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"
56 //#define DUMP_SOLUTION
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 ONLY_BRIGGS_REMATS /* only remats without parameters (or only with ignored params) */
62 #define REMAT_WHILE_LIVE /* only remat values that are live */
63 //#define NO_ENLARGE_L1V3N355 /* do not remat after the death of some operand */
64 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
65 #define MAY_DIE_AT_REMAT /* allow values to die after a pre remat */
66 #define NO_SINGLE_USE_REMATS /* do not repair schedule */
67 //#define KEEPALIVE_SPILLS
68 //#define KEEPALIVE_RELOADS
69 #define GOODWIN_REDUCTION
70 //#define NO_MEMCOPIES
71 //#define VERIFY_DOMINANCE
72 #define WITH_MEMOPERANDS
76 #define LPP_SERVER "i44pc52"
77 #define LPP_SOLVER "cplex"
80 #define COST_MEMOPERAND 7
84 #define ILP_TIMEOUT 300
88 typedef struct _spill_ilp_t {
89 const arch_register_class_t *cls;
91 const be_chordal_env_t *chordal_env;
96 pset *all_possible_remats;
101 set *values; /**< for collecting all definitions of values before running ssa-construction */
105 #ifdef WITH_MEMOPERANDS
108 DEBUG_ONLY(firm_dbg_module_t * dbg);
111 typedef int ilp_var_t;
112 typedef int ilp_cst_t;
114 typedef struct _spill_bb_t {
119 typedef struct _remat_t {
120 const ir_node *op; /**< for copy_irn */
121 const ir_node *value; /**< the value which is being recomputed by this remat */
122 const ir_node *proj; /**< not NULL if the above op produces a tuple */
123 int cost; /**< cost of this remat */
124 int inverse; /**< nonzero if this is an inverse remat */
128 * Data to be attached to each IR node. For remats this contains the ilp_var
129 * for this remat and for normal ops this contains the ilp_vars for
130 * reloading each operand
132 typedef struct _op_t {
137 const remat_t *remat; /** the remat this op belongs to */
138 int pre; /** 1, if this is a pressure-increasing remat */
142 ir_node *op; /** the operation this live range belongs to */
151 typedef struct _defs_t {
152 const ir_node *value;
153 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
154 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
157 typedef struct _remat_info_t {
158 const ir_node *irn; /**< the irn to which these remats belong */
159 pset *remats; /**< possible remats for this value */
160 pset *remats_by_operand; /**< remats with this value as operand */
163 typedef struct _keyval_t {
168 typedef struct _spill_t {
177 #ifdef WITH_MEMOPERANDS
178 typedef struct _memoperand_t {
179 ir_node *irn; /**< the irn */
180 unsigned int pos; /**< the position of the argument */
181 ilp_var_t ilp; /**< the ilp var for this memory operand */
186 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
188 return chordal_has_class(si->chordal_env, irn);
193 cmp_remat(const void *a, const void *b)
195 const keyval_t *p = a;
196 const keyval_t *q = b;
197 const remat_t *r = p->val;
198 const remat_t *s = q->val;
202 return !(r == s || r->op == s->op);
206 cmp_remat(const void *a, const void *b)
208 const remat_t *r = a;
209 const remat_t *s = a;
211 return !(r == s || r->op == s->op);
215 cmp_spill(const void *a, const void *b, size_t size)
217 const spill_t *p = a;
218 const spill_t *q = b;
220 // return !(p->irn == q->irn && p->bb == q->bb);
221 return !(p->irn == q->irn);
224 #ifdef WITH_MEMOPERANDS
226 cmp_memoperands(const void *a, const void *b, size_t size)
228 const memoperand_t *p = a;
229 const memoperand_t *q = b;
231 return !(p->irn == q->irn && p->pos == q->pos);
236 set_find_keyval(set * set, const void * key)
241 return set_find(set, &query, sizeof(query), HASH_PTR(key));
245 set_insert_keyval(set * set, void * key, void * val)
251 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
255 set_find_def(set * set, const ir_node * value)
260 return set_find(set, &query, sizeof(query), HASH_PTR(value));
264 set_insert_def(set * set, const ir_node * value)
271 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
274 #ifdef WITH_MEMOPERANDS
275 static memoperand_t *
276 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
283 return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
286 static memoperand_t *
287 set_find_memoperand(set * set, const ir_node * irn, unsigned int pos)
291 query.irn = (ir_node*)irn;
293 return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
299 set_find_spill(set * set, const ir_node * value)
303 query.irn = (ir_node*)value;
304 return set_find(set, &query, sizeof(query), HASH_PTR(value));
307 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
308 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
309 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
310 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
311 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
314 cmp_remat_info(const void *a, const void *b, size_t size)
316 const remat_info_t *p = a;
317 const remat_info_t *q = b;
319 return !(p->irn == q->irn);
323 cmp_defs(const void *a, const void *b, size_t size)
328 return !(p->value == q->value);
332 cmp_keyval(const void *a, const void *b, size_t size)
334 const keyval_t *p = a;
335 const keyval_t *q = b;
337 return !(p->key == q->key);
341 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
344 #ifndef EXECFREQ_LOOPDEPH
345 return get_block_execfreq(si->chordal_env->exec_freq, get_block(irn)) + FUDGE;
348 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
350 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
355 get_cost(const spill_ilp_t * si, const ir_node * irn)
357 if(be_is_Spill(irn)) {
359 } else if(be_is_Reload(irn)){
362 return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
367 * Checks, whether node and its operands have suitable reg classes
370 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
373 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
374 int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
378 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
381 for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
382 ir_node *op = get_irn_n(irn, n);
383 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
386 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
393 * Try to create a remat from @p op with destination value @p dest_value
395 static INLINE remat_t *
396 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
398 remat_t *remat = NULL;
400 // if(!mode_is_datab(get_irn_mode(dest_value)))
403 if(dest_value == op) {
404 const ir_node *proj = NULL;
406 if(is_Proj(dest_value)) {
407 op = get_irn_n(op, 0);
411 if(!is_rematerializable(si, op))
414 remat = obstack_alloc(si->obst, sizeof(*remat));
416 remat->cost = get_cost(si, op);
417 remat->value = dest_value;
421 arch_inverse_t inverse;
424 /* get the index of the operand we want to retrieve by the inverse op */
425 for (n = get_irn_arity(op)-1; n>=0; --n) {
426 ir_node *arg = get_irn_n(op, n);
428 if(arg == dest_value) break;
432 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", n, op));
434 /* else ask the backend to give an inverse op */
435 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
438 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
440 assert(inverse.n > 0 && "inverse op should have at least one node");
442 for(i=inverse.n-1; i>=0; --i) {
443 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
447 remat = obstack_alloc(si->obst, sizeof(*remat));
448 remat->op = inverse.nodes[0];
449 remat->cost = inverse.costs;
450 remat->value = dest_value;
451 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
454 assert(is_Proj(remat->proj));
456 assert(0 && "I can not handle remats with more than 2 nodes");
463 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
465 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
473 add_remat(const spill_ilp_t * si, const remat_t * remat)
475 remat_info_t *remat_info,
480 assert(remat->value);
482 query.irn = remat->value;
484 query.remats_by_operand = NULL;
485 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
487 if(remat_info->remats == NULL) {
488 remat_info->remats = new_pset(cmp_remat, 4096);
490 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
492 /* insert the remat into the remats_be_operand set of each argument of the remat op */
493 for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
494 ir_node *arg = get_irn_n(remat->op, n);
498 query.remats_by_operand = NULL;
499 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
501 if(remat_info->remats_by_operand == NULL) {
502 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
504 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
508 #ifdef NO_SINGLE_USE_REMATS
510 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
512 const ir_edge_t *edge = get_irn_out_edge_first(irn);
516 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
519 edge = get_irn_out_edge_next(irn, edge);
526 #ifdef ONLY_BRIGGS_REMATS
528 get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
531 unsigned int ret = 0;
533 for(n=get_irn_arity(irn)-1; n>=0; --n) {
534 if(has_reg_class(si, irn)) ++ret;
542 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
547 if( has_reg_class(si, op)
548 #ifdef NO_SINGLE_USE_REMATS
549 && (get_irn_n_nonremat_edges(si, op) > 1)
551 #ifdef ONLY_BRIGGS_REMATS
552 && (get_irn_n_nonignore_args(si, op) == 0)
555 remat = get_remat_from_op(si, op, op);
557 add_remat(si, remat);
561 #if defined(COLLECT_INVERSE_REMATS) && !defined(ONLY_BRIGGS_REMATS)
562 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
564 for (n = get_irn_arity(op)-1; n>=0; --n) {
565 ir_node *arg = get_irn_n(op, n);
567 if(has_reg_class(si, arg)) {
568 /* try to get an inverse remat */
569 remat = get_remat_from_op(si, arg, op);
571 add_remat(si, remat);
580 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
583 ir_node *def_block = get_nodes_block(val);
589 /* if pos is at end of a basic block */
591 ret = (pos == def_block || block_dominates(def_block, pos));
592 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
596 /* else if this is a normal operation */
597 block = get_nodes_block(pos);
598 if(block == def_block) {
599 if(!sched_is_scheduled(val)) return 1;
601 ret = sched_comes_after(val, pos);
602 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
606 ret = block_dominates(def_block, block);
607 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
611 static INLINE ir_node *
612 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
614 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
618 * Returns first non-Phi node of block @p bb
620 static INLINE ir_node *
621 sched_block_first_nonphi(const ir_node * bb)
623 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
627 sched_skip_proj_predicator(const ir_node * irn, void * data)
629 return (is_Proj(irn));
632 static INLINE ir_node *
633 sched_next_nonproj(const ir_node * irn, int forward)
635 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
639 * Returns next operation node (non-Proj) after @p irn
640 * or the basic block of this node
642 static INLINE ir_node *
643 sched_next_op(const ir_node * irn)
645 ir_node *next = sched_next(irn);
650 return sched_next_nonproj(next, 1);
654 * Returns previous operation node (non-Proj) before @p irn
655 * or the basic block of this node
657 static INLINE ir_node *
658 sched_prev_op(const ir_node * irn)
660 ir_node *prev = sched_prev(irn);
665 return sched_next_nonproj(prev, 0);
669 sched_put_after(ir_node * insert, ir_node * irn)
671 if(is_Block(insert)) {
672 insert = sched_block_first_nonphi(insert);
674 insert = sched_next_op(insert);
676 sched_add_before(insert, irn);
680 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
682 if(is_Block(insert)) {
683 insert = sched_block_last_noncf(si, insert);
685 insert = sched_next_nonproj(insert, 0);
686 insert = sched_prev(insert);
688 sched_add_after(insert, irn);
692 * Tells you whether a @p remat can be placed before the irn @p pos
695 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
697 const ir_node *op = remat->op;
703 prev = sched_block_last_noncf(si, pos);
704 prev = sched_next_nonproj(prev, 0);
706 prev = sched_prev_op(pos);
708 /* do not remat if the rematted value is defined immediately before this op */
709 if(prev == remat->op) {
714 /* this should be just fine, the following OP will be using this value, right? */
