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"
39 #include <lpp/lpp_net.h>
40 #include <lpp/lpp_cplex.h>
41 //#include <lc_pset.h>
42 #include <libcore/lc_bitset.h>
46 #include "besched_t.h"
51 #include "bespillremat.h"
53 #include "bepressurestat.h"
55 #include "bechordal_t.h"
58 #include <libcore/lc_opts.h>
59 #include <libcore/lc_opts_enum.h>
60 #endif /* WITH_LIBCORE */
62 #define DUMP_PROBLEM 1
64 #define DUMP_SOLUTION 4
66 static int dump_flags = 0;
67 static int activate_log = 0;
70 static const lc_opt_enum_mask_items_t dump_items[] = {
71 { "problem", DUMP_PROBLEM },
73 { "solution", DUMP_SOLUTION },
77 static lc_opt_enum_mask_var_t dump_var = {
78 &dump_flags, dump_items
81 static const lc_opt_table_entry_t options[] = {
82 LC_OPT_ENT_ENUM_MASK("dump", "dump ifg before, after or after each cloud", &dump_var),
83 LC_OPT_ENT_BOOL ("log", "activate the lpp log", &activate_log),
87 void be_spill_remat_register_options(lc_opt_entry_t *grp)
89 lc_opt_entry_t *my_grp = lc_opt_get_grp(grp, "remat");
90 lc_opt_add_table(my_grp, options);
94 //#define DUMP_SOLUTION
96 //#define KEEPALIVE /* keep alive all inserted remats and dump graph with remats */
97 //#define COLLECT_REMATS /* enable rematerialization */
98 //#define COLLECT_INVERSE_REMATS /* enable placement of inverse remats */
99 //#define ONLY_BRIGGS_REMATS /* only remats without parameters (or only with ignored params) */
100 #define REMAT_WHILE_LIVE /* only remat values that are live */
101 //#define NO_ENLARGE_L1V3N355 /* do not remat after the death of some operand */
102 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
103 #define MAY_DIE_AT_REMAT /* allow values to die after a pre remat */
104 #define NO_SINGLE_USE_REMATS /* do not repair schedule */
105 //#define KEEPALIVE_SPILLS
106 //#define KEEPALIVE_RELOADS
107 #define GOODWIN_REDUCTION
108 //#define NO_MEMCOPIES
109 //#define VERIFY_DOMINANCE
110 #define WITH_MEMOPERANDS
113 //#define SOLVE_LOCAL
114 #define LPP_SERVER "i44pc52"
115 #define LPP_SOLVER "cplex"
118 #define COST_MEMOPERAND 7
119 #define COST_STORE 50
122 #define ILP_TIMEOUT 300
126 typedef struct _spill_ilp_t {
127 const arch_register_class_t *cls;
129 const be_chordal_env_t *chordal_env;
132 struct obstack *obst;
134 pset *all_possible_remats;
139 set *values; /**< for collecting all definitions of values before running ssa-construction */
143 #ifdef WITH_MEMOPERANDS
146 DEBUG_ONLY(firm_dbg_module_t * dbg);
149 typedef int ilp_var_t;
150 typedef int ilp_cst_t;
152 typedef struct _spill_bb_t {
157 typedef struct _remat_t {
158 const ir_node *op; /**< for copy_irn */
159 const ir_node *value; /**< the value which is being recomputed by this remat */
160 const ir_node *proj; /**< not NULL if the above op produces a tuple */
161 int cost; /**< cost of this remat */
162 int inverse; /**< nonzero if this is an inverse remat */
166 * Data to be attached to each IR node. For remats this contains the ilp_var
167 * for this remat and for normal ops this contains the ilp_vars for
168 * reloading each operand
170 typedef struct _op_t {
175 const remat_t *remat; /** the remat this op belongs to */
176 int pre; /** 1, if this is a pressure-increasing remat */
180 ir_node *op; /** the operation this live range belongs to */
189 typedef struct _defs_t {
190 const ir_node *value;
191 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
192 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
195 typedef struct _remat_info_t {
196 const ir_node *irn; /**< the irn to which these remats belong */
197 pset *remats; /**< possible remats for this value */
198 pset *remats_by_operand; /**< remats with this value as operand */
201 typedef struct _keyval_t {
206 typedef struct _spill_t {
215 #ifdef WITH_MEMOPERANDS
216 typedef struct _memoperand_t {
217 ir_node *irn; /**< the irn */
218 unsigned int pos; /**< the position of the argument */
219 ilp_var_t ilp; /**< the ilp var for this memory operand */
224 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
226 return chordal_has_class(si->chordal_env, irn);
231 cmp_remat(const void *a, const void *b)
233 const keyval_t *p = a;
234 const keyval_t *q = b;
235 const remat_t *r = p->val;
236 const remat_t *s = q->val;
240 return !(r == s || r->op == s->op);
244 cmp_remat(const void *a, const void *b)
246 const remat_t *r = a;
247 const remat_t *s = a;
249 return !(r == s || r->op == s->op);
253 cmp_spill(const void *a, const void *b, size_t size)
255 const spill_t *p = a;
256 const spill_t *q = b;
258 // return !(p->irn == q->irn && p->bb == q->bb);
259 return !(p->irn == q->irn);
262 #ifdef WITH_MEMOPERANDS
264 cmp_memoperands(const void *a, const void *b, size_t size)
266 const memoperand_t *p = a;
267 const memoperand_t *q = b;
269 return !(p->irn == q->irn && p->pos == q->pos);
274 set_find_keyval(set * set, const void * key)
279 return set_find(set, &query, sizeof(query), HASH_PTR(key));
283 set_insert_keyval(set * set, void * key, void * val)
289 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
293 set_find_def(set * set, const ir_node * value)
298 return set_find(set, &query, sizeof(query), HASH_PTR(value));
302 set_insert_def(set * set, const ir_node * value)
309 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
312 #ifdef WITH_MEMOPERANDS
313 static memoperand_t *
314 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
321 return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
324 static memoperand_t *
325 set_find_memoperand(set * set, const ir_node * irn, unsigned int pos)
329 query.irn = (ir_node*)irn;
331 return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
337 set_find_spill(set * set, const ir_node * value)
341 query.irn = (ir_node*)value;
342 return set_find(set, &query, sizeof(query), HASH_PTR(value));
345 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
346 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
347 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
348 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
349 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
352 cmp_remat_info(const void *a, const void *b, size_t size)
354 const remat_info_t *p = a;
355 const remat_info_t *q = b;
357 return !(p->irn == q->irn);
361 cmp_defs(const void *a, const void *b, size_t size)
366 return !(p->value == q->value);
370 cmp_keyval(const void *a, const void *b, size_t size)
372 const keyval_t *p = a;
373 const keyval_t *q = b;
375 return !(p->key == q->key);
379 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
382 #ifndef EXECFREQ_LOOPDEPH
383 return get_block_execfreq(si->chordal_env->exec_freq, get_block(irn)) + FUDGE;
386 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
388 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
393 get_cost(const spill_ilp_t * si, const ir_node * irn)
395 if(be_is_Spill(irn)) {
397 } else if(be_is_Reload(irn)){
400 return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
405 * Checks, whether node and its operands have suitable reg classes
408 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
411 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
412 int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
416 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
419 for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
420 ir_node *op = get_irn_n(irn, n);
421 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
424 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
431 * Try to create a remat from @p op with destination value @p dest_value
433 static INLINE remat_t *
434 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
436 remat_t *remat = NULL;
438 // if(!mode_is_datab(get_irn_mode(dest_value)))
441 if(dest_value == op) {
442 const ir_node *proj = NULL;
444 if(is_Proj(dest_value)) {
445 op = get_irn_n(op, 0);
449 if(!is_rematerializable(si, op))
452 remat = obstack_alloc(si->obst, sizeof(*remat));
454 remat->cost = get_cost(si, op);
455 remat->value = dest_value;
459 arch_inverse_t inverse;
462 /* get the index of the operand we want to retrieve by the inverse op */
463 for (n = get_irn_arity(op)-1; n>=0; --n) {
464 ir_node *arg = get_irn_n(op, n);
466 if(arg == dest_value) break;
470 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", n, op));
472 /* else ask the backend to give an inverse op */
473 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
476 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
478 assert(inverse.n > 0 && "inverse op should have at least one node");
480 for(i=inverse.n-1; i>=0; --i) {
481 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
485 remat = obstack_alloc(si->obst, sizeof(*remat));
486 remat->op = inverse.nodes[0];
487 remat->cost = inverse.costs;
488 remat->value = dest_value;
489 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
492 assert(is_Proj(remat->proj));
494 assert(0 && "I can not handle remats with more than 2 nodes");
501 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
503 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
511 add_remat(const spill_ilp_t * si, const remat_t * remat)
513 remat_info_t *remat_info,
518 assert(remat->value);
520 query.irn = remat->value;
522 query.remats_by_operand = NULL;
523 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
525 if(remat_info->remats == NULL) {
526 remat_info->remats = new_pset(cmp_remat, 4096);
528 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
530 /* insert the remat into the remats_be_operand set of each argument of the remat op */
531 for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
532 ir_node *arg = get_irn_n(remat->op, n);
536 query.remats_by_operand = NULL;
537 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
539 if(remat_info->remats_by_operand == NULL) {
540 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
542 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
546 #ifdef NO_SINGLE_USE_REMATS
548 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
550 const ir_edge_t *edge = get_irn_out_edge_first(irn);
554 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
557 edge = get_irn_out_edge_next(irn, edge);
564 #ifdef ONLY_BRIGGS_REMATS
566 get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
569 unsigned int ret = 0;
571 for(n=get_irn_arity(irn)-1; n>=0; --n) {
572 if(has_reg_class(si, irn)) ++ret;
580 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
585 if( has_reg_class(si, op)
586 #ifdef NO_SINGLE_USE_REMATS
587 && (get_irn_n_nonremat_edges(si, op) > 1)
589 #ifdef ONLY_BRIGGS_REMATS
590 && (get_irn_n_nonignore_args(si, op) == 0)
593 remat = get_remat_from_op(si, op, op);
595 add_remat(si, remat);
599 #if defined(COLLECT_INVERSE_REMATS) && !defined(ONLY_BRIGGS_REMATS)
600 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
602 for (n = get_irn_arity(op)-1; n>=0; --n) {
603 ir_node *arg = get_irn_n(op, n);
605 if(has_reg_class(si, arg)) {
606 /* try to get an inverse remat */
607 remat = get_remat_from_op(si, arg, op);
609 add_remat(si, remat);
618 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
621 ir_node *def_block = get_nodes_block(val);
627 /* if pos is at end of a basic block */
629 ret = (pos == def_block || block_dominates(def_block, pos));
630 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
634 /* else if this is a normal operation */
635 block = get_nodes_block(pos);
636 if(block == def_block) {
637 if(!sched_is_scheduled(val)) return 1;
639 ret = sched_comes_after(val, pos);
640 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
644 ret = block_dominates(def_block, block);
645 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
649 static INLINE ir_node *
650 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
652 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
656 * Returns first non-Phi node of block @p bb
658 static INLINE ir_node *
659 sched_block_first_nonphi(const ir_node * bb)
661 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
665 sched_skip_proj_predicator(const ir_node * irn, void * data)
667 return (is_Proj(irn));
670 static INLINE ir_node *
671 sched_next_nonproj(const ir_node * irn, int forward)
673 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
677 * Returns next operation node (non-Proj) after @p irn
678 * or the basic block of this node
680 static INLINE ir_node *
681 sched_next_op(const ir_node * irn)
683 ir_node *next = sched_next(irn);
688 return sched_next_nonproj(next, 1);
692 * Returns previous operation node (non-Proj) before @p irn
