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"
52 #include "bespillremat.h"
54 #include "bepressurestat.h"
56 #include "bechordal_t.h"
59 #include <libcore/lc_opts.h>
60 #include <libcore/lc_opts_enum.h>
61 #endif /* WITH_LIBCORE */
63 #define DUMP_PROBLEM 1
65 #define DUMP_SOLUTION 4
67 #define KEEPALIVE_REMATS 1
68 #define KEEPALIVE_SPILLS 2
69 #define KEEPALIVE_RELOADS 4
71 #define VERIFY_MEMINTERF 1
72 #define VERIFY_DOMINANCE 2
75 #define REMATS_BRIGGS 1
76 #define REMATS_NOINVERSE 2
79 static int opt_dump_flags = 0;
80 static int opt_log = 0;
81 static int opt_keep_alive = 0;
82 static int opt_goodwin = 1;
83 static int opt_memcopies = 1;
84 static int opt_memoperands = 1;
85 static int opt_verify = VERIFY_MEMINTERF;
86 static int opt_remats = REMATS_ALL;
87 static int opt_repair_schedule = 0;
88 static int opt_no_enlarge_liveness = 0;
89 static int opt_remat_while_live = 1;
90 static int opt_timeout = 300;
91 static double opt_cost_reload = 8.0;
92 static double opt_cost_memoperand = 7.0;
93 static double opt_cost_spill = 50.0;
94 static double opt_cost_remat = 1.0;
98 static const lc_opt_enum_mask_items_t dump_items[] = {
99 { "problem", DUMP_PROBLEM },
101 { "solution", DUMP_SOLUTION },
105 static lc_opt_enum_mask_var_t dump_var = {
106 &opt_dump_flags, dump_items
109 static const lc_opt_enum_mask_items_t keepalive_items[] = {
110 { "remats", KEEPALIVE_REMATS },
111 { "spills", KEEPALIVE_SPILLS },
112 { "reloads", KEEPALIVE_RELOADS },
116 static lc_opt_enum_mask_var_t keep_alive_var = {
117 &opt_keep_alive, keepalive_items
120 static const lc_opt_enum_mask_items_t remats_items[] = {
121 { "none", REMATS_NONE },
122 { "briggs", REMATS_BRIGGS },
123 { "noinverse", REMATS_NOINVERSE },
124 { "all", REMATS_ALL },
128 static lc_opt_enum_mask_var_t remats_var = {
129 &opt_remats, remats_items
132 static const lc_opt_table_entry_t options[] = {
133 LC_OPT_ENT_ENUM_MASK("keepalive", "keep alive remats, spills or reloads", &keep_alive_var),
135 LC_OPT_ENT_BOOL ("goodwin", "activate goodwin reduction", &opt_goodwin),
136 LC_OPT_ENT_BOOL ("memcopies", "activate memcopy handling", &opt_memcopies),
137 LC_OPT_ENT_BOOL ("memoperands", "activate memoperands", &opt_memoperands),
138 LC_OPT_ENT_ENUM_INT ("remats", "type of remats to insert (none, briggs, noinverse or all)",&remats_var),
139 LC_OPT_ENT_BOOL ("repair_schedule", "repair the schedule by rematting once used nodes",&opt_repair_schedule),
140 LC_OPT_ENT_BOOL ("no_enlage_liveness", "do not enlarge liveness of operands of remats",&opt_no_enlarge_liveness),
141 LC_OPT_ENT_BOOL ("remat_while_live", "remat only values that can be used by real ops", &opt_remat_while_live),
143 LC_OPT_ENT_ENUM_MASK("dump", "dump problem, mps or solution", &dump_var),
144 LC_OPT_ENT_BOOL ("log", "activate the lpp log", &opt_log),
145 LC_OPT_ENT_INT ("timeout", "ILP solver timeout", &opt_timeout),
147 LC_OPT_ENT_DBL ("cost_reload", "cost of a reload", &opt_cost_reload),
148 LC_OPT_ENT_DBL ("cost_memoperand", "cost of a memory operand", &opt_cost_memoperand),
149 LC_OPT_ENT_DBL ("cost_spill", "cost of a spill instruction", &opt_cost_spill),
150 LC_OPT_ENT_DBL ("cost_remat", "cost of a rematerialization", &opt_cost_remat),
154 void be_spill_remat_register_options(lc_opt_entry_t *grp)
156 lc_opt_entry_t *my_grp = lc_opt_get_grp(grp, "remat");
157 lc_opt_add_table(my_grp, options);
162 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
163 //#define SCHEDULE_PHIM /* insert phim nodes into schedule */
166 //#define SOLVE_LOCAL
167 #define LPP_SERVER "i44pc52"
168 #define LPP_SOLVER "cplex"
171 #define MAX_PATHS INT_MAX
174 typedef struct _spill_ilp_t {
175 const arch_register_class_t *cls;
177 const be_chordal_env_t *chordal_env;
180 struct obstack *obst;
182 pset *all_possible_remats;
185 set *values; /**< for collecting all definitions of values before running ssa-construction */
190 DEBUG_ONLY(firm_dbg_module_t * dbg);
193 typedef int ilp_var_t;
194 typedef int ilp_cst_t;
196 typedef struct _spill_bb_t {
201 typedef struct _remat_t {
202 const ir_node *op; /**< for copy_irn */
203 const ir_node *value; /**< the value which is being recomputed by this remat */
204 const ir_node *proj; /**< not NULL if the above op produces a tuple */
205 int cost; /**< cost of this remat */
206 int inverse; /**< nonzero if this is an inverse remat */
210 * Data to be attached to each IR node. For remats this contains the ilp_var
211 * for this remat and for normal ops this contains the ilp_vars for
212 * reloading each operand
214 typedef struct _op_t {
219 const remat_t *remat; /** the remat this op belongs to */
220 int pre; /** 1, if this is a pressure-increasing remat */
224 ir_node *op; /** the operation this live range belongs to */
233 typedef struct _defs_t {
234 const ir_node *value;
235 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
236 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
239 typedef struct _remat_info_t {
240 const ir_node *irn; /**< the irn to which these remats belong */
241 pset *remats; /**< possible remats for this value */
242 pset *remats_by_operand; /**< remats with this value as operand */
245 typedef struct _keyval_t {
250 typedef struct _spill_t {
259 typedef struct _memoperand_t {
260 ir_node *irn; /**< the irn */
261 unsigned int pos; /**< the position of the argument */
262 ilp_var_t ilp; /**< the ilp var for this memory operand */
266 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
268 return chordal_has_class(si->chordal_env, irn);
273 cmp_remat(const void *a, const void *b)
275 const keyval_t *p = a;
276 const keyval_t *q = b;
277 const remat_t *r = p->val;
278 const remat_t *s = q->val;
282 return !(r == s || r->op == s->op);
286 cmp_remat(const void *a, const void *b)
288 const remat_t *r = a;
289 const remat_t *s = a;
291 return !(r == s || r->op == s->op);
295 cmp_spill(const void *a, const void *b, size_t size)
297 const spill_t *p = a;
298 const spill_t *q = b;
300 // return !(p->irn == q->irn && p->bb == q->bb);
301 return !(p->irn == q->irn);
305 cmp_memoperands(const void *a, const void *b, size_t size)
307 const memoperand_t *p = a;
308 const memoperand_t *q = b;
310 return !(p->irn == q->irn && p->pos == q->pos);
314 set_find_keyval(set * set, const void * key)
319 return set_find(set, &query, sizeof(query), HASH_PTR(key));
323 set_insert_keyval(set * set, void * key, void * val)
329 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
333 set_find_def(set * set, const ir_node * value)
338 return set_find(set, &query, sizeof(query), HASH_PTR(value));
342 set_insert_def(set * set, const ir_node * value)
349 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
352 static memoperand_t *
353 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
360 return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
363 static memoperand_t *
364 set_find_memoperand(set * set, const ir_node * irn, unsigned int pos)
368 query.irn = (ir_node*)irn;
370 return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
375 set_find_spill(set * set, const ir_node * value)
379 query.irn = (ir_node*)value;
380 return set_find(set, &query, sizeof(query), HASH_PTR(value));
383 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
384 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
385 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
386 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
387 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
390 cmp_remat_info(const void *a, const void *b, size_t size)
392 const remat_info_t *p = a;
393 const remat_info_t *q = b;
395 return !(p->irn == q->irn);
399 cmp_defs(const void *a, const void *b, size_t size)
404 return !(p->value == q->value);
408 cmp_keyval(const void *a, const void *b, size_t size)
410 const keyval_t *p = a;
411 const keyval_t *q = b;
413 return !(p->key == q->key);
417 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
420 #ifndef EXECFREQ_LOOPDEPH
421 return get_block_execfreq(si->chordal_env->exec_freq, get_block(irn)) + FUDGE;
424 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
426 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
431 get_cost(const spill_ilp_t * si, const ir_node * irn)
433 if(be_is_Spill(irn)) {
434 return opt_cost_spill;
435 } else if(be_is_Reload(irn)){
436 return opt_cost_reload;
438 return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
443 * Checks, whether node and its operands have suitable reg classes
446 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
449 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
450 int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
454 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
457 for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
458 ir_node *op = get_irn_n(irn, n);
459 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
462 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
469 * Try to create a remat from @p op with destination value @p dest_value
471 static INLINE remat_t *
472 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
474 remat_t *remat = NULL;
476 // if(!mode_is_datab(get_irn_mode(dest_value)))
479 if(dest_value == op) {
480 const ir_node *proj = NULL;
482 if(is_Proj(dest_value)) {
483 op = get_Proj_pred(op);
487 if(!is_rematerializable(si, op))
490 remat = obstack_alloc(si->obst, sizeof(*remat));
492 remat->cost = get_cost(si, op);
493 remat->value = dest_value;
497 arch_inverse_t inverse;
500 /* get the index of the operand we want to retrieve by the inverse op */
501 for (n = get_irn_arity(op)-1; n>=0; --n) {
502 ir_node *arg = get_irn_n(op, n);
504 if(arg == dest_value) break;
508 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", n, op));
510 /* else ask the backend to give an inverse op */
511 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
514 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
516 assert(inverse.n > 0 && "inverse op should have at least one node");
518 for(i=inverse.n-1; i>=0; --i) {
519 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
523 remat = obstack_alloc(si->obst, sizeof(*remat));
524 remat->op = inverse.nodes[0];
525 remat->cost = inverse.costs;
526 remat->value = dest_value;
527 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
530 assert(is_Proj(remat->proj));
532 assert(0 && "I can not handle remats with more than 2 nodes");
539 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
541 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
549 add_remat(const spill_ilp_t * si, const remat_t * remat)
551 remat_info_t *remat_info,
556 assert(remat->value);
558 query.irn = remat->value;
560 query.remats_by_operand = NULL;
561 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
563 if(remat_info->remats == NULL) {
564 remat_info->remats = new_pset(cmp_remat, 4096);
566 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
568 /* insert the remat into the remats_be_operand set of each argument of the remat op */
569 for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
570 ir_node *arg = get_irn_n(remat->op, n);
574 query.remats_by_operand = NULL;
575 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
577 if(remat_info->remats_by_operand == NULL) {
578 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
580 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
585 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
587 const ir_edge_t *edge = get_irn_out_edge_first(irn);
591 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
594 edge = get_irn_out_edge_next(irn, edge);
601 get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
607 irn = get_Proj_pred(irn);
609 for(n=get_irn_arity(irn)-1; n>=0; --n) {
610 const ir_node *arg = get_irn_n(irn, n);
612 if(has_reg_class(si, arg)) ++ret;
619 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
624 if( has_reg_class(si, op)
625 && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1)
626 && (opt_remats != REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0)
628 remat = get_remat_from_op(si, op, op);
630 add_remat(si, remat);
634 if(opt_remats == REMATS_ALL) {
635 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
637 for (n = get_irn_arity(op)-1; n>=0; --n) {
638 ir_node *arg = get_irn_n(op, n);
640 if(has_reg_class(si, arg)) {
641 /* try to get an inverse remat */
642 remat = get_remat_from_op(si, arg, op);
644 add_remat(si, remat);
652 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
655 ir_node *def_block = get_nodes_block(val);
661 /* if pos is at end of a basic block */
663 ret = (pos == def_block || block_dominates(def_block, pos));
664 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
668 /* else if this is a normal operation */
