1 /** vim: set sw=4 ts=4:
4 * @author Adam M. Szalkowski & Sebastian Hack
6 * ILP based spilling & rematerialization
8 * Copyright (C) 2006 Universitaet Karlsruhe
9 * Released under the GPL
37 #include <lpp/lpp_net.h>
38 #include <lpp/lpp_cplex.h>
39 //#include <lc_pset.h>
40 #include <libcore/lc_bitset.h>
44 #include "besched_t.h"
49 #include "bespillremat.h"
52 #include "bechordal_t.h"
58 //#define KEEPALIVE /* keep alive all inserted remats and dump graph with remats */
59 #define COLLECT_REMATS /* enable rematerialization */
60 #define COLLECT_INVERSE_REMATS /* enable placement of inverse remats */
61 #define REMAT_WHILE_LIVE /* only remat values that are live */
62 //#define NO_ENLARGE_L1V3N355 /* do not remat after the death of some operand */
63 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
64 //#define MAY_DIE_AT_PRE_REMAT /* allow values to die after a pre remat */
65 #define CHECK_POST_REMAT /* check pressure after post remats (conservative but otherwise we can temporarily exceed the register pressure) */
66 #define NO_SINGLE_USE_REMATS /* do not repair schedule */
67 //#define KEEPALIVE_SPILLS
68 //#define KEEPALIVE_RELOADS
69 #define GOODWIN_REDUCTION
73 #define LPP_SERVER "i44pc52"
74 #define LPP_SOLVER "cplex"
80 #define ILP_TIMEOUT 90
84 typedef struct _spill_ilp_t {
85 const arch_register_class_t *cls;
87 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 {
113 typedef struct _remat_t {
114 const ir_node *op; /**< for copy_irn */
115 const ir_node *proj; /**< not NULL if the above op produces a tuple */
116 const ir_node *value; /**< the value which is being recomputed by this remat */
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 */
142 typedef struct _defs_t {
144 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
145 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
148 typedef struct _remat_info_t {
149 const ir_node *irn; /**< the irn to which these remats belong */
150 pset *remats; /**< possible remats for this value */
151 pset *remats_by_operand; /**< remats with this value as operand */
154 typedef struct _keyval_t {
159 typedef struct _spill_t {
169 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
171 return chordal_has_class(si->chordal_env, irn);
176 cmp_remat(const void *a, const void *b)
178 const keyval_t *p = a;
179 const keyval_t *q = b;
180 const remat_t *r = p->val;
181 const remat_t *s = q->val;
185 return !(r == s || r->op == s->op);
189 cmp_remat(const void *a, const void *b)
191 const remat_t *r = a;
192 const remat_t *s = a;
194 return !(r == s || r->op == s->op);
198 cmp_spill(const void *a, const void *b, size_t size)
200 const spill_t *p = a;
201 const spill_t *q = b;
203 // return !(p->irn == q->irn && p->bb == q->bb);
204 return !(p->irn == q->irn);
208 set_find_keyval(set * set, void * key)
213 return set_find(set, &query, sizeof(query), HASH_PTR(key));
217 set_insert_keyval(set * set, void * key, void * val)
223 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
227 set_find_def(set * set, ir_node * value)
232 return set_find(set, &query, sizeof(query), HASH_PTR(value));
236 set_insert_def(set * set, ir_node * value)
243 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
247 set_find_spill(set * set, ir_node * value)
252 return set_find(set, &query, sizeof(query), HASH_PTR(value));
255 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
256 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
257 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
258 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
259 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
262 cmp_remat_info(const void *a, const void *b, size_t size)
264 const remat_info_t *p = a;
265 const remat_info_t *q = b;
267 return !(p->irn == q->irn);
271 cmp_defs(const void *a, const void *b, size_t size)
276 return !(p->value == q->value);
280 cmp_keyval(const void *a, const void *b, size_t size)
282 const keyval_t *p = a;
283 const keyval_t *q = b;
285 return !(p->key == q->key);
289 execution_frequency(const spill_ilp_t * si, const ir_node * irn)
293 return get_block_execfreq(si->execfreqs, irn);
295 return get_block_execfreq(si->execfreqs, get_nodes_block(irn));
299 return exp(get_loop_depth(get_irn_loop(irn)) * log(10));
301 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10));
306 * Checks, whether node and its operands have suitable reg classes
309 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
313 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
314 int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
318 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
321 for (i = 0, n = get_irn_arity(irn); i < n && remat; ++i) {
322 ir_node *op = get_irn_n(irn, i);
323 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
326 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
333 * Try to create a remat from @p op with destination value @p dest_value
335 static INLINE remat_t *
336 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
338 remat_t *remat = NULL;
340 // if(!mode_is_datab(get_irn_mode(dest_value)))
343 if(dest_value == op) {
344 const ir_node *proj = NULL;
346 if(is_Proj(dest_value)) {
347 op = get_irn_n(op, 0);
351 if(!is_rematerializable(si, op))
354 remat = obstack_alloc(si->obst, sizeof(*remat));
356 remat->cost = arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, op);
357 remat->value = dest_value;
361 arch_inverse_t inverse;
365 /* get the index of the operand we want to retrieve by the inverse op */
366 for (i = 0, n = get_irn_arity(op); i < n; ++i) {
367 ir_node *arg = get_irn_n(op, i);
369 if(arg == dest_value) break;
371 if(i == n) return NULL;
373 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", i, op));
375 /* else ask the backend to give an inverse op */
376 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, i, &inverse, si->obst)) {
379 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
381 assert(inverse.n > 0 && "inverse op should have at least one node");
383 for(i=0; i<inverse.n; ++i) {
384 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
388 remat = obstack_alloc(si->obst, sizeof(*remat));
389 remat->op = inverse.nodes[0];
390 remat->cost = inverse.costs;
391 remat->value = dest_value;
392 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
395 assert(is_Proj(remat->proj));
397 assert(0 && "I can not handle remats with more than 2 nodes");
404 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
406 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
414 add_remat(const spill_ilp_t * si, const remat_t * remat)
416 remat_info_t *remat_info,
422 assert(remat->value);
424 query.irn = remat->value;
426 query.remats_by_operand = NULL;
427 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
429 if(remat_info->remats == NULL) {
430 remat_info->remats = new_pset(cmp_remat, 4096);
432 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
434 /* insert the remat into the remats_be_operand set of each argument of the remat op */
435 for (i = 0, n = get_irn_arity(remat->op); i < n; ++i) {
436 ir_node *arg = get_irn_n(remat->op, i);
440 query.remats_by_operand = NULL;
441 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
443 if(remat_info->remats_by_operand == NULL) {
444 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
446 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
451 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
453 const ir_edge_t *edge = get_irn_out_edge_first(irn);
457 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
460 edge = get_irn_out_edge_next(irn, edge);
467 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
473 #ifdef NO_SINGLE_USE_REMATS
474 if(has_reg_class(si, op) && (get_irn_n_nonremat_edges(si, op) > 1)) {
476 if(has_reg_class(si, op)) {
478 remat = get_remat_from_op(si, op, op);
480 add_remat(si, remat);
484 #ifdef COLLECT_INVERSE_REMATS
485 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
487 for (i = 0, n = get_irn_arity(op); i < n; ++i) {
488 ir_node *arg = get_irn_n(op, i);
490 if(has_reg_class(si, arg)) {
491 /* try to get an inverse remat */
492 remat = get_remat_from_op(si, arg, op);
494 add_remat(si, remat);
503 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
506 ir_node *def_block = get_nodes_block(val);
512 /* if pos is at end of a basic block */
514 ret = (pos == def_block || block_dominates(def_block, pos));
515 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
519 /* else if this is a normal operation */
520 block = get_nodes_block(pos);
521 if(block == def_block) {
522 if(!sched_is_scheduled(val)) return 1;
524 ret = sched_comes_after(val, pos);
525 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