693 * or the basic block of this node
695 static INLINE ir_node *
696 sched_prev_op(const ir_node * irn)
698 ir_node *prev = sched_prev(irn);
703 return sched_next_nonproj(prev, 0);
707 sched_put_after(ir_node * insert, ir_node * irn)
709 if(is_Block(insert)) {
710 insert = sched_block_first_nonphi(insert);
712 insert = sched_next_op(insert);
714 sched_add_before(insert, irn);
718 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
720 if(is_Block(insert)) {
721 insert = sched_block_last_noncf(si, insert);
723 insert = sched_next_nonproj(insert, 0);
724 insert = sched_prev(insert);
726 sched_add_after(insert, irn);
730 * Tells you whether a @p remat can be placed before the irn @p pos
733 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
735 const ir_node *op = remat->op;
741 prev = sched_block_last_noncf(si, pos);
742 prev = sched_next_nonproj(prev, 0);
744 prev = sched_prev_op(pos);
746 /* do not remat if the rematted value is defined immediately before this op */
747 if(prev == remat->op) {
752 /* this should be just fine, the following OP will be using this value, right? */
754 /* only remat AFTER the real definition of a value (?) */
755 if(!value_is_defined_before(si, pos, remat->value)) {
756 // ir_fprintf(stderr, "error(not defined)");
761 for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
762 const ir_node *arg = get_irn_n(op, n);
764 #ifdef NO_ENLARGE_L1V3N355
765 if(has_reg_class(si, arg) && live) {
766 res &= pset_find_ptr(live, arg)?1:0;
768 res &= value_is_defined_before(si, pos, arg);
771 res &= value_is_defined_before(si, pos, arg);
779 * Tells you whether a @p remat can be placed after the irn @p pos
782 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
785 pos = sched_block_first_nonphi(pos);
787 pos = sched_next_op(pos);
790 /* only remat AFTER the real definition of a value (?) */
791 if(!value_is_defined_before(si, pos, remat->value)) {
795 return can_remat_before(si, remat, pos, live);
799 * Collect potetially rematerializable OPs
802 walker_remat_collector(ir_node * irn, void * data)
804 spill_ilp_t *si = data;
806 if(!is_Block(irn) && !is_Phi(irn)) {
807 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
808 get_remats_from_op(si, irn);
813 * Inserts a copy of @p irn before @p pos
816 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
821 bb = is_Block(pos)?pos:get_nodes_block(pos);
822 copy = exact_copy(irn);
824 _set_phi_class(copy, NULL);
825 set_nodes_block(copy, bb);
826 sched_put_before(si, pos, copy);
832 * Inserts a copy of @p irn after @p pos
835 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
840 bb = is_Block(pos)?pos:get_nodes_block(pos);
841 copy = exact_copy(irn);
843 _set_phi_class(copy, NULL);
844 set_nodes_block(copy, bb);
845 sched_put_after(pos, copy);
851 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
855 if(can_remat_after(si, remat, pos, live)) {
860 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
862 copy = insert_copy_after(si, remat->op, pos);
864 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
865 op = obstack_alloc(si->obst, sizeof(*op));
867 op->attr.remat.remat = remat;
868 op->attr.remat.pre = 0;
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 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
893 if(can_remat_before(si, remat, pos, live)) {
898 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
900 copy = insert_copy_before(si, remat->op, pos);
902 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
903 op = obstack_alloc(si->obst, sizeof(*op));
905 op->attr.remat.remat = remat;
906 op->attr.remat.pre = 1;
907 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
909 set_irn_link(copy, op);
910 pset_insert_ptr(si->all_possible_remats, copy);
912 proj_copy = insert_copy_after(si, remat->proj, copy);
913 set_irn_n(proj_copy, 0, copy);
914 set_irn_link(proj_copy, op);
915 pset_insert_ptr(si->all_possible_remats, proj_copy);
927 get_block_n_succs(const ir_node *block) {
928 const ir_edge_t *edge;
930 assert(edges_activated(current_ir_graph));
932 edge = get_block_succ_first(block);
936 edge = get_block_succ_next(block, edge);
941 is_merge_edge(const ir_node * bb)
943 #ifdef GOODWIN_REDUCTION
944 return get_block_n_succs(bb) == 1;
951 is_diverge_edge(const ir_node * bb)
953 #ifdef GOODWIN_REDUCTION
954 return get_Block_n_cfgpreds(bb) == 1;
961 walker_regclass_copy_insertor(ir_node * irn, void * data)
963 spill_ilp_t *si = data;
965 if(is_Phi(irn) && has_reg_class(si, irn)) {
968 for(n=get_irn_arity(irn)-1; n>=0; --n) {
969 ir_node *phi_arg = get_irn_n(irn, n);
970 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
972 if(!has_reg_class(si, phi_arg)) {
973 ir_node *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
974 ir_node *pos = sched_block_last_noncf(si, bb);
975 op_t *op = obstack_alloc(si->obst, sizeof(*op));
977 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
978 sched_add_after(pos, copy);
979 set_irn_n(irn, n, copy);
982 op->attr.live_range.args.reloads = NULL;
983 op->attr.live_range.ilp = ILP_UNDEF;
984 set_irn_link(copy, op);
992 * Insert (so far unused) remats into the irg to
993 * recompute the potential liveness of all values
996 walker_remat_insertor(ir_node * bb, void * data)
998 spill_ilp_t *si = data;
999 spill_bb_t *spill_bb;
1002 pset *live = pset_new_ptr_default();
1004 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1006 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1007 ir_node *value = be_lv_get_irn(si->lv, bb, i);
1009 /* add remats at end of block */
1010 if (has_reg_class(si, value)) {
1011 pset_insert_ptr(live, value);
1015 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1016 set_irn_link(bb, spill_bb);
1018 irn = sched_last(bb);
1019 while(!sched_is_end(irn)) {
1026 next = sched_prev(irn);
1028 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
1030 if(is_Phi(irn) || is_Proj(irn)) {
1033 if(has_reg_class(si, irn)) {
1034 pset_remove_ptr(live, irn);
1037 op = obstack_alloc(si->obst, sizeof(*op));
1039 op->attr.live_range.args.reloads = NULL;
1040 op->attr.live_range.ilp = ILP_UNDEF;
1041 set_irn_link(irn, op);
1047 op = obstack_alloc(si->obst, sizeof(*op));
1049 op->attr.live_range.ilp = ILP_UNDEF;
1050 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1051 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1052 set_irn_link(irn, op);
1054 args = pset_new_ptr_default();
1056 /* collect arguments of op */
1057 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1058 ir_node *arg = get_irn_n(irn, n);
1060 pset_insert_ptr(args, arg);
1063 /* set args of op already live in epilog */
1064 pset_foreach(args, arg) {
1065 if(has_reg_class(si, arg)) {
1066 pset_insert_ptr(live, arg);
1069 /* delete defined value from live set */
1070 if(has_reg_class(si, irn)) {
1071 pset_remove_ptr(live, irn);
1075 remat_args = pset_new_ptr_default();
1077 /* insert all possible remats before irn */
1078 pset_foreach(args, arg) {
1079 remat_info_t *remat_info,
1083 /* continue if the operand has the wrong reg class
1085 if(!has_reg_class(si, arg))
1089 query.remats = NULL;
1090 query.remats_by_operand = NULL;
1091 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1097 if(remat_info->remats) {
1098 pset_foreach(remat_info->remats, remat) {
1099 ir_node *remat_irn = NULL;
1101 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
1102 #ifdef REMAT_WHILE_LIVE
1103 if(pset_find_ptr(live, remat->value)) {
1104 remat_irn = insert_remat_before(si, remat, irn, live);
1107 remat_irn = insert_remat_before(si, remat, irn, live);
1110 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1111 ir_node *remat_arg = get_irn_n(remat_irn, n);
1113 if(!has_reg_class(si, remat_arg)) continue;
1115 pset_insert_ptr(remat_args, remat_arg);
1122 /* now we add remat args to op's args because they could also die at this op */
1123 pset_foreach(args,arg) {
1124 if(pset_find_ptr(remat_args, arg)) {
1125 pset_remove_ptr(remat_args, arg);
1128 pset_foreach(remat_args,arg) {
1129 pset_insert_ptr(args, arg);
1132 /* insert all possible remats after irn */
1133 pset_foreach(args, arg) {
1134 remat_info_t *remat_info,
1138 /* continue if the operand has the wrong reg class */
1139 if(!has_reg_class(si, arg))
1143 query.remats = NULL;
1144 query.remats_by_operand = NULL;
1145 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1151 /* do not place post remats after jumps */
1152 if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
1154 if(remat_info->remats_by_operand) {
1155 pset_foreach(remat_info->remats_by_operand, remat) {
1156 /* do not insert remats producing the same value as one of the operands */
1157 if(!pset_find_ptr(args, remat->value)) {
1158 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
1159 #ifdef REMAT_WHILE_LIVE
1160 if(pset_find_ptr(live, remat->value)) {
1161 insert_remat_after(si, remat, irn, live);
1164 insert_remat_after(si, remat, irn, live);
1171 del_pset(remat_args);
1176 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_in, i) {
1177 ir_node *value = be_lv_get_irn(si->lv, bb, i);
1179 /* add remats at end if successor has multiple predecessors */
1180 if(is_merge_edge(bb)) {
1181 /* add remats at end of block */
1182 if (be_is_live_end(si->lv, bb, value) && has_reg_class(si, value)) {
1183 remat_info_t *remat_info,
1188 query.remats = NULL;
1189 query.remats_by_operand = NULL;
1190 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1192 if(remat_info && remat_info->remats) {
1193 pset_foreach(remat_info->remats, remat) {
1194 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1196 insert_remat_before(si, remat, bb, NULL);
1201 if(is_diverge_edge(bb)) {
1202 /* add remat2s at beginning of block */
1203 if ((be_is_live_in(si->lv, bb, value) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) {
1204 remat_info_t *remat_info,
1209 query.remats = NULL;
1210 query.remats_by_operand = NULL;
1211 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1213 if(remat_info && remat_info->remats) {
1214 pset_foreach(remat_info->remats, remat) {
1215 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at beginning of block %+F\n", remat->op, bb));
1217 /* put the remat here if all its args are available */
1218 insert_remat_after(si, remat, bb, NULL);
1228 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1231 luke_endwalker(ir_node * bb, void * data)
1233 spill_ilp_t *si = (spill_ilp_t*)data;
1239 spill_bb_t *spill_bb = get_irn_link(bb);
1243 live = pset_new_ptr_default();
1244 use_end = pset_new_ptr_default();
1246 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1247 irn = be_lv_get_irn(si->lv, bb, i);
1248 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1251 pset_insert_ptr(live, irn);
1252 op = get_irn_link(irn);
1253 assert(!op->is_remat);
1257 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1258 /* their reg_out must always be set */
1259 sched_foreach_reverse(bb, irn) {
1262 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1264 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1265 ir_node *irn_arg = get_irn_n(irn, n);
1267 if(has_reg_class(si, irn_arg)) {
1268 pset_insert_ptr(use_end, irn_arg);
1273 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1274 //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1275 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1277 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1279 /* if this is a merge edge we can reload at the end of this block */
1280 if(is_merge_edge(bb)) {
1281 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1282 } else if(pset_count(use_end)){
1283 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1285 spill_bb->reloads = NULL;
1288 pset_foreach(live,irn) {
1292 int default_spilled;
1295 /* handle values used by control flow nodes later separately */
1296 if(pset_find_ptr(use_end, irn)) continue;
1299 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1301 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1303 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1304 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1305 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1307 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1308 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1310 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1311 /* by default spill value right after definition */
1312 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1313 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1315 if(is_merge_edge(bb)) {
