260 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1262 spill->reg_in = ILP_UNDEF;
1263 spill->mem_in = ILP_UNDEF;
1265 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1266 end_use_req = lpp_add_cst(si->lpp, buf, lpp_equal, 1);
1267 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1275 next_post_remat(const ir_node * irn)
1280 irn = sched_block_first_nonphi(irn);
1282 irn = sched_next_op(irn);
1285 if(sched_is_end(irn))
1288 op = (op_t*)get_irn_link(irn);
1289 if(op->is_remat && !op->attr.remat.pre) {
1298 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1304 ret = sched_block_last_noncf(si, irn);
1305 ret = sched_next(ret);
1306 ret = sched_prev_op(ret);
1308 ret = sched_prev_op(irn);
1311 if(sched_is_end(ret) || is_Phi(ret))
1314 op = (op_t*)get_irn_link(ret);
1315 if(op->is_remat && op->attr.remat.pre) {
1323 * Find a remat of value @p value in the epilog of @p pos
1326 find_post_remat(const ir_node * value, const ir_node * pos)
1328 while((pos = next_post_remat(pos)) != NULL) {
1331 op = get_irn_link(pos);
1332 assert(op->is_remat && !op->attr.remat.pre);
1334 if(op->attr.remat.remat->value == value)
1335 return (ir_node*)pos;
1338 const ir_edge_t *edge;
1339 foreach_out_edge(pos, edge) {
1340 ir_node *proj = get_edge_src_irn(edge);
1341 assert(is_Proj(proj));
1351 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1353 spill_bb_t *spill_bb = get_irn_link(bb);
1359 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1361 double spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1363 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1365 spill->reg_out = ILP_UNDEF;
1366 spill->reg_in = ILP_UNDEF;
1367 spill->mem_in = ILP_UNDEF;
1369 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1370 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1372 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1373 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1380 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1385 live_foreach(bb, li) {
1386 irn = (ir_node *) li->irn;
1388 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1389 pset_insert_ptr(live, irn);
1393 irn = sched_last(bb);
1395 /* all values eaten by control flow operations are also live until the end of the block */
1396 sched_foreach_reverse(bb, irn) {
1399 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1401 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1402 ir_node *arg = get_irn_n(irn,i);
1404 if(has_reg_class(si, arg)) {
1405 pset_insert_ptr(live, arg);
1412 * Inserts ILP-constraints and variables for memory copying before the given position
1415 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1417 const ir_node *succ;
1418 const ir_edge_t *edge;
1419 spill_bb_t *spill_bb = get_irn_link(block);
1428 assert(edges_activated(current_ir_graph));
1430 edge = get_block_succ_first(block);
1436 edge = get_block_succ_next(block, edge);
1437 /* next block can only contain phis, if this is a merge edge */
1440 ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1441 copyreg = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1443 ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1444 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1446 pset_foreach(live, tmp) {
1449 op_t *op = get_irn_link(irn);
1450 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1452 spill = set_find_spill(spill_bb->ilp, tmp);
1455 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1457 lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1459 sched_foreach(succ, phi) {
1460 const ir_node *to_copy;
1462 spill_t *to_copy_spill;
1463 op_t *phi_op = get_irn_link(phi);
1464 ilp_var_t reload = ILP_UNDEF;
1467 if(!is_Phi(phi)) break;
1468 if(!has_reg_class(si, phi)) continue;
1470 to_copy = get_irn_n(phi, pos);
1472 to_copy_op = get_irn_link(to_copy);
1474 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1475 assert(to_copy_spill);
1477 if(spill_bb->reloads) {
1478 keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1481 reload = PTR_TO_INT(keyval->val);
1485 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N", block, to_copy);
1486 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1488 /* copy - reg_out - reload - remat - live_range <= 0 */
1489 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1490 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1491 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1492 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1493 foreach_pre_remat(si, block, tmp) {
1494 op_t *remat_op = get_irn_link(tmp);
1495 if(remat_op->attr.remat.remat->value == to_copy) {
1496 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1500 ir_snprintf(buf, sizeof(buf), "copyreq_%N_%N", block, to_copy);
1501 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1503 /* copy - reg_out - copyreg <= 0 */
1504 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1505 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1506 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1512 * Walk all irg blocks and emit this ILP
1515 luke_blockwalker(ir_node * bb, void * data)
1517 spill_ilp_t *si = (spill_ilp_t*)data;
1522 spill_bb_t *spill_bb = get_irn_link(bb);
1525 pset *defs = pset_new_ptr_default();
1528 live = pset_new_ptr_default();
1530 /****************************************
1531 * B A S I C B L O C K E N D
1532 ***************************************/
1535 /* init live values at end of block */
1536 get_live_end(si, bb, live);
1538 pset_foreach(live, irn) {
1540 ilp_var_t reload = ILP_UNDEF;
1542 spill = set_find_spill(spill_bb->ilp, irn);
1545 if(spill_bb->reloads) {
1546 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1549 reload = PTR_TO_INT(keyval->val);
1553 op = get_irn_link(irn);
1554 assert(!op->is_remat);
1556 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1557 op->attr.live_range.ilp = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1558 op->attr.live_range.op = bb;
1560 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1561 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1563 /* reg_out - reload - remat - live_range <= 0 */
1564 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1565 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1566 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1567 foreach_pre_remat(si, bb, tmp) {
1568 op_t *remat_op = get_irn_link(tmp);
1569 if(remat_op->attr.remat.remat->value == irn) {
1570 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1573 /* maybe we should also assure that reg_out >= live_range etc. */
1576 #ifndef NO_MEMCOPIES
1577 insert_mem_copy_position(si, live, bb);
1581 * start new live ranges for values used by remats at end of block
1582 * and assure the remat args are available
1584 foreach_pre_remat(si, bb, tmp) {
1585 op_t *remat_op = get_irn_link(tmp);
1588 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1589 ir_node *remat_arg = get_irn_n(tmp, n);
1590 op_t *arg_op = get_irn_link(remat_arg);
1593 if(!has_reg_class(si, remat_arg)) continue;
1595 /* if value is becoming live through use by remat */
1596 if(!pset_find_ptr(live, remat_arg)) {
1597 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1598 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1600 arg_op->attr.live_range.ilp = prev_lr;
1601 arg_op->attr.live_range.op = bb;
1603 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1605 pset_insert_ptr(live, remat_arg);
1606 add_to_spill_bb(si, bb, remat_arg);
1609 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1610 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1611 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1613 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1614 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1616 /* use reload placed for this argument */
1617 if(spill_bb->reloads) {