716 /* only remat AFTER the real definition of a value (?) */
717 if(!value_is_defined_before(si, pos, remat->value)) {
718 // ir_fprintf(stderr, "error(not defined)");
723 for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
724 const ir_node *arg = get_irn_n(op, n);
726 #ifdef NO_ENLARGE_L1V3N355
727 if(has_reg_class(si, arg) && live) {
728 res &= pset_find_ptr(live, arg)?1:0;
730 res &= value_is_defined_before(si, pos, arg);
733 res &= value_is_defined_before(si, pos, arg);
741 * Tells you whether a @p remat can be placed after the irn @p pos
744 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
747 pos = sched_block_first_nonphi(pos);
749 pos = sched_next_op(pos);
752 /* only remat AFTER the real definition of a value (?) */
753 if(!value_is_defined_before(si, pos, remat->value)) {
757 return can_remat_before(si, remat, pos, live);
761 * Collect potetially rematerializable OPs
764 walker_remat_collector(ir_node * irn, void * data)
766 spill_ilp_t *si = data;
768 if(!is_Block(irn) && !is_Phi(irn)) {
769 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
770 get_remats_from_op(si, irn);
775 * Inserts a copy of @p irn before @p pos
778 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
783 bb = is_Block(pos)?pos:get_nodes_block(pos);
784 copy = exact_copy(irn);
786 _set_phi_class(copy, NULL);
787 set_nodes_block(copy, bb);
788 sched_put_before(si, pos, copy);
794 * Inserts a copy of @p irn after @p pos
797 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
802 bb = is_Block(pos)?pos:get_nodes_block(pos);
803 copy = exact_copy(irn);
805 _set_phi_class(copy, NULL);
806 set_nodes_block(copy, bb);
807 sched_put_after(pos, copy);
813 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
817 if(can_remat_after(si, remat, pos, live)) {
822 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
824 copy = insert_copy_after(si, remat->op, pos);
826 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
827 op = obstack_alloc(si->obst, sizeof(*op));
829 op->attr.remat.remat = remat;
830 op->attr.remat.pre = 0;
831 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
833 set_irn_link(copy, op);
834 pset_insert_ptr(si->all_possible_remats, copy);
836 proj_copy = insert_copy_after(si, remat->proj, copy);
837 set_irn_n(proj_copy, 0, copy);
838 set_irn_link(proj_copy, op);
839 pset_insert_ptr(si->all_possible_remats, proj_copy);
851 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
855 if(can_remat_before(si, remat, pos, live)) {
860 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
862 copy = insert_copy_before(si, remat->op, pos);
864 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
865 op = obstack_alloc(si->obst, sizeof(*op));
867 op->attr.remat.remat = remat;
868 op->attr.remat.pre = 1;
869 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
871 set_irn_link(copy, op);
872 pset_insert_ptr(si->all_possible_remats, copy);
874 proj_copy = insert_copy_after(si, remat->proj, copy);
875 set_irn_n(proj_copy, 0, copy);
876 set_irn_link(proj_copy, op);
877 pset_insert_ptr(si->all_possible_remats, proj_copy);
889 get_block_n_succs(const ir_node *block) {
890 const ir_edge_t *edge;
892 assert(edges_activated(current_ir_graph));
894 edge = get_block_succ_first(block);
898 edge = get_block_succ_next(block, edge);
903 is_merge_edge(const ir_node * bb)
905 #ifdef GOODWIN_REDUCTION
906 return get_block_n_succs(bb) == 1;
913 is_diverge_edge(const ir_node * bb)
915 #ifdef GOODWIN_REDUCTION
916 return get_Block_n_cfgpreds(bb) == 1;
923 walker_regclass_copy_insertor(ir_node * irn, void * data)
925 spill_ilp_t *si = data;
927 if(is_Phi(irn) && has_reg_class(si, irn)) {
930 for(n=get_irn_arity(irn)-1; n>=0; --n) {
931 ir_node *phi_arg = get_irn_n(irn, n);
932 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
934 if(!has_reg_class(si, phi_arg)) {
935 ir_node *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
936 ir_node *pos = sched_block_last_noncf(si, bb);
937 op_t *op = obstack_alloc(si->obst, sizeof(*op));
939 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
940 sched_add_after(pos, copy);
941 set_irn_n(irn, n, copy);
944 op->attr.live_range.args.reloads = NULL;
945 op->attr.live_range.ilp = ILP_UNDEF;
946 set_irn_link(copy, op);
954 * Insert (so far unused) remats into the irg to
955 * recompute the potential liveness of all values
958 walker_remat_insertor(ir_node * bb, void * data)
960 spill_ilp_t *si = data;
961 spill_bb_t *spill_bb;
964 pset *live = pset_new_ptr_default();
966 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
968 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
969 ir_node *value = be_lv_get_irn(si->lv, bb, i);
971 /* add remats at end of block */
972 if (has_reg_class(si, value)) {
973 pset_insert_ptr(live, value);
977 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
978 set_irn_link(bb, spill_bb);
980 irn = sched_last(bb);
981 while(!sched_is_end(irn)) {
988 next = sched_prev(irn);
990 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
992 if(is_Phi(irn) || is_Proj(irn)) {
995 if(has_reg_class(si, irn)) {
996 pset_remove_ptr(live, irn);
999 op = obstack_alloc(si->obst, sizeof(*op));
1001 op->attr.live_range.args.reloads = NULL;
1002 op->attr.live_range.ilp = ILP_UNDEF;
1003 set_irn_link(irn, op);
1009 op = obstack_alloc(si->obst, sizeof(*op));
1011 op->attr.live_range.ilp = ILP_UNDEF;
1012 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1013 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1014 set_irn_link(irn, op);
1016 args = pset_new_ptr_default();
1018 /* collect arguments of op */
1019 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1020 ir_node *arg = get_irn_n(irn, n);
1022 pset_insert_ptr(args, arg);
1025 /* set args of op already live in epilog */
1026 pset_foreach(args, arg) {
1027 if(has_reg_class(si, arg)) {
1028 pset_insert_ptr(live, arg);
1031 /* delete defined value from live set */
1032 if(has_reg_class(si, irn)) {
1033 pset_remove_ptr(live, irn);
1037 remat_args = pset_new_ptr_default();
1039 /* insert all possible remats before irn */
1040 pset_foreach(args, arg) {
1041 remat_info_t *remat_info,
1045 /* continue if the operand has the wrong reg class
1047 if(!has_reg_class(si, arg))
1051 query.remats = NULL;
1052 query.remats_by_operand = NULL;
1053 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1059 if(remat_info->remats) {
1060 pset_foreach(remat_info->remats, remat) {
1061 ir_node *remat_irn = NULL;
1063 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
1064 #ifdef REMAT_WHILE_LIVE
1065 if(pset_find_ptr(live, remat->value)) {
1066 remat_irn = insert_remat_before(si, remat, irn, live);
1069 remat_irn = insert_remat_before(si, remat, irn, live);
1072 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1073 ir_node *remat_arg = get_irn_n(remat_irn, n);
1075 if(!has_reg_class(si, remat_arg)) continue;
1077 pset_insert_ptr(remat_args, remat_arg);
1084 /* now we add remat args to op's args because they could also die at this op */
1085 pset_foreach(args,arg) {
1086 if(pset_find_ptr(remat_args, arg)) {
1087 pset_remove_ptr(remat_args, arg);
1090 pset_foreach(remat_args,arg) {
1091 pset_insert_ptr(args, arg);
1094 /* insert all possible remats after irn */
1095 pset_foreach(args, arg) {
1096 remat_info_t *remat_info,
1100 /* continue if the operand has the wrong reg class */
1101 if(!has_reg_class(si, arg))
1105 query.remats = NULL;
1106 query.remats_by_operand = NULL;
1107 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1113 /* do not place post remats after jumps */
1114 if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
1116 if(remat_info->remats_by_operand) {
1117 pset_foreach(remat_info->remats_by_operand, remat) {
1118 /* do not insert remats producing the same value as one of the operands */
1119 if(!pset_find_ptr(args, remat->value)) {
1120 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
1121 #ifdef REMAT_WHILE_LIVE
1122 if(pset_find_ptr(live, remat->value)) {
1123 insert_remat_after(si, remat, irn, live);
1126 insert_remat_after(si, remat, irn, live);
1133 del_pset(remat_args);
1138 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_in, i) {
1139 ir_node *value = be_lv_get_irn(si->lv, bb, i);
1141 /* add remats at end if successor has multiple predecessors */
1142 if(is_merge_edge(bb)) {
1143 /* add remats at end of block */
1144 if (be_is_live_end(si->lv, bb, value) && has_reg_class(si, value)) {
1145 remat_info_t *remat_info,
1150 query.remats = NULL;
1151 query.remats_by_operand = NULL;
1152 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1154 if(remat_info && remat_info->remats) {
1155 pset_foreach(remat_info->remats, remat) {
1156 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1158 insert_remat_before(si, remat, bb, NULL);
1163 if(is_diverge_edge(bb)) {
1164 /* add remat2s at beginning of block */
1165 if ((be_is_live_in(si->lv, bb, value) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) {
1166 remat_info_t *remat_info,
1171 query.remats = NULL;
1172 query.remats_by_operand = NULL;
1173 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1175 if(remat_info && remat_info->remats) {
1176 pset_foreach(remat_info->remats, remat) {
1177 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at beginning of block %+F\n", remat->op, bb));
1179 /* put the remat here if all its args are available */
1180 insert_remat_after(si, remat, bb, NULL);
1190 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1193 luke_endwalker(ir_node * bb, void * data)
1195 spill_ilp_t *si = (spill_ilp_t*)data;
1201 spill_bb_t *spill_bb = get_irn_link(bb);
1205 live = pset_new_ptr_default();
1206 use_end = pset_new_ptr_default();
1208 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1209 irn = be_lv_get_irn(si->lv, bb, i);
1210 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1213 pset_insert_ptr(live, irn);
1214 op = get_irn_link(irn);
1215 assert(!op->is_remat);
1219 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1220 /* their reg_out must always be set */
1221 sched_foreach_reverse(bb, irn) {
1224 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1226 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1227 ir_node *irn_arg = get_irn_n(irn, n);
1229 if(has_reg_class(si, irn_arg)) {
1230 pset_insert_ptr(use_end, irn_arg);
1235 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1236 //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1237 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1239 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1241 /* if this is a merge edge we can reload at the end of this block */
1242 if(is_merge_edge(bb)) {
1243 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1244 } else if(pset_count(use_end)){
1245 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1247 spill_bb->reloads = NULL;
1250 pset_foreach(live,irn) {
1256 /* handle values used by control flow nodes later separately */
1257 if(pset_find_ptr(use_end, irn)) continue;
1260 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1262 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1264 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1265 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1266 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1268 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1269 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1271 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1272 /* by default spill value right after definition */
1273 be_is_live_in(si->lv, bb, irn) {
1274 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, 0.0);
1276 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, 1.0);
1279 if(is_merge_edge(bb)) {
1283 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1284 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb), 0.0);
1285 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1287 /* reload <= mem_out */
1288 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1289 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1290 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1293 spill->reg_in = ILP_UNDEF;
1294 spill->mem_in = ILP_UNDEF;
1297 pset_foreach(use_end,irn) {
1301 ilp_cst_t end_use_req,
1306 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1308 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1310 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1311 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1312 /* if irn is used at the end of the block, then it is live anyway */
