1 /** vim: set sw=4 ts=4:
4 * @author Adam M. Szalkowski & Sebastian Hack
6 * ILP based spilling & rematerialization
8 * Copyright (C) 2006 Universitaet Karlsruhe
9 * Released under the GPL
37 #include <lpp/lpp_net.h>
38 #include <lpp/lpp_cplex.h>
39 //#include <lc_pset.h>
40 #include <libcore/lc_bitset.h>
44 #include "besched_t.h"
49 #include "bespillremat.h"
51 #include "bepressurestat.h"
53 #include "bechordal_t.h"
59 //#define KEEPALIVE /* keep alive all inserted remats and dump graph with remats */
60 #define COLLECT_REMATS /* enable rematerialization */
61 #define COLLECT_INVERSE_REMATS /* enable placement of inverse remats */
62 #define REMAT_WHILE_LIVE /* only remat values that are live */
63 //#define NO_ENLARGE_L1V3N355 /* do not remat after the death of some operand */
64 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
65 //#define MAY_DIE_AT_PRE_REMAT /* allow values to die after a pre remat */
66 #define CHECK_POST_REMAT /* check pressure after post remats (conservative but otherwise we can temporarily exceed the register pressure) */
67 #define NO_SINGLE_USE_REMATS /* do not repair schedule */
68 //#define KEEPALIVE_SPILLS
69 //#define KEEPALIVE_RELOADS
70 #define GOODWIN_REDUCTION
75 #define LPP_SERVER "i44pc52"
76 #define LPP_SOLVER "cplex"
82 #define ILP_TIMEOUT 120
86 typedef struct _spill_ilp_t {
87 const arch_register_class_t *cls;
89 const be_chordal_env_t *chordal_env;
93 pset *all_possible_remats;
98 set *values; /**< for collecting all definitions of values before running ssa-construction */
103 DEBUG_ONLY(firm_dbg_module_t * dbg);
106 typedef int ilp_var_t;
107 typedef int ilp_cst_t;
109 typedef struct _spill_bb_t {
114 typedef struct _remat_t {
115 const ir_node *op; /**< for copy_irn */
116 const ir_node *value; /**< the value which is being recomputed by this remat */
117 ir_node *proj; /**< not NULL if the above op produces a tuple */
118 int cost; /**< cost of this remat */
119 int inverse; /**< nonzero if this is an inverse remat */
123 * Data to be attached to each IR node. For remats this contains the ilp_var
124 * for this remat and for normal ops this contains the ilp_vars for
125 * reloading each operand
127 typedef struct _op_t {
132 remat_t *remat; /** the remat this op belongs to */
133 int pre; /** 1, if this is a pressure-increasing remat */
137 ir_node *op; /** the operation this live range belongs to */
146 typedef struct _defs_t {
148 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
149 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
152 typedef struct _remat_info_t {
153 const ir_node *irn; /**< the irn to which these remats belong */
154 pset *remats; /**< possible remats for this value */
155 pset *remats_by_operand; /**< remats with this value as operand */
158 typedef struct _keyval_t {
163 typedef struct _spill_t {
173 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
175 return chordal_has_class(si->chordal_env, irn);
180 cmp_remat(const void *a, const void *b)
182 const keyval_t *p = a;
183 const keyval_t *q = b;
184 const remat_t *r = p->val;
185 const remat_t *s = q->val;
189 return !(r == s || r->op == s->op);
193 cmp_remat(const void *a, const void *b)
195 const remat_t *r = a;
196 const remat_t *s = a;
198 return !(r == s || r->op == s->op);
202 cmp_spill(const void *a, const void *b, size_t size)
204 const spill_t *p = a;
205 const spill_t *q = b;
207 // return !(p->irn == q->irn && p->bb == q->bb);
208 return !(p->irn == q->irn);
212 set_find_keyval(set * set, void * key)
217 return set_find(set, &query, sizeof(query), HASH_PTR(key));
221 set_insert_keyval(set * set, void * key, void * val)
227 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
231 set_find_def(set * set, ir_node * value)
236 return set_find(set, &query, sizeof(query), HASH_PTR(value));
240 set_insert_def(set * set, ir_node * value)
247 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
251 set_find_spill(set * set, ir_node * value)
256 return set_find(set, &query, sizeof(query), HASH_PTR(value));
259 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
260 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
261 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
262 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
263 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
266 cmp_remat_info(const void *a, const void *b, size_t size)
268 const remat_info_t *p = a;
269 const remat_info_t *q = b;
271 return !(p->irn == q->irn);
275 cmp_defs(const void *a, const void *b, size_t size)
280 return !(p->value == q->value);
284 cmp_keyval(const void *a, const void *b, size_t size)
286 const keyval_t *p = a;
287 const keyval_t *q = b;
289 return !(p->key == q->key);
293 execution_frequency(const spill_ilp_t * si, const ir_node * irn)
298 return get_block_execfreq(si->execfreqs, irn) + FUDGE;
300 return get_block_execfreq(si->execfreqs, get_nodes_block(irn)) + FUDGE;
304 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
306 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
311 get_cost(const spill_ilp_t * si, const ir_node * irn)
313 if(be_is_Spill(irn)) {
315 } else if(be_is_Reload(irn)){
318 return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
324 * Checks, whether node and its operands have suitable reg classes
327 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
331 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
332 int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
336 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
339 for (i = 0, n = get_irn_arity(irn); i < n && remat; ++i) {
340 ir_node *op = get_irn_n(irn, i);
341 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
344 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
351 * Try to create a remat from @p op with destination value @p dest_value
353 static INLINE remat_t *
354 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
356 remat_t *remat = NULL;
358 // if(!mode_is_datab(get_irn_mode(dest_value)))
361 if(dest_value == op) {
362 const ir_node *proj = NULL;
364 if(is_Proj(dest_value)) {
365 op = get_irn_n(op, 0);
369 if(!is_rematerializable(si, op))
372 remat = obstack_alloc(si->obst, sizeof(*remat));
374 remat->cost = get_cost(si, op);
375 remat->value = dest_value;
379 arch_inverse_t inverse;
383 /* get the index of the operand we want to retrieve by the inverse op */
384 for (i = 0, n = get_irn_arity(op); i < n; ++i) {
385 ir_node *arg = get_irn_n(op, i);
387 if(arg == dest_value) break;
389 if(i == n) return NULL;
391 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", i, op));
393 /* else ask the backend to give an inverse op */
394 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, i, &inverse, si->obst)) {
397 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
399 assert(inverse.n > 0 && "inverse op should have at least one node");
401 for(i=0; i<inverse.n; ++i) {
402 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
406 remat = obstack_alloc(si->obst, sizeof(*remat));
407 remat->op = inverse.nodes[0];
408 remat->cost = inverse.costs;
409 remat->value = dest_value;
410 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
413 assert(is_Proj(remat->proj));
415 assert(0 && "I can not handle remats with more than 2 nodes");
422 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
424 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
432 add_remat(const spill_ilp_t * si, const remat_t * remat)
434 remat_info_t *remat_info,
440 assert(remat->value);
442 query.irn = remat->value;
444 query.remats_by_operand = NULL;
445 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
447 if(remat_info->remats == NULL) {
448 remat_info->remats = new_pset(cmp_remat, 4096);
450 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
452 /* insert the remat into the remats_be_operand set of each argument of the remat op */
453 for (i = 0, n = get_irn_arity(remat->op); i < n; ++i) {
454 ir_node *arg = get_irn_n(remat->op, i);
458 query.remats_by_operand = NULL;
459 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
461 if(remat_info->remats_by_operand == NULL) {
462 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
464 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
469 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
471 const ir_edge_t *edge = get_irn_out_edge_first(irn);
475 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
478 edge = get_irn_out_edge_next(irn, edge);
485 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
491 #ifdef NO_SINGLE_USE_REMATS
492 if(has_reg_class(si, op) && (get_irn_n_nonremat_edges(si, op) > 1)) {
494 if(has_reg_class(si, op)) {
496 remat = get_remat_from_op(si, op, op);
498 add_remat(si, remat);
502 #ifdef COLLECT_INVERSE_REMATS
503 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
505 for (i = 0, n = get_irn_arity(op); i < n; ++i) {
506 ir_node *arg = get_irn_n(op, i);
508 if(has_reg_class(si, arg)) {
509 /* try to get an inverse remat */
510 remat = get_remat_from_op(si, arg, op);
512 add_remat(si, remat);
521 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
524 ir_node *def_block = get_nodes_block(val);
530 /* if pos is at end of a basic block */
532 ret = (pos == def_block || block_dominates(def_block, pos));
533 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
537 /* else if this is a normal operation */
538 block = get_nodes_block(pos);
539 if(block == def_block) {
540 if(!sched_is_scheduled(val)) return 1;
542 ret = sched_comes_after(val, pos);
543 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
547 ret = block_dominates(def_block, block);
548 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
552 static INLINE ir_node *
553 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
555 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
559 * Returns first non-Phi node of block @p bb
561 static INLINE ir_node *
562 sched_block_first_nonphi(const ir_node * bb)
564 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
568 sched_skip_proj_predicator(const ir_node * irn, void * data)
570 return (is_Proj(irn));
573 static INLINE ir_node *
574 sched_next_nonproj(const ir_node * irn, int forward)
