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>
47 #include "besched_t.h"
48 #include "bessaconstr.h"
53 #include "bespillremat.h"
55 #include "bepressurestat.h"
56 #include "beprofile.h"
57 #include "bespilloptions.h"
59 #include "bechordal_t.h"
61 #include <libcore/lc_opts.h>
62 #include <libcore/lc_opts_enum.h>
64 #define DUMP_PROBLEM 1
66 #define DUMP_SOLUTION 4
68 #define DUMP_PRESSURE 16
70 #define KEEPALIVE_REMATS 1
71 #define KEEPALIVE_SPILLS 2
72 #define KEEPALIVE_RELOADS 4
74 #define VERIFY_MEMINTERF 1
75 #define VERIFY_DOMINANCE 2
78 #define REMATS_BRIGGS 1
79 #define REMATS_NOINVERSE 2
82 static unsigned opt_dump_flags = 0;
83 static int opt_log = 0;
84 static unsigned opt_keep_alive = 0;
85 static int opt_goodwin = 1;
86 static int opt_memcopies = 1;
87 static int opt_memoperands = 1;
88 static int opt_verify = VERIFY_MEMINTERF;
89 static unsigned opt_remats = REMATS_ALL;
90 static int opt_repair_schedule = 0;
91 static int opt_no_enlarge_liveness = 0;
92 static int opt_remat_while_live = 1;
93 static int opt_timeout = 300;
94 static double opt_cost_reload = 8.0;
95 static double opt_cost_memoperand = 7.0;
96 static double opt_cost_spill = 15.0;
97 static double opt_cost_remat = 1.0;
100 static const lc_opt_enum_mask_items_t dump_items[] = {
101 { "problem", DUMP_PROBLEM },
103 { "solution", DUMP_SOLUTION },
104 { "stats", DUMP_STATS },
105 { "pressure", DUMP_PRESSURE },
109 static lc_opt_enum_mask_var_t dump_var = {
110 &opt_dump_flags, dump_items
113 static const lc_opt_enum_mask_items_t keepalive_items[] = {
114 { "remats", KEEPALIVE_REMATS },
115 { "spills", KEEPALIVE_SPILLS },
116 { "reloads", KEEPALIVE_RELOADS },
120 static lc_opt_enum_mask_var_t keep_alive_var = {
121 &opt_keep_alive, keepalive_items
124 static const lc_opt_enum_mask_items_t remats_items[] = {
125 { "none", REMATS_NONE },
126 { "briggs", REMATS_BRIGGS },
127 { "noinverse", REMATS_NOINVERSE },
128 { "all", REMATS_ALL },
132 static lc_opt_enum_mask_var_t remats_var = {
133 &opt_remats, remats_items
136 static const lc_opt_table_entry_t options[] = {
137 LC_OPT_ENT_ENUM_MASK("keepalive", "keep alive inserted nodes", &keep_alive_var),
139 LC_OPT_ENT_BOOL ("goodwin", "activate goodwin reduction", &opt_goodwin),
140 LC_OPT_ENT_BOOL ("memcopies", "activate memcopy handling", &opt_memcopies),
141 LC_OPT_ENT_BOOL ("memoperands", "activate memoperands", &opt_memoperands),
142 LC_OPT_ENT_ENUM_INT ("remats", "type of remats to insert", &remats_var),
143 LC_OPT_ENT_BOOL ("repair_schedule", "repair the schedule by rematting once used nodes",&opt_repair_schedule),
144 LC_OPT_ENT_BOOL ("no_enlage_liveness", "do not enlarge liveness of operands of remats",&opt_no_enlarge_liveness),
145 LC_OPT_ENT_BOOL ("remat_while_live", "only remat where rematted value was live", &opt_remat_while_live),
147 LC_OPT_ENT_ENUM_MASK("dump", "dump problem, solution or statistical data", &dump_var),
148 LC_OPT_ENT_BOOL ("log", "activate the lpp log", &opt_log),
149 LC_OPT_ENT_INT ("timeout", "ILP solver timeout", &opt_timeout),
151 LC_OPT_ENT_DBL ("cost_reload", "cost of a reload", &opt_cost_reload),
152 LC_OPT_ENT_DBL ("cost_memoperand", "cost of a memory operand", &opt_cost_memoperand),
153 LC_OPT_ENT_DBL ("cost_spill", "cost of a spill instruction", &opt_cost_spill),
154 LC_OPT_ENT_DBL ("cost_remat", "cost of a rematerialization", &opt_cost_remat),
158 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
159 //#define SCHEDULE_PHIM /* insert phim nodes into schedule */
162 //#define SOLVE_LOCAL
163 #define LPP_SERVER "i44pc52"
164 #define LPP_SOLVER "cplex"
167 #define MAX_PATHS INT_MAX
170 typedef struct _spill_ilp_t {
171 const arch_register_class_t *cls;
176 struct obstack *obst;
178 pset *all_possible_remats;
181 set *values; /**< for collecting all definitions of values before running ssa-construction */
187 #ifndef SCHEDULE_PHIM
190 DEBUG_ONLY(firm_dbg_module_t * dbg);
193 typedef int ilp_var_t;
194 typedef int ilp_cst_t;
196 typedef struct _spill_bb_t {
201 typedef struct _remat_t {
202 const ir_node *op; /**< for copy_irn */
203 const ir_node *value; /**< the value which is being recomputed by this remat */
204 const ir_node *proj; /**< not NULL if the above op produces a tuple */
205 int cost; /**< cost of this remat */
206 int inverse; /**< nonzero if this is an inverse remat */
210 * Data to be attached to each IR node. For remats this contains the ilp_var
211 * for this remat and for normal ops this contains the ilp_vars for
212 * reloading each operand
214 typedef struct _op_t {
219 const remat_t *remat; /** the remat this op belongs to */
220 int pre; /** 1, if this is a pressure-increasing remat */
224 ir_node *op; /** the operation this live range belongs to */
233 typedef struct _defs_t {
234 const ir_node *value;
235 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
236 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
239 typedef struct _remat_info_t {
240 const ir_node *irn; /**< the irn to which these remats belong */
241 pset *remats; /**< possible remats for this value */
242 pset *remats_by_operand; /**< remats with this value as operand */
245 typedef struct _keyval_t {
250 typedef struct _spill_t {
259 typedef struct _memoperand_t {
260 ir_node *irn; /**< the irn */
261 unsigned int pos; /**< the position of the argument */
262 ilp_var_t ilp; /**< the ilp var for this memory operand */
266 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
268 return arch_irn_consider_in_reg_alloc(si->birg->main_env->arch_env,
274 cmp_remat(const void *a, const void *b)
276 const keyval_t *p = a;
277 const keyval_t *q = b;
278 const remat_t *r = p->val;
279 const remat_t *s = q->val;
283 return !(r == s || r->op == s->op);
287 cmp_remat(const void *a, const void *b)
289 const remat_t *r = a;
290 const remat_t *s = a;
292 return !(r == s || r->op == s->op);
296 cmp_spill(const void *a, const void *b, size_t size)
298 const spill_t *p = a;
299 const spill_t *q = b;
301 // return !(p->irn == q->irn && p->bb == q->bb);
302 return !(p->irn == q->irn);
306 cmp_memoperands(const void *a, const void *b, size_t size)
308 const memoperand_t *p = a;
309 const memoperand_t *q = b;
311 return !(p->irn == q->irn && p->pos == q->pos);
315 set_find_keyval(set * set, const void * key)
320 return set_find(set, &query, sizeof(query), HASH_PTR(key));
324 set_insert_keyval(set * set, void * key, void * val)
330 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
334 set_find_def(set * set, const ir_node * value)
339 return set_find(set, &query, sizeof(query), HASH_PTR(value));
343 set_insert_def(set * set, const ir_node * value)
350 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
353 static memoperand_t *
354 set_insert_memoperand(set * set, ir_node * irn, unsigned int pos, ilp_var_t ilp)
361 return set_insert(set, &query, sizeof(query), HASH_PTR(irn)+pos);
364 static memoperand_t *
365 set_find_memoperand(set * set, const ir_node * irn, unsigned int pos)
369 query.irn = (ir_node*)irn;
371 return set_find(set, &query, sizeof(query), HASH_PTR(irn)+pos);
376 set_find_spill(set * set, const ir_node * value)
380 query.irn = (ir_node*)value;
381 return set_find(set, &query, sizeof(query), HASH_PTR(value));
384 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
385 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
386 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
387 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
388 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
391 cmp_remat_info(const void *a, const void *b, size_t size)
393 const remat_info_t *p = a;
394 const remat_info_t *q = b;
396 return !(p->irn == q->irn);
400 cmp_defs(const void *a, const void *b, size_t size)
405 return !(p->value == q->value);
409 cmp_keyval(const void *a, const void *b, size_t size)
411 const keyval_t *p = a;
412 const keyval_t *q = b;
414 return !(p->key == q->key);
418 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
421 if(be_profile_has_data())
422 return ((double)be_profile_get_block_execcount(get_block(irn))) + FUDGE;
424 #ifndef EXECFREQ_LOOPDEPH
425 return get_block_execfreq(si->birg->exec_freq, get_block(irn)) + FUDGE;
428 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
430 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
435 get_cost(const spill_ilp_t * si, const ir_node * irn)
437 if(be_is_Spill(irn)) {
438 return opt_cost_spill;
439 } else if(be_is_Reload(irn)){
440 return opt_cost_reload;
442 return arch_get_op_estimated_cost(si->birg->main_env->arch_env, irn);
447 * Checks, whether node and its operands have suitable reg classes
450 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
453 const arch_env_t *arch_env = si->birg->main_env->arch_env;
454 int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
458 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
461 for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
462 ir_node *op = get_irn_n(irn, n);
463 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
466 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
473 * Try to create a remat from @p op with destination value @p dest_value
475 static INLINE remat_t *
476 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
478 remat_t *remat = NULL;
480 // if(!mode_is_datab(get_irn_mode(dest_value)))
483 if(dest_value == op) {
484 const ir_node *proj = NULL;
486 if(is_Proj(dest_value)) {
487 op = get_Proj_pred(op);
491 if(!is_rematerializable(si, op))
494 remat = obstack_alloc(si->obst, sizeof(*remat));
496 remat->cost = (int)get_cost(si, op);
497 remat->value = dest_value;
501 arch_inverse_t inverse;
504 /* get the index of the operand we want to retrieve by the inverse op */
505 for (n = get_irn_arity(op)-1; n>=0; --n) {
506 ir_node *arg = get_irn_n(op, n);
508 if(arg == dest_value) break;
512 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", n, op));
514 /* else ask the backend to give an inverse op */
515 if(arch_get_inverse(si->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
518 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
520 assert(inverse.n > 0 && "inverse op should have at least one node");
522 for(i=inverse.n-1; i>=0; --i) {
523 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
527 remat = obstack_alloc(si->obst, sizeof(*remat));
528 remat->op = inverse.nodes[0];
529 remat->cost = inverse.costs;
530 remat->value = dest_value;
531 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
534 // Matze: commented this out, this doesn't seem to be true if
535 // the inverse is a simple operation with only 1 result...