1319 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1320 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb), 0.0);
1321 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1323 /* reload <= mem_out */
1324 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1325 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1326 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1329 spill->reg_in = ILP_UNDEF;
1330 spill->mem_in = ILP_UNDEF;
1333 pset_foreach(use_end,irn) {
1337 ilp_cst_t end_use_req,
1340 int default_spilled;
1343 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1345 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1347 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1348 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1349 /* if irn is used at the end of the block, then it is live anyway */
1350 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1352 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1353 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1355 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1356 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1357 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1359 /* reload for use be control flow op */
1360 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1361 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb), 1.0);
1362 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1364 /* reload <= mem_out */
1365 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1366 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1367 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1369 spill->reg_in = ILP_UNDEF;
1370 spill->mem_in = ILP_UNDEF;
1372 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1373 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1374 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1382 next_post_remat(const ir_node * irn)
1388 next = sched_block_first_nonphi(irn);
1390 next = sched_next_op(irn);
1393 if(sched_is_end(next))
1396 op = get_irn_link(next);
1397 if(op->is_remat && !op->attr.remat.pre) {
1406 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1412 ret = sched_block_last_noncf(si, irn);
1413 ret = sched_next(ret);
1414 ret = sched_prev_op(ret);
1416 ret = sched_prev_op(irn);
1419 if(sched_is_end(ret) || is_Phi(ret))
1422 op = (op_t*)get_irn_link(ret);
1423 if(op->is_remat && op->attr.remat.pre) {
1431 * Find a remat of value @p value in the epilog of @p pos
1434 find_post_remat(const ir_node * value, const ir_node * pos)
1436 while((pos = next_post_remat(pos)) != NULL) {
1439 op = get_irn_link(pos);
1440 assert(op->is_remat && !op->attr.remat.pre);
1442 if(op->attr.remat.remat->value == value)
1443 return (ir_node*)pos;
1446 const ir_edge_t *edge;
1447 foreach_out_edge(pos, edge) {
1448 ir_node *proj = get_edge_src_irn(edge);
1449 assert(is_Proj(proj));
1459 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1461 spill_bb_t *spill_bb = get_irn_link(bb);
1465 int default_spilled;
1468 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1470 double spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1472 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1474 spill->reg_out = ILP_UNDEF;
1475 spill->reg_in = ILP_UNDEF;
1476 spill->mem_in = ILP_UNDEF;
1478 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1479 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1481 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1482 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1483 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1490 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1495 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1496 irn = be_lv_get_irn(si->lv, bb, i);
1498 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1499 pset_insert_ptr(live, irn);
1503 irn = sched_last(bb);
1505 /* all values eaten by control flow operations are also live until the end of the block */
1506 sched_foreach_reverse(bb, irn) {
1509 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1511 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1512 ir_node *arg = get_irn_n(irn,i);
1514 if(has_reg_class(si, arg)) {
1515 pset_insert_ptr(live, arg);
1522 * Inserts ILP-constraints and variables for memory copying before the given position
1525 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1527 const ir_node *succ;
1528 const ir_edge_t *edge;
1529 spill_bb_t *spill_bb = get_irn_link(block);
1538 assert(edges_activated(current_ir_graph));
1540 edge = get_block_succ_first(block);
1546 edge = get_block_succ_next(block, edge);
1547 /* next block can only contain phis, if this is a merge edge */
1550 ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1551 copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1553 ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1554 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1556 pset_foreach(live, tmp) {
1559 op_t *op = get_irn_link(irn);
1560 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1562 spill = set_find_spill(spill_bb->ilp, tmp);
1565 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1567 lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1569 sched_foreach(succ, phi) {
1570 const ir_node *to_copy;
1572 spill_t *to_copy_spill;
1573 op_t *phi_op = get_irn_link(phi);
1574 ilp_var_t reload = ILP_UNDEF;
1577 if(!is_Phi(phi)) break;
1578 if(!has_reg_class(si, phi)) continue;
1580 to_copy = get_irn_n(phi, pos);
1582 to_copy_op = get_irn_link(to_copy);
1584 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1585 assert(to_copy_spill);
1587 if(spill_bb->reloads) {
1588 keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1591 reload = PTR_TO_INT(keyval->val);
1595 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1596 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1598 /* copy - reg_out - reload - remat - live_range <= 0 */
1599 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1600 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1601 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1602 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1603 foreach_pre_remat(si, block, tmp) {
1604 op_t *remat_op = get_irn_link(tmp);
1605 if(remat_op->attr.remat.remat->value == to_copy) {
1606 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1610 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1611 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1613 /* copy - reg_out - copyreg <= 0 */
1614 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1615 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1616 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1622 * Walk all irg blocks and emit this ILP
1625 luke_blockwalker(ir_node * bb, void * data)
1627 spill_ilp_t *si = (spill_ilp_t*)data;
1632 spill_bb_t *spill_bb = get_irn_link(bb);
1635 pset *defs = pset_new_ptr_default();
1636 #ifdef WITH_MEMOPERANDS
1637 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
1641 live = pset_new_ptr_default();
1643 /****************************************
1644 * B A S I C B L O C K E N D
1645 ***************************************/
1648 /* init live values at end of block */
1649 get_live_end(si, bb, live);
1651 pset_foreach(live, irn) {
1653 ilp_var_t reload = ILP_UNDEF;
1655 spill = set_find_spill(spill_bb->ilp, irn);
1658 if(spill_bb->reloads) {
1659 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1662 reload = PTR_TO_INT(keyval->val);
1666 op = get_irn_link(irn);
1667 assert(!op->is_remat);
1669 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1670 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1671 op->attr.live_range.op = bb;
1673 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1674 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1676 /* reg_out - reload - remat - live_range <= 0 */
1677 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1678 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1679 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1680 foreach_pre_remat(si, bb, tmp) {
1681 op_t *remat_op = get_irn_link(tmp);
1682 if(remat_op->attr.remat.remat->value == irn) {
1683 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1686 /* maybe we should also assure that reg_out >= live_range etc. */
1689 #ifndef NO_MEMCOPIES
1690 insert_mem_copy_position(si, live, bb);
1694 * start new live ranges for values used by remats at end of block
1695 * and assure the remat args are available
1697 foreach_pre_remat(si, bb, tmp) {
1698 op_t *remat_op = get_irn_link(tmp);
1701 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1702 ir_node *remat_arg = get_irn_n(tmp, n);
1703 op_t *arg_op = get_irn_link(remat_arg);
1706 if(!has_reg_class(si, remat_arg)) continue;
1708 /* if value is becoming live through use by remat */
1709 if(!pset_find_ptr(live, remat_arg)) {
1710 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1711 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1713 arg_op->attr.live_range.ilp = prev_lr;
1714 arg_op->attr.live_range.op = bb;
1716 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1718 pset_insert_ptr(live, remat_arg);
1719 add_to_spill_bb(si, bb, remat_arg);
1722 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1723 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1724 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1726 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1727 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1729 /* use reload placed for this argument */
1730 if(spill_bb->reloads) {
1731 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1734 ilp_var_t reload = PTR_TO_INT(keyval->val);
1736 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1741 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1746 /**************************************
1747 * B A S I C B L O C K B O D Y
1748 **************************************/
1750 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1756 ilp_cst_t check_pre,
1762 #ifdef WITH_MEMOPERANDS
1763 ilp_cst_t one_memoperand;
1766 /* iterate only until first phi */
1770 op = get_irn_link(irn);
1772 if(op->is_remat) continue;
1773 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1775 /* collect defined values */
1776 if(has_reg_class(si, irn)) {
1777 pset_insert_ptr(defs, irn);
1781 if(is_Proj(irn)) continue;
1784 * init set of irn's arguments
1785 * and all possibly used values around this op
1786 * and values defined by post remats
1788 args = new_set(cmp_keyval, get_irn_arity(irn));
1789 used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1790 remat_defs = pset_new_ptr(pset_count(live));
1792 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1793 ir_node *irn_arg = get_irn_n(irn, n);
1794 if(has_reg_class(si, irn_arg)) {
1795 set_insert_keyval(args, irn_arg, (void*)n);
1796 pset_insert_ptr(used, irn_arg);
1799 foreach_post_remat(irn, tmp) {
1800 op_t *remat_op = get_irn_link(tmp);
1802 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1804 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1805 ir_node *remat_arg = get_irn_n(tmp, n);
1806 if(has_reg_class(si, remat_arg)) {
1807 pset_insert_ptr(used, remat_arg);
1811 foreach_pre_remat(si, irn, tmp) {
1812 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1813 ir_node *remat_arg = get_irn_n(tmp, n);
1814 if(has_reg_class(si, remat_arg)) {
1815 pset_insert_ptr(used, remat_arg);
1820 /**********************************
1821 * I N E P I L O G O F irn
1822 **********************************/
1824 /* ensure each dying value is used by only one post remat */
1825 pset_foreach(used, tmp) {
1826 ir_node *value = tmp;
1827 op_t *value_op = get_irn_link(value);
1832 foreach_post_remat(irn, remat) {
1833 op_t *remat_op = get_irn_link(remat);
1835 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1836 ir_node *remat_arg = get_irn_n(remat, n);
1838 /* if value is used by this remat add it to constraint */
1839 if(remat_arg == value) {
1841 /* sum remat2s <= 1 + n_remats*live_range */
1842 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1843 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
1847 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1853 // value_op->attr.live_range.ilp != ILP_UNDEF
1854 if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
1855 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1859 /* ensure at least one value dies at post remat */
1860 foreach_post_remat(irn, tmp) {
1861 op_t *remat_op = get_irn_link(tmp);
1862 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
