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 "irnodeset.h"
37 #include "irbackedge_t.h"
41 #include <lpp/lpp_net.h>
42 #include <lpp/lpp_cplex.h>
46 #include "besched_t.h"
47 #include "bessaconstr.h"
52 #include "bespillremat.h"
54 #include "bepressurestat.h"
55 #include "beprofile.h"
56 #include "bespilloptions.h"
58 #include "bechordal_t.h"
60 #include <libcore/lc_opts.h>
61 #include <libcore/lc_opts_enum.h>
63 #define DUMP_PROBLEM 1
65 #define DUMP_SOLUTION 4
67 #define DUMP_PRESSURE 16
69 #define KEEPALIVE_REMATS 1
70 #define KEEPALIVE_SPILLS 2
71 #define KEEPALIVE_RELOADS 4
73 #define VERIFY_MEMINTERF 1
74 #define VERIFY_DOMINANCE 2
77 #define REMATS_BRIGGS 1
78 #define REMATS_NOINVERSE 2
81 static unsigned opt_dump_flags = 0;
82 static int opt_log = 0;
83 static unsigned opt_keep_alive = 0;
84 static int opt_goodwin = 1;
85 static int opt_memcopies = 1;
86 static int opt_memoperands = 1;
87 static int opt_verify = VERIFY_MEMINTERF;
88 static unsigned opt_remats = REMATS_ALL;
89 static int opt_repair_schedule = 0;
90 static int opt_no_enlarge_liveness = 0;
91 static int opt_remat_while_live = 1;
92 static int opt_timeout = 300;
93 static double opt_cost_reload = 8.0;
94 static double opt_cost_memoperand = 7.0;
95 static double opt_cost_spill = 15.0;
96 static double opt_cost_remat = 1.0;
99 static const lc_opt_enum_mask_items_t dump_items[] = {
100 { "problem", DUMP_PROBLEM },
102 { "solution", DUMP_SOLUTION },
103 { "stats", DUMP_STATS },
104 { "pressure", DUMP_PRESSURE },
108 static lc_opt_enum_mask_var_t dump_var = {
109 &opt_dump_flags, dump_items
112 static const lc_opt_enum_mask_items_t keepalive_items[] = {
113 { "remats", KEEPALIVE_REMATS },
114 { "spills", KEEPALIVE_SPILLS },
115 { "reloads", KEEPALIVE_RELOADS },
119 static lc_opt_enum_mask_var_t keep_alive_var = {
120 &opt_keep_alive, keepalive_items
123 static const lc_opt_enum_mask_items_t remats_items[] = {
124 { "none", REMATS_NONE },
125 { "briggs", REMATS_BRIGGS },
126 { "noinverse", REMATS_NOINVERSE },
127 { "all", REMATS_ALL },
131 static lc_opt_enum_mask_var_t remats_var = {
132 &opt_remats, remats_items
135 static const lc_opt_table_entry_t options[] = {
136 LC_OPT_ENT_ENUM_MASK("keepalive", "keep alive inserted nodes", &keep_alive_var),
138 LC_OPT_ENT_BOOL ("goodwin", "activate goodwin reduction", &opt_goodwin),
139 LC_OPT_ENT_BOOL ("memcopies", "activate memcopy handling", &opt_memcopies),
140 LC_OPT_ENT_BOOL ("memoperands", "activate memoperands", &opt_memoperands),
141 LC_OPT_ENT_ENUM_INT ("remats", "type of remats to insert", &remats_var),
142 LC_OPT_ENT_BOOL ("repair_schedule", "repair the schedule by rematting once used nodes",&opt_repair_schedule),
143 LC_OPT_ENT_BOOL ("no_enlage_liveness", "do not enlarge liveness of operands of remats",&opt_no_enlarge_liveness),
144 LC_OPT_ENT_BOOL ("remat_while_live", "only remat where rematted value was live", &opt_remat_while_live),
146 LC_OPT_ENT_ENUM_MASK("dump", "dump problem, solution or statistical data", &dump_var),
147 LC_OPT_ENT_BOOL ("log", "activate the lpp log", &opt_log),
148 LC_OPT_ENT_INT ("timeout", "ILP solver timeout", &opt_timeout),
150 LC_OPT_ENT_DBL ("cost_reload", "cost of a reload", &opt_cost_reload),
151 LC_OPT_ENT_DBL ("cost_memoperand", "cost of a memory operand", &opt_cost_memoperand),
152 LC_OPT_ENT_DBL ("cost_spill", "cost of a spill instruction", &opt_cost_spill),
153 LC_OPT_ENT_DBL ("cost_remat", "cost of a rematerialization", &opt_cost_remat),
157 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
158 //#define SCHEDULE_PHIM /* insert phim nodes into schedule */
161 //#define SOLVE_LOCAL
162 #define LPP_SERVER "i44pc52"
163 #define LPP_SOLVER "cplex"
166 #define MAX_PATHS INT_MAX
169 typedef struct _spill_ilp_t {
170 const arch_register_class_t *cls;
175 struct obstack *obst;
177 pset *all_possible_remats;
180 set *values; /**< for collecting all definitions of values before running ssa-construction */
186 #ifndef SCHEDULE_PHIM
189 DEBUG_ONLY(firm_dbg_module_t * dbg);
192 typedef int ilp_var_t;
193 typedef int ilp_cst_t;
195 typedef struct _spill_bb_t {
200 typedef struct _remat_t {
201 const ir_node *op; /**< for copy_irn */
202 const ir_node *value; /**< the value which is being recomputed by this remat */
203 const ir_node *proj; /**< not NULL if the above op produces a tuple */
204 int cost; /**< cost of this remat */
205 int inverse; /**< nonzero if this is an inverse remat */
209 * Data to be attached to each IR node. For remats this contains the ilp_var
210 * for this remat and for normal ops this contains the ilp_vars for
211 * reloading each operand
213 typedef struct _op_t {
218 const remat_t *remat; /** the remat this op belongs to */
219 int pre; /** 1, if this is a pressure-increasing remat */
223 ir_node *op; /** the operation this live range belongs to */
232 typedef struct _defs_t {
233 const ir_node *value;
234 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
235 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
238 typedef struct _remat_info_t {
239 const ir_node *irn; /**< the irn to which these remats belong */
240 pset *remats; /**< possible remats for this value */
241 pset *remats_by_operand; /**< remats with this value as operand */
244 typedef struct _keyval_t {
249 typedef struct _spill_t {
258 typedef struct _memoperand_t {
259 ir_node *irn; /**< the irn */
260 unsigned int pos; /**< the position of the argument */
261 ilp_var_t ilp; /**< the ilp var for this memory operand */
265 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
267 return arch_irn_consider_in_reg_alloc(si->birg->main_env->arch_env,
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 if(be_profile_has_data())
421 return ((double)be_profile_get_block_execcount(get_block(irn))) + FUDGE;
423 #ifndef EXECFREQ_LOOPDEPH
424 return get_block_execfreq(si->birg->exec_freq, get_block(irn)) + FUDGE;
427 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
429 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
434 get_cost(const spill_ilp_t * si, const ir_node * irn)
436 if(be_is_Spill(irn)) {
437 return opt_cost_spill;
438 } else if(be_is_Reload(irn)){
439 return opt_cost_reload;
441 return arch_get_op_estimated_cost(si->birg->main_env->arch_env, irn);
446 * Checks, whether node and its operands have suitable reg classes
449 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
452 const arch_env_t *arch_env = si->birg->main_env->arch_env;
453 int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
457 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
460 for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
461 ir_node *op = get_irn_n(irn, n);
462 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
465 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
472 * Try to create a remat from @p op with destination value @p dest_value
474 static INLINE remat_t *
475 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
477 remat_t *remat = NULL;
479 // if(!mode_is_datab(get_irn_mode(dest_value)))
482 if(dest_value == op) {
483 const ir_node *proj = NULL;
485 if(is_Proj(dest_value)) {
486 op = get_Proj_pred(op);
490 if(!is_rematerializable(si, op))
493 remat = obstack_alloc(si->obst, sizeof(*remat));
495 remat->cost = (int)get_cost(si, op);
496 remat->value = dest_value;
500 arch_inverse_t inverse;
503 /* get the index of the operand we want to retrieve by the inverse op */
504 for (n = get_irn_arity(op)-1; n>=0; --n) {
505 ir_node *arg = get_irn_n(op, n);
507 if(arg == dest_value) break;
511 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", n, op));
513 /* else ask the backend to give an inverse op */
514 if(arch_get_inverse(si->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
517 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
519 assert(inverse.n > 0 && "inverse op should have at least one node");
521 for(i=inverse.n-1; i>=0; --i) {
522 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
526 remat = obstack_alloc(si->obst, sizeof(*remat));
527 remat->op = inverse.nodes[0];
528 remat->cost = inverse.costs;
529 remat->value = dest_value;
530 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
533 // Matze: commented this out, this doesn't seem to be true if
534 // the inverse is a simple operation with only 1 result...