576 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
580 * Returns next operation node (non-Proj) after @p irn
581 * or the basic block of this node
583 static INLINE ir_node *
584 sched_next_op(const ir_node * irn)
586 ir_node *next = sched_next(irn);
591 return sched_next_nonproj(next, 1);
595 * Returns previous operation node (non-Proj) before @p irn
596 * or the basic block of this node
598 static INLINE ir_node *
599 sched_prev_op(const ir_node * irn)
601 ir_node *prev = sched_prev(irn);
606 return sched_next_nonproj(prev, 0);
610 sched_put_after(ir_node * insert, ir_node * irn)
612 if(is_Block(insert)) {
613 insert = sched_block_first_nonphi(insert);
615 insert = sched_next_op(insert);
617 sched_add_before(insert, irn);
621 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
623 if(is_Block(insert)) {
624 insert = sched_block_last_noncf(si, insert);
626 insert = sched_next_nonproj(insert, 0);
627 insert = sched_prev(insert);
629 sched_add_after(insert, irn);
633 * Tells you whether a @p remat can be placed before the irn @p pos
636 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
638 const ir_node *op = remat->op;
645 prev = sched_block_last_noncf(si, pos);
646 prev = sched_next_nonproj(prev, 0);
648 prev = sched_prev_op(pos);
650 /* do not remat if the rematted value is defined immediately before this op */
651 if(prev == remat->op) {
656 /* this should be just fine, the following OP will be using this value, right? */
658 /* only remat AFTER the real definition of a value (?) */
659 if(!value_is_defined_before(si, pos, remat->value)) {
660 // ir_fprintf(stderr, "error(not defined)");
665 for(i=0, n=get_irn_arity(op); i<n && res; ++i) {
666 const ir_node *arg = get_irn_n(op, i);
668 #ifdef NO_ENLARGE_L1V3N355
669 if(has_reg_class(si, arg) && live) {
670 res &= pset_find_ptr(live, arg)?1:0;
672 res &= value_is_defined_before(si, pos, arg);
675 res &= value_is_defined_before(si, pos, arg);
683 * Tells you whether a @p remat can be placed after the irn @p pos
686 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
689 pos = sched_block_first_nonphi(pos);
691 pos = sched_next_op(pos);
694 /* only remat AFTER the real definition of a value (?) */
695 if(!value_is_defined_before(si, pos, remat->value)) {
699 return can_remat_before(si, remat, pos, live);
703 * Collect potetially rematerializable OPs
706 walker_remat_collector(ir_node * irn, void * data)
708 spill_ilp_t *si = data;
710 if(!is_Block(irn) && !is_Phi(irn)) {
711 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
712 get_remats_from_op(si, irn);
717 * Inserts a copy of @p irn before @p pos
720 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
725 bb = is_Block(pos)?pos:get_nodes_block(pos);
726 copy = exact_copy(irn);
727 set_nodes_block(copy, bb);
728 sched_put_before(si, pos, copy);
734 * Inserts a copy of @p irn after @p pos
737 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
742 bb = is_Block(pos)?pos:get_nodes_block(pos);
743 copy = exact_copy(irn);
744 set_nodes_block(copy, bb);
745 sched_put_after(pos, copy);
751 insert_remat_after(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
755 if(can_remat_after(si, remat, pos, live)) {
760 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
762 copy = insert_copy_after(si, remat->op, pos);
764 // ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", remat->value, pos);
765 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
766 op = obstack_alloc(si->obst, sizeof(*op));
768 op->attr.remat.remat = remat;
769 op->attr.remat.pre = 0;
770 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
772 set_irn_link(copy, op);
773 pset_insert_ptr(si->all_possible_remats, copy);
775 proj_copy = insert_copy_after(si, remat->proj, copy);
776 set_irn_n(proj_copy, 0, copy);
777 set_irn_link(proj_copy, op);
778 pset_insert_ptr(si->all_possible_remats, proj_copy);
786 insert_remat_before(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
790 if(can_remat_before(si, remat, pos, live)) {
795 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
797 copy = insert_copy_before(si, remat->op, pos);
799 // ir_snprintf(buf, sizeof(buf), "remat_%N_%N", remat->value, pos);
800 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
801 op = obstack_alloc(si->obst, sizeof(*op));
803 op->attr.remat.remat = remat;
804 op->attr.remat.pre = 1;
805 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
807 set_irn_link(copy, op);
808 pset_insert_ptr(si->all_possible_remats, copy);
810 proj_copy = insert_copy_after(si, remat->proj, copy);
811 set_irn_n(proj_copy, 0, copy);
812 set_irn_link(proj_copy, op);
813 pset_insert_ptr(si->all_possible_remats, proj_copy);
821 get_block_n_succs(const ir_node *block) {
822 const ir_edge_t *edge;
824 assert(edges_activated(current_ir_graph));
826 edge = get_block_succ_first(block);
830 edge = get_block_succ_next(block, edge);
835 is_merge_edge(const ir_node * bb)
837 #ifdef GOODWIN_REDUCTION
838 return get_block_n_succs(bb) == 1;
845 is_diverge_edge(const ir_node * bb)
847 #ifdef GOODWIN_REDUCTION
848 return get_Block_n_cfgpreds(bb) == 1;
855 walker_regclass_copy_insertor(ir_node * irn, void * data)
857 spill_ilp_t *si = data;
859 if(is_Phi(irn) && has_reg_class(si, irn)) {
862 for(n=get_irn_arity(irn)-1; n>=0; --n) {
863 ir_node *phi_arg = get_irn_n(irn, n);
864 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
866 if(!has_reg_class(si, phi_arg)) {
867 ir_node *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
868 ir_node *pos = sched_block_last_noncf(si, bb);
869 op_t *op = obstack_alloc(si->obst, sizeof(*op));
871 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
872 sched_add_after(pos, copy);
873 set_irn_n(irn, n, copy);
876 op->attr.live_range.args.reloads = NULL;
877 op->attr.live_range.ilp = ILP_UNDEF;
878 set_irn_link(copy, op);
886 * Insert (so far unused) remats into the irg to
887 * recompute the potential liveness of all values
890 walker_remat_insertor(ir_node * bb, void * data)
892 spill_ilp_t *si = data;
893 spill_bb_t *spill_bb;
897 pset *live = pset_new_ptr_default();
899 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
901 live_foreach(bb, li) {
902 ir_node *value = (ir_node *) li->irn;
904 /* add remats at end of block */
905 if (live_is_end(li) && has_reg_class(si, value)) {
906 pset_insert_ptr(live, value);
910 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
911 set_irn_link(bb, spill_bb);
913 irn = sched_last(bb);
914 while(!sched_is_end(irn)) {
920 next = sched_prev(irn);
922 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
924 if(is_Phi(irn) || is_Proj(irn)) {
927 if(has_reg_class(si, irn)) {
928 pset_remove_ptr(live, irn);
931 op = obstack_alloc(si->obst, sizeof(*op));
933 op->attr.live_range.args.reloads = NULL;
934 op->attr.live_range.ilp = ILP_UNDEF;
935 set_irn_link(irn, op);
941 op = obstack_alloc(si->obst, sizeof(*op));
943 op->attr.live_range.ilp = ILP_UNDEF;
944 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
945 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
946 set_irn_link(irn, op);
948 args = pset_new_ptr_default();
950 /* collect arguments of op */
951 for (n = get_irn_arity(irn)-1; n>=0; --n) {
952 ir_node *arg = get_irn_n(irn, n);
954 pset_insert_ptr(args, arg);
957 /* set args of op already live in epilog */
958 pset_foreach(args, arg) {
959 if(has_reg_class(si, arg)) {
960 pset_insert_ptr(live, arg);
964 /* insert all possible remats after irn */
965 pset_foreach(args, arg) {
966 remat_info_t *remat_info,
970 /* continue if the operand has the wrong reg class */
971 if(!has_reg_class(si, arg))
976 query.remats_by_operand = NULL;
977 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
983 /* do not place post remats after jumps */
984 if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
986 if(remat_info->remats_by_operand) {
987 pset_foreach(remat_info->remats_by_operand, remat) {
988 /* do not insert remats producing the same value as one of the operands */
989 if(!pset_find_ptr(args, remat->value)) {
990 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
991 #ifdef REMAT_WHILE_LIVE
992 if(pset_find_ptr(live, remat->value)) {
993 insert_remat_after(si, remat, irn, live);
996 insert_remat_after(si, remat, irn, live);
1003 /* delete defined value from live set */
1004 if(has_reg_class(si, irn)) {
1005 pset_remove_ptr(live, irn);
1008 /* insert all possible remats before irn */
1009 pset_foreach(args, arg) {
1010 remat_info_t *remat_info,
1014 /* continue if the operand has the wrong reg class
1016 if(!has_reg_class(si, arg))
1020 query.remats = NULL;
1021 query.remats_by_operand = NULL;
1022 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1028 if(remat_info->remats) {
1029 pset_foreach(remat_info->remats, remat) {
1030 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
1031 #ifdef REMAT_WHILE_LIVE
1032 if(pset_find_ptr(live, remat->value)) {
1033 insert_remat_before(si, remat, irn, live);
1036 insert_remat_before(si, remat, irn, live);
1046 live_foreach(bb, li) {
1047 ir_node *value = (ir_node *) li->irn;
1049 /* add remats at end if successor has multiple predecessors */
1050 if(is_merge_edge(bb)) {
1051 /* add remats at end of block */
1052 if (live_is_end(li) && has_reg_class(si, value)) {
1053 remat_info_t *remat_info,
1058 query.remats = NULL;
1059 query.remats_by_operand = NULL;
1060 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1062 if(remat_info && remat_info->remats) {
1063 pset_foreach(remat_info->remats, remat) {
1064 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1066 insert_remat_before(si, remat, bb, NULL);
1071 if(is_diverge_edge(bb)) {
1072 /* add remat2s at beginning of block */
1073 if ((live_is_in(li) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) {
1074 remat_info_t *remat_info,
1079 query.remats = NULL;
1080 query.remats_by_operand = NULL;
1081 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1083 if(remat_info && remat_info->remats) {
1084 pset_foreach(remat_info->remats, remat) {
1085 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at beginning of block %+F\n", remat->op, bb));
1087 /* put the remat here if all its args are available */