536 //assert(is_Proj(remat->proj));
538 assert(0 && "I can not handle remats with more than 2 nodes");
545 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
547 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
555 add_remat(const spill_ilp_t * si, const remat_t * remat)
557 remat_info_t *remat_info,
562 assert(remat->value);
564 query.irn = remat->value;
566 query.remats_by_operand = NULL;
567 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
569 if(remat_info->remats == NULL) {
570 remat_info->remats = new_pset(cmp_remat, 4096);
572 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
574 /* insert the remat into the remats_be_operand set of each argument of the remat op */
575 for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
576 ir_node *arg = get_irn_n(remat->op, n);
580 query.remats_by_operand = NULL;
581 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
583 if(remat_info->remats_by_operand == NULL) {
584 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
586 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
591 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
593 const ir_edge_t *edge = get_irn_out_edge_first(irn);
597 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
600 edge = get_irn_out_edge_next(irn, edge);
607 get_irn_n_nonignore_args(const spill_ilp_t * si, const ir_node * irn)
613 irn = get_Proj_pred(irn);
615 for(n=get_irn_arity(irn)-1; n>=0; --n) {
616 const ir_node *arg = get_irn_n(irn, n);
618 if(has_reg_class(si, arg)) ++ret;
625 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
630 if( has_reg_class(si, op)
631 && (opt_repair_schedule || get_irn_n_nonremat_edges(si, op) > 1)
632 && (opt_remats != REMATS_BRIGGS || get_irn_n_nonignore_args(si, op) == 0)
634 remat = get_remat_from_op(si, op, op);
636 add_remat(si, remat);
640 if(opt_remats == REMATS_ALL) {
641 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
643 for (n = get_irn_arity(op)-1; n>=0; --n) {
644 ir_node *arg = get_irn_n(op, n);
646 if(has_reg_class(si, arg)) {
647 /* try to get an inverse remat */
648 remat = get_remat_from_op(si, arg, op);
650 add_remat(si, remat);
658 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
661 ir_node *def_block = get_nodes_block(val);
667 /* if pos is at end of a basic block */
669 ret = (pos == def_block || block_dominates(def_block, pos));
670 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
674 /* else if this is a normal operation */
675 block = get_nodes_block(pos);
676 if(block == def_block) {
677 if(!sched_is_scheduled(val)) return 1;
679 ret = sched_comes_after(val, pos);
680 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
684 ret = block_dominates(def_block, block);
685 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
689 static INLINE ir_node *
690 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
692 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->birg->main_env->arch_env);
696 * Returns first non-Phi node of block @p bb
698 static INLINE ir_node *
699 sched_block_first_nonphi(const ir_node * bb)
701 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
705 sched_skip_proj_predicator(const ir_node * irn, void * data)
707 return (is_Proj(irn));
710 static INLINE ir_node *
711 sched_next_nonproj(const ir_node * irn, int forward)
713 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
717 * Returns next operation node (non-Proj) after @p irn
718 * or the basic block of this node
720 static INLINE ir_node *
721 sched_next_op(const ir_node * irn)
723 ir_node *next = sched_next(irn);
728 return sched_next_nonproj(next, 1);
732 * Returns previous operation node (non-Proj) before @p irn
733 * or the basic block of this node
735 static INLINE ir_node *
736 sched_prev_op(const ir_node * irn)
738 ir_node *prev = sched_prev(irn);
743 return sched_next_nonproj(prev, 0);
747 sched_put_after(ir_node * insert, ir_node * irn)
749 if(is_Block(insert)) {
750 insert = sched_block_first_nonphi(insert);
752 insert = sched_next_op(insert);
755 sched_add_before(insert, irn);
759 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
761 if(is_Block(insert)) {
762 insert = sched_block_last_noncf(si, insert);
764 insert = sched_next_nonproj(insert, 0);
765 insert = sched_prev(insert);
768 sched_add_after(insert, irn);
772 next_post_remat(const ir_node * irn)
778 next = sched_block_first_nonphi(irn);
780 next = sched_next_op(irn);
783 if(sched_is_end(next))
786 op = get_irn_link(next);
787 if(op->is_remat && !op->attr.remat.pre) {
796 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
802 ret = sched_block_last_noncf(si, irn);
803 ret = sched_next(ret);
804 ret = sched_prev_op(ret);
806 ret = sched_prev_op(irn);
809 if(sched_is_end(ret) || is_Phi(ret))
812 op = (op_t*)get_irn_link(ret);
813 if(op->is_remat && op->attr.remat.pre) {
821 * Tells you whether a @p remat can be placed before the irn @p pos
824 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
826 const ir_node *op = remat->op;
832 prev = sched_block_last_noncf(si, pos);
833 prev = sched_next_nonproj(prev, 0);
835 prev = sched_prev_op(pos);
837 /* do not remat if the rematted value is defined immediately before this op */
838 if(prev == remat->op) {
843 /* this should be just fine, the following OP will be using this value, right? */
845 /* only remat AFTER the real definition of a value (?) */
846 if(!value_is_defined_before(si, pos, remat->value)) {
847 // ir_fprintf(stderr, "error(not defined)");
852 for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
853 const ir_node *arg = get_irn_n(op, n);
855 if(opt_no_enlarge_liveness) {
856 if(has_reg_class(si, arg) && live) {
857 res &= pset_find_ptr((pset*)live, arg)?1:0;
859 res &= value_is_defined_before(si, pos, arg);
862 res &= value_is_defined_before(si, pos, arg);
870 * Tells you whether a @p remat can be placed after the irn @p pos
873 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
876 pos = sched_block_first_nonphi(pos);
878 pos = sched_next_op(pos);
881 /* only remat AFTER the real definition of a value (?) */
882 if(!value_is_defined_before(si, pos, remat->value)) {
886 return can_remat_before(si, remat, pos, live);
890 * Collect potetially rematerializable OPs
893 walker_remat_collector(ir_node * irn, void * data)
895 spill_ilp_t *si = data;
897 if(!is_Block(irn) && !is_Phi(irn)) {
898 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
899 get_remats_from_op(si, irn);
904 * Inserts a copy of @p irn before @p pos
907 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
912 bb = is_Block(pos)?pos:get_nodes_block(pos);
913 copy = exact_copy(irn);
915 set_phi_class(si->pc, copy, NULL);
916 set_nodes_block(copy, bb);
917 sched_put_before(si, pos, copy);
923 * Inserts a copy of @p irn after @p pos
926 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
931 bb = is_Block(pos)?pos:get_nodes_block(pos);
932 copy = exact_copy(irn);
934 set_phi_class(si->pc, copy, NULL);
935 set_nodes_block(copy, bb);
936 sched_put_after(pos, copy);
942 insert_remat_after(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
946 if(can_remat_after(si, remat, pos, live)) {
951 DBG((si->dbg, LEVEL_3, "\t >inserting remat2 %+F\n", remat->op));
953 copy = insert_copy_after(si, remat->op, pos);
955 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
956 op = obstack_alloc(si->obst, sizeof(*op));
958 op->attr.remat.remat = remat;
959 op->attr.remat.pre = 0;
960 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
962 set_irn_link(copy, op);
963 pset_insert_ptr(si->all_possible_remats, copy);
965 proj_copy = insert_copy_after(si, remat->proj, copy);
966 set_irn_n(proj_copy, 0, copy);
967 set_irn_link(proj_copy, op);
968 pset_insert_ptr(si->all_possible_remats, proj_copy);
980 insert_remat_before(spill_ilp_t * si, const remat_t * remat, ir_node * pos, const pset * live)
984 if(can_remat_before(si, remat, pos, live)) {
989 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
991 copy = insert_copy_before(si, remat->op, pos);
993 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
994 op = obstack_alloc(si->obst, sizeof(*op));
996 op->attr.remat.remat = remat;
997 op->attr.remat.pre = 1;
998 op->attr.remat.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos), 0.0);
1000 set_irn_link(copy, op);
1001 pset_insert_ptr(si->all_possible_remats, copy);
1003 proj_copy = insert_copy_after(si, remat->proj, copy);
1004 set_irn_n(proj_copy, 0, copy);
1005 set_irn_link(proj_copy, op);
1006 pset_insert_ptr(si->all_possible_remats, proj_copy);
1018 get_block_n_succs(const ir_node *block) {
1019 const ir_edge_t *edge;
1021 assert(edges_activated(current_ir_graph));
1023 edge = get_block_succ_first(block);
1027 edge = get_block_succ_next(block, edge);
1028 return edge ? 2 : 1;
1032 is_start_block(const ir_node * bb)
1034 return get_irg_start_block(get_irn_irg(bb)) == bb;
1038 is_merge_edge(const ir_node * bb)
1040 if(is_start_block(bb))
1044 return get_block_n_succs(bb) == 1;
1050 is_diverge_edge(const ir_node * bb)
1052 if(is_start_block(bb))
1056 return get_Block_n_cfgpreds(bb) == 1;
1062 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1067 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1068 irn = be_lv_get_irn(si->lv, bb, i);
1070 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1071 pset_insert_ptr(live, irn);
1075 irn = sched_last(bb);
1077 /* all values eaten by control flow operations are also live until the end of the block */
1078 sched_foreach_reverse(bb, irn) {
1081 if(!sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) break;
1083 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1084 ir_node *arg = get_irn_n(irn,i);
1086 if(has_reg_class(si, arg)) {
1087 pset_insert_ptr(live, arg);
1092 * find values that are used by remats at end of block
1093 * and insert them into live set
1095 foreach_pre_remat(si, bb, irn) {
1098 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1099 ir_node *remat_arg = get_irn_n(irn, n);
1101 if(!has_reg_class(si, remat_arg)) continue;
1103 /* if value is becoming live through use by remat */
1104 if(!pset_find_ptr(live, remat_arg)) {
1105 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, irn));
1107 pset_insert_ptr(live, remat_arg);
1114 walker_regclass_copy_insertor(ir_node * irn, void * data)
1116 spill_ilp_t *si = data;
1118 if(is_Phi(irn) && has_reg_class(si, irn)) {
1121 for(n=get_irn_arity(irn)-1; n>=0; --n) {
1122 ir_node *phi_arg = get_irn_n(irn, n);
1123 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
1125 if(!has_reg_class(si, phi_arg)) {
1126 ir_node *copy = be_new_Copy(si->cls, si->birg->irg, bb, phi_arg);
1127 ir_node *pos = sched_block_last_noncf(si, bb);
1128 op_t *op = obstack_alloc(si->obst, sizeof(*op));
1130 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
1131 sched_add_after(pos, copy);
1132 set_irn_n(irn, n, copy);
1135 op->attr.live_range.args.reloads = NULL;
1136 op->attr.live_range.ilp = ILP_UNDEF;
1137 set_irn_link(copy, op);
1144 * Insert (so far unused) remats into the irg to
1145 * recompute the potential liveness of all values
1148 walker_remat_insertor(ir_node * bb, void * data)
1150 spill_ilp_t *si = data;
1157 /* skip start block, no remats to do there */
1158 if(is_start_block(bb)) return;
1160 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
1162 live = pset_new_ptr_default();
1163 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1164 ir_node *value = be_lv_get_irn(si->lv, bb, i);
1166 /* add remats at end of block */
1167 if (has_reg_class(si, value)) {
1168 pset_insert_ptr(live, value);
1172 irn = sched_last(bb);
1173 while(!sched_is_end(irn)) {
1179 next = sched_prev(irn);
1181 /* delete defined value from live set */
1182 if(has_reg_class(si, irn)) {
1183 pset_remove_ptr(live, irn);
1186 if(is_Phi(irn) || is_Proj(irn)) {
1191 args = pset_new_ptr_default();
1192 used = pset_new_ptr_default();
1194 /* collect arguments of op and set args of op already live in epilog */
1195 for (n = get_irn_arity(irn)-1; n>=0; --n) {
1196 ir_node *arg = get_irn_n(irn, n);
1198 pset_insert_ptr(args, arg);
1199 if(has_reg_class(si, arg)) {
1200 pset_insert_ptr(live, arg);
1201 pset_insert_ptr(used, arg);
1205 /* insert all possible remats before irn */
1206 pset_foreach(args, arg) {
1207 remat_info_t *remat_info,
1210 /* continue if the operand has the wrong reg class */
1211 if(!has_reg_class(si, arg))
1215 query.remats = NULL;
1216 query.remats_by_operand = NULL;
1217 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1223 if(remat_info->remats) {
1224 pset_foreach(remat_info->remats, remat) {
1225 ir_node *remat_irn = NULL;
1227 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
1228 remat_irn = insert_remat_before(si, remat, irn, live);
1231 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1232 ir_node *remat_arg = get_irn_n(remat_irn, n);
1234 /* collect args of remats which are not args of op */
1235 if(has_reg_class(si, remat_arg) && !pset_find_ptr(args, remat_arg)) {
1236 pset_insert_ptr(used, remat_arg);
1244 /* do not place post remats after jumps */
1245 if(sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) {
1251 /* insert all possible remats after irn */
1252 post_remats = pset_new_ptr_default();
1253 pset_foreach(used, arg) {
1254 remat_info_t *remat_info,
1257 /* continue if the operand has the wrong reg class */
1258 if(!has_reg_class(si, arg))
1262 query.remats = NULL;
1263 query.remats_by_operand = NULL;
1264 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1270 if(remat_info->remats_by_operand) {
1271 pset_foreach(remat_info->remats_by_operand, remat) {
1272 /* do not insert remats producing the same value as one of the operands */
1273 if(!pset_find_ptr(args, remat->value)) {
1274 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
1276 /* only remat values that can be used by real ops */
1277 if(!opt_remat_while_live || pset_find_ptr(live, remat->value)) {
1278 pset_insert_ptr(post_remats, remat);
1284 pset_foreach(post_remats, remat) {
1285 insert_remat_after(si, remat, irn, live);
1287 del_pset(post_remats);
1294 /* add remats at end if successor has multiple predecessors */
1295 if(is_merge_edge(bb)) {
1296 pset *live_out = pset_new_ptr_default();
1299 get_live_end(si, bb, live_out);
1301 /* add remats at end of block */
1302 pset_foreach(live_out, value) {
1303 remat_info_t *remat_info,
1307 query.remats = NULL;
1308 query.remats_by_operand = NULL;
1309 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1311 if(remat_info && remat_info->remats) {
1312 pset_foreach(remat_info->remats, remat) {
1313 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1315 insert_remat_before(si, remat, bb, live_out);
1322 if(is_diverge_edge(bb)) {
1323 pset *live_in = pset_new_ptr_default();
1326 be_lv_foreach(si->lv, bb, be_lv_state_in, i) {
1327 value = be_lv_get_irn(si->lv, bb, i);
1329 if(has_reg_class(si, value)) {
1330 pset_insert_ptr(live_in, value);
1333 /* add phis to live_in */
