1 /** vim: set sw=4 ts=4:
4 * @author Adam M. Szalkowski & Sebastian Hack
6 * ILP based spilling & rematerialization
8 * Copyright (C) 2006 Universitaet Karlsruhe
9 * Released under the GPL
32 #include "phiclass_t.h"
38 #include <lpp/lpp_net.h>
39 #include <lpp/lpp_cplex.h>
40 //#include <lc_pset.h>
41 #include <libcore/lc_bitset.h>
45 #include "besched_t.h"
50 #include "bespillremat.h"
52 #include "bepressurestat.h"
54 #include "bechordal_t.h"
60 //#define KEEPALIVE /* keep alive all inserted remats and dump graph with remats */
61 #define COLLECT_REMATS /* enable rematerialization */
62 #define COLLECT_INVERSE_REMATS /* enable placement of inverse remats */
63 #define REMAT_WHILE_LIVE /* only remat values that are live */
64 //#define NO_ENLARGE_L1V3N355 /* do not remat after the death of some operand */
65 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
66 #define MAY_DIE_AT_REMAT /* allow values to die after a pre remat */
67 #define NO_SINGLE_USE_REMATS /* do not repair schedule */
68 //#define KEEPALIVE_SPILLS
69 //#define KEEPALIVE_RELOADS
70 #define GOODWIN_REDUCTION
71 //#define NO_MEMCOPIES
72 //#define VERIFY_DOMINANCE
76 #define LPP_SERVER "i44pc52"
77 #define LPP_SOLVER "cplex"
83 #define ILP_TIMEOUT 300
87 typedef struct _spill_ilp_t {
88 const arch_register_class_t *cls;
90 const be_chordal_env_t *chordal_env;
94 pset *all_possible_remats;
99 set *values; /**< for collecting all definitions of values before running ssa-construction */
103 DEBUG_ONLY(firm_dbg_module_t * dbg);
106 typedef int ilp_var_t;
107 typedef int ilp_cst_t;
109 typedef struct _spill_bb_t {
114 typedef struct _remat_t {
115 const ir_node *op; /**< for copy_irn */
116 const ir_node *value; /**< the value which is being recomputed by this remat */
117 ir_node *proj; /**< not NULL if the above op produces a tuple */
118 int cost; /**< cost of this remat */
119 int inverse; /**< nonzero if this is an inverse remat */
123 * Data to be attached to each IR node. For remats this contains the ilp_var
124 * for this remat and for normal ops this contains the ilp_vars for
125 * reloading each operand
127 typedef struct _op_t {
132 remat_t *remat; /** the remat this op belongs to */
133 int pre; /** 1, if this is a pressure-increasing remat */
137 ir_node *op; /** the operation this live range belongs to */
146 typedef struct _defs_t {
148 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
149 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
152 typedef struct _remat_info_t {
153 const ir_node *irn; /**< the irn to which these remats belong */
154 pset *remats; /**< possible remats for this value */
155 pset *remats_by_operand; /**< remats with this value as operand */
158 typedef struct _keyval_t {
163 typedef struct _spill_t {
173 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
175 return chordal_has_class(si->chordal_env, irn);
180 cmp_remat(const void *a, const void *b)
182 const keyval_t *p = a;
183 const keyval_t *q = b;
184 const remat_t *r = p->val;
185 const remat_t *s = q->val;
189 return !(r == s || r->op == s->op);
193 cmp_remat(const void *a, const void *b)
195 const remat_t *r = a;
196 const remat_t *s = a;
198 return !(r == s || r->op == s->op);
202 cmp_spill(const void *a, const void *b, size_t size)
204 const spill_t *p = a;
205 const spill_t *q = b;
207 // return !(p->irn == q->irn && p->bb == q->bb);
208 return !(p->irn == q->irn);
212 set_find_keyval(set * set, void * key)
217 return set_find(set, &query, sizeof(query), HASH_PTR(key));
221 set_insert_keyval(set * set, void * key, void * val)
227 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
231 set_find_def(set * set, ir_node * value)
236 return set_find(set, &query, sizeof(query), HASH_PTR(value));
240 set_insert_def(set * set, ir_node * value)
247 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
251 set_find_spill(set * set, ir_node * value)
256 return set_find(set, &query, sizeof(query), HASH_PTR(value));
259 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
260 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
261 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
262 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
263 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
266 cmp_remat_info(const void *a, const void *b, size_t size)
268 const remat_info_t *p = a;
269 const remat_info_t *q = b;
271 return !(p->irn == q->irn);
275 cmp_defs(const void *a, const void *b, size_t size)
280 return !(p->value == q->value);
284 cmp_keyval(const void *a, const void *b, size_t size)
286 const keyval_t *p = a;
287 const keyval_t *q = b;
289 return !(p->key == q->key);
293 execution_frequency(const spill_ilp_t *si, const ir_node * irn)
296 #ifndef EXECFREQ_LOOPDEPH
297 return get_block_execfreq(si->chordal_env->exec_freq, get_block(irn)) + FUDGE;
300 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
302 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
307 get_cost(const spill_ilp_t * si, const ir_node * irn)
309 if(be_is_Spill(irn)) {
311 } else if(be_is_Reload(irn)){
314 return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
319 * Checks, whether node and its operands have suitable reg classes
322 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
325 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
326 int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
330 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
333 for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
334 ir_node *op = get_irn_n(irn, n);
335 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
338 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
345 * Try to create a remat from @p op with destination value @p dest_value
347 static INLINE remat_t *
348 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
350 remat_t *remat = NULL;
352 // if(!mode_is_datab(get_irn_mode(dest_value)))
355 if(dest_value == op) {
356 const ir_node *proj = NULL;
358 if(is_Proj(dest_value)) {
359 op = get_irn_n(op, 0);
363 if(!is_rematerializable(si, op))
366 remat = obstack_alloc(si->obst, sizeof(*remat));
368 remat->cost = get_cost(si, op);
369 remat->value = dest_value;
373 arch_inverse_t inverse;
376 /* get the index of the operand we want to retrieve by the inverse op */
377 for (n = get_irn_arity(op)-1; n>=0; --n) {
378 ir_node *arg = get_irn_n(op, n);
380 if(arg == dest_value) break;
384 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", n, op));
386 /* else ask the backend to give an inverse op */
387 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
390 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
392 assert(inverse.n > 0 && "inverse op should have at least one node");
394 for(i=inverse.n-1; i>=0; --i) {
395 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
399 remat = obstack_alloc(si->obst, sizeof(*remat));
400 remat->op = inverse.nodes[0];
401 remat->cost = inverse.costs;
402 remat->value = dest_value;
403 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
406 assert(is_Proj(remat->proj));
408 assert(0 && "I can not handle remats with more than 2 nodes");
415 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
417 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
425 add_remat(const spill_ilp_t * si, const remat_t * remat)
427 remat_info_t *remat_info,
432 assert(remat->value);
434 query.irn = remat->value;
436 query.remats_by_operand = NULL;
437 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
439 if(remat_info->remats == NULL) {
440 remat_info->remats = new_pset(cmp_remat, 4096);
442 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
444 /* insert the remat into the remats_be_operand set of each argument of the remat op */
445 for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
446 ir_node *arg = get_irn_n(remat->op, n);
450 query.remats_by_operand = NULL;
451 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
453 if(remat_info->remats_by_operand == NULL) {
454 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
456 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
461 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
463 const ir_edge_t *edge = get_irn_out_edge_first(irn);
467 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
470 edge = get_irn_out_edge_next(irn, edge);
477 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
482 #ifdef NO_SINGLE_USE_REMATS
483 if(has_reg_class(si, op) && (get_irn_n_nonremat_edges(si, op) > 1)) {
485 if(has_reg_class(si, op)) {
487 remat = get_remat_from_op(si, op, op);
489 add_remat(si, remat);
493 #ifdef COLLECT_INVERSE_REMATS
494 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
496 for (n = get_irn_arity(op)-1; n>=0; --n) {
497 ir_node *arg = get_irn_n(op, n);
499 if(has_reg_class(si, arg)) {
500 /* try to get an inverse remat */
501 remat = get_remat_from_op(si, arg, op);
503 add_remat(si, remat);
512 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
515 ir_node *def_block = get_nodes_block(val);
521 /* if pos is at end of a basic block */
523 ret = (pos == def_block || block_dominates(def_block, pos));
524 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
528 /* else if this is a normal operation */
529 block = get_nodes_block(pos);
530 if(block == def_block) {
531 if(!sched_is_scheduled(val)) return 1;
533 ret = sched_comes_after(val, pos);
534 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
538 ret = block_dominates(def_block, block);
539 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
543 static INLINE ir_node *
544 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
546 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
550 * Returns first non-Phi node of block @p bb
552 static INLINE ir_node *
553 sched_block_first_nonphi(const ir_node * bb)
555 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
559 sched_skip_proj_predicator(const ir_node * irn, void * data)
561 return (is_Proj(irn));
564 static INLINE ir_node *
565 sched_next_nonproj(const ir_node * irn, int forward)
567 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
571 * Returns next operation node (non-Proj) after @p irn
572 * or the basic block of this node
574 static INLINE ir_node *
575 sched_next_op(const ir_node * irn)
577 ir_node *next = sched_next(irn);
582 return sched_next_nonproj(next, 1);
586 * Returns previous operation node (non-Proj) before @p irn
587 * or the basic block of this node
589 static INLINE ir_node *
590 sched_prev_op(const ir_node * irn)
592 ir_node *prev = sched_prev(irn);
597 return sched_next_nonproj(prev, 0);
601 sched_put_after(ir_node * insert, ir_node * irn)
603 if(is_Block(insert)) {
604 insert = sched_block_first_nonphi(insert);
606 insert = sched_next_op(insert);
608 sched_add_before(insert, irn);
612 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
614 if(is_Block(insert)) {
615 insert = sched_block_last_noncf(si, insert);
617 insert = sched_next_nonproj(insert, 0);
618 insert = sched_prev(insert);
620 sched_add_after(insert, irn);
624 * Tells you whether a @p remat can be placed before the irn @p pos
627 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
629 const ir_node *op = remat->op;
635 prev = sched_block_last_noncf(si, pos);
636 prev = sched_next_nonproj(prev, 0);
638 prev = sched_prev_op(pos);
640 /* do not remat if the rematted value is defined immediately before this op */
641 if(prev == remat->op) {
646 /* this should be just fine, the following OP will be using this value, right? */
648 /* only remat AFTER the real definition of a value (?) */
649 if(!value_is_defined_before(si, pos, remat->value)) {
650 // ir_fprintf(stderr, "error(not defined)");
655 for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
656 const ir_node *arg = get_irn_n(op, n);
658 #ifdef NO_ENLARGE_L1V3N355
659 if(has_reg_class(si, arg) && live) {
