1 /** vim: set sw=4 ts=4:
4 * @author Adam M. Szalkowski & Sebastian Hack
6 * ILP based spilling & rematerialization
8 * Copyright (C) 2006 Universitaet Karlsruhe
9 * Released under the GPL
32 #include "phiclass_t.h"
39 #include <lpp/lpp_net.h>
40 #include <lpp/lpp_cplex.h>
41 //#include <lc_pset.h>
42 //#include <libcore/lc_bitset.h>
46 #include "besched_t.h"
52 #include "bespillremat.h"
54 #include "bepressurestat.h"
56 #include "bechordal_t.h"
59 #include <libcore/lc_opts.h>
60 #include <libcore/lc_opts_enum.h>
61 #endif /* WITH_LIBCORE */
63 #define DUMP_PROBLEM 1
65 #define DUMP_SOLUTION 4
67 #define KEEPALIVE_REMATS 1
68 #define KEEPALIVE_SPILLS 2
69 #define KEEPALIVE_RELOADS 4
71 #define VERIFY_MEMINTERF 1
72 #define VERIFY_DOMINANCE 2
75 #define REMATS_BRIGGS 1
76 #define REMATS_NOINVERSE 2
79 static int opt_dump_flags = 0;
80 static int opt_log = 0;
81 static int opt_keep_alive = 0;
82 static int opt_goodwin = 1;
83 static int opt_memcopies = 1;
84 static int opt_memoperands = 1;
85 static int opt_verify = VERIFY_MEMINTERF;
86 static int opt_remats = REMATS_ALL;
87 static int opt_repair_schedule = 0;
88 static int opt_no_enlarge_liveness = 0;
89 static int opt_remat_while_live = 1;
90 static int opt_timeout = 300;
91 static double opt_cost_reload = 8.0;
92 static double opt_cost_memoperand = 7.0;
93 static double opt_cost_spill = 50.0;
94 static double opt_cost_remat = 1.0;
98 static const lc_opt_enum_mask_items_t dump_items[] = {
99 { "problem", DUMP_PROBLEM },
101 { "solution", DUMP_SOLUTION },
105 static lc_opt_enum_mask_var_t dump_var = {
106 &opt_dump_flags, dump_items
109 static const lc_opt_enum_mask_items_t keepalive_items[] = {
110 { "remats", KEEPALIVE_REMATS },
111 { "spills", KEEPALIVE_SPILLS },
112 { "reloads", KEEPALIVE_RELOADS },
116 static lc_opt_enum_mask_var_t keep_alive_var = {
117 &opt_keep_alive, keepalive_items
120 static const lc_opt_enum_mask_items_t remats_items[] = {
121 { "none", REMATS_NONE },
122 { "briggs", REMATS_BRIGGS },
123 { "noinverse", REMATS_NOINVERSE },
124 { "all", REMATS_ALL },
128 static lc_opt_enum_mask_var_t remats_var = {
129 &opt_remats, remats_items
132 static const lc_opt_table_entry_t options[] = {
133 LC_OPT_ENT_ENUM_MASK("keepalive", "keep alive remats, spills or reloads", &keep_alive_var),
135 LC_OPT_ENT_BOOL ("goodwin", "activate goodwin reduction", &opt_goodwin),
136 LC_OPT_ENT_BOOL ("memcopies", "activate memcopy handling", &opt_memcopies),
137 LC_OPT_ENT_BOOL ("memoperands", "activate memoperands", &opt_memoperands),
138 LC_OPT_ENT_ENUM_INT ("remats", "type of remats to insert (none, briggs, noinverse or all)",&remats_var),
139 LC_OPT_ENT_BOOL ("repair_schedule", "repair the schedule by rematting once used nodes",&opt_repair_schedule),
140 LC_OPT_ENT_BOOL ("no_enlage_liveness", "do not enlarge liveness of operands of remats",&opt_no_enlarge_liveness),
141 LC_OPT_ENT_BOOL ("remat_while_live", "remat only values that can be used by real ops", &opt_remat_while_live),
143 LC_OPT_ENT_ENUM_MASK("dump", "dump problem, mps or solution", &dump_var),
144 LC_OPT_ENT_BOOL ("log", "activate the lpp log", &opt_log),
145 LC_OPT_ENT_INT ("timeout", "ILP solver timeout", &opt_timeout),
147 LC_OPT_ENT_DBL ("cost_reload", "cost of a reload", &opt_cost_reload),
148 LC_OPT_ENT_DBL ("cost_memoperand", "cost of a memory operand", &opt_cost_memoperand),
149 LC_OPT_ENT_DBL ("cost_spill", "cost of a spill instruction", &opt_cost_spill),
150 LC_OPT_ENT_DBL ("cost_remat", "cost of a rematerialization", &opt_cost_remat),
154 void be_spill_remat_register_options(lc_opt_entry_t *grp)
156 lc_opt_entry_t *my_grp = lc_opt_get_grp(grp, "remat");
157 lc_opt_add_table(my_grp, options);
162 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
163 //#define SCHEDULE_PHIM /* insert phim nodes into schedule */
166 //#define SOLVE_LOCAL
167 #define LPP_SERVER "i44pc23"
168 #define LPP_SOLVER "cplex"
171 #define MAX_PATHS INT_MAX
174 typedef struct _spill_ilp_t {
175 const arch_register_class_t *cls;
177 const be_chordal_env_t *chordal_env;
180 struct obstack *obst;
182 pset *all_possible_remats;
185 set *values; /**< for collecting all definitions of values before running ssa-construction */
190 DEBUG_ONLY(firm_dbg_module_t * dbg);
193 typedef int ilp_var_t;
194 typedef int ilp_cst_t;
196 typedef struct _spill_bb_t {
201 typedef struct _remat_t {
202 const ir_node *op; /**< for copy_irn */
203 const ir_node *value; /**< the value which is being recomputed by this remat */
204 const ir_node *proj; /**< not NULL if the above op produces a tuple */
205 int cost; /**< cost of this remat */
206 int inverse; /**< nonzero if this is an inverse remat */
210 * Data to be attached to each IR node. For remats this contains the ilp_var
211 * for this remat and for normal ops this contains the ilp_vars for
212 * reloading each operand
214 typedef struct _op_t {
219 const remat_t *remat; /** the remat this op belongs to */
220 int pre; /** 1, if this is a pressure-increasing remat */
224 ir_node *op; /** the operation this live range belongs to */
233 typedef struct _defs_t {
234 const ir_node *value;
235 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
236 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
239 typedef struct _remat_info_t {
240 const ir_node *irn; /**< the irn to which these remats belong */
241 pset *remats; /**< possible remats for this value */
242 pset *remats_by_operand; /**< remats with this value as operand */
245 typedef struct _keyval_t {
250 typedef struct _spill_t {
259 typedef struct _memoperand_t {
260 ir_node *irn; /**< the irn */
261 unsigned int pos; /**< the position of the argument */
262 ilp_var_t ilp; /**< the ilp var for this memory operand */
266 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
268 return chordal_has_class(si->chordal_env, irn);
273 cmp_remat(const void *a, const void *b)
275 const keyval_t *p = a;
276 const keyval_t *q = b;
277 const remat_t *r = p->val;
278 const remat_t *s = q->val;
282 return !(r == s || r->op == s->op);
286 cmp_remat(const void *a, const void *b)
288 const remat_t *r = a;
289 const remat_t *s = a;
291 return !(r == s || r->op == s->op);
295 cmp_spill(const void *a, const void *b, size_t size)
297 const spill_t *p = a;
298 const spill_t *q = b;
300 // return !(p->irn == q->irn && p->bb == q->bb);
301 return !(p->irn == q->irn);
305 cmp_memoperands(const void *a, const void *b, size_t size)
307 const memoperand_t *p = a;
308 const memoperand_t *q = b;
310 return !(p->irn == q->irn && p->pos == q->pos);
314 set_find_keyval(set * set, const void * key)
319 return set_find(set, &query, sizeof(query), HASH_PTR(key));
323 set_insert_keyval(set * set, void * key, void * val)
329 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
333 set_find_def(set * set, const ir_node * value)
338 return set_find(set, &query, sizeof(query), HASH_PTR(value));
342 set_insert_def(set * set, const ir_node * value)
349 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
352 static memoperand_t *
353 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
360 return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
363 static memoperand_t *
364 set_find_memoperand(set * set, const ir_node * irn, unsigned int pos)
368 query.irn = (ir_node*)irn;
370 return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
375 set_find_spill(set * set, const ir_node * value)
379 query.irn = (ir_node*)value;
380 return set_find(set, &query, sizeof(query), HASH_PTR(value));
383 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
384 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
385 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
386 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
387 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
390 cmp_remat_info(const void *a, const void *b, size_t size)
392 const remat_info_t *p = a;
393 const remat_info_t *q = b;
395 return !(p->irn == q->irn);
399 cmp_defs(const void *a, const void *b, size_t size)
404 return !(p->value == q->value);
408 cmp_keyval(const void *a, const void *b, size_t size)
410 const keyval_t *p = a;
411 const keyval_t *q = b;
413 return !(p->key == q->key);
417 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
420 #ifndef EXECFREQ_LOOPDEPH
421 return get_block_execfreq(si->chordal_env->exec_freq, get_block(irn)) + FUDGE;
424 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
426 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
431 get_cost(const spill_ilp_t * si, const ir_node * irn)
433 if(be_is_Spill(irn)) {
434 return opt_cost_spill;
435 } else if(be_is_Reload(irn)){
436 return opt_cost_reload;
438 return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
443 * Checks, whether node and its operands have suitable reg classes
446 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
449 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
450 int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
454 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
457 for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
458 ir_node *op = get_irn_n(irn, n);
459 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
462 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
469 * Try to create a remat from @p op with destination value @p dest_value
471 static INLINE remat_t *
472 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
474 remat_t *remat = NULL;
476 // if(!mode_is_datab(get_irn_mode(dest_value)))
479 if(dest_value == op) {
480 const ir_node *proj = NULL;
482 if(is_Proj(dest_value)) {
483 op = get_Proj_pred(op);
487 if(!is_rematerializable(si, op))
490 remat = obstack_alloc(si->obst, sizeof(*remat));
492 remat->cost = get_cost(si, op);
493 remat->value = dest_value;
497 arch_inverse_t inverse;
500 /* get the index of the operand we want to retrieve by the inverse op */
501 for (n = get_irn_arity(op)-1; n>=0; --n) {
502 ir_node *arg = get_irn_n(op, n);
504 if(arg == dest_value) break;
508 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", n, op));
510 /* else ask the backend to give an inverse op */
511 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
514 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
516 assert(inverse.n > 0 && "inverse op should have at least one node");
518 for(i=inverse.n-1; i>=0; --i) {
519 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
523 remat = obstack_alloc(si->obst, sizeof(*remat));
524 remat->op = inverse.nodes[0];
525 remat->cost = inverse.costs;
526 remat->value = dest_value;
527 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
530 assert(is_Proj(remat->proj));
532 assert(0 && "I can not handle remats with more than 2 nodes");
539 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
541 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
549 add_remat(const spill_ilp_t * si, const remat_t * remat)
551 remat_info_t *remat_info,
556 assert(remat->value);
558 query.irn = remat->value;
560 query.remats_by_operand = NULL;
561 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
563 if(remat_info->remats == NULL) {
564 remat_info->remats = new_pset(cmp_remat, 4096);
566 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
568 /* insert the remat into the remats_be_operand set of each argument of the remat op */
569 for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
570 ir_node *arg = get_irn_n(remat->op, n);
574 query.remats_by_operand = NULL;
575 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
577 if(remat_info->remats_by_operand == NULL) {
578 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
580 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
585 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
587 const ir_edge_t *edge = get_irn_out_edge_first(irn);
591 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
594 edge = get_irn_out_edge_next(irn, edge);
601 get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
607 irn = get_Proj_pred(irn);
609 for(n=get_irn_arity(irn)-1; n>=0; --n) {
610 const ir_node *arg = get_irn_n(irn, n);
612 if(has_reg_class(si, arg)) ++ret;
619 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
624 if( has_reg_class(si, op)
625 && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1)
626 && (opt_remats != REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0)
628 remat = get_remat_from_op(si, op, op);
630 add_remat(si, remat);
634 if(opt_remats == REMATS_ALL) {
635 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
637 for (n = get_irn_arity(op)-1; n>=0; --n) {
638 ir_node *arg = get_irn_n(op, n);
640 if(has_reg_class(si, arg)) {
641 /* try to get an inverse remat */
642 remat = get_remat_from_op(si, arg, op);
644 add_remat(si, remat);
652 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
655 ir_node *def_block = get_nodes_block(val);
