1 /** vim: set sw=4 ts=4:
4 * @author Adam M. Szalkowski & Sebastian Hack
6 * ILP based spilling & rematerialization
8 * Copyright (C) 2006 Universitaet Karlsruhe
9 * Released under the GPL
32 #include "phiclass_t.h"
38 #include <lpp/lpp_net.h>
39 #include <lpp/lpp_cplex.h>
40 //#include <lc_pset.h>
41 #include <libcore/lc_bitset.h>
45 #include "besched_t.h"
50 #include "bespillremat.h"
52 #include "bepressurestat.h"
54 #include "bechordal_t.h"
60 //#define KEEPALIVE /* keep alive all inserted remats and dump graph with remats */
61 #define COLLECT_REMATS /* enable rematerialization */
62 #define COLLECT_INVERSE_REMATS /* enable placement of inverse remats */
63 #define REMAT_WHILE_LIVE /* only remat values that are live */
64 //#define NO_ENLARGE_L1V3N355 /* do not remat after the death of some operand */
65 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
66 #define MAY_DIE_AT_REMAT /* allow values to die after a pre remat */
67 #define NO_SINGLE_USE_REMATS /* do not repair schedule */
68 //#define KEEPALIVE_SPILLS
69 //#define KEEPALIVE_RELOADS
70 #define GOODWIN_REDUCTION
71 //#define NO_MEMCOPIES
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 */
102 DEBUG_ONLY(firm_dbg_module_t * dbg);
105 typedef int ilp_var_t;
106 typedef int ilp_cst_t;
108 typedef struct _spill_bb_t {
113 typedef struct _remat_t {
114 const ir_node *op; /**< for copy_irn */
115 const ir_node *value; /**< the value which is being recomputed by this remat */
116 ir_node *proj; /**< not NULL if the above op produces a tuple */
117 int cost; /**< cost of this remat */
118 int inverse; /**< nonzero if this is an inverse remat */
122 * Data to be attached to each IR node. For remats this contains the ilp_var
123 * for this remat and for normal ops this contains the ilp_vars for
124 * reloading each operand
126 typedef struct _op_t {
131 remat_t *remat; /** the remat this op belongs to */
132 int pre; /** 1, if this is a pressure-increasing remat */
136 ir_node *op; /** the operation this live range belongs to */
145 typedef struct _defs_t {
147 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
148 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
151 typedef struct _remat_info_t {
152 const ir_node *irn; /**< the irn to which these remats belong */
153 pset *remats; /**< possible remats for this value */
154 pset *remats_by_operand; /**< remats with this value as operand */
157 typedef struct _keyval_t {
162 typedef struct _spill_t {
172 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
174 return chordal_has_class(si->chordal_env, irn);
179 cmp_remat(const void *a, const void *b)
181 const keyval_t *p = a;
182 const keyval_t *q = b;
183 const remat_t *r = p->val;
184 const remat_t *s = q->val;
188 return !(r == s || r->op == s->op);
192 cmp_remat(const void *a, const void *b)
194 const remat_t *r = a;
195 const remat_t *s = a;
197 return !(r == s || r->op == s->op);
201 cmp_spill(const void *a, const void *b, size_t size)
203 const spill_t *p = a;
204 const spill_t *q = b;
206 // return !(p->irn == q->irn && p->bb == q->bb);
207 return !(p->irn == q->irn);
211 set_find_keyval(set * set, void * key)
216 return set_find(set, &query, sizeof(query), HASH_PTR(key));
220 set_insert_keyval(set * set, void * key, void * val)
226 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
230 set_find_def(set * set, ir_node * value)
235 return set_find(set, &query, sizeof(query), HASH_PTR(value));
239 set_insert_def(set * set, ir_node * value)
246 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
250 set_find_spill(set * set, ir_node * value)
255 return set_find(set, &query, sizeof(query), HASH_PTR(value));
258 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
259 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
260 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
261 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
262 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
265 cmp_remat_info(const void *a, const void *b, size_t size)
267 const remat_info_t *p = a;
268 const remat_info_t *q = b;
270 return !(p->irn == q->irn);
274 cmp_defs(const void *a, const void *b, size_t size)
279 return !(p->value == q->value);
283 cmp_keyval(const void *a, const void *b, size_t size)
285 const keyval_t *p = a;
286 const keyval_t *q = b;
288 return !(p->key == q->key);
292 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
295 #ifndef EXECFREQ_LOOPDEPH
296 return get_block_execfreq(si->chordal_env->exec_freq, get_block(irn)) + FUDGE;
299 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
301 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
306 get_cost(const spill_ilp_t * si, const ir_node * irn)
308 if(be_is_Spill(irn)) {
310 } else if(be_is_Reload(irn)){
313 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(used, 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 // value_op->attr.live_range.ilp != ILP_UNDEF
1733 if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
1734 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1738 /* ensure at least one value dies at post remat */
1739 foreach_post_remat(irn, tmp) {
1740 op_t *remat_op = get_irn_link(tmp);
1741 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
1744 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
1745 remat_arg = get_irn_n(tmp, n);
1747 if(has_reg_class(si, remat_arg)) {
1749 /* does arg always die at this op? */
1750 if(!pset_find_ptr(live, remat_arg))
1751 goto skip_one_must_die;
1753 pset_insert_ptr(remat_args, remat_arg);
1757 /* remat + \sum live_range(remat_arg) <= |args| */
1758 ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
1759 cst = lpp_add_cst(si->lpp, buf, lpp_less, pset_count(remat_args));
1760 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1762 pset_foreach(remat_args, remat_arg) {
1763 op_t *arg_op = get_irn_link(remat_arg);
1765 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1769 del_pset(remat_args);
1772 /* new live ranges for values from L\U defined by post remats */
1773 pset_foreach(live, tmp) {
1774 ir_node *value = tmp;
1775 op_t *value_op = get_irn_link(value);
1777 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
1778 ilp_var_t prev_lr = ILP_UNDEF;
1781 if(pset_find_ptr(remat_defs, value)) {
1783 /* next_live_range <= prev_live_range + sum remat2s */
1784 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1785 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1787 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1788 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1790 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1791 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1793 foreach_post_remat(irn, remat) {
1794 op_t *remat_op = get_irn_link(remat);
1796 /* if value is being rematerialized by this remat */
1797 if(value == remat_op->attr.remat.remat->value) {
1798 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1802 value_op->attr.live_range.ilp = prev_lr;
1803 value_op->attr.live_range.op = irn;
1808 /* requirements for post remats and start live ranges from L/U' for values dying here */
1809 foreach_post_remat(irn, tmp) {
1810 op_t *remat_op = get_irn_link(tmp);
1813 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1814 ir_node *remat_arg = get_irn_n(tmp, n);
1815 op_t *arg_op = get_irn_link(remat_arg);
1817 if(!has_reg_class(si, remat_arg)) continue;
1819 /* only for values in L\U (TODO and D?), the others are handled with post_use */
1820 if(!pset_find_ptr(used, remat_arg)) {
1821 /* remat <= live_range(remat_arg) */
1822 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
1823 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1825 /* if value is becoming live through use by remat2 */
1826 if(!pset_find_ptr(live, remat_arg)) {