660 res &= pset_find_ptr(live, arg)?1:0;
662 res &= value_is_defined_before(si, pos, arg);
665 res &= value_is_defined_before(si, pos, arg);
673 * Tells you whether a @p remat can be placed after the irn @p pos
676 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
679 pos = sched_block_first_nonphi(pos);
681 pos = sched_next_op(pos);
684 /* only remat AFTER the real definition of a value (?) */
685 if(!value_is_defined_before(si, pos, remat->value)) {
689 return can_remat_before(si, remat, pos, live);
693 * Collect potetially rematerializable OPs
696 walker_remat_collector(ir_node * irn, void * data)
698 spill_ilp_t *si = data;
700 if(!is_Block(irn) && !is_Phi(irn)) {
701 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
702 get_remats_from_op(si, irn);
707 * Inserts a copy of @p irn before @p pos
710 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
715 bb = is_Block(pos)?pos:get_nodes_block(pos);
716 copy = exact_copy(irn);
718 _set_phi_class(copy, NULL);
719 set_nodes_block(copy, bb);
720 sched_put_before(si, pos, copy);
726 * Inserts a copy of @p irn after @p pos
729 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
734 bb = is_Block(pos)?pos:get_nodes_block(pos);
735 copy = exact_copy(irn);
737 _set_phi_class(copy, NULL);
738 set_nodes_block(copy, bb);
739 sched_put_after(pos, copy);
745 insert_remat_after(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
749 if(can_remat_after(si, remat, pos, live)) {
754 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
756 copy = insert_copy_after(si, remat->op, pos);
758 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
759 op = obstack_alloc(si->obst, sizeof(*op));
761 op->attr.remat.remat = remat;
762 op->attr.remat.pre = 0;
763 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
765 set_irn_link(copy, op);
766 pset_insert_ptr(si->all_possible_remats, copy);
768 proj_copy = insert_copy_after(si, remat->proj, copy);
769 set_irn_n(proj_copy, 0, copy);
770 set_irn_link(proj_copy, op);
771 pset_insert_ptr(si->all_possible_remats, proj_copy);
783 insert_remat_before(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
787 if(can_remat_before(si, remat, pos, live)) {
792 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
794 copy = insert_copy_before(si, remat->op, pos);
796 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
797 op = obstack_alloc(si->obst, sizeof(*op));
799 op->attr.remat.remat = remat;
800 op->attr.remat.pre = 1;
801 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
803 set_irn_link(copy, op);
804 pset_insert_ptr(si->all_possible_remats, copy);
806 proj_copy = insert_copy_after(si, remat->proj, copy);
807 set_irn_n(proj_copy, 0, copy);
808 set_irn_link(proj_copy, op);
809 pset_insert_ptr(si->all_possible_remats, proj_copy);
821 get_block_n_succs(const ir_node *block) {
822 const ir_edge_t *edge;
824 assert(edges_activated(current_ir_graph));
826 edge = get_block_succ_first(block);
830 edge = get_block_succ_next(block, edge);
835 is_merge_edge(const ir_node * bb)
837 #ifdef GOODWIN_REDUCTION
838 return get_block_n_succs(bb) == 1;
845 is_diverge_edge(const ir_node * bb)
847 #ifdef GOODWIN_REDUCTION
848 return get_Block_n_cfgpreds(bb) == 1;
855 walker_regclass_copy_insertor(ir_node * irn, void * data)
857 spill_ilp_t *si = data;
859 if(is_Phi(irn) && has_reg_class(si, irn)) {
862 for(n=get_irn_arity(irn)-1; n>=0; --n) {
863 ir_node *phi_arg = get_irn_n(irn, n);
864 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
866 if(!has_reg_class(si, phi_arg)) {
867 ir_node *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
868 ir_node *pos = sched_block_last_noncf(si, bb);
869 op_t *op = obstack_alloc(si->obst, sizeof(*op));
871 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
872 sched_add_after(pos, copy);
873 set_irn_n(irn, n, copy);
876 op->attr.live_range.args.reloads = NULL;
877 op->attr.live_range.ilp = ILP_UNDEF;
878 set_irn_link(copy, op);
886 * Insert (so far unused) remats into the irg to
887 * recompute the potential liveness of all values
890 walker_remat_insertor(ir_node * bb, void * data)
892 spill_ilp_t *si = data;
893 spill_bb_t *spill_bb;
897 pset *live = pset_new_ptr_default();
899 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
901 live_foreach(bb, li) {
902 ir_node *value = (ir_node *) li->irn;
904 /* add remats at end of block */
905 if (live_is_end(li) && has_reg_class(si, value)) {
906 pset_insert_ptr(live, value);
910 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
911 set_irn_link(bb, spill_bb);
913 irn = sched_last(bb);
914 while(!sched_is_end(irn)) {
921 next = sched_prev(irn);
923 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
925 if(is_Phi(irn) || is_Proj(irn)) {
928 if(has_reg_class(si, irn)) {
929 pset_remove_ptr(live, irn);
932 op = obstack_alloc(si->obst, sizeof(*op));
934 op->attr.live_range.args.reloads = NULL;
935 op->attr.live_range.ilp = ILP_UNDEF;
936 set_irn_link(irn, op);
942 op = obstack_alloc(si->obst, sizeof(*op));
944 op->attr.live_range.ilp = ILP_UNDEF;
945 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
946 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
947 set_irn_link(irn, op);
949 args = pset_new_ptr_default();
951 /* collect arguments of op */
952 for (n = get_irn_arity(irn)-1; n>=0; --n) {
953 ir_node *arg = get_irn_n(irn, n);
955 pset_insert_ptr(args, arg);
958 /* set args of op already live in epilog */
959 pset_foreach(args, arg) {
960 if(has_reg_class(si, arg)) {
961 pset_insert_ptr(live, arg);
964 /* delete defined value from live set */
965 if(has_reg_class(si, irn)) {
966 pset_remove_ptr(live, irn);
970 remat_args = pset_new_ptr_default();
972 /* insert all possible remats before irn */
973 pset_foreach(args, arg) {
974 remat_info_t *remat_info,
978 /* continue if the operand has the wrong reg class
980 if(!has_reg_class(si, arg))
985 query.remats_by_operand = NULL;
986 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
992 if(remat_info->remats) {
993 pset_foreach(remat_info->remats, remat) {
994 ir_node *remat_irn = NULL;
996 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
997 #ifdef REMAT_WHILE_LIVE
998 if(pset_find_ptr(live, remat->value)) {
999 remat_irn = insert_remat_before(si, remat, irn, live);
1002 remat_irn = insert_remat_before(si, remat, irn, live);
1005 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1006 ir_node *remat_arg = get_irn_n(remat_irn, n);
1008 if(!has_reg_class(si, remat_arg)) continue;
1010 pset_insert_ptr(remat_args, remat_arg);
1017 /* now we add remat args to op's args because they could also die at this op */
1018 pset_foreach(args,arg) {
1019 if(pset_find_ptr(remat_args, arg)) {
1020 pset_remove_ptr(remat_args, arg);
1023 pset_foreach(remat_args,arg) {
1024 pset_insert_ptr(args, arg);
1027 /* insert all possible remats after irn */
1028 pset_foreach(args, arg) {
1029 remat_info_t *remat_info,
1033 /* continue if the operand has the wrong reg class */
1034 if(!has_reg_class(si, arg))
1038 query.remats = NULL;
1039 query.remats_by_operand = NULL;
1040 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1046 /* do not place post remats after jumps */
1047 if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
1049 if(remat_info->remats_by_operand) {
1050 pset_foreach(remat_info->remats_by_operand, remat) {
1051 /* do not insert remats producing the same value as one of the operands */
1052 if(!pset_find_ptr(args, remat->value)) {
1053 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
1054 #ifdef REMAT_WHILE_LIVE
1055 if(pset_find_ptr(live, remat->value)) {
1056 insert_remat_after(si, remat, irn, live);
1059 insert_remat_after(si, remat, irn, live);
1066 del_pset(remat_args);
1071 live_foreach(bb, li) {
1072 ir_node *value = (ir_node *) li->irn;
1074 /* add remats at end if successor has multiple predecessors */
1075 if(is_merge_edge(bb)) {
1076 /* add remats at end of block */
1077 if (live_is_end(li) && has_reg_class(si, value)) {
1078 remat_info_t *remat_info,
1083 query.remats = NULL;
1084 query.remats_by_operand = NULL;
1085 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1087 if(remat_info && remat_info->remats) {
1088 pset_foreach(remat_info->remats, remat) {
1089 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1091 insert_remat_before(si, remat, bb, NULL);
1096 if(is_diverge_edge(bb)) {
1097 /* add remat2s at beginning of block */
1098 if ((live_is_in(li) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) {
1099 remat_info_t *remat_info,
1104 query.remats = NULL;
1105 query.remats_by_operand = NULL;
1106 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1108 if(remat_info && remat_info->remats) {
1109 pset_foreach(remat_info->remats, remat) {
1110 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at beginning of block %+F\n", remat->op, bb));
1112 /* put the remat here if all its args are available */
1113 insert_remat_after(si, remat, bb, NULL);
1123 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1126 luke_endwalker(ir_node * bb, void * data)
1128 spill_ilp_t *si = (spill_ilp_t*)data;
1135 spill_bb_t *spill_bb = get_irn_link(bb);
1138 live = pset_new_ptr_default();
1139 use_end = pset_new_ptr_default();
1141 live_foreach(bb, li) {
1142 irn = (ir_node *) li->irn;
1143 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1146 pset_insert_ptr(live, irn);
1147 op = get_irn_link(irn);
1148 assert(!op->is_remat);
1152 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1153 /* their reg_out must always be set */
1154 sched_foreach_reverse(bb, irn) {
1157 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1159 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1160 ir_node *irn_arg = get_irn_n(irn, n);
1162 if(has_reg_class(si, irn_arg)) {
1163 pset_insert_ptr(use_end, irn_arg);
1168 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1169 //cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1170 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1172 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1174 /* if this is a merge edge we can reload at the end of this block */
1175 if(is_merge_edge(bb)) {
1176 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1177 } else if(pset_count(use_end)){
1178 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1180 spill_bb->reloads = NULL;
1183 pset_foreach(live,irn) {
1189 /* handle values used by control flow nodes later separately */
1190 if(pset_find_ptr(use_end, irn)) continue;
1193 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1195 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1197 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1198 spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1199 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1201 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1202 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1204 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1205 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1207 if(is_merge_edge(bb)) {
1211 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1212 reload = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1213 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1215 /* reload <= mem_out */
1216 rel_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1217 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1218 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1221 spill->reg_in = ILP_UNDEF;
1222 spill->mem_in = ILP_UNDEF;
1225 pset_foreach(use_end,irn) {
1229 ilp_cst_t end_use_req,
1234 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1236 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1238 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1239 spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1240 /* if irn is used at the end of the block, then it is live anyway */