1618 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1621 ilp_var_t reload = PTR_TO_INT(keyval->val);
1623 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1628 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1633 /**************************************
1634 * B A S I C B L O C K B O D Y
1635 **************************************/
1637 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1643 ilp_cst_t check_pre,
1650 /* iterate only until first phi */
1654 op = get_irn_link(irn);
1656 if(op->is_remat) continue;
1657 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1659 /* collect defined values */
1660 if(has_reg_class(si, irn)) {
1661 pset_insert_ptr(defs, irn);
1665 if(is_Proj(irn)) continue;
1668 * init set of irn's arguments
1669 * and all possibly used values around this op
1670 * and values defined by post remats
1672 args = new_set(cmp_keyval, get_irn_arity(irn));
1673 used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1674 remat_defs = pset_new_ptr(pset_count(live));
1676 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1677 ir_node *irn_arg = get_irn_n(irn, n);
1678 if(has_reg_class(si, irn_arg)) {
1679 set_insert_keyval(args, irn_arg, (void*)n);
1680 pset_insert_ptr(used, irn_arg);
1683 foreach_post_remat(irn, tmp) {
1684 op_t *remat_op = get_irn_link(tmp);
1686 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1688 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1689 ir_node *remat_arg = get_irn_n(tmp, n);
1690 if(has_reg_class(si, remat_arg)) {
1691 pset_insert_ptr(used, remat_arg);
1695 foreach_pre_remat(si, irn, tmp) {
1696 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1697 ir_node *remat_arg = get_irn_n(tmp, n);
1698 if(has_reg_class(si, remat_arg)) {
1699 pset_insert_ptr(used, remat_arg);
1704 /**********************************
1705 * I N E P I L O G O F irn
1706 **********************************/
1708 /* ensure each dying value is used by only one post remat */
1709 pset_foreach(live, tmp) {
1710 ir_node *value = tmp;
1711 op_t *value_op = get_irn_link(value);
1716 foreach_post_remat(irn, remat) {
1717 op_t *remat_op = get_irn_link(remat);
1719 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1720 ir_node *remat_arg = get_irn_n(remat, n);
1722 /* if value is used by this remat add it to constraint */
1723 if(remat_arg == value) {
1725 /* sum remat2s <= 1 + n_remats*live_range */
1726 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1727 cst = lpp_add_cst(si->lpp, buf, lpp_less, 1.0);
1731 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1737 if(value_op->attr.live_range.ilp != ILP_UNDEF && cst != ILP_UNDEF) {
1738 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1744 /* new live ranges for values from L\U defined by post remats */
1745 pset_foreach(live, tmp) {
1746 ir_node *value = tmp;
1747 op_t *value_op = get_irn_link(value);
1749 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
1750 ilp_var_t prev_lr = ILP_UNDEF;
1753 if(pset_find_ptr(remat_defs, value)) {
1755 /* next_live_range <= prev_live_range + sum remat2s */
1756 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1757 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1759 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1760 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1762 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1763 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1765 foreach_post_remat(irn, remat) {
1766 op_t *remat_op = get_irn_link(remat);
1768 /* if value is being rematerialized by this remat */
1769 if(value == remat_op->attr.remat.remat->value) {
1770 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1774 value_op->attr.live_range.ilp = prev_lr;
1775 value_op->attr.live_range.op = irn;
1780 /* requirements for post remats and start live ranges from L/U' for values dying here */
1781 foreach_post_remat(irn, tmp) {
1782 op_t *remat_op = get_irn_link(tmp);
1785 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1786 ir_node *remat_arg = get_irn_n(tmp, n);
1787 op_t *arg_op = get_irn_link(remat_arg);
1789 if(!has_reg_class(si, remat_arg)) continue;
1791 /* only for values in L\U (TODO and D?), the others are handled with post_use */
1792 if(!pset_find_ptr(used, remat_arg)) {
1793 /* remat <= live_rang(remat_arg) */
1794 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
1795 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1797 /* if value is becoming live through use by remat2 */
1798 if(!pset_find_ptr(live, remat_arg)) {
1801 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1802 lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1804 arg_op->attr.live_range.ilp = lr;
1805 arg_op->attr.live_range.op = irn;
1807 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
1809 pset_insert_ptr(live, remat_arg);
1810 add_to_spill_bb(si, bb, remat_arg);
1813 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1814 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1819 d = pset_count(defs);
1820 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
1822 /* count how many regs irn needs for arguments */
1823 u = set_count(args);
1826 /* check the register pressure in the epilog */
1827 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
1828 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
1829 check_post = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - d);
1831 /* add L\U' to check_post */
1832 pset_foreach(live, tmp) {
1833 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
1834 /* if a live value is not used by irn */
1835 tmp_op = get_irn_link(tmp);
1836 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
1840 /***********************************************************
1841 * 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
1842 **********************************************************/
1845 pset_foreach(used, tmp) {
1851 op_t *arg_op = get_irn_link(arg);
1854 spill = add_to_spill_bb(si, bb, arg);
1856 /* new live range for each used value */
1857 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
1858 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1860 /* the epilog stuff - including post_use, check_post, check_post_remat */
1861 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
1862 post_use = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1864 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
1866 /* arg is live throughout epilog if the next live_range is in a register */
1867 if(pset_find_ptr(live, arg)) {
1868 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
1870 /* post_use >= next_lr + remat */
1871 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1872 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1873 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
1874 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1878 /* if value is not an arg of op and not possibly defined by post remat
1879 * then it may only die and not become live
1881 if(!set_find_keyval(args, arg)) {
1882 /* post_use <= prev_lr */
1883 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
1884 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1885 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1886 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1888 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
1889 /* next_lr <= prev_lr */
1890 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
1891 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1892 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1893 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1899 /* forall post remat which use arg add a similar cst */
1900 foreach_post_remat(irn, remat) {
1903 for (n=get_irn_arity(remat)-1; n>=0; --n) {
1904 ir_node *remat_arg = get_irn_n(remat, n);
1905 op_t *remat_op = get_irn_link(remat);