1313 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1315 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1316 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1318 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1319 be_is_live_in(si->lv, bb, irn) {
1320 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, 0.0);
1322 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, 1.0);
1325 /* reload for use be control flow op */
1326 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1327 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb), 1.0);
1328 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1330 /* reload <= mem_out */
1331 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1332 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1333 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1335 spill->reg_in = ILP_UNDEF;
1336 spill->mem_in = ILP_UNDEF;
1338 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1339 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1340 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1348 next_post_remat(const ir_node * irn)
1354 next = sched_block_first_nonphi(irn);
1356 next = sched_next_op(irn);
1359 if(sched_is_end(next))
1362 op = get_irn_link(next);
1363 if(op->is_remat && !op->attr.remat.pre) {
1372 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1378 ret = sched_block_last_noncf(si, irn);
1379 ret = sched_next(ret);
1380 ret = sched_prev_op(ret);
1382 ret = sched_prev_op(irn);
1385 if(sched_is_end(ret) || is_Phi(ret))
1388 op = (op_t*)get_irn_link(ret);
1389 if(op->is_remat && op->attr.remat.pre) {
1397 * Find a remat of value @p value in the epilog of @p pos
1400 find_post_remat(const ir_node * value, const ir_node * pos)
1402 while((pos = next_post_remat(pos)) != NULL) {
1405 op = get_irn_link(pos);
1406 assert(op->is_remat && !op->attr.remat.pre);
1408 if(op->attr.remat.remat->value == value)
1409 return (ir_node*)pos;
1412 const ir_edge_t *edge;
1413 foreach_out_edge(pos, edge) {
1414 ir_node *proj = get_edge_src_irn(edge);
1415 assert(is_Proj(proj));
1425 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1427 spill_bb_t *spill_bb = get_irn_link(bb);
1433 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1435 double spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1437 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1439 spill->reg_out = ILP_UNDEF;
1440 spill->reg_in = ILP_UNDEF;
1441 spill->mem_in = ILP_UNDEF;
1443 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1444 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1446 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1447 be_is_live_in(si->lv, bb, irn) {
1448 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, 0.0);
1450 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, 1.0);
1458 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1463 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1464 irn = be_lv_get_irn(si->lv, bb, i);
1466 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1467 pset_insert_ptr(live, irn);
1471 irn = sched_last(bb);
1473 /* all values eaten by control flow operations are also live until the end of the block */
1474 sched_foreach_reverse(bb, irn) {
1477 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1479 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1480 ir_node *arg = get_irn_n(irn,i);
1482 if(has_reg_class(si, arg)) {
1483 pset_insert_ptr(live, arg);
1490 * Inserts ILP-constraints and variables for memory copying before the given position
1493 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1495 const ir_node *succ;
1496 const ir_edge_t *edge;
1497 spill_bb_t *spill_bb = get_irn_link(block);
1506 assert(edges_activated(current_ir_graph));
1508 edge = get_block_succ_first(block);
1514 edge = get_block_succ_next(block, edge);
1515 /* next block can only contain phis, if this is a merge edge */
1518 ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1519 copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1521 ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1522 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1524 pset_foreach(live, tmp) {
1527 op_t *op = get_irn_link(irn);
1528 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1530 spill = set_find_spill(spill_bb->ilp, tmp);
1533 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1535 lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1537 sched_foreach(succ, phi) {
1538 const ir_node *to_copy;
1540 spill_t *to_copy_spill;
1541 op_t *phi_op = get_irn_link(phi);
1542 ilp_var_t reload = ILP_UNDEF;
1545 if(!is_Phi(phi)) break;
1546 if(!has_reg_class(si, phi)) continue;
1548 to_copy = get_irn_n(phi, pos);
1550 to_copy_op = get_irn_link(to_copy);
1552 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1553 assert(to_copy_spill);
1555 if(spill_bb->reloads) {
1556 keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1559 reload = PTR_TO_INT(keyval->val);
1563 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1564 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1566 /* copy - reg_out - reload - remat - live_range <= 0 */
1567 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1568 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1569 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1570 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1571 foreach_pre_remat(si, block, tmp) {
1572 op_t *remat_op = get_irn_link(tmp);
1573 if(remat_op->attr.remat.remat->value == to_copy) {
1574 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1578 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1579 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1581 /* copy - reg_out - copyreg <= 0 */
1582 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1583 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1584 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1590 * Walk all irg blocks and emit this ILP
1593 luke_blockwalker(ir_node * bb, void * data)
1595 spill_ilp_t *si = (spill_ilp_t*)data;
1600 spill_bb_t *spill_bb = get_irn_link(bb);
1603 pset *defs = pset_new_ptr_default();
1604 #ifdef WITH_MEMOPERANDS
1605 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
1609 live = pset_new_ptr_default();
1611 /****************************************
1612 * B A S I C B L O C K E N D
1613 ***************************************/
1616 /* init live values at end of block */
1617 get_live_end(si, bb, live);
1619 pset_foreach(live, irn) {
1621 ilp_var_t reload = ILP_UNDEF;
1623 spill = set_find_spill(spill_bb->ilp, irn);
1626 if(spill_bb->reloads) {
1627 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1630 reload = PTR_TO_INT(keyval->val);
1634 op = get_irn_link(irn);
1635 assert(!op->is_remat);
1637 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1638 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1639 op->attr.live_range.op = bb;
1641 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1642 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1644 /* reg_out - reload - remat - live_range <= 0 */
1645 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1646 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1647 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1648 foreach_pre_remat(si, bb, tmp) {
1649 op_t *remat_op = get_irn_link(tmp);
1650 if(remat_op->attr.remat.remat->value == irn) {
1651 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1654 /* maybe we should also assure that reg_out >= live_range etc. */
1657 #ifndef NO_MEMCOPIES
1658 insert_mem_copy_position(si, live, bb);
1662 * start new live ranges for values used by remats at end of block
1663 * and assure the remat args are available
1665 foreach_pre_remat(si, bb, tmp) {
1666 op_t *remat_op = get_irn_link(tmp);
1669 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1670 ir_node *remat_arg = get_irn_n(tmp, n);
1671 op_t *arg_op = get_irn_link(remat_arg);
1674 if(!has_reg_class(si, remat_arg)) continue;
1676 /* if value is becoming live through use by remat */
1677 if(!pset_find_ptr(live, remat_arg)) {
1678 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1679 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1681 arg_op->attr.live_range.ilp = prev_lr;
1682 arg_op->attr.live_range.op = bb;
1684 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1686 pset_insert_ptr(live, remat_arg);
1687 add_to_spill_bb(si, bb, remat_arg);
1690 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1691 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1692 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1694 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1695 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1697 /* use reload placed for this argument */
1698 if(spill_bb->reloads) {
1699 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1702 ilp_var_t reload = PTR_TO_INT(keyval->val);
1704 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1709 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1714 /**************************************
1715 * B A S I C B L O C K B O D Y
1716 **************************************/
1718 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1724 ilp_cst_t check_pre,
1730 #ifdef WITH_MEMOPERANDS
1731 ilp_cst_t one_memoperand;
1734 /* iterate only until first phi */
1738 op = get_irn_link(irn);
1740 if(op->is_remat) continue;
1741 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1743 /* collect defined values */
1744 if(has_reg_class(si, irn)) {
1745 pset_insert_ptr(defs, irn);
1749 if(is_Proj(irn)) continue;
1752 * init set of irn's arguments
1753 * and all possibly used values around this op
1754 * and values defined by post remats
1756 args = new_set(cmp_keyval, get_irn_arity(irn));
1757 used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1758 remat_defs = pset_new_ptr(pset_count(live));
1760 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1761 ir_node *irn_arg = get_irn_n(irn, n);
1762 if(has_reg_class(si, irn_arg)) {
1763 set_insert_keyval(args, irn_arg, (void*)n);
1764 pset_insert_ptr(used, irn_arg);
1767 foreach_post_remat(irn, tmp) {
1768 op_t *remat_op = get_irn_link(tmp);
1770 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1772 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1773 ir_node *remat_arg = get_irn_n(tmp, n);
1774 if(has_reg_class(si, remat_arg)) {
1775 pset_insert_ptr(used, remat_arg);
1779 foreach_pre_remat(si, irn, tmp) {
1780 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1781 ir_node *remat_arg = get_irn_n(tmp, n);
1782 if(has_reg_class(si, remat_arg)) {
1783 pset_insert_ptr(used, remat_arg);
1788 /**********************************
1789 * I N E P I L O G O F irn
1790 **********************************/
1792 /* ensure each dying value is used by only one post remat */
1793 pset_foreach(used, tmp) {
1794 ir_node *value = tmp;
1795 op_t *value_op = get_irn_link(value);
1800 foreach_post_remat(irn, remat) {
1801 op_t *remat_op = get_irn_link(remat);
1803 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1804 ir_node *remat_arg = get_irn_n(remat, n);
1806 /* if value is used by this remat add it to constraint */
1807 if(remat_arg == value) {
1809 /* sum remat2s <= 1 + n_remats*live_range */
1810 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1811 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
1815 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1821 // value_op->attr.live_range.ilp != ILP_UNDEF
1822 if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
1823 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1827 /* ensure at least one value dies at post remat */
1828 foreach_post_remat(irn, tmp) {
1829 op_t *remat_op = get_irn_link(tmp);
1830 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
1833 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
1834 remat_arg = get_irn_n(tmp, n);
1836 if(has_reg_class(si, remat_arg)) {
1838 /* does arg always die at this op? */
1839 if(!pset_find_ptr(live, remat_arg))
1840 goto skip_one_must_die;
1842 pset_insert_ptr(remat_args, remat_arg);
1846 /* remat + \sum live_range(remat_arg) <= |args| */
1847 ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