669 block = get_nodes_block(pos);
670 if(block == def_block) {
671 if(!sched_is_scheduled(val)) return 1;
673 ret = sched_comes_after(val, pos);
674 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
678 ret = block_dominates(def_block, block);
679 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
683 static INLINE ir_node *
684 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
686 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
690 * Returns first non-Phi node of block @p bb
692 static INLINE ir_node *
693 sched_block_first_nonphi(const ir_node * bb)
695 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
699 sched_skip_proj_predicator(const ir_node * irn, void * data)
701 return (is_Proj(irn));
704 static INLINE ir_node *
705 sched_next_nonproj(const ir_node * irn, int forward)
707 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
711 * Returns next operation node (non-Proj) after @p irn
712 * or the basic block of this node
714 static INLINE ir_node *
715 sched_next_op(const ir_node * irn)
717 ir_node *next = sched_next(irn);
722 return sched_next_nonproj(next, 1);
726 * Returns previous operation node (non-Proj) before @p irn
727 * or the basic block of this node
729 static INLINE ir_node *
730 sched_prev_op(const ir_node * irn)
732 ir_node *prev = sched_prev(irn);
737 return sched_next_nonproj(prev, 0);
741 sched_put_after(ir_node * insert, ir_node * irn)
743 if(is_Block(insert)) {
744 insert = sched_block_first_nonphi(insert);
746 insert = sched_next_op(insert);
748 sched_add_before(insert, irn);
752 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
754 if(is_Block(insert)) {
755 insert = sched_block_last_noncf(si, insert);
757 insert = sched_next_nonproj(insert, 0);
758 insert = sched_prev(insert);
760 sched_add_after(insert, irn);
764 * Tells you whether a @p remat can be placed before the irn @p pos
767 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
769 const ir_node *op = remat->op;
775 prev = sched_block_last_noncf(si, pos);
776 prev = sched_next_nonproj(prev, 0);
778 prev = sched_prev_op(pos);
780 /* do not remat if the rematted value is defined immediately before this op */
781 if(prev == remat->op) {
786 /* this should be just fine, the following OP will be using this value, right? */
788 /* only remat AFTER the real definition of a value (?) */
789 if(!value_is_defined_before(si, pos, remat->value)) {
790 // ir_fprintf(stderr, "error(not defined)");
795 for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
796 const ir_node *arg = get_irn_n(op, n);
798 if(opt_no_enlarge_liveness) {
799 if(has_reg_class(si, arg) && live) {
800 res &= pset_find_ptr((pset*)live, arg)?1:0;
802 res &= value_is_defined_before(si, pos, arg);
805 res &= value_is_defined_before(si, pos, arg);
813 * Tells you whether a @p remat can be placed after the irn @p pos
816 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
819 pos = sched_block_first_nonphi(pos);
821 pos = sched_next_op(pos);
824 /* only remat AFTER the real definition of a value (?) */
825 if(!value_is_defined_before(si, pos, remat->value)) {
829 return can_remat_before(si, remat, pos, live);
833 * Collect potetially rematerializable OPs
836 walker_remat_collector(ir_node * irn, void * data)
838 spill_ilp_t *si = data;
840 if(!is_Block(irn) && !is_Phi(irn)) {
841 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
842 get_remats_from_op(si, irn);
847 * Inserts a copy of @p irn before @p pos
850 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
855 bb = is_Block(pos)?pos:get_nodes_block(pos);
856 copy = exact_copy(irn);
858 _set_phi_class(copy, NULL);
859 set_nodes_block(copy, bb);
860 sched_put_before(si, pos, copy);
866 * Inserts a copy of @p irn after @p pos
869 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
874 bb = is_Block(pos)?pos:get_nodes_block(pos);
875 copy = exact_copy(irn);
877 _set_phi_class(copy, NULL);
878 set_nodes_block(copy, bb);
879 sched_put_after(pos, copy);
885 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
889 if(can_remat_after(si, remat, pos, live)) {
894 DBG((si->dbg, LEVEL_3, "\t >inserting remat2 %+F\n", remat->op));
896 copy = insert_copy_after(si, remat->op, pos);
898 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
899 op = obstack_alloc(si->obst, sizeof(*op));
901 op->attr.remat.remat = remat;
902 op->attr.remat.pre = 0;
903 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
905 set_irn_link(copy, op);
906 pset_insert_ptr(si->all_possible_remats, copy);
908 proj_copy = insert_copy_after(si, remat->proj, copy);
909 set_irn_n(proj_copy, 0, copy);
910 set_irn_link(proj_copy, op);
911 pset_insert_ptr(si->all_possible_remats, proj_copy);
923 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
927 if(can_remat_before(si, remat, pos, live)) {
932 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
934 copy = insert_copy_before(si, remat->op, pos);
936 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
937 op = obstack_alloc(si->obst, sizeof(*op));
939 op->attr.remat.remat = remat;
940 op->attr.remat.pre = 1;
941 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
943 set_irn_link(copy, op);
944 pset_insert_ptr(si->all_possible_remats, copy);
946 proj_copy = insert_copy_after(si, remat->proj, copy);
947 set_irn_n(proj_copy, 0, copy);
948 set_irn_link(proj_copy, op);
949 pset_insert_ptr(si->all_possible_remats, proj_copy);
961 get_block_n_succs(const ir_node *block) {
962 const ir_edge_t *edge;
964 assert(edges_activated(current_ir_graph));
966 edge = get_block_succ_first(block);
970 edge = get_block_succ_next(block, edge);
975 is_start_block(const ir_node * bb)
977 return get_irg_start_block(get_irn_irg(bb)) == bb;
981 is_merge_edge(const ir_node * bb)
983 if(is_start_block(bb))
987 return get_block_n_succs(bb) == 1;
993 is_diverge_edge(const ir_node * bb)
995 if(is_start_block(bb))
999 return get_Block_n_cfgpreds(bb) == 1;
1005 walker_regclass_copy_insertor(ir_node * irn, void * data)
1007 spill_ilp_t *si = data;
1009 if(is_Phi(irn) && has_reg_class(si, irn)) {
1012 for(n=get_irn_arity(irn)-1; n>=0; --n) {
1013 ir_node *phi_arg = get_irn_n(irn, n);
1014 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
1016 if(!has_reg_class(si, phi_arg)) {
1017 ir_node *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
1018 ir_node *pos = sched_block_last_noncf(si, bb);
1019 op_t *op = obstack_alloc(si->obst, sizeof(*op));
1021 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
1022 sched_add_after(pos, copy);
1023 set_irn_n(irn, n, copy);
1026 op->attr.live_range.args.reloads = NULL;
1027 op->attr.live_range.ilp = ILP_UNDEF;
1028 set_irn_link(copy, op);
1035 * Insert (so far unused) remats into the irg to
1036 * recompute the potential liveness of all values
1039 walker_remat_insertor(ir_node * bb, void * data)
1041 spill_ilp_t *si = data;
1046 /* skip start block, no remats to do there */
1047 if(is_start_block(bb)) return;
1049 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1051 live = pset_new_ptr_default();
1052 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1053 ir_node *value = be_lv_get_irn(si->lv, bb, i);
1055 /* add remats at end of block */
1056 if (has_reg_class(si, value)) {
1057 pset_insert_ptr(live, value);
1061 irn = sched_last(bb);
1062 while(!sched_is_end(irn)) {
1068 next = sched_prev(irn);
1070 if(is_Phi(irn) || is_Proj(irn)) {
1075 args = pset_new_ptr_default();
1077 /* collect arguments of op */
1078 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1079 ir_node *arg = get_irn_n(irn, n);
1081 pset_insert_ptr(args, arg);
1084 /* set args of op already live in epilog */
1085 pset_foreach(args, arg) {
1086 if(has_reg_class(si, arg)) {
1087 pset_insert_ptr(live, arg);
1090 /* delete defined value from live set */
1091 if(has_reg_class(si, irn)) {
1092 pset_remove_ptr(live, irn);
1095 remat_args = pset_new_ptr_default();
1097 /* insert all possible remats before irn */
1098 pset_foreach(args, arg) {
1099 remat_info_t *remat_info,
1103 /* continue if the operand has the wrong reg class
1105 if(!has_reg_class(si, arg))
1109 query.remats = NULL;
1110 query.remats_by_operand = NULL;
1111 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1117 if(remat_info->remats) {
1118 pset_foreach(remat_info->remats, remat) {
1119 ir_node *remat_irn = NULL;
1121 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
1122 if(opt_remat_while_live) {
1123 if(pset_find_ptr(live, remat->value)) {
1124 remat_irn = insert_remat_before(si, remat, irn, live);
1127 remat_irn = insert_remat_before(si, remat, irn, live);
1130 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1131 ir_node *remat_arg = get_irn_n(remat_irn, n);
1133 if(!has_reg_class(si, remat_arg)) continue;
1135 pset_insert_ptr(remat_args, remat_arg);
1142 /* now we add remat args to op's args because they could also die at this op */
1143 pset_foreach(args,arg) {
1144 if(pset_find_ptr(remat_args, arg)) {
1145 pset_remove_ptr(remat_args, arg);
1148 pset_foreach(remat_args,arg) {
1149 pset_insert_ptr(args, arg);
1152 /* insert all possible remats after irn */
1153 pset_foreach(args, arg) {
1154 remat_info_t *remat_info,
1158 /* continue if the operand has the wrong reg class */
1159 if(!has_reg_class(si, arg))
1163 query.remats = NULL;
1164 query.remats_by_operand = NULL;
1165 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1171 /* do not place post remats after jumps */
1172 if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
1174 if(remat_info->remats_by_operand) {
1175 pset_foreach(remat_info->remats_by_operand, remat) {
1176 /* do not insert remats producing the same value as one of the operands */
1177 if(!pset_find_ptr(args, remat->value)) {
1178 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
1179 if(opt_remat_while_live) {
1180 if(pset_find_ptr(live, remat->value)) {
1181 insert_remat_after(si, remat, irn, live);
1184 insert_remat_after(si, remat, irn, live);
1191 del_pset(remat_args);
1196 /* add remats at end if successor has multiple predecessors */
1197 if(is_merge_edge(bb)) {
1198 pset *live_out = pset_new_ptr_default();
1201 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1202 value = be_lv_get_irn(si->lv, bb, i);
1204 if (has_reg_class(si, value)) {
1205 pset_insert_ptr(live_out, value);
1209 /* add remats at end of block */
1210 pset_foreach(live_out, value) {
1211 remat_info_t *remat_info,
1216 query.remats = NULL;
1217 query.remats_by_operand = NULL;
1218 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1220 if(remat_info && remat_info->remats) {
1221 pset_foreach(remat_info->remats, remat) {
1222 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1224 insert_remat_before(si, remat, bb, live_out);
1231 if(is_diverge_edge(bb)) {
1232 pset *live_in = pset_new_ptr_default();
1235 be_lv_foreach(si->lv, bb, be_lv_state_in, i) {
1236 value = be_lv_get_irn(si->lv, bb, i);
1238 if(has_reg_class(si, value)) {
1239 pset_insert_ptr(live_in, value);
1242 sched_foreach(bb, value) {
1243 if(!is_Phi(value)) break;
1245 if(has_reg_class(si, value)) {
1246 pset_insert_ptr(live_in, value);
1250 /* add remat2s at beginning of block */
1251 pset_foreach(live_in, value) {
1252 remat_info_t *remat_info,
1257 query.remats = NULL;
1258 query.remats_by_operand = NULL;
1259 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1261 if(remat_info && remat_info->remats_by_operand) {
1262 pset_foreach(remat_info->remats_by_operand, remat) {
1263 DBG((si->dbg, LEVEL_4, "\t considering remat2 %+F at beginning of block %+F\n", remat->op, bb));
1265 /* put the remat here if all its args are available */
1266 insert_remat_after(si, remat, bb, live_in);
1276 can_be_copied(const ir_node * bb, const ir_node * irn)
1278 assert(is_merge_edge(bb));
1280 const ir_edge_t *edge = get_block_succ_first(bb);
1281 const ir_node *next_bb = edge->src;
1282 int pos = edge->pos;
1285 sched_foreach(next_bb, phi) {
1286 const ir_node *phi_arg;
1288 if(!is_Phi(phi)) break;
1290 phi_arg = get_irn_n(phi, pos);
1292 if(phi_arg == irn) {
1300 * Initialize additional node info
1303 luke_initializer(ir_node * bb, void * data)
1305 spill_ilp_t *si = (spill_ilp_t*)data;
1306 spill_bb_t *spill_bb;
1309 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1310 set_irn_link(bb, spill_bb);
1312 sched_foreach(bb, irn) {
1315 op = obstack_alloc(si->obst, sizeof(*op));
1317 op->attr.live_range.ilp = ILP_UNDEF;
1320 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1321 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1323 } else if(!is_Proj(irn)) {
1324 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1325 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1327 op->attr.live_range.args.reloads = NULL;
1329 set_irn_link(irn, op);