535 //assert(is_Proj(remat->proj));
537 assert(0 && "I can not handle remats with more than 2 nodes");
544 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
546 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
554 add_remat(const spill_ilp_t * si, const remat_t * remat)
556 remat_info_t *remat_info,
561 assert(remat->value);
563 query.irn = remat->value;
565 query.remats_by_operand = NULL;
566 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
568 if(remat_info->remats == NULL) {
569 remat_info->remats = new_pset(cmp_remat, 4096);
571 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
573 /* insert the remat into the remats_be_operand set of each argument of the remat op */
574 for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
575 ir_node *arg = get_irn_n(remat->op, n);
579 query.remats_by_operand = NULL;
580 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
582 if(remat_info->remats_by_operand == NULL) {
583 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
585 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
590 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
592 const ir_edge_t *edge = get_irn_out_edge_first(irn);
596 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
599 edge = get_irn_out_edge_next(irn, edge);
606 get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
612 irn = get_Proj_pred(irn);
614 for(n=get_irn_arity(irn)-1; n>=0; --n) {
615 const ir_node *arg = get_irn_n(irn, n);
617 if(has_reg_class(si, arg)) ++ret;
624 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
629 if( has_reg_class(si, op)
630 && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1)
631 && (opt_remats != REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0)
633 remat = get_remat_from_op(si, op, op);
635 add_remat(si, remat);
639 if(opt_remats == REMATS_ALL) {
640 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
642 for (n = get_irn_arity(op)-1; n>=0; --n) {
643 ir_node *arg = get_irn_n(op, n);
645 if(has_reg_class(si, arg)) {
646 /* try to get an inverse remat */
647 remat = get_remat_from_op(si, arg, op);
649 add_remat(si, remat);
657 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
660 ir_node *def_block = get_nodes_block(val);
666 /* if pos is at end of a basic block */
668 ret = (pos == def_block || block_dominates(def_block, pos));
669 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
673 /* else if this is a normal operation */
674 block = get_nodes_block(pos);
675 if(block == def_block) {
676 if(!sched_is_scheduled(val)) return 1;
678 ret = sched_comes_after(val, pos);
679 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
683 ret = block_dominates(def_block, block);
684 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
688 static INLINE ir_node *
689 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
691 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->birg->main_env->arch_env);
695 * Returns first non-Phi node of block @p bb
697 static INLINE ir_node *
698 sched_block_first_nonphi(const ir_node * bb)
700 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
704 sched_skip_proj_predicator(const ir_node * irn, void * data)
706 return (is_Proj(irn));
709 static INLINE ir_node *
710 sched_next_nonproj(const ir_node * irn, int forward)
712 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
716 * Returns next operation node (non-Proj) after @p irn
717 * or the basic block of this node
719 static INLINE ir_node *
720 sched_next_op(const ir_node * irn)
722 ir_node *next = sched_next(irn);
727 return sched_next_nonproj(next, 1);
731 * Returns previous operation node (non-Proj) before @p irn
732 * or the basic block of this node
734 static INLINE ir_node *
735 sched_prev_op(const ir_node * irn)
737 ir_node *prev = sched_prev(irn);
742 return sched_next_nonproj(prev, 0);
746 sched_put_after(ir_node * insert, ir_node * irn)
748 if(is_Block(insert)) {
749 insert = sched_block_first_nonphi(insert);
751 insert = sched_next_op(insert);
754 sched_add_before(insert, irn);
758 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
760 if(is_Block(insert)) {
761 insert = sched_block_last_noncf(si, insert);
763 insert = sched_next_nonproj(insert, 0);
764 insert = sched_prev(insert);
767 sched_add_after(insert, irn);
771 next_post_remat(const ir_node * irn)
777 next = sched_block_first_nonphi(irn);
779 next = sched_next_op(irn);
782 if(sched_is_end(next))
785 op = get_irn_link(next);
786 if(op->is_remat && !op->attr.remat.pre) {
795 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
801 ret = sched_block_last_noncf(si, irn);
802 ret = sched_next(ret);
803 ret = sched_prev_op(ret);
805 ret = sched_prev_op(irn);
808 if(sched_is_end(ret) || is_Phi(ret))
811 op = (op_t*)get_irn_link(ret);
812 if(op->is_remat && op->attr.remat.pre) {
820 * Tells you whether a @p remat can be placed before the irn @p pos
823 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
825 const ir_node *op = remat->op;
831 prev = sched_block_last_noncf(si, pos);
832 prev = sched_next_nonproj(prev, 0);
834 prev = sched_prev_op(pos);
836 /* do not remat if the rematted value is defined immediately before this op */
837 if(prev == remat->op) {
842 /* this should be just fine, the following OP will be using this value, right? */
844 /* only remat AFTER the real definition of a value (?) */
845 if(!value_is_defined_before(si, pos, remat->value)) {
846 // ir_fprintf(stderr, "error(not defined)");
851 for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
852 const ir_node *arg = get_irn_n(op, n);
854 if(opt_no_enlarge_liveness) {
855 if(has_reg_class(si, arg) && live) {
856 res &= pset_find_ptr((pset*)live, arg)?1:0;
858 res &= value_is_defined_before(si, pos, arg);
861 res &= value_is_defined_before(si, pos, arg);
869 * Tells you whether a @p remat can be placed after the irn @p pos
872 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
875 pos = sched_block_first_nonphi(pos);
877 pos = sched_next_op(pos);
880 /* only remat AFTER the real definition of a value (?) */
881 if(!value_is_defined_before(si, pos, remat->value)) {
885 return can_remat_before(si, remat, pos, live);
889 * Collect potetially rematerializable OPs
892 walker_remat_collector(ir_node * irn, void * data)
894 spill_ilp_t *si = data;
896 if(!is_Block(irn) && !is_Phi(irn)) {
897 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
898 get_remats_from_op(si, irn);
903 * Inserts a copy of @p irn before @p pos
906 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
911 bb = is_Block(pos)?pos:get_nodes_block(pos);
912 copy = exact_copy(irn);
914 set_phi_class(si->pc, copy, NULL);
915 set_nodes_block(copy, bb);
916 sched_put_before(si, pos, copy);
922 * Inserts a copy of @p irn after @p pos
925 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
930 bb = is_Block(pos)?pos:get_nodes_block(pos);
931 copy = exact_copy(irn);
933 set_phi_class(si->pc, copy, NULL);
934 set_nodes_block(copy, bb);
935 sched_put_after(pos, copy);
941 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
945 if(can_remat_after(si, remat, pos, live)) {
950 DBG((si->dbg, LEVEL_3, "\t >inserting remat2 %+F\n", remat->op));
952 copy = insert_copy_after(si, remat->op, pos);
954 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
955 op = obstack_alloc(si->obst, sizeof(*op));
957 op->attr.remat.remat = remat;
958 op->attr.remat.pre = 0;
959 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
961 set_irn_link(copy, op);
962 pset_insert_ptr(si->all_possible_remats, copy);
964 proj_copy = insert_copy_after(si, remat->proj, copy);
965 set_irn_n(proj_copy, 0, copy);
966 set_irn_link(proj_copy, op);
967 pset_insert_ptr(si->all_possible_remats, proj_copy);
979 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
983 if(can_remat_before(si, remat, pos, live)) {
988 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
990 copy = insert_copy_before(si, remat->op, pos);
992 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
993 op = obstack_alloc(si->obst, sizeof(*op));
995 op->attr.remat.remat = remat;
996 op->attr.remat.pre = 1;
997 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
999 set_irn_link(copy, op);
1000 pset_insert_ptr(si->all_possible_remats, copy);
1002 proj_copy = insert_copy_after(si, remat->proj, copy);
1003 set_irn_n(proj_copy, 0, copy);
1004 set_irn_link(proj_copy, op);
1005 pset_insert_ptr(si->all_possible_remats, proj_copy);
1017 get_block_n_succs(const ir_node *block) {
1018 const ir_edge_t *edge;
1020 assert(edges_activated(current_ir_graph));
1022 edge = get_block_succ_first(block);
1026 edge = get_block_succ_next(block, edge);
1027 return edge ? 2 : 1;
1031 is_start_block(const ir_node * bb)
1033 return get_irg_start_block(get_irn_irg(bb)) == bb;
1037 is_merge_edge(const ir_node * bb)
1039 if(is_start_block(bb))
1043 return get_block_n_succs(bb) == 1;
1049 is_diverge_edge(const ir_node * bb)
1051 if(is_start_block(bb))
1055 return get_Block_n_cfgpreds(bb) == 1;
1061 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1066 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1067 irn = be_lv_get_irn(si->lv, bb, i);
1069 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1070 pset_insert_ptr(live, irn);
1074 irn = sched_last(bb);
1076 /* all values eaten by control flow operations are also live until the end of the block */
1077 sched_foreach_reverse(bb, irn) {
1080 if(!sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) break;
1082 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1083 ir_node *arg = get_irn_n(irn,i);
1085 if(has_reg_class(si, arg)) {
1086 pset_insert_ptr(live, arg);
1091 * find values that are used by remats at end of block
1092 * and insert them into live set
1094 foreach_pre_remat(si, bb, irn) {
1097 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1098 ir_node *remat_arg = get_irn_n(irn, n);
1100 if(!has_reg_class(si, remat_arg)) continue;
1102 /* if value is becoming live through use by remat */
1103 if(!pset_find_ptr(live, remat_arg)) {
1104 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, irn));
1106 pset_insert_ptr(live, remat_arg);
1113 walker_regclass_copy_insertor(ir_node * irn, void * data)
1115 spill_ilp_t *si = data;
1117 if(is_Phi(irn) && has_reg_class(si, irn)) {
1120 for(n=get_irn_arity(irn)-1; n>=0; --n) {
1121 ir_node *phi_arg = get_irn_n(irn, n);
1122 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
1124 if(!has_reg_class(si, phi_arg)) {
1125 ir_node *copy = be_new_Copy(si->cls, si->birg->irg, bb, phi_arg);
1126 ir_node *pos = sched_block_last_noncf(si, bb);
1127 op_t *op = obstack_alloc(si->obst, sizeof(*op));
1129 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
1130 sched_add_after(pos, copy);
1131 set_irn_n(irn, n, copy);
1134 op->attr.live_range.args.reloads = NULL;
1135 op->attr.live_range.ilp = ILP_UNDEF;
1136 set_irn_link(copy, op);
1143 * Insert (so far unused) remats into the irg to
1144 * recompute the potential liveness of all values
1147 walker_remat_insertor(ir_node * bb, void * data)
1149 spill_ilp_t *si = data;
1156 /* skip start block, no remats to do there */
1157 if(is_start_block(bb)) return;
1159 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1161 live = pset_new_ptr_default();
1162 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1163 ir_node *value = be_lv_get_irn(si->lv, bb, i);
1165 /* add remats at end of block */
1166 if (has_reg_class(si, value)) {
1167 pset_insert_ptr(live, value);
1171 irn = sched_last(bb);
1172 while(!sched_is_end(irn)) {
1178 next = sched_prev(irn);
1180 /* delete defined value from live set */
1181 if(has_reg_class(si, irn)) {
1182 pset_remove_ptr(live, irn);
1185 if(is_Phi(irn) || is_Proj(irn)) {
1190 args = pset_new_ptr_default();
1191 used = pset_new_ptr_default();
1193 /* collect arguments of op and set args of op already live in epilog */
1194 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1195 ir_node *arg = get_irn_n(irn, n);
1197 pset_insert_ptr(args, arg);
1198 if(has_reg_class(si, arg)) {
1199 pset_insert_ptr(live, arg);
1200 pset_insert_ptr(used, arg);
1204 /* insert all possible remats before irn */
1205 pset_foreach(args, arg) {
1206 remat_info_t *remat_info,
1209 /* continue if the operand has the wrong reg class */
1210 if(!has_reg_class(si, arg))
1214 query.remats = NULL;
1215 query.remats_by_operand = NULL;
1216 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1222 if(remat_info->remats) {
1223 pset_foreach(remat_info->remats, remat) {
1224 ir_node *remat_irn = NULL;
1226 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
1227 remat_irn = insert_remat_before(si, remat, irn, live);
1230 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1231 ir_node *remat_arg = get_irn_n(remat_irn, n);
1233 /* collect args of remats which are not args of op */
1234 if(has_reg_class(si, remat_arg) && !pset_find_ptr(args, remat_arg)) {
1235 pset_insert_ptr(used, remat_arg);
1243 /* do not place post remats after jumps */
1244 if(sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) {
1250 /* insert all possible remats after irn */
1251 post_remats = pset_new_ptr_default();
1252 pset_foreach(used, arg) {
1253 remat_info_t *remat_info,
1256 /* continue if the operand has the wrong reg class */
1257 if(!has_reg_class(si, arg))
1261 query.remats = NULL;
1262 query.remats_by_operand = NULL;
1263 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1269 if(remat_info->remats_by_operand) {
1270 pset_foreach(remat_info->remats_by_operand, remat) {
1271 /* do not insert remats producing the same value as one of the operands */
1272 if(!pset_find_ptr(args, remat->value)) {
1273 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
1275 /* only remat values that can be used by real ops */
1276 if(!opt_remat_while_live || pset_find_ptr(live, remat->value)) {
1277 pset_insert_ptr(post_remats, remat);
1283 pset_foreach(post_remats, remat) {
1284 insert_remat_after(si, remat, irn, live);
1286 del_pset(post_remats);
1293 /* add remats at end if successor has multiple predecessors */
1294 if(is_merge_edge(bb)) {
1295 pset *live_out = pset_new_ptr_default();
1298 get_live_end(si, bb, live_out);
1300 /* add remats at end of block */
1301 pset_foreach(live_out, value) {
1302 remat_info_t *remat_info,
1306 query.remats = NULL;
1307 query.remats_by_operand = NULL;
1308 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1310 if(remat_info && remat_info->remats) {
1311 pset_foreach(remat_info->remats, remat) {
1312 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1314 insert_remat_before(si, remat, bb, live_out);
1321 if(is_diverge_edge(bb)) {
1322 pset *live_in = pset_new_ptr_default();
1325 be_lv_foreach(si->lv, bb, be_lv_state_in, i) {