529 ret = block_dominates(def_block, block);
530 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
534 static INLINE ir_node *
535 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
537 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
541 * Returns first non-Phi node of block @p bb
543 static INLINE ir_node *
544 sched_block_first_nonphi(const ir_node * bb)
546 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
550 sched_skip_proj_predicator(const ir_node * irn, void * data)
552 return (is_Proj(irn));
555 static INLINE ir_node *
556 sched_next_nonproj(const ir_node * irn, int forward)
558 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
562 * Returns next operation node (non-Proj) after @p irn
563 * or the basic block of this node
565 static INLINE ir_node *
566 sched_next_op(const ir_node * irn)
568 ir_node *next = sched_next(irn);
573 return sched_next_nonproj(next, 1);
577 * Returns previous operation node (non-Proj) before @p irn
578 * or the basic block of this node
580 static INLINE ir_node *
581 sched_prev_op(const ir_node * irn)
583 ir_node *prev = sched_prev(irn);
588 return sched_next_nonproj(prev, 0);
592 sched_put_after(ir_node * insert, ir_node * irn)
594 if(is_Block(insert)) {
595 insert = sched_block_first_nonphi(insert);
597 insert = sched_next_op(insert);
599 sched_add_before(insert, irn);
603 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
605 if(is_Block(insert)) {
606 insert = sched_block_last_noncf(si, insert);
608 insert = sched_next_nonproj(insert, 0);
609 insert = sched_prev(insert);
611 sched_add_after(insert, irn);
615 * Tells you whether a @p remat can be placed before the irn @p pos
618 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
620 const ir_node *op = remat->op;
627 prev = sched_block_last_noncf(si, pos);
628 prev = sched_next_nonproj(prev, 0);
630 prev = sched_prev_op(pos);
632 /* do not remat if the rematted value is defined immediately before this op */
633 if(prev == remat->op) {
638 /* this should be just fine, the following OP will be using this value, right? */
640 /* only remat AFTER the real definition of a value (?) */
641 if(!value_is_defined_before(si, pos, remat->value)) {
642 // ir_fprintf(stderr, "error(not defined)");
647 for(i=0, n=get_irn_arity(op); i<n && res; ++i) {
648 const ir_node *arg = get_irn_n(op, i);
650 #ifdef NO_ENLARGE_L1V3N355
651 if(has_reg_class(si, arg) && live) {
652 res &= pset_find_ptr(live, arg)?1:0;
654 res &= value_is_defined_before(si, pos, arg);
657 res &= value_is_defined_before(si, pos, arg);
665 * Tells you whether a @p remat can be placed after the irn @p pos
668 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
671 pos = sched_block_first_nonphi(pos);
673 pos = sched_next_op(pos);
676 /* only remat AFTER the real definition of a value (?) */
677 if(!value_is_defined_before(si, pos, remat->value)) {
681 return can_remat_before(si, remat, pos, live);
685 * Collect potetially rematerializable OPs
688 walker_remat_collector(ir_node * irn, void * data)
690 spill_ilp_t *si = data;
692 if(!is_Block(irn) && !is_Phi(irn)) {
693 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
694 get_remats_from_op(si, irn);
699 * Inserts a copy of @p irn before @p pos
702 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
707 bb = is_Block(pos)?pos:get_nodes_block(pos);
708 copy = exact_copy(irn);
709 set_nodes_block(copy, bb);
710 sched_put_before(si, pos, copy);
716 * Inserts a copy of @p irn after @p pos
719 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
724 bb = is_Block(pos)?pos:get_nodes_block(pos);
725 copy = exact_copy(irn);
726 set_nodes_block(copy, bb);
727 sched_put_after(pos, copy);
733 insert_remat_after(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
737 if(can_remat_after(si, remat, pos, live)) {
742 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
744 copy = insert_copy_after(si, remat->op, pos);
746 // ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", remat->value, pos);
747 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
748 op = obstack_alloc(si->obst, sizeof(*op));
750 op->attr.remat.remat = remat;
751 op->attr.remat.pre = 0;
752 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
754 set_irn_link(copy, op);
755 pset_insert_ptr(si->all_possible_remats, copy);
757 proj_copy = insert_copy_after(si, remat->proj, copy);
758 set_irn_n(proj_copy, 0, copy);
759 set_irn_link(proj_copy, op);
760 pset_insert_ptr(si->all_possible_remats, proj_copy);
768 insert_remat_before(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
772 if(can_remat_before(si, remat, pos, live)) {
777 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
779 copy = insert_copy_before(si, remat->op, pos);
781 // ir_snprintf(buf, sizeof(buf), "remat_%N_%N", remat->value, pos);
782 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
783 op = obstack_alloc(si->obst, sizeof(*op));
785 op->attr.remat.remat = remat;
786 op->attr.remat.pre = 1;
787 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
789 set_irn_link(copy, op);
790 pset_insert_ptr(si->all_possible_remats, copy);
792 proj_copy = insert_copy_after(si, remat->proj, copy);
793 set_irn_n(proj_copy, 0, copy);
794 set_irn_link(proj_copy, op);
795 pset_insert_ptr(si->all_possible_remats, proj_copy);
803 get_block_n_succs(const ir_node *block) {
804 const ir_edge_t *edge;
806 assert(edges_activated(current_ir_graph));
808 edge = get_block_succ_first(block);
812 edge = get_block_succ_next(block, edge);
817 is_merge_edge(const ir_node * bb)
819 #ifdef GOODWIN_REDUCTION
820 return get_block_n_succs(bb) == 1;
827 is_diverge_edge(const ir_node * bb)
829 #ifdef GOODWIN_REDUCTION
830 return get_Block_n_cfgpreds(bb) == 1;
837 * Insert (so far unused) remats into the irg to
838 * recompute the potential liveness of all values
841 walker_remat_insertor(ir_node * bb, void * data)
843 spill_ilp_t *si = data;
844 spill_bb_t *spill_bb;
849 pset *live = pset_new_ptr_default();
851 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
853 live_foreach(bb, li) {
854 ir_node *value = (ir_node *) li->irn;
856 /* add remats at end of block */
857 if (live_is_end(li) && has_reg_class(si, value)) {
858 pset_insert_ptr(live, value);
862 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
863 set_irn_link(bb, spill_bb);
865 irn = sched_last(bb);
866 while(!sched_is_end(irn)) {
872 next = sched_prev(irn);
874 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
876 if(is_Phi(irn) || is_Proj(irn)) {
879 if(has_reg_class(si, irn)) {
880 pset_remove_ptr(live, irn);
883 op = obstack_alloc(si->obst, sizeof(*op));
885 op->attr.live_range.reloads = NULL;
886 op->attr.live_range.ilp = ILP_UNDEF;
887 set_irn_link(irn, op);
893 op = obstack_alloc(si->obst, sizeof(*op));
895 op->attr.live_range.ilp = ILP_UNDEF;
896 op->attr.live_range.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.reloads) * get_irn_arity(irn));
897 memset(op->attr.live_range.reloads, 0xFF, sizeof(*op->attr.live_range.reloads) * get_irn_arity(irn));
898 set_irn_link(irn, op);
900 args = pset_new_ptr_default();
902 /* collect arguments of op */
903 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
904 ir_node *arg = get_irn_n(irn, i);
906 pset_insert_ptr(args, arg);
909 /* set args of op live in epilog */
910 pset_foreach(args, arg) {
911 if(has_reg_class(si, arg)) {
912 pset_insert_ptr(live, arg);
916 /* insert all possible remats after irn */
917 pset_foreach(args, arg) {
918 remat_info_t *remat_info,
922 /* continue if the operand has the wrong reg class
924 if(!has_reg_class(si, arg))
929 query.remats_by_operand = NULL;
930 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
936 /* do not place post remats after jumps */
937 if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
939 if(remat_info->remats_by_operand) {
940 pset_foreach(remat_info->remats_by_operand, remat) {
941 /* do not insert remats producing the same value as one of the operands */
942 if(!pset_find_ptr(args, remat->value)) {
943 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
944 #ifdef REMAT_WHILE_LIVE
945 if(pset_find_ptr(live, remat->value)) {
946 insert_remat_after(si, remat, irn, live);
949 insert_remat_after(si, remat, irn, live);
956 /* delete defined value from live set */
957 if(has_reg_class(si, irn)) {
958 pset_remove_ptr(live, irn);
961 /* insert all possible remats before irn */
962 pset_foreach(args, arg) {
963 remat_info_t *remat_info,
967 /* continue if the operand has the wrong reg class
969 if(!has_reg_class(si, arg))
974 query.remats_by_operand = NULL;
975 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
981 if(remat_info->remats) {
982 pset_foreach(remat_info->remats, remat) {
