1 /** vim: set sw=4 ts=4:
4 * @author Adam M. Szalkowski & Sebastian Hack
6 * ILP based spilling & rematerialization
8 * Copyright (C) 2006 Universitaet Karlsruhe
9 * Released under the GPL
32 #include "phiclass_t.h"
39 #include <lpp/lpp_net.h>
40 #include <lpp/lpp_cplex.h>
41 //#include <lc_pset.h>
42 //#include <libcore/lc_bitset.h>
46 #include "besched_t.h"
52 #include "bespillremat.h"
54 #include "bepressurestat.h"
55 #include "beprofile.h"
57 #include "bechordal_t.h"
60 #include <libcore/lc_opts.h>
61 #include <libcore/lc_opts_enum.h>
62 #endif /* WITH_LIBCORE */
64 #define DUMP_PROBLEM 1
66 #define DUMP_SOLUTION 4
68 #define DUMP_PRESSURE 16
70 #define KEEPALIVE_REMATS 1
71 #define KEEPALIVE_SPILLS 2
72 #define KEEPALIVE_RELOADS 4
74 #define VERIFY_MEMINTERF 1
75 #define VERIFY_DOMINANCE 2
78 #define REMATS_BRIGGS 1
79 #define REMATS_NOINVERSE 2
82 static int opt_dump_flags = 0;
83 static int opt_log = 0;
84 static int opt_keep_alive = 0;
85 static int opt_goodwin = 1;
86 static int opt_memcopies = 1;
87 static int opt_memoperands = 1;
88 static int opt_verify = VERIFY_MEMINTERF;
89 static int opt_remats = REMATS_ALL;
90 static int opt_repair_schedule = 0;
91 static int opt_no_enlarge_liveness = 0;
92 static int opt_timeout = 300;
93 static double opt_cost_reload = 8.0;
94 static double opt_cost_memoperand = 7.0;
95 static double opt_cost_spill = 30.0;
96 static double opt_cost_remat = 1.0;
100 static const lc_opt_enum_mask_items_t dump_items[] = {
101 { "problem", DUMP_PROBLEM },
103 { "solution", DUMP_SOLUTION },
104 { "stats", DUMP_STATS },
105 { "pressure", DUMP_PRESSURE },
109 static lc_opt_enum_mask_var_t dump_var = {
110 &opt_dump_flags, dump_items
113 static const lc_opt_enum_mask_items_t keepalive_items[] = {
114 { "remats", KEEPALIVE_REMATS },
115 { "spills", KEEPALIVE_SPILLS },
116 { "reloads", KEEPALIVE_RELOADS },
120 static lc_opt_enum_mask_var_t keep_alive_var = {
121 &opt_keep_alive, keepalive_items
124 static const lc_opt_enum_mask_items_t remats_items[] = {
125 { "none", REMATS_NONE },
126 { "briggs", REMATS_BRIGGS },
127 { "noinverse", REMATS_NOINVERSE },
128 { "all", REMATS_ALL },
132 static lc_opt_enum_mask_var_t remats_var = {
133 &opt_remats, remats_items
136 static const lc_opt_table_entry_t options[] = {
137 LC_OPT_ENT_ENUM_MASK("keepalive", "keep alive remats, spills or reloads", &keep_alive_var),
139 LC_OPT_ENT_BOOL ("goodwin", "activate goodwin reduction", &opt_goodwin),
140 LC_OPT_ENT_BOOL ("memcopies", "activate memcopy handling", &opt_memcopies),
141 LC_OPT_ENT_BOOL ("memoperands", "activate memoperands", &opt_memoperands),
142 LC_OPT_ENT_ENUM_INT ("remats", "type of remats to insert (none, briggs, noinverse or all)",&remats_var),
143 LC_OPT_ENT_BOOL ("repair_schedule", "repair the schedule by rematting once used nodes",&opt_repair_schedule),
144 LC_OPT_ENT_BOOL ("no_enlage_liveness", "do not enlarge liveness of operands of remats",&opt_no_enlarge_liveness),
146 LC_OPT_ENT_ENUM_MASK("dump", "dump problem, mps, solution, stats or pressure", &dump_var),
147 LC_OPT_ENT_BOOL ("log", "activate the lpp log", &opt_log),
148 LC_OPT_ENT_INT ("timeout", "ILP solver timeout", &opt_timeout),
150 LC_OPT_ENT_DBL ("cost_reload", "cost of a reload", &opt_cost_reload),
151 LC_OPT_ENT_DBL ("cost_memoperand", "cost of a memory operand", &opt_cost_memoperand),
152 LC_OPT_ENT_DBL ("cost_spill", "cost of a spill instruction", &opt_cost_spill),
153 LC_OPT_ENT_DBL ("cost_remat", "cost of a rematerialization", &opt_cost_remat),
157 void be_spill_remat_register_options(lc_opt_entry_t *grp)
159 lc_opt_entry_t *my_grp = lc_opt_get_grp(grp, "remat");
160 lc_opt_add_table(my_grp, options);
165 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
166 //#define SCHEDULE_PHIM /* insert phim nodes into schedule */
169 //#define SOLVE_LOCAL
170 #define LPP_SERVER "i44pc52"
171 #define LPP_SOLVER "cplex"
174 #define MAX_PATHS INT_MAX
177 typedef struct _spill_ilp_t {
178 const arch_register_class_t *cls;
180 const be_chordal_env_t *chordal_env;
183 struct obstack *obst;
185 pset *all_possible_remats;
188 set *values; /**< for collecting all definitions of values before running ssa-construction */
193 DEBUG_ONLY(firm_dbg_module_t * dbg);
196 typedef int ilp_var_t;
197 typedef int ilp_cst_t;
199 typedef struct _spill_bb_t {
204 typedef struct _remat_t {
205 const ir_node *op; /**< for copy_irn */
206 const ir_node *value; /**< the value which is being recomputed by this remat */
207 const ir_node *proj; /**< not NULL if the above op produces a tuple */
208 int cost; /**< cost of this remat */
209 int inverse; /**< nonzero if this is an inverse remat */
213 * Data to be attached to each IR node. For remats this contains the ilp_var
214 * for this remat and for normal ops this contains the ilp_vars for
215 * reloading each operand
217 typedef struct _op_t {
222 const remat_t *remat; /** the remat this op belongs to */
223 int pre; /** 1, if this is a pressure-increasing remat */
227 ir_node *op; /** the operation this live range belongs to */
236 typedef struct _defs_t {
237 const ir_node *value;
238 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
239 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
242 typedef struct _remat_info_t {
243 const ir_node *irn; /**< the irn to which these remats belong */
244 pset *remats; /**< possible remats for this value */
245 pset *remats_by_operand; /**< remats with this value as operand */
248 typedef struct _keyval_t {
253 typedef struct _spill_t {
262 typedef struct _memoperand_t {
263 ir_node *irn; /**< the irn */
264 unsigned int pos; /**< the position of the argument */
265 ilp_var_t ilp; /**< the ilp var for this memory operand */
269 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
271 return chordal_has_class(si->chordal_env, irn);
276 cmp_remat(const void *a, const void *b)
278 const keyval_t *p = a;
279 const keyval_t *q = b;
280 const remat_t *r = p->val;
281 const remat_t *s = q->val;
285 return !(r == s || r->op == s->op);
289 cmp_remat(const void *a, const void *b)
291 const remat_t *r = a;
292 const remat_t *s = a;
294 return !(r == s || r->op == s->op);
298 cmp_spill(const void *a, const void *b, size_t size)
300 const spill_t *p = a;
301 const spill_t *q = b;
303 // return !(p->irn == q->irn && p->bb == q->bb);
304 return !(p->irn == q->irn);
308 cmp_memoperands(const void *a, const void *b, size_t size)
310 const memoperand_t *p = a;
311 const memoperand_t *q = b;
313 return !(p->irn == q->irn && p->pos == q->pos);
317 set_find_keyval(set * set, const void * key)
322 return set_find(set, &query, sizeof(query), HASH_PTR(key));
326 set_insert_keyval(set * set, void * key, void * val)
332 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
336 set_find_def(set * set, const ir_node * value)
341 return set_find(set, &query, sizeof(query), HASH_PTR(value));
345 set_insert_def(set * set, const ir_node * value)
352 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
355 static memoperand_t *
356 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
363 return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
366 static memoperand_t *
367 set_find_memoperand(set * set, const ir_node * irn, unsigned int pos)
371 query.irn = (ir_node*)irn;
373 return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
378 set_find_spill(set * set, const ir_node * value)
382 query.irn = (ir_node*)value;
383 return set_find(set, &query, sizeof(query), HASH_PTR(value));
386 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
387 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
388 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
389 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
390 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
393 cmp_remat_info(const void *a, const void *b, size_t size)
395 const remat_info_t *p = a;
396 const remat_info_t *q = b;
398 return !(p->irn == q->irn);
402 cmp_defs(const void *a, const void *b, size_t size)
407 return !(p->value == q->value);
411 cmp_keyval(const void *a, const void *b, size_t size)
413 const keyval_t *p = a;
414 const keyval_t *q = b;
416 return !(p->key == q->key);
420 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
423 if(be_profile_has_data())
424 return ((double)be_profile_get_block_execcount(get_block(irn))) + FUDGE;
426 #ifndef EXECFREQ_LOOPDEPH
427 return get_block_execfreq(si->chordal_env->exec_freq, get_block(irn)) + FUDGE;
430 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
432 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
437 get_cost(const spill_ilp_t * si, const ir_node * irn)
439 if(be_is_Spill(irn)) {
440 return opt_cost_spill;
441 } else if(be_is_Reload(irn)){
442 return opt_cost_reload;
444 return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
449 * Checks, whether node and its operands have suitable reg classes
452 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
455 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
456 int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
460 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
463 for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
464 ir_node *op = get_irn_n(irn, n);
465 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
468 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
475 * Try to create a remat from @p op with destination value @p dest_value
477 static INLINE remat_t *
478 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
480 remat_t *remat = NULL;
482 // if(!mode_is_datab(get_irn_mode(dest_value)))
485 if(dest_value == op) {
486 const ir_node *proj = NULL;
488 if(is_Proj(dest_value)) {
489 op = get_Proj_pred(op);
493 if(!is_rematerializable(si, op))
496 remat = obstack_alloc(si->obst, sizeof(*remat));
498 remat->cost = get_cost(si, op);
499 remat->value = dest_value;
503 arch_inverse_t inverse;
506 /* get the index of the operand we want to retrieve by the inverse op */
507 for (n = get_irn_arity(op)-1; n>=0; --n) {
508 ir_node *arg = get_irn_n(op, n);
510 if(arg == dest_value) break;
514 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", n, op));
516 /* else ask the backend to give an inverse op */
517 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
520 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
522 assert(inverse.n > 0 && "inverse op should have at least one node");
524 for(i=inverse.n-1; i>=0; --i) {
525 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
529 remat = obstack_alloc(si->obst, sizeof(*remat));
530 remat->op = inverse.nodes[0];
531 remat->cost = inverse.costs;
532 remat->value = dest_value;
533 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
536 assert(is_Proj(remat->proj));
538 assert(0 && "I can not handle remats with more than 2 nodes");
545 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
547 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
555 add_remat(const spill_ilp_t * si, const remat_t * remat)
557 remat_info_t *remat_info,
562 assert(remat->value);
564 query.irn = remat->value;
566 query.remats_by_operand = NULL;
567 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
569 if(remat_info->remats == NULL) {
570 remat_info->remats = new_pset(cmp_remat, 4096);
572 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
574 /* insert the remat into the remats_be_operand set of each argument of the remat op */
575 for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
576 ir_node *arg = get_irn_n(remat->op, n);
580 query.remats_by_operand = NULL;
581 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
583 if(remat_info->remats_by_operand == NULL) {
584 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
586 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
591 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
593 const ir_edge_t *edge = get_irn_out_edge_first(irn);
597 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
600 edge = get_irn_out_edge_next(irn, edge);
607 get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
613 irn = get_Proj_pred(irn);
615 for(n=get_irn_arity(irn)-1; n>=0; --n) {
616 const ir_node *arg = get_irn_n(irn, n);
618 if(has_reg_class(si, arg)) ++ret;
625 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
630 if( has_reg_class(si, op)
631 && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1)
632 && (opt_remats != REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0)
634 remat = get_remat_from_op(si, op, op);
636 add_remat(si, remat);
640 if(opt_remats == REMATS_ALL) {
641 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
643 for (n = get_irn_arity(op)-1; n>=0; --n) {
644 ir_node *arg = get_irn_n(op, n);
646 if(has_reg_class(si, arg)) {
647 /* try to get an inverse remat */
648 remat = get_remat_from_op(si, arg, op);
650 add_remat(si, remat);
658 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
661 ir_node *def_block = get_nodes_block(val);
667 /* if pos is at end of a basic block */
669 ret = (pos == def_block || block_dominates(def_block, pos));
670 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
674 /* else if this is a normal operation */