1829 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1830 lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1832 arg_op->attr.live_range.ilp = lr;
1833 arg_op->attr.live_range.op = irn;
1835 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
1837 pset_insert_ptr(live, remat_arg);
1838 add_to_spill_bb(si, bb, remat_arg);
1841 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1842 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1847 d = pset_count(defs);
1848 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
1850 /* count how many regs irn needs for arguments */
1851 u = set_count(args);
1854 /* check the register pressure in the epilog */
1855 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
1856 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
1857 check_post = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - d);
1859 /* add L\U' to check_post */
1860 pset_foreach(live, tmp) {
1861 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
1862 /* if a live value is not used by irn */
1863 tmp_op = get_irn_link(tmp);
1864 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
1868 /***********************************************************
1869 * 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
1870 **********************************************************/
1873 pset_foreach(used, tmp) {
1879 op_t *arg_op = get_irn_link(arg);
1882 spill = add_to_spill_bb(si, bb, arg);
1884 /* new live range for each used value */
1885 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
1886 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1888 /* the epilog stuff - including post_use, check_post, check_post_remat */
1889 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
1890 post_use = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1892 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
1894 /* arg is live throughout epilog if the next live_range is in a register */
1895 if(pset_find_ptr(live, arg)) {
1896 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
1898 /* post_use >= next_lr + remat */
1899 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1900 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1901 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
1902 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1906 /* if value is not an arg of op and not possibly defined by post remat
1907 * then it may only die and not become live
1909 if(!set_find_keyval(args, arg)) {
1910 /* post_use <= prev_lr */
1911 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
1912 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1913 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1914 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1916 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
1917 /* next_lr <= prev_lr */
1918 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
1919 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1920 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1921 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1927 /* forall post remat which use arg add a similar cst */
1928 foreach_post_remat(irn, remat) {
1931 for (n=get_irn_arity(remat)-1; n>=0; --n) {
1932 ir_node *remat_arg = get_irn_n(remat, n);
1933 op_t *remat_op = get_irn_link(remat);
1935 if(remat_arg == arg) {
1936 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
1938 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1939 cst = lpp_add_cst(si->lpp, buf, lpp_greater, 0.0);
1940 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1941 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1946 /* new live range begins for each used value */
1947 arg_op->attr.live_range.ilp = prev_lr;
1948 arg_op->attr.live_range.op = irn;
1950 /*if(!pset_find_ptr(live, arg)) {
1951 pset_insert_ptr(live, arg);
1952 add_to_spill_bb(si, bb, arg);
1954 pset_insert_ptr(live, arg);
1958 /* just to be sure */
1959 check_post = ILP_UNDEF;
1968 /* check the register pressure in the prolog */
1969 /* sum_{L\U} lr <= k - |U| */
1970 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
1971 check_pre = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - u);
1973 /* for the prolog remove defined values from the live set */
1974 pset_foreach(defs, tmp) {
1975 pset_remove_ptr(live, tmp);
1978 /***********************************************************
1979 * 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
1980 **********************************************************/
1983 set_foreach(args, keyval) {
1985 ir_node *arg = keyval->key;
1986 int i = PTR_TO_INT(keyval->val);
1987 op_t *arg_op = get_irn_link(arg);
1989 spill = set_find_spill(spill_bb->ilp, arg);
1992 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
1993 op->attr.live_range.args.reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1995 /* reload <= mem_out */
1996 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
1997 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1998 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
1999 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2001 /* requirement: arg must be in register for use */
2002 /* reload + remat + live_range == 1 */
2003 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2004 cst = lpp_add_cst(si->lpp, buf, lpp_equal, 1.0);
2006 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2007 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2008 foreach_pre_remat(si, irn, tmp) {
2009 op_t *remat_op = get_irn_link(tmp);
2010 if(remat_op->attr.remat.remat->value == arg) {
2011 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2016 /* iterate over L\U */
2017 pset_foreach(live, tmp) {
2018 if(!set_find_keyval(args, tmp)) {
2019 /* if a live value is not used by irn */
2020 tmp_op = get_irn_link(tmp);
2021 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2026 /* requirements for remats */
2027 /* start new live ranges for values used by remats */
2028 foreach_pre_remat(si, irn, tmp) {
2029 op_t *remat_op = get_irn_link(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 *arg_op = get_irn_link(remat_arg);
2037 if(!has_reg_class(si, remat_arg)) continue;
2039 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2040 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2041 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2043 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2044 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2046 /* if remat arg is also used by current op then we can use reload placed for this argument */
2047 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2048 int index = (int)keyval->val;
2050 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2058 /*************************
2059 * D O N E W I T H O P
2060 *************************/
2062 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2064 pset_foreach(live, tmp) {
2065 assert(has_reg_class(si, tmp));
2068 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2069 ir_node *arg = get_irn_n(irn, n);
2071 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2074 del_pset(remat_defs);
2078 defs = pset_new_ptr_default();
2083 /***************************************
2084 * B E G I N N I N G O F B L O C K
2085 ***************************************/
2088 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2089 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2091 pset_foreach(live, irn) {
2092 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2095 /* construct mem_outs for all values */
2097 set_foreach(spill_bb->ilp, spill) {
2098 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2099 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2101 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2102 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2104 if(pset_find_ptr(live, spill->irn)) {