1334 sched_foreach(bb, value) {
1335 if(!is_Phi(value)) break;
1337 if(has_reg_class(si, value)) {
1338 pset_insert_ptr(live_in, value);
1342 /* add remat2s at beginning of block */
1343 post_remats = pset_new_ptr_default();
1344 pset_foreach(live_in, value) {
1345 remat_info_t *remat_info,
1349 query.remats = NULL;
1350 query.remats_by_operand = NULL;
1351 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1353 if(remat_info && remat_info->remats_by_operand) {
1354 pset_foreach(remat_info->remats_by_operand, remat) {
1355 DBG((si->dbg, LEVEL_4, "\t considering remat2 %+F at beginning of block %+F\n", remat->op, bb));
1357 /* put the remat here if all its args are available and result is still live */
1358 if(!opt_remat_while_live || pset_find_ptr(live_in, remat->value)) {
1359 pset_insert_ptr(post_remats, remat);
1364 pset_foreach(post_remats, remat) {
1365 insert_remat_after(si, remat, bb, live_in);
1367 del_pset(post_remats);
1373 can_be_copied(const ir_node * bb, const ir_node * irn)
1375 const ir_edge_t *edge = get_block_succ_first(bb);
1376 const ir_node *next_bb = edge->src;
1377 int pos = edge->pos;
1380 assert(is_merge_edge(bb));
1382 sched_foreach(next_bb, phi) {
1383 const ir_node *phi_arg;
1385 if(!is_Phi(phi)) break;
1387 phi_arg = get_irn_n(phi, pos);
1389 if(phi_arg == irn) {
1397 * Initialize additional node info
1400 luke_initializer(ir_node * bb, void * data)
1402 spill_ilp_t *si = (spill_ilp_t*)data;
1403 spill_bb_t *spill_bb;
1406 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
1407 set_irn_link(bb, spill_bb);
1409 sched_foreach(bb, irn) {
1412 op = obstack_alloc(si->obst, sizeof(*op));
1414 op->attr.live_range.ilp = ILP_UNDEF;
1417 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1418 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(irn));
1420 } else if(!is_Proj(irn)) {
1421 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1422 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
1424 op->attr.live_range.args.reloads = NULL;
1426 set_irn_link(irn, op);
1432 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1435 luke_endwalker(ir_node * bb, void * data)
1437 spill_ilp_t *si = (spill_ilp_t*)data;
1443 spill_bb_t *spill_bb = get_irn_link(bb);
1446 live = pset_new_ptr_default();
1447 use_end = pset_new_ptr_default();
1449 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
1450 irn = be_lv_get_irn(si->lv, bb, i);
1451 if (has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1452 pset_insert_ptr(live, irn);
1456 * find values that are used by remats at end of block
1457 * and insert them into live set
1459 foreach_pre_remat(si, bb, irn) {
1462 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1463 ir_node *remat_arg = get_irn_n(irn, n);
1465 if(has_reg_class(si, remat_arg)) {
1466 pset_insert_ptr(live, remat_arg);
1471 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1472 /* their reg_out must always be set */
1473 sched_foreach_reverse(bb, irn) {
1476 if(!sched_skip_cf_predicator(irn, si->birg->main_env->arch_env)) break;
1478 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1479 ir_node *irn_arg = get_irn_n(irn, n);
1481 if(has_reg_class(si, irn_arg)) {
1482 pset_insert_ptr(use_end, irn_arg);
1487 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1488 //cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1489 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1491 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1493 /* if this is a merge edge we can reload at the end of this block */
1494 if(is_merge_edge(bb)) {
1495 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1496 } else if(pset_count(use_end)){
1497 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1499 spill_bb->reloads = NULL;
1502 pset_foreach(live,irn) {
1506 int default_spilled;
1509 /* handle values used by control flow nodes later separately */
1510 if(pset_find_ptr(use_end, irn)) continue;
1513 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1515 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1517 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1518 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1519 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1521 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1522 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1524 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1525 /* by default spill value right after definition */
1526 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1527 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1529 if(is_merge_edge(bb)) {
1533 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1534 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), can_be_copied(bb, irn));
1535 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1537 /* reload <= mem_out */
1538 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1539 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1540 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1543 spill->reg_in = ILP_UNDEF;
1544 spill->mem_in = ILP_UNDEF;
1547 pset_foreach(use_end,irn) {
1551 ilp_cst_t end_use_req,
1554 int default_spilled;
1557 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1559 spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1561 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1562 spill->reg_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1564 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1565 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1567 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1568 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1569 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1571 /* reload for use be control flow op */
1572 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1573 reload = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_reload*execution_frequency(si, bb), 1.0);
1574 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1576 /* reload <= mem_out */
1577 rel_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1578 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1579 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1581 spill->reg_in = ILP_UNDEF;
1582 spill->mem_in = ILP_UNDEF;
1584 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1585 end_use_req = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1);
1586 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1595 * Find a remat of value @p value in the epilog of @p pos
1598 find_post_remat(const ir_node * value, const ir_node * pos)
1600 while((pos = next_post_remat(pos)) != NULL) {
1603 op = get_irn_link(pos);
1604 assert(op->is_remat && !op->attr.remat.pre);
1606 if(op->attr.remat.remat->value == value)
1607 return (ir_node*)pos;
1610 const ir_edge_t *edge;
1611 foreach_out_edge(pos, edge) {
1612 ir_node *proj = get_edge_src_irn(edge);
1613 assert(is_Proj(proj));
1624 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1626 spill_bb_t *spill_bb = get_irn_link(bb);
1630 int default_spilled;
1633 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1635 double spill_cost = is_Unknown(irn)?0.0001:opt_cost_spill*execution_frequency(si, bb);
1637 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1639 spill->reg_out = ILP_UNDEF;
1640 spill->reg_in = ILP_UNDEF;
1641 spill->mem_in = ILP_UNDEF;
1643 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1644 spill->mem_out = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1646 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1647 default_spilled = be_is_live_in(si->lv, bb, irn) || is_Phi(irn);
1648 spill->spill = lpp_add_var_default(si->lpp, buf, lpp_binary, spill_cost, !default_spilled);
1655 * Inserts ILP-constraints and variables for memory copying before the given position
1658 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1660 const ir_node *succ;
1661 const ir_edge_t *edge;
1662 spill_bb_t *spill_bb = get_irn_link(block);
1671 assert(edges_activated(current_ir_graph));
1673 edge = get_block_succ_first(block);
1679 edge = get_block_succ_next(block, edge);
1680 /* next block can only contain phis, if this is a merge edge */
1683 ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1684 copyreg = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
1686 ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1687 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
1689 pset_foreach(live, tmp) {
1692 op_t *op = get_irn_link(irn);
1693 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1695 spill = set_find_spill(spill_bb->ilp, tmp);
1698 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1700 lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1702 sched_foreach(succ, phi) {
1703 const ir_node *to_copy;
1705 spill_t *to_copy_spill;
1706 op_t *phi_op = get_irn_link(phi);
1707 ilp_var_t reload = ILP_UNDEF;
1710 if(!is_Phi(phi)) break;
1711 if(!has_reg_class(si, phi)) continue;
1713 to_copy = get_irn_n(phi, pos);
1714 to_copy_op = get_irn_link(to_copy);
1716 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1717 assert(to_copy_spill);
1719 if(spill_bb->reloads) {
1720 keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1723 reload = PTR_TO_INT(keyval->val);
1727 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1728 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1730 /* copy - reg_out - reload - remat - live_range <= 0 */
1731 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1732 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1733 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1734 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1735 foreach_pre_remat(si, block, tmp) {
1736 op_t *remat_op = get_irn_link(tmp);
1737 if(remat_op->attr.remat.remat->value == to_copy) {
1738 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1742 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1743 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1745 /* copy - reg_out - copyreg <= 0 */
1746 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1747 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1748 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1754 * Walk all irg blocks and emit this ILP
1757 luke_blockwalker(ir_node * bb, void * data)
1759 spill_ilp_t *si = (spill_ilp_t*)data;
1764 spill_bb_t *spill_bb = get_irn_link(bb);
1767 pset *defs = pset_new_ptr_default();
1768 const arch_env_t *arch_env = si->birg->main_env->arch_env;
1770 live = pset_new_ptr_default();
1772 /****************************************
1773 * B A S I C B L O C K E N D
1774 ***************************************/
1777 /* init live values at end of block */
1778 get_live_end(si, bb, live);
1780 pset_foreach(live, irn) {
1782 ilp_var_t reload = ILP_UNDEF;
1784 spill = set_find_spill(spill_bb->ilp, irn);
1787 if(spill_bb->reloads) {
1788 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1791 reload = PTR_TO_INT(keyval->val);
1795 op = get_irn_link(irn);
1796 assert(!op->is_remat);
1798 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1799 op->attr.live_range.ilp = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
1800 op->attr.live_range.op = bb;
1802 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1803 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
1805 /* reg_out - reload - remat - live_range <= 0 */
1806 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1807 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1808 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1809 foreach_pre_remat(si, bb, tmp) {
1810 op_t *remat_op = get_irn_link(tmp);
1811 if(remat_op->attr.remat.remat->value == irn) {
1812 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1815 ir_snprintf(buf, sizeof(buf), "reg_out2_%N_%N", bb, irn);
1816 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
1818 /* value may only die at bb end if it is used for a mem copy */
1819 /* reg_out + \sum copy - reload - remat - live_range >= 0 */
1820 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1821 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1822 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1823 foreach_pre_remat(si, bb, tmp) {
1824 op_t *remat_op = get_irn_link(tmp);
1825 if(remat_op->attr.remat.remat->value == irn) {
1826 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1829 if(is_merge_edge(bb)) {
1830 const ir_edge_t *edge = get_block_succ_first(bb);
1831 const ir_node *next_bb = edge->src;
1832 int pos = edge->pos;
1835 sched_foreach(next_bb, phi) {
1836 const ir_node *phi_arg;
1838 if(!is_Phi(phi)) break;
1840 phi_arg = get_irn_n(phi, pos);
1842 if(phi_arg == irn) {
1843 op_t *phi_op = get_irn_link(phi);
1844 ilp_var_t copy = phi_op->attr.live_range.args.copies[pos];
1846 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
1853 insert_mem_copy_position(si, live, bb);
1856 * assure the remat args are available
1858 foreach_pre_remat(si, bb, tmp) {
1859 op_t *remat_op = get_irn_link(tmp);
1862 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1863 ir_node *remat_arg = get_irn_n(tmp, n);
1864 op_t *arg_op = get_irn_link(remat_arg);
1866 if(!has_reg_class(si, remat_arg)) continue;
1868 spill = set_find_spill(spill_bb->ilp, remat_arg);
1871 /* arguments of remats have to be live until the very end of the block
1872 * remat = reg_out(remat_arg) and (reload(remat_arg) or live_range(remat_arg)),
1873 * no remats, they could be in wrong order
1876 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1877 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1879 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 3.0);
1880 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, -2.0);
1881 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1883 /* use reload placed for this argument */
1884 if(spill_bb->reloads) {
1885 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1888 ilp_var_t reload = PTR_TO_INT(keyval->val);
1890 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1895 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1900 /**************************************
1901 * B A S I C B L O C K B O D Y
1902 **************************************/
1904 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1910 ilp_cst_t check_pre,
1916 ilp_cst_t one_memoperand = -1;
1918 /* iterate only until first phi */
1922 op = get_irn_link(irn);
1924 if(op->is_remat) continue;
1926 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1928 /* collect defined values */
1929 if(has_reg_class(si, irn)) {
1930 pset_insert_ptr(defs, irn);
1934 if(is_Proj(irn)) continue;
1937 * init set of irn's arguments
1938 * and all possibly used values around this op
1939 * and values defined by post remats
1941 args = new_set(cmp_keyval, get_irn_arity(irn));
1942 used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1943 remat_defs = pset_new_ptr(pset_count(live));
1945 if(!is_start_block(bb) || !be_is_Barrier(irn)) {
1946 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1947 ir_node *irn_arg = get_irn_n(irn, n);