983 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
984 #ifdef REMAT_WHILE_LIVE
985 if(pset_find_ptr(live, remat->value)) {
986 insert_remat_before(si, remat, irn, live);
989 insert_remat_before(si, remat, irn, live);
999 live_foreach(bb, li) {
1000 ir_node *value = (ir_node *) li->irn;
1002 /* add remats at end if successor has multiple predecessors */
1003 if(is_merge_edge(bb)) {
1004 /* add remats at end of block */
1005 if (live_is_end(li) && has_reg_class(si, value)) {
1006 remat_info_t *remat_info,
1011 query.remats = NULL;
1012 query.remats_by_operand = NULL;
1013 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1015 if(remat_info && remat_info->remats) {
1016 pset_foreach(remat_info->remats, remat) {
1017 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1019 insert_remat_before(si, remat, bb, NULL);
1025 if(is_diverge_edge(bb) > 1) {
1026 /* add remat2s at beginning of block */
1027 if ((live_is_in(li) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) {
1028 remat_info_t *remat_info,
1033 query.remats = NULL;
1034 query.remats_by_operand = NULL;
1035 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1037 if(remat_info && remat_info->remats) {
1038 pset_foreach(remat_info->remats, remat) {
1039 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at beginning of block %+F\n", remat->op, bb));
1041 /* put the remat here if all its args are available */
1042 insert_remat_after(si, remat, bb, NULL);
1052 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1055 luke_endwalker(ir_node * bb, void * data)
1057 spill_ilp_t *si = (spill_ilp_t*)data;
1064 spill_bb_t *spill_bb = get_irn_link(bb);
1067 live = pset_new_ptr_default();
1068 use_end = pset_new_ptr_default();
1070 live_foreach(bb, li) {
1071 irn = (ir_node *) li->irn;
1072 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1075 pset_insert_ptr(live, irn);
1076 op = get_irn_link(irn);
1077 assert(!op->is_remat);
1081 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1082 /* their reg_out is unimportant because it can always be set */
1083 sched_foreach_reverse(bb, irn) {
1087 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1089 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
1090 ir_node *irn_arg = get_irn_n(irn, i);
1091 if(has_reg_class(si, irn_arg)) {
1092 pset_insert_ptr(use_end, irn);
1097 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1098 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1100 spill_bb->ilp = new_set(cmp_spill, 16);
1102 live_foreach(bb, li) {
1103 irn = (ir_node *) li->irn;
1104 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1109 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1111 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1112 spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1113 /* if irn is used at the end of the block, then it is live anyway */
1114 if(!pset_find_ptr(use_end, irn))
1115 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1117 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1118 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1120 if(is_diverge_edge(bb)) {
1121 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1122 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE*execution_frequency(si, bb));
1124 spill->spill = ILP_UNDEF;
1127 spill->reg_in = ILP_UNDEF;
1128 spill->mem_in = ILP_UNDEF;
1137 next_post_remat(const ir_node * irn)
1142 irn = sched_block_first_nonphi(irn);
1144 irn = sched_next_op(irn);
1147 if(sched_is_end(irn))
1150 op = (op_t*)get_irn_link(irn);
1151 if(op->is_remat && !op->attr.remat.pre) {
1160 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1166 ret = sched_block_last_noncf(si, irn);
1167 ret = sched_next(ret);
1168 ret = sched_prev_op(ret);
1170 ret = sched_prev_op(irn);
1173 if(sched_is_end(ret) || is_Phi(ret))
1176 op = (op_t*)get_irn_link(ret);
1177 if(op->is_remat && op->attr.remat.pre) {
1185 * Find a remat of value @p value in the epilog of @p pos
1188 find_post_remat(const ir_node * value, const ir_node * pos)
1190 while((pos = next_post_remat(pos)) != NULL) {
1193 op = get_irn_link(pos);
1194 assert(op->is_remat && !op->attr.remat.pre);
1196 if(op->attr.remat.remat->value == value)
1197 return (ir_node*)pos;
1200 const ir_edge_t *edge;
1201 foreach_out_edge(pos, edge) {
1202 ir_node *proj = get_edge_src_irn(edge);
1203 assert(is_Proj(proj));
1213 * Find a remat of value @p value in the prolog of @p pos
1216 find_pre_remat(const spill_ilp_t * si, const ir_node * value, const ir_node * pos)
1218 while((pos = next_pre_remat(si,pos)) != NULL) {
1221 op = get_irn_link(pos);
1222 assert(op->is_remat && op->attr.remat.pre);
1224 if(op->attr.remat.remat->value == value)
1225 return (ir_node*)pos;
1232 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1234 spill_bb_t *spill_bb = get_irn_link(bb);
1240 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1242 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1244 spill->reg_out = ILP_UNDEF;
1245 spill->reg_in = ILP_UNDEF;
1246 spill->mem_in = ILP_UNDEF;
1248 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1249 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1251 if(is_diverge_edge(bb)) {
1252 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1253 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE*execution_frequency(si, bb));
1255 spill->spill = ILP_UNDEF;
1263 * Walk all irg blocks and emit this ILP
1266 luke_blockwalker(ir_node * bb, void * data)
1268 spill_ilp_t *si = (spill_ilp_t*)data;
1274 spill_bb_t *spill_bb = get_irn_link(bb);
1280 live = pset_new_ptr_default();
1282 /* do something at the end of the block */
1284 /* init live values at end of block */
1285 live_foreach(bb, li) {
1286 ir_node *irn = (ir_node *) li->irn;
1288 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1289 pset_insert_ptr(live, irn);
1293 if(is_merge_edge(bb)) {
1294 spill_bb->reloads = obstack_alloc(si->obst, pset_count(live) * sizeof(*spill_bb->reloads));
1295 memset(spill_bb->reloads, 0xFF, pset_count(live) * sizeof(*spill_bb->reloads));
1297 spill_bb->reloads = NULL;
1301 live_foreach(bb, li) {
1302 ir_node *irn = (ir_node *) li->irn;
1305 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1306 spill = set_find_spill(spill_bb->ilp, irn);
1309 if(spill_bb->reloads) {
1310 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1311 spill_bb->reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1313 /* reload <= mem_out */
1314 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1315 lpp_set_factor_fast(si->lpp, cst, spill_bb->reloads[i], 1.0);
1316 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
1319 op = get_irn_link(irn);
1320 assert(!op->is_remat);
1322 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1323 op->attr.live_range.ilp = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1324 op->attr.live_range.op = bb;
1326 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1327 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1329 /* reg_out - reload - remat - live_range <= 0 */
1330 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1331 if(spill_bb->reloads) lpp_set_factor_fast(si->lpp, cst, spill_bb->reloads[i], -1.0);
1332 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1333 foreach_pre_remat(si, bb, tmp) {
1334 op_t *remat_op = get_irn_link(tmp);
1335 if(remat_op->attr.remat.remat->value == irn) {
1336 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1343 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1345 sched_foreach_reverse(bb, irn) {
1351 ilp_cst_t check_pre,
1353 #ifdef CHECK_POST_REMAT
1354 ilp_cst_t check_post_remat;
1356 set *args = new_set(cmp_keyval, get_irn_arity(irn));
1362 op = get_irn_link(irn);
1364 if(op->is_remat) continue;
1365 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1367 if(has_reg_class(si, irn)) {
1368 assert(pset_find_ptr(live, irn));
1369 pset_remove_ptr(live, irn);
1372 /* init set of irn's arguments */
1373 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
1374 ir_node *irn_arg = get_irn_n(irn, i);
1375 if(has_reg_class(si, irn_arg)) {
1376 set_insert_keyval(args, irn_arg, (void*)i);
1380 #ifdef CHECK_POST_REMAT
1381 /* check the register pressure after the epilog */
1382 ir_snprintf(buf, sizeof(buf), "check_post_remat_%N", irn);
1383 check_post_remat = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1385 /* iterate over L\U */
1386 pset_foreach(live, tmp) {
1387 if(!set_find_keyval(args, tmp)) {
1388 /* if a live value is not used by irn */
1389 tmp_op = get_irn_link(tmp);
1390 // assert(tmp_op->attr.live_range.op != irn);
1391 lpp_set_factor_fast(si->lpp, check_post_remat, tmp_op->attr.live_range.ilp, 1.0);
1394 /* iterate over following remats and remove possibly defined values again from check_post_remat */
