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
36 #include "irbackedge_t.h"
40 #include <lpp/lpp_net.h>
41 #include <lpp/lpp_cplex.h>
45 #include "besched_t.h"
51 #include "bespillremat.h"
53 #include "bepressurestat.h"
54 #include "beprofile.h"
55 #include "bespilloptions.h"
57 #include "bechordal_t.h"
59 #include <libcore/lc_opts.h>
60 #include <libcore/lc_opts_enum.h>
62 #define DUMP_PROBLEM 1
64 #define DUMP_SOLUTION 4
66 #define DUMP_PRESSURE 16
68 #define KEEPALIVE_REMATS 1
69 #define KEEPALIVE_SPILLS 2
70 #define KEEPALIVE_RELOADS 4
72 #define VERIFY_MEMINTERF 1
73 #define VERIFY_DOMINANCE 2
76 #define REMATS_BRIGGS 1
77 #define REMATS_NOINVERSE 2
80 static unsigned opt_dump_flags = 0;
81 static int opt_log = 0;
82 static unsigned opt_keep_alive = 0;
83 static int opt_goodwin = 1;
84 static int opt_memcopies = 1;
85 static int opt_memoperands = 1;
86 static int opt_verify = VERIFY_MEMINTERF;
87 static unsigned opt_remats = REMATS_ALL;
88 static int opt_repair_schedule = 0;
89 static int opt_no_enlarge_liveness = 0;
90 static int opt_remat_while_live = 1;
91 static int opt_timeout = 300;
92 static double opt_cost_reload = 8.0;
93 static double opt_cost_memoperand = 7.0;
94 static double opt_cost_spill = 15.0;
95 static double opt_cost_remat = 1.0;
98 static const lc_opt_enum_mask_items_t dump_items[] = {
99 { "problem", DUMP_PROBLEM },
101 { "solution", DUMP_SOLUTION },
102 { "stats", DUMP_STATS },
103 { "pressure", DUMP_PRESSURE },
107 static lc_opt_enum_mask_var_t dump_var = {
108 &opt_dump_flags, dump_items
111 static const lc_opt_enum_mask_items_t keepalive_items[] = {
112 { "remats", KEEPALIVE_REMATS },
113 { "spills", KEEPALIVE_SPILLS },
114 { "reloads", KEEPALIVE_RELOADS },
118 static lc_opt_enum_mask_var_t keep_alive_var = {
119 &opt_keep_alive, keepalive_items
122 static const lc_opt_enum_mask_items_t remats_items[] = {
123 { "none", REMATS_NONE },
124 { "briggs", REMATS_BRIGGS },
125 { "noinverse", REMATS_NOINVERSE },
126 { "all", REMATS_ALL },
130 static lc_opt_enum_mask_var_t remats_var = {
131 &opt_remats, remats_items
134 static const lc_opt_table_entry_t options[] = {
135 LC_OPT_ENT_ENUM_MASK("keepalive", "keep alive inserted nodes", &keep_alive_var),
137 LC_OPT_ENT_BOOL ("goodwin", "activate goodwin reduction", &opt_goodwin),
138 LC_OPT_ENT_BOOL ("memcopies", "activate memcopy handling", &opt_memcopies),
139 LC_OPT_ENT_BOOL ("memoperands", "activate memoperands", &opt_memoperands),
140 LC_OPT_ENT_ENUM_INT ("remats", "type of remats to insert", &remats_var),
141 LC_OPT_ENT_BOOL ("repair_schedule", "repair the schedule by rematting once used nodes",&opt_repair_schedule),
142 LC_OPT_ENT_BOOL ("no_enlage_liveness", "do not enlarge liveness of operands of remats",&opt_no_enlarge_liveness),
143 LC_OPT_ENT_BOOL ("remat_while_live", "only remat where rematted value was live", &opt_remat_while_live),
145 LC_OPT_ENT_ENUM_MASK("dump", "dump problem, solution or statistical data", &dump_var),
146 LC_OPT_ENT_BOOL ("log", "activate the lpp log", &opt_log),
147 LC_OPT_ENT_INT ("timeout", "ILP solver timeout", &opt_timeout),
149 LC_OPT_ENT_DBL ("cost_reload", "cost of a reload", &opt_cost_reload),
150 LC_OPT_ENT_DBL ("cost_memoperand", "cost of a memory operand", &opt_cost_memoperand),
151 LC_OPT_ENT_DBL ("cost_spill", "cost of a spill instruction", &opt_cost_spill),
152 LC_OPT_ENT_DBL ("cost_remat", "cost of a rematerialization", &opt_cost_remat),
156 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
157 //#define SCHEDULE_PHIM /* insert phim nodes into schedule */
160 //#define SOLVE_LOCAL
161 #define LPP_SERVER "i44pc52"
162 #define LPP_SOLVER "cplex"
165 #define MAX_PATHS INT_MAX
168 typedef struct _spill_ilp_t {
169 const arch_register_class_t *cls;
174 struct obstack *obst;
176 pset *all_possible_remats;
179 set *values; /**< for collecting all definitions of values before running ssa-construction */
185 #ifndef SCHEDULE_PHIM
188 DEBUG_ONLY(firm_dbg_module_t * dbg);
191 typedef int ilp_var_t;
192 typedef int ilp_cst_t;
194 typedef struct _spill_bb_t {
199 typedef struct _remat_t {
200 const ir_node *op; /**< for copy_irn */
201 const ir_node *value; /**< the value which is being recomputed by this remat */
202 const ir_node *proj; /**< not NULL if the above op produces a tuple */
203 int cost; /**< cost of this remat */
204 int inverse; /**< nonzero if this is an inverse remat */
208 * Data to be attached to each IR node. For remats this contains the ilp_var
209 * for this remat and for normal ops this contains the ilp_vars for
210 * reloading each operand
212 typedef struct _op_t {
217 const remat_t *remat; /** the remat this op belongs to */
218 int pre; /** 1, if this is a pressure-increasing remat */
222 ir_node *op; /** the operation this live range belongs to */
231 typedef struct _defs_t {
232 const ir_node *value;
233 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
234 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
237 typedef struct _remat_info_t {
238 const ir_node *irn; /**< the irn to which these remats belong */
239 pset *remats; /**< possible remats for this value */
240 pset *remats_by_operand; /**< remats with this value as operand */
243 typedef struct _keyval_t {
248 typedef struct _spill_t {
257 typedef struct _memoperand_t {
258 ir_node *irn; /**< the irn */
259 unsigned int pos; /**< the position of the argument */
260 ilp_var_t ilp; /**< the ilp var for this memory operand */
264 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
266 return arch_irn_consider_in_reg_alloc(si->birg->main_env->arch_env,
272 cmp_remat(const void *a, const void *b)
274 const keyval_t *p = a;
275 const keyval_t *q = b;
276 const remat_t *r = p->val;
277 const remat_t *s = q->val;
281 return !(r == s || r->op == s->op);
285 cmp_remat(const void *a, const void *b)
287 const remat_t *r = a;
288 const remat_t *s = a;
290 return !(r == s || r->op == s->op);
294 cmp_spill(const void *a, const void *b, size_t size)
296 const spill_t *p = a;
297 const spill_t *q = b;
299 // return !(p->irn == q->irn && p->bb == q->bb);
300 return !(p->irn == q->irn);
304 cmp_memoperands(const void *a, const void *b, size_t size)
306 const memoperand_t *p = a;
307 const memoperand_t *q = b;
309 return !(p->irn == q->irn && p->pos == q->pos);
313 set_find_keyval(set * set, const void * key)
318 return set_find(set, &query, sizeof(query), HASH_PTR(key));
322 set_insert_keyval(set * set, void * key, void * val)
328 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
332 set_find_def(set * set, const ir_node * value)
337 return set_find(set, &query, sizeof(query), HASH_PTR(value));
341 set_insert_def(set * set, const ir_node * value)
348 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
351 static memoperand_t *
352 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
359 return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
362 static memoperand_t *
363 set_find_memoperand(set * set, const ir_node * irn, unsigned int pos)
367 query.irn = (ir_node*)irn;
369 return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
374 set_find_spill(set * set, const ir_node * value)
378 query.irn = (ir_node*)value;
379 return set_find(set, &query, sizeof(query), HASH_PTR(value));
382 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
383 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
384 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
385 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
386 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
389 cmp_remat_info(const void *a, const void *b, size_t size)
391 const remat_info_t *p = a;
392 const remat_info_t *q = b;
394 return !(p->irn == q->irn);
398 cmp_defs(const void *a, const void *b, size_t size)
403 return !(p->value == q->value);
407 cmp_keyval(const void *a, const void *b, size_t size)
409 const keyval_t *p = a;
410 const keyval_t *q = b;
412 return !(p->key == q->key);
416 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
419 if(be_profile_has_data())
420 return ((double)be_profile_get_block_execcount(get_block(irn))) + FUDGE;
422 #ifndef EXECFREQ_LOOPDEPH
423 return get_block_execfreq(si->birg->exec_freq, get_block(irn)) + FUDGE;
426 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
428 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
433 get_cost(const spill_ilp_t * si, const ir_node * irn)
435 if(be_is_Spill(irn)) {
436 return opt_cost_spill;
437 } else if(be_is_Reload(irn)){
438 return opt_cost_reload;
440 return arch_get_op_estimated_cost(si->birg->main_env->arch_env, irn);
445 * Checks, whether node and its operands have suitable reg classes
448 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
451 const arch_env_t *arch_env = si->birg->main_env->arch_env;
452 int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
456 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
459 for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
460 ir_node *op = get_irn_n(irn, n);
461 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
464 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
471 * Try to create a remat from @p op with destination value @p dest_value
473 static INLINE remat_t *
474 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
476 remat_t *remat = NULL;
478 // if(!mode_is_datab(get_irn_mode(dest_value)))
481 if(dest_value == op) {
482 const ir_node *proj = NULL;
484 if(is_Proj(dest_value)) {
485 op = get_Proj_pred(op);
489 if(!is_rematerializable(si, op))
492 remat = obstack_alloc(si->obst, sizeof(*remat));
494 remat->cost = (int)get_cost(si, op);
495 remat->value = dest_value;
499 arch_inverse_t inverse;
502 /* get the index of the operand we want to retrieve by the inverse op */
503 for (n = get_irn_arity(op)-1; n>=0; --n) {
504 ir_node *arg = get_irn_n(op, n);
506 if(arg == dest_value) break;
510 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", n, op));
512 /* else ask the backend to give an inverse op */
513 if(arch_get_inverse(si->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
516 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
518 assert(inverse.n > 0 && "inverse op should have at least one node");
520 for(i=inverse.n-1; i>=0; --i) {
521 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
525 remat = obstack_alloc(si->obst, sizeof(*remat));
526 remat->op = inverse.nodes[0];
527 remat->cost = inverse.costs;
528 remat->value = dest_value;
529 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
532 // Matze: commented this out, this doesn't seem to be true if
533 // the inverse is a simple operation with only 1 result...