1088 insert_remat_after(si, remat, bb, NULL);
1098 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1101 luke_endwalker(ir_node * bb, void * data)
1103 spill_ilp_t *si = (spill_ilp_t*)data;
1110 spill_bb_t *spill_bb = get_irn_link(bb);
1113 live = pset_new_ptr_default();
1114 use_end = pset_new_ptr_default();
1116 live_foreach(bb, li) {
1117 irn = (ir_node *) li->irn;
1118 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1121 pset_insert_ptr(live, irn);
1122 op = get_irn_link(irn);
1123 assert(!op->is_remat);
1127 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1128 /* their reg_out must always be set */
1129 sched_foreach_reverse(bb, irn) {
1132 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1134 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1135 ir_node *irn_arg = get_irn_n(irn, n);
1137 if(has_reg_class(si, irn_arg)) {
1138 pset_insert_ptr(use_end, irn_arg);
1143 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1144 //cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1145 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1147 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1149 /* if this is a merge edge we can reload at the end of this block */
1150 if(is_merge_edge(bb)) {
1151 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1152 } else if(pset_count(use_end)){
1153 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1155 spill_bb->reloads = NULL;
1158 pset_foreach(live,irn) {
1164 /* handle values used by control flow nodes later separately */
1165 if(pset_find_ptr(use_end, irn)) continue;
1168 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1170 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1172 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1173 spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1174 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1176 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1177 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1179 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1180 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1182 if(is_merge_edge(bb)) {
1186 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1187 reload = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1188 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1190 /* reload <= mem_out */
1191 rel_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1192 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1193 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1196 spill->reg_in = ILP_UNDEF;
1197 spill->mem_in = ILP_UNDEF;
1200 pset_foreach(use_end,irn) {
1204 ilp_cst_t end_use_req,
1209 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1211 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1213 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1214 spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1215 /* if irn is used at the end of the block, then it is live anyway */
1216 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1218 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1219 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1221 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1222 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1224 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1225 reload = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1226 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1228 /* reload <= mem_out */
1229 rel_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1230 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1231 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1233 spill->reg_in = ILP_UNDEF;
1234 spill->mem_in = ILP_UNDEF;
1236 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1237 end_use_req = lpp_add_cst(si->lpp, buf, lpp_equal, 1);
1238 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1246 next_post_remat(const ir_node * irn)
1251 irn = sched_block_first_nonphi(irn);
1253 irn = sched_next_op(irn);
1256 if(sched_is_end(irn))
1259 op = (op_t*)get_irn_link(irn);
1260 if(op->is_remat && !op->attr.remat.pre) {
1269 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1275 ret = sched_block_last_noncf(si, irn);
1276 ret = sched_next(ret);
1277 ret = sched_prev_op(ret);
1279 ret = sched_prev_op(irn);
1282 if(sched_is_end(ret) || is_Phi(ret))
1285 op = (op_t*)get_irn_link(ret);
1286 if(op->is_remat && op->attr.remat.pre) {
1294 * Find a remat of value @p value in the epilog of @p pos
1297 find_post_remat(const ir_node * value, const ir_node * pos)
1299 while((pos = next_post_remat(pos)) != NULL) {
1302 op = get_irn_link(pos);
1303 assert(op->is_remat && !op->attr.remat.pre);
1305 if(op->attr.remat.remat->value == value)
1306 return (ir_node*)pos;
1309 const ir_edge_t *edge;
1310 foreach_out_edge(pos, edge) {
1311 ir_node *proj = get_edge_src_irn(edge);
1312 assert(is_Proj(proj));
1322 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1324 spill_bb_t *spill_bb = get_irn_link(bb);
1330 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1332 double spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1334 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1336 spill->reg_out = ILP_UNDEF;
1337 spill->reg_in = ILP_UNDEF;
1338 spill->mem_in = ILP_UNDEF;
1340 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1341 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1343 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1344 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1351 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1356 live_foreach(bb, li) {
1357 irn = (ir_node *) li->irn;
1359 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1360 pset_insert_ptr(live, irn);
1364 irn = sched_last(bb);
1366 /* all values eaten by control flow operations are also live until the end of the block */
1367 sched_foreach_reverse(bb, irn) {
1370 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1372 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1373 ir_node *arg = get_irn_n(irn,i);
1375 if(has_reg_class(si, arg)) {
1376 pset_insert_ptr(live, arg);
1383 * Walk all irg blocks and emit this ILP
1386 luke_blockwalker(ir_node * bb, void * data)
1388 spill_ilp_t *si = (spill_ilp_t*)data;
1393 spill_bb_t *spill_bb = get_irn_link(bb);
1398 pset *defs = pset_new_ptr_default();
1401 live = pset_new_ptr_default();
1403 /****************************************
1404 * B A S I C B L O C K E N D
1405 ***************************************/
1408 /* init live values at end of block */
1409 get_live_end(si, bb, live);
1411 pset_foreach(live, irn) {
1413 ilp_var_t reload = ILP_UNDEF;
1415 spill = set_find_spill(spill_bb->ilp, irn);
1418 if(spill_bb->reloads) {
1419 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1422 reload = PTR_TO_INT(keyval->val);
1426 op = get_irn_link(irn);
1427 assert(!op->is_remat);
1429 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1430 op->attr.live_range.ilp = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1431 op->attr.live_range.op = bb;
1433 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1434 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1436 /* reg_out - reload - remat - live_range <= 0 */
1437 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1438 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1439 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1440 foreach_pre_remat(si, bb, tmp) {
1441 op_t *remat_op = get_irn_link(tmp);
1442 if(remat_op->attr.remat.remat->value == irn) {
1443 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1446 /* maybe we should also assure that reg_out >= live_range etc. */
1450 * start new live ranges for values used by remats at end of block
1451 * and assure the remat args are available
1453 foreach_pre_remat(si, bb, tmp) {
1454 op_t *remat_op = get_irn_link(tmp);
1457 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1458 ir_node *remat_arg = get_irn_n(tmp, n);
1459 op_t *arg_op = get_irn_link(remat_arg);
1462 if(!has_reg_class(si, remat_arg)) continue;
1464 /* if value is becoming live through use by remat */
1465 if(!pset_find_ptr(live, remat_arg)) {
1466 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1467 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1469 arg_op->attr.live_range.ilp = prev_lr;
1470 arg_op->attr.live_range.op = bb;
1472 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1474 pset_insert_ptr(live, remat_arg);
1475 add_to_spill_bb(si, bb, remat_arg);
1478 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1479 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1480 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1482 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1483 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1485 /* use reload placed for this argument */
1486 if(spill_bb->reloads) {
1487 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1490 ilp_var_t reload = PTR_TO_INT(keyval->val);
1492 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1497 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1502 /**************************************
1503 * B A S I C B L O C K B O D Y
1504 **************************************/
1506 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1512 ilp_cst_t check_pre,
1514 #ifdef CHECK_POST_REMAT
1515 ilp_cst_t check_post_remat;
1520 /* iterate only until first phi */
1524 op = get_irn_link(irn);
1526 if(op->is_remat) continue;
1527 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1529 /* collect defined values */
1530 if(has_reg_class(si, irn)) {
1531 pset_insert_ptr(defs, irn);
1535 if(is_Proj(irn)) continue;
1538 /* remove defined values from live set */
1539 if(has_reg_class(si, irn)) {
1540 assert(pset_find_ptr(live, irn));
1541 pset_remove_ptr(live, irn);
1545 /* init set of irn's arguments */
1546 args = new_set(cmp_keyval, get_irn_arity(irn));
1547 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1548 ir_node *irn_arg = get_irn_n(irn, n);
1549 if(has_reg_class(si, irn_arg)) {
1550 set_insert_keyval(args, irn_arg, (void*)n);
1554 /**********************************
1555 * I N E P I L O G O F irn
1556 **********************************/