661 /* if pos is at end of a basic block */
663 ret = (pos == def_block || block_dominates(def_block, pos));
664 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
668 /* else if this is a normal operation */
669 block = get_nodes_block(pos);
670 if(block == def_block) {
671 if(!sched_is_scheduled(val)) return 1;
673 ret = sched_comes_after(val, pos);
674 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
678 ret = block_dominates(def_block, block);
679 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
683 static INLINE ir_node *
684 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
686 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
690 * Returns first non-Phi node of block @p bb
692 static INLINE ir_node *
693 sched_block_first_nonphi(const ir_node * bb)
695 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
699 sched_skip_proj_predicator(const ir_node * irn, void * data)
701 return (is_Proj(irn));
704 static INLINE ir_node *
705 sched_next_nonproj(const ir_node * irn, int forward)
707 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
711 * Returns next operation node (non-Proj) after @p irn
712 * or the basic block of this node
714 static INLINE ir_node *
715 sched_next_op(const ir_node * irn)
717 ir_node *next = sched_next(irn);
722 return sched_next_nonproj(next, 1);
726 * Returns previous operation node (non-Proj) before @p irn
727 * or the basic block of this node
729 static INLINE ir_node *
730 sched_prev_op(const ir_node * irn)
732 ir_node *prev = sched_prev(irn);
737 return sched_next_nonproj(prev, 0);
741 sched_put_after(ir_node * insert, ir_node * irn)
743 if(is_Block(insert)) {
744 insert = sched_block_first_nonphi(insert);
746 insert = sched_next_op(insert);
748 sched_add_before(insert, irn);
752 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
754 if(is_Block(insert)) {
755 insert = sched_block_last_noncf(si, insert);
757 insert = sched_next_nonproj(insert, 0);
758 insert = sched_prev(insert);
760 sched_add_after(insert, irn);
764 * Tells you whether a @p remat can be placed before the irn @p pos
767 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
769 const ir_node *op = remat->op;
775 prev = sched_block_last_noncf(si, pos);
776 prev = sched_next_nonproj(prev, 0);
778 prev = sched_prev_op(pos);
780 /* do not remat if the rematted value is defined immediately before this op */
781 if(prev == remat->op) {
786 /* this should be just fine, the following OP will be using this value, right? */
788 /* only remat AFTER the real definition of a value (?) */
789 if(!value_is_defined_before(si, pos, remat->value)) {
790 // ir_fprintf(stderr, "error(not defined)");
795 for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
796 const ir_node *arg = get_irn_n(op, n);
798 if(opt_no_enlarge_liveness) {
799 if(has_reg_class(si, arg) && live) {
800 res &= pset_find_ptr((pset*)live, arg)?1:0;
802 res &= value_is_defined_before(si, pos, arg);
805 res &= value_is_defined_before(si, pos, arg);
813 * Tells you whether a @p remat can be placed after the irn @p pos
816 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
819 pos = sched_block_first_nonphi(pos);
821 pos = sched_next_op(pos);
824 /* only remat AFTER the real definition of a value (?) */
825 if(!value_is_defined_before(si, pos, remat->value)) {
829 return can_remat_before(si, remat, pos, live);
833 * Collect potetially rematerializable OPs
836 walker_remat_collector(ir_node * irn, void * data)
838 spill_ilp_t *si = data;
840 if(!is_Block(irn) && !is_Phi(irn)) {
841 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
842 get_remats_from_op(si, irn);
847 * Inserts a copy of @p irn before @p pos
850 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
855 bb = is_Block(pos)?pos:get_nodes_block(pos);
856 copy = exact_copy(irn);
858 _set_phi_class(copy, NULL);
859 set_nodes_block(copy, bb);
860 sched_put_before(si, pos, copy);
866 * Inserts a copy of @p irn after @p pos
869 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
874 bb = is_Block(pos)?pos:get_nodes_block(pos);
875 copy = exact_copy(irn);
877 _set_phi_class(copy, NULL);
878 set_nodes_block(copy, bb);
879 sched_put_after(pos, copy);
885 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
889 if(can_remat_after(si, remat, pos, live)) {
894 DBG((si->dbg, LEVEL_3, "\t >inserting remat2 %+F\n", remat->op));
896 copy = insert_copy_after(si, remat->op, pos);
898 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
899 op = obstack_alloc(si->obst, sizeof(*op));
901 op->attr.remat.remat = remat;
902 op->attr.remat.pre = 0;
903 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
905 set_irn_link(copy, op);
906 pset_insert_ptr(si->all_possible_remats, copy);
908 proj_copy = insert_copy_after(si, remat->proj, copy);
909 set_irn_n(proj_copy, 0, copy);
910 set_irn_link(proj_copy, op);
911 pset_insert_ptr(si->all_possible_remats, proj_copy);
923 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
927 if(can_remat_before(si, remat, pos, live)) {
932 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
934 copy = insert_copy_before(si, remat->op, pos);
936 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
937 op = obstack_alloc(si->obst, sizeof(*op));
939 op->attr.remat.remat = remat;
940 op->attr.remat.pre = 1;
941 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
943 set_irn_link(copy, op);
944 pset_insert_ptr(si->all_possible_remats, copy);
946 proj_copy = insert_copy_after(si, remat->proj, copy);
947 set_irn_n(proj_copy, 0, copy);
948 set_irn_link(proj_copy, op);
949 pset_insert_ptr(si->all_possible_remats, proj_copy);
961 get_block_n_succs(const ir_node *block) {
962 const ir_edge_t *edge;
964 assert(edges_activated(current_ir_graph));
966 edge = get_block_succ_first(block);
970 edge = get_block_succ_next(block, edge);
975 is_start_block(const ir_node * bb)
977 return get_irg_start_block(get_irn_irg(bb)) == bb;
981 is_before_frame(const ir_node * bb, const ir_node * irn)
983 const ir_node *frame = get_irg_frame(get_irn_irg(bb));
985 if(is_start_block(bb) && sched_get_time_step(frame) >= sched_get_time_step(irn))
992 is_merge_edge(const ir_node * bb)
994 if(is_start_block(bb))
998 return get_block_n_succs(bb) == 1;
1004 is_diverge_edge(const ir_node * bb)
1006 if(is_start_block(bb))
1010 return get_Block_n_cfgpreds(bb) == 1;
1016 walker_regclass_copy_insertor(ir_node * irn, void * data)
1018 spill_ilp_t *si = data;
1020 if(is_Phi(irn) && has_reg_class(si, irn)) {
1023 for(n=get_irn_arity(irn)-1; n>=0; --n) {
1024 ir_node *phi_arg = get_irn_n(irn, n);
1025 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
1027 if(!has_reg_class(si, phi_arg)) {
1028 ir_node *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
1029 ir_node *pos = sched_block_last_noncf(si, bb);
1030 op_t *op = obstack_alloc(si->obst, sizeof(*op));
1032 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
1033 sched_add_after(pos, copy);
1034 set_irn_n(irn, n, copy);
1037 op->attr.live_range.args.reloads = NULL;
1038 op->attr.live_range.ilp = ILP_UNDEF;
1039 set_irn_link(copy, op);
1047 * Insert (so far unused) remats into the irg to
1048 * recompute the potential liveness of all values
1051 walker_remat_insertor(ir_node * bb, void * data)
1053 spill_ilp_t *si = data;
1054 spill_bb_t *spill_bb;
1057 pset *live = pset_new_ptr_default();
1059 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1061 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1062 ir_node *value = be_lv_get_irn(si->lv, bb, i);
1064 /* add remats at end of block */
1065 if (has_reg_class(si, value)) {
1066 pset_insert_ptr(live, value);
1070 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1071 set_irn_link(bb, spill_bb);
1073 irn = sched_last(bb);
1074 while(!sched_is_end(irn)) {
1081 next = sched_prev(irn);
1083 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
1085 if(is_Phi(irn) || is_Proj(irn)) {
1088 if(has_reg_class(si, irn)) {
1089 pset_remove_ptr(live, irn);
1092 op = obstack_alloc(si->obst, sizeof(*op));
1094 op->attr.live_range.args.reloads = NULL;
1095 op->attr.live_range.ilp = ILP_UNDEF;
1096 set_irn_link(irn, op);
1102 op = obstack_alloc(si->obst, sizeof(*op));
1104 op->attr.live_range.ilp = ILP_UNDEF;
1105 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1106 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1107 set_irn_link(irn, op);
1109 args = pset_new_ptr_default();
1111 /* collect arguments of op */
1112 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1113 ir_node *arg = get_irn_n(irn, n);
1115 pset_insert_ptr(args, arg);
1118 /* set args of op already live in epilog */
1119 pset_foreach(args, arg) {
1120 if(has_reg_class(si, arg)) {
1121 pset_insert_ptr(live, arg);
1124 /* delete defined value from live set */
1125 if(has_reg_class(si, irn)) {
1126 pset_remove_ptr(live, irn);
1130 remat_args = pset_new_ptr_default();
1132 /* insert all possible remats before irn */
1133 pset_foreach(args, arg) {
1134 remat_info_t *remat_info,
1138 /* continue if the operand has the wrong reg class
1140 if(!has_reg_class(si, arg))
1144 query.remats = NULL;
1145 query.remats_by_operand = NULL;
1146 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1152 if(remat_info->remats) {
1153 pset_foreach(remat_info->remats, remat) {
1154 ir_node *remat_irn = NULL;
1156 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
1157 if(opt_remat_while_live) {
1158 if(pset_find_ptr(live, remat->value)) {
1159 remat_irn = insert_remat_before(si, remat, irn, live);
1162 remat_irn = insert_remat_before(si, remat, irn, live);
1165 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1166 ir_node *remat_arg = get_irn_n(remat_irn, n);
1168 if(!has_reg_class(si, remat_arg)) continue;
1170 pset_insert_ptr(remat_args, remat_arg);
1177 /* now we add remat args to op's args because they could also die at this op */
1178 pset_foreach(args,arg) {
1179 if(pset_find_ptr(remat_args, arg)) {
1180 pset_remove_ptr(remat_args, arg);
1183 pset_foreach(remat_args,arg) {
1184 pset_insert_ptr(args, arg);
1187 /* insert all possible remats after irn */
1188 pset_foreach(args, arg) {
1189 remat_info_t *remat_info,
1193 /* continue if the operand has the wrong reg class */
1194 if(!has_reg_class(si, arg))
1198 query.remats = NULL;
1199 query.remats_by_operand = NULL;
1200 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1206 /* do not place post remats after jumps */
1207 if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
1209 if(remat_info->remats_by_operand) {
1210 pset_foreach(remat_info->remats_by_operand, remat) {
1211 /* do not insert remats producing the same value as one of the operands */
1212 if(!pset_find_ptr(args, remat->value)) {
1213 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
1214 if(opt_remat_while_live) {
1215 if(pset_find_ptr(live, remat->value)) {
1216 insert_remat_after(si, remat, irn, live);
1219 insert_remat_after(si, remat, irn, live);
1226 del_pset(remat_args);
1231 /* add remats at end if successor has multiple predecessors */
1232 if(is_merge_edge(bb)) {
1233 pset *live_out = pset_new_ptr_default();
1236 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1237 value = be_lv_get_irn(si->lv, bb, i);
1239 if (has_reg_class(si, value)) {
1240 pset_insert_ptr(live_out, value);
1244 /* add remats at end of block */
1245 pset_foreach(live_out, value) {
1246 remat_info_t *remat_info,
1251 query.remats = NULL;
1252 query.remats_by_operand = NULL;
1253 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1255 if(remat_info && remat_info->remats) {
1256 pset_foreach(remat_info->remats, remat) {
1257 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1259 insert_remat_before(si, remat, bb, live_out);
1266 if(is_diverge_edge(bb)) {
1267 pset *live_in = pset_new_ptr_default();
1270 be_lv_foreach(si->lv, bb, be_lv_state_in, i) {
1271 value = be_lv_get_irn(si->lv, bb, i);
1273 if(has_reg_class(si, value)) {
1274 pset_insert_ptr(live_in, value);
1277 sched_foreach(bb, value) {
1278 if(!is_Phi(value)) break;
1280 if(has_reg_class(si, value)) {
1281 pset_insert_ptr(live_in, value);
1285 /* add remat2s at beginning of block */
1286 pset_foreach(live_in, value) {
1287 remat_info_t *remat_info,
1292 query.remats = NULL;
1293 query.remats_by_operand = NULL;
1294 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1296 if(remat_info && remat_info->remats_by_operand) {
1297 pset_foreach(remat_info->remats_by_operand, remat) {
1298 DBG((si->dbg, LEVEL_4, "\t considering remat2 %+F at beginning of block %+F\n", remat->op, bb));
1300 /* put the remat here if all its args are available */
1301 insert_remat_after(si, remat, bb, live_in);
1311 can_be_copied(const ir_node * bb, const ir_node * irn)
1313 assert(is_merge_edge(bb));
1315 const ir_edge_t *edge = get_block_succ_first(bb);
1316 const ir_node *next_bb = edge->src;