534 //assert(is_Proj(remat->proj));
536 assert(0 && "I can not handle remats with more than 2 nodes");
543 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
545 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
553 add_remat(const spill_ilp_t * si, const remat_t * remat)
555 remat_info_t *remat_info,
560 assert(remat->value);
562 query.irn = remat->value;
564 query.remats_by_operand = NULL;
565 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
567 if(remat_info->remats == NULL) {
568 remat_info->remats = new_pset(cmp_remat, 4096);
570 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
572 /* insert the remat into the remats_be_operand set of each argument of the remat op */
573 for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
574 ir_node *arg = get_irn_n(remat->op, n);
578 query.remats_by_operand = NULL;
579 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
581 if(remat_info->remats_by_operand == NULL) {
582 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
584 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
589 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
591 const ir_edge_t *edge = get_irn_out_edge_first(irn);
595 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
598 edge = get_irn_out_edge_next(irn, edge);
605 get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
611 irn = get_Proj_pred(irn);
613 for(n=get_irn_arity(irn)-1; n>=0; --n) {
614 const ir_node *arg = get_irn_n(irn, n);
616 if(has_reg_class(si, arg)) ++ret;
623 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
628 if( has_reg_class(si, op)
629 && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1)
630 && (opt_remats != REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0)
632 remat = get_remat_from_op(si, op, op);
634 add_remat(si, remat);
638 if(opt_remats == REMATS_ALL) {
639 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
641 for (n = get_irn_arity(op)-1; n>=0; --n) {
642 ir_node *arg = get_irn_n(op, n);
644 if(has_reg_class(si, arg)) {
645 /* try to get an inverse remat */
646 remat = get_remat_from_op(si, arg, op);
648 add_remat(si, remat);
656 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
659 ir_node *def_block = get_nodes_block(val);
665 /* if pos is at end of a basic block */
667 ret = (pos == def_block || block_dominates(def_block, pos));
668 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
672 /* else if this is a normal operation */
673 block = get_nodes_block(pos);
674 if(block == def_block) {
675 if(!sched_is_scheduled(val)) return 1;
677 ret = sched_comes_after(val, pos);
678 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
682 ret = block_dominates(def_block, block);
683 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
687 static INLINE ir_node *
688 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
690 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->birg->main_env->arch_env);
694 * Returns first non-Phi node of block @p bb
696 static INLINE ir_node *
697 sched_block_first_nonphi(const ir_node * bb)
699 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
703 sched_skip_proj_predicator(const ir_node * irn, void * data)
705 return (is_Proj(irn));
708 static INLINE ir_node *
709 sched_next_nonproj(const ir_node * irn, int forward)
711 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
715 * Returns next operation node (non-Proj) after @p irn
716 * or the basic block of this node
718 static INLINE ir_node *
719 sched_next_op(const ir_node * irn)
721 ir_node *next = sched_next(irn);
726 return sched_next_nonproj(next, 1);
730 * Returns previous operation node (non-Proj) before @p irn
731 * or the basic block of this node
733 static INLINE ir_node *
734 sched_prev_op(const ir_node * irn)
736 ir_node *prev = sched_prev(irn);
741 return sched_next_nonproj(prev, 0);
745 sched_put_after(ir_node * insert, ir_node * irn)
747 if(is_Block(insert)) {
748 insert = sched_block_first_nonphi(insert);
750 insert = sched_next_op(insert);
753 sched_add_before(insert, irn);
757 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
759 if(is_Block(insert)) {
760 insert = sched_block_last_noncf(si, insert);
762 insert = sched_next_nonproj(insert, 0);
763 insert = sched_prev(insert);
766 sched_add_after(insert, irn);
770 next_post_remat(const ir_node * irn)
776 next = sched_block_first_nonphi(irn);
778 next = sched_next_op(irn);
781 if(sched_is_end(next))
784 op = get_irn_link(next);
785 if(op->is_remat && !op->attr.remat.pre) {
794 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
800 ret = sched_block_last_noncf(si, irn);
801 ret = sched_next(ret);
802 ret = sched_prev_op(ret);
804 ret = sched_prev_op(irn);
807 if(sched_is_end(ret) || is_Phi(ret))
810 op = (op_t*)get_irn_link(ret);
811 if(op->is_remat && op->attr.remat.pre) {
819 * Tells you whether a @p remat can be placed before the irn @p pos
822 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
824 const ir_node *op = remat->op;
830 prev = sched_block_last_noncf(si, pos);
831 prev = sched_next_nonproj(prev, 0);
833 prev = sched_prev_op(pos);
835 /* do not remat if the rematted value is defined immediately before this op */
836 if(prev == remat->op) {
841 /* this should be just fine, the following OP will be using this value, right? */
843 /* only remat AFTER the real definition of a value (?) */
844 if(!value_is_defined_before(si, pos, remat->value)) {
845 // ir_fprintf(stderr, "error(not defined)");
850 for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
851 const ir_node *arg = get_irn_n(op, n);
853 if(opt_no_enlarge_liveness) {
854 if(has_reg_class(si, arg) && live) {
855 res &= pset_find_ptr((pset*)live, arg)?1:0;
857 res &= value_is_defined_before(si, pos, arg);
860 res &= value_is_defined_before(si, pos, arg);
868 * Tells you whether a @p remat can be placed after the irn @p pos
871 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
874 pos = sched_block_first_nonphi(pos);
876 pos = sched_next_op(pos);
879 /* only remat AFTER the real definition of a value (?) */
880 if(!value_is_defined_before(si, pos, remat->value)) {
884 return can_remat_before(si, remat, pos, live);
888 * Collect potetially rematerializable OPs
891 walker_remat_collector(ir_node * irn, void * data)
893 spill_ilp_t *si = data;
895 if(!is_Block(irn) && !is_Phi(irn)) {
896 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
897 get_remats_from_op(si, irn);
902 * Inserts a copy of @p irn before @p pos
905 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
910 bb = is_Block(pos)?pos:get_nodes_block(pos);
911 copy = exact_copy(irn);
913 set_phi_class(si->pc, copy, NULL);
914 set_nodes_block(copy, bb);
915 sched_put_before(si, pos, copy);
921 * Inserts a copy of @p irn after @p pos
924 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
929 bb = is_Block(pos)?pos:get_nodes_block(pos);
930 copy = exact_copy(irn);
932 set_phi_class(si->pc, copy, NULL);
933 set_nodes_block(copy, bb);
934 sched_put_after(pos, copy);
940 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
944 if(can_remat_after(si, remat, pos, live)) {
949 DBG((si->dbg, LEVEL_3, "\t >inserting remat2 %+F\n", remat->op));
951 copy = insert_copy_after(si, remat->op, pos);
953 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
954 op = obstack_alloc(si->obst, sizeof(*op));
956 op->attr.remat.remat = remat;
957 op->attr.remat.pre = 0;
958 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
960 set_irn_link(copy, op);
961 pset_insert_ptr(si->all_possible_remats, copy);
963 proj_copy = insert_copy_after(si, remat->proj, copy);
964 set_irn_n(proj_copy, 0, copy);
965 set_irn_link(proj_copy, op);
966 pset_insert_ptr(si->all_possible_remats, proj_copy);
978 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
982 if(can_remat_before(si, remat, pos, live)) {
987 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
989 copy = insert_copy_before(si, remat->op, pos);
991 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
992 op = obstack_alloc(si->obst, sizeof(*op));
994 op->attr.remat.remat = remat;
995 op->attr.remat.pre = 1;
996 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
998 set_irn_link(copy, op);
999 pset_insert_ptr(si->all_possible_remats, copy);
1001 proj_copy = insert_copy_after(si, remat->proj, copy);
1002 set_irn_n(proj_copy, 0, copy);
1003 set_irn_link(proj_copy, op);
1004 pset_insert_ptr(si->all_possible_remats, proj_copy);
1016 get_block_n_succs(const ir_node *block) {
1017 const ir_edge_t *edge;
1019 assert(edges_activated(current_ir_graph));
1021 edge = get_block_succ_first(block);
1025 edge = get_block_succ_next(block, edge);
1026 return edge ? 2 : 1;
1030 is_start_block(const ir_node * bb)
1032 return get_irg_start_block(get_irn_irg(bb)) == bb;
1036 is_merge_edge(const ir_node * bb)
1038 if(is_start_block(bb))
1042 return get_block_n_succs(bb) == 1;
1048 is_diverge_edge(const ir_node * bb)
1050 if(is_start_block(bb))
1054 return get_Block_n_cfgpreds(bb) == 1;
1060 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1065 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1066 irn = be_lv_get_irn(si->lv, bb, i);
1068 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1069 pset_insert_ptr(live, irn);
1073 irn = sched_last(bb);
1075 /* all values eaten by control flow operations are also live until the end of the block */
1076 sched_foreach_reverse(bb, irn) {
1079 if(!sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) break;
1081 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1082 ir_node *arg = get_irn_n(irn,i);
1084 if(has_reg_class(si, arg)) {
1085 pset_insert_ptr(live, arg);
1090 * find values that are used by remats at end of block
1091 * and insert them into live set
1093 foreach_pre_remat(si, bb, irn) {
1096 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1097 ir_node *remat_arg = get_irn_n(irn, n);
1099 if(!has_reg_class(si, remat_arg)) continue;
1101 /* if value is becoming live through use by remat */
1102 if(!pset_find_ptr(live, remat_arg)) {
1103 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, irn));
1105 pset_insert_ptr(live, remat_arg);
1112 walker_regclass_copy_insertor(ir_node * irn, void * data)
1114 spill_ilp_t *si = data;
1116 if(is_Phi(irn) && has_reg_class(si, irn)) {
1119 for(n=get_irn_arity(irn)-1; n>=0; --n) {
1120 ir_node *phi_arg = get_irn_n(irn, n);
1121 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
1123 if(!has_reg_class(si, phi_arg)) {
1124 ir_node *copy = be_new_Copy(si->cls, si->birg->irg, bb, phi_arg);
1125 ir_node *pos = sched_block_last_noncf(si, bb);
1126 op_t *op = obstack_alloc(si->obst, sizeof(*op));
1128 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
1129 sched_add_after(pos, copy);
1130 set_irn_n(irn, n, copy);
1133 op->attr.live_range.args.reloads = NULL;
1134 op->attr.live_range.ilp = ILP_UNDEF;
1135 set_irn_link(copy, op);
1142 * Insert (so far unused) remats into the irg to
1143 * recompute the potential liveness of all values
1146 walker_remat_insertor(ir_node * bb, void * data)
1148 spill_ilp_t *si = data;
1155 /* skip start block, no remats to do there */
1156 if(is_start_block(bb)) return;
1158 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1160 live = pset_new_ptr_default();
1161 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1162 ir_node *value = be_lv_get_irn(si->lv, bb, i);
1164 /* add remats at end of block */
1165 if (has_reg_class(si, value)) {
1166 pset_insert_ptr(live, value);
1170 irn = sched_last(bb);
1171 while(!sched_is_end(irn)) {
1177 next = sched_prev(irn);
1179 /* delete defined value from live set */
1180 if(has_reg_class(si, irn)) {
1181 pset_remove_ptr(live, irn);
1184 if(is_Phi(irn) || is_Proj(irn)) {
1189 args = pset_new_ptr_default();
1190 used = pset_new_ptr_default();
1192 /* collect arguments of op and set args of op already live in epilog */
1193 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1194 ir_node *arg = get_irn_n(irn, n);
1196 pset_insert_ptr(args, arg);
1197 if(has_reg_class(si, arg)) {
1198 pset_insert_ptr(live, arg);
1199 pset_insert_ptr(used, arg);
1203 /* insert all possible remats before irn */
1204 pset_foreach(args, arg) {
1205 remat_info_t *remat_info,
1208 /* continue if the operand has the wrong reg class */
1209 if(!has_reg_class(si, arg))
1213 query.remats = NULL;
1214 query.remats_by_operand = NULL;
1215 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1221 if(remat_info->remats) {
1222 pset_foreach(remat_info->remats, remat) {
1223 ir_node *remat_irn = NULL;
1225 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
1226 remat_irn = insert_remat_before(si, remat, irn, live);
1229 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1230 ir_node *remat_arg = get_irn_n(remat_irn, n);
1232 /* collect args of remats which are not args of op */
1233 if(has_reg_class(si, remat_arg) && !pset_find_ptr(args, remat_arg)) {
1234 pset_insert_ptr(used, remat_arg);
1242 /* do not place post remats after jumps */
1243 if(sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) {
1249 /* insert all possible remats after irn */
1250 post_remats = pset_new_ptr_default();
1251 pset_foreach(used, arg) {
1252 remat_info_t *remat_info,
1255 /* continue if the operand has the wrong reg class */
1256 if(!has_reg_class(si, arg))
1260 query.remats = NULL;
1261 query.remats_by_operand = NULL;
1262 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1268 if(remat_info->remats_by_operand) {
1269 pset_foreach(remat_info->remats_by_operand, remat) {
1270 /* do not insert remats producing the same value as one of the operands */
1271 if(!pset_find_ptr(args, remat->value)) {
1272 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
1274 /* only remat values that can be used by real ops */
1275 if(!opt_remat_while_live || pset_find_ptr(live, remat->value)) {
1276 pset_insert_ptr(post_remats, remat);
1282 pset_foreach(post_remats, remat) {
1283 insert_remat_after(si, remat, irn, live);
1285 del_pset(post_remats);
1292 /* add remats at end if successor has multiple predecessors */
1293 if(is_merge_edge(bb)) {
1294 pset *live_out = pset_new_ptr_default();
1297 get_live_end(si, bb, live_out);
1299 /* add remats at end of block */
1300 pset_foreach(live_out, value) {
1301 remat_info_t *remat_info,
1305 query.remats = NULL;
1306 query.remats_by_operand = NULL;
1307 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1309 if(remat_info && remat_info->remats) {
1310 pset_foreach(remat_info->remats, remat) {
1311 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1313 insert_remat_before(si, remat, bb, live_out);
1320 if(is_diverge_edge(bb)) {
1321 pset *live_in = pset_new_ptr_default();
1324 be_lv_foreach(si->lv, bb, be_lv_state_in, i) {
