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"
38 #include <lpp/lpp_net.h>
39 #include <lpp/lpp_cplex.h>
40 //#include <lc_pset.h>
41 #include <libcore/lc_bitset.h>
45 #include "besched_t.h"
50 #include "bespillremat.h"
52 #include "bepressurestat.h"
54 #include "bechordal_t.h"
60 //#define KEEPALIVE /* keep alive all inserted remats and dump graph with remats */
61 #define COLLECT_REMATS /* enable rematerialization */
62 #define COLLECT_INVERSE_REMATS /* enable placement of inverse remats */
63 #define REMAT_WHILE_LIVE /* only remat values that are live */
64 //#define NO_ENLARGE_L1V3N355 /* do not remat after the death of some operand */
65 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
66 #define MAY_DIE_AT_REMAT /* allow values to die after a pre remat */
67 #define NO_SINGLE_USE_REMATS /* do not repair schedule */
68 //#define KEEPALIVE_SPILLS
69 //#define KEEPALIVE_RELOADS
70 #define GOODWIN_REDUCTION
71 //#define NO_MEMCOPIES
72 //#define VERIFY_DOMINANCE
73 #define WITH_MEMOPERANDS
77 #define LPP_SERVER "i44pc52"
78 #define LPP_SOLVER "cplex"
81 #define COST_MEMOPERAND 7
85 #define ILP_TIMEOUT 300
89 typedef struct _spill_ilp_t {
90 const arch_register_class_t *cls;
92 const be_chordal_env_t *chordal_env;
97 pset *all_possible_remats;
102 set *values; /**< for collecting all definitions of values before running ssa-construction */
106 #ifdef WITH_MEMOPERANDS
109 DEBUG_ONLY(firm_dbg_module_t * dbg);
112 typedef int ilp_var_t;
113 typedef int ilp_cst_t;
115 typedef struct _spill_bb_t {
120 typedef struct _remat_t {
121 const ir_node *op; /**< for copy_irn */
122 const ir_node *value; /**< the value which is being recomputed by this remat */
123 ir_node *proj; /**< not NULL if the above op produces a tuple */
124 int cost; /**< cost of this remat */
125 int inverse; /**< nonzero if this is an inverse remat */
129 * Data to be attached to each IR node. For remats this contains the ilp_var
130 * for this remat and for normal ops this contains the ilp_vars for
131 * reloading each operand
133 typedef struct _op_t {
138 remat_t *remat; /** the remat this op belongs to */
139 int pre; /** 1, if this is a pressure-increasing remat */
143 ir_node *op; /** the operation this live range belongs to */
152 typedef struct _defs_t {
154 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
155 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
158 typedef struct _remat_info_t {
159 const ir_node *irn; /**< the irn to which these remats belong */
160 pset *remats; /**< possible remats for this value */
161 pset *remats_by_operand; /**< remats with this value as operand */
164 typedef struct _keyval_t {
169 typedef struct _spill_t {
178 #ifdef WITH_MEMOPERANDS
179 typedef struct _memoperand_t {
180 const ir_node *irn; /**< the irn */
181 unsigned int pos; /**< the position of the argument */
182 ilp_var_t ilp; /**< the ilp var for this memory operand */
187 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
189 return chordal_has_class(si->chordal_env, irn);
194 cmp_remat(const void *a, const void *b)
196 const keyval_t *p = a;
197 const keyval_t *q = b;
198 const remat_t *r = p->val;
199 const remat_t *s = q->val;
203 return !(r == s || r->op == s->op);
207 cmp_remat(const void *a, const void *b)
209 const remat_t *r = a;
210 const remat_t *s = a;
212 return !(r == s || r->op == s->op);
216 cmp_spill(const void *a, const void *b, size_t size)
218 const spill_t *p = a;
219 const spill_t *q = b;
221 // return !(p->irn == q->irn && p->bb == q->bb);
222 return !(p->irn == q->irn);
225 #ifdef WITH_MEMOPERANDS
227 cmp_memoperands(const void *a, const void *b, size_t size)
229 const memoperand_t *p = a;
230 const memoperand_t *q = b;
232 return !(p->irn == q->irn && p->pos == q->pos);
237 set_find_keyval(set * set, void * key)
242 return set_find(set, &query, sizeof(query), HASH_PTR(key));
246 set_insert_keyval(set * set, void * key, void * val)
252 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
256 set_find_def(set * set, ir_node * value)
261 return set_find(set, &query, sizeof(query), HASH_PTR(value));
265 set_insert_def(set * set, ir_node * value)
272 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
275 #ifdef WITH_MEMOPERANDS
276 static memoperand_t *
277 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
284 return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
287 static memoperand_t *
288 set_find_memoperand(set * set, ir_node * irn, unsigned int pos)
294 return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
300 set_find_spill(set * set, ir_node * value)
305 return set_find(set, &query, sizeof(query), HASH_PTR(value));
308 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
309 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
310 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
311 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
312 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
315 cmp_remat_info(const void *a, const void *b, size_t size)
317 const remat_info_t *p = a;
318 const remat_info_t *q = b;
320 return !(p->irn == q->irn);
324 cmp_defs(const void *a, const void *b, size_t size)
329 return !(p->value == q->value);
333 cmp_keyval(const void *a, const void *b, size_t size)
335 const keyval_t *p = a;
336 const keyval_t *q = b;
338 return !(p->key == q->key);
342 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
345 #ifndef EXECFREQ_LOOPDEPH
346 return get_block_execfreq(si->chordal_env->exec_freq, get_block(irn)) + FUDGE;
349 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
351 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
356 get_cost(const spill_ilp_t * si, const ir_node * irn)
358 if(be_is_Spill(irn)) {
360 } else if(be_is_Reload(irn)){
363 return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
368 * Checks, whether node and its operands have suitable reg classes
371 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
374 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
375 int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
379 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
382 for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
383 ir_node *op = get_irn_n(irn, n);
384 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
387 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
394 * Try to create a remat from @p op with destination value @p dest_value
396 static INLINE remat_t *
397 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
399 remat_t *remat = NULL;
401 // if(!mode_is_datab(get_irn_mode(dest_value)))
404 if(dest_value == op) {
405 const ir_node *proj = NULL;
407 if(is_Proj(dest_value)) {
408 op = get_irn_n(op, 0);
412 if(!is_rematerializable(si, op))
415 remat = obstack_alloc(si->obst, sizeof(*remat));
417 remat->cost = get_cost(si, op);
418 remat->value = dest_value;
422 arch_inverse_t inverse;
425 /* get the index of the operand we want to retrieve by the inverse op */
426 for (n = get_irn_arity(op)-1; n>=0; --n) {
427 ir_node *arg = get_irn_n(op, n);
429 if(arg == dest_value) break;
433 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", n, op));
435 /* else ask the backend to give an inverse op */
436 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
439 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
441 assert(inverse.n > 0 && "inverse op should have at least one node");
443 for(i=inverse.n-1; i>=0; --i) {
444 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
448 remat = obstack_alloc(si->obst, sizeof(*remat));
449 remat->op = inverse.nodes[0];
450 remat->cost = inverse.costs;
451 remat->value = dest_value;
452 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
455 assert(is_Proj(remat->proj));
457 assert(0 && "I can not handle remats with more than 2 nodes");
464 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
466 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
474 add_remat(const spill_ilp_t * si, const remat_t * remat)
476 remat_info_t *remat_info,
481 assert(remat->value);
483 query.irn = remat->value;
485 query.remats_by_operand = NULL;
486 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
488 if(remat_info->remats == NULL) {
489 remat_info->remats = new_pset(cmp_remat, 4096);
491 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
493 /* insert the remat into the remats_be_operand set of each argument of the remat op */
494 for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
495 ir_node *arg = get_irn_n(remat->op, n);
499 query.remats_by_operand = NULL;
500 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
502 if(remat_info->remats_by_operand == NULL) {
503 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
505 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
510 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
512 const ir_edge_t *edge = get_irn_out_edge_first(irn);
516 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
519 edge = get_irn_out_edge_next(irn, edge);
526 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
531 #ifdef NO_SINGLE_USE_REMATS
532 if(has_reg_class(si, op) && (get_irn_n_nonremat_edges(si, op) > 1)) {
534 if(has_reg_class(si, op)) {
536 remat = get_remat_from_op(si, op, op);
538 add_remat(si, remat);
542 #ifdef COLLECT_INVERSE_REMATS
543 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
545 for (n = get_irn_arity(op)-1; n>=0; --n) {
546 ir_node *arg = get_irn_n(op, n);
548 if(has_reg_class(si, arg)) {
549 /* try to get an inverse remat */
550 remat = get_remat_from_op(si, arg, op);
552 add_remat(si, remat);
561 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
564 ir_node *def_block = get_nodes_block(val);
570 /* if pos is at end of a basic block */
572 ret = (pos == def_block || block_dominates(def_block, pos));
573 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
577 /* else if this is a normal operation */
578 block = get_nodes_block(pos);
579 if(block == def_block) {
580 if(!sched_is_scheduled(val)) return 1;
582 ret = sched_comes_after(val, pos);
583 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
587 ret = block_dominates(def_block, block);
588 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
592 static INLINE ir_node *
593 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
595 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
599 * Returns first non-Phi node of block @p bb
601 static INLINE ir_node *
602 sched_block_first_nonphi(const ir_node * bb)
604 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
608 sched_skip_proj_predicator(const ir_node * irn, void * data)
610 return (is_Proj(irn));
613 static INLINE ir_node *
614 sched_next_nonproj(const ir_node * irn, int forward)
616 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
620 * Returns next operation node (non-Proj) after @p irn
621 * or the basic block of this node
623 static INLINE ir_node *
624 sched_next_op(const ir_node * irn)
626 ir_node *next = sched_next(irn);
631 return sched_next_nonproj(next, 1);
635 * Returns previous operation node (non-Proj) before @p irn
636 * or the basic block of this node
638 static INLINE ir_node *
639 sched_prev_op(const ir_node * irn)
641 ir_node *prev = sched_prev(irn);
646 return sched_next_nonproj(prev, 0);
650 sched_put_after(ir_node * insert, ir_node * irn)
652 if(is_Block(insert)) {
653 insert = sched_block_first_nonphi(insert);
655 insert = sched_next_op(insert);
657 sched_add_before(insert, irn);
661 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
663 if(is_Block(insert)) {
664 insert = sched_block_last_noncf(si, insert);
666 insert = sched_next_nonproj(insert, 0);
667 insert = sched_prev(insert);
669 sched_add_after(insert, irn);
673 * Tells you whether a @p remat can be placed before the irn @p pos
676 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
678 const ir_node *op = remat->op;
684 prev = sched_block_last_noncf(si, pos);
685 prev = sched_next_nonproj(prev, 0);
687 prev = sched_prev_op(pos);
689 /* do not remat if the rematted value is defined immediately before this op */
690 if(prev == remat->op) {
695 /* this should be just fine, the following OP will be using this value, right? */
697 /* only remat AFTER the real definition of a value (?) */
698 if(!value_is_defined_before(si, pos, remat->value)) {
699 // ir_fprintf(stderr, "error(not defined)");