1558 /* new live ranges for values from L\U defined by post remats */
1559 pset_foreach(live, tmp) {
1560 ir_node *value = tmp;//remat_op->attr.remat.remat->value;
1561 op_t *value_op = get_irn_link(value);
1563 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
1564 ilp_var_t prev_lr = ILP_UNDEF;
1568 foreach_post_remat(irn, remat) {
1569 op_t *remat_op = get_irn_link(remat);
1571 /* if value is being rematerialized by this remat */
1572 if(value == remat_op->attr.remat.remat->value) {
1573 if(cst == ILP_UNDEF) {
1574 /* next_live_range <= prev_live_range + sum remat2s */
1575 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1576 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1578 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1579 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1581 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1582 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1585 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1589 if(prev_lr != ILP_UNDEF) {
1590 value_op->attr.live_range.ilp = prev_lr;
1591 value_op->attr.live_range.op = irn;
1596 /* requirements for post remats and start live ranges from L/U for values dying here */
1597 foreach_post_remat(irn, tmp) {
1598 op_t *remat_op = get_irn_link(tmp);
1601 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1602 ir_node *remat_arg = get_irn_n(tmp, n);
1603 op_t *arg_op = get_irn_link(remat_arg);
1605 if(!has_reg_class(si, remat_arg)) continue;
1607 /* only for values in L\U (TODO and D?), the others are handled with post_use */
1608 if(!set_find_keyval(args, remat_arg)) {
1609 /* remat <= live_rang(remat_arg) */
1610 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
1611 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1613 /* if value is becoming live through use by remat2 */
1614 if(!pset_find_ptr(live, remat_arg)) {
1617 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1618 lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1620 arg_op->attr.live_range.ilp = lr;
1621 arg_op->attr.live_range.op = irn;
1623 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
1625 pset_insert_ptr(live, remat_arg);
1626 add_to_spill_bb(si, bb, remat_arg);
1629 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1630 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1635 #ifdef CHECK_POST_REMAT
1636 /* check the register pressure after the epilog */
1637 ir_snprintf(buf, sizeof(buf), "check_post_remat_%N", irn);
1638 check_post_remat = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1640 /* iterate over L\U */
1641 pset_foreach(live, tmp) {
1642 if(!set_find_keyval(args, tmp)) {
1643 /* if a live value is not used by irn */
1644 tmp_op = get_irn_link(tmp);
1645 lpp_set_factor_fast(si->lpp, check_post_remat, tmp_op->attr.live_range.ilp, 1.0);
1648 /* iterate over remats in epilog and remove possibly defined values again from check_post_remat */
1649 foreach_post_remat(irn, tmp) {
1650 op_t *remat_op = get_irn_link(tmp);
1651 const ir_node *value = remat_op->attr.remat.remat->value;
1652 op_t *val_op = get_irn_link(value);
1654 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
1656 /* values that are defined by remats are not counted */
1657 #ifdef REMAT_WHILE_LIVE
1658 assert(val_op->attr.live_range.ilp);
1659 lpp_set_factor_fast(si->lpp, check_post_remat, val_op->attr.live_range.ilp, 0.0);
1661 if(val_op->attr.live_range.ilp != ILP_UNDEF) {
1662 lpp_set_factor_fast(si->lpp, check_post_remat, val_op->attr.live_range.ilp, 0.0);
1664 #endif /* REMAT_WHILE_LIVE */
1666 /* instead add these post remats to the check */
1667 lpp_set_factor_fast(si->lpp, check_post_remat, remat_op->attr.remat.ilp, 1.0);
1670 #if 0 /* defs are still in live */
1671 /* get count of values in my register class defined by irn */
1672 /* also add defined values to check_post_remat; do this before iterating over args */
1673 pset_foreach(defs,tmp) {
1674 op_t *def_op = get_irn_link(tmp);
1676 lpp_set_factor_fast(si->lpp, check_post_remat, def_op->attr.live_range.ilp, 1.0);
1679 #endif /* CHECK_POST_REMAT */
1682 d = pset_count(defs);
1683 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
1685 /* count how many regs irn needs for arguments */
1686 u = set_count(args);
1689 /* check the register pressure in the epilog */
1690 /* sum_{L\U} lr + sum_{U} post_use <= k - |D| */
1691 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
1692 check_post = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - d);
1694 /* add L\U to check_post */
1695 pset_foreach(live, tmp) {
1696 if(!set_find_keyval(args, tmp) && !pset_find_ptr(defs, tmp)) {
1697 /* if a live value is not used by irn */
1698 tmp_op = get_irn_link(tmp);
1699 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
1703 /***********************************************************
1704 * I T E R A T I O N O V E R A R G S F O R E P I L O G
1705 **********************************************************/
1708 set_foreach(args, keyval) {
1713 ir_node *arg = keyval->key;
1714 op_t *arg_op = get_irn_link(arg);
1716 spill = add_to_spill_bb(si, bb, arg);
1718 /* new live range for each argument */
1719 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
1720 next_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1722 i = (int)keyval->val;
1724 /* the epilog stuff - including post_use, check_post, check_post_remat */
1725 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
1726 post_use = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1728 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
1730 /* arg is live throughout epilog if the next live_range is in a register */
1731 if(pset_find_ptr(live, arg)) {
1732 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
1734 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1735 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1736 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
1737 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1739 #ifdef CHECK_POST_REMAT
1740 //lpp_set_factor_fast(si->lpp, check_post_remat, arg_op->attr.live_range.ilp, 1.0);
1741 lpp_set_factor_fast(si->lpp, check_post_remat, post_use, 1.0);
1745 /*forall remat2 which use arg add a similar cst*/
1746 foreach_post_remat(irn, tmp) {
1749 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1750 ir_node *remat_arg = get_irn_n(tmp, n);
1751 op_t *remat_op = get_irn_link(tmp);
1753 if(remat_arg == arg) {
1754 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
1756 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1757 cst = lpp_add_cst(si->lpp, buf, lpp_greater, 0.0);
1758 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1759 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1764 /* new live range begins for each argument */
1765 arg_op->attr.live_range.ilp = next_lr;
1766 arg_op->attr.live_range.op = irn;
1768 pset_insert_ptr(live, arg);
1771 /* just to be sure */
1772 check_post = ILP_UNDEF;
1773 #ifdef CHECK_POST_REMAT
1774 check_post_remat = ILP_UNDEF;
1784 /* check the register pressure in the prolog */
1785 /* sum_{L\U} lr <= k - |U| */
1786 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
1787 check_pre = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - u);
1789 /* for the prolog remove defined values from the live set */
1790 pset_foreach(defs, tmp) {
1791 pset_remove_ptr(live, tmp);
1794 /***********************************************************
1795 * 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
1796 **********************************************************/
1799 set_foreach(args, keyval) {
1801 ir_node *arg = keyval->key;
1802 int i = PTR_TO_INT(keyval->val);
1803 op_t *arg_op = get_irn_link(arg);
1805 spill = set_find_spill(spill_bb->ilp, arg);
1808 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
1809 op->attr.live_range.args.reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1811 /* reload <= mem_out */
1812 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
1813 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1814 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
1815 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
1817 /* requirement: arg must be in register for use */
1818 /* reload + remat + live_range == 1 */
1819 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
1820 cst = lpp_add_cst(si->lpp, buf, lpp_equal, 1.0);
1822 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1823 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
1824 foreach_pre_remat(si, irn, tmp) {
1825 op_t *remat_op = get_irn_link(tmp);
1826 if(remat_op->attr.remat.remat->value == arg) {
1827 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1832 /* iterate over L\U */
1833 pset_foreach(live, tmp) {
1834 if(!set_find_keyval(args, tmp)) {
1835 /* if a live value is not used by irn */
1836 tmp_op = get_irn_link(tmp);
1837 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
1842 /* TODO allow new live ranges even if value does not die after remat??? MAY_DIE_AT_PRE_REMAT */
1844 /* requirements for remats */
1845 /* start new live ranges for values used by remats */
1846 foreach_pre_remat(si, irn, tmp) {
1847 op_t *remat_op = get_irn_link(tmp);
1850 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1851 ir_node *remat_arg = get_irn_n(tmp, n);
1852 op_t *arg_op = get_irn_link(remat_arg);
1855 if(!has_reg_class(si, remat_arg)) continue;
1857 /* if value is becoming live through use by remat */
1858 if(!pset_find_ptr(live, remat_arg)) {
1859 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1860 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1862 arg_op->attr.live_range.ilp = prev_lr;
1863 arg_op->attr.live_range.op = irn;
1865 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat %+F\n", remat_arg, tmp));
1867 pset_insert_ptr(live, remat_arg);
1868 add_to_spill_bb(si, bb, remat_arg);
1871 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1872 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1873 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1875 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1876 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1878 /* if remat arg is also used by current op then we can use reload placed for this argument */