1907 if(remat_arg == arg) {
1908 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
1910 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1911 cst = lpp_add_cst(si->lpp, buf, lpp_greater, 0.0);
1912 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1913 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1918 /* new live range begins for each used value */
1919 arg_op->attr.live_range.ilp = prev_lr;
1920 arg_op->attr.live_range.op = irn;
1922 /*if(!pset_find_ptr(live, arg)) {
1923 pset_insert_ptr(live, arg);
1924 add_to_spill_bb(si, bb, arg);
1926 pset_insert_ptr(live, arg);
1930 /* just to be sure */
1931 check_post = ILP_UNDEF;
1940 /* check the register pressure in the prolog */
1941 /* sum_{L\U} lr <= k - |U| */
1942 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
1943 check_pre = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - u);
1945 /* for the prolog remove defined values from the live set */
1946 pset_foreach(defs, tmp) {
1947 pset_remove_ptr(live, tmp);
1950 /***********************************************************
1951 * 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
1952 **********************************************************/
1955 set_foreach(args, keyval) {
1957 ir_node *arg = keyval->key;
1958 int i = PTR_TO_INT(keyval->val);
1959 op_t *arg_op = get_irn_link(arg);
1961 spill = set_find_spill(spill_bb->ilp, arg);
1964 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
1965 op->attr.live_range.args.reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1967 /* reload <= mem_out */
1968 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
1969 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1970 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
1971 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
1973 /* requirement: arg must be in register for use */
1974 /* reload + remat + live_range == 1 */
1975 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
1976 cst = lpp_add_cst(si->lpp, buf, lpp_equal, 1.0);
1978 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1979 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
1980 foreach_pre_remat(si, irn, tmp) {
1981 op_t *remat_op = get_irn_link(tmp);
1982 if(remat_op->attr.remat.remat->value == arg) {
1983 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1988 /* iterate over L\U */
1989 pset_foreach(live, tmp) {
1990 if(!set_find_keyval(args, tmp)) {
1991 /* if a live value is not used by irn */
1992 tmp_op = get_irn_link(tmp);
1993 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
1998 /* requirements for remats */
1999 /* start new live ranges for values used by remats */
2000 foreach_pre_remat(si, irn, tmp) {
2001 op_t *remat_op = get_irn_link(tmp);
2004 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2005 ir_node *remat_arg = get_irn_n(tmp, n);
2006 op_t *arg_op = get_irn_link(remat_arg);
2009 if(!has_reg_class(si, remat_arg)) continue;
2011 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2012 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2013 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2015 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2016 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2018 /* if remat arg is also used by current op then we can use reload placed for this argument */
2019 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2020 int index = (int)keyval->val;
2022 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2030 /*************************
2031 * D O N E W I T H O P
2032 *************************/
2034 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2036 pset_foreach(live, tmp) {
2037 assert(has_reg_class(si, tmp));
2040 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2041 ir_node *arg = get_irn_n(irn, n);
2043 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2046 del_pset(remat_defs);
2050 defs = pset_new_ptr_default();
2055 /***************************************
2056 * B E G I N N I N G O F B L O C K
2057 ***************************************/
2060 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2061 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2063 pset_foreach(live, irn) {
2064 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2067 /* construct mem_outs for all values */
2069 set_foreach(spill_bb->ilp, spill) {
2070 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2071 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2073 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2074 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2076 if(pset_find_ptr(live, spill->irn)) {
2077 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2079 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2080 spill->mem_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2081 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2083 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2085 op_t *op = get_irn_link(spill->irn);
2087 /* do we have to copy a phi argument? */
2088 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2089 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2091 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2092 const ir_node *arg = get_irn_n(spill->irn, n);
2098 /* argument already done? */
2099 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2101 /* get sum of execution frequencies of blocks with the same phi argument */
2102 for(m=n; m>=0; --m) {
2103 const ir_node *arg2 = get_irn_n(spill->irn, m);
2106 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2110 /* copies are not for free */
2111 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2112 var = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE * freq);
2114 for(m=n; m>=0; --m) {
2115 const ir_node *arg2 = get_irn_n(spill->irn, m);
2118 op->attr.live_range.args.copies[m] = var;
2122 /* copy <= mem_in */
2123 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2124 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2125 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2126 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2133 /* L\U is empty at bb start */
2134 /* arg is live throughout epilog if it is reg_in into this block */
2136 /* check the register pressure at the beginning of the block
2139 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2140 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
2142 pset_foreach(live, irn) {
2145 spill = set_find_spill(spill_bb->ilp, irn);
2148 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2149 spill->reg_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2151 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2153 /* spill + mem_in <= 1 */
2154 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2155 nospill = lpp_add_cst(si->lpp, buf, lpp_less, 1);
2157 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2158 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2161 foreach_post_remat(bb, irn) {
2162 op_t *remat_op = get_irn_link(irn);
2164 DBG((si->dbg, LEVEL_4, "\t next post remat: %+F\n", irn));
2165 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
2167 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2170 /* forall post remats add requirements */
2171 foreach_post_remat(bb, tmp) {
2174 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2175 ir_node *remat_arg = get_irn_n(tmp, n);
2176 op_t *remat_op = get_irn_link(tmp);
2178 if(!has_reg_class(si, remat_arg)) continue;
2180 spill = set_find_spill(spill_bb->ilp, remat_arg);
2183 /* remat <= reg_in_argument */
2184 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2185 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2186 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2187 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2191 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2192 pset_foreach(live, irn) {