1865 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
1866 remat_arg = get_irn_n(tmp, n);
1868 if(has_reg_class(si, remat_arg)) {
1870 /* does arg always die at this op? */
1871 if(!pset_find_ptr(live, remat_arg))
1872 goto skip_one_must_die;
1874 pset_insert_ptr(remat_args, remat_arg);
1878 /* remat + \sum live_range(remat_arg) <= |args| */
1879 ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
1880 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
1881 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1883 pset_foreach(remat_args, remat_arg) {
1884 op_t *arg_op = get_irn_link(remat_arg);
1886 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1890 del_pset(remat_args);
1893 /* new live ranges for values from L\U defined by post remats */
1894 pset_foreach(live, tmp) {
1895 ir_node *value = tmp;
1896 op_t *value_op = get_irn_link(value);
1898 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
1899 ilp_var_t prev_lr = ILP_UNDEF;
1902 if(pset_find_ptr(remat_defs, value)) {
1904 /* next_live_range <= prev_live_range + sum remat2s */
1905 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1906 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1908 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1909 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1911 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1912 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1914 foreach_post_remat(irn, remat) {
1915 op_t *remat_op = get_irn_link(remat);
1917 /* if value is being rematerialized by this remat */
1918 if(value == remat_op->attr.remat.remat->value) {
1919 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1923 value_op->attr.live_range.ilp = prev_lr;
1924 value_op->attr.live_range.op = irn;
1929 /* requirements for post remats and start live ranges from L/U' for values dying here */
1930 foreach_post_remat(irn, tmp) {
1931 op_t *remat_op = get_irn_link(tmp);
1934 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1935 ir_node *remat_arg = get_irn_n(tmp, n);
1936 op_t *arg_op = get_irn_link(remat_arg);
1938 if(!has_reg_class(si, remat_arg)) continue;
1940 /* only for values in L\U (TODO and D?), the others are handled with post_use */
1941 if(!pset_find_ptr(used, remat_arg)) {
1942 /* remat <= live_range(remat_arg) */
1943 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
1944 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1946 /* if value is becoming live through use by remat2 */
1947 if(!pset_find_ptr(live, remat_arg)) {
1950 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1951 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1953 arg_op->attr.live_range.ilp = lr;
1954 arg_op->attr.live_range.op = irn;
1956 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
1958 pset_insert_ptr(live, remat_arg);
1959 add_to_spill_bb(si, bb, remat_arg);
1962 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1963 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1968 d = pset_count(defs);
1969 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
1971 /* count how many regs irn needs for arguments */
1972 u = set_count(args);
1975 /* check the register pressure in the epilog */
1976 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
1977 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
1978 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
1980 /* add L\U' to check_post */
1981 pset_foreach(live, tmp) {
1982 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
1983 /* if a live value is not used by irn */
1984 tmp_op = get_irn_link(tmp);
1985 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
1989 /***********************************************************
1990 * 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
1991 **********************************************************/
1994 pset_foreach(used, tmp) {
2000 op_t *arg_op = get_irn_link(arg);
2003 spill = add_to_spill_bb(si, bb, arg);
2005 /* new live range for each used value */
2006 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2007 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2009 /* the epilog stuff - including post_use, check_post, check_post_remat */
2010 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2011 post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2013 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2015 /* arg is live throughout epilog if the next live_range is in a register */
2016 if(pset_find_ptr(live, arg)) {
2017 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
2019 /* post_use >= next_lr + remat */
2020 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2021 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2022 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2023 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2027 /* if value is not an arg of op and not possibly defined by post remat
2028 * then it may only die and not become live
2030 if(!set_find_keyval(args, arg)) {
2031 /* post_use <= prev_lr */
2032 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2033 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2034 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2035 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2037 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2038 /* next_lr <= prev_lr */
2039 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2040 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2041 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2042 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2047 /* forall post remat which use arg add a similar cst */
2048 foreach_post_remat(irn, remat) {
2051 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2052 ir_node *remat_arg = get_irn_n(remat, n);
2053 op_t *remat_op = get_irn_link(remat);
2055 if(remat_arg == arg) {
2056 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
2058 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2059 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2060 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2061 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2066 #ifdef WITH_MEMOPERANDS
2067 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2068 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2069 ilp_var_t memoperand;
2071 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2072 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, COST_MEMOPERAND*execution_frequency(si, bb), 0.0);
2073 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2075 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2076 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2078 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2079 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2080 // if(arg_op->attr.live_range.ilp != ILP_UNDEF)
2081 // lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2086 /* new live range begins for each used value */
2087 arg_op->attr.live_range.ilp = prev_lr;
2088 arg_op->attr.live_range.op = irn;
2090 /*if(!pset_find_ptr(live, arg)) {
2091 pset_insert_ptr(live, arg);
2092 add_to_spill_bb(si, bb, arg);
2094 pset_insert_ptr(live, arg);
2098 /* just to be sure */
2099 check_post = ILP_UNDEF;
2108 /* check the register pressure in the prolog */
2109 /* sum_{L\U} lr <= k - |U| */
2110 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2111 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2113 /* for the prolog remove defined values from the live set */
2114 pset_foreach(defs, tmp) {
2115 pset_remove_ptr(live, tmp);
2118 #ifdef WITH_MEMOPERANDS
2119 ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2120 one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2123 /***********************************************************
2124 * 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
2125 **********************************************************/
2128 set_foreach(args, keyval) {
2130 const ir_node *arg = keyval->key;
2131 int i = PTR_TO_INT(keyval->val);
2132 op_t *arg_op = get_irn_link(arg);
2133 ilp_cst_t requirements;
2134 #ifdef WITH_MEMOPERANDS
2138 spill = set_find_spill(spill_bb->ilp, arg);
2141 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2142 op->attr.live_range.args.reloads[i] = lpp_add_var_default(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb), 1.0);
2144 /* reload <= mem_out */
2145 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2146 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2147 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2148 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2150 /* requirement: arg must be in register for use */
2151 /* reload + remat + live_range == 1 */
2152 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2153 requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2155 lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2156 lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2157 foreach_pre_remat(si, irn, tmp) {
2158 op_t *remat_op = get_irn_link(tmp);
2159 if(remat_op->attr.remat.remat->value == arg) {
2160 lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2164 #ifdef WITH_MEMOPERANDS
2166 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2167 if(get_irn_n(irn, n) == arg) {
2171 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2172 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2173 memoperand_t *memoperand;
2174 memoperand = set_find_memoperand(si->memoperands, irn, n);
2176 /* memoperand <= mem_out */
2177 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2178 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2179 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2180 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2182 /* the memoperand is only sufficient if it is used once by the op */
2183 if(n_memoperands == 1)
2184 lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2186 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2188 /* we have one more free register if we use a memory operand */
2189 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2195 /* iterate over L\U */
2196 pset_foreach(live, tmp) {
2197 if(!set_find_keyval(args, tmp)) {
2198 /* if a live value is not used by irn */
2199 tmp_op = get_irn_link(tmp);
2200 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2205 /* requirements for remats */
2206 /* start new live ranges for values used by remats */
2207 foreach_pre_remat(si, irn, tmp) {
2208 op_t *remat_op = get_irn_link(tmp);
2211 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2212 ir_node *remat_arg = get_irn_n(tmp, n);
2213 op_t *arg_op = get_irn_link(remat_arg);
2215 if(!has_reg_class(si, remat_arg)) continue;
2217 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2218 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2219 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2221 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2222 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2224 /* if remat arg is also used by current op then we can use reload placed for this argument */
2225 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2226 int index = (int)keyval->val;
2228 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2236 /*************************
2237 * D O N E W I T H O P
2238 *************************/
2240 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2242 pset_foreach(live, tmp) {
2243 assert(has_reg_class(si, tmp));
2246 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2247 ir_node *arg = get_irn_n(irn, n);
2249 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2252 del_pset(remat_defs);
2256 defs = pset_new_ptr_default();
2261 /***************************************
2262 * B E G I N N I N G O F B L O C K
2263 ***************************************/
2266 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2267 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2269 pset_foreach(live, irn) {
2270 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2273 /* construct mem_outs for all values */
2275 set_foreach(spill_bb->ilp, spill) {
2276 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2277 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2279 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2280 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2282 if(pset_find_ptr(live, spill->irn)) {
2283 int default_spilled;