1948 if(has_reg_class(si, irn_arg)) {
1949 set_insert_keyval(args, irn_arg, (void*)n);
1950 pset_insert_ptr(used, irn_arg);
1953 foreach_post_remat(irn, tmp) {
1954 op_t *remat_op = get_irn_link(tmp);
1956 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1958 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1959 ir_node *remat_arg = get_irn_n(tmp, n);
1960 if(has_reg_class(si, remat_arg)) {
1961 pset_insert_ptr(used, remat_arg);
1965 foreach_pre_remat(si, irn, tmp) {
1966 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1967 ir_node *remat_arg = get_irn_n(tmp, n);
1968 if(has_reg_class(si, remat_arg)) {
1969 pset_insert_ptr(used, remat_arg);
1975 /**********************************
1976 * I N E P I L O G O F irn
1977 **********************************/
1979 /* ensure each dying value is used by only one post remat */
1980 pset_foreach(used, tmp) {
1981 ir_node *value = tmp;
1982 op_t *value_op = get_irn_link(value);
1987 foreach_post_remat(irn, remat) {
1988 op_t *remat_op = get_irn_link(remat);
1990 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1991 ir_node *remat_arg = get_irn_n(remat, n);
1993 /* if value is used by this remat add it to constraint */
1994 if(remat_arg == value) {
1996 /* sum remat2s <= 1 + n_remats*live_range */
1997 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1998 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2002 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2008 if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
2009 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
2013 /* ensure at least one value dies at post remat */
2014 foreach_post_remat(irn, tmp) {
2015 op_t *remat_op = get_irn_link(tmp);
2016 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
2019 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2020 remat_arg = get_irn_n(tmp, n);
2022 if(has_reg_class(si, remat_arg)) {
2024 /* does arg always die at this op? */
2025 if(!pset_find_ptr(live, remat_arg))
2026 goto skip_one_must_die;
2028 pset_insert_ptr(remat_args, remat_arg);
2032 /* remat + \sum live_range(remat_arg) <= |args| */
2033 ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
2034 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2035 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2037 pset_foreach(remat_args, remat_arg) {
2038 op_t *arg_op = get_irn_link(remat_arg);
2040 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2044 del_pset(remat_args);
2047 /* new live ranges for values from L\U defined by post remats */
2048 pset_foreach(live, tmp) {
2049 ir_node *value = tmp;
2050 op_t *value_op = get_irn_link(value);
2052 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
2053 ilp_var_t prev_lr = ILP_UNDEF;
2056 if(pset_find_ptr(remat_defs, value)) {
2058 /* next_live_range <= prev_live_range + sum remat2s */
2059 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
2060 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2062 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
2063 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2065 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
2066 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2068 foreach_post_remat(irn, remat) {
2069 op_t *remat_op = get_irn_link(remat);
2071 /* if value is being rematerialized by this remat */
2072 if(value == remat_op->attr.remat.remat->value) {
2073 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2077 value_op->attr.live_range.ilp = prev_lr;
2078 value_op->attr.live_range.op = irn;
2083 /* requirements for post remats and start live ranges from L/U' for values dying here */
2084 foreach_post_remat(irn, tmp) {
2085 op_t *remat_op = get_irn_link(tmp);
2088 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2089 ir_node *remat_arg = get_irn_n(tmp, n);
2090 op_t *arg_op = get_irn_link(remat_arg);
2092 if(!has_reg_class(si, remat_arg)) continue;
2094 /* only for values in L\U (TODO and D?), the others are handled with post_use */
2095 if(!pset_find_ptr(used, remat_arg)) {
2096 /* remat <= live_range(remat_arg) */
2097 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
2098 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2100 /* if value is becoming live through use by remat2 */
2101 if(!pset_find_ptr(live, remat_arg)) {
2104 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
2105 lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2107 arg_op->attr.live_range.ilp = lr;
2108 arg_op->attr.live_range.op = irn;
2110 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
2112 pset_insert_ptr(live, remat_arg);
2113 add_to_spill_bb(si, bb, remat_arg);
2116 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2117 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2122 d = pset_count(defs);
2123 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
2125 /* count how many regs irn needs for arguments */
2126 u = set_count(args);
2129 /* check the register pressure in the epilog */
2130 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
2131 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
2132 check_post = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - d);
2134 /* add L\U' to check_post */
2135 pset_foreach(live, tmp) {
2136 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
2137 /* if a live value is not used by irn */
2138 tmp_op = get_irn_link(tmp);
2139 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
2143 /***********************************************************
2144 * 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
2145 **********************************************************/
2148 pset_foreach(used, tmp) {
2154 op_t *arg_op = get_irn_link(arg);
2157 spill = add_to_spill_bb(si, bb, arg);
2159 /* new live range for each used value */
2160 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
2161 prev_lr = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2163 /* the epilog stuff - including post_use, check_post, check_post_remat */
2164 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
2165 post_use = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2167 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
2169 /* arg is live throughout epilog if the next live_range is in a register */
2170 if(pset_find_ptr(live, arg)) {
2171 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
2173 /* post_use >= next_lr + remat */
2174 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2175 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2176 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2177 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2180 /* forall post remat which use arg add a similar cst */
2181 foreach_post_remat(irn, remat) {
2184 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2185 ir_node *remat_arg = get_irn_n(remat, n);
2186 op_t *remat_op = get_irn_link(remat);
2188 if(remat_arg == arg) {
2189 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
2191 /* post_use >= remat */
2192 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
2193 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2194 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
2195 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2200 /* if value is not an arg of op and not possibly defined by post remat
2201 * then it may only die and not become live
2203 if(!set_find_keyval(args, arg)) {
2204 /* post_use <= prev_lr */
2205 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
2206 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2207 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2208 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2210 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
2211 /* next_lr <= prev_lr */
2212 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
2213 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2214 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2215 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
2219 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2220 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2221 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2222 ilp_var_t memoperand;
2224 ir_snprintf(buf, sizeof(buf), "memoperand_%N_%d", irn, n);
2225 memoperand = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_memoperand*execution_frequency(si, bb), 0.0);
2226 set_insert_memoperand(si->memoperands, irn, n, memoperand);
2228 ir_snprintf(buf, sizeof(buf), "nolivepost_%N_%d", irn, n);
2229 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2231 lpp_set_factor_fast(si->lpp, cst, memoperand, 1.0);
2232 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
2237 /* new live range begins for each used value */
2238 arg_op->attr.live_range.ilp = prev_lr;
2239 arg_op->attr.live_range.op = irn;
2241 pset_insert_ptr(live, arg);
2244 /* just to be sure */
2245 check_post = ILP_UNDEF;
2247 /* allow original defintions to be removed */
2248 if(opt_repair_schedule) {
2249 pset_foreach(defs, tmp) {
2250 op_t *tmp_op = get_irn_link(tmp);
2251 spill_t *spill = set_find_spill(spill_bb->ilp, tmp);
2256 ir_snprintf(buf, sizeof(buf), "delete_%N", tmp);
2257 delete = lpp_add_var_default(si->lpp, buf, lpp_binary, -1.0*get_cost(si, irn)*execution_frequency(si, bb), 0.0);
2259 /* op may not be killed if its first live_range is 1 */
2260 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2261 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2262 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2263 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2265 /* op may not be killed if it is spilled after the definition */
2266 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2267 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2268 lpp_set_factor_fast(si->lpp, cst, delete, 1.0);
2269 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2274 ir_snprintf(buf, sizeof(buf), "keep_%N", tmp);
2275 keep = lpp_add_var_default(si->lpp, buf, lpp_binary, get_cost(si, irn)*execution_frequency(si, bb), 1.0);
2277 /* op may not be killed if its first live_range is 1 */
2278 ir_snprintf(buf, sizeof(buf), "killorig-lr_%N", tmp);
2279 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2280 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2281 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, -1.0);
2283 /* op may not be killed if it is spilled after the definition */
2284 ir_snprintf(buf, sizeof(buf), "killorig-spill_%N", tmp);
2285 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 0.0);
2286 lpp_set_factor_fast(si->lpp, cst, keep, 1.0);
2287 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2292 pset_foreach(defs, tmp) {
2293 op_t *tmp_op = get_irn_link(tmp);
2294 spill_t *spill = set_find_spill(spill_bb->ilp, tmp);
2297 /* live_range or spill should be 1
2298 TODO: lr should be live until first use */
2299 ir_snprintf(buf, sizeof(buf), "nokillorig_%N", tmp);
2300 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_greater, 1.0);
2301 lpp_set_factor_fast(si->lpp, cst, tmp_op->attr.live_range.ilp, 1.0);
2302 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2312 /* check the register pressure in the prolog */
2313 /* sum_{L\U} lr <= k - |U| */
2314 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
2315 check_pre = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs - u);
2317 /* for the prolog remove defined values from the live set */
2318 pset_foreach(defs, tmp) {
2319 pset_remove_ptr(live, tmp);
2322 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2323 ir_snprintf(buf, sizeof(buf), "one_memoperand_%N", irn);
2324 one_memoperand = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2327 /***********************************************************
2328 * 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
2329 **********************************************************/
2332 set_foreach(args, keyval) {
2334 const ir_node *arg = keyval->key;
2335 int i = PTR_TO_INT(keyval->val);
2336 op_t *arg_op = get_irn_link(arg);
2337 ilp_cst_t requirements;
2340 spill = set_find_spill(spill_bb->ilp, arg);
2343 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
2344 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);
2346 /* reload <= mem_out */
2347 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
2348 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2349 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2350 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2352 /* requirement: arg must be in register for use */
2353 /* reload + remat + live_range == 1 */
2354 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2355 requirements = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 1.0);
2357 lpp_set_factor_fast(si->lpp, requirements, arg_op->attr.live_range.ilp, 1.0);
2358 lpp_set_factor_fast(si->lpp, requirements, op->attr.live_range.args.reloads[i], 1.0);
2359 foreach_pre_remat(si, irn, tmp) {
2360 op_t *remat_op = get_irn_link(tmp);
2361 if(remat_op->attr.remat.remat->value == arg) {
2362 lpp_set_factor_fast(si->lpp, requirements, remat_op->attr.remat.ilp, 1.0);
2366 if(opt_memoperands && (!is_start_block(bb) || be_is_Barrier(irn))) {
2368 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2369 if(get_irn_n(irn, n) == arg) {
2373 for(n = get_irn_arity(irn)-1; n>=0; --n) {
2374 if(get_irn_n(irn, n) == arg && arch_possible_memory_operand(arch_env, irn, n)) {
2375 memoperand_t *memoperand;
2376 memoperand = set_find_memoperand(si->memoperands, irn, n);
2378 /* memoperand <= mem_out */
2379 ir_snprintf(buf, sizeof(buf), "req_memoperand_%N_%d", irn, n);
2380 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2381 lpp_set_factor_fast(si->lpp, cst, memoperand->ilp, 1.0);
2382 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2384 /* the memoperand is only sufficient if it is used once by the op */
2385 if(n_memoperands == 1)
2386 lpp_set_factor_fast(si->lpp, requirements, memoperand->ilp, 1.0);
2388 lpp_set_factor_fast(si->lpp, one_memoperand, memoperand->ilp, 1.0);
2390 /* we have one more free register if we use a memory operand */
2391 lpp_set_factor_fast(si->lpp, check_pre, memoperand->ilp, -1.0);
2397 /* iterate over L\U */
2398 pset_foreach(live, tmp) {
2399 if(!set_find_keyval(args, tmp)) {
2400 /* if a live value is not used by irn */
2401 tmp_op = get_irn_link(tmp);
2402 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2406 /* requirements for remats */
2407 foreach_pre_remat(si, irn, tmp) {
2408 op_t *remat_op = get_irn_link(tmp);
2411 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2412 ir_node *remat_arg = get_irn_n(tmp, n);
2413 op_t *arg_op = get_irn_link(remat_arg);
2415 if(!has_reg_class(si, remat_arg)) continue;
2417 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2418 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2419 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2421 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2422 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2424 /* if remat arg is also used by current op then we can use reload placed for this argument */