1241 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1243 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1244 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1246 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1247 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1249 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1250 reload = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1251 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1253 /* reload <= mem_out */
1254 rel_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1255 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1256 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1258 spill->reg_in = ILP_UNDEF;
1259 spill->mem_in = ILP_UNDEF;
1261 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1262 end_use_req = lpp_add_cst(si->lpp, buf, lpp_equal, 1);
1263 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1271 next_post_remat(const ir_node * irn)
1276 irn = sched_block_first_nonphi(irn);
1278 irn = sched_next_op(irn);
1281 if(sched_is_end(irn))
1284 op = (op_t*)get_irn_link(irn);
1285 if(op->is_remat && !op->attr.remat.pre) {
1294 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1300 ret = sched_block_last_noncf(si, irn);
1301 ret = sched_next(ret);
1302 ret = sched_prev_op(ret);
1304 ret = sched_prev_op(irn);
1307 if(sched_is_end(ret) || is_Phi(ret))
1310 op = (op_t*)get_irn_link(ret);
1311 if(op->is_remat && op->attr.remat.pre) {
1319 * Find a remat of value @p value in the epilog of @p pos
1322 find_post_remat(const ir_node * value, const ir_node * pos)
1324 while((pos = next_post_remat(pos)) != NULL) {
1327 op = get_irn_link(pos);
1328 assert(op->is_remat && !op->attr.remat.pre);
1330 if(op->attr.remat.remat->value == value)
1331 return (ir_node*)pos;
1334 const ir_edge_t *edge;
1335 foreach_out_edge(pos, edge) {
1336 ir_node *proj = get_edge_src_irn(edge);
1337 assert(is_Proj(proj));
1347 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1349 spill_bb_t *spill_bb = get_irn_link(bb);
1355 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1357 double spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(si, bb);
1359 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1361 spill->reg_out = ILP_UNDEF;
1362 spill->reg_in = ILP_UNDEF;
1363 spill->mem_in = ILP_UNDEF;
1365 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1366 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1368 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1369 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1376 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1381 live_foreach(bb, li) {
1382 irn = (ir_node *) li->irn;
1384 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1385 pset_insert_ptr(live, irn);
1389 irn = sched_last(bb);
1391 /* all values eaten by control flow operations are also live until the end of the block */
1392 sched_foreach_reverse(bb, irn) {
1395 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1397 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1398 ir_node *arg = get_irn_n(irn,i);
1400 if(has_reg_class(si, arg)) {
1401 pset_insert_ptr(live, arg);
1408 * Inserts ILP-constraints and variables for memory copying before the given position
1411 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1413 const ir_node *succ;
1414 const ir_edge_t *edge;
1415 spill_bb_t *spill_bb = get_irn_link(block);
1424 assert(edges_activated(current_ir_graph));
1426 edge = get_block_succ_first(block);
1432 edge = get_block_succ_next(block, edge);
1433 /* next block can only contain phis, if this is a merge edge */
1436 ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1437 copyreg = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1439 ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1440 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1442 pset_foreach(live, tmp) {
1445 op_t *op = get_irn_link(irn);
1446 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1448 spill = set_find_spill(spill_bb->ilp, tmp);
1451 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1453 lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1455 sched_foreach(succ, phi) {
1456 const ir_node *to_copy;
1458 spill_t *to_copy_spill;
1459 op_t *phi_op = get_irn_link(phi);
1460 ilp_var_t reload = ILP_UNDEF;
1463 if(!is_Phi(phi)) break;
1464 if(!has_reg_class(si, phi)) continue;
1466 to_copy = get_irn_n(phi, pos);
1468 to_copy_op = get_irn_link(to_copy);
1470 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1471 assert(to_copy_spill);
1473 if(spill_bb->reloads) {
1474 keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1477 reload = PTR_TO_INT(keyval->val);
1481 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N_%N", block, phi, to_copy);
1482 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1484 /* copy - reg_out - reload - remat - live_range <= 0 */
1485 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1486 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1487 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1488 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1489 foreach_pre_remat(si, block, tmp) {
1490 op_t *remat_op = get_irn_link(tmp);
1491 if(remat_op->attr.remat.remat->value == to_copy) {
1492 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1496 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N_%N", block, phi, to_copy);
1497 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1499 /* copy - reg_out - copyreg <= 0 */
1500 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1501 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1502 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1508 * Walk all irg blocks and emit this ILP
1511 luke_blockwalker(ir_node * bb, void * data)
1513 spill_ilp_t *si = (spill_ilp_t*)data;
1518 spill_bb_t *spill_bb = get_irn_link(bb);
1521 pset *defs = pset_new_ptr_default();
1524 live = pset_new_ptr_default();
1526 /****************************************
1527 * B A S I C B L O C K E N D
1528 ***************************************/
1531 /* init live values at end of block */
1532 get_live_end(si, bb, live);
1534 pset_foreach(live, irn) {
1536 ilp_var_t reload = ILP_UNDEF;
1538 spill = set_find_spill(spill_bb->ilp, irn);
1541 if(spill_bb->reloads) {
1542 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1545 reload = PTR_TO_INT(keyval->val);
1549 op = get_irn_link(irn);
1550 assert(!op->is_remat);
1552 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1553 op->attr.live_range.ilp = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1554 op->attr.live_range.op = bb;
1556 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1557 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1559 /* reg_out - reload - remat - live_range <= 0 */
1560 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1561 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1562 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1563 foreach_pre_remat(si, bb, tmp) {
1564 op_t *remat_op = get_irn_link(tmp);
1565 if(remat_op->attr.remat.remat->value == irn) {
1566 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1569 /* maybe we should also assure that reg_out >= live_range etc. */
1572 #ifndef NO_MEMCOPIES
1573 insert_mem_copy_position(si, live, bb);
1577 * start new live ranges for values used by remats at end of block
1578 * and assure the remat args are available
1580 foreach_pre_remat(si, bb, tmp) {
1581 op_t *remat_op = get_irn_link(tmp);
1584 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1585 ir_node *remat_arg = get_irn_n(tmp, n);
1586 op_t *arg_op = get_irn_link(remat_arg);
1589 if(!has_reg_class(si, remat_arg)) continue;
1591 /* if value is becoming live through use by remat */
1592 if(!pset_find_ptr(live, remat_arg)) {
1593 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1594 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1596 arg_op->attr.live_range.ilp = prev_lr;
1597 arg_op->attr.live_range.op = bb;
1599 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1601 pset_insert_ptr(live, remat_arg);
1602 add_to_spill_bb(si, bb, remat_arg);
1605 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1606 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1607 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1609 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1610 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1612 /* use reload placed for this argument */
1613 if(spill_bb->reloads) {
1614 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1617 ilp_var_t reload = PTR_TO_INT(keyval->val);
1619 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1624 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1629 /**************************************
1630 * B A S I C B L O C K B O D Y
1631 **************************************/
1633 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1639 ilp_cst_t check_pre,
1646 /* iterate only until first phi */
1650 op = get_irn_link(irn);
1652 if(op->is_remat) continue;
1653 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1655 /* collect defined values */
1656 if(has_reg_class(si, irn)) {
1657 pset_insert_ptr(defs, irn);
1661 if(is_Proj(irn)) continue;
1664 * init set of irn's arguments
1665 * and all possibly used values around this op
1666 * and values defined by post remats
1668 args = new_set(cmp_keyval, get_irn_arity(irn));
1669 used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1670 remat_defs = pset_new_ptr(pset_count(live));
1672 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1673 ir_node *irn_arg = get_irn_n(irn, n);
1674 if(has_reg_class(si, irn_arg)) {
1675 set_insert_keyval(args, irn_arg, (void*)n);
1676 pset_insert_ptr(used, irn_arg);
1679 foreach_post_remat(irn, tmp) {
1680 op_t *remat_op = get_irn_link(tmp);
1682 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1684 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1685 ir_node *remat_arg = get_irn_n(tmp, n);
1686 if(has_reg_class(si, remat_arg)) {
1687 pset_insert_ptr(used, remat_arg);
1691 foreach_pre_remat(si, irn, tmp) {
1692 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1693 ir_node *remat_arg = get_irn_n(tmp, n);
1694 if(has_reg_class(si, remat_arg)) {
1695 pset_insert_ptr(used, remat_arg);
1700 /**********************************
1701 * I N E P I L O G O F irn
1702 **********************************/
1704 /* ensure each dying value is used by only one post remat */
1705 pset_foreach(used, tmp) {
1706 ir_node *value = tmp;
1707 op_t *value_op = get_irn_link(value);
1712 foreach_post_remat(irn, remat) {
1713 op_t *remat_op = get_irn_link(remat);
1715 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1716 ir_node *remat_arg = get_irn_n(remat, n);
1718 /* if value is used by this remat add it to constraint */
1719 if(remat_arg == value) {
1721 /* sum remat2s <= 1 + n_remats*live_range */
1722 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1723 cst = lpp_add_cst(si->lpp, buf, lpp_less, 1.0);
1727 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1733 // value_op->attr.live_range.ilp != ILP_UNDEF
1734 if(pset_find_ptr(live, value) && cst != ILP_UNDEF) {
1735 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1739 /* ensure at least one value dies at post remat */
1740 foreach_post_remat(irn, tmp) {
1741 op_t *remat_op = get_irn_link(tmp);
1742 pset *remat_args = pset_new_ptr(get_irn_arity(tmp));
1745 for(n=get_irn_arity(tmp)-1; n>=0; --n) {
1746 remat_arg = get_irn_n(tmp, n);
1748 if(has_reg_class(si, remat_arg)) {
1750 /* does arg always die at this op? */
1751 if(!pset_find_ptr(live, remat_arg))
1752 goto skip_one_must_die;
1754 pset_insert_ptr(remat_args, remat_arg);
1758 /* remat + \sum live_range(remat_arg) <= |args| */
1759 ir_snprintf(buf, sizeof(buf), "one_must_die_%+F", tmp);
1760 cst = lpp_add_cst(si->lpp, buf, lpp_less, pset_count(remat_args));