704 for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
705 const ir_node *arg = get_irn_n(op, n);
707 #ifdef NO_ENLARGE_L1V3N355
708 if(has_reg_class(si, arg) && live) {
709 res &= pset_find_ptr(live, arg)?1:0;
711 res &= value_is_defined_before(si, pos, arg);
714 res &= value_is_defined_before(si, pos, arg);
722 * Tells you whether a @p remat can be placed after the irn @p pos
725 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
728 pos = sched_block_first_nonphi(pos);
730 pos = sched_next_op(pos);
733 /* only remat AFTER the real definition of a value (?) */
734 if(!value_is_defined_before(si, pos, remat->value)) {
738 return can_remat_before(si, remat, pos, live);
742 * Collect potetially rematerializable OPs
745 walker_remat_collector(ir_node * irn, void * data)
747 spill_ilp_t *si = data;
749 if(!is_Block(irn) && !is_Phi(irn)) {
750 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
751 get_remats_from_op(si, irn);
756 * Inserts a copy of @p irn before @p pos
759 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
764 bb = is_Block(pos)?pos:get_nodes_block(pos);
765 copy = exact_copy(irn);
767 _set_phi_class(copy, NULL);
768 set_nodes_block(copy, bb);
769 sched_put_before(si, pos, copy);
775 * Inserts a copy of @p irn after @p pos
778 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
783 bb = is_Block(pos)?pos:get_nodes_block(pos);
784 copy = exact_copy(irn);
786 _set_phi_class(copy, NULL);
787 set_nodes_block(copy, bb);
788 sched_put_after(pos, copy);
794 insert_remat_after(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
798 if(can_remat_after(si, remat, pos, live)) {
803 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
805 copy = insert_copy_after(si, remat->op, pos);
807 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
808 op = obstack_alloc(si->obst, sizeof(*op));
810 op->attr.remat.remat = remat;
811 op->attr.remat.pre = 0;
812 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
814 set_irn_link(copy, op);
815 pset_insert_ptr(si->all_possible_remats, copy);
817 proj_copy = insert_copy_after(si, remat->proj, copy);
818 set_irn_n(proj_copy, 0, copy);
819 set_irn_link(proj_copy, op);
820 pset_insert_ptr(si->all_possible_remats, proj_copy);
832 insert_remat_before(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
836 if(can_remat_before(si, remat, pos, live)) {
841 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
843 copy = insert_copy_before(si, remat->op, pos);
845 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
846 op = obstack_alloc(si->obst, sizeof(*op));
848 op->attr.remat.remat = remat;
849 op->attr.remat.pre = 1;
850 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
852 set_irn_link(copy, op);
853 pset_insert_ptr(si->all_possible_remats, copy);
855 proj_copy = insert_copy_after(si, remat->proj, copy);
856 set_irn_n(proj_copy, 0, copy);
857 set_irn_link(proj_copy, op);
858 pset_insert_ptr(si->all_possible_remats, proj_copy);
870 get_block_n_succs(const ir_node *block) {
871 const ir_edge_t *edge;
873 assert(edges_activated(current_ir_graph));
875 edge = get_block_succ_first(block);
879 edge = get_block_succ_next(block, edge);
884 is_merge_edge(const ir_node * bb)
886 #ifdef GOODWIN_REDUCTION
887 return get_block_n_succs(bb) == 1;
894 is_diverge_edge(const ir_node * bb)
896 #ifdef GOODWIN_REDUCTION
897 return get_Block_n_cfgpreds(bb) == 1;
904 walker_regclass_copy_insertor(ir_node * irn, void * data)
906 spill_ilp_t *si = data;
908 if(is_Phi(irn) && has_reg_class(si, irn)) {
911 for(n=get_irn_arity(irn)-1; n>=0; --n) {
912 ir_node *phi_arg = get_irn_n(irn, n);
913 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
915 if(!has_reg_class(si, phi_arg)) {
916 ir_node *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
917 ir_node *pos = sched_block_last_noncf(si, bb);
918 op_t *op = obstack_alloc(si->obst, sizeof(*op));
920 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
921 sched_add_after(pos, copy);
922 set_irn_n(irn, n, copy);
925 op->attr.live_range.args.reloads = NULL;
926 op->attr.live_range.ilp = ILP_UNDEF;
927 set_irn_link(copy, op);
935 * Insert (so far unused) remats into the irg to
936 * recompute the potential liveness of all values
939 walker_remat_insertor(ir_node * bb, void * data)
941 spill_ilp_t *si = data;
942 spill_bb_t *spill_bb;
945 pset *live = pset_new_ptr_default();
947 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
949 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
950 ir_node *value = be_lv_get_irn(si->lv, bb, i);
952 /* add remats at end of block */
953 if (has_reg_class(si, value)) {
954 pset_insert_ptr(live, value);
958 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
959 set_irn_link(bb, spill_bb);
961 irn = sched_last(bb);
962 while(!sched_is_end(irn)) {
969 next = sched_prev(irn);
971 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
973 if(is_Phi(irn) || is_Proj(irn)) {
976 if(has_reg_class(si, irn)) {
977 pset_remove_ptr(live, irn);
980 op = obstack_alloc(si->obst, sizeof(*op));
982 op->attr.live_range.args.reloads = NULL;
983 op->attr.live_range.ilp = ILP_UNDEF;
984 set_irn_link(irn, op);
990 op = obstack_alloc(si->obst, sizeof(*op));
992 op->attr.live_range.ilp = ILP_UNDEF;
993 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
994 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
995 set_irn_link(irn, op);
997 args = pset_new_ptr_default();
999 /* collect arguments of op */
1000 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1001 ir_node *arg = get_irn_n(irn, n);
1003 pset_insert_ptr(args, arg);
1006 /* set args of op already live in epilog */
1007 pset_foreach(args, arg) {
1008 if(has_reg_class(si, arg)) {
1009 pset_insert_ptr(live, arg);
1012 /* delete defined value from live set */
1013 if(has_reg_class(si, irn)) {
1014 pset_remove_ptr(live, irn);
1018 remat_args = pset_new_ptr_default();
1020 /* insert all possible remats before irn */
1021 pset_foreach(args, arg) {
1022 remat_info_t *remat_info,
1026 /* continue if the operand has the wrong reg class
1028 if(!has_reg_class(si, arg))
1032 query.remats = NULL;
1033 query.remats_by_operand = NULL;
1034 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1040 if(remat_info->remats) {
1041 pset_foreach(remat_info->remats, remat) {
1042 ir_node *remat_irn = NULL;
1044 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
1045 #ifdef REMAT_WHILE_LIVE
1046 if(pset_find_ptr(live, remat->value)) {
1047 remat_irn = insert_remat_before(si, remat, irn, live);
1050 remat_irn = insert_remat_before(si, remat, irn, live);
1053 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1054 ir_node *remat_arg = get_irn_n(remat_irn, n);
1056 if(!has_reg_class(si, remat_arg)) continue;
1058 pset_insert_ptr(remat_args, remat_arg);
1065 /* now we add remat args to op's args because they could also die at this op */
1066 pset_foreach(args,arg) {
1067 if(pset_find_ptr(remat_args, arg)) {
1068 pset_remove_ptr(remat_args, arg);
1071 pset_foreach(remat_args,arg) {
1072 pset_insert_ptr(args, arg);
1075 /* insert all possible remats after irn */
1076 pset_foreach(args, arg) {
1077 remat_info_t *remat_info,
1081 /* continue if the operand has the wrong reg class */
1082 if(!has_reg_class(si, arg))
1086 query.remats = NULL;
1087 query.remats_by_operand = NULL;
1088 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1094 /* do not place post remats after jumps */
1095 if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
1097 if(remat_info->remats_by_operand) {
1098 pset_foreach(remat_info->remats_by_operand, remat) {
1099 /* do not insert remats producing the same value as one of the operands */
1100 if(!pset_find_ptr(args, remat->value)) {
1101 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
1102 #ifdef REMAT_WHILE_LIVE
1103 if(pset_find_ptr(live, remat->value)) {
1104 insert_remat_after(si, remat, irn, live);
1107 insert_remat_after(si, remat, irn, live);
1114 del_pset(remat_args);
1119 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_in, i) {
1120 ir_node *value = be_lv_get_irn(si->lv, bb, i);
1122 /* add remats at end if successor has multiple predecessors */
1123 if(is_merge_edge(bb)) {
1124 /* add remats at end of block */
1125 if (be_is_live_end(si->lv, bb, value) && has_reg_class(si, value)) {
1126 remat_info_t *remat_info,
1131 query.remats = NULL;
1132 query.remats_by_operand = NULL;
1133 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1135 if(remat_info && remat_info->remats) {
1136 pset_foreach(remat_info->remats, remat) {
1137 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1139 insert_remat_before(si, remat, bb, NULL);
1144 if(is_diverge_edge(bb)) {
1145 /* add remat2s at beginning of block */
1146 if ((be_is_live_in(si->lv, bb, value) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) {
1147 remat_info_t *remat_info,
1152 query.remats = NULL;
1153 query.remats_by_operand = NULL;
1154 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1156 if(remat_info && remat_info->remats) {
1157 pset_foreach(remat_info->remats, remat) {
1158 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at beginning of block %+F\n", remat->op, bb));
1160 /* put the remat here if all its args are available */
1161 insert_remat_after(si, remat, bb, NULL);
1171 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1174 luke_endwalker(ir_node * bb, void * data)
1176 spill_ilp_t *si = (spill_ilp_t*)data;
1182 spill_bb_t *spill_bb = get_irn_link(bb);
1186 live = pset_new_ptr_default();
1187 use_end = pset_new_ptr_default();
1189 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1190 irn = be_lv_get_irn(si->lv, bb, i);
1191 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1194 pset_insert_ptr(live, irn);
1195 op = get_irn_link(irn);
1196 assert(!op->is_remat);
1200 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1201 /* their reg_out must always be set */
1202 sched_foreach_reverse(bb, irn) {
1205 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1207 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1208 ir_node *irn_arg = get_irn_n(irn, n);
1210 if(has_reg_class(si, irn_arg)) {
1211 pset_insert_ptr(use_end, irn_arg);
1216 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1217 //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1218 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1220 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1222 /* if this is a merge edge we can reload at the end of this block */
1223 if(is_merge_edge(bb)) {
1224 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1225 } else if(pset_count(use_end)){
1226 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1228 spill_bb->reloads = NULL;
1231 pset_foreach(live,irn) {
1237 /* handle values used by control flow nodes later separately */
1238 if(pset_find_ptr(use_end, irn)) continue;
1241 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1243 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1245 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1246 spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1247 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1249 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1250 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1252 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1253 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1255 if(is_merge_edge(bb)) {
1259 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1260 reload = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1261 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1263 /* reload <= mem_out */
1264 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1265 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1266 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1269 spill->reg_in = ILP_UNDEF;
1270 spill->mem_in = ILP_UNDEF;
1273 pset_foreach(use_end,irn) {
1277 ilp_cst_t end_use_req,
1282 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1284 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1286 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1287 spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1288 /* if irn is used at the end of the block, then it is live anyway */