1395 foreach_post_remat(irn, tmp) {
1396 op_t *remat_op = get_irn_link(tmp);
1397 const ir_node *value = remat_op->attr.remat.remat->value;
1398 op_t *val_op = get_irn_link(value);
1400 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
1402 /* values that are defined by remat2s are not counted */
1403 #ifdef REMAT_WHILE_LIVE
1404 assert(val_op->attr.live_range.ilp);
1405 lpp_set_factor_fast(si->lpp, check_post_remat, val_op->attr.live_range.ilp, 0.0);
1407 if(val_op->attr.live_range.ilp != ILP_UNDEF) {
1408 lpp_set_factor_fast(si->lpp, check_post_remat, val_op->attr.live_range.ilp, 0.0);
1410 #endif /* REMAT_WHILE_LIVE */
1412 #endif /* CHECK_POST_REMAT */
1415 /* new live ranges for values from L\U defined by remat2s or used by remats */
1416 pset_foreach(live, tmp) {
1417 ir_node *value = tmp;//remat_op->attr.remat.remat->value;
1418 op_t *value_op = get_irn_link(value);
1420 if(!set_find_keyval(args, value)) {
1421 ilp_var_t prev_lr = ILP_UNDEF;
1425 foreach_post_remat(irn, remat) {
1426 op_t *remat_op = get_irn_link(remat);
1428 /* if value is being rematerialized by this remat */
1429 if(value == remat_op->attr.remat.remat->value) {
1430 if(cst == ILP_UNDEF) {
1431 /* next_live_range <= prev_live_range + sum remat2s */
1432 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1433 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1434 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1435 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1436 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1437 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1440 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1444 #ifdef MAY_DIE_AT_PRE_REMAT
1445 if(cst == ILP_UNDEF) {
1446 foreach_pre_remat(si, irn, remat) {
1450 for (i = 0, n = get_irn_arity(remat); i < n; ++i) {
1451 ir_node *remat_arg = get_irn_n(remat, i);
1453 /* if value is being used by this remat */
1454 if(value == remat_arg) {
1455 /* next_live_range <= prev_live_range */
1456 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1457 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1459 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1460 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1461 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1462 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1465 /* TODO check afterwards whether lr dies after a pre-remat (should not happen) */
1472 if(prev_lr != ILP_UNDEF) {
1473 value_op->attr.live_range.ilp = prev_lr;
1474 value_op->attr.live_range.op = irn;
1479 /* get count of values in my register class defined by irn */
1480 /* also add defined values to check_post_remat; do this before iterating over args */
1481 if(get_irn_mode(irn) == mode_T) {
1482 ir_node *proj = sched_next(irn);
1483 op_t *proj_op = get_irn_link(proj);
1485 while(is_Proj(proj)) {
1486 if(has_reg_class(si, proj)) {
1488 #ifdef CHECK_POST_REMAT
1489 lpp_set_factor_fast(si->lpp, check_post_remat, proj_op->attr.live_range.ilp, 1.0);
1492 proj = sched_next(proj);
1493 proj_op = get_irn_link(proj);
1496 if(has_reg_class(si, irn)) {
1498 #ifdef CHECK_POST_REMAT
1499 lpp_set_factor_fast(si->lpp, check_post_remat, op->attr.live_range.ilp, 1.0);
1503 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
1505 /* count how many regs irn needs for arguments */
1506 k = set_count(args);
1508 /* check the register pressure in the prolog */
1509 /* sum_{L\U} lr <= n - |U| */
1510 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
1511 check_pre = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - k);
1513 /* check the register pressure in the epilog */
1514 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
1515 check_post = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - d);
1517 set_foreach(args, keyval) {
1523 ir_node *arg = keyval->key;
1525 spill = add_to_spill_bb(si, bb, arg);
1527 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
1528 next_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1530 i = (int)keyval->val;
1533 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
1534 op->attr.live_range.reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1536 /* reload <= mem_out */
1537 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1538 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.reloads[i], 1.0);
1539 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
1541 arg_op = get_irn_link(arg);
1543 /* requirement: arg must be in register for use */
1544 /* reload + remat + live_range == 1 */
1545 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
1546 cst = lpp_add_cst(si->lpp, buf, lpp_equal, 1.0);
1548 lpp_set_factor_fast(si->lpp, cst, next_lr, 1.0);
1549 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.reloads[i], 1.0);
1550 foreach_pre_remat(si, irn, tmp) {
1551 op_t *remat_op = get_irn_link(tmp);
1552 if(remat_op->attr.remat.remat->value == arg) {
1553 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1557 /* the epilog stuff - including post_use, post, post_remat */
1558 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
1559 post_use = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1561 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
1563 /* arg is live throughout epilog if the next live_range is in a register */
1564 if(pset_find_ptr(live, arg)) {
1565 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
1567 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1568 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1569 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
1570 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1572 #ifdef CHECK_POST_REMAT
1573 lpp_set_factor_fast(si->lpp, check_post_remat, arg_op->attr.live_range.ilp, 1.0);
1577 /*forall remat2 which use arg add a similar cst*/
1578 foreach_post_remat(irn, tmp) {
1582 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1583 ir_node *remat_arg = get_irn_n(tmp, i);
1584 op_t *remat_op = get_irn_link(tmp);
1586 if(remat_arg == arg) {
1587 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
1589 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1590 cst = lpp_add_cst(si->lpp, buf, lpp_greater, 0.0);
1591 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1592 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1597 /* new live range begins for each argument */
1598 arg_op->attr.live_range.ilp = next_lr;
1599 arg_op->attr.live_range.op = irn;
1601 pset_insert_ptr(live, arg);
1604 /* start new live ranges for values used by remats */
1605 foreach_pre_remat(si, irn, tmp) {
1609 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1610 ir_node *remat_arg = get_irn_n(tmp, i);
1611 op_t *arg_op = get_irn_link(remat_arg);
1614 if(!has_reg_class(si, remat_arg)) continue;
1616 /* if value is becoming live through use by remat */
1617 if(!pset_find_ptr(live, remat_arg)) {
1618 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1619 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1621 arg_op->attr.live_range.ilp = prev_lr;
1622 arg_op->attr.live_range.op = irn;
1624 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat %+F\n", remat_arg, tmp));
1626 /* TODO ist das hier die richtige Stelle???? */
1627 pset_insert_ptr(live, remat_arg);
1628 add_to_spill_bb(si, bb, remat_arg);
1630 /* TODO check afterwards whether lr dies after a pre-remat (should not happen) */
1634 /* iterate over L\U */
1635 pset_foreach(live, tmp) {
1636 if(!set_find_keyval(args, tmp)) {
1637 /* if a live value is not used by irn */
1638 tmp_op = get_irn_link(tmp);
1639 // assert(tmp_op->attr.live_range.op != irn);
1640 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
1641 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
1645 /* requirements for remats */
1646 foreach_pre_remat(si, irn, tmp) {
1647 op_t *remat_op = get_irn_link(tmp);
1651 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1652 ir_node *remat_arg = get_irn_n(tmp, i);
1653 op_t *arg_op = get_irn_link(remat_arg);
1655 if(!has_reg_class(si, remat_arg)) continue;
1657 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1658 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1659 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1661 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1662 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1664 /* if remat arg is also used by current op then we can use reload placed for this argument */
1665 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
1666 int index = (int)keyval->val;
1668 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.reloads[index], -1.0);
1673 /* requirements for remats2
1675 * TODO unsure if this does the right thing.