1325 value = be_lv_get_irn(si->lv, bb, i);
1327 if(has_reg_class(si, value)) {
1328 pset_insert_ptr(live_in, value);
1331 /* add phis to live_in */
1332 sched_foreach(bb, value) {
1333 if(!is_Phi(value)) break;
1335 if(has_reg_class(si, value)) {
1336 pset_insert_ptr(live_in, value);
1340 /* add remat2s at beginning of block */
1341 post_remats = pset_new_ptr_default();
1342 pset_foreach(live_in, value) {
1343 remat_info_t *remat_info,
1347 query.remats = NULL;
1348 query.remats_by_operand = NULL;
1349 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1351 if(remat_info && remat_info->remats_by_operand) {
1352 pset_foreach(remat_info->remats_by_operand, remat) {
1353 DBG((si->dbg, LEVEL_4, "\t considering remat2 %+F at beginning of block %+F\n", remat->op, bb));
1355 /* put the remat here if all its args are available and result is still live */
1356 if(!opt_remat_while_live || pset_find_ptr(live_in, remat->value)) {
1357 pset_insert_ptr(post_remats, remat);
1362 pset_foreach(post_remats, remat) {
1363 insert_remat_after(si, remat, bb, live_in);
1365 del_pset(post_remats);
1371 can_be_copied(const ir_node * bb, const ir_node * irn)
1373 const ir_edge_t *edge = get_block_succ_first(bb);
1374 const ir_node *next_bb = edge->src;
1375 int pos = edge->pos;
1378 assert(is_merge_edge(bb));
1380 sched_foreach(next_bb, phi) {
1381 const ir_node *phi_arg;
1383 if(!is_Phi(phi)) break;
1385 phi_arg = get_irn_n(phi, pos);
1387 if(phi_arg == irn) {
1395 * Initialize additional node info
1398 luke_initializer(ir_node * bb, void * data)
1400 spill_ilp_t *si = (spill_ilp_t*)data;
1401 spill_bb_t *spill_bb;
1404 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1405 set_irn_link(bb, spill_bb);
1407 sched_foreach(bb, irn) {
1410 op = obstack_alloc(si->obst, sizeof(*op));
1412 op->attr.live_range.ilp = ILP_UNDEF;
1415 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1416 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1418 } else if(!is_Proj(irn)) {
1419 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1420 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1422 op->attr.live_range.args.reloads = NULL;
1424 set_irn_link(irn, op);
1430 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1433 luke_endwalker(ir_node * bb, void * data)
1435 spill_ilp_t *si = (spill_ilp_t*)data;
1441 spill_bb_t *spill_bb = get_irn_link(bb);
1444 live = pset_new_ptr_default();
1445 use_end = pset_new_ptr_default();
1447 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1448 irn = be_lv_get_irn(si->lv, bb, i);
1449 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1450 pset_insert_ptr(live, irn);
1454 * find values that are used by remats at end of block
1455 * and insert them into live set
1457 foreach_pre_remat(si, bb, irn) {
1460 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1461 ir_node *remat_arg = get_irn_n(irn, n);
1463 if(has_reg_class(si, remat_arg)) {
1464 pset_insert_ptr(live, remat_arg);
1469 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1470 /* their reg_out must always be set */
1471 sched_foreach_reverse(bb, irn) {
1474 if(!sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) break;
1476 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1477 ir_node *irn_arg = get_irn_n(irn, n);
1479 if(has_reg_class(si, irn_arg)) {
1480 pset_insert_ptr(use_end, irn_arg);
1485 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1486 //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1487 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1489 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1491 /* if this is a merge edge we can reload at the end of this block */
1492 if(is_merge_edge(bb)) {
1493 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1494 } else if(pset_count(use_end)){
1495 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1497 spill_bb->reloads = NULL;
1500 pset_foreach(live,irn) {
1504 int default_spilled;
1507 /* handle values used by control flow nodes later separately */
1508 if(pset_find_ptr(use_end, irn)) continue;
1511 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1513 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1515 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1516 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1517 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1519 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1520 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1522 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1523 /* by default spill value right after definition */
1524 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1525 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1527 if(is_merge_edge(bb)) {
1531 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1532 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), can_be_copied(bb, irn));
1533 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1535 /* reload <= mem_out */
1536 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1537 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1538 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1541 spill->reg_in = ILP_UNDEF;
1542 spill->mem_in = ILP_UNDEF;
1545 pset_foreach(use_end,irn) {
1549 ilp_cst_t end_use_req,
1552 int default_spilled;
1555 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1557 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1559 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1560 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1562 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1563 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1565 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1566 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1567 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1569 /* reload for use be control flow op */
1570 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1571 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1572 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1574 /* reload <= mem_out */
1575 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1576 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1577 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1579 spill->reg_in = ILP_UNDEF;
1580 spill->mem_in = ILP_UNDEF;
1582 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1583 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1584 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1593 * Find a remat of value @p value in the epilog of @p pos
1596 find_post_remat(const ir_node * value, const ir_node * pos)
1598 while((pos = next_post_remat(pos)) != NULL) {
1601 op = get_irn_link(pos);
1602 assert(op->is_remat && !op->attr.remat.pre);
1604 if(op->attr.remat.remat->value == value)
1605 return (ir_node*)pos;
1608 const ir_edge_t *edge;
1609 foreach_out_edge(pos, edge) {
1610 ir_node *proj = get_edge_src_irn(edge);
1611 assert(is_Proj(proj));
1622 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1624 spill_bb_t *spill_bb = get_irn_link(bb);
1628 int default_spilled;
1631 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1633 double spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1635 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1637 spill->reg_out = ILP_UNDEF;
1638 spill->reg_in = ILP_UNDEF;
1639 spill->mem_in = ILP_UNDEF;
1641 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1642 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1644 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1645 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1646 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1653 * Inserts ILP-constraints and variables for memory copying before the given position
1656 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1658 const ir_node *succ;
1659 const ir_edge_t *edge;
1660 spill_bb_t *spill_bb = get_irn_link(block);
1669 assert(edges_activated(current_ir_graph));
1671 edge = get_block_succ_first(block);
1677 edge = get_block_succ_next(block, edge);
1678 /* next block can only contain phis, if this is a merge edge */
1681 ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1682 copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1684 ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1685 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1687 pset_foreach(live, tmp) {
1690 op_t *op = get_irn_link(irn);
1691 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1693 spill = set_find_spill(spill_bb->ilp, tmp);
1696 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1698 lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1700 sched_foreach(succ, phi) {
1701 const ir_node *to_copy;
1703 spill_t *to_copy_spill;
1704 op_t *phi_op = get_irn_link(phi);
1705 ilp_var_t reload = ILP_UNDEF;
1708 if(!is_Phi(phi)) break;
1709 if(!has_reg_class(si, phi)) continue;
1711 to_copy = get_irn_n(phi, pos);
1712 to_copy_op = get_irn_link(to_copy);
1714 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1715 assert(to_copy_spill);
1717 if(spill_bb->reloads) {
1718 keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1721 reload = PTR_TO_INT(keyval->val);
1725 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1726 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1728 /* copy - reg_out - reload - remat - live_range <= 0 */
1729 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1730 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1731 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1732 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1733 foreach_pre_remat(si, block, tmp) {
1734 op_t *remat_op = get_irn_link(tmp);
1735 if(remat_op->attr.remat.remat->value == to_copy) {
1736 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1740 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1741 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1743 /* copy - reg_out - copyreg <= 0 */
1744 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1745 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1746 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1752 * Walk all irg blocks and emit this ILP
1755 luke_blockwalker(ir_node * bb, void * data)
1757 spill_ilp_t *si = (spill_ilp_t*)data;
1762 spill_bb_t *spill_bb = get_irn_link(bb);
1765 pset *defs = pset_new_ptr_default();
1766 const arch_env_t *arch_env = si->birg->main_env->arch_env;
1768 live = pset_new_ptr_default();
1770 /****************************************
1771 * B A S I C B L O C K E N D
1772 ***************************************/
1775 /* init live values at end of block */
1776 get_live_end(si, bb, live);
1778 pset_foreach(live, irn) {
1780 ilp_var_t reload = ILP_UNDEF;
1782 spill = set_find_spill(spill_bb->ilp, irn);
1785 if(spill_bb->reloads) {
1786 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1789 reload = PTR_TO_INT(keyval->val);
1793 op = get_irn_link(irn);
1794 assert(!op->is_remat);
1796 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1797 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1798 op->attr.live_range.op = bb;
1800 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1801 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1803 /* reg_out - reload - remat - live_range <= 0 */
1804 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1805 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1806 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1807 foreach_pre_remat(si, bb, tmp) {
1808 op_t *remat_op = get_irn_link(tmp);
1809 if(remat_op->attr.remat.remat->value == irn) {
1810 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1813 ir_snprintf(buf, sizeof(buf), "reg_out2_%N_%N", bb, irn);
1814 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
1816 /* value may only die at bb end if it is used for a mem copy */
1817 /* reg_out + \sum copy - reload - remat - live_range >= 0 */
1818 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1819 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1820 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1821 foreach_pre_remat(si, bb, tmp) {
1822 op_t *remat_op = get_irn_link(tmp);
1823 if(remat_op->attr.remat.remat->value == irn) {
1824 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1827 if(is_merge_edge(bb)) {
1828 const ir_edge_t *edge = get_block_succ_first(bb);
1829 const ir_node *next_bb = edge->src;
1830 int pos = edge->pos;
1833 sched_foreach(next_bb, phi) {
1834 const ir_node *phi_arg;
1836 if(!is_Phi(phi)) break;
1838 phi_arg = get_irn_n(phi, pos);
1840 if(phi_arg == irn) {
1841 op_t *phi_op = get_irn_link(phi);
1842 ilp_var_t copy = phi_op->attr.live_range.args.copies[pos];
1844 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
1851 insert_mem_copy_position(si, live, bb);
1854 * assure the remat args are available
1856 foreach_pre_remat(si, bb, tmp) {
1857 op_t *remat_op = get_irn_link(tmp);
1860 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1861 ir_node *remat_arg = get_irn_n(tmp, n);
1862 op_t *arg_op = get_irn_link(remat_arg);
1864 if(!has_reg_class(si, remat_arg)) continue;
1866 spill = set_find_spill(spill_bb->ilp, remat_arg);
1869 /* arguments of remats have to be live until the very end of the block
1870 * remat = reg_out(remat_arg) and (reload(remat_arg) or live_range(remat_arg)),
1871 * no remats, they could be in wrong order
1874 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1875 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1877 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 3.0);
1878 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, -2.0);
1879 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1881 /* use reload placed for this argument */
1882 if(spill_bb->reloads) {
1883 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1886 ilp_var_t reload = PTR_TO_INT(keyval->val);
1888 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1893 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1898 /**************************************
1899 * B A S I C B L O C K B O D Y
1900 **************************************/
1902 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1908 ilp_cst_t check_pre,
1914 ilp_cst_t one_memoperand = -1;
1916 /* iterate only until first phi */
1920 op = get_irn_link(irn);
1922 if(op->is_remat) continue;
1924 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1926 /* collect defined values */
1927 if(has_reg_class(si, irn)) {
1928 pset_insert_ptr(defs, irn);
1932 if(is_Proj(irn)) continue;
1935 * init set of irn's arguments
1936 * and all possibly used values around this op
1937 * and values defined by post remats
1939 args = new_set(cmp_keyval, get_irn_arity(irn));
1940 used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1941 remat_defs = pset_new_ptr(pset_count(live));