2284 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2286 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2287 default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2288 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2289 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2291 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2293 op_t *op = get_irn_link(spill->irn);
2295 /* do we have to copy a phi argument? */
2296 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2297 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2299 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2300 const ir_node *arg = get_irn_n(spill->irn, n);
2306 /* argument already done? */
2307 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2309 /* get sum of execution frequencies of blocks with the same phi argument */
2310 for(m=n; m>=0; --m) {
2311 const ir_node *arg2 = get_irn_n(spill->irn, m);
2314 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2318 /* copies are not for free */
2319 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2320 var = lpp_add_var_default(si->lpp, buf, lpp_binary, COST_STORE * freq, 1.0);
2322 for(m=n; m>=0; --m) {
2323 const ir_node *arg2 = get_irn_n(spill->irn, m);
2326 op->attr.live_range.args.copies[m] = var;
2331 /* copy <= mem_in */
2332 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2333 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2334 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2335 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2343 /* L\U is empty at bb start */
2344 /* arg is live throughout epilog if it is reg_in into this block */
2346 /* check the register pressure at the beginning of the block
2349 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2350 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2352 pset_foreach(live, irn) {
2355 spill = set_find_spill(spill_bb->ilp, irn);
2358 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2359 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2361 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2363 /* spill + mem_in <= 1 */
2364 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2365 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2367 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2368 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2371 foreach_post_remat(bb, irn) {
2372 op_t *remat_op = get_irn_link(irn);
2374 DBG((si->dbg, LEVEL_4, "\t next post remat: %+F\n", irn));
2375 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
2377 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2380 /* forall post remats add requirements */
2381 foreach_post_remat(bb, tmp) {
2384 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2385 ir_node *remat_arg = get_irn_n(tmp, n);
2386 op_t *remat_op = get_irn_link(tmp);
2388 if(!has_reg_class(si, remat_arg)) continue;
2390 spill = set_find_spill(spill_bb->ilp, remat_arg);
2393 /* remat <= reg_in_argument */
2394 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2395 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2396 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2397 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2401 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2402 pset_foreach(live, irn) {
2406 spill = set_find_spill(spill_bb->ilp, irn);
2407 assert(spill && spill->irn == irn);
2409 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2410 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2413 ir_node *phi_arg = get_Phi_pred(irn, n);
2414 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2415 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2417 op_t *op = get_irn_link(irn);
2419 /* although the phi is in the right regclass one or more of
2420 * its arguments can be in a different one or at least to
2423 if(has_reg_class(si, phi_arg)) {
2424 /* mem_in < mem_out_arg + copy */
2425 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2426 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2428 /* reg_in < reg_out_arg */
2429 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2430 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2432 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2433 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2435 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2438 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2439 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2440 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2444 /* else assure the value arrives on all paths in the same resource */
2446 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2449 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2450 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2453 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2454 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2455 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2456 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2458 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2459 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2461 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2464 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2465 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2470 /* first live ranges from reg_ins */
2471 pset_foreach(live, irn) {
2472 op_t *op = get_irn_link(irn);
2474 spill = set_find_spill(spill_bb->ilp, irn);
2475 assert(spill && spill->irn == irn);
2477 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2478 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2479 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2480 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2482 foreach_post_remat(bb, tmp) {
2483 op_t *remat_op = get_irn_link(tmp);
2485 if(remat_op->attr.remat.remat->value == irn) {
2486 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2491 /* walk forward now and compute constraints for placing spills */
2492 /* this must only be done for values that are not defined in this block */
2493 /* TODO are these values at start of block? if yes, just check whether this is a diverge edge and skip the loop */
2494 pset_foreach(live, irn) {
2496 * if value is defined in this block we can anways place the spill directly after the def
2497 * -> no constraint necessary
2499 if(!is_Phi(irn) && get_nodes_block(irn) == bb) continue;
2502 spill = set_find_spill(spill_bb->ilp, irn);
2505 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2506 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2508 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2509 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2512 sched_foreach_op(bb, tmp) {
2513 op_t *op = get_irn_link(tmp);
2515 if(is_Phi(tmp)) continue;
2516 assert(!is_Proj(tmp));
2519 const ir_node *value = op->attr.remat.remat->value;
2522 /* only collect remats up to the first real use of a value */
2523 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2528 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2529 ir_node *arg = get_irn_n(tmp, n);
2532 /* if a value is used stop collecting remats */
2545 typedef struct _irnlist_t {
2546 struct list_head list;
2550 typedef struct _interference_t {
2551 struct list_head blocklist;
2557 cmp_interference(const void *a, const void *b, size_t size)
2559 const interference_t *p = a;
2560 const interference_t *q = b;
2562 return !(p->a == q->a && p->b == q->b);
2565 static interference_t *
2566 set_find_interference(set * set, ir_node * a, ir_node * b)
2568 interference_t query;
2570 query.a = (a>b)?a:b;
2571 query.b = (a>b)?b:a;
2573 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2576 static interference_t *
2577 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2579 interference_t query,
2581 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2585 result = set_find_interference(set, a, b);
2588 list_add(&list->list, &result->blocklist);
2592 query.a = (a>b)?a:b;
2593 query.b = (a>b)?b:a;
2595 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2597 INIT_LIST_HEAD(&result->blocklist);
2598 list_add(&list->list, &result->blocklist);
2604 values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
2606 const ir_edge_t *edge;
2608 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2609 /* both values are live in, so they interfere */
2613 /* ensure a dominates b */
2614 if(value_dominates(b,a)) {
2620 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2623 /* the following code is stolen from bera.c */
2624 if(be_is_live_end(si->lv, bb, a))
2627 foreach_out_edge(a, edge) {
2628 const ir_node *user = edge->src;
2629 if(get_nodes_block(user) == bb
2632 && value_dominates(b, user))
2640 * Walk all irg blocks and collect interfering values inside of phi classes
2643 luke_interferencewalker(ir_node * bb, void * data)
2645 spill_ilp_t *si = (spill_ilp_t*)data;
2648 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2649 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
2650 op_t *a_op = get_irn_link(a);
2653 /* a is only interesting if it is in my register class and if it is inside a phi class */
2654 if (has_reg_class(si, a) && get_phi_class(a)) {
2658 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)) {
2659 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
2660 op_t *b_op = get_irn_link(b);
2663 /* a and b are only interesting if they are in the same phi class */
2664 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2668 if(values_interfere_in_block(si, bb, a, b)) {
2669 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2670 set_insert_interference(si, si->interferences, a, b, bb);
2678 static unsigned int copy_path_id = 0;
2681 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2688 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2689 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2691 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2693 pset_foreach(copies, ptr) {
2694 copy = PTR_TO_INT(ptr);
2695 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2700 * @parameter copies contains a path of copies which lead us to irn
2701 * @parameter visited contains a set of nodes already visited on this path
2704 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2706 const ir_edge_t *edge;
2707 op_t *op = get_irn_link(irn);
2708 pset *visited_users = pset_new_ptr_default();
2711 if(op->is_remat) return 0;
2713 pset_insert_ptr(visited, irn);
2717 pset *visited_operands = pset_new_ptr(get_irn_arity(irn));
2719 /* visit all operands */
2720 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2721 ir_node *arg = get_irn_n(irn, n);
2722 ilp_var_t copy = op->attr.live_range.args.copies[n];
2724 if(!has_reg_class(si, arg)) continue;
2725 if(pset_find_ptr(visited_operands, arg)) continue;
2726 pset_insert_ptr(visited_operands, arg);
2729 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2730 del_pset(visited_operands);
2731 del_pset(visited_users);
2732 pset_remove_ptr(visited, irn);
2735 pset_insert(copies, INT_TO_PTR(copy), copy);
2736 write_copy_path_cst(si, copies, any_interfere);
2737 pset_remove(copies, INT_TO_PTR(copy), copy);
2738 } else if(!pset_find_ptr(visited, arg)) {
2739 pset_insert(copies, INT_TO_PTR(copy), copy);
2740 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
2741 pset_remove(copies, INT_TO_PTR(copy), copy);
2743 /*if(paths > MAX_PATHS) {
2744 if(pset_count(copies) == 0) {
2748 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2749 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2750 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2751 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2752 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2756 del_pset(visited_operands);
2757 del_pset(visited_users);
2758 pset_remove_ptr(visited, irn);
2761 } else if(pset_count(copies) == 0) {
2767 del_pset(visited_operands);
2770 /* visit all uses which are phis */
2771 foreach_out_edge(irn, edge) {
2772 ir_node *user = edge->src;
2773 int pos = edge->pos;
2774 op_t *op = get_irn_link(user);
2777 if(!is_Phi(user)) continue;
2778 if(!has_reg_class(si, user)) continue;
2779 if(pset_find_ptr(visited_users, user)) continue;
2780 pset_insert_ptr(visited_users, user);
2782 copy = op->attr.live_range.args.copies[pos];
2784 if(user == target) {
2785 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2786 del_pset(visited_users);
2787 pset_remove_ptr(visited, irn);
2790 pset_insert(copies, INT_TO_PTR(copy), copy);
2791 write_copy_path_cst(si, copies, any_interfere);
2792 pset_remove(copies, INT_TO_PTR(copy), copy);
