1 /** vim: set sw=4 ts=4:
4 * @author Adam M. Szalkowski & Sebastian Hack
6 * ILP based spilling & rematerialization
8 * Copyright (C) 2006 Universitaet Karlsruhe
9 * Released under the GPL
32 #include "phiclass_t.h"
37 #include <lpp/lpp_net.h>
38 #include <lpp/lpp_cplex.h>
39 //#include <lc_pset.h>
40 #include <libcore/lc_bitset.h>
44 #include "besched_t.h"
49 #include "bespillremat.h"
51 #include "bepressurestat.h"
53 #include "bechordal_t.h"
59 //#define KEEPALIVE /* keep alive all inserted remats and dump graph with remats */
60 #define COLLECT_REMATS /* enable rematerialization */
61 #define COLLECT_INVERSE_REMATS /* enable placement of inverse remats */
62 #define REMAT_WHILE_LIVE /* only remat values that are live */
63 //#define NO_ENLARGE_L1V3N355 /* do not remat after the death of some operand */
64 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
65 #define MAY_DIE_AT_REMAT /* allow values to die after a pre remat */
66 #define NO_SINGLE_USE_REMATS /* do not repair schedule */
67 //#define KEEPALIVE_SPILLS
68 //#define KEEPALIVE_RELOADS
69 #define GOODWIN_REDUCTION
70 //#define NO_MEMCOPIES
74 #define LPP_SERVER "i44pc52"
75 #define LPP_SOLVER "cplex"
81 #define ILP_TIMEOUT 120
85 typedef struct _spill_ilp_t {
86 const arch_register_class_t *cls;
88 const be_chordal_env_t *chordal_env;
92 pset *all_possible_remats;
97 set *values; /**< for collecting all definitions of values before running ssa-construction */
101 DEBUG_ONLY(firm_dbg_module_t * dbg);
104 typedef int ilp_var_t;
105 typedef int ilp_cst_t;
107 typedef struct _spill_bb_t {
112 typedef struct _remat_t {
113 const ir_node *op; /**< for copy_irn */
114 const ir_node *value; /**< the value which is being recomputed by this remat */
115 ir_node *proj; /**< not NULL if the above op produces a tuple */
116 int cost; /**< cost of this remat */
117 int inverse; /**< nonzero if this is an inverse remat */
121 * Data to be attached to each IR node. For remats this contains the ilp_var
122 * for this remat and for normal ops this contains the ilp_vars for
123 * reloading each operand
125 typedef struct _op_t {
130 remat_t *remat; /** the remat this op belongs to */
131 int pre; /** 1, if this is a pressure-increasing remat */
135 ir_node *op; /** the operation this live range belongs to */
144 typedef struct _defs_t {
146 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
147 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
150 typedef struct _remat_info_t {
151 const ir_node *irn; /**< the irn to which these remats belong */
152 pset *remats; /**< possible remats for this value */
153 pset *remats_by_operand; /**< remats with this value as operand */
156 typedef struct _keyval_t {
161 typedef struct _spill_t {
171 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
173 return chordal_has_class(si->chordal_env, irn);
178 cmp_remat(const void *a, const void *b)
180 const keyval_t *p = a;
181 const keyval_t *q = b;
182 const remat_t *r = p->val;
183 const remat_t *s = q->val;
187 return !(r == s || r->op == s->op);
191 cmp_remat(const void *a, const void *b)
193 const remat_t *r = a;
194 const remat_t *s = a;
196 return !(r == s || r->op == s->op);
200 cmp_spill(const void *a, const void *b, size_t size)
202 const spill_t *p = a;
203 const spill_t *q = b;
205 // return !(p->irn == q->irn && p->bb == q->bb);
206 return !(p->irn == q->irn);
210 set_find_keyval(set * set, void * key)
215 return set_find(set, &query, sizeof(query), HASH_PTR(key));
219 set_insert_keyval(set * set, void * key, void * val)
225 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
229 set_find_def(set * set, ir_node * value)
234 return set_find(set, &query, sizeof(query), HASH_PTR(value));
238 set_insert_def(set * set, ir_node * value)
245 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
249 set_find_spill(set * set, ir_node * value)
254 return set_find(set, &query, sizeof(query), HASH_PTR(value));
257 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
258 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
259 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
260 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
261 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
264 cmp_remat_info(const void *a, const void *b, size_t size)
266 const remat_info_t *p = a;
267 const remat_info_t *q = b;
269 return !(p->irn == q->irn);
273 cmp_defs(const void *a, const void *b, size_t size)
278 return !(p->value == q->value);
282 cmp_keyval(const void *a, const void *b, size_t size)
284 const keyval_t *p = a;
285 const keyval_t *q = b;
287 return !(p->key == q->key);
291 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
294 #ifndef EXECFREQ_LOOPDEPH
295 return get_block_execfreq(si->chordal_env->exec_freq, get_block(irn)) + FUDGE;
298 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
300 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
305 get_cost(const spill_ilp_t * si, const ir_node * irn)
307 if(be_is_Spill(irn)) {
309 } else if(be_is_Reload(irn)){
312 return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
318 * Checks, whether node and its operands have suitable reg classes
321 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
324 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
325 int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
329 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
332 for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
333 ir_node *op = get_irn_n(irn, n);
334 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
337 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
344 * Try to create a remat from @p op with destination value @p dest_value
346 static INLINE remat_t *
347 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
349 remat_t *remat = NULL;
351 // if(!mode_is_datab(get_irn_mode(dest_value)))
354 if(dest_value == op) {
355 const ir_node *proj = NULL;
357 if(is_Proj(dest_value)) {
358 op = get_irn_n(op, 0);
362 if(!is_rematerializable(si, op))
365 remat = obstack_alloc(si->obst, sizeof(*remat));
367 remat->cost = get_cost(si, op);
368 remat->value = dest_value;
372 arch_inverse_t inverse;
375 /* get the index of the operand we want to retrieve by the inverse op */
376 for (n = get_irn_arity(op)-1; n>=0; --n) {
377 ir_node *arg = get_irn_n(op, n);
379 if(arg == dest_value) break;
383 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", n, op));
385 /* else ask the backend to give an inverse op */
386 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
389 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
391 assert(inverse.n > 0 && "inverse op should have at least one node");
393 for(i=inverse.n-1; i>=0; --i) {
394 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
398 remat = obstack_alloc(si->obst, sizeof(*remat));
399 remat->op = inverse.nodes[0];
400 remat->cost = inverse.costs;
401 remat->value = dest_value;
402 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
405 assert(is_Proj(remat->proj));
407 assert(0 && "I can not handle remats with more than 2 nodes");
414 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
416 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
424 add_remat(const spill_ilp_t * si, const remat_t * remat)
426 remat_info_t *remat_info,
431 assert(remat->value);
433 query.irn = remat->value;
435 query.remats_by_operand = NULL;
436 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
438 if(remat_info->remats == NULL) {
439 remat_info->remats = new_pset(cmp_remat, 4096);
441 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
443 /* insert the remat into the remats_be_operand set of each argument of the remat op */
444 for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
445 ir_node *arg = get_irn_n(remat->op, n);
449 query.remats_by_operand = NULL;
450 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
452 if(remat_info->remats_by_operand == NULL) {
453 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
455 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
460 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
462 const ir_edge_t *edge = get_irn_out_edge_first(irn);
466 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
469 edge = get_irn_out_edge_next(irn, edge);
476 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
481 #ifdef NO_SINGLE_USE_REMATS
482 if(has_reg_class(si, op) && (get_irn_n_nonremat_edges(si, op) > 1)) {
484 if(has_reg_class(si, op)) {
486 remat = get_remat_from_op(si, op, op);
488 add_remat(si, remat);
492 #ifdef COLLECT_INVERSE_REMATS
493 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
495 for (n = get_irn_arity(op)-1; n>=0; --n) {
496 ir_node *arg = get_irn_n(op, n);
498 if(has_reg_class(si, arg)) {
499 /* try to get an inverse remat */
500 remat = get_remat_from_op(si, arg, op);
502 add_remat(si, remat);
511 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
514 ir_node *def_block = get_nodes_block(val);
520 /* if pos is at end of a basic block */
522 ret = (pos == def_block || block_dominates(def_block, pos));
523 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
527 /* else if this is a normal operation */
528 block = get_nodes_block(pos);
529 if(block == def_block) {
530 if(!sched_is_scheduled(val)) return 1;
532 ret = sched_comes_after(val, pos);
533 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
537 ret = block_dominates(def_block, block);
538 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
542 static INLINE ir_node *
543 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
545 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
549 * Returns first non-Phi node of block @p bb
551 static INLINE ir_node *
552 sched_block_first_nonphi(const ir_node * bb)
554 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
558 sched_skip_proj_predicator(const ir_node * irn, void * data)
560 return (is_Proj(irn));
563 static INLINE ir_node *
564 sched_next_nonproj(const ir_node * irn, int forward)
566 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
570 * Returns next operation node (non-Proj) after @p irn
571 * or the basic block of this node
573 static INLINE ir_node *
574 sched_next_op(const ir_node * irn)
576 ir_node *next = sched_next(irn);
581 return sched_next_nonproj(next, 1);
585 * Returns previous operation node (non-Proj) before @p irn
586 * or the basic block of this node
588 static INLINE ir_node *
589 sched_prev_op(const ir_node * irn)
591 ir_node *prev = sched_prev(irn);
596 return sched_next_nonproj(prev, 0);
600 sched_put_after(ir_node * insert, ir_node * irn)
602 if(is_Block(insert)) {
603 insert = sched_block_first_nonphi(insert);
605 insert = sched_next_op(insert);
607 sched_add_before(insert, irn);
611 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
613 if(is_Block(insert)) {
614 insert = sched_block_last_noncf(si, insert);
616 insert = sched_next_nonproj(insert, 0);
617 insert = sched_prev(insert);
619 sched_add_after(insert, irn);
623 * Tells you whether a @p remat can be placed before the irn @p pos
626 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
628 const ir_node *op = remat->op;
634 prev = sched_block_last_noncf(si, pos);
635 prev = sched_next_nonproj(prev, 0);
637 prev = sched_prev_op(pos);
