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
37 #include <lpp/lpp_net.h>
38 #include <lpp/lpp_cplex.h>
39 //#include <lc_pset.h>
40 #include <libcore/lc_bitset.h>
44 #include "besched_t.h"
49 #include "bespillremat.h"
51 #include "bepressurestat.h"
53 #include "bechordal_t.h"
59 //#define KEEPALIVE /* keep alive all inserted remats and dump graph with remats */
60 #define COLLECT_REMATS /* enable rematerialization */
61 #define COLLECT_INVERSE_REMATS /* enable placement of inverse remats */
62 #define REMAT_WHILE_LIVE /* only remat values that are live */
63 //#define NO_ENLARGE_L1V3N355 /* do not remat after the death of some operand */
64 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
65 //#define MAY_DIE_AT_PRE_REMAT /* allow values to die after a pre remat */
66 #define CHECK_POST_REMAT /* check pressure after post remats (conservative but otherwise we can temporarily exceed the register pressure) */
67 #define NO_SINGLE_USE_REMATS /* do not repair schedule */
68 //#define KEEPALIVE_SPILLS
69 //#define KEEPALIVE_RELOADS
70 #define GOODWIN_REDUCTION
74 #define LPP_SERVER "i44pc52"
75 #define LPP_SOLVER "cplex"
81 #define ILP_TIMEOUT 120
85 typedef struct _spill_ilp_t {
86 const arch_register_class_t *cls;
88 const be_chordal_env_t *chordal_env;
93 pset *all_possible_remats;
98 set *values; /**< for collecting all definitions of values before running ssa-construction */
102 DEBUG_ONLY(firm_dbg_module_t * dbg);
105 typedef int ilp_var_t;
106 typedef int ilp_cst_t;
108 typedef struct _spill_bb_t {
114 typedef struct _remat_t {
115 const ir_node *op; /**< for copy_irn */
116 const ir_node *proj; /**< not NULL if the above op produces a tuple */
117 const ir_node *value; /**< the value which is being recomputed by this remat */
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 */
143 typedef struct _defs_t {
145 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
146 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
149 typedef struct _remat_info_t {
150 const ir_node *irn; /**< the irn to which these remats belong */
151 pset *remats; /**< possible remats for this value */
152 pset *remats_by_operand; /**< remats with this value as operand */
155 typedef struct _keyval_t {
160 typedef struct _spill_t {
170 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
172 return chordal_has_class(si->chordal_env, irn);
177 cmp_remat(const void *a, const void *b)
179 const keyval_t *p = a;
180 const keyval_t *q = b;
181 const remat_t *r = p->val;
182 const remat_t *s = q->val;
186 return !(r == s || r->op == s->op);
190 cmp_remat(const void *a, const void *b)
192 const remat_t *r = a;
193 const remat_t *s = a;
195 return !(r == s || r->op == s->op);
199 cmp_spill(const void *a, const void *b, size_t size)
201 const spill_t *p = a;
202 const spill_t *q = b;
204 // return !(p->irn == q->irn && p->bb == q->bb);
205 return !(p->irn == q->irn);
209 set_find_keyval(set * set, void * key)
214 return set_find(set, &query, sizeof(query), HASH_PTR(key));
218 set_insert_keyval(set * set, void * key, void * val)
224 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
228 set_find_def(set * set, ir_node * value)
233 return set_find(set, &query, sizeof(query), HASH_PTR(value));
237 set_insert_def(set * set, ir_node * value)
244 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
248 set_find_spill(set * set, ir_node * value)
253 return set_find(set, &query, sizeof(query), HASH_PTR(value));
256 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
257 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
258 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
259 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
260 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
263 cmp_remat_info(const void *a, const void *b, size_t size)
265 const remat_info_t *p = a;
266 const remat_info_t *q = b;
268 return !(p->irn == q->irn);
272 cmp_defs(const void *a, const void *b, size_t size)
277 return !(p->value == q->value);
281 cmp_keyval(const void *a, const void *b, size_t size)
283 const keyval_t *p = a;
284 const keyval_t *q = b;
286 return !(p->key == q->key);
290 execution_frequency(const spill_ilp_t * si, const ir_node * irn)
295 return get_block_execfreq(si->execfreqs, irn) + FUDGE;
297 return get_block_execfreq(si->execfreqs, get_nodes_block(irn)) + FUDGE;
301 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
303 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
308 get_cost(const spill_ilp_t * si, const ir_node * irn)
310 if(be_is_Spill(irn)) {
312 } else if(be_is_Reload(irn)){
315 return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
321 * Checks, whether node and its operands have suitable reg classes
324 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
328 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
329 int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
333 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
336 for (i = 0, n = get_irn_arity(irn); i < n && remat; ++i) {
337 ir_node *op = get_irn_n(irn, i);
338 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
341 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
348 * Try to create a remat from @p op with destination value @p dest_value
350 static INLINE remat_t *
351 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
353 remat_t *remat = NULL;
355 // if(!mode_is_datab(get_irn_mode(dest_value)))
358 if(dest_value == op) {
359 const ir_node *proj = NULL;
361 if(is_Proj(dest_value)) {
362 op = get_irn_n(op, 0);
366 if(!is_rematerializable(si, op))
369 remat = obstack_alloc(si->obst, sizeof(*remat));
371 remat->cost = get_cost(si, op);
372 remat->value = dest_value;
376 arch_inverse_t inverse;
380 /* get the index of the operand we want to retrieve by the inverse op */
381 for (i = 0, n = get_irn_arity(op); i < n; ++i) {
382 ir_node *arg = get_irn_n(op, i);
384 if(arg == dest_value) break;
386 if(i == n) return NULL;
388 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", i, op));
390 /* else ask the backend to give an inverse op */
391 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, i, &inverse, si->obst)) {
394 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
396 assert(inverse.n > 0 && "inverse op should have at least one node");
398 for(i=0; i<inverse.n; ++i) {
399 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
403 remat = obstack_alloc(si->obst, sizeof(*remat));
404 remat->op = inverse.nodes[0];
405 remat->cost = inverse.costs;
406 remat->value = dest_value;
407 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
410 assert(is_Proj(remat->proj));
412 assert(0 && "I can not handle remats with more than 2 nodes");
419 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
421 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
429 add_remat(const spill_ilp_t * si, const remat_t * remat)
431 remat_info_t *remat_info,
437 assert(remat->value);
439 query.irn = remat->value;
441 query.remats_by_operand = NULL;
442 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
444 if(remat_info->remats == NULL) {
445 remat_info->remats = new_pset(cmp_remat, 4096);
447 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
449 /* insert the remat into the remats_be_operand set of each argument of the remat op */
450 for (i = 0, n = get_irn_arity(remat->op); i < n; ++i) {
451 ir_node *arg = get_irn_n(remat->op, i);
455 query.remats_by_operand = NULL;
456 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
458 if(remat_info->remats_by_operand == NULL) {
459 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
461 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
466 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
468 const ir_edge_t *edge = get_irn_out_edge_first(irn);
472 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
475 edge = get_irn_out_edge_next(irn, edge);
482 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
488 #ifdef NO_SINGLE_USE_REMATS
489 if(has_reg_class(si, op) && (get_irn_n_nonremat_edges(si, op) > 1)) {
491 if(has_reg_class(si, op)) {
493 remat = get_remat_from_op(si, op, op);
495 add_remat(si, remat);
499 #ifdef COLLECT_INVERSE_REMATS
500 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
502 for (i = 0, n = get_irn_arity(op); i < n; ++i) {
503 ir_node *arg = get_irn_n(op, i);
505 if(has_reg_class(si, arg)) {
506 /* try to get an inverse remat */
507 remat = get_remat_from_op(si, arg, op);
509 add_remat(si, remat);
518 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
521 ir_node *def_block = get_nodes_block(val);
527 /* if pos is at end of a basic block */
529 ret = (pos == def_block || block_dominates(def_block, pos));
