1 /** vim: set sw=4 ts=4:
4 * @author Adam M. Szalkowski & Sebastian Hack
6 * ILP based spilling & rematerialization
8 * Copyright (C) 2006 Universitaet Karlsruhe
9 * Released under the GPL
32 #include "phiclass_t.h"
39 #include <lpp/lpp_net.h>
40 #include <lpp/lpp_cplex.h>
41 //#include <lc_pset.h>
42 #include <libcore/lc_bitset.h>
46 #include "besched_t.h"
51 #include "bespillremat.h"
53 #include "bepressurestat.h"
55 #include "bechordal_t.h"
58 #include <libcore/lc_opts.h>
59 #include <libcore/lc_opts_enum.h>
60 #endif /* WITH_LIBCORE */
62 #define DUMP_PROBLEM 1
64 #define DUMP_SOLUTION 4
66 #define KEEPALIVE_REMATS 1
67 #define KEEPALIVE_SPILLS 2
68 #define KEEPALIVE_RELOADS 4
70 #define VERIFY_MEMINTERF 1
71 #define VERIFY_DOMINANCE 2
74 #define REMATS_BRIGGS 1
75 #define REMATS_NOINVERSE 2
78 static int opt_dump_flags = 0;
79 static int opt_log = 0;
80 static int opt_keep_alive = 0;
81 static int opt_goodwin = 1;
82 static int opt_memcopies = 1;
83 static int opt_memoperands = 1;
84 static int opt_verify = VERIFY_MEMINTERF;
85 static int opt_remats = REMATS_ALL;
86 static int opt_repair_schedule = 0;
87 static int opt_no_enlarge_liveness = 0;
88 static int opt_remat_while_live = 1;
89 static int opt_timeout = 300;
90 static double opt_cost_reload = 8.0
91 static double opt_cost_memoperand = 7.0
92 static double opt_cost_spill = 50.0
93 static double opt_cost_remat = 1.0
97 static const lc_opt_enum_mask_items_t dump_items[] = {
98 { "problem", DUMP_PROBLEM },
100 { "solution", DUMP_SOLUTION },
104 static lc_opt_enum_mask_var_t dump_var = {
105 &opt_dump_flags, dump_items
108 static const lc_opt_enum_mask_items_t keepalive_items[] = {
109 { "remats", KEEPALIVE_REMATS },
110 { "spills", KEEPALIVE_SPILLS },
111 { "reloads", KEEPALIVE_RELOADS },
115 static lc_opt_enum_mask_var_t dump_var = {
116 &opt_keep_alive, keepalive_items
119 static const lc_opt_enum_mask_items_t remats_items[] = {
120 { "none", REMATS_NONE },
121 { "briggs", REMATS_BRIGGS },
122 { "noinverse", REMATS_NOINVERSE },
123 { "ALL", REMATS_ALL },
127 static lc_opt_enum_mask_var_t dump_var = {
128 &opt_remats, remats_items
131 static const lc_opt_table_entry_t options[] = {
132 LC_OPT_ENT_ENUM_MASK("keepalive", "keep alive remats, spills or reloads", &keep_alive_var),
134 LC_OPT_ENT_BOOL ("goodwin", "activate goodwin reduction", &opt_goodwin),
135 LC_OPT_ENT_BOOL ("memcopies", "activate memcopy handling", &opt_memcopies),
136 LC_OPT_ENT_BOOL ("memoperands", "activate memoperands", &opt_memoperands),
137 LC_OPT_ENT_ENUM_INT ("remat", "type of remats to insert (none, briggs, noinverse or all)", &remats_var),
138 LC_OPT_ENT_BOOL ("repair_schedule", "repair the schedule by rematting once used nodes",&opt_repair_schedule),
139 LC_OPT_ENT_BOOL ("no_enlage_liveness", "do not enlarge liveness of operands of remats",&opt_enlarge_liveness),
140 LC_OPT_ENT_BOOL ("remat_while_live", "remat only values that can be used by real ops", &opt_remat_while_live),
142 LC_OPT_ENT_ENUM_MASK("dump", "dump ifg before, after or after each cloud", &dump_var),
143 LC_OPT_ENT_BOOL ("log", "activate the lpp log", &opt_log),
144 LC_OPT_ENT_INT ("timeout", "ILP solver timeout", &opt_timeout),
146 LC_OPT_ENT_DBL ("cost_reload", "cost of a reload", &opt_cost_reload),
147 LC_OPT_ENT_DBL ("cost_memoperand", "cost of a memory operand", &opt_cost_memoperand),
148 LC_OPT_ENT_DBL ("cost_spill", "cost of a spill instruction", &opt_cost_spill),
149 LC_OPT_ENT_DBL ("cost_remat", "cost of a rematerialization", &opt_cost_remat),
153 void be_spill_remat_register_options(lc_opt_entry_t *grp)
155 lc_opt_entry_t *my_grp = lc_opt_get_grp(grp, "remat");
156 lc_opt_add_table(my_grp, options);
160 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
163 //#define SOLVE_LOCAL
164 #define LPP_SERVER "i44pc52"
165 #define LPP_SOLVER "cplex"
171 typedef struct _spill_ilp_t {
172 const arch_register_class_t *cls;
174 const be_chordal_env_t *chordal_env;
177 struct obstack *obst;
179 pset *all_possible_remats;
182 set *values; /**< for collecting all definitions of values before running ssa-construction */
187 DEBUG_ONLY(firm_dbg_module_t * dbg);
190 typedef int ilp_var_t;
191 typedef int ilp_cst_t;
193 typedef struct _spill_bb_t {
198 typedef struct _remat_t {
199 const ir_node *op; /**< for copy_irn */
200 const ir_node *value; /**< the value which is being recomputed by this remat */
201 const ir_node *proj; /**< not NULL if the above op produces a tuple */
202 int cost; /**< cost of this remat */
203 int inverse; /**< nonzero if this is an inverse remat */
207 * Data to be attached to each IR node. For remats this contains the ilp_var
208 * for this remat and for normal ops this contains the ilp_vars for
209 * reloading each operand
211 typedef struct _op_t {
216 const remat_t *remat; /** the remat this op belongs to */
217 int pre; /** 1, if this is a pressure-increasing remat */
221 ir_node *op; /** the operation this live range belongs to */
230 typedef struct _defs_t {
231 const ir_node *value;
232 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
233 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
236 typedef struct _remat_info_t {
237 const ir_node *irn; /**< the irn to which these remats belong */
238 pset *remats; /**< possible remats for this value */
239 pset *remats_by_operand; /**< remats with this value as operand */
242 typedef struct _keyval_t {
247 typedef struct _spill_t {
256 typedef struct _memoperand_t {
257 ir_node *irn; /**< the irn */
258 unsigned int pos; /**< the position of the argument */
259 ilp_var_t ilp; /**< the ilp var for this memory operand */
263 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
265 return chordal_has_class(si->chordal_env, irn);
270 cmp_remat(const void *a, const void *b)
272 const keyval_t *p = a;
273 const keyval_t *q = b;
274 const remat_t *r = p->val;
275 const remat_t *s = q->val;
279 return !(r == s || r->op == s->op);
283 cmp_remat(const void *a, const void *b)
285 const remat_t *r = a;
286 const remat_t *s = a;
288 return !(r == s || r->op == s->op);
292 cmp_spill(const void *a, const void *b, size_t size)
294 const spill_t *p = a;
295 const spill_t *q = b;
297 // return !(p->irn == q->irn && p->bb == q->bb);
298 return !(p->irn == q->irn);
302 cmp_memoperands(const void *a, const void *b, size_t size)
304 const memoperand_t *p = a;
305 const memoperand_t *q = b;
307 return !(p->irn == q->irn && p->pos == q->pos);
311 set_find_keyval(set * set, const void * key)
316 return set_find(set, &query, sizeof(query), HASH_PTR(key));
320 set_insert_keyval(set * set, void * key, void * val)
326 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
330 set_find_def(set * set, const ir_node * value)
335 return set_find(set, &query, sizeof(query), HASH_PTR(value));
339 set_insert_def(set * set, const ir_node * value)
346 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
349 static memoperand_t *
350 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
357 return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
360 static memoperand_t *
361 set_find_memoperand(set * set, const ir_node * irn, unsigned int pos)
365 query.irn = (ir_node*)irn;
367 return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
372 set_find_spill(set * set, const ir_node * value)
376 query.irn = (ir_node*)value;
377 return set_find(set, &query, sizeof(query), HASH_PTR(value));
380 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
381 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
382 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
383 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
384 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
387 cmp_remat_info(const void *a, const void *b, size_t size)
389 const remat_info_t *p = a;
390 const remat_info_t *q = b;
392 return !(p->irn == q->irn);
396 cmp_defs(const void *a, const void *b, size_t size)
401 return !(p->value == q->value);
405 cmp_keyval(const void *a, const void *b, size_t size)
407 const keyval_t *p = a;
408 const keyval_t *q = b;
410 return !(p->key == q->key);
414 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
417 #ifndef EXECFREQ_LOOPDEPH
418 return get_block_execfreq(si->chordal_env->exec_freq, get_block(irn)) + FUDGE;
421 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
423 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
428 get_cost(const spill_ilp_t * si, const ir_node * irn)
430 if(be_is_Spill(irn)) {
431 return opt_cost_spill;
432 } else if(be_is_Reload(irn)){
433 return opt_cost_reload;
435 return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
440 * Checks, whether node and its operands have suitable reg classes
443 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
446 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
447 int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
451 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
454 for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
455 ir_node *op = get_irn_n(irn, n);
456 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
459 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
466 * Try to create a remat from @p op with destination value @p dest_value
468 static INLINE remat_t *
469 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
471 remat_t *remat = NULL;
473 // if(!mode_is_datab(get_irn_mode(dest_value)))
476 if(dest_value == op) {
477 const ir_node *proj = NULL;
479 if(is_Proj(dest_value)) {
480 op = get_irn_n(op, 0);
484 if(!is_rematerializable(si, op))
487 remat = obstack_alloc(si->obst, sizeof(*remat));
489 remat->cost = get_cost(si, op);
490 remat->value = dest_value;
494 arch_inverse_t inverse;
497 /* get the index of the operand we want to retrieve by the inverse op */
498 for (n = get_irn_arity(op)-1; n>=0; --n) {
499 ir_node *arg = get_irn_n(op, n);
501 if(arg == dest_value) break;
505 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", n, op));
507 /* else ask the backend to give an inverse op */
508 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
511 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
513 assert(inverse.n > 0 && "inverse op should have at least one node");
515 for(i=inverse.n-1; i>=0; --i) {
516 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
520 remat = obstack_alloc(si->obst, sizeof(*remat));
521 remat->op = inverse.nodes[0];
522 remat->cost = inverse.costs;
523 remat->value = dest_value;
524 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
527 assert(is_Proj(remat->proj));
529 assert(0 && "I can not handle remats with more than 2 nodes");
536 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
538 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
546 add_remat(const spill_ilp_t * si, const remat_t * remat)
548 remat_info_t *remat_info,
553 assert(remat->value);
555 query.irn = remat->value;
557 query.remats_by_operand = NULL;
558 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
560 if(remat_info->remats == NULL) {
561 remat_info->remats = new_pset(cmp_remat, 4096);
563 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
565 /* insert the remat into the remats_be_operand set of each argument of the remat op */
566 for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
567 ir_node *arg = get_irn_n(remat->op, n);
571 query.remats_by_operand = NULL;
572 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
574 if(remat_info->remats_by_operand == NULL) {
575 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
577 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
582 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
584 const ir_edge_t *edge = get_irn_out_edge_first(irn);
588 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
591 edge = get_irn_out_edge_next(irn, edge);
598 get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
601 unsigned int ret = 0;
603 for(n=get_irn_arity(irn)-1; n>=0; --n) {
604 if(has_reg_class(si, irn)) ++ret;
611 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