675 block = get_nodes_block(pos);
676 if(block == def_block) {
677 if(!sched_is_scheduled(val)) return 1;
679 ret = sched_comes_after(val, pos);
680 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
684 ret = block_dominates(def_block, block);
685 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
689 static INLINE ir_node *
690 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
692 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
696 * Returns first non-Phi node of block @p bb
698 static INLINE ir_node *
699 sched_block_first_nonphi(const ir_node * bb)
701 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
705 sched_skip_proj_predicator(const ir_node * irn, void * data)
707 return (is_Proj(irn));
710 static INLINE ir_node *
711 sched_next_nonproj(const ir_node * irn, int forward)
713 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
717 * Returns next operation node (non-Proj) after @p irn
718 * or the basic block of this node
720 static INLINE ir_node *
721 sched_next_op(const ir_node * irn)
723 ir_node *next = sched_next(irn);
728 return sched_next_nonproj(next, 1);
732 * Returns previous operation node (non-Proj) before @p irn
733 * or the basic block of this node
735 static INLINE ir_node *
736 sched_prev_op(const ir_node * irn)
738 ir_node *prev = sched_prev(irn);
743 return sched_next_nonproj(prev, 0);
747 sched_put_after(ir_node * insert, ir_node * irn)
749 if(is_Block(insert)) {
750 insert = sched_block_first_nonphi(insert);
752 insert = sched_next_op(insert);
754 sched_add_before(insert, irn);
758 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
760 if(is_Block(insert)) {
761 insert = sched_block_last_noncf(si, insert);
763 insert = sched_next_nonproj(insert, 0);
764 insert = sched_prev(insert);
766 sched_add_after(insert, irn);
770 * Tells you whether a @p remat can be placed before the irn @p pos
773 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
775 const ir_node *op = remat->op;
781 prev = sched_block_last_noncf(si, pos);
782 prev = sched_next_nonproj(prev, 0);
784 prev = sched_prev_op(pos);
786 /* do not remat if the rematted value is defined immediately before this op */
787 if(prev == remat->op) {
792 /* this should be just fine, the following OP will be using this value, right? */
794 /* only remat AFTER the real definition of a value (?) */
795 if(!value_is_defined_before(si, pos, remat->value)) {
796 // ir_fprintf(stderr, "error(not defined)");
801 for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
802 const ir_node *arg = get_irn_n(op, n);
804 if(opt_no_enlarge_liveness) {
805 if(has_reg_class(si, arg) && live) {
806 res &= pset_find_ptr((pset*)live, arg)?1:0;
808 res &= value_is_defined_before(si, pos, arg);
811 res &= value_is_defined_before(si, pos, arg);
819 * Tells you whether a @p remat can be placed after the irn @p pos
822 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
825 pos = sched_block_first_nonphi(pos);
827 pos = sched_next_op(pos);
830 /* only remat AFTER the real definition of a value (?) */
831 if(!value_is_defined_before(si, pos, remat->value)) {
835 return can_remat_before(si, remat, pos, live);
839 * Collect potetially rematerializable OPs
842 walker_remat_collector(ir_node * irn, void * data)
844 spill_ilp_t *si = data;
846 if(!is_Block(irn) && !is_Phi(irn)) {
847 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
848 get_remats_from_op(si, irn);
853 * Inserts a copy of @p irn before @p pos
856 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
861 bb = is_Block(pos)?pos:get_nodes_block(pos);
862 copy = exact_copy(irn);
864 _set_phi_class(copy, NULL);
865 set_nodes_block(copy, bb);
866 sched_put_before(si, pos, copy);
872 * Inserts a copy of @p irn after @p pos
875 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
880 bb = is_Block(pos)?pos:get_nodes_block(pos);
881 copy = exact_copy(irn);
883 _set_phi_class(copy, NULL);
884 set_nodes_block(copy, bb);
885 sched_put_after(pos, copy);
891 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
895 if(can_remat_after(si, remat, pos, live)) {
900 DBG((si->dbg, LEVEL_3, "\t >inserting remat2 %+F\n", remat->op));
902 copy = insert_copy_after(si, remat->op, pos);
904 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
905 op = obstack_alloc(si->obst, sizeof(*op));
907 op->attr.remat.remat = remat;
908 op->attr.remat.pre = 0;
909 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
911 set_irn_link(copy, op);
912 pset_insert_ptr(si->all_possible_remats, copy);
914 proj_copy = insert_copy_after(si, remat->proj, copy);
915 set_irn_n(proj_copy, 0, copy);
916 set_irn_link(proj_copy, op);
917 pset_insert_ptr(si->all_possible_remats, proj_copy);
929 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
933 if(can_remat_before(si, remat, pos, live)) {
938 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
940 copy = insert_copy_before(si, remat->op, pos);
942 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
943 op = obstack_alloc(si->obst, sizeof(*op));
945 op->attr.remat.remat = remat;
946 op->attr.remat.pre = 1;
947 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
949 set_irn_link(copy, op);
950 pset_insert_ptr(si->all_possible_remats, copy);
952 proj_copy = insert_copy_after(si, remat->proj, copy);
953 set_irn_n(proj_copy, 0, copy);
954 set_irn_link(proj_copy, op);
955 pset_insert_ptr(si->all_possible_remats, proj_copy);
967 get_block_n_succs(const ir_node *block) {
968 const ir_edge_t *edge;
970 assert(edges_activated(current_ir_graph));
972 edge = get_block_succ_first(block);
976 edge = get_block_succ_next(block, edge);
981 is_start_block(const ir_node * bb)
983 return get_irg_start_block(get_irn_irg(bb)) == bb;
987 is_merge_edge(const ir_node * bb)
989 if(is_start_block(bb))
993 return get_block_n_succs(bb) == 1;
999 is_diverge_edge(const ir_node * bb)
1001 if(is_start_block(bb))
1005 return get_Block_n_cfgpreds(bb) == 1;
1011 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1016 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1017 irn = be_lv_get_irn(si->lv, bb, i);
1019 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1020 pset_insert_ptr(live, irn);
1024 irn = sched_last(bb);
1026 /* all values eaten by control flow operations are also live until the end of the block */
1027 sched_foreach_reverse(bb, irn) {
1030 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1032 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1033 ir_node *arg = get_irn_n(irn,i);
1035 if(has_reg_class(si, arg)) {
1036 pset_insert_ptr(live, arg);
1043 walker_regclass_copy_insertor(ir_node * irn, void * data)
1045 spill_ilp_t *si = data;
1047 if(is_Phi(irn) && has_reg_class(si, irn)) {
1050 for(n=get_irn_arity(irn)-1; n>=0; --n) {
1051 ir_node *phi_arg = get_irn_n(irn, n);
1052 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
1054 if(!has_reg_class(si, phi_arg)) {
1055 ir_node *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
1056 ir_node *pos = sched_block_last_noncf(si, bb);
1057 op_t *op = obstack_alloc(si->obst, sizeof(*op));
1059 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
1060 sched_add_after(pos, copy);
1061 set_irn_n(irn, n, copy);
1064 op->attr.live_range.args.reloads = NULL;
1065 op->attr.live_range.ilp = ILP_UNDEF;
1066 set_irn_link(copy, op);
1073 * Insert (so far unused) remats into the irg to
1074 * recompute the potential liveness of all values
1077 walker_remat_insertor(ir_node * bb, void * data)
1079 spill_ilp_t *si = data;
1086 /* skip start block, no remats to do there */
1087 if(is_start_block(bb)) return;
1089 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1091 live = pset_new_ptr_default();
1092 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1093 ir_node *value = be_lv_get_irn(si->lv, bb, i);
1095 /* add remats at end of block */
1096 if (has_reg_class(si, value)) {
1097 pset_insert_ptr(live, value);
1101 irn = sched_last(bb);
1102 while(!sched_is_end(irn)) {
1108 next = sched_prev(irn);
1110 /* delete defined value from live set */
1111 if(has_reg_class(si, irn)) {
1112 pset_remove_ptr(live, irn);
1115 if(is_Phi(irn) || is_Proj(irn)) {
1120 args = pset_new_ptr_default();
1121 used = pset_new_ptr_default();
1123 /* collect arguments of op and set args of op already live in epilog */
1124 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1125 ir_node *arg = get_irn_n(irn, n);
1127 pset_insert_ptr(args, arg);
1128 if(has_reg_class(si, arg)) {
1129 pset_insert_ptr(live, arg);
1130 pset_insert_ptr(used, arg);
1134 /* insert all possible remats before irn */
1135 pset_foreach(args, arg) {
1136 remat_info_t *remat_info,
1139 /* continue if the operand has the wrong reg class */
1140 if(!has_reg_class(si, arg))
1144 query.remats = NULL;
1145 query.remats_by_operand = NULL;
1146 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1152 if(remat_info->remats) {
1153 pset_foreach(remat_info->remats, remat) {
1154 ir_node *remat_irn = NULL;
1156 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
1157 remat_irn = insert_remat_before(si, remat, irn, live);
1160 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1161 ir_node *remat_arg = get_irn_n(remat_irn, n);
1163 /* collect args of remats which are not args of op */
1164 if(has_reg_class(si, remat_arg) && !pset_find_ptr(args, remat_arg)) {
1165 pset_insert_ptr(used, remat_arg);
1173 /* do not place post remats after jumps */
1174 if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) {
1180 /* insert all possible remats after irn */
1181 post_remats = pset_new_ptr_default();
1182 pset_foreach(used, arg) {
1183 remat_info_t *remat_info,
1186 /* continue if the operand has the wrong reg class */
1187 if(!has_reg_class(si, arg))
1191 query.remats = NULL;
1192 query.remats_by_operand = NULL;
1193 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1199 if(remat_info->remats_by_operand) {
1200 pset_foreach(remat_info->remats_by_operand, remat) {
1201 /* do not insert remats producing the same value as one of the operands */
1202 if(!pset_find_ptr(args, remat->value)) {
1203 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
1205 /* only remat values that can be used be real ops */
1206 if(pset_find_ptr(live, remat->value)) {
1207 pset_insert_ptr(post_remats, remat);
1213 pset_foreach(post_remats, remat) {
1214 insert_remat_after(si, remat, irn, live);
1216 del_pset(post_remats);
1223 /* add remats at end if successor has multiple predecessors */
1224 if(is_merge_edge(bb)) {
1225 pset *live_out = pset_new_ptr_default();
1228 get_live_end(si, bb, live_out);
1230 /* add remats at end of block */
1231 pset_foreach(live_out, value) {
1232 remat_info_t *remat_info,
1236 query.remats = NULL;
1237 query.remats_by_operand = NULL;
1238 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1240 if(remat_info && remat_info->remats) {
1241 pset_foreach(remat_info->remats, remat) {
1242 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1244 insert_remat_before(si, remat, bb, live_out);
1251 if(is_diverge_edge(bb)) {
1252 pset *live_in = pset_new_ptr_default();
1255 be_lv_foreach(si->lv, bb, be_lv_state_in, i) {
1256 value = be_lv_get_irn(si->lv, bb, i);
1258 if(has_reg_class(si, value)) {
1259 pset_insert_ptr(live_in, value);
1262 /* add phis to live_in */
1263 sched_foreach(bb, value) {
1264 if(!is_Phi(value)) break;
1266 if(has_reg_class(si, value)) {
1267 pset_insert_ptr(live_in, value);
1271 /* add remat2s at beginning of block */
1272 post_remats = pset_new_ptr_default();
1273 pset_foreach(live_in, value) {
1274 remat_info_t *remat_info,
1278 query.remats = NULL;
1279 query.remats_by_operand = NULL;
1280 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1282 if(remat_info && remat_info->remats_by_operand) {
1283 pset_foreach(remat_info->remats_by_operand, remat) {
1284 DBG((si->dbg, LEVEL_4, "\t considering remat2 %+F at beginning of block %+F\n", remat->op, bb));
1286 /* put the remat here if all its args are available and result is still live */
1287 if(pset_find_ptr(live_in, remat->value)) {
1288 pset_insert_ptr(post_remats, remat);
1293 pset_foreach(post_remats, remat) {
1294 insert_remat_after(si, remat, bb, live_in);
1296 del_pset(post_remats);
1302 can_be_copied(const ir_node * bb, const ir_node * irn)
1304 const ir_edge_t *edge = get_block_succ_first(bb);
1305 const ir_node *next_bb = edge->src;
1306 int pos = edge->pos;
1309 assert(is_merge_edge(bb));
1311 sched_foreach(next_bb, phi) {
1312 const ir_node *phi_arg;
1314 if(!is_Phi(phi)) break;
1316 phi_arg = get_irn_n(phi, pos);
1318 if(phi_arg == irn) {
1326 * Initialize additional node info
1329 luke_initializer(ir_node * bb, void * data)
1331 spill_ilp_t *si = (spill_ilp_t*)data;
1332 spill_bb_t *spill_bb;
1335 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1336 set_irn_link(bb, spill_bb);
1338 sched_foreach(bb, irn) {
1341 op = obstack_alloc(si->obst, sizeof(*op));
1343 op->attr.live_range.ilp = ILP_UNDEF;
1346 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1347 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1349 } else if(!is_Proj(irn)) {