1943 if(!is_start_block(bb) || !be_is_Barrier(irn)) {
1944 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1945 ir_node *irn_arg = get_irn_n(irn, n);
1946 if(has_reg_class(si, irn_arg)) {
1947 set_insert_keyval(args, irn_arg, (void*)n);
1948 pset_insert_ptr(used, irn_arg);
1951 foreach_post_remat(irn, tmp) {
1952 op_t *remat_op = get_irn_link(tmp);
1954 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1956 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1957 ir_node *remat_arg = get_irn_n(tmp, n);
1958 if(has_reg_class(si, remat_arg)) {
1959 pset_insert_ptr(used, remat_arg);
1963 foreach_pre_remat(si, irn, tmp) {
1964 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1965 ir_node *remat_arg = get_irn_n(tmp, n);
1966 if(has_reg_class(si, remat_arg)) {
1967 pset_insert_ptr(used, remat_arg);
1973 /**********************************
1974 * I N E P I L O G O F irn
1975 **********************************/
1977 /* ensure each dying value is used by only one post remat */
1978 pset_foreach(used, tmp) {
1979 ir_node *value = tmp;
1980 op_t *value_op = get_irn_link(value);
1985 foreach_post_remat(irn, remat) {
1986 op_t *remat_op = get_irn_link(remat);
1988 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1989 ir_node *remat_arg = get_irn_n(remat, n);
1991 /* if value is used by this remat add it to constraint */
1992 if(remat_arg == value) {
1994 /* sum remat2s <= 1 + n_remats*live_range */
1995 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1996 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2000 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2006 if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
2007 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
2011 /* ensure at least one value dies at post remat */
2012 foreach_post_remat(irn, tmp) {
2013 op_t *remat_op = get_irn_link(tmp);
2014 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
2017 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2018 remat_arg = get_irn_n(tmp, n);
2020 if(has_reg_class(si, remat_arg)) {
2022 /* does arg always die at this op? */
2023 if(!pset_find_ptr(live, remat_arg))
2024 goto skip_one_must_die;
2026 pset_insert_ptr(remat_args, remat_arg);
2030 /* remat + \sum live_range(remat_arg) <= |args| */
2031 ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
2032 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2033 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2035 pset_foreach(remat_args, remat_arg) {
2036 op_t *arg_op = get_irn_link(remat_arg);
2038 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2042 del_pset(remat_args);
2045 /* new live ranges for values from L\U defined by post remats */
2046 pset_foreach(live, tmp) {
2047 ir_node *value = tmp;
2048 op_t *value_op = get_irn_link(value);
2050 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
2051 ilp_var_t prev_lr = ILP_UNDEF;
2054 if(pset_find_ptr(remat_defs, value)) {
2056 /* next_live_range <= prev_live_range + sum remat2s */
2057 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
2058 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2060 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
2061 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2063 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
2064 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2066 foreach_post_remat(irn, remat) {
2067 op_t *remat_op = get_irn_link(remat);
2069 /* if value is being rematerialized by this remat */
2070 if(value == remat_op->attr.remat.remat->value) {
2071 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2075 value_op->attr.live_range.ilp = prev_lr;
2076 value_op->attr.live_range.op = irn;
2081 /* requirements for post remats and start live ranges from L/U' for values dying here */
2082 foreach_post_remat(irn, tmp) {
2083 op_t *remat_op = get_irn_link(tmp);
2086 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2087 ir_node *remat_arg = get_irn_n(tmp, n);
2088 op_t *arg_op = get_irn_link(remat_arg);
2090 if(!has_reg_class(si, remat_arg)) continue;
2092 /* only for values in L\U (TODO and D?), the others are handled with post_use */
2093 if(!pset_find_ptr(used, remat_arg)) {
2094 /* remat <= live_range(remat_arg) */
2095 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
2096 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2098 /* if value is becoming live through use by remat2 */
2099 if(!pset_find_ptr(live, remat_arg)) {
2102 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
2103 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2105 arg_op->attr.live_range.ilp = lr;
2106 arg_op->attr.live_range.op = irn;
2108 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
2110 pset_insert_ptr(live, remat_arg);
2111 add_to_spill_bb(si, bb, remat_arg);
2114 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2115 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2120 d = pset_count(defs);
2121 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
2123 /* count how many regs irn needs for arguments */
2124 u = set_count(args);
2127 /* check the register pressure in the epilog */
2128 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
2129 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
2130 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
2132 /* add L\U' to check_post */
2133 pset_foreach(live, tmp) {
2134 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
2135 /* if a live value is not used by irn */
2136 tmp_op = get_irn_link(tmp);
2137 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
2141 /***********************************************************
2142 * 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
2143 **********************************************************/
2146 pset_foreach(used, tmp) {
2152 op_t *arg_op = get_irn_link(arg);
2155 spill = add_to_spill_bb(si, bb, arg);
2157 /* new live range for each used value */
2158 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2159 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2161 /* the epilog stuff - including post_use, check_post, check_post_remat */
2162 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2163 post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2165 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2167 /* arg is live throughout epilog if the next live_range is in a register */
2168 if(pset_find_ptr(live, arg)) {
2169 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
2171 /* post_use >= next_lr + remat */
2172 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2173 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2174 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2175 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2178 /* forall post remat which use arg add a similar cst */
2179 foreach_post_remat(irn, remat) {
2182 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2183 ir_node *remat_arg = get_irn_n(remat, n);
2184 op_t *remat_op = get_irn_link(remat);
2186 if(remat_arg == arg) {
2187 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
2189 /* post_use >= remat */
2190 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2191 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2192 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2193 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2198 /* if value is not an arg of op and not possibly defined by post remat
2199 * then it may only die and not become live
2201 if(!set_find_keyval(args, arg)) {
2202 /* post_use <= prev_lr */
2203 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2204 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2205 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2206 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2208 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2209 /* next_lr <= prev_lr */
2210 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2211 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2212 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2213 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2217 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2218 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2219 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2220 ilp_var_t memoperand;
2222 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2223 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0);
2224 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2226 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2227 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2229 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2230 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2235 /* new live range begins for each used value */
2236 arg_op->attr.live_range.ilp = prev_lr;
2237 arg_op->attr.live_range.op = irn;
2239 pset_insert_ptr(live, arg);
2242 /* just to be sure */
2243 check_post = ILP_UNDEF;
2245 /* allow original defintions to be removed */
2246 if(opt_repair_schedule) {
2247 pset_foreach(defs, tmp) {
2248 op_t *tmp_op = get_irn_link(tmp);
2249 spill_t *spill = set_find_spill(spill_bb->ilp, tmp);
2254 ir_snprintf(buf, sizeof(buf), "delete_%N", tmp);
2255 delete = lpp_add_var_default(si->lpp, buf, lpp_binary, -1.0*get_cost(si, irn)*execution_frequency(si, bb), 0.0);
2257 /* op may not be killed if its first live_range is 1 */
2258 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2259 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2260 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2261 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2263 /* op may not be killed if it is spilled after the definition */
2264 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2265 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2266 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2267 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2272 ir_snprintf(buf, sizeof(buf), "keep_%N", tmp);
2273 keep = lpp_add_var_default(si->lpp, buf, lpp_binary, get_cost(si, irn)*execution_frequency(si, bb), 1.0);
2275 /* op may not be killed if its first live_range is 1 */
2276 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2277 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2278 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2279 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, -1.0);
2281 /* op may not be killed if it is spilled after the definition */
2282 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2283 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2284 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2285 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2290 pset_foreach(defs, tmp) {
2291 op_t *tmp_op = get_irn_link(tmp);
2292 spill_t *spill = set_find_spill(spill_bb->ilp, tmp);
2295 /* live_range or spill should be 1
2296 TODO: lr should be live until first use */
2297 ir_snprintf(buf, sizeof(buf), "nokillorig_%N", tmp);
2298 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 1.0);
2299 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2300 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2310 /* check the register pressure in the prolog */
2311 /* sum_{L\U} lr <= k - |U| */
2312 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2313 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2315 /* for the prolog remove defined values from the live set */
2316 pset_foreach(defs, tmp) {
2317 pset_remove_ptr(live, tmp);
2320 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2321 ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2322 one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2325 /***********************************************************
2326 * 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
2327 **********************************************************/
2330 set_foreach(args, keyval) {
2332 const ir_node *arg = keyval->key;
2333 int i = PTR_TO_INT(keyval->val);
2334 op_t *arg_op = get_irn_link(arg);
2335 ilp_cst_t requirements;
2338 spill = set_find_spill(spill_bb->ilp, arg);
2341 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2342 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);
2344 /* reload <= mem_out */
2345 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2346 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2347 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2348 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2350 /* requirement: arg must be in register for use */
2351 /* reload + remat + live_range == 1 */
2352 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2353 requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2355 lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2356 lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2357 foreach_pre_remat(si, irn, tmp) {
2358 op_t *remat_op = get_irn_link(tmp);
2359 if(remat_op->attr.remat.remat->value == arg) {
2360 lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2364 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2366 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2367 if(get_irn_n(irn, n) == arg) {
2371 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2372 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2373 memoperand_t *memoperand;
2374 memoperand = set_find_memoperand(si->memoperands, irn, n);
2376 /* memoperand <= mem_out */
2377 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2378 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2379 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2380 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2382 /* the memoperand is only sufficient if it is used once by the op */
2383 if(n_memoperands == 1)
2384 lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2386 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2388 /* we have one more free register if we use a memory operand */
2389 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2395 /* iterate over L\U */
2396 pset_foreach(live, tmp) {
2397 if(!set_find_keyval(args, tmp)) {
2398 /* if a live value is not used by irn */
2399 tmp_op = get_irn_link(tmp);
2400 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2404 /* requirements for remats */
2405 foreach_pre_remat(si, irn, tmp) {
2406 op_t *remat_op = get_irn_link(tmp);
2409 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2410 ir_node *remat_arg = get_irn_n(tmp, n);
2411 op_t *arg_op = get_irn_link(remat_arg);
2413 if(!has_reg_class(si, remat_arg)) continue;