639 /* do not remat if the rematted value is defined immediately before this op */
640 if(prev == remat->op) {
645 /* this should be just fine, the following OP will be using this value, right? */
647 /* only remat AFTER the real definition of a value (?) */
648 if(!value_is_defined_before(si, pos, remat->value)) {
649 // ir_fprintf(stderr, "error(not defined)");
654 for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
655 const ir_node *arg = get_irn_n(op, n);
657 #ifdef NO_ENLARGE_L1V3N355
658 if(has_reg_class(si, arg) && live) {
659 res &= pset_find_ptr(live, arg)?1:0;
661 res &= value_is_defined_before(si, pos, arg);
664 res &= value_is_defined_before(si, pos, arg);
672 * Tells you whether a @p remat can be placed after the irn @p pos
675 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
678 pos = sched_block_first_nonphi(pos);
680 pos = sched_next_op(pos);
683 /* only remat AFTER the real definition of a value (?) */
684 if(!value_is_defined_before(si, pos, remat->value)) {
688 return can_remat_before(si, remat, pos, live);
692 * Collect potetially rematerializable OPs
695 walker_remat_collector(ir_node * irn, void * data)
697 spill_ilp_t *si = data;
699 if(!is_Block(irn) && !is_Phi(irn)) {
700 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
701 get_remats_from_op(si, irn);
706 * Inserts a copy of @p irn before @p pos
709 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
714 bb = is_Block(pos)?pos:get_nodes_block(pos);
715 copy = exact_copy(irn);
717 _set_phi_class(copy, NULL);
718 set_nodes_block(copy, bb);
719 sched_put_before(si, pos, copy);
725 * Inserts a copy of @p irn after @p pos
728 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
733 bb = is_Block(pos)?pos:get_nodes_block(pos);
734 copy = exact_copy(irn);
736 _set_phi_class(copy, NULL);
737 set_nodes_block(copy, bb);
738 sched_put_after(pos, copy);
744 insert_remat_after(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
748 if(can_remat_after(si, remat, pos, live)) {
753 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
755 copy = insert_copy_after(si, remat->op, pos);
757 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
758 op = obstack_alloc(si->obst, sizeof(*op));
760 op->attr.remat.remat = remat;
761 op->attr.remat.pre = 0;
762 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
764 set_irn_link(copy, op);
765 pset_insert_ptr(si->all_possible_remats, copy);
767 proj_copy = insert_copy_after(si, remat->proj, copy);
768 set_irn_n(proj_copy, 0, copy);
769 set_irn_link(proj_copy, op);
770 pset_insert_ptr(si->all_possible_remats, proj_copy);
782 insert_remat_before(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
786 if(can_remat_before(si, remat, pos, live)) {
791 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
793 copy = insert_copy_before(si, remat->op, pos);
795 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
796 op = obstack_alloc(si->obst, sizeof(*op));
798 op->attr.remat.remat = remat;
799 op->attr.remat.pre = 1;
800 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
802 set_irn_link(copy, op);
803 pset_insert_ptr(si->all_possible_remats, copy);
805 proj_copy = insert_copy_after(si, remat->proj, copy);
806 set_irn_n(proj_copy, 0, copy);
807 set_irn_link(proj_copy, op);
808 pset_insert_ptr(si->all_possible_remats, proj_copy);
820 get_block_n_succs(const ir_node *block) {
821 const ir_edge_t *edge;
823 assert(edges_activated(current_ir_graph));
825 edge = get_block_succ_first(block);
829 edge = get_block_succ_next(block, edge);
834 is_merge_edge(const ir_node * bb)
836 #ifdef GOODWIN_REDUCTION
837 return get_block_n_succs(bb) == 1;
844 is_diverge_edge(const ir_node * bb)
846 #ifdef GOODWIN_REDUCTION
847 return get_Block_n_cfgpreds(bb) == 1;
854 walker_regclass_copy_insertor(ir_node * irn, void * data)
856 spill_ilp_t *si = data;
858 if(is_Phi(irn) && has_reg_class(si, irn)) {
861 for(n=get_irn_arity(irn)-1; n>=0; --n) {
862 ir_node *phi_arg = get_irn_n(irn, n);
863 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
865 if(!has_reg_class(si, phi_arg)) {
866 ir_node *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
867 ir_node *pos = sched_block_last_noncf(si, bb);
868 op_t *op = obstack_alloc(si->obst, sizeof(*op));
870 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
871 sched_add_after(pos, copy);
872 set_irn_n(irn, n, copy);
875 op->attr.live_range.args.reloads = NULL;
876 op->attr.live_range.ilp = ILP_UNDEF;
877 set_irn_link(copy, op);
885 * Insert (so far unused) remats into the irg to
886 * recompute the potential liveness of all values
889 walker_remat_insertor(ir_node * bb, void * data)
891 spill_ilp_t *si = data;
892 spill_bb_t *spill_bb;
896 pset *live = pset_new_ptr_default();
898 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
900 live_foreach(bb, li) {
901 ir_node *value = (ir_node *) li->irn;
903 /* add remats at end of block */
904 if (live_is_end(li) && has_reg_class(si, value)) {
905 pset_insert_ptr(live, value);
909 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
910 set_irn_link(bb, spill_bb);
912 irn = sched_last(bb);
913 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);
963 /* delete defined value from live set */
964 if(has_reg_class(si, irn)) {
965 pset_remove_ptr(live, irn);
969 remat_args = pset_new_ptr_default();
971 /* insert all possible remats before irn */
972 pset_foreach(args, arg) {
973 remat_info_t *remat_info,
977 /* continue if the operand has the wrong reg class
979 if(!has_reg_class(si, arg))
984 query.remats_by_operand = NULL;
985 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
991 if(remat_info->remats) {
992 pset_foreach(remat_info->remats, remat) {
993 ir_node *remat_irn = NULL;
995 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
996 #ifdef REMAT_WHILE_LIVE
997 if(pset_find_ptr(live, remat->value)) {
998 remat_irn = insert_remat_before(si, remat, irn, live);
1001 remat_irn = insert_remat_before(si, remat, irn, live);
1004 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1005 ir_node *remat_arg = get_irn_n(remat_irn, n);
1007 if(!has_reg_class(si, remat_arg)) continue;
1009 pset_insert_ptr(remat_args, remat_arg);
1016 /* now we add remat args to op's args because they could also die at this op */
1017 pset_foreach(args,arg) {
1018 if(pset_find_ptr(remat_args, arg)) {
1019 pset_remove_ptr(remat_args, arg);
1022 pset_foreach(remat_args,arg) {
1023 pset_insert_ptr(args, arg);
1026 /* insert all possible remats after irn */
1027 pset_foreach(args, arg) {
1028 remat_info_t *remat_info,
1032 /* continue if the operand has the wrong reg class */
1033 if(!has_reg_class(si, arg))
1037 query.remats = NULL;
1038 query.remats_by_operand = NULL;
1039 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1045 /* do not place post remats after jumps */
1046 if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
1048 if(remat_info->remats_by_operand) {
1049 pset_foreach(remat_info->remats_by_operand, remat) {
1050 /* do not insert remats producing the same value as one of the operands */
1051 if(!pset_find_ptr(args, remat->value)) {
1052 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
1053 #ifdef REMAT_WHILE_LIVE
1054 if(pset_find_ptr(live, remat->value)) {
1055 insert_remat_after(si, remat, irn, live);
1058 insert_remat_after(si, remat, irn, live);
1065 del_pset(remat_args);
1070 live_foreach(bb, li) {
1071 ir_node *value = (ir_node *) li->irn;
1073 /* add remats at end if successor has multiple predecessors */
1074 if(is_merge_edge(bb)) {
1075 /* add remats at end of block */
1076 if (live_is_end(li) && has_reg_class(si, value)) {
1077 remat_info_t *remat_info,
1082 query.remats = NULL;
1083 query.remats_by_operand = NULL;
1084 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1086 if(remat_info && remat_info->remats) {
1087 pset_foreach(remat_info->remats, remat) {
1088 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1090 insert_remat_before(si, remat, bb, NULL);
1095 if(is_diverge_edge(bb)) {
1096 /* add remat2s at beginning of block */
1097 if ((live_is_in(li) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) {
1098 remat_info_t *remat_info,
1103 query.remats = NULL;
1104 query.remats_by_operand = NULL;
1105 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1107 if(remat_info && remat_info->remats) {
1108 pset_foreach(remat_info->remats, remat) {
1109 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at beginning of block %+F\n", remat->op, bb));
1111 /* put the remat here if all its args are available */
1112 insert_remat_after(si, remat, bb, NULL);
1122 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1125 luke_endwalker(ir_node * bb, void * data)
1127 spill_ilp_t *si = (spill_ilp_t*)data;
1134 spill_bb_t *spill_bb = get_irn_link(bb);
1137 live = pset_new_ptr_default();
1138 use_end = pset_new_ptr_default();
1140 live_foreach(bb, li) {
1141 irn = (ir_node *) li->irn;
1142 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1145 pset_insert_ptr(live, irn);
1146 op = get_irn_link(irn);
1147 assert(!op->is_remat);
1151 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1152 /* their reg_out must always be set */
1153 sched_foreach_reverse(bb, irn) {
1156 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1158 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1159 ir_node *irn_arg = get_irn_n(irn, n);
1161 if(has_reg_class(si, irn_arg)) {
1162 pset_insert_ptr(use_end, irn_arg);
1167 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1168 //cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1169 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1171 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1173 /* if this is a merge edge we can reload at the end of this block */
1174 if(is_merge_edge(bb)) {
1175 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1176 } else if(pset_count(use_end)){
1177 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1179 spill_bb->reloads = NULL;
1182 pset_foreach(live,irn) {
1188 /* handle values used by control flow nodes later separately */
1189 if(pset_find_ptr(use_end, irn)) continue;
1192 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1194 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1196 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1197 spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1198 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1200 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1201 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1203 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1204 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1206 if(is_merge_edge(bb)) {
1210 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1211 reload = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1212 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1214 /* reload <= mem_out */
1215 rel_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1216 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1217 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1220 spill->reg_in = ILP_UNDEF;
1221 spill->mem_in = ILP_UNDEF;
1224 pset_foreach(use_end,irn) {
1228 ilp_cst_t end_use_req,