1761 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1763 pset_foreach(remat_args, remat_arg) {
1764 op_t *arg_op = get_irn_link(remat_arg);
1766 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1770 del_pset(remat_args);
1773 /* new live ranges for values from L\U defined by post remats */
1774 pset_foreach(live, tmp) {
1775 ir_node *value = tmp;
1776 op_t *value_op = get_irn_link(value);
1778 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
1779 ilp_var_t prev_lr = ILP_UNDEF;
1782 if(pset_find_ptr(remat_defs, value)) {
1784 /* next_live_range <= prev_live_range + sum remat2s */
1785 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1786 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1788 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1789 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1791 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1792 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1794 foreach_post_remat(irn, remat) {
1795 op_t *remat_op = get_irn_link(remat);
1797 /* if value is being rematerialized by this remat */
1798 if(value == remat_op->attr.remat.remat->value) {
1799 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1803 value_op->attr.live_range.ilp = prev_lr;
1804 value_op->attr.live_range.op = irn;
1809 /* requirements for post remats and start live ranges from L/U' for values dying here */
1810 foreach_post_remat(irn, tmp) {
1811 op_t *remat_op = get_irn_link(tmp);
1814 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1815 ir_node *remat_arg = get_irn_n(tmp, n);
1816 op_t *arg_op = get_irn_link(remat_arg);
1818 if(!has_reg_class(si, remat_arg)) continue;
1820 /* only for values in L\U (TODO and D?), the others are handled with post_use */
1821 if(!pset_find_ptr(used, remat_arg)) {
1822 /* remat <= live_range(remat_arg) */
1823 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
1824 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1826 /* if value is becoming live through use by remat2 */
1827 if(!pset_find_ptr(live, remat_arg)) {
1830 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1831 lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1833 arg_op->attr.live_range.ilp = lr;
1834 arg_op->attr.live_range.op = irn;
1836 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
1838 pset_insert_ptr(live, remat_arg);
1839 add_to_spill_bb(si, bb, remat_arg);
1842 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1843 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1848 d = pset_count(defs);
1849 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
1851 /* count how many regs irn needs for arguments */
1852 u = set_count(args);
1855 /* check the register pressure in the epilog */
1856 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
1857 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
1858 check_post = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - d);
1860 /* add L\U' to check_post */
1861 pset_foreach(live, tmp) {
1862 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
1863 /* if a live value is not used by irn */
1864 tmp_op = get_irn_link(tmp);
1865 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
1869 /***********************************************************
1870 * 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
1871 **********************************************************/
1874 pset_foreach(used, tmp) {
1880 op_t *arg_op = get_irn_link(arg);
1883 spill = add_to_spill_bb(si, bb, arg);
1885 /* new live range for each used value */
1886 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
1887 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1889 /* the epilog stuff - including post_use, check_post, check_post_remat */
1890 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
1891 post_use = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1893 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
1895 /* arg is live throughout epilog if the next live_range is in a register */
1896 if(pset_find_ptr(live, arg)) {
1897 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
1899 /* post_use >= next_lr + remat */
1900 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1901 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1902 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
1903 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1907 /* if value is not an arg of op and not possibly defined by post remat
1908 * then it may only die and not become live
1910 if(!set_find_keyval(args, arg)) {
1911 /* post_use <= prev_lr */
1912 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
1913 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1914 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1915 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1917 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
1918 /* next_lr <= prev_lr */
1919 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
1920 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1921 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1922 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1928 /* forall post remat which use arg add a similar cst */
1929 foreach_post_remat(irn, remat) {
1932 for (n=get_irn_arity(remat)-1; n>=0; --n) {
1933 ir_node *remat_arg = get_irn_n(remat, n);
1934 op_t *remat_op = get_irn_link(remat);
1936 if(remat_arg == arg) {
1937 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
1939 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1940 cst = lpp_add_cst(si->lpp, buf, lpp_greater, 0.0);
1941 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1942 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1947 /* new live range begins for each used value */
1948 arg_op->attr.live_range.ilp = prev_lr;
1949 arg_op->attr.live_range.op = irn;
1951 /*if(!pset_find_ptr(live, arg)) {
1952 pset_insert_ptr(live, arg);
1953 add_to_spill_bb(si, bb, arg);
1955 pset_insert_ptr(live, arg);
1959 /* just to be sure */
1960 check_post = ILP_UNDEF;
1969 /* check the register pressure in the prolog */
1970 /* sum_{L\U} lr <= k - |U| */
1971 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
1972 check_pre = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - u);
1974 /* for the prolog remove defined values from the live set */
1975 pset_foreach(defs, tmp) {
1976 pset_remove_ptr(live, tmp);
1979 /***********************************************************
1980 * 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
1981 **********************************************************/
1984 set_foreach(args, keyval) {
1986 ir_node *arg = keyval->key;
1987 int i = PTR_TO_INT(keyval->val);
1988 op_t *arg_op = get_irn_link(arg);
1990 spill = set_find_spill(spill_bb->ilp, arg);
1993 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
1994 op->attr.live_range.args.reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1996 /* reload <= mem_out */
1997 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
1998 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1999 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2000 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
2002 /* requirement: arg must be in register for use */
2003 /* reload + remat + live_range == 1 */
2004 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
2005 cst = lpp_add_cst(si->lpp, buf, lpp_equal, 1.0);
2007 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
2008 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
2009 foreach_pre_remat(si, irn, tmp) {
2010 op_t *remat_op = get_irn_link(tmp);
2011 if(remat_op->attr.remat.remat->value == arg) {
2012 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2017 /* iterate over L\U */
2018 pset_foreach(live, tmp) {
2019 if(!set_find_keyval(args, tmp)) {
2020 /* if a live value is not used by irn */
2021 tmp_op = get_irn_link(tmp);
2022 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2027 /* requirements for remats */
2028 /* start new live ranges for values used by remats */
2029 foreach_pre_remat(si, irn, tmp) {
2030 op_t *remat_op = get_irn_link(tmp);
2033 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2034 ir_node *remat_arg = get_irn_n(tmp, n);
2035 op_t *arg_op = get_irn_link(remat_arg);
2038 if(!has_reg_class(si, remat_arg)) continue;
2040 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2041 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2042 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2044 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2045 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2047 /* if remat arg is also used by current op then we can use reload placed for this argument */
2048 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2049 int index = (int)keyval->val;
2051 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2059 /*************************
2060 * D O N E W I T H O P
2061 *************************/
2063 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2065 pset_foreach(live, tmp) {
2066 assert(has_reg_class(si, tmp));
2069 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2070 ir_node *arg = get_irn_n(irn, n);
2072 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2075 del_pset(remat_defs);
2079 defs = pset_new_ptr_default();
2084 /***************************************
2085 * B E G I N N I N G O F B L O C K
2086 ***************************************/
2089 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2090 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2092 pset_foreach(live, irn) {
2093 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2096 /* construct mem_outs for all values */
2098 set_foreach(spill_bb->ilp, spill) {
2099 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2100 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2102 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2103 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2105 if(pset_find_ptr(live, spill->irn)) {
2106 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2108 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2109 spill->mem_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2110 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2112 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2114 op_t *op = get_irn_link(spill->irn);
2116 /* do we have to copy a phi argument? */
2117 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2118 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2120 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2121 const ir_node *arg = get_irn_n(spill->irn, n);
2127 /* argument already done? */
2128 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2130 /* get sum of execution frequencies of blocks with the same phi argument */
2131 for(m=n; m>=0; --m) {
2132 const ir_node *arg2 = get_irn_n(spill->irn, m);
2135 freq += execution_frequency(si, get_Block_cfgpred_block(bb, m));
2139 /* copies are not for free */
2140 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2141 var = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE * freq);
2143 for(m=n; m>=0; --m) {
2144 const ir_node *arg2 = get_irn_n(spill->irn, m);
2147 op->attr.live_range.args.copies[m] = var;
2151 /* copy <= mem_in */
2152 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2153 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2154 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2155 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2162 /* L\U is empty at bb start */
2163 /* arg is live throughout epilog if it is reg_in into this block */
2165 /* check the register pressure at the beginning of the block
2168 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2169 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
2171 pset_foreach(live, irn) {
2174 spill = set_find_spill(spill_bb->ilp, irn);