1326 value = be_lv_get_irn(si->lv, bb, i);
1328 if(has_reg_class(si, value)) {
1329 pset_insert_ptr(live_in, value);
1332 /* add phis to live_in */
1333 sched_foreach(bb, value) {
1334 if(!is_Phi(value)) break;
1336 if(has_reg_class(si, value)) {
1337 pset_insert_ptr(live_in, value);
1341 /* add remat2s at beginning of block */
1342 post_remats = pset_new_ptr_default();
1343 pset_foreach(live_in, value) {
1344 remat_info_t *remat_info,
1348 query.remats = NULL;
1349 query.remats_by_operand = NULL;
1350 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1352 if(remat_info && remat_info->remats_by_operand) {
1353 pset_foreach(remat_info->remats_by_operand, remat) {
1354 DBG((si->dbg, LEVEL_4, "\t considering remat2 %+F at beginning of block %+F\n", remat->op, bb));
1356 /* put the remat here if all its args are available and result is still live */
1357 if(!opt_remat_while_live || pset_find_ptr(live_in, remat->value)) {
1358 pset_insert_ptr(post_remats, remat);
1363 pset_foreach(post_remats, remat) {
1364 insert_remat_after(si, remat, bb, live_in);
1366 del_pset(post_remats);
1372 can_be_copied(const ir_node * bb, const ir_node * irn)
1374 const ir_edge_t *edge = get_block_succ_first(bb);
1375 const ir_node *next_bb = edge->src;
1376 int pos = edge->pos;
1379 assert(is_merge_edge(bb));
1381 sched_foreach(next_bb, phi) {
1382 const ir_node *phi_arg;
1384 if(!is_Phi(phi)) break;
1386 phi_arg = get_irn_n(phi, pos);
1388 if(phi_arg == irn) {
1396 * Initialize additional node info
1399 luke_initializer(ir_node * bb, void * data)
1401 spill_ilp_t *si = (spill_ilp_t*)data;
1402 spill_bb_t *spill_bb;
1405 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1406 set_irn_link(bb, spill_bb);
1408 sched_foreach(bb, irn) {
1411 op = obstack_alloc(si->obst, sizeof(*op));
1413 op->attr.live_range.ilp = ILP_UNDEF;
1416 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1417 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1419 } else if(!is_Proj(irn)) {
1420 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1421 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1423 op->attr.live_range.args.reloads = NULL;
1425 set_irn_link(irn, op);
1431 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1434 luke_endwalker(ir_node * bb, void * data)
1436 spill_ilp_t *si = (spill_ilp_t*)data;
1442 spill_bb_t *spill_bb = get_irn_link(bb);
1445 live = pset_new_ptr_default();
1446 use_end = pset_new_ptr_default();
1448 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1449 irn = be_lv_get_irn(si->lv, bb, i);
1450 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1451 pset_insert_ptr(live, irn);
1455 * find values that are used by remats at end of block
1456 * and insert them into live set
1458 foreach_pre_remat(si, bb, irn) {
1461 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1462 ir_node *remat_arg = get_irn_n(irn, n);
1464 if(has_reg_class(si, remat_arg)) {
1465 pset_insert_ptr(live, remat_arg);
1470 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1471 /* their reg_out must always be set */
1472 sched_foreach_reverse(bb, irn) {
1475 if(!sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) break;
1477 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1478 ir_node *irn_arg = get_irn_n(irn, n);
1480 if(has_reg_class(si, irn_arg)) {
1481 pset_insert_ptr(use_end, irn_arg);
1486 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1487 //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1488 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1490 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1492 /* if this is a merge edge we can reload at the end of this block */
1493 if(is_merge_edge(bb)) {
1494 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1495 } else if(pset_count(use_end)){
1496 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1498 spill_bb->reloads = NULL;
1501 pset_foreach(live,irn) {
1505 int default_spilled;
1508 /* handle values used by control flow nodes later separately */
1509 if(pset_find_ptr(use_end, irn)) continue;
1512 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1514 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1516 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1517 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1518 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1520 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1521 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1523 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1524 /* by default spill value right after definition */
1525 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1526 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1528 if(is_merge_edge(bb)) {
1532 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1533 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), can_be_copied(bb, irn));
1534 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1536 /* reload <= mem_out */
1537 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1538 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1539 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1542 spill->reg_in = ILP_UNDEF;
1543 spill->mem_in = ILP_UNDEF;
1546 pset_foreach(use_end,irn) {
1550 ilp_cst_t end_use_req,
1553 int default_spilled;
1556 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1558 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1560 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1561 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1563 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1564 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1566 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1567 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1568 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1570 /* reload for use be control flow op */
1571 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1572 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1573 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1575 /* reload <= mem_out */
1576 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1577 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1578 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1580 spill->reg_in = ILP_UNDEF;
1581 spill->mem_in = ILP_UNDEF;
1583 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1584 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1585 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1594 * Find a remat of value @p value in the epilog of @p pos
1597 find_post_remat(const ir_node * value, const ir_node * pos)
1599 while((pos = next_post_remat(pos)) != NULL) {
1602 op = get_irn_link(pos);
1603 assert(op->is_remat && !op->attr.remat.pre);
1605 if(op->attr.remat.remat->value == value)
1606 return (ir_node*)pos;
1609 const ir_edge_t *edge;
1610 foreach_out_edge(pos, edge) {
1611 ir_node *proj = get_edge_src_irn(edge);
1612 assert(is_Proj(proj));
1623 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1625 spill_bb_t *spill_bb = get_irn_link(bb);
1629 int default_spilled;
1632 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1634 double spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1636 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1638 spill->reg_out = ILP_UNDEF;
1639 spill->reg_in = ILP_UNDEF;
1640 spill->mem_in = ILP_UNDEF;
1642 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1643 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1645 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1646 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1647 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1654 * Inserts ILP-constraints and variables for memory copying before the given position
1657 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1659 const ir_node *succ;
1660 const ir_edge_t *edge;
1661 spill_bb_t *spill_bb = get_irn_link(block);
1670 assert(edges_activated(current_ir_graph));
1672 edge = get_block_succ_first(block);
1678 edge = get_block_succ_next(block, edge);
1679 /* next block can only contain phis, if this is a merge edge */
1682 ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1683 copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1685 ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1686 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1688 pset_foreach(live, tmp) {
1691 op_t *op = get_irn_link(irn);
1692 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1694 spill = set_find_spill(spill_bb->ilp, tmp);
1697 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1699 lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1701 sched_foreach(succ, phi) {
1702 const ir_node *to_copy;
1704 spill_t *to_copy_spill;
1705 op_t *phi_op = get_irn_link(phi);
1706 ilp_var_t reload = ILP_UNDEF;
1709 if(!is_Phi(phi)) break;
1710 if(!has_reg_class(si, phi)) continue;
1712 to_copy = get_irn_n(phi, pos);
1713 to_copy_op = get_irn_link(to_copy);
1715 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1716 assert(to_copy_spill);
1718 if(spill_bb->reloads) {
1719 keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1722 reload = PTR_TO_INT(keyval->val);
1726 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1727 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1729 /* copy - reg_out - reload - remat - live_range <= 0 */
1730 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1731 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1732 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1733 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1734 foreach_pre_remat(si, block, tmp) {
1735 op_t *remat_op = get_irn_link(tmp);
1736 if(remat_op->attr.remat.remat->value == to_copy) {
1737 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1741 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1742 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1744 /* copy - reg_out - copyreg <= 0 */
1745 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1746 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1747 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1753 * Walk all irg blocks and emit this ILP
1756 luke_blockwalker(ir_node * bb, void * data)
1758 spill_ilp_t *si = (spill_ilp_t*)data;
1763 spill_bb_t *spill_bb = get_irn_link(bb);
1766 pset *defs = pset_new_ptr_default();
1767 const arch_env_t *arch_env = si->birg->main_env->arch_env;
1769 live = pset_new_ptr_default();
1771 /****************************************
1772 * B A S I C B L O C K E N D
1773 ***************************************/
1776 /* init live values at end of block */
1777 get_live_end(si, bb, live);
1779 pset_foreach(live, irn) {
1781 ilp_var_t reload = ILP_UNDEF;
1783 spill = set_find_spill(spill_bb->ilp, irn);
1786 if(spill_bb->reloads) {
1787 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1790 reload = PTR_TO_INT(keyval->val);
1794 op = get_irn_link(irn);
1795 assert(!op->is_remat);
1797 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1798 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1799 op->attr.live_range.op = bb;
1801 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1802 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1804 /* reg_out - reload - remat - live_range <= 0 */
1805 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1806 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1807 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1808 foreach_pre_remat(si, bb, tmp) {
1809 op_t *remat_op = get_irn_link(tmp);
1810 if(remat_op->attr.remat.remat->value == irn) {
1811 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1814 ir_snprintf(buf, sizeof(buf), "reg_out2_%N_%N", bb, irn);
1815 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
1817 /* value may only die at bb end if it is used for a mem copy */
1818 /* reg_out + \sum copy - reload - remat - live_range >= 0 */
1819 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1820 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1821 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1822 foreach_pre_remat(si, bb, tmp) {
1823 op_t *remat_op = get_irn_link(tmp);
1824 if(remat_op->attr.remat.remat->value == irn) {
1825 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1828 if(is_merge_edge(bb)) {
1829 const ir_edge_t *edge = get_block_succ_first(bb);
1830 const ir_node *next_bb = edge->src;
1831 int pos = edge->pos;
1834 sched_foreach(next_bb, phi) {
1835 const ir_node *phi_arg;
1837 if(!is_Phi(phi)) break;
1839 phi_arg = get_irn_n(phi, pos);
1841 if(phi_arg == irn) {
1842 op_t *phi_op = get_irn_link(phi);
1843 ilp_var_t copy = phi_op->attr.live_range.args.copies[pos];
1845 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
1852 insert_mem_copy_position(si, live, bb);
1855 * assure the remat args are available
1857 foreach_pre_remat(si, bb, tmp) {
1858 op_t *remat_op = get_irn_link(tmp);
1861 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1862 ir_node *remat_arg = get_irn_n(tmp, n);
1863 op_t *arg_op = get_irn_link(remat_arg);
1865 if(!has_reg_class(si, remat_arg)) continue;
1867 spill = set_find_spill(spill_bb->ilp, remat_arg);
1870 /* arguments of remats have to be live until the very end of the block
1871 * remat = reg_out(remat_arg) and (reload(remat_arg) or live_range(remat_arg)),
1872 * no remats, they could be in wrong order
1875 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1876 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1878 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 3.0);
1879 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, -2.0);
1880 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1882 /* use reload placed for this argument */
1883 if(spill_bb->reloads) {
1884 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1887 ilp_var_t reload = PTR_TO_INT(keyval->val);
1889 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1894 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1899 /**************************************
1900 * B A S I C B L O C K B O D Y
1901 **************************************/
1903 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1909 ilp_cst_t check_pre,
1915 ilp_cst_t one_memoperand = -1;
1917 /* iterate only until first phi */
1921 op = get_irn_link(irn);
1923 if(op->is_remat) continue;
1925 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1927 /* collect defined values */
1928 if(has_reg_class(si, irn)) {
1929 pset_insert_ptr(defs, irn);
1933 if(is_Proj(irn)) continue;
1936 * init set of irn's arguments
1937 * and all possibly used values around this op
1938 * and values defined by post remats
1940 args = new_set(cmp_keyval, get_irn_arity(irn));
1941 used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1942 remat_defs = pset_new_ptr(pset_count(live));