2415 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2416 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2417 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2419 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2420 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2422 /* if remat arg is also used by current op then we can use reload placed for this argument */
2423 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2424 int index = (int)keyval->val;
2426 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2434 /*************************
2435 * D O N E W I T H O P
2436 *************************/
2438 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2440 pset_foreach(live, tmp) {
2441 assert(has_reg_class(si, tmp));
2445 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2446 ir_node *arg = get_irn_n(irn, n);
2448 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2452 del_pset(remat_defs);
2456 defs = pset_new_ptr_default();
2458 /* skip everything above barrier in start block */
2459 if(is_start_block(bb) && be_is_Barrier(irn)) {
2460 assert(pset_count(live) == 0);
2469 /***************************************
2470 * B E G I N N I N G O F B L O C K
2471 ***************************************/
2474 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2475 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2477 pset_foreach(live, irn) {
2478 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2481 /* construct mem_outs for all values */
2482 set_foreach(spill_bb->ilp, spill) {
2483 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2484 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2486 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2487 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2489 if(pset_find_ptr(live, spill->irn)) {
2490 int default_spilled;
2491 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2493 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2494 default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2495 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2496 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2498 if(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2500 op_t *op = get_irn_link(spill->irn);
2502 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2503 const ir_node *arg = get_irn_n(spill->irn, n);
2509 /* argument already done? */
2510 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2512 /* get sum of execution frequencies of blocks with the same phi argument */
2513 for(m=n; m>=0; --m) {
2514 const ir_node *arg2 = get_irn_n(spill->irn, m);
2517 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2521 /* copies are not for free */
2522 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2523 var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
2525 for(m=n; m>=0; --m) {
2526 const ir_node *arg2 = get_irn_n(spill->irn, m);
2529 op->attr.live_range.args.copies[m] = var;
2534 /* copy <= mem_in */
2535 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2536 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2537 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2538 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2545 foreach_post_remat(bb, tmp) {
2547 op_t *remat_op = get_irn_link(tmp);
2548 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
2551 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2552 remat_arg = get_irn_n(tmp, n);
2554 if(has_reg_class(si, remat_arg)) {
2555 pset_insert_ptr(remat_args, remat_arg);
2559 /* remat + \sum live_range(remat_arg) <= |args| */
2560 ir_snprintf(buf, sizeof(buf), "one_must_die_%N", tmp);
2561 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2562 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2564 pset_foreach(remat_args, remat_arg) {
2565 if(pset_find_ptr(live, remat_arg)) {
2566 op_t *remat_arg_op = get_irn_link(remat_arg);
2567 lpp_set_factor_fast(si->lpp, cst, remat_arg_op->attr.live_range.ilp, 1.0);
2570 del_pset(remat_args);
2573 foreach_post_remat(bb, tmp) {
2576 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2577 ir_node *remat_arg = get_irn_n(tmp, n);
2579 /* if value is becoming live through use by remat2 */
2580 if(has_reg_class(si, remat_arg) && !pset_find_ptr(live, remat_arg)) {
2581 op_t *remat_arg_op = get_irn_link(remat_arg);
2584 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 at bb start %+F\n", remat_arg, tmp));
2586 pset_insert_ptr(live, remat_arg);
2587 spill = add_to_spill_bb(si, bb, remat_arg);
2588 remat_arg_op->attr.live_range.ilp = ILP_UNDEF;
2590 /* we need reg_in and mem_in for this value; they will be referenced later */
2591 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", remat_arg, bb);
2592 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2593 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", remat_arg, bb);
2594 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2597 /* optimization: all memory stuff should be 0, for we do not want to insert reloads for remats */
2598 ir_snprintf(buf, sizeof(buf), "nomem_%N_%N", remat_arg, bb);
2599 nomem = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0.0);
2600 lpp_set_factor_fast(si->lpp, nomem, spill->spill, 1.0);
2605 /* L\U is empty at bb start */
2606 /* arg is live throughout epilog if it is reg_in into this block */
2608 /* check the register pressure at the beginning of the block
2611 /* reg_in entspricht post_use */
2613 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2614 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2616 pset_foreach(live, irn) {
2619 spill = set_find_spill(spill_bb->ilp, irn);
2622 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2623 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2625 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2627 /* spill + mem_in <= 1 */
2628 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2629 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2631 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2632 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2634 } /* post_remats are NOT included in register pressure check because
2635 they do not increase regpressure */
2637 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2638 pset_foreach(live, irn) {
2642 spill = set_find_spill(spill_bb->ilp, irn);
2643 assert(spill && spill->irn == irn);
2645 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2646 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2649 ir_node *phi_arg = get_Phi_pred(irn, n);
2650 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2651 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2653 op_t *op = get_irn_link(irn);
2655 /* although the phi is in the right regclass one or more of
2656 * its arguments can be in a different one or at least to
2659 if(has_reg_class(si, phi_arg)) {
2660 /* mem_in < mem_out_arg + copy */
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);
2664 /* reg_in < reg_out_arg */
2665 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2666 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2668 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2669 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2671 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2674 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2676 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2678 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2682 /* else assure the value arrives on all paths in the same resource */
2684 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2687 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2688 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2691 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2692 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2693 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2694 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2696 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2697 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2699 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2702 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2703 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2708 foreach_post_remat(bb, tmp) {
2711 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2712 ir_node *remat_arg = get_irn_n(tmp, n);
2713 op_t *remat_op = get_irn_link(tmp);
2715 if(!has_reg_class(si, remat_arg)) continue;
2717 spill = set_find_spill(spill_bb->ilp, remat_arg);
2720 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2721 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2722 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2723 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2727 pset_foreach(live, irn) {
2728 const op_t *op = get_irn_link(irn);
2729 const ir_node *remat;
2734 foreach_post_remat(bb, remat) {
2737 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2738 const ir_node *arg = get_irn_n(remat, n);
2741 const op_t *remat_op = get_irn_link(remat);
2743 if(cst == ILP_UNDEF) {
2744 /* sum remat2s <= 1 + n_remats*live_range */
2745 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", irn, bb);
2746 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2748 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2754 if(cst != ILP_UNDEF && op->attr.live_range.ilp != ILP_UNDEF) {
2755 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -n_remats);
2759 /* first live ranges from reg_ins */
2760 pset_foreach(live, irn) {
2761 op_t *op = get_irn_link(irn);
2763 if(op->attr.live_range.ilp != ILP_UNDEF) {
2765 spill = set_find_spill(spill_bb->ilp, irn);
2766 assert(spill && spill->irn == irn);
2768 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2769 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2770 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2771 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2773 foreach_post_remat(bb, tmp) {
2774 op_t *remat_op = get_irn_link(tmp);
2776 if(remat_op->attr.remat.remat->value == irn) {
2777 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2783 /* walk forward now and compute constraints for placing spills */
2784 /* this must only be done for values that are not defined in this block */
2785 pset_foreach(live, irn) {
2787 * if value is defined in this block we can anways place the spill directly after the def
2788 * -> no constraint necessary
2790 if(!is_Phi(irn) && get_nodes_block(irn) == bb) {
2795 spill = set_find_spill(spill_bb->ilp, irn);
2798 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2799 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2801 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2802 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2805 sched_foreach_op(bb, tmp) {
2806 op_t *op = get_irn_link(tmp);
2808 if(is_Phi(tmp)) continue;
2809 assert(!is_Proj(tmp));
2812 const ir_node *value = op->attr.remat.remat->value;
2815 /* only collect remats up to the first real use of a value */
2816 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2821 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2822 ir_node *arg = get_irn_n(tmp, n);
2825 /* if a value is used stop collecting remats */
2838 typedef struct _irnlist_t {
2839 struct list_head list;
2843 typedef struct _interference_t {
2844 struct list_head blocklist;
2850 cmp_interference(const void *a, const void *b, size_t size)
2852 const interference_t *p = a;
2853 const interference_t *q = b;
2855 return !(p->a == q->a && p->b == q->b);
2858 static interference_t *
2859 set_find_interference(set * set, ir_node * a, ir_node * b)
2861 interference_t query;
2863 query.a = (a>b)?a:b;
2864 query.b = (a>b)?b:a;
2866 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2869 static interference_t *
2870 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2872 interference_t query,
2874 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2878 result = set_find_interference(set, a, b);
2881 list_add(&list->list, &result->blocklist);
2885 query.a = (a>b)?a:b;
2886 query.b = (a>b)?b:a;
2888 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2890 INIT_LIST_HEAD(&result->blocklist);
2891 list_add(&list->list, &result->blocklist);
2897 values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
2899 const ir_edge_t *edge;
2901 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2902 /* both values are live in, so they interfere */
2906 /* ensure a dominates b */
2907 if(value_dominates(b,a)) {
2913 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2916 /* the following code is stolen from bera.c */
2917 if(be_is_live_end(si->lv, bb, a))
2920 foreach_out_edge(a, edge) {
2921 const ir_node *user = edge->src;
2922 if(get_nodes_block(user) == bb
2925 && !pset_find_ptr(si->inverse_ops, user)
2926 && value_dominates(b, user))
2934 * Walk all irg blocks and collect interfering values inside of phi classes
2937 luke_interferencewalker(ir_node * bb, void * data)
2939 spill_ilp_t *si = (spill_ilp_t*)data;
2942 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2943 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
2944 op_t *a_op = get_irn_link(a);
2947 /* a is only interesting if it is in my register class and if it is inside a phi class */
2948 if (has_reg_class(si, a) && get_phi_class(si->pc, a)) {
2949 if (a_op->is_remat || pset_find_ptr(si->inverse_ops, a))
2952 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)) {
2953 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
2954 op_t *b_op = get_irn_link(b);
2956 /* a and b are only interesting if they are in the same phi class */
2957 if (has_reg_class(si, b) && get_phi_class(si->pc, a) == get_phi_class(si->pc, b)) {
2958 if (b_op->is_remat || pset_find_ptr(si->inverse_ops, b))
2961 if (values_interfere_in_block(si, bb, a, b)) {
2962 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2963 set_insert_interference(si, si->interferences, a, b, bb);
2971 static unsigned int copy_path_id = 0;
2974 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2981 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2982 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2984 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2986 pset_foreach(copies, ptr) {
2987 copy = PTR_TO_INT(ptr);
2988 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2993 * @parameter copies contains a path of copies which lead us to irn
2994 * @parameter visited contains a set of nodes already visited on this path
2997 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2999 const ir_edge_t *edge;
3000 op_t *op = get_irn_link(irn);
3001 pset *visited_users = pset_new_ptr_default();
3004 if(op->is_remat) return 0;
3006 pset_insert_ptr(visited, irn);
3010 pset *visited_operands = pset_new_ptr(get_irn_arity(irn));
3012 /* visit all operands */
3013 for(n=get_irn_arity(irn)-1; n>=0; --n) {
3014 ir_node *arg = get_irn_n(irn, n);