1289 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1291 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1292 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1294 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1295 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1297 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1298 reload = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1299 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1301 /* reload <= mem_out */
1302 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1303 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1304 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1306 spill->reg_in = ILP_UNDEF;
1307 spill->mem_in = ILP_UNDEF;
1309 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1310 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1311 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1319 next_post_remat(const ir_node * irn)
1324 irn = sched_block_first_nonphi(irn);
1326 irn = sched_next_op(irn);
1329 if(sched_is_end(irn))
1332 op = (op_t*)get_irn_link(irn);
1333 if(op->is_remat && !op->attr.remat.pre) {
1342 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1348 ret = sched_block_last_noncf(si, irn);
1349 ret = sched_next(ret);
1350 ret = sched_prev_op(ret);
1352 ret = sched_prev_op(irn);
1355 if(sched_is_end(ret) || is_Phi(ret))
1358 op = (op_t*)get_irn_link(ret);
1359 if(op->is_remat && op->attr.remat.pre) {
1367 * Find a remat of value @p value in the epilog of @p pos
1370 find_post_remat(const ir_node * value, const ir_node * pos)
1372 while((pos = next_post_remat(pos)) != NULL) {
1375 op = get_irn_link(pos);
1376 assert(op->is_remat && !op->attr.remat.pre);
1378 if(op->attr.remat.remat->value == value)
1379 return (ir_node*)pos;
1382 const ir_edge_t *edge;
1383 foreach_out_edge(pos, edge) {
1384 ir_node *proj = get_edge_src_irn(edge);
1385 assert(is_Proj(proj));
1395 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1397 spill_bb_t *spill_bb = get_irn_link(bb);
1403 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1405 double spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1407 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1409 spill->reg_out = ILP_UNDEF;
1410 spill->reg_in = ILP_UNDEF;
1411 spill->mem_in = ILP_UNDEF;
1413 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1414 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1416 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1417 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1424 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1429 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1430 irn = be_lv_get_irn(si->lv, bb, i);
1432 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1433 pset_insert_ptr(live, irn);
1437 irn = sched_last(bb);
1439 /* all values eaten by control flow operations are also live until the end of the block */
1440 sched_foreach_reverse(bb, irn) {
1443 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1445 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1446 ir_node *arg = get_irn_n(irn,i);
1448 if(has_reg_class(si, arg)) {
1449 pset_insert_ptr(live, arg);
1456 * Inserts ILP-constraints and variables for memory copying before the given position
1459 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1461 const ir_node *succ;
1462 const ir_edge_t *edge;
1463 spill_bb_t *spill_bb = get_irn_link(block);
1472 assert(edges_activated(current_ir_graph));
1474 edge = get_block_succ_first(block);
1480 edge = get_block_succ_next(block, edge);
1481 /* next block can only contain phis, if this is a merge edge */
1484 ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1485 copyreg = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1487 ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1488 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1490 pset_foreach(live, tmp) {
1493 op_t *op = get_irn_link(irn);
1494 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1496 spill = set_find_spill(spill_bb->ilp, tmp);
1499 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1501 lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1503 sched_foreach(succ, phi) {
1504 const ir_node *to_copy;
1506 spill_t *to_copy_spill;
1507 op_t *phi_op = get_irn_link(phi);
1508 ilp_var_t reload = ILP_UNDEF;
1511 if(!is_Phi(phi)) break;
1512 if(!has_reg_class(si, phi)) continue;
1514 to_copy = get_irn_n(phi, pos);
1516 to_copy_op = get_irn_link(to_copy);
1518 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1519 assert(to_copy_spill);
1521 if(spill_bb->reloads) {
1522 keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1525 reload = PTR_TO_INT(keyval->val);
1529 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1530 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1532 /* copy - reg_out - reload - remat - live_range <= 0 */
1533 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1534 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1535 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1536 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1537 foreach_pre_remat(si, block, tmp) {
1538 op_t *remat_op = get_irn_link(tmp);
1539 if(remat_op->attr.remat.remat->value == to_copy) {
1540 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1544 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1545 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1547 /* copy - reg_out - copyreg <= 0 */
1548 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1549 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1550 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1556 * Walk all irg blocks and emit this ILP
1559 luke_blockwalker(ir_node * bb, void * data)
1561 spill_ilp_t *si = (spill_ilp_t*)data;
1566 spill_bb_t *spill_bb = get_irn_link(bb);
1569 pset *defs = pset_new_ptr_default();
1570 #ifdef WITH_MEMOPERANDS
1571 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
1575 live = pset_new_ptr_default();
1577 /****************************************
1578 * B A S I C B L O C K E N D
1579 ***************************************/
1582 /* init live values at end of block */
1583 get_live_end(si, bb, live);
1585 pset_foreach(live, irn) {
1587 ilp_var_t reload = ILP_UNDEF;
1589 spill = set_find_spill(spill_bb->ilp, irn);
1592 if(spill_bb->reloads) {
1593 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1596 reload = PTR_TO_INT(keyval->val);
1600 op = get_irn_link(irn);
1601 assert(!op->is_remat);
1603 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1604 op->attr.live_range.ilp = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1605 op->attr.live_range.op = bb;
1607 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1608 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1610 /* reg_out - reload - remat - live_range <= 0 */
1611 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1612 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1613 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1614 foreach_pre_remat(si, bb, tmp) {
1615 op_t *remat_op = get_irn_link(tmp);
1616 if(remat_op->attr.remat.remat->value == irn) {
1617 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1620 /* maybe we should also assure that reg_out >= live_range etc. */
1623 #ifndef NO_MEMCOPIES
1624 insert_mem_copy_position(si, live, bb);
1628 * start new live ranges for values used by remats at end of block
1629 * and assure the remat args are available
1631 foreach_pre_remat(si, bb, tmp) {
1632 op_t *remat_op = get_irn_link(tmp);
1635 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1636 ir_node *remat_arg = get_irn_n(tmp, n);
1637 op_t *arg_op = get_irn_link(remat_arg);
1640 if(!has_reg_class(si, remat_arg)) continue;
1642 /* if value is becoming live through use by remat */
1643 if(!pset_find_ptr(live, remat_arg)) {
1644 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1645 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1647 arg_op->attr.live_range.ilp = prev_lr;
1648 arg_op->attr.live_range.op = bb;
1650 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1652 pset_insert_ptr(live, remat_arg);
1653 add_to_spill_bb(si, bb, remat_arg);
1656 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1657 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1658 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1660 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1661 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1663 /* use reload placed for this argument */
1664 if(spill_bb->reloads) {
1665 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1668 ilp_var_t reload = PTR_TO_INT(keyval->val);
1670 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1675 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1680 /**************************************
1681 * B A S I C B L O C K B O D Y
1682 **************************************/
1684 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1690 ilp_cst_t check_pre,
1696 #ifdef WITH_MEMOPERANDS
1697 ilp_cst_t one_memoperand;
1700 /* iterate only until first phi */
1704 op = get_irn_link(irn);
1706 if(op->is_remat) continue;
1707 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1709 /* collect defined values */
1710 if(has_reg_class(si, irn)) {
1711 pset_insert_ptr(defs, irn);
1715 if(is_Proj(irn)) continue;
1718 * init set of irn's arguments
1719 * and all possibly used values around this op
1720 * and values defined by post remats
1722 args = new_set(cmp_keyval, get_irn_arity(irn));
1723 used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1724 remat_defs = pset_new_ptr(pset_count(live));
1726 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1727 ir_node *irn_arg = get_irn_n(irn, n);
1728 if(has_reg_class(si, irn_arg)) {
1729 set_insert_keyval(args, irn_arg, (void*)n);
1730 pset_insert_ptr(used, irn_arg);
1733 foreach_post_remat(irn, tmp) {
1734 op_t *remat_op = get_irn_link(tmp);
1736 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1738 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1739 ir_node *remat_arg = get_irn_n(tmp, n);
1740 if(has_reg_class(si, remat_arg)) {
1741 pset_insert_ptr(used, remat_arg);
1745 foreach_pre_remat(si, irn, tmp) {
1746 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1747 ir_node *remat_arg = get_irn_n(tmp, n);
1748 if(has_reg_class(si, remat_arg)) {
1749 pset_insert_ptr(used, remat_arg);
1754 /**********************************
1755 * I N E P I L O G O F irn
1756 **********************************/
1758 /* ensure each dying value is used by only one post remat */
1759 pset_foreach(used, tmp) {
1760 ir_node *value = tmp;
1761 op_t *value_op = get_irn_link(value);
1766 foreach_post_remat(irn, remat) {
1767 op_t *remat_op = get_irn_link(remat);
1769 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1770 ir_node *remat_arg = get_irn_n(remat, n);
1772 /* if value is used by this remat add it to constraint */
1773 if(remat_arg == value) {
1775 /* sum remat2s <= 1 + n_remats*live_range */
1776 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1777 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
1781 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1787 // value_op->attr.live_range.ilp != ILP_UNDEF
1788 if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
1789 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1793 /* ensure at least one value dies at post remat */
1794 foreach_post_remat(irn, tmp) {
1795 op_t *remat_op = get_irn_link(tmp);
1796 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
1799 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
1800 remat_arg = get_irn_n(tmp, n);
1802 if(has_reg_class(si, remat_arg)) {
1804 /* does arg always die at this op? */
1805 if(!pset_find_ptr(live, remat_arg))
1806 goto skip_one_must_die;
1808 pset_insert_ptr(remat_args, remat_arg);
1812 /* remat + \sum live_range(remat_arg) <= |args| */
1813 ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
1814 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
1815 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1817 pset_foreach(remat_args, remat_arg) {
1818 op_t *arg_op = get_irn_link(remat_arg);
1820 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1824 del_pset(remat_args);
1827 /* new live ranges for values from L\U defined by post remats */
1828 pset_foreach(live, tmp) {
1829 ir_node *value = tmp;
1830 op_t *value_op = get_irn_link(value);
1832 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
1833 ilp_var_t prev_lr = ILP_UNDEF;