1676 * should insert values into set if they do not become live through remat and
1679 foreach_post_remat(irn, tmp) {
1680 op_t *remat_op = get_irn_link(tmp);
1684 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1685 ir_node *remat_arg = get_irn_n(tmp, i);
1686 op_t *arg_op = get_irn_link(remat_arg);
1688 if(!has_reg_class(si, remat_arg)) continue;
1690 /* only for values in L\U, the others are handled with post_use */
1691 if(!set_find_keyval(args, remat_arg)) {
1692 /* remat <= live_rang(remat_arg) */
1693 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
1694 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1696 /* if value is becoming live through use by remat2 */
1697 if(!pset_find_ptr(live, remat_arg)) {
1700 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1701 lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1703 arg_op->attr.live_range.ilp = lr;
1704 arg_op->attr.live_range.op = irn;
1706 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
1708 pset_insert_ptr(live, remat_arg);
1709 add_to_spill_bb(si, bb, remat_arg);
1712 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1713 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1718 #ifdef CHECK_POST_REMAT
1719 /* iterate over following remats and add them to check_post_remat */
1720 foreach_post_remat(irn, tmp) {
1721 op_t *remat_op = get_irn_link(tmp);
1723 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
1725 lpp_set_factor_fast(si->lpp, check_post_remat, remat_op->attr.remat.ilp, 1.0);
1731 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
1733 pset_foreach(live, tmp) {
1734 assert(has_reg_class(si, tmp));
1737 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
1738 ir_node *arg = get_irn_n(irn, i);
1740 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
1748 /* do something at the beginning of the block */
1750 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
1751 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
1753 pset_foreach(live, irn) {
1754 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
1757 /* construct mem_outs for all values */
1759 set_foreach(spill_bb->ilp, spill) {
1760 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
1761 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1763 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
1764 if(spill->spill != ILP_UNDEF) {
1765 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
1768 if(pset_find_ptr(live, spill->irn)) {
1769 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
1771 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
1772 spill->mem_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1774 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
1779 /* L\U is empty at bb start */
1780 /* arg is live throughout epilog if it is reg_in into this block */
1782 /* check the register pressure at the beginning of the block
1785 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
1786 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1788 pset_foreach(live, irn) {
1789 spill = set_find_spill(spill_bb->ilp, irn);
1792 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
1793 spill->reg_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1795 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
1797 foreach_post_remat(bb, irn) {
1798 op_t *remat_op = get_irn_link(irn);
1800 DBG((si->dbg, LEVEL_4, "\t next post remat: %+F\n", irn));
1801 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
1803 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1806 /* forall remat2 add requirements */
1807 foreach_post_remat(bb, tmp) {
1811 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1812 ir_node *remat_arg = get_irn_n(tmp, i);
1813 op_t *remat_op = get_irn_link(tmp);
1815 if(!has_reg_class(si, remat_arg)) continue;
1817 spill = set_find_spill(spill_bb->ilp, remat_arg);
1820 /* TODO verify this is placed correctly */
1821 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
1822 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1823 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
1824 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1828 /* mem_in/reg_in for live_in values, especially phis and their arguments */
1829 pset_foreach(live, irn) {
1834 spill = set_find_spill(spill_bb->ilp, irn);
1835 assert(spill && spill->irn == irn);
1837 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
1838 for (i = 0, n = get_Phi_n_preds(irn); i < n; ++i) {
1841 ir_node *phi_arg = get_Phi_pred(irn, i);
1842 ir_node *bb_p = get_Block_cfgpred_block(bb, i);
1843 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
1846 /* although the phi is in the right regclass one or more of
1847 * its arguments can be in a different one or at least to
1850 if(has_reg_class(si, phi_arg)) {
1851 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
1852 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1853 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
1854 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1856 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
1857 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
1859 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
1862 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
1863 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
1867 /* else assure the value arrives on all paths in the same resource */
1869 for (i = 0, n = get_Block_n_cfgpreds(bb); i < n; ++i) {
1872 ir_node *bb_p = get_Block_cfgpred_block(bb, i);
1873 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
1876 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
1877 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1878 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
1879 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1881 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
1882 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
1884 spill_p = set_find_spill(spill_bb_p->ilp, irn);
1887 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
1888 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
1893 /* first live ranges from reg_ins */
1894 pset_foreach(live, irn) {
1895 op_t *op = get_irn_link(irn);
1897 spill = set_find_spill(spill_bb->ilp, irn);
1898 assert(spill && spill->irn == irn);
1900 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
1901 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1902 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1903 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
1905 foreach_post_remat(bb, tmp) {
1906 op_t *remat_op = get_irn_link(tmp);
1908 if(remat_op->attr.remat.remat->value == irn) {
1909 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1914 /* walk forward now and compute constraints for placing spills */
1915 /* this must only be done for values that are not defined in this block */
1916 if(is_diverge_edge(bb)) {
1917 pset_foreach(live, irn) {
1918 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
1919 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1921 spill = set_find_spill(spill_bb->ilp, irn);
1924 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
1925 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
1927 sched_foreach_op(bb, tmp) {
1928 op_t *op = get_irn_link(tmp);
1930 if(is_Phi(tmp)) continue;
1931 assert(!is_Proj(tmp));
1934 ir_node *value = op->attr.remat.remat->value;
1937 /* only collect remats up to the first use of a value */
1938 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
1944 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1945 ir_node *arg = get_irn_n(tmp, i);
1948 /* if a value is used stop collecting remats */
1954 if(cst == ILP_UNDEF) break;
1960 /* if a value is used by a mem-phi, then mem_in of this value is 0 (has to be spilled again into a different slot)
1961 mem_in(phi) -> not mem_in(orig_value) TODO: how does this depend on a certain predecessor?