1848 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
1849 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1851 pset_foreach(remat_args, remat_arg) {
1852 op_t *arg_op = get_irn_link(remat_arg);
1854 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1858 del_pset(remat_args);
1861 /* new live ranges for values from L\U defined by post remats */
1862 pset_foreach(live, tmp) {
1863 ir_node *value = tmp;
1864 op_t *value_op = get_irn_link(value);
1866 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
1867 ilp_var_t prev_lr = ILP_UNDEF;
1870 if(pset_find_ptr(remat_defs, value)) {
1872 /* next_live_range <= prev_live_range + sum remat2s */
1873 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1874 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1876 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1877 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1879 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1880 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1882 foreach_post_remat(irn, remat) {
1883 op_t *remat_op = get_irn_link(remat);
1885 /* if value is being rematerialized by this remat */
1886 if(value == remat_op->attr.remat.remat->value) {
1887 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1891 value_op->attr.live_range.ilp = prev_lr;
1892 value_op->attr.live_range.op = irn;
1897 /* requirements for post remats and start live ranges from L/U' for values dying here */
1898 foreach_post_remat(irn, tmp) {
1899 op_t *remat_op = get_irn_link(tmp);
1902 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1903 ir_node *remat_arg = get_irn_n(tmp, n);
1904 op_t *arg_op = get_irn_link(remat_arg);
1906 if(!has_reg_class(si, remat_arg)) continue;
1908 /* only for values in L\U (TODO and D?), the others are handled with post_use */
1909 if(!pset_find_ptr(used, remat_arg)) {
1910 /* remat <= live_range(remat_arg) */
1911 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
1912 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1914 /* if value is becoming live through use by remat2 */
1915 if(!pset_find_ptr(live, remat_arg)) {
1918 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1919 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1921 arg_op->attr.live_range.ilp = lr;
1922 arg_op->attr.live_range.op = irn;
1924 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
1926 pset_insert_ptr(live, remat_arg);
1927 add_to_spill_bb(si, bb, remat_arg);
1930 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1931 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1936 d = pset_count(defs);
1937 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
1939 /* count how many regs irn needs for arguments */
1940 u = set_count(args);
1943 /* check the register pressure in the epilog */
1944 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
1945 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
1946 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
1948 /* add L\U' to check_post */
1949 pset_foreach(live, tmp) {
1950 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
1951 /* if a live value is not used by irn */
1952 tmp_op = get_irn_link(tmp);
1953 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
1957 /***********************************************************
1958 * 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
1959 **********************************************************/
1962 pset_foreach(used, tmp) {
1968 op_t *arg_op = get_irn_link(arg);
1971 spill = add_to_spill_bb(si, bb, arg);
1973 /* new live range for each used value */
1974 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
1975 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1977 /* the epilog stuff - including post_use, check_post, check_post_remat */
1978 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
1979 post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1981 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
1983 /* arg is live throughout epilog if the next live_range is in a register */
1984 if(pset_find_ptr(live, arg)) {
1985 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
1987 /* post_use >= next_lr + remat */
1988 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1989 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1990 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
1991 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1995 /* if value is not an arg of op and not possibly defined by post remat
1996 * then it may only die and not become live
1998 if(!set_find_keyval(args, arg)) {
1999 /* post_use <= prev_lr */
2000 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2001 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2002 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2003 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2005 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2006 /* next_lr <= prev_lr */
2007 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2008 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2009 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2010 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2015 /* forall post remat which use arg add a similar cst */
2016 foreach_post_remat(irn, remat) {
2019 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2020 ir_node *remat_arg = get_irn_n(remat, n);
2021 op_t *remat_op = get_irn_link(remat);
2023 if(remat_arg == arg) {
2024 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
2026 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2027 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2028 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2029 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2034 #ifdef WITH_MEMOPERANDS
2035 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2036 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2037 ilp_var_t memoperand;
2039 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2040 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, COST_MEMOPERAND*execution_frequency(si, bb), 0.0);
2041 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2043 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2044 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2046 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2047 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2048 // if(arg_op->attr.live_range.ilp != ILP_UNDEF)
2049 // lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2054 /* new live range begins for each used value */
2055 arg_op->attr.live_range.ilp = prev_lr;
2056 arg_op->attr.live_range.op = irn;
2058 /*if(!pset_find_ptr(live, arg)) {
2059 pset_insert_ptr(live, arg);
2060 add_to_spill_bb(si, bb, arg);
2062 pset_insert_ptr(live, arg);
2066 /* just to be sure */
2067 check_post = ILP_UNDEF;
2076 /* check the register pressure in the prolog */
2077 /* sum_{L\U} lr <= k - |U| */
2078 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2079 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2081 /* for the prolog remove defined values from the live set */
2082 pset_foreach(defs, tmp) {
2083 pset_remove_ptr(live, tmp);
2086 #ifdef WITH_MEMOPERANDS
2087 ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2088 one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2091 /***********************************************************
2092 * 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
2093 **********************************************************/
2096 set_foreach(args, keyval) {
2098 const ir_node *arg = keyval->key;
2099 int i = PTR_TO_INT(keyval->val);
2100 op_t *arg_op = get_irn_link(arg);
2101 ilp_cst_t requirements;
2102 #ifdef WITH_MEMOPERANDS
2106 spill = set_find_spill(spill_bb->ilp, arg);
2109 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2110 op->attr.live_range.args.reloads[i] = lpp_add_var_default(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb), 1.0);
2112 /* reload <= mem_out */
2113 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2114 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2115 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2116 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2118 /* requirement: arg must be in register for use */
2119 /* reload + remat + live_range == 1 */
2120 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2121 requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2123 lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2124 lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2125 foreach_pre_remat(si, irn, tmp) {
2126 op_t *remat_op = get_irn_link(tmp);
2127 if(remat_op->attr.remat.remat->value == arg) {
2128 lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2132 #ifdef WITH_MEMOPERANDS
2134 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2135 if(get_irn_n(irn, n) == arg) {
2139 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2140 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2141 memoperand_t *memoperand;
2142 memoperand = set_find_memoperand(si->memoperands, irn, n);
2144 /* memoperand <= mem_out */
2145 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2146 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2147 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2148 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2150 /* the memoperand is only sufficient if it is used once by the op */
2151 if(n_memoperands == 1)
2152 lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2154 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2156 /* we have one more free register if we use a memory operand */
2157 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2163 /* iterate over L\U */
2164 pset_foreach(live, tmp) {
2165 if(!set_find_keyval(args, tmp)) {
2166 /* if a live value is not used by irn */
2167 tmp_op = get_irn_link(tmp);
2168 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2173 /* requirements for remats */
2174 /* start new live ranges for values used by remats */
2175 foreach_pre_remat(si, irn, tmp) {
2176 op_t *remat_op = get_irn_link(tmp);
2179 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2180 ir_node *remat_arg = get_irn_n(tmp, n);
2181 op_t *arg_op = get_irn_link(remat_arg);
2183 if(!has_reg_class(si, remat_arg)) continue;
2185 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2186 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2187 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2189 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2190 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2192 /* if remat arg is also used by current op then we can use reload placed for this argument */
2193 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2194 int index = (int)keyval->val;
2196 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2204 /*************************
2205 * D O N E W I T H O P
2206 *************************/
2208 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2210 pset_foreach(live, tmp) {
2211 assert(has_reg_class(si, tmp));
2214 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2215 ir_node *arg = get_irn_n(irn, n);
2217 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2220 del_pset(remat_defs);
2224 defs = pset_new_ptr_default();
2229 /***************************************
2230 * B E G I N N I N G O F B L O C K
2231 ***************************************/
2234 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2235 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2237 pset_foreach(live, irn) {
2238 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2241 /* construct mem_outs for all values */
2243 set_foreach(spill_bb->ilp, spill) {
2244 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2245 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2247 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2248 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2250 if(pset_find_ptr(live, spill->irn)) {
2251 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2253 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2254 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2255 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2257 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2259 op_t *op = get_irn_link(spill->irn);
2261 /* do we have to copy a phi argument? */