616 if( has_reg_class(si, op)
617 && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1)
618 && (opt_remats != REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0)
620 remat = get_remat_from_op(si, op, op);
622 add_remat(si, remat);
626 if(opt_remats == REMATS_ALL) {
627 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
629 for (n = get_irn_arity(op)-1; n>=0; --n) {
630 ir_node *arg = get_irn_n(op, n);
632 if(has_reg_class(si, arg)) {
633 /* try to get an inverse remat */
634 remat = get_remat_from_op(si, arg, op);
636 add_remat(si, remat);
644 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
647 ir_node *def_block = get_nodes_block(val);
653 /* if pos is at end of a basic block */
655 ret = (pos == def_block || block_dominates(def_block, pos));
656 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
660 /* else if this is a normal operation */
661 block = get_nodes_block(pos);
662 if(block == def_block) {
663 if(!sched_is_scheduled(val)) return 1;
665 ret = sched_comes_after(val, pos);
666 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
670 ret = block_dominates(def_block, block);
671 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
675 static INLINE ir_node *
676 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
678 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
682 * Returns first non-Phi node of block @p bb
684 static INLINE ir_node *
685 sched_block_first_nonphi(const ir_node * bb)
687 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
691 sched_skip_proj_predicator(const ir_node * irn, void * data)
693 return (is_Proj(irn));
696 static INLINE ir_node *
697 sched_next_nonproj(const ir_node * irn, int forward)
699 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
703 * Returns next operation node (non-Proj) after @p irn
704 * or the basic block of this node
706 static INLINE ir_node *
707 sched_next_op(const ir_node * irn)
709 ir_node *next = sched_next(irn);
714 return sched_next_nonproj(next, 1);
718 * Returns previous operation node (non-Proj) before @p irn
719 * or the basic block of this node
721 static INLINE ir_node *
722 sched_prev_op(const ir_node * irn)
724 ir_node *prev = sched_prev(irn);
729 return sched_next_nonproj(prev, 0);
733 sched_put_after(ir_node * insert, ir_node * irn)
735 if(is_Block(insert)) {
736 insert = sched_block_first_nonphi(insert);
738 insert = sched_next_op(insert);
740 sched_add_before(insert, irn);
744 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
746 if(is_Block(insert)) {
747 insert = sched_block_last_noncf(si, insert);
749 insert = sched_next_nonproj(insert, 0);
750 insert = sched_prev(insert);
752 sched_add_after(insert, irn);
756 * Tells you whether a @p remat can be placed before the irn @p pos
759 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
761 const ir_node *op = remat->op;
767 prev = sched_block_last_noncf(si, pos);
768 prev = sched_next_nonproj(prev, 0);
770 prev = sched_prev_op(pos);
772 /* do not remat if the rematted value is defined immediately before this op */
773 if(prev == remat->op) {
778 /* this should be just fine, the following OP will be using this value, right? */
780 /* only remat AFTER the real definition of a value (?) */
781 if(!value_is_defined_before(si, pos, remat->value)) {
782 // ir_fprintf(stderr, "error(not defined)");
787 for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
788 const ir_node *arg = get_irn_n(op, n);
790 if(opt_no_enlarge_liveness) {
791 if(has_reg_class(si, arg) && live) {
792 res &= pset_find_ptr(live, arg)?1:0;
794 res &= value_is_defined_before(si, pos, arg);
797 res &= value_is_defined_before(si, pos, arg);
805 * Tells you whether a @p remat can be placed after the irn @p pos
808 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
811 pos = sched_block_first_nonphi(pos);
813 pos = sched_next_op(pos);
816 /* only remat AFTER the real definition of a value (?) */
817 if(!value_is_defined_before(si, pos, remat->value)) {
821 return can_remat_before(si, remat, pos, live);
825 * Collect potetially rematerializable OPs
828 walker_remat_collector(ir_node * irn, void * data)
830 spill_ilp_t *si = data;
832 if(!is_Block(irn) && !is_Phi(irn)) {
833 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
834 get_remats_from_op(si, irn);
839 * Inserts a copy of @p irn before @p pos
842 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
847 bb = is_Block(pos)?pos:get_nodes_block(pos);
848 copy = exact_copy(irn);
850 _set_phi_class(copy, NULL);
851 set_nodes_block(copy, bb);
852 sched_put_before(si, pos, copy);
858 * Inserts a copy of @p irn after @p pos
861 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
866 bb = is_Block(pos)?pos:get_nodes_block(pos);
867 copy = exact_copy(irn);
869 _set_phi_class(copy, NULL);
870 set_nodes_block(copy, bb);
871 sched_put_after(pos, copy);
877 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
881 if(can_remat_after(si, remat, pos, live)) {
886 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
888 copy = insert_copy_after(si, remat->op, pos);
890 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
891 op = obstack_alloc(si->obst, sizeof(*op));
893 op->attr.remat.remat = remat;
894 op->attr.remat.pre = 0;
895 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
897 set_irn_link(copy, op);
898 pset_insert_ptr(si->all_possible_remats, copy);
900 proj_copy = insert_copy_after(si, remat->proj, copy);
901 set_irn_n(proj_copy, 0, copy);
902 set_irn_link(proj_copy, op);
903 pset_insert_ptr(si->all_possible_remats, proj_copy);
915 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
919 if(can_remat_before(si, remat, pos, live)) {
924 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
926 copy = insert_copy_before(si, remat->op, pos);
928 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
929 op = obstack_alloc(si->obst, sizeof(*op));
931 op->attr.remat.remat = remat;
932 op->attr.remat.pre = 1;
933 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
935 set_irn_link(copy, op);
936 pset_insert_ptr(si->all_possible_remats, copy);
938 proj_copy = insert_copy_after(si, remat->proj, copy);
939 set_irn_n(proj_copy, 0, copy);
940 set_irn_link(proj_copy, op);
941 pset_insert_ptr(si->all_possible_remats, proj_copy);
953 get_block_n_succs(const ir_node *block) {
954 const ir_edge_t *edge;
956 assert(edges_activated(current_ir_graph));
958 edge = get_block_succ_first(block);
962 edge = get_block_succ_next(block, edge);
967 is_merge_edge(const ir_node * bb)
970 return get_block_n_succs(bb) == 1;
976 is_diverge_edge(const ir_node * bb)
979 return get_Block_n_cfgpreds(bb) == 1;
985 walker_regclass_copy_insertor(ir_node * irn, void * data)
987 spill_ilp_t *si = data;
989 if(is_Phi(irn) && has_reg_class(si, irn)) {
992 for(n=get_irn_arity(irn)-1; n>=0; --n) {
993 ir_node *phi_arg = get_irn_n(irn, n);
994 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
996 if(!has_reg_class(si, phi_arg)) {
997 ir_node *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
998 ir_node *pos = sched_block_last_noncf(si, bb);
999 op_t *op = obstack_alloc(si->obst, sizeof(*op));
1001 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
1002 sched_add_after(pos, copy);
1003 set_irn_n(irn, n, copy);
1006 op->attr.live_range.args.reloads = NULL;
1007 op->attr.live_range.ilp = ILP_UNDEF;
1008 set_irn_link(copy, op);
1016 * Insert (so far unused) remats into the irg to
1017 * recompute the potential liveness of all values
1020 walker_remat_insertor(ir_node * bb, void * data)
1022 spill_ilp_t *si = data;
1023 spill_bb_t *spill_bb;
1026 pset *live = pset_new_ptr_default();
1028 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1030 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1031 ir_node *value = be_lv_get_irn(si->lv, bb, i);
1033 /* add remats at end of block */
1034 if (has_reg_class(si, value)) {
1035 pset_insert_ptr(live, value);
1039 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1040 set_irn_link(bb, spill_bb);
1042 irn = sched_last(bb);
1043 while(!sched_is_end(irn)) {
1050 next = sched_prev(irn);
1052 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
1054 if(is_Phi(irn) || is_Proj(irn)) {
1057 if(has_reg_class(si, irn)) {
1058 pset_remove_ptr(live, irn);
1061 op = obstack_alloc(si->obst, sizeof(*op));
1063 op->attr.live_range.args.reloads = NULL;
1064 op->attr.live_range.ilp = ILP_UNDEF;
1065 set_irn_link(irn, op);
1071 op = obstack_alloc(si->obst, sizeof(*op));
1073 op->attr.live_range.ilp = ILP_UNDEF;
1074 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1075 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1076 set_irn_link(irn, op);
1078 args = pset_new_ptr_default();
1080 /* collect arguments of op */
1081 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1082 ir_node *arg = get_irn_n(irn, n);
1084 pset_insert_ptr(args, arg);
1087 /* set args of op already live in epilog */
1088 pset_foreach(args, arg) {
1089 if(has_reg_class(si, arg)) {
1090 pset_insert_ptr(live, arg);
1093 /* delete defined value from live set */
1094 if(has_reg_class(si, irn)) {
1095 pset_remove_ptr(live, irn);
1099 remat_args = pset_new_ptr_default();
1101 /* insert all possible remats before irn */
1102 pset_foreach(args, arg) {
1103 remat_info_t *remat_info,
1107 /* continue if the operand has the wrong reg class
1109 if(!has_reg_class(si, arg))
1113 query.remats = NULL;
1114 query.remats_by_operand = NULL;
1115 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1121 if(remat_info->remats) {
1122 pset_foreach(remat_info->remats, remat) {
1123 ir_node *remat_irn = NULL;
1125 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
1126 if(opt_remat_while_live) {
1127 if(pset_find_ptr(live, remat->value)) {
1128 remat_irn = insert_remat_before(si, remat, irn, live);
1131 remat_irn = insert_remat_before(si, remat, irn, live);
1134 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1135 ir_node *remat_arg = get_irn_n(remat_irn, n);
1137 if(!has_reg_class(si, remat_arg)) continue;
1139 pset_insert_ptr(remat_args, remat_arg);
1146 /* now we add remat args to op's args because they could also die at this op */
1147 pset_foreach(args,arg) {
1148 if(pset_find_ptr(remat_args, arg)) {
1149 pset_remove_ptr(remat_args, arg);
1152 pset_foreach(remat_args,arg) {
1153 pset_insert_ptr(args, arg);
1156 /* insert all possible remats after irn */
1157 pset_foreach(args, arg) {
1158 remat_info_t *remat_info,
1162 /* continue if the operand has the wrong reg class */
1163 if(!has_reg_class(si, arg))
1167 query.remats = NULL;
1168 query.remats_by_operand = NULL;
1169 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1175 /* do not place post remats after jumps */
1176 if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
1178 if(remat_info->remats_by_operand) {
1179 pset_foreach(remat_info->remats_by_operand, remat) {
1180 /* do not insert remats producing the same value as one of the operands */
1181 if(!pset_find_ptr(args, remat->value)) {
1182 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
1183 if(opt_remat_while_live) {
1184 if(pset_find_ptr(live, remat->value)) {
1185 insert_remat_after(si, remat, irn, live);
1188 insert_remat_after(si, remat, irn, live);
1195 del_pset(remat_args);
1200 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_in, i) {
1201 ir_node *value = be_lv_get_irn(si->lv, bb, i);
1203 /* add remats at end if successor has multiple predecessors */
1204 if(is_merge_edge(bb)) {
1205 /* add remats at end of block */
1206 if (be_is_live_end(si->lv, bb, value) && has_reg_class(si, value)) {
1207 remat_info_t *remat_info,
1212 query.remats = NULL;
1213 query.remats_by_operand = NULL;
1214 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1216 if(remat_info && remat_info->remats) {
1217 pset_foreach(remat_info->remats, remat) {
1218 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1220 insert_remat_before(si, remat, bb, NULL);
1225 if(is_diverge_edge(bb)) {
1226 /* add remat2s at beginning of block */