2177 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2178 spill->reg_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2180 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2182 /* spill + mem_in <= 1 */
2183 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2184 nospill = lpp_add_cst(si->lpp, buf, lpp_less, 1);
2186 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2187 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2190 foreach_post_remat(bb, irn) {
2191 op_t *remat_op = get_irn_link(irn);
2193 DBG((si->dbg, LEVEL_4, "\t next post remat: %+F\n", irn));
2194 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
2196 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2199 /* forall post remats add requirements */
2200 foreach_post_remat(bb, tmp) {
2203 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2204 ir_node *remat_arg = get_irn_n(tmp, n);
2205 op_t *remat_op = get_irn_link(tmp);
2207 if(!has_reg_class(si, remat_arg)) continue;
2209 spill = set_find_spill(spill_bb->ilp, remat_arg);
2212 /* remat <= reg_in_argument */
2213 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2214 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2215 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2216 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2220 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2221 pset_foreach(live, irn) {
2225 spill = set_find_spill(spill_bb->ilp, irn);
2226 assert(spill && spill->irn == irn);
2228 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2229 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2232 ir_node *phi_arg = get_Phi_pred(irn, n);
2233 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2234 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2236 op_t *op = get_irn_link(irn);
2238 /* although the phi is in the right regclass one or more of
2239 * its arguments can be in a different one or at least to
2242 if(has_reg_class(si, phi_arg)) {
2243 /* mem_in < mem_out_arg + copy */
2244 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2245 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2247 /* reg_in < reg_out_arg */
2248 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2249 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2251 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2252 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2254 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2257 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2258 lpp_set_factor_fast(si->lpp, mem_in, op->attr.live_range.args.copies[n], -1.0);
2259 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2263 /* else assure the value arrives on all paths in the same resource */
2265 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2268 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2269 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2272 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2273 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2274 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2275 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2277 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2278 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2280 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2283 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2284 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2289 /* first live ranges from reg_ins */
2290 pset_foreach(live, irn) {
2291 op_t *op = get_irn_link(irn);
2293 spill = set_find_spill(spill_bb->ilp, irn);
2294 assert(spill && spill->irn == irn);
2296 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2297 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2298 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2299 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2301 foreach_post_remat(bb, tmp) {
2302 op_t *remat_op = get_irn_link(tmp);
2304 if(remat_op->attr.remat.remat->value == irn) {
2305 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2310 /* walk forward now and compute constraints for placing spills */
2311 /* this must only be done for values that are not defined in this block */
2312 /* TODO are these values at start of block? if yes, just check whether this is a diverge edge and skip the loop */
2313 pset_foreach(live, irn) {
2315 * if value is defined in this block we can anways place the spill directly after the def
2316 * -> no constraint necessary
2318 if(!is_Phi(irn) && get_nodes_block(irn) == bb) continue;
2321 spill = set_find_spill(spill_bb->ilp, irn);
2324 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2325 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2327 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2328 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2331 sched_foreach_op(bb, tmp) {
2332 op_t *op = get_irn_link(tmp);
2334 if(is_Phi(tmp)) continue;
2335 assert(!is_Proj(tmp));
2338 ir_node *value = op->attr.remat.remat->value;
2341 /* only collect remats up to the first real use of a value */
2342 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2347 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2348 ir_node *arg = get_irn_n(tmp, n);
2351 /* if a value is used stop collecting remats */
2364 typedef struct _irnlist_t {
2365 struct list_head list;
2369 typedef struct _interference_t {
2370 struct list_head blocklist;
2376 cmp_interference(const void *a, const void *b, size_t size)
2378 const interference_t *p = a;
2379 const interference_t *q = b;
2381 return !(p->a == q->a && p->b == q->b);
2384 static interference_t *
2385 set_find_interference(set * set, ir_node * a, ir_node * b)
2387 interference_t query;
2389 query.a = (a>b)?a:b;
2390 query.b = (a>b)?b:a;
2392 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2395 static interference_t *
2396 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2398 interference_t query,
2400 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2404 result = set_find_interference(set, a, b);
2407 list_add(&list->list, &result->blocklist);
2411 query.a = (a>b)?a:b;
2412 query.b = (a>b)?b:a;
2414 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2416 INIT_LIST_HEAD(&result->blocklist);
2417 list_add(&list->list, &result->blocklist);
2423 values_interfere_in_block(ir_node * bb, ir_node * a, ir_node * b)
2425 const ir_edge_t *edge;
2427 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2428 /* both values are live in, so they interfere */
2432 /* ensure a dominates b */
2433 if(value_dominates(b,a)) {
2439 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2442 /* the following code is stolen from bera.c */
2443 if(is_live_end(bb, a))
2446 foreach_out_edge(a, edge) {
2447 const ir_node *user = edge->src;
2448 if(get_nodes_block(user) == bb
2451 && value_dominates(b, user))
2459 * Walk all irg blocks and collect interfering values inside of phi classes
2462 luke_interferencewalker(ir_node * bb, void * data)
2464 spill_ilp_t *si = (spill_ilp_t*)data;
2468 live_foreach(bb, li1) {
2469 ir_node *a = (ir_node *) li1->irn;
2470 op_t *a_op = get_irn_link(a);
2472 if(a_op->is_remat) continue;
2474 /* a is only interesting if it is in my register class and if it is inside a phi class */
2475 if (has_reg_class(si, a) && get_phi_class(a)) {
2476 for(li2=li1->next; li2; li2 = li2->next) {
2477 ir_node *b = (ir_node *) li2->irn;
2478 op_t *b_op = get_irn_link(b);
2480 if(b_op->is_remat) continue;
2482 /* a and b are only interesting if they are in the same phi class */
2483 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2484 if(values_interfere_in_block(bb, a, b)) {
2485 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2486 set_insert_interference(si, si->interferences, a, b, bb);
2494 static unsigned int copy_path_id = 0;
2497 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2504 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2505 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2507 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2509 pset_foreach(copies, ptr) {
2510 copy = PTR_TO_INT(ptr);
2511 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2516 * @parameter copies contains a path of copies which lead us to irn
2517 * @parameter visited contains a set of nodes already visited on this path
2520 find_copy_path(spill_ilp_t * si, ir_node * irn, ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2523 op_t *op = get_irn_link(irn);
2524 pset *visited_users = pset_new_ptr_default();
2527 if(op->is_remat) return;
2529 pset_insert_ptr(visited, irn);
2533 pset *visited_operands = pset_new_ptr(get_irn_arity(irn));
2535 /* visit all operands */
2536 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2537 ir_node *arg = get_irn_n(irn, n);
2538 ilp_var_t copy = op->attr.live_range.args.copies[n];
2540 if(!has_reg_class(si, arg)) continue;
2541 if(pset_find_ptr(visited_operands, arg)) continue;
2542 pset_insert_ptr(visited_operands, arg);
2545 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2546 del_pset(visited_operands);
2547 del_pset(visited_users);
2548 pset_remove_ptr(visited, irn);
2551 pset_insert(copies, INT_TO_PTR(copy), copy);
2552 write_copy_path_cst(si, copies, any_interfere);
2553 pset_remove(copies, INT_TO_PTR(copy), copy);
2554 } else if(!pset_find_ptr(visited, arg)) {
2555 pset_insert(copies, INT_TO_PTR(copy), copy);
2556 paths += find_copy_path(si, arg, target, any_interfere, copies, visited);
2557 pset_remove(copies, INT_TO_PTR(copy), copy);
2559 /*if(paths > MAX_PATHS) {
2560 if(pset_count(copies) == 0) {
2564 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2565 cst = lpp_add_cst(si->lpp, buf, lpp_equal, 0);
2566 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2567 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2568 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2572 del_pset(visited_operands);
2573 del_pset(visited_users);
2574 pset_remove_ptr(visited, irn);
2577 } else if(pset_count(copies) == 0) {
2583 del_pset(visited_operands);
2586 /* visit all uses which are phis */
2587 foreach_out_edge(irn, edge) {
2588 ir_node *user = edge->src;
2589 int pos = edge->pos;
2590 op_t *op = get_irn_link(user);
2593 if(!is_Phi(user)) continue;
2594 if(!has_reg_class(si, user)) continue;
2595 if(pset_find_ptr(visited_users, user)) continue;
2596 pset_insert_ptr(visited_users, user);
2598 copy = op->attr.live_range.args.copies[pos];
2600 if(user == target) {
2601 if(++paths > MAX_PATHS && pset_count(copies) != 0) {
2602 del_pset(visited_users);
2603 pset_remove_ptr(visited, irn);
2606 pset_insert(copies, INT_TO_PTR(copy), copy);
2607 write_copy_path_cst(si, copies, any_interfere);
2608 pset_remove(copies, INT_TO_PTR(copy), copy);
2609 } else if(!pset_find_ptr(visited, user)) {
2610 pset_insert(copies, INT_TO_PTR(copy), copy);
2611 paths += find_copy_path(si, user, target, any_interfere, copies, visited);
2612 pset_remove(copies, INT_TO_PTR(copy), copy);
2614 /*if(paths > MAX_PATHS) {
2615 if(pset_count(copies) == 0) {
2619 ir_snprintf(buf, sizeof(buf), "always_copy-%d-%d", any_interfere, copy);
2620 cst = lpp_add_cst(si->lpp, buf, lpp_equal, 0);
2621 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2622 lpp_set_factor_fast(si->lpp, cst, copy, 1.0);
2623 DBG((si->dbg, LEVEL_1, "ALWAYS COPYING %d FOR INTERFERENCE %d\n", copy, any_interfere));
2627 del_pset(visited_users);
2628 pset_remove_ptr(visited, irn);
2631 } else if(pset_count(copies) == 0) {
2637 del_pset(visited_users);
2638 pset_remove_ptr(visited, irn);
2643 gen_copy_constraints(spill_ilp_t * si, ir_node * a, ir_node * b, ilp_var_t any_interfere)
2645 pset * copies = pset_new_ptr_default();
2646 pset * visited = pset_new_ptr_default();
2648 find_copy_path(si, a, b, any_interfere, copies, visited);
2656 memcopyhandler(spill_ilp_t * si)
2658 interference_t *interference;