1944 if(!is_start_block(bb) || !be_is_Barrier(irn)) {
1945 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1946 ir_node *irn_arg = get_irn_n(irn, n);
1947 if(has_reg_class(si, irn_arg)) {
1948 set_insert_keyval(args, irn_arg, (void*)n);
1949 pset_insert_ptr(used, irn_arg);
1952 foreach_post_remat(irn, tmp) {
1953 op_t *remat_op = get_irn_link(tmp);
1955 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1957 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1958 ir_node *remat_arg = get_irn_n(tmp, n);
1959 if(has_reg_class(si, remat_arg)) {
1960 pset_insert_ptr(used, remat_arg);
1964 foreach_pre_remat(si, irn, tmp) {
1965 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1966 ir_node *remat_arg = get_irn_n(tmp, n);
1967 if(has_reg_class(si, remat_arg)) {
1968 pset_insert_ptr(used, remat_arg);
1974 /**********************************
1975 * I N E P I L O G O F irn
1976 **********************************/
1978 /* ensure each dying value is used by only one post remat */
1979 pset_foreach(used, tmp) {
1980 ir_node *value = tmp;
1981 op_t *value_op = get_irn_link(value);
1986 foreach_post_remat(irn, remat) {
1987 op_t *remat_op = get_irn_link(remat);
1989 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1990 ir_node *remat_arg = get_irn_n(remat, n);
1992 /* if value is used by this remat add it to constraint */
1993 if(remat_arg == value) {
1995 /* sum remat2s <= 1 + n_remats*live_range */
1996 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1997 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2001 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2007 if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
2008 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
2012 /* ensure at least one value dies at post remat */
2013 foreach_post_remat(irn, tmp) {
2014 op_t *remat_op = get_irn_link(tmp);
2015 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
2018 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2019 remat_arg = get_irn_n(tmp, n);
2021 if(has_reg_class(si, remat_arg)) {
2023 /* does arg always die at this op? */
2024 if(!pset_find_ptr(live, remat_arg))
2025 goto skip_one_must_die;
2027 pset_insert_ptr(remat_args, remat_arg);
2031 /* remat + \sum live_range(remat_arg) <= |args| */
2032 ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
2033 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2034 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2036 pset_foreach(remat_args, remat_arg) {
2037 op_t *arg_op = get_irn_link(remat_arg);
2039 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2043 del_pset(remat_args);
2046 /* new live ranges for values from L\U defined by post remats */
2047 pset_foreach(live, tmp) {
2048 ir_node *value = tmp;
2049 op_t *value_op = get_irn_link(value);
2051 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
2052 ilp_var_t prev_lr = ILP_UNDEF;
2055 if(pset_find_ptr(remat_defs, value)) {
2057 /* next_live_range <= prev_live_range + sum remat2s */
2058 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
2059 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2061 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
2062 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2064 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
2065 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2067 foreach_post_remat(irn, remat) {
2068 op_t *remat_op = get_irn_link(remat);
2070 /* if value is being rematerialized by this remat */
2071 if(value == remat_op->attr.remat.remat->value) {
2072 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2076 value_op->attr.live_range.ilp = prev_lr;
2077 value_op->attr.live_range.op = irn;
2082 /* requirements for post remats and start live ranges from L/U' for values dying here */
2083 foreach_post_remat(irn, tmp) {
2084 op_t *remat_op = get_irn_link(tmp);
2087 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2088 ir_node *remat_arg = get_irn_n(tmp, n);
2089 op_t *arg_op = get_irn_link(remat_arg);
2091 if(!has_reg_class(si, remat_arg)) continue;
2093 /* only for values in L\U (TODO and D?), the others are handled with post_use */
2094 if(!pset_find_ptr(used, remat_arg)) {
2095 /* remat <= live_range(remat_arg) */
2096 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
2097 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2099 /* if value is becoming live through use by remat2 */
2100 if(!pset_find_ptr(live, remat_arg)) {
2103 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
2104 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2106 arg_op->attr.live_range.ilp = lr;
2107 arg_op->attr.live_range.op = irn;
2109 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
2111 pset_insert_ptr(live, remat_arg);
2112 add_to_spill_bb(si, bb, remat_arg);
2115 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2116 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2121 d = pset_count(defs);
2122 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
2124 /* count how many regs irn needs for arguments */
2125 u = set_count(args);
2128 /* check the register pressure in the epilog */
2129 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
2130 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
2131 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
2133 /* add L\U' to check_post */
2134 pset_foreach(live, tmp) {
2135 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
2136 /* if a live value is not used by irn */
2137 tmp_op = get_irn_link(tmp);
2138 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
2142 /***********************************************************
2143 * 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
2144 **********************************************************/
2147 pset_foreach(used, tmp) {
2153 op_t *arg_op = get_irn_link(arg);
2156 spill = add_to_spill_bb(si, bb, arg);
2158 /* new live range for each used value */
2159 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2160 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2162 /* the epilog stuff - including post_use, check_post, check_post_remat */
2163 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2164 post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2166 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2168 /* arg is live throughout epilog if the next live_range is in a register */
2169 if(pset_find_ptr(live, arg)) {
2170 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
2172 /* post_use >= next_lr + 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, arg_op->attr.live_range.ilp, 1.0);
2179 /* forall post remat which use arg add a similar cst */
2180 foreach_post_remat(irn, remat) {
2183 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2184 ir_node *remat_arg = get_irn_n(remat, n);
2185 op_t *remat_op = get_irn_link(remat);
2187 if(remat_arg == arg) {
2188 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
2190 /* post_use >= remat */
2191 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2192 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2193 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2194 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2199 /* if value is not an arg of op and not possibly defined by post remat
2200 * then it may only die and not become live
2202 if(!set_find_keyval(args, arg)) {
2203 /* post_use <= prev_lr */
2204 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2205 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2206 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2207 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2209 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2210 /* next_lr <= prev_lr */
2211 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2212 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2213 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2214 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2218 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2219 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2220 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2221 ilp_var_t memoperand;
2223 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2224 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0);
2225 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2227 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2228 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2230 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2231 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2236 /* new live range begins for each used value */
2237 arg_op->attr.live_range.ilp = prev_lr;
2238 arg_op->attr.live_range.op = irn;
2240 pset_insert_ptr(live, arg);
2243 /* just to be sure */
2244 check_post = ILP_UNDEF;
2246 /* allow original defintions to be removed */
2247 if(opt_repair_schedule) {
2248 pset_foreach(defs, tmp) {
2249 op_t *tmp_op = get_irn_link(tmp);
2250 spill_t *spill = set_find_spill(spill_bb->ilp, tmp);
2255 ir_snprintf(buf, sizeof(buf), "delete_%N", tmp);
2256 delete = lpp_add_var_default(si->lpp, buf, lpp_binary, -1.0*get_cost(si, irn)*execution_frequency(si, bb), 0.0);
2258 /* op may not be killed if its first live_range is 1 */
2259 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2260 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2261 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2262 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2264 /* op may not be killed if it is spilled after the definition */
2265 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2266 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2267 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2268 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2273 ir_snprintf(buf, sizeof(buf), "keep_%N", tmp);
2274 keep = lpp_add_var_default(si->lpp, buf, lpp_binary, get_cost(si, irn)*execution_frequency(si, bb), 1.0);
2276 /* op may not be killed if its first live_range is 1 */
2277 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2278 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2279 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2280 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, -1.0);
2282 /* op may not be killed if it is spilled after the definition */
2283 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2284 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2285 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2286 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2291 pset_foreach(defs, tmp) {
2292 op_t *tmp_op = get_irn_link(tmp);
2293 spill_t *spill = set_find_spill(spill_bb->ilp, tmp);
2296 /* live_range or spill should be 1
2297 TODO: lr should be live until first use */
2298 ir_snprintf(buf, sizeof(buf), "nokillorig_%N", tmp);
2299 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 1.0);
2300 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2301 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2311 /* check the register pressure in the prolog */
2312 /* sum_{L\U} lr <= k - |U| */
2313 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2314 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2316 /* for the prolog remove defined values from the live set */
2317 pset_foreach(defs, tmp) {
2318 pset_remove_ptr(live, tmp);
2321 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2322 ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2323 one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2326 /***********************************************************
2327 * 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
2328 **********************************************************/
2331 set_foreach(args, keyval) {
2333 const ir_node *arg = keyval->key;
2334 int i = PTR_TO_INT(keyval->val);
2335 op_t *arg_op = get_irn_link(arg);
2336 ilp_cst_t requirements;
2339 spill = set_find_spill(spill_bb->ilp, arg);
2342 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2343 op->attr.live_range.args.reloads[i] = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
2345 /* reload <= mem_out */
2346 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2347 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2348 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2349 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2351 /* requirement: arg must be in register for use */
2352 /* reload + remat + live_range == 1 */
2353 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2354 requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2356 lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2357 lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2358 foreach_pre_remat(si, irn, tmp) {
2359 op_t *remat_op = get_irn_link(tmp);
2360 if(remat_op->attr.remat.remat->value == arg) {
2361 lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2365 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2367 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2368 if(get_irn_n(irn, n) == arg) {
2372 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2373 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2374 memoperand_t *memoperand;
2375 memoperand = set_find_memoperand(si->memoperands, irn, n);
2377 /* memoperand <= mem_out */
2378 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2379 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2380 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2381 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2383 /* the memoperand is only sufficient if it is used once by the op */
2384 if(n_memoperands == 1)
2385 lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2387 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2389 /* we have one more free register if we use a memory operand */
2390 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2396 /* iterate over L\U */
2397 pset_foreach(live, tmp) {
2398 if(!set_find_keyval(args, tmp)) {
2399 /* if a live value is not used by irn */
2400 tmp_op = get_irn_link(tmp);
2401 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2405 /* requirements for remats */
2406 foreach_pre_remat(si, irn, tmp) {
2407 op_t *remat_op = get_irn_link(tmp);
2410 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2411 ir_node *remat_arg = get_irn_n(tmp, n);
2412 op_t *arg_op = get_irn_link(remat_arg);
2414 if(!has_reg_class(si, remat_arg)) continue;
2416 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2417 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2418 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2420 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2421 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2423 /* if remat arg is also used by current op then we can use reload placed for this argument */