1227 if ((be_is_live_in(si->lv, bb, value) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) {
1228 remat_info_t *remat_info,
1233 query.remats = NULL;
1234 query.remats_by_operand = NULL;
1235 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1237 if(remat_info && remat_info->remats) {
1238 pset_foreach(remat_info->remats, remat) {
1239 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at beginning of block %+F\n", remat->op, bb));
1241 /* put the remat here if all its args are available */
1242 insert_remat_after(si, remat, bb, NULL);
1252 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1255 luke_endwalker(ir_node * bb, void * data)
1257 spill_ilp_t *si = (spill_ilp_t*)data;
1263 spill_bb_t *spill_bb = get_irn_link(bb);
1267 live = pset_new_ptr_default();
1268 use_end = pset_new_ptr_default();
1270 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1271 irn = be_lv_get_irn(si->lv, bb, i);
1272 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1275 pset_insert_ptr(live, irn);
1276 op = get_irn_link(irn);
1277 assert(!op->is_remat);
1281 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1282 /* their reg_out must always be set */
1283 sched_foreach_reverse(bb, irn) {
1286 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1288 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1289 ir_node *irn_arg = get_irn_n(irn, n);
1291 if(has_reg_class(si, irn_arg)) {
1292 pset_insert_ptr(use_end, irn_arg);
1297 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1298 //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1299 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1301 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1303 /* if this is a merge edge we can reload at the end of this block */
1304 if(is_merge_edge(bb)) {
1305 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1306 } else if(pset_count(use_end)){
1307 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1309 spill_bb->reloads = NULL;
1312 pset_foreach(live,irn) {
1316 int default_spilled;
1319 /* handle values used by control flow nodes later separately */
1320 if(pset_find_ptr(use_end, irn)) continue;
1323 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1325 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1327 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1328 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1329 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1331 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1332 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1334 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1335 /* by default spill value right after definition */
1336 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1337 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1339 if(is_merge_edge(bb)) {
1343 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1344 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 0.0);
1345 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1347 /* reload <= mem_out */
1348 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1349 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1350 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1353 spill->reg_in = ILP_UNDEF;
1354 spill->mem_in = ILP_UNDEF;
1357 pset_foreach(use_end,irn) {
1361 ilp_cst_t end_use_req,
1364 int default_spilled;
1367 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1369 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1371 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1372 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1373 /* if irn is used at the end of the block, then it is live anyway */
1374 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1376 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1377 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1379 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1380 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1381 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1383 /* reload for use be control flow op */
1384 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1385 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1386 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1388 /* reload <= mem_out */
1389 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1390 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1391 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1393 spill->reg_in = ILP_UNDEF;
1394 spill->mem_in = ILP_UNDEF;
1396 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1397 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1398 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1406 next_post_remat(const ir_node * irn)
1412 next = sched_block_first_nonphi(irn);
1414 next = sched_next_op(irn);
1417 if(sched_is_end(next))
1420 op = get_irn_link(next);
1421 if(op->is_remat && !op->attr.remat.pre) {
1430 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1436 ret = sched_block_last_noncf(si, irn);
1437 ret = sched_next(ret);
1438 ret = sched_prev_op(ret);
1440 ret = sched_prev_op(irn);
1443 if(sched_is_end(ret) || is_Phi(ret))
1446 op = (op_t*)get_irn_link(ret);
1447 if(op->is_remat && op->attr.remat.pre) {
1455 * Find a remat of value @p value in the epilog of @p pos
1458 find_post_remat(const ir_node * value, const ir_node * pos)
1460 while((pos = next_post_remat(pos)) != NULL) {
1463 op = get_irn_link(pos);
1464 assert(op->is_remat && !op->attr.remat.pre);
1466 if(op->attr.remat.remat->value == value)
1467 return (ir_node*)pos;
1470 const ir_edge_t *edge;
1471 foreach_out_edge(pos, edge) {
1472 ir_node *proj = get_edge_src_irn(edge);
1473 assert(is_Proj(proj));
1483 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1485 spill_bb_t *spill_bb = get_irn_link(bb);
1489 int default_spilled;
1492 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1494 double spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1496 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1498 spill->reg_out = ILP_UNDEF;
1499 spill->reg_in = ILP_UNDEF;
1500 spill->mem_in = ILP_UNDEF;
1502 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1503 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1505 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1506 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1507 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1514 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1519 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1520 irn = be_lv_get_irn(si->lv, bb, i);
1522 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1523 pset_insert_ptr(live, irn);
1527 irn = sched_last(bb);
1529 /* all values eaten by control flow operations are also live until the end of the block */
1530 sched_foreach_reverse(bb, irn) {
1533 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1535 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1536 ir_node *arg = get_irn_n(irn,i);
1538 if(has_reg_class(si, arg)) {
1539 pset_insert_ptr(live, arg);
1546 * Inserts ILP-constraints and variables for memory copying before the given position
1549 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1551 const ir_node *succ;
1552 const ir_edge_t *edge;
1553 spill_bb_t *spill_bb = get_irn_link(block);
1562 assert(edges_activated(current_ir_graph));
1564 edge = get_block_succ_first(block);
1570 edge = get_block_succ_next(block, edge);
1571 /* next block can only contain phis, if this is a merge edge */
1574 ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1575 copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1577 ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1578 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1580 pset_foreach(live, tmp) {
1583 op_t *op = get_irn_link(irn);
1584 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1586 spill = set_find_spill(spill_bb->ilp, tmp);
1589 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1591 lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1593 sched_foreach(succ, phi) {
1594 const ir_node *to_copy;
1596 spill_t *to_copy_spill;
1597 op_t *phi_op = get_irn_link(phi);
1598 ilp_var_t reload = ILP_UNDEF;
1601 if(!is_Phi(phi)) break;
1602 if(!has_reg_class(si, phi)) continue;
1604 to_copy = get_irn_n(phi, pos);
1606 to_copy_op = get_irn_link(to_copy);
1608 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1609 assert(to_copy_spill);
1611 if(spill_bb->reloads) {
1612 keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1615 reload = PTR_TO_INT(keyval->val);
1619 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1620 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1622 /* copy - reg_out - reload - remat - live_range <= 0 */
1623 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1624 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1625 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1626 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1627 foreach_pre_remat(si, block, tmp) {
1628 op_t *remat_op = get_irn_link(tmp);
1629 if(remat_op->attr.remat.remat->value == to_copy) {
1630 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1634 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1635 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1637 /* copy - reg_out - copyreg <= 0 */
1638 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1639 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1640 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1646 * Walk all irg blocks and emit this ILP
1649 luke_blockwalker(ir_node * bb, void * data)
1651 spill_ilp_t *si = (spill_ilp_t*)data;
1656 spill_bb_t *spill_bb = get_irn_link(bb);
1659 pset *defs = pset_new_ptr_default();
1660 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
1663 live = pset_new_ptr_default();
1665 /****************************************
1666 * B A S I C B L O C K E N D
1667 ***************************************/
1670 /* init live values at end of block */
1671 get_live_end(si, bb, live);
1673 pset_foreach(live, irn) {
1675 ilp_var_t reload = ILP_UNDEF;
1677 spill = set_find_spill(spill_bb->ilp, irn);
1680 if(spill_bb->reloads) {
1681 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1684 reload = PTR_TO_INT(keyval->val);
1688 op = get_irn_link(irn);
1689 assert(!op->is_remat);
1691 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1692 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1693 op->attr.live_range.op = bb;
1695 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1696 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1698 /* reg_out - reload - remat - live_range <= 0 */
1699 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1700 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1701 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1702 foreach_pre_remat(si, bb, tmp) {
1703 op_t *remat_op = get_irn_link(tmp);
1704 if(remat_op->attr.remat.remat->value == irn) {
1705 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1708 /* maybe we should also assure that reg_out >= live_range etc. */
1712 insert_mem_copy_position(si, live, bb);
1715 * start new live ranges for values used by remats at end of block
1716 * and assure the remat args are available
1718 foreach_pre_remat(si, bb, tmp) {
1719 op_t *remat_op = get_irn_link(tmp);
1722 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1723 ir_node *remat_arg = get_irn_n(tmp, n);
1724 op_t *arg_op = get_irn_link(remat_arg);
1727 if(!has_reg_class(si, remat_arg)) continue;
1729 /* if value is becoming live through use by remat */
1730 if(!pset_find_ptr(live, remat_arg)) {
1731 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1732 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1734 arg_op->attr.live_range.ilp = prev_lr;
1735 arg_op->attr.live_range.op = bb;
1737 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1739 pset_insert_ptr(live, remat_arg);
1740 add_to_spill_bb(si, bb, remat_arg);