1879 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
1880 int index = (int)keyval->val;
1882 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
1890 /*************************
1891 * D O N E W I T H O P
1892 *************************/
1894 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
1896 pset_foreach(live, tmp) {
1897 assert(has_reg_class(si, tmp));
1900 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1901 ir_node *arg = get_irn_n(irn, n);
1903 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
1908 defs = pset_new_ptr_default();
1913 /***************************************
1914 * B E G I N N I N G O F B L O C K
1915 ***************************************/
1918 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
1919 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
1921 pset_foreach(live, irn) {
1922 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
1925 /* construct mem_outs for all values */
1927 set_foreach(spill_bb->ilp, spill) {
1928 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
1929 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1931 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
1932 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
1934 if(pset_find_ptr(live, spill->irn)) {
1935 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
1937 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
1938 spill->mem_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1939 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
1941 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
1943 op_t *op = get_irn_link(spill->irn);
1945 /* do we have to copy a phi argument? */
1946 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
1947 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
1949 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
1950 const ir_node *arg = get_irn_n(spill->irn, n);
1956 /* argument already done? */
1957 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
1959 /* get sum of execution frequencies of blocks with the same phi argument */
1960 for(m=n; m>=0; --m) {
1961 const ir_node *arg2 = get_irn_n(spill->irn, m);
1964 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
1968 /* copies are not for free */
1969 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
1970 var = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE * freq);
1972 for(m=n; m>=0; --m) {
1973 const ir_node *arg2 = get_irn_n(spill->irn, m);
1976 op->attr.live_range.args.copies[m] = var;
1980 /* copy <= mem_in */
1981 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
1982 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1983 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
1984 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
1991 /* L\U is empty at bb start */
1992 /* arg is live throughout epilog if it is reg_in into this block */
1994 /* check the register pressure at the beginning of the block
1997 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
1998 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
2000 pset_foreach(live, irn) {
2003 spill = set_find_spill(spill_bb->ilp, irn);
2006 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2007 spill->reg_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2009 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2011 /* spill + mem_in <= 1 */
2012 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2013 nospill = lpp_add_cst(si->lpp, buf, lpp_less, 1);
2015 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2016 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2019 foreach_post_remat(bb, irn) {
2020 op_t *remat_op = get_irn_link(irn);
2022 DBG((si->dbg, LEVEL_4, "\t next post remat: %+F\n", irn));
2023 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
2025 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2028 /* forall post remats add requirements */
2029 foreach_post_remat(bb, tmp) {
2032 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2033 ir_node *remat_arg = get_irn_n(tmp, n);
2034 op_t *remat_op = get_irn_link(tmp);
2036 if(!has_reg_class(si, remat_arg)) continue;
2038 spill = set_find_spill(spill_bb->ilp, remat_arg);
2041 /* remat <= reg_in_argument */
2042 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2043 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2044 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2045 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2049 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2050 pset_foreach(live, irn) {
2054 spill = set_find_spill(spill_bb->ilp, irn);
2055 assert(spill && spill->irn == irn);
2057 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2058 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2061 ir_node *phi_arg = get_Phi_pred(irn, n);
2062 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2063 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2066 /* although the phi is in the right regclass one or more of
2067 * its arguments can be in a different one or at least to
2070 if(has_reg_class(si, phi_arg)) {
2071 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2072 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2073 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2074 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2076 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2077 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2079 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2082 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2083 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2087 /* else assure the value arrives on all paths in the same resource */
2089 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2092 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2093 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2096 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2097 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2098 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2099 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2101 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2102 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2104 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2107 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2108 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2113 /* first live ranges from reg_ins */
2114 pset_foreach(live, irn) {
2115 op_t *op = get_irn_link(irn);
2117 spill = set_find_spill(spill_bb->ilp, irn);
2118 assert(spill && spill->irn == irn);
2120 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2121 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2122 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2123 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2125 foreach_post_remat(bb, tmp) {
2126 op_t *remat_op = get_irn_link(tmp);
2128 if(remat_op->attr.remat.remat->value == irn) {
2129 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2134 /* walk forward now and compute constraints for placing spills */
2135 /* this must only be done for values that are not defined in this block */
2136 /* TODO are these values at start of block? if yes, just check whether this is a diverge edge and skip the loop */
2137 pset_foreach(live, irn) {
2139 * if value is defined in this block we can anways place the spill directly after the def
2140 * -> no constraint necessary
2142 if(!is_Phi(irn) && get_nodes_block(irn) == bb) continue;
2145 spill = set_find_spill(spill_bb->ilp, irn);
2148 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2149 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2151 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2152 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2155 sched_foreach_op(bb, tmp) {
2156 op_t *op = get_irn_link(tmp);
2158 if(is_Phi(tmp)) continue;
2159 assert(!is_Proj(tmp));
2162 ir_node *value = op->attr.remat.remat->value;
2165 /* only collect remats up to the first use of a value */
2166 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2171 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2172 ir_node *arg = get_irn_n(tmp, n);
2175 /* if a value is used stop collecting remats */
2181 if(cst == ILP_UNDEF) break;
2189 typedef struct _irnlist_t {
2190 struct list_head list;
2194 typedef struct _interference_t {
2195 struct list_head blocklist;
2201 cmp_interference(const void *a, const void *b, size_t size)
2203 const interference_t *p = a;
2204 const interference_t *q = b;
2206 return !(p->a == q->a && p->b == q->b);
2209 static interference_t *
2210 set_find_interference(set * set, ir_node * a, ir_node * b)
2212 interference_t query;
2214 query.a = (a>b)?a:b;
2215 query.b = (a>b)?b:a;
2217 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2220 static interference_t *
2221 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2223 interference_t query,
2225 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2229 result = set_find_interference(set, a, b);
2232 list_add(&list->list, &result->blocklist);
2236 query.a = (a>b)?a:b;
2237 query.b = (a>b)?b:a;
2239 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2241 INIT_LIST_HEAD(&result->blocklist);
2242 list_add(&list->list, &result->blocklist);
2248 values_interfere_in_block(ir_node * bb, ir_node * a, ir_node * b)
2250 const ir_edge_t *edge;
2252 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2253 /* both values are live in, so they interfere */
2257 /* ensure a dominates b */
2258 if(value_dominates(b,a)) {
2264 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2267 /* the following code is stolen from bera.c */
2268 if(is_live_end(bb, a))
2271 foreach_out_edge(a, edge) {
2272 const ir_node *user = edge->src;
2273 if(get_nodes_block(user) == bb
2276 && value_dominates(b, user))
2284 * Walk all irg blocks and collect interfering values inside of phi classes
2287 luke_interferencewalker(ir_node * bb, void * data)
2289 spill_ilp_t *si = (spill_ilp_t*)data;
2293 live_foreach(bb, li1) {
2294 ir_node *a = (ir_node *) li1->irn;
2295 op_t *a_op = get_irn_link(a);
2297 if(a_op->is_remat) continue;