1233 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1235 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1237 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1238 spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1239 /* if irn is used at the end of the block, then it is live anyway */
1240 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1242 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1243 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1245 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1246 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1248 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1249 reload = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1250 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1252 /* reload <= mem_out */
1253 rel_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1254 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1255 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1257 spill->reg_in = ILP_UNDEF;
1258 spill->mem_in = ILP_UNDEF;
1260 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1261 end_use_req = lpp_add_cst(si->lpp, buf, lpp_equal, 1);
1262 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1270 next_post_remat(const ir_node * irn)
1275 irn = sched_block_first_nonphi(irn);
1277 irn = sched_next_op(irn);
1280 if(sched_is_end(irn))
1283 op = (op_t*)get_irn_link(irn);
1284 if(op->is_remat && !op->attr.remat.pre) {
1293 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1299 ret = sched_block_last_noncf(si, irn);
1300 ret = sched_next(ret);
1301 ret = sched_prev_op(ret);
1303 ret = sched_prev_op(irn);
1306 if(sched_is_end(ret) || is_Phi(ret))
1309 op = (op_t*)get_irn_link(ret);
1310 if(op->is_remat && op->attr.remat.pre) {
1318 * Find a remat of value @p value in the epilog of @p pos
1321 find_post_remat(const ir_node * value, const ir_node * pos)
1323 while((pos = next_post_remat(pos)) != NULL) {
1326 op = get_irn_link(pos);
1327 assert(op->is_remat && !op->attr.remat.pre);
1329 if(op->attr.remat.remat->value == value)
1330 return (ir_node*)pos;
1333 const ir_edge_t *edge;
1334 foreach_out_edge(pos, edge) {
1335 ir_node *proj = get_edge_src_irn(edge);
1336 assert(is_Proj(proj));
1346 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1348 spill_bb_t *spill_bb = get_irn_link(bb);
1354 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1356 double spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1358 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1360 spill->reg_out = ILP_UNDEF;
1361 spill->reg_in = ILP_UNDEF;
1362 spill->mem_in = ILP_UNDEF;
1364 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1365 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1367 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1368 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1375 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1380 live_foreach(bb, li) {
1381 irn = (ir_node *) li->irn;
1383 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1384 pset_insert_ptr(live, irn);
1388 irn = sched_last(bb);
1390 /* all values eaten by control flow operations are also live until the end of the block */
1391 sched_foreach_reverse(bb, irn) {
1394 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1396 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1397 ir_node *arg = get_irn_n(irn,i);
1399 if(has_reg_class(si, arg)) {
1400 pset_insert_ptr(live, arg);
1407 * Inserts ILP-constraints and variables for memory copying before the given position
1410 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1412 const ir_node *succ;
1413 const ir_edge_t *edge;
1414 spill_bb_t *spill_bb = get_irn_link(block);
1423 assert(edges_activated(current_ir_graph));
1425 edge = get_block_succ_first(block);
1431 edge = get_block_succ_next(block, edge);
1432 /* next block can only contain phis, if this is a merge edge */
1435 ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1436 copyreg = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1438 ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1439 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1441 pset_foreach(live, tmp) {
1444 op_t *op = get_irn_link(irn);
1445 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1447 spill = set_find_spill(spill_bb->ilp, tmp);
1450 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1452 lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1454 sched_foreach(succ, phi) {
1455 const ir_node *to_copy;
1457 spill_t *to_copy_spill;
1458 op_t *phi_op = get_irn_link(phi);
1459 ilp_var_t reload = ILP_UNDEF;
1462 if(!is_Phi(phi)) break;
1463 if(!has_reg_class(si, phi)) continue;
1465 to_copy = get_irn_n(phi, pos);
1467 to_copy_op = get_irn_link(to_copy);
1469 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1470 assert(to_copy_spill);
1472 if(spill_bb->reloads) {
1473 keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1476 reload = PTR_TO_INT(keyval->val);
1480 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N", block, to_copy);
1481 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1483 /* copy - reg_out - reload - remat - live_range <= 0 */
1484 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1485 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1486 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1487 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1488 foreach_pre_remat(si, block, tmp) {
1489 op_t *remat_op = get_irn_link(tmp);
1490 if(remat_op->attr.remat.remat->value == to_copy) {
1491 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1495 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N", block, to_copy);
1496 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1498 /* copy - reg_out - copyreg <= 0 */
1499 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1500 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1501 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1507 * Walk all irg blocks and emit this ILP
1510 luke_blockwalker(ir_node * bb, void * data)
1512 spill_ilp_t *si = (spill_ilp_t*)data;
1517 spill_bb_t *spill_bb = get_irn_link(bb);
1520 pset *defs = pset_new_ptr_default();
1523 live = pset_new_ptr_default();
1525 /****************************************
1526 * B A S I C B L O C K E N D
1527 ***************************************/
1530 /* init live values at end of block */
1531 get_live_end(si, bb, live);
1533 pset_foreach(live, irn) {
1535 ilp_var_t reload = ILP_UNDEF;
1537 spill = set_find_spill(spill_bb->ilp, irn);
1540 if(spill_bb->reloads) {
1541 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1544 reload = PTR_TO_INT(keyval->val);
1548 op = get_irn_link(irn);
1549 assert(!op->is_remat);
1551 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1552 op->attr.live_range.ilp = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1553 op->attr.live_range.op = bb;
1555 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1556 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1558 /* reg_out - reload - remat - live_range <= 0 */
1559 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1560 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1561 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1562 foreach_pre_remat(si, bb, tmp) {
1563 op_t *remat_op = get_irn_link(tmp);
1564 if(remat_op->attr.remat.remat->value == irn) {
1565 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1568 /* maybe we should also assure that reg_out >= live_range etc. */
1571 #ifndef NO_MEMCOPIES
1572 insert_mem_copy_position(si, live, bb);
1576 * start new live ranges for values used by remats at end of block
1577 * and assure the remat args are available
1579 foreach_pre_remat(si, bb, tmp) {
1580 op_t *remat_op = get_irn_link(tmp);
1583 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1584 ir_node *remat_arg = get_irn_n(tmp, n);
1585 op_t *arg_op = get_irn_link(remat_arg);
1588 if(!has_reg_class(si, remat_arg)) continue;
1590 /* if value is becoming live through use by remat */
1591 if(!pset_find_ptr(live, remat_arg)) {
1592 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1593 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1595 arg_op->attr.live_range.ilp = prev_lr;
1596 arg_op->attr.live_range.op = bb;
1598 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1600 pset_insert_ptr(live, remat_arg);
1601 add_to_spill_bb(si, bb, remat_arg);
1604 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1605 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1606 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1608 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1609 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1611 /* use reload placed for this argument */
1612 if(spill_bb->reloads) {
1613 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1616 ilp_var_t reload = PTR_TO_INT(keyval->val);
1618 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1623 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1628 /**************************************
1629 * B A S I C B L O C K B O D Y
1630 **************************************/
1632 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1638 ilp_cst_t check_pre,
1645 /* iterate only until first phi */
1649 op = get_irn_link(irn);
1651 if(op->is_remat) continue;
1652 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1654 /* collect defined values */
1655 if(has_reg_class(si, irn)) {
1656 pset_insert_ptr(defs, irn);
1660 if(is_Proj(irn)) continue;
1663 * init set of irn's arguments
1664 * and all possibly used values around this op
1665 * and values defined by post remats
1667 args = new_set(cmp_keyval, get_irn_arity(irn));
1668 used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1669 remat_defs = pset_new_ptr(pset_count(live));
1671 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1672 ir_node *irn_arg = get_irn_n(irn, n);
1673 if(has_reg_class(si, irn_arg)) {
1674 set_insert_keyval(args, irn_arg, (void*)n);
1675 pset_insert_ptr(used, irn_arg);
1678 foreach_post_remat(irn, tmp) {
1679 op_t *remat_op = get_irn_link(tmp);
1681 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1683 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1684 ir_node *remat_arg = get_irn_n(tmp, n);
1685 if(has_reg_class(si, remat_arg)) {
1686 pset_insert_ptr(used, remat_arg);
1690 foreach_pre_remat(si, irn, tmp) {
1691 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1692 ir_node *remat_arg = get_irn_n(tmp, n);
1693 if(has_reg_class(si, remat_arg)) {
1694 pset_insert_ptr(used, remat_arg);
1699 /**********************************
1700 * I N E P I L O G O F irn
1701 **********************************/
1703 /* ensure each dying value is used by only one post remat */
1704 pset_foreach(live, tmp) {
1705 ir_node *value = tmp;
1706 op_t *value_op = get_irn_link(value);
1711 foreach_post_remat(irn, remat) {
1712 op_t *remat_op = get_irn_link(remat);
1714 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1715 ir_node *remat_arg = get_irn_n(remat, n);
1717 /* if value is used by this remat add it to constraint */
1718 if(remat_arg == value) {
1720 /* sum remat2s <= 1 + n_remats*live_range */
1721 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1722 cst = lpp_add_cst(si->lpp, buf, lpp_less, 1.0);