1743 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1744 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1745 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1747 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1748 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1750 /* use reload placed for this argument */
1751 if(spill_bb->reloads) {
1752 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1755 ilp_var_t reload = PTR_TO_INT(keyval->val);
1757 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1762 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1767 /**************************************
1768 * B A S I C B L O C K B O D Y
1769 **************************************/
1771 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1777 ilp_cst_t check_pre,
1783 ilp_cst_t one_memoperand;
1785 /* iterate only until first phi */
1789 op = get_irn_link(irn);
1791 if(op->is_remat) continue;
1792 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1794 /* collect defined values */
1795 if(has_reg_class(si, irn)) {
1796 pset_insert_ptr(defs, irn);
1800 if(is_Proj(irn)) continue;
1803 * init set of irn's arguments
1804 * and all possibly used values around this op
1805 * and values defined by post remats
1807 args = new_set(cmp_keyval, get_irn_arity(irn));
1808 used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1809 remat_defs = pset_new_ptr(pset_count(live));
1811 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1812 ir_node *irn_arg = get_irn_n(irn, n);
1813 if(has_reg_class(si, irn_arg)) {
1814 set_insert_keyval(args, irn_arg, (void*)n);
1815 pset_insert_ptr(used, irn_arg);
1818 foreach_post_remat(irn, tmp) {
1819 op_t *remat_op = get_irn_link(tmp);
1821 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1823 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1824 ir_node *remat_arg = get_irn_n(tmp, n);
1825 if(has_reg_class(si, remat_arg)) {
1826 pset_insert_ptr(used, remat_arg);
1830 foreach_pre_remat(si, irn, tmp) {
1831 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1832 ir_node *remat_arg = get_irn_n(tmp, n);
1833 if(has_reg_class(si, remat_arg)) {
1834 pset_insert_ptr(used, remat_arg);
1839 /**********************************
1840 * I N E P I L O G O F irn
1841 **********************************/
1843 /* ensure each dying value is used by only one post remat */
1844 pset_foreach(used, tmp) {
1845 ir_node *value = tmp;
1846 op_t *value_op = get_irn_link(value);
1851 foreach_post_remat(irn, remat) {
1852 op_t *remat_op = get_irn_link(remat);
1854 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1855 ir_node *remat_arg = get_irn_n(remat, n);
1857 /* if value is used by this remat add it to constraint */
1858 if(remat_arg == value) {
1860 /* sum remat2s <= 1 + n_remats*live_range */
1861 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1862 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
1866 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1872 // value_op->attr.live_range.ilp != ILP_UNDEF
1873 if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
1874 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1878 /* ensure at least one value dies at post remat */
1879 foreach_post_remat(irn, tmp) {
1880 op_t *remat_op = get_irn_link(tmp);
1881 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
1884 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
1885 remat_arg = get_irn_n(tmp, n);
1887 if(has_reg_class(si, remat_arg)) {
1889 /* does arg always die at this op? */
1890 if(!pset_find_ptr(live, remat_arg))
1891 goto skip_one_must_die;
1893 pset_insert_ptr(remat_args, remat_arg);
1897 /* remat + \sum live_range(remat_arg) <= |args| */
1898 ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
1899 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
1900 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1902 pset_foreach(remat_args, remat_arg) {
1903 op_t *arg_op = get_irn_link(remat_arg);
1905 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1909 del_pset(remat_args);
1912 /* new live ranges for values from L\U defined by post remats */
1913 pset_foreach(live, tmp) {
1914 ir_node *value = tmp;
1915 op_t *value_op = get_irn_link(value);
1917 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
1918 ilp_var_t prev_lr = ILP_UNDEF;
1921 if(pset_find_ptr(remat_defs, value)) {
1923 /* next_live_range <= prev_live_range + sum remat2s */
1924 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1925 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1927 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1928 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1930 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1931 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1933 foreach_post_remat(irn, remat) {
1934 op_t *remat_op = get_irn_link(remat);
1936 /* if value is being rematerialized by this remat */
1937 if(value == remat_op->attr.remat.remat->value) {
1938 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1942 value_op->attr.live_range.ilp = prev_lr;
1943 value_op->attr.live_range.op = irn;
1948 /* requirements for post remats and start live ranges from L/U' for values dying here */
1949 foreach_post_remat(irn, tmp) {
1950 op_t *remat_op = get_irn_link(tmp);
1953 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1954 ir_node *remat_arg = get_irn_n(tmp, n);
1955 op_t *arg_op = get_irn_link(remat_arg);
1957 if(!has_reg_class(si, remat_arg)) continue;
1959 /* only for values in L\U (TODO and D?), the others are handled with post_use */
1960 if(!pset_find_ptr(used, remat_arg)) {
1961 /* remat <= live_range(remat_arg) */
1962 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
1963 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1965 /* if value is becoming live through use by remat2 */
1966 if(!pset_find_ptr(live, remat_arg)) {
1969 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1970 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1972 arg_op->attr.live_range.ilp = lr;
1973 arg_op->attr.live_range.op = irn;
1975 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
1977 pset_insert_ptr(live, remat_arg);
1978 add_to_spill_bb(si, bb, remat_arg);
1981 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1982 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1987 d = pset_count(defs);
1988 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
1990 /* count how many regs irn needs for arguments */
1991 u = set_count(args);
1994 /* check the register pressure in the epilog */
1995 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
1996 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
1997 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
1999 /* add L\U' to check_post */
2000 pset_foreach(live, tmp) {
2001 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
2002 /* if a live value is not used by irn */
2003 tmp_op = get_irn_link(tmp);
2004 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
2008 /***********************************************************
2009 * 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
2010 **********************************************************/
2013 pset_foreach(used, tmp) {
2019 op_t *arg_op = get_irn_link(arg);
2022 spill = add_to_spill_bb(si, bb, arg);
2024 /* new live range for each used value */
2025 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2026 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2028 /* the epilog stuff - including post_use, check_post, check_post_remat */
2029 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2030 post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2032 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2034 /* arg is live throughout epilog if the next live_range is in a register */
2035 if(pset_find_ptr(live, arg)) {
2036 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
2038 /* post_use >= next_lr + remat */
2039 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2040 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2041 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2042 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2046 /* if value is not an arg of op and not possibly defined by post remat
2047 * then it may only die and not become live
2049 if(!set_find_keyval(args, arg)) {
2050 /* post_use <= prev_lr */
2051 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2052 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2053 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2054 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2056 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2057 /* next_lr <= prev_lr */
2058 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2059 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2060 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2061 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2066 /* forall post remat which use arg add a similar cst */
2067 foreach_post_remat(irn, remat) {
2070 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2071 ir_node *remat_arg = get_irn_n(remat, n);
2072 op_t *remat_op = get_irn_link(remat);
2074 if(remat_arg == arg) {
2075 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
2077 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2078 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2079 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2080 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2085 if(opt_memoperands) {
2086 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2087 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2088 ilp_var_t memoperand;
2090 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2091 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0);
2092 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2094 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2095 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2097 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2098 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2099 // if(arg_op->attr.live_range.ilp != ILP_UNDEF)
2100 // lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2105 /* new live range begins for each used value */
2106 arg_op->attr.live_range.ilp = prev_lr;
2107 arg_op->attr.live_range.op = irn;
2109 /*if(!pset_find_ptr(live, arg)) {
2110 pset_insert_ptr(live, arg);
2111 add_to_spill_bb(si, bb, arg);
2113 pset_insert_ptr(live, arg);
2117 /* just to be sure */
2118 check_post = ILP_UNDEF;
2127 /* check the register pressure in the prolog */
2128 /* sum_{L\U} lr <= k - |U| */
2129 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2130 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2132 /* for the prolog remove defined values from the live set */
2133 pset_foreach(defs, tmp) {
2134 pset_remove_ptr(live, tmp);
2137 if(opt_memoperands) {
2138 ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2139 one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2142 /***********************************************************
2143 * 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
2144 **********************************************************/
2147 set_foreach(args, keyval) {
2149 const ir_node *arg = keyval->key;
2150 int i = PTR_TO_INT(keyval->val);
2151 op_t *arg_op = get_irn_link(arg);
2152 ilp_cst_t requirements;
2155 spill = set_find_spill(spill_bb->ilp, arg);
2158 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2159 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);
2161 /* reload <= mem_out */
2162 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2163 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2164 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2165 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2167 /* requirement: arg must be in register for use */
2168 /* reload + remat + live_range == 1 */
2169 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2170 requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2172 lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2173 lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2174 foreach_pre_remat(si, irn, tmp) {