2196 spill = set_find_spill(spill_bb->ilp, irn);
2197 assert(spill && spill->irn == irn);
2199 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2200 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2203 ir_node *phi_arg = get_Phi_pred(irn, n);
2204 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2205 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2208 /* although the phi is in the right regclass one or more of
2209 * its arguments can be in a different one or at least to
2212 if(has_reg_class(si, phi_arg)) {
2213 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2214 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2215 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2216 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2218 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2219 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2221 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2224 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2225 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2229 /* else assure the value arrives on all paths in the same resource */
2231 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2234 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2235 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2238 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2239 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2240 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2241 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2243 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2244 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2246 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2249 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2250 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2255 /* first live ranges from reg_ins */
2256 pset_foreach(live, irn) {
2257 op_t *op = get_irn_link(irn);
2259 spill = set_find_spill(spill_bb->ilp, irn);
2260 assert(spill && spill->irn == irn);
2262 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2263 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2264 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2265 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2267 foreach_post_remat(bb, tmp) {
2268 op_t *remat_op = get_irn_link(tmp);
2270 if(remat_op->attr.remat.remat->value == irn) {
2271 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2276 /* walk forward now and compute constraints for placing spills */
2277 /* this must only be done for values that are not defined in this block */
2278 /* TODO are these values at start of block? if yes, just check whether this is a diverge edge and skip the loop */
2279 pset_foreach(live, irn) {
2281 * if value is defined in this block we can anways place the spill directly after the def
2282 * -> no constraint necessary
2284 if(!is_Phi(irn) && get_nodes_block(irn) == bb) continue;
2287 spill = set_find_spill(spill_bb->ilp, irn);
2290 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2291 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2293 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2294 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2297 sched_foreach_op(bb, tmp) {
2298 op_t *op = get_irn_link(tmp);
2300 if(is_Phi(tmp)) continue;
2301 assert(!is_Proj(tmp));
2304 ir_node *value = op->attr.remat.remat->value;
2307 /* only collect remats up to the first use of a value */
2308 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2313 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2314 ir_node *arg = get_irn_n(tmp, n);
2317 /* if a value is used stop collecting remats */
2323 if(cst == ILP_UNDEF) break;
2331 typedef struct _irnlist_t {
2332 struct list_head list;
2336 typedef struct _interference_t {
2337 struct list_head blocklist;
2343 cmp_interference(const void *a, const void *b, size_t size)
2345 const interference_t *p = a;
2346 const interference_t *q = b;
2348 return !(p->a == q->a && p->b == q->b);
2351 static interference_t *
2352 set_find_interference(set * set, ir_node * a, ir_node * b)
2354 interference_t query;
2356 query.a = (a>b)?a:b;
2357 query.b = (a>b)?b:a;
2359 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2362 static interference_t *
2363 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2365 interference_t query,
2367 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2371 result = set_find_interference(set, a, b);
2374 list_add(&list->list, &result->blocklist);
2378 query.a = (a>b)?a:b;
2379 query.b = (a>b)?b:a;
2381 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2383 INIT_LIST_HEAD(&result->blocklist);
2384 list_add(&list->list, &result->blocklist);
2390 values_interfere_in_block(ir_node * bb, ir_node * a, ir_node * b)
2392 const ir_edge_t *edge;
2394 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2395 /* both values are live in, so they interfere */
2399 /* ensure a dominates b */
2400 if(value_dominates(b,a)) {
2406 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2409 /* the following code is stolen from bera.c */
2410 if(is_live_end(bb, a))
2413 foreach_out_edge(a, edge) {
2414 const ir_node *user = edge->src;
2415 if(get_nodes_block(user) == bb
2418 && value_dominates(b, user))
2426 * Walk all irg blocks and collect interfering values inside of phi classes
2429 luke_interferencewalker(ir_node * bb, void * data)
2431 spill_ilp_t *si = (spill_ilp_t*)data;
2435 live_foreach(bb, li1) {
2436 ir_node *a = (ir_node *) li1->irn;
2437 op_t *a_op = get_irn_link(a);
2439 if(a_op->is_remat) continue;
2441 /* a is only interesting if it is in my register class and if it is inside a phi class */
2442 if (has_reg_class(si, a) && get_phi_class(a)) {
2443 for(li2=li1->next; li2; li2 = li2->next) {
2444 ir_node *b = (ir_node *) li2->irn;
2445 op_t *b_op = get_irn_link(b);
2447 if(b_op->is_remat) continue;
2449 /* a and b are only interesting if they are in the same phi class */
2450 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2451 if(values_interfere_in_block(bb, a, b)) {
2452 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2453 set_insert_interference(si, si->interferences, a, b, bb);
2461 static unsigned int copy_path_id = 0;
2464 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2471 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2472 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2474 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2476 pset_foreach(copies, ptr) {
2477 copy = PTR_TO_INT(ptr);
2478 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2483 * @parameter copies contains a path of copies which lead us to irn
2484 * @parameter visited contains a set of nodes already visited on this path
2487 find_copy_path(spill_ilp_t * si, ir_node * irn, ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2490 op_t *op = get_irn_link(irn);
2492 if(op->is_remat) return;
2494 pset_insert_ptr(visited, irn);
2499 /* visit all operands */
2500 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2501 ir_node *arg = get_irn_n(irn, n);
2502 ilp_var_t copy = op->attr.live_range.args.copies[n];
2504 if(!has_reg_class(si, arg)) continue;
2507 pset_insert(copies, INT_TO_PTR(copy), copy);
2508 write_copy_path_cst(si, copies, any_interfere);
2509 pset_remove(copies, INT_TO_PTR(copy), copy);
2511 if(!pset_find_ptr(visited, arg)) {
2512 pset_insert(copies, INT_TO_PTR(copy), copy);
2513 find_copy_path(si, arg, target, any_interfere, copies, visited);
2514 pset_remove(copies, INT_TO_PTR(copy), copy);
2520 /* visit all uses which are phis */
2521 foreach_out_edge(irn, edge) {
2522 ir_node *user = edge->src;
2523 int pos = edge->pos;
2524 op_t *op = get_irn_link(user);
2527 if(!is_Phi(user)) continue;
2528 if(!has_reg_class(si, user)) continue;