2425 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2426 int index = (int)keyval->val;
2428 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2436 /*************************
2437 * D O N E W I T H O P
2438 *************************/
2440 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2442 pset_foreach(live, tmp) {
2443 assert(has_reg_class(si, tmp));
2447 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2448 ir_node *arg = get_irn_n(irn, n);
2450 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2454 del_pset(remat_defs);
2458 defs = pset_new_ptr_default();
2460 /* skip everything above barrier in start block */
2461 if(is_start_block(bb) && be_is_Barrier(irn)) {
2462 assert(pset_count(live) == 0);
2471 /***************************************
2472 * B E G I N N I N G O F B L O C K
2473 ***************************************/
2476 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2477 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2479 pset_foreach(live, irn) {
2480 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2483 /* construct mem_outs for all values */
2484 set_foreach(spill_bb->ilp, spill) {
2485 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2486 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2488 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2489 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2491 if(pset_find_ptr(live, spill->irn)) {
2492 int default_spilled;
2493 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2495 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2496 default_spilled = be_is_live_in(si->lv, bb, spill->irn) || is_Phi(spill->irn);
2497 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, default_spilled);
2498 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2500 if(opt_memcopies && is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2502 op_t *op = get_irn_link(spill->irn);
2504 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2505 const ir_node *arg = get_irn_n(spill->irn, n);
2511 /* argument already done? */
2512 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2514 /* get sum of execution frequencies of blocks with the same phi argument */
2515 for(m=n; m>=0; --m) {
2516 const ir_node *arg2 = get_irn_n(spill->irn, m);
2519 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2523 /* copies are not for free */
2524 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2525 var = lpp_add_var_default(si->lpp, buf, lpp_binary, opt_cost_spill * freq, 1.0);
2527 for(m=n; m>=0; --m) {
2528 const ir_node *arg2 = get_irn_n(spill->irn, m);
2531 op->attr.live_range.args.copies[m] = var;
2536 /* copy <= mem_in */
2537 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2538 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2539 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2540 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2547 foreach_post_remat(bb, tmp) {
2549 op_t *remat_op = get_irn_link(tmp);
2550 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
2553 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2554 remat_arg = get_irn_n(tmp, n);
2556 if(has_reg_class(si, remat_arg)) {
2557 pset_insert_ptr(remat_args, remat_arg);
2561 /* remat + \sum live_range(remat_arg) <= |args| */
2562 ir_snprintf(buf, sizeof(buf), "one_must_die_%N", tmp);
2563 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, pset_count(remat_args));
2564 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2566 pset_foreach(remat_args, remat_arg) {
2567 if(pset_find_ptr(live, remat_arg)) {
2568 op_t *remat_arg_op = get_irn_link(remat_arg);
2569 lpp_set_factor_fast(si->lpp, cst, remat_arg_op->attr.live_range.ilp, 1.0);
2572 del_pset(remat_args);
2575 foreach_post_remat(bb, tmp) {
2578 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
2579 ir_node *remat_arg = get_irn_n(tmp, n);
2581 /* if value is becoming live through use by remat2 */
2582 if(has_reg_class(si, remat_arg) && !pset_find_ptr(live, remat_arg)) {
2583 op_t *remat_arg_op = get_irn_link(remat_arg);
2586 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 at bb start %+F\n", remat_arg, tmp));
2588 pset_insert_ptr(live, remat_arg);
2589 spill = add_to_spill_bb(si, bb, remat_arg);
2590 remat_arg_op->attr.live_range.ilp = ILP_UNDEF;
2592 /* we need reg_in and mem_in for this value; they will be referenced later */
2593 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", remat_arg, bb);
2594 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2595 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", remat_arg, bb);
2596 spill->mem_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
2599 /* optimization: all memory stuff should be 0, for we do not want to insert reloads for remats */
2600 ir_snprintf(buf, sizeof(buf), "nomem_%N_%N", remat_arg, bb);
2601 nomem = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0.0);
2602 lpp_set_factor_fast(si->lpp, nomem, spill->spill, 1.0);
2607 /* L\U is empty at bb start */
2608 /* arg is live throughout epilog if it is reg_in into this block */
2610 /* check the register pressure at the beginning of the block
2613 /* reg_in entspricht post_use */
2615 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2616 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, si->n_regs);
2618 pset_foreach(live, irn) {
2621 spill = set_find_spill(spill_bb->ilp, irn);
2624 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2625 spill->reg_in = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 0.0);
2627 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2629 /* spill + mem_in <= 1 */
2630 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2631 nospill = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
2633 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2634 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2636 } /* post_remats are NOT included in register pressure check because
2637 they do not increase regpressure */
2639 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2640 pset_foreach(live, irn) {
2644 spill = set_find_spill(spill_bb->ilp, irn);
2645 assert(spill && spill->irn == irn);
2647 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2648 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2651 ir_node *phi_arg = get_Phi_pred(irn, n);
2652 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2653 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2655 op_t *op = get_irn_link(irn);
2657 /* although the phi is in the right regclass one or more of
2658 * its arguments can be in a different one or at least to
2661 if(has_reg_class(si, phi_arg)) {
2662 /* mem_in < mem_out_arg + copy */
2663 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2664 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2666 /* reg_in < reg_out_arg */
2667 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2668 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2670 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2671 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2673 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2676 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2678 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2680 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2684 /* else assure the value arrives on all paths in the same resource */
2686 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2689 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2690 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2693 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2694 mem_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2695 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2696 reg_in = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2698 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2699 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2701 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2704 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2705 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2710 foreach_post_remat(bb, tmp) {
2713 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2714 ir_node *remat_arg = get_irn_n(tmp, n);
2715 op_t *remat_op = get_irn_link(tmp);
2717 if(!has_reg_class(si, remat_arg)) continue;
2719 spill = set_find_spill(spill_bb->ilp, remat_arg);
2722 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2723 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2724 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2725 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2729 pset_foreach(live, irn) {
2730 const op_t *op = get_irn_link(irn);
2731 const ir_node *remat;
2736 foreach_post_remat(bb, remat) {
2739 for (n=get_irn_arity(remat)-1; n>=0; --n) {
2740 const ir_node *arg = get_irn_n(remat, n);
2743 const op_t *remat_op = get_irn_link(remat);
2745 if(cst == ILP_UNDEF) {
2746 /* sum remat2s <= 1 + n_remats*live_range */
2747 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", irn, bb);
2748 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1.0);
2750 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2756 if(cst != ILP_UNDEF && op->attr.live_range.ilp != ILP_UNDEF) {
2757 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -n_remats);
2761 /* first live ranges from reg_ins */
2762 pset_foreach(live, irn) {
2763 op_t *op = get_irn_link(irn);
2765 if(op->attr.live_range.ilp != ILP_UNDEF) {
2767 spill = set_find_spill(spill_bb->ilp, irn);
2768 assert(spill && spill->irn == irn);
2770 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2771 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2772 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2773 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2775 foreach_post_remat(bb, tmp) {
2776 op_t *remat_op = get_irn_link(tmp);
2778 if(remat_op->attr.remat.remat->value == irn) {
2779 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2785 /* walk forward now and compute constraints for placing spills */
2786 /* this must only be done for values that are not defined in this block */
2787 pset_foreach(live, irn) {
2789 * if value is defined in this block we can anways place the spill directly after the def
2790 * -> no constraint necessary
2792 if(!is_Phi(irn) && get_nodes_block(irn) == bb) {
2797 spill = set_find_spill(spill_bb->ilp, irn);
2800 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2801 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0.0);
2803 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2804 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2807 sched_foreach_op(bb, tmp) {
2808 op_t *op = get_irn_link(tmp);
2810 if(is_Phi(tmp)) continue;
2811 assert(!is_Proj(tmp));
2814 const ir_node *value = op->attr.remat.remat->value;
2817 /* only collect remats up to the first real use of a value */
2818 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2823 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2824 ir_node *arg = get_irn_n(tmp, n);
2827 /* if a value is used stop collecting remats */
2840 typedef struct _irnlist_t {
2841 struct list_head list;
2845 typedef struct _interference_t {
2846 struct list_head blocklist;
2852 cmp_interference(const void *a, const void *b, size_t size)
2854 const interference_t *p = a;
2855 const interference_t *q = b;
2857 return !(p->a == q->a && p->b == q->b);
2860 static interference_t *
2861 set_find_interference(set * set, ir_node * a, ir_node * b)
2863 interference_t query;
2865 query.a = (a>b)?a:b;
2866 query.b = (a>b)?b:a;
2868 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2871 static interference_t *
2872 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2874 interference_t query,
2876 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2880 result = set_find_interference(set, a, b);
2883 list_add(&list->list, &result->blocklist);
2887 query.a = (a>b)?a:b;
2888 query.b = (a>b)?b:a;
2890 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2892 INIT_LIST_HEAD(&result->blocklist);
2893 list_add(&list->list, &result->blocklist);
2899 values_interfere_in_block(const spill_ilp_t * si, const ir_node * bb, const ir_node * a, const ir_node * b)
2901 const ir_edge_t *edge;
2903 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2904 /* both values are live in, so they interfere */
2908 /* ensure a dominates b */
2909 if(value_dominates(b,a)) {
2915 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2918 /* the following code is stolen from bera.c */
2919 if(be_is_live_end(si->lv, bb, a))
2922 foreach_out_edge(a, edge) {
2923 const ir_node *user = edge->src;
2924 if(get_nodes_block(user) == bb
2927 && !pset_find_ptr(si->inverse_ops, user)
2928 && value_dominates(b, user))
2936 * Walk all irg blocks and collect interfering values inside of phi classes
2939 luke_interferencewalker(ir_node * bb, void * data)
2941 spill_ilp_t *si = (spill_ilp_t*)data;
2944 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
2945 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
2946 op_t *a_op = get_irn_link(a);
2949 /* a is only interesting if it is in my register class and if it is inside a phi class */
2950 if (has_reg_class(si, a) && get_phi_class(si->pc, a)) {
2951 if (a_op->is_remat || pset_find_ptr(si->inverse_ops, a))
2954 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)) {
2955 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
2956 op_t *b_op = get_irn_link(b);
2958 /* a and b are only interesting if they are in the same phi class */
2959 if (has_reg_class(si, b) && get_phi_class(si->pc, a) == get_phi_class(si->pc, b)) {
2960 if (b_op->is_remat || pset_find_ptr(si->inverse_ops, b))
2963 if (values_interfere_in_block(si, bb, a, b)) {
2964 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2965 set_insert_interference(si, si->interferences, a, b, bb);
2973 static unsigned int copy_path_id = 0;
2976 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2983 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2984 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
2986 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2988 pset_foreach(copies, ptr) {
2989 copy = PTR_TO_INT(ptr);
2990 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2995 * @parameter copies contains a path of copies which lead us to irn
2996 * @parameter visited contains a set of nodes already visited on this path
2999 find_copy_path(spill_ilp_t * si, const ir_node * irn, const ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
3001 const ir_edge_t *edge;
3002 op_t *op = get_irn_link(irn);
3003 pset *visited_users = pset_new_ptr_default();
3006 if(op->is_remat) return 0;
3008 pset_insert_ptr(visited, irn);
3012 pset *visited_operands = pset_new_ptr(get_irn_arity(irn));
3014 /* visit all operands */