1317 int pos = edge->pos;
1320 sched_foreach(next_bb, phi) {
1321 const ir_node *phi_arg;
1323 if(!is_Phi(phi)) break;
1325 phi_arg = get_irn_n(phi, pos);
1327 if(phi_arg == irn) {
1335 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1338 luke_endwalker(ir_node * bb, void * data)
1340 spill_ilp_t *si = (spill_ilp_t*)data;
1346 spill_bb_t *spill_bb = get_irn_link(bb);
1350 live = pset_new_ptr_default();
1351 use_end = pset_new_ptr_default();
1353 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1354 irn = be_lv_get_irn(si->lv, bb, i);
1355 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1358 pset_insert_ptr(live, irn);
1359 op = get_irn_link(irn);
1360 assert(!op->is_remat);
1364 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1365 /* their reg_out must always be set */
1366 sched_foreach_reverse(bb, irn) {
1369 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1371 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1372 ir_node *irn_arg = get_irn_n(irn, n);
1374 if(has_reg_class(si, irn_arg)) {
1375 pset_insert_ptr(use_end, irn_arg);
1380 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1381 //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1382 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1384 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1386 /* if this is a merge edge we can reload at the end of this block */
1387 if(is_merge_edge(bb)) {
1388 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1389 } else if(pset_count(use_end)){
1390 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1392 spill_bb->reloads = NULL;
1395 pset_foreach(live,irn) {
1399 int default_spilled;
1402 /* handle values used by control flow nodes later separately */
1403 if(pset_find_ptr(use_end, irn)) continue;
1406 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1408 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1410 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1411 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1412 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1414 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1415 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1417 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1418 /* by default spill value right after definition */
1419 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1420 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1422 if(is_merge_edge(bb)) {
1426 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1427 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), can_be_copied(bb, irn));
1428 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1430 /* reload <= mem_out */
1431 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1432 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1433 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1436 spill->reg_in = ILP_UNDEF;
1437 spill->mem_in = ILP_UNDEF;
1440 pset_foreach(use_end,irn) {
1444 ilp_cst_t end_use_req,
1447 int default_spilled;
1450 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1452 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1454 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1455 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1456 /* if irn is used at the end of the block, then it is live anyway */
1457 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1459 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1460 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1462 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1463 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1464 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1466 /* reload for use be control flow op */
1467 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1468 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1469 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1471 /* reload <= mem_out */
1472 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1473 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1474 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1476 spill->reg_in = ILP_UNDEF;
1477 spill->mem_in = ILP_UNDEF;
1479 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1480 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1481 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1489 next_post_remat(const ir_node * irn)
1495 next = sched_block_first_nonphi(irn);
1497 next = sched_next_op(irn);
1500 if(sched_is_end(next))
1503 op = get_irn_link(next);
1504 if(op->is_remat && !op->attr.remat.pre) {
1513 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1519 ret = sched_block_last_noncf(si, irn);
1520 ret = sched_next(ret);
1521 ret = sched_prev_op(ret);
1523 ret = sched_prev_op(irn);
1526 if(sched_is_end(ret) || is_Phi(ret))
1529 op = (op_t*)get_irn_link(ret);
1530 if(op->is_remat && op->attr.remat.pre) {
1538 * Find a remat of value @p value in the epilog of @p pos
1541 find_post_remat(const ir_node * value, const ir_node * pos)
1543 while((pos = next_post_remat(pos)) != NULL) {
1546 op = get_irn_link(pos);
1547 assert(op->is_remat && !op->attr.remat.pre);
1549 if(op->attr.remat.remat->value == value)
1550 return (ir_node*)pos;
1553 const ir_edge_t *edge;
1554 foreach_out_edge(pos, edge) {
1555 ir_node *proj = get_edge_src_irn(edge);
1556 assert(is_Proj(proj));
1566 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1568 spill_bb_t *spill_bb = get_irn_link(bb);
1572 int default_spilled;
1575 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1577 double spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1579 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1581 spill->reg_out = ILP_UNDEF;
1582 spill->reg_in = ILP_UNDEF;
1583 spill->mem_in = ILP_UNDEF;
1585 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1586 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1588 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1589 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1590 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1597 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1602 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1603 irn = be_lv_get_irn(si->lv, bb, i);
1605 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1606 pset_insert_ptr(live, irn);
1610 irn = sched_last(bb);
1612 /* all values eaten by control flow operations are also live until the end of the block */
1613 sched_foreach_reverse(bb, irn) {
1616 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1618 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1619 ir_node *arg = get_irn_n(irn,i);
1621 if(has_reg_class(si, arg)) {
1622 pset_insert_ptr(live, arg);
1629 * Inserts ILP-constraints and variables for memory copying before the given position
1632 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1634 const ir_node *succ;
1635 const ir_edge_t *edge;
1636 spill_bb_t *spill_bb = get_irn_link(block);
1645 assert(edges_activated(current_ir_graph));
1647 edge = get_block_succ_first(block);
1653 edge = get_block_succ_next(block, edge);
1654 /* next block can only contain phis, if this is a merge edge */
1657 ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1658 copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1660 ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1661 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1663 pset_foreach(live, tmp) {
1666 op_t *op = get_irn_link(irn);
1667 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1669 spill = set_find_spill(spill_bb->ilp, tmp);
1672 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1674 lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1676 sched_foreach(succ, phi) {
1677 const ir_node *to_copy;
1679 spill_t *to_copy_spill;
1680 op_t *phi_op = get_irn_link(phi);
1681 ilp_var_t reload = ILP_UNDEF;
1684 if(!is_Phi(phi)) break;
1685 if(!has_reg_class(si, phi)) continue;
1687 to_copy = get_irn_n(phi, pos);
1689 to_copy_op = get_irn_link(to_copy);
1691 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1692 assert(to_copy_spill);
1694 if(spill_bb->reloads) {
1695 keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1698 reload = PTR_TO_INT(keyval->val);
1702 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1703 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1705 /* copy - reg_out - reload - remat - live_range <= 0 */
1706 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1707 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1708 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1709 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1710 foreach_pre_remat(si, block, tmp) {
1711 op_t *remat_op = get_irn_link(tmp);
1712 if(remat_op->attr.remat.remat->value == to_copy) {
1713 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1717 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1718 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1720 /* copy - reg_out - copyreg <= 0 */
1721 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1722 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1723 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1729 * Walk all irg blocks and emit this ILP
1732 luke_blockwalker(ir_node * bb, void * data)
1734 spill_ilp_t *si = (spill_ilp_t*)data;
1739 spill_bb_t *spill_bb = get_irn_link(bb);
1742 pset *defs = pset_new_ptr_default();
1743 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
1746 live = pset_new_ptr_default();
1748 /****************************************
1749 * B A S I C B L O C K E N D
1750 ***************************************/
1753 /* init live values at end of block */
1754 get_live_end(si, bb, live);
1756 pset_foreach(live, irn) {
1758 ilp_var_t reload = ILP_UNDEF;
1760 spill = set_find_spill(spill_bb->ilp, irn);
1763 if(spill_bb->reloads) {
1764 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1767 reload = PTR_TO_INT(keyval->val);
1771 op = get_irn_link(irn);
1772 assert(!op->is_remat);
1774 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1775 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1776 op->attr.live_range.op = bb;
1778 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1779 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1781 /* reg_out - reload - remat - live_range <= 0 */
1782 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1783 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1784 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1785 foreach_pre_remat(si, bb, tmp) {
1786 op_t *remat_op = get_irn_link(tmp);
1787 if(remat_op->attr.remat.remat->value == irn) {
1788 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1791 ir_snprintf(buf, sizeof(buf), "reg_out2_%N_%N", bb, irn);
1792 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
1794 /* value may only die at bb end if it is used for a mem copy */
1795 /* reg_out + \sum copy - reload - remat - live_range >= 0 */
1796 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1797 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1798 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1799 foreach_pre_remat(si, bb, tmp) {
1800 op_t *remat_op = get_irn_link(tmp);
1801 if(remat_op->attr.remat.remat->value == irn) {
1802 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1805 if(is_merge_edge(bb)) {
1806 const ir_edge_t *edge = get_block_succ_first(bb);
1807 const ir_node *next_bb = edge->src;
1808 int pos = edge->pos;
1811 sched_foreach(next_bb, phi) {
1812 const ir_node *phi_arg;
1814 if(!is_Phi(phi)) break;
1816 phi_arg = get_irn_n(phi, pos);
1818 if(phi_arg == irn) {
1819 op_t *phi_op = get_irn_link(phi);
1820 ilp_var_t copy = phi_op->attr.live_range.args.copies[pos];
1822 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
1829 insert_mem_copy_position(si, live, bb);
1832 * start new live ranges for values used by remats at end of block
1833 * and assure the remat args are available
1835 foreach_pre_remat(si, bb, tmp) {
1836 op_t *remat_op = get_irn_link(tmp);
1839 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1840 ir_node *remat_arg = get_irn_n(tmp, n);
1841 op_t *arg_op = get_irn_link(remat_arg);
1844 if(!has_reg_class(si, remat_arg)) continue;
1846 /* if value is becoming live through use by remat */
1847 if(!pset_find_ptr(live, remat_arg)) {
1848 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1849 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1851 arg_op->attr.live_range.ilp = prev_lr;
1852 arg_op->attr.live_range.op = bb;
1854 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1856 pset_insert_ptr(live, remat_arg);
1857 add_to_spill_bb(si, bb, remat_arg);
1860 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1861 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1862 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1864 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1865 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1867 /* use reload placed for this argument */