1350 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1351 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1353 op->attr.live_range.args.reloads = NULL;
1355 set_irn_link(irn, op);
1361 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1364 luke_endwalker(ir_node * bb, void * data)
1366 spill_ilp_t *si = (spill_ilp_t*)data;
1372 spill_bb_t *spill_bb = get_irn_link(bb);
1375 live = pset_new_ptr_default();
1376 use_end = pset_new_ptr_default();
1378 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1379 irn = be_lv_get_irn(si->lv, bb, i);
1380 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1381 pset_insert_ptr(live, irn);
1385 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1386 /* their reg_out must always be set */
1387 sched_foreach_reverse(bb, irn) {
1390 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1392 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1393 ir_node *irn_arg = get_irn_n(irn, n);
1395 if(has_reg_class(si, irn_arg)) {
1396 pset_insert_ptr(use_end, irn_arg);
1401 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1402 //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1403 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1405 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1407 /* if this is a merge edge we can reload at the end of this block */
1408 if(is_merge_edge(bb)) {
1409 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1410 } else if(pset_count(use_end)){
1411 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1413 spill_bb->reloads = NULL;
1416 pset_foreach(live,irn) {
1420 int default_spilled;
1423 /* handle values used by control flow nodes later separately */
1424 if(pset_find_ptr(use_end, irn)) continue;
1427 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1429 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1431 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1432 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1433 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1435 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1436 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1438 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1439 /* by default spill value right after definition */
1440 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1441 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1443 if(is_merge_edge(bb)) {
1447 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1448 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), can_be_copied(bb, irn));
1449 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1451 /* reload <= mem_out */
1452 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1453 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1454 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1457 spill->reg_in = ILP_UNDEF;
1458 spill->mem_in = ILP_UNDEF;
1461 pset_foreach(use_end,irn) {
1465 ilp_cst_t end_use_req,
1468 int default_spilled;
1471 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1473 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1475 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1476 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1477 /* if irn is used at the end of the block, then it is live anyway */
1478 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1480 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1481 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1483 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1484 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1485 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1487 /* reload for use be control flow op */
1488 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1489 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1490 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1492 /* reload <= mem_out */
1493 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1494 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1495 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1497 spill->reg_in = ILP_UNDEF;
1498 spill->mem_in = ILP_UNDEF;
1500 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1501 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1502 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1510 next_post_remat(const ir_node * irn)
1516 next = sched_block_first_nonphi(irn);
1518 next = sched_next_op(irn);
1521 if(sched_is_end(next))
1524 op = get_irn_link(next);
1525 if(op->is_remat && !op->attr.remat.pre) {
1534 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1540 ret = sched_block_last_noncf(si, irn);
1541 ret = sched_next(ret);
1542 ret = sched_prev_op(ret);
1544 ret = sched_prev_op(irn);
1547 if(sched_is_end(ret) || is_Phi(ret))
1550 op = (op_t*)get_irn_link(ret);
1551 if(op->is_remat && op->attr.remat.pre) {
1559 * Find a remat of value @p value in the epilog of @p pos
1562 find_post_remat(const ir_node * value, const ir_node * pos)
1564 while((pos = next_post_remat(pos)) != NULL) {
1567 op = get_irn_link(pos);
1568 assert(op->is_remat && !op->attr.remat.pre);
1570 if(op->attr.remat.remat->value == value)
1571 return (ir_node*)pos;
1574 const ir_edge_t *edge;
1575 foreach_out_edge(pos, edge) {
1576 ir_node *proj = get_edge_src_irn(edge);
1577 assert(is_Proj(proj));
1587 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1589 spill_bb_t *spill_bb = get_irn_link(bb);
1593 int default_spilled;
1596 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1598 double spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1600 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1602 spill->reg_out = ILP_UNDEF;
1603 spill->reg_in = ILP_UNDEF;
1604 spill->mem_in = ILP_UNDEF;
1606 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1607 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1609 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1610 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1611 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1618 * Inserts ILP-constraints and variables for memory copying before the given position
1621 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1623 const ir_node *succ;
1624 const ir_edge_t *edge;
1625 spill_bb_t *spill_bb = get_irn_link(block);
1634 assert(edges_activated(current_ir_graph));
1636 edge = get_block_succ_first(block);
1642 edge = get_block_succ_next(block, edge);
1643 /* next block can only contain phis, if this is a merge edge */
1646 ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1647 copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1649 ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1650 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1652 pset_foreach(live, tmp) {
1655 op_t *op = get_irn_link(irn);
1656 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1658 spill = set_find_spill(spill_bb->ilp, tmp);
1661 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1663 lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1665 sched_foreach(succ, phi) {
1666 const ir_node *to_copy;
1668 spill_t *to_copy_spill;
1669 op_t *phi_op = get_irn_link(phi);
1670 ilp_var_t reload = ILP_UNDEF;
1673 if(!is_Phi(phi)) break;
1674 if(!has_reg_class(si, phi)) continue;
1676 to_copy = get_irn_n(phi, pos);
1677 to_copy_op = get_irn_link(to_copy);
1679 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1680 assert(to_copy_spill);
1682 if(spill_bb->reloads) {
1683 keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1686 reload = PTR_TO_INT(keyval->val);
1690 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1691 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1693 /* copy - reg_out - reload - remat - live_range <= 0 */
1694 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1695 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1696 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1697 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1698 foreach_pre_remat(si, block, tmp) {
1699 op_t *remat_op = get_irn_link(tmp);
1700 if(remat_op->attr.remat.remat->value == to_copy) {
1701 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1705 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1706 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1708 /* copy - reg_out - copyreg <= 0 */
1709 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1710 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1711 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1717 * Walk all irg blocks and emit this ILP
1720 luke_blockwalker(ir_node * bb, void * data)
1722 spill_ilp_t *si = (spill_ilp_t*)data;
1727 spill_bb_t *spill_bb = get_irn_link(bb);
1730 pset *defs = pset_new_ptr_default();
1731 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
1733 live = pset_new_ptr_default();
1735 /****************************************
1736 * B A S I C B L O C K E N D
1737 ***************************************/
1740 /* init live values at end of block */
1741 get_live_end(si, bb, live);
1743 pset_foreach(live, irn) {
1745 ilp_var_t reload = ILP_UNDEF;
1747 spill = set_find_spill(spill_bb->ilp, irn);
1750 if(spill_bb->reloads) {
1751 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1754 reload = PTR_TO_INT(keyval->val);
1758 op = get_irn_link(irn);
1759 assert(!op->is_remat);
1761 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1762 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1763 op->attr.live_range.op = bb;
1765 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1766 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1768 /* reg_out - reload - remat - live_range <= 0 */
1769 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1770 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1771 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1772 foreach_pre_remat(si, bb, tmp) {
1773 op_t *remat_op = get_irn_link(tmp);
1774 if(remat_op->attr.remat.remat->value == irn) {
1775 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1778 ir_snprintf(buf, sizeof(buf), "reg_out2_%N_%N", bb, irn);
1779 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
1781 /* value may only die at bb end if it is used for a mem copy */
1782 /* reg_out + \sum copy - reload - remat - live_range >= 0 */
1783 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1784 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1785 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1786 foreach_pre_remat(si, bb, tmp) {
1787 op_t *remat_op = get_irn_link(tmp);
1788 if(remat_op->attr.remat.remat->value == irn) {
1789 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1792 if(is_merge_edge(bb)) {
1793 const ir_edge_t *edge = get_block_succ_first(bb);
1794 const ir_node *next_bb = edge->src;
1795 int pos = edge->pos;
1798 sched_foreach(next_bb, phi) {
1799 const ir_node *phi_arg;
1801 if(!is_Phi(phi)) break;
1803 phi_arg = get_irn_n(phi, pos);
1805 if(phi_arg == irn) {
1806 op_t *phi_op = get_irn_link(phi);
1807 ilp_var_t copy = phi_op->attr.live_range.args.copies[pos];
1809 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
1816 insert_mem_copy_position(si, live, bb);
1819 * start new live ranges for values used by remats at end of block
1820 * and assure the remat args are available
1822 foreach_pre_remat(si, bb, tmp) {
1823 op_t *remat_op = get_irn_link(tmp);
1826 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1827 ir_node *remat_arg = get_irn_n(tmp, n);
1828 op_t *arg_op = get_irn_link(remat_arg);
1831 if(!has_reg_class(si, remat_arg)) continue;
1833 /* if value is becoming live through use by remat */
1834 if(!pset_find_ptr(live, remat_arg)) {
1835 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1836 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1838 arg_op->attr.live_range.ilp = prev_lr;
1839 arg_op->attr.live_range.op = bb;
1841 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1843 pset_insert_ptr(live, remat_arg);
1844 add_to_spill_bb(si, bb, remat_arg);
1847 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1848 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1849 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1851 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1852 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1854 /* use reload placed for this argument */
1855 if(spill_bb->reloads) {
1856 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1859 ilp_var_t reload = PTR_TO_INT(keyval->val);
1861 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1866 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1871 /**************************************
1872 * B A S I C B L O C K B O D Y
1873 **************************************/
1875 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1881 ilp_cst_t check_pre,
1887 ilp_cst_t one_memoperand;
1889 /* iterate only until first phi */
1893 op = get_irn_link(irn);
1895 if(op->is_remat) continue;
1897 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1899 /* collect defined values */
1900 if(has_reg_class(si, irn)) {
1901 pset_insert_ptr(defs, irn);
1905 if(is_Proj(irn)) continue;
1908 * init set of irn's arguments
1909 * and all possibly used values around this op
1910 * and values defined by post remats