530 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
534 /* else if this is a normal operation */
535 block = get_nodes_block(pos);
536 if(block == def_block) {
537 if(!sched_is_scheduled(val)) return 1;
539 ret = sched_comes_after(val, pos);
540 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
544 ret = block_dominates(def_block, block);
545 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
549 static INLINE ir_node *
550 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
552 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
556 * Returns first non-Phi node of block @p bb
558 static INLINE ir_node *
559 sched_block_first_nonphi(const ir_node * bb)
561 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
565 sched_skip_proj_predicator(const ir_node * irn, void * data)
567 return (is_Proj(irn));
570 static INLINE ir_node *
571 sched_next_nonproj(const ir_node * irn, int forward)
573 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
577 * Returns next operation node (non-Proj) after @p irn
578 * or the basic block of this node
580 static INLINE ir_node *
581 sched_next_op(const ir_node * irn)
583 ir_node *next = sched_next(irn);
588 return sched_next_nonproj(next, 1);
592 * Returns previous operation node (non-Proj) before @p irn
593 * or the basic block of this node
595 static INLINE ir_node *
596 sched_prev_op(const ir_node * irn)
598 ir_node *prev = sched_prev(irn);
603 return sched_next_nonproj(prev, 0);
607 sched_put_after(ir_node * insert, ir_node * irn)
609 if(is_Block(insert)) {
610 insert = sched_block_first_nonphi(insert);
612 insert = sched_next_op(insert);
614 sched_add_before(insert, irn);
618 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
620 if(is_Block(insert)) {
621 insert = sched_block_last_noncf(si, insert);
623 insert = sched_next_nonproj(insert, 0);
624 insert = sched_prev(insert);
626 sched_add_after(insert, irn);
630 * Tells you whether a @p remat can be placed before the irn @p pos
633 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
635 const ir_node *op = remat->op;
642 prev = sched_block_last_noncf(si, pos);
643 prev = sched_next_nonproj(prev, 0);
645 prev = sched_prev_op(pos);
647 /* do not remat if the rematted value is defined immediately before this op */
648 if(prev == remat->op) {
653 /* this should be just fine, the following OP will be using this value, right? */
655 /* only remat AFTER the real definition of a value (?) */
656 if(!value_is_defined_before(si, pos, remat->value)) {
657 // ir_fprintf(stderr, "error(not defined)");
662 for(i=0, n=get_irn_arity(op); i<n && res; ++i) {
663 const ir_node *arg = get_irn_n(op, i);
665 #ifdef NO_ENLARGE_L1V3N355
666 if(has_reg_class(si, arg) && live) {
667 res &= pset_find_ptr(live, arg)?1:0;
669 res &= value_is_defined_before(si, pos, arg);
672 res &= value_is_defined_before(si, pos, arg);
680 * Tells you whether a @p remat can be placed after the irn @p pos
683 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
686 pos = sched_block_first_nonphi(pos);
688 pos = sched_next_op(pos);
691 /* only remat AFTER the real definition of a value (?) */
692 if(!value_is_defined_before(si, pos, remat->value)) {
696 return can_remat_before(si, remat, pos, live);
700 * Collect potetially rematerializable OPs
703 walker_remat_collector(ir_node * irn, void * data)
705 spill_ilp_t *si = data;
707 if(!is_Block(irn) && !is_Phi(irn)) {
708 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
709 get_remats_from_op(si, irn);
714 * Inserts a copy of @p irn before @p pos
717 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
722 bb = is_Block(pos)?pos:get_nodes_block(pos);
723 copy = exact_copy(irn);
724 set_nodes_block(copy, bb);
725 sched_put_before(si, pos, copy);
731 * Inserts a copy of @p irn after @p pos
734 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
739 bb = is_Block(pos)?pos:get_nodes_block(pos);
740 copy = exact_copy(irn);
741 set_nodes_block(copy, bb);
742 sched_put_after(pos, copy);
748 insert_remat_after(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
752 if(can_remat_after(si, remat, pos, live)) {
757 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
759 copy = insert_copy_after(si, remat->op, pos);
761 // ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", remat->value, pos);
762 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
763 op = obstack_alloc(si->obst, sizeof(*op));
765 op->attr.remat.remat = remat;
766 op->attr.remat.pre = 0;
767 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
769 set_irn_link(copy, op);
770 pset_insert_ptr(si->all_possible_remats, copy);
772 proj_copy = insert_copy_after(si, remat->proj, copy);
773 set_irn_n(proj_copy, 0, copy);
774 set_irn_link(proj_copy, op);
775 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", remat->value, pos);
797 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
798 op = obstack_alloc(si->obst, sizeof(*op));
800 op->attr.remat.remat = remat;
801 op->attr.remat.pre = 1;
802 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
804 set_irn_link(copy, op);
805 pset_insert_ptr(si->all_possible_remats, copy);
807 proj_copy = insert_copy_after(si, remat->proj, copy);
808 set_irn_n(proj_copy, 0, copy);
809 set_irn_link(proj_copy, op);
810 pset_insert_ptr(si->all_possible_remats, proj_copy);
818 get_block_n_succs(const ir_node *block) {
819 const ir_edge_t *edge;
821 assert(edges_activated(current_ir_graph));
823 edge = get_block_succ_first(block);
827 edge = get_block_succ_next(block, edge);
832 is_merge_edge(const ir_node * bb)
834 #ifdef GOODWIN_REDUCTION
835 return get_block_n_succs(bb) == 1;
842 is_diverge_edge(const ir_node * bb)
844 #ifdef GOODWIN_REDUCTION
845 return get_Block_n_cfgpreds(bb) == 1;
852 * Insert (so far unused) remats into the irg to
853 * recompute the potential liveness of all values
856 walker_remat_insertor(ir_node * bb, void * data)
858 spill_ilp_t *si = data;
859 spill_bb_t *spill_bb;
864 pset *live = pset_new_ptr_default();
866 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
868 live_foreach(bb, li) {
869 ir_node *value = (ir_node *) li->irn;
871 /* add remats at end of block */
872 if (live_is_end(li) && has_reg_class(si, value)) {
873 pset_insert_ptr(live, value);
877 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
878 set_irn_link(bb, spill_bb);
880 irn = sched_last(bb);
881 while(!sched_is_end(irn)) {
887 next = sched_prev(irn);
889 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
891 if(is_Phi(irn) || is_Proj(irn)) {
894 if(has_reg_class(si, irn)) {
895 pset_remove_ptr(live, irn);
898 op = obstack_alloc(si->obst, sizeof(*op));
900 op->attr.live_range.reloads = NULL;
901 op->attr.live_range.ilp = ILP_UNDEF;
902 set_irn_link(irn, op);
908 op = obstack_alloc(si->obst, sizeof(*op));
910 op->attr.live_range.ilp = ILP_UNDEF;
911 op->attr.live_range.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.reloads) * get_irn_arity(irn));
912 memset(op->attr.live_range.reloads, 0xFF, sizeof(*op->attr.live_range.reloads) * get_irn_arity(irn));
913 set_irn_link(irn, op);
915 args = pset_new_ptr_default();
917 /* collect arguments of op */
918 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
919 ir_node *arg = get_irn_n(irn, i);
921 pset_insert_ptr(args, arg);
924 /* set args of op live in epilog */
925 pset_foreach(args, arg) {
926 if(has_reg_class(si, arg)) {
927 pset_insert_ptr(live, arg);
931 /* insert all possible remats after irn */
932 pset_foreach(args, arg) {
933 remat_info_t *remat_info,
937 /* continue if the operand has the wrong reg class
939 if(!has_reg_class(si, arg))
944 query.remats_by_operand = NULL;
945 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
951 /* do not place post remats after jumps */
952 if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
954 if(remat_info->remats_by_operand) {
955 pset_foreach(remat_info->remats_by_operand, remat) {
956 /* do not insert remats producing the same value as one of the operands */
957 if(!pset_find_ptr(args, remat->value)) {
958 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
959 #ifdef REMAT_WHILE_LIVE
960 if(pset_find_ptr(live, remat->value)) {
961 insert_remat_after(si, remat, irn, live);
964 insert_remat_after(si, remat, irn, live);
971 /* delete defined value from live set */
972 if(has_reg_class(si, irn)) {
973 pset_remove_ptr(live, irn);
976 /* insert all possible remats before irn */
977 pset_foreach(args, arg) {
978 remat_info_t *remat_info,
982 /* continue if the operand has the wrong reg class
984 if(!has_reg_class(si, arg))
989 query.remats_by_operand = NULL;
990 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
996 if(remat_info->remats) {