2262 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2263 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2265 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2266 const ir_node *arg = get_irn_n(spill->irn, n);
2272 /* argument already done? */
2273 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2275 /* get sum of execution frequencies of blocks with the same phi argument */
2276 for(m=n; m>=0; --m) {
2277 const ir_node *arg2 = get_irn_n(spill->irn, m);
2280 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2284 /* copies are not for free */
2285 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2286 var = lpp_add_var_default(si->lpp, buf, lpp_binary, COST_STORE * freq, 1.0);
2288 for(m=n; m>=0; --m) {
2289 const ir_node *arg2 = get_irn_n(spill->irn, m);
2292 op->attr.live_range.args.copies[m] = var;
2296 /* copy <= mem_in */
2297 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2298 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2299 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2300 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2307 /* L\U is empty at bb start */
2308 /* arg is live throughout epilog if it is reg_in into this block */
2310 /* check the register pressure at the beginning of the block
2313 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2314 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2316 pset_foreach(live, irn) {
2319 spill = set_find_spill(spill_bb->ilp, irn);
2322 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2323 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2325 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2327 /* spill + mem_in <= 1 */
2328 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2329 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2331 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2332 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2335 foreach_post_remat(bb, irn) {
2336 op_t *remat_op = get_irn_link(irn);
2338 DBG((si->dbg, LEVEL_4, "\t next post remat: %+F\n", irn));
2339 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
2341 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2344 /* forall post remats add requirements */
2345 foreach_post_remat(bb, tmp) {
2348 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2349 ir_node *remat_arg = get_irn_n(tmp, n);
2350 op_t *remat_op = get_irn_link(tmp);
2352 if(!has_reg_class(si, remat_arg)) continue;
2354 spill = set_find_spill(spill_bb->ilp, remat_arg);
2357 /* remat <= reg_in_argument */
2358 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2359 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2360 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2361 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2365 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2366 pset_foreach(live, irn) {
2370 spill = set_find_spill(spill_bb->ilp, irn);
2371 assert(spill && spill->irn == irn);
2373 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2374 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2377 ir_node *phi_arg = get_Phi_pred(irn, n);
2378 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2379 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2381 op_t *op = get_irn_link(irn);
2383 /* although the phi is in the right regclass one or more of
2384 * its arguments can be in a different one or at least to
2387 if(has_reg_class(si, phi_arg)) {
2388 /* mem_in < mem_out_arg + copy */
2389 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2390 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2392 /* reg_in < reg_out_arg */
2393 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2394 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2396 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2397 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2399 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2402 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2403 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2404 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2408 /* else assure the value arrives on all paths in the same resource */
2410 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2413 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2414 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2417 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2418 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2419 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2420 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2422 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2423 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2425 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2428 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2429 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2434 /* first live ranges from reg_ins */
2435 pset_foreach(live, irn) {
2436 op_t *op = get_irn_link(irn);
2438 spill = set_find_spill(spill_bb->ilp, irn);
2439 assert(spill && spill->irn == irn);
2441 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2442 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2443 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2444 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2446 foreach_post_remat(bb, tmp) {
2447 op_t *remat_op = get_irn_link(tmp);
2449 if(remat_op->attr.remat.remat->value == irn) {
2450 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2455 /* walk forward now and compute constraints for placing spills */
2456 /* this must only be done for values that are not defined in this block */
2457 /* TODO are these values at start of block? if yes, just check whether this is a diverge edge and skip the loop */
2458 pset_foreach(live, irn) {
2460 * if value is defined in this block we can anways place the spill directly after the def
2461 * -> no constraint necessary
2463 if(!is_Phi(irn) && get_nodes_block(irn) == bb) continue;
2466 spill = set_find_spill(spill_bb->ilp, irn);
2469 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2470 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2472 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2473 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2476 sched_foreach_op(bb, tmp) {
2477 op_t *op = get_irn_link(tmp);
2479 if(is_Phi(tmp)) continue;
2480 assert(!is_Proj(tmp));
2483 const ir_node *value = op->attr.remat.remat->value;
2486 /* only collect remats up to the first real use of a value */
2487 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2492 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2493 ir_node *arg = get_irn_n(tmp, n);
2496 /* if a value is used stop collecting remats */
2509 typedef struct _irnlist_t {
2510 struct list_head list;
2514 typedef struct _interference_t {
2515 struct list_head blocklist;
2521 cmp_interference(const void *a, const void *b, size_t size)
2523 const interference_t *p = a;
2524 const interference_t *q = b;
2526 return !(p->a == q->a && p->b == q->b);
2529 static interference_t *
2530 set_find_interference(set * set, ir_node * a, ir_node * b)
2532 interference_t query;
2534 query.a = (a>b)?a:b;
2535 query.b = (a>b)?b:a;
2537 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2540 static interference_t *
2541 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2543 interference_t query,
2545 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2549 result = set_find_interference(set, a, b);
2552 list_add(&list->list, &result->blocklist);
2556 query.a = (a>b)?a:b;
2557 query.b = (a>b)?b:a;
2559 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2561 INIT_LIST_HEAD(&result->blocklist);
2562 list_add(&list->list, &result->blocklist);
2568 values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
2570 const ir_edge_t *edge;
2572 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2573 /* both values are live in, so they interfere */
2577 /* ensure a dominates b */
2578 if(value_dominates(b,a)) {
2584 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2587 /* the following code is stolen from bera.c */
2588 if(be_is_live_end(si->lv, bb, a))
2591 foreach_out_edge(a, edge) {
2592 const ir_node *user = edge->src;
2593 if(get_nodes_block(user) == bb
2596 && value_dominates(b, user))
2604 * Walk all irg blocks and collect interfering values inside of phi classes
2607 luke_interferencewalker(ir_node * bb, void * data)
2609 spill_ilp_t *si = (spill_ilp_t*)data;
2612 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2613 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
2614 op_t *a_op = get_irn_link(a);
2617 /* a is only interesting if it is in my register class and if it is inside a phi class */
2618 if (has_reg_class(si, a) && get_phi_class(a)) {
2622 for(l2=_be_lv_next_irn(si->lv, bb, 0xff, l1+1); l2>=0; l2=_be_lv_next_irn(si->lv, bb, 0xff, l2+1)) {
2623 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
2624 op_t *b_op = get_irn_link(b);
2627 /* a and b are only interesting if they are in the same phi class */
2628 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2632 if(values_interfere_in_block(si, bb, a, b)) {
2633 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2634 set_insert_interference(si, si->interferences, a, b, bb);
2642 static unsigned int copy_path_id = 0;
2645 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2652 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2653 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2655 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2657 pset_foreach(copies, ptr) {
2658 copy = PTR_TO_INT(ptr);
2659 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2664 * @parameter copies contains a path of copies which lead us to irn
2665 * @parameter visited contains a set of nodes already visited on this path
2668 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2670 const ir_edge_t *edge;
2671 op_t *op = get_irn_link(irn);
2672 pset *visited_users = pset_new_ptr_default();
2675 if(op->is_remat) return 0;
2677 pset_insert_ptr(visited, irn);
2681 pset *visited_operands = pset_new_ptr(get_irn_arity(irn));
2683 /* visit all operands */
2684 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2685 ir_node *arg = get_irn_n(irn, n);
2686 ilp_var_t copy = op->attr.live_range.args.copies[n];
2688 if(!has_reg_class(si, arg)) continue;
2689 if(pset_find_ptr(visited_operands, arg)) continue;
2690 pset_insert_ptr(visited_operands, arg);
2693 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2694 del_pset(visited_operands);
2695 del_pset(visited_users);
2696 pset_remove_ptr(visited, irn);
2699 pset_insert(copies, INT_TO_PTR(copy), copy);
2700 write_copy_path_cst(si, copies, any_interfere);
2701 pset_remove(copies, INT_TO_PTR(copy), copy);
2702 } else if(!pset_find_ptr(visited, arg)) {
2703 pset_insert(copies, INT_TO_PTR(copy), copy);
2704 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
2705 pset_remove(copies, INT_TO_PTR(copy), copy);
2707 /*if(paths > MAX_PATHS) {
2708 if(pset_count(copies) == 0) {
2712 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2713 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2714 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2715 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2716 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2720 del_pset(visited_operands);
2721 del_pset(visited_users);
2722 pset_remove_ptr(visited, irn);
2725 } else if(pset_count(copies) == 0) {
2731 del_pset(visited_operands);
2734 /* visit all uses which are phis */
2735 foreach_out_edge(irn, edge) {
2736 ir_node *user = edge->src;
2737 int pos = edge->pos;
2738 op_t *op = get_irn_link(user);
2741 if(!is_Phi(user)) continue;
2742 if(!has_reg_class(si, user)) continue;
2743 if(pset_find_ptr(visited_users, user)) continue;
2744 pset_insert_ptr(visited_users, user);
2746 copy = op->attr.live_range.args.copies[pos];
2748 if(user == target) {
2749 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2750 del_pset(visited_users);
2751 pset_remove_ptr(visited, irn);
2754 pset_insert(copies, INT_TO_PTR(copy), copy);
2755 write_copy_path_cst(si, copies, any_interfere);
2756 pset_remove(copies, INT_TO_PTR(copy), copy);
2757 } else if(!pset_find_ptr(visited, user)) {
2758 pset_insert(copies, INT_TO_PTR(copy), copy);
2759 paths += find_copy_path(si, user, target, any_interfere, copies, visited);