3015 ilp_var_t copy = op->attr.live_range.args.copies[n];
3017 if(!has_reg_class(si, arg)) continue;
3018 if(pset_find_ptr(visited_operands, arg)) continue;
3019 pset_insert_ptr(visited_operands, arg);
3022 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3023 del_pset(visited_operands);
3024 del_pset(visited_users);
3025 pset_remove_ptr(visited, irn);
3028 pset_insert(copies, INT_TO_PTR(copy), copy);
3029 write_copy_path_cst(si, copies, any_interfere);
3030 pset_remove(copies, INT_TO_PTR(copy), copy);
3031 } else if(!pset_find_ptr(visited, arg)) {
3032 pset_insert(copies, INT_TO_PTR(copy), copy);
3033 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
3034 pset_remove(copies, INT_TO_PTR(copy), copy);
3036 if(paths > MAX_PATHS) {
3037 if(pset_count(copies) == 0) {
3041 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3042 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3043 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3044 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3045 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3049 del_pset(visited_operands);
3050 del_pset(visited_users);
3051 pset_remove_ptr(visited, irn);
3054 } else if(pset_count(copies) == 0) {
3060 del_pset(visited_operands);
3063 /* visit all uses which are phis */
3064 foreach_out_edge(irn, edge) {
3065 ir_node *user = edge->src;
3066 int pos = edge->pos;
3067 op_t *op = get_irn_link(user);
3070 if(!is_Phi(user)) continue;
3071 if(!has_reg_class(si, user)) continue;
3072 if(pset_find_ptr(visited_users, user)) continue;
3073 pset_insert_ptr(visited_users, user);
3075 copy = op->attr.live_range.args.copies[pos];
3077 if(user == target) {
3078 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3079 del_pset(visited_users);
3080 pset_remove_ptr(visited, irn);
3083 pset_insert(copies, INT_TO_PTR(copy), copy);
3084 write_copy_path_cst(si, copies, any_interfere);
3085 pset_remove(copies, INT_TO_PTR(copy), copy);
3086 } else if(!pset_find_ptr(visited, user)) {
3087 pset_insert(copies, INT_TO_PTR(copy), copy);
3088 paths += find_copy_path(si, user, target, any_interfere, copies, visited);
3089 pset_remove(copies, INT_TO_PTR(copy), copy);
3091 if(paths > MAX_PATHS) {
3092 if(pset_count(copies) == 0) {
3096 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3097 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3098 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3099 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3100 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3104 del_pset(visited_users);
3105 pset_remove_ptr(visited, irn);
3108 } else if(pset_count(copies) == 0) {
3114 del_pset(visited_users);
3115 pset_remove_ptr(visited, irn);
3120 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
3122 pset * copies = pset_new_ptr_default();
3123 pset * visited = pset_new_ptr_default();
3125 find_copy_path(si, a, b, any_interfere, copies, visited);
3133 memcopyhandler(spill_ilp_t * si)
3135 interference_t *interference;
3137 /* teste Speicherwerte auf Interferenz */
3139 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
3140 irg_block_walk_graph(si->birg->irg, luke_interferencewalker, NULL, si);
3142 /* now lets emit the ILP unequations for the crap */
3143 set_foreach(si->interferences, interference) {
3145 ilp_var_t interfere, any_interfere;
3146 ilp_cst_t any_interfere_cst, cst;
3147 const ir_node *a = interference->a;
3148 const ir_node *b = interference->b;
3150 /* any_interf <= \sum interf */
3151 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
3152 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3153 any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3155 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
3157 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
3158 const ir_node *bb = irnlist->irn;
3159 spill_bb_t *spill_bb = get_irn_link(bb);
3164 spilla = set_find_spill(spill_bb->ilp, a);
3167 spillb = set_find_spill(spill_bb->ilp, b);
3170 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
3171 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
3172 /* 2: - mem_in_a - spill_a + interfere <= 0 */
3173 /* 3: - mem_in_b - spill_b + interfere <= 0 */
3174 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
3175 interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3177 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
3178 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
3180 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
3181 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
3182 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
3183 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
3184 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
3186 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
3187 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3189 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3190 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
3191 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
3193 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
3194 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3196 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3197 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
3198 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
3201 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
3203 /* any_interfere >= interf */
3204 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
3205 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3207 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3208 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3211 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
3212 gen_copy_constraints(si,a,b,any_interfere);
3220 return fabs(x) < 0.00001;
3223 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
3225 spill_ilp_t *si = get_irg_link(current_ir_graph);
3227 if(pset_find_ptr(si->all_possible_remats, n)) {
3228 op_t *op = (op_t*)get_irn_link(n);
3229 assert(op && op->is_remat);
3231 if(!op->attr.remat.remat->inverse) {
3232 if(op->attr.remat.pre) {
3233 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
3235 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
3240 op_t *op = (op_t*)get_irn_link(n);
3241 assert(op && op->is_remat);
3243 if(op->attr.remat.pre) {
3244 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
3246 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
3257 dump_graph_with_remats(ir_graph * irg, const char * suffix)
3259 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
3260 be_dump(irg, suffix, dump_ir_block_graph_sched);
3261 set_dump_node_vcgattr_hook(NULL);
3265 * Edge hook to dump the schedule edges with annotated register pressure.
3268 sched_pressure_edge_hook(FILE *F, ir_node *irn)
3270 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
3271 ir_node *prev = sched_prev(irn);
3272 fprintf(F, "edge:{sourcename:\"");
3274 fprintf(F, "\" targetname:\"");
3276 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
3277 fprintf(F, "\" color:magenta}\n");
3283 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
3285 DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3287 dump_consts_local(0);
3288 set_dump_node_edge_hook(sched_pressure_edge_hook);
3289 dump_ir_block_graph(irg, suffix);
3290 set_dump_node_edge_hook(old_edge_hook);
3294 walker_pressure_annotator(ir_node * bb, void * data)
3296 spill_ilp_t *si = data;
3299 pset *live = pset_new_ptr_default();
3302 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3303 irn = be_lv_get_irn(si->lv, bb, i);
3305 if (has_reg_class(si, irn)) {
3306 pset_insert_ptr(live, irn);
3310 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3312 sched_foreach_reverse(bb, irn) {
3314 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3318 if(has_reg_class(si, irn)) {
3319 pset_remove_ptr(live, irn);
3320 if(is_Proj(irn)) ++projs;
3323 if(!is_Proj(irn)) projs = 0;
3325 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3326 ir_node *arg = get_irn_n(irn, n);
3328 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3330 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3337 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3339 be_dump(si->birg->irg, suffix, dump_ir_block_graph_sched_pressure);
3343 connect_all_remats_with_keep(spill_ilp_t * si)
3351 n_remats = pset_count(si->all_possible_remats);
3353 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3356 pset_foreach(si->all_possible_remats, irn) {
3361 si->keep = be_new_Keep(si->cls, si->birg->irg, get_irg_end_block(si->birg->irg), n_remats, ins);
3363 obstack_free(si->obst, ins);
3368 connect_all_spills_with_keep(spill_ilp_t * si)
3377 n_spills = pset_count(si->spills);
3379 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3382 pset_foreach(si->spills, irn) {
3387 keep = be_new_Keep(si->cls, si->birg->irg, get_irg_end_block(si->birg->irg), n_spills, ins);
3389 obstack_free(si->obst, ins);
3393 /** insert a spill at an arbitrary position */
3394 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3396 ir_node *bl = is_Block(insert) ? insert : get_nodes_block(insert);
3397 ir_graph *irg = get_irn_irg(bl);
3398 ir_node *frame = get_irg_frame(irg);
3401 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
3402 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3404 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3407 * search the right insertion point. a spill of a phi cannot be put
3408 * directly after the phi, if there are some phis behind the one which
3409 * is spilled. Also, a spill of a Proj must be after all Projs of the
3412 * Here's one special case:
3413 * If the spill is in the start block, the spill must be after the frame
3414 * pointer is set up. This is done by setting insert to the end of the block
3415 * which is its default initialization (see above).
3418 if (bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3421 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3424 sched_add_after(insert, spill);
3429 delete_remat(spill_ilp_t * si, ir_node * remat) {
3431 ir_node *bad = get_irg_bad(si->birg->irg);
3433 sched_remove(remat);
3435 /* kill links to operands */
3436 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3437 set_irn_n(remat, n, bad);
3442 clean_remat_info(spill_ilp_t * si)
3446 remat_info_t *remat_info;
3447 ir_node *bad = get_irg_bad(si->birg->irg);
3449 set_foreach(si->remat_info, remat_info) {
3450 if(!remat_info->remats) continue;
3452 pset_foreach(remat_info->remats, remat)
3454 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3455 if(sched_is_scheduled(remat->proj)) {
3456 sched_remove((ir_node*)remat->proj);
3458 set_irn_n((ir_node*)remat->proj, -1, bad);
3459 set_irn_n((ir_node*)remat->proj, 0, bad);
3462 if(get_irn_n_edges(remat->op) == 0) {
3463 if(sched_is_scheduled(remat->op)) {
3464 sched_remove((ir_node*)remat->op);
3466 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3467 set_irn_n((ir_node*)remat->op, n, bad);
3472 if(remat_info->remats) del_pset(remat_info->remats);
3473 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3478 delete_unnecessary_remats(spill_ilp_t * si)
3480 if(opt_keep_alive & KEEPALIVE_REMATS) {
3482 ir_node *bad = get_irg_bad(si->birg->irg);
3485 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3486 ir_node *keep_arg = get_irn_n(si->keep, n);
3487 op_t *arg_op = get_irn_link(keep_arg);
3490 assert(arg_op->is_remat);
3492 name = si->lpp->vars[arg_op->attr.remat.ilp];
3494 if(is_zero(name->value)) {
3495 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
3496 /* TODO check whether reload is preferred over remat (could be bug) */
3497 delete_remat(si, keep_arg);
3499 if(!arg_op->attr.remat.remat->inverse) {
3500 if(arg_op->attr.remat.pre) {
3501 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3503 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3506 if(arg_op->attr.remat.pre) {
3507 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3509 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3514 set_irn_n(si->keep, n, bad);
3517 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
3522 pset_foreach(si->all_possible_remats, remat) {
3523 op_t *remat_op = get_irn_link(remat);
3524 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
3526 if(is_zero(name->value)) {
3527 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
3528 /* TODO check whether reload is preferred over remat (could be bug) */
3529 delete_remat(si, remat);
3531 if(!remat_op->attr.remat.remat->inverse) {
3532 if(remat_op->attr.remat.pre) {
3533 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3535 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3538 if(remat_op->attr.remat.pre) {
3539 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3541 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3550 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3552 pset *spills = pset_new_ptr_default();
3554 const ir_node *next;
3557 defs = set_find_def(si->values, value);
3559 if(defs && defs->spills) {
3560 for(next = defs->spills; next; next = get_irn_link(next)) {
3561 pset_insert_ptr(spills, next);
3569 new_r_PhiM_nokeep(ir_graph * irg, ir_node *block, int arity, ir_node **in)
3573 assert( get_irn_arity(block) == arity );
3575 res = new_ir_node(NULL, irg, block, op_Phi, mode_M, arity, in);
3576 res->attr.phi_backedge = new_backedge_arr(irg->obst, arity);
3582 * @param before The node after which the spill will be placed in the schedule
3585 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3589 const arch_env_t *arch_env = si->birg->main_env->arch_env;
3591 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3593 spill = be_spill2(arch_env, irn, before);
3595 defs = set_insert_def(si->values, value);
3598 /* enter into the linked list */
3599 set_irn_link(spill, defs->spills);
3600 defs->spills = spill;
3602 if(opt_keep_alive & KEEPALIVE_SPILLS)
3603 pset_insert_ptr(si->spills, spill);
3609 * @param before The Phi node which has to be spilled
3612 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3619 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3621 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3622 ins[n] = si->m_unknown;
3625 mem_phi = new_r_PhiM_nokeep(si->birg->irg, get_nodes_block(phi), get_irn_arity(phi), ins);
3627 defs = set_insert_def(si->values, phi);
3630 /* enter into the linked list */
3631 set_irn_link(mem_phi, defs->spills);
3632 defs->spills = mem_phi;
3634 #ifdef SCHEDULE_PHIM
3635 sched_add_after(phi, mem_phi);
3637 pset_insert_ptr(si->phims, mem_phi);
3640 if(opt_keep_alive & KEEPALIVE_SPILLS)
3641 pset_insert_ptr(si->spills, mem_phi);
3648 * Add remat to list of defs, destroys link field!