1964 /* mem_in of mem-phi has associated costs (but first one is free) */
1965 /* define n_mem_copies as positive integer in each predecessor block,
1966 #mem_in into this block from predecessor block - 1 weighted with SPILL_COST*execfreq(predecessor)
1976 * Speicherkopienminimierung: teste Speicherwerte auf Interferenz
1977 * und weise Spillkontexte zu. Sorge bei Phis dafuer, dass gleiche
1978 * Kontexte zusammenfliessen (Operanden und Ergebnis hat gleichen
1985 return fabs(x) < 0.00001;
1990 is_spilled(const spill_ilp_t * si, const live_range_t * lr)
1992 return !is_zero(lpp_get_var_sol(si->lpp, lr->in_mem_var));
1997 is_mem_phi(const ir_node * phi, void *data)
1999 spill_ilp_t *si = data;
2000 // return is_spilled(si, get_use_head(si, phi)->closest_use);
2005 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
2007 spill_ilp_t *si = get_irg_link(current_ir_graph);
2009 if(pset_find_ptr(si->all_possible_remats, n)) {
2010 op_t *op = (op_t*)get_irn_link(n);
2011 assert(op && op->is_remat);
2013 if(!op->attr.remat.remat->inverse) {
2014 if(op->attr.remat.pre) {
2015 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
2017 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
2022 op_t *op = (op_t*)get_irn_link(n);
2023 assert(op && op->is_remat);
2025 if(op->attr.remat.pre) {
2026 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
2028 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
2039 dump_graph_with_remats(ir_graph * irg, const char * suffix)
2041 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
2042 be_dump(irg, suffix, dump_ir_block_graph_sched);
2043 set_dump_node_vcgattr_hook(NULL);
2048 * Edge hook to dump the schedule edges with annotated register pressure.
2051 sched_pressure_edge_hook(FILE *F, ir_node *irn)
2053 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
2054 ir_node *prev = sched_prev(irn);
2055 fprintf(F, "edge:{sourcename:\"");
2057 fprintf(F, "\" targetname:\"");
2059 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
2060 fprintf(F, "\" color:magenta}\n");
2066 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
2068 DUMP_NODE_EDGE_FUNC old = get_dump_node_edge_hook();
2070 dump_consts_local(0);
2071 set_dump_node_edge_hook(sched_pressure_edge_hook);
2072 dump_ir_block_graph(irg, suffix);
2073 set_dump_node_edge_hook(old);
2077 walker_pressure_annotator(ir_node * bb, void * data)
2079 spill_ilp_t *si = data;
2084 pset *live = pset_new_ptr_default();
2087 live_foreach(bb, li) {
2088 irn = (ir_node *) li->irn;
2090 if (live_is_end(li) && has_reg_class(si, irn)) {
2091 pset_insert_ptr(live, irn);
2095 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
2097 sched_foreach_reverse(bb, irn) {
2099 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
2103 if(has_reg_class(si, irn)) {
2104 pset_remove_ptr(live, irn);
2105 if(is_Proj(irn)) ++projs;
2108 if(!is_Proj(irn)) projs = 0;
2110 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
2111 ir_node *arg = get_irn_n(irn, i);
2113 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
2115 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
2122 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
2124 be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
2129 connect_all_remats_with_keep(spill_ilp_t * si)
2137 n_remats = pset_count(si->all_possible_remats);
2139 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
2142 pset_foreach(si->all_possible_remats, irn) {
2147 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
2149 obstack_free(si->obst, ins);
2155 connect_all_spills_with_keep(spill_ilp_t * si)
2164 n_spills = pset_count(si->spills);
2166 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
2169 pset_foreach(si->spills, irn) {
2174 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
2176 obstack_free(si->obst, ins);
2180 /** insert a spill at an arbitrary position */
2181 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert, ir_node *ctx)
2183 ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
2184 ir_graph *irg = get_irn_irg(bl);
2185 ir_node *frame = get_irg_frame(irg);
2189 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
2190 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
2192 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn, ctx);
2195 * search the right insertion point. a spill of a phi cannot be put
2196 * directly after the phi, if there are some phis behind the one which
2197 * is spilled. Also, a spill of a Proj must be after all Projs of the
2200 * Here's one special case:
2201 * If the spill is in the start block, the spill must be after the frame
2202 * pointer is set up. This is done by setting insert to the end of the block
2203 * which is its default initialization (see above).
2206 if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
2209 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
2212 sched_add_after(insert, spill);
2217 delete_remat(spill_ilp_t * si, ir_node * remat) {
2220 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2222 sched_remove(remat);
2224 /* kill links to operands */
2225 for (i = -1, n = get_irn_arity(remat); i < n; ++i) {
2226 set_irn_n(remat, i, bad);
2231 clean_remat_info(spill_ilp_t * si)
2236 remat_info_t *remat_info;
2237 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2239 set_foreach(si->remat_info, remat_info) {
2240 if(!remat_info->remats) continue;
2242 pset_foreach(remat_info->remats, remat)
2244 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
2245 set_irn_n(remat->proj, -1, bad);
2246 set_irn_n(remat->proj, 0, bad);
2249 if(get_irn_n_edges(remat->op) == 0) {
2250 for (i = -1, n = get_irn_arity(remat->op); i < n; ++i) {
2251 set_irn_n(remat->op, i, bad);
2256 if(remat_info->remats) del_pset(remat_info->remats);
2257 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
2262 delete_unnecessary_remats(spill_ilp_t * si)
2267 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2270 ir_node *end = get_irg_end(si->chordal_env->irg);
2273 for (i = 0, n = get_irn_arity(si->keep); i < n; ++i) {
2274 ir_node *keep_arg = get_irn_n(si->keep, i);
2275 op_t *arg_op = get_irn_link(keep_arg);
2278 assert(arg_op->is_remat);
2280 name = si->lpp->vars[arg_op->attr.remat.ilp];
2282 if(is_zero(name->value)) {
2283 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
2284 /* TODO check whether reload is preferred over remat (could be bug) */
2285 delete_remat(si, keep_arg);
2287 if(!arg_op->attr.remat.remat->inverse) {
2288 if(arg_op->attr.remat.pre) {
2289 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
2291 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
2294 if(arg_op->attr.remat.pre) {
2295 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
2297 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
2302 set_irn_n(si->keep, i, bad);
2305 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
2306 ir_node *end_arg = get_End_keepalive(end, i);
2308 if(end_arg != si->keep) {
2309 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
2312 keeps = obstack_finish(si->obst);
2313 set_End_keepalives(end, n-1, keeps);
2314 obstack_free(si->obst, keeps);
2317 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
2322 pset_foreach(si->all_possible_remats, remat) {
2323 op_t *remat_op = get_irn_link(remat);
2324 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
2326 if(is_zero(name->value)) {
2327 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
2328 /* TODO check whether reload is preferred over remat (could be bug) */
2329 delete_remat(si, remat);
2331 if(!remat_op->attr.remat.remat->inverse) {
2332 if(remat_op->attr.remat.pre) {
2333 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
2335 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
2338 if(remat_op->attr.remat.pre) {
2339 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
2341 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
2350 * @param before The node after which the spill will be placed in the schedule
2352 /* TODO set context properly */
2354 insert_spill(spill_ilp_t * si, const ir_node * irn, const ir_node * value, const ir_node * before)
2358 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
2360 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
2362 spill = be_spill2(arch_env, irn, before, irn);
2364 defs = set_insert_def(si->values, value);
2367 /* enter into the linked list */
2368 set_irn_link(spill, defs->spills);
2369 defs->spills = spill;
2371 #ifdef KEEPALIVE_SPILLS
2372 pset_insert_ptr(si->spills, spill);
2379 * @param before The Phi node which has to be spilled
2382 insert_mem_phi(spill_ilp_t * si, const ir_node * phi)
2390 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
2392 for(i=0,n=get_irn_arity(phi); i<n; ++i) {
2393 ins[i] = si->m_unknown;
2396 mem_phi = new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
2398 defs = set_insert_def(si->values, phi);
2401 /* enter into the linked list */
2402 set_irn_link(mem_phi, defs->spills);
2403 defs->spills = mem_phi;
2405 #ifdef KEEPALIVE_SPILLS
2406 pset_insert_ptr(si->spills, mem_phi);
2413 * Add remat to list of defs, destroys link field!