2299 /* a is only interesting if it is in my register class and if it is inside a phi class */
2300 if (has_reg_class(si, a) && get_phi_class(a)) {
2301 for(li2=li1->next; li2; li2 = li2->next) {
2302 ir_node *b = (ir_node *) li2->irn;
2303 op_t *b_op = get_irn_link(b);
2305 if(b_op->is_remat) continue;
2307 /* a and b are only interesting if they are in the same phi class */
2308 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2309 if(values_interfere_in_block(bb, a, b)) {
2310 //DBG((si->dbg, LEVEL_1, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2311 ir_fprintf(stderr, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b);
2312 set_insert_interference(si, si->interferences, a, b, bb);
2320 static unsigned int copy_path_id = 0;
2323 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2330 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2331 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2333 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2335 pset_foreach(copies, ptr) {
2336 copy = PTR_TO_INT(ptr);
2337 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2342 * @parameter copies contains a path of copies which lead us to irn
2343 * @parameter visited contains a set of nodes already visited on this path
2346 find_copy_path(spill_ilp_t * si, ir_node * irn, ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2349 op_t *op = get_irn_link(irn);
2351 if(op->is_remat) return;
2353 pset_insert_ptr(visited, irn);
2358 /* visit all operands */
2359 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2360 ir_node *arg = get_irn_n(irn, n);
2361 ilp_var_t copy = op->attr.live_range.args.copies[n];
2363 if(!has_reg_class(si, arg)) continue;
2366 pset_insert(copies, INT_TO_PTR(copy), copy);
2367 write_copy_path_cst(si, copies, any_interfere);
2368 pset_remove(copies, INT_TO_PTR(copy), copy);
2370 if(!pset_find_ptr(visited, arg)) {
2371 pset_insert(copies, INT_TO_PTR(copy), copy);
2372 find_copy_path(si, arg, target, any_interfere, copies, visited);
2373 pset_remove(copies, INT_TO_PTR(copy), copy);
2379 /* visit all uses which are phis */
2380 foreach_out_edge(irn, edge) {
2381 ir_node *user = edge->src;
2382 int pos = edge->pos;
2383 op_t *op = get_irn_link(user);
2386 if(!is_Phi(user)) continue;
2387 if(!has_reg_class(si, user)) continue;
2389 copy = op->attr.live_range.args.copies[pos];
2391 if(user == target) {
2392 pset_insert(copies, INT_TO_PTR(copy), copy);
2393 write_copy_path_cst(si, copies, any_interfere);
2394 pset_remove(copies, INT_TO_PTR(copy), copy);
2396 if(!pset_find_ptr(visited, user)) {
2397 pset_insert(copies, INT_TO_PTR(copy), copy);
2398 find_copy_path(si, user, target, any_interfere, copies, visited);
2399 pset_remove(copies, INT_TO_PTR(copy), copy);
2404 pset_remove_ptr(visited, irn);
2408 gen_copy_constraints(spill_ilp_t * si, ir_node * a, ir_node * b, ilp_var_t any_interfere)
2410 pset * copies = pset_new_ptr_default();
2411 pset * visited = pset_new_ptr_default();
2413 find_copy_path(si, a, b, any_interfere, copies, visited);
2421 memcopyhandler(spill_ilp_t * si)
2423 interference_t *interference;
2425 /* teste Speicherwerte auf Interferenz */
2427 /* analyze phi classes */
2428 phi_class_compute(si->chordal_env->irg);
2430 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
2431 irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
2433 // phi_class_free(si->chordal_env->irg);
2435 /* now lets emit the ILP unequations for the crap */
2436 set_foreach(si->interferences, interference) {
2438 ilp_var_t interfere,
2440 ilp_cst_t any_interfere_cst,
2442 const ir_node *a = interference->a;
2443 const ir_node *b = interference->b;
2444 struct list_head *pos;
2446 /* any_interf <= \sum interf */
2447 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
2448 any_interfere_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2449 any_interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2451 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
2453 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
2454 const ir_node *bb = irnlist->irn;
2455 spill_bb_t *spill_bb = get_irn_link(bb);
2462 spilla = set_find_spill(spill_bb->ilp, a);
2466 spillb = set_find_spill(spill_bb->ilp, b);
2469 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
2470 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
2471 /* 2: - mem_in_a - spill_a + interfere <= 0 */
2472 /* 3: - mem_in_b - spill_b + interfere <= 0 */
2473 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
2474 interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2476 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
2477 cst = lpp_add_cst(si->lpp, buf, lpp_less, 1);
2479 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
2480 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
2481 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
2482 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
2483 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
2485 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
2486 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2488 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2489 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
2490 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
2492 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
2493 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2495 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2496 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
2497 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
2500 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
2502 /* any_interfere >= interf */
2503 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
2504 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2506 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2507 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2510 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
2511 gen_copy_constraints(si,a,b,any_interfere);
2518 memcopyinsertor(spill_ilp_t * si)
2520 /* weise Spillkontexte zu. Sorge bei Phis dafuer, dass gleiche
2521 * Kontexte zusammenfliessen (Operanden und Ergebnis hat gleichen
2537 return fabs(x) < 0.00001;
2541 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
2543 spill_ilp_t *si = get_irg_link(current_ir_graph);
2545 if(pset_find_ptr(si->all_possible_remats, n)) {
2546 op_t *op = (op_t*)get_irn_link(n);
2547 assert(op && op->is_remat);
2549 if(!op->attr.remat.remat->inverse) {
2550 if(op->attr.remat.pre) {
2551 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
2553 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
2558 op_t *op = (op_t*)get_irn_link(n);
2559 assert(op && op->is_remat);
2561 if(op->attr.remat.pre) {
2562 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
2564 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
2575 dump_graph_with_remats(ir_graph * irg, const char * suffix)
2577 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
2578 be_dump(irg, suffix, dump_ir_block_graph_sched);
2579 set_dump_node_vcgattr_hook(NULL);
2584 * Edge hook to dump the schedule edges with annotated register pressure.
2587 sched_pressure_edge_hook(FILE *F, ir_node *irn)
2589 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
2590 ir_node *prev = sched_prev(irn);
2591 fprintf(F, "edge:{sourcename:\"");
2593 fprintf(F, "\" targetname:\"");
2595 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
2596 fprintf(F, "\" color:magenta}\n");
2602 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
2604 DUMP_NODE_EDGE_FUNC old = get_dump_node_edge_hook();
2606 dump_consts_local(0);
2607 set_dump_node_edge_hook(sched_pressure_edge_hook);
2608 dump_ir_block_graph(irg, suffix);
2609 set_dump_node_edge_hook(old);
2613 walker_pressure_annotator(ir_node * bb, void * data)
2615 spill_ilp_t *si = data;
2619 pset *live = pset_new_ptr_default();
2622 live_foreach(bb, li) {
2623 irn = (ir_node *) li->irn;
2625 if (live_is_end(li) && has_reg_class(si, irn)) {
2626 pset_insert_ptr(live, irn);
2630 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
2632 sched_foreach_reverse(bb, irn) {
2634 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
2638 if(has_reg_class(si, irn)) {
2639 pset_remove_ptr(live, irn);
2640 if(is_Proj(irn)) ++projs;
2643 if(!is_Proj(irn)) projs = 0;
2645 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2646 ir_node *arg = get_irn_n(irn, n);
2648 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
2650 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
2657 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
2659 be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
2664 connect_all_remats_with_keep(spill_ilp_t * si)
2672 n_remats = pset_count(si->all_possible_remats);
2674 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
2677 pset_foreach(si->all_possible_remats, irn) {
2682 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
2684 obstack_free(si->obst, ins);
2690 connect_all_spills_with_keep(spill_ilp_t * si)
2699 n_spills = pset_count(si->spills);
2701 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
2704 pset_foreach(si->spills, irn) {
2709 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
2711 obstack_free(si->obst, ins);
2715 /** insert a spill at an arbitrary position */
2716 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert, ir_node *ctx)
2718 ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
2719 ir_graph *irg = get_irn_irg(bl);
2720 ir_node *frame = get_irg_frame(irg);
2724 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
2725 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
2727 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn, ctx);
2730 * search the right insertion point. a spill of a phi cannot be put
2731 * directly after the phi, if there are some phis behind the one which
2732 * is spilled. Also, a spill of a Proj must be after all Projs of the
2735 * Here's one special case:
2736 * If the spill is in the start block, the spill must be after the frame
2737 * pointer is set up. This is done by setting insert to the end of the block
2738 * which is its default initialization (see above).