1836 if(pset_find_ptr(remat_defs, value)) {
1838 /* next_live_range <= prev_live_range + sum remat2s */
1839 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1840 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1842 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1843 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1845 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1846 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1848 foreach_post_remat(irn, remat) {
1849 op_t *remat_op = get_irn_link(remat);
1851 /* if value is being rematerialized by this remat */
1852 if(value == remat_op->attr.remat.remat->value) {
1853 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1857 value_op->attr.live_range.ilp = prev_lr;
1858 value_op->attr.live_range.op = irn;
1863 /* requirements for post remats and start live ranges from L/U' for values dying here */
1864 foreach_post_remat(irn, tmp) {
1865 op_t *remat_op = get_irn_link(tmp);
1868 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1869 ir_node *remat_arg = get_irn_n(tmp, n);
1870 op_t *arg_op = get_irn_link(remat_arg);
1872 if(!has_reg_class(si, remat_arg)) continue;
1874 /* only for values in L\U (TODO and D?), the others are handled with post_use */
1875 if(!pset_find_ptr(used, remat_arg)) {
1876 /* remat <= live_range(remat_arg) */
1877 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
1878 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1880 /* if value is becoming live through use by remat2 */
1881 if(!pset_find_ptr(live, remat_arg)) {
1884 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1885 lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1887 arg_op->attr.live_range.ilp = lr;
1888 arg_op->attr.live_range.op = irn;
1890 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
1892 pset_insert_ptr(live, remat_arg);
1893 add_to_spill_bb(si, bb, remat_arg);
1896 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1897 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1902 d = pset_count(defs);
1903 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
1905 /* count how many regs irn needs for arguments */
1906 u = set_count(args);
1909 /* check the register pressure in the epilog */
1910 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
1911 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
1912 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
1914 /* add L\U' to check_post */
1915 pset_foreach(live, tmp) {
1916 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
1917 /* if a live value is not used by irn */
1918 tmp_op = get_irn_link(tmp);
1919 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
1923 /***********************************************************
1924 * 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
1925 **********************************************************/
1928 pset_foreach(used, tmp) {
1934 op_t *arg_op = get_irn_link(arg);
1937 spill = add_to_spill_bb(si, bb, arg);
1939 /* new live range for each used value */
1940 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
1941 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1943 /* the epilog stuff - including post_use, check_post, check_post_remat */
1944 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
1945 post_use = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1947 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
1949 /* arg is live throughout epilog if the next live_range is in a register */
1950 if(pset_find_ptr(live, arg)) {
1951 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
1953 /* post_use >= next_lr + remat */
1954 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1955 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1956 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
1957 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1961 /* if value is not an arg of op and not possibly defined by post remat
1962 * then it may only die and not become live
1964 if(!set_find_keyval(args, arg)) {
1965 /* post_use <= prev_lr */
1966 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
1967 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1968 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1969 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1971 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
1972 /* next_lr <= prev_lr */
1973 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
1974 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1975 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1976 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1981 /* forall post remat which use arg add a similar cst */
1982 foreach_post_remat(irn, remat) {
1985 for (n=get_irn_arity(remat)-1; n>=0; --n) {
1986 ir_node *remat_arg = get_irn_n(remat, n);
1987 op_t *remat_op = get_irn_link(remat);
1989 if(remat_arg == arg) {
1990 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
1992 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1993 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1994 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
1995 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2000 #ifdef WITH_MEMOPERANDS
2001 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2002 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2003 ilp_var_t memoperand;
2005 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2006 memoperand = lpp_add_var(si->lpp, buf, lpp_binary, COST_MEMOPERAND*execution_frequency(si, bb));
2007 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2009 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2010 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2012 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2013 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2014 // if(arg_op->attr.live_range.ilp != ILP_UNDEF)
2015 // lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2020 /* new live range begins for each used value */
2021 arg_op->attr.live_range.ilp = prev_lr;
2022 arg_op->attr.live_range.op = irn;
2024 /*if(!pset_find_ptr(live, arg)) {
2025 pset_insert_ptr(live, arg);
2026 add_to_spill_bb(si, bb, arg);
2028 pset_insert_ptr(live, arg);
2032 /* just to be sure */
2033 check_post = ILP_UNDEF;
2042 /* check the register pressure in the prolog */
2043 /* sum_{L\U} lr <= k - |U| */
2044 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2045 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2047 /* for the prolog remove defined values from the live set */
2048 pset_foreach(defs, tmp) {
2049 pset_remove_ptr(live, tmp);
2052 #ifdef WITH_MEMOPERANDS
2053 ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2054 one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2057 /***********************************************************
2058 * 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
2059 **********************************************************/
2062 set_foreach(args, keyval) {
2064 ir_node *arg = keyval->key;
2065 int i = PTR_TO_INT(keyval->val);
2066 op_t *arg_op = get_irn_link(arg);
2067 ilp_cst_t requirements;
2068 #ifdef WITH_MEMOPERANDS
2072 spill = set_find_spill(spill_bb->ilp, arg);
2075 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2076 op->attr.live_range.args.reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
2078 /* reload <= mem_out */
2079 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2080 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2081 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2082 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2084 /* requirement: arg must be in register for use */
2085 /* reload + remat + live_range == 1 */
2086 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2087 requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2089 lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2090 lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2091 foreach_pre_remat(si, irn, tmp) {
2092 op_t *remat_op = get_irn_link(tmp);
2093 if(remat_op->attr.remat.remat->value == arg) {
2094 lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2098 #ifdef WITH_MEMOPERANDS
2100 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2101 if(get_irn_n(irn, n) == arg) {
2105 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2106 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2107 memoperand_t *memoperand;
2108 memoperand = set_find_memoperand(si->memoperands, irn, n);
2110 /* memoperand <= mem_out */
2111 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2112 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2113 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2114 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2116 /* the memoperand is only sufficient if it is used once by the op */
2117 if(n_memoperands == 1)
2118 lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2120 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2122 /* we have one more free register if we use a memory operand */
2123 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2129 /* iterate over L\U */
2130 pset_foreach(live, tmp) {
2131 if(!set_find_keyval(args, tmp)) {
2132 /* if a live value is not used by irn */
2133 tmp_op = get_irn_link(tmp);
2134 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2139 /* requirements for remats */
2140 /* start new live ranges for values used by remats */
2141 foreach_pre_remat(si, irn, tmp) {
2142 op_t *remat_op = get_irn_link(tmp);
2145 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2146 ir_node *remat_arg = get_irn_n(tmp, n);
2147 op_t *arg_op = get_irn_link(remat_arg);
2150 if(!has_reg_class(si, remat_arg)) continue;
2152 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2153 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2154 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2156 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2157 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2159 /* if remat arg is also used by current op then we can use reload placed for this argument */
2160 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2161 int index = (int)keyval->val;
2163 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2171 /*************************
2172 * D O N E W I T H O P
2173 *************************/
2175 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2177 pset_foreach(live, tmp) {
2178 assert(has_reg_class(si, tmp));
2181 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2182 ir_node *arg = get_irn_n(irn, n);
2184 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2187 del_pset(remat_defs);
2191 defs = pset_new_ptr_default();
2196 /***************************************
2197 * B E G I N N I N G O F B L O C K
2198 ***************************************/
2201 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2202 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2204 pset_foreach(live, irn) {
2205 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2208 /* construct mem_outs for all values */
2210 set_foreach(spill_bb->ilp, spill) {
2211 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2212 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2214 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2215 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2217 if(pset_find_ptr(live, spill->irn)) {
2218 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2220 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2221 spill->mem_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2222 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2224 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2226 op_t *op = get_irn_link(spill->irn);
2228 /* do we have to copy a phi argument? */
2229 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2230 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2232 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2233 const ir_node *arg = get_irn_n(spill->irn, n);
2239 /* argument already done? */
2240 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2242 /* get sum of execution frequencies of blocks with the same phi argument */
2243 for(m=n; m>=0; --m) {
2244 const ir_node *arg2 = get_irn_n(spill->irn, m);
2247 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2251 /* copies are not for free */