2530 copy = op->attr.live_range.args.copies[pos];
2532 if(user == target) {
2533 pset_insert(copies, INT_TO_PTR(copy), copy);
2534 write_copy_path_cst(si, copies, any_interfere);
2535 pset_remove(copies, INT_TO_PTR(copy), copy);
2537 if(!pset_find_ptr(visited, user)) {
2538 pset_insert(copies, INT_TO_PTR(copy), copy);
2539 find_copy_path(si, user, target, any_interfere, copies, visited);
2540 pset_remove(copies, INT_TO_PTR(copy), copy);
2545 pset_remove_ptr(visited, irn);
2549 gen_copy_constraints(spill_ilp_t * si, ir_node * a, ir_node * b, ilp_var_t any_interfere)
2551 pset * copies = pset_new_ptr_default();
2552 pset * visited = pset_new_ptr_default();
2554 find_copy_path(si, a, b, any_interfere, copies, visited);
2562 memcopyhandler(spill_ilp_t * si)
2564 interference_t *interference;
2566 /* teste Speicherwerte auf Interferenz */
2568 /* analyze phi classes */
2569 phi_class_compute(si->chordal_env->irg);
2571 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
2572 irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
2574 // phi_class_free(si->chordal_env->irg);
2576 /* now lets emit the ILP unequations for the crap */
2577 set_foreach(si->interferences, interference) {
2579 ilp_var_t interfere,
2581 ilp_cst_t any_interfere_cst,
2583 const ir_node *a = interference->a;
2584 const ir_node *b = interference->b;
2586 /* any_interf <= \sum interf */
2587 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
2588 any_interfere_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2589 any_interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2591 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
2593 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
2594 const ir_node *bb = irnlist->irn;
2595 spill_bb_t *spill_bb = get_irn_link(bb);
2602 spilla = set_find_spill(spill_bb->ilp, a);
2606 spillb = set_find_spill(spill_bb->ilp, b);
2609 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
2610 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
2611 /* 2: - mem_in_a - spill_a + interfere <= 0 */
2612 /* 3: - mem_in_b - spill_b + interfere <= 0 */
2613 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
2614 interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2616 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
2617 cst = lpp_add_cst(si->lpp, buf, lpp_less, 1);
2619 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
2620 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
2621 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
2622 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
2623 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
2625 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
2626 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2628 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2629 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
2630 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
2632 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
2633 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2635 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2636 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
2637 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
2640 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
2642 /* any_interfere >= interf */
2643 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
2644 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2646 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2647 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2650 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
2651 gen_copy_constraints(si,a,b,any_interfere);
2658 memcopyinsertor(spill_ilp_t * si)
2660 /* weise Spillkontexte zu. Sorge bei Phis dafuer, dass gleiche
2661 * Kontexte zusammenfliessen (Operanden und Ergebnis hat gleichen
2677 return fabs(x) < 0.00001;
2681 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
2683 spill_ilp_t *si = get_irg_link(current_ir_graph);
2685 if(pset_find_ptr(si->all_possible_remats, n)) {
2686 op_t *op = (op_t*)get_irn_link(n);
2687 assert(op && op->is_remat);
2689 if(!op->attr.remat.remat->inverse) {
2690 if(op->attr.remat.pre) {
2691 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
2693 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
2698 op_t *op = (op_t*)get_irn_link(n);
2699 assert(op && op->is_remat);
2701 if(op->attr.remat.pre) {
2702 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
2704 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
2715 dump_graph_with_remats(ir_graph * irg, const char * suffix)
2717 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
2718 be_dump(irg, suffix, dump_ir_block_graph_sched);
2719 set_dump_node_vcgattr_hook(NULL);
2724 * Edge hook to dump the schedule edges with annotated register pressure.
2727 sched_pressure_edge_hook(FILE *F, ir_node *irn)
2729 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
2730 ir_node *prev = sched_prev(irn);
2731 fprintf(F, "edge:{sourcename:\"");
2733 fprintf(F, "\" targetname:\"");
2735 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
2736 fprintf(F, "\" color:magenta}\n");
2742 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
2744 DUMP_NODE_EDGE_FUNC old = get_dump_node_edge_hook();
2746 dump_consts_local(0);
2747 set_dump_node_edge_hook(sched_pressure_edge_hook);
2748 dump_ir_block_graph(irg, suffix);
2749 set_dump_node_edge_hook(old);
2753 walker_pressure_annotator(ir_node * bb, void * data)
2755 spill_ilp_t *si = data;
2759 pset *live = pset_new_ptr_default();
2762 live_foreach(bb, li) {
2763 irn = (ir_node *) li->irn;
2765 if (live_is_end(li) && has_reg_class(si, irn)) {
2766 pset_insert_ptr(live, irn);
2770 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
2772 sched_foreach_reverse(bb, irn) {
2774 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
2778 if(has_reg_class(si, irn)) {
2779 pset_remove_ptr(live, irn);
2780 if(is_Proj(irn)) ++projs;
2783 if(!is_Proj(irn)) projs = 0;
2785 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2786 ir_node *arg = get_irn_n(irn, n);
2788 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
2790 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
2797 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
2799 be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
2804 connect_all_remats_with_keep(spill_ilp_t * si)
2812 n_remats = pset_count(si->all_possible_remats);
2814 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
2817 pset_foreach(si->all_possible_remats, irn) {
2822 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
2824 obstack_free(si->obst, ins);
2830 connect_all_spills_with_keep(spill_ilp_t * si)
2839 n_spills = pset_count(si->spills);
2841 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
2844 pset_foreach(si->spills, irn) {
2849 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
2851 obstack_free(si->obst, ins);
2855 /** insert a spill at an arbitrary position */
2856 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert, ir_node *ctx)
2858 ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
2859 ir_graph *irg = get_irn_irg(bl);
2860 ir_node *frame = get_irg_frame(irg);
2864 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
2865 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
2867 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn, ctx);
2870 * search the right insertion point. a spill of a phi cannot be put
2871 * directly after the phi, if there are some phis behind the one which
2872 * is spilled. Also, a spill of a Proj must be after all Projs of the
2875 * Here's one special case:
2876 * If the spill is in the start block, the spill must be after the frame
2877 * pointer is set up. This is done by setting insert to the end of the block
2878 * which is its default initialization (see above).