997 pset_foreach(remat_info->remats, remat) {
998 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
999 #ifdef REMAT_WHILE_LIVE
1000 if(pset_find_ptr(live, remat->value)) {
1001 insert_remat_before(si, remat, irn, live);
1004 insert_remat_before(si, remat, irn, live);
1014 live_foreach(bb, li) {
1015 ir_node *value = (ir_node *) li->irn;
1017 /* add remats at end if successor has multiple predecessors */
1018 if(is_merge_edge(bb)) {
1019 /* add remats at end of block */
1020 if (live_is_end(li) && has_reg_class(si, value)) {
1021 remat_info_t *remat_info,
1026 query.remats = NULL;
1027 query.remats_by_operand = NULL;
1028 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1030 if(remat_info && remat_info->remats) {
1031 pset_foreach(remat_info->remats, remat) {
1032 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1034 insert_remat_before(si, remat, bb, NULL);
1039 if(is_diverge_edge(bb)) {
1040 /* add remat2s at beginning of block */
1041 if ((live_is_in(li) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) {
1042 remat_info_t *remat_info,
1047 query.remats = NULL;
1048 query.remats_by_operand = NULL;
1049 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1051 if(remat_info && remat_info->remats) {
1052 pset_foreach(remat_info->remats, remat) {
1053 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at beginning of block %+F\n", remat->op, bb));
1055 /* put the remat here if all its args are available */
1056 insert_remat_after(si, remat, bb, NULL);
1066 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1069 luke_endwalker(ir_node * bb, void * data)
1071 spill_ilp_t *si = (spill_ilp_t*)data;
1078 spill_bb_t *spill_bb = get_irn_link(bb);
1081 live = pset_new_ptr_default();
1082 use_end = pset_new_ptr_default();
1084 live_foreach(bb, li) {
1085 irn = (ir_node *) li->irn;
1086 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1089 pset_insert_ptr(live, irn);
1090 op = get_irn_link(irn);
1091 assert(!op->is_remat);
1095 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1096 /* their reg_out is unimportant because it can always be set */
1097 sched_foreach_reverse(bb, irn) {
1101 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1103 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
1104 ir_node *irn_arg = get_irn_n(irn, i);
1105 if(has_reg_class(si, irn_arg)) {
1106 pset_insert_ptr(use_end, irn);
1111 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1112 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1114 spill_bb->ilp = new_set(cmp_spill, 16);
1116 live_foreach(bb, li) {
1117 irn = (ir_node *) li->irn;
1118 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1123 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1125 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1126 spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1127 /* if irn is used at the end of the block, then it is live anyway */
1128 if(!pset_find_ptr(use_end, irn))
1129 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1131 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1132 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1134 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1135 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE*execution_frequency(si, bb));
1137 spill->reg_in = ILP_UNDEF;
1138 spill->mem_in = ILP_UNDEF;
1147 next_post_remat(const ir_node * irn)
1152 irn = sched_block_first_nonphi(irn);
1154 irn = sched_next_op(irn);
1157 if(sched_is_end(irn))
1160 op = (op_t*)get_irn_link(irn);
1161 if(op->is_remat && !op->attr.remat.pre) {
1170 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1176 ret = sched_block_last_noncf(si, irn);
1177 ret = sched_next(ret);
1178 ret = sched_prev_op(ret);
1180 ret = sched_prev_op(irn);
1183 if(sched_is_end(ret) || is_Phi(ret))
1186 op = (op_t*)get_irn_link(ret);
1187 if(op->is_remat && op->attr.remat.pre) {
1195 * Find a remat of value @p value in the epilog of @p pos
1198 find_post_remat(const ir_node * value, const ir_node * pos)
1200 while((pos = next_post_remat(pos)) != NULL) {
1203 op = get_irn_link(pos);
1204 assert(op->is_remat && !op->attr.remat.pre);
1206 if(op->attr.remat.remat->value == value)
1207 return (ir_node*)pos;
1210 const ir_edge_t *edge;
1211 foreach_out_edge(pos, edge) {
1212 ir_node *proj = get_edge_src_irn(edge);
1213 assert(is_Proj(proj));
1223 * Find a remat of value @p value in the prolog of @p pos
1226 find_pre_remat(const spill_ilp_t * si, const ir_node * value, const ir_node * pos)
1228 while((pos = next_pre_remat(si,pos)) != NULL) {
1231 op = get_irn_link(pos);
1232 assert(op->is_remat && op->attr.remat.pre);
1234 if(op->attr.remat.remat->value == value)
1235 return (ir_node*)pos;
1242 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1244 spill_bb_t *spill_bb = get_irn_link(bb);
1250 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1252 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1254 spill->reg_out = ILP_UNDEF;
1255 spill->reg_in = ILP_UNDEF;
1256 spill->mem_in = ILP_UNDEF;
1258 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1259 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1261 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1262 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE*execution_frequency(si, bb));
1269 * Walk all irg blocks and emit this ILP
1272 luke_blockwalker(ir_node * bb, void * data)
1274 spill_ilp_t *si = (spill_ilp_t*)data;
1280 spill_bb_t *spill_bb = get_irn_link(bb);
1286 live = pset_new_ptr_default();
1288 /* do something at the end of the block */
1290 /* init live values at end of block */
1291 live_foreach(bb, li) {
1292 ir_node *irn = (ir_node *) li->irn;
1294 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1295 pset_insert_ptr(live, irn);
1299 if(is_merge_edge(bb)) {
1300 spill_bb->reloads = obstack_alloc(si->obst, pset_count(live) * sizeof(*spill_bb->reloads));
1301 memset(spill_bb->reloads, 0xFF, pset_count(live) * sizeof(*spill_bb->reloads));
1303 spill_bb->reloads = NULL;
1307 live_foreach(bb, li) {
1308 ir_node *irn = (ir_node *) li->irn;
1311 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1312 spill = set_find_spill(spill_bb->ilp, irn);
1315 if(spill_bb->reloads) {
1316 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1317 spill_bb->reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1319 /* reload <= mem_out */
1320 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1321 lpp_set_factor_fast(si->lpp, cst, spill_bb->reloads[i], 1.0);
1322 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
1325 op = get_irn_link(irn);
1326 assert(!op->is_remat);
1328 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1329 op->attr.live_range.ilp = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1330 op->attr.live_range.op = bb;
1332 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1333 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1335 /* reg_out - reload - remat - live_range <= 0 */
1336 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1337 if(spill_bb->reloads) lpp_set_factor_fast(si->lpp, cst, spill_bb->reloads[i], -1.0);
1338 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1339 foreach_pre_remat(si, bb, tmp) {
1340 op_t *remat_op = get_irn_link(tmp);
1341 if(remat_op->attr.remat.remat->value == irn) {
1342 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1349 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1351 sched_foreach_reverse(bb, irn) {
1357 ilp_cst_t check_pre,
1359 #ifdef CHECK_POST_REMAT
1360 ilp_cst_t check_post_remat;
1362 set *args = new_set(cmp_keyval, get_irn_arity(irn));
1368 op = get_irn_link(irn);
1370 if(op->is_remat) continue;
1371 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1373 if(has_reg_class(si, irn)) {
1374 assert(pset_find_ptr(live, irn));
1375 pset_remove_ptr(live, irn);
1378 /* init set of irn's arguments */
1379 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
1380 ir_node *irn_arg = get_irn_n(irn, i);
1381 if(has_reg_class(si, irn_arg)) {
1382 set_insert_keyval(args, irn_arg, (void*)i);
1386 #ifdef CHECK_POST_REMAT
1387 /* check the register pressure after the epilog */
1388 ir_snprintf(buf, sizeof(buf), "check_post_remat_%N", irn);
1389 check_post_remat = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1391 /* iterate over L\U */
1392 pset_foreach(live, tmp) {
1393 if(!set_find_keyval(args, tmp)) {
1394 /* if a live value is not used by irn */
1395 tmp_op = get_irn_link(tmp);
1396 // assert(tmp_op->attr.live_range.op != irn);
1397 lpp_set_factor_fast(si->lpp, check_post_remat, tmp_op->attr.live_range.ilp, 1.0);
1400 /* iterate over following remats and remove possibly defined values again from check_post_remat */