3651 insert_remat(spill_ilp_t * si, ir_node * remat)
3654 op_t *remat_op = get_irn_link(remat);
3656 assert(remat_op->is_remat);
3658 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3661 /* enter into the linked list */
3662 set_irn_link(remat, defs->remats);
3663 defs->remats = remat;
3668 * Add reload before operation and add to list of defs
3671 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3676 const arch_env_t *arch_env = si->birg->main_env->arch_env;
3678 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3680 defs = set_find_def(si->values, value);
3682 spill = defs->spills;
3683 assert(spill && "no spill placed before reload");
3685 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3687 /* enter into the linked list */
3688 set_irn_link(reload, defs->remats);
3689 defs->remats = reload;
3694 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3697 ir_node *value = get_irn_n(memoperand->irn, memoperand->pos);
3699 const arch_env_t *arch_env = si->birg->main_env->arch_env;
3701 DBG((si->dbg, LEVEL_2, "\t inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3703 defs = set_find_def(si->values, value);
3705 spill = defs->spills;
3706 assert(spill && "no spill placed before reload");
3708 arch_perform_memory_operand(arch_env, memoperand->irn, spill, memoperand->pos);
3711 void insert_memoperands(spill_ilp_t * si)
3713 memoperand_t *memoperand;
3716 set_foreach(si->memoperands, memoperand) {
3717 name = si->lpp->vars[memoperand->ilp];
3718 if(!is_zero(name->value)) {
3719 perform_memory_operand(si, memoperand);
3725 walker_spill_placer(ir_node * bb, void * data) {
3726 spill_ilp_t *si = (spill_ilp_t*)data;
3728 spill_bb_t *spill_bb = get_irn_link(bb);
3729 pset *spills_to_do = pset_new_ptr_default();
3732 set_foreach(spill_bb->ilp, spill) {
3735 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3736 name = si->lpp->vars[spill->mem_in];
3737 if(!is_zero(name->value)) {
3740 mem_phi = insert_mem_phi(si, spill->irn);
3742 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3746 name = si->lpp->vars[spill->spill];
3747 if(!is_zero(name->value)) {
3748 /* place spill directly after definition */
3749 if(get_nodes_block(spill->irn) == bb) {
3750 insert_spill(si, spill->irn, spill->irn, spill->irn);
3754 /* place spill at bb start */
3755 if(spill->reg_in > 0) {
3756 name = si->lpp->vars[spill->reg_in];
3757 if(!is_zero(name->value)) {
3758 insert_spill(si, spill->irn, spill->irn, bb);
3762 /* place spill after a remat */
3763 pset_insert_ptr(spills_to_do, spill->irn);
3766 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3769 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3770 op_t *op = get_irn_link(irn);
3772 if(be_is_Spill(irn)) continue;
3775 /* TODO fix this if we want to support remats with more than two nodes */
3776 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3777 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3779 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3782 if(pset_find_ptr(spills_to_do, irn)) {
3783 pset_remove_ptr(spills_to_do, irn);
3785 insert_spill(si, irn, irn, irn);
3791 assert(pset_count(spills_to_do) == 0);
3793 /* afterwards free data in block */
3794 del_pset(spills_to_do);
3798 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3800 ir_node *insert_pos = bb;
3802 const arch_env_t *arch_env = si->birg->main_env->arch_env;
3804 /* find last definition of arg value in block */
3809 defs = set_find_def(si->values, value);
3811 if(defs && defs->remats) {
3812 for(next = defs->remats; next; next = get_irn_link(next)) {
3813 if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3814 last = sched_get_time_step(next);
3820 if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3821 last = sched_get_time_step(value);
3825 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
3827 spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3833 phim_fixer(spill_ilp_t *si) {
3836 set_foreach(si->values, defs) {
3837 const ir_node *phi = defs->value;
3838 op_t *op = get_irn_link(phi);
3839 ir_node *phi_m = NULL;
3840 ir_node *next = defs->spills;
3843 if(!is_Phi(phi)) continue;
3846 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3850 next = get_irn_link(next);
3853 if(!phi_m) continue;
3855 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3856 ir_node *value = get_irn_n(phi, n);
3857 defs_t *val_defs = set_find_def(si->values, value);
3859 /* a spill of this value */
3864 ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3865 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3867 if(!is_zero(name->value)) {
3868 spill = insert_mem_copy(si, pred, value);
3870 spill = val_defs->spills;
3873 spill = val_defs->spills;
3876 assert(spill && "no spill placed before PhiM");
3877 set_irn_n(phi_m, n, spill);
3883 walker_reload_placer(ir_node * bb, void * data) {
3884 spill_ilp_t *si = (spill_ilp_t*)data;
3886 spill_bb_t *spill_bb = get_irn_link(bb);
3888 /* reloads at end of block */
3889 if(spill_bb->reloads) {
3892 set_foreach(spill_bb->reloads, keyval) {
3893 ir_node *irn = (ir_node*)keyval->key;
3894 ilp_var_t reload = PTR_TO_INT(keyval->val);
3897 name = si->lpp->vars[reload];
3898 if(!is_zero(name->value)) {
3900 ir_node *insert_pos = bb;
3901 ir_node *prev = sched_block_last_noncf(si, bb);
3902 op_t *prev_op = get_irn_link(prev);
3904 while(be_is_Spill(prev)) {
3905 prev = sched_prev(prev);
3908 prev_op = get_irn_link(prev);
3910 /* insert reload before pre-remats */
3911 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3912 && prev_op->is_remat && prev_op->attr.remat.pre) {
3916 prev = sched_prev(prev);
3917 } while(be_is_Spill(prev));
3919 prev_op = get_irn_link(prev);
3923 reload = insert_reload(si, irn, insert_pos);
3925 if(opt_keep_alive & KEEPALIVE_RELOADS)
3926 pset_insert_ptr(si->spills, reload);
3931 /* walk and insert more reloads and collect remats */
3932 sched_foreach_reverse(bb, irn) {
3933 op_t *op = get_irn_link(irn);
3935 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3936 if(is_Phi(irn)) break;
3939 if(get_irn_mode(irn) != mode_T) {
3940 insert_remat(si, irn);
3945 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3946 ir_node *arg = get_irn_n(irn, n);
3948 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3951 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3952 if(!is_zero(name->value)) {
3954 ir_node *insert_pos = irn;
3955 ir_node *prev = sched_prev(insert_pos);
3958 while(be_is_Spill(prev)) {
3959 prev = sched_prev(prev);
3962 prev_op = get_irn_link(prev);
3964 /* insert reload before pre-remats */
3965 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3966 && prev_op->is_remat && prev_op->attr.remat.pre) {
3970 prev = sched_prev(prev);
3971 } while(be_is_Spill(prev));
3973 prev_op = get_irn_link(prev);
3977 reload = insert_reload(si, arg, insert_pos);
3979 assert(reload && "no reload returned");
3980 set_irn_n(irn, n, reload);
3982 if(opt_keep_alive & KEEPALIVE_RELOADS)
3983 pset_insert_ptr(si->spills, reload);
3990 del_set(spill_bb->ilp);
3991 if(spill_bb->reloads) del_set(spill_bb->reloads);
3995 walker_collect_used(ir_node * irn, void * data)
3997 bitset_t *used = data;
3999 bitset_set(used, get_irn_idx(irn));
4002 struct kill_helper {
4008 walker_kill_unused(ir_node * bb, void * data)
4010 struct kill_helper *kh = data;
4011 ir_node *bad = get_irg_bad(get_irn_irg(bb));
4015 for(irn=sched_first(bb); !sched_is_end(irn);) {
4016 ir_node *next = sched_next(irn);
4019 if(!bitset_is_set(kh->used, get_irn_idx(irn))) {
4020 if(be_is_Spill(irn) || be_is_Reload(irn)) {
4021 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)));
4023 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
4029 set_nodes_block(irn, bad);
4030 for (n=get_irn_arity(irn)-1; n>=0; --n) {
4031 set_irn_n(irn, n, bad);
4038 #ifndef SCHEDULE_PHIM
4040 kill_unused_phims(spill_ilp_t * si, struct kill_helper * kh)
4043 ir_node *bad = get_irg_bad(si->birg->irg);
4046 pset_foreach(si->phims, phi) {
4047 if(!bitset_is_set(kh->used, get_irn_idx(phi))) {
4049 set_nodes_block(phi, bad);
4050 for (n=get_irn_arity(phi)-1; n>=0; --n) {
4051 set_irn_n(phi, n, bad);
4059 kill_all_unused_values_in_schedule(spill_ilp_t * si)
4061 struct kill_helper kh;
4063 kh.used = bitset_malloc(get_irg_last_idx(si->birg->irg));
4066 irg_walk_graph(si->birg->irg, walker_collect_used, NULL, kh.used);
4067 #ifndef SCHEDULE_PHIM
4068 kill_unused_phims(si, &kh);
4070 irg_block_walk_graph(si->birg->irg, walker_kill_unused, NULL, &kh);
4072 bitset_free(kh.used);
4076 print_irn_pset(pset * p)
4080 pset_foreach(p, irn) {
4081 ir_printf("%+F\n", irn);
4086 dump_phi_class(spill_ilp_t *si, ir_node **phiclass, const char * file)
4088 FILE *f = fopen(file, "w");
4090 interference_t *interference;
4093 set_break(si->interferences);
4095 ir_fprintf(f, "digraph phiclass {\n");
4097 for (i = ARR_LEN(phiclass) - 1; i >= 0; --i) {
4100 ir_fprintf(f, " %F%N [shape=box]\n", irn, irn);
4103 for (i = ARR_LEN(phiclass) - 1; i >= 0; --i) {
4110 for (n = get_irn_arity(irn) - 1; n >= 0; --n) {
4111 ir_node *arg = get_irn_n(irn, n);
4113 ir_fprintf(f, " %F%N -> %F%N\n", irn, irn, arg, arg);
4117 set_foreach(si->interferences, interference) {
4118 const ir_node *a = interference->a;
4119 const ir_node *b = interference->b;
4120 if (get_phi_class(si->pc, (ir_node *)a) == phiclass) {
4121 ir_fprintf(f, " %F%N -> %F%N [color=red,dir=none,style=bold]\n", a, a, b, b);
4130 rewire_uses(spill_ilp_t * si)
4133 pset *ignore = pset_new_ptr(1);
4134 be_dom_front_info_t *dom_front = si->birg->dom_front;
4136 pset_insert_ptr(ignore, get_irg_end(si->birg->irg));
4138 /* then fix uses of spills */
4139 set_foreach(si->values, defs) {
4142 const ir_node *next = defs->remats;
4145 reloads = pset_new_ptr_default();
4148 if(be_is_Reload(next)) {
4149 pset_insert_ptr(reloads, next);
4153 next = get_irn_link(next);
4156 spills = get_spills_for_value(si, defs->value);
4157 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));
4158 if(pset_count(spills) > 1) {
4159 //assert(pset_count(reloads) > 0);
4160 // print_irn_pset(spills);