2793 } else if(!pset_find_ptr(visited, user)) {
2794 pset_insert(copies, INT_TO_PTR(copy), copy);
2795 paths += find_copy_path(si, user, target, any_interfere, copies, visited);
2796 pset_remove(copies, INT_TO_PTR(copy), copy);
2798 /*if(paths > MAX_PATHS) {
2799 if(pset_count(copies) == 0) {
2803 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2804 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2805 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2806 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2807 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2811 del_pset(visited_users);
2812 pset_remove_ptr(visited, irn);
2815 } else if(pset_count(copies) == 0) {
2821 del_pset(visited_users);
2822 pset_remove_ptr(visited, irn);
2827 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
2829 pset * copies = pset_new_ptr_default();
2830 pset * visited = pset_new_ptr_default();
2832 find_copy_path(si, a, b, any_interfere, copies, visited);
2840 memcopyhandler(spill_ilp_t * si)
2842 interference_t *interference;
2844 /* teste Speicherwerte auf Interferenz */
2846 /* analyze phi classes */
2847 phi_class_compute(si->chordal_env->irg);
2849 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
2850 irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
2852 /* now lets emit the ILP unequations for the crap */
2853 set_foreach(si->interferences, interference) {
2855 ilp_var_t interfere,
2857 ilp_cst_t any_interfere_cst,
2859 const ir_node *a = interference->a;
2860 const ir_node *b = interference->b;
2862 /* any_interf <= \sum interf */
2863 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
2864 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2865 any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2867 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
2869 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
2870 const ir_node *bb = irnlist->irn;
2871 spill_bb_t *spill_bb = get_irn_link(bb);
2876 spilla = set_find_spill(spill_bb->ilp, a);
2879 spillb = set_find_spill(spill_bb->ilp, b);
2882 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
2883 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
2884 /* 2: - mem_in_a - spill_a + interfere <= 0 */
2885 /* 3: - mem_in_b - spill_b + interfere <= 0 */
2886 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
2887 interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2889 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
2890 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2892 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
2893 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
2894 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
2895 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
2896 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
2898 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
2899 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2901 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2902 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
2903 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
2905 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
2906 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2908 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2909 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
2910 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
2913 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
2915 /* any_interfere >= interf */
2916 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
2917 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2919 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2920 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2923 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
2924 gen_copy_constraints(si,a,b,any_interfere);
2932 return fabs(x) < 0.00001;
2936 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
2938 spill_ilp_t *si = get_irg_link(current_ir_graph);
2940 if(pset_find_ptr(si->all_possible_remats, n)) {
2941 op_t *op = (op_t*)get_irn_link(n);
2942 assert(op && op->is_remat);
2944 if(!op->attr.remat.remat->inverse) {
2945 if(op->attr.remat.pre) {
2946 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
2948 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
2953 op_t *op = (op_t*)get_irn_link(n);
2954 assert(op && op->is_remat);
2956 if(op->attr.remat.pre) {
2957 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
2959 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
2970 dump_graph_with_remats(ir_graph * irg, const char * suffix)
2972 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
2973 be_dump(irg, suffix, dump_ir_block_graph_sched);
2974 set_dump_node_vcgattr_hook(NULL);
2979 * Edge hook to dump the schedule edges with annotated register pressure.
2982 sched_pressure_edge_hook(FILE *F, ir_node *irn)
2984 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
2985 ir_node *prev = sched_prev(irn);
2986 fprintf(F, "edge:{sourcename:\"");
2988 fprintf(F, "\" targetname:\"");
2990 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
2991 fprintf(F, "\" color:magenta}\n");
2997 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
2999 DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3001 dump_consts_local(0);
3002 set_dump_node_edge_hook(sched_pressure_edge_hook);
3003 dump_ir_block_graph(irg, suffix);
3004 set_dump_node_edge_hook(old_edge_hook);
3008 walker_pressure_annotator(ir_node * bb, void * data)
3010 spill_ilp_t *si = data;
3013 pset *live = pset_new_ptr_default();
3016 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3017 irn = be_lv_get_irn(si->lv, bb, i);
3019 if (has_reg_class(si, irn)) {
3020 pset_insert_ptr(live, irn);
3024 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3026 sched_foreach_reverse(bb, irn) {
3028 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3032 if(has_reg_class(si, irn)) {
3033 pset_remove_ptr(live, irn);
3034 if(is_Proj(irn)) ++projs;
3037 if(!is_Proj(irn)) projs = 0;
3039 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3040 ir_node *arg = get_irn_n(irn, n);
3042 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3044 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3051 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3053 be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
3058 connect_all_remats_with_keep(spill_ilp_t * si)
3066 n_remats = pset_count(si->all_possible_remats);
3068 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3071 pset_foreach(si->all_possible_remats, irn) {
3076 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
3078 obstack_free(si->obst, ins);
3084 connect_all_spills_with_keep(spill_ilp_t * si)
3093 n_spills = pset_count(si->spills);
3095 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3098 pset_foreach(si->spills, irn) {
3103 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
3105 obstack_free(si->obst, ins);
3109 /** insert a spill at an arbitrary position */
3110 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3112 ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
3113 ir_graph *irg = get_irn_irg(bl);
3114 ir_node *frame = get_irg_frame(irg);
3118 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
3119 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3121 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3124 * search the right insertion point. a spill of a phi cannot be put
3125 * directly after the phi, if there are some phis behind the one which
3126 * is spilled. Also, a spill of a Proj must be after all Projs of the
3129 * Here's one special case:
3130 * If the spill is in the start block, the spill must be after the frame
3131 * pointer is set up. This is done by setting insert to the end of the block
3132 * which is its default initialization (see above).
3135 if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3138 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3141 sched_add_after(insert, spill);
3146 delete_remat(spill_ilp_t * si, ir_node * remat) {
3148 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3150 sched_remove(remat);
3152 /* kill links to operands */
3153 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3154 set_irn_n(remat, n, bad);
3159 clean_remat_info(spill_ilp_t * si)
3163 remat_info_t *remat_info;
3164 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3166 set_foreach(si->remat_info, remat_info) {
3167 if(!remat_info->remats) continue;
3169 pset_foreach(remat_info->remats, remat)
3171 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3172 set_irn_n((ir_node*)remat->proj, -1, bad);
3173 set_irn_n((ir_node*)remat->proj, 0, bad);
3176 if(get_irn_n_edges(remat->op) == 0) {
3177 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3178 set_irn_n((ir_node*)remat->op, n, bad);
3183 if(remat_info->remats) del_pset(remat_info->remats);
3184 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3189 delete_unnecessary_remats(spill_ilp_t * si)
3193 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3196 ir_node *end = get_irg_end(si->chordal_env->irg);
3199 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3200 ir_node *keep_arg = get_irn_n(si->keep, n);
3201 op_t *arg_op = get_irn_link(keep_arg);
3204 assert(arg_op->is_remat);
3206 name = si->lpp->vars[arg_op->attr.remat.ilp];
3208 if(is_zero(name->value)) {
3209 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
3210 /* TODO check whether reload is preferred over remat (could be bug) */
3211 delete_remat(si, keep_arg);
3213 if(!arg_op->attr.remat.remat->inverse) {
3214 if(arg_op->attr.remat.pre) {
3215 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3217 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3220 if(arg_op->attr.remat.pre) {
3221 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3223 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3228 set_irn_n(si->keep, n, bad);
3231 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
3232 ir_node *end_arg = get_End_keepalive(end, i);
3234 if(end_arg != si->keep) {
3235 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
3238 keeps = obstack_finish(si->obst);
3239 set_End_keepalives(end, n-1, keeps);
3240 obstack_free(si->obst, keeps);
3243 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
3248 pset_foreach(si->all_possible_remats, remat) {
3249 op_t *remat_op = get_irn_link(remat);
3250 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
3252 if(is_zero(name->value)) {
3253 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
3254 /* TODO check whether reload is preferred over remat (could be bug) */
3255 delete_remat(si, remat);
3257 if(!remat_op->attr.remat.remat->inverse) {
3258 if(remat_op->attr.remat.pre) {
3259 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3261 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3264 if(remat_op->attr.remat.pre) {
3265 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3267 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3276 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3278 pset *spills = pset_new_ptr_default();
3280 const ir_node *next;
3283 defs = set_find_def(si->values, value);
3285 if(defs && defs->spills) {
3286 for(next = defs->spills; next; next = get_irn_link(next)) {
3287 pset_insert_ptr(spills, next);
3295 * @param before The node after which the spill will be placed in the schedule
3297 /* TODO set context properly */
3299 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3303 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3305 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3307 spill = be_spill2(arch_env, irn, before);
3309 defs = set_insert_def(si->values, value);
3312 /* enter into the linked list */
3313 set_irn_link(spill, defs->spills);
3314 defs->spills = spill;
3316 #ifdef KEEPALIVE_SPILLS
3317 pset_insert_ptr(si->spills, spill);
3324 * @param before The Phi node which has to be spilled
3327 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3334 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3336 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3337 ins[n] = si->m_unknown;
3340 mem_phi = new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3342 defs = set_insert_def(si->values, phi);
3345 /* enter into the linked list */
3346 set_irn_link(mem_phi, defs->spills);
3347 defs->spills = mem_phi;
3349 sched_add_after(phi, mem_phi);
3351 #ifdef KEEPALIVE_SPILLS
3352 pset_insert_ptr(si->spills, mem_phi);
3360 * Add remat to list of defs, destroys link field!