1868 if(spill_bb->reloads) {
1869 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1872 ilp_var_t reload = PTR_TO_INT(keyval->val);
1874 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1879 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1884 /**************************************
1885 * B A S I C B L O C K B O D Y
1886 **************************************/
1888 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1894 ilp_cst_t check_pre,
1900 ilp_cst_t one_memoperand;
1902 /* iterate only until first phi */
1906 op = get_irn_link(irn);
1908 if(op->is_remat) continue;
1909 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1911 /* collect defined values */
1912 if(has_reg_class(si, irn)) {
1913 pset_insert_ptr(defs, irn);
1917 if(is_Proj(irn)) continue;
1920 * init set of irn's arguments
1921 * and all possibly used values around this op
1922 * and values defined by post remats
1924 args = new_set(cmp_keyval, get_irn_arity(irn));
1925 used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1926 remat_defs = pset_new_ptr(pset_count(live));
1928 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1929 ir_node *irn_arg = get_irn_n(irn, n);
1930 if(has_reg_class(si, irn_arg)) {
1931 set_insert_keyval(args, irn_arg, (void*)n);
1932 pset_insert_ptr(used, irn_arg);
1935 foreach_post_remat(irn, tmp) {
1936 op_t *remat_op = get_irn_link(tmp);
1938 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1940 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1941 ir_node *remat_arg = get_irn_n(tmp, n);
1942 if(has_reg_class(si, remat_arg)) {
1943 pset_insert_ptr(used, remat_arg);
1947 foreach_pre_remat(si, irn, tmp) {
1948 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1949 ir_node *remat_arg = get_irn_n(tmp, n);
1950 if(has_reg_class(si, remat_arg)) {
1951 pset_insert_ptr(used, remat_arg);
1956 /**********************************
1957 * I N E P I L O G O F irn
1958 **********************************/
1960 /* ensure each dying value is used by only one post remat */
1961 pset_foreach(used, tmp) {
1962 ir_node *value = tmp;
1963 op_t *value_op = get_irn_link(value);
1968 foreach_post_remat(irn, remat) {
1969 op_t *remat_op = get_irn_link(remat);
1971 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1972 ir_node *remat_arg = get_irn_n(remat, n);
1974 /* if value is used by this remat add it to constraint */
1975 if(remat_arg == value) {
1977 /* sum remat2s <= 1 + n_remats*live_range */
1978 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1979 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
1983 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1989 if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
1990 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1994 /* ensure at least one value dies at post remat */
1995 foreach_post_remat(irn, tmp) {
1996 op_t *remat_op = get_irn_link(tmp);
1997 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
2000 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2001 remat_arg = get_irn_n(tmp, n);
2003 if(has_reg_class(si, remat_arg)) {
2005 /* does arg always die at this op? */
2006 if(!pset_find_ptr(live, remat_arg))
2007 goto skip_one_must_die;
2009 pset_insert_ptr(remat_args, remat_arg);
2013 /* remat + \sum live_range(remat_arg) <= |args| */
2014 ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
2015 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2016 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2018 pset_foreach(remat_args, remat_arg) {
2019 op_t *arg_op = get_irn_link(remat_arg);
2021 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2025 del_pset(remat_args);
2028 /* new live ranges for values from L\U defined by post remats */
2029 pset_foreach(live, tmp) {
2030 ir_node *value = tmp;
2031 op_t *value_op = get_irn_link(value);
2033 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
2034 ilp_var_t prev_lr = ILP_UNDEF;
2037 if(pset_find_ptr(remat_defs, value)) {
2039 /* next_live_range <= prev_live_range + sum remat2s */
2040 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
2041 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2043 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
2044 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2046 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
2047 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2049 foreach_post_remat(irn, remat) {
2050 op_t *remat_op = get_irn_link(remat);
2052 /* if value is being rematerialized by this remat */
2053 if(value == remat_op->attr.remat.remat->value) {
2054 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2058 value_op->attr.live_range.ilp = prev_lr;
2059 value_op->attr.live_range.op = irn;
2064 /* requirements for post remats and start live ranges from L/U' for values dying here */
2065 foreach_post_remat(irn, tmp) {
2066 op_t *remat_op = get_irn_link(tmp);
2069 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2070 ir_node *remat_arg = get_irn_n(tmp, n);
2071 op_t *arg_op = get_irn_link(remat_arg);
2073 if(!has_reg_class(si, remat_arg)) continue;
2075 /* only for values in L\U (TODO and D?), the others are handled with post_use */
2076 if(!pset_find_ptr(used, remat_arg)) {
2077 /* remat <= live_range(remat_arg) */
2078 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
2079 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2081 /* if value is becoming live through use by remat2 */
2082 if(!pset_find_ptr(live, remat_arg)) {
2085 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
2086 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2088 arg_op->attr.live_range.ilp = lr;
2089 arg_op->attr.live_range.op = irn;
2091 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
2093 pset_insert_ptr(live, remat_arg);
2094 add_to_spill_bb(si, bb, remat_arg);
2097 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2098 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2103 d = pset_count(defs);
2104 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
2106 /* count how many regs irn needs for arguments */
2107 u = set_count(args);
2110 /* check the register pressure in the epilog */
2111 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
2112 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
2113 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
2115 /* add L\U' to check_post */
2116 pset_foreach(live, tmp) {
2117 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
2118 /* if a live value is not used by irn */
2119 tmp_op = get_irn_link(tmp);
2120 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
2124 /***********************************************************
2125 * 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
2126 **********************************************************/
2129 pset_foreach(used, tmp) {
2135 op_t *arg_op = get_irn_link(arg);
2138 spill = add_to_spill_bb(si, bb, arg);
2140 /* new live range for each used value */
2141 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2142 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, is_before_frame(bb, irn)?1.0:0.0);
2144 /* the epilog stuff - including post_use, check_post, check_post_remat */
2145 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2146 post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, is_before_frame(bb, irn)?1.0:0.0);
2148 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2150 /* arg is live throughout epilog if the next live_range is in a register */
2151 if(pset_find_ptr(live, arg)) {
2152 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
2154 /* post_use >= next_lr + remat */
2155 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2156 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2157 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2158 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2161 /* forall post remat which use arg add a similar cst */
2162 foreach_post_remat(irn, remat) {
2165 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2166 ir_node *remat_arg = get_irn_n(remat, n);
2167 op_t *remat_op = get_irn_link(remat);
2169 if(remat_arg == arg) {
2170 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
2172 /* post_use >= remat */
2173 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2174 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2175 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2176 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2181 /* if value is not an arg of op and not possibly defined by post remat
2182 * then it may only die and not become live
2184 if(!set_find_keyval(args, arg)) {
2185 /* post_use <= prev_lr */
2186 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2187 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2188 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2189 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2191 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2192 /* next_lr <= prev_lr */
2193 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2194 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2195 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2196 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2200 if(opt_memoperands) {
2201 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2202 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2203 ilp_var_t memoperand;
2205 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2206 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0);
2207 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2209 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2210 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2212 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2213 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2218 /* new live range begins for each used value */
2219 arg_op->attr.live_range.ilp = prev_lr;
2220 arg_op->attr.live_range.op = irn;
2222 pset_insert_ptr(live, arg);
2225 /* just to be sure */
2226 check_post = ILP_UNDEF;
2235 /* check the register pressure in the prolog */
2236 /* sum_{L\U} lr <= k - |U| */
2237 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2238 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2240 /* for the prolog remove defined values from the live set */
2241 pset_foreach(defs, tmp) {
2242 pset_remove_ptr(live, tmp);
2245 if(opt_memoperands) {
2246 ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2247 one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2250 /***********************************************************
2251 * 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
2252 **********************************************************/
2255 set_foreach(args, keyval) {
2257 const ir_node *arg = keyval->key;
2258 int i = PTR_TO_INT(keyval->val);
2259 op_t *arg_op = get_irn_link(arg);
2260 ilp_cst_t requirements;
2263 spill = set_find_spill(spill_bb->ilp, arg);
2266 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2267 op->attr.live_range.args.reloads[i] = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), is_before_frame(bb, irn)?0.0:1.0);
2269 /* reload <= mem_out */
2270 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2271 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2272 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2273 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2275 /* requirement: arg must be in register for use */
2276 /* reload + remat + live_range == 1 */
2277 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2278 requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2280 lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2281 lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2282 foreach_pre_remat(si, irn, tmp) {
2283 op_t *remat_op = get_irn_link(tmp);
2284 if(remat_op->attr.remat.remat->value == arg) {
2285 lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2289 if(opt_memoperands) {
2291 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2292 if(get_irn_n(irn, n) == arg) {
2296 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2297 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2298 memoperand_t *memoperand;
2299 memoperand = set_find_memoperand(si->memoperands, irn, n);
2301 /* memoperand <= mem_out */
2302 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2303 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2304 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2305 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2307 /* the memoperand is only sufficient if it is used once by the op */
2308 if(n_memoperands == 1)
2309 lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2311 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2313 /* we have one more free register if we use a memory operand */