1401 foreach_post_remat(irn, tmp) {
1402 op_t *remat_op = get_irn_link(tmp);
1403 const ir_node *value = remat_op->attr.remat.remat->value;
1404 op_t *val_op = get_irn_link(value);
1406 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
1408 /* values that are defined by remat2s are not counted */
1409 #ifdef REMAT_WHILE_LIVE
1410 assert(val_op->attr.live_range.ilp);
1411 lpp_set_factor_fast(si->lpp, check_post_remat, val_op->attr.live_range.ilp, 0.0);
1413 if(val_op->attr.live_range.ilp != ILP_UNDEF) {
1414 lpp_set_factor_fast(si->lpp, check_post_remat, val_op->attr.live_range.ilp, 0.0);
1416 #endif /* REMAT_WHILE_LIVE */
1418 #endif /* CHECK_POST_REMAT */
1421 /* new live ranges for values from L\U defined by remat2s or used by remats */
1422 pset_foreach(live, tmp) {
1423 ir_node *value = tmp;//remat_op->attr.remat.remat->value;
1424 op_t *value_op = get_irn_link(value);
1426 if(!set_find_keyval(args, value)) {
1427 ilp_var_t prev_lr = ILP_UNDEF;
1431 foreach_post_remat(irn, remat) {
1432 op_t *remat_op = get_irn_link(remat);
1434 /* if value is being rematerialized by this remat */
1435 if(value == remat_op->attr.remat.remat->value) {
1436 if(cst == ILP_UNDEF) {
1437 /* next_live_range <= prev_live_range + sum remat2s */
1438 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1439 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1440 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1441 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1442 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1443 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1446 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1450 #ifdef MAY_DIE_AT_PRE_REMAT
1451 if(cst == ILP_UNDEF) {
1452 foreach_pre_remat(si, irn, remat) {
1456 for (i = 0, n = get_irn_arity(remat); i < n; ++i) {
1457 ir_node *remat_arg = get_irn_n(remat, i);
1459 /* if value is being used by this remat */
1460 if(value == remat_arg) {
1461 /* next_live_range <= prev_live_range */
1462 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1463 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1465 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1466 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1467 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1468 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1471 /* TODO check afterwards whether lr dies after a pre-remat (should not happen) */
1478 if(prev_lr != ILP_UNDEF) {
1479 value_op->attr.live_range.ilp = prev_lr;
1480 value_op->attr.live_range.op = irn;
1485 /* get count of values in my register class defined by irn */
1486 /* also add defined values to check_post_remat; do this before iterating over args */
1487 if(get_irn_mode(irn) == mode_T) {
1488 ir_node *proj = sched_next(irn);
1489 op_t *proj_op = get_irn_link(proj);
1491 while(is_Proj(proj)) {
1492 if(has_reg_class(si, proj)) {
1494 #ifdef CHECK_POST_REMAT
1495 lpp_set_factor_fast(si->lpp, check_post_remat, proj_op->attr.live_range.ilp, 1.0);
1498 proj = sched_next(proj);
1499 proj_op = get_irn_link(proj);
1502 if(has_reg_class(si, irn)) {
1504 #ifdef CHECK_POST_REMAT
1505 lpp_set_factor_fast(si->lpp, check_post_remat, op->attr.live_range.ilp, 1.0);
1509 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
1511 /* count how many regs irn needs for arguments */
1512 k = set_count(args);
1514 /* check the register pressure in the prolog */
1515 /* sum_{L\U} lr <= n - |U| */
1516 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
1517 check_pre = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - k);
1519 /* check the register pressure in the epilog */
1520 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
1521 check_post = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - d);
1523 set_foreach(args, keyval) {
1529 ir_node *arg = keyval->key;
1531 spill = add_to_spill_bb(si, bb, arg);
1533 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
1534 next_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1536 i = (int)keyval->val;
1539 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
1540 op->attr.live_range.reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1542 /* reload <= mem_out */
1543 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1544 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.reloads[i], 1.0);
1545 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
1547 arg_op = get_irn_link(arg);
1549 /* requirement: arg must be in register for use */
1550 /* reload + remat + live_range == 1 */
1551 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
1552 cst = lpp_add_cst(si->lpp, buf, lpp_equal, 1.0);
1554 lpp_set_factor_fast(si->lpp, cst, next_lr, 1.0);
1555 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.reloads[i], 1.0);
1556 foreach_pre_remat(si, irn, tmp) {
1557 op_t *remat_op = get_irn_link(tmp);
1558 if(remat_op->attr.remat.remat->value == arg) {
1559 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1563 /* the epilog stuff - including post_use, post, post_remat */
1564 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
1565 post_use = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1567 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
1569 /* arg is live throughout epilog if the next live_range is in a register */
1570 if(pset_find_ptr(live, arg)) {
1571 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
1573 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1574 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1575 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
1576 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1578 #ifdef CHECK_POST_REMAT
1579 lpp_set_factor_fast(si->lpp, check_post_remat, arg_op->attr.live_range.ilp, 1.0);
1583 /*forall remat2 which use arg add a similar cst*/
1584 foreach_post_remat(irn, tmp) {
1588 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1589 ir_node *remat_arg = get_irn_n(tmp, i);
1590 op_t *remat_op = get_irn_link(tmp);
1592 if(remat_arg == arg) {
1593 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
1595 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1596 cst = lpp_add_cst(si->lpp, buf, lpp_greater, 0.0);
1597 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1598 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1603 /* new live range begins for each argument */
1604 arg_op->attr.live_range.ilp = next_lr;
1605 arg_op->attr.live_range.op = irn;
1607 pset_insert_ptr(live, arg);
1610 /* start new live ranges for values used by remats */
1611 foreach_pre_remat(si, irn, tmp) {
1615 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1616 ir_node *remat_arg = get_irn_n(tmp, i);
1617 op_t *arg_op = get_irn_link(remat_arg);
1620 if(!has_reg_class(si, remat_arg)) continue;
1622 /* if value is becoming live through use by remat */
1623 if(!pset_find_ptr(live, remat_arg)) {
1624 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1625 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1627 arg_op->attr.live_range.ilp = prev_lr;
1628 arg_op->attr.live_range.op = irn;
1630 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat %+F\n", remat_arg, tmp));
1632 /* TODO ist das hier die richtige Stelle???? */
1633 pset_insert_ptr(live, remat_arg);
1634 add_to_spill_bb(si, bb, remat_arg);
1636 /* TODO check afterwards whether lr dies after a pre-remat (should not happen) */
1640 /* iterate over L\U */
1641 pset_foreach(live, tmp) {
1642 if(!set_find_keyval(args, tmp)) {
1643 /* if a live value is not used by irn */
1644 tmp_op = get_irn_link(tmp);
1645 // assert(tmp_op->attr.live_range.op != irn);
1646 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
1647 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
1651 /* requirements for remats */
1652 foreach_pre_remat(si, irn, tmp) {
1653 op_t *remat_op = get_irn_link(tmp);
1657 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1658 ir_node *remat_arg = get_irn_n(tmp, i);
1659 op_t *arg_op = get_irn_link(remat_arg);
1661 if(!has_reg_class(si, remat_arg)) continue;
1663 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1664 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1665 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1667 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1668 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1670 /* if remat arg is also used by current op then we can use reload placed for this argument */
1671 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
1672 int index = (int)keyval->val;
1674 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.reloads[index], -1.0);
1679 /* requirements for remats2
1681 * TODO unsure if this does the right thing.