2660 /* teste Speicherwerte auf Interferenz */
2662 /* analyze phi classes */
2663 phi_class_compute(si->chordal_env->irg);
2665 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
2666 irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
2668 /* now lets emit the ILP unequations for the crap */
2669 set_foreach(si->interferences, interference) {
2671 ilp_var_t interfere,
2673 ilp_cst_t any_interfere_cst,
2675 const ir_node *a = interference->a;
2676 const ir_node *b = interference->b;
2678 /* any_interf <= \sum interf */
2679 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
2680 any_interfere_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2681 any_interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2683 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
2685 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
2686 const ir_node *bb = irnlist->irn;
2687 spill_bb_t *spill_bb = get_irn_link(bb);
2694 spilla = set_find_spill(spill_bb->ilp, a);
2698 spillb = set_find_spill(spill_bb->ilp, b);
2701 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
2702 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
2703 /* 2: - mem_in_a - spill_a + interfere <= 0 */
2704 /* 3: - mem_in_b - spill_b + interfere <= 0 */
2705 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
2706 interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2708 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
2709 cst = lpp_add_cst(si->lpp, buf, lpp_less, 1);
2711 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
2712 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
2713 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
2714 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
2715 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
2717 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
2718 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2720 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2721 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
2722 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
2724 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
2725 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2727 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2728 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
2729 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
2732 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
2734 /* any_interfere >= interf */
2735 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
2736 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2738 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2739 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2742 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
2743 gen_copy_constraints(si,a,b,any_interfere);
2751 return fabs(x) < 0.00001;
2755 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
2757 spill_ilp_t *si = get_irg_link(current_ir_graph);
2759 if(pset_find_ptr(si->all_possible_remats, n)) {
2760 op_t *op = (op_t*)get_irn_link(n);
2761 assert(op && op->is_remat);
2763 if(!op->attr.remat.remat->inverse) {
2764 if(op->attr.remat.pre) {
2765 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
2767 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
2772 op_t *op = (op_t*)get_irn_link(n);
2773 assert(op && op->is_remat);
2775 if(op->attr.remat.pre) {
2776 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
2778 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
2789 dump_graph_with_remats(ir_graph * irg, const char * suffix)
2791 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
2792 be_dump(irg, suffix, dump_ir_block_graph_sched);
2793 set_dump_node_vcgattr_hook(NULL);
2798 * Edge hook to dump the schedule edges with annotated register pressure.
2801 sched_pressure_edge_hook(FILE *F, ir_node *irn)
2803 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
2804 ir_node *prev = sched_prev(irn);
2805 fprintf(F, "edge:{sourcename:\"");
2807 fprintf(F, "\" targetname:\"");
2809 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
2810 fprintf(F, "\" color:magenta}\n");
2816 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
2818 DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
2820 dump_consts_local(0);
2821 set_dump_node_edge_hook(sched_pressure_edge_hook);
2822 dump_ir_block_graph(irg, suffix);
2823 set_dump_node_edge_hook(old_edge_hook);
2827 walker_pressure_annotator(ir_node * bb, void * data)
2829 spill_ilp_t *si = data;
2833 pset *live = pset_new_ptr_default();
2836 live_foreach(bb, li) {
2837 irn = (ir_node *) li->irn;
2839 if (live_is_end(li) && has_reg_class(si, irn)) {
2840 pset_insert_ptr(live, irn);
2844 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
2846 sched_foreach_reverse(bb, irn) {
2848 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
2852 if(has_reg_class(si, irn)) {
2853 pset_remove_ptr(live, irn);
2854 if(is_Proj(irn)) ++projs;
2857 if(!is_Proj(irn)) projs = 0;
2859 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2860 ir_node *arg = get_irn_n(irn, n);
2862 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
2864 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
2871 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
2873 be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
2878 connect_all_remats_with_keep(spill_ilp_t * si)
2886 n_remats = pset_count(si->all_possible_remats);
2888 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
2891 pset_foreach(si->all_possible_remats, irn) {
2896 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
2898 obstack_free(si->obst, ins);
2904 connect_all_spills_with_keep(spill_ilp_t * si)
2913 n_spills = pset_count(si->spills);
2915 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
2918 pset_foreach(si->spills, irn) {
2923 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
2925 obstack_free(si->obst, ins);
2929 /** insert a spill at an arbitrary position */
2930 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert, ir_node *ctx)
2932 ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
2933 ir_graph *irg = get_irn_irg(bl);
2934 ir_node *frame = get_irg_frame(irg);
2938 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
2939 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
2941 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn, ctx);
2944 * search the right insertion point. a spill of a phi cannot be put
2945 * directly after the phi, if there are some phis behind the one which
2946 * is spilled. Also, a spill of a Proj must be after all Projs of the
2949 * Here's one special case:
2950 * If the spill is in the start block, the spill must be after the frame
2951 * pointer is set up. This is done by setting insert to the end of the block
2952 * which is its default initialization (see above).
2955 if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
2958 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
2961 sched_add_after(insert, spill);
2966 delete_remat(spill_ilp_t * si, ir_node * remat) {
2968 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2970 sched_remove(remat);
2972 /* kill links to operands */
2973 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
2974 set_irn_n(remat, n, bad);
2979 clean_remat_info(spill_ilp_t * si)
2983 remat_info_t *remat_info;
2984 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2986 set_foreach(si->remat_info, remat_info) {
2987 if(!remat_info->remats) continue;
2989 pset_foreach(remat_info->remats, remat)
2991 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
2992 set_irn_n(remat->proj, -1, bad);
2993 set_irn_n(remat->proj, 0, bad);
2996 if(get_irn_n_edges(remat->op) == 0) {
2997 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
2998 set_irn_n(remat->op, n, bad);
3003 if(remat_info->remats) del_pset(remat_info->remats);
3004 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
3009 delete_unnecessary_remats(spill_ilp_t * si)
3013 ir_node *bad = get_irg_bad(si->chordal_env->irg);
3016 ir_node *end = get_irg_end(si->chordal_env->irg);
3019 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
3020 ir_node *keep_arg = get_irn_n(si->keep, n);
3021 op_t *arg_op = get_irn_link(keep_arg);
3024 assert(arg_op->is_remat);
3026 name = si->lpp->vars[arg_op->attr.remat.ilp];
3028 if(is_zero(name->value)) {
3029 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
3030 /* TODO check whether reload is preferred over remat (could be bug) */
3031 delete_remat(si, keep_arg);
3033 if(!arg_op->attr.remat.remat->inverse) {
3034 if(arg_op->attr.remat.pre) {
3035 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
3037 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
3040 if(arg_op->attr.remat.pre) {
3041 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
3043 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
3048 set_irn_n(si->keep, n, bad);
3051 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
3052 ir_node *end_arg = get_End_keepalive(end, i);
3054 if(end_arg != si->keep) {
3055 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
3058 keeps = obstack_finish(si->obst);
3059 set_End_keepalives(end, n-1, keeps);
3060 obstack_free(si->obst, keeps);
3063 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
3068 pset_foreach(si->all_possible_remats, remat) {
3069 op_t *remat_op = get_irn_link(remat);
3070 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
3072 if(is_zero(name->value)) {
3073 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
3074 /* TODO check whether reload is preferred over remat (could be bug) */
3075 delete_remat(si, remat);
3077 if(!remat_op->attr.remat.remat->inverse) {
3078 if(remat_op->attr.remat.pre) {
3079 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3081 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3084 if(remat_op->attr.remat.pre) {
3085 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3087 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3096 get_spills_for_value(spill_ilp_t * si, ir_node * value)
3098 pset *spills = pset_new_ptr_default();
3103 defs = set_find_def(si->values, value);
3105 if(defs && defs->spills) {
3106 for(next = defs->spills; next; next = get_irn_link(next)) {
3107 pset_insert_ptr(spills, next);
3115 get_remats_for_value(spill_ilp_t * si, ir_node * value)
3117 pset *remats = pset_new_ptr_default();
3122 pset_insert_ptr(remats, value);
3123 defs = set_find_def(si->values, value);
3125 if(defs && defs->remats) {
3126 for(next = defs->remats; next; next = get_irn_link(next)) {
3127 pset_insert_ptr(remats, next);
3136 * @param before The node after which the spill will be placed in the schedule
3138 /* TODO set context properly */
3140 insert_spill(spill_ilp_t * si, ir_node * irn, ir_node * value, ir_node * before)
3144 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3146 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3148 spill = be_spill2(arch_env, irn, before, irn);
3150 defs = set_insert_def(si->values, value);
3153 /* enter into the linked list */
3154 set_irn_link(spill, defs->spills);
3155 defs->spills = spill;
3157 #ifdef KEEPALIVE_SPILLS
3158 pset_insert_ptr(si->spills, spill);
3165 * @param before The Phi node which has to be spilled
3168 insert_mem_phi(spill_ilp_t * si, const ir_node * phi)
3174 op_t *op = get_irn_link(phi);
3176 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3178 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3179 ins[n] = si->m_unknown;
3182 mem_phi = new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3184 defs = set_insert_def(si->values, phi);
3187 /* enter into the linked list */
3188 set_irn_link(mem_phi, defs->spills);
3189 defs->spills = mem_phi;
3191 sched_add_after(phi, mem_phi);
3193 #ifdef KEEPALIVE_SPILLS
3194 pset_insert_ptr(si->spills, mem_phi);
3202 * Add remat to list of defs, destroys link field!