2314 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2320 /* iterate over L\U */
2321 pset_foreach(live, tmp) {
2322 if(!set_find_keyval(args, tmp)) {
2323 /* if a live value is not used by irn */
2324 tmp_op = get_irn_link(tmp);
2325 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2330 /* requirements for remats */
2331 foreach_pre_remat(si, irn, tmp) {
2332 op_t *remat_op = get_irn_link(tmp);
2335 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2336 ir_node *remat_arg = get_irn_n(tmp, n);
2337 op_t *arg_op = get_irn_link(remat_arg);
2339 if(!has_reg_class(si, remat_arg)) continue;
2341 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2342 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2343 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2345 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2346 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2348 /* if remat arg is also used by current op then we can use reload placed for this argument */
2349 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2350 int index = (int)keyval->val;
2352 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2360 /*************************
2361 * D O N E W I T H O P
2362 *************************/
2364 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2366 pset_foreach(live, tmp) {
2367 assert(has_reg_class(si, tmp));
2370 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2371 ir_node *arg = get_irn_n(irn, n);
2373 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2376 del_pset(remat_defs);
2380 defs = pset_new_ptr_default();
2385 /***************************************
2386 * B E G I N N I N G O F B L O C K
2387 ***************************************/
2390 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2391 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2393 pset_foreach(live, irn) {
2394 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2397 /* construct mem_outs for all values */
2399 set_foreach(spill_bb->ilp, spill) {
2400 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2401 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2403 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2404 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2406 if(pset_find_ptr(live, spill->irn)) {
2407 int default_spilled;
2408 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2410 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2411 default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2412 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2413 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2415 if(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2417 op_t *op = get_irn_link(spill->irn);
2419 /* do we have to copy a phi argument? */
2420 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2421 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2423 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2424 const ir_node *arg = get_irn_n(spill->irn, n);
2430 /* argument already done? */
2431 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2433 /* get sum of execution frequencies of blocks with the same phi argument */
2434 for(m=n; m>=0; --m) {
2435 const ir_node *arg2 = get_irn_n(spill->irn, m);
2438 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2442 /* copies are not for free */
2443 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2444 var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
2446 for(m=n; m>=0; --m) {
2447 const ir_node *arg2 = get_irn_n(spill->irn, m);
2450 op->attr.live_range.args.copies[m] = var;
2455 /* copy <= mem_in */
2456 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2457 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2458 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2459 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2466 foreach_post_remat(bb, tmp) {
2468 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
2469 op_t *remat_op = get_irn_link(tmp);
2472 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2473 remat_arg = get_irn_n(tmp, n);
2474 if(has_reg_class(si, remat_arg)) {
2475 pset_insert_ptr(remat_args, remat_arg);
2478 assert(pset_count(remat_args) > 0 && "post remats should have at least one arg");
2480 /* remat + \sum live_range(remat_arg) <= |args| */
2481 ir_snprintf(buf, sizeof(buf), "one_must_die_%N", tmp);
2482 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2483 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2485 pset_foreach(remat_args, remat_arg) {
2486 /* if value is becoming live through use by remat2 */
2487 if(!pset_find_ptr(live, remat_arg)) {
2488 op_t *remat_arg_op = get_irn_link(remat_arg);
2491 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 at bb start %+F\n", remat_arg, tmp));
2493 pset_insert_ptr(live, remat_arg);
2494 spill = add_to_spill_bb(si, bb, remat_arg);
2495 remat_arg_op->attr.live_range.ilp = ILP_UNDEF;
2497 /* we need reg_in and mem_in for this value; they will be referenced later */
2498 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", remat_arg, bb);
2499 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2500 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", remat_arg, bb);
2501 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2504 /* optimization: all memory stuff should be 0, for we do not want to insert reloads for remats */
2505 ir_snprintf(buf, sizeof(buf), "nomem_%N_%N", remat_arg, bb);
2506 nomem = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0.0);
2508 lpp_set_factor_fast(si->lpp, nomem, spill->spill, 1.0);
2509 if(spill_bb->reloads) {
2510 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
2513 ilp_var_t reload = PTR_TO_INT(keyval->val);
2514 lpp_set_factor_fast(si->lpp, nomem, reload, 1.0);
2518 op_t *remat_arg_op = get_irn_link(remat_arg);
2519 lpp_set_factor_fast(si->lpp, cst, remat_arg_op->attr.live_range.ilp, 1.0);
2522 del_pset(remat_args);
2525 /* L\U is empty at bb start */
2526 /* arg is live throughout epilog if it is reg_in into this block */
2528 /* check the register pressure at the beginning of the block
2531 /* reg_in entspricht post_use */
2533 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2534 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2536 pset_foreach(live, irn) {
2539 spill = set_find_spill(spill_bb->ilp, irn);
2542 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2543 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2545 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2547 /* spill + mem_in <= 1 */
2548 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2549 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2551 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2552 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2554 } /* post_remats are NOT included in register pressure check because
2555 they do not increase regpressure */
2557 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2558 pset_foreach(live, irn) {
2562 spill = set_find_spill(spill_bb->ilp, irn);
2563 assert(spill && spill->irn == irn);
2565 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2566 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2569 ir_node *phi_arg = get_Phi_pred(irn, n);
2570 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2571 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2573 op_t *op = get_irn_link(irn);
2575 /* although the phi is in the right regclass one or more of
2576 * its arguments can be in a different one or at least to
2579 if(has_reg_class(si, phi_arg)) {
2580 /* mem_in < mem_out_arg + copy */
2581 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2582 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2584 /* reg_in < reg_out_arg */
2585 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2586 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2588 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2589 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2591 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2594 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2596 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2598 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2602 /* else assure the value arrives on all paths in the same resource */
2604 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2607 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2608 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2611 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2612 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2613 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2614 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2616 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2617 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2619 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2622 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2623 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2628 foreach_post_remat(bb, tmp) {
2631 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2632 ir_node *remat_arg = get_irn_n(tmp, n);
2633 op_t *remat_op = get_irn_link(tmp);
2635 if(!has_reg_class(si, remat_arg)) continue;
2637 spill = set_find_spill(spill_bb->ilp, remat_arg);
2640 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2641 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2642 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2643 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2647 pset_foreach(live, irn) {
2648 const op_t *op = get_irn_link(irn);
2649 const ir_node *remat;
2654 foreach_post_remat(bb, remat) {
2657 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2658 const ir_node *arg = get_irn_n(remat, n);
2661 const op_t *remat_op = get_irn_link(remat);
2663 if(cst == ILP_UNDEF) {
2664 /* sum remat2s <= 1 + n_remats*live_range */
2665 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", irn, bb);
2666 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2668 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2674 if(cst != ILP_UNDEF && op->attr.live_range.ilp != ILP_UNDEF) {
2675 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -n_remats);
2679 /* first live ranges from reg_ins */
2680 pset_foreach(live, irn) {
2681 op_t *op = get_irn_link(irn);
2683 if(op->attr.live_range.ilp != ILP_UNDEF) {
2685 spill = set_find_spill(spill_bb->ilp, irn);
2686 assert(spill && spill->irn == irn);
2688 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2689 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2690 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2691 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2693 foreach_post_remat(bb, tmp) {
2694 op_t *remat_op = get_irn_link(tmp);
2696 if(remat_op->attr.remat.remat->value == irn) {
2697 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2703 /* walk forward now and compute constraints for placing spills */
2704 /* this must only be done for values that are not defined in this block */
2705 /* TODO are these values at start of block? if yes, just check whether this is a diverge edge and skip the loop */
2706 pset_foreach(live, irn) {
2708 * if value is defined in this block we can anways place the spill directly after the def
2709 * -> no constraint necessary
2711 if(!is_Phi(irn) && get_nodes_block(irn) == bb) continue;
2714 spill = set_find_spill(spill_bb->ilp, irn);
2717 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2718 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2720 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2721 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2724 sched_foreach_op(bb, tmp) {
2725 op_t *op = get_irn_link(tmp);
2727 if(is_Phi(tmp)) continue;
2728 assert(!is_Proj(tmp));
2731 const ir_node *value = op->attr.remat.remat->value;
2734 /* only collect remats up to the first real use of a value */
2735 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2740 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2741 ir_node *arg = get_irn_n(tmp, n);
2744 /* if a value is used stop collecting remats */
2757 typedef struct _irnlist_t {
2758 struct list_head list;
2762 typedef struct _interference_t {
2763 struct list_head blocklist;
2769 cmp_interference(const void *a, const void *b, size_t size)
2771 const interference_t *p = a;
2772 const interference_t *q = b;
2774 return !(p->a == q->a && p->b == q->b);
2777 static interference_t *
2778 set_find_interference(set * set, ir_node * a, ir_node * b)
2780 interference_t query;
2782 query.a = (a>b)?a:b;
2783 query.b = (a>b)?b:a;
2785 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2788 static interference_t *
2789 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2791 interference_t query,