2760 pset_remove(copies, INT_TO_PTR(copy), copy);
2762 /*if(paths > MAX_PATHS) {
2763 if(pset_count(copies) == 0) {
2767 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2768 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2769 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2770 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2771 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2775 del_pset(visited_users);
2776 pset_remove_ptr(visited, irn);
2779 } else if(pset_count(copies) == 0) {
2785 del_pset(visited_users);
2786 pset_remove_ptr(visited, irn);
2791 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
2793 pset * copies = pset_new_ptr_default();
2794 pset * visited = pset_new_ptr_default();
2796 find_copy_path(si, a, b, any_interfere, copies, visited);
2804 memcopyhandler(spill_ilp_t * si)
2806 interference_t *interference;
2808 /* teste Speicherwerte auf Interferenz */
2810 /* analyze phi classes */
2811 phi_class_compute(si->chordal_env->irg);
2813 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
2814 irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
2816 /* now lets emit the ILP unequations for the crap */
2817 set_foreach(si->interferences, interference) {
2819 ilp_var_t interfere,
2821 ilp_cst_t any_interfere_cst,
2823 const ir_node *a = interference->a;
2824 const ir_node *b = interference->b;
2826 /* any_interf <= \sum interf */
2827 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
2828 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2829 any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2831 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
2833 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
2834 const ir_node *bb = irnlist->irn;
2835 spill_bb_t *spill_bb = get_irn_link(bb);
2840 spilla = set_find_spill(spill_bb->ilp, a);
2843 spillb = set_find_spill(spill_bb->ilp, b);
2846 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
2847 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
2848 /* 2: - mem_in_a - spill_a + interfere <= 0 */
2849 /* 3: - mem_in_b - spill_b + interfere <= 0 */
2850 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
2851 interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2853 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
2854 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2856 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
2857 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
2858 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
2859 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
2860 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
2862 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
2863 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2865 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2866 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
2867 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
2869 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
2870 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2872 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2873 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
2874 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
2877 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
2879 /* any_interfere >= interf */
2880 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
2881 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2883 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2884 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2887 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
2888 gen_copy_constraints(si,a,b,any_interfere);
2896 return fabs(x) < 0.00001;
2900 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
2902 spill_ilp_t *si = get_irg_link(current_ir_graph);
2904 if(pset_find_ptr(si->all_possible_remats, n)) {
2905 op_t *op = (op_t*)get_irn_link(n);
2906 assert(op && op->is_remat);
2908 if(!op->attr.remat.remat->inverse) {
2909 if(op->attr.remat.pre) {
2910 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
2912 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
2917 op_t *op = (op_t*)get_irn_link(n);
2918 assert(op && op->is_remat);
2920 if(op->attr.remat.pre) {
2921 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
2923 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
2934 dump_graph_with_remats(ir_graph * irg, const char * suffix)
2936 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
2937 be_dump(irg, suffix, dump_ir_block_graph_sched);
2938 set_dump_node_vcgattr_hook(NULL);
2943 * Edge hook to dump the schedule edges with annotated register pressure.
2946 sched_pressure_edge_hook(FILE *F, ir_node *irn)
2948 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
2949 ir_node *prev = sched_prev(irn);
2950 fprintf(F, "edge:{sourcename:\"");
2952 fprintf(F, "\" targetname:\"");
2954 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
2955 fprintf(F, "\" color:magenta}\n");
2961 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
2963 DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
2965 dump_consts_local(0);
2966 set_dump_node_edge_hook(sched_pressure_edge_hook);
2967 dump_ir_block_graph(irg, suffix);
2968 set_dump_node_edge_hook(old_edge_hook);
2972 walker_pressure_annotator(ir_node * bb, void * data)
2974 spill_ilp_t *si = data;
2977 pset *live = pset_new_ptr_default();
2980 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
2981 irn = be_lv_get_irn(si->lv, bb, i);
2983 if (has_reg_class(si, irn)) {
2984 pset_insert_ptr(live, irn);
2988 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
2990 sched_foreach_reverse(bb, irn) {
2992 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
2996 if(has_reg_class(si, irn)) {
2997 pset_remove_ptr(live, irn);
2998 if(is_Proj(irn)) ++projs;
3001 if(!is_Proj(irn)) projs = 0;
3003 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3004 ir_node *arg = get_irn_n(irn, n);
3006 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3008 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3015 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3017 be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
3022 connect_all_remats_with_keep(spill_ilp_t * si)
3030 n_remats = pset_count(si->all_possible_remats);
3032 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3035 pset_foreach(si->all_possible_remats, irn) {
3040 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
3042 obstack_free(si->obst, ins);
3048 connect_all_spills_with_keep(spill_ilp_t * si)
3057 n_spills = pset_count(si->spills);
3059 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3062 pset_foreach(si->spills, irn) {
3067 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
3069 obstack_free(si->obst, ins);
3073 /** insert a spill at an arbitrary position */
3074 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3076 ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
3077 ir_graph *irg = get_irn_irg(bl);
3078 ir_node *frame = get_irg_frame(irg);
3082 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
3083 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3085 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3088 * search the right insertion point. a spill of a phi cannot be put
3089 * directly after the phi, if there are some phis behind the one which
3090 * is spilled. Also, a spill of a Proj must be after all Projs of the
3093 * Here's one special case:
3094 * If the spill is in the start block, the spill must be after the frame
3095 * pointer is set up. This is done by setting insert to the end of the block
3096 * which is its default initialization (see above).
3099 if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3102 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3105 sched_add_after(insert, spill);
3110 delete_remat(spill_ilp_t * si, ir_node * remat) {
3112 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3114 sched_remove(remat);
3116 /* kill links to operands */
3117 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3118 set_irn_n(remat, n, bad);
3123 clean_remat_info(spill_ilp_t * si)
3127 remat_info_t *remat_info;
3128 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3130 set_foreach(si->remat_info, remat_info) {
3131 if(!remat_info->remats) continue;
3133 pset_foreach(remat_info->remats, remat)
3135 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3136 set_irn_n((ir_node*)remat->proj, -1, bad);
3137 set_irn_n((ir_node*)remat->proj, 0, bad);
3140 if(get_irn_n_edges(remat->op) == 0) {
3141 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3142 set_irn_n((ir_node*)remat->op, n, bad);
3147 if(remat_info->remats) del_pset(remat_info->remats);
3148 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3153 delete_unnecessary_remats(spill_ilp_t * si)
3157 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3160 ir_node *end = get_irg_end(si->chordal_env->irg);
3163 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3164 ir_node *keep_arg = get_irn_n(si->keep, n);
3165 op_t *arg_op = get_irn_link(keep_arg);
3168 assert(arg_op->is_remat);
3170 name = si->lpp->vars[arg_op->attr.remat.ilp];
3172 if(is_zero(name->value)) {
3173 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
3174 /* TODO check whether reload is preferred over remat (could be bug) */
3175 delete_remat(si, keep_arg);
3177 if(!arg_op->attr.remat.remat->inverse) {
3178 if(arg_op->attr.remat.pre) {
3179 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3181 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3184 if(arg_op->attr.remat.pre) {
3185 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3187 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3192 set_irn_n(si->keep, n, bad);
3195 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
3196 ir_node *end_arg = get_End_keepalive(end, i);
3198 if(end_arg != si->keep) {
3199 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
3202 keeps = obstack_finish(si->obst);
3203 set_End_keepalives(end, n-1, keeps);
3204 obstack_free(si->obst, keeps);
3207 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
3212 pset_foreach(si->all_possible_remats, remat) {
3213 op_t *remat_op = get_irn_link(remat);
3214 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
3216 if(is_zero(name->value)) {
3217 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
3218 /* TODO check whether reload is preferred over remat (could be bug) */
3219 delete_remat(si, remat);
3221 if(!remat_op->attr.remat.remat->inverse) {
3222 if(remat_op->attr.remat.pre) {
3223 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3225 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3228 if(remat_op->attr.remat.pre) {
3229 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3231 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3240 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3242 pset *spills = pset_new_ptr_default();
3244 const ir_node *next;
3247 defs = set_find_def(si->values, value);
3249 if(defs && defs->spills) {
3250 for(next = defs->spills; next; next = get_irn_link(next)) {
3251 pset_insert_ptr(spills, next);
3259 * @param before The node after which the spill will be placed in the schedule
3261 /* TODO set context properly */
3263 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3267 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3269 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3271 spill = be_spill2(arch_env, irn, before);
3273 defs = set_insert_def(si->values, value);
3276 /* enter into the linked list */
3277 set_irn_link(spill, defs->spills);
3278 defs->spills = spill;
3280 #ifdef KEEPALIVE_SPILLS
3281 pset_insert_ptr(si->spills, spill);
3288 * @param before The Phi node which has to be spilled
3291 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3298 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3300 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3301 ins[n] = si->m_unknown;
3304 mem_phi = new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3306 defs = set_insert_def(si->values, phi);
3309 /* enter into the linked list */
3310 set_irn_link(mem_phi, defs->spills);
3311 defs->spills = mem_phi;
3313 sched_add_after(phi, mem_phi);
3315 #ifdef KEEPALIVE_SPILLS
3316 pset_insert_ptr(si->spills, mem_phi);
3324 * Add remat to list of defs, destroys link field!