1682 * should insert values into set if they do not become live through remat and
1685 foreach_post_remat(irn, tmp) {
1686 op_t *remat_op = get_irn_link(tmp);
1690 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1691 ir_node *remat_arg = get_irn_n(tmp, i);
1692 op_t *arg_op = get_irn_link(remat_arg);
1694 if(!has_reg_class(si, remat_arg)) continue;
1696 /* only for values in L\U, the others are handled with post_use */
1697 if(!set_find_keyval(args, remat_arg)) {
1698 /* remat <= live_rang(remat_arg) */
1699 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
1700 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1702 /* if value is becoming live through use by remat2 */
1703 if(!pset_find_ptr(live, remat_arg)) {
1706 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1707 lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1709 arg_op->attr.live_range.ilp = lr;
1710 arg_op->attr.live_range.op = irn;
1712 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
1714 pset_insert_ptr(live, remat_arg);
1715 add_to_spill_bb(si, bb, remat_arg);
1718 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1719 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1724 #ifdef CHECK_POST_REMAT
1725 /* iterate over following remats and add them to check_post_remat */
1726 foreach_post_remat(irn, tmp) {
1727 op_t *remat_op = get_irn_link(tmp);
1729 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
1731 lpp_set_factor_fast(si->lpp, check_post_remat, remat_op->attr.remat.ilp, 1.0);
1737 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
1739 pset_foreach(live, tmp) {
1740 assert(has_reg_class(si, tmp));
1743 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
1744 ir_node *arg = get_irn_n(irn, i);
1746 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
1754 /* do something at the beginning of the block */
1756 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
1757 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
1759 pset_foreach(live, irn) {
1760 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
1763 /* construct mem_outs for all values */
1765 set_foreach(spill_bb->ilp, spill) {
1766 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
1767 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1769 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
1770 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
1772 if(pset_find_ptr(live, spill->irn)) {
1773 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
1775 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
1776 spill->mem_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1778 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
1783 /* L\U is empty at bb start */
1784 /* arg is live throughout epilog if it is reg_in into this block */
1786 /* check the register pressure at the beginning of the block
1789 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
1790 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1792 pset_foreach(live, irn) {
1793 spill = set_find_spill(spill_bb->ilp, irn);
1796 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
1797 spill->reg_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1799 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
1801 foreach_post_remat(bb, irn) {
1802 op_t *remat_op = get_irn_link(irn);
1804 DBG((si->dbg, LEVEL_4, "\t next post remat: %+F\n", irn));
1805 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
1807 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1810 /* forall remat2 add requirements */
1811 foreach_post_remat(bb, tmp) {
1815 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1816 ir_node *remat_arg = get_irn_n(tmp, i);
1817 op_t *remat_op = get_irn_link(tmp);
1819 if(!has_reg_class(si, remat_arg)) continue;
1821 spill = set_find_spill(spill_bb->ilp, remat_arg);
1824 /* TODO verify this is placed correctly */
1825 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
1826 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1827 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
1828 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1832 /* mem_in/reg_in for live_in values, especially phis and their arguments */
1833 pset_foreach(live, irn) {
1838 spill = set_find_spill(spill_bb->ilp, irn);
1839 assert(spill && spill->irn == irn);
1841 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
1842 for (i = 0, n = get_Phi_n_preds(irn); i < n; ++i) {
1845 ir_node *phi_arg = get_Phi_pred(irn, i);
1846 ir_node *bb_p = get_Block_cfgpred_block(bb, i);
1847 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
1850 /* although the phi is in the right regclass one or more of
1851 * its arguments can be in a different one or at least to
1854 if(has_reg_class(si, phi_arg)) {
1855 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
1856 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1857 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
1858 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1860 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
1861 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
1863 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
1866 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
1867 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
1871 /* else assure the value arrives on all paths in the same resource */
1873 for (i = 0, n = get_Block_n_cfgpreds(bb); i < n; ++i) {
1876 ir_node *bb_p = get_Block_cfgpred_block(bb, i);
1877 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
1880 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
1881 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1882 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
1883 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1885 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
1886 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
1888 spill_p = set_find_spill(spill_bb_p->ilp, irn);
1891 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
1892 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
1897 /* first live ranges from reg_ins */
1898 pset_foreach(live, irn) {
1899 op_t *op = get_irn_link(irn);
1901 spill = set_find_spill(spill_bb->ilp, irn);
1902 assert(spill && spill->irn == irn);
1904 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
1905 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1906 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1907 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
1909 foreach_post_remat(bb, tmp) {
1910 op_t *remat_op = get_irn_link(tmp);
1912 if(remat_op->attr.remat.remat->value == irn) {
1913 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1918 /* walk forward now and compute constraints for placing spills */
1919 /* this must only be done for values that are not defined in this block */
1920 /* TODO are these values at start of block? if yes, just check whether this is a diverge edge and skip the loop */
1921 pset_foreach(live, irn) {
1922 spill = set_find_spill(spill_bb->ilp, irn);
1925 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
1926 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1928 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
1929 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
1931 sched_foreach_op(bb, tmp) {
1932 op_t *op = get_irn_link(tmp);
1934 if(is_Phi(tmp)) continue;
1935 assert(!is_Proj(tmp));
1938 ir_node *value = op->attr.remat.remat->value;
1941 /* only collect remats up to the first use of a value */
1942 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
1948 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1949 ir_node *arg = get_irn_n(tmp, i);
1952 /* if a value is used stop collecting remats */
1958 if(cst == ILP_UNDEF) break;
1963 /* if a value is used by a mem-phi, then mem_in of this value is 0 (has to be spilled again into a different slot)
1964 mem_in(phi) -> not mem_in(orig_value) TODO: how does this depend on a certain predecessor?