1335 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1338 luke_endwalker(ir_node * bb, void * data)
1340 spill_ilp_t *si = (spill_ilp_t*)data;
1346 spill_bb_t *spill_bb = get_irn_link(bb);
1349 live = pset_new_ptr_default();
1350 use_end = pset_new_ptr_default();
1352 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1353 irn = be_lv_get_irn(si->lv, bb, i);
1354 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1355 pset_insert_ptr(live, irn);
1359 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1360 /* their reg_out must always be set */
1361 sched_foreach_reverse(bb, irn) {
1364 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1366 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1367 ir_node *irn_arg = get_irn_n(irn, n);
1369 if(has_reg_class(si, irn_arg)) {
1370 pset_insert_ptr(use_end, irn_arg);
1375 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1376 //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1377 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1379 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1381 /* if this is a merge edge we can reload at the end of this block */
1382 if(is_merge_edge(bb)) {
1383 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1384 } else if(pset_count(use_end)){
1385 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1387 spill_bb->reloads = NULL;
1390 pset_foreach(live,irn) {
1394 int default_spilled;
1397 /* handle values used by control flow nodes later separately */
1398 if(pset_find_ptr(use_end, irn)) continue;
1401 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1403 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1405 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1406 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1407 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1409 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1410 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1412 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1413 /* by default spill value right after definition */
1414 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1415 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1417 if(is_merge_edge(bb)) {
1421 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1422 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), can_be_copied(bb, irn));
1423 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1425 /* reload <= mem_out */
1426 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1427 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1428 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1431 spill->reg_in = ILP_UNDEF;
1432 spill->mem_in = ILP_UNDEF;
1435 pset_foreach(use_end,irn) {
1439 ilp_cst_t end_use_req,
1442 int default_spilled;
1445 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1447 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1449 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1450 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1451 /* if irn is used at the end of the block, then it is live anyway */
1452 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1454 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1455 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1457 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1458 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1459 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1461 /* reload for use be control flow op */
1462 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1463 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1464 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1466 /* reload <= mem_out */
1467 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1468 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1469 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1471 spill->reg_in = ILP_UNDEF;
1472 spill->mem_in = ILP_UNDEF;
1474 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1475 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1476 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1484 next_post_remat(const ir_node * irn)
1490 next = sched_block_first_nonphi(irn);
1492 next = sched_next_op(irn);
1495 if(sched_is_end(next))
1498 op = get_irn_link(next);
1499 if(op->is_remat && !op->attr.remat.pre) {
1508 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1514 ret = sched_block_last_noncf(si, irn);
1515 ret = sched_next(ret);
1516 ret = sched_prev_op(ret);
1518 ret = sched_prev_op(irn);
1521 if(sched_is_end(ret) || is_Phi(ret))
1524 op = (op_t*)get_irn_link(ret);
1525 if(op->is_remat && op->attr.remat.pre) {
1533 * Find a remat of value @p value in the epilog of @p pos
1536 find_post_remat(const ir_node * value, const ir_node * pos)
1538 while((pos = next_post_remat(pos)) != NULL) {
1541 op = get_irn_link(pos);
1542 assert(op->is_remat && !op->attr.remat.pre);
1544 if(op->attr.remat.remat->value == value)
1545 return (ir_node*)pos;
1548 const ir_edge_t *edge;
1549 foreach_out_edge(pos, edge) {
1550 ir_node *proj = get_edge_src_irn(edge);
1551 assert(is_Proj(proj));
1561 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1563 spill_bb_t *spill_bb = get_irn_link(bb);
1567 int default_spilled;
1570 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1572 double spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1574 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1576 spill->reg_out = ILP_UNDEF;
1577 spill->reg_in = ILP_UNDEF;
1578 spill->mem_in = ILP_UNDEF;
1580 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1581 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1583 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1584 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1585 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1592 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1597 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1598 irn = be_lv_get_irn(si->lv, bb, i);
1600 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1601 pset_insert_ptr(live, irn);
1605 irn = sched_last(bb);
1607 /* all values eaten by control flow operations are also live until the end of the block */
1608 sched_foreach_reverse(bb, irn) {
1611 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1613 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1614 ir_node *arg = get_irn_n(irn,i);
1616 if(has_reg_class(si, arg)) {
1617 pset_insert_ptr(live, arg);
1624 * Inserts ILP-constraints and variables for memory copying before the given position
1627 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1629 const ir_node *succ;
1630 const ir_edge_t *edge;
1631 spill_bb_t *spill_bb = get_irn_link(block);
1640 assert(edges_activated(current_ir_graph));
1642 edge = get_block_succ_first(block);
1648 edge = get_block_succ_next(block, edge);
1649 /* next block can only contain phis, if this is a merge edge */
1652 ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1653 copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1655 ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1656 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1658 pset_foreach(live, tmp) {
1661 op_t *op = get_irn_link(irn);
1662 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1664 spill = set_find_spill(spill_bb->ilp, tmp);
1667 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1669 lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1671 sched_foreach(succ, phi) {
1672 const ir_node *to_copy;
1674 spill_t *to_copy_spill;
1675 op_t *phi_op = get_irn_link(phi);
1676 ilp_var_t reload = ILP_UNDEF;
1679 if(!is_Phi(phi)) break;
1680 if(!has_reg_class(si, phi)) continue;
1682 to_copy = get_irn_n(phi, pos);
1683 to_copy_op = get_irn_link(to_copy);
1685 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1686 assert(to_copy_spill);
1688 if(spill_bb->reloads) {
1689 keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1692 reload = PTR_TO_INT(keyval->val);
1696 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1697 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1699 /* copy - reg_out - reload - remat - live_range <= 0 */
1700 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1701 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1702 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1703 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1704 foreach_pre_remat(si, block, tmp) {
1705 op_t *remat_op = get_irn_link(tmp);
1706 if(remat_op->attr.remat.remat->value == to_copy) {
1707 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1711 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1712 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1714 /* copy - reg_out - copyreg <= 0 */
1715 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1716 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1717 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1723 * Walk all irg blocks and emit this ILP
1726 luke_blockwalker(ir_node * bb, void * data)
1728 spill_ilp_t *si = (spill_ilp_t*)data;
1733 spill_bb_t *spill_bb = get_irn_link(bb);
1736 pset *defs = pset_new_ptr_default();
1737 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
1739 live = pset_new_ptr_default();
1741 /****************************************
1742 * B A S I C B L O C K E N D
1743 ***************************************/
1746 /* init live values at end of block */
1747 get_live_end(si, bb, live);
1749 pset_foreach(live, irn) {
1751 ilp_var_t reload = ILP_UNDEF;
1753 spill = set_find_spill(spill_bb->ilp, irn);
1756 if(spill_bb->reloads) {
1757 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1760 reload = PTR_TO_INT(keyval->val);
1764 op = get_irn_link(irn);
1765 assert(!op->is_remat);
1767 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1768 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1769 op->attr.live_range.op = bb;
1771 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1772 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1774 /* reg_out - reload - remat - live_range <= 0 */
1775 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1776 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1777 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1778 foreach_pre_remat(si, bb, tmp) {
1779 op_t *remat_op = get_irn_link(tmp);
1780 if(remat_op->attr.remat.remat->value == irn) {
1781 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1784 ir_snprintf(buf, sizeof(buf), "reg_out2_%N_%N", bb, irn);
1785 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
1787 /* value may only die at bb end if it is used for a mem copy */
1788 /* reg_out + \sum copy - reload - remat - live_range >= 0 */
1789 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1790 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1791 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1792 foreach_pre_remat(si, bb, tmp) {
1793 op_t *remat_op = get_irn_link(tmp);
1794 if(remat_op->attr.remat.remat->value == irn) {
1795 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1798 if(is_merge_edge(bb)) {
1799 const ir_edge_t *edge = get_block_succ_first(bb);
1800 const ir_node *next_bb = edge->src;
1801 int pos = edge->pos;
1804 sched_foreach(next_bb, phi) {
1805 const ir_node *phi_arg;
1807 if(!is_Phi(phi)) break;
1809 phi_arg = get_irn_n(phi, pos);
1811 if(phi_arg == irn) {
1812 op_t *phi_op = get_irn_link(phi);
1813 ilp_var_t copy = phi_op->attr.live_range.args.copies[pos];
1815 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
1822 insert_mem_copy_position(si, live, bb);
1825 * start new live ranges for values used by remats at end of block
1826 * and assure the remat args are available
1828 foreach_pre_remat(si, bb, tmp) {
1829 op_t *remat_op = get_irn_link(tmp);
1832 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1833 ir_node *remat_arg = get_irn_n(tmp, n);
1834 op_t *arg_op = get_irn_link(remat_arg);
1837 if(!has_reg_class(si, remat_arg)) continue;
1839 /* if value is becoming live through use by remat */
1840 if(!pset_find_ptr(live, remat_arg)) {
1841 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1842 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1844 arg_op->attr.live_range.ilp = prev_lr;
1845 arg_op->attr.live_range.op = bb;
1847 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1849 pset_insert_ptr(live, remat_arg);
1850 add_to_spill_bb(si, bb, remat_arg);
1853 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1854 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1855 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1857 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1858 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1860 /* use reload placed for this argument */
1861 if(spill_bb->reloads) {