3327 insert_remat(spill_ilp_t * si, ir_node * remat)
3330 op_t *remat_op = get_irn_link(remat);
3332 assert(remat_op->is_remat);
3334 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3337 /* enter into the linked list */
3338 set_irn_link(remat, defs->remats);
3339 defs->remats = remat;
3344 * Add reload before operation and add to list of defs
3347 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3352 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3354 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3356 defs = set_find_def(si->values, value);
3358 spill = defs->spills;
3359 assert(spill && "no spill placed before reload");
3361 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3363 /* enter into the linked list */
3364 set_irn_link(reload, defs->remats);
3365 defs->remats = reload;
3370 #ifdef WITH_MEMOPERANDS
3371 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3375 ir_node *value = get_irn_n(memoperand->irn, memoperand->pos);
3377 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3379 DBG((si->dbg, LEVEL_2, "\t inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3381 defs = set_find_def(si->values, value);
3383 spill = defs->spills;
3384 assert(spill && "no spill placed before reload");
3386 reload = be_reload(arch_env, si->cls, memoperand->irn, get_irn_mode(value), spill);
3388 arch_perform_memory_operand(arch_env, memoperand->irn, reload, memoperand->pos);
3389 sched_remove(reload);
3392 void insert_memoperands(spill_ilp_t * si)
3394 memoperand_t *memoperand;
3397 set_foreach(si->memoperands, memoperand) {
3398 name = si->lpp->vars[memoperand->ilp];
3399 if(!is_zero(name->value)) {
3400 perform_memory_operand(si, memoperand);
3407 walker_spill_placer(ir_node * bb, void * data) {
3408 spill_ilp_t *si = (spill_ilp_t*)data;
3410 spill_bb_t *spill_bb = get_irn_link(bb);
3411 pset *spills_to_do = pset_new_ptr_default();
3414 set_foreach(spill_bb->ilp, spill) {
3417 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3418 name = si->lpp->vars[spill->mem_in];
3419 if(!is_zero(name->value)) {
3422 mem_phi = insert_mem_phi(si, spill->irn);
3424 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3428 name = si->lpp->vars[spill->spill];
3429 if(!is_zero(name->value)) {
3430 /* place spill directly after definition */
3431 if(get_nodes_block(spill->irn) == bb) {
3432 insert_spill(si, spill->irn, spill->irn, spill->irn);
3436 /* place spill at bb start */
3437 if(spill->reg_in > 0) {
3438 name = si->lpp->vars[spill->reg_in];
3439 if(!is_zero(name->value)) {
3440 insert_spill(si, spill->irn, spill->irn, bb);
3444 /* place spill after a remat */
3445 pset_insert_ptr(spills_to_do, spill->irn);
3448 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3451 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3452 op_t *op = get_irn_link(irn);
3454 if(be_is_Spill(irn)) continue;
3457 /* TODO fix this if we want to support remats with more than two nodes */
3458 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3459 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3461 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3464 if(pset_find_ptr(spills_to_do, irn)) {
3465 pset_remove_ptr(spills_to_do, irn);
3467 insert_spill(si, irn, irn, irn);
3473 assert(pset_count(spills_to_do) == 0);
3475 /* afterwards free data in block */
3476 del_pset(spills_to_do);
3480 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3482 ir_node *insert_pos = bb;
3484 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3486 /* find last definition of arg value in block */
3491 defs = set_find_def(si->values, value);
3493 if(defs && defs->remats) {
3494 for(next = defs->remats; next; next = get_irn_link(next)) {
3495 if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3496 last = sched_get_time_step(next);
3502 if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3503 last = sched_get_time_step(value);
3507 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
3509 spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3515 phim_fixer(spill_ilp_t *si) {
3518 set_foreach(si->values, defs) {
3519 const ir_node *phi = defs->value;
3520 op_t *op = get_irn_link(phi);
3521 ir_node *phi_m = NULL;
3522 ir_node *next = defs->spills;
3525 if(!is_Phi(phi)) continue;
3528 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3532 next = get_irn_link(next);
3535 if(!phi_m) continue;
3537 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3538 ir_node *value = get_irn_n(phi, n);
3539 defs_t *val_defs = set_find_def(si->values, value);
3541 /* a spill of this value */
3545 #ifndef NO_MEMCOPIES
3546 ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3547 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3549 if(!is_zero(name->value)) {
3550 spill = insert_mem_copy(si, pred, value);
3552 spill = val_defs->spills;
3555 spill = val_defs->spills;
3557 assert(spill && "no spill placed before PhiM");
3558 set_irn_n(phi_m, n, spill);
3564 walker_reload_placer(ir_node * bb, void * data) {
3565 spill_ilp_t *si = (spill_ilp_t*)data;
3567 spill_bb_t *spill_bb = get_irn_link(bb);
3569 /* reloads at end of block */
3570 if(spill_bb->reloads) {
3573 set_foreach(spill_bb->reloads, keyval) {
3574 ir_node *irn = (ir_node*)keyval->key;
3575 ilp_var_t reload = PTR_TO_INT(keyval->val);
3578 name = si->lpp->vars[reload];
3579 if(!is_zero(name->value)) {
3581 ir_node *insert_pos = bb;
3582 ir_node *prev = sched_block_last_noncf(si, bb);
3583 op_t *prev_op = get_irn_link(prev);
3585 while(be_is_Spill(prev)) {
3586 prev = sched_prev(prev);
3589 prev_op = get_irn_link(prev);
3591 /* insert reload before pre-remats */
3592 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3593 && prev_op->is_remat && prev_op->attr.remat.pre) {
3597 prev = sched_prev(prev);
3598 } while(be_is_Spill(prev));
3600 prev_op = get_irn_link(prev);
3604 reload = insert_reload(si, irn, insert_pos);
3606 #ifdef KEEPALIVE_RELOADS
3607 pset_insert_ptr(si->spills, reload);
3613 /* walk and insert more reloads and collect remats */
3614 sched_foreach_reverse(bb, irn) {
3615 op_t *op = get_irn_link(irn);
3617 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3618 if(is_Phi(irn)) break;
3621 if(get_irn_mode(irn) != mode_T) {
3622 insert_remat(si, irn);
3627 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3628 ir_node *arg = get_irn_n(irn, n);
3630 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3633 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3634 if(!is_zero(name->value)) {
3636 ir_node *insert_pos = irn;
3637 ir_node *prev = sched_prev(insert_pos);
3640 while(be_is_Spill(prev)) {
3641 prev = sched_prev(prev);
3644 prev_op = get_irn_link(prev);
3646 /* insert reload before pre-remats */
3647 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3648 && prev_op->is_remat && prev_op->attr.remat.pre) {
3652 prev = sched_prev(prev);
3653 } while(be_is_Spill(prev));
3655 prev_op = get_irn_link(prev);
3659 reload = insert_reload(si, arg, insert_pos);
3661 set_irn_n(irn, n, reload);
3663 #ifdef KEEPALIVE_RELOADS
3664 pset_insert_ptr(si->spills, reload);
3672 del_set(spill_bb->ilp);
3673 if(spill_bb->reloads) del_set(spill_bb->reloads);
3677 walker_collect_used(ir_node * irn, void * data)
3679 lc_bitset_t *used = data;
3681 lc_bitset_set(used, get_irn_idx(irn));
3684 struct kill_helper {
3690 walker_kill_unused(ir_node * bb, void * data)
3692 struct kill_helper *kh = data;
3693 ir_node *bad = get_irg_bad(get_irn_irg(bb));
3697 for(irn=sched_first(bb); !sched_is_end(irn);) {
3698 ir_node *next = sched_next(irn);
3701 if(!lc_bitset_is_set(kh->used, get_irn_idx(irn))) {
3702 if(be_is_Spill(irn) || be_is_Reload(irn)) {
3703 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)));
3705 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3711 set_nodes_block(irn, bad);
3712 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3713 set_irn_n(irn, n, bad);
3721 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3723 struct kill_helper kh;
3725 kh.used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3728 irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3729 irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3731 lc_bitset_free(kh.used);
3735 print_irn_pset(pset * p)
3739 pset_foreach(p, irn) {
3740 ir_printf("%+F\n", irn);
3745 dump_phi_class(spill_ilp_t * si, pset * phiclass, const char * file)
3747 FILE *f = fopen(file, "w");
3749 interference_t *interference;
3751 pset_break(phiclass);
3752 set_break(si->interferences);
3754 ir_fprintf(f, "digraph phiclass {\n");
3756 pset_foreach(phiclass, irn) {
3758 ir_fprintf(f, " %F%N [shape=box]\n",irn,irn);
3761 pset_foreach(phiclass, irn) {
3764 if(!is_Phi(irn)) continue;
3766 for(n=get_irn_arity(irn)-1; n>=0; --n) {
3767 ir_node *arg = get_irn_n(irn, n);
3769 ir_fprintf(f, " %F%N -> %F%N\n",irn,irn,arg,arg);
3773 set_foreach(si->interferences, interference) {
3774 const ir_node *a = interference->a;
3775 const ir_node *b = interference->b;
3776 if(get_phi_class(a) == phiclass) {
3777 ir_fprintf(f, " %F%N -> %F%N [color=red,dir=none,style=bold]\n",a,a,b,b);
3786 rewire_uses(spill_ilp_t * si)
3788 dom_front_info_t *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
3790 pset *ignore = pset_new_ptr(1);