3015 for(n=get_irn_arity(irn)-1; n>=0; --n) {
3016 ir_node *arg = get_irn_n(irn, n);
3017 ilp_var_t copy = op->attr.live_range.args.copies[n];
3019 if(!has_reg_class(si, arg)) continue;
3020 if(pset_find_ptr(visited_operands, arg)) continue;
3021 pset_insert_ptr(visited_operands, arg);
3024 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3025 del_pset(visited_operands);
3026 del_pset(visited_users);
3027 pset_remove_ptr(visited, irn);
3030 pset_insert(copies, INT_TO_PTR(copy), copy);
3031 write_copy_path_cst(si, copies, any_interfere);
3032 pset_remove(copies, INT_TO_PTR(copy), copy);
3033 } else if(!pset_find_ptr(visited, arg)) {
3034 pset_insert(copies, INT_TO_PTR(copy), copy);
3035 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
3036 pset_remove(copies, INT_TO_PTR(copy), copy);
3038 if(paths > MAX_PATHS) {
3039 if(pset_count(copies) == 0) {
3043 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3044 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3045 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3046 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3047 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3051 del_pset(visited_operands);
3052 del_pset(visited_users);
3053 pset_remove_ptr(visited, irn);
3056 } else if(pset_count(copies) == 0) {
3062 del_pset(visited_operands);
3065 /* visit all uses which are phis */
3066 foreach_out_edge(irn, edge) {
3067 ir_node *user = edge->src;
3068 int pos = edge->pos;
3069 op_t *op = get_irn_link(user);
3072 if(!is_Phi(user)) continue;
3073 if(!has_reg_class(si, user)) continue;
3074 if(pset_find_ptr(visited_users, user)) continue;
3075 pset_insert_ptr(visited_users, user);
3077 copy = op->attr.live_range.args.copies[pos];
3079 if(user == target) {
3080 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
3081 del_pset(visited_users);
3082 pset_remove_ptr(visited, irn);
3085 pset_insert(copies, INT_TO_PTR(copy), copy);
3086 write_copy_path_cst(si, copies, any_interfere);
3087 pset_remove(copies, INT_TO_PTR(copy), copy);
3088 } else if(!pset_find_ptr(visited, user)) {
3089 pset_insert(copies, INT_TO_PTR(copy), copy);
3090 paths += find_copy_path(si, user, target, any_interfere, copies, visited);
3091 pset_remove(copies, INT_TO_PTR(copy), copy);
3093 if(paths > MAX_PATHS) {
3094 if(pset_count(copies) == 0) {
3098 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
3099 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_equal, 0);
3100 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3101 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
3102 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
3106 del_pset(visited_users);
3107 pset_remove_ptr(visited, irn);
3110 } else if(pset_count(copies) == 0) {
3116 del_pset(visited_users);
3117 pset_remove_ptr(visited, irn);
3122 gen_copy_constraints(spill_ilp_t * si, const ir_node * a, const ir_node * b, ilp_var_t any_interfere)
3124 pset * copies = pset_new_ptr_default();
3125 pset * visited = pset_new_ptr_default();
3127 find_copy_path(si, a, b, any_interfere, copies, visited);
3135 memcopyhandler(spill_ilp_t * si)
3137 interference_t *interference;
3139 /* teste Speicherwerte auf Interferenz */
3141 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
3142 irg_block_walk_graph(si->birg->irg, luke_interferencewalker, NULL, si);
3144 /* now lets emit the ILP unequations for the crap */
3145 set_foreach(si->interferences, interference) {
3147 ilp_var_t interfere, any_interfere;
3148 ilp_cst_t any_interfere_cst, cst;
3149 const ir_node *a = interference->a;
3150 const ir_node *b = interference->b;
3152 /* any_interf <= \sum interf */
3153 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
3154 any_interfere_cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3155 any_interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3157 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
3159 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
3160 const ir_node *bb = irnlist->irn;
3161 spill_bb_t *spill_bb = get_irn_link(bb);
3166 spilla = set_find_spill(spill_bb->ilp, a);
3169 spillb = set_find_spill(spill_bb->ilp, b);
3172 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
3173 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
3174 /* 2: - mem_in_a - spill_a + interfere <= 0 */
3175 /* 3: - mem_in_b - spill_b + interfere <= 0 */
3176 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
3177 interfere = lpp_add_var_default(si->lpp, buf, lpp_binary, 0.0, 1.0);
3179 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
3180 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 1);
3182 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
3183 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
3184 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
3185 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
3186 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
3188 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
3189 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3191 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3192 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
3193 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
3195 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
3196 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3198 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3199 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
3200 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
3203 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
3205 /* any_interfere >= interf */
3206 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
3207 cst = lpp_add_cst_uniq(si->lpp, buf, lpp_less, 0);
3209 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
3210 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
3213 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
3214 gen_copy_constraints(si,a,b,any_interfere);
3222 return fabs(x) < 0.00001;
3225 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
3227 spill_ilp_t *si = get_irg_link(current_ir_graph);
3229 if(pset_find_ptr(si->all_possible_remats, n)) {
3230 op_t *op = (op_t*)get_irn_link(n);
3231 assert(op && op->is_remat);
3233 if(!op->attr.remat.remat->inverse) {
3234 if(op->attr.remat.pre) {
3235 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
3237 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
3242 op_t *op = (op_t*)get_irn_link(n);
3243 assert(op && op->is_remat);
3245 if(op->attr.remat.pre) {
3246 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
3248 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
3259 dump_graph_with_remats(ir_graph * irg, const char * suffix)
3261 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
3262 be_dump(irg, suffix, dump_ir_block_graph_sched);
3263 set_dump_node_vcgattr_hook(NULL);
3267 * Edge hook to dump the schedule edges with annotated register pressure.
3270 sched_pressure_edge_hook(FILE *F, ir_node *irn)
3272 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
3273 ir_node *prev = sched_prev(irn);
3274 fprintf(F, "edge:{sourcename:\"");
3276 fprintf(F, "\" targetname:\"");
3278 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
3279 fprintf(F, "\" color:magenta}\n");
3285 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
3287 DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
3289 dump_consts_local(0);
3290 set_dump_node_edge_hook(sched_pressure_edge_hook);
3291 dump_ir_block_graph(irg, suffix);
3292 set_dump_node_edge_hook(old_edge_hook);
3296 walker_pressure_annotator(ir_node * bb, void * data)
3298 spill_ilp_t *si = data;
3301 pset *live = pset_new_ptr_default();
3304 be_lv_foreach(si->lv, bb, be_lv_state_end, i) {
3305 irn = be_lv_get_irn(si->lv, bb, i);
3307 if (has_reg_class(si, irn)) {
3308 pset_insert_ptr(live, irn);
3312 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
3314 sched_foreach_reverse(bb, irn) {
3316 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
3320 if(has_reg_class(si, irn)) {
3321 pset_remove_ptr(live, irn);
3322 if(is_Proj(irn)) ++projs;
3325 if(!is_Proj(irn)) projs = 0;
3327 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3328 ir_node *arg = get_irn_n(irn, n);
3330 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
3332 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
3339 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
3341 be_dump(si->birg->irg, suffix, dump_ir_block_graph_sched_pressure);
3345 connect_all_remats_with_keep(spill_ilp_t * si)
3353 n_remats = pset_count(si->all_possible_remats);
3355 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
3358 pset_foreach(si->all_possible_remats, irn) {
3363 si->keep = be_new_Keep(si->cls, si->birg->irg, get_irg_end_block(si->birg->irg), n_remats, ins);
3365 obstack_free(si->obst, ins);
3370 connect_all_spills_with_keep(spill_ilp_t * si)
3379 n_spills = pset_count(si->spills);
3381 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
3384 pset_foreach(si->spills, irn) {
3389 keep = be_new_Keep(si->cls, si->birg->irg, get_irg_end_block(si->birg->irg), n_spills, ins);
3391 obstack_free(si->obst, ins);
3395 /** insert a spill at an arbitrary position */
3396 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert)
3398 ir_node *bl = is_Block(insert) ? insert : get_nodes_block(insert);
3399 ir_graph *irg = get_irn_irg(bl);
3400 ir_node *frame = get_irg_frame(irg);
3403 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
3404 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
3406 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn);
3409 * search the right insertion point. a spill of a phi cannot be put
3410 * directly after the phi, if there are some phis behind the one which
3411 * is spilled. Also, a spill of a Proj must be after all Projs of the
3414 * Here's one special case:
3415 * If the spill is in the start block, the spill must be after the frame
3416 * pointer is set up. This is done by setting insert to the end of the block
3417 * which is its default initialization (see above).
3420 if (bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
3423 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
3426 sched_add_after(insert, spill);
3431 delete_remat(spill_ilp_t * si, ir_node * remat) {
3433 ir_node *bad = get_irg_bad(si->birg->irg);
3435 sched_remove(remat);
3437 /* kill links to operands */
3438 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
3439 set_irn_n(remat, n, bad);
3444 clean_remat_info(spill_ilp_t * si)
3448 remat_info_t *remat_info;
3449 ir_node *bad = get_irg_bad(si->birg->irg);
3451 set_foreach(si->remat_info, remat_info) {
3452 if(!remat_info->remats) continue;
3454 pset_foreach(remat_info->remats, remat)
3456 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
3457 if(sched_is_scheduled(remat->proj)) {
3458 sched_remove((ir_node*)remat->proj);
3460 set_irn_n((ir_node*)remat->proj, -1, bad);
3461 set_irn_n((ir_node*)remat->proj, 0, bad);
3464 if(get_irn_n_edges(remat->op) == 0) {
3465 if(sched_is_scheduled(remat->op)) {
3466 sched_remove((ir_node*)remat->op);
3468 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
3469 set_irn_n((ir_node*)remat->op, n, bad);
3474 if(remat_info->remats) del_pset(remat_info->remats);
3475 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3480 delete_unnecessary_remats(spill_ilp_t * si)
3482 if(opt_keep_alive & KEEPALIVE_REMATS) {
3484 ir_node *bad = get_irg_bad(si->birg->irg);
3487 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3488 ir_node *keep_arg = get_irn_n(si->keep, n);
3489 op_t *arg_op = get_irn_link(keep_arg);
3492 assert(arg_op->is_remat);
3494 name = si->lpp->vars[arg_op->attr.remat.ilp];
3496 if(is_zero(name->value)) {
3497 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
3498 /* TODO check whether reload is preferred over remat (could be bug) */
3499 delete_remat(si, keep_arg);
3501 if(!arg_op->attr.remat.remat->inverse) {
3502 if(arg_op->attr.remat.pre) {
3503 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3505 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3508 if(arg_op->attr.remat.pre) {
3509 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3511 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3516 set_irn_n(si->keep, n, bad);
3519 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
3524 pset_foreach(si->all_possible_remats, remat) {
3525 op_t *remat_op = get_irn_link(remat);
3526 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
3528 if(is_zero(name->value)) {
3529 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
3530 /* TODO check whether reload is preferred over remat (could be bug) */
3531 delete_remat(si, remat);
3533 if(!remat_op->attr.remat.remat->inverse) {
3534 if(remat_op->attr.remat.pre) {
3535 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3537 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3540 if(remat_op->attr.remat.pre) {
3541 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3543 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3552 get_spills_for_value(spill_ilp_t * si, const ir_node * value)
3554 pset *spills = pset_new_ptr_default();
3556 const ir_node *next;
3559 defs = set_find_def(si->values, value);
3561 if(defs && defs->spills) {
3562 for(next = defs->spills; next; next = get_irn_link(next)) {
3563 pset_insert_ptr(spills, next);
3571 new_r_PhiM_nokeep(ir_graph * irg, ir_node *block, int arity, ir_node **in)
3575 assert( get_irn_arity(block) == arity );
3577 res = new_ir_node(NULL, irg, block, op_Phi, mode_M, arity, in);
3578 res->attr.phi_backedge = new_backedge_arr(irg->obst, arity);
3584 * @param before The node after which the spill will be placed in the schedule
3587 insert_spill(spill_ilp_t * si, ir_node * irn, const ir_node * value, ir_node * before)
3591 const arch_env_t *arch_env = si->birg->main_env->arch_env;
3593 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3595 spill = be_spill2(arch_env, irn, before);
3597 defs = set_insert_def(si->values, value);
3600 /* enter into the linked list */
3601 set_irn_link(spill, defs->spills);
3602 defs->spills = spill;
3604 if(opt_keep_alive & KEEPALIVE_SPILLS)
3605 pset_insert_ptr(si->spills, spill);
3611 * @param before The Phi node which has to be spilled
3614 insert_mem_phi(spill_ilp_t * si, ir_node * phi)
3621 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3623 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3624 ins[n] = si->m_unknown;
3627 mem_phi = new_r_PhiM_nokeep(si->birg->irg, get_nodes_block(phi), get_irn_arity(phi), ins);
3629 defs = set_insert_def(si->values, phi);
3632 /* enter into the linked list */
3633 set_irn_link(mem_phi, defs->spills);
3634 defs->spills = mem_phi;
3636 #ifdef SCHEDULE_PHIM
3637 sched_add_after(phi, mem_phi);
3639 pset_insert_ptr(si->phims, mem_phi);
3642 if(opt_keep_alive & KEEPALIVE_SPILLS)
3643 pset_insert_ptr(si->spills, mem_phi);
3650 * Add remat to list of defs, destroys link field!