1967 /* mem_in of mem-phi has associated costs (but first one is free) */
1968 /* define n_mem_copies as positive integer in each predecessor block,
1969 #mem_in into this block from predecessor block - 1 weighted with SPILL_COST*execfreq(predecessor)
1979 * Speicherkopienminimierung: teste Speicherwerte auf Interferenz
1980 * und weise Spillkontexte zu. Sorge bei Phis dafuer, dass gleiche
1981 * Kontexte zusammenfliessen (Operanden und Ergebnis hat gleichen
1988 return fabs(x) < 0.00001;
1993 is_spilled(const spill_ilp_t * si, const live_range_t * lr)
1995 return !is_zero(lpp_get_var_sol(si->lpp, lr->in_mem_var));
2000 is_mem_phi(const ir_node * phi, void *data)
2002 spill_ilp_t *si = data;
2003 // return is_spilled(si, get_use_head(si, phi)->closest_use);
2008 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
2010 spill_ilp_t *si = get_irg_link(current_ir_graph);
2012 if(pset_find_ptr(si->all_possible_remats, n)) {
2013 op_t *op = (op_t*)get_irn_link(n);
2014 assert(op && op->is_remat);
2016 if(!op->attr.remat.remat->inverse) {
2017 if(op->attr.remat.pre) {
2018 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
2020 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
2025 op_t *op = (op_t*)get_irn_link(n);
2026 assert(op && op->is_remat);
2028 if(op->attr.remat.pre) {
2029 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
2031 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
2042 dump_graph_with_remats(ir_graph * irg, const char * suffix)
2044 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
2045 be_dump(irg, suffix, dump_ir_block_graph_sched);
2046 set_dump_node_vcgattr_hook(NULL);
2051 * Edge hook to dump the schedule edges with annotated register pressure.
2054 sched_pressure_edge_hook(FILE *F, ir_node *irn)
2056 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
2057 ir_node *prev = sched_prev(irn);
2058 fprintf(F, "edge:{sourcename:\"");
2060 fprintf(F, "\" targetname:\"");
2062 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
2063 fprintf(F, "\" color:magenta}\n");
2069 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
2071 DUMP_NODE_EDGE_FUNC old = get_dump_node_edge_hook();
2073 dump_consts_local(0);
2074 set_dump_node_edge_hook(sched_pressure_edge_hook);
2075 dump_ir_block_graph(irg, suffix);
2076 set_dump_node_edge_hook(old);
2080 walker_pressure_annotator(ir_node * bb, void * data)
2082 spill_ilp_t *si = data;
2087 pset *live = pset_new_ptr_default();
2090 live_foreach(bb, li) {
2091 irn = (ir_node *) li->irn;
2093 if (live_is_end(li) && has_reg_class(si, irn)) {
2094 pset_insert_ptr(live, irn);
2098 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
2100 sched_foreach_reverse(bb, irn) {
2102 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
2106 if(has_reg_class(si, irn)) {
2107 pset_remove_ptr(live, irn);
2108 if(is_Proj(irn)) ++projs;
2111 if(!is_Proj(irn)) projs = 0;
2113 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
2114 ir_node *arg = get_irn_n(irn, i);
2116 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
2118 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
2125 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
2127 be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
2132 connect_all_remats_with_keep(spill_ilp_t * si)
2140 n_remats = pset_count(si->all_possible_remats);
2142 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
2145 pset_foreach(si->all_possible_remats, irn) {
2150 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
2152 obstack_free(si->obst, ins);
2158 connect_all_spills_with_keep(spill_ilp_t * si)
2167 n_spills = pset_count(si->spills);
2169 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
2172 pset_foreach(si->spills, irn) {
2177 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
2179 obstack_free(si->obst, ins);
2183 /** insert a spill at an arbitrary position */
2184 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert, ir_node *ctx)
2186 ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
2187 ir_graph *irg = get_irn_irg(bl);
2188 ir_node *frame = get_irg_frame(irg);
2192 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
2193 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
2195 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn, ctx);
2198 * search the right insertion point. a spill of a phi cannot be put
2199 * directly after the phi, if there are some phis behind the one which
2200 * is spilled. Also, a spill of a Proj must be after all Projs of the
2203 * Here's one special case:
2204 * If the spill is in the start block, the spill must be after the frame
2205 * pointer is set up. This is done by setting insert to the end of the block
2206 * which is its default initialization (see above).
2209 if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
2212 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
2215 sched_add_after(insert, spill);
2220 delete_remat(spill_ilp_t * si, ir_node * remat) {
2223 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2225 sched_remove(remat);
2227 /* kill links to operands */
2228 for (i = -1, n = get_irn_arity(remat); i < n; ++i) {
2229 set_irn_n(remat, i, bad);
2234 clean_remat_info(spill_ilp_t * si)
2239 remat_info_t *remat_info;
2240 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2242 set_foreach(si->remat_info, remat_info) {
2243 if(!remat_info->remats) continue;
2245 pset_foreach(remat_info->remats, remat)
2247 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
2248 set_irn_n(remat->proj, -1, bad);
2249 set_irn_n(remat->proj, 0, bad);
2252 if(get_irn_n_edges(remat->op) == 0) {
2253 for (i = -1, n = get_irn_arity(remat->op); i < n; ++i) {
2254 set_irn_n(remat->op, i, bad);
2259 if(remat_info->remats) del_pset(remat_info->remats);
2260 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
2265 delete_unnecessary_remats(spill_ilp_t * si)
2270 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2273 ir_node *end = get_irg_end(si->chordal_env->irg);
2276 for (i = 0, n = get_irn_arity(si->keep); i < n; ++i) {
2277 ir_node *keep_arg = get_irn_n(si->keep, i);
2278 op_t *arg_op = get_irn_link(keep_arg);
2281 assert(arg_op->is_remat);
2283 name = si->lpp->vars[arg_op->attr.remat.ilp];
2285 if(is_zero(name->value)) {
2286 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
2287 /* TODO check whether reload is preferred over remat (could be bug) */
2288 delete_remat(si, keep_arg);
2290 if(!arg_op->attr.remat.remat->inverse) {
2291 if(arg_op->attr.remat.pre) {
2292 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
2294 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
2297 if(arg_op->attr.remat.pre) {
2298 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
2300 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
2305 set_irn_n(si->keep, i, bad);
2308 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
2309 ir_node *end_arg = get_End_keepalive(end, i);
2311 if(end_arg != si->keep) {
2312 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
2315 keeps = obstack_finish(si->obst);
2316 set_End_keepalives(end, n-1, keeps);
2317 obstack_free(si->obst, keeps);
2320 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
2325 pset_foreach(si->all_possible_remats, remat) {
2326 op_t *remat_op = get_irn_link(remat);
2327 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
2329 if(is_zero(name->value)) {
2330 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
2331 /* TODO check whether reload is preferred over remat (could be bug) */
2332 delete_remat(si, remat);
2334 if(!remat_op->attr.remat.remat->inverse) {
2335 if(remat_op->attr.remat.pre) {
2336 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
2338 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
2341 if(remat_op->attr.remat.pre) {
2342 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
2344 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
2353 * @param before The node after which the spill will be placed in the schedule
2355 /* TODO set context properly */
2357 insert_spill(spill_ilp_t * si, const ir_node * irn, const ir_node * value, const ir_node * before)
2361 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
2363 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
2365 spill = be_spill2(arch_env, irn, before, irn);
2367 defs = set_insert_def(si->values, value);
2370 /* enter into the linked list */
2371 set_irn_link(spill, defs->spills);
2372 defs->spills = spill;
2374 #ifdef KEEPALIVE_SPILLS
2375 pset_insert_ptr(si->spills, spill);
2382 * @param before The Phi node which has to be spilled
2385 insert_mem_phi(spill_ilp_t * si, const ir_node * phi)
2393 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
2395 for(i=0,n=get_irn_arity(phi); i<n; ++i) {
2396 ins[i] = si->m_unknown;
2399 mem_phi = new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
2401 defs = set_insert_def(si->values, phi);
2404 /* enter into the linked list */
2405 set_irn_link(mem_phi, defs->spills);
2406 defs->spills = mem_phi;
2408 sched_add_after(phi, mem_phi);
2410 #ifdef KEEPALIVE_SPILLS
2411 pset_insert_ptr(si->spills, mem_phi);
2418 * Add remat to list of defs, destroys link field!