1912 args = new_set(cmp_keyval, get_irn_arity(irn));
1913 used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1914 remat_defs = pset_new_ptr(pset_count(live));
1916 if(!is_start_block(bb) || !be_is_Barrier(irn)) {
1917 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1918 ir_node *irn_arg = get_irn_n(irn, n);
1919 if(has_reg_class(si, irn_arg)) {
1920 set_insert_keyval(args, irn_arg, (void*)n);
1921 pset_insert_ptr(used, irn_arg);
1924 foreach_post_remat(irn, tmp) {
1925 op_t *remat_op = get_irn_link(tmp);
1927 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1929 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1930 ir_node *remat_arg = get_irn_n(tmp, n);
1931 if(has_reg_class(si, remat_arg)) {
1932 pset_insert_ptr(used, remat_arg);
1936 foreach_pre_remat(si, irn, tmp) {
1937 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1938 ir_node *remat_arg = get_irn_n(tmp, n);
1939 if(has_reg_class(si, remat_arg)) {
1940 pset_insert_ptr(used, remat_arg);
1946 /**********************************
1947 * I N E P I L O G O F irn
1948 **********************************/
1950 /* ensure each dying value is used by only one post remat */
1951 pset_foreach(used, tmp) {
1952 ir_node *value = tmp;
1953 op_t *value_op = get_irn_link(value);
1958 foreach_post_remat(irn, remat) {
1959 op_t *remat_op = get_irn_link(remat);
1961 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1962 ir_node *remat_arg = get_irn_n(remat, n);
1964 /* if value is used by this remat add it to constraint */
1965 if(remat_arg == value) {
1967 /* sum remat2s <= 1 + n_remats*live_range */
1968 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1969 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
1973 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1979 if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
1980 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1984 /* ensure at least one value dies at post remat */
1985 foreach_post_remat(irn, tmp) {
1986 op_t *remat_op = get_irn_link(tmp);
1987 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
1990 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
1991 remat_arg = get_irn_n(tmp, n);
1993 if(has_reg_class(si, remat_arg)) {
1995 /* does arg always die at this op? */
1996 if(!pset_find_ptr(live, remat_arg))
1997 goto skip_one_must_die;
1999 pset_insert_ptr(remat_args, remat_arg);
2003 /* remat + \sum live_range(remat_arg) <= |args| */
2004 ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
2005 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2006 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2008 pset_foreach(remat_args, remat_arg) {
2009 op_t *arg_op = get_irn_link(remat_arg);
2011 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2015 del_pset(remat_args);
2018 /* new live ranges for values from L\U defined by post remats */
2019 pset_foreach(live, tmp) {
2020 ir_node *value = tmp;
2021 op_t *value_op = get_irn_link(value);
2023 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
2024 ilp_var_t prev_lr = ILP_UNDEF;
2027 if(pset_find_ptr(remat_defs, value)) {
2029 /* next_live_range <= prev_live_range + sum remat2s */
2030 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
2031 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2033 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
2034 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2036 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
2037 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2039 foreach_post_remat(irn, remat) {
2040 op_t *remat_op = get_irn_link(remat);
2042 /* if value is being rematerialized by this remat */
2043 if(value == remat_op->attr.remat.remat->value) {
2044 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2048 value_op->attr.live_range.ilp = prev_lr;
2049 value_op->attr.live_range.op = irn;
2054 /* requirements for post remats and start live ranges from L/U' for values dying here */
2055 foreach_post_remat(irn, tmp) {
2056 op_t *remat_op = get_irn_link(tmp);
2059 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2060 ir_node *remat_arg = get_irn_n(tmp, n);
2061 op_t *arg_op = get_irn_link(remat_arg);
2063 if(!has_reg_class(si, remat_arg)) continue;
2065 /* only for values in L\U (TODO and D?), the others are handled with post_use */
2066 if(!pset_find_ptr(used, remat_arg)) {
2067 /* remat <= live_range(remat_arg) */
2068 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
2069 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2071 /* if value is becoming live through use by remat2 */
2072 if(!pset_find_ptr(live, remat_arg)) {
2075 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
2076 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2078 arg_op->attr.live_range.ilp = lr;
2079 arg_op->attr.live_range.op = irn;
2081 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
2083 pset_insert_ptr(live, remat_arg);
2084 add_to_spill_bb(si, bb, remat_arg);
2087 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2088 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2093 d = pset_count(defs);
2094 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
2096 /* count how many regs irn needs for arguments */
2097 u = set_count(args);
2100 /* check the register pressure in the epilog */
2101 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
2102 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
2103 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
2105 /* add L\U' to check_post */
2106 pset_foreach(live, tmp) {
2107 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
2108 /* if a live value is not used by irn */
2109 tmp_op = get_irn_link(tmp);
2110 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
2114 /***********************************************************
2115 * 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
2116 **********************************************************/
2119 pset_foreach(used, tmp) {
2125 op_t *arg_op = get_irn_link(arg);
2128 spill = add_to_spill_bb(si, bb, arg);
2130 /* new live range for each used value */
2131 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2132 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2134 /* the epilog stuff - including post_use, check_post, check_post_remat */
2135 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2136 post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2138 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2140 /* arg is live throughout epilog if the next live_range is in a register */
2141 if(pset_find_ptr(live, arg)) {
2142 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
2144 /* post_use >= next_lr + remat */
2145 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2146 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2147 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2148 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2151 /* forall post remat which use arg add a similar cst */
2152 foreach_post_remat(irn, remat) {
2155 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2156 ir_node *remat_arg = get_irn_n(remat, n);
2157 op_t *remat_op = get_irn_link(remat);
2159 if(remat_arg == arg) {
2160 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
2162 /* post_use >= remat */
2163 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2164 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2165 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2166 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2171 /* if value is not an arg of op and not possibly defined by post remat
2172 * then it may only die and not become live
2174 if(!set_find_keyval(args, arg)) {
2175 /* post_use <= prev_lr */
2176 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2177 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2178 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2179 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2181 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2182 /* next_lr <= prev_lr */
2183 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2184 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2185 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2186 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2190 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2191 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2192 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2193 ilp_var_t memoperand;
2195 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2196 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0);
2197 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2199 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2200 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2202 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2203 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2208 /* new live range begins for each used value */
2209 arg_op->attr.live_range.ilp = prev_lr;
2210 arg_op->attr.live_range.op = irn;
2212 pset_insert_ptr(live, arg);
2215 /* just to be sure */
2216 check_post = ILP_UNDEF;
2218 /* allow original defintions to be removed */
2219 if(opt_repair_schedule) {
2220 pset_foreach(defs, tmp) {
2221 op_t *tmp_op = get_irn_link(tmp);
2222 spill_t *spill = set_find_spill(spill_bb->ilp, tmp);
2227 ir_snprintf(buf, sizeof(buf), "delete_%N", tmp);
2228 delete = lpp_add_var_default(si->lpp, buf, lpp_binary, -1.0*get_cost(si, irn)*execution_frequency(si, bb), 0.0);
2230 /* op may not be killed if its first live_range is 1 */
2231 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2232 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2233 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2234 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2236 /* op may not be killed if it is spilled after the definition */
2237 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2238 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2239 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2240 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2245 ir_snprintf(buf, sizeof(buf), "keep_%N", tmp);
2246 keep = lpp_add_var_default(si->lpp, buf, lpp_binary, get_cost(si, irn)*execution_frequency(si, bb), 1.0);
2248 /* op may not be killed if its first live_range is 1 */
2249 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2250 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2251 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2252 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, -1.0);
2254 /* op may not be killed if it is spilled after the definition */
2255 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2256 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2257 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2258 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2262 pset_foreach(defs, tmp) {
2263 op_t *tmp_op = get_irn_link(tmp);
2264 spill_t *spill = set_find_spill(spill_bb->ilp, tmp);
2267 /* live_range or spill should be 1
2268 TODO: lr should be live until first use */
2269 ir_snprintf(buf, sizeof(buf), "nokillorig_%N", tmp);
2270 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 1.0);
2271 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2272 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2282 /* check the register pressure in the prolog */
2283 /* sum_{L\U} lr <= k - |U| */
2284 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2285 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2287 /* for the prolog remove defined values from the live set */
2288 pset_foreach(defs, tmp) {
2289 pset_remove_ptr(live, tmp);
2292 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2293 ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2294 one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2297 /***********************************************************
2298 * 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
2299 **********************************************************/
2302 set_foreach(args, keyval) {
2304 const ir_node *arg = keyval->key;
2305 int i = PTR_TO_INT(keyval->val);
2306 op_t *arg_op = get_irn_link(arg);
2307 ilp_cst_t requirements;
2310 spill = set_find_spill(spill_bb->ilp, arg);
2313 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2314 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);
2316 /* reload <= mem_out */
2317 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2318 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2319 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2320 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2322 /* requirement: arg must be in register for use */
2323 /* reload + remat + live_range == 1 */
2324 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2325 requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2327 lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2328 lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2329 foreach_pre_remat(si, irn, tmp) {
2330 op_t *remat_op = get_irn_link(tmp);
2331 if(remat_op->attr.remat.remat->value == arg) {