2175 op_t *remat_op = get_irn_link(tmp);
2176 if(remat_op->attr.remat.remat->value == arg) {
2177 lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2181 if(opt_memoperands) {
2183 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2184 if(get_irn_n(irn, n) == arg) {
2188 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2189 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2190 memoperand_t *memoperand;
2191 memoperand = set_find_memoperand(si->memoperands, irn, n);
2193 /* memoperand <= mem_out */
2194 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2195 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2196 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2197 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2199 /* the memoperand is only sufficient if it is used once by the op */
2200 if(n_memoperands == 1)
2201 lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2203 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2205 /* we have one more free register if we use a memory operand */
2206 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2212 /* iterate over L\U */
2213 pset_foreach(live, tmp) {
2214 if(!set_find_keyval(args, tmp)) {
2215 /* if a live value is not used by irn */
2216 tmp_op = get_irn_link(tmp);
2217 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2222 /* requirements for remats */
2223 /* start new live ranges for values used by remats */
2224 foreach_pre_remat(si, irn, tmp) {
2225 op_t *remat_op = get_irn_link(tmp);
2228 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2229 ir_node *remat_arg = get_irn_n(tmp, n);
2230 op_t *arg_op = get_irn_link(remat_arg);
2232 if(!has_reg_class(si, remat_arg)) continue;
2234 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2235 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2236 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2238 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2239 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2241 /* if remat arg is also used by current op then we can use reload placed for this argument */
2242 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2243 int index = (int)keyval->val;
2245 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2253 /*************************
2254 * D O N E W I T H O P
2255 *************************/
2257 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2259 pset_foreach(live, tmp) {
2260 assert(has_reg_class(si, tmp));
2263 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2264 ir_node *arg = get_irn_n(irn, n);
2266 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2269 del_pset(remat_defs);
2273 defs = pset_new_ptr_default();
2278 /***************************************
2279 * B E G I N N I N G O F B L O C K
2280 ***************************************/
2283 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2284 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2286 pset_foreach(live, irn) {
2287 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2290 /* construct mem_outs for all values */
2292 set_foreach(spill_bb->ilp, spill) {
2293 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2294 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2296 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2297 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2299 if(pset_find_ptr(live, spill->irn)) {
2300 int default_spilled;
2301 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2303 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2304 default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2305 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2306 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2308 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2310 op_t *op = get_irn_link(spill->irn);
2312 /* do we have to copy a phi argument? */
2313 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2314 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2316 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2317 const ir_node *arg = get_irn_n(spill->irn, n);
2323 /* argument already done? */
2324 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2326 /* get sum of execution frequencies of blocks with the same phi argument */
2327 for(m=n; m>=0; --m) {
2328 const ir_node *arg2 = get_irn_n(spill->irn, m);
2331 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2335 /* copies are not for free */
2336 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2337 var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
2339 for(m=n; m>=0; --m) {
2340 const ir_node *arg2 = get_irn_n(spill->irn, m);
2343 op->attr.live_range.args.copies[m] = var;
2348 /* copy <= mem_in */
2349 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2350 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2351 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2352 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2360 /* L\U is empty at bb start */
2361 /* arg is live throughout epilog if it is reg_in into this block */
2363 /* check the register pressure at the beginning of the block
2366 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2367 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2369 pset_foreach(live, irn) {
2372 spill = set_find_spill(spill_bb->ilp, irn);
2375 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2376 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2378 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2380 /* spill + mem_in <= 1 */
2381 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2382 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2384 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2385 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2388 foreach_post_remat(bb, irn) {
2389 op_t *remat_op = get_irn_link(irn);
2391 DBG((si->dbg, LEVEL_4, "\t next post remat: %+F\n", irn));
2392 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
2394 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2397 /* forall post remats add requirements */
2398 foreach_post_remat(bb, tmp) {
2401 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2402 ir_node *remat_arg = get_irn_n(tmp, n);
2403 op_t *remat_op = get_irn_link(tmp);
2405 if(!has_reg_class(si, remat_arg)) continue;
2407 spill = set_find_spill(spill_bb->ilp, remat_arg);
2410 /* remat <= reg_in_argument */
2411 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2412 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2413 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2414 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2418 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2419 pset_foreach(live, irn) {
2423 spill = set_find_spill(spill_bb->ilp, irn);
2424 assert(spill && spill->irn == irn);
2426 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2427 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2430 ir_node *phi_arg = get_Phi_pred(irn, n);
2431 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2432 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2434 op_t *op = get_irn_link(irn);
2436 /* although the phi is in the right regclass one or more of
2437 * its arguments can be in a different one or at least to
2440 if(has_reg_class(si, phi_arg)) {
2441 /* mem_in < mem_out_arg + copy */
2442 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2443 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2445 /* reg_in < reg_out_arg */
2446 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2447 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2449 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2450 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2452 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2455 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2456 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2457 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2461 /* else assure the value arrives on all paths in the same resource */
2463 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2466 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2467 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2470 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2471 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2472 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2473 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2475 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2476 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2478 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2481 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2482 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2487 /* first live ranges from reg_ins */
2488 pset_foreach(live, irn) {
2489 op_t *op = get_irn_link(irn);
2491 spill = set_find_spill(spill_bb->ilp, irn);
2492 assert(spill && spill->irn == irn);
2494 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2495 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2496 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2497 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2499 foreach_post_remat(bb, tmp) {
2500 op_t *remat_op = get_irn_link(tmp);
2502 if(remat_op->attr.remat.remat->value == irn) {
2503 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2508 /* walk forward now and compute constraints for placing spills */
2509 /* this must only be done for values that are not defined in this block */
2510 /* TODO are these values at start of block? if yes, just check whether this is a diverge edge and skip the loop */
2511 pset_foreach(live, irn) {
2513 * if value is defined in this block we can anways place the spill directly after the def
2514 * -> no constraint necessary
2516 if(!is_Phi(irn) && get_nodes_block(irn) == bb) continue;
2519 spill = set_find_spill(spill_bb->ilp, irn);
2522 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2523 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2525 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2526 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2529 sched_foreach_op(bb, tmp) {
2530 op_t *op = get_irn_link(tmp);
2532 if(is_Phi(tmp)) continue;
2533 assert(!is_Proj(tmp));
2536 const ir_node *value = op->attr.remat.remat->value;
2539 /* only collect remats up to the first real use of a value */
2540 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2545 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2546 ir_node *arg = get_irn_n(tmp, n);
2549 /* if a value is used stop collecting remats */
2562 typedef struct _irnlist_t {
2563 struct list_head list;
2567 typedef struct _interference_t {
2568 struct list_head blocklist;
2574 cmp_interference(const void *a, const void *b, size_t size)
2576 const interference_t *p = a;
2577 const interference_t *q = b;
2579 return !(p->a == q->a && p->b == q->b);
2582 static interference_t *
2583 set_find_interference(set * set, ir_node * a, ir_node * b)
2585 interference_t query;
2587 query.a = (a>b)?a:b;
2588 query.b = (a>b)?b:a;
2590 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2593 static interference_t *
2594 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2596 interference_t query,
2598 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2602 result = set_find_interference(set, a, b);
2605 list_add(&list->list, &result->blocklist);
2609 query.a = (a>b)?a:b;
2610 query.b = (a>b)?b:a;
2612 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2614 INIT_LIST_HEAD(&result->blocklist);
2615 list_add(&list->list, &result->blocklist);
2621 values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
2623 const ir_edge_t *edge;
2625 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2626 /* both values are live in, so they interfere */
2630 /* ensure a dominates b */
2631 if(value_dominates(b,a)) {