3653 insert_remat(spill_ilp_t * si, ir_node * remat)
3656 op_t *remat_op = get_irn_link(remat);
3658 assert(remat_op->is_remat);
3660 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3663 /* enter into the linked list */
3664 set_irn_link(remat, defs->remats);
3665 defs->remats = remat;
3670 * Add reload before operation and add to list of defs
3673 insert_reload(spill_ilp_t * si, const ir_node * value, ir_node * after)
3678 const arch_env_t *arch_env = si->birg->main_env->arch_env;
3680 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3682 defs = set_find_def(si->values, value);
3684 spill = defs->spills;
3685 assert(spill && "no spill placed before reload");
3687 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3689 /* enter into the linked list */
3690 set_irn_link(reload, defs->remats);
3691 defs->remats = reload;
3696 void perform_memory_operand(spill_ilp_t * si, memoperand_t * memoperand)
3699 ir_node *value = get_irn_n(memoperand->irn, memoperand->pos);
3701 const arch_env_t *arch_env = si->birg->main_env->arch_env;
3703 DBG((si->dbg, LEVEL_2, "\t inserting memory operand for value %+F at %+F\n", value, memoperand->irn));
3705 defs = set_find_def(si->values, value);
3707 spill = defs->spills;
3708 assert(spill && "no spill placed before reload");
3710 arch_perform_memory_operand(arch_env, memoperand->irn, spill, memoperand->pos);
3713 void insert_memoperands(spill_ilp_t * si)
3715 memoperand_t *memoperand;
3718 set_foreach(si->memoperands, memoperand) {
3719 name = si->lpp->vars[memoperand->ilp];
3720 if(!is_zero(name->value)) {
3721 perform_memory_operand(si, memoperand);
3727 walker_spill_placer(ir_node * bb, void * data) {
3728 spill_ilp_t *si = (spill_ilp_t*)data;
3730 spill_bb_t *spill_bb = get_irn_link(bb);
3731 pset *spills_to_do = pset_new_ptr_default();
3734 set_foreach(spill_bb->ilp, spill) {
3737 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3738 name = si->lpp->vars[spill->mem_in];
3739 if(!is_zero(name->value)) {
3742 mem_phi = insert_mem_phi(si, spill->irn);
3744 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3748 name = si->lpp->vars[spill->spill];
3749 if(!is_zero(name->value)) {
3750 /* place spill directly after definition */
3751 if(get_nodes_block(spill->irn) == bb) {
3752 insert_spill(si, spill->irn, spill->irn, spill->irn);
3756 /* place spill at bb start */
3757 if(spill->reg_in > 0) {
3758 name = si->lpp->vars[spill->reg_in];
3759 if(!is_zero(name->value)) {
3760 insert_spill(si, spill->irn, spill->irn, bb);
3764 /* place spill after a remat */
3765 pset_insert_ptr(spills_to_do, spill->irn);
3768 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3771 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3772 op_t *op = get_irn_link(irn);
3774 if(be_is_Spill(irn)) continue;
3777 /* TODO fix this if we want to support remats with more than two nodes */
3778 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3779 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3781 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3784 if(pset_find_ptr(spills_to_do, irn)) {
3785 pset_remove_ptr(spills_to_do, irn);
3787 insert_spill(si, irn, irn, irn);
3793 assert(pset_count(spills_to_do) == 0);
3795 /* afterwards free data in block */
3796 del_pset(spills_to_do);
3800 insert_mem_copy(spill_ilp_t * si, ir_node * bb, ir_node * value)
3802 ir_node *insert_pos = bb;
3804 const arch_env_t *arch_env = si->birg->main_env->arch_env;
3806 /* find last definition of arg value in block */
3811 defs = set_find_def(si->values, value);
3813 if(defs && defs->remats) {
3814 for(next = defs->remats; next; next = get_irn_link(next)) {
3815 if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3816 last = sched_get_time_step(next);
3822 if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3823 last = sched_get_time_step(value);
3827 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
3829 spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos);
3835 phim_fixer(spill_ilp_t *si) {
3838 set_foreach(si->values, defs) {
3839 const ir_node *phi = defs->value;
3840 op_t *op = get_irn_link(phi);
3841 ir_node *phi_m = NULL;
3842 ir_node *next = defs->spills;
3845 if(!is_Phi(phi)) continue;
3848 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3852 next = get_irn_link(next);
3855 if(!phi_m) continue;
3857 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3858 ir_node *value = get_irn_n(phi, n);
3859 defs_t *val_defs = set_find_def(si->values, value);
3861 /* a spill of this value */
3866 ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3867 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3869 if(!is_zero(name->value)) {
3870 spill = insert_mem_copy(si, pred, value);
3872 spill = val_defs->spills;
3875 spill = val_defs->spills;
3878 assert(spill && "no spill placed before PhiM");
3879 set_irn_n(phi_m, n, spill);
3885 walker_reload_placer(ir_node * bb, void * data) {
3886 spill_ilp_t *si = (spill_ilp_t*)data;
3888 spill_bb_t *spill_bb = get_irn_link(bb);
3890 /* reloads at end of block */
3891 if(spill_bb->reloads) {
3894 set_foreach(spill_bb->reloads, keyval) {
3895 ir_node *irn = (ir_node*)keyval->key;
3896 ilp_var_t reload = PTR_TO_INT(keyval->val);
3899 name = si->lpp->vars[reload];
3900 if(!is_zero(name->value)) {
3902 ir_node *insert_pos = bb;
3903 ir_node *prev = sched_block_last_noncf(si, bb);
3904 op_t *prev_op = get_irn_link(prev);
3906 while(be_is_Spill(prev)) {
3907 prev = sched_prev(prev);
3910 prev_op = get_irn_link(prev);
3912 /* insert reload before pre-remats */
3913 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3914 && prev_op->is_remat && prev_op->attr.remat.pre) {
3918 prev = sched_prev(prev);
3919 } while(be_is_Spill(prev));
3921 prev_op = get_irn_link(prev);
3925 reload = insert_reload(si, irn, insert_pos);
3927 if(opt_keep_alive & KEEPALIVE_RELOADS)
3928 pset_insert_ptr(si->spills, reload);
3933 /* walk and insert more reloads and collect remats */
3934 sched_foreach_reverse(bb, irn) {
3935 op_t *op = get_irn_link(irn);
3937 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3938 if(is_Phi(irn)) break;
3941 if(get_irn_mode(irn) != mode_T) {
3942 insert_remat(si, irn);
3947 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3948 ir_node *arg = get_irn_n(irn, n);
3950 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3953 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3954 if(!is_zero(name->value)) {
3956 ir_node *insert_pos = irn;
3957 ir_node *prev = sched_prev(insert_pos);
3960 while(be_is_Spill(prev)) {
3961 prev = sched_prev(prev);
3964 prev_op = get_irn_link(prev);
3966 /* insert reload before pre-remats */
3967 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3968 && prev_op->is_remat && prev_op->attr.remat.pre) {
3972 prev = sched_prev(prev);
3973 } while(be_is_Spill(prev));
3975 prev_op = get_irn_link(prev);
3979 reload = insert_reload(si, arg, insert_pos);
3981 assert(reload && "no reload returned");
3982 set_irn_n(irn, n, reload);
3984 if(opt_keep_alive & KEEPALIVE_RELOADS)
3985 pset_insert_ptr(si->spills, reload);
3992 del_set(spill_bb->ilp);
3993 if(spill_bb->reloads) del_set(spill_bb->reloads);
3997 walker_collect_used(ir_node * irn, void * data)
3999 bitset_t *used = data;
4001 bitset_set(used, get_irn_idx(irn));
4004 struct kill_helper {
4010 walker_kill_unused(ir_node * bb, void * data)
4012 struct kill_helper *kh = data;
4013 ir_node *bad = get_irg_bad(get_irn_irg(bb));
4017 for(irn=sched_first(bb); !sched_is_end(irn);) {
4018 ir_node *next = sched_next(irn);
4021 if(!bitset_is_set(kh->used, get_irn_idx(irn))) {
4022 if(be_is_Spill(irn) || be_is_Reload(irn)) {
4023 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)));
4025 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
4031 set_nodes_block(irn, bad);
4032 for (n=get_irn_arity(irn)-1; n>=0; --n) {
4033 set_irn_n(irn, n, bad);
4040 #ifndef SCHEDULE_PHIM
4042 kill_unused_phims(spill_ilp_t * si, struct kill_helper * kh)
4045 ir_node *bad = get_irg_bad(si->birg->irg);
4048 pset_foreach(si->phims, phi) {
4049 if(!bitset_is_set(kh->used, get_irn_idx(phi))) {
4051 set_nodes_block(phi, bad);
4052 for (n=get_irn_arity(phi)-1; n>=0; --n) {
4053 set_irn_n(phi, n, bad);
4061 kill_all_unused_values_in_schedule(spill_ilp_t * si)
4063 struct kill_helper kh;
4065 kh.used = bitset_malloc(get_irg_last_idx(si->birg->irg));
4068 irg_walk_graph(si->birg->irg, walker_collect_used, NULL, kh.used);
4069 #ifndef SCHEDULE_PHIM
4070 kill_unused_phims(si, &kh);
4072 irg_block_walk_graph(si->birg->irg, walker_kill_unused, NULL, &kh);
4074 bitset_free(kh.used);
4078 print_irn_pset(pset * p)
4082 pset_foreach(p, irn) {
4083 ir_printf("%+F\n", irn);
4088 dump_phi_class(spill_ilp_t *si, ir_node **phiclass, const char * file)
4090 FILE *f = fopen(file, "w");
4092 interference_t *interference;
4095 set_break(si->interferences);
4097 ir_fprintf(f, "digraph phiclass {\n");
4099 for (i = ARR_LEN(phiclass) - 1; i >= 0; --i) {
4102 ir_fprintf(f, " %F%N [shape=box]\n", irn, irn);
4105 for (i = ARR_LEN(phiclass) - 1; i >= 0; --i) {
4112 for (n = get_irn_arity(irn) - 1; n >= 0; --n) {
4113 ir_node *arg = get_irn_n(irn, n);
4115 ir_fprintf(f, " %F%N -> %F%N\n", irn, irn, arg, arg);
4119 set_foreach(si->interferences, interference) {
4120 const ir_node *a = interference->a;
4121 const ir_node *b = interference->b;
4122 if (get_phi_class(si->pc, (ir_node *)a) == phiclass) {
4123 ir_fprintf(f, " %F%N -> %F%N [color=red,dir=none,style=bold]\n", a, a, b, b);
4132 rewire_uses(spill_ilp_t * si)
4135 ir_nodeset_t ignore;
4137 ir_nodeset_init(&ignore);
4138 ir_nodeset_insert(&ignore, get_irg_end(si->birg->irg));
4140 /* then fix uses of spills */
4141 set_foreach(si->values, defs) {
4144 const ir_node *next = defs->remats;
4147 reloads = pset_new_ptr_default();
4150 if(be_is_Reload(next)) {
4151 pset_insert_ptr(reloads, next);
4155 next = get_irn_link(next);
4158 spills = get_spills_for_value(si, defs->value);
4159 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));
4160 if(pset_count(spills) > 1) {
4161 be_ssa_construction_env_t senv;
4163 //assert(pset_count(reloads) > 0);
4164 // print_irn_pset(spills);
4165 // print_irn_pset(reloads);
4167 be_ssa_construction_init(&senv, si->birg);
4168 be_ssa_construction_set_ignore_uses(&senv, &ignore);
4169 pset_foreach(spills, node) {
4170 be_ssa_construction_add_copy(&senv, node);
4172 pset_foreach(spills, node) {
4173 be_ssa_construction_fix_users(&senv, node);
4175 be_ssa_construction_update_liveness_phis(&senv, si->lv);