1726 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1732 if(value_op->attr.live_range.ilp != ILP_UNDEF && cst != ILP_UNDEF) {
1733 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1739 /* new live ranges for values from L\U defined by post remats */
1740 pset_foreach(live, tmp) {
1741 ir_node *value = tmp;
1742 op_t *value_op = get_irn_link(value);
1744 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
1745 ilp_var_t prev_lr = ILP_UNDEF;
1748 if(pset_find_ptr(remat_defs, value)) {
1750 /* next_live_range <= prev_live_range + sum remat2s */
1751 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1752 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1754 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1755 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1757 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1758 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1760 foreach_post_remat(irn, remat) {
1761 op_t *remat_op = get_irn_link(remat);
1763 /* if value is being rematerialized by this remat */
1764 if(value == remat_op->attr.remat.remat->value) {
1765 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1769 value_op->attr.live_range.ilp = prev_lr;
1770 value_op->attr.live_range.op = irn;
1775 /* requirements for post remats and start live ranges from L/U' for values dying here */
1776 foreach_post_remat(irn, tmp) {
1777 op_t *remat_op = get_irn_link(tmp);
1780 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1781 ir_node *remat_arg = get_irn_n(tmp, n);
1782 op_t *arg_op = get_irn_link(remat_arg);
1784 if(!has_reg_class(si, remat_arg)) continue;
1786 /* only for values in L\U (TODO and D?), the others are handled with post_use */
1787 if(!pset_find_ptr(used, remat_arg)) {
1788 /* remat <= live_range(remat_arg) */
1789 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
1790 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1792 /* if value is becoming live through use by remat2 */
1793 if(!pset_find_ptr(live, remat_arg)) {
1796 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1797 lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1799 arg_op->attr.live_range.ilp = lr;
1800 arg_op->attr.live_range.op = irn;
1802 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
1804 pset_insert_ptr(live, remat_arg);
1805 add_to_spill_bb(si, bb, remat_arg);
1808 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1809 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1814 d = pset_count(defs);
1815 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
1817 /* count how many regs irn needs for arguments */
1818 u = set_count(args);
1821 /* check the register pressure in the epilog */
1822 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
1823 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
1824 check_post = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - d);
1826 /* add L\U' to check_post */
1827 pset_foreach(live, tmp) {
1828 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
1829 /* if a live value is not used by irn */
1830 tmp_op = get_irn_link(tmp);
1831 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
1835 /***********************************************************
1836 * 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
1837 **********************************************************/
1840 pset_foreach(used, tmp) {
1846 op_t *arg_op = get_irn_link(arg);
1849 spill = add_to_spill_bb(si, bb, arg);
1851 /* new live range for each used value */
1852 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
1853 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1855 /* the epilog stuff - including post_use, check_post, check_post_remat */
1856 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
1857 post_use = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1859 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
1861 /* arg is live throughout epilog if the next live_range is in a register */
1862 if(pset_find_ptr(live, arg)) {
1863 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
1865 /* post_use >= next_lr + remat */
1866 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1867 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1868 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
1869 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1873 /* if value is not an arg of op and not possibly defined by post remat
1874 * then it may only die and not become live
1876 if(!set_find_keyval(args, arg)) {
1877 /* post_use <= prev_lr */
1878 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
1879 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1880 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1881 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1883 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
1884 /* next_lr <= prev_lr */
1885 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
1886 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1887 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1888 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1894 /* forall post remat which use arg add a similar cst */
1895 foreach_post_remat(irn, remat) {
1898 for (n=get_irn_arity(remat)-1; n>=0; --n) {
1899 ir_node *remat_arg = get_irn_n(remat, n);
1900 op_t *remat_op = get_irn_link(remat);
1902 if(remat_arg == arg) {
1903 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
1905 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1906 cst = lpp_add_cst(si->lpp, buf, lpp_greater, 0.0);
1907 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1908 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1913 /* new live range begins for each used value */
1914 arg_op->attr.live_range.ilp = prev_lr;
1915 arg_op->attr.live_range.op = irn;
1917 /*if(!pset_find_ptr(live, arg)) {
1918 pset_insert_ptr(live, arg);
1919 add_to_spill_bb(si, bb, arg);
1921 pset_insert_ptr(live, arg);
1925 /* just to be sure */
1926 check_post = ILP_UNDEF;
1935 /* check the register pressure in the prolog */
1936 /* sum_{L\U} lr <= k - |U| */
1937 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
1938 check_pre = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - u);
1940 /* for the prolog remove defined values from the live set */
1941 pset_foreach(defs, tmp) {
1942 pset_remove_ptr(live, tmp);
1945 /***********************************************************
1946 * 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
1947 **********************************************************/
1950 set_foreach(args, keyval) {
1952 ir_node *arg = keyval->key;
1953 int i = PTR_TO_INT(keyval->val);
1954 op_t *arg_op = get_irn_link(arg);
1956 spill = set_find_spill(spill_bb->ilp, arg);
1959 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
1960 op->attr.live_range.args.reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1962 /* reload <= mem_out */
1963 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
1964 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1965 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
1966 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
1968 /* requirement: arg must be in register for use */
1969 /* reload + remat + live_range == 1 */
1970 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
1971 cst = lpp_add_cst(si->lpp, buf, lpp_equal, 1.0);
1973 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1974 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
1975 foreach_pre_remat(si, irn, tmp) {
1976 op_t *remat_op = get_irn_link(tmp);
1977 if(remat_op->attr.remat.remat->value == arg) {
1978 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1983 /* iterate over L\U */
1984 pset_foreach(live, tmp) {
1985 if(!set_find_keyval(args, tmp)) {
1986 /* if a live value is not used by irn */
1987 tmp_op = get_irn_link(tmp);
1988 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
1993 /* requirements for remats */
1994 /* start new live ranges for values used by remats */
1995 foreach_pre_remat(si, irn, tmp) {
1996 op_t *remat_op = get_irn_link(tmp);
1999 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2000 ir_node *remat_arg = get_irn_n(tmp, n);
2001 op_t *arg_op = get_irn_link(remat_arg);
2004 if(!has_reg_class(si, remat_arg)) continue;
2006 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2007 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2008 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2010 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2011 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2013 /* if remat arg is also used by current op then we can use reload placed for this argument */
2014 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2015 int index = (int)keyval->val;
2017 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2025 /*************************
2026 * D O N E W I T H O P
2027 *************************/
2029 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2031 pset_foreach(live, tmp) {
2032 assert(has_reg_class(si, tmp));
2035 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2036 ir_node *arg = get_irn_n(irn, n);
2038 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2041 del_pset(remat_defs);
2045 defs = pset_new_ptr_default();
2050 /***************************************
2051 * B E G I N N I N G O F B L O C K
2052 ***************************************/
2055 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2056 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2058 pset_foreach(live, irn) {
2059 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2062 /* construct mem_outs for all values */
2064 set_foreach(spill_bb->ilp, spill) {
2065 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2066 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2068 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2069 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2071 if(pset_find_ptr(live, spill->irn)) {
2072 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2074 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2075 spill->mem_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2076 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2078 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2080 op_t *op = get_irn_link(spill->irn);
2082 /* do we have to copy a phi argument? */
2083 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2084 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2086 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2087 const ir_node *arg = get_irn_n(spill->irn, n);
2093 /* argument already done? */
2094 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2096 /* get sum of execution frequencies of blocks with the same phi argument */
2097 for(m=n; m>=0; --m) {
2098 const ir_node *arg2 = get_irn_n(spill->irn, m);
2101 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2105 /* copies are not for free */
2106 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2107 var = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE * freq);
2109 for(m=n; m>=0; --m) {
2110 const ir_node *arg2 = get_irn_n(spill->irn, m);
2113 op->attr.live_range.args.copies[m] = var;
2117 /* copy <= mem_in */