2252 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2253 var = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE * freq);
2255 for(m=n; m>=0; --m) {
2256 const ir_node *arg2 = get_irn_n(spill->irn, m);
2259 op->attr.live_range.args.copies[m] = var;
2263 /* copy <= mem_in */
2264 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2265 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2266 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2267 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2274 /* L\U is empty at bb start */
2275 /* arg is live throughout epilog if it is reg_in into this block */
2277 /* check the register pressure at the beginning of the block
2280 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2281 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2283 pset_foreach(live, irn) {
2286 spill = set_find_spill(spill_bb->ilp, irn);
2289 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2290 spill->reg_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2292 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2294 /* spill + mem_in <= 1 */
2295 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2296 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2298 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2299 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2302 foreach_post_remat(bb, irn) {
2303 op_t *remat_op = get_irn_link(irn);
2305 DBG((si->dbg, LEVEL_4, "\t next post remat: %+F\n", irn));
2306 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
2308 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2311 /* forall post remats add requirements */
2312 foreach_post_remat(bb, tmp) {
2315 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2316 ir_node *remat_arg = get_irn_n(tmp, n);
2317 op_t *remat_op = get_irn_link(tmp);
2319 if(!has_reg_class(si, remat_arg)) continue;
2321 spill = set_find_spill(spill_bb->ilp, remat_arg);
2324 /* remat <= reg_in_argument */
2325 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2326 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2327 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2328 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2332 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2333 pset_foreach(live, irn) {
2337 spill = set_find_spill(spill_bb->ilp, irn);
2338 assert(spill && spill->irn == irn);
2340 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2341 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2344 ir_node *phi_arg = get_Phi_pred(irn, n);
2345 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2346 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2348 op_t *op = get_irn_link(irn);
2350 /* although the phi is in the right regclass one or more of
2351 * its arguments can be in a different one or at least to
2354 if(has_reg_class(si, phi_arg)) {
2355 /* mem_in < mem_out_arg + copy */
2356 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2357 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2359 /* reg_in < reg_out_arg */
2360 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2361 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2363 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2364 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2366 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2369 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2370 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2371 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2375 /* else assure the value arrives on all paths in the same resource */
2377 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2380 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2381 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2384 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2385 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2386 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2387 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2389 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2390 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2392 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2395 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2396 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2401 /* first live ranges from reg_ins */
2402 pset_foreach(live, irn) {
2403 op_t *op = get_irn_link(irn);
2405 spill = set_find_spill(spill_bb->ilp, irn);
2406 assert(spill && spill->irn == irn);
2408 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2409 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2410 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2411 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2413 foreach_post_remat(bb, tmp) {
2414 op_t *remat_op = get_irn_link(tmp);
2416 if(remat_op->attr.remat.remat->value == irn) {
2417 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2422 /* walk forward now and compute constraints for placing spills */
2423 /* this must only be done for values that are not defined in this block */
2424 /* TODO are these values at start of block? if yes, just check whether this is a diverge edge and skip the loop */
2425 pset_foreach(live, irn) {
2427 * if value is defined in this block we can anways place the spill directly after the def
2428 * -> no constraint necessary
2430 if(!is_Phi(irn) && get_nodes_block(irn) == bb) continue;
2433 spill = set_find_spill(spill_bb->ilp, irn);
2436 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2437 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2439 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2440 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2443 sched_foreach_op(bb, tmp) {
2444 op_t *op = get_irn_link(tmp);
2446 if(is_Phi(tmp)) continue;
2447 assert(!is_Proj(tmp));
2450 ir_node *value = op->attr.remat.remat->value;
2453 /* only collect remats up to the first real use of a value */
2454 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2459 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2460 ir_node *arg = get_irn_n(tmp, n);
2463 /* if a value is used stop collecting remats */
2470 /* if the semicolon is missing, the microsoft compiler complains */
2477 typedef struct _irnlist_t {
2478 struct list_head list;
2482 typedef struct _interference_t {
2483 struct list_head blocklist;
2489 cmp_interference(const void *a, const void *b, size_t size)
2491 const interference_t *p = a;
2492 const interference_t *q = b;
2494 return !(p->a == q->a && p->b == q->b);
2497 static interference_t *
2498 set_find_interference(set * set, ir_node * a, ir_node * b)
2500 interference_t query;
2502 query.a = (a>b)?a:b;
2503 query.b = (a>b)?b:a;
2505 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2508 static interference_t *
2509 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2511 interference_t query,
2513 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2517 result = set_find_interference(set, a, b);
2520 list_add(&list->list, &result->blocklist);
2524 query.a = (a>b)?a:b;
2525 query.b = (a>b)?b:a;
2527 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2529 INIT_LIST_HEAD(&result->blocklist);
2530 list_add(&list->list, &result->blocklist);
2536 values_interfere_in_block(const spill_ilp_t * si, ir_node * bb, ir_node * a, ir_node * b)
2538 const ir_edge_t *edge;
2540 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2541 /* both values are live in, so they interfere */
2545 /* ensure a dominates b */
2546 if(value_dominates(b,a)) {
2552 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2555 /* the following code is stolen from bera.c */
2556 if(be_is_live_end(si->lv, bb, a))
2559 foreach_out_edge(a, edge) {
2560 const ir_node *user = edge->src;
2561 if(get_nodes_block(user) == bb
2564 && value_dominates(b, user))
2572 * Walk all irg blocks and collect interfering values inside of phi classes
2575 luke_interferencewalker(ir_node * bb, void * data)
2577 spill_ilp_t *si = (spill_ilp_t*)data;
2580 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2581 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
2582 op_t *a_op = get_irn_link(a);
2584 if(a_op->is_remat) continue;
2586 /* a is only interesting if it is in my register class and if it is inside a phi class */
2587 if (has_reg_class(si, a) && get_phi_class(a)) {
2588 for(l2=_be_lv_next_irn(si->lv, bb, 0xff, l1+1); l2>=0; l2=_be_lv_next_irn(si->lv, bb, 0xff, l2+1)) {
2589 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
2590 op_t *b_op = get_irn_link(b);
2592 if(b_op->is_remat) continue;
2594 /* a and b are only interesting if they are in the same phi class */
2595 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2596 if(values_interfere_in_block(si, bb, a, b)) {
2597 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2598 set_insert_interference(si, si->interferences, a, b, bb);
2606 static unsigned int copy_path_id = 0;
2609 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2616 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2617 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2619 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2621 pset_foreach(copies, ptr) {
2622 copy = PTR_TO_INT(ptr);
2623 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2628 * @parameter copies contains a path of copies which lead us to irn
2629 * @parameter visited contains a set of nodes already visited on this path
2632 find_copy_path(spill_ilp_t * si, ir_node * irn, ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2635 op_t *op = get_irn_link(irn);
2636 pset *visited_users = pset_new_ptr_default();
2639 if(op->is_remat) return;
2641 pset_insert_ptr(visited, irn);
2645 pset *visited_operands = pset_new_ptr(get_irn_arity(irn));
2647 /* visit all operands */
2648 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2649 ir_node *arg = get_irn_n(irn, n);
2650 ilp_var_t copy = op->attr.live_range.args.copies[n];
2652 if(!has_reg_class(si, arg)) continue;
2653 if(pset_find_ptr(visited_operands, arg)) continue;
2654 pset_insert_ptr(visited_operands, arg);
2657 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2658 del_pset(visited_operands);
2659 del_pset(visited_users);
2660 pset_remove_ptr(visited, irn);
2663 pset_insert(copies, INT_TO_PTR(copy), copy);
2664 write_copy_path_cst(si, copies, any_interfere);
2665 pset_remove(copies, INT_TO_PTR(copy), copy);
2666 } else if(!pset_find_ptr(visited, arg)) {
2667 pset_insert(copies, INT_TO_PTR(copy), copy);
2668 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
2669 pset_remove(copies, INT_TO_PTR(copy), copy);
2671 /*if(paths > MAX_PATHS) {
2672 if(pset_count(copies) == 0) {
2676 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2677 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2678 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2679 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2680 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2684 del_pset(visited_operands);
2685 del_pset(visited_users);
2686 pset_remove_ptr(visited, irn);
2689 } else if(pset_count(copies) == 0) {
2695 del_pset(visited_operands);
2698 /* visit all uses which are phis */
2699 foreach_out_edge(irn, edge) {
2700 ir_node *user = edge->src;
2701 int pos = edge->pos;
2702 op_t *op = get_irn_link(user);
2705 if(!is_Phi(user)) continue;
2706 if(!has_reg_class(si, user)) continue;
2707 if(pset_find_ptr(visited_users, user)) continue;
2708 pset_insert_ptr(visited_users, user);
2710 copy = op->attr.live_range.args.copies[pos];
2712 if(user == target) {
2713 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2714 del_pset(visited_users);
2715 pset_remove_ptr(visited, irn);
2718 pset_insert(copies, INT_TO_PTR(copy), copy);
2719 write_copy_path_cst(si, copies, any_interfere);
2720 pset_remove(copies, INT_TO_PTR(copy), copy);
2721 } else if(!pset_find_ptr(visited, user)) {
2722 pset_insert(copies, INT_TO_PTR(copy), copy);
2723 paths += find_copy_path(si, user, target, any_interfere, copies, visited);
2724 pset_remove(copies, INT_TO_PTR(copy), copy);
2726 /*if(paths > MAX_PATHS) {
2727 if(pset_count(copies) == 0) {
2731 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2732 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2733 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2734 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2735 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2739 del_pset(visited_users);