2637 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2640 /* the following code is stolen from bera.c */
2641 if(be_is_live_end(si->lv, bb, a))
2644 foreach_out_edge(a, edge) {
2645 const ir_node *user = edge->src;
2646 if(get_nodes_block(user) == bb
2649 && value_dominates(b, user))
2657 * Walk all irg blocks and collect interfering values inside of phi classes
2660 luke_interferencewalker(ir_node * bb, void * data)
2662 spill_ilp_t *si = (spill_ilp_t*)data;
2665 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2666 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
2667 op_t *a_op = get_irn_link(a);
2670 /* a is only interesting if it is in my register class and if it is inside a phi class */
2671 if (has_reg_class(si, a) && get_phi_class(a)) {
2675 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)) {
2676 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
2677 op_t *b_op = get_irn_link(b);
2680 /* a and b are only interesting if they are in the same phi class */
2681 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2685 if(values_interfere_in_block(si, bb, a, b)) {
2686 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2687 set_insert_interference(si, si->interferences, a, b, bb);
2695 static unsigned int copy_path_id = 0;
2698 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2705 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2706 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2708 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2710 pset_foreach(copies, ptr) {
2711 copy = PTR_TO_INT(ptr);
2712 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2717 * @parameter copies contains a path of copies which lead us to irn
2718 * @parameter visited contains a set of nodes already visited on this path
2721 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2723 const ir_edge_t *edge;
2724 op_t *op = get_irn_link(irn);
2725 pset *visited_users = pset_new_ptr_default();
2728 if(op->is_remat) return 0;
2730 pset_insert_ptr(visited, irn);
2734 pset *visited_operands = pset_new_ptr(get_irn_arity(irn));
2736 /* visit all operands */
2737 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2738 ir_node *arg = get_irn_n(irn, n);
2739 ilp_var_t copy = op->attr.live_range.args.copies[n];
2741 if(!has_reg_class(si, arg)) continue;
2742 if(pset_find_ptr(visited_operands, arg)) continue;
2743 pset_insert_ptr(visited_operands, arg);
2746 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2747 del_pset(visited_operands);
2748 del_pset(visited_users);
2749 pset_remove_ptr(visited, irn);
2752 pset_insert(copies, INT_TO_PTR(copy), copy);
2753 write_copy_path_cst(si, copies, any_interfere);
2754 pset_remove(copies, INT_TO_PTR(copy), copy);
2755 } else if(!pset_find_ptr(visited, arg)) {
2756 pset_insert(copies, INT_TO_PTR(copy), copy);
2757 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
2758 pset_remove(copies, INT_TO_PTR(copy), copy);
2760 /*if(paths > MAX_PATHS) {
2761 if(pset_count(copies) == 0) {
2765 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2766 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2767 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2768 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2769 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2773 del_pset(visited_operands);
2774 del_pset(visited_users);
2775 pset_remove_ptr(visited, irn);
2778 } else if(pset_count(copies) == 0) {
2784 del_pset(visited_operands);
2787 /* visit all uses which are phis */
2788 foreach_out_edge(irn, edge) {
2789 ir_node *user = edge->src;
2790 int pos = edge->pos;
2791 op_t *op = get_irn_link(user);
2794 if(!is_Phi(user)) continue;
2795 if(!has_reg_class(si, user)) continue;
2796 if(pset_find_ptr(visited_users, user)) continue;
2797 pset_insert_ptr(visited_users, user);
2799 copy = op->attr.live_range.args.copies[pos];
2801 if(user == target) {
2802 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2803 del_pset(visited_users);
2804 pset_remove_ptr(visited, irn);
2807 pset_insert(copies, INT_TO_PTR(copy), copy);
2808 write_copy_path_cst(si, copies, any_interfere);
2809 pset_remove(copies, INT_TO_PTR(copy), copy);
2810 } else if(!pset_find_ptr(visited, user)) {
2811 pset_insert(copies, INT_TO_PTR(copy), copy);
2812 paths += find_copy_path(si, user, target, any_interfere, copies, visited);
2813 pset_remove(copies, INT_TO_PTR(copy), copy);
2815 /*if(paths > MAX_PATHS) {
2816 if(pset_count(copies) == 0) {
2820 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2821 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2822 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2823 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2824 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2828 del_pset(visited_users);
2829 pset_remove_ptr(visited, irn);
2832 } else if(pset_count(copies) == 0) {
2838 del_pset(visited_users);
2839 pset_remove_ptr(visited, irn);
2844 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
2846 pset * copies = pset_new_ptr_default();
2847 pset * visited = pset_new_ptr_default();
2849 find_copy_path(si, a, b, any_interfere, copies, visited);
2857 memcopyhandler(spill_ilp_t * si)
2859 interference_t *interference;
2861 /* teste Speicherwerte auf Interferenz */
2863 /* analyze phi classes */
2864 phi_class_compute(si->chordal_env->irg);
2866 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
2867 irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
2869 /* now lets emit the ILP unequations for the crap */
2870 set_foreach(si->interferences, interference) {
2872 ilp_var_t interfere,
2874 ilp_cst_t any_interfere_cst,
2876 const ir_node *a = interference->a;
2877 const ir_node *b = interference->b;
2879 /* any_interf <= \sum interf */
2880 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
2881 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2882 any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2884 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
2886 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
2887 const ir_node *bb = irnlist->irn;
2888 spill_bb_t *spill_bb = get_irn_link(bb);
2893 spilla = set_find_spill(spill_bb->ilp, a);
2896 spillb = set_find_spill(spill_bb->ilp, b);
2899 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
2900 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
2901 /* 2: - mem_in_a - spill_a + interfere <= 0 */
2902 /* 3: - mem_in_b - spill_b + interfere <= 0 */
2903 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
2904 interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2906 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
2907 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2909 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
2910 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
2911 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
2912 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
2913 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
2915 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
2916 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2918 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2919 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
2920 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
2922 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
2923 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2925 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2926 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
2927 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
2930 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
2932 /* any_interfere >= interf */
2933 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
2934 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2936 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2937 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2940 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
2941 gen_copy_constraints(si,a,b,any_interfere);
2949 return fabs(x) < 0.00001;
2952 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
2954 spill_ilp_t *si = get_irg_link(current_ir_graph);
2956 if(pset_find_ptr(si->all_possible_remats, n)) {
2957 op_t *op = (op_t*)get_irn_link(n);
2958 assert(op && op->is_remat);
2960 if(!op->attr.remat.remat->inverse) {
2961 if(op->attr.remat.pre) {
2962 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
2964 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
2969 op_t *op = (op_t*)get_irn_link(n);
2970 assert(op && op->is_remat);
2972 if(op->attr.remat.pre) {
2973 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
2975 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
2986 dump_graph_with_remats(ir_graph * irg, const char * suffix)
2988 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
2989 be_dump(irg, suffix, dump_ir_block_graph_sched);
2990 set_dump_node_vcgattr_hook(NULL);
2994 * Edge hook to dump the schedule edges with annotated register pressure.
2997 sched_pressure_edge_hook(FILE *F, ir_node *irn)
2999 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
3000 ir_node *prev = sched_prev(irn);
3001 fprintf(F, "edge:{sourcename:\"");
3003 fprintf(F, "\" targetname:\"");
3005 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
3006 fprintf(F, "\" color:magenta}\n");
3012 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
3014 DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3016 dump_consts_local(0);
3017 set_dump_node_edge_hook(sched_pressure_edge_hook);
3018 dump_ir_block_graph(irg, suffix);
3019 set_dump_node_edge_hook(old_edge_hook);
3023 walker_pressure_annotator(ir_node * bb, void * data)
3025 spill_ilp_t *si = data;
3028 pset *live = pset_new_ptr_default();
3031 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3032 irn = be_lv_get_irn(si->lv, bb, i);
3034 if (has_reg_class(si, irn)) {
3035 pset_insert_ptr(live, irn);
3039 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3041 sched_foreach_reverse(bb, irn) {
3043 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3047 if(has_reg_class(si, irn)) {
3048 pset_remove_ptr(live, irn);
3049 if(is_Proj(irn)) ++projs;
3052 if(!is_Proj(irn)) projs = 0;
3054 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3055 ir_node *arg = get_irn_n(irn, n);
3057 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3059 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3066 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3068 be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
3072 connect_all_remats_with_keep(spill_ilp_t * si)
3080 n_remats = pset_count(si->all_possible_remats);
3082 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3085 pset_foreach(si->all_possible_remats, irn) {
3090 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
3092 obstack_free(si->obst, ins);
3098 connect_all_spills_with_keep(spill_ilp_t * si)
3107 n_spills = pset_count(si->spills);
3109 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3112 pset_foreach(si->spills, irn) {
3117 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
3119 obstack_free(si->obst, ins);
3123 /** insert a spill at an arbitrary position */
3124 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3126 ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
3127 ir_graph *irg = get_irn_irg(bl);
3128 ir_node *frame = get_irg_frame(irg);
3132 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
3133 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3135 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3138 * search the right insertion point. a spill of a phi cannot be put
3139 * directly after the phi, if there are some phis behind the one which
3140 * is spilled. Also, a spill of a Proj must be after all Projs of the
3143 * Here's one special case:
3144 * If the spill is in the start block, the spill must be after the frame
3145 * pointer is set up. This is done by setting insert to the end of the block
3146 * which is its default initialization (see above).