2793 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2797 result = set_find_interference(set, a, b);
2800 list_add(&list->list, &result->blocklist);
2804 query.a = (a>b)?a:b;
2805 query.b = (a>b)?b:a;
2807 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2809 INIT_LIST_HEAD(&result->blocklist);
2810 list_add(&list->list, &result->blocklist);
2816 values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
2818 const ir_edge_t *edge;
2820 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2821 /* both values are live in, so they interfere */
2825 /* ensure a dominates b */
2826 if(value_dominates(b,a)) {
2832 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2835 /* the following code is stolen from bera.c */
2836 if(be_is_live_end(si->lv, bb, a))
2839 foreach_out_edge(a, edge) {
2840 const ir_node *user = edge->src;
2841 if(get_nodes_block(user) == bb
2844 && !pset_find_ptr(si->inverse_ops, user)
2845 && value_dominates(b, user))
2853 * Walk all irg blocks and collect interfering values inside of phi classes
2856 luke_interferencewalker(ir_node * bb, void * data)
2858 spill_ilp_t *si = (spill_ilp_t*)data;
2861 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2862 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
2863 op_t *a_op = get_irn_link(a);
2866 /* a is only interesting if it is in my register class and if it is inside a phi class */
2867 if (has_reg_class(si, a) && get_phi_class(a)) {
2868 if(a_op->is_remat || pset_find_ptr(si->inverse_ops, a))
2871 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)) {
2872 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
2873 op_t *b_op = get_irn_link(b);
2876 /* a and b are only interesting if they are in the same phi class */
2877 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2878 if(b_op->is_remat || pset_find_ptr(si->inverse_ops, b))
2881 if(values_interfere_in_block(si, bb, a, b)) {
2882 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2883 set_insert_interference(si, si->interferences, a, b, bb);
2891 static unsigned int copy_path_id = 0;
2894 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2901 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2902 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2904 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2906 pset_foreach(copies, ptr) {
2907 copy = PTR_TO_INT(ptr);
2908 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2913 * @parameter copies contains a path of copies which lead us to irn
2914 * @parameter visited contains a set of nodes already visited on this path
2917 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2919 const ir_edge_t *edge;
2920 op_t *op = get_irn_link(irn);
2921 pset *visited_users = pset_new_ptr_default();
2924 if(op->is_remat) return 0;
2926 pset_insert_ptr(visited, irn);
2930 pset *visited_operands = pset_new_ptr(get_irn_arity(irn));
2932 /* visit all operands */
2933 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2934 ir_node *arg = get_irn_n(irn, n);
2935 ilp_var_t copy = op->attr.live_range.args.copies[n];
2937 if(!has_reg_class(si, arg)) continue;
2938 if(pset_find_ptr(visited_operands, arg)) continue;
2939 pset_insert_ptr(visited_operands, arg);
2942 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2943 del_pset(visited_operands);
2944 del_pset(visited_users);
2945 pset_remove_ptr(visited, irn);
2948 pset_insert(copies, INT_TO_PTR(copy), copy);
2949 write_copy_path_cst(si, copies, any_interfere);
2950 pset_remove(copies, INT_TO_PTR(copy), copy);
2951 } else if(!pset_find_ptr(visited, arg)) {
2952 pset_insert(copies, INT_TO_PTR(copy), copy);
2953 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
2954 pset_remove(copies, INT_TO_PTR(copy), copy);
2956 if(paths > MAX_PATHS) {
2957 if(pset_count(copies) == 0) {
2961 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2962 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
2963 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2964 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2965 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2969 del_pset(visited_operands);
2970 del_pset(visited_users);
2971 pset_remove_ptr(visited, irn);
2974 } else if(pset_count(copies) == 0) {
2980 del_pset(visited_operands);
2983 /* visit all uses which are phis */
2984 foreach_out_edge(irn, edge) {
2985 ir_node *user = edge->src;
2986 int pos = edge->pos;
2987 op_t *op = get_irn_link(user);
2990 if(!is_Phi(user)) continue;
2991 if(!has_reg_class(si, user)) continue;
2992 if(pset_find_ptr(visited_users, user)) continue;
2993 pset_insert_ptr(visited_users, user);
2995 copy = op->attr.live_range.args.copies[pos];
2997 if(user == target) {
2998 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2999 del_pset(visited_users);
3000 pset_remove_ptr(visited, irn);
3003 pset_insert(copies, INT_TO_PTR(copy), copy);
3004 write_copy_path_cst(si, copies, any_interfere);
3005 pset_remove(copies, INT_TO_PTR(copy), copy);
3006 } else if(!pset_find_ptr(visited, user)) {
3007 pset_insert(copies, INT_TO_PTR(copy), copy);
3008 paths += find_copy_path(si, user, target, any_interfere, copies, visited);
3009 pset_remove(copies, INT_TO_PTR(copy), copy);
3011 if(paths > MAX_PATHS) {
3012 if(pset_count(copies) == 0) {
3016 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3017 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3018 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3019 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3020 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3024 del_pset(visited_users);
3025 pset_remove_ptr(visited, irn);
3028 } else if(pset_count(copies) == 0) {
3034 del_pset(visited_users);
3035 pset_remove_ptr(visited, irn);
3040 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
3042 pset * copies = pset_new_ptr_default();
3043 pset * visited = pset_new_ptr_default();
3045 find_copy_path(si, a, b, any_interfere, copies, visited);
3053 memcopyhandler(spill_ilp_t * si)
3055 interference_t *interference;
3057 /* teste Speicherwerte auf Interferenz */
3059 /* analyze phi classes */
3060 phi_class_compute(si->chordal_env->irg);
3062 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
3063 irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
3065 /* now lets emit the ILP unequations for the crap */
3066 set_foreach(si->interferences, interference) {
3068 ilp_var_t interfere,
3070 ilp_cst_t any_interfere_cst,
3072 const ir_node *a = interference->a;
3073 const ir_node *b = interference->b;
3075 /* any_interf <= \sum interf */
3076 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
3077 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3078 any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3080 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
3082 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
3083 const ir_node *bb = irnlist->irn;
3084 spill_bb_t *spill_bb = get_irn_link(bb);
3089 spilla = set_find_spill(spill_bb->ilp, a);
3092 spillb = set_find_spill(spill_bb->ilp, b);
3095 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
3096 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
3097 /* 2: - mem_in_a - spill_a + interfere <= 0 */
3098 /* 3: - mem_in_b - spill_b + interfere <= 0 */
3099 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
3100 interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3102 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
3103 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
3105 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
3106 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
3107 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
3108 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
3109 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
3111 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
3112 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3114 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3115 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
3116 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
3118 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
3119 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3121 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3122 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
3123 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
3126 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
3128 /* any_interfere >= interf */
3129 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
3130 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3132 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3133 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3136 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
3137 gen_copy_constraints(si,a,b,any_interfere);
3145 return fabs(x) < 0.00001;
3148 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
3150 spill_ilp_t *si = get_irg_link(current_ir_graph);
3152 if(pset_find_ptr(si->all_possible_remats, n)) {
3153 op_t *op = (op_t*)get_irn_link(n);
3154 assert(op && op->is_remat);
3156 if(!op->attr.remat.remat->inverse) {
3157 if(op->attr.remat.pre) {
3158 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
3160 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
3165 op_t *op = (op_t*)get_irn_link(n);
3166 assert(op && op->is_remat);
3168 if(op->attr.remat.pre) {
3169 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
3171 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
3182 dump_graph_with_remats(ir_graph * irg, const char * suffix)
3184 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
3185 be_dump(irg, suffix, dump_ir_block_graph_sched);
3186 set_dump_node_vcgattr_hook(NULL);
3190 * Edge hook to dump the schedule edges with annotated register pressure.
3193 sched_pressure_edge_hook(FILE *F, ir_node *irn)
3195 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
3196 ir_node *prev = sched_prev(irn);
3197 fprintf(F, "edge:{sourcename:\"");
3199 fprintf(F, "\" targetname:\"");
3201 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
3202 fprintf(F, "\" color:magenta}\n");
3208 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
3210 DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3212 dump_consts_local(0);
3213 set_dump_node_edge_hook(sched_pressure_edge_hook);
3214 dump_ir_block_graph(irg, suffix);
3215 set_dump_node_edge_hook(old_edge_hook);
3219 walker_pressure_annotator(ir_node * bb, void * data)
3221 spill_ilp_t *si = data;
3224 pset *live = pset_new_ptr_default();
3227 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3228 irn = be_lv_get_irn(si->lv, bb, i);
3230 if (has_reg_class(si, irn)) {
3231 pset_insert_ptr(live, irn);
3235 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3237 sched_foreach_reverse(bb, irn) {
3239 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3243 if(has_reg_class(si, irn)) {
3244 pset_remove_ptr(live, irn);
3245 if(is_Proj(irn)) ++projs;
3248 if(!is_Proj(irn)) projs = 0;
3250 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3251 ir_node *arg = get_irn_n(irn, n);
3253 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3255 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3262 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3264 be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
3268 connect_all_remats_with_keep(spill_ilp_t * si)
3276 n_remats = pset_count(si->all_possible_remats);
3278 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3281 pset_foreach(si->all_possible_remats, irn) {
3286 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
3288 obstack_free(si->obst, ins);
3293 connect_all_spills_with_keep(spill_ilp_t * si)
3302 n_spills = pset_count(si->spills);
3304 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3307 pset_foreach(si->spills, irn) {
3312 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
3314 obstack_free(si->obst, ins);
3318 /** insert a spill at an arbitrary position */
3319 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3321 ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
3322 ir_graph *irg = get_irn_irg(bl);
3323 ir_node *frame = get_irg_frame(irg);
3327 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
3328 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3330 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3333 * search the right insertion point. a spill of a phi cannot be put
3334 * directly after the phi, if there are some phis behind the one which
3335 * is spilled. Also, a spill of a Proj must be after all Projs of the
3338 * Here's one special case:
3339 * If the spill is in the start block, the spill must be after the frame
3340 * pointer is set up. This is done by setting insert to the end of the block
3341 * which is its default initialization (see above).