2332 lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2336 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2338 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2339 if(get_irn_n(irn, n) == arg) {
2343 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2344 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2345 memoperand_t *memoperand;
2346 memoperand = set_find_memoperand(si->memoperands, irn, n);
2348 /* memoperand <= mem_out */
2349 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2350 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2351 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2352 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2354 /* the memoperand is only sufficient if it is used once by the op */
2355 if(n_memoperands == 1)
2356 lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2358 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2360 /* we have one more free register if we use a memory operand */
2361 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2367 /* iterate over L\U */
2368 pset_foreach(live, tmp) {
2369 if(!set_find_keyval(args, tmp)) {
2370 /* if a live value is not used by irn */
2371 tmp_op = get_irn_link(tmp);
2372 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2376 /* requirements for remats */
2377 foreach_pre_remat(si, irn, tmp) {
2378 op_t *remat_op = get_irn_link(tmp);
2381 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2382 ir_node *remat_arg = get_irn_n(tmp, n);
2383 op_t *arg_op = get_irn_link(remat_arg);
2385 if(!has_reg_class(si, remat_arg)) continue;
2387 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2388 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2389 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2391 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2392 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2394 /* if remat arg is also used by current op then we can use reload placed for this argument */
2395 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2396 int index = (int)keyval->val;
2398 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2406 /*************************
2407 * D O N E W I T H O P
2408 *************************/
2410 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2412 pset_foreach(live, tmp) {
2413 assert(has_reg_class(si, tmp));
2416 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2417 ir_node *arg = get_irn_n(irn, n);
2419 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2422 del_pset(remat_defs);
2426 defs = pset_new_ptr_default();
2428 /* skip everything above barrier in start block */
2429 if(is_start_block(bb) && be_is_Barrier(irn)) {
2430 assert(pset_count(live) == 0);
2439 /***************************************
2440 * B E G I N N I N G O F B L O C K
2441 ***************************************/
2444 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2445 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2447 pset_foreach(live, irn) {
2448 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2451 /* construct mem_outs for all values */
2452 set_foreach(spill_bb->ilp, spill) {
2453 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2454 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2456 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2457 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2459 if(pset_find_ptr(live, spill->irn)) {
2460 int default_spilled;
2461 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2463 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2464 default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2465 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2466 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2468 if(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2470 op_t *op = get_irn_link(spill->irn);
2472 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2473 const ir_node *arg = get_irn_n(spill->irn, n);
2479 /* argument already done? */
2480 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2482 /* get sum of execution frequencies of blocks with the same phi argument */
2483 for(m=n; m>=0; --m) {
2484 const ir_node *arg2 = get_irn_n(spill->irn, m);
2487 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2491 /* copies are not for free */
2492 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2493 var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
2495 for(m=n; m>=0; --m) {
2496 const ir_node *arg2 = get_irn_n(spill->irn, m);
2499 op->attr.live_range.args.copies[m] = var;
2504 /* copy <= mem_in */
2505 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2506 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2507 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2508 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2515 foreach_post_remat(bb, tmp) {
2517 op_t *remat_op = get_irn_link(tmp);
2518 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
2521 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2522 remat_arg = get_irn_n(tmp, n);
2524 if(has_reg_class(si, remat_arg)) {
2525 pset_insert_ptr(remat_args, remat_arg);
2529 /* remat + \sum live_range(remat_arg) <= |args| */
2530 ir_snprintf(buf, sizeof(buf), "one_must_die_%N", tmp);
2531 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2532 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2534 pset_foreach(remat_args, remat_arg) {
2535 if(pset_find_ptr(live, remat_arg)) {
2536 op_t *remat_arg_op = get_irn_link(remat_arg);
2537 lpp_set_factor_fast(si->lpp, cst, remat_arg_op->attr.live_range.ilp, 1.0);
2540 del_pset(remat_args);
2543 foreach_post_remat(bb, tmp) {
2546 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2547 ir_node *remat_arg = get_irn_n(tmp, n);
2549 /* if value is becoming live through use by remat2 */
2550 if(has_reg_class(si, remat_arg) && !pset_find_ptr(live, remat_arg)) {
2551 op_t *remat_arg_op = get_irn_link(remat_arg);
2554 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 at bb start %+F\n", remat_arg, tmp));
2556 pset_insert_ptr(live, remat_arg);
2557 spill = add_to_spill_bb(si, bb, remat_arg);
2558 remat_arg_op->attr.live_range.ilp = ILP_UNDEF;
2560 /* we need reg_in and mem_in for this value; they will be referenced later */
2561 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", remat_arg, bb);
2562 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2563 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", remat_arg, bb);
2564 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2567 /* optimization: all memory stuff should be 0, for we do not want to insert reloads for remats */
2568 ir_snprintf(buf, sizeof(buf), "nomem_%N_%N", remat_arg, bb);
2569 nomem = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0.0);
2570 lpp_set_factor_fast(si->lpp, nomem, spill->spill, 1.0);
2575 /* L\U is empty at bb start */
2576 /* arg is live throughout epilog if it is reg_in into this block */
2578 /* check the register pressure at the beginning of the block
2581 /* reg_in entspricht post_use */
2583 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2584 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2586 pset_foreach(live, irn) {
2589 spill = set_find_spill(spill_bb->ilp, irn);
2592 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2593 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2595 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2597 /* spill + mem_in <= 1 */
2598 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2599 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2601 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2602 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2604 } /* post_remats are NOT included in register pressure check because
2605 they do not increase regpressure */
2607 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2608 pset_foreach(live, irn) {
2612 spill = set_find_spill(spill_bb->ilp, irn);
2613 assert(spill && spill->irn == irn);
2615 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2616 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2619 ir_node *phi_arg = get_Phi_pred(irn, n);
2620 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2621 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2623 op_t *op = get_irn_link(irn);
2625 /* although the phi is in the right regclass one or more of
2626 * its arguments can be in a different one or at least to
2629 if(has_reg_class(si, phi_arg)) {
2630 /* mem_in < mem_out_arg + copy */
2631 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2632 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2634 /* reg_in < reg_out_arg */
2635 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2636 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2638 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2639 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2641 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2644 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2646 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2648 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2652 /* else assure the value arrives on all paths in the same resource */
2654 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2657 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2658 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2661 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2662 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2663 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2664 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2666 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2667 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2669 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2672 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2673 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2678 foreach_post_remat(bb, tmp) {
2681 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2682 ir_node *remat_arg = get_irn_n(tmp, n);
2683 op_t *remat_op = get_irn_link(tmp);
2685 if(!has_reg_class(si, remat_arg)) continue;
2687 spill = set_find_spill(spill_bb->ilp, remat_arg);
2690 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2691 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2692 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2693 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2697 pset_foreach(live, irn) {
2698 const op_t *op = get_irn_link(irn);
2699 const ir_node *remat;
2704 foreach_post_remat(bb, remat) {
2707 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2708 const ir_node *arg = get_irn_n(remat, n);
2711 const op_t *remat_op = get_irn_link(remat);
2713 if(cst == ILP_UNDEF) {
2714 /* sum remat2s <= 1 + n_remats*live_range */
2715 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", irn, bb);
2716 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2718 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2724 if(cst != ILP_UNDEF && op->attr.live_range.ilp != ILP_UNDEF) {
2725 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -n_remats);
2729 /* first live ranges from reg_ins */
2730 pset_foreach(live, irn) {
2731 op_t *op = get_irn_link(irn);
2733 if(op->attr.live_range.ilp != ILP_UNDEF) {
2735 spill = set_find_spill(spill_bb->ilp, irn);
2736 assert(spill && spill->irn == irn);
2738 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2739 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2740 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2741 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2743 foreach_post_remat(bb, tmp) {
2744 op_t *remat_op = get_irn_link(tmp);
2746 if(remat_op->attr.remat.remat->value == irn) {
2747 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2753 /* walk forward now and compute constraints for placing spills */
2754 /* this must only be done for values that are not defined in this block */
2755 pset_foreach(live, irn) {
2757 * if value is defined in this block we can anways place the spill directly after the def
2758 * -> no constraint necessary
2760 if(!is_Phi(irn) && get_nodes_block(irn) == bb) {
2765 spill = set_find_spill(spill_bb->ilp, irn);
2768 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2769 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2771 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2772 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2775 sched_foreach_op(bb, tmp) {
2776 op_t *op = get_irn_link(tmp);
2778 if(is_Phi(tmp)) continue;
2779 assert(!is_Proj(tmp));
2782 const ir_node *value = op->attr.remat.remat->value;
2785 /* only collect remats up to the first real use of a value */
2786 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2791 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2792 ir_node *arg = get_irn_n(tmp, n);
2795 /* if a value is used stop collecting remats */
2808 typedef struct _irnlist_t {
2809 struct list_head list;
2813 typedef struct _interference_t {
2814 struct list_head blocklist;
2820 cmp_interference(const void *a, const void *b, size_t size)
2822 const interference_t *p = a;
2823 const interference_t *q = b;
2825 return !(p->a == q->a && p->b == q->b);
2828 static interference_t *