2421 insert_remat(spill_ilp_t * si, ir_node * remat)
2424 op_t *remat_op = get_irn_link(remat);
2426 assert(remat_op->is_remat);
2428 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
2431 /* enter into the linked list */
2432 set_irn_link(remat, defs->remats);
2433 defs->remats = remat;
2438 collect_spills(spill_ilp_t * si, ir_node * value, pset * spills, pset * visited)
2443 defs = set_find_def(si->values, value);
2445 if(defs && defs->spills) {
2446 for(next = defs->spills; next; next = get_irn_link(next)) {
2447 pset_insert_ptr(spills, next);
2449 } else if (is_Phi(value)) {
2451 if(!pset_find_ptr(visited, value)) {
2455 pset_insert_ptr(visited, value);
2456 for(i=0, n=get_irn_arity(value); i<n; ++i) {
2457 ir_node *arg = get_irn_n(value, i);
2459 collect_spills(si, arg, spills, visited);
2463 // assert(0 && "Phi operand not spilled");
2469 get_spills_for_value(spill_ilp_t * si, ir_node * value)
2471 pset *spills = pset_new_ptr_default();
2472 // pset *visited = pset_new_ptr_default();
2474 // collect_spills(si, value, spills, visited);
2475 // del_pset(visited);
2479 defs = set_find_def(si->values, value);
2481 if(defs && defs->spills) {
2482 for(next = defs->spills; next; next = get_irn_link(next)) {
2483 pset_insert_ptr(spills, next);
2491 * Add reload before operation and add to list of defs
2494 insert_reload(spill_ilp_t * si, const ir_node * value, const ir_node * after)
2499 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
2501 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
2503 defs = set_find_def(si->values, value);
2504 /* get a spill of this value */
2506 if((!defs || !defs->spills) && is_Phi(value)) {
2509 spills = get_spills_for_value(si, value);
2511 spill = pset_first(spills);
2515 defs = set_insert_def(si->values, value);
2517 defs->spills = spill;
2518 set_irn_link(spill, NULL);
2520 spill = defs->spills;
2523 spill = defs->spills;
2524 assert(spill && "no spill placed before reload");
2526 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
2528 /* enter into the linked list */
2529 set_irn_link(reload, defs->remats);
2530 defs->remats = reload;
2536 walker_spill_placer(ir_node * bb, void * data) {
2537 spill_ilp_t *si = (spill_ilp_t*)data;
2539 spill_bb_t *spill_bb = get_irn_link(bb);
2540 pset *spills_to_do = pset_new_ptr_default();
2543 set_foreach(spill_bb->ilp, spill) {
2546 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2547 name = si->lpp->vars[spill->mem_in];
2548 if(!is_zero(name->value)) {
2551 mem_phi = insert_mem_phi(si, spill->irn);
2553 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
2557 name = si->lpp->vars[spill->spill];
2558 if(!is_zero(name->value)) {
2559 if(spill->reg_in > 0) {
2560 name = si->lpp->vars[spill->reg_in];
2561 if(!is_zero(name->value)) {
2562 insert_spill(si, spill->irn, spill->irn, bb);
2566 pset_insert_ptr(spills_to_do, spill->irn);
2569 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
2572 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
2573 op_t *op = get_irn_link(irn);
2575 if(be_is_Spill(irn)) continue;
2578 /* TODO fix this if we want to support remats with more than two nodes */
2579 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
2580 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
2582 insert_spill(si, irn, op->attr.remat.remat->value, irn);
2585 if(pset_find_ptr(spills_to_do, irn)) {
2586 pset_remove_ptr(spills_to_do, irn);
2588 insert_spill(si, irn, irn, irn);
2594 assert(pset_count(spills_to_do) == 0);
2596 /* afterwards free data in block */
2597 del_pset(spills_to_do);
2601 phim_fixer(spill_ilp_t *si) {
2604 set_foreach(si->values, defs) {
2605 const ir_node *phi = defs->value;
2606 ir_node *phi_m = NULL;
2607 ir_node *next = defs->spills;
2611 if(!is_Phi(phi)) continue;
2614 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
2618 next = get_irn_link(next);
2621 if(!phi_m) continue;
2623 for(i=0,n=get_irn_arity(phi); i<n; ++i) {
2624 const ir_node *value = get_irn_n(phi, i);
2625 defs_t *val_defs = set_find_def(si->values, value);
2627 /* get a spill of this value */
2628 ir_node *spill = val_defs->spills;
2630 assert(spill && "no spill placed before PhiM");
2632 set_irn_n(phi_m, i, spill);
2638 walker_reload_placer(ir_node * bb, void * data) {
2639 spill_ilp_t *si = (spill_ilp_t*)data;
2641 spill_bb_t *spill_bb = get_irn_link(bb);
2645 sched_foreach_reverse(bb, irn) {
2646 op_t *op = get_irn_link(irn);
2648 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
2649 if(is_Phi(irn)) break;
2652 if(get_irn_mode(irn) != mode_T) {
2653 insert_remat(si, irn);
2658 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
2659 ir_node *arg = get_irn_n(irn, i);
2661 if(op->attr.live_range.reloads && op->attr.live_range.reloads[i] != ILP_UNDEF) {
2664 name = si->lpp->vars[op->attr.live_range.reloads[i]];
2665 if(!is_zero(name->value)) {
2667 ir_node *insert_pos = irn;
2668 ir_node *prev = insert_pos;
2672 prev = sched_prev(prev);
2673 } while(be_is_Spill(prev));
2675 prev_op = get_irn_link(prev);
2677 /* insert reload before pre-remats */
2678 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
2679 && prev_op->is_remat && prev_op->attr.remat.pre) {
2683 prev = sched_prev(prev);
2684 } while(be_is_Spill(prev));
2686 prev_op = get_irn_link(prev);
2690 reload = insert_reload(si, arg, insert_pos);
2692 set_irn_n(irn, i, reload);
2694 #ifdef KEEPALIVE_RELOADS
2695 pset_insert_ptr(si->spills, reload);
2703 /* reloads at end of block */
2704 if(spill_bb->reloads) {
2706 live_foreach(bb, li) {
2707 ir_node *irn = (ir_node *) li->irn;
2709 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
2712 name = si->lpp->vars[spill_bb->reloads[i]];
2713 if(!is_zero(name->value)) {
2715 ir_node *insert_pos = bb;
2716 ir_node *prev = sched_prev(insert_pos);
2717 op_t *prev_op = get_irn_link(prev);
2719 /* insert reload before pre-remats */
2720 while(!sched_is_end(prev) && !be_is_Reload(prev) && !be_is_Spill(prev)
2721 && prev_op->is_remat && prev_op->attr.remat.pre) {
2724 prev = sched_prev(insert_pos);
2725 prev_op = get_irn_link(prev);
2728 reload = insert_reload(si, irn, insert_pos);
2730 #ifdef KEEPALIVE_RELOADS
2731 pset_insert_ptr(si->spills, reload);
2739 del_set(spill_bb->ilp);
2743 walker_collect_used(ir_node * irn, void * data)
2745 lc_bitset_t *used = data;
2747 lc_bitset_set(used, get_irn_idx(irn));
2750 struct kill_helper {
2756 walker_kill_unused(ir_node * bb, void * data)
2758 struct kill_helper *kh = data;
2759 const ir_node *bad = get_irg_bad(get_irn_irg(bb));
2763 for(irn=sched_first(bb); !sched_is_end(irn);) {
2764 ir_node *next = sched_next(irn);
2768 if(!lc_bitset_is_set(kh->used, get_irn_idx(irn))) {
2769 if(be_is_Spill(irn) || be_is_Reload(irn)) {
2770 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)));
2771 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
2776 set_nodes_block(irn, bad);
2777 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
2778 set_irn_n(irn, i, bad);
2786 kill_all_unused_values_in_schedule(spill_ilp_t * si)
2788 struct kill_helper kh;
2790 kh.used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
2793 irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
2794 irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