1862 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1865 ilp_var_t reload = PTR_TO_INT(keyval->val);
1867 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1872 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1877 /**************************************
1878 * B A S I C B L O C K B O D Y
1879 **************************************/
1881 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1887 ilp_cst_t check_pre,
1893 ilp_cst_t one_memoperand;
1895 /* iterate only until first phi */
1899 op = get_irn_link(irn);
1901 if(op->is_remat) continue;
1903 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1905 /* collect defined values */
1906 if(has_reg_class(si, irn)) {
1907 pset_insert_ptr(defs, irn);
1911 if(is_Proj(irn)) continue;
1914 * init set of irn's arguments
1915 * and all possibly used values around this op
1916 * and values defined by post remats
1918 args = new_set(cmp_keyval, get_irn_arity(irn));
1919 used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1920 remat_defs = pset_new_ptr(pset_count(live));
1922 if(!is_start_block(bb) || !be_is_Barrier(irn)) {
1923 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1924 ir_node *irn_arg = get_irn_n(irn, n);
1925 if(has_reg_class(si, irn_arg)) {
1926 set_insert_keyval(args, irn_arg, (void*)n);
1927 pset_insert_ptr(used, irn_arg);
1930 foreach_post_remat(irn, tmp) {
1931 op_t *remat_op = get_irn_link(tmp);
1933 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1935 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1936 ir_node *remat_arg = get_irn_n(tmp, n);
1937 if(has_reg_class(si, remat_arg)) {
1938 pset_insert_ptr(used, remat_arg);
1942 foreach_pre_remat(si, irn, tmp) {
1943 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1944 ir_node *remat_arg = get_irn_n(tmp, n);
1945 if(has_reg_class(si, remat_arg)) {
1946 pset_insert_ptr(used, remat_arg);
1952 /**********************************
1953 * I N E P I L O G O F irn
1954 **********************************/
1956 /* ensure each dying value is used by only one post remat */
1957 pset_foreach(used, tmp) {
1958 ir_node *value = tmp;
1959 op_t *value_op = get_irn_link(value);
1964 foreach_post_remat(irn, remat) {
1965 op_t *remat_op = get_irn_link(remat);
1967 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1968 ir_node *remat_arg = get_irn_n(remat, n);
1970 /* if value is used by this remat add it to constraint */
1971 if(remat_arg == value) {
1973 /* sum remat2s <= 1 + n_remats*live_range */
1974 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1975 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
1979 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1985 if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
1986 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1990 /* ensure at least one value dies at post remat */
1991 foreach_post_remat(irn, tmp) {
1992 op_t *remat_op = get_irn_link(tmp);
1993 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
1996 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
1997 remat_arg = get_irn_n(tmp, n);
1999 if(has_reg_class(si, remat_arg)) {
2001 /* does arg always die at this op? */
2002 if(!pset_find_ptr(live, remat_arg))
2003 goto skip_one_must_die;
2005 pset_insert_ptr(remat_args, remat_arg);
2009 /* remat + \sum live_range(remat_arg) <= |args| */
2010 ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
2011 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2012 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2014 pset_foreach(remat_args, remat_arg) {
2015 op_t *arg_op = get_irn_link(remat_arg);
2017 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2021 del_pset(remat_args);
2024 /* new live ranges for values from L\U defined by post remats */
2025 pset_foreach(live, tmp) {
2026 ir_node *value = tmp;
2027 op_t *value_op = get_irn_link(value);
2029 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
2030 ilp_var_t prev_lr = ILP_UNDEF;
2033 if(pset_find_ptr(remat_defs, value)) {
2035 /* next_live_range <= prev_live_range + sum remat2s */
2036 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
2037 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2039 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
2040 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2042 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
2043 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2045 foreach_post_remat(irn, remat) {
2046 op_t *remat_op = get_irn_link(remat);
2048 /* if value is being rematerialized by this remat */
2049 if(value == remat_op->attr.remat.remat->value) {
2050 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2054 value_op->attr.live_range.ilp = prev_lr;
2055 value_op->attr.live_range.op = irn;
2060 /* requirements for post remats and start live ranges from L/U' for values dying here */
2061 foreach_post_remat(irn, tmp) {
2062 op_t *remat_op = get_irn_link(tmp);
2065 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2066 ir_node *remat_arg = get_irn_n(tmp, n);
2067 op_t *arg_op = get_irn_link(remat_arg);
2069 if(!has_reg_class(si, remat_arg)) continue;
2071 /* only for values in L\U (TODO and D?), the others are handled with post_use */
2072 if(!pset_find_ptr(used, remat_arg)) {
2073 /* remat <= live_range(remat_arg) */
2074 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
2075 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2077 /* if value is becoming live through use by remat2 */
2078 if(!pset_find_ptr(live, remat_arg)) {
2081 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
2082 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2084 arg_op->attr.live_range.ilp = lr;
2085 arg_op->attr.live_range.op = irn;
2087 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
2089 pset_insert_ptr(live, remat_arg);
2090 add_to_spill_bb(si, bb, remat_arg);
2093 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2094 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2099 d = pset_count(defs);
2100 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
2102 /* count how many regs irn needs for arguments */
2103 u = set_count(args);
2106 /* check the register pressure in the epilog */
2107 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
2108 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
2109 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
2111 /* add L\U' to check_post */
2112 pset_foreach(live, tmp) {
2113 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
2114 /* if a live value is not used by irn */
2115 tmp_op = get_irn_link(tmp);
2116 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
2120 /***********************************************************
2121 * 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
2122 **********************************************************/
2125 pset_foreach(used, tmp) {
2131 op_t *arg_op = get_irn_link(arg);
2134 spill = add_to_spill_bb(si, bb, arg);
2136 /* new live range for each used value */
2137 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2138 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2140 /* the epilog stuff - including post_use, check_post, check_post_remat */
2141 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2142 post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2144 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2146 /* arg is live throughout epilog if the next live_range is in a register */
2147 if(pset_find_ptr(live, arg)) {
2148 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
2150 /* post_use >= next_lr + remat */
2151 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2152 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2153 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2154 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2157 /* forall post remat which use arg add a similar cst */
2158 foreach_post_remat(irn, remat) {
2161 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2162 ir_node *remat_arg = get_irn_n(remat, n);
2163 op_t *remat_op = get_irn_link(remat);
2165 if(remat_arg == arg) {
2166 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
2168 /* post_use >= remat */
2169 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2170 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2171 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2172 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2177 /* if value is not an arg of op and not possibly defined by post remat
2178 * then it may only die and not become live
2180 if(!set_find_keyval(args, arg)) {
2181 /* post_use <= prev_lr */
2182 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2183 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2184 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2185 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2187 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2188 /* next_lr <= prev_lr */
2189 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2190 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2191 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2192 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2196 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2197 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2198 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2199 ilp_var_t memoperand;
2201 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2202 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0);
2203 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2205 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2206 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2208 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2209 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2214 /* new live range begins for each used value */
2215 arg_op->attr.live_range.ilp = prev_lr;
2216 arg_op->attr.live_range.op = irn;
2218 pset_insert_ptr(live, arg);
2221 /* just to be sure */
2222 check_post = ILP_UNDEF;
2229 /* check the register pressure in the prolog */
2230 /* sum_{L\U} lr <= k - |U| */
2231 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2232 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2234 /* for the prolog remove defined values from the live set */
2235 pset_foreach(defs, tmp) {
2236 pset_remove_ptr(live, tmp);
2239 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2240 ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2241 one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2244 /***********************************************************
2245 * 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
2246 **********************************************************/
2249 set_foreach(args, keyval) {
2251 const ir_node *arg = keyval->key;
2252 int i = PTR_TO_INT(keyval->val);
2253 op_t *arg_op = get_irn_link(arg);
2254 ilp_cst_t requirements;
2257 spill = set_find_spill(spill_bb->ilp, arg);
2260 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2261 op->attr.live_range.args.reloads[i] = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
2263 /* reload <= mem_out */
2264 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2265 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2266 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2267 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2269 /* requirement: arg must be in register for use */
2270 /* reload + remat + live_range == 1 */
2271 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2272 requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2274 lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2275 lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2276 foreach_pre_remat(si, irn, tmp) {
2277 op_t *remat_op = get_irn_link(tmp);
2278 if(remat_op->attr.remat.remat->value == arg) {
2279 lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2283 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2285 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2286 if(get_irn_n(irn, n) == arg) {
2290 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2291 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2292 memoperand_t *memoperand;
2293 memoperand = set_find_memoperand(si->memoperands, irn, n);
2295 /* memoperand <= mem_out */
2296 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2297 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2298 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2299 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2301 /* the memoperand is only sufficient if it is used once by the op */
2302 if(n_memoperands == 1)
2303 lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2305 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2307 /* we have one more free register if we use a memory operand */