2118 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2119 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2120 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2121 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2128 /* L\U is empty at bb start */
2129 /* arg is live throughout epilog if it is reg_in into this block */
2131 /* check the register pressure at the beginning of the block
2134 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2135 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
2137 pset_foreach(live, irn) {
2140 spill = set_find_spill(spill_bb->ilp, irn);
2143 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2144 spill->reg_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2146 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2148 /* spill + mem_in <= 1 */
2149 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2150 nospill = lpp_add_cst(si->lpp, buf, lpp_less, 1);
2152 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2153 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2156 foreach_post_remat(bb, irn) {
2157 op_t *remat_op = get_irn_link(irn);
2159 DBG((si->dbg, LEVEL_4, "\t next post remat: %+F\n", irn));
2160 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
2162 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2165 /* forall post remats add requirements */
2166 foreach_post_remat(bb, tmp) {
2169 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2170 ir_node *remat_arg = get_irn_n(tmp, n);
2171 op_t *remat_op = get_irn_link(tmp);
2173 if(!has_reg_class(si, remat_arg)) continue;
2175 spill = set_find_spill(spill_bb->ilp, remat_arg);
2178 /* remat <= reg_in_argument */
2179 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2180 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2181 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2182 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2186 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2187 pset_foreach(live, irn) {
2191 spill = set_find_spill(spill_bb->ilp, irn);
2192 assert(spill && spill->irn == irn);
2194 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2195 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2198 ir_node *phi_arg = get_Phi_pred(irn, n);
2199 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2200 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2203 /* although the phi is in the right regclass one or more of
2204 * its arguments can be in a different one or at least to
2207 if(has_reg_class(si, phi_arg)) {
2208 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2209 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2210 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2211 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2213 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2214 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2216 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2219 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2220 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2224 /* else assure the value arrives on all paths in the same resource */
2226 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2229 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2230 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2233 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2234 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2235 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2236 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2238 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2239 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2241 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2244 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2245 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2250 /* first live ranges from reg_ins */
2251 pset_foreach(live, irn) {
2252 op_t *op = get_irn_link(irn);
2254 spill = set_find_spill(spill_bb->ilp, irn);
2255 assert(spill && spill->irn == irn);
2257 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2258 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2259 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2260 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2262 foreach_post_remat(bb, tmp) {
2263 op_t *remat_op = get_irn_link(tmp);
2265 if(remat_op->attr.remat.remat->value == irn) {
2266 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2271 /* walk forward now and compute constraints for placing spills */
2272 /* this must only be done for values that are not defined in this block */
2273 /* TODO are these values at start of block? if yes, just check whether this is a diverge edge and skip the loop */
2274 pset_foreach(live, irn) {
2276 * if value is defined in this block we can anways place the spill directly after the def
2277 * -> no constraint necessary
2279 if(!is_Phi(irn) && get_nodes_block(irn) == bb) continue;
2282 spill = set_find_spill(spill_bb->ilp, irn);
2285 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2286 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2288 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2289 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2292 sched_foreach_op(bb, tmp) {
2293 op_t *op = get_irn_link(tmp);
2295 if(is_Phi(tmp)) continue;
2296 assert(!is_Proj(tmp));
2299 ir_node *value = op->attr.remat.remat->value;
2302 /* only collect remats up to the first use of a value */
2303 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2308 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2309 ir_node *arg = get_irn_n(tmp, n);
2312 /* if a value is used stop collecting remats */
2318 if(cst == ILP_UNDEF) break;
2326 typedef struct _irnlist_t {
2327 struct list_head list;
2331 typedef struct _interference_t {
2332 struct list_head blocklist;
2338 cmp_interference(const void *a, const void *b, size_t size)
2340 const interference_t *p = a;
2341 const interference_t *q = b;
2343 return !(p->a == q->a && p->b == q->b);
2346 static interference_t *
2347 set_find_interference(set * set, ir_node * a, ir_node * b)
2349 interference_t query;
2351 query.a = (a>b)?a:b;
2352 query.b = (a>b)?b:a;
2354 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2357 static interference_t *
2358 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2360 interference_t query,
2362 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2366 result = set_find_interference(set, a, b);
2369 list_add(&list->list, &result->blocklist);
2373 query.a = (a>b)?a:b;
2374 query.b = (a>b)?b:a;
2376 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2378 INIT_LIST_HEAD(&result->blocklist);
2379 list_add(&list->list, &result->blocklist);
2385 values_interfere_in_block(ir_node * bb, ir_node * a, ir_node * b)
2387 const ir_edge_t *edge;
2389 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2390 /* both values are live in, so they interfere */
2394 /* ensure a dominates b */
2395 if(value_dominates(b,a)) {
2401 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2404 /* the following code is stolen from bera.c */
2405 if(is_live_end(bb, a))
2408 foreach_out_edge(a, edge) {
2409 const ir_node *user = edge->src;
2410 if(get_nodes_block(user) == bb
2413 && value_dominates(b, user))
2421 * Walk all irg blocks and collect interfering values inside of phi classes
2424 luke_interferencewalker(ir_node * bb, void * data)
2426 spill_ilp_t *si = (spill_ilp_t*)data;
2430 live_foreach(bb, li1) {
2431 ir_node *a = (ir_node *) li1->irn;
2432 op_t *a_op = get_irn_link(a);
2434 if(a_op->is_remat) continue;
2436 /* a is only interesting if it is in my register class and if it is inside a phi class */
2437 if (has_reg_class(si, a) && get_phi_class(a)) {
2438 for(li2=li1->next; li2; li2 = li2->next) {
2439 ir_node *b = (ir_node *) li2->irn;
2440 op_t *b_op = get_irn_link(b);
2442 if(b_op->is_remat) continue;
2444 /* a and b are only interesting if they are in the same phi class */
2445 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2446 if(values_interfere_in_block(bb, a, b)) {
2447 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2448 set_insert_interference(si, si->interferences, a, b, bb);
2456 static unsigned int copy_path_id = 0;
2459 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2466 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2467 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2469 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2471 pset_foreach(copies, ptr) {
2472 copy = PTR_TO_INT(ptr);
2473 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2478 * @parameter copies contains a path of copies which lead us to irn
2479 * @parameter visited contains a set of nodes already visited on this path
2482 find_copy_path(spill_ilp_t * si, ir_node * irn, ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2485 op_t *op = get_irn_link(irn);
2487 if(op->is_remat) return;
2489 pset_insert_ptr(visited, irn);
2494 /* visit all operands */
2495 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2496 ir_node *arg = get_irn_n(irn, n);
2497 ilp_var_t copy = op->attr.live_range.args.copies[n];
2499 if(!has_reg_class(si, arg)) continue;
2502 pset_insert(copies, INT_TO_PTR(copy), copy);
2503 write_copy_path_cst(si, copies, any_interfere);
2504 pset_remove(copies, INT_TO_PTR(copy), copy);
2506 if(!pset_find_ptr(visited, arg)) {
2507 pset_insert(copies, INT_TO_PTR(copy), copy);
2508 find_copy_path(si, arg, target, any_interfere, copies, visited);
2509 pset_remove(copies, INT_TO_PTR(copy), copy);
2515 /* visit all uses which are phis */
2516 foreach_out_edge(irn, edge) {
2517 ir_node *user = edge->src;
2518 int pos = edge->pos;
2519 op_t *op = get_irn_link(user);
2522 if(!is_Phi(user)) continue;
2523 if(!has_reg_class(si, user)) continue;
2525 copy = op->attr.live_range.args.copies[pos];
2527 if(user == target) {
2528 pset_insert(copies, INT_TO_PTR(copy), copy);
2529 write_copy_path_cst(si, copies, any_interfere);
2530 pset_remove(copies, INT_TO_PTR(copy), copy);
2532 if(!pset_find_ptr(visited, user)) {
2533 pset_insert(copies, INT_TO_PTR(copy), copy);
2534 find_copy_path(si, user, target, any_interfere, copies, visited);
2535 pset_remove(copies, INT_TO_PTR(copy), copy);
2540 pset_remove_ptr(visited, irn);
2544 gen_copy_constraints(spill_ilp_t * si, ir_node * a, ir_node * b, ilp_var_t any_interfere)
2546 pset * copies = pset_new_ptr_default();
2547 pset * visited = pset_new_ptr_default();
2549 find_copy_path(si, a, b, any_interfere, copies, visited);
2557 memcopyhandler(spill_ilp_t * si)
2559 interference_t *interference;
2561 /* teste Speicherwerte auf Interferenz */
2563 /* analyze phi classes */
2564 phi_class_compute(si->chordal_env->irg);
2566 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
2567 irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
2569 // phi_class_free(si->chordal_env->irg);
2571 /* now lets emit the ILP unequations for the crap */
2572 set_foreach(si->interferences, interference) {
2574 ilp_var_t interfere,
2576 ilp_cst_t any_interfere_cst,
2578 const ir_node *a = interference->a;
2579 const ir_node *b = interference->b;
2581 /* any_interf <= \sum interf */
2582 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