2881 if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
2884 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
2887 sched_add_after(insert, spill);
2892 delete_remat(spill_ilp_t * si, ir_node * remat) {
2894 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2896 sched_remove(remat);
2898 /* kill links to operands */
2899 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
2900 set_irn_n(remat, n, bad);
2905 clean_remat_info(spill_ilp_t * si)
2909 remat_info_t *remat_info;
2910 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2912 set_foreach(si->remat_info, remat_info) {
2913 if(!remat_info->remats) continue;
2915 pset_foreach(remat_info->remats, remat)
2917 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
2918 set_irn_n(remat->proj, -1, bad);
2919 set_irn_n(remat->proj, 0, bad);
2922 if(get_irn_n_edges(remat->op) == 0) {
2923 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
2924 set_irn_n(remat->op, n, bad);
2929 if(remat_info->remats) del_pset(remat_info->remats);
2930 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
2935 delete_unnecessary_remats(spill_ilp_t * si)
2939 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2942 ir_node *end = get_irg_end(si->chordal_env->irg);
2945 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
2946 ir_node *keep_arg = get_irn_n(si->keep, n);
2947 op_t *arg_op = get_irn_link(keep_arg);
2950 assert(arg_op->is_remat);
2952 name = si->lpp->vars[arg_op->attr.remat.ilp];
2954 if(is_zero(name->value)) {
2955 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
2956 /* TODO check whether reload is preferred over remat (could be bug) */
2957 delete_remat(si, keep_arg);
2959 if(!arg_op->attr.remat.remat->inverse) {
2960 if(arg_op->attr.remat.pre) {
2961 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
2963 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
2966 if(arg_op->attr.remat.pre) {
2967 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
2969 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
2974 set_irn_n(si->keep, n, bad);
2977 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
2978 ir_node *end_arg = get_End_keepalive(end, i);
2980 if(end_arg != si->keep) {
2981 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
2984 keeps = obstack_finish(si->obst);
2985 set_End_keepalives(end, n-1, keeps);
2986 obstack_free(si->obst, keeps);
2989 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
2994 pset_foreach(si->all_possible_remats, remat) {
2995 op_t *remat_op = get_irn_link(remat);
2996 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
2998 if(is_zero(name->value)) {
2999 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
3000 /* TODO check whether reload is preferred over remat (could be bug) */
3001 delete_remat(si, remat);
3003 if(!remat_op->attr.remat.remat->inverse) {
3004 if(remat_op->attr.remat.pre) {
3005 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3007 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3010 if(remat_op->attr.remat.pre) {
3011 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3013 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3022 get_spills_for_value(spill_ilp_t * si, ir_node * value)
3024 pset *spills = pset_new_ptr_default();
3029 defs = set_find_def(si->values, value);
3031 if(defs && defs->spills) {
3032 for(next = defs->spills; next; next = get_irn_link(next)) {
3033 pset_insert_ptr(spills, next);
3041 get_remats_for_value(spill_ilp_t * si, ir_node * value)
3043 pset *remats = pset_new_ptr_default();
3048 pset_insert_ptr(remats, value);
3049 defs = set_find_def(si->values, value);
3051 if(defs && defs->remats) {
3052 for(next = defs->remats; next; next = get_irn_link(next)) {
3053 pset_insert_ptr(remats, next);
3062 * @param before The node after which the spill will be placed in the schedule
3064 /* TODO set context properly */
3066 insert_spill(spill_ilp_t * si, ir_node * irn, ir_node * value, ir_node * before)
3070 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3072 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3074 spill = be_spill2(arch_env, irn, before, irn);
3076 defs = set_insert_def(si->values, value);
3079 /* enter into the linked list */
3080 set_irn_link(spill, defs->spills);
3081 defs->spills = spill;
3083 #ifdef KEEPALIVE_SPILLS
3084 pset_insert_ptr(si->spills, spill);
3091 * @param before The Phi node which has to be spilled
3094 insert_mem_phi(spill_ilp_t * si, const ir_node * phi)
3100 op_t *op = get_irn_link(phi);
3102 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3104 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3105 ins[n] = si->m_unknown;
3108 mem_phi = new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3110 defs = set_insert_def(si->values, phi);
3113 /* enter into the linked list */
3114 set_irn_link(mem_phi, defs->spills);
3115 defs->spills = mem_phi;
3117 sched_add_after(phi, mem_phi);
3119 #ifdef KEEPALIVE_SPILLS
3120 pset_insert_ptr(si->spills, mem_phi);
3128 * Add remat to list of defs, destroys link field!
3131 insert_remat(spill_ilp_t * si, ir_node * remat)
3134 op_t *remat_op = get_irn_link(remat);
3136 assert(remat_op->is_remat);
3138 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3141 /* enter into the linked list */
3142 set_irn_link(remat, defs->remats);
3143 defs->remats = remat;
3148 * Add reload before operation and add to list of defs
3151 insert_reload(spill_ilp_t * si, const ir_node * value, const ir_node * after)
3156 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3158 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3160 defs = set_find_def(si->values, value);
3162 spill = defs->spills;
3163 assert(spill && "no spill placed before reload");
3165 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3167 /* enter into the linked list */
3168 set_irn_link(reload, defs->remats);
3169 defs->remats = reload;
3175 walker_spill_placer(ir_node * bb, void * data) {
3176 spill_ilp_t *si = (spill_ilp_t*)data;
3178 spill_bb_t *spill_bb = get_irn_link(bb);
3179 pset *spills_to_do = pset_new_ptr_default();
3182 set_foreach(spill_bb->ilp, spill) {
3185 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3186 name = si->lpp->vars[spill->mem_in];
3187 if(!is_zero(name->value)) {
3190 mem_phi = insert_mem_phi(si, spill->irn);
3192 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3196 name = si->lpp->vars[spill->spill];
3197 if(!is_zero(name->value)) {
3198 /* place spill directly after definition */
3199 if(get_nodes_block(spill->irn) == bb) {
3200 insert_spill(si, spill->irn, spill->irn, spill->irn);
3204 /* place spill at bb start */
3205 if(spill->reg_in > 0) {
3206 name = si->lpp->vars[spill->reg_in];
3207 if(!is_zero(name->value)) {
3208 insert_spill(si, spill->irn, spill->irn, bb);
3212 /* place spill after a remat */
3213 pset_insert_ptr(spills_to_do, spill->irn);
3216 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3219 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3220 op_t *op = get_irn_link(irn);
3222 if(be_is_Spill(irn)) continue;
3225 /* TODO fix this if we want to support remats with more than two nodes */