3149 if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3152 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3155 sched_add_after(insert, spill);
3160 delete_remat(spill_ilp_t * si, ir_node * remat) {
3162 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3164 sched_remove(remat);
3166 /* kill links to operands */
3167 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3168 set_irn_n(remat, n, bad);
3173 clean_remat_info(spill_ilp_t * si)
3177 remat_info_t *remat_info;
3178 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3180 set_foreach(si->remat_info, remat_info) {
3181 if(!remat_info->remats) continue;
3183 pset_foreach(remat_info->remats, remat)
3185 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3186 set_irn_n((ir_node*)remat->proj, -1, bad);
3187 set_irn_n((ir_node*)remat->proj, 0, bad);
3190 if(get_irn_n_edges(remat->op) == 0) {
3191 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3192 set_irn_n((ir_node*)remat->op, n, bad);
3197 if(remat_info->remats) del_pset(remat_info->remats);
3198 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3203 delete_unnecessary_remats(spill_ilp_t * si)
3205 if(opt_keep_alive & KEEPALIVE_REMATS) {
3207 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3210 ir_node *end = get_irg_end(si->chordal_env->irg);
3213 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3214 ir_node *keep_arg = get_irn_n(si->keep, n);
3215 op_t *arg_op = get_irn_link(keep_arg);
3218 assert(arg_op->is_remat);
3220 name = si->lpp->vars[arg_op->attr.remat.ilp];
3222 if(is_zero(name->value)) {
3223 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
3224 /* TODO check whether reload is preferred over remat (could be bug) */
3225 delete_remat(si, keep_arg);
3227 if(!arg_op->attr.remat.remat->inverse) {
3228 if(arg_op->attr.remat.pre) {
3229 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3231 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3234 if(arg_op->attr.remat.pre) {
3235 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3237 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3242 set_irn_n(si->keep, n, bad);
3245 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
3246 ir_node *end_arg = get_End_keepalive(end, i);
3248 if(end_arg != si->keep) {
3249 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
3252 keeps = obstack_finish(si->obst);
3253 set_End_keepalives(end, n-1, keeps);
3254 obstack_free(si->obst, keeps);
3257 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
3262 pset_foreach(si->all_possible_remats, remat) {
3263 op_t *remat_op = get_irn_link(remat);
3264 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
3266 if(is_zero(name->value)) {
3267 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
3268 /* TODO check whether reload is preferred over remat (could be bug) */
3269 delete_remat(si, remat);
3271 if(!remat_op->attr.remat.remat->inverse) {
3272 if(remat_op->attr.remat.pre) {
3273 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3275 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3278 if(remat_op->attr.remat.pre) {
3279 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3281 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3290 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3292 pset *spills = pset_new_ptr_default();
3294 const ir_node *next;
3297 defs = set_find_def(si->values, value);
3299 if(defs && defs->spills) {
3300 for(next = defs->spills; next; next = get_irn_link(next)) {
3301 pset_insert_ptr(spills, next);
3309 * @param before The node after which the spill will be placed in the schedule
3311 /* TODO set context properly */
3313 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3317 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3319 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3321 spill = be_spill2(arch_env, irn, before);
3323 defs = set_insert_def(si->values, value);
3326 /* enter into the linked list */
3327 set_irn_link(spill, defs->spills);
3328 defs->spills = spill;
3330 if(opt_keep_alive & KEEPALIVE_SPILLS)
3331 pset_insert_ptr(si->spills, spill);
3337 * @param before The Phi node which has to be spilled
3340 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3347 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3349 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3350 ins[n] = si->m_unknown;
3353 mem_phi = new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3355 defs = set_insert_def(si->values, phi);
3358 /* enter into the linked list */
3359 set_irn_link(mem_phi, defs->spills);
3360 defs->spills = mem_phi;
3362 sched_add_after(phi, mem_phi);
3364 if(opt_keep_alive & KEEPALIVE_SPILLS)
3365 pset_insert_ptr(si->spills, mem_phi);
3372 * Add remat to list of defs, destroys link field!
3375 insert_remat(spill_ilp_t * si, ir_node * remat)
3378 op_t *remat_op = get_irn_link(remat);
3380 assert(remat_op->is_remat);
3382 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3385 /* enter into the linked list */
3386 set_irn_link(remat, defs->remats);
3387 defs->remats = remat;
3392 * Add reload before operation and add to list of defs
3395 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3400 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3402 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3404 defs = set_find_def(si->values, value);
3406 spill = defs->spills;
3407 assert(spill && "no spill placed before reload");
3409 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3411 /* enter into the linked list */
3412 set_irn_link(reload, defs->remats);
3413 defs->remats = reload;
3418 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3422 ir_node *value = get_irn_n(memoperand->irn, memoperand->pos);
3424 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3426 DBG((si->dbg, LEVEL_2, "\t inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3428 defs = set_find_def(si->values, value);
3430 spill = defs->spills;
3431 assert(spill && "no spill placed before reload");
3433 reload = be_reload(arch_env, si->cls, memoperand->irn, get_irn_mode(value), spill);
3435 arch_perform_memory_operand(arch_env, memoperand->irn, reload, memoperand->pos);
3436 sched_remove(reload);
3439 void insert_memoperands(spill_ilp_t * si)
3441 memoperand_t *memoperand;
3444 set_foreach(si->memoperands, memoperand) {
3445 name = si->lpp->vars[memoperand->ilp];
3446 if(!is_zero(name->value)) {
3447 perform_memory_operand(si, memoperand);
3453 walker_spill_placer(ir_node * bb, void * data) {
3454 spill_ilp_t *si = (spill_ilp_t*)data;
3456 spill_bb_t *spill_bb = get_irn_link(bb);
3457 pset *spills_to_do = pset_new_ptr_default();
3460 set_foreach(spill_bb->ilp, spill) {
3463 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3464 name = si->lpp->vars[spill->mem_in];
3465 if(!is_zero(name->value)) {
3468 mem_phi = insert_mem_phi(si, spill->irn);
3470 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3474 name = si->lpp->vars[spill->spill];
3475 if(!is_zero(name->value)) {
3476 /* place spill directly after definition */
3477 if(get_nodes_block(spill->irn) == bb) {
3478 insert_spill(si, spill->irn, spill->irn, spill->irn);
3482 /* place spill at bb start */
3483 if(spill->reg_in > 0) {
3484 name = si->lpp->vars[spill->reg_in];
3485 if(!is_zero(name->value)) {
3486 insert_spill(si, spill->irn, spill->irn, bb);
3490 /* place spill after a remat */
3491 pset_insert_ptr(spills_to_do, spill->irn);
3494 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3497 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3498 op_t *op = get_irn_link(irn);
3500 if(be_is_Spill(irn)) continue;
3503 /* TODO fix this if we want to support remats with more than two nodes */
3504 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3505 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3507 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3510 if(pset_find_ptr(spills_to_do, irn)) {
3511 pset_remove_ptr(spills_to_do, irn);
3513 insert_spill(si, irn, irn, irn);
3519 assert(pset_count(spills_to_do) == 0);
3521 /* afterwards free data in block */
3522 del_pset(spills_to_do);
3526 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3528 ir_node *insert_pos = bb;
3530 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3532 /* find last definition of arg value in block */
3537 defs = set_find_def(si->values, value);
3539 if(defs && defs->remats) {
3540 for(next = defs->remats; next; next = get_irn_link(next)) {
3541 if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3542 last = sched_get_time_step(next);
3548 if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3549 last = sched_get_time_step(value);
3553 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
3555 spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3561 phim_fixer(spill_ilp_t *si) {
3564 set_foreach(si->values, defs) {
3565 const ir_node *phi = defs->value;
3566 op_t *op = get_irn_link(phi);
3567 ir_node *phi_m = NULL;
3568 ir_node *next = defs->spills;
3571 if(!is_Phi(phi)) continue;
3574 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3578 next = get_irn_link(next);
3581 if(!phi_m) continue;
3583 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3584 ir_node *value = get_irn_n(phi, n);
3585 defs_t *val_defs = set_find_def(si->values, value);
3587 /* a spill of this value */
3592 ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3593 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3595 if(!is_zero(name->value)) {
3596 spill = insert_mem_copy(si, pred, value);
3598 spill = val_defs->spills;
3601 spill = val_defs->spills;
3604 assert(spill && "no spill placed before PhiM");
3605 set_irn_n(phi_m, n, spill);
3611 walker_reload_placer(ir_node * bb, void * data) {
3612 spill_ilp_t *si = (spill_ilp_t*)data;
3614 spill_bb_t *spill_bb = get_irn_link(bb);
3616 /* reloads at end of block */
3617 if(spill_bb->reloads) {
3620 set_foreach(spill_bb->reloads, keyval) {
3621 ir_node *irn = (ir_node*)keyval->key;
3622 ilp_var_t reload = PTR_TO_INT(keyval->val);
3625 name = si->lpp->vars[reload];
3626 if(!is_zero(name->value)) {
3628 ir_node *insert_pos = bb;
3629 ir_node *prev = sched_block_last_noncf(si, bb);
3630 op_t *prev_op = get_irn_link(prev);
3632 while(be_is_Spill(prev)) {
3633 prev = sched_prev(prev);
3636 prev_op = get_irn_link(prev);
3638 /* insert reload before pre-remats */
3639 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3640 && prev_op->is_remat && prev_op->attr.remat.pre) {
3644 prev = sched_prev(prev);
3645 } while(be_is_Spill(prev));
3647 prev_op = get_irn_link(prev);
3651 reload = insert_reload(si, irn, insert_pos);
3653 if(opt_keep_alive & KEEPALIVE_RELOADS)
3654 pset_insert_ptr(si->spills, reload);
3659 /* walk and insert more reloads and collect remats */
3660 sched_foreach_reverse(bb, irn) {
3661 op_t *op = get_irn_link(irn);
3663 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3664 if(is_Phi(irn)) break;
3667 if(get_irn_mode(irn) != mode_T) {
3668 insert_remat(si, irn);
3673 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3674 ir_node *arg = get_irn_n(irn, n);
3676 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3679 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3680 if(!is_zero(name->value)) {
3682 ir_node *insert_pos = irn;
3683 ir_node *prev = sched_prev(insert_pos);
3686 while(be_is_Spill(prev)) {
3687 prev = sched_prev(prev);
3690 prev_op = get_irn_link(prev);
3692 /* insert reload before pre-remats */
3693 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3694 && prev_op->is_remat && prev_op->attr.remat.pre) {
3698 prev = sched_prev(prev);
3699 } while(be_is_Spill(prev));
3701 prev_op = get_irn_link(prev);
3705 reload = insert_reload(si, arg, insert_pos);
3707 set_irn_n(irn, n, reload);