3792 pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
3794 /* then fix uses of spills */
3795 set_foreach(si->values, defs) {
3798 const ir_node *next = defs->remats;
3801 reloads = pset_new_ptr_default();
3804 if(be_is_Reload(next)) {
3805 pset_insert_ptr(reloads, next);
3809 next = get_irn_link(next);
3812 spills = get_spills_for_value(si, defs->value);
3813 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));
3814 if(pset_count(spills) > 1) {
3815 //assert(pset_count(reloads) > 0);
3816 // print_irn_pset(spills);
3817 // print_irn_pset(reloads);
3819 be_ssa_constr_set_ignore(dfi, si->lv, spills, ignore);
3826 /* first fix uses of remats and reloads */
3827 set_foreach(si->values, defs) {
3829 const ir_node *next = defs->remats;
3832 nodes = pset_new_ptr_default();
3833 pset_insert_ptr(nodes, defs->value);
3836 pset_insert_ptr(nodes, next);
3837 next = get_irn_link(next);
3840 if(pset_count(nodes) > 1) {
3841 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
3842 be_ssa_constr_set(dfi, si->lv, nodes);
3849 // remove_unused_defs(si);
3851 be_free_dominance_frontiers(dfi);
3856 writeback_results(spill_ilp_t * si)
3858 /* walk through the graph and collect all spills, reloads and remats for a value */
3860 si->values = new_set(cmp_defs, 4096);
3862 DBG((si->dbg, LEVEL_1, "Applying results\n"));
3863 delete_unnecessary_remats(si);
3864 si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
3865 irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
3866 irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
3867 #ifdef WITH_MEMOPERANDS
3868 insert_memoperands(si);
3872 /* clean the remat info! there are still back-edges leading there! */
3873 clean_remat_info(si);
3877 connect_all_spills_with_keep(si);
3879 del_set(si->values);
3883 get_n_regs(spill_ilp_t * si)
3885 int arch_n_regs = arch_register_class_n_regs(si->cls);
3889 for(i=0; i<arch_n_regs; i++) {
3890 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
3895 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
3900 walker_reload_mover(ir_node * bb, void * data)
3902 spill_ilp_t *si = data;
3905 sched_foreach(bb, tmp) {
3906 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
3907 ir_node *reload = tmp;
3910 /* move reload upwards */
3912 int pressure = (int)get_irn_link(reload);
3913 if(pressure < si->n_regs) {
3914 irn = sched_prev(reload);
3915 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
3916 sched_remove(reload);
3917 pressure = (int)get_irn_link(irn);
3919 while(pressure < si->n_regs) {
3920 if( sched_is_end(irn) ||
3921 (be_is_Reload(irn) && has_reg_class(si, irn)) ||
3922 /* do not move reload before its spill */
3923 (irn == be_get_Reload_mem(reload)) ||
3924 /* do not move before phi */
3927 set_irn_link(irn, INT_TO_PTR(pressure+1));
3928 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
3929 irn = sched_prev(irn);
3931 pressure = (int)get_irn_link(irn);
3934 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
3935 sched_put_after(irn, reload);
3942 move_reloads_upward(spill_ilp_t * si)
3944 irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
3949 * Walk all irg blocks and check for interfering spills inside of phi classes
3952 luke_meminterferencechecker(ir_node * bb, void * data)
3954 spill_ilp_t *si = (spill_ilp_t*)data;
3957 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
3958 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
3960 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
3962 /* a is only interesting if it is in my register class and if it is inside a phi class */
3963 if (has_reg_class(si, a) && get_phi_class(a)) {
3964 for(l2=_be_lv_next_irn(si->lv, bb, 0xff, l1+1); l2>=0; l2=_be_lv_next_irn(si->lv, bb, 0xff, l2+1)) {
3965 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
3967 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
3969 /* a and b are only interesting if they are in the same phi class */
3970 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
3971 if(values_interfere_in_block(si, bb, a, b)) {
3972 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
3981 verify_phiclasses(spill_ilp_t * si)
3983 /* analyze phi classes */
3984 phi_class_compute(si->chordal_env->irg);
3986 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
3987 irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
3991 walker_spillslotassigner(ir_node * irn, void * data)
3995 if(!be_is_Spill(irn)) return;
3997 /* set spill context to phi class if it has one ;) */
4000 // Matze: not needed anymore
4001 cls = get_phi_class(irn);
4003 be_set_Spill_context(irn, cls);
4005 be_set_Spill_context(irn, irn);
4011 assign_spillslots(spill_ilp_t * si)
4013 DBG((si->dbg, LEVEL_2, "\t calling spill slot assigner\n"));
4014 irg_walk_graph(si->chordal_env->irg, walker_spillslotassigner, NULL, si);
4018 be_spill_remat(const be_chordal_env_t * chordal_env)
4020 char problem_name[256];
4021 char dump_suffix[256];
4022 char dump_suffix2[256];
4023 struct obstack obst;
4026 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
4027 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
4028 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
4030 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4031 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4033 #ifdef VERIFY_DOMINANCE
4034 be_check_dominance(chordal_env->irg);
4037 obstack_init(&obst);
4038 si.chordal_env = chordal_env;
4040 si.cls = chordal_env->cls;
4041 si.lpp = new_lpp(problem_name, lpp_minimize);
4042 si.remat_info = new_set(cmp_remat_info, 4096);
4043 si.interferences = new_set(cmp_interference, 32);
4044 #ifdef WITH_MEMOPERANDS
4045 si.memoperands = new_set(cmp_memoperands, 128);
4047 si.all_possible_remats = pset_new_ptr_default();
4048 si.spills = pset_new_ptr_default();
4049 si.inverse_ops = pset_new_ptr_default();
4050 si.lv = chordal_env->lv;
4054 si.n_regs = get_n_regs(&si);
4056 set_irg_link(chordal_env->irg, &si);
4057 compute_doms(chordal_env->irg);
4059 /* compute phi classes */
4060 // phi_class_compute(chordal_env->irg);
4062 be_analyze_regpressure(chordal_env, "-pre");
4064 #ifdef COLLECT_REMATS
4065 /* collect remats */
4066 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4067 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
4070 /* insert possible remats */
4071 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4072 irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
4073 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4076 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4077 connect_all_remats_with_keep(&si);
4078 /* dump graph with inserted remats */
4079 dump_graph_with_remats(chordal_env->irg, dump_suffix);
4082 /* insert copies for phi arguments not in my regclass */
4083 irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
4085 /* recompute liveness */
4086 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4087 be_liveness_recompute(si.lv);
4091 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4092 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4093 irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
4095 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4096 irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
4098 #ifndef NO_MEMCOPIES
4099 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4100 memcopyhandler(&si);
4108 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4109 if ((f = fopen(buf, "wt")) != NULL) {
4110 lpp_dump_plain(si.lpp, f);
4117 DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4119 lpp_set_time_limit(si.lpp, ILP_TIMEOUT);
4122 lpp_set_log(si.lpp, stdout);
4125 lpp_solve_cplex(si.lpp);
4127 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4129 assert(lpp_is_sol_valid(si.lpp)
4130 && "solution of ILP must be valid");
4132 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));
4134 #ifdef DUMP_SOLUTION
4139 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4140 if ((f = fopen(buf, "wt")) != NULL) {
4142 for (i = 0; i < si.lpp->var_next; ++i) {
4143 lpp_name_t *name = si.lpp->vars[i];
4144 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4151 writeback_results(&si);
4155 kill_all_unused_values_in_schedule(&si);
4157 #if defined(KEEPALIVE_SPILLS) || defined(KEEPALIVE_RELOADS)
4158 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
4161 // move reloads upwards
4162 be_liveness_recompute(si.lv);
4163 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4164 move_reloads_upward(&si);
4166 #ifndef NO_MEMCOPIES
4167 verify_phiclasses(&si);
4168 assign_spillslots(&si);
4171 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4173 dump_pressure_graph(&si, dump_suffix2);
4175 be_analyze_regpressure(chordal_env, "-post");
4177 #ifdef VERIFY_DOMINANCE
4178 be_check_dominance(chordal_env->irg);
4181 free_dom(chordal_env->irg);
4182 del_set(si.interferences);
4183 del_pset(si.inverse_ops);
4184 del_pset(si.all_possible_remats);
4185 #ifdef WITH_MEMOPERANDS
4186 del_set(si.memoperands);
4188 del_pset(si.spills);
4190 obstack_free(&obst, NULL);
4191 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4194 #else /* WITH_ILP */
4197 only_that_you_can_compile_without_WITH_ILP_defined(void)
4201 #endif /* WITH_ILP */