2829 set_find_interference(set * set, ir_node * a, ir_node * b)
2831 interference_t query;
2833 query.a = (a>b)?a:b;
2834 query.b = (a>b)?b:a;
2836 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2839 static interference_t *
2840 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2842 interference_t query,
2844 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2848 result = set_find_interference(set, a, b);
2851 list_add(&list->list, &result->blocklist);
2855 query.a = (a>b)?a:b;
2856 query.b = (a>b)?b:a;
2858 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2860 INIT_LIST_HEAD(&result->blocklist);
2861 list_add(&list->list, &result->blocklist);
2867 values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
2869 const ir_edge_t *edge;
2871 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2872 /* both values are live in, so they interfere */
2876 /* ensure a dominates b */
2877 if(value_dominates(b,a)) {
2883 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2886 /* the following code is stolen from bera.c */
2887 if(be_is_live_end(si->lv, bb, a))
2890 foreach_out_edge(a, edge) {
2891 const ir_node *user = edge->src;
2892 if(get_nodes_block(user) == bb
2895 && !pset_find_ptr(si->inverse_ops, user)
2896 && value_dominates(b, user))
2904 * Walk all irg blocks and collect interfering values inside of phi classes
2907 luke_interferencewalker(ir_node * bb, void * data)
2909 spill_ilp_t *si = (spill_ilp_t*)data;
2912 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2913 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
2914 op_t *a_op = get_irn_link(a);
2917 /* a is only interesting if it is in my register class and if it is inside a phi class */
2918 if (has_reg_class(si, a) && get_phi_class(a)) {
2919 if(a_op->is_remat || pset_find_ptr(si->inverse_ops, a))
2922 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)) {
2923 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
2924 op_t *b_op = get_irn_link(b);
2927 /* a and b are only interesting if they are in the same phi class */
2928 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2929 if(b_op->is_remat || pset_find_ptr(si->inverse_ops, b))
2932 if(values_interfere_in_block(si, bb, a, b)) {
2933 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2934 set_insert_interference(si, si->interferences, a, b, bb);
2942 static unsigned int copy_path_id = 0;
2945 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2952 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2953 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2955 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2957 pset_foreach(copies, ptr) {
2958 copy = PTR_TO_INT(ptr);
2959 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2964 * @parameter copies contains a path of copies which lead us to irn
2965 * @parameter visited contains a set of nodes already visited on this path
2968 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2970 const ir_edge_t *edge;
2971 op_t *op = get_irn_link(irn);
2972 pset *visited_users = pset_new_ptr_default();
2975 if(op->is_remat) return 0;
2977 pset_insert_ptr(visited, irn);
2981 pset *visited_operands = pset_new_ptr(get_irn_arity(irn));
2983 /* visit all operands */
2984 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2985 ir_node *arg = get_irn_n(irn, n);
2986 ilp_var_t copy = op->attr.live_range.args.copies[n];
2988 if(!has_reg_class(si, arg)) continue;
2989 if(pset_find_ptr(visited_operands, arg)) continue;
2990 pset_insert_ptr(visited_operands, arg);
2993 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2994 del_pset(visited_operands);
2995 del_pset(visited_users);
2996 pset_remove_ptr(visited, irn);
2999 pset_insert(copies, INT_TO_PTR(copy), copy);
3000 write_copy_path_cst(si, copies, any_interfere);
3001 pset_remove(copies, INT_TO_PTR(copy), copy);
3002 } else if(!pset_find_ptr(visited, arg)) {
3003 pset_insert(copies, INT_TO_PTR(copy), copy);
3004 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
3005 pset_remove(copies, INT_TO_PTR(copy), copy);
3007 if(paths > MAX_PATHS) {
3008 if(pset_count(copies) == 0) {
3012 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3013 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3014 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3015 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3016 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3020 del_pset(visited_operands);
3021 del_pset(visited_users);
3022 pset_remove_ptr(visited, irn);
3025 } else if(pset_count(copies) == 0) {
3031 del_pset(visited_operands);
3034 /* visit all uses which are phis */
3035 foreach_out_edge(irn, edge) {
3036 ir_node *user = edge->src;
3037 int pos = edge->pos;
3038 op_t *op = get_irn_link(user);
3041 if(!is_Phi(user)) continue;
3042 if(!has_reg_class(si, user)) continue;
3043 if(pset_find_ptr(visited_users, user)) continue;
3044 pset_insert_ptr(visited_users, user);
3046 copy = op->attr.live_range.args.copies[pos];
3048 if(user == target) {
3049 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3050 del_pset(visited_users);
3051 pset_remove_ptr(visited, irn);
3054 pset_insert(copies, INT_TO_PTR(copy), copy);
3055 write_copy_path_cst(si, copies, any_interfere);
3056 pset_remove(copies, INT_TO_PTR(copy), copy);
3057 } else if(!pset_find_ptr(visited, user)) {
3058 pset_insert(copies, INT_TO_PTR(copy), copy);
3059 paths += find_copy_path(si, user, target, any_interfere, copies, visited);
3060 pset_remove(copies, INT_TO_PTR(copy), copy);
3062 if(paths > MAX_PATHS) {
3063 if(pset_count(copies) == 0) {
3067 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3068 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3069 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3070 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3071 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3075 del_pset(visited_users);
3076 pset_remove_ptr(visited, irn);
3079 } else if(pset_count(copies) == 0) {
3085 del_pset(visited_users);
3086 pset_remove_ptr(visited, irn);
3091 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
3093 pset * copies = pset_new_ptr_default();
3094 pset * visited = pset_new_ptr_default();
3096 find_copy_path(si, a, b, any_interfere, copies, visited);
3104 memcopyhandler(spill_ilp_t * si)
3106 interference_t *interference;
3108 /* teste Speicherwerte auf Interferenz */
3110 /* analyze phi classes */
3111 phi_class_compute(si->chordal_env->irg);
3113 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
3114 irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
3116 /* now lets emit the ILP unequations for the crap */
3117 set_foreach(si->interferences, interference) {
3119 ilp_var_t interfere,
3121 ilp_cst_t any_interfere_cst,
3123 const ir_node *a = interference->a;
3124 const ir_node *b = interference->b;
3126 /* any_interf <= \sum interf */
3127 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
3128 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3129 any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3131 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
3133 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
3134 const ir_node *bb = irnlist->irn;
3135 spill_bb_t *spill_bb = get_irn_link(bb);
3140 spilla = set_find_spill(spill_bb->ilp, a);
3143 spillb = set_find_spill(spill_bb->ilp, b);
3146 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
3147 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
3148 /* 2: - mem_in_a - spill_a + interfere <= 0 */
3149 /* 3: - mem_in_b - spill_b + interfere <= 0 */
3150 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
3151 interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3153 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
3154 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
3156 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
3157 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
3158 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
3159 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
3160 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
3162 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
3163 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3165 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3166 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
3167 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
3169 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
3170 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3172 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3173 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
3174 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
3177 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
3179 /* any_interfere >= interf */
3180 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
3181 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3183 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3184 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3187 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
3188 gen_copy_constraints(si,a,b,any_interfere);
3196 return fabs(x) < 0.00001;
3199 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
3201 spill_ilp_t *si = get_irg_link(current_ir_graph);
3203 if(pset_find_ptr(si->all_possible_remats, n)) {
3204 op_t *op = (op_t*)get_irn_link(n);
3205 assert(op && op->is_remat);
3207 if(!op->attr.remat.remat->inverse) {
3208 if(op->attr.remat.pre) {
3209 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
3211 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
3216 op_t *op = (op_t*)get_irn_link(n);
3217 assert(op && op->is_remat);
3219 if(op->attr.remat.pre) {
3220 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
3222 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
3233 dump_graph_with_remats(ir_graph * irg, const char * suffix)
3235 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
3236 be_dump(irg, suffix, dump_ir_block_graph_sched);
3237 set_dump_node_vcgattr_hook(NULL);
3241 * Edge hook to dump the schedule edges with annotated register pressure.
3244 sched_pressure_edge_hook(FILE *F, ir_node *irn)
3246 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
3247 ir_node *prev = sched_prev(irn);
3248 fprintf(F, "edge:{sourcename:\"");
3250 fprintf(F, "\" targetname:\"");
3252 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
3253 fprintf(F, "\" color:magenta}\n");
3259 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
3261 DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3263 dump_consts_local(0);
3264 set_dump_node_edge_hook(sched_pressure_edge_hook);
3265 dump_ir_block_graph(irg, suffix);
3266 set_dump_node_edge_hook(old_edge_hook);
3270 walker_pressure_annotator(ir_node * bb, void * data)
3272 spill_ilp_t *si = data;
3275 pset *live = pset_new_ptr_default();
3278 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3279 irn = be_lv_get_irn(si->lv, bb, i);
3281 if (has_reg_class(si, irn)) {
3282 pset_insert_ptr(live, irn);
3286 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3288 sched_foreach_reverse(bb, irn) {
3290 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3294 if(has_reg_class(si, irn)) {
3295 pset_remove_ptr(live, irn);
3296 if(is_Proj(irn)) ++projs;
3299 if(!is_Proj(irn)) projs = 0;
3301 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3302 ir_node *arg = get_irn_n(irn, n);
3304 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3306 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3313 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3315 be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
3319 connect_all_remats_with_keep(spill_ilp_t * si)
3327 n_remats = pset_count(si->all_possible_remats);
3329 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3332 pset_foreach(si->all_possible_remats, irn) {
3337 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
3339 obstack_free(si->obst, ins);
3344 connect_all_spills_with_keep(spill_ilp_t * si)
3353 n_spills = pset_count(si->spills);
3355 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3358 pset_foreach(si->spills, irn) {
3363 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
3365 obstack_free(si->obst, ins);
3369 /** insert a spill at an arbitrary position */
3370 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3372 ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
3373 ir_graph *irg = get_irn_irg(bl);
3374 ir_node *frame = get_irg_frame(irg);
3378 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
3379 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3381 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3384 * search the right insertion point. a spill of a phi cannot be put
3385 * directly after the phi, if there are some phis behind the one which
3386 * is spilled. Also, a spill of a Proj must be after all Projs of the
3389 * Here's one special case:
3390 * If the spill is in the start block, the spill must be after the frame
3391 * pointer is set up. This is done by setting insert to the end of the block
3392 * which is its default initialization (see above).