2424 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2425 int index = (int)keyval->val;
2427 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2435 /*************************
2436 * D O N E W I T H O P
2437 *************************/
2439 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2441 pset_foreach(live, tmp) {
2442 assert(has_reg_class(si, tmp));
2446 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2447 ir_node *arg = get_irn_n(irn, n);
2449 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2453 del_pset(remat_defs);
2457 defs = pset_new_ptr_default();
2459 /* skip everything above barrier in start block */
2460 if(is_start_block(bb) && be_is_Barrier(irn)) {
2461 assert(pset_count(live) == 0);
2470 /***************************************
2471 * B E G I N N I N G O F B L O C K
2472 ***************************************/
2475 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2476 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2478 pset_foreach(live, irn) {
2479 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2482 /* construct mem_outs for all values */
2483 set_foreach(spill_bb->ilp, spill) {
2484 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2485 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2487 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2488 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2490 if(pset_find_ptr(live, spill->irn)) {
2491 int default_spilled;
2492 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2494 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2495 default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2496 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2497 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2499 if(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2501 op_t *op = get_irn_link(spill->irn);
2503 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2504 const ir_node *arg = get_irn_n(spill->irn, n);
2510 /* argument already done? */
2511 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2513 /* get sum of execution frequencies of blocks with the same phi argument */
2514 for(m=n; m>=0; --m) {
2515 const ir_node *arg2 = get_irn_n(spill->irn, m);
2518 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2522 /* copies are not for free */
2523 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2524 var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
2526 for(m=n; m>=0; --m) {
2527 const ir_node *arg2 = get_irn_n(spill->irn, m);
2530 op->attr.live_range.args.copies[m] = var;
2535 /* copy <= mem_in */
2536 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2537 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2538 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2539 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2546 foreach_post_remat(bb, tmp) {
2548 op_t *remat_op = get_irn_link(tmp);
2549 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
2552 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2553 remat_arg = get_irn_n(tmp, n);
2555 if(has_reg_class(si, remat_arg)) {
2556 pset_insert_ptr(remat_args, remat_arg);
2560 /* remat + \sum live_range(remat_arg) <= |args| */
2561 ir_snprintf(buf, sizeof(buf), "one_must_die_%N", tmp);
2562 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2563 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2565 pset_foreach(remat_args, remat_arg) {
2566 if(pset_find_ptr(live, remat_arg)) {
2567 op_t *remat_arg_op = get_irn_link(remat_arg);
2568 lpp_set_factor_fast(si->lpp, cst, remat_arg_op->attr.live_range.ilp, 1.0);
2571 del_pset(remat_args);
2574 foreach_post_remat(bb, tmp) {
2577 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2578 ir_node *remat_arg = get_irn_n(tmp, n);
2580 /* if value is becoming live through use by remat2 */
2581 if(has_reg_class(si, remat_arg) && !pset_find_ptr(live, remat_arg)) {
2582 op_t *remat_arg_op = get_irn_link(remat_arg);
2585 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 at bb start %+F\n", remat_arg, tmp));
2587 pset_insert_ptr(live, remat_arg);
2588 spill = add_to_spill_bb(si, bb, remat_arg);
2589 remat_arg_op->attr.live_range.ilp = ILP_UNDEF;
2591 /* we need reg_in and mem_in for this value; they will be referenced later */
2592 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", remat_arg, bb);
2593 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2594 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", remat_arg, bb);
2595 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2598 /* optimization: all memory stuff should be 0, for we do not want to insert reloads for remats */
2599 ir_snprintf(buf, sizeof(buf), "nomem_%N_%N", remat_arg, bb);
2600 nomem = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0.0);
2601 lpp_set_factor_fast(si->lpp, nomem, spill->spill, 1.0);
2606 /* L\U is empty at bb start */
2607 /* arg is live throughout epilog if it is reg_in into this block */
2609 /* check the register pressure at the beginning of the block
2612 /* reg_in entspricht post_use */
2614 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2615 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2617 pset_foreach(live, irn) {
2620 spill = set_find_spill(spill_bb->ilp, irn);
2623 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2624 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2626 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2628 /* spill + mem_in <= 1 */
2629 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2630 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2632 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2633 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2635 } /* post_remats are NOT included in register pressure check because
2636 they do not increase regpressure */
2638 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2639 pset_foreach(live, irn) {
2643 spill = set_find_spill(spill_bb->ilp, irn);
2644 assert(spill && spill->irn == irn);
2646 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2647 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2650 ir_node *phi_arg = get_Phi_pred(irn, n);
2651 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2652 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2654 op_t *op = get_irn_link(irn);
2656 /* although the phi is in the right regclass one or more of
2657 * its arguments can be in a different one or at least to
2660 if(has_reg_class(si, phi_arg)) {
2661 /* mem_in < mem_out_arg + copy */
2662 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2663 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2665 /* reg_in < reg_out_arg */
2666 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2667 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2669 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2670 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2672 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2675 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2677 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2679 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2683 /* else assure the value arrives on all paths in the same resource */
2685 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2688 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2689 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2692 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2693 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2694 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2695 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2697 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2698 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2700 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2703 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2704 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2709 foreach_post_remat(bb, tmp) {
2712 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2713 ir_node *remat_arg = get_irn_n(tmp, n);
2714 op_t *remat_op = get_irn_link(tmp);
2716 if(!has_reg_class(si, remat_arg)) continue;
2718 spill = set_find_spill(spill_bb->ilp, remat_arg);
2721 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2722 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2723 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2724 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2728 pset_foreach(live, irn) {
2729 const op_t *op = get_irn_link(irn);
2730 const ir_node *remat;
2735 foreach_post_remat(bb, remat) {
2738 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2739 const ir_node *arg = get_irn_n(remat, n);
2742 const op_t *remat_op = get_irn_link(remat);
2744 if(cst == ILP_UNDEF) {
2745 /* sum remat2s <= 1 + n_remats*live_range */
2746 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", irn, bb);
2747 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2749 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2755 if(cst != ILP_UNDEF && op->attr.live_range.ilp != ILP_UNDEF) {
2756 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -n_remats);
2760 /* first live ranges from reg_ins */
2761 pset_foreach(live, irn) {
2762 op_t *op = get_irn_link(irn);
2764 if(op->attr.live_range.ilp != ILP_UNDEF) {
2766 spill = set_find_spill(spill_bb->ilp, irn);
2767 assert(spill && spill->irn == irn);
2769 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2770 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2771 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2772 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2774 foreach_post_remat(bb, tmp) {
2775 op_t *remat_op = get_irn_link(tmp);
2777 if(remat_op->attr.remat.remat->value == irn) {
2778 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2784 /* walk forward now and compute constraints for placing spills */
2785 /* this must only be done for values that are not defined in this block */
2786 pset_foreach(live, irn) {
2788 * if value is defined in this block we can anways place the spill directly after the def
2789 * -> no constraint necessary
2791 if(!is_Phi(irn) && get_nodes_block(irn) == bb) {
2796 spill = set_find_spill(spill_bb->ilp, irn);
2799 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2800 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2802 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2803 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2806 sched_foreach_op(bb, tmp) {
2807 op_t *op = get_irn_link(tmp);
2809 if(is_Phi(tmp)) continue;
2810 assert(!is_Proj(tmp));
2813 const ir_node *value = op->attr.remat.remat->value;
2816 /* only collect remats up to the first real use of a value */
2817 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2822 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2823 ir_node *arg = get_irn_n(tmp, n);
2826 /* if a value is used stop collecting remats */
2839 typedef struct _irnlist_t {
2840 struct list_head list;
2844 typedef struct _interference_t {
2845 struct list_head blocklist;
2851 cmp_interference(const void *a, const void *b, size_t size)
2853 const interference_t *p = a;
2854 const interference_t *q = b;
2856 return !(p->a == q->a && p->b == q->b);
2859 static interference_t *
2860 set_find_interference(set * set, ir_node * a, ir_node * b)
2862 interference_t query;
2864 query.a = (a>b)?a:b;
2865 query.b = (a>b)?b:a;
2867 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2870 static interference_t *
2871 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2873 interference_t query,
2875 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2879 result = set_find_interference(set, a, b);
2882 list_add(&list->list, &result->blocklist);
2886 query.a = (a>b)?a:b;
2887 query.b = (a>b)?b:a;
2889 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2891 INIT_LIST_HEAD(&result->blocklist);
2892 list_add(&list->list, &result->blocklist);
2898 values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
2900 const ir_edge_t *edge;
2902 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2903 /* both values are live in, so they interfere */
2907 /* ensure a dominates b */
2908 if(value_dominates(b,a)) {
2914 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2917 /* the following code is stolen from bera.c */
2918 if(be_is_live_end(si->lv, bb, a))
2921 foreach_out_edge(a, edge) {
2922 const ir_node *user = edge->src;
2923 if(get_nodes_block(user) == bb
2926 && !pset_find_ptr(si->inverse_ops, user)
2927 && value_dominates(b, user))
2935 * Walk all irg blocks and collect interfering values inside of phi classes
2938 luke_interferencewalker(ir_node * bb, void * data)
2940 spill_ilp_t *si = (spill_ilp_t*)data;
2943 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2944 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
2945 op_t *a_op = get_irn_link(a);
2948 /* a is only interesting if it is in my register class and if it is inside a phi class */
2949 if (has_reg_class(si, a) && get_phi_class(si->pc, a)) {
2950 if (a_op->is_remat || pset_find_ptr(si->inverse_ops, a))
2953 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)) {
2954 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
2955 op_t *b_op = get_irn_link(b);
2957 /* a and b are only interesting if they are in the same phi class */
2958 if (has_reg_class(si, b) && get_phi_class(si->pc, a) == get_phi_class(si->pc, b)) {
2959 if (b_op->is_remat || pset_find_ptr(si->inverse_ops, b))
2962 if (values_interfere_in_block(si, bb, a, b)) {
2963 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2964 set_insert_interference(si, si->interferences, a, b, bb);
2972 static unsigned int copy_path_id = 0;
2975 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2982 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2983 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2985 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2987 pset_foreach(copies, ptr) {
2988 copy = PTR_TO_INT(ptr);
2989 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2994 * @parameter copies contains a path of copies which lead us to irn
2995 * @parameter visited contains a set of nodes already visited on this path
2998 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
3000 const ir_edge_t *edge;
3001 op_t *op = get_irn_link(irn);
3002 pset *visited_users = pset_new_ptr_default();
3005 if(op->is_remat) return 0;
3007 pset_insert_ptr(visited, irn);
3011 pset *visited_operands = pset_new_ptr(get_irn_arity(irn));
3013 /* visit all operands */
3014 for(n=get_irn_arity(irn)-1; n>=0; --n) {
3015 ir_node *arg = get_irn_n(irn, n);