3709 if(opt_keep_alive & KEEPALIVE_RELOADS)
3710 pset_insert_ptr(si->spills, reload);
3717 del_set(spill_bb->ilp);
3718 if(spill_bb->reloads) del_set(spill_bb->reloads);
3722 walker_collect_used(ir_node * irn, void * data)
3724 lc_bitset_t *used = data;
3726 lc_bitset_set(used, get_irn_idx(irn));
3729 struct kill_helper {
3735 walker_kill_unused(ir_node * bb, void * data)
3737 struct kill_helper *kh = data;
3738 ir_node *bad = get_irg_bad(get_irn_irg(bb));
3742 for(irn=sched_first(bb); !sched_is_end(irn);) {
3743 ir_node *next = sched_next(irn);
3746 if(!lc_bitset_is_set(kh->used, get_irn_idx(irn))) {
3747 if(be_is_Spill(irn) || be_is_Reload(irn)) {
3748 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)));
3750 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3756 set_nodes_block(irn, bad);
3757 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3758 set_irn_n(irn, n, bad);
3766 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3768 struct kill_helper kh;
3770 kh.used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3773 irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3774 irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3776 lc_bitset_free(kh.used);
3780 print_irn_pset(pset * p)
3784 pset_foreach(p, irn) {
3785 ir_printf("%+F\n", irn);
3790 dump_phi_class(spill_ilp_t * si, pset * phiclass, const char * file)
3792 FILE *f = fopen(file, "w");
3794 interference_t *interference;
3796 pset_break(phiclass);
3797 set_break(si->interferences);
3799 ir_fprintf(f, "digraph phiclass {\n");
3801 pset_foreach(phiclass, irn) {
3803 ir_fprintf(f, " %F%N [shape=box]\n",irn,irn);
3806 pset_foreach(phiclass, irn) {
3809 if(!is_Phi(irn)) continue;
3811 for(n=get_irn_arity(irn)-1; n>=0; --n) {
3812 ir_node *arg = get_irn_n(irn, n);
3814 ir_fprintf(f, " %F%N -> %F%N\n",irn,irn,arg,arg);
3818 set_foreach(si->interferences, interference) {
3819 const ir_node *a = interference->a;
3820 const ir_node *b = interference->b;
3821 if(get_phi_class(a) == phiclass) {
3822 ir_fprintf(f, " %F%N -> %F%N [color=red,dir=none,style=bold]\n",a,a,b,b);
3831 rewire_uses(spill_ilp_t * si)
3833 dom_front_info_t *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
3835 pset *ignore = pset_new_ptr(1);
3837 pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
3839 /* then fix uses of spills */
3840 set_foreach(si->values, defs) {
3843 const ir_node *next = defs->remats;
3846 reloads = pset_new_ptr_default();
3849 if(be_is_Reload(next)) {
3850 pset_insert_ptr(reloads, next);
3854 next = get_irn_link(next);
3857 spills = get_spills_for_value(si, defs->value);
3858 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));
3859 if(pset_count(spills) > 1) {
3860 //assert(pset_count(reloads) > 0);
3861 // print_irn_pset(spills);
3862 // print_irn_pset(reloads);
3864 be_ssa_constr_set_ignore(dfi, si->lv, spills, ignore);
3871 /* first fix uses of remats and reloads */
3872 set_foreach(si->values, defs) {
3874 const ir_node *next = defs->remats;
3877 nodes = pset_new_ptr_default();
3878 pset_insert_ptr(nodes, defs->value);
3881 pset_insert_ptr(nodes, next);
3882 next = get_irn_link(next);
3885 if(pset_count(nodes) > 1) {
3886 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
3887 be_ssa_constr_set(dfi, si->lv, nodes);
3894 // remove_unused_defs(si);
3896 be_free_dominance_frontiers(dfi);
3901 writeback_results(spill_ilp_t * si)
3903 /* walk through the graph and collect all spills, reloads and remats for a value */
3905 si->values = new_set(cmp_defs, 4096);
3907 DBG((si->dbg, LEVEL_1, "Applying results\n"));
3908 delete_unnecessary_remats(si);
3909 si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
3910 irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
3911 irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
3913 insert_memoperands(si);
3916 /* clean the remat info! there are still back-edges leading there! */
3917 clean_remat_info(si);
3921 connect_all_spills_with_keep(si);
3923 del_set(si->values);
3927 get_n_regs(spill_ilp_t * si)
3929 int arch_n_regs = arch_register_class_n_regs(si->cls);
3933 for(i=0; i<arch_n_regs; i++) {
3934 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
3939 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
3944 walker_reload_mover(ir_node * bb, void * data)
3946 spill_ilp_t *si = data;
3949 sched_foreach(bb, tmp) {
3950 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
3951 ir_node *reload = tmp;
3954 /* move reload upwards */
3956 int pressure = (int)get_irn_link(reload);
3957 if(pressure < si->n_regs) {
3958 irn = sched_prev(reload);
3959 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
3960 sched_remove(reload);
3961 pressure = (int)get_irn_link(irn);
3963 while(pressure < si->n_regs) {
3964 if( sched_is_end(irn) ||
3965 (be_is_Reload(irn) && has_reg_class(si, irn)) ||
3966 /* do not move reload before its spill */
3967 (irn == be_get_Reload_mem(reload)) ||
3968 /* do not move before phi */
3971 set_irn_link(irn, INT_TO_PTR(pressure+1));
3972 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
3973 irn = sched_prev(irn);
3975 pressure = (int)get_irn_link(irn);
3978 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
3979 sched_put_after(irn, reload);
3986 move_reloads_upward(spill_ilp_t * si)
3988 irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
3993 * Walk all irg blocks and check for interfering spills inside of phi classes
3996 luke_meminterferencechecker(ir_node * bb, void * data)
3998 spill_ilp_t *si = (spill_ilp_t*)data;
4001 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
4002 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
4004 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
4006 /* a is only interesting if it is in my register class and if it is inside a phi class */
4007 if (has_reg_class(si, a) && get_phi_class(a)) {
4008 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)) {
4009 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
4011 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
4013 /* a and b are only interesting if they are in the same phi class */
4014 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
4015 if(values_interfere_in_block(si, bb, a, b)) {
4016 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4025 verify_phiclasses(spill_ilp_t * si)
4027 /* analyze phi classes */
4028 phi_class_compute(si->chordal_env->irg);
4030 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4031 irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
4035 walker_spillslotassigner(ir_node * irn, void * data)
4039 if(!be_is_Spill(irn)) return;
4041 /* set spill context to phi class if it has one ;) */
4044 // Matze: not needed anymore
4045 cls = get_phi_class(irn);
4047 be_set_Spill_context(irn, cls);
4049 be_set_Spill_context(irn, irn);
4055 assign_spillslots(spill_ilp_t * si)
4057 DBG((si->dbg, LEVEL_2, "\t calling spill slot assigner\n"));
4058 irg_walk_graph(si->chordal_env->irg, walker_spillslotassigner, NULL, si);
4062 be_spill_remat(const be_chordal_env_t * chordal_env)
4065 char problem_name[256];
4066 char dump_suffix[256];
4067 char dump_suffix2[256];
4068 struct obstack obst;
4071 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
4072 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
4073 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
4075 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4076 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4078 if(opt_verify & VERIFY_DOMINANCE)
4079 be_check_dominance(chordal_env->irg);
4081 obstack_init(&obst);
4082 si.chordal_env = chordal_env;
4084 si.cls = chordal_env->cls;
4085 si.lpp = new_lpp(problem_name, lpp_minimize);
4086 si.remat_info = new_set(cmp_remat_info, 4096);
4087 si.interferences = new_set(cmp_interference, 32);
4088 si.memoperands = new_set(cmp_memoperands, 128);
4089 si.all_possible_remats = pset_new_ptr_default();
4090 si.spills = pset_new_ptr_default();
4091 si.inverse_ops = pset_new_ptr_default();
4092 si.lv = chordal_env->lv;
4094 si.n_regs = get_n_regs(&si);
4096 set_irg_link(chordal_env->irg, &si);
4097 compute_doms(chordal_env->irg);
4099 /* compute phi classes */
4100 // phi_class_compute(chordal_env->irg);
4102 be_analyze_regpressure(chordal_env, "-pre");
4105 /* collect remats */
4106 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4107 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
4110 /* insert possible remats */
4111 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4112 irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
4113 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4115 if(opt_keep_alive & KEEPALIVE_REMATS) {
4116 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4117 connect_all_remats_with_keep(&si);
4118 /* dump graph with inserted remats */
4119 dump_graph_with_remats(chordal_env->irg, dump_suffix);
4122 /* insert copies for phi arguments not in my regclass */
4123 irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
4125 /* recompute liveness */
4126 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4127 be_liveness_recompute(si.lv);
4131 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4132 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4133 irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
4135 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4136 irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
4139 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4140 memcopyhandler(&si);
4143 if(opt_dump_flags & DUMP_PROBLEM) {
4145 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4146 if ((f = fopen(buf, "wt")) != NULL) {
4147 lpp_dump_plain(si.lpp, f);
4152 if(opt_dump_flags & DUMP_MPS) {
4155 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4156 if((f = fopen(buf, "wt")) != NULL) {
4157 mps_write_mps(si.lpp, s_mps_fixed, f);
4161 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4162 if((f = fopen(buf, "wt")) != NULL) {
4163 mps_write_mst(si.lpp, s_mps_fixed, f);
4168 lpp_check_startvals(si.lpp);
4171 DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4172 lpp_set_time_limit(si.lpp, opt_timeout);
4175 lpp_set_log(si.lpp, stdout);
4178 lpp_solve_cplex(si.lpp);
4180 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4182 assert(lpp_is_sol_valid(si.lpp)
4183 && "solution of ILP must be valid");
4185 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));
4187 if(opt_dump_flags & DUMP_SOLUTION) {
4191 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4192 if ((f = fopen(buf, "wt")) != NULL) {
4194 for (i = 0; i < si.lpp->var_next; ++i) {
4195 lpp_name_t *name = si.lpp->vars[i];
4196 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4202 writeback_results(&si);
4206 kill_all_unused_values_in_schedule(&si);
4208 if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
4209 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
4211 // move reloads upwards
4212 be_liveness_recompute(si.lv);
4213 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4214 move_reloads_upward(&si);
4217 verify_phiclasses(&si);
4218 assign_spillslots(&si);
4221 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4223 dump_pressure_graph(&si, dump_suffix2);
4225 be_analyze_regpressure(chordal_env, "-post");
4227 if(opt_verify & VERIFY_DOMINANCE)
4228 be_check_dominance(chordal_env->irg);
4230 free_dom(chordal_env->irg);
4231 del_set(si.interferences);
4232 del_pset(si.inverse_ops);
4233 del_pset(si.all_possible_remats);
4234 del_set(si.memoperands);
4235 del_pset(si.spills);
4237 obstack_free(&obst, NULL);
4238 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4241 #else /* WITH_ILP */
4244 only_that_you_can_compile_without_WITH_ILP_defined(void)
4248 #endif /* WITH_ILP */