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
36 #include <lpp/lpp_net.h>
37 #include <lpp/lpp_cplex.h>
38 //#include <lc_pset.h>
39 #include <libcore/lc_bitset.h>
43 #include "besched_t.h"
48 #include "bespillremat.h"
51 #include "bechordal_t.h"
58 #define COLLECT_REMATS
59 #define REMAT_WHILE_LIVE
60 #define NO_ENLARGE_L1V3N355
61 //#define EXECFREQ_LOOPDEPH
65 #define LPP_SERVER "i44pc52"
66 #define LPP_SOLVER "cplex"
72 #define ILP_TIMEOUT 20
76 typedef struct _spill_ilp_t {
77 const arch_register_class_t *cls;
79 const be_chordal_env_t *chordal_env;
84 pset *all_possible_remats;
88 set *values; /**< for collecting all definitions of values before running ssa-construction */
90 DEBUG_ONLY(firm_dbg_module_t * dbg);
93 typedef int ilp_var_t;
94 typedef int ilp_cst_t;
96 typedef struct _spill_bb_t {
102 typedef struct _remat_t {
103 const ir_node *op; /**< for copy_irn */
104 const ir_node *proj; /**< not NULL if the above op produces a tuple */
105 const ir_node *value; /**< the value which is being recomputed by this remat */
106 int cost; /**< cost of this remat */
110 * Data to be attached to each IR node. For remats this contains the ilp_var
111 * for this remat and for normal ops this contains the ilp_vars for
112 * reloading each operand
114 typedef struct _op_t {
119 remat_t *remat; /** the remat this op belongs to */
120 int pre; /** 1, if this is a pressure-increasing remat */
124 ir_node *op; /** the operation this live range belongs to */
130 typedef struct _defs_t {
132 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
133 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
136 typedef struct _remat_info_t {
137 const ir_node *irn; /**< the irn to which these remats belong */
138 pset *remats; /**< possible remats for this value */
139 pset *remats_by_operand; /**< remats with this value as operand */
142 typedef struct _keyval_t {
147 typedef struct _spill_t {
157 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
159 return chordal_has_class(si->chordal_env, irn);
164 cmp_remat(const void *a, const void *b)
166 const keyval_t *p = a;
167 const keyval_t *q = b;
168 const remat_t *r = p->val;
169 const remat_t *s = q->val;
173 return !(r == s || r->op == s->op);
177 cmp_remat(const void *a, const void *b)
179 const remat_t *r = a;
180 const remat_t *s = a;
182 return !(r == s || r->op == s->op);
186 cmp_spill(const void *a, const void *b, size_t size)
188 const spill_t *p = a;
189 const spill_t *q = b;
191 // return !(p->irn == q->irn && p->bb == q->bb);
192 return !(p->irn == q->irn);
196 set_find_keyval(set * set, void * key)
201 return set_find(set, &query, sizeof(query), HASH_PTR(key));
205 set_insert_keyval(set * set, void * key, void * val)
211 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
215 set_find_def(set * set, ir_node * value)
220 return set_find(set, &query, sizeof(query), HASH_PTR(value));
224 set_insert_def(set * set, ir_node * value)
231 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
235 set_find_spill(set * set, ir_node * value)
240 return set_find(set, &query, sizeof(query), HASH_PTR(value));
243 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
244 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
245 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
246 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
247 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
250 cmp_remat_info(const void *a, const void *b, size_t size)
252 const remat_info_t *p = a;
253 const remat_info_t *q = b;
255 return !(p->irn == q->irn);
259 cmp_defs(const void *a, const void *b, size_t size)
264 return !(p->value == q->value);
268 cmp_keyval(const void *a, const void *b, size_t size)
270 const keyval_t *p = a;
271 const keyval_t *q = b;
273 return !(p->key == q->key);
277 execution_frequency(const spill_ilp_t * si, const ir_node * irn)
279 #ifdef EXECFREQ_LOOPDEPH
281 return expf((float)get_loop_depth(get_irn_loop(irn)) * logf(10));
283 return expf((float)get_loop_depth(get_irn_loop(get_nodes_block(irn))) * logf(10));
286 return get_block_execfreq(si->execfreqs, irn);
288 return get_block_execfreq(si->execfreqs, get_nodes_block(irn));
294 * Checks, whether node and its operands have suitable reg classes
297 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
301 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
302 int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
306 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
309 for (i = 0, n = get_irn_arity(irn); i < n && remat; ++i) {
310 ir_node *op = get_irn_n(irn, i);
311 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
314 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
321 * Try to create a remat from @p op with destination value @p dest_value
323 static INLINE remat_t *
324 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
326 remat_t *remat = NULL;
328 // if(!mode_is_datab(get_irn_mode(dest_value)))
331 if(dest_value == op) {
332 const ir_node *proj = NULL;
334 if(is_Proj(dest_value)) {
335 op = get_irn_n(op, 0);
339 if(!is_rematerializable(si, op))
342 remat = obstack_alloc(si->obst, sizeof(*remat));
344 remat->cost = COST_REMAT; /* TODO ask backend for real cost */
345 remat->value = dest_value;
348 arch_inverse_t inverse;
352 /* get the index of the operand we want to retrieve by the inverse op */
353 for (i = 0, n = get_irn_arity(op); i < n; ++i) {
354 ir_node *arg = get_irn_n(op, i);
356 if(arg == dest_value) break;
358 if(i == n) return NULL;