2308 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2314 /* iterate over L\U */
2315 pset_foreach(live, tmp) {
2316 if(!set_find_keyval(args, tmp)) {
2317 /* if a live value is not used by irn */
2318 tmp_op = get_irn_link(tmp);
2319 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2323 /* requirements for remats */
2324 foreach_pre_remat(si, irn, tmp) {
2325 op_t *remat_op = get_irn_link(tmp);
2328 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2329 ir_node *remat_arg = get_irn_n(tmp, n);
2330 op_t *arg_op = get_irn_link(remat_arg);
2332 if(!has_reg_class(si, remat_arg)) continue;
2334 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2335 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2336 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2338 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2339 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2341 /* if remat arg is also used by current op then we can use reload placed for this argument */
2342 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2343 int index = (int)keyval->val;
2345 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2353 /*************************
2354 * D O N E W I T H O P
2355 *************************/
2357 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2359 pset_foreach(live, tmp) {
2360 assert(has_reg_class(si, tmp));
2363 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2364 ir_node *arg = get_irn_n(irn, n);
2366 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2369 del_pset(remat_defs);
2373 defs = pset_new_ptr_default();
2375 /* skip everything above barrier in start block */
2376 if(is_start_block(bb) && be_is_Barrier(irn)) {
2377 assert(pset_count(live) == 0);
2386 /***************************************
2387 * B E G I N N I N G O F B L O C K
2388 ***************************************/
2391 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2392 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2394 pset_foreach(live, irn) {
2395 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2398 /* construct mem_outs for all values */
2399 set_foreach(spill_bb->ilp, spill) {
2400 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2401 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2403 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2404 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2406 if(pset_find_ptr(live, spill->irn)) {
2407 int default_spilled;
2408 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2410 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2411 default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2412 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2413 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2415 if(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2417 op_t *op = get_irn_link(spill->irn);
2419 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2420 const ir_node *arg = get_irn_n(spill->irn, n);
2426 /* argument already done? */
2427 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2429 /* get sum of execution frequencies of blocks with the same phi argument */
2430 for(m=n; m>=0; --m) {
2431 const ir_node *arg2 = get_irn_n(spill->irn, m);
2434 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2438 /* copies are not for free */
2439 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2440 var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
2442 for(m=n; m>=0; --m) {
2443 const ir_node *arg2 = get_irn_n(spill->irn, m);
2446 op->attr.live_range.args.copies[m] = var;
2451 /* copy <= mem_in */
2452 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2453 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2454 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2455 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2462 foreach_post_remat(bb, tmp) {
2464 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
2465 op_t *remat_op = get_irn_link(tmp);
2468 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2469 remat_arg = get_irn_n(tmp, n);
2470 if(has_reg_class(si, remat_arg)) {
2471 pset_insert_ptr(remat_args, remat_arg);
2474 assert(pset_count(remat_args) > 0 && "post remats should have at least one arg");
2476 /* remat + \sum live_range(remat_arg) <= |args| */
2477 ir_snprintf(buf, sizeof(buf), "one_must_die_%N", tmp);
2478 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2479 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2481 pset_foreach(remat_args, remat_arg) {
2482 /* if value is becoming live through use by remat2 */
2483 if(!pset_find_ptr(live, remat_arg)) {
2484 op_t *remat_arg_op = get_irn_link(remat_arg);
2487 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 at bb start %+F\n", remat_arg, tmp));
2489 pset_insert_ptr(live, remat_arg);
2490 spill = add_to_spill_bb(si, bb, remat_arg);
2491 remat_arg_op->attr.live_range.ilp = ILP_UNDEF;
2493 /* we need reg_in and mem_in for this value; they will be referenced later */
2494 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", remat_arg, bb);
2495 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2496 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", remat_arg, bb);
2497 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2500 /* optimization: all memory stuff should be 0, for we do not want to insert reloads for remats */
2501 ir_snprintf(buf, sizeof(buf), "nomem_%N_%N", remat_arg, bb);
2502 nomem = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0.0);
2504 lpp_set_factor_fast(si->lpp, nomem, spill->spill, 1.0);
2505 if(spill_bb->reloads) {
2506 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
2509 ilp_var_t reload = PTR_TO_INT(keyval->val);
2510 lpp_set_factor_fast(si->lpp, nomem, reload, 1.0);
2514 op_t *remat_arg_op = get_irn_link(remat_arg);
2515 lpp_set_factor_fast(si->lpp, cst, remat_arg_op->attr.live_range.ilp, 1.0);
2518 del_pset(remat_args);
2521 /* L\U is empty at bb start */
2522 /* arg is live throughout epilog if it is reg_in into this block */
2524 /* check the register pressure at the beginning of the block
2527 /* reg_in entspricht post_use */
2529 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2530 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2532 pset_foreach(live, irn) {
2535 spill = set_find_spill(spill_bb->ilp, irn);
2538 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2539 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2541 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2543 /* spill + mem_in <= 1 */
2544 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2545 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2547 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2548 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2550 } /* post_remats are NOT included in register pressure check because
2551 they do not increase regpressure */
2553 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2554 pset_foreach(live, irn) {
2558 spill = set_find_spill(spill_bb->ilp, irn);
2559 assert(spill && spill->irn == irn);
2561 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2562 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2565 ir_node *phi_arg = get_Phi_pred(irn, n);
2566 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2567 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2569 op_t *op = get_irn_link(irn);
2571 /* although the phi is in the right regclass one or more of
2572 * its arguments can be in a different one or at least to
2575 if(has_reg_class(si, phi_arg)) {
2576 /* mem_in < mem_out_arg + copy */
2577 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2578 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2580 /* reg_in < reg_out_arg */
2581 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2582 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2584 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2585 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2587 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2590 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2592 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2594 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2598 /* else assure the value arrives on all paths in the same resource */
2600 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2603 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2604 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2607 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2608 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2609 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2610 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2612 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2613 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2615 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2618 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2619 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2624 foreach_post_remat(bb, tmp) {
2627 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2628 ir_node *remat_arg = get_irn_n(tmp, n);
2629 op_t *remat_op = get_irn_link(tmp);
2631 if(!has_reg_class(si, remat_arg)) continue;
2633 spill = set_find_spill(spill_bb->ilp, remat_arg);
2636 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2637 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2638 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2639 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2643 pset_foreach(live, irn) {
2644 const op_t *op = get_irn_link(irn);
2645 const ir_node *remat;
2650 foreach_post_remat(bb, remat) {
2653 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2654 const ir_node *arg = get_irn_n(remat, n);
2657 const op_t *remat_op = get_irn_link(remat);
2659 if(cst == ILP_UNDEF) {
2660 /* sum remat2s <= 1 + n_remats*live_range */
2661 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", irn, bb);
2662 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2664 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2670 if(cst != ILP_UNDEF && op->attr.live_range.ilp != ILP_UNDEF) {
2671 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -n_remats);
2675 /* first live ranges from reg_ins */
2676 pset_foreach(live, irn) {
2677 op_t *op = get_irn_link(irn);
2679 if(op->attr.live_range.ilp != ILP_UNDEF) {
2681 spill = set_find_spill(spill_bb->ilp, irn);
2682 assert(spill && spill->irn == irn);
2684 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2685 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2686 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2687 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2689 foreach_post_remat(bb, tmp) {
2690 op_t *remat_op = get_irn_link(tmp);
2692 if(remat_op->attr.remat.remat->value == irn) {
2693 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2699 /* walk forward now and compute constraints for placing spills */
2700 /* this must only be done for values that are not defined in this block */
2701 pset_foreach(live, irn) {
2703 * if value is defined in this block we can anways place the spill directly after the def
2704 * -> no constraint necessary
2706 if(!is_Phi(irn) && get_nodes_block(irn) == bb) {
2711 spill = set_find_spill(spill_bb->ilp, irn);
2714 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2715 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2717 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2718 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2721 sched_foreach_op(bb, tmp) {
2722 op_t *op = get_irn_link(tmp);
2724 if(is_Phi(tmp)) continue;
2725 assert(!is_Proj(tmp));
2728 const ir_node *value = op->attr.remat.remat->value;
2731 /* only collect remats up to the first real use of a value */
2732 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2737 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2738 ir_node *arg = get_irn_n(tmp, n);
2741 /* if a value is used stop collecting remats */
2754 typedef struct _irnlist_t {
2755 struct list_head list;
2759 typedef struct _interference_t {
2760 struct list_head blocklist;
2766 cmp_interference(const void *a, const void *b, size_t size)
2768 const interference_t *p = a;
2769 const interference_t *q = b;
2771 return !(p->a == q->a && p->b == q->b);
2774 static interference_t *
2775 set_find_interference(set * set, ir_node * a, ir_node * b)
2777 interference_t query;
2779 query.a = (a>b)?a:b;
2780 query.b = (a>b)?b:a;
2782 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2785 static interference_t *
2786 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2788 interference_t query,
2790 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2794 result = set_find_interference(set, a, b);