4176 pset_foreach(spills, node) {
4177 be_liveness_update(si->lv, node);
4179 be_ssa_construction_destroy(&senv);
4186 /* first fix uses of remats and reloads */
4187 set_foreach(si->values, defs) {
4188 const ir_node *next = defs->remats;
4192 be_ssa_construction_env_t senv;
4194 be_ssa_construction_init(&senv, si->birg);
4196 if(sched_is_scheduled(defs->value)) {
4197 be_ssa_construction_add_copy(&senv, (ir_node*) defs->value);
4201 next = defs->remats;
4203 be_ssa_construction_add_copy(&senv, (ir_node*) next);
4204 next = get_irn_link(next);
4207 if(sched_is_scheduled(defs->value)) {
4208 be_ssa_construction_fix_users(&senv, (ir_node*) defs->value);
4211 next = defs->remats;
4213 be_ssa_construction_fix_users(&senv, (ir_node*) next);
4214 next = get_irn_link(next);
4217 be_ssa_construction_update_liveness_phis(&senv, si->lv);
4218 if(sched_is_scheduled(defs->value)) {
4219 be_liveness_update(si->lv, (ir_node*) defs->value);
4222 next = defs->remats;
4224 be_liveness_update(si->lv, (ir_node*) next);
4225 next = get_irn_link(next);
4228 be_ssa_construction_destroy(&senv);
4232 ir_nodeset_destroy(&ignore);
4233 // remove_unused_defs(si);
4238 writeback_results(spill_ilp_t * si)
4240 /* walk through the graph and collect all spills, reloads and remats for a value */
4242 si->values = new_set(cmp_defs, 4096);
4244 DBG((si->dbg, LEVEL_1, "Applying results\n"));
4245 delete_unnecessary_remats(si);
4246 si->m_unknown = new_r_Unknown(si->birg->irg, mode_M);
4247 irg_block_walk_graph(si->birg->irg, walker_spill_placer, NULL, si);
4248 irg_block_walk_graph(si->birg->irg, walker_reload_placer, NULL, si);
4250 insert_memoperands(si);
4253 /* clean the remat info! there are still back-edges leading there! */
4254 clean_remat_info(si);
4258 connect_all_spills_with_keep(si);
4260 del_set(si->values);
4264 get_n_regs(spill_ilp_t * si)
4266 int arch_n_regs = arch_register_class_n_regs(si->cls);
4268 bitset_t *arch_regs = bitset_malloc(arch_n_regs);
4269 bitset_t *abi_regs = bitset_malloc(arch_n_regs);
4271 arch_put_non_ignore_regs(si->birg->main_env->arch_env, si->cls, arch_regs);
4272 be_abi_put_ignore_regs(si->birg->abi, si->cls, abi_regs);
4274 bitset_andnot(arch_regs, abi_regs);
4275 arch_n_regs = bitset_popcnt(arch_regs);
4277 bitset_free(arch_regs);
4278 bitset_free(abi_regs);
4280 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", arch_n_regs, si->cls->name));
4285 walker_reload_mover(ir_node * bb, void * data)
4287 spill_ilp_t *si = data;
4290 sched_foreach(bb, tmp) {
4291 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
4292 ir_node *reload = tmp;
4295 /* move reload upwards */
4297 int pressure = (int)get_irn_link(reload);
4298 if(pressure < si->n_regs) {
4299 irn = sched_prev(reload);
4300 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
4301 sched_remove(reload);
4302 pressure = (int)get_irn_link(irn);
4304 while(pressure < si->n_regs) {
4305 if( sched_is_end(irn) ||
4306 (be_is_Reload(irn) && has_reg_class(si, irn)) ||
4307 /* do not move reload before its spill */
4308 (irn == be_get_Reload_mem(reload)) ||
4309 /* do not move before phi */
4312 set_irn_link(irn, INT_TO_PTR(pressure+1));
4313 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
4314 irn = sched_prev(irn);
4316 pressure = (int)get_irn_link(irn);
4319 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
4320 sched_put_after(irn, reload);
4327 move_reloads_upward(spill_ilp_t * si)
4329 irg_block_walk_graph(si->birg->irg, walker_reload_mover, NULL, si);
4334 * Walk all irg blocks and check for interfering spills inside of phi classes
4337 luke_meminterferencechecker(ir_node * bb, void * data)
4339 spill_ilp_t *si = (spill_ilp_t*)data;
4342 be_lv_foreach(si->lv, bb, be_lv_state_end | be_lv_state_out | be_lv_state_in, l1) {
4343 ir_node *a = be_lv_get_irn(si->lv, bb, l1);
4345 if (! be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T))
4348 /* a is only interesting if it is in my register class and if it is inside a phi class */
4349 if (has_reg_class(si, a) && get_phi_class(si->pc, a)) {
4350 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)) {
4351 ir_node *b = be_lv_get_irn(si->lv, bb, l2);
4353 if (! be_is_Spill(b) && (! is_Phi(b) || get_irn_mode(b) != mode_T))
4356 /* a and b are only interesting if they are in the same phi class */
4357 if (has_reg_class(si, b) && get_phi_class(si->pc, a) == get_phi_class(si->pc, b)) {
4358 if (values_interfere_in_block(si, bb, a, b)) {
4359 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
4368 verify_phiclasses(spill_ilp_t * si)
4370 /* analyze phi classes */
4371 phi_class_free(si->pc);
4372 si->pc = phi_class_new_from_irg(si->birg->irg, 0);
4374 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
4375 irg_block_walk_graph(si->birg->irg, luke_meminterferencechecker, NULL, si);
4379 be_spill_remat(be_irg_t *birg, const arch_register_class_t *cls)
4382 char problem_name[256];
4383 char dump_suffix[256];
4384 char dump_suffix2[256];
4385 struct obstack obst;
4387 ir_graph *irg = be_get_birg_irg(birg);
4389 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", irg, cls->name);
4390 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", cls->name);
4391 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", cls->name);
4393 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
4394 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
4396 if(opt_verify & VERIFY_DOMINANCE)
4397 be_check_dominance(irg);
4399 be_assure_dom_front(birg);
4400 be_assure_liveness(birg);
4402 obstack_init(&obst);
4406 si.lpp = new_lpp(problem_name, lpp_minimize);
4407 si.remat_info = new_set(cmp_remat_info, 4096);
4408 si.interferences = new_set(cmp_interference, 32);
4409 si.memoperands = new_set(cmp_memoperands, 128);
4410 si.all_possible_remats = pset_new_ptr_default();
4411 si.spills = pset_new_ptr_default();
4412 si.inverse_ops = pset_new_ptr_default();
4415 si.n_regs = get_n_regs(&si);
4417 set_irg_link(irg, &si);
4420 /* compute phi classes */
4421 // phi_class_compute(irg);
4423 if(opt_dump_flags & DUMP_STATS)
4424 be_analyze_regpressure(birg, cls, "-pre");
4426 DBG((si.dbg, LEVEL_2, "\t initializing\n"));
4427 irg_block_walk_graph(irg, luke_initializer, NULL, &si);
4430 /* collect remats */
4431 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
4432 irg_walk_graph(irg, walker_remat_collector, NULL, &si);
4435 /* insert possible remats */
4436 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
4437 irg_block_walk_graph(irg, walker_remat_insertor, NULL, &si);
4438 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
4440 if(opt_keep_alive & KEEPALIVE_REMATS) {
4441 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
4442 connect_all_remats_with_keep(&si);
4443 /* dump graph with inserted remats */
4444 dump_graph_with_remats(irg, dump_suffix);
4447 /* insert copies for phi arguments not in my regclass */
4448 irg_walk_graph(irg, walker_regclass_copy_insertor, NULL, &si);
4450 /* recompute liveness */
4451 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
4452 be_liveness_recompute(si.lv);
4455 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
4456 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
4457 irg_block_walk_graph(irg, luke_endwalker, NULL, &si);
4459 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
4460 irg_block_walk_graph(irg, luke_blockwalker, NULL, &si);
4462 si.pc = phi_class_new_from_irg(birg->irg, 0);
4463 if (opt_memcopies) {
4464 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
4465 memcopyhandler(&si);
4468 if (opt_dump_flags & DUMP_PROBLEM) {
4470 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
4471 if ((f = fopen(buf, "wt")) != NULL) {
4472 lpp_dump_plain(si.lpp, f);
4477 if (opt_dump_flags & DUMP_MPS) {
4480 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mps", problem_name);
4481 if ((f = fopen(buf, "wt")) != NULL) {
4482 mps_write_mps(si.lpp, s_mps_fixed, f);
4486 ir_snprintf(buf, sizeof(buf), "%s-spillremat.mst", problem_name);
4487 if ((f = fopen(buf, "wt")) != NULL) {
4488 mps_write_mst(si.lpp, s_mps_fixed, f);
4493 lpp_check_startvals(si.lpp);
4496 DBG((si.dbg, LEVEL_1, "\tSolving %s (%d variables, %d constraints)\n", problem_name, si.lpp->var_next, si.lpp->cst_next));
4497 lpp_set_time_limit(si.lpp, opt_timeout);
4500 lpp_set_log(si.lpp, stdout);
4503 lpp_solve_cplex(si.lpp);
4505 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
4507 assert(lpp_is_sol_valid(si.lpp)
4508 && "solution of ILP must be valid");
4510 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));
4512 if(opt_dump_flags & DUMP_SOLUTION) {
4516 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
4517 if ((f = fopen(buf, "wt")) != NULL) {
4519 for (i = 0; i < si.lpp->var_next; ++i) {
4520 lpp_name_t *name = si.lpp->vars[i];
4521 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
4527 #ifndef SCHEDULE_PHIM
4528 si.phims = pset_new_ptr_default();
4530 writeback_results(&si);
4535 kill_all_unused_values_in_schedule(&si);
4537 #if !defined(SCHEDULE_PHIM) && defined(SOLVE)
4541 if(opt_keep_alive & (KEEPALIVE_SPILLS | KEEPALIVE_RELOADS))
4542 be_dump(irg, "-spills-placed", dump_ir_block_graph);
4544 // move reloads upwards
4545 be_liveness_recompute(si.lv);
4546 irg_block_walk_graph(irg, walker_pressure_annotator, NULL, &si);
4547 move_reloads_upward(&si);
4550 verify_phiclasses(&si);
4553 irg_block_walk_graph(irg, walker_pressure_annotator, NULL, &si);
4555 if(opt_dump_flags & DUMP_PRESSURE)
4556 dump_pressure_graph(&si, dump_suffix2);
4558 if(opt_dump_flags & DUMP_STATS)
4559 be_analyze_regpressure(birg, cls, "-post");
4561 if(opt_verify & VERIFY_DOMINANCE)
4562 be_check_dominance(irg);
4565 del_set(si.interferences);
4566 del_pset(si.inverse_ops);
4567 del_pset(si.all_possible_remats);
4568 del_set(si.memoperands);
4569 del_pset(si.spills);
4571 phi_class_free(si.pc);
4572 obstack_free(&obst, NULL);
4573 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4576 void be_init_spillremat(void)
4578 static be_spiller_t remat_spiller = {
4581 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
4582 lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
4583 lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
4584 lc_opt_entry_t *remat_grp = lc_opt_get_grp(chordal_grp, "remat");
4586 be_register_spiller("remat", &remat_spiller);
4587 lc_opt_add_table(remat_grp, options);
4590 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillremat);
4592 #else /* WITH_ILP */
4595 only_that_you_can_compile_without_WITH_ILP_defined(void)
4599 #endif /* WITH_ILP */