3395 if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3398 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3401 sched_add_after(insert, spill);
3406 delete_remat(spill_ilp_t * si, ir_node * remat) {
3408 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3410 sched_remove(remat);
3412 /* kill links to operands */
3413 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3414 set_irn_n(remat, n, bad);
3419 clean_remat_info(spill_ilp_t * si)
3423 remat_info_t *remat_info;
3424 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3426 set_foreach(si->remat_info, remat_info) {
3427 if(!remat_info->remats) continue;
3429 pset_foreach(remat_info->remats, remat)
3431 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3432 if(sched_is_scheduled(remat->proj)) {
3433 sched_remove((ir_node*)remat->proj);
3435 set_irn_n((ir_node*)remat->proj, -1, bad);
3436 set_irn_n((ir_node*)remat->proj, 0, bad);
3439 if(get_irn_n_edges(remat->op) == 0) {
3440 if(sched_is_scheduled(remat->op)) {
3441 sched_remove((ir_node*)remat->op);
3443 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3444 set_irn_n((ir_node*)remat->op, n, bad);
3449 if(remat_info->remats) del_pset(remat_info->remats);
3450 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3455 delete_unnecessary_remats(spill_ilp_t * si)
3457 if(opt_keep_alive & KEEPALIVE_REMATS) {
3459 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3462 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3463 ir_node *keep_arg = get_irn_n(si->keep, n);
3464 op_t *arg_op = get_irn_link(keep_arg);
3467 assert(arg_op->is_remat);
3469 name = si->lpp->vars[arg_op->attr.remat.ilp];
3471 if(is_zero(name->value)) {
3472 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
3473 /* TODO check whether reload is preferred over remat (could be bug) */
3474 delete_remat(si, keep_arg);
3476 if(!arg_op->attr.remat.remat->inverse) {
3477 if(arg_op->attr.remat.pre) {
3478 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3480 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3483 if(arg_op->attr.remat.pre) {
3484 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3486 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3491 set_irn_n(si->keep, n, bad);
3494 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
3499 pset_foreach(si->all_possible_remats, remat) {
3500 op_t *remat_op = get_irn_link(remat);
3501 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
3503 if(is_zero(name->value)) {
3504 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
3505 /* TODO check whether reload is preferred over remat (could be bug) */
3506 delete_remat(si, remat);
3508 if(!remat_op->attr.remat.remat->inverse) {
3509 if(remat_op->attr.remat.pre) {
3510 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3512 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3515 if(remat_op->attr.remat.pre) {
3516 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3518 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3527 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3529 pset *spills = pset_new_ptr_default();
3531 const ir_node *next;
3534 defs = set_find_def(si->values, value);
3536 if(defs && defs->spills) {
3537 for(next = defs->spills; next; next = get_irn_link(next)) {
3538 pset_insert_ptr(spills, next);
3546 * @param before The node after which the spill will be placed in the schedule
3549 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3553 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3555 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3557 spill = be_spill2(arch_env, irn, before);
3559 defs = set_insert_def(si->values, value);
3562 /* enter into the linked list */
3563 set_irn_link(spill, defs->spills);
3564 defs->spills = spill;
3566 if(opt_keep_alive & KEEPALIVE_SPILLS)
3567 pset_insert_ptr(si->spills, spill);
3573 * @param before The Phi node which has to be spilled
3576 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3583 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3585 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3586 ins[n] = si->m_unknown;
3589 mem_phi = new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3591 defs = set_insert_def(si->values, phi);
3594 /* enter into the linked list */
3595 set_irn_link(mem_phi, defs->spills);
3596 defs->spills = mem_phi;
3598 #ifdef SCHEDULE_PHIM
3599 sched_add_after(phi, mem_phi);
3602 if(opt_keep_alive & KEEPALIVE_SPILLS)
3603 pset_insert_ptr(si->spills, mem_phi);
3610 * Add remat to list of defs, destroys link field!
3613 insert_remat(spill_ilp_t * si, ir_node * remat)
3616 op_t *remat_op = get_irn_link(remat);
3618 assert(remat_op->is_remat);
3620 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3623 /* enter into the linked list */
3624 set_irn_link(remat, defs->remats);
3625 defs->remats = remat;
3630 * Add reload before operation and add to list of defs
3633 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3638 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3640 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3642 defs = set_find_def(si->values, value);
3644 spill = defs->spills;
3645 assert(spill && "no spill placed before reload");
3647 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3649 /* enter into the linked list */
3650 set_irn_link(reload, defs->remats);
3651 defs->remats = reload;
3656 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3659 ir_node *value = get_irn_n(memoperand->irn, memoperand->pos);
3661 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3663 DBG((si->dbg, LEVEL_2, "\t inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3665 defs = set_find_def(si->values, value);
3667 spill = defs->spills;
3668 assert(spill && "no spill placed before reload");
3670 arch_perform_memory_operand(arch_env, memoperand->irn, spill, memoperand->pos);
3673 void insert_memoperands(spill_ilp_t * si)
3675 memoperand_t *memoperand;
3678 set_foreach(si->memoperands, memoperand) {
3679 name = si->lpp->vars[memoperand->ilp];
3680 if(!is_zero(name->value)) {
3681 perform_memory_operand(si, memoperand);
3687 walker_spill_placer(ir_node * bb, void * data) {
3688 spill_ilp_t *si = (spill_ilp_t*)data;
3690 spill_bb_t *spill_bb = get_irn_link(bb);
3691 pset *spills_to_do = pset_new_ptr_default();
3694 set_foreach(spill_bb->ilp, spill) {
3697 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3698 name = si->lpp->vars[spill->mem_in];
3699 if(!is_zero(name->value)) {
3702 mem_phi = insert_mem_phi(si, spill->irn);
3704 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3708 name = si->lpp->vars[spill->spill];
3709 if(!is_zero(name->value)) {
3710 /* place spill directly after definition */
3711 if(get_nodes_block(spill->irn) == bb) {
3712 insert_spill(si, spill->irn, spill->irn, spill->irn);
3716 /* place spill at bb start */
3717 if(spill->reg_in > 0) {
3718 name = si->lpp->vars[spill->reg_in];
3719 if(!is_zero(name->value)) {
3720 insert_spill(si, spill->irn, spill->irn, bb);
3724 /* place spill after a remat */
3725 pset_insert_ptr(spills_to_do, spill->irn);
3728 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3731 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3732 op_t *op = get_irn_link(irn);
3734 if(be_is_Spill(irn)) continue;
3737 /* TODO fix this if we want to support remats with more than two nodes */
3738 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3739 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3741 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3744 if(pset_find_ptr(spills_to_do, irn)) {
3745 pset_remove_ptr(spills_to_do, irn);
3747 insert_spill(si, irn, irn, irn);
3753 assert(pset_count(spills_to_do) == 0);
3755 /* afterwards free data in block */
3756 del_pset(spills_to_do);
3760 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3762 ir_node *insert_pos = bb;
3764 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3766 /* find last definition of arg value in block */
3771 defs = set_find_def(si->values, value);
3773 if(defs && defs->remats) {
3774 for(next = defs->remats; next; next = get_irn_link(next)) {
3775 if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3776 last = sched_get_time_step(next);
3782 if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3783 last = sched_get_time_step(value);
3787 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
3789 spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3795 phim_fixer(spill_ilp_t *si) {
3798 set_foreach(si->values, defs) {
3799 const ir_node *phi = defs->value;
3800 op_t *op = get_irn_link(phi);
3801 ir_node *phi_m = NULL;
3802 ir_node *next = defs->spills;
3805 if(!is_Phi(phi)) continue;
3808 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3812 next = get_irn_link(next);
3815 if(!phi_m) continue;
3817 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3818 ir_node *value = get_irn_n(phi, n);
3819 defs_t *val_defs = set_find_def(si->values, value);
3821 /* a spill of this value */
3826 ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3827 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3829 if(!is_zero(name->value)) {
3830 spill = insert_mem_copy(si, pred, value);
3832 spill = val_defs->spills;
3835 spill = val_defs->spills;
3838 assert(spill && "no spill placed before PhiM");
3839 set_irn_n(phi_m, n, spill);
3845 walker_reload_placer(ir_node * bb, void * data) {
3846 spill_ilp_t *si = (spill_ilp_t*)data;
3848 spill_bb_t *spill_bb = get_irn_link(bb);
3850 /* reloads at end of block */
3851 if(spill_bb->reloads) {
3854 set_foreach(spill_bb->reloads, keyval) {
3855 ir_node *irn = (ir_node*)keyval->key;
3856 ilp_var_t reload = PTR_TO_INT(keyval->val);
3859 name = si->lpp->vars[reload];
3860 if(!is_zero(name->value)) {
3862 ir_node *insert_pos = bb;
3863 ir_node *prev = sched_block_last_noncf(si, bb);
3864 op_t *prev_op = get_irn_link(prev);
3866 while(be_is_Spill(prev)) {
3867 prev = sched_prev(prev);
3870 prev_op = get_irn_link(prev);
3872 /* insert reload before pre-remats */
3873 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3874 && prev_op->is_remat && prev_op->attr.remat.pre) {
3878 prev = sched_prev(prev);
3879 } while(be_is_Spill(prev));
3881 prev_op = get_irn_link(prev);
3885 reload = insert_reload(si, irn, insert_pos);
3887 if(opt_keep_alive & KEEPALIVE_RELOADS)
3888 pset_insert_ptr(si->spills, reload);
3893 /* walk and insert more reloads and collect remats */
3894 sched_foreach_reverse(bb, irn) {
3895 op_t *op = get_irn_link(irn);
3897 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3898 if(is_Phi(irn)) break;
3901 if(get_irn_mode(irn) != mode_T) {
3902 insert_remat(si, irn);
3907 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3908 ir_node *arg = get_irn_n(irn, n);
3910 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3913 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3914 if(!is_zero(name->value)) {
3916 ir_node *insert_pos = irn;
3917 ir_node *prev = sched_prev(insert_pos);
3920 while(be_is_Spill(prev)) {
3921 prev = sched_prev(prev);
3924 prev_op = get_irn_link(prev);
3926 /* insert reload before pre-remats */
3927 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3928 && prev_op->is_remat && prev_op->attr.remat.pre) {
3932 prev = sched_prev(prev);
3933 } while(be_is_Spill(prev));
3935 prev_op = get_irn_link(prev);
3939 reload = insert_reload(si, arg, insert_pos);
3941 set_irn_n(irn, n, reload);
3943 if(opt_keep_alive & KEEPALIVE_RELOADS)
3944 pset_insert_ptr(si->spills, reload);
3951 del_set(spill_bb->ilp);
3952 if(spill_bb->reloads) del_set(spill_bb->reloads);
3956 walker_collect_used(ir_node * irn, void * data)