2797 list_add(&list->list, &result->blocklist);
2801 query.a = (a>b)?a:b;
2802 query.b = (a>b)?b:a;
2804 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2806 INIT_LIST_HEAD(&result->blocklist);
2807 list_add(&list->list, &result->blocklist);
2813 values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
2815 const ir_edge_t *edge;
2817 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2818 /* both values are live in, so they interfere */
2822 /* ensure a dominates b */
2823 if(value_dominates(b,a)) {
2829 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2832 /* the following code is stolen from bera.c */
2833 if(be_is_live_end(si->lv, bb, a))
2836 foreach_out_edge(a, edge) {
2837 const ir_node *user = edge->src;
2838 if(get_nodes_block(user) == bb
2841 && !pset_find_ptr(si->inverse_ops, user)
2842 && value_dominates(b, user))
2850 * Walk all irg blocks and collect interfering values inside of phi classes
2853 luke_interferencewalker(ir_node * bb, void * data)
2855 spill_ilp_t *si = (spill_ilp_t*)data;
2858 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2859 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
2860 op_t *a_op = get_irn_link(a);
2863 /* a is only interesting if it is in my register class and if it is inside a phi class */
2864 if (has_reg_class(si, a) && get_phi_class(a)) {
2865 if(a_op->is_remat || pset_find_ptr(si->inverse_ops, a))
2868 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)) {
2869 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
2870 op_t *b_op = get_irn_link(b);
2873 /* a and b are only interesting if they are in the same phi class */
2874 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2875 if(b_op->is_remat || pset_find_ptr(si->inverse_ops, b))
2878 if(values_interfere_in_block(si, bb, a, b)) {
2879 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2880 set_insert_interference(si, si->interferences, a, b, bb);
2888 static unsigned int copy_path_id = 0;
2891 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2898 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2899 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2901 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2903 pset_foreach(copies, ptr) {
2904 copy = PTR_TO_INT(ptr);
2905 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2910 * @parameter copies contains a path of copies which lead us to irn
2911 * @parameter visited contains a set of nodes already visited on this path
2914 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2916 const ir_edge_t *edge;
2917 op_t *op = get_irn_link(irn);
2918 pset *visited_users = pset_new_ptr_default();
2921 if(op->is_remat) return 0;
2923 pset_insert_ptr(visited, irn);
2927 pset *visited_operands = pset_new_ptr(get_irn_arity(irn));
2929 /* visit all operands */
2930 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2931 ir_node *arg = get_irn_n(irn, n);
2932 ilp_var_t copy = op->attr.live_range.args.copies[n];
2934 if(!has_reg_class(si, arg)) continue;
2935 if(pset_find_ptr(visited_operands, arg)) continue;
2936 pset_insert_ptr(visited_operands, arg);
2939 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2940 del_pset(visited_operands);
2941 del_pset(visited_users);
2942 pset_remove_ptr(visited, irn);
2945 pset_insert(copies, INT_TO_PTR(copy), copy);
2946 write_copy_path_cst(si, copies, any_interfere);
2947 pset_remove(copies, INT_TO_PTR(copy), copy);
2948 } else if(!pset_find_ptr(visited, arg)) {
2949 pset_insert(copies, INT_TO_PTR(copy), copy);
2950 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
2951 pset_remove(copies, INT_TO_PTR(copy), copy);
2953 if(paths > MAX_PATHS) {
2954 if(pset_count(copies) == 0) {
2958 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2959 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2960 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2961 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2962 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2966 del_pset(visited_operands);
2967 del_pset(visited_users);
2968 pset_remove_ptr(visited, irn);
2971 } else if(pset_count(copies) == 0) {
2977 del_pset(visited_operands);
2980 /* visit all uses which are phis */
2981 foreach_out_edge(irn, edge) {
2982 ir_node *user = edge->src;
2983 int pos = edge->pos;
2984 op_t *op = get_irn_link(user);
2987 if(!is_Phi(user)) continue;
2988 if(!has_reg_class(si, user)) continue;
2989 if(pset_find_ptr(visited_users, user)) continue;
2990 pset_insert_ptr(visited_users, user);
2992 copy = op->attr.live_range.args.copies[pos];
2994 if(user == target) {
2995 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2996 del_pset(visited_users);
2997 pset_remove_ptr(visited, irn);
3000 pset_insert(copies, INT_TO_PTR(copy), copy);
3001 write_copy_path_cst(si, copies, any_interfere);
3002 pset_remove(copies, INT_TO_PTR(copy), copy);
3003 } else if(!pset_find_ptr(visited, user)) {
3004 pset_insert(copies, INT_TO_PTR(copy), copy);
3005 paths += find_copy_path(si, user, target, any_interfere, copies, visited);
3006 pset_remove(copies, INT_TO_PTR(copy), copy);
3008 if(paths > MAX_PATHS) {
3009 if(pset_count(copies) == 0) {
3013 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3014 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3015 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3016 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3017 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3021 del_pset(visited_users);
3022 pset_remove_ptr(visited, irn);
3025 } else if(pset_count(copies) == 0) {
3031 del_pset(visited_users);
3032 pset_remove_ptr(visited, irn);
3037 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
3039 pset * copies = pset_new_ptr_default();
3040 pset * visited = pset_new_ptr_default();
3042 find_copy_path(si, a, b, any_interfere, copies, visited);
3050 memcopyhandler(spill_ilp_t * si)
3052 interference_t *interference;
3054 /* teste Speicherwerte auf Interferenz */
3056 /* analyze phi classes */
3057 phi_class_compute(si->chordal_env->irg);
3059 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
3060 irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
3062 /* now lets emit the ILP unequations for the crap */
3063 set_foreach(si->interferences, interference) {
3065 ilp_var_t interfere,
3067 ilp_cst_t any_interfere_cst,
3069 const ir_node *a = interference->a;
3070 const ir_node *b = interference->b;
3072 /* any_interf <= \sum interf */
3073 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
3074 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3075 any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3077 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
3079 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
3080 const ir_node *bb = irnlist->irn;
3081 spill_bb_t *spill_bb = get_irn_link(bb);
3086 spilla = set_find_spill(spill_bb->ilp, a);
3089 spillb = set_find_spill(spill_bb->ilp, b);
3092 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
3093 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
3094 /* 2: - mem_in_a - spill_a + interfere <= 0 */
3095 /* 3: - mem_in_b - spill_b + interfere <= 0 */
3096 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
3097 interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3099 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
3100 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
3102 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
3103 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
3104 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
3105 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
3106 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
3108 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
3109 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3111 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3112 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
3113 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
3115 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
3116 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3118 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3119 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
3120 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
3123 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
3125 /* any_interfere >= interf */
3126 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
3127 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3129 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3130 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3133 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
3134 gen_copy_constraints(si,a,b,any_interfere);
3142 return fabs(x) < 0.00001;
3145 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
3147 spill_ilp_t *si = get_irg_link(current_ir_graph);
3149 if(pset_find_ptr(si->all_possible_remats, n)) {
3150 op_t *op = (op_t*)get_irn_link(n);
3151 assert(op && op->is_remat);
3153 if(!op->attr.remat.remat->inverse) {
3154 if(op->attr.remat.pre) {
3155 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
3157 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
3162 op_t *op = (op_t*)get_irn_link(n);
3163 assert(op && op->is_remat);
3165 if(op->attr.remat.pre) {
3166 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
3168 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
3179 dump_graph_with_remats(ir_graph * irg, const char * suffix)
3181 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
3182 be_dump(irg, suffix, dump_ir_block_graph_sched);
3183 set_dump_node_vcgattr_hook(NULL);
3187 * Edge hook to dump the schedule edges with annotated register pressure.
3190 sched_pressure_edge_hook(FILE *F, ir_node *irn)
3192 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
3193 ir_node *prev = sched_prev(irn);
3194 fprintf(F, "edge:{sourcename:\"");
3196 fprintf(F, "\" targetname:\"");
3198 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
3199 fprintf(F, "\" color:magenta}\n");
3205 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
3207 DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3209 dump_consts_local(0);
3210 set_dump_node_edge_hook(sched_pressure_edge_hook);
3211 dump_ir_block_graph(irg, suffix);
3212 set_dump_node_edge_hook(old_edge_hook);
3216 walker_pressure_annotator(ir_node * bb, void * data)
3218 spill_ilp_t *si = data;
3221 pset *live = pset_new_ptr_default();
3224 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3225 irn = be_lv_get_irn(si->lv, bb, i);
3227 if (has_reg_class(si, irn)) {
3228 pset_insert_ptr(live, irn);
3232 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3234 sched_foreach_reverse(bb, irn) {
3236 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3240 if(has_reg_class(si, irn)) {
3241 pset_remove_ptr(live, irn);
3242 if(is_Proj(irn)) ++projs;
3245 if(!is_Proj(irn)) projs = 0;
3247 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3248 ir_node *arg = get_irn_n(irn, n);
3250 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3252 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3259 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3261 be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
3265 connect_all_remats_with_keep(spill_ilp_t * si)
3273 n_remats = pset_count(si->all_possible_remats);
3275 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3278 pset_foreach(si->all_possible_remats, irn) {
3283 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
3285 obstack_free(si->obst, ins);
3290 connect_all_spills_with_keep(spill_ilp_t * si)
3299 n_spills = pset_count(si->spills);
3301 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3304 pset_foreach(si->spills, irn) {
3309 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
3311 obstack_free(si->obst, ins);
3315 /** insert a spill at an arbitrary position */
3316 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3318 ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
3319 ir_graph *irg = get_irn_irg(bl);
3320 ir_node *frame = get_irg_frame(irg);
3324 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
3325 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3327 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3330 * search the right insertion point. a spill of a phi cannot be put
3331 * directly after the phi, if there are some phis behind the one which
3332 * is spilled. Also, a spill of a Proj must be after all Projs of the
3335 * Here's one special case:
3336 * If the spill is in the start block, the spill must be after the frame
3337 * pointer is set up. This is done by setting insert to the end of the block
3338 * which is its default initialization (see above).