2416 insert_remat(spill_ilp_t * si, ir_node * remat)
2419 op_t *remat_op = get_irn_link(remat);
2421 assert(remat_op->is_remat);
2423 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
2426 /* enter into the linked list */
2427 set_irn_link(remat, defs->remats);
2428 defs->remats = remat;
2433 collect_spills(spill_ilp_t * si, ir_node * value, pset * spills, pset * visited)
2438 defs = set_find_def(si->values, value);
2440 if(defs && defs->spills) {
2441 for(next = defs->spills; next; next = get_irn_link(next)) {
2442 pset_insert_ptr(spills, next);
2444 } else if (is_Phi(value)) {
2446 if(!pset_find_ptr(visited, value)) {
2450 pset_insert_ptr(visited, value);
2451 for(i=0, n=get_irn_arity(value); i<n; ++i) {
2452 ir_node *arg = get_irn_n(value, i);
2454 collect_spills(si, arg, spills, visited);
2458 // assert(0 && "Phi operand not spilled");
2464 get_spills_for_value(spill_ilp_t * si, ir_node * value)
2466 pset *spills = pset_new_ptr_default();
2467 // pset *visited = pset_new_ptr_default();
2469 // collect_spills(si, value, spills, visited);
2470 // del_pset(visited);
2474 defs = set_find_def(si->values, value);
2476 if(defs && defs->spills) {
2477 for(next = defs->spills; next; next = get_irn_link(next)) {
2478 pset_insert_ptr(spills, next);
2486 * Add reload before operation and add to list of defs
2489 insert_reload(spill_ilp_t * si, const ir_node * value, const ir_node * after)
2494 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
2496 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
2498 defs = set_find_def(si->values, value);
2499 /* get a spill of this value */
2501 if((!defs || !defs->spills) && is_Phi(value)) {
2504 spills = get_spills_for_value(si, value);
2506 spill = pset_first(spills);
2510 defs = set_insert_def(si->values, value);
2512 defs->spills = spill;
2513 set_irn_link(spill, NULL);
2515 spill = defs->spills;
2518 spill = defs->spills;
2519 assert(spill && "no spill placed before reload");
2521 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
2523 /* enter into the linked list */
2524 set_irn_link(reload, defs->remats);
2525 defs->remats = reload;
2531 walker_spill_placer(ir_node * bb, void * data) {
2532 spill_ilp_t *si = (spill_ilp_t*)data;
2534 spill_bb_t *spill_bb = get_irn_link(bb);
2535 pset *spills_to_do = pset_new_ptr_default();
2538 set_foreach(spill_bb->ilp, spill) {
2541 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2542 name = si->lpp->vars[spill->mem_in];
2543 if(!is_zero(name->value)) {
2546 mem_phi = insert_mem_phi(si, spill->irn);
2548 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
2552 if(spill->spill != ILP_UNDEF) {
2553 name = si->lpp->vars[spill->spill];
2554 if(!is_zero(name->value)) {
2555 if(spill->reg_in > 0) {
2556 name = si->lpp->vars[spill->reg_in];
2557 if(!is_zero(name->value)) {
2558 insert_spill(si, spill->irn, spill->irn, bb);
2562 pset_insert_ptr(spills_to_do, spill->irn);
2566 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
2569 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
2570 op_t *op = get_irn_link(irn);
2572 if(be_is_Spill(irn)) continue;
2575 /* TODO fix this if we want to support remats with more than two nodes */
2576 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
2577 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
2579 insert_spill(si, irn, op->attr.remat.remat->value, irn);
2582 if(pset_find_ptr(spills_to_do, irn)) {
2583 pset_remove_ptr(spills_to_do, irn);
2585 insert_spill(si, irn, irn, irn);
2591 assert(pset_count(spills_to_do) == 0);
2593 /* afterwards free data in block */
2594 del_pset(spills_to_do);
2598 phim_fixer(spill_ilp_t *si) {
2601 set_foreach(si->values, defs) {
2602 const ir_node *phi = defs->value;
2603 ir_node *phi_m = defs->spills;
2607 if(!is_Phi(phi)) continue;
2608 if(!phi_m || !is_Phi(phi_m) || get_irn_mode(phi_m) != mode_M) continue;
2610 for(i=0,n=get_irn_arity(phi); i<n; ++i) {
2611 const ir_node *value = get_irn_n(phi, i);
2612 defs_t *val_defs = set_find_def(si->values, value);
2614 /* get a spill of this value */
2615 ir_node *spill = val_defs->spills;
2617 assert(spill && "no spill placed before PhiM");
2619 set_irn_n(phi_m, i, spill);
2625 walker_reload_placer(ir_node * bb, void * data) {
2626 spill_ilp_t *si = (spill_ilp_t*)data;
2628 spill_bb_t *spill_bb = get_irn_link(bb);
2632 sched_foreach_reverse(bb, irn) {
2633 op_t *op = get_irn_link(irn);
2635 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
2636 if(is_Phi(irn)) break;
2639 if(get_irn_mode(irn) != mode_T) {
2640 insert_remat(si, irn);
2645 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
2646 ir_node *arg = get_irn_n(irn, i);
2648 if(op->attr.live_range.reloads && op->attr.live_range.reloads[i] != ILP_UNDEF) {
2651 name = si->lpp->vars[op->attr.live_range.reloads[i]];
2652 if(!is_zero(name->value)) {
2654 ir_node *insert_pos = irn;
2655 ir_node *prev = sched_prev(insert_pos);
2656 op_t *prev_op = get_irn_link(prev);
2658 /* insert reload before pre-remats */
2659 while(!sched_is_end(prev) && !be_is_Reload(prev) && !be_is_Spill(prev)
2660 && prev_op->is_remat && prev_op->attr.remat.pre) {
2663 prev = sched_prev(insert_pos);
2664 prev_op = get_irn_link(prev);
2667 reload = insert_reload(si, arg, insert_pos);
2669 set_irn_n(irn, i, reload);
2671 #ifdef KEEPALIVE_RELOADS
2672 pset_insert_ptr(si->spills, reload);
2680 /* reloads at end of block */
2681 if(spill_bb->reloads) {
2683 live_foreach(bb, li) {
2684 ir_node *irn = (ir_node *) li->irn;
2686 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
2689 name = si->lpp->vars[spill_bb->reloads[i]];
2690 if(!is_zero(name->value)) {
2692 ir_node *insert_pos = bb;
2693 ir_node *prev = sched_prev(insert_pos);
2694 op_t *prev_op = get_irn_link(prev);
2696 /* insert reload before pre-remats */
2697 while(!sched_is_end(prev) && !be_is_Reload(prev) && !be_is_Spill(prev)
2698 && prev_op->is_remat && prev_op->attr.remat.pre) {
2701 prev = sched_prev(insert_pos);
2702 prev_op = get_irn_link(prev);
2705 reload = insert_reload(si, irn, insert_pos);
2707 #ifdef KEEPALIVE_RELOADS
2708 pset_insert_ptr(si->spills, reload);
2716 del_set(spill_bb->ilp);
2720 walker_collect_used(ir_node * irn, void * data)
2722 lc_bitset_t *used = data;
2724 lc_bitset_set(used, get_irn_idx(irn));
2728 walker_kill_unused(ir_node * bb, void * data)
2730 lc_bitset_t *used = data;
2731 const ir_node *bad = get_irg_bad(get_irn_irg(bb));
2735 for(irn=sched_first(bb); !sched_is_end(irn);) {
2736 ir_node *next = sched_next(irn);
2740 if(!lc_bitset_is_set(used, get_irn_idx(irn))) {
2741 assert(!be_is_Spill(irn) && !be_is_Reload(irn) && "something is fishy, spill or remat is unused");
2745 set_nodes_block(irn, bad);
2746 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
2747 set_irn_n(irn, i, bad);
2755 kill_all_unused_values_in_schedule(spill_ilp_t * si)
2757 lc_bitset_t *used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
2759 irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, used);
2760 irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, used);