3016 ilp_var_t copy = op->attr.live_range.args.copies[n];
3018 if(!has_reg_class(si, arg)) continue;
3019 if(pset_find_ptr(visited_operands, arg)) continue;
3020 pset_insert_ptr(visited_operands, arg);
3023 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3024 del_pset(visited_operands);
3025 del_pset(visited_users);
3026 pset_remove_ptr(visited, irn);
3029 pset_insert(copies, INT_TO_PTR(copy), copy);
3030 write_copy_path_cst(si, copies, any_interfere);
3031 pset_remove(copies, INT_TO_PTR(copy), copy);
3032 } else if(!pset_find_ptr(visited, arg)) {
3033 pset_insert(copies, INT_TO_PTR(copy), copy);
3034 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
3035 pset_remove(copies, INT_TO_PTR(copy), copy);
3037 if(paths > MAX_PATHS) {
3038 if(pset_count(copies) == 0) {
3042 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3043 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3044 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3045 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3046 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3050 del_pset(visited_operands);
3051 del_pset(visited_users);
3052 pset_remove_ptr(visited, irn);
3055 } else if(pset_count(copies) == 0) {
3061 del_pset(visited_operands);
3064 /* visit all uses which are phis */
3065 foreach_out_edge(irn, edge) {
3066 ir_node *user = edge->src;
3067 int pos = edge->pos;
3068 op_t *op = get_irn_link(user);
3071 if(!is_Phi(user)) continue;
3072 if(!has_reg_class(si, user)) continue;
3073 if(pset_find_ptr(visited_users, user)) continue;
3074 pset_insert_ptr(visited_users, user);
3076 copy = op->attr.live_range.args.copies[pos];
3078 if(user == target) {
3079 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3080 del_pset(visited_users);
3081 pset_remove_ptr(visited, irn);
3084 pset_insert(copies, INT_TO_PTR(copy), copy);
3085 write_copy_path_cst(si, copies, any_interfere);
3086 pset_remove(copies, INT_TO_PTR(copy), copy);
3087 } else if(!pset_find_ptr(visited, user)) {
3088 pset_insert(copies, INT_TO_PTR(copy), copy);
3089 paths += find_copy_path(si, user, target, any_interfere, copies, visited);
3090 pset_remove(copies, INT_TO_PTR(copy), copy);
3092 if(paths > MAX_PATHS) {
3093 if(pset_count(copies) == 0) {
3097 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3098 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3099 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3100 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3101 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3105 del_pset(visited_users);
3106 pset_remove_ptr(visited, irn);
3109 } else if(pset_count(copies) == 0) {
3115 del_pset(visited_users);
3116 pset_remove_ptr(visited, irn);
3121 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
3123 pset * copies = pset_new_ptr_default();
3124 pset * visited = pset_new_ptr_default();
3126 find_copy_path(si, a, b, any_interfere, copies, visited);
3134 memcopyhandler(spill_ilp_t * si)
3136 interference_t *interference;
3138 /* teste Speicherwerte auf Interferenz */
3140 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
3141 irg_block_walk_graph(si->birg->irg, luke_interferencewalker, NULL, si);
3143 /* now lets emit the ILP unequations for the crap */
3144 set_foreach(si->interferences, interference) {
3146 ilp_var_t interfere, any_interfere;
3147 ilp_cst_t any_interfere_cst, cst;
3148 const ir_node *a = interference->a;
3149 const ir_node *b = interference->b;
3151 /* any_interf <= \sum interf */
3152 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
3153 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3154 any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3156 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
3158 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
3159 const ir_node *bb = irnlist->irn;
3160 spill_bb_t *spill_bb = get_irn_link(bb);
3165 spilla = set_find_spill(spill_bb->ilp, a);
3168 spillb = set_find_spill(spill_bb->ilp, b);
3171 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
3172 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
3173 /* 2: - mem_in_a - spill_a + interfere <= 0 */
3174 /* 3: - mem_in_b - spill_b + interfere <= 0 */
3175 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
3176 interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3178 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
3179 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
3181 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
3182 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
3183 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
3184 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
3185 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
3187 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
3188 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3190 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3191 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
3192 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
3194 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
3195 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3197 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3198 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
3199 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
3202 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
3204 /* any_interfere >= interf */
3205 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
3206 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3208 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3209 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3212 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
3213 gen_copy_constraints(si,a,b,any_interfere);
3221 return fabs(x) < 0.00001;
3224 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
3226 spill_ilp_t *si = get_irg_link(current_ir_graph);
3228 if(pset_find_ptr(si->all_possible_remats, n)) {
3229 op_t *op = (op_t*)get_irn_link(n);
3230 assert(op && op->is_remat);
3232 if(!op->attr.remat.remat->inverse) {
3233 if(op->attr.remat.pre) {
3234 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
3236 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
3241 op_t *op = (op_t*)get_irn_link(n);
3242 assert(op && op->is_remat);
3244 if(op->attr.remat.pre) {
3245 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
3247 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
3258 dump_graph_with_remats(ir_graph * irg, const char * suffix)
3260 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
3261 be_dump(irg, suffix, dump_ir_block_graph_sched);
3262 set_dump_node_vcgattr_hook(NULL);
3266 * Edge hook to dump the schedule edges with annotated register pressure.
3269 sched_pressure_edge_hook(FILE *F, ir_node *irn)
3271 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
3272 ir_node *prev = sched_prev(irn);
3273 fprintf(F, "edge:{sourcename:\"");
3275 fprintf(F, "\" targetname:\"");
3277 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
3278 fprintf(F, "\" color:magenta}\n");
3284 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
3286 DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3288 dump_consts_local(0);
3289 set_dump_node_edge_hook(sched_pressure_edge_hook);
3290 dump_ir_block_graph(irg, suffix);
3291 set_dump_node_edge_hook(old_edge_hook);
3295 walker_pressure_annotator(ir_node * bb, void * data)
3297 spill_ilp_t *si = data;
3300 pset *live = pset_new_ptr_default();
3303 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3304 irn = be_lv_get_irn(si->lv, bb, i);
3306 if (has_reg_class(si, irn)) {
3307 pset_insert_ptr(live, irn);
3311 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3313 sched_foreach_reverse(bb, irn) {
3315 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3319 if(has_reg_class(si, irn)) {
3320 pset_remove_ptr(live, irn);
3321 if(is_Proj(irn)) ++projs;
3324 if(!is_Proj(irn)) projs = 0;
3326 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3327 ir_node *arg = get_irn_n(irn, n);
3329 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3331 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3338 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3340 be_dump(si->birg->irg, suffix, dump_ir_block_graph_sched_pressure);
3344 connect_all_remats_with_keep(spill_ilp_t * si)
3352 n_remats = pset_count(si->all_possible_remats);
3354 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3357 pset_foreach(si->all_possible_remats, irn) {
3362 si->keep = be_new_Keep(si->cls, si->birg->irg, get_irg_end_block(si->birg->irg), n_remats, ins);
3364 obstack_free(si->obst, ins);
3369 connect_all_spills_with_keep(spill_ilp_t * si)
3378 n_spills = pset_count(si->spills);
3380 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3383 pset_foreach(si->spills, irn) {
3388 keep = be_new_Keep(si->cls, si->birg->irg, get_irg_end_block(si->birg->irg), n_spills, ins);
3390 obstack_free(si->obst, ins);
3394 /** insert a spill at an arbitrary position */
3395 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3397 ir_node *bl = is_Block(insert) ? insert : get_nodes_block(insert);
3398 ir_graph *irg = get_irn_irg(bl);
3399 ir_node *frame = get_irg_frame(irg);
3402 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
3403 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3405 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3408 * search the right insertion point. a spill of a phi cannot be put
3409 * directly after the phi, if there are some phis behind the one which
3410 * is spilled. Also, a spill of a Proj must be after all Projs of the
3413 * Here's one special case:
3414 * If the spill is in the start block, the spill must be after the frame
3415 * pointer is set up. This is done by setting insert to the end of the block
3416 * which is its default initialization (see above).
3419 if (bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3422 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3425 sched_add_after(insert, spill);
3430 delete_remat(spill_ilp_t * si, ir_node * remat) {
3432 ir_node *bad = get_irg_bad(si->birg->irg);
3434 sched_remove(remat);
3436 /* kill links to operands */
3437 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3438 set_irn_n(remat, n, bad);
3443 clean_remat_info(spill_ilp_t * si)
3447 remat_info_t *remat_info;
3448 ir_node *bad = get_irg_bad(si->birg->irg);
3450 set_foreach(si->remat_info, remat_info) {
3451 if(!remat_info->remats) continue;
3453 pset_foreach(remat_info->remats, remat)
3455 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3456 if(sched_is_scheduled(remat->proj)) {
3457 sched_remove((ir_node*)remat->proj);
3459 set_irn_n((ir_node*)remat->proj, -1, bad);
3460 set_irn_n((ir_node*)remat->proj, 0, bad);
3463 if(get_irn_n_edges(remat->op) == 0) {
3464 if(sched_is_scheduled(remat->op)) {
3465 sched_remove((ir_node*)remat->op);
3467 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3468 set_irn_n((ir_node*)remat->op, n, bad);
3473 if(remat_info->remats) del_pset(remat_info->remats);
3474 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3479 delete_unnecessary_remats(spill_ilp_t * si)
3481 if(opt_keep_alive & KEEPALIVE_REMATS) {
3483 ir_node *bad = get_irg_bad(si->birg->irg);
3486 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3487 ir_node *keep_arg = get_irn_n(si->keep, n);
3488 op_t *arg_op = get_irn_link(keep_arg);
3491 assert(arg_op->is_remat);
3493 name = si->lpp->vars[arg_op->attr.remat.ilp];
3495 if(is_zero(name->value)) {
3496 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
3497 /* TODO check whether reload is preferred over remat (could be bug) */
3498 delete_remat(si, keep_arg);
3500 if(!arg_op->attr.remat.remat->inverse) {
3501 if(arg_op->attr.remat.pre) {
3502 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3504 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3507 if(arg_op->attr.remat.pre) {
3508 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3510 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3515 set_irn_n(si->keep, n, bad);
3518 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
3523 pset_foreach(si->all_possible_remats, remat) {
3524 op_t *remat_op = get_irn_link(remat);
3525 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
3527 if(is_zero(name->value)) {
3528 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
3529 /* TODO check whether reload is preferred over remat (could be bug) */
3530 delete_remat(si, remat);
3532 if(!remat_op->attr.remat.remat->inverse) {
3533 if(remat_op->attr.remat.pre) {
3534 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3536 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3539 if(remat_op->attr.remat.pre) {
3540 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3542 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3551 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3553 pset *spills = pset_new_ptr_default();
3555 const ir_node *next;
3558 defs = set_find_def(si->values, value);
3560 if(defs && defs->spills) {
3561 for(next = defs->spills; next; next = get_irn_link(next)) {
3562 pset_insert_ptr(spills, next);
3570 new_r_PhiM_nokeep(ir_graph * irg, ir_node *block, int arity, ir_node **in)
3574 assert( get_irn_arity(block) == arity );
3576 res = new_ir_node(NULL, irg, block, op_Phi, mode_M, arity, in);
3577 res->attr.phi_backedge = new_backedge_arr(irg->obst, arity);
3583 * @param before The node after which the spill will be placed in the schedule
3586 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3590 const arch_env_t *arch_env = si->birg->main_env->arch_env;
3592 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3594 spill = be_spill2(arch_env, irn, before);
3596 defs = set_insert_def(si->values, value);
3599 /* enter into the linked list */
3600 set_irn_link(spill, defs->spills);
3601 defs->spills = spill;
3603 if(opt_keep_alive & KEEPALIVE_SPILLS)
3604 pset_insert_ptr(si->spills, spill);
3610 * @param before The Phi node which has to be spilled
3613 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3620 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3622 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3623 ins[n] = si->m_unknown;
3626 mem_phi = new_r_PhiM_nokeep(si->birg->irg, get_nodes_block(phi), get_irn_arity(phi), ins);
3628 defs = set_insert_def(si->values, phi);
3631 /* enter into the linked list */
3632 set_irn_link(mem_phi, defs->spills);
3633 defs->spills = mem_phi;
3635 #ifdef SCHEDULE_PHIM
3636 sched_add_after(phi, mem_phi);
3638 pset_insert_ptr(si->phims, mem_phi);
3641 if(opt_keep_alive & KEEPALIVE_SPILLS)
3642 pset_insert_ptr(si->spills, mem_phi);
3649 * Add remat to list of defs, destroys link field!