3226 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3227 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3229 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3232 if(pset_find_ptr(spills_to_do, irn)) {
3233 pset_remove_ptr(spills_to_do, irn);
3235 insert_spill(si, irn, irn, irn);
3241 assert(pset_count(spills_to_do) == 0);
3243 /* afterwards free data in block */
3244 del_pset(spills_to_do);
3248 insert_mem_copy(spill_ilp_t * si, const ir_node * bb, const ir_node * value)
3250 ir_node *insert_pos = bb;
3252 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3254 /* find last definition of arg value in block */
3259 defs = set_find_def(si->values, value);
3261 if(defs && defs->remats) {
3262 for(next = defs->remats; next; next = get_irn_link(next)) {
3263 if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3264 last = sched_get_time_step(next);
3270 if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3271 last = sched_get_time_step(value);
3275 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
3277 spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos, value);
3283 phim_fixer(spill_ilp_t *si) {
3286 set_foreach(si->values, defs) {
3287 const ir_node *phi = defs->value;
3288 op_t *op = get_irn_link(phi);
3289 ir_node *phi_m = NULL;
3290 ir_node *next = defs->spills;
3293 if(!is_Phi(phi)) continue;
3296 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3300 next = get_irn_link(next);
3303 if(!phi_m) continue;
3305 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3306 const ir_node *value = get_irn_n(phi, n);
3307 defs_t *val_defs = set_find_def(si->values, value);
3308 ir_node *arg = get_irn_n(phi_m, n);
3310 /* get a spill of this value */
3311 ir_node *spill = val_defs->spills;
3314 #ifndef NO_MEMCOPIES
3315 ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3316 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3318 if(!is_zero(name->value)) {
3319 spill = insert_mem_copy(si, pred, value);
3321 assert(spill && "no spill placed before PhiM");
3324 assert(spill && "no spill placed before PhiM");
3326 set_irn_n(phi_m, n, spill);
3332 walker_reload_placer(ir_node * bb, void * data) {
3333 spill_ilp_t *si = (spill_ilp_t*)data;
3335 spill_bb_t *spill_bb = get_irn_link(bb);
3339 /* reloads at end of block */
3340 if(spill_bb->reloads) {
3343 set_foreach(spill_bb->reloads, keyval) {
3344 ir_node *irn = (ir_node*)keyval->key;
3345 ilp_var_t reload = PTR_TO_INT(keyval->val);
3348 name = si->lpp->vars[reload];
3349 if(!is_zero(name->value)) {
3351 ir_node *insert_pos = bb;
3352 ir_node *prev = sched_block_last_noncf(si, bb);
3353 op_t *prev_op = get_irn_link(prev);
3355 while(be_is_Spill(prev)) {
3356 prev = sched_prev(prev);
3359 prev_op = get_irn_link(prev);
3361 /* insert reload before pre-remats */
3362 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3363 && prev_op->is_remat && prev_op->attr.remat.pre) {
3367 prev = sched_prev(prev);
3368 } while(be_is_Spill(prev));
3370 prev_op = get_irn_link(prev);
3374 reload = insert_reload(si, irn, insert_pos);
3376 #ifdef KEEPALIVE_RELOADS
3377 pset_insert_ptr(si->spills, reload);
3383 /* walk and insert more reloads and collect remats */
3384 sched_foreach_reverse(bb, irn) {
3385 op_t *op = get_irn_link(irn);
3387 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3388 if(is_Phi(irn)) break;
3391 if(get_irn_mode(irn) != mode_T) {
3392 insert_remat(si, irn);
3397 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3398 ir_node *arg = get_irn_n(irn, n);
3400 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3403 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3404 if(!is_zero(name->value)) {
3406 ir_node *insert_pos = irn;
3407 ir_node *prev = sched_prev(insert_pos);
3410 while(be_is_Spill(prev)) {
3411 prev = sched_prev(prev);
3414 prev_op = get_irn_link(prev);
3416 /* insert reload before pre-remats */
3417 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3418 && prev_op->is_remat && prev_op->attr.remat.pre) {
3422 prev = sched_prev(prev);
3423 } while(be_is_Spill(prev));
3425 prev_op = get_irn_link(prev);
3429 reload = insert_reload(si, arg, insert_pos);
3431 set_irn_n(irn, n, reload);
3433 #ifdef KEEPALIVE_RELOADS
3434 pset_insert_ptr(si->spills, reload);
3442 del_set(spill_bb->ilp);
3443 if(spill_bb->reloads) del_set(spill_bb->reloads);
3447 walker_collect_used(ir_node * irn, void * data)
3449 lc_bitset_t *used = data;
3451 lc_bitset_set(used, get_irn_idx(irn));
3454 struct kill_helper {
3460 walker_kill_unused(ir_node * bb, void * data)
3462 struct kill_helper *kh = data;
3463 const ir_node *bad = get_irg_bad(get_irn_irg(bb));
3467 for(irn=sched_first(bb); !sched_is_end(irn);) {
3468 ir_node *next = sched_next(irn);
3471 if(!lc_bitset_is_set(kh->used, get_irn_idx(irn))) {
3472 if(be_is_Spill(irn) || be_is_Reload(irn)) {
3473 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)));
3475 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3481 set_nodes_block(irn, bad);
3482 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3483 set_irn_n(irn, n, bad);
3491 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3493 struct kill_helper kh;
3495 kh.used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3498 irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3499 irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3501 lc_bitset_free(kh.used);
3505 print_irn_pset(pset * p)
3509 pset_foreach(p, irn) {
3510 ir_printf("%+F\n", irn);
3515 rewire_uses(spill_ilp_t * si)
3517 dom_front_info_t *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
3519 pset *ignore = pset_new_ptr(1);
3521 pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
3523 /* then fix uses of spills */
3524 set_foreach(si->values, defs) {
3527 ir_node *next = defs->remats;
3530 reloads = pset_new_ptr_default();
3533 if(be_is_Reload(next)) {
3534 pset_insert_ptr(reloads, next);
3538 next = get_irn_link(next);
3541 spills = get_spills_for_value(si, defs->value);
3542 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));
3543 if(pset_count(spills) > 1) {
3544 //assert(pset_count(reloads) > 0);
3545 // print_irn_pset(spills);
3546 // print_irn_pset(reloads);
3548 be_ssa_constr_set_ignore(dfi, spills, ignore);
3555 /* first fix uses of remats and reloads */
3556 set_foreach(si->values, defs) {
3558 ir_node *next = defs->remats;
3561 nodes = pset_new_ptr_default();
3562 pset_insert_ptr(nodes, defs->value);
3565 pset_insert_ptr(nodes, next);
3566 next = get_irn_link(next);
3569 if(pset_count(nodes) > 1) {
3570 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
3571 be_ssa_constr_set(dfi, nodes);
3578 // remove_unused_defs(si);
3580 be_free_dominance_frontiers(dfi);
3585 writeback_results(spill_ilp_t * si)
3587 /* walk through the graph and collect all spills, reloads and remats for a value */
3589 si->values = new_set(cmp_defs, 4096);
3591 DBG((si->dbg, LEVEL_1, "Applying results\n"));
3592 delete_unnecessary_remats(si);
3593 si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
3594 irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
3595 irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