3341 if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3344 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3347 sched_add_after(insert, spill);
3352 delete_remat(spill_ilp_t * si, ir_node * remat) {
3354 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3356 sched_remove(remat);
3358 /* kill links to operands */
3359 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3360 set_irn_n(remat, n, bad);
3365 clean_remat_info(spill_ilp_t * si)
3369 remat_info_t *remat_info;
3370 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3372 set_foreach(si->remat_info, remat_info) {
3373 if(!remat_info->remats) continue;
3375 pset_foreach(remat_info->remats, remat)
3377 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3378 set_irn_n((ir_node*)remat->proj, -1, bad);
3379 set_irn_n((ir_node*)remat->proj, 0, bad);
3382 if(get_irn_n_edges(remat->op) == 0) {
3383 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3384 set_irn_n((ir_node*)remat->op, n, bad);
3389 if(remat_info->remats) del_pset(remat_info->remats);
3390 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3395 delete_unnecessary_remats(spill_ilp_t * si)
3397 if(opt_keep_alive & KEEPALIVE_REMATS) {
3399 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3402 // ir_node *end = get_irg_end(si->chordal_env->irg);
3405 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3406 ir_node *keep_arg = get_irn_n(si->keep, n);
3407 op_t *arg_op = get_irn_link(keep_arg);
3410 assert(arg_op->is_remat);
3412 name = si->lpp->vars[arg_op->attr.remat.ilp];
3414 if(is_zero(name->value)) {
3415 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
3416 /* TODO check whether reload is preferred over remat (could be bug) */
3417 delete_remat(si, keep_arg);
3419 if(!arg_op->attr.remat.remat->inverse) {
3420 if(arg_op->attr.remat.pre) {
3421 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3423 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3426 if(arg_op->attr.remat.pre) {
3427 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3429 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3434 set_irn_n(si->keep, n, bad);
3437 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
3438 ir_node *end_arg = get_End_keepalive(end, i);
3440 if(end_arg != si->keep) {
3441 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
3444 keeps = obstack_finish(si->obst);
3445 set_End_keepalives(end, n-1, keeps);
3446 obstack_free(si->obst, keeps);
3449 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
3454 pset_foreach(si->all_possible_remats, remat) {
3455 op_t *remat_op = get_irn_link(remat);
3456 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
3458 if(is_zero(name->value)) {
3459 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
3460 /* TODO check whether reload is preferred over remat (could be bug) */
3461 delete_remat(si, remat);
3463 if(!remat_op->attr.remat.remat->inverse) {
3464 if(remat_op->attr.remat.pre) {
3465 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3467 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3470 if(remat_op->attr.remat.pre) {
3471 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3473 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3482 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3484 pset *spills = pset_new_ptr_default();
3486 const ir_node *next;
3489 defs = set_find_def(si->values, value);
3491 if(defs && defs->spills) {
3492 for(next = defs->spills; next; next = get_irn_link(next)) {
3493 pset_insert_ptr(spills, next);
3501 * @param before The node after which the spill will be placed in the schedule
3504 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3508 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3510 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3512 spill = be_spill2(arch_env, irn, before);
3514 defs = set_insert_def(si->values, value);
3517 /* enter into the linked list */
3518 set_irn_link(spill, defs->spills);
3519 defs->spills = spill;
3521 if(opt_keep_alive & KEEPALIVE_SPILLS)
3522 pset_insert_ptr(si->spills, spill);
3528 * @param before The Phi node which has to be spilled
3531 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3538 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3540 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3541 ins[n] = si->m_unknown;
3544 mem_phi = new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3546 defs = set_insert_def(si->values, phi);
3549 /* enter into the linked list */
3550 set_irn_link(mem_phi, defs->spills);
3551 defs->spills = mem_phi;
3553 #ifdef SCHEDULE_PHIM
3554 sched_add_after(phi, mem_phi);
3557 if(opt_keep_alive & KEEPALIVE_SPILLS)
3558 pset_insert_ptr(si->spills, mem_phi);
3565 * Add remat to list of defs, destroys link field!
3568 insert_remat(spill_ilp_t * si, ir_node * remat)
3571 op_t *remat_op = get_irn_link(remat);
3573 assert(remat_op->is_remat);
3575 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3578 /* enter into the linked list */
3579 set_irn_link(remat, defs->remats);
3580 defs->remats = remat;
3585 * Add reload before operation and add to list of defs
3588 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3593 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3595 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3597 defs = set_find_def(si->values, value);
3599 spill = defs->spills;
3600 assert(spill && "no spill placed before reload");
3602 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3604 /* enter into the linked list */
3605 set_irn_link(reload, defs->remats);
3606 defs->remats = reload;
3611 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3614 ir_node *value = get_irn_n(memoperand->irn, memoperand->pos);
3616 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3618 DBG((si->dbg, LEVEL_2, "\t inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3620 defs = set_find_def(si->values, value);
3622 spill = defs->spills;
3623 assert(spill && "no spill placed before reload");
3625 arch_perform_memory_operand(arch_env, memoperand->irn, spill, memoperand->pos);
3628 void insert_memoperands(spill_ilp_t * si)
3630 memoperand_t *memoperand;
3633 set_foreach(si->memoperands, memoperand) {
3634 name = si->lpp->vars[memoperand->ilp];
3635 if(!is_zero(name->value)) {
3636 perform_memory_operand(si, memoperand);
3642 walker_spill_placer(ir_node * bb, void * data) {
3643 spill_ilp_t *si = (spill_ilp_t*)data;
3645 spill_bb_t *spill_bb = get_irn_link(bb);
3646 pset *spills_to_do = pset_new_ptr_default();
3649 set_foreach(spill_bb->ilp, spill) {
3652 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3653 name = si->lpp->vars[spill->mem_in];
3654 if(!is_zero(name->value)) {
3657 mem_phi = insert_mem_phi(si, spill->irn);
3659 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3663 name = si->lpp->vars[spill->spill];
3664 if(!is_zero(name->value)) {
3665 /* place spill directly after definition */
3666 if(get_nodes_block(spill->irn) == bb) {
3667 insert_spill(si, spill->irn, spill->irn, spill->irn);
3671 /* place spill at bb start */
3672 if(spill->reg_in > 0) {
3673 name = si->lpp->vars[spill->reg_in];
3674 if(!is_zero(name->value)) {
3675 insert_spill(si, spill->irn, spill->irn, bb);
3679 /* place spill after a remat */
3680 pset_insert_ptr(spills_to_do, spill->irn);
3683 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3686 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3687 op_t *op = get_irn_link(irn);
3689 if(be_is_Spill(irn)) continue;
3692 /* TODO fix this if we want to support remats with more than two nodes */
3693 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3694 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3696 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3699 if(pset_find_ptr(spills_to_do, irn)) {
3700 pset_remove_ptr(spills_to_do, irn);
3702 insert_spill(si, irn, irn, irn);
3708 assert(pset_count(spills_to_do) == 0);
3710 /* afterwards free data in block */
3711 del_pset(spills_to_do);
3715 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3717 ir_node *insert_pos = bb;
3719 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3721 /* find last definition of arg value in block */
3726 defs = set_find_def(si->values, value);
3728 if(defs && defs->remats) {
3729 for(next = defs->remats; next; next = get_irn_link(next)) {
3730 if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3731 last = sched_get_time_step(next);
3737 if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3738 last = sched_get_time_step(value);
3742 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
3744 spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3750 phim_fixer(spill_ilp_t *si) {
3753 set_foreach(si->values, defs) {
3754 const ir_node *phi = defs->value;
3755 op_t *op = get_irn_link(phi);
3756 ir_node *phi_m = NULL;
3757 ir_node *next = defs->spills;
3760 if(!is_Phi(phi)) continue;
3763 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3767 next = get_irn_link(next);
3770 if(!phi_m) continue;
3772 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3773 ir_node *value = get_irn_n(phi, n);
3774 defs_t *val_defs = set_find_def(si->values, value);
3776 /* a spill of this value */
3781 ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3782 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3784 if(!is_zero(name->value)) {
3785 spill = insert_mem_copy(si, pred, value);
3787 spill = val_defs->spills;
3790 spill = val_defs->spills;
3793 assert(spill && "no spill placed before PhiM");
3794 set_irn_n(phi_m, n, spill);
3800 walker_reload_placer(ir_node * bb, void * data) {
3801 spill_ilp_t *si = (spill_ilp_t*)data;
3803 spill_bb_t *spill_bb = get_irn_link(bb);
3805 /* reloads at end of block */
3806 if(spill_bb->reloads) {
3809 set_foreach(spill_bb->reloads, keyval) {
3810 ir_node *irn = (ir_node*)keyval->key;
3811 ilp_var_t reload = PTR_TO_INT(keyval->val);
3814 name = si->lpp->vars[reload];
3815 if(!is_zero(name->value)) {
3817 ir_node *insert_pos = bb;
3818 ir_node *prev = sched_block_last_noncf(si, bb);
3819 op_t *prev_op = get_irn_link(prev);
3821 while(be_is_Spill(prev)) {
3822 prev = sched_prev(prev);
3825 prev_op = get_irn_link(prev);
3827 /* insert reload before pre-remats */
3828 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3829 && prev_op->is_remat && prev_op->attr.remat.pre) {
3833 prev = sched_prev(prev);
3834 } while(be_is_Spill(prev));
3836 prev_op = get_irn_link(prev);
3840 reload = insert_reload(si, irn, insert_pos);
3842 if(opt_keep_alive & KEEPALIVE_RELOADS)
3843 pset_insert_ptr(si->spills, reload);
3848 /* walk and insert more reloads and collect remats */
3849 sched_foreach_reverse(bb, irn) {
3850 op_t *op = get_irn_link(irn);
3852 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3853 if(is_Phi(irn)) break;
3856 if(get_irn_mode(irn) != mode_T) {
3857 insert_remat(si, irn);
3862 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3863 ir_node *arg = get_irn_n(irn, n);
3865 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3868 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3869 if(!is_zero(name->value)) {
3871 ir_node *insert_pos = irn;
3872 ir_node *prev = sched_prev(insert_pos);
3875 while(be_is_Spill(prev)) {
3876 prev = sched_prev(prev);
3879 prev_op = get_irn_link(prev);
3881 /* insert reload before pre-remats */
3882 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3883 && prev_op->is_remat && prev_op->attr.remat.pre) {
3887 prev = sched_prev(prev);
3888 } while(be_is_Spill(prev));
3890 prev_op = get_irn_link(prev);
3894 reload = insert_reload(si, arg, insert_pos);
3896 set_irn_n(irn, n, reload);
3898 if(opt_keep_alive & KEEPALIVE_RELOADS)