2796 lc_bitset_free(kh.used);
2800 print_irn_pset(pset * p)
2804 pset_foreach(p, irn) {
2805 ir_printf("%+F\n", irn);
2810 rewire_uses(spill_ilp_t * si)
2812 dom_front_info_t *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
2814 pset *ignore = pset_new_ptr(1);
2816 pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
2818 /* then fix uses of spills */
2819 set_foreach(si->values, defs) {
2822 ir_node *next = defs->remats;
2825 reloads = pset_new_ptr_default();
2828 if(be_is_Reload(next)) {
2829 pset_insert_ptr(reloads, next);
2833 next = get_irn_link(next);
2836 spills = get_spills_for_value(si, defs->value);
2837 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));
2838 if(pset_count(spills) > 1) {
2839 //assert(pset_count(reloads) > 0);
2840 // print_irn_pset(spills);
2841 // print_irn_pset(reloads);
2843 be_ssa_constr_set_ignore(dfi, spills, ignore);
2850 /* first fix uses of remats and reloads */
2851 set_foreach(si->values, defs) {
2853 ir_node *next = defs->remats;
2856 nodes = pset_new_ptr_default();
2857 pset_insert_ptr(nodes, defs->value);
2860 pset_insert_ptr(nodes, next);
2861 next = get_irn_link(next);
2864 if(pset_count(nodes) > 1) {
2865 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
2866 be_ssa_constr_set(dfi, nodes);
2873 // remove_unused_defs(si);
2875 be_free_dominance_frontiers(dfi);
2879 writeback_results(spill_ilp_t * si)
2881 /* walk through the graph and collect all spills, reloads and remats for a value */
2883 si->values = new_set(cmp_defs, 4096);
2885 DBG((si->dbg, LEVEL_1, "Applying results\n"));
2886 delete_unnecessary_remats(si);
2887 si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
2888 irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
2890 irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
2892 /* clean the remat info! there are still back-edges leading there! */
2893 clean_remat_info(si);
2897 connect_all_spills_with_keep(si);
2899 del_set(si->values);
2903 get_n_regs(spill_ilp_t * si)
2905 int arch_n_regs = arch_register_class_n_regs(si->cls);
2909 for(i=0; i<arch_n_regs; i++) {
2910 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
2915 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
2920 walker_reload_mover(ir_node * bb, void * data)
2922 spill_ilp_t *si = data;
2925 sched_foreach(bb, tmp) {
2926 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
2927 ir_node *reload = tmp;
2930 /* move reload upwards */
2932 int pressure = (int)get_irn_link(reload);
2933 if(pressure < si->n_regs) {
2934 irn = sched_prev(reload);
2935 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
2936 sched_remove(reload);
2937 pressure = (int)get_irn_link(irn);
2939 while(pressure < si->n_regs) {
2940 if(sched_is_end(irn) || (be_is_Reload(irn) && has_reg_class(si, irn))) break;
2942 set_irn_link(irn, INT_TO_PTR(pressure+1));
2943 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
2944 irn = sched_prev(irn);
2946 pressure = (int)get_irn_link(irn);
2949 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
2950 sched_put_after(irn, reload);
2957 move_reloads_upward(spill_ilp_t * si)
2959 irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
2963 be_spill_remat(const be_chordal_env_t * chordal_env)
2965 char problem_name[256];
2966 char dump_suffix[256];
2967 char dump_suffix2[256];
2968 char dump_suffix3[256];
2969 struct obstack obst;
2972 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
2973 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
2974 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
2975 ir_snprintf(dump_suffix3, sizeof(dump_suffix3), "-%s-reloads_moved", chordal_env->cls->name);
2977 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
2978 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
2980 obstack_init(&obst);
2981 si.chordal_env = chordal_env;
2983 si.senv = be_new_spill_env(chordal_env, is_mem_phi, &si);
2984 si.cls = chordal_env->cls;
2985 si.lpp = new_lpp(problem_name, lpp_minimize);
2986 si.remat_info = new_set(cmp_remat_info, 4096);
2987 si.all_possible_remats = pset_new_ptr_default();
2988 si.spills = pset_new_ptr_default();
2989 si.inverse_ops = pset_new_ptr_default();
2990 #ifndef EXECFREQ_LOOPDEPH
2991 si.execfreqs = compute_execfreq(chordal_env->irg);
2993 si.execfreqs = NULL;
2998 si.n_regs = get_n_regs(&si);
3000 set_irg_link(chordal_env->irg, &si);
3001 compute_doms(chordal_env->irg);
3003 be_analyze_regpressure(chordal_env, "-pre");
3005 #ifdef COLLECT_REMATS
3006 /* collect remats */
3007 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
3008 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
3011 /* insert possible remats */
3012 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
3013 irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
3014 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
3017 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
3018 connect_all_remats_with_keep(&si);
3019 /* dump graph with inserted remats */
3020 dump_graph_with_remats(chordal_env->irg, dump_suffix);
3024 /* recompute liveness */
3025 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
3026 be_liveness(chordal_env->irg);
3030 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
3031 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
3032 irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
3034 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
3035 irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
3042 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
3043 if ((f = fopen(buf, "wt")) != NULL) {
3044 lpp_dump_plain(si.lpp, f);
3051 DBG((si.dbg, LEVEL_1, "\tSolving %F\n", chordal_env->irg));
3053 lpp_set_time_limit(si.lpp, ILP_TIMEOUT);
3057 lpp_solve_cplex(si.lpp);
3059 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
3061 assert(lpp_is_sol_valid(si.lpp)
3062 && "solution of ILP must be valid");
3064 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));
3066 #ifdef DUMP_SOLUTION
3071 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
3072 if ((f = fopen(buf, "wt")) != NULL) {
3074 for (i = 0; i < si.lpp->var_next; ++i) {
3075 lpp_name_t *name = si.lpp->vars[i];
3076 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
3083 writeback_results(&si);
3087 kill_all_unused_values_in_schedule(&si);
3089 #if defined(KEEPALIVE_SPILLS) || defined(KEEPALIVE_RELOADS)
3090 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
3093 be_liveness(chordal_env->irg);
3094 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
3096 dump_pressure_graph(&si, dump_suffix2);
3098 // TODO fix temporarily exceeded regpressure due to remat2s
3100 // TODO insert copys to fix interferences in memory
3102 // move reloads upwards
3103 move_reloads_upward(&si);
3104 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
3105 dump_pressure_graph(&si, dump_suffix3);
3107 be_analyze_regpressure(chordal_env, "-post");
3109 free_dom(chordal_env->irg);
3110 del_pset(si.inverse_ops);
3111 del_pset(si.all_possible_remats);
3112 del_pset(si.spills);
3113 #ifndef EXECFREQ_LOOPDEPH
3114 free_execfreq(si.execfreqs);
3117 obstack_free(&obst, NULL);
3121 #else /* WITH_ILP */
3124 only_that_you_can_compile_without_WITH_ILP_defined(void)
3128 #endif /* WITH_ILP */