3344 if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3347 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3350 sched_add_after(insert, spill);
3355 delete_remat(spill_ilp_t * si, ir_node * remat) {
3357 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3359 sched_remove(remat);
3361 /* kill links to operands */
3362 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3363 set_irn_n(remat, n, bad);
3368 clean_remat_info(spill_ilp_t * si)
3372 remat_info_t *remat_info;
3373 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3375 set_foreach(si->remat_info, remat_info) {
3376 if(!remat_info->remats) continue;
3378 pset_foreach(remat_info->remats, remat)
3380 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3381 set_irn_n((ir_node*)remat->proj, -1, bad);
3382 set_irn_n((ir_node*)remat->proj, 0, bad);
3385 if(get_irn_n_edges(remat->op) == 0) {
3386 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3387 set_irn_n((ir_node*)remat->op, n, bad);
3392 if(remat_info->remats) del_pset(remat_info->remats);
3393 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3398 delete_unnecessary_remats(spill_ilp_t * si)
3400 if(opt_keep_alive & KEEPALIVE_REMATS) {
3402 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3405 // ir_node *end = get_irg_end(si->chordal_env->irg);
3408 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3409 ir_node *keep_arg = get_irn_n(si->keep, n);
3410 op_t *arg_op = get_irn_link(keep_arg);
3413 assert(arg_op->is_remat);
3415 name = si->lpp->vars[arg_op->attr.remat.ilp];
3417 if(is_zero(name->value)) {
3418 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
3419 /* TODO check whether reload is preferred over remat (could be bug) */
3420 delete_remat(si, keep_arg);
3422 if(!arg_op->attr.remat.remat->inverse) {
3423 if(arg_op->attr.remat.pre) {
3424 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3426 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3429 if(arg_op->attr.remat.pre) {
3430 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3432 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3437 set_irn_n(si->keep, n, bad);
3440 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
3441 ir_node *end_arg = get_End_keepalive(end, i);
3443 if(end_arg != si->keep) {
3444 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
3447 keeps = obstack_finish(si->obst);
3448 set_End_keepalives(end, n-1, keeps);
3449 obstack_free(si->obst, keeps);
3452 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
3457 pset_foreach(si->all_possible_remats, remat) {
3458 op_t *remat_op = get_irn_link(remat);
3459 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
3461 if(is_zero(name->value)) {
3462 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
3463 /* TODO check whether reload is preferred over remat (could be bug) */
3464 delete_remat(si, remat);
3466 if(!remat_op->attr.remat.remat->inverse) {
3467 if(remat_op->attr.remat.pre) {
3468 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3470 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3473 if(remat_op->attr.remat.pre) {
3474 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3476 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3485 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3487 pset *spills = pset_new_ptr_default();
3489 const ir_node *next;
3492 defs = set_find_def(si->values, value);
3494 if(defs && defs->spills) {
3495 for(next = defs->spills; next; next = get_irn_link(next)) {
3496 pset_insert_ptr(spills, next);
3504 * @param before The node after which the spill will be placed in the schedule
3507 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3511 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3513 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3515 spill = be_spill2(arch_env, irn, before);
3517 defs = set_insert_def(si->values, value);
3520 /* enter into the linked list */
3521 set_irn_link(spill, defs->spills);
3522 defs->spills = spill;
3524 if(opt_keep_alive & KEEPALIVE_SPILLS)
3525 pset_insert_ptr(si->spills, spill);
3531 * @param before The Phi node which has to be spilled
3534 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3541 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3543 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3544 ins[n] = si->m_unknown;
3547 mem_phi = new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3549 defs = set_insert_def(si->values, phi);
3552 /* enter into the linked list */
3553 set_irn_link(mem_phi, defs->spills);
3554 defs->spills = mem_phi;
3556 #ifdef SCHEDULE_PHIM
3557 sched_add_after(phi, mem_phi);
3560 if(opt_keep_alive & KEEPALIVE_SPILLS)
3561 pset_insert_ptr(si->spills, mem_phi);
3568 * Add remat to list of defs, destroys link field!
3571 insert_remat(spill_ilp_t * si, ir_node * remat)
3574 op_t *remat_op = get_irn_link(remat);
3576 assert(remat_op->is_remat);
3578 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3581 /* enter into the linked list */
3582 set_irn_link(remat, defs->remats);
3583 defs->remats = remat;
3588 * Add reload before operation and add to list of defs
3591 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3596 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3598 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3600 defs = set_find_def(si->values, value);
3602 spill = defs->spills;
3603 assert(spill && "no spill placed before reload");
3605 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3607 /* enter into the linked list */
3608 set_irn_link(reload, defs->remats);
3609 defs->remats = reload;
3614 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3617 ir_node *value = get_irn_n(memoperand->irn, memoperand->pos);
3619 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3621 DBG((si->dbg, LEVEL_2, "\t inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3623 defs = set_find_def(si->values, value);
3625 spill = defs->spills;
3626 assert(spill && "no spill placed before reload");
3628 arch_perform_memory_operand(arch_env, memoperand->irn, spill, memoperand->pos);
3631 void insert_memoperands(spill_ilp_t * si)
3633 memoperand_t *memoperand;
3636 set_foreach(si->memoperands, memoperand) {
3637 name = si->lpp->vars[memoperand->ilp];
3638 if(!is_zero(name->value)) {
3639 perform_memory_operand(si, memoperand);
3645 walker_spill_placer(ir_node * bb, void * data) {
3646 spill_ilp_t *si = (spill_ilp_t*)data;
3648 spill_bb_t *spill_bb = get_irn_link(bb);
3649 pset *spills_to_do = pset_new_ptr_default();
3652 set_foreach(spill_bb->ilp, spill) {
3655 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3656 name = si->lpp->vars[spill->mem_in];
3657 if(!is_zero(name->value)) {
3660 mem_phi = insert_mem_phi(si, spill->irn);
3662 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3666 name = si->lpp->vars[spill->spill];
3667 if(!is_zero(name->value)) {
3668 /* place spill directly after definition */
3669 if(get_nodes_block(spill->irn) == bb) {
3670 insert_spill(si, spill->irn, spill->irn, spill->irn);
3674 /* place spill at bb start */
3675 if(spill->reg_in > 0) {
3676 name = si->lpp->vars[spill->reg_in];
3677 if(!is_zero(name->value)) {
3678 insert_spill(si, spill->irn, spill->irn, bb);
3682 /* place spill after a remat */
3683 pset_insert_ptr(spills_to_do, spill->irn);
3686 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3689 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3690 op_t *op = get_irn_link(irn);
3692 if(be_is_Spill(irn)) continue;
3695 /* TODO fix this if we want to support remats with more than two nodes */
3696 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3697 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3699 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3702 if(pset_find_ptr(spills_to_do, irn)) {
3703 pset_remove_ptr(spills_to_do, irn);
3705 insert_spill(si, irn, irn, irn);
3711 assert(pset_count(spills_to_do) == 0);
3713 /* afterwards free data in block */
3714 del_pset(spills_to_do);
3718 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3720 ir_node *insert_pos = bb;
3722 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3724 /* find last definition of arg value in block */
3729 defs = set_find_def(si->values, value);
3731 if(defs && defs->remats) {
3732 for(next = defs->remats; next; next = get_irn_link(next)) {
3733 if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3734 last = sched_get_time_step(next);
3740 if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3741 last = sched_get_time_step(value);
3745 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
3747 spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3753 phim_fixer(spill_ilp_t *si) {
3756 set_foreach(si->values, defs) {
3757 const ir_node *phi = defs->value;
3758 op_t *op = get_irn_link(phi);
3759 ir_node *phi_m = NULL;
3760 ir_node *next = defs->spills;
3763 if(!is_Phi(phi)) continue;
3766 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3770 next = get_irn_link(next);
3773 if(!phi_m) continue;
3775 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3776 ir_node *value = get_irn_n(phi, n);
3777 defs_t *val_defs = set_find_def(si->values, value);
3779 /* a spill of this value */
3784 ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3785 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3787 if(!is_zero(name->value)) {
3788 spill = insert_mem_copy(si, pred, value);
3790 spill = val_defs->spills;
3793 spill = val_defs->spills;
3796 assert(spill && "no spill placed before PhiM");
3797 set_irn_n(phi_m, n, spill);
3803 walker_reload_placer(ir_node * bb, void * data) {
3804 spill_ilp_t *si = (spill_ilp_t*)data;
3806 spill_bb_t *spill_bb = get_irn_link(bb);
3808 /* reloads at end of block */
3809 if(spill_bb->reloads) {
3812 set_foreach(spill_bb->reloads, keyval) {
3813 ir_node *irn = (ir_node*)keyval->key;
3814 ilp_var_t reload = PTR_TO_INT(keyval->val);
3817 name = si->lpp->vars[reload];
3818 if(!is_zero(name->value)) {
3820 ir_node *insert_pos = bb;
3821 ir_node *prev = sched_block_last_noncf(si, bb);
3822 op_t *prev_op = get_irn_link(prev);
3824 while(be_is_Spill(prev)) {
3825 prev = sched_prev(prev);
3828 prev_op = get_irn_link(prev);
3830 /* insert reload before pre-remats */
3831 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3832 && prev_op->is_remat && prev_op->attr.remat.pre) {
3836 prev = sched_prev(prev);
3837 } while(be_is_Spill(prev));
3839 prev_op = get_irn_link(prev);
3843 reload = insert_reload(si, irn, insert_pos);
3845 if(opt_keep_alive & KEEPALIVE_RELOADS)
3846 pset_insert_ptr(si->spills, reload);
3851 /* walk and insert more reloads and collect remats */
3852 sched_foreach_reverse(bb, irn) {
3853 op_t *op = get_irn_link(irn);
3855 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3856 if(is_Phi(irn)) break;
3859 if(get_irn_mode(irn) != mode_T) {
3860 insert_remat(si, irn);
3865 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3866 ir_node *arg = get_irn_n(irn, n);
3868 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3871 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3872 if(!is_zero(name->value)) {
3874 ir_node *insert_pos = irn;
3875 ir_node *prev = sched_prev(insert_pos);
3878 while(be_is_Spill(prev)) {
3879 prev = sched_prev(prev);
3882 prev_op = get_irn_link(prev);
3884 /* insert reload before pre-remats */
3885 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3886 && prev_op->is_remat && prev_op->attr.remat.pre) {
3890 prev = sched_prev(prev);
3891 } while(be_is_Spill(prev));
3893 prev_op = get_irn_link(prev);
3897 reload = insert_reload(si, arg, insert_pos);
3899 set_irn_n(irn, n, reload);
3901 if(opt_keep_alive & KEEPALIVE_RELOADS)