3363 insert_remat(spill_ilp_t * si, ir_node * remat)
3366 op_t *remat_op = get_irn_link(remat);
3368 assert(remat_op->is_remat);
3370 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3373 /* enter into the linked list */
3374 set_irn_link(remat, defs->remats);
3375 defs->remats = remat;
3380 * Add reload before operation and add to list of defs
3383 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3388 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3390 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3392 defs = set_find_def(si->values, value);
3394 spill = defs->spills;
3395 assert(spill && "no spill placed before reload");
3397 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3399 /* enter into the linked list */
3400 set_irn_link(reload, defs->remats);
3401 defs->remats = reload;
3406 #ifdef WITH_MEMOPERANDS
3407 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3411 ir_node *value = get_irn_n(memoperand->irn, memoperand->pos);
3413 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3415 DBG((si->dbg, LEVEL_2, "\t inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3417 defs = set_find_def(si->values, value);
3419 spill = defs->spills;
3420 assert(spill && "no spill placed before reload");
3422 reload = be_reload(arch_env, si->cls, memoperand->irn, get_irn_mode(value), spill);
3424 arch_perform_memory_operand(arch_env, memoperand->irn, reload, memoperand->pos);
3425 sched_remove(reload);
3428 void insert_memoperands(spill_ilp_t * si)
3430 memoperand_t *memoperand;
3433 set_foreach(si->memoperands, memoperand) {
3434 name = si->lpp->vars[memoperand->ilp];
3435 if(!is_zero(name->value)) {
3436 perform_memory_operand(si, memoperand);
3443 walker_spill_placer(ir_node * bb, void * data) {
3444 spill_ilp_t *si = (spill_ilp_t*)data;
3446 spill_bb_t *spill_bb = get_irn_link(bb);
3447 pset *spills_to_do = pset_new_ptr_default();
3450 set_foreach(spill_bb->ilp, spill) {
3453 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3454 name = si->lpp->vars[spill->mem_in];
3455 if(!is_zero(name->value)) {
3458 mem_phi = insert_mem_phi(si, spill->irn);
3460 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3464 name = si->lpp->vars[spill->spill];
3465 if(!is_zero(name->value)) {
3466 /* place spill directly after definition */
3467 if(get_nodes_block(spill->irn) == bb) {
3468 insert_spill(si, spill->irn, spill->irn, spill->irn);
3472 /* place spill at bb start */
3473 if(spill->reg_in > 0) {
3474 name = si->lpp->vars[spill->reg_in];
3475 if(!is_zero(name->value)) {
3476 insert_spill(si, spill->irn, spill->irn, bb);
3480 /* place spill after a remat */
3481 pset_insert_ptr(spills_to_do, spill->irn);
3484 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3487 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3488 op_t *op = get_irn_link(irn);
3490 if(be_is_Spill(irn)) continue;
3493 /* TODO fix this if we want to support remats with more than two nodes */
3494 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3495 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3497 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3500 if(pset_find_ptr(spills_to_do, irn)) {
3501 pset_remove_ptr(spills_to_do, irn);
3503 insert_spill(si, irn, irn, irn);
3509 assert(pset_count(spills_to_do) == 0);
3511 /* afterwards free data in block */
3512 del_pset(spills_to_do);
3516 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3518 ir_node *insert_pos = bb;
3520 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3522 /* find last definition of arg value in block */
3527 defs = set_find_def(si->values, value);
3529 if(defs && defs->remats) {
3530 for(next = defs->remats; next; next = get_irn_link(next)) {
3531 if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3532 last = sched_get_time_step(next);
3538 if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3539 last = sched_get_time_step(value);
3543 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
3545 spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3551 phim_fixer(spill_ilp_t *si) {
3554 set_foreach(si->values, defs) {
3555 const ir_node *phi = defs->value;
3556 op_t *op = get_irn_link(phi);
3557 ir_node *phi_m = NULL;
3558 ir_node *next = defs->spills;
3561 if(!is_Phi(phi)) continue;
3564 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3568 next = get_irn_link(next);
3571 if(!phi_m) continue;
3573 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3574 ir_node *value = get_irn_n(phi, n);
3575 defs_t *val_defs = set_find_def(si->values, value);
3577 /* a spill of this value */
3581 #ifndef NO_MEMCOPIES
3582 ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3583 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3585 if(!is_zero(name->value)) {
3586 spill = insert_mem_copy(si, pred, value);
3588 spill = val_defs->spills;
3591 spill = val_defs->spills;
3593 assert(spill && "no spill placed before PhiM");
3594 set_irn_n(phi_m, n, spill);
3600 walker_reload_placer(ir_node * bb, void * data) {
3601 spill_ilp_t *si = (spill_ilp_t*)data;
3603 spill_bb_t *spill_bb = get_irn_link(bb);
3605 /* reloads at end of block */
3606 if(spill_bb->reloads) {
3609 set_foreach(spill_bb->reloads, keyval) {
3610 ir_node *irn = (ir_node*)keyval->key;
3611 ilp_var_t reload = PTR_TO_INT(keyval->val);
3614 name = si->lpp->vars[reload];
3615 if(!is_zero(name->value)) {
3617 ir_node *insert_pos = bb;
3618 ir_node *prev = sched_block_last_noncf(si, bb);
3619 op_t *prev_op = get_irn_link(prev);
3621 while(be_is_Spill(prev)) {
3622 prev = sched_prev(prev);
3625 prev_op = get_irn_link(prev);
3627 /* insert reload before pre-remats */
3628 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3629 && prev_op->is_remat && prev_op->attr.remat.pre) {
3633 prev = sched_prev(prev);
3634 } while(be_is_Spill(prev));
3636 prev_op = get_irn_link(prev);
3640 reload = insert_reload(si, irn, insert_pos);
3642 #ifdef KEEPALIVE_RELOADS
3643 pset_insert_ptr(si->spills, reload);
3649 /* walk and insert more reloads and collect remats */
3650 sched_foreach_reverse(bb, irn) {
3651 op_t *op = get_irn_link(irn);
3653 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3654 if(is_Phi(irn)) break;
3657 if(get_irn_mode(irn) != mode_T) {
3658 insert_remat(si, irn);
3663 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3664 ir_node *arg = get_irn_n(irn, n);
3666 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3669 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3670 if(!is_zero(name->value)) {
3672 ir_node *insert_pos = irn;
3673 ir_node *prev = sched_prev(insert_pos);
3676 while(be_is_Spill(prev)) {
3677 prev = sched_prev(prev);
3680 prev_op = get_irn_link(prev);
3682 /* insert reload before pre-remats */
3683 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3684 && prev_op->is_remat && prev_op->attr.remat.pre) {
3688 prev = sched_prev(prev);
3689 } while(be_is_Spill(prev));
3691 prev_op = get_irn_link(prev);
3695 reload = insert_reload(si, arg, insert_pos);
3697 set_irn_n(irn, n, reload);
3699 #ifdef KEEPALIVE_RELOADS
3700 pset_insert_ptr(si->spills, reload);
3708 del_set(spill_bb->ilp);
3709 if(spill_bb->reloads) del_set(spill_bb->reloads);
3713 walker_collect_used(ir_node * irn, void * data)
3715 lc_bitset_t *used = data;
3717 lc_bitset_set(used, get_irn_idx(irn));
3720 struct kill_helper {
3726 walker_kill_unused(ir_node * bb, void * data)
3728 struct kill_helper *kh = data;
3729 ir_node *bad = get_irg_bad(get_irn_irg(bb));
3733 for(irn=sched_first(bb); !sched_is_end(irn);) {
3734 ir_node *next = sched_next(irn);
3737 if(!lc_bitset_is_set(kh->used, get_irn_idx(irn))) {
3738 if(be_is_Spill(irn) || be_is_Reload(irn)) {
3739 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)));
3741 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3747 set_nodes_block(irn, bad);
3748 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3749 set_irn_n(irn, n, bad);
3757 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3759 struct kill_helper kh;
3761 kh.used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3764 irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3765 irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3767 lc_bitset_free(kh.used);
3771 print_irn_pset(pset * p)
3775 pset_foreach(p, irn) {
3776 ir_printf("%+F\n", irn);
3781 dump_phi_class(spill_ilp_t * si, pset * phiclass, const char * file)
3783 FILE *f = fopen(file, "w");
3785 interference_t *interference;
3787 pset_break(phiclass);
3788 set_break(si->interferences);
3790 ir_fprintf(f, "digraph phiclass {\n");
3792 pset_foreach(phiclass, irn) {
3794 ir_fprintf(f, " %F%N [shape=box]\n",irn,irn);
3797 pset_foreach(phiclass, irn) {
3800 if(!is_Phi(irn)) continue;
3802 for(n=get_irn_arity(irn)-1; n>=0; --n) {
3803 ir_node *arg = get_irn_n(irn, n);
3805 ir_fprintf(f, " %F%N -> %F%N\n",irn,irn,arg,arg);
3809 set_foreach(si->interferences, interference) {
3810 const ir_node *a = interference->a;
3811 const ir_node *b = interference->b;
3812 if(get_phi_class(a) == phiclass) {
3813 ir_fprintf(f, " %F%N -> %F%N [color=red,dir=none,style=bold]\n",a,a,b,b);
3822 rewire_uses(spill_ilp_t * si)
3824 dom_front_info_t *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
3826 pset *ignore = pset_new_ptr(1);
3828 pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
3830 /* then fix uses of spills */
3831 set_foreach(si->values, defs) {
3834 const ir_node *next = defs->remats;