3652 insert_remat(spill_ilp_t * si, ir_node * remat)
3655 op_t *remat_op = get_irn_link(remat);
3657 assert(remat_op->is_remat);
3659 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3662 /* enter into the linked list */
3663 set_irn_link(remat, defs->remats);
3664 defs->remats = remat;
3669 * Add reload before operation and add to list of defs
3672 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3677 const arch_env_t *arch_env = si->birg->main_env->arch_env;
3679 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3681 defs = set_find_def(si->values, value);
3683 spill = defs->spills;
3684 assert(spill && "no spill placed before reload");
3686 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3688 /* enter into the linked list */
3689 set_irn_link(reload, defs->remats);
3690 defs->remats = reload;
3695 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3698 ir_node *value = get_irn_n(memoperand->irn, memoperand->pos);
3700 const arch_env_t *arch_env = si->birg->main_env->arch_env;
3702 DBG((si->dbg, LEVEL_2, "\t inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3704 defs = set_find_def(si->values, value);
3706 spill = defs->spills;
3707 assert(spill && "no spill placed before reload");
3709 arch_perform_memory_operand(arch_env, memoperand->irn, spill, memoperand->pos);
3712 void insert_memoperands(spill_ilp_t * si)
3714 memoperand_t *memoperand;
3717 set_foreach(si->memoperands, memoperand) {
3718 name = si->lpp->vars[memoperand->ilp];
3719 if(!is_zero(name->value)) {
3720 perform_memory_operand(si, memoperand);
3726 walker_spill_placer(ir_node * bb, void * data) {
3727 spill_ilp_t *si = (spill_ilp_t*)data;
3729 spill_bb_t *spill_bb = get_irn_link(bb);
3730 pset *spills_to_do = pset_new_ptr_default();
3733 set_foreach(spill_bb->ilp, spill) {
3736 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3737 name = si->lpp->vars[spill->mem_in];
3738 if(!is_zero(name->value)) {
3741 mem_phi = insert_mem_phi(si, spill->irn);
3743 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3747 name = si->lpp->vars[spill->spill];
3748 if(!is_zero(name->value)) {
3749 /* place spill directly after definition */
3750 if(get_nodes_block(spill->irn) == bb) {
3751 insert_spill(si, spill->irn, spill->irn, spill->irn);
3755 /* place spill at bb start */
3756 if(spill->reg_in > 0) {
3757 name = si->lpp->vars[spill->reg_in];
3758 if(!is_zero(name->value)) {
3759 insert_spill(si, spill->irn, spill->irn, bb);
3763 /* place spill after a remat */
3764 pset_insert_ptr(spills_to_do, spill->irn);
3767 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3770 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3771 op_t *op = get_irn_link(irn);
3773 if(be_is_Spill(irn)) continue;
3776 /* TODO fix this if we want to support remats with more than two nodes */
3777 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3778 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3780 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3783 if(pset_find_ptr(spills_to_do, irn)) {
3784 pset_remove_ptr(spills_to_do, irn);
3786 insert_spill(si, irn, irn, irn);
3792 assert(pset_count(spills_to_do) == 0);
3794 /* afterwards free data in block */
3795 del_pset(spills_to_do);
3799 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3801 ir_node *insert_pos = bb;
3803 const arch_env_t *arch_env = si->birg->main_env->arch_env;
3805 /* find last definition of arg value in block */
3810 defs = set_find_def(si->values, value);
3812 if(defs && defs->remats) {
3813 for(next = defs->remats; next; next = get_irn_link(next)) {
3814 if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3815 last = sched_get_time_step(next);
3821 if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3822 last = sched_get_time_step(value);
3826 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
3828 spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3834 phim_fixer(spill_ilp_t *si) {
3837 set_foreach(si->values, defs) {
3838 const ir_node *phi = defs->value;
3839 op_t *op = get_irn_link(phi);
3840 ir_node *phi_m = NULL;
3841 ir_node *next = defs->spills;
3844 if(!is_Phi(phi)) continue;
3847 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3851 next = get_irn_link(next);
3854 if(!phi_m) continue;
3856 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3857 ir_node *value = get_irn_n(phi, n);
3858 defs_t *val_defs = set_find_def(si->values, value);
3860 /* a spill of this value */
3865 ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3866 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3868 if(!is_zero(name->value)) {
3869 spill = insert_mem_copy(si, pred, value);
3871 spill = val_defs->spills;
3874 spill = val_defs->spills;
3877 assert(spill && "no spill placed before PhiM");
3878 set_irn_n(phi_m, n, spill);
3884 walker_reload_placer(ir_node * bb, void * data) {
3885 spill_ilp_t *si = (spill_ilp_t*)data;
3887 spill_bb_t *spill_bb = get_irn_link(bb);
3889 /* reloads at end of block */
3890 if(spill_bb->reloads) {
3893 set_foreach(spill_bb->reloads, keyval) {
3894 ir_node *irn = (ir_node*)keyval->key;
3895 ilp_var_t reload = PTR_TO_INT(keyval->val);
3898 name = si->lpp->vars[reload];
3899 if(!is_zero(name->value)) {
3901 ir_node *insert_pos = bb;
3902 ir_node *prev = sched_block_last_noncf(si, bb);
3903 op_t *prev_op = get_irn_link(prev);
3905 while(be_is_Spill(prev)) {
3906 prev = sched_prev(prev);
3909 prev_op = get_irn_link(prev);
3911 /* insert reload before pre-remats */
3912 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3913 && prev_op->is_remat && prev_op->attr.remat.pre) {
3917 prev = sched_prev(prev);
3918 } while(be_is_Spill(prev));
3920 prev_op = get_irn_link(prev);
3924 reload = insert_reload(si, irn, insert_pos);
3926 if(opt_keep_alive & KEEPALIVE_RELOADS)
3927 pset_insert_ptr(si->spills, reload);
3932 /* walk and insert more reloads and collect remats */
3933 sched_foreach_reverse(bb, irn) {
3934 op_t *op = get_irn_link(irn);
3936 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3937 if(is_Phi(irn)) break;
3940 if(get_irn_mode(irn) != mode_T) {
3941 insert_remat(si, irn);
3946 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3947 ir_node *arg = get_irn_n(irn, n);
3949 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3952 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3953 if(!is_zero(name->value)) {
3955 ir_node *insert_pos = irn;
3956 ir_node *prev = sched_prev(insert_pos);
3959 while(be_is_Spill(prev)) {
3960 prev = sched_prev(prev);
3963 prev_op = get_irn_link(prev);
3965 /* insert reload before pre-remats */
3966 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3967 && prev_op->is_remat && prev_op->attr.remat.pre) {
3971 prev = sched_prev(prev);
3972 } while(be_is_Spill(prev));
3974 prev_op = get_irn_link(prev);
3978 reload = insert_reload(si, arg, insert_pos);
3980 assert(reload && "no reload returned");
3981 set_irn_n(irn, n, reload);
3983 if(opt_keep_alive & KEEPALIVE_RELOADS)
3984 pset_insert_ptr(si->spills, reload);
3991 del_set(spill_bb->ilp);
3992 if(spill_bb->reloads) del_set(spill_bb->reloads);
3996 walker_collect_used(ir_node * irn, void * data)
3998 bitset_t *used = data;
4000 bitset_set(used, get_irn_idx(irn));
4003 struct kill_helper {
4009 walker_kill_unused(ir_node * bb, void * data)
4011 struct kill_helper *kh = data;
4012 ir_node *bad = get_irg_bad(get_irn_irg(bb));
4016 for(irn=sched_first(bb); !sched_is_end(irn);) {
4017 ir_node *next = sched_next(irn);
4020 if(!bitset_is_set(kh->used, get_irn_idx(irn))) {
4021 if(be_is_Spill(irn) || be_is_Reload(irn)) {
4022 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)));
4024 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
4030 set_nodes_block(irn, bad);
4031 for (n=get_irn_arity(irn)-1; n>=0; --n) {
4032 set_irn_n(irn, n, bad);
4039 #ifndef SCHEDULE_PHIM
4041 kill_unused_phims(spill_ilp_t * si, struct kill_helper * kh)
4044 ir_node *bad = get_irg_bad(si->birg->irg);
4047 pset_foreach(si->phims, phi) {
4048 if(!bitset_is_set(kh->used, get_irn_idx(phi))) {
4050 set_nodes_block(phi, bad);
4051 for (n=get_irn_arity(phi)-1; n>=0; --n) {
4052 set_irn_n(phi, n, bad);
4060 kill_all_unused_values_in_schedule(spill_ilp_t * si)
4062 struct kill_helper kh;
4064 kh.used = bitset_malloc(get_irg_last_idx(si->birg->irg));
4067 irg_walk_graph(si->birg->irg, walker_collect_used, NULL, kh.used);
4068 #ifndef SCHEDULE_PHIM
4069 kill_unused_phims(si, &kh);
4071 irg_block_walk_graph(si->birg->irg, walker_kill_unused, NULL, &kh);
4073 bitset_free(kh.used);
4077 print_irn_pset(pset * p)
4081 pset_foreach(p, irn) {
4082 ir_printf("%+F\n", irn);
4087 dump_phi_class(spill_ilp_t *si, ir_node **phiclass, const char * file)
4089 FILE *f = fopen(file, "w");
4091 interference_t *interference;
4094 set_break(si->interferences);
4096 ir_fprintf(f, "digraph phiclass {\n");
4098 for (i = ARR_LEN(phiclass) - 1; i >= 0; --i) {
4101 ir_fprintf(f, " %F%N [shape=box]\n", irn, irn);
4104 for (i = ARR_LEN(phiclass) - 1; i >= 0; --i) {
4111 for (n = get_irn_arity(irn) - 1; n >= 0; --n) {
4112 ir_node *arg = get_irn_n(irn, n);
4114 ir_fprintf(f, " %F%N -> %F%N\n", irn, irn, arg, arg);
4118 set_foreach(si->interferences, interference) {
4119 const ir_node *a = interference->a;
4120 const ir_node *b = interference->b;
4121 if (get_phi_class(si->pc, (ir_node *)a) == phiclass) {
4122 ir_fprintf(f, " %F%N -> %F%N [color=red,dir=none,style=bold]\n", a, a, b, b);
4131 rewire_uses(spill_ilp_t * si)
4134 ir_nodeset_t ignore;
4136 ir_nodeset_init(&ignore);
4137 ir_nodeset_insert(&ignore, get_irg_end(si->birg->irg));
4139 /* then fix uses of spills */
4140 set_foreach(si->values, defs) {
4143 const ir_node *next = defs->remats;
4146 reloads = pset_new_ptr_default();
4149 if(be_is_Reload(next)) {
4150 pset_insert_ptr(reloads, next);
4154 next = get_irn_link(next);
4157 spills = get_spills_for_value(si, defs->value);
4158 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));
4159 if(pset_count(spills) > 1) {
4160 be_ssa_construction_env_t senv;
4162 //assert(pset_count(reloads) > 0);
4163 // print_irn_pset(spills);
4164 // print_irn_pset(reloads);
4166 be_ssa_construction_init(&senv, si->birg);
4167 be_ssa_construction_set_ignore_uses(&senv, &ignore);
4168 pset_foreach(spills, node) {
4169 be_ssa_construction_add_copy(&senv, node);
4171 pset_foreach(spills, node) {
4172 be_ssa_construction_fix_users(&senv, node);
4174 be_ssa_construction_update_liveness_phis(&senv, si->lv);
4175 pset_foreach(spills, node) {
4176 be_liveness_update(si->lv, node);