3902 pset_insert_ptr(si->spills, reload);
3909 del_set(spill_bb->ilp);
3910 if(spill_bb->reloads) del_set(spill_bb->reloads);
3914 walker_collect_used(ir_node * irn, void * data)
3916 bitset_t *used = data;
3918 bitset_set(used, get_irn_idx(irn));
3921 struct kill_helper {
3927 walker_kill_unused(ir_node * bb, void * data)
3929 struct kill_helper *kh = data;
3930 ir_node *bad = get_irg_bad(get_irn_irg(bb));
3934 for(irn=sched_first(bb); !sched_is_end(irn);) {
3935 ir_node *next = sched_next(irn);
3938 if(!bitset_is_set(kh->used, get_irn_idx(irn))) {
3939 if(be_is_Spill(irn) || be_is_Reload(irn)) {
3940 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)));
3942 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3948 set_nodes_block(irn, bad);
3949 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3950 set_irn_n(irn, n, bad);
3958 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3960 struct kill_helper kh;
3962 kh.used = bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3965 irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3966 irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3968 bitset_free(kh.used);
3972 print_irn_pset(pset * p)
3976 pset_foreach(p, irn) {
3977 ir_printf("%+F\n", irn);
3982 dump_phi_class(spill_ilp_t * si, pset * phiclass, const char * file)
3984 FILE *f = fopen(file, "w");
3986 interference_t *interference;
3988 pset_break(phiclass);
3989 set_break(si->interferences);
3991 ir_fprintf(f, "digraph phiclass {\n");
3993 pset_foreach(phiclass, irn) {
3995 ir_fprintf(f, " %F%N [shape=box]\n",irn,irn);
3998 pset_foreach(phiclass, irn) {
4001 if(!is_Phi(irn)) continue;
4003 for(n=get_irn_arity(irn)-1; n>=0; --n) {
4004 ir_node *arg = get_irn_n(irn, n);
4006 ir_fprintf(f, " %F%N -> %F%N\n",irn,irn,arg,arg);
4010 set_foreach(si->interferences, interference) {
4011 const ir_node *a = interference->a;
4012 const ir_node *b = interference->b;
4013 if(get_phi_class(a) == phiclass) {
4014 ir_fprintf(f, " %F%N -> %F%N [color=red,dir=none,style=bold]\n",a,a,b,b);
4023 rewire_uses(spill_ilp_t * si)
4025 dom_front_info_t *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
4027 pset *ignore = pset_new_ptr(1);
4029 pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
4031 /* then fix uses of spills */
4032 set_foreach(si->values, defs) {
4035 const ir_node *next = defs->remats;
4038 reloads = pset_new_ptr_default();
4041 if(be_is_Reload(next)) {
4042 pset_insert_ptr(reloads, next);
4046 next = get_irn_link(next);
4049 spills = get_spills_for_value(si, defs->value);
4050 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));
4051 if(pset_count(spills) > 1) {
4052 //assert(pset_count(reloads) > 0);
4053 // print_irn_pset(spills);
4054 // print_irn_pset(reloads);
4056 be_ssa_constr_set_ignore(dfi, si->lv, spills, ignore);
4063 /* first fix uses of remats and reloads */
4064 set_foreach(si->values, defs) {
4066 const ir_node *next = defs->remats;
4069 nodes = pset_new_ptr_default();
4070 pset_insert_ptr(nodes, defs->value);
4073 pset_insert_ptr(nodes, next);
4074 next = get_irn_link(next);
4077 if(pset_count(nodes) > 1) {
4078 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
4079 be_ssa_constr_set(dfi, si->lv, nodes);
4086 // remove_unused_defs(si);
4088 be_free_dominance_frontiers(dfi);
4093 writeback_results(spill_ilp_t * si)
4095 /* walk through the graph and collect all spills, reloads and remats for a value */
4097 si->values = new_set(cmp_defs, 4096);
4099 DBG((si->dbg, LEVEL_1, "Applying results\n"));
4100 delete_unnecessary_remats(si);
4101 si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
4102 irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
4103 irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
4105 insert_memoperands(si);
4108 /* clean the remat info! there are still back-edges leading there! */
4109 clean_remat_info(si);
4113 connect_all_spills_with_keep(si);
4115 del_set(si->values);
4119 get_n_regs(spill_ilp_t * si)
4121 int arch_n_regs = arch_register_class_n_regs(si->cls);
4123 bitset_t *arch_regs = bitset_malloc(arch_n_regs);
4124 bitset_t *abi_regs = bitset_malloc(arch_n_regs);
4126 arch_put_non_ignore_regs(si->chordal_env->birg->main_env->arch_env, si->cls, arch_regs);
4127 be_abi_put_ignore_regs(si->chordal_env->birg->abi, si->cls, abi_regs);
4129 bitset_andnot(arch_regs, abi_regs);
4130 arch_n_regs = bitset_popcnt(arch_regs);
4132 bitset_free(arch_regs);
4133 bitset_free(abi_regs);
4135 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", arch_n_regs, si->cls->name));
4140 walker_reload_mover(ir_node * bb, void * data)
4142 spill_ilp_t *si = data;
4145 sched_foreach(bb, tmp) {
4146 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
4147 ir_node *reload = tmp;
4150 /* move reload upwards */
4152 int pressure = (int)get_irn_link(reload);
4153 if(pressure < si->n_regs) {
4154 irn = sched_prev(reload);
4155 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
4156 sched_remove(reload);
4157 pressure = (int)get_irn_link(irn);
4159 while(pressure < si->n_regs) {
4160 if( sched_is_end(irn) ||
4161 (be_is_Reload(irn) && has_reg_class(si, irn)) ||
4162 /* do not move reload before its spill */
4163 (irn == be_get_Reload_mem(reload)) ||
4164 /* do not move before phi */
4167 set_irn_link(irn, INT_TO_PTR(pressure+1));
4168 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
4169 irn = sched_prev(irn);
4171 pressure = (int)get_irn_link(irn);
4174 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
4175 sched_put_after(irn, reload);
4182 move_reloads_upward(spill_ilp_t * si)
4184 irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
4189 * Walk all irg blocks and check for interfering spills inside of phi classes
4192 luke_meminterferencechecker(ir_node * bb, void * data)
4194 spill_ilp_t *si = (spill_ilp_t*)data;
4197 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
4198 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
4200 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
4202 /* a is only interesting if it is in my register class and if it is inside a phi class */
4203 if (has_reg_class(si, a) && get_phi_class(a)) {
4204 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)) {
4205 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
4207 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
4209 /* a and b are only interesting if they are in the same phi class */
4210 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
4211 if(values_interfere_in_block(si, bb, a, b)) {
4212 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4221 verify_phiclasses(spill_ilp_t * si)
4223 /* analyze phi classes */
4224 phi_class_compute(si->chordal_env->irg);
4226 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4227 irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
4231 be_spill_remat(const be_chordal_env_t * chordal_env)
4234 char problem_name[256];
4235 char dump_suffix[256];
4236 char dump_suffix2[256];
4237 struct obstack obst;
4240 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
4241 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
4242 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
4244 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4245 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4247 if(opt_verify & VERIFY_DOMINANCE)
4248 be_check_dominance(chordal_env->irg);
4250 obstack_init(&obst);
4251 si.chordal_env = chordal_env;
4253 si.cls = chordal_env->cls;
4254 si.lpp = new_lpp(problem_name, lpp_minimize);
4255 si.remat_info = new_set(cmp_remat_info, 4096);
4256 si.interferences = new_set(cmp_interference, 32);
4257 si.memoperands = new_set(cmp_memoperands, 128);
4258 si.all_possible_remats = pset_new_ptr_default();
4259 si.spills = pset_new_ptr_default();
4260 si.inverse_ops = pset_new_ptr_default();
4261 si.lv = chordal_env->lv;
4263 si.n_regs = get_n_regs(&si);
4265 set_irg_link(chordal_env->irg, &si);
4266 compute_doms(chordal_env->irg);
4268 /* compute phi classes */
4269 // phi_class_compute(chordal_env->irg);
4271 be_analyze_regpressure(chordal_env, "-pre");
4274 /* collect remats */
4275 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4276 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
4279 /* insert possible remats */
4280 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4281 irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
4282 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4284 if(opt_keep_alive & KEEPALIVE_REMATS) {
4285 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4286 connect_all_remats_with_keep(&si);
4287 /* dump graph with inserted remats */
4288 dump_graph_with_remats(chordal_env->irg, dump_suffix);
4291 /* insert copies for phi arguments not in my regclass */
4292 irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
4294 /* recompute liveness */
4295 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4296 be_liveness_recompute(si.lv);
4300 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4301 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4302 irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
4304 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4305 irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
4308 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4309 memcopyhandler(&si);
4312 if(opt_dump_flags & DUMP_PROBLEM) {
4314 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4315 if ((f = fopen(buf, "wt")) != NULL) {
4316 lpp_dump_plain(si.lpp, f);
4321 if(opt_dump_flags & DUMP_MPS) {
4324 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4325 if((f = fopen(buf, "wt")) != NULL) {
4326 mps_write_mps(si.lpp, s_mps_fixed, f);
4330 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4331 if((f = fopen(buf, "wt")) != NULL) {
4332 mps_write_mst(si.lpp, s_mps_fixed, f);
4337 lpp_check_startvals(si.lpp);
4340 DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4341 lpp_set_time_limit(si.lpp, opt_timeout);
4344 lpp_set_log(si.lpp, stdout);
4347 lpp_solve_cplex(si.lpp);
4349 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4351 assert(lpp_is_sol_valid(si.lpp)
4352 && "solution of ILP must be valid");
4354 DBG((si.dbg, LEVEL_1, "\t%s: iterations: %d, solution time: %g, objective function: %g\n", problem_name, si.lpp->iterations, si.lpp->sol_time, is_zero(si.lpp->objval)?0.0:si.lpp->objval));
4356 if(opt_dump_flags & DUMP_SOLUTION) {
4360 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4361 if ((f = fopen(buf, "wt")) != NULL) {
4363 for (i = 0; i < si.lpp->var_next; ++i) {
4364 lpp_name_t *name = si.lpp->vars[i];
4365 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4371 writeback_results(&si);
4375 kill_all_unused_values_in_schedule(&si);
4377 if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
4378 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
4380 // move reloads upwards
4381 be_liveness_recompute(si.lv);
4382 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4383 move_reloads_upward(&si);
4386 verify_phiclasses(&si);
4389 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4391 dump_pressure_graph(&si, dump_suffix2);
4393 be_analyze_regpressure(chordal_env, "-post");
4395 if(opt_verify & VERIFY_DOMINANCE)
4396 be_check_dominance(chordal_env->irg);
4398 free_dom(chordal_env->irg);
4399 del_set(si.interferences);
4400 del_pset(si.inverse_ops);
4401 del_pset(si.all_possible_remats);
4402 del_set(si.memoperands);
4403 del_pset(si.spills);
4405 obstack_free(&obst, NULL);
4406 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4409 #else /* WITH_ILP */
4412 only_that_you_can_compile_without_WITH_ILP_defined(void)
4416 #endif /* WITH_ILP */