2762 lc_bitset_free(used);
2766 print_irn_pset(pset * p)
2770 pset_foreach(p, irn) {
2771 ir_printf("%+F\n", irn);
2776 rewire_uses(spill_ilp_t * si)
2778 dom_front_info_t *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
2781 /* then fix uses of spills */
2782 set_foreach(si->values, defs) {
2785 ir_node *next = defs->remats;
2788 reloads = pset_new_ptr_default();
2791 if(be_is_Reload(next)) {
2792 pset_insert_ptr(reloads, next);
2796 next = get_irn_link(next);
2799 spills = get_spills_for_value(si, defs->value);
2800 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));
2801 if(pset_count(spills) > 1) {
2802 //assert(pset_count(reloads) > 0);
2803 // print_irn_pset(spills);
2804 // print_irn_pset(reloads);
2806 be_ssa_constr_set(dfi, spills);
2813 /* first fix uses of remats and reloads */
2814 set_foreach(si->values, defs) {
2816 ir_node *next = defs->remats;
2819 nodes = pset_new_ptr_default();
2820 pset_insert_ptr(nodes, defs->value);
2823 pset_insert_ptr(nodes, next);
2824 next = get_irn_link(next);
2827 if(pset_count(nodes) > 1) {
2828 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
2829 be_ssa_constr_set(dfi, nodes);
2836 // remove_unused_defs(si);
2838 be_free_dominance_frontiers(dfi);
2842 writeback_results(spill_ilp_t * si)
2844 /* walk through the graph and collect all spills, reloads and remats for a value */
2846 si->values = new_set(cmp_defs, 4096);
2848 DBG((si->dbg, LEVEL_1, "Applying results\n"));
2849 delete_unnecessary_remats(si);
2850 si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
2851 irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
2853 irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
2855 /* clean the remat info! there are still back-edges leading there! */
2856 clean_remat_info(si);
2860 connect_all_spills_with_keep(si);
2862 del_set(si->values);
2866 get_n_regs(spill_ilp_t * si)
2868 int arch_n_regs = arch_register_class_n_regs(si->cls);
2872 for(i=0; i<arch_n_regs; i++) {
2873 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
2878 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
2883 walker_reload_mover(ir_node * bb, void * data)
2885 spill_ilp_t *si = data;
2888 sched_foreach(bb, tmp) {
2889 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
2890 ir_node *reload = tmp;
2893 /* move reload upwards */
2895 int pressure = (int)get_irn_link(reload);
2896 if(pressure < si->n_regs) {
2897 irn = sched_prev(reload);
2898 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
2899 sched_remove(reload);
2900 pressure = (int)get_irn_link(irn);
2902 while(pressure < si->n_regs) {
2903 if(sched_is_end(irn) || (be_is_Reload(irn) && has_reg_class(si, irn))) break;
2905 set_irn_link(irn, INT_TO_PTR(pressure+1));
2906 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
2907 irn = sched_prev(irn);
2909 pressure = (int)get_irn_link(irn);
2912 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
2913 sched_put_after(irn, reload);
2920 move_reloads_upward(spill_ilp_t * si)
2922 irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
2926 be_spill_remat(const be_chordal_env_t * chordal_env)
2928 char problem_name[256];
2929 char dump_suffix[256];
2930 char dump_suffix2[256];
2931 char dump_suffix3[256];
2932 struct obstack obst;
2935 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
2936 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
2937 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
2938 ir_snprintf(dump_suffix3, sizeof(dump_suffix3), "-%s-reloads_moved", chordal_env->cls->name);
2940 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
2941 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
2943 obstack_init(&obst);
2944 si.chordal_env = chordal_env;
2946 si.senv = be_new_spill_env(chordal_env, is_mem_phi, &si);
2947 si.cls = chordal_env->cls;
2948 si.lpp = new_lpp(problem_name, lpp_minimize);
2949 si.remat_info = new_set(cmp_remat_info, 4096);
2950 si.all_possible_remats = pset_new_ptr_default();
2951 si.spills = pset_new_ptr_default();
2952 si.inverse_ops = pset_new_ptr_default();
2953 #ifndef EXECFREQ_LOOPDEPH
2954 si.execfreqs = compute_execfreq(chordal_env->irg);
2956 si.execfreqs = NULL;
2961 si.n_regs = get_n_regs(&si);
2963 set_irg_link(chordal_env->irg, &si);
2964 compute_doms(chordal_env->irg);
2966 #ifdef COLLECT_REMATS
2967 /* collect remats */
2968 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
2969 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
2972 /* insert possible remats */
2973 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
2974 irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
2975 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
2978 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
2979 connect_all_remats_with_keep(&si);
2980 /* dump graph with inserted remats */
2981 dump_graph_with_remats(chordal_env->irg, dump_suffix);
2985 /* recompute liveness */
2986 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
2987 be_liveness(chordal_env->irg);
2991 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
2992 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
2993 irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
2995 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
2996 irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
3003 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
3004 if ((f = fopen(buf, "wt")) != NULL) {
3005 lpp_dump_plain(si.lpp, f);
3012 DBG((si.dbg, LEVEL_1, "\tSolving %F\n", chordal_env->irg));
3013 lpp_set_time_limit(si.lpp, ILP_TIMEOUT);
3016 lpp_solve_cplex(si.lpp);
3018 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
3020 assert(lpp_is_sol_valid(si.lpp)
3021 && "solution of ILP must be valid");
3023 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));
3025 #ifdef DUMP_SOLUTION
3030 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
3031 if ((f = fopen(buf, "wt")) != NULL) {
3033 for (i = 0; i < si.lpp->var_next; ++i) {
3034 lpp_name_t *name = si.lpp->vars[i];
3035 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
3042 writeback_results(&si);
3046 kill_all_unused_values_in_schedule(&si);
3048 #if defined(KEEPALIVE_SPILLS) || defined(KEEPALIVE_RELOADS)
3049 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
3052 be_liveness(chordal_env->irg);
3053 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
3055 dump_pressure_graph(&si, dump_suffix2);
3057 // TODO fix temporarily exceeded regpressure due to remat2s
3059 // TODO insert copys to fix interferences in memory
3061 // move reloads upwards
3062 move_reloads_upward(&si);
3063 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
3064 dump_pressure_graph(&si, dump_suffix3);
3066 free_dom(chordal_env->irg);
3067 del_pset(si.inverse_ops);
3068 del_pset(si.all_possible_remats);
3069 del_pset(si.spills);
3070 #ifndef EXECFREQ_LOOPDEPH
3071 free_execfreq(si.execfreqs);
3074 obstack_free(&obst, NULL);
3078 #else /* WITH_ILP */
3081 only_that_you_can_compile_without_WITH_ILP_defined(void)
3085 #endif /* WITH_ILP */