3958 bitset_t *used = data;
3960 bitset_set(used, get_irn_idx(irn));
3963 struct kill_helper {
3969 walker_kill_unused(ir_node * bb, void * data)
3971 struct kill_helper *kh = data;
3972 ir_node *bad = get_irg_bad(get_irn_irg(bb));
3976 for(irn=sched_first(bb); !sched_is_end(irn);) {
3977 ir_node *next = sched_next(irn);
3980 if(!bitset_is_set(kh->used, get_irn_idx(irn))) {
3981 if(be_is_Spill(irn) || be_is_Reload(irn)) {
3982 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)));
3984 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3990 set_nodes_block(irn, bad);
3991 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3992 set_irn_n(irn, n, bad);
4000 kill_all_unused_values_in_schedule(spill_ilp_t * si)
4002 struct kill_helper kh;
4004 kh.used = bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
4007 irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
4008 irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
4010 bitset_free(kh.used);
4014 print_irn_pset(pset * p)
4018 pset_foreach(p, irn) {
4019 ir_printf("%+F\n", irn);
4024 dump_phi_class(spill_ilp_t * si, pset * phiclass, const char * file)
4026 FILE *f = fopen(file, "w");
4028 interference_t *interference;
4030 pset_break(phiclass);
4031 set_break(si->interferences);
4033 ir_fprintf(f, "digraph phiclass {\n");
4035 pset_foreach(phiclass, irn) {
4037 ir_fprintf(f, " %F%N [shape=box]\n",irn,irn);
4040 pset_foreach(phiclass, irn) {
4043 if(!is_Phi(irn)) continue;
4045 for(n=get_irn_arity(irn)-1; n>=0; --n) {
4046 ir_node *arg = get_irn_n(irn, n);
4048 ir_fprintf(f, " %F%N -> %F%N\n",irn,irn,arg,arg);
4052 set_foreach(si->interferences, interference) {
4053 const ir_node *a = interference->a;
4054 const ir_node *b = interference->b;
4055 if(get_phi_class(a) == phiclass) {
4056 ir_fprintf(f, " %F%N -> %F%N [color=red,dir=none,style=bold]\n",a,a,b,b);
4065 rewire_uses(spill_ilp_t * si)
4067 dom_front_info_t *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
4069 pset *ignore = pset_new_ptr(1);
4071 pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
4073 /* then fix uses of spills */
4074 set_foreach(si->values, defs) {
4077 const ir_node *next = defs->remats;
4080 reloads = pset_new_ptr_default();
4083 if(be_is_Reload(next)) {
4084 pset_insert_ptr(reloads, next);
4088 next = get_irn_link(next);
4091 spills = get_spills_for_value(si, defs->value);
4092 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));
4093 if(pset_count(spills) > 1) {
4094 //assert(pset_count(reloads) > 0);
4095 // print_irn_pset(spills);
4096 // print_irn_pset(reloads);
4098 be_ssa_constr_set_ignore(dfi, si->lv, spills, ignore);
4105 /* first fix uses of remats and reloads */
4106 set_foreach(si->values, defs) {
4108 const ir_node *next = defs->remats;
4112 nodes = pset_new_ptr_default();
4113 if(sched_is_scheduled(defs->value)) {
4114 pset_insert_ptr(nodes, defs->value);
4119 pset_insert_ptr(nodes, next);
4120 next = get_irn_link(next);
4123 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-orig_kept, defs->value));
4124 be_ssa_constr_set(dfi, si->lv, nodes);
4130 // remove_unused_defs(si);
4132 be_free_dominance_frontiers(dfi);
4137 writeback_results(spill_ilp_t * si)
4139 /* walk through the graph and collect all spills, reloads and remats for a value */
4141 si->values = new_set(cmp_defs, 4096);
4143 DBG((si->dbg, LEVEL_1, "Applying results\n"));
4144 delete_unnecessary_remats(si);
4145 si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
4146 irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
4147 irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
4149 insert_memoperands(si);
4152 /* clean the remat info! there are still back-edges leading there! */
4153 clean_remat_info(si);
4157 connect_all_spills_with_keep(si);
4159 del_set(si->values);
4163 get_n_regs(spill_ilp_t * si)
4165 int arch_n_regs = arch_register_class_n_regs(si->cls);
4167 bitset_t *arch_regs = bitset_malloc(arch_n_regs);
4168 bitset_t *abi_regs = bitset_malloc(arch_n_regs);
4170 arch_put_non_ignore_regs(si->chordal_env->birg->main_env->arch_env, si->cls, arch_regs);
4171 be_abi_put_ignore_regs(si->chordal_env->birg->abi, si->cls, abi_regs);
4173 bitset_andnot(arch_regs, abi_regs);
4174 arch_n_regs = bitset_popcnt(arch_regs);
4176 bitset_free(arch_regs);
4177 bitset_free(abi_regs);
4179 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", arch_n_regs, si->cls->name));
4184 walker_reload_mover(ir_node * bb, void * data)
4186 spill_ilp_t *si = data;
4189 sched_foreach(bb, tmp) {
4190 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
4191 ir_node *reload = tmp;
4194 /* move reload upwards */
4196 int pressure = (int)get_irn_link(reload);
4197 if(pressure < si->n_regs) {
4198 irn = sched_prev(reload);
4199 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
4200 sched_remove(reload);
4201 pressure = (int)get_irn_link(irn);
4203 while(pressure < si->n_regs) {
4204 if( sched_is_end(irn) ||
4205 (be_is_Reload(irn) && has_reg_class(si, irn)) ||
4206 /* do not move reload before its spill */
4207 (irn == be_get_Reload_mem(reload)) ||
4208 /* do not move before phi */
4211 set_irn_link(irn, INT_TO_PTR(pressure+1));
4212 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
4213 irn = sched_prev(irn);
4215 pressure = (int)get_irn_link(irn);
4218 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
4219 sched_put_after(irn, reload);
4226 move_reloads_upward(spill_ilp_t * si)
4228 irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
4233 * Walk all irg blocks and check for interfering spills inside of phi classes
4236 luke_meminterferencechecker(ir_node * bb, void * data)
4238 spill_ilp_t *si = (spill_ilp_t*)data;
4241 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
4242 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
4244 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
4246 /* a is only interesting if it is in my register class and if it is inside a phi class */
4247 if (has_reg_class(si, a) && get_phi_class(a)) {
4248 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)) {
4249 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
4251 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
4253 /* a and b are only interesting if they are in the same phi class */
4254 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
4255 if(values_interfere_in_block(si, bb, a, b)) {
4256 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4265 verify_phiclasses(spill_ilp_t * si)
4267 /* analyze phi classes */
4268 phi_class_compute(si->chordal_env->irg);
4270 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4271 irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
4275 be_spill_remat(const be_chordal_env_t * chordal_env)
4278 char problem_name[256];
4279 char dump_suffix[256];
4280 char dump_suffix2[256];
4281 struct obstack obst;
4284 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
4285 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
4286 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
4288 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4289 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4291 if(opt_verify & VERIFY_DOMINANCE)
4292 be_check_dominance(chordal_env->irg);
4294 obstack_init(&obst);
4295 si.chordal_env = chordal_env;
4297 si.cls = chordal_env->cls;
4298 si.lpp = new_lpp(problem_name, lpp_minimize);
4299 si.remat_info = new_set(cmp_remat_info, 4096);
4300 si.interferences = new_set(cmp_interference, 32);
4301 si.memoperands = new_set(cmp_memoperands, 128);
4302 si.all_possible_remats = pset_new_ptr_default();
4303 si.spills = pset_new_ptr_default();
4304 si.inverse_ops = pset_new_ptr_default();
4305 si.lv = chordal_env->lv;
4307 si.n_regs = get_n_regs(&si);
4309 set_irg_link(chordal_env->irg, &si);
4310 compute_doms(chordal_env->irg);
4312 /* compute phi classes */
4313 // phi_class_compute(chordal_env->irg);
4315 if(opt_dump_flags & DUMP_STATS)
4316 be_analyze_regpressure(chordal_env, "-pre");
4318 DBG((si.dbg, LEVEL_2, "\t initializing\n"));
4319 irg_block_walk_graph(chordal_env->irg, luke_initializer, NULL, &si);
4322 /* collect remats */
4323 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4324 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
4327 /* insert possible remats */
4328 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4329 irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
4330 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4332 if(opt_keep_alive & KEEPALIVE_REMATS) {
4333 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4334 connect_all_remats_with_keep(&si);
4335 /* dump graph with inserted remats */
4336 dump_graph_with_remats(chordal_env->irg, dump_suffix);
4339 /* insert copies for phi arguments not in my regclass */
4340 irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
4342 /* recompute liveness */
4343 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4344 be_liveness_recompute(si.lv);
4347 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4348 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4349 irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
4351 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4352 irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
4355 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4356 memcopyhandler(&si);
4359 if(opt_dump_flags & DUMP_PROBLEM) {
4361 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4362 if ((f = fopen(buf, "wt")) != NULL) {
4363 lpp_dump_plain(si.lpp, f);
4368 if(opt_dump_flags & DUMP_MPS) {
4371 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4372 if((f = fopen(buf, "wt")) != NULL) {
4373 mps_write_mps(si.lpp, s_mps_fixed, f);
4377 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4378 if((f = fopen(buf, "wt")) != NULL) {
4379 mps_write_mst(si.lpp, s_mps_fixed, f);
4384 lpp_check_startvals(si.lpp);
4387 DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4388 lpp_set_time_limit(si.lpp, opt_timeout);
4391 lpp_set_log(si.lpp, stdout);
4394 lpp_solve_cplex(si.lpp);
4396 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4398 assert(lpp_is_sol_valid(si.lpp)
4399 && "solution of ILP must be valid");
4401 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));
4403 if(opt_dump_flags & DUMP_SOLUTION) {
4407 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4408 if ((f = fopen(buf, "wt")) != NULL) {
4410 for (i = 0; i < si.lpp->var_next; ++i) {
4411 lpp_name_t *name = si.lpp->vars[i];
4412 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4418 writeback_results(&si);
4422 kill_all_unused_values_in_schedule(&si);
4424 if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
4425 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
4427 // move reloads upwards
4428 be_liveness_recompute(si.lv);
4429 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4430 move_reloads_upward(&si);
4433 verify_phiclasses(&si);
4436 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4438 if(opt_dump_flags & DUMP_PRESSURE)
4439 dump_pressure_graph(&si, dump_suffix2);
4441 if(opt_dump_flags & DUMP_STATS)
4442 be_analyze_regpressure(chordal_env, "-post");
4444 if(opt_verify & VERIFY_DOMINANCE)
4445 be_check_dominance(chordal_env->irg);
4447 free_dom(chordal_env->irg);
4448 del_set(si.interferences);
4449 del_pset(si.inverse_ops);
4450 del_pset(si.all_possible_remats);
4451 del_set(si.memoperands);
4452 del_pset(si.spills);
4454 obstack_free(&obst, NULL);
4455 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4458 #else /* WITH_ILP */
4461 only_that_you_can_compile_without_WITH_ILP_defined(void)
4465 #endif /* WITH_ILP */