2105 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2107 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2108 spill->mem_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2109 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2111 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2113 op_t *op = get_irn_link(spill->irn);
2115 /* do we have to copy a phi argument? */
2116 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2117 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2119 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2120 const ir_node *arg = get_irn_n(spill->irn, n);
2126 /* argument already done? */
2127 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2129 /* get sum of execution frequencies of blocks with the same phi argument */
2130 for(m=n; m>=0; --m) {
2131 const ir_node *arg2 = get_irn_n(spill->irn, m);
2134 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2138 /* copies are not for free */
2139 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2140 var = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE * freq);
2142 for(m=n; m>=0; --m) {
2143 const ir_node *arg2 = get_irn_n(spill->irn, m);
2146 op->attr.live_range.args.copies[m] = var;
2150 /* copy <= mem_in */
2151 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2152 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2153 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2154 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2161 /* L\U is empty at bb start */
2162 /* arg is live throughout epilog if it is reg_in into this block */
2164 /* check the register pressure at the beginning of the block
2167 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2168 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
2170 pset_foreach(live, irn) {
2173 spill = set_find_spill(spill_bb->ilp, irn);
2176 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2177 spill->reg_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2179 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2181 /* spill + mem_in <= 1 */
2182 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2183 nospill = lpp_add_cst(si->lpp, buf, lpp_less, 1);
2185 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2186 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2189 foreach_post_remat(bb, irn) {
2190 op_t *remat_op = get_irn_link(irn);
2192 DBG((si->dbg, LEVEL_4, "\t next post remat: %+F\n", irn));
2193 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
2195 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2198 /* forall post remats add requirements */
2199 foreach_post_remat(bb, tmp) {
2202 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2203 ir_node *remat_arg = get_irn_n(tmp, n);
2204 op_t *remat_op = get_irn_link(tmp);
2206 if(!has_reg_class(si, remat_arg)) continue;
2208 spill = set_find_spill(spill_bb->ilp, remat_arg);
2211 /* remat <= reg_in_argument */
2212 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2213 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2214 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2215 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2219 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2220 pset_foreach(live, irn) {
2224 spill = set_find_spill(spill_bb->ilp, irn);
2225 assert(spill && spill->irn == irn);
2227 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2228 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2231 ir_node *phi_arg = get_Phi_pred(irn, n);
2232 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2233 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2236 /* although the phi is in the right regclass one or more of
2237 * its arguments can be in a different one or at least to
2240 if(has_reg_class(si, phi_arg)) {
2241 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2242 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2243 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2244 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2246 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2247 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2249 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2252 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2253 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2257 /* else assure the value arrives on all paths in the same resource */
2259 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2262 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2263 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2266 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2267 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2268 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2269 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2271 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2272 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2274 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2277 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2278 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2283 /* first live ranges from reg_ins */
2284 pset_foreach(live, irn) {
2285 op_t *op = get_irn_link(irn);
2287 spill = set_find_spill(spill_bb->ilp, irn);
2288 assert(spill && spill->irn == irn);
2290 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2291 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2292 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2293 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2295 foreach_post_remat(bb, tmp) {
2296 op_t *remat_op = get_irn_link(tmp);
2298 if(remat_op->attr.remat.remat->value == irn) {
2299 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2304 /* walk forward now and compute constraints for placing spills */
2305 /* this must only be done for values that are not defined in this block */
2306 /* TODO are these values at start of block? if yes, just check whether this is a diverge edge and skip the loop */
2307 pset_foreach(live, irn) {
2309 * if value is defined in this block we can anways place the spill directly after the def
2310 * -> no constraint necessary
2312 if(!is_Phi(irn) && get_nodes_block(irn) == bb) continue;
2315 spill = set_find_spill(spill_bb->ilp, irn);
2318 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2319 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2321 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2322 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2325 sched_foreach_op(bb, tmp) {
2326 op_t *op = get_irn_link(tmp);
2328 if(is_Phi(tmp)) continue;
2329 assert(!is_Proj(tmp));
2332 ir_node *value = op->attr.remat.remat->value;
2335 /* only collect remats up to the first use of a value */
2336 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2341 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2342 ir_node *arg = get_irn_n(tmp, n);
2345 /* if a value is used stop collecting remats */
2351 if(cst == ILP_UNDEF) break;
2359 typedef struct _irnlist_t {
2360 struct list_head list;
2364 typedef struct _interference_t {
2365 struct list_head blocklist;
2371 cmp_interference(const void *a, const void *b, size_t size)
2373 const interference_t *p = a;
2374 const interference_t *q = b;
2376 return !(p->a == q->a && p->b == q->b);
2379 static interference_t *
2380 set_find_interference(set * set, ir_node * a, ir_node * b)
2382 interference_t query;
2384 query.a = (a>b)?a:b;
2385 query.b = (a>b)?b:a;
2387 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2390 static interference_t *
2391 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2393 interference_t query,
2395 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2399 result = set_find_interference(set, a, b);
2402 list_add(&list->list, &result->blocklist);
2406 query.a = (a>b)?a:b;
2407 query.b = (a>b)?b:a;