2740 pset_remove_ptr(visited, irn);
2743 } else if(pset_count(copies) == 0) {
2749 del_pset(visited_users);
2750 pset_remove_ptr(visited, irn);
2755 gen_copy_constraints(spill_ilp_t * si, ir_node * a, ir_node * b, ilp_var_t any_interfere)
2757 pset * copies = pset_new_ptr_default();
2758 pset * visited = pset_new_ptr_default();
2760 find_copy_path(si, a, b, any_interfere, copies, visited);
2768 memcopyhandler(spill_ilp_t * si)
2770 interference_t *interference;
2772 /* teste Speicherwerte auf Interferenz */
2774 /* analyze phi classes */
2775 phi_class_compute(si->chordal_env->irg);
2777 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
2778 irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
2780 /* now lets emit the ILP unequations for the crap */
2781 set_foreach(si->interferences, interference) {
2783 ilp_var_t interfere,
2785 ilp_cst_t any_interfere_cst,
2787 const ir_node *a = interference->a;
2788 const ir_node *b = interference->b;
2790 /* any_interf <= \sum interf */
2791 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
2792 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2793 any_interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2795 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
2797 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
2798 const ir_node *bb = irnlist->irn;
2799 spill_bb_t *spill_bb = get_irn_link(bb);
2806 spilla = set_find_spill(spill_bb->ilp, a);
2810 spillb = set_find_spill(spill_bb->ilp, b);
2813 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
2814 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
2815 /* 2: - mem_in_a - spill_a + interfere <= 0 */
2816 /* 3: - mem_in_b - spill_b + interfere <= 0 */
2817 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
2818 interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2820 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
2821 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2823 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
2824 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
2825 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
2826 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
2827 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
2829 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
2830 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2832 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2833 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
2834 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
2836 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
2837 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2839 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2840 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
2841 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
2844 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
2846 /* any_interfere >= interf */
2847 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
2848 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2850 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2851 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2854 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
2855 gen_copy_constraints(si,a,b,any_interfere);
2863 return fabs(x) < 0.00001;
2867 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
2869 spill_ilp_t *si = get_irg_link(current_ir_graph);
2871 if(pset_find_ptr(si->all_possible_remats, n)) {
2872 op_t *op = (op_t*)get_irn_link(n);
2873 assert(op && op->is_remat);
2875 if(!op->attr.remat.remat->inverse) {
2876 if(op->attr.remat.pre) {
2877 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
2879 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
2884 op_t *op = (op_t*)get_irn_link(n);
2885 assert(op && op->is_remat);
2887 if(op->attr.remat.pre) {
2888 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
2890 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
2901 dump_graph_with_remats(ir_graph * irg, const char * suffix)
2903 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
2904 be_dump(irg, suffix, dump_ir_block_graph_sched);
2905 set_dump_node_vcgattr_hook(NULL);
2910 * Edge hook to dump the schedule edges with annotated register pressure.
2913 sched_pressure_edge_hook(FILE *F, ir_node *irn)
2915 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
2916 ir_node *prev = sched_prev(irn);
2917 fprintf(F, "edge:{sourcename:\"");
2919 fprintf(F, "\" targetname:\"");
2921 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
2922 fprintf(F, "\" color:magenta}\n");
2928 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
2930 DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
2932 dump_consts_local(0);
2933 set_dump_node_edge_hook(sched_pressure_edge_hook);
2934 dump_ir_block_graph(irg, suffix);
2935 set_dump_node_edge_hook(old_edge_hook);
2939 walker_pressure_annotator(ir_node * bb, void * data)
2941 spill_ilp_t *si = data;
2944 pset *live = pset_new_ptr_default();
2947 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
2948 irn = be_lv_get_irn(si->lv, bb, i);
2950 if (has_reg_class(si, irn)) {
2951 pset_insert_ptr(live, irn);
2955 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
2957 sched_foreach_reverse(bb, irn) {
2959 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
2963 if(has_reg_class(si, irn)) {
2964 pset_remove_ptr(live, irn);
2965 if(is_Proj(irn)) ++projs;
2968 if(!is_Proj(irn)) projs = 0;
2970 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2971 ir_node *arg = get_irn_n(irn, n);
2973 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
2975 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
2982 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
2984 be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
2989 connect_all_remats_with_keep(spill_ilp_t * si)
2997 n_remats = pset_count(si->all_possible_remats);
2999 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3002 pset_foreach(si->all_possible_remats, irn) {
3007 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
3009 obstack_free(si->obst, ins);
3015 connect_all_spills_with_keep(spill_ilp_t * si)
3024 n_spills = pset_count(si->spills);
3026 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3029 pset_foreach(si->spills, irn) {
3034 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
3036 obstack_free(si->obst, ins);
3040 /** insert a spill at an arbitrary position */
3041 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert, ir_node *ctx)
3043 ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
3044 ir_graph *irg = get_irn_irg(bl);
3045 ir_node *frame = get_irg_frame(irg);
3049 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
3050 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3052 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn, ctx);
3055 * search the right insertion point. a spill of a phi cannot be put
3056 * directly after the phi, if there are some phis behind the one which
3057 * is spilled. Also, a spill of a Proj must be after all Projs of the
3060 * Here's one special case:
3061 * If the spill is in the start block, the spill must be after the frame
3062 * pointer is set up. This is done by setting insert to the end of the block
3063 * which is its default initialization (see above).
3066 if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3069 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3072 sched_add_after(insert, spill);
3077 delete_remat(spill_ilp_t * si, ir_node * remat) {
3079 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3081 sched_remove(remat);
3083 /* kill links to operands */
3084 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3085 set_irn_n(remat, n, bad);
3090 clean_remat_info(spill_ilp_t * si)
3094 remat_info_t *remat_info;
3095 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3097 set_foreach(si->remat_info, remat_info) {
3098 if(!remat_info->remats) continue;
3100 pset_foreach(remat_info->remats, remat)
3102 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3103 set_irn_n(remat->proj, -1, bad);
3104 set_irn_n(remat->proj, 0, bad);
3107 if(get_irn_n_edges(remat->op) == 0) {
3108 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3109 set_irn_n(remat->op, n, bad);
3114 if(remat_info->remats) del_pset(remat_info->remats);
3115 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3120 delete_unnecessary_remats(spill_ilp_t * si)
3124 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3127 ir_node *end = get_irg_end(si->chordal_env->irg);
3130 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3131 ir_node *keep_arg = get_irn_n(si->keep, n);
3132 op_t *arg_op = get_irn_link(keep_arg);
3135 assert(arg_op->is_remat);
3137 name = si->lpp->vars[arg_op->attr.remat.ilp];
3139 if(is_zero(name->value)) {
3140 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
3141 /* TODO check whether reload is preferred over remat (could be bug) */
3142 delete_remat(si, keep_arg);
3144 if(!arg_op->attr.remat.remat->inverse) {
3145 if(arg_op->attr.remat.pre) {
3146 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3148 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3151 if(arg_op->attr.remat.pre) {
3152 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3154 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3159 set_irn_n(si->keep, n, bad);
3162 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
3163 ir_node *end_arg = get_End_keepalive(end, i);
3165 if(end_arg != si->keep) {
3166 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
3169 keeps = obstack_finish(si->obst);
3170 set_End_keepalives(end, n-1, keeps);
3171 obstack_free(si->obst, keeps);
3174 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
3179 pset_foreach(si->all_possible_remats, remat) {
3180 op_t *remat_op = get_irn_link(remat);
3181 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
3183 if(is_zero(name->value)) {
3184 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
3185 /* TODO check whether reload is preferred over remat (could be bug) */
3186 delete_remat(si, remat);
3188 if(!remat_op->attr.remat.remat->inverse) {
3189 if(remat_op->attr.remat.pre) {
3190 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3192 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3195 if(remat_op->attr.remat.pre) {
3196 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3198 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3207 get_spills_for_value(spill_ilp_t * si, ir_node * value)
3209 pset *spills = pset_new_ptr_default();
3214 defs = set_find_def(si->values, value);
3216 if(defs && defs->spills) {
3217 for(next = defs->spills; next; next = get_irn_link(next)) {
3218 pset_insert_ptr(spills, next);
3226 get_remats_for_value(spill_ilp_t * si, ir_node * value)
3228 pset *remats = pset_new_ptr_default();
3233 pset_insert_ptr(remats, value);
3234 defs = set_find_def(si->values, value);
3236 if(defs && defs->remats) {
3237 for(next = defs->remats; next; next = get_irn_link(next)) {
3238 pset_insert_ptr(remats, next);
3247 * @param before The node after which the spill will be placed in the schedule
3249 /* TODO set context properly */
3251 insert_spill(spill_ilp_t * si, ir_node * irn, ir_node * value, ir_node * before)
3255 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3257 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3259 spill = be_spill2(arch_env, irn, before, irn);
3261 defs = set_insert_def(si->values, value);
3264 /* enter into the linked list */
3265 set_irn_link(spill, defs->spills);
3266 defs->spills = spill;
3268 #ifdef KEEPALIVE_SPILLS
3269 pset_insert_ptr(si->spills, spill);
3276 * @param before The Phi node which has to be spilled
3279 insert_mem_phi(spill_ilp_t * si, const ir_node * phi)
3285 op_t *op = get_irn_link(phi);
3287 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3289 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3290 ins[n] = si->m_unknown;
3293 mem_phi = new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3295 defs = set_insert_def(si->values, phi);
3298 /* enter into the linked list */
3299 set_irn_link(mem_phi, defs->spills);
3300 defs->spills = mem_phi;
3302 sched_add_after(phi, mem_phi);
3304 #ifdef KEEPALIVE_SPILLS
3305 pset_insert_ptr(si->spills, mem_phi);
3313 * Add remat to list of defs, destroys link field!