4161 // print_irn_pset(reloads);
4163 be_ssa_constr_set_ignore(dom_front, si->lv, spills, ignore);
4170 /* first fix uses of remats and reloads */
4171 set_foreach(si->values, defs) {
4173 const ir_node *next = defs->remats;
4177 nodes = pset_new_ptr_default();
4178 if(sched_is_scheduled(defs->value)) {
4179 pset_insert_ptr(nodes, defs->value);
4184 pset_insert_ptr(nodes, next);
4185 next = get_irn_link(next);
4188 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-orig_kept, defs->value));
4189 be_ssa_constr_set_ignore(dom_front, si->lv, nodes, NULL);
4195 // remove_unused_defs(si);
4200 writeback_results(spill_ilp_t * si)
4202 /* walk through the graph and collect all spills, reloads and remats for a value */
4204 si->values = new_set(cmp_defs, 4096);
4206 DBG((si->dbg, LEVEL_1, "Applying results\n"));
4207 delete_unnecessary_remats(si);
4208 si->m_unknown = new_r_Unknown(si->birg->irg, mode_M);
4209 irg_block_walk_graph(si->birg->irg, walker_spill_placer, NULL, si);
4210 irg_block_walk_graph(si->birg->irg, walker_reload_placer, NULL, si);
4212 insert_memoperands(si);
4215 /* clean the remat info! there are still back-edges leading there! */
4216 clean_remat_info(si);
4220 connect_all_spills_with_keep(si);
4222 del_set(si->values);
4226 get_n_regs(spill_ilp_t * si)
4228 int arch_n_regs = arch_register_class_n_regs(si->cls);
4230 bitset_t *arch_regs = bitset_malloc(arch_n_regs);
4231 bitset_t *abi_regs = bitset_malloc(arch_n_regs);
4233 arch_put_non_ignore_regs(si->birg->main_env->arch_env, si->cls, arch_regs);
4234 be_abi_put_ignore_regs(si->birg->abi, si->cls, abi_regs);
4236 bitset_andnot(arch_regs, abi_regs);
4237 arch_n_regs = bitset_popcnt(arch_regs);
4239 bitset_free(arch_regs);
4240 bitset_free(abi_regs);
4242 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", arch_n_regs, si->cls->name));
4247 walker_reload_mover(ir_node * bb, void * data)
4249 spill_ilp_t *si = data;
4252 sched_foreach(bb, tmp) {
4253 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
4254 ir_node *reload = tmp;
4257 /* move reload upwards */
4259 int pressure = (int)get_irn_link(reload);
4260 if(pressure < si->n_regs) {
4261 irn = sched_prev(reload);
4262 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
4263 sched_remove(reload);
4264 pressure = (int)get_irn_link(irn);
4266 while(pressure < si->n_regs) {
4267 if( sched_is_end(irn) ||
4268 (be_is_Reload(irn) && has_reg_class(si, irn)) ||
4269 /* do not move reload before its spill */
4270 (irn == be_get_Reload_mem(reload)) ||
4271 /* do not move before phi */
4274 set_irn_link(irn, INT_TO_PTR(pressure+1));
4275 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
4276 irn = sched_prev(irn);
4278 pressure = (int)get_irn_link(irn);
4281 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
4282 sched_put_after(irn, reload);
4289 move_reloads_upward(spill_ilp_t * si)
4291 irg_block_walk_graph(si->birg->irg, walker_reload_mover, NULL, si);
4296 * Walk all irg blocks and check for interfering spills inside of phi classes
4299 luke_meminterferencechecker(ir_node * bb, void * data)
4301 spill_ilp_t *si = (spill_ilp_t*)data;
4304 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
4305 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
4307 if (! be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T))
4310 /* a is only interesting if it is in my register class and if it is inside a phi class */
4311 if (has_reg_class(si, a) && get_phi_class(si->pc, a)) {
4312 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)) {
4313 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
4315 if (! be_is_Spill(b) && (! is_Phi(b) || get_irn_mode(b) != mode_T))
4318 /* a and b are only interesting if they are in the same phi class */
4319 if (has_reg_class(si, b) && get_phi_class(si->pc, a) == get_phi_class(si->pc, b)) {
4320 if (values_interfere_in_block(si, bb, a, b)) {
4321 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4330 verify_phiclasses(spill_ilp_t * si)
4332 /* analyze phi classes */
4333 phi_class_free(si->pc);
4334 si->pc = phi_class_new_from_irg(si->birg->irg, 0);
4336 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4337 irg_block_walk_graph(si->birg->irg, luke_meminterferencechecker, NULL, si);
4341 be_spill_remat(be_irg_t *birg, const arch_register_class_t *cls)
4344 char problem_name[256];
4345 char dump_suffix[256];
4346 char dump_suffix2[256];
4347 struct obstack obst;
4349 ir_graph *irg = be_get_birg_irg(birg);
4351 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", irg, cls->name);
4352 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", cls->name);
4353 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", cls->name);
4355 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4356 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4358 if(opt_verify & VERIFY_DOMINANCE)
4359 be_check_dominance(irg);
4361 be_assure_dom_front(birg);
4362 be_assure_liveness(birg);
4364 obstack_init(&obst);
4368 si.lpp = new_lpp(problem_name, lpp_minimize);
4369 si.remat_info = new_set(cmp_remat_info, 4096);
4370 si.interferences = new_set(cmp_interference, 32);
4371 si.memoperands = new_set(cmp_memoperands, 128);
4372 si.all_possible_remats = pset_new_ptr_default();
4373 si.spills = pset_new_ptr_default();
4374 si.inverse_ops = pset_new_ptr_default();
4377 si.n_regs = get_n_regs(&si);
4379 set_irg_link(irg, &si);
4382 /* compute phi classes */
4383 // phi_class_compute(irg);
4385 if(opt_dump_flags & DUMP_STATS)
4386 be_analyze_regpressure(birg, cls, "-pre");
4388 DBG((si.dbg, LEVEL_2, "\t initializing\n"));
4389 irg_block_walk_graph(irg, luke_initializer, NULL, &si);
4392 /* collect remats */
4393 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4394 irg_walk_graph(irg, walker_remat_collector, NULL, &si);
4397 /* insert possible remats */
4398 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4399 irg_block_walk_graph(irg, walker_remat_insertor, NULL, &si);
4400 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4402 if(opt_keep_alive & KEEPALIVE_REMATS) {
4403 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4404 connect_all_remats_with_keep(&si);
4405 /* dump graph with inserted remats */
4406 dump_graph_with_remats(irg, dump_suffix);
4409 /* insert copies for phi arguments not in my regclass */
4410 irg_walk_graph(irg, walker_regclass_copy_insertor, NULL, &si);
4412 /* recompute liveness */
4413 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4414 be_liveness_recompute(si.lv);
4417 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4418 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4419 irg_block_walk_graph(irg, luke_endwalker, NULL, &si);
4421 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4422 irg_block_walk_graph(irg, luke_blockwalker, NULL, &si);
4424 si.pc = phi_class_new_from_irg(birg->irg, 0);
4425 if (opt_memcopies) {
4426 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4427 memcopyhandler(&si);
4430 if (opt_dump_flags & DUMP_PROBLEM) {
4432 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4433 if ((f = fopen(buf, "wt")) != NULL) {
4434 lpp_dump_plain(si.lpp, f);
4439 if (opt_dump_flags & DUMP_MPS) {
4442 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4443 if ((f = fopen(buf, "wt")) != NULL) {
4444 mps_write_mps(si.lpp, s_mps_fixed, f);
4448 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4449 if ((f = fopen(buf, "wt")) != NULL) {
4450 mps_write_mst(si.lpp, s_mps_fixed, f);
4455 lpp_check_startvals(si.lpp);
4458 DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4459 lpp_set_time_limit(si.lpp, opt_timeout);
4462 lpp_set_log(si.lpp, stdout);
4465 lpp_solve_cplex(si.lpp);
4467 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4469 assert(lpp_is_sol_valid(si.lpp)
4470 && "solution of ILP must be valid");
4472 DBG((si.dbg, LEVEL_1, "\t%s: iterations: %d, solution time: %g, objective function: %g, best bound: %g\n", problem_name, si.lpp->iterations, si.lpp->sol_time, is_zero(si.lpp->objval)?0.0:si.lpp->objval, is_zero(si.lpp->best_bound)?0.0:si.lpp->best_bound));
4474 if(opt_dump_flags & DUMP_SOLUTION) {
4478 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4479 if ((f = fopen(buf, "wt")) != NULL) {
4481 for (i = 0; i < si.lpp->var_next; ++i) {
4482 lpp_name_t *name = si.lpp->vars[i];
4483 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4489 #ifndef SCHEDULE_PHIM
4490 si.phims = pset_new_ptr_default();
4492 writeback_results(&si);
4497 kill_all_unused_values_in_schedule(&si);
4499 #if !defined(SCHEDULE_PHIM) && defined(SOLVE)
4503 if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
4504 be_dump(irg, "-spills-placed", dump_ir_block_graph);
4506 // move reloads upwards
4507 be_liveness_recompute(si.lv);
4508 irg_block_walk_graph(irg, walker_pressure_annotator, NULL, &si);
4509 move_reloads_upward(&si);
4512 verify_phiclasses(&si);
4515 irg_block_walk_graph(irg, walker_pressure_annotator, NULL, &si);
4517 if(opt_dump_flags & DUMP_PRESSURE)
4518 dump_pressure_graph(&si, dump_suffix2);
4520 if(opt_dump_flags & DUMP_STATS)
4521 be_analyze_regpressure(birg, cls, "-post");
4523 if(opt_verify & VERIFY_DOMINANCE)
4524 be_check_dominance(irg);
4527 del_set(si.interferences);
4528 del_pset(si.inverse_ops);
4529 del_pset(si.all_possible_remats);
4530 del_set(si.memoperands);
4531 del_pset(si.spills);
4533 phi_class_free(si.pc);
4534 obstack_free(&obst, NULL);
4535 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4538 void be_init_spillremat(void)
4540 static be_spiller_t remat_spiller = {
4543 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
4544 lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
4545 lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
4546 lc_opt_entry_t *remat_grp = lc_opt_get_grp(chordal_grp, "remat");
4548 be_register_spiller("remat", &remat_spiller);
4549 lc_opt_add_table(remat_grp, options);
4552 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillremat);
4554 #else /* WITH_ILP */
4557 only_that_you_can_compile_without_WITH_ILP_defined(void)
4561 #endif /* WITH_ILP */