2409 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2411 INIT_LIST_HEAD(&result->blocklist);
2412 list_add(&list->list, &result->blocklist);
2418 values_interfere_in_block(ir_node * bb, ir_node * a, ir_node * b)
2420 const ir_edge_t *edge;
2422 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2423 /* both values are live in, so they interfere */
2427 /* ensure a dominates b */
2428 if(value_dominates(b,a)) {
2434 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2437 /* the following code is stolen from bera.c */
2438 if(is_live_end(bb, a))
2441 foreach_out_edge(a, edge) {
2442 const ir_node *user = edge->src;
2443 if(get_nodes_block(user) == bb
2446 && value_dominates(b, user))
2454 * Walk all irg blocks and collect interfering values inside of phi classes
2457 luke_interferencewalker(ir_node * bb, void * data)
2459 spill_ilp_t *si = (spill_ilp_t*)data;
2463 live_foreach(bb, li1) {
2464 ir_node *a = (ir_node *) li1->irn;
2465 op_t *a_op = get_irn_link(a);
2467 if(a_op->is_remat) continue;
2469 /* a is only interesting if it is in my register class and if it is inside a phi class */
2470 if (has_reg_class(si, a) && get_phi_class(a)) {
2471 for(li2=li1->next; li2; li2 = li2->next) {
2472 ir_node *b = (ir_node *) li2->irn;
2473 op_t *b_op = get_irn_link(b);
2475 if(b_op->is_remat) continue;
2477 /* a and b are only interesting if they are in the same phi class */
2478 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2479 if(values_interfere_in_block(bb, a, b)) {
2480 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2481 set_insert_interference(si, si->interferences, a, b, bb);
2489 static unsigned int copy_path_id = 0;
2492 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2499 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2500 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2502 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2504 pset_foreach(copies, ptr) {
2505 copy = PTR_TO_INT(ptr);
2506 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2511 * @parameter copies contains a path of copies which lead us to irn
2512 * @parameter visited contains a set of nodes already visited on this path
2515 find_copy_path(spill_ilp_t * si, ir_node * irn, ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2518 op_t *op = get_irn_link(irn);
2520 if(op->is_remat) return;
2522 pset_insert_ptr(visited, irn);
2527 /* visit all operands */
2528 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2529 ir_node *arg = get_irn_n(irn, n);
2530 ilp_var_t copy = op->attr.live_range.args.copies[n];
2532 if(!has_reg_class(si, arg)) continue;
2535 pset_insert(copies, INT_TO_PTR(copy), copy);
2536 write_copy_path_cst(si, copies, any_interfere);
2537 pset_remove(copies, INT_TO_PTR(copy), copy);
2539 if(!pset_find_ptr(visited, arg)) {
2540 pset_insert(copies, INT_TO_PTR(copy), copy);
2541 find_copy_path(si, arg, target, any_interfere, copies, visited);
2542 pset_remove(copies, INT_TO_PTR(copy), copy);
2548 /* visit all uses which are phis */
2549 foreach_out_edge(irn, edge) {
2550 ir_node *user = edge->src;
2551 int pos = edge->pos;
2552 op_t *op = get_irn_link(user);
2555 if(!is_Phi(user)) continue;
2556 if(!has_reg_class(si, user)) continue;
2558 copy = op->attr.live_range.args.copies[pos];
2560 if(user == target) {
2561 pset_insert(copies, INT_TO_PTR(copy), copy);
2562 write_copy_path_cst(si, copies, any_interfere);
2563 pset_remove(copies, INT_TO_PTR(copy), copy);
2565 if(!pset_find_ptr(visited, user)) {
2566 pset_insert(copies, INT_TO_PTR(copy), copy);
2567 find_copy_path(si, user, target, any_interfere, copies, visited);
2568 pset_remove(copies, INT_TO_PTR(copy), copy);
2573 pset_remove_ptr(visited, irn);
2577 gen_copy_constraints(spill_ilp_t * si, ir_node * a, ir_node * b, ilp_var_t any_interfere)
2579 pset * copies = pset_new_ptr_default();
2580 pset * visited = pset_new_ptr_default();
2582 find_copy_path(si, a, b, any_interfere, copies, visited);
2590 memcopyhandler(spill_ilp_t * si)
2592 interference_t *interference;
2594 /* teste Speicherwerte auf Interferenz */
2596 /* analyze phi classes */
2597 phi_class_compute(si->chordal_env->irg);
2599 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
2600 irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
2602 // phi_class_free(si->chordal_env->irg);
2604 /* now lets emit the ILP unequations for the crap */
2605 set_foreach(si->interferences, interference) {
2607 ilp_var_t interfere,
2609 ilp_cst_t any_interfere_cst,
2611 const ir_node *a = interference->a;
2612 const ir_node *b = interference->b;
2614 /* any_interf <= \sum interf */
2615 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
2616 any_interfere_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2617 any_interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2619 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
2621 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
2622 const ir_node *bb = irnlist->irn;
2623 spill_bb_t *spill_bb = get_irn_link(bb);
2630 spilla = set_find_spill(spill_bb->ilp, a);
2634 spillb = set_find_spill(spill_bb->ilp, b);
2637 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
2638 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
2639 /* 2: - mem_in_a - spill_a + interfere <= 0 */
2640 /* 3: - mem_in_b - spill_b + interfere <= 0 */
2641 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
2642 interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2644 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
2645 cst = lpp_add_cst(si->lpp, buf, lpp_less, 1);
2647 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
2648 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
2649 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
2650 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
2651 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
2653 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
2654 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2656 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2657 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
2658 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
2660 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
2661 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2663 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2664 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
2665 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
2668 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
2670 /* any_interfere >= interf */
2671 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
2672 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2674 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2675 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2678 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
2679 gen_copy_constraints(si,a,b,any_interfere);
2687 return fabs(x) < 0.00001;
2691 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
2693 spill_ilp_t *si = get_irg_link(current_ir_graph);
2695 if(pset_find_ptr(si->all_possible_remats, n)) {
2696 op_t *op = (op_t*)get_irn_link(n);
2697 assert(op && op->is_remat);
2699 if(!op->attr.remat.remat->inverse) {
2700 if(op->attr.remat.pre) {
2701 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
2703 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
2708 op_t *op = (op_t*)get_irn_link(n);
2709 assert(op && op->is_remat);
2711 if(op->attr.remat.pre) {
2712 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
2714 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
2725 dump_graph_with_remats(ir_graph * irg, const char * suffix)
2727 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
2728 be_dump(irg, suffix, dump_ir_block_graph_sched);
2729 set_dump_node_vcgattr_hook(NULL);
2734 * Edge hook to dump the schedule edges with annotated register pressure.