3316 insert_remat(spill_ilp_t * si, ir_node * remat)
3319 op_t *remat_op = get_irn_link(remat);
3321 assert(remat_op->is_remat);
3323 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3326 /* enter into the linked list */
3327 set_irn_link(remat, defs->remats);
3328 defs->remats = remat;
3333 * Add reload before operation and add to list of defs
3336 insert_reload(spill_ilp_t * si, const ir_node * value, const ir_node * after)
3341 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3343 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3345 defs = set_find_def(si->values, value);
3347 spill = defs->spills;
3348 assert(spill && "no spill placed before reload");
3350 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3352 /* enter into the linked list */
3353 set_irn_link(reload, defs->remats);
3354 defs->remats = reload;
3359 #ifdef WITH_MEMOPERANDS
3360 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3364 ir_node *value = get_irn_n(memoperand->irn, memoperand->pos);
3366 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3368 DBG((si->dbg, LEVEL_2, "\t inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3370 defs = set_find_def(si->values, value);
3372 spill = defs->spills;
3373 assert(spill && "no spill placed before reload");
3375 reload = be_reload(arch_env, si->cls, memoperand->irn, get_irn_mode(value), spill);
3377 arch_perform_memory_operand(arch_env, memoperand->irn, reload, memoperand->pos);
3378 sched_remove(reload);
3381 void insert_memoperands(spill_ilp_t * si)
3383 memoperand_t *memoperand;
3386 set_foreach(si->memoperands, memoperand) {
3387 name = si->lpp->vars[memoperand->ilp];
3388 if(!is_zero(name->value)) {
3389 perform_memory_operand(si, memoperand);
3396 walker_spill_placer(ir_node * bb, void * data) {
3397 spill_ilp_t *si = (spill_ilp_t*)data;
3399 spill_bb_t *spill_bb = get_irn_link(bb);
3400 pset *spills_to_do = pset_new_ptr_default();
3403 set_foreach(spill_bb->ilp, spill) {
3406 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3407 name = si->lpp->vars[spill->mem_in];
3408 if(!is_zero(name->value)) {
3411 mem_phi = insert_mem_phi(si, spill->irn);
3413 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3417 name = si->lpp->vars[spill->spill];
3418 if(!is_zero(name->value)) {
3419 /* place spill directly after definition */
3420 if(get_nodes_block(spill->irn) == bb) {
3421 insert_spill(si, spill->irn, spill->irn, spill->irn);
3425 /* place spill at bb start */
3426 if(spill->reg_in > 0) {
3427 name = si->lpp->vars[spill->reg_in];
3428 if(!is_zero(name->value)) {
3429 insert_spill(si, spill->irn, spill->irn, bb);
3433 /* place spill after a remat */
3434 pset_insert_ptr(spills_to_do, spill->irn);
3437 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3440 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3441 op_t *op = get_irn_link(irn);
3443 if(be_is_Spill(irn)) continue;
3446 /* TODO fix this if we want to support remats with more than two nodes */
3447 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3448 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3450 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3453 if(pset_find_ptr(spills_to_do, irn)) {
3454 pset_remove_ptr(spills_to_do, irn);
3456 insert_spill(si, irn, irn, irn);
3462 assert(pset_count(spills_to_do) == 0);
3464 /* afterwards free data in block */
3465 del_pset(spills_to_do);
3469 insert_mem_copy(spill_ilp_t * si, const ir_node * bb, const ir_node * value)
3471 ir_node *insert_pos = bb;
3473 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3475 /* find last definition of arg value in block */
3480 defs = set_find_def(si->values, value);
3482 if(defs && defs->remats) {
3483 for(next = defs->remats; next; next = get_irn_link(next)) {
3484 if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3485 last = sched_get_time_step(next);
3491 if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3492 last = sched_get_time_step(value);
3496 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
3498 spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos, value);
3504 phim_fixer(spill_ilp_t *si) {
3507 set_foreach(si->values, defs) {
3508 const ir_node *phi = defs->value;
3509 op_t *op = get_irn_link(phi);
3510 ir_node *phi_m = NULL;
3511 ir_node *next = defs->spills;
3514 if(!is_Phi(phi)) continue;
3517 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3521 next = get_irn_link(next);
3524 if(!phi_m) continue;
3526 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3527 const ir_node *value = get_irn_n(phi, n);
3528 defs_t *val_defs = set_find_def(si->values, value);
3529 ir_node *arg = get_irn_n(phi_m, n);
3531 /* a spill of this value */
3535 #ifndef NO_MEMCOPIES
3536 ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3537 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3539 if(!is_zero(name->value)) {
3540 spill = insert_mem_copy(si, pred, value);
3542 spill = val_defs->spills;
3545 spill = val_defs->spills;
3547 assert(spill && "no spill placed before PhiM");
3548 set_irn_n(phi_m, n, spill);
3554 walker_reload_placer(ir_node * bb, void * data) {
3555 spill_ilp_t *si = (spill_ilp_t*)data;
3557 spill_bb_t *spill_bb = get_irn_link(bb);
3560 /* reloads at end of block */
3561 if(spill_bb->reloads) {
3564 set_foreach(spill_bb->reloads, keyval) {
3565 ir_node *irn = (ir_node*)keyval->key;
3566 ilp_var_t reload = PTR_TO_INT(keyval->val);
3569 name = si->lpp->vars[reload];
3570 if(!is_zero(name->value)) {
3572 ir_node *insert_pos = bb;
3573 ir_node *prev = sched_block_last_noncf(si, bb);
3574 op_t *prev_op = get_irn_link(prev);
3576 while(be_is_Spill(prev)) {
3577 prev = sched_prev(prev);
3580 prev_op = get_irn_link(prev);
3582 /* insert reload before pre-remats */
3583 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3584 && prev_op->is_remat && prev_op->attr.remat.pre) {
3588 prev = sched_prev(prev);
3589 } while(be_is_Spill(prev));
3591 prev_op = get_irn_link(prev);
3595 reload = insert_reload(si, irn, insert_pos);
3597 #ifdef KEEPALIVE_RELOADS
3598 pset_insert_ptr(si->spills, reload);
3604 /* walk and insert more reloads and collect remats */
3605 sched_foreach_reverse(bb, irn) {
3606 op_t *op = get_irn_link(irn);
3608 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3609 if(is_Phi(irn)) break;
3612 if(get_irn_mode(irn) != mode_T) {
3613 insert_remat(si, irn);
3618 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3619 ir_node *arg = get_irn_n(irn, n);
3621 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3624 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3625 if(!is_zero(name->value)) {
3627 ir_node *insert_pos = irn;
3628 ir_node *prev = sched_prev(insert_pos);
3631 while(be_is_Spill(prev)) {
3632 prev = sched_prev(prev);
3635 prev_op = get_irn_link(prev);
3637 /* insert reload before pre-remats */
3638 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3639 && prev_op->is_remat && prev_op->attr.remat.pre) {
3643 prev = sched_prev(prev);
3644 } while(be_is_Spill(prev));
3646 prev_op = get_irn_link(prev);
3650 reload = insert_reload(si, arg, insert_pos);
3652 set_irn_n(irn, n, reload);
3654 #ifdef KEEPALIVE_RELOADS
3655 pset_insert_ptr(si->spills, reload);
3663 del_set(spill_bb->ilp);
3664 if(spill_bb->reloads) del_set(spill_bb->reloads);
3668 walker_collect_used(ir_node * irn, void * data)
3670 lc_bitset_t *used = data;
3672 lc_bitset_set(used, get_irn_idx(irn));
3675 struct kill_helper {
3681 walker_kill_unused(ir_node * bb, void * data)
3683 struct kill_helper *kh = data;
3684 const ir_node *bad = get_irg_bad(get_irn_irg(bb));
3688 for(irn=sched_first(bb); !sched_is_end(irn);) {
3689 ir_node *next = sched_next(irn);
3692 if(!lc_bitset_is_set(kh->used, get_irn_idx(irn))) {
3693 if(be_is_Spill(irn) || be_is_Reload(irn)) {
3694 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)));
3696 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3702 set_nodes_block(irn, bad);
3703 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3704 set_irn_n(irn, n, bad);
3712 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3714 struct kill_helper kh;
3716 kh.used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3719 irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3720 irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3722 lc_bitset_free(kh.used);
3726 print_irn_pset(pset * p)
3730 pset_foreach(p, irn) {
3731 ir_printf("%+F\n", irn);
3736 dump_phi_class(spill_ilp_t * si, pset * phiclass, const char * file)
3738 FILE *f = fopen(file, "w");
3740 interference_t *interference;
3742 pset_break(phiclass);
3743 set_break(si->interferences);
3745 ir_fprintf(f, "digraph phiclass {\n");
3747 pset_foreach(phiclass, irn) {
3749 ir_fprintf(f, " %F%N [shape=box]\n",irn,irn);
3752 pset_foreach(phiclass, irn) {
3755 if(!is_Phi(irn)) continue;
3757 for(n=get_irn_arity(irn)-1; n>=0; --n) {
3758 ir_node *arg = get_irn_n(irn, n);
3760 ir_fprintf(f, " %F%N -> %F%N\n",irn,irn,arg,arg);
3764 set_foreach(si->interferences, interference) {
3765 const ir_node *a = interference->a;
3766 const ir_node *b = interference->b;
3767 if(get_phi_class(a) == phiclass) {
3768 ir_fprintf(f, " %F%N -> %F%N [color=red,dir=none,style=bold]\n",a,a,b,b);
3777 rewire_uses(spill_ilp_t * si)