2583 any_interfere_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2584 any_interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2586 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
2588 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
2589 const ir_node *bb = irnlist->irn;
2590 spill_bb_t *spill_bb = get_irn_link(bb);
2597 spilla = set_find_spill(spill_bb->ilp, a);
2601 spillb = set_find_spill(spill_bb->ilp, b);
2604 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
2605 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
2606 /* 2: - mem_in_a - spill_a + interfere <= 0 */
2607 /* 3: - mem_in_b - spill_b + interfere <= 0 */
2608 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
2609 interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2611 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
2612 cst = lpp_add_cst(si->lpp, buf, lpp_less, 1);
2614 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
2615 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
2616 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
2617 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
2618 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
2620 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
2621 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2623 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2624 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
2625 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
2627 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
2628 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2630 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2631 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
2632 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
2635 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
2637 /* any_interfere >= interf */
2638 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
2639 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2641 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2642 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2645 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
2646 gen_copy_constraints(si,a,b,any_interfere);
2654 return fabs(x) < 0.00001;
2658 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
2660 spill_ilp_t *si = get_irg_link(current_ir_graph);
2662 if(pset_find_ptr(si->all_possible_remats, n)) {
2663 op_t *op = (op_t*)get_irn_link(n);
2664 assert(op && op->is_remat);
2666 if(!op->attr.remat.remat->inverse) {
2667 if(op->attr.remat.pre) {
2668 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
2670 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
2675 op_t *op = (op_t*)get_irn_link(n);
2676 assert(op && op->is_remat);
2678 if(op->attr.remat.pre) {
2679 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
2681 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
2692 dump_graph_with_remats(ir_graph * irg, const char * suffix)
2694 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
2695 be_dump(irg, suffix, dump_ir_block_graph_sched);
2696 set_dump_node_vcgattr_hook(NULL);
2701 * Edge hook to dump the schedule edges with annotated register pressure.
2704 sched_pressure_edge_hook(FILE *F, ir_node *irn)
2706 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
2707 ir_node *prev = sched_prev(irn);
2708 fprintf(F, "edge:{sourcename:\"");
2710 fprintf(F, "\" targetname:\"");
2712 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
2713 fprintf(F, "\" color:magenta}\n");
2719 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
2721 DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
2723 dump_consts_local(0);
2724 set_dump_node_edge_hook(sched_pressure_edge_hook);
2725 dump_ir_block_graph(irg, suffix);
2726 set_dump_node_edge_hook(old_edge_hook);
2730 walker_pressure_annotator(ir_node * bb, void * data)
2732 spill_ilp_t *si = data;
2736 pset *live = pset_new_ptr_default();
2739 live_foreach(bb, li) {
2740 irn = (ir_node *) li->irn;
2742 if (live_is_end(li) && has_reg_class(si, irn)) {
2743 pset_insert_ptr(live, irn);
2747 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
2749 sched_foreach_reverse(bb, irn) {
2751 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
2755 if(has_reg_class(si, irn)) {
2756 pset_remove_ptr(live, irn);
2757 if(is_Proj(irn)) ++projs;
2760 if(!is_Proj(irn)) projs = 0;
2762 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2763 ir_node *arg = get_irn_n(irn, n);
2765 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
2767 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
2774 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
2776 be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
2781 connect_all_remats_with_keep(spill_ilp_t * si)
2789 n_remats = pset_count(si->all_possible_remats);
2791 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
2794 pset_foreach(si->all_possible_remats, irn) {
2799 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
2801 obstack_free(si->obst, ins);
2807 connect_all_spills_with_keep(spill_ilp_t * si)
2816 n_spills = pset_count(si->spills);
2818 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
2821 pset_foreach(si->spills, irn) {
2826 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
2828 obstack_free(si->obst, ins);
2832 /** insert a spill at an arbitrary position */
2833 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert, ir_node *ctx)
2835 ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
2836 ir_graph *irg = get_irn_irg(bl);
2837 ir_node *frame = get_irg_frame(irg);
2841 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
2842 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
2844 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn, ctx);
2847 * search the right insertion point. a spill of a phi cannot be put
2848 * directly after the phi, if there are some phis behind the one which
2849 * is spilled. Also, a spill of a Proj must be after all Projs of the
2852 * Here's one special case:
2853 * If the spill is in the start block, the spill must be after the frame
2854 * pointer is set up. This is done by setting insert to the end of the block
2855 * which is its default initialization (see above).
2858 if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
2861 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
2864 sched_add_after(insert, spill);
2869 delete_remat(spill_ilp_t * si, ir_node * remat) {
2871 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2873 sched_remove(remat);
2875 /* kill links to operands */
2876 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
2877 set_irn_n(remat, n, bad);
2882 clean_remat_info(spill_ilp_t * si)
2886 remat_info_t *remat_info;
2887 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2889 set_foreach(si->remat_info, remat_info) {
2890 if(!remat_info->remats) continue;
2892 pset_foreach(remat_info->remats, remat)
2894 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
2895 set_irn_n(remat->proj, -1, bad);
2896 set_irn_n(remat->proj, 0, bad);
2899 if(get_irn_n_edges(remat->op) == 0) {
2900 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
2901 set_irn_n(remat->op, n, bad);
2906 if(remat_info->remats) del_pset(remat_info->remats);
2907 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
2912 delete_unnecessary_remats(spill_ilp_t * si)
2916 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2919 ir_node *end = get_irg_end(si->chordal_env->irg);
2922 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
2923 ir_node *keep_arg = get_irn_n(si->keep, n);
2924 op_t *arg_op = get_irn_link(keep_arg);
2927 assert(arg_op->is_remat);
2929 name = si->lpp->vars[arg_op->attr.remat.ilp];
2931 if(is_zero(name->value)) {
2932 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
2933 /* TODO check whether reload is preferred over remat (could be bug) */
2934 delete_remat(si, keep_arg);
2936 if(!arg_op->attr.remat.remat->inverse) {
2937 if(arg_op->attr.remat.pre) {
2938 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
2940 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
2943 if(arg_op->attr.remat.pre) {
2944 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
2946 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
2951 set_irn_n(si->keep, n, bad);
2954 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
2955 ir_node *end_arg = get_End_keepalive(end, i);
2957 if(end_arg != si->keep) {
2958 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
2961 keeps = obstack_finish(si->obst);
2962 set_End_keepalives(end, n-1, keeps);
2963 obstack_free(si->obst, keeps);
2966 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
2971 pset_foreach(si->all_possible_remats, remat) {
2972 op_t *remat_op = get_irn_link(remat);
2973 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
2975 if(is_zero(name->value)) {
2976 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
2977 /* TODO check whether reload is preferred over remat (could be bug) */
2978 delete_remat(si, remat);
2980 if(!remat_op->attr.remat.remat->inverse) {
2981 if(remat_op->attr.remat.pre) {
2982 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
2984 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
2987 if(remat_op->attr.remat.pre) {
2988 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
2990 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
2999 get_spills_for_value(spill_ilp_t * si, ir_node * value)
3001 pset *spills = pset_new_ptr_default();
3006 defs = set_find_def(si->values, value);
3008 if(defs && defs->spills) {
3009 for(next = defs->spills; next; next = get_irn_link(next)) {
3010 pset_insert_ptr(spills, next);
3018 get_remats_for_value(spill_ilp_t * si, ir_node * value)
3020 pset *remats = pset_new_ptr_default();
3025 pset_insert_ptr(remats, value);
3026 defs = set_find_def(si->values, value);
3028 if(defs && defs->remats) {
3029 for(next = defs->remats; next; next = get_irn_link(next)) {
3030 pset_insert_ptr(remats, next);
3039 * @param before The node after which the spill will be placed in the schedule
3041 /* TODO set context properly */
3043 insert_spill(spill_ilp_t * si, ir_node * irn, ir_node * value, ir_node * before)
3047 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3049 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3051 spill = be_spill2(arch_env, irn, before, irn);
3053 defs = set_insert_def(si->values, value);
3056 /* enter into the linked list */
3057 set_irn_link(spill, defs->spills);
3058 defs->spills = spill;
3060 #ifdef KEEPALIVE_SPILLS
3061 pset_insert_ptr(si->spills, spill);
3068 * @param before The Phi node which has to be spilled
3071 insert_mem_phi(spill_ilp_t * si, const ir_node * phi)
3077 op_t *op = get_irn_link(phi);
3079 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3081 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3082 ins[n] = si->m_unknown;
3085 mem_phi = new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3087 defs = set_insert_def(si->values, phi);
3090 /* enter into the linked list */
3091 set_irn_link(mem_phi, defs->spills);
3092 defs->spills = mem_phi;
3094 sched_add_after(phi, mem_phi);
3096 #ifdef KEEPALIVE_SPILLS
3097 pset_insert_ptr(si->spills, mem_phi);
3105 * Add remat to list of defs, destroys link field!