2737 sched_pressure_edge_hook(FILE *F, ir_node *irn)
2739 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
2740 ir_node *prev = sched_prev(irn);
2741 fprintf(F, "edge:{sourcename:\"");
2743 fprintf(F, "\" targetname:\"");
2745 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
2746 fprintf(F, "\" color:magenta}\n");
2752 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
2754 DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
2756 dump_consts_local(0);
2757 set_dump_node_edge_hook(sched_pressure_edge_hook);
2758 dump_ir_block_graph(irg, suffix);
2759 set_dump_node_edge_hook(old_edge_hook);
2763 walker_pressure_annotator(ir_node * bb, void * data)
2765 spill_ilp_t *si = data;
2769 pset *live = pset_new_ptr_default();
2772 live_foreach(bb, li) {
2773 irn = (ir_node *) li->irn;
2775 if (live_is_end(li) && has_reg_class(si, irn)) {
2776 pset_insert_ptr(live, irn);
2780 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
2782 sched_foreach_reverse(bb, irn) {
2784 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
2788 if(has_reg_class(si, irn)) {
2789 pset_remove_ptr(live, irn);
2790 if(is_Proj(irn)) ++projs;
2793 if(!is_Proj(irn)) projs = 0;
2795 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2796 ir_node *arg = get_irn_n(irn, n);
2798 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
2800 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
2807 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
2809 be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
2814 connect_all_remats_with_keep(spill_ilp_t * si)
2822 n_remats = pset_count(si->all_possible_remats);
2824 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
2827 pset_foreach(si->all_possible_remats, irn) {
2832 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
2834 obstack_free(si->obst, ins);
2840 connect_all_spills_with_keep(spill_ilp_t * si)
2849 n_spills = pset_count(si->spills);
2851 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
2854 pset_foreach(si->spills, irn) {
2859 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
2861 obstack_free(si->obst, ins);
2865 /** insert a spill at an arbitrary position */
2866 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert, ir_node *ctx)
2868 ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
2869 ir_graph *irg = get_irn_irg(bl);
2870 ir_node *frame = get_irg_frame(irg);
2874 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
2875 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
2877 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn, ctx);
2880 * search the right insertion point. a spill of a phi cannot be put
2881 * directly after the phi, if there are some phis behind the one which
2882 * is spilled. Also, a spill of a Proj must be after all Projs of the
2885 * Here's one special case:
2886 * If the spill is in the start block, the spill must be after the frame
2887 * pointer is set up. This is done by setting insert to the end of the block
2888 * which is its default initialization (see above).
2891 if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
2894 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
2897 sched_add_after(insert, spill);
2902 delete_remat(spill_ilp_t * si, ir_node * remat) {
2904 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2906 sched_remove(remat);
2908 /* kill links to operands */
2909 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
2910 set_irn_n(remat, n, bad);
2915 clean_remat_info(spill_ilp_t * si)
2919 remat_info_t *remat_info;
2920 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2922 set_foreach(si->remat_info, remat_info) {
2923 if(!remat_info->remats) continue;
2925 pset_foreach(remat_info->remats, remat)
2927 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
2928 set_irn_n(remat->proj, -1, bad);
2929 set_irn_n(remat->proj, 0, bad);
2932 if(get_irn_n_edges(remat->op) == 0) {
2933 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
2934 set_irn_n(remat->op, n, bad);
2939 if(remat_info->remats) del_pset(remat_info->remats);
2940 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
2945 delete_unnecessary_remats(spill_ilp_t * si)
2949 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2952 ir_node *end = get_irg_end(si->chordal_env->irg);
2955 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
2956 ir_node *keep_arg = get_irn_n(si->keep, n);
2957 op_t *arg_op = get_irn_link(keep_arg);
2960 assert(arg_op->is_remat);
2962 name = si->lpp->vars[arg_op->attr.remat.ilp];
2964 if(is_zero(name->value)) {
2965 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
2966 /* TODO check whether reload is preferred over remat (could be bug) */
2967 delete_remat(si, keep_arg);
2969 if(!arg_op->attr.remat.remat->inverse) {
2970 if(arg_op->attr.remat.pre) {
2971 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
2973 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
2976 if(arg_op->attr.remat.pre) {
2977 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
2979 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
2984 set_irn_n(si->keep, n, bad);
2987 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
2988 ir_node *end_arg = get_End_keepalive(end, i);
2990 if(end_arg != si->keep) {
2991 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
2994 keeps = obstack_finish(si->obst);
2995 set_End_keepalives(end, n-1, keeps);
2996 obstack_free(si->obst, keeps);
2999 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
3004 pset_foreach(si->all_possible_remats, remat) {
3005 op_t *remat_op = get_irn_link(remat);
3006 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
3008 if(is_zero(name->value)) {
3009 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
3010 /* TODO check whether reload is preferred over remat (could be bug) */
3011 delete_remat(si, remat);
3013 if(!remat_op->attr.remat.remat->inverse) {
3014 if(remat_op->attr.remat.pre) {
3015 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3017 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3020 if(remat_op->attr.remat.pre) {