3899 pset_insert_ptr(si->spills, reload);
3906 del_set(spill_bb->ilp);
3907 if(spill_bb->reloads) del_set(spill_bb->reloads);
3911 walker_collect_used(ir_node * irn, void * data)
3913 bitset_t *used = data;
3915 bitset_set(used, get_irn_idx(irn));
3918 struct kill_helper {
3924 walker_kill_unused(ir_node * bb, void * data)
3926 struct kill_helper *kh = data;
3927 ir_node *bad = get_irg_bad(get_irn_irg(bb));
3931 for(irn=sched_first(bb); !sched_is_end(irn);) {
3932 ir_node *next = sched_next(irn);
3935 if(!bitset_is_set(kh->used, get_irn_idx(irn))) {
3936 if(be_is_Spill(irn) || be_is_Reload(irn)) {
3937 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)));
3939 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3945 set_nodes_block(irn, bad);
3946 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3947 set_irn_n(irn, n, bad);
3955 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3957 struct kill_helper kh;
3959 kh.used = bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3962 irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3963 irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3965 bitset_free(kh.used);
3969 print_irn_pset(pset * p)
3973 pset_foreach(p, irn) {
3974 ir_printf("%+F\n", irn);
3979 dump_phi_class(spill_ilp_t * si, pset * phiclass, const char * file)
3981 FILE *f = fopen(file, "w");
3983 interference_t *interference;
3985 pset_break(phiclass);
3986 set_break(si->interferences);
3988 ir_fprintf(f, "digraph phiclass {\n");
3990 pset_foreach(phiclass, irn) {
3992 ir_fprintf(f, " %F%N [shape=box]\n",irn,irn);
3995 pset_foreach(phiclass, irn) {
3998 if(!is_Phi(irn)) continue;
4000 for(n=get_irn_arity(irn)-1; n>=0; --n) {
4001 ir_node *arg = get_irn_n(irn, n);
4003 ir_fprintf(f, " %F%N -> %F%N\n",irn,irn,arg,arg);
4007 set_foreach(si->interferences, interference) {
4008 const ir_node *a = interference->a;
4009 const ir_node *b = interference->b;
4010 if(get_phi_class(a) == phiclass) {
4011 ir_fprintf(f, " %F%N -> %F%N [color=red,dir=none,style=bold]\n",a,a,b,b);
4020 rewire_uses(spill_ilp_t * si)
4022 dom_front_info_t *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
4024 pset *ignore = pset_new_ptr(1);
4026 pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
4028 /* then fix uses of spills */
4029 set_foreach(si->values, defs) {
4032 const ir_node *next = defs->remats;
4035 reloads = pset_new_ptr_default();
4038 if(be_is_Reload(next)) {
4039 pset_insert_ptr(reloads, next);
4043 next = get_irn_link(next);
4046 spills = get_spills_for_value(si, defs->value);
4047 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));
4048 if(pset_count(spills) > 1) {
4049 //assert(pset_count(reloads) > 0);
4050 // print_irn_pset(spills);
4051 // print_irn_pset(reloads);
4053 be_ssa_constr_set_ignore(dfi, si->lv, spills, ignore);
4060 /* first fix uses of remats and reloads */
4061 set_foreach(si->values, defs) {
4063 const ir_node *next = defs->remats;
4066 nodes = pset_new_ptr_default();
4067 pset_insert_ptr(nodes, defs->value);
4070 pset_insert_ptr(nodes, next);
4071 next = get_irn_link(next);
4074 if(pset_count(nodes) > 1) {
4075 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
4076 be_ssa_constr_set(dfi, si->lv, nodes);
4083 // remove_unused_defs(si);
4085 be_free_dominance_frontiers(dfi);
4090 writeback_results(spill_ilp_t * si)
4092 /* walk through the graph and collect all spills, reloads and remats for a value */
4094 si->values = new_set(cmp_defs, 4096);
4096 DBG((si->dbg, LEVEL_1, "Applying results\n"));
4097 delete_unnecessary_remats(si);
4098 si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
4099 irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
4100 irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
4102 insert_memoperands(si);
4105 /* clean the remat info! there are still back-edges leading there! */
4106 clean_remat_info(si);
4110 connect_all_spills_with_keep(si);
4112 del_set(si->values);
4116 get_n_regs(spill_ilp_t * si)
4118 int arch_n_regs = arch_register_class_n_regs(si->cls);
4120 bitset_t *arch_regs = bitset_malloc(arch_n_regs);
4121 bitset_t *abi_regs = bitset_malloc(arch_n_regs);
4123 arch_put_non_ignore_regs(si->chordal_env->birg->main_env->arch_env, si->cls, arch_regs);
4124 be_abi_put_ignore_regs(si->chordal_env->birg->abi, si->cls, abi_regs);
4126 bitset_andnot(arch_regs, abi_regs);
4127 arch_n_regs = bitset_popcnt(arch_regs);
4129 bitset_free(arch_regs);
4130 bitset_free(abi_regs);
4132 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", arch_n_regs, si->cls->name));
4137 walker_reload_mover(ir_node * bb, void * data)
4139 spill_ilp_t *si = data;
4142 sched_foreach(bb, tmp) {
4143 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
4144 ir_node *reload = tmp;
4147 /* move reload upwards */
4149 int pressure = (int)get_irn_link(reload);
4150 if(pressure < si->n_regs) {
4151 irn = sched_prev(reload);
4152 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
4153 sched_remove(reload);
4154 pressure = (int)get_irn_link(irn);
4156 while(pressure < si->n_regs) {
4157 if( sched_is_end(irn) ||
4158 (be_is_Reload(irn) && has_reg_class(si, irn)) ||
4159 /* do not move reload before its spill */
4160 (irn == be_get_Reload_mem(reload)) ||
4161 /* do not move before phi */
4164 set_irn_link(irn, INT_TO_PTR(pressure+1));
4165 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
4166 irn = sched_prev(irn);
4168 pressure = (int)get_irn_link(irn);
4171 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
4172 sched_put_after(irn, reload);
4179 move_reloads_upward(spill_ilp_t * si)
4181 irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
4186 * Walk all irg blocks and check for interfering spills inside of phi classes
4189 luke_meminterferencechecker(ir_node * bb, void * data)
4191 spill_ilp_t *si = (spill_ilp_t*)data;
4194 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
4195 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
4197 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
4199 /* a is only interesting if it is in my register class and if it is inside a phi class */
4200 if (has_reg_class(si, a) && get_phi_class(a)) {
4201 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)) {
4202 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
4204 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
4206 /* a and b are only interesting if they are in the same phi class */
4207 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
4208 if(values_interfere_in_block(si, bb, a, b)) {
4209 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4218 verify_phiclasses(spill_ilp_t * si)
4220 /* analyze phi classes */
4221 phi_class_compute(si->chordal_env->irg);
4223 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4224 irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
4228 be_spill_remat(const be_chordal_env_t * chordal_env)
4231 char problem_name[256];
4232 char dump_suffix[256];
4233 char dump_suffix2[256];
4234 struct obstack obst;
4237 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
4238 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
4239 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
4241 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4242 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4244 if(opt_verify & VERIFY_DOMINANCE)
4245 be_check_dominance(chordal_env->irg);
4247 obstack_init(&obst);
4248 si.chordal_env = chordal_env;
4250 si.cls = chordal_env->cls;
4251 si.lpp = new_lpp(problem_name, lpp_minimize);
4252 si.remat_info = new_set(cmp_remat_info, 4096);
4253 si.interferences = new_set(cmp_interference, 32);
4254 si.memoperands = new_set(cmp_memoperands, 128);
4255 si.all_possible_remats = pset_new_ptr_default();
4256 si.spills = pset_new_ptr_default();
4257 si.inverse_ops = pset_new_ptr_default();
4258 si.lv = chordal_env->lv;
4260 si.n_regs = get_n_regs(&si);
4262 set_irg_link(chordal_env->irg, &si);
4263 compute_doms(chordal_env->irg);
4265 /* compute phi classes */
4266 // phi_class_compute(chordal_env->irg);
4268 be_analyze_regpressure(chordal_env, "-pre");
4270 DBG((si.dbg, LEVEL_2, "\t initializing\n"));
4271 irg_block_walk_graph(chordal_env->irg, luke_initializer, NULL, &si);
4274 /* collect remats */
4275 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4276 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
4279 /* insert possible remats */
4280 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4281 irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
4282 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4284 if(opt_keep_alive & KEEPALIVE_REMATS) {
4285 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4286 connect_all_remats_with_keep(&si);
4287 /* dump graph with inserted remats */
4288 dump_graph_with_remats(chordal_env->irg, dump_suffix);
4291 /* insert copies for phi arguments not in my regclass */
4292 irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
4294 /* recompute liveness */
4295 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4296 be_liveness_recompute(si.lv);
4299 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4300 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4301 irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
4303 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4304 irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
4307 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4308 memcopyhandler(&si);
4311 if(opt_dump_flags & DUMP_PROBLEM) {
4313 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4314 if ((f = fopen(buf, "wt")) != NULL) {
4315 lpp_dump_plain(si.lpp, f);
4320 if(opt_dump_flags & DUMP_MPS) {
4323 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4324 if((f = fopen(buf, "wt")) != NULL) {
4325 mps_write_mps(si.lpp, s_mps_fixed, f);
4329 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4330 if((f = fopen(buf, "wt")) != NULL) {
4331 mps_write_mst(si.lpp, s_mps_fixed, f);
4336 lpp_check_startvals(si.lpp);
4339 DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4340 lpp_set_time_limit(si.lpp, opt_timeout);
4343 lpp_set_log(si.lpp, stdout);
4346 lpp_solve_cplex(si.lpp);
4348 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4350 assert(lpp_is_sol_valid(si.lpp)
4351 && "solution of ILP must be valid");
4353 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));
4355 if(opt_dump_flags & DUMP_SOLUTION) {
4359 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4360 if ((f = fopen(buf, "wt")) != NULL) {
4362 for (i = 0; i < si.lpp->var_next; ++i) {
4363 lpp_name_t *name = si.lpp->vars[i];
4364 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4370 writeback_results(&si);
4374 kill_all_unused_values_in_schedule(&si);
4376 if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
4377 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
4379 // move reloads upwards
4380 be_liveness_recompute(si.lv);
4381 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4382 move_reloads_upward(&si);
4385 verify_phiclasses(&si);
4388 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4390 dump_pressure_graph(&si, dump_suffix2);
4392 be_analyze_regpressure(chordal_env, "-post");
4394 if(opt_verify & VERIFY_DOMINANCE)
4395 be_check_dominance(chordal_env->irg);
4397 free_dom(chordal_env->irg);
4398 del_set(si.interferences);
4399 del_pset(si.inverse_ops);
4400 del_pset(si.all_possible_remats);
4401 del_set(si.memoperands);
4402 del_pset(si.spills);
4404 obstack_free(&obst, NULL);
4405 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4408 #else /* WITH_ILP */
4411 only_that_you_can_compile_without_WITH_ILP_defined(void)
4415 #endif /* WITH_ILP */