3108 insert_remat(spill_ilp_t * si, ir_node * remat)
3111 op_t *remat_op = get_irn_link(remat);
3113 assert(remat_op->is_remat);
3115 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3118 /* enter into the linked list */
3119 set_irn_link(remat, defs->remats);
3120 defs->remats = remat;
3125 * Add reload before operation and add to list of defs
3128 insert_reload(spill_ilp_t * si, const ir_node * value, const ir_node * after)
3133 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3135 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3137 defs = set_find_def(si->values, value);
3139 spill = defs->spills;
3140 assert(spill && "no spill placed before reload");
3142 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3144 /* enter into the linked list */
3145 set_irn_link(reload, defs->remats);
3146 defs->remats = reload;
3152 walker_spill_placer(ir_node * bb, void * data) {
3153 spill_ilp_t *si = (spill_ilp_t*)data;
3155 spill_bb_t *spill_bb = get_irn_link(bb);
3156 pset *spills_to_do = pset_new_ptr_default();
3159 set_foreach(spill_bb->ilp, spill) {
3162 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3163 name = si->lpp->vars[spill->mem_in];
3164 if(!is_zero(name->value)) {
3167 mem_phi = insert_mem_phi(si, spill->irn);
3169 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3173 name = si->lpp->vars[spill->spill];
3174 if(!is_zero(name->value)) {
3175 /* place spill directly after definition */
3176 if(get_nodes_block(spill->irn) == bb) {
3177 insert_spill(si, spill->irn, spill->irn, spill->irn);
3181 /* place spill at bb start */
3182 if(spill->reg_in > 0) {
3183 name = si->lpp->vars[spill->reg_in];
3184 if(!is_zero(name->value)) {
3185 insert_spill(si, spill->irn, spill->irn, bb);
3189 /* place spill after a remat */
3190 pset_insert_ptr(spills_to_do, spill->irn);
3193 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3196 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3197 op_t *op = get_irn_link(irn);
3199 if(be_is_Spill(irn)) continue;
3202 /* TODO fix this if we want to support remats with more than two nodes */
3203 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3204 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3206 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3209 if(pset_find_ptr(spills_to_do, irn)) {
3210 pset_remove_ptr(spills_to_do, irn);
3212 insert_spill(si, irn, irn, irn);
3218 assert(pset_count(spills_to_do) == 0);
3220 /* afterwards free data in block */
3221 del_pset(spills_to_do);
3225 insert_mem_copy(spill_ilp_t * si, const ir_node * bb, const ir_node * value)
3227 ir_node *insert_pos = bb;
3229 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3231 /* find last definition of arg value in block */
3236 defs = set_find_def(si->values, value);
3238 if(defs && defs->remats) {
3239 for(next = defs->remats; next; next = get_irn_link(next)) {
3240 if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3241 last = sched_get_time_step(next);
3247 if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3248 last = sched_get_time_step(value);
3252 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
3254 spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos, value);
3260 phim_fixer(spill_ilp_t *si) {
3263 set_foreach(si->values, defs) {
3264 const ir_node *phi = defs->value;
3265 op_t *op = get_irn_link(phi);
3266 ir_node *phi_m = NULL;
3267 ir_node *next = defs->spills;
3270 if(!is_Phi(phi)) continue;
3273 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3277 next = get_irn_link(next);
3280 if(!phi_m) continue;
3282 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3283 const ir_node *value = get_irn_n(phi, n);
3284 defs_t *val_defs = set_find_def(si->values, value);
3285 ir_node *arg = get_irn_n(phi_m, n);
3287 /* get a spill of this value */
3288 ir_node *spill = val_defs->spills;
3291 #ifndef NO_MEMCOPIES
3292 ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3293 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3295 if(!is_zero(name->value)) {
3296 spill = insert_mem_copy(si, pred, value);
3298 assert(spill && "no spill placed before PhiM");
3301 assert(spill && "no spill placed before PhiM");
3303 set_irn_n(phi_m, n, spill);
3309 walker_reload_placer(ir_node * bb, void * data) {
3310 spill_ilp_t *si = (spill_ilp_t*)data;
3312 spill_bb_t *spill_bb = get_irn_link(bb);
3316 /* reloads at end of block */
3317 if(spill_bb->reloads) {
3320 set_foreach(spill_bb->reloads, keyval) {
3321 ir_node *irn = (ir_node*)keyval->key;
3322 ilp_var_t reload = PTR_TO_INT(keyval->val);
3325 name = si->lpp->vars[reload];
3326 if(!is_zero(name->value)) {
3328 ir_node *insert_pos = bb;
3329 ir_node *prev = sched_block_last_noncf(si, bb);
3330 op_t *prev_op = get_irn_link(prev);
3332 while(be_is_Spill(prev)) {
3333 prev = sched_prev(prev);
3336 prev_op = get_irn_link(prev);
3338 /* insert reload before pre-remats */
3339 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3340 && prev_op->is_remat && prev_op->attr.remat.pre) {
3344 prev = sched_prev(prev);
3345 } while(be_is_Spill(prev));
3347 prev_op = get_irn_link(prev);
3351 reload = insert_reload(si, irn, insert_pos);
3353 #ifdef KEEPALIVE_RELOADS
3354 pset_insert_ptr(si->spills, reload);
3360 /* walk and insert more reloads and collect remats */
3361 sched_foreach_reverse(bb, irn) {
3362 op_t *op = get_irn_link(irn);
3364 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3365 if(is_Phi(irn)) break;
3368 if(get_irn_mode(irn) != mode_T) {
3369 insert_remat(si, irn);
3374 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3375 ir_node *arg = get_irn_n(irn, n);
3377 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3380 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3381 if(!is_zero(name->value)) {
3383 ir_node *insert_pos = irn;
3384 ir_node *prev = sched_prev(insert_pos);
3387 while(be_is_Spill(prev)) {
3388 prev = sched_prev(prev);
3391 prev_op = get_irn_link(prev);
3393 /* insert reload before pre-remats */
3394 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3395 && prev_op->is_remat && prev_op->attr.remat.pre) {
3399 prev = sched_prev(prev);
3400 } while(be_is_Spill(prev));
3402 prev_op = get_irn_link(prev);
3406 reload = insert_reload(si, arg, insert_pos);
3408 set_irn_n(irn, n, reload);
3410 #ifdef KEEPALIVE_RELOADS
3411 pset_insert_ptr(si->spills, reload);
3419 del_set(spill_bb->ilp);
3420 if(spill_bb->reloads) del_set(spill_bb->reloads);
3424 walker_collect_used(ir_node * irn, void * data)
3426 lc_bitset_t *used = data;
3428 lc_bitset_set(used, get_irn_idx(irn));
3431 struct kill_helper {
3437 walker_kill_unused(ir_node * bb, void * data)
3439 struct kill_helper *kh = data;
3440 const ir_node *bad = get_irg_bad(get_irn_irg(bb));
3444 for(irn=sched_first(bb); !sched_is_end(irn);) {
3445 ir_node *next = sched_next(irn);
3448 if(!lc_bitset_is_set(kh->used, get_irn_idx(irn))) {
3449 if(be_is_Spill(irn) || be_is_Reload(irn)) {
3450 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)));
3452 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3458 set_nodes_block(irn, bad);
3459 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3460 set_irn_n(irn, n, bad);
3468 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3470 struct kill_helper kh;
3472 kh.used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3475 irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3476 irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3478 lc_bitset_free(kh.used);
3482 print_irn_pset(pset * p)
3486 pset_foreach(p, irn) {
3487 ir_printf("%+F\n", irn);
3492 rewire_uses(spill_ilp_t * si)
3494 dom_front_info_t *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
3496 pset *ignore = pset_new_ptr(1);
3498 pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
3500 /* then fix uses of spills */
3501 set_foreach(si->values, defs) {
3504 ir_node *next = defs->remats;
3507 reloads = pset_new_ptr_default();
3510 if(be_is_Reload(next)) {
3511 pset_insert_ptr(reloads, next);
3515 next = get_irn_link(next);
3518 spills = get_spills_for_value(si, defs->value);
3519 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));