3021 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3023 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3032 get_spills_for_value(spill_ilp_t * si, ir_node * value)
3034 pset *spills = pset_new_ptr_default();
3039 defs = set_find_def(si->values, value);
3041 if(defs && defs->spills) {
3042 for(next = defs->spills; next; next = get_irn_link(next)) {
3043 pset_insert_ptr(spills, next);
3051 get_remats_for_value(spill_ilp_t * si, ir_node * value)
3053 pset *remats = pset_new_ptr_default();
3058 pset_insert_ptr(remats, value);
3059 defs = set_find_def(si->values, value);
3061 if(defs && defs->remats) {
3062 for(next = defs->remats; next; next = get_irn_link(next)) {
3063 pset_insert_ptr(remats, next);
3072 * @param before The node after which the spill will be placed in the schedule
3074 /* TODO set context properly */
3076 insert_spill(spill_ilp_t * si, ir_node * irn, ir_node * value, ir_node * before)
3080 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3082 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3084 spill = be_spill2(arch_env, irn, before, irn);
3086 defs = set_insert_def(si->values, value);
3089 /* enter into the linked list */
3090 set_irn_link(spill, defs->spills);
3091 defs->spills = spill;
3093 #ifdef KEEPALIVE_SPILLS
3094 pset_insert_ptr(si->spills, spill);
3101 * @param before The Phi node which has to be spilled
3104 insert_mem_phi(spill_ilp_t * si, const ir_node * phi)
3110 op_t *op = get_irn_link(phi);
3112 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3114 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3115 ins[n] = si->m_unknown;
3118 mem_phi = new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3120 defs = set_insert_def(si->values, phi);
3123 /* enter into the linked list */
3124 set_irn_link(mem_phi, defs->spills);
3125 defs->spills = mem_phi;
3127 sched_add_after(phi, mem_phi);
3129 #ifdef KEEPALIVE_SPILLS
3130 pset_insert_ptr(si->spills, mem_phi);
3138 * Add remat to list of defs, destroys link field!
3141 insert_remat(spill_ilp_t * si, ir_node * remat)
3144 op_t *remat_op = get_irn_link(remat);
3146 assert(remat_op->is_remat);
3148 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3151 /* enter into the linked list */
3152 set_irn_link(remat, defs->remats);
3153 defs->remats = remat;
3158 * Add reload before operation and add to list of defs
3161 insert_reload(spill_ilp_t * si, const ir_node * value, const ir_node * after)
3166 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3168 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3170 defs = set_find_def(si->values, value);
3172 spill = defs->spills;
3173 assert(spill && "no spill placed before reload");
3175 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3177 /* enter into the linked list */
3178 set_irn_link(reload, defs->remats);
3179 defs->remats = reload;
3185 walker_spill_placer(ir_node * bb, void * data) {
3186 spill_ilp_t *si = (spill_ilp_t*)data;
3188 spill_bb_t *spill_bb = get_irn_link(bb);
3189 pset *spills_to_do = pset_new_ptr_default();
3192 set_foreach(spill_bb->ilp, spill) {
3195 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3196 name = si->lpp->vars[spill->mem_in];
3197 if(!is_zero(name->value)) {
3200 mem_phi = insert_mem_phi(si, spill->irn);
3202 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3206 name = si->lpp->vars[spill->spill];
3207 if(!is_zero(name->value)) {
3208 /* place spill directly after definition */
3209 if(get_nodes_block(spill->irn) == bb) {
3210 insert_spill(si, spill->irn, spill->irn, spill->irn);
3214 /* place spill at bb start */
3215 if(spill->reg_in > 0) {
3216 name = si->lpp->vars[spill->reg_in];
3217 if(!is_zero(name->value)) {
3218 insert_spill(si, spill->irn, spill->irn, bb);
3222 /* place spill after a remat */
3223 pset_insert_ptr(spills_to_do, spill->irn);
3226 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3229 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3230 op_t *op = get_irn_link(irn);
3232 if(be_is_Spill(irn)) continue;
3235 /* TODO fix this if we want to support remats with more than two nodes */
3236 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3237 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3239 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3242 if(pset_find_ptr(spills_to_do, irn)) {
3243 pset_remove_ptr(spills_to_do, irn);
3245 insert_spill(si, irn, irn, irn);
3251 assert(pset_count(spills_to_do) == 0);
3253 /* afterwards free data in block */
3254 del_pset(spills_to_do);
3258 insert_mem_copy(spill_ilp_t * si, const ir_node * bb, const ir_node * value)
3260 ir_node *insert_pos = bb;
3262 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3264 /* find last definition of arg value in block */
3269 defs = set_find_def(si->values, value);
3271 if(defs && defs->remats) {
3272 for(next = defs->remats; next; next = get_irn_link(next)) {
3273 if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3274 last = sched_get_time_step(next);
3280 if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3281 last = sched_get_time_step(value);
3285 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
3287 spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos, value);
3293 phim_fixer(spill_ilp_t *si) {
3296 set_foreach(si->values, defs) {
3297 const ir_node *phi = defs->value;
3298 op_t *op = get_irn_link(phi);
3299 ir_node *phi_m = NULL;
3300 ir_node *next = defs->spills;
3303 if(!is_Phi(phi)) continue;
3306 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3310 next = get_irn_link(next);
3313 if(!phi_m) continue;
3315 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3316 const ir_node *value = get_irn_n(phi, n);
3317 defs_t *val_defs = set_find_def(si->values, value);
3318 ir_node *arg = get_irn_n(phi_m, n);
3320 /* get a spill of this value */
3321 ir_node *spill = val_defs->spills;
3324 #ifndef NO_MEMCOPIES
3325 ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3326 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3328 if(!is_zero(name->value)) {
3329 spill = insert_mem_copy(si, pred, value);
3331 assert(spill && "no spill placed before PhiM");