2741 if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
2744 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
2747 sched_add_after(insert, spill);
2752 delete_remat(spill_ilp_t * si, ir_node * remat) {
2754 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2756 sched_remove(remat);
2758 /* kill links to operands */
2759 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
2760 set_irn_n(remat, n, bad);
2765 clean_remat_info(spill_ilp_t * si)
2769 remat_info_t *remat_info;
2770 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2772 set_foreach(si->remat_info, remat_info) {
2773 if(!remat_info->remats) continue;
2775 pset_foreach(remat_info->remats, remat)
2777 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
2778 set_irn_n(remat->proj, -1, bad);
2779 set_irn_n(remat->proj, 0, bad);
2782 if(get_irn_n_edges(remat->op) == 0) {
2783 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
2784 set_irn_n(remat->op, n, bad);
2789 if(remat_info->remats) del_pset(remat_info->remats);
2790 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
2795 delete_unnecessary_remats(spill_ilp_t * si)
2799 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2802 ir_node *end = get_irg_end(si->chordal_env->irg);
2805 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
2806 ir_node *keep_arg = get_irn_n(si->keep, n);
2807 op_t *arg_op = get_irn_link(keep_arg);
2810 assert(arg_op->is_remat);
2812 name = si->lpp->vars[arg_op->attr.remat.ilp];
2814 if(is_zero(name->value)) {
2815 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
2816 /* TODO check whether reload is preferred over remat (could be bug) */
2817 delete_remat(si, keep_arg);
2819 if(!arg_op->attr.remat.remat->inverse) {
2820 if(arg_op->attr.remat.pre) {
2821 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
2823 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
2826 if(arg_op->attr.remat.pre) {
2827 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
2829 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
2834 set_irn_n(si->keep, n, bad);
2837 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
2838 ir_node *end_arg = get_End_keepalive(end, i);
2840 if(end_arg != si->keep) {
2841 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
2844 keeps = obstack_finish(si->obst);
2845 set_End_keepalives(end, n-1, keeps);
2846 obstack_free(si->obst, keeps);
2849 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
2854 pset_foreach(si->all_possible_remats, remat) {
2855 op_t *remat_op = get_irn_link(remat);
2856 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
2858 if(is_zero(name->value)) {
2859 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
2860 /* TODO check whether reload is preferred over remat (could be bug) */
2861 delete_remat(si, remat);
2863 if(!remat_op->attr.remat.remat->inverse) {
2864 if(remat_op->attr.remat.pre) {
2865 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
2867 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
2870 if(remat_op->attr.remat.pre) {
2871 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
2873 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
2882 * @param before The node after which the spill will be placed in the schedule
2884 /* TODO set context properly */
2886 insert_spill(spill_ilp_t * si, ir_node * irn, ir_node * value, ir_node * before)
2890 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
2892 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
2894 spill = be_spill2(arch_env, irn, before, irn);
2896 defs = set_insert_def(si->values, value);
2899 /* enter into the linked list */
2900 set_irn_link(spill, defs->spills);
2901 defs->spills = spill;
2903 #ifdef KEEPALIVE_SPILLS
2904 pset_insert_ptr(si->spills, spill);
2911 * @param before The Phi node which has to be spilled
2914 insert_mem_phi(spill_ilp_t * si, const ir_node * phi)
2921 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
2923 for(n=get_irn_arity(phi)-1; n>=0; --n) {
2924 ins[n] = si->m_unknown;
2927 mem_phi = new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
2929 defs = set_insert_def(si->values, phi);
2932 /* enter into the linked list */
2933 set_irn_link(mem_phi, defs->spills);
2934 defs->spills = mem_phi;
2936 sched_add_after(phi, mem_phi);
2938 #ifdef KEEPALIVE_SPILLS
2939 pset_insert_ptr(si->spills, mem_phi);
2946 * Add remat to list of defs, destroys link field!
2949 insert_remat(spill_ilp_t * si, ir_node * remat)
2952 op_t *remat_op = get_irn_link(remat);
2954 assert(remat_op->is_remat);
2956 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
2959 /* enter into the linked list */
2960 set_irn_link(remat, defs->remats);
2961 defs->remats = remat;
2966 collect_spills(spill_ilp_t * si, ir_node * value, pset * spills, pset * visited)
2971 defs = set_find_def(si->values, value);
2973 if(defs && defs->spills) {
2974 for(next = defs->spills; next; next = get_irn_link(next)) {
2975 pset_insert_ptr(spills, next);
2977 } else if (is_Phi(value)) {
2979 if(!pset_find_ptr(visited, value)) {
2983 pset_insert_ptr(visited, value);
2984 for(i=0, n=get_irn_arity(value); i<n; ++i) {
2985 ir_node *arg = get_irn_n(value, i);
2987 collect_spills(si, arg, spills, visited);
2991 // assert(0 && "Phi operand not spilled");
2997 get_spills_for_value(spill_ilp_t * si, ir_node * value)
2999 pset *spills = pset_new_ptr_default();
3000 // pset *visited = pset_new_ptr_default();
3002 // collect_spills(si, value, spills, visited);
3003 // del_pset(visited);
3007 defs = set_find_def(si->values, value);
3009 if(defs && defs->spills) {
3010 for(next = defs->spills; next; next = get_irn_link(next)) {
3011 pset_insert_ptr(spills, next);
3019 * Add reload before operation and add to list of defs
3022 insert_reload(spill_ilp_t * si, const ir_node * value, const ir_node * after)
3027 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3029 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3031 defs = set_find_def(si->values, value);
3032 /* get a spill of this value */
3034 if((!defs || !defs->spills) && is_Phi(value)) {
3037 spills = get_spills_for_value(si, value);
3039 spill = pset_first(spills);
3043 defs = set_insert_def(si->values, value);
3045 defs->spills = spill;
3046 set_irn_link(spill, NULL);
3048 spill = defs->spills;
3051 spill = defs->spills;
3052 assert(spill && "no spill placed before reload");
3054 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3056 /* enter into the linked list */
3057 set_irn_link(reload, defs->remats);
3058 defs->remats = reload;
3064 walker_spill_placer(ir_node * bb, void * data) {
3065 spill_ilp_t *si = (spill_ilp_t*)data;
3067 spill_bb_t *spill_bb = get_irn_link(bb);
3068 pset *spills_to_do = pset_new_ptr_default();
3071 set_foreach(spill_bb->ilp, spill) {
3074 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3075 name = si->lpp->vars[spill->mem_in];
3076 if(!is_zero(name->value)) {
3079 mem_phi = insert_mem_phi(si, spill->irn);
3081 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3085 name = si->lpp->vars[spill->spill];
3086 if(!is_zero(name->value)) {
3087 /* place spill directly after definition */
3088 if(get_nodes_block(spill->irn) == bb) {
3089 insert_spill(si, spill->irn, spill->irn, spill->irn);
3093 /* place spill at bb start */
3094 if(spill->reg_in > 0) {
3095 name = si->lpp->vars[spill->reg_in];
3096 if(!is_zero(name->value)) {
3097 insert_spill(si, spill->irn, spill->irn, bb);
3101 /* place spill after a remat */
3102 pset_insert_ptr(spills_to_do, spill->irn);
3105 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3108 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3109 op_t *op = get_irn_link(irn);
3111 if(be_is_Spill(irn)) continue;
3114 /* TODO fix this if we want to support remats with more than two nodes */
3115 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3116 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3118 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3121 if(pset_find_ptr(spills_to_do, irn)) {
3122 pset_remove_ptr(spills_to_do, irn);
3124 insert_spill(si, irn, irn, irn);
3130 assert(pset_count(spills_to_do) == 0);
3132 /* afterwards free data in block */
3133 del_pset(spills_to_do);
3137 phim_fixer(spill_ilp_t *si) {
3140 set_foreach(si->values, defs) {
3141 const ir_node *phi = defs->value;
3142 ir_node *phi_m = NULL;
3143 ir_node *next = defs->spills;
3146 if(!is_Phi(phi)) continue;
3149 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3153 next = get_irn_link(next);
3156 if(!phi_m) continue;
3158 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3159 const ir_node *value = get_irn_n(phi, n);
3160 defs_t *val_defs = set_find_def(si->values, value);
3162 /* get a spill of this value */
3163 ir_node *spill = val_defs->spills;
3165 assert(spill && "no spill placed before PhiM");
3167 set_irn_n(phi_m, n, spill);
3173 walker_reload_placer(ir_node * bb, void * data) {
3174 spill_ilp_t *si = (spill_ilp_t*)data;
3176 spill_bb_t *spill_bb = get_irn_link(bb);
3180 /* reloads at end of block */
3181 if(spill_bb->reloads) {
3184 set_foreach(spill_bb->reloads, keyval) {
3185 ir_node *irn = (ir_node*)keyval->key;
3186 ilp_var_t reload = PTR_TO_INT(keyval->val);
3189 name = si->lpp->vars[reload];
3190 if(!is_zero(name->value)) {
3192 ir_node *insert_pos = bb;
3193 ir_node *prev = sched_block_last_noncf(si, bb);
3194 op_t *prev_op = get_irn_link(prev);
3196 /* insert reload before pre-remats */
3197 while(!sched_is_end(prev) && !be_is_Reload(prev) && !be_is_Spill(prev)
3198 && prev_op->is_remat && prev_op->attr.remat.pre) {
3201 prev = sched_prev(insert_pos);
3202 prev_op = get_irn_link(prev);
3205 reload = insert_reload(si, irn, insert_pos);
3207 #ifdef KEEPALIVE_RELOADS