3520 if(pset_count(spills) > 1) {
3521 //assert(pset_count(reloads) > 0);
3522 // print_irn_pset(spills);
3523 // print_irn_pset(reloads);
3525 be_ssa_constr_set_ignore(dfi, spills, ignore);
3532 /* first fix uses of remats and reloads */
3533 set_foreach(si->values, defs) {
3535 ir_node *next = defs->remats;
3538 nodes = pset_new_ptr_default();
3539 pset_insert_ptr(nodes, defs->value);
3542 pset_insert_ptr(nodes, next);
3543 next = get_irn_link(next);
3546 if(pset_count(nodes) > 1) {
3547 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
3548 be_ssa_constr_set(dfi, nodes);
3555 // remove_unused_defs(si);
3557 be_free_dominance_frontiers(dfi);
3562 writeback_results(spill_ilp_t * si)
3564 /* walk through the graph and collect all spills, reloads and remats for a value */
3566 si->values = new_set(cmp_defs, 4096);
3568 DBG((si->dbg, LEVEL_1, "Applying results\n"));
3569 delete_unnecessary_remats(si);
3570 si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
3571 irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
3572 irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
3575 /* clean the remat info! there are still back-edges leading there! */
3576 clean_remat_info(si);
3580 connect_all_spills_with_keep(si);
3582 del_set(si->values);
3586 get_n_regs(spill_ilp_t * si)
3588 int arch_n_regs = arch_register_class_n_regs(si->cls);
3592 for(i=0; i<arch_n_regs; i++) {
3593 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
3598 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
3603 walker_reload_mover(ir_node * bb, void * data)
3605 spill_ilp_t *si = data;
3608 sched_foreach(bb, tmp) {
3609 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
3610 ir_node *reload = tmp;
3613 /* move reload upwards */
3615 int pressure = (int)get_irn_link(reload);
3616 if(pressure < si->n_regs) {
3617 irn = sched_prev(reload);
3618 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
3619 sched_remove(reload);
3620 pressure = (int)get_irn_link(irn);
3622 while(pressure < si->n_regs) {
3623 if(sched_is_end(irn) || (be_is_Reload(irn) && has_reg_class(si, irn))) break;
3625 set_irn_link(irn, INT_TO_PTR(pressure+1));
3626 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
3627 irn = sched_prev(irn);
3629 pressure = (int)get_irn_link(irn);
3632 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
3633 sched_put_after(irn, reload);
3640 move_reloads_upward(spill_ilp_t * si)
3642 irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
3647 * Walk all irg blocks and check for interfering spills inside of phi classes
3650 luke_meminterferencechecker(ir_node * bb, void * data)
3652 spill_ilp_t *si = (spill_ilp_t*)data;
3656 live_foreach(bb, li1) {
3657 ir_node *a = (ir_node *) li1->irn;
3659 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
3661 /* a is only interesting if it is inside a phi class */
3662 if (get_phi_class(a)) {
3663 for(li2=li1->next; li2; li2 = li2->next) {
3664 ir_node *b = (ir_node *) li2->irn;
3666 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
3668 /* a and b are only interesting if they are in the same phi class */
3669 if(get_phi_class(a) == get_phi_class(b)) {
3670 if(values_interfere_in_block(bb, a, b)) {
3671 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
3680 verify_phiclasses(spill_ilp_t * si)
3682 /* analyze phi classes */
3683 phi_class_compute(si->chordal_env->irg);
3685 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
3686 irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
3690 walker_spillslotassigner(ir_node * irn, void * data)
3692 spill_ilp_t *si = (spill_ilp_t*)data;
3695 if(!be_is_Spill(irn)) return;
3697 /* set spill context to phi class if it has one ;) */
3699 cls = get_phi_class(irn);
3701 be_set_Spill_context(irn, cls);
3703 be_set_Spill_context(irn, irn);
3708 assign_spillslots(spill_ilp_t * si)
3710 DBG((si->dbg, LEVEL_2, "\t calling spill slot assigner\n"));
3711 irg_walk_graph(si->chordal_env->irg, walker_spillslotassigner, NULL, si);
3715 be_spill_remat(const be_chordal_env_t * chordal_env)
3717 char problem_name[256];
3718 char dump_suffix[256];
3719 char dump_suffix2[256];
3720 char dump_suffix3[256];
3721 struct obstack obst;
3724 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
3725 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
3726 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
3728 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
3729 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
3731 obstack_init(&obst);
3732 si.chordal_env = chordal_env;
3734 si.cls = chordal_env->cls;
3735 si.lpp = new_lpp(problem_name, lpp_minimize);
3736 si.remat_info = new_set(cmp_remat_info, 4096);
3737 si.interferences = new_set(cmp_interference, 32);
3738 si.all_possible_remats = pset_new_ptr_default();
3739 si.spills = pset_new_ptr_default();
3740 si.inverse_ops = pset_new_ptr_default();
3744 si.n_regs = get_n_regs(&si);
3746 set_irg_link(chordal_env->irg, &si);
3747 compute_doms(chordal_env->irg);
3749 /* compute phi classes */
3750 // phi_class_compute(chordal_env->irg);
3752 be_analyze_regpressure(chordal_env, "-pre");
3754 #ifdef COLLECT_REMATS
3755 /* collect remats */
3756 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
3757 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
3760 /* insert possible remats */
3761 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
3762 irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
3763 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
3766 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
3767 connect_all_remats_with_keep(&si);
3768 /* dump graph with inserted remats */
3769 dump_graph_with_remats(chordal_env->irg, dump_suffix);
3772 /* insert copies for phi arguments not in my regclass */
3773 irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
3775 /* recompute liveness */
3776 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
3777 be_liveness(chordal_env->irg);
3781 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
3782 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
3783 irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
3785 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
3786 irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
3788 #ifndef NO_MEMCOPIES
3789 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
3790 memcopyhandler(&si);
3798 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
3799 if ((f = fopen(buf, "wt")) != NULL) {
3800 lpp_dump_plain(si.lpp, f);
3807 DBG((si.dbg, LEVEL_1, "\tSolving %F\n", chordal_env->irg));
3809 lpp_set_time_limit(si.lpp, ILP_TIMEOUT);
3813 lpp_solve_cplex(si.lpp);
3815 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
3817 assert(lpp_is_sol_valid(si.lpp)
3818 && "solution of ILP must be valid");
3820 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));
3822 #ifdef DUMP_SOLUTION
3827 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
3828 if ((f = fopen(buf, "wt")) != NULL) {
3830 for (i = 0; i < si.lpp->var_next; ++i) {
3831 lpp_name_t *name = si.lpp->vars[i];
3832 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
3839 writeback_results(&si);
3843 kill_all_unused_values_in_schedule(&si);
3845 #if defined(KEEPALIVE_SPILLS) || defined(KEEPALIVE_RELOADS)
3846 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
3849 // move reloads upwards
3850 be_liveness(chordal_env->irg);
3851 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
3852 move_reloads_upward(&si);
3854 #ifndef NO_MEMCOPIES
3855 verify_phiclasses(&si);
3856 assign_spillslots(&si);
3859 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
3861 dump_pressure_graph(&si, dump_suffix2);
3863 be_analyze_regpressure(chordal_env, "-post");
3865 free_dom(chordal_env->irg);
3866 del_set(si.interferences);
3867 del_pset(si.inverse_ops);
3868 del_pset(si.all_possible_remats);
3869 del_pset(si.spills);
3871 obstack_free(&obst, NULL);
3872 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
3875 #else /* WITH_ILP */
3878 only_that_you_can_compile_without_WITH_ILP_defined(void)
3882 #endif /* WITH_ILP */