3334 assert(spill && "no spill placed before PhiM");
3336 set_irn_n(phi_m, n, spill);
3342 walker_reload_placer(ir_node * bb, void * data) {
3343 spill_ilp_t *si = (spill_ilp_t*)data;
3345 spill_bb_t *spill_bb = get_irn_link(bb);
3349 /* reloads at end of block */
3350 if(spill_bb->reloads) {
3353 set_foreach(spill_bb->reloads, keyval) {
3354 ir_node *irn = (ir_node*)keyval->key;
3355 ilp_var_t reload = PTR_TO_INT(keyval->val);
3358 name = si->lpp->vars[reload];
3359 if(!is_zero(name->value)) {
3361 ir_node *insert_pos = bb;
3362 ir_node *prev = sched_block_last_noncf(si, bb);
3363 op_t *prev_op = get_irn_link(prev);
3365 while(be_is_Spill(prev)) {
3366 prev = sched_prev(prev);
3369 prev_op = get_irn_link(prev);
3371 /* insert reload before pre-remats */
3372 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3373 && prev_op->is_remat && prev_op->attr.remat.pre) {
3377 prev = sched_prev(prev);
3378 } while(be_is_Spill(prev));
3380 prev_op = get_irn_link(prev);
3384 reload = insert_reload(si, irn, insert_pos);
3386 #ifdef KEEPALIVE_RELOADS
3387 pset_insert_ptr(si->spills, reload);
3393 /* walk and insert more reloads and collect remats */
3394 sched_foreach_reverse(bb, irn) {
3395 op_t *op = get_irn_link(irn);
3397 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3398 if(is_Phi(irn)) break;
3401 if(get_irn_mode(irn) != mode_T) {
3402 insert_remat(si, irn);
3407 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3408 ir_node *arg = get_irn_n(irn, n);
3410 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3413 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3414 if(!is_zero(name->value)) {
3416 ir_node *insert_pos = irn;
3417 ir_node *prev = sched_prev(insert_pos);
3420 while(be_is_Spill(prev)) {
3421 prev = sched_prev(prev);
3424 prev_op = get_irn_link(prev);
3426 /* insert reload before pre-remats */
3427 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3428 && prev_op->is_remat && prev_op->attr.remat.pre) {
3432 prev = sched_prev(prev);
3433 } while(be_is_Spill(prev));
3435 prev_op = get_irn_link(prev);
3439 reload = insert_reload(si, arg, insert_pos);
3441 set_irn_n(irn, n, reload);
3443 #ifdef KEEPALIVE_RELOADS
3444 pset_insert_ptr(si->spills, reload);
3452 del_set(spill_bb->ilp);
3453 if(spill_bb->reloads) del_set(spill_bb->reloads);
3457 walker_collect_used(ir_node * irn, void * data)
3459 lc_bitset_t *used = data;
3461 lc_bitset_set(used, get_irn_idx(irn));
3464 struct kill_helper {
3470 walker_kill_unused(ir_node * bb, void * data)
3472 struct kill_helper *kh = data;
3473 const ir_node *bad = get_irg_bad(get_irn_irg(bb));
3477 for(irn=sched_first(bb); !sched_is_end(irn);) {
3478 ir_node *next = sched_next(irn);
3481 if(!lc_bitset_is_set(kh->used, get_irn_idx(irn))) {
3482 if(be_is_Spill(irn) || be_is_Reload(irn)) {
3483 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)));
3485 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3491 set_nodes_block(irn, bad);
3492 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3493 set_irn_n(irn, n, bad);
3501 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3503 struct kill_helper kh;
3505 kh.used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3508 irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3509 irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3511 lc_bitset_free(kh.used);
3515 print_irn_pset(pset * p)
3519 pset_foreach(p, irn) {
3520 ir_printf("%+F\n", irn);
3525 rewire_uses(spill_ilp_t * si)
3527 dom_front_info_t *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
3529 pset *ignore = pset_new_ptr(1);
3531 pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
3533 /* then fix uses of spills */
3534 set_foreach(si->values, defs) {
3537 ir_node *next = defs->remats;
3540 reloads = pset_new_ptr_default();
3543 if(be_is_Reload(next)) {
3544 pset_insert_ptr(reloads, next);
3548 next = get_irn_link(next);
3551 spills = get_spills_for_value(si, defs->value);
3552 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));
3553 if(pset_count(spills) > 1) {
3554 //assert(pset_count(reloads) > 0);
3555 // print_irn_pset(spills);
3556 // print_irn_pset(reloads);
3558 be_ssa_constr_set_ignore(dfi, spills, ignore);
3565 /* first fix uses of remats and reloads */
3566 set_foreach(si->values, defs) {
3568 ir_node *next = defs->remats;
3571 nodes = pset_new_ptr_default();
3572 pset_insert_ptr(nodes, defs->value);
3575 pset_insert_ptr(nodes, next);
3576 next = get_irn_link(next);
3579 if(pset_count(nodes) > 1) {
3580 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
3581 be_ssa_constr_set(dfi, nodes);
3588 // remove_unused_defs(si);
3590 be_free_dominance_frontiers(dfi);
3595 writeback_results(spill_ilp_t * si)
3597 /* walk through the graph and collect all spills, reloads and remats for a value */
3599 si->values = new_set(cmp_defs, 4096);
3601 DBG((si->dbg, LEVEL_1, "Applying results\n"));
3602 delete_unnecessary_remats(si);
3603 si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
3604 irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
3605 irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
3608 /* clean the remat info! there are still back-edges leading there! */
3609 clean_remat_info(si);
3613 connect_all_spills_with_keep(si);
3615 del_set(si->values);
3619 get_n_regs(spill_ilp_t * si)
3621 int arch_n_regs = arch_register_class_n_regs(si->cls);
3625 for(i=0; i<arch_n_regs; i++) {
3626 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
3631 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
3636 walker_reload_mover(ir_node * bb, void * data)
3638 spill_ilp_t *si = data;
3641 sched_foreach(bb, tmp) {
3642 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
3643 ir_node *reload = tmp;
3646 /* move reload upwards */
3648 int pressure = (int)get_irn_link(reload);
3649 if(pressure < si->n_regs) {
3650 irn = sched_prev(reload);
3651 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
3652 sched_remove(reload);
3653 pressure = (int)get_irn_link(irn);
3655 while(pressure < si->n_regs) {