3208 pset_insert_ptr(si->spills, reload);
3214 /* walk and insert more reloads and collect remats */
3215 sched_foreach_reverse(bb, irn) {
3216 op_t *op = get_irn_link(irn);
3218 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3219 if(is_Phi(irn)) break;
3222 if(get_irn_mode(irn) != mode_T) {
3223 insert_remat(si, irn);
3228 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3229 ir_node *arg = get_irn_n(irn, n);
3231 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3234 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3235 if(!is_zero(name->value)) {
3237 ir_node *insert_pos = irn;
3238 ir_node *prev = insert_pos;
3242 prev = sched_prev(prev);
3243 } while(be_is_Spill(prev));
3245 prev_op = get_irn_link(prev);
3247 /* insert reload before pre-remats */
3248 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3249 && prev_op->is_remat && prev_op->attr.remat.pre) {
3253 prev = sched_prev(prev);
3254 } while(be_is_Spill(prev));
3256 prev_op = get_irn_link(prev);
3260 reload = insert_reload(si, arg, insert_pos);
3262 set_irn_n(irn, n, reload);
3264 #ifdef KEEPALIVE_RELOADS
3265 pset_insert_ptr(si->spills, reload);
3273 del_set(spill_bb->ilp);
3274 if(spill_bb->reloads) del_set(spill_bb->reloads);
3278 walker_collect_used(ir_node * irn, void * data)
3280 lc_bitset_t *used = data;
3282 lc_bitset_set(used, get_irn_idx(irn));
3285 struct kill_helper {
3291 walker_kill_unused(ir_node * bb, void * data)
3293 struct kill_helper *kh = data;
3294 const ir_node *bad = get_irg_bad(get_irn_irg(bb));
3298 for(irn=sched_first(bb); !sched_is_end(irn);) {
3299 ir_node *next = sched_next(irn);
3302 if(!lc_bitset_is_set(kh->used, get_irn_idx(irn))) {
3303 if(be_is_Spill(irn) || be_is_Reload(irn)) {
3304 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)));
3306 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3312 set_nodes_block(irn, bad);
3313 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3314 set_irn_n(irn, n, bad);
3322 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3324 struct kill_helper kh;
3326 kh.used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3329 irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3330 irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3332 lc_bitset_free(kh.used);
3336 print_irn_pset(pset * p)
3340 pset_foreach(p, irn) {
3341 ir_printf("%+F\n", irn);
3346 rewire_uses(spill_ilp_t * si)
3348 dom_front_info_t *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
3350 pset *ignore = pset_new_ptr(1);
3352 pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
3354 /* then fix uses of spills */
3355 set_foreach(si->values, defs) {
3358 ir_node *next = defs->remats;
3361 reloads = pset_new_ptr_default();
3364 if(be_is_Reload(next)) {
3365 pset_insert_ptr(reloads, next);
3369 next = get_irn_link(next);
3372 spills = get_spills_for_value(si, defs->value);
3373 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));
3374 if(pset_count(spills) > 1) {
3375 //assert(pset_count(reloads) > 0);
3376 // print_irn_pset(spills);
3377 // print_irn_pset(reloads);
3379 be_ssa_constr_set_ignore(dfi, spills, ignore);
3386 /* first fix uses of remats and reloads */
3387 set_foreach(si->values, defs) {
3389 ir_node *next = defs->remats;
3392 nodes = pset_new_ptr_default();
3393 pset_insert_ptr(nodes, defs->value);
3396 pset_insert_ptr(nodes, next);
3397 next = get_irn_link(next);
3400 if(pset_count(nodes) > 1) {
3401 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
3402 be_ssa_constr_set(dfi, nodes);
3409 // remove_unused_defs(si);
3411 be_free_dominance_frontiers(dfi);
3415 writeback_results(spill_ilp_t * si)
3417 /* walk through the graph and collect all spills, reloads and remats for a value */
3419 si->values = new_set(cmp_defs, 4096);
3421 DBG((si->dbg, LEVEL_1, "Applying results\n"));
3422 delete_unnecessary_remats(si);
3423 si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
3424 irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
3426 irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
3428 /* clean the remat info! there are still back-edges leading there! */
3429 clean_remat_info(si);
3433 connect_all_spills_with_keep(si);
3435 del_set(si->values);
3439 get_n_regs(spill_ilp_t * si)
3441 int arch_n_regs = arch_register_class_n_regs(si->cls);
3445 for(i=0; i<arch_n_regs; i++) {
3446 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
3451 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
3456 walker_reload_mover(ir_node * bb, void * data)
3458 spill_ilp_t *si = data;
3461 sched_foreach(bb, tmp) {
3462 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
3463 ir_node *reload = tmp;
3466 /* move reload upwards */
3468 int pressure = (int)get_irn_link(reload);
3469 if(pressure < si->n_regs) {
3470 irn = sched_prev(reload);
3471 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
3472 sched_remove(reload);
3473 pressure = (int)get_irn_link(irn);
3475 while(pressure < si->n_regs) {
3476 if(sched_is_end(irn) || (be_is_Reload(irn) && has_reg_class(si, irn))) break;
3478 set_irn_link(irn, INT_TO_PTR(pressure+1));
3479 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
3480 irn = sched_prev(irn);
3482 pressure = (int)get_irn_link(irn);
3485 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
3486 sched_put_after(irn, reload);
3493 move_reloads_upward(spill_ilp_t * si)
3495 irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
3499 be_spill_remat(const be_chordal_env_t * chordal_env)
3501 char problem_name[256];
3502 char dump_suffix[256];
3503 char dump_suffix2[256];
3504 char dump_suffix3[256];
3505 struct obstack obst;
3508 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
3509 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
3510 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
3512 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
3513 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
3515 obstack_init(&obst);
3516 si.chordal_env = chordal_env;
3518 si.cls = chordal_env->cls;
3519 si.lpp = new_lpp(problem_name, lpp_minimize);
3520 si.remat_info = new_set(cmp_remat_info, 4096);
3521 si.interferences = new_set(cmp_interference, 4096);
3522 si.all_possible_remats = pset_new_ptr_default();
3523 si.spills = pset_new_ptr_default();
3524 si.inverse_ops = pset_new_ptr_default();
3525 #ifndef EXECFREQ_LOOPDEPH
3526 si.execfreqs = compute_execfreq(chordal_env->irg);
3528 si.execfreqs = NULL;
3533 si.n_regs = get_n_regs(&si);
3535 set_irg_link(chordal_env->irg, &si);
3536 compute_doms(chordal_env->irg);
3538 /* compute phi classes */
3539 // phi_class_compute(chordal_env->irg);
3541 be_analyze_regpressure(chordal_env, "-pre");
3543 #ifdef COLLECT_REMATS
3544 /* collect remats */
3545 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
3546 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
3549 /* insert possible remats */
3550 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
3551 irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
3552 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
3555 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
3556 connect_all_remats_with_keep(&si);
3557 /* dump graph with inserted remats */
3558 dump_graph_with_remats(chordal_env->irg, dump_suffix);
3561 /* insert copies for phi arguments not in my regclass */
3562 irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
3564 /* recompute liveness */
3565 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
3566 be_liveness(chordal_env->irg);
3570 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
3571 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
3572 irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
3574 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
3575 irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
3577 #ifndef NO_MEMCOPIES
3578 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
3579 memcopyhandler(&si);
3587 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
3588 if ((f = fopen(buf, "wt")) != NULL) {
3589 lpp_dump_plain(si.lpp, f);
3596 DBG((si.dbg, LEVEL_1, "\tSolving %F\n", chordal_env->irg));
3598 lpp_set_time_limit(si.lpp, ILP_TIMEOUT);
3602 lpp_solve_cplex(si.lpp);
3604 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
3606 assert(lpp_is_sol_valid(si.lpp)
3607 && "solution of ILP must be valid");
3609 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));
3611 #ifdef DUMP_SOLUTION
3616 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
3617 if ((f = fopen(buf, "wt")) != NULL) {
3619 for (i = 0; i < si.lpp->var_next; ++i) {
3620 lpp_name_t *name = si.lpp->vars[i];
3621 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
3628 writeback_results(&si);
3632 kill_all_unused_values_in_schedule(&si);
3634 #if defined(KEEPALIVE_SPILLS) || defined(KEEPALIVE_RELOADS)
3635 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
3638 // move reloads upwards
3639 be_liveness(chordal_env->irg);
3640 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
3641 move_reloads_upward(&si);
3643 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
3645 dump_pressure_graph(&si, dump_suffix2);
3647 // TODO fix temporarily exceeded regpressure due to remat2s
3649 // TODO insert copys to fix interferences in memory
3651 be_analyze_regpressure(chordal_env, "-post");
3653 free_dom(chordal_env->irg);
3654 del_set(si.interferences);
3655 del_pset(si.inverse_ops);
3656 del_pset(si.all_possible_remats);
3657 del_pset(si.spills);
3658 #ifndef EXECFREQ_LOOPDEPH
3659 free_execfreq(si.execfreqs);
3662 obstack_free(&obst, NULL);
3663 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
3666 #else /* WITH_ILP */
3669 only_that_you_can_compile_without_WITH_ILP_defined(void)
3673 #endif /* WITH_ILP */