3837 reloads = pset_new_ptr_default();
3840 if(be_is_Reload(next)) {
3841 pset_insert_ptr(reloads, next);
3845 next = get_irn_link(next);
3848 spills = get_spills_for_value(si, defs->value);
3849 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));
3850 if(pset_count(spills) > 1) {
3851 //assert(pset_count(reloads) > 0);
3852 // print_irn_pset(spills);
3853 // print_irn_pset(reloads);
3855 be_ssa_constr_set_ignore(dfi, si->lv, spills, ignore);
3862 /* first fix uses of remats and reloads */
3863 set_foreach(si->values, defs) {
3865 const ir_node *next = defs->remats;
3868 nodes = pset_new_ptr_default();
3869 pset_insert_ptr(nodes, defs->value);
3872 pset_insert_ptr(nodes, next);
3873 next = get_irn_link(next);
3876 if(pset_count(nodes) > 1) {
3877 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
3878 be_ssa_constr_set(dfi, si->lv, nodes);
3885 // remove_unused_defs(si);
3887 be_free_dominance_frontiers(dfi);
3892 writeback_results(spill_ilp_t * si)
3894 /* walk through the graph and collect all spills, reloads and remats for a value */
3896 si->values = new_set(cmp_defs, 4096);
3898 DBG((si->dbg, LEVEL_1, "Applying results\n"));
3899 delete_unnecessary_remats(si);
3900 si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
3901 irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
3902 irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
3903 #ifdef WITH_MEMOPERANDS
3904 insert_memoperands(si);
3908 /* clean the remat info! there are still back-edges leading there! */
3909 clean_remat_info(si);
3913 connect_all_spills_with_keep(si);
3915 del_set(si->values);
3919 get_n_regs(spill_ilp_t * si)
3921 int arch_n_regs = arch_register_class_n_regs(si->cls);
3925 for(i=0; i<arch_n_regs; i++) {
3926 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
3931 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
3936 walker_reload_mover(ir_node * bb, void * data)
3938 spill_ilp_t *si = data;
3941 sched_foreach(bb, tmp) {
3942 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
3943 ir_node *reload = tmp;
3946 /* move reload upwards */
3948 int pressure = (int)get_irn_link(reload);
3949 if(pressure < si->n_regs) {
3950 irn = sched_prev(reload);
3951 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
3952 sched_remove(reload);
3953 pressure = (int)get_irn_link(irn);
3955 while(pressure < si->n_regs) {
3956 if( sched_is_end(irn) ||
3957 (be_is_Reload(irn) && has_reg_class(si, irn)) ||
3958 /* do not move reload before its spill */
3959 (irn == be_get_Reload_mem(reload)) ||
3960 /* do not move before phi */
3963 set_irn_link(irn, INT_TO_PTR(pressure+1));
3964 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
3965 irn = sched_prev(irn);
3967 pressure = (int)get_irn_link(irn);
3970 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
3971 sched_put_after(irn, reload);
3978 move_reloads_upward(spill_ilp_t * si)
3980 irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
3985 * Walk all irg blocks and check for interfering spills inside of phi classes
3988 luke_meminterferencechecker(ir_node * bb, void * data)
3990 spill_ilp_t *si = (spill_ilp_t*)data;
3993 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
3994 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
3996 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
3998 /* a is only interesting if it is in my register class and if it is inside a phi class */
3999 if (has_reg_class(si, a) && get_phi_class(a)) {
4000 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)) {
4001 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
4003 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
4005 /* a and b are only interesting if they are in the same phi class */
4006 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
4007 if(values_interfere_in_block(si, bb, a, b)) {
4008 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4017 verify_phiclasses(spill_ilp_t * si)
4019 /* analyze phi classes */
4020 phi_class_compute(si->chordal_env->irg);
4022 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4023 irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
4027 walker_spillslotassigner(ir_node * irn, void * data)
4031 if(!be_is_Spill(irn)) return;
4033 /* set spill context to phi class if it has one ;) */
4036 // Matze: not needed anymore
4037 cls = get_phi_class(irn);
4039 be_set_Spill_context(irn, cls);
4041 be_set_Spill_context(irn, irn);
4047 assign_spillslots(spill_ilp_t * si)
4049 DBG((si->dbg, LEVEL_2, "\t calling spill slot assigner\n"));
4050 irg_walk_graph(si->chordal_env->irg, walker_spillslotassigner, NULL, si);
4054 be_spill_remat(const be_chordal_env_t * chordal_env)
4057 char problem_name[256];
4058 char dump_suffix[256];
4059 char dump_suffix2[256];
4060 struct obstack obst;
4063 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
4064 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
4065 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
4067 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4068 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4070 #ifdef VERIFY_DOMINANCE
4071 be_check_dominance(chordal_env->irg);
4074 obstack_init(&obst);
4075 si.chordal_env = chordal_env;
4077 si.cls = chordal_env->cls;
4078 si.lpp = new_lpp(problem_name, lpp_minimize);
4079 si.remat_info = new_set(cmp_remat_info, 4096);
4080 si.interferences = new_set(cmp_interference, 32);
4081 #ifdef WITH_MEMOPERANDS
4082 si.memoperands = new_set(cmp_memoperands, 128);
4084 si.all_possible_remats = pset_new_ptr_default();
4085 si.spills = pset_new_ptr_default();
4086 si.inverse_ops = pset_new_ptr_default();
4087 si.lv = chordal_env->lv;
4091 si.n_regs = get_n_regs(&si);
4093 set_irg_link(chordal_env->irg, &si);
4094 compute_doms(chordal_env->irg);
4096 /* compute phi classes */
4097 // phi_class_compute(chordal_env->irg);
4099 be_analyze_regpressure(chordal_env, "-pre");
4101 #ifdef COLLECT_REMATS
4102 /* collect remats */
4103 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4104 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
4107 /* insert possible remats */
4108 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4109 irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
4110 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4113 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4114 connect_all_remats_with_keep(&si);
4115 /* dump graph with inserted remats */
4116 dump_graph_with_remats(chordal_env->irg, dump_suffix);
4119 /* insert copies for phi arguments not in my regclass */
4120 irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
4122 /* recompute liveness */
4123 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4124 be_liveness_recompute(si.lv);
4128 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4129 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4130 irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
4132 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4133 irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
4135 #ifndef NO_MEMCOPIES
4136 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4137 memcopyhandler(&si);
4140 if(dump_flags & DUMP_PROBLEM) {
4142 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4143 if ((f = fopen(buf, "wt")) != NULL) {
4144 lpp_dump_plain(si.lpp, f);
4149 if(dump_flags & DUMP_MPS) {
4152 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4153 if((f = fopen(buf, "wt")) != NULL) {
4154 mps_write_mps(si.lpp, s_mps_fixed, f);
4158 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4159 if((f = fopen(buf, "wt")) != NULL) {
4160 mps_write_mst(si.lpp, s_mps_fixed, f);
4165 lpp_check_startvals(si.lpp);
4168 DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4170 lpp_set_time_limit(si.lpp, ILP_TIMEOUT);
4174 lpp_set_log(si.lpp, stdout);
4177 lpp_solve_cplex(si.lpp);
4179 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4181 assert(lpp_is_sol_valid(si.lpp)
4182 && "solution of ILP must be valid");
4184 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));
4186 if(dump_flags & DUMP_SOLUTION) {
4190 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4191 if ((f = fopen(buf, "wt")) != NULL) {
4193 for (i = 0; i < si.lpp->var_next; ++i) {
4194 lpp_name_t *name = si.lpp->vars[i];
4195 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4201 writeback_results(&si);
4205 kill_all_unused_values_in_schedule(&si);
4207 #if defined(KEEPALIVE_SPILLS) || defined(KEEPALIVE_RELOADS)
4208 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
4211 // move reloads upwards
4212 be_liveness_recompute(si.lv);
4213 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4214 move_reloads_upward(&si);
4216 #ifndef NO_MEMCOPIES
4217 verify_phiclasses(&si);
4218 assign_spillslots(&si);
4221 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4223 dump_pressure_graph(&si, dump_suffix2);
4225 be_analyze_regpressure(chordal_env, "-post");
4227 #ifdef VERIFY_DOMINANCE
4228 be_check_dominance(chordal_env->irg);
4231 free_dom(chordal_env->irg);
4232 del_set(si.interferences);
4233 del_pset(si.inverse_ops);
4234 del_pset(si.all_possible_remats);
4235 #ifdef WITH_MEMOPERANDS
4236 del_set(si.memoperands);
4238 del_pset(si.spills);
4240 obstack_free(&obst, NULL);
4241 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4244 #else /* WITH_ILP */
4247 only_that_you_can_compile_without_WITH_ILP_defined(void)
4251 #endif /* WITH_ILP */