3656 if(sched_is_end(irn) || (be_is_Reload(irn) && has_reg_class(si, irn))) break;
3658 set_irn_link(irn, INT_TO_PTR(pressure+1));
3659 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
3660 irn = sched_prev(irn);
3662 pressure = (int)get_irn_link(irn);
3665 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
3666 sched_put_after(irn, reload);
3673 move_reloads_upward(spill_ilp_t * si)
3675 irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
3680 * Walk all irg blocks and check for interfering spills inside of phi classes
3683 luke_meminterferencechecker(ir_node * bb, void * data)
3685 spill_ilp_t *si = (spill_ilp_t*)data;
3689 live_foreach(bb, li1) {
3690 ir_node *a = (ir_node *) li1->irn;
3692 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
3694 /* a is only interesting if it is inside a phi class */
3695 if (get_phi_class(a)) {
3696 for(li2=li1->next; li2; li2 = li2->next) {
3697 ir_node *b = (ir_node *) li2->irn;
3699 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
3701 /* a and b are only interesting if they are in the same phi class */
3702 if(get_phi_class(a) == get_phi_class(b)) {
3703 if(values_interfere_in_block(bb, a, b)) {
3704 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
3713 verify_phiclasses(spill_ilp_t * si)
3715 /* analyze phi classes */
3716 phi_class_compute(si->chordal_env->irg);
3718 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
3719 irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
3723 walker_spillslotassigner(ir_node * irn, void * data)
3725 spill_ilp_t *si = (spill_ilp_t*)data;
3728 if(!be_is_Spill(irn)) return;
3730 /* set spill context to phi class if it has one ;) */
3732 cls = get_phi_class(irn);
3734 be_set_Spill_context(irn, cls);
3736 be_set_Spill_context(irn, irn);
3741 assign_spillslots(spill_ilp_t * si)
3743 DBG((si->dbg, LEVEL_2, "\t calling spill slot assigner\n"));
3744 irg_walk_graph(si->chordal_env->irg, walker_spillslotassigner, NULL, si);
3748 be_spill_remat(const be_chordal_env_t * chordal_env)
3750 char problem_name[256];
3751 char dump_suffix[256];
3752 char dump_suffix2[256];
3753 struct obstack obst;
3756 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
3757 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
3758 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
3760 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
3761 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
3763 be_check_dominance(chordal_env->irg);
3765 obstack_init(&obst);
3766 si.chordal_env = chordal_env;
3768 si.cls = chordal_env->cls;
3769 si.lpp = new_lpp(problem_name, lpp_minimize);
3770 si.remat_info = new_set(cmp_remat_info, 4096);
3771 si.interferences = new_set(cmp_interference, 32);
3772 si.all_possible_remats = pset_new_ptr_default();
3773 si.spills = pset_new_ptr_default();
3774 si.inverse_ops = pset_new_ptr_default();
3778 si.n_regs = get_n_regs(&si);
3780 set_irg_link(chordal_env->irg, &si);
3781 compute_doms(chordal_env->irg);
3783 /* compute phi classes */
3784 // phi_class_compute(chordal_env->irg);
3786 be_analyze_regpressure(chordal_env, "-pre");
3788 #ifdef COLLECT_REMATS
3789 /* collect remats */
3790 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
3791 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
3794 /* insert possible remats */
3795 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
3796 irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
3797 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
3800 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
3801 connect_all_remats_with_keep(&si);
3802 /* dump graph with inserted remats */
3803 dump_graph_with_remats(chordal_env->irg, dump_suffix);
3806 /* insert copies for phi arguments not in my regclass */
3807 irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
3809 /* recompute liveness */
3810 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
3811 be_liveness(chordal_env->irg);
3815 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
3816 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
3817 irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
3819 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
3820 irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
3822 #ifndef NO_MEMCOPIES
3823 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
3824 memcopyhandler(&si);
3832 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
3833 if ((f = fopen(buf, "wt")) != NULL) {
3834 lpp_dump_plain(si.lpp, f);
3841 DBG((si.dbg, LEVEL_1, "\tSolving %F\n", chordal_env->irg));
3843 lpp_set_time_limit(si.lpp, ILP_TIMEOUT);
3847 lpp_solve_cplex(si.lpp);
3849 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
3851 assert(lpp_is_sol_valid(si.lpp)
3852 && "solution of ILP must be valid");
3854 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));
3856 #ifdef DUMP_SOLUTION
3861 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
3862 if ((f = fopen(buf, "wt")) != NULL) {
3864 for (i = 0; i < si.lpp->var_next; ++i) {
3865 lpp_name_t *name = si.lpp->vars[i];
3866 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
3873 writeback_results(&si);
3877 kill_all_unused_values_in_schedule(&si);
3879 #if defined(KEEPALIVE_SPILLS) || defined(KEEPALIVE_RELOADS)
3880 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
3883 // move reloads upwards
3884 be_liveness(chordal_env->irg);
3885 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
3886 move_reloads_upward(&si);
3888 #ifndef NO_MEMCOPIES
3889 verify_phiclasses(&si);
3890 assign_spillslots(&si);
3893 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
3895 dump_pressure_graph(&si, dump_suffix2);
3897 be_analyze_regpressure(chordal_env, "-post");
3899 be_check_dominance(chordal_env->irg);
3901 free_dom(chordal_env->irg);
3902 del_set(si.interferences);
3903 del_pset(si.inverse_ops);
3904 del_pset(si.all_possible_remats);
3905 del_pset(si.spills);
3907 obstack_free(&obst, NULL);
3908 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
3911 #else /* WITH_ILP */
3914 only_that_you_can_compile_without_WITH_ILP_defined(void)
3918 #endif /* WITH_ILP */