3205 insert_remat(spill_ilp_t * si, ir_node * remat)
3208 op_t *remat_op = get_irn_link(remat);
3210 assert(remat_op->is_remat);
3212 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3215 /* enter into the linked list */
3216 set_irn_link(remat, defs->remats);
3217 defs->remats = remat;
3222 * Add reload before operation and add to list of defs
3225 insert_reload(spill_ilp_t * si, const ir_node * value, const ir_node * after)
3230 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3232 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3234 defs = set_find_def(si->values, value);
3236 spill = defs->spills;
3237 assert(spill && "no spill placed before reload");
3239 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3241 /* enter into the linked list */
3242 set_irn_link(reload, defs->remats);
3243 defs->remats = reload;
3249 walker_spill_placer(ir_node * bb, void * data) {
3250 spill_ilp_t *si = (spill_ilp_t*)data;
3252 spill_bb_t *spill_bb = get_irn_link(bb);
3253 pset *spills_to_do = pset_new_ptr_default();
3256 set_foreach(spill_bb->ilp, spill) {
3259 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3260 name = si->lpp->vars[spill->mem_in];
3261 if(!is_zero(name->value)) {
3264 mem_phi = insert_mem_phi(si, spill->irn);
3266 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3270 name = si->lpp->vars[spill->spill];
3271 if(!is_zero(name->value)) {
3272 /* place spill directly after definition */
3273 if(get_nodes_block(spill->irn) == bb) {
3274 insert_spill(si, spill->irn, spill->irn, spill->irn);
3278 /* place spill at bb start */
3279 if(spill->reg_in > 0) {
3280 name = si->lpp->vars[spill->reg_in];
3281 if(!is_zero(name->value)) {
3282 insert_spill(si, spill->irn, spill->irn, bb);
3286 /* place spill after a remat */
3287 pset_insert_ptr(spills_to_do, spill->irn);
3290 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3293 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3294 op_t *op = get_irn_link(irn);
3296 if(be_is_Spill(irn)) continue;
3299 /* TODO fix this if we want to support remats with more than two nodes */
3300 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3301 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3303 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3306 if(pset_find_ptr(spills_to_do, irn)) {
3307 pset_remove_ptr(spills_to_do, irn);
3309 insert_spill(si, irn, irn, irn);
3315 assert(pset_count(spills_to_do) == 0);
3317 /* afterwards free data in block */
3318 del_pset(spills_to_do);
3322 insert_mem_copy(spill_ilp_t * si, const ir_node * bb, const ir_node * value)
3324 ir_node *insert_pos = bb;
3326 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3328 /* find last definition of arg value in block */
3333 defs = set_find_def(si->values, value);
3335 if(defs && defs->remats) {
3336 for(next = defs->remats; next; next = get_irn_link(next)) {
3337 if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3338 last = sched_get_time_step(next);
3344 if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3345 last = sched_get_time_step(value);
3349 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
3351 spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos, value);
3357 phim_fixer(spill_ilp_t *si) {
3360 set_foreach(si->values, defs) {
3361 const ir_node *phi = defs->value;
3362 op_t *op = get_irn_link(phi);
3363 ir_node *phi_m = NULL;
3364 ir_node *next = defs->spills;
3367 if(!is_Phi(phi)) continue;
3370 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3374 next = get_irn_link(next);
3377 if(!phi_m) continue;
3379 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3380 const ir_node *value = get_irn_n(phi, n);
3381 defs_t *val_defs = set_find_def(si->values, value);
3382 ir_node *arg = get_irn_n(phi_m, n);
3384 /* a spill of this value */
3388 #ifndef NO_MEMCOPIES
3389 ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3390 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3392 if(!is_zero(name->value)) {
3393 spill = insert_mem_copy(si, pred, value);
3395 spill = val_defs->spills;
3398 spill = val_defs->spills;
3400 assert(spill && "no spill placed before PhiM");
3401 set_irn_n(phi_m, n, spill);
3407 walker_reload_placer(ir_node * bb, void * data) {
3408 spill_ilp_t *si = (spill_ilp_t*)data;
3410 spill_bb_t *spill_bb = get_irn_link(bb);
3414 /* reloads at end of block */
3415 if(spill_bb->reloads) {
3418 set_foreach(spill_bb->reloads, keyval) {
3419 ir_node *irn = (ir_node*)keyval->key;
3420 ilp_var_t reload = PTR_TO_INT(keyval->val);
3423 name = si->lpp->vars[reload];
3424 if(!is_zero(name->value)) {
3426 ir_node *insert_pos = bb;
3427 ir_node *prev = sched_block_last_noncf(si, bb);
3428 op_t *prev_op = get_irn_link(prev);
3430 while(be_is_Spill(prev)) {
3431 prev = sched_prev(prev);
3434 prev_op = get_irn_link(prev);
3436 /* insert reload before pre-remats */
3437 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3438 && prev_op->is_remat && prev_op->attr.remat.pre) {
3442 prev = sched_prev(prev);
3443 } while(be_is_Spill(prev));
3445 prev_op = get_irn_link(prev);
3449 reload = insert_reload(si, irn, insert_pos);
3451 #ifdef KEEPALIVE_RELOADS
3452 pset_insert_ptr(si->spills, reload);
3458 /* walk and insert more reloads and collect remats */
3459 sched_foreach_reverse(bb, irn) {
3460 op_t *op = get_irn_link(irn);
3462 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3463 if(is_Phi(irn)) break;
3466 if(get_irn_mode(irn) != mode_T) {
3467 insert_remat(si, irn);
3472 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3473 ir_node *arg = get_irn_n(irn, n);
3475 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3478 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3479 if(!is_zero(name->value)) {
3481 ir_node *insert_pos = irn;
3482 ir_node *prev = sched_prev(insert_pos);
3485 while(be_is_Spill(prev)) {
3486 prev = sched_prev(prev);
3489 prev_op = get_irn_link(prev);
3491 /* insert reload before pre-remats */
3492 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3493 && prev_op->is_remat && prev_op->attr.remat.pre) {
3497 prev = sched_prev(prev);
3498 } while(be_is_Spill(prev));
3500 prev_op = get_irn_link(prev);
3504 reload = insert_reload(si, arg, insert_pos);
3506 set_irn_n(irn, n, reload);
3508 #ifdef KEEPALIVE_RELOADS
3509 pset_insert_ptr(si->spills, reload);
3517 del_set(spill_bb->ilp);
3518 if(spill_bb->reloads) del_set(spill_bb->reloads);
3522 walker_collect_used(ir_node * irn, void * data)
3524 lc_bitset_t *used = data;
3526 lc_bitset_set(used, get_irn_idx(irn));
3529 struct kill_helper {
3535 walker_kill_unused(ir_node * bb, void * data)
3537 struct kill_helper *kh = data;
3538 const ir_node *bad = get_irg_bad(get_irn_irg(bb));
3542 for(irn=sched_first(bb); !sched_is_end(irn);) {
3543 ir_node *next = sched_next(irn);
3546 if(!lc_bitset_is_set(kh->used, get_irn_idx(irn))) {
3547 if(be_is_Spill(irn) || be_is_Reload(irn)) {
3548 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)));
3550 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3556 set_nodes_block(irn, bad);
3557 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3558 set_irn_n(irn, n, bad);
3566 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3568 struct kill_helper kh;
3570 kh.used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3573 irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3574 irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3576 lc_bitset_free(kh.used);
3580 print_irn_pset(pset * p)
3584 pset_foreach(p, irn) {
3585 ir_printf("%+F\n", irn);
3590 dump_phi_class(spill_ilp_t * si, pset * phiclass, const char * file)
3592 FILE *f = fopen(file, "w");
3594 interference_t *interference;
3596 pset_break(phiclass);
3597 set_break(si->interferences);
3599 ir_fprintf(f, "digraph phiclass {\n");
3601 pset_foreach(phiclass, irn) {
3603 ir_fprintf(f, " %F%N [shape=box]\n",irn,irn);
3606 pset_foreach(phiclass, irn) {
3609 if(!is_Phi(irn)) continue;
3611 for(n=get_irn_arity(irn)-1; n>=0; --n) {
3612 ir_node *arg = get_irn_n(irn, n);
3614 ir_fprintf(f, " %F%N -> %F%N\n",irn,irn,arg,arg);
3618 set_foreach(si->interferences, interference) {
3619 const ir_node *a = interference->a;
3620 const ir_node *b = interference->b;
3621 if(get_phi_class(a) == phiclass) {
3622 ir_fprintf(f, " %F%N -> %F%N [color=red,dir=none,style=bold]\n",a,a,b,b);
3631 rewire_uses(spill_ilp_t * si)
3633 dom_front_info_t *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
3635 pset *ignore = pset_new_ptr(1);
3637 pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
3639 /* then fix uses of spills */
3640 set_foreach(si->values, defs) {
3643 ir_node *next = defs->remats;
3646 reloads = pset_new_ptr_default();