4178 be_ssa_construction_destroy(&senv);
4185 /* first fix uses of remats and reloads */
4186 set_foreach(si->values, defs) {
4187 const ir_node *next = defs->remats;
4191 be_ssa_construction_env_t senv;
4193 be_ssa_construction_init(&senv, si->birg);
4195 if(sched_is_scheduled(defs->value)) {
4196 be_ssa_construction_add_copy(&senv, (ir_node*) defs->value);
4200 next = defs->remats;
4202 be_ssa_construction_add_copy(&senv, (ir_node*) next);
4203 next = get_irn_link(next);
4206 if(sched_is_scheduled(defs->value)) {
4207 be_ssa_construction_fix_users(&senv, (ir_node*) defs->value);
4210 next = defs->remats;
4212 be_ssa_construction_fix_users(&senv, (ir_node*) next);
4213 next = get_irn_link(next);
4216 be_ssa_construction_update_liveness_phis(&senv, si->lv);
4217 if(sched_is_scheduled(defs->value)) {
4218 be_liveness_update(si->lv, (ir_node*) defs->value);
4221 next = defs->remats;
4223 be_liveness_update(si->lv, (ir_node*) next);
4224 next = get_irn_link(next);
4227 be_ssa_construction_destroy(&senv);
4231 ir_nodeset_destroy(&ignore);
4232 // remove_unused_defs(si);
4237 writeback_results(spill_ilp_t * si)
4239 /* walk through the graph and collect all spills, reloads and remats for a value */
4241 si->values = new_set(cmp_defs, 4096);
4243 DBG((si->dbg, LEVEL_1, "Applying results\n"));
4244 delete_unnecessary_remats(si);
4245 si->m_unknown = new_r_Unknown(si->birg->irg, mode_M);
4246 irg_block_walk_graph(si->birg->irg, walker_spill_placer, NULL, si);
4247 irg_block_walk_graph(si->birg->irg, walker_reload_placer, NULL, si);
4249 insert_memoperands(si);
4252 /* clean the remat info! there are still back-edges leading there! */
4253 clean_remat_info(si);
4257 connect_all_spills_with_keep(si);
4259 del_set(si->values);
4263 get_n_regs(spill_ilp_t * si)
4265 int arch_n_regs = arch_register_class_n_regs(si->cls);
4267 bitset_t *arch_regs = bitset_malloc(arch_n_regs);
4268 bitset_t *abi_regs = bitset_malloc(arch_n_regs);
4270 arch_put_non_ignore_regs(si->birg->main_env->arch_env, si->cls, arch_regs);
4271 be_abi_put_ignore_regs(si->birg->abi, si->cls, abi_regs);
4273 bitset_andnot(arch_regs, abi_regs);
4274 arch_n_regs = bitset_popcnt(arch_regs);
4276 bitset_free(arch_regs);
4277 bitset_free(abi_regs);
4279 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", arch_n_regs, si->cls->name));
4284 walker_reload_mover(ir_node * bb, void * data)
4286 spill_ilp_t *si = data;
4289 sched_foreach(bb, tmp) {
4290 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
4291 ir_node *reload = tmp;
4294 /* move reload upwards */
4296 int pressure = (int)get_irn_link(reload);
4297 if(pressure < si->n_regs) {
4298 irn = sched_prev(reload);
4299 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
4300 sched_remove(reload);
4301 pressure = (int)get_irn_link(irn);
4303 while(pressure < si->n_regs) {
4304 if( sched_is_end(irn) ||
4305 (be_is_Reload(irn) && has_reg_class(si, irn)) ||
4306 /* do not move reload before its spill */
4307 (irn == be_get_Reload_mem(reload)) ||
4308 /* do not move before phi */
4311 set_irn_link(irn, INT_TO_PTR(pressure+1));
4312 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
4313 irn = sched_prev(irn);
4315 pressure = (int)get_irn_link(irn);
4318 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
4319 sched_put_after(irn, reload);
4326 move_reloads_upward(spill_ilp_t * si)
4328 irg_block_walk_graph(si->birg->irg, walker_reload_mover, NULL, si);
4333 * Walk all irg blocks and check for interfering spills inside of phi classes
4336 luke_meminterferencechecker(ir_node * bb, void * data)
4338 spill_ilp_t *si = (spill_ilp_t*)data;
4341 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
4342 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
4344 if (! be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T))
4347 /* a is only interesting if it is in my register class and if it is inside a phi class */
4348 if (has_reg_class(si, a) && get_phi_class(si->pc, a)) {
4349 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)) {
4350 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
4352 if (! be_is_Spill(b) && (! is_Phi(b) || get_irn_mode(b) != mode_T))
4355 /* a and b are only interesting if they are in the same phi class */
4356 if (has_reg_class(si, b) && get_phi_class(si->pc, a) == get_phi_class(si->pc, b)) {
4357 if (values_interfere_in_block(si, bb, a, b)) {
4358 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4367 verify_phiclasses(spill_ilp_t * si)
4369 /* analyze phi classes */
4370 phi_class_free(si->pc);
4371 si->pc = phi_class_new_from_irg(si->birg->irg, 0);
4373 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4374 irg_block_walk_graph(si->birg->irg, luke_meminterferencechecker, NULL, si);
4378 be_spill_remat(be_irg_t *birg, const arch_register_class_t *cls)
4381 char problem_name[256];
4382 char dump_suffix[256];
4383 char dump_suffix2[256];
4384 struct obstack obst;
4386 ir_graph *irg = be_get_birg_irg(birg);
4388 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", irg, cls->name);
4389 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", cls->name);
4390 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", cls->name);
4392 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4393 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4395 if(opt_verify & VERIFY_DOMINANCE)
4396 be_check_dominance(irg);
4398 be_assure_dom_front(birg);
4399 be_assure_liveness(birg);
4401 obstack_init(&obst);
4405 si.lpp = new_lpp(problem_name, lpp_minimize);
4406 si.remat_info = new_set(cmp_remat_info, 4096);
4407 si.interferences = new_set(cmp_interference, 32);
4408 si.memoperands = new_set(cmp_memoperands, 128);
4409 si.all_possible_remats = pset_new_ptr_default();
4410 si.spills = pset_new_ptr_default();
4411 si.inverse_ops = pset_new_ptr_default();
4414 si.n_regs = get_n_regs(&si);
4416 set_irg_link(irg, &si);
4419 /* compute phi classes */
4420 // phi_class_compute(irg);
4422 if(opt_dump_flags & DUMP_STATS)
4423 be_analyze_regpressure(birg, cls, "-pre");
4425 DBG((si.dbg, LEVEL_2, "\t initializing\n"));
4426 irg_block_walk_graph(irg, luke_initializer, NULL, &si);
4429 /* collect remats */
4430 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4431 irg_walk_graph(irg, walker_remat_collector, NULL, &si);
4434 /* insert possible remats */
4435 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4436 irg_block_walk_graph(irg, walker_remat_insertor, NULL, &si);
4437 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4439 if(opt_keep_alive & KEEPALIVE_REMATS) {
4440 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4441 connect_all_remats_with_keep(&si);
4442 /* dump graph with inserted remats */
4443 dump_graph_with_remats(irg, dump_suffix);
4446 /* insert copies for phi arguments not in my regclass */
4447 irg_walk_graph(irg, walker_regclass_copy_insertor, NULL, &si);
4449 /* recompute liveness */
4450 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4451 be_liveness_recompute(si.lv);
4454 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4455 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4456 irg_block_walk_graph(irg, luke_endwalker, NULL, &si);
4458 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4459 irg_block_walk_graph(irg, luke_blockwalker, NULL, &si);
4461 si.pc = phi_class_new_from_irg(birg->irg, 0);
4462 if (opt_memcopies) {
4463 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4464 memcopyhandler(&si);
4467 if (opt_dump_flags & DUMP_PROBLEM) {
4469 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4470 if ((f = fopen(buf, "wt")) != NULL) {
4471 lpp_dump_plain(si.lpp, f);
4476 if (opt_dump_flags & DUMP_MPS) {
4479 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4480 if ((f = fopen(buf, "wt")) != NULL) {
4481 mps_write_mps(si.lpp, s_mps_fixed, f);
4485 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4486 if ((f = fopen(buf, "wt")) != NULL) {
4487 mps_write_mst(si.lpp, s_mps_fixed, f);
4492 lpp_check_startvals(si.lpp);
4495 DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4496 lpp_set_time_limit(si.lpp, opt_timeout);
4499 lpp_set_log(si.lpp, stdout);
4502 lpp_solve_cplex(si.lpp);
4504 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4506 assert(lpp_is_sol_valid(si.lpp)
4507 && "solution of ILP must be valid");
4509 DBG((si.dbg, LEVEL_1, "\t%s: iterations: %d, solution time: %g, objective function: %g, best bound: %g\n", problem_name, si.lpp->iterations, si.lpp->sol_time, is_zero(si.lpp->objval)?0.0:si.lpp->objval, is_zero(si.lpp->best_bound)?0.0:si.lpp->best_bound));
4511 if(opt_dump_flags & DUMP_SOLUTION) {
4515 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4516 if ((f = fopen(buf, "wt")) != NULL) {
4518 for (i = 0; i < si.lpp->var_next; ++i) {
4519 lpp_name_t *name = si.lpp->vars[i];
4520 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4526 #ifndef SCHEDULE_PHIM
4527 si.phims = pset_new_ptr_default();
4529 writeback_results(&si);
4534 kill_all_unused_values_in_schedule(&si);
4536 #if !defined(SCHEDULE_PHIM) && defined(SOLVE)
4540 if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
4541 be_dump(irg, "-spills-placed", dump_ir_block_graph);
4543 // move reloads upwards
4544 be_liveness_recompute(si.lv);
4545 irg_block_walk_graph(irg, walker_pressure_annotator, NULL, &si);
4546 move_reloads_upward(&si);
4549 verify_phiclasses(&si);
4552 irg_block_walk_graph(irg, walker_pressure_annotator, NULL, &si);
4554 if(opt_dump_flags & DUMP_PRESSURE)
4555 dump_pressure_graph(&si, dump_suffix2);
4557 if(opt_dump_flags & DUMP_STATS)
4558 be_analyze_regpressure(birg, cls, "-post");
4560 if(opt_verify & VERIFY_DOMINANCE)
4561 be_check_dominance(irg);
4564 del_set(si.interferences);
4565 del_pset(si.inverse_ops);
4566 del_pset(si.all_possible_remats);
4567 del_set(si.memoperands);
4568 del_pset(si.spills);
4570 phi_class_free(si.pc);
4571 obstack_free(&obst, NULL);
4572 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4575 void be_init_spillremat(void)
4577 static be_spiller_t remat_spiller = {
4580 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
4581 lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
4582 lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
4583 lc_opt_entry_t *remat_grp = lc_opt_get_grp(chordal_grp, "remat");
4585 be_register_spiller("remat", &remat_spiller);
4586 lc_opt_add_table(remat_grp, options);
4589 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillremat);
4591 #else /* WITH_ILP */
4594 only_that_you_can_compile_without_WITH_ILP_defined(void)
4598 #endif /* WITH_ILP */