3598 /* clean the remat info! there are still back-edges leading there! */
3599 clean_remat_info(si);
3603 connect_all_spills_with_keep(si);
3605 del_set(si->values);
3609 get_n_regs(spill_ilp_t * si)
3611 int arch_n_regs = arch_register_class_n_regs(si->cls);
3615 for(i=0; i<arch_n_regs; i++) {
3616 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
3621 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
3626 walker_reload_mover(ir_node * bb, void * data)
3628 spill_ilp_t *si = data;
3631 sched_foreach(bb, tmp) {
3632 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
3633 ir_node *reload = tmp;
3636 /* move reload upwards */
3638 int pressure = (int)get_irn_link(reload);
3639 if(pressure < si->n_regs) {
3640 irn = sched_prev(reload);
3641 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
3642 sched_remove(reload);
3643 pressure = (int)get_irn_link(irn);
3645 while(pressure < si->n_regs) {
3646 if(sched_is_end(irn) || (be_is_Reload(irn) && has_reg_class(si, irn))) break;
3648 set_irn_link(irn, INT_TO_PTR(pressure+1));
3649 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
3650 irn = sched_prev(irn);
3652 pressure = (int)get_irn_link(irn);
3655 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
3656 sched_put_after(irn, reload);
3663 move_reloads_upward(spill_ilp_t * si)
3665 irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
3670 * Walk all irg blocks and check for interfering spills inside of phi classes
3673 luke_meminterferencechecker(ir_node * bb, void * data)
3675 spill_ilp_t *si = (spill_ilp_t*)data;
3679 live_foreach(bb, li1) {
3680 ir_node *a = (ir_node *) li1->irn;
3682 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
3684 /* a is only interesting if it is inside a phi class */
3685 if (get_phi_class(a)) {
3686 for(li2=li1->next; li2; li2 = li2->next) {
3687 ir_node *b = (ir_node *) li2->irn;
3689 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
3691 /* a and b are only interesting if they are in the same phi class */
3692 if(get_phi_class(a) == get_phi_class(b)) {
3693 if(values_interfere_in_block(bb, a, b)) {
3694 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
3703 verify_phiclasses(spill_ilp_t * si)
3705 /* analyze phi classes */
3706 phi_class_compute(si->chordal_env->irg);
3708 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
3709 irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
3713 walker_spillslotassigner(ir_node * irn, void * data)
3715 spill_ilp_t *si = (spill_ilp_t*)data;
3718 if(!be_is_Spill(irn)) return;
3720 /* set spill context to phi class if it has one ;) */
3722 cls = get_phi_class(irn);
3724 be_set_Spill_context(irn, cls);
3726 be_set_Spill_context(irn, irn);
3731 assign_spillslots(spill_ilp_t * si)
3733 DBG((si->dbg, LEVEL_2, "\t calling spill slot assigner\n"));
3734 irg_walk_graph(si->chordal_env->irg, walker_spillslotassigner, NULL, si);
3738 be_spill_remat(const be_chordal_env_t * chordal_env)
3740 char problem_name[256];
3741 char dump_suffix[256];
3742 char dump_suffix2[256];
3743 char dump_suffix3[256];
3744 struct obstack obst;
3747 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
3748 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
3749 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
3751 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
3752 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
3754 obstack_init(&obst);
3755 si.chordal_env = chordal_env;
3757 si.cls = chordal_env->cls;
3758 si.lpp = new_lpp(problem_name, lpp_minimize);
3759 si.remat_info = new_set(cmp_remat_info, 4096);
3760 si.interferences = new_set(cmp_interference, 32);
3761 si.all_possible_remats = pset_new_ptr_default();
3762 si.spills = pset_new_ptr_default();
3763 si.inverse_ops = pset_new_ptr_default();
3764 #ifndef EXECFREQ_LOOPDEPH
3765 si.execfreqs = compute_execfreq(chordal_env->irg);
3767 si.execfreqs = NULL;
3772 si.n_regs = get_n_regs(&si);
3774 set_irg_link(chordal_env->irg, &si);
3775 compute_doms(chordal_env->irg);
3777 /* compute phi classes */
3778 // phi_class_compute(chordal_env->irg);
3780 be_analyze_regpressure(chordal_env, "-pre");
3782 #ifdef COLLECT_REMATS
3783 /* collect remats */
3784 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
3785 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
3788 /* insert possible remats */
3789 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
3790 irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
3791 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
3794 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
3795 connect_all_remats_with_keep(&si);
3796 /* dump graph with inserted remats */
3797 dump_graph_with_remats(chordal_env->irg, dump_suffix);
3800 /* insert copies for phi arguments not in my regclass */
3801 irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
3803 /* recompute liveness */
3804 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
3805 be_liveness(chordal_env->irg);
3809 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
3810 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
3811 irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
3813 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
3814 irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
3816 #ifndef NO_MEMCOPIES
3817 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
3818 memcopyhandler(&si);
3826 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
3827 if ((f = fopen(buf, "wt")) != NULL) {
3828 lpp_dump_plain(si.lpp, f);
3835 DBG((si.dbg, LEVEL_1, "\tSolving %F\n", chordal_env->irg));
3837 lpp_set_time_limit(si.lpp, ILP_TIMEOUT);
3841 lpp_solve_cplex(si.lpp);
3843 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
3845 assert(lpp_is_sol_valid(si.lpp)
3846 && "solution of ILP must be valid");
3848 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));
3850 #ifdef DUMP_SOLUTION
3855 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
3856 if ((f = fopen(buf, "wt")) != NULL) {
3858 for (i = 0; i < si.lpp->var_next; ++i) {
3859 lpp_name_t *name = si.lpp->vars[i];
3860 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
3867 writeback_results(&si);
3871 kill_all_unused_values_in_schedule(&si);
3873 #if defined(KEEPALIVE_SPILLS) || defined(KEEPALIVE_RELOADS)
3874 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
3877 // move reloads upwards
3878 be_liveness(chordal_env->irg);
3879 //irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
3880 //move_reloads_upward(&si);
3882 #ifndef NO_MEMCOPIES
3883 verify_phiclasses(&si);
3884 assign_spillslots(&si);
3887 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
3889 dump_pressure_graph(&si, dump_suffix2);
3891 be_analyze_regpressure(chordal_env, "-post");
3893 free_dom(chordal_env->irg);
3894 del_set(si.interferences);
3895 del_pset(si.inverse_ops);
3896 del_pset(si.all_possible_remats);
3897 del_pset(si.spills);
3898 #ifndef EXECFREQ_LOOPDEPH
3899 free_execfreq(si.execfreqs);
3902 obstack_free(&obst, NULL);
3903 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
3906 #else /* WITH_ILP */
3909 only_that_you_can_compile_without_WITH_ILP_defined(void)
3913 #endif /* WITH_ILP */