3649 if(be_is_Reload(next)) {
3650 pset_insert_ptr(reloads, next);
3654 next = get_irn_link(next);
3657 spills = get_spills_for_value(si, defs->value);
3658 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));
3659 if(pset_count(spills) > 1) {
3660 //assert(pset_count(reloads) > 0);
3661 // print_irn_pset(spills);
3662 // print_irn_pset(reloads);
3664 be_ssa_constr_set_ignore(dfi, spills, ignore);
3671 /* first fix uses of remats and reloads */
3672 set_foreach(si->values, defs) {
3674 ir_node *next = defs->remats;
3677 nodes = pset_new_ptr_default();
3678 pset_insert_ptr(nodes, defs->value);
3681 pset_insert_ptr(nodes, next);
3682 next = get_irn_link(next);
3685 if(pset_count(nodes) > 1) {
3686 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
3687 be_ssa_constr_set(dfi, nodes);
3694 // remove_unused_defs(si);
3696 be_free_dominance_frontiers(dfi);
3701 writeback_results(spill_ilp_t * si)
3703 /* walk through the graph and collect all spills, reloads and remats for a value */
3705 si->values = new_set(cmp_defs, 4096);
3707 DBG((si->dbg, LEVEL_1, "Applying results\n"));
3708 delete_unnecessary_remats(si);
3709 si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
3710 irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
3711 irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
3714 /* clean the remat info! there are still back-edges leading there! */
3715 clean_remat_info(si);
3719 connect_all_spills_with_keep(si);
3721 del_set(si->values);
3725 get_n_regs(spill_ilp_t * si)
3727 int arch_n_regs = arch_register_class_n_regs(si->cls);
3731 for(i=0; i<arch_n_regs; i++) {
3732 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
3737 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
3742 walker_reload_mover(ir_node * bb, void * data)
3744 spill_ilp_t *si = data;
3747 sched_foreach(bb, tmp) {
3748 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
3749 ir_node *reload = tmp;
3752 /* move reload upwards */
3754 int pressure = (int)get_irn_link(reload);
3755 if(pressure < si->n_regs) {
3756 irn = sched_prev(reload);
3757 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
3758 sched_remove(reload);
3759 pressure = (int)get_irn_link(irn);
3761 while(pressure < si->n_regs) {
3762 if( sched_is_end(irn) ||
3763 (be_is_Reload(irn) && has_reg_class(si, irn)) ||
3764 /* do not move reload before its spill */
3765 (irn == be_get_Reload_mem(reload)) ) break;
3767 set_irn_link(irn, INT_TO_PTR(pressure+1));
3768 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
3769 irn = sched_prev(irn);
3771 pressure = (int)get_irn_link(irn);
3774 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
3775 sched_put_after(irn, reload);
3782 move_reloads_upward(spill_ilp_t * si)
3784 irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
3789 * Walk all irg blocks and check for interfering spills inside of phi classes
3792 luke_meminterferencechecker(ir_node * bb, void * data)
3794 spill_ilp_t *si = (spill_ilp_t*)data;
3798 live_foreach(bb, li1) {
3799 ir_node *a = (ir_node *) li1->irn;
3801 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
3803 /* a is only interesting if it is inside a phi class */
3804 if (get_phi_class(a)) {
3805 for(li2=li1->next; li2; li2 = li2->next) {
3806 ir_node *b = (ir_node *) li2->irn;
3808 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
3810 /* a and b are only interesting if they are in the same phi class */
3811 if(get_phi_class(a) == get_phi_class(b)) {
3812 if(values_interfere_in_block(bb, a, b)) {
3813 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
3822 verify_phiclasses(spill_ilp_t * si)
3824 /* analyze phi classes */
3825 phi_class_compute(si->chordal_env->irg);
3827 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
3828 irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
3832 walker_spillslotassigner(ir_node * irn, void * data)
3834 spill_ilp_t *si = (spill_ilp_t*)data;
3837 if(!be_is_Spill(irn)) return;
3839 /* set spill context to phi class if it has one ;) */
3841 cls = get_phi_class(irn);
3843 be_set_Spill_context(irn, cls);
3845 be_set_Spill_context(irn, irn);
3850 assign_spillslots(spill_ilp_t * si)
3852 DBG((si->dbg, LEVEL_2, "\t calling spill slot assigner\n"));
3853 irg_walk_graph(si->chordal_env->irg, walker_spillslotassigner, NULL, si);
3857 be_spill_remat(const be_chordal_env_t * chordal_env)
3859 char problem_name[256];
3860 char dump_suffix[256];
3861 char dump_suffix2[256];
3862 struct obstack obst;
3865 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
3866 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
3867 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
3869 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
3870 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
3872 #ifdef VERIFY_DOMINANCE
3873 be_check_dominance(chordal_env->irg);
3876 obstack_init(&obst);
3877 si.chordal_env = chordal_env;
3879 si.cls = chordal_env->cls;
3880 si.lpp = new_lpp(problem_name, lpp_minimize);
3881 si.remat_info = new_set(cmp_remat_info, 4096);
3882 si.interferences = new_set(cmp_interference, 32);
3883 si.all_possible_remats = pset_new_ptr_default();
3884 si.spills = pset_new_ptr_default();
3885 si.inverse_ops = pset_new_ptr_default();
3889 si.n_regs = get_n_regs(&si);
3891 set_irg_link(chordal_env->irg, &si);
3892 compute_doms(chordal_env->irg);
3894 /* compute phi classes */
3895 // phi_class_compute(chordal_env->irg);
3897 be_analyze_regpressure(chordal_env, "-pre");
3899 #ifdef COLLECT_REMATS
3900 /* collect remats */
3901 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
3902 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
3905 /* insert possible remats */
3906 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
3907 irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
3908 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
3911 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
3912 connect_all_remats_with_keep(&si);
3913 /* dump graph with inserted remats */
3914 dump_graph_with_remats(chordal_env->irg, dump_suffix);
3917 /* insert copies for phi arguments not in my regclass */
3918 irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
3920 /* recompute liveness */
3921 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
3922 be_liveness(chordal_env->irg);
3926 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
3927 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
3928 irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
3930 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
3931 irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
3933 #ifndef NO_MEMCOPIES
3934 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
3935 memcopyhandler(&si);
3943 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
3944 if ((f = fopen(buf, "wt")) != NULL) {
3945 lpp_dump_plain(si.lpp, f);
3952 DBG((si.dbg, LEVEL_1, "\tSolving %F\n", chordal_env->irg));
3954 lpp_set_time_limit(si.lpp, ILP_TIMEOUT);
3958 lpp_solve_cplex(si.lpp);
3960 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
3962 assert(lpp_is_sol_valid(si.lpp)
3963 && "solution of ILP must be valid");
3965 DBG((si.dbg, LEVEL_1, "\t%s: iterations: %d, solution time: %g, objective function: %g\n", problem_name, si.lpp->iterations, si.lpp->sol_time, is_zero(si.lpp->objval)?0.0:si.lpp->objval));
3967 #ifdef DUMP_SOLUTION
3972 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
3973 if ((f = fopen(buf, "wt")) != NULL) {
3975 for (i = 0; i < si.lpp->var_next; ++i) {
3976 lpp_name_t *name = si.lpp->vars[i];
3977 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
3984 writeback_results(&si);
3988 kill_all_unused_values_in_schedule(&si);
3990 #if defined(KEEPALIVE_SPILLS) || defined(KEEPALIVE_RELOADS)
3991 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
3994 // move reloads upwards
3995 be_liveness(chordal_env->irg);
3996 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
3997 move_reloads_upward(&si);
3999 #ifndef NO_MEMCOPIES
4000 verify_phiclasses(&si);
4001 assign_spillslots(&si);
4004 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
4006 dump_pressure_graph(&si, dump_suffix2);
4008 be_analyze_regpressure(chordal_env, "-post");
4010 #ifdef VERIFY_DOMINANCE
4011 be_check_dominance(chordal_env->irg);
4014 free_dom(chordal_env->irg);
4015 del_set(si.interferences);
4016 del_pset(si.inverse_ops);
4017 del_pset(si.all_possible_remats);
4018 del_pset(si.spills);
4020 obstack_free(&obst, NULL);
4021 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
4024 #else /* WITH_ILP */
4027 only_that_you_can_compile_without_WITH_ILP_defined(void)
4031 #endif /* WITH_ILP */