3779 dom_front_info_t *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
3781 pset *ignore = pset_new_ptr(1);
3783 pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
3785 /* then fix uses of spills */
3786 set_foreach(si->values, defs) {
3789 ir_node *next = defs->remats;
3792 reloads = pset_new_ptr_default();
3795 if(be_is_Reload(next)) {
3796 pset_insert_ptr(reloads, next);
3800 next = get_irn_link(next);
3803 spills = get_spills_for_value(si, defs->value);
3804 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));
3805 if(pset_count(spills) > 1) {
3806 //assert(pset_count(reloads) > 0);
3807 // print_irn_pset(spills);
3808 // print_irn_pset(reloads);
3810 be_ssa_constr_set_ignore(dfi, si->lv, spills, ignore);
3817 /* first fix uses of remats and reloads */
3818 set_foreach(si->values, defs) {
3820 ir_node *next = defs->remats;
3823 nodes = pset_new_ptr_default();
3824 pset_insert_ptr(nodes, defs->value);
3827 pset_insert_ptr(nodes, next);
3828 next = get_irn_link(next);
3831 if(pset_count(nodes) > 1) {
3832 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
3833 be_ssa_constr_set(dfi, si->lv, nodes);
3840 // remove_unused_defs(si);
3842 be_free_dominance_frontiers(dfi);
3847 writeback_results(spill_ilp_t * si)
3849 /* walk through the graph and collect all spills, reloads and remats for a value */
3851 si->values = new_set(cmp_defs, 4096);
3853 DBG((si->dbg, LEVEL_1, "Applying results\n"));
3854 delete_unnecessary_remats(si);
3855 si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
3856 irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
3857 irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
3858 #ifdef WITH_MEMOPERANDS
3859 insert_memoperands(si);
3863 /* clean the remat info! there are still back-edges leading there! */
3864 clean_remat_info(si);
3868 connect_all_spills_with_keep(si);
3870 del_set(si->values);
3874 get_n_regs(spill_ilp_t * si)
3876 int arch_n_regs = arch_register_class_n_regs(si->cls);
3880 for(i=0; i<arch_n_regs; i++) {
3881 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
3886 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
3891 walker_reload_mover(ir_node * bb, void * data)
3893 spill_ilp_t *si = data;
3896 sched_foreach(bb, tmp) {
3897 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
3898 ir_node *reload = tmp;
3901 /* move reload upwards */
3903 int pressure = (int)get_irn_link(reload);
3904 if(pressure < si->n_regs) {
3905 irn = sched_prev(reload);
3906 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
3907 sched_remove(reload);
3908 pressure = (int)get_irn_link(irn);
3910 while(pressure < si->n_regs) {
3911 if( sched_is_end(irn) ||
3912 (be_is_Reload(irn) && has_reg_class(si, irn)) ||
3913 /* do not move reload before its spill */
3914 (irn == be_get_Reload_mem(reload)) ) break;
3916 set_irn_link(irn, INT_TO_PTR(pressure+1));
3917 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
3918 irn = sched_prev(irn);
3920 pressure = (int)get_irn_link(irn);
3923 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
3924 sched_put_after(irn, reload);
3931 move_reloads_upward(spill_ilp_t * si)
3933 irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
3938 * Walk all irg blocks and check for interfering spills inside of phi classes
3941 luke_meminterferencechecker(ir_node * bb, void * data)
3943 spill_ilp_t *si = (spill_ilp_t*)data;
3946 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
3947 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
3949 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
3951 /* a is only interesting if it is in my register class and if it is inside a phi class */
3952 if (has_reg_class(si, a) && get_phi_class(a)) {
3953 for(l2=_be_lv_next_irn(si->lv, bb, 0xff, l1+1); l2>=0; l2=_be_lv_next_irn(si->lv, bb, 0xff, l2+1)) {
3954 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
3956 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
3958 /* a and b are only interesting if they are in the same phi class */
3959 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
3960 if(values_interfere_in_block(si, bb, a, b)) {
3961 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
3970 verify_phiclasses(spill_ilp_t * si)
3972 /* analyze phi classes */
3973 phi_class_compute(si->chordal_env->irg);
3975 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
3976 irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
3980 walker_spillslotassigner(ir_node * irn, void * data)
3982 spill_ilp_t *si = (spill_ilp_t*)data;
3985 if(!be_is_Spill(irn)) return;
3987 /* set spill context to phi class if it has one ;) */
3989 cls = get_phi_class(irn);
3991 be_set_Spill_context(irn, cls);
3993 be_set_Spill_context(irn, irn);
3998 assign_spillslots(spill_ilp_t * si)
4000 DBG((si->dbg, LEVEL_2, "\t calling spill slot assigner\n"));
4001 irg_walk_graph(si->chordal_env->irg, walker_spillslotassigner, NULL, si);
4005 be_spill_remat(const be_chordal_env_t * chordal_env)
4007 char problem_name[256];
4008 char dump_suffix[256];
4009 char dump_suffix2[256];
4010 struct obstack obst;
4013 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
4014 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
4015 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
4017 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4018 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4020 #ifdef VERIFY_DOMINANCE
4021 be_check_dominance(chordal_env->irg);
4024 obstack_init(&obst);
4025 si.chordal_env = chordal_env;
4027 si.cls = chordal_env->cls;
4028 si.lpp = new_lpp(problem_name, lpp_minimize);
4029 si.remat_info = new_set(cmp_remat_info, 4096);
4030 si.interferences = new_set(cmp_interference, 32);
4031 #ifdef WITH_MEMOPERANDS
4032 si.memoperands = new_set(cmp_memoperands, 128);
4034 si.all_possible_remats = pset_new_ptr_default();
4035 si.spills = pset_new_ptr_default();
4036 si.inverse_ops = pset_new_ptr_default();
4037 si.lv = chordal_env->lv;
4041 si.n_regs = get_n_regs(&si);
4043 set_irg_link(chordal_env->irg, &si);
4044 compute_doms(chordal_env->irg);
4046 /* compute phi classes */
4047 // phi_class_compute(chordal_env->irg);
4049 be_analyze_regpressure(chordal_env, "-pre");
4051 #ifdef COLLECT_REMATS
4052 /* collect remats */
4053 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4054 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
4057 /* insert possible remats */
4058 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4059 irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
4060 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4063 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4064 connect_all_remats_with_keep(&si);
4065 /* dump graph with inserted remats */
4066 dump_graph_with_remats(chordal_env->irg, dump_suffix);
4069 /* insert copies for phi arguments not in my regclass */
4070 irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
4072 /* recompute liveness */
4073 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4074 be_liveness_recompute(si.lv);
4078 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4079 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4080 irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
4082 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4083 irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
4085 #ifndef NO_MEMCOPIES
4086 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4087 memcopyhandler(&si);
4095 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4096 if ((f = fopen(buf, "wt")) != NULL) {
4097 lpp_dump_plain(si.lpp, f);
4104 DBG((si.dbg, LEVEL_1, "\tSolving %F\n", chordal_env->irg));
4106 lpp_set_time_limit(si.lpp, ILP_TIMEOUT);
4110 lpp_solve_cplex(si.lpp);
4112 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4114 assert(lpp_is_sol_valid(si.lpp)
4115 && "solution of ILP must be valid");
4117 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));
4119 #ifdef DUMP_SOLUTION
4124 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4125 if ((f = fopen(buf, "wt")) != NULL) {
4127 for (i = 0; i < si.lpp->var_next; ++i) {
4128 lpp_name_t *name = si.lpp->vars[i];
4129 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4136 writeback_results(&si);
4140 kill_all_unused_values_in_schedule(&si);
4142 #if defined(KEEPALIVE_SPILLS) || defined(KEEPALIVE_RELOADS)
4143 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
4146 // move reloads upwards
4147 be_liveness_recompute(si.lv);
4148 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4149 move_reloads_upward(&si);
4151 #ifndef NO_MEMCOPIES
4152 verify_phiclasses(&si);
4153 assign_spillslots(&si);
4156 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4158 dump_pressure_graph(&si, dump_suffix2);
4160 be_analyze_regpressure(chordal_env, "-post");
4162 #ifdef VERIFY_DOMINANCE
4163 be_check_dominance(chordal_env->irg);
4166 free_dom(chordal_env->irg);
4167 del_set(si.interferences);
4168 del_pset(si.inverse_ops);
4169 del_pset(si.all_possible_remats);
4170 #ifdef WITH_MEMOPERANDS
4171 del_set(si.memoperands);
4173 del_pset(si.spills);
4175 obstack_free(&obst, NULL);
4176 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4179 #else /* WITH_ILP */
4182 only_that_you_can_compile_without_WITH_ILP_defined(void)
4186 #endif /* WITH_ILP */