360 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", i, op));
362 /* else ask the backend to give an inverse op */
363 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, i, &inverse, si->obst)) {
364 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
366 assert(inverse.n > 0 && "inverse op should have at least one node");
369 remat = obstack_alloc(si->obst, sizeof(*remat));
370 remat->op = inverse.nodes[0];
371 remat->cost = inverse.costs;
372 remat->value = dest_value;
373 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
375 assert(is_Proj(remat->proj));
377 DBG((si->dbg, LEVEL_1, "\t I can not handle remats with %d nodes\n", inverse.n));
384 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
386 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
394 add_remat(const spill_ilp_t * si, const remat_t * remat)
396 remat_info_t *remat_info,
402 assert(remat->value);
404 query.irn = remat->value;
406 query.remats_by_operand = NULL;
407 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
409 if(remat_info->remats == NULL) {
410 remat_info->remats = new_pset(cmp_remat, 4096);
412 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
414 /* insert the remat into the remats_be_operand set of each argument of the remat op */
415 for (i = 0, n = get_irn_arity(remat->op); i < n; ++i) {
416 ir_node *arg = get_irn_n(remat->op, i);
420 query.remats_by_operand = NULL;
421 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
423 if(remat_info->remats_by_operand == NULL) {
424 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
426 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
431 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
437 if(has_reg_class(si, op)) {
438 remat = get_remat_from_op(si, op, op);
440 add_remat(si, remat);
444 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
446 for (i = 0, n = get_irn_arity(op); i < n; ++i) {
447 ir_node *arg = get_irn_n(op, i);
449 if(has_reg_class(si, arg)) {
450 /* try to get an inverse remat */
451 remat = get_remat_from_op(si, arg, op);
453 add_remat(si, remat);
461 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
464 ir_node *def_block = get_nodes_block(val);
470 /* if pos is at end of a basic block */
472 ret = (pos == def_block || block_dominates(def_block, pos));
473 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
477 /* else if this is a normal operation */
478 block = get_nodes_block(pos);
479 if(block == def_block) {
480 ret = sched_comes_after(val, pos);
481 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
485 ret = block_dominates(def_block, block);
486 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
490 static INLINE ir_node *
491 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
493 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
497 * Returns first non-Phi node of block @p bb
499 static INLINE ir_node *
500 sched_block_first_nonphi(const ir_node * bb)
502 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
506 sched_skip_proj_predicator(const ir_node * irn, void * data)
508 return (is_Proj(irn));
511 static INLINE ir_node *
512 sched_next_nonproj(const ir_node * irn, int forward)
514 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
518 * Returns next operation node (non-Proj) after @p irn
519 * or the basic block of this node
521 static INLINE ir_node *
522 sched_next_op(const ir_node * irn)
524 ir_node *next = sched_next(irn);
529 return sched_next_nonproj(next, 1);
533 * Returns previous operation node (non-Proj) before @p irn
534 * or the basic block of this node
536 static INLINE ir_node *
537 sched_prev_op(const ir_node * irn)
539 ir_node *prev = sched_prev(irn);
544 return sched_next_nonproj(prev, 0);
548 sched_put_after(ir_node * insert, ir_node * irn)
550 if(is_Block(insert)) {
551 insert = sched_block_first_nonphi(insert);
553 insert = sched_next_op(insert);
555 sched_add_before(insert, irn);
559 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
561 if(is_Block(insert)) {
562 insert = sched_block_last_noncf(si, insert);
564 insert = sched_next_nonproj(insert, 0);
565 insert = sched_prev(insert);
567 sched_add_after(insert, irn);
571 * Tells you whether a @p remat can be placed before the irn @p pos
574 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
576 const ir_node *op = remat->op;
583 prev = sched_block_last_noncf(si, pos);
584 prev = sched_next_nonproj(prev, 0);
586 prev = sched_prev_op(pos);
588 /* do not remat if the rematted value is defined immediately before this op */
589 if(prev == remat->op) {
594 /* this should be just fine, the following OP will be using this value, right? */
596 /* only remat AFTER the real definition of a value (?) */
597 if(!value_is_defined_before(si, pos, remat->value)) {
598 // ir_fprintf(stderr, "error(not defined)");
603 for(i=0, n=get_irn_arity(op); i<n && res; ++i) {
604 const ir_node *arg = get_irn_n(op, i);
606 #ifdef NO_ENLARGE_L1V3N355
607 if(has_reg_class(si, arg) && live) {
608 res &= pset_find_ptr(live, arg)?1:0;
610 res &= value_is_defined_before(si, pos, arg);
613 res &= value_is_defined_before(si, pos, arg);
621 * Tells you whether a @p remat can be placed after the irn @p pos
624 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
627 pos = sched_block_first_nonphi(pos);
629 pos = sched_next_op(pos);
632 /* only remat AFTER the real definition of a value (?) */
633 if(!value_is_defined_before(si, pos, remat->value)) {
637 return can_remat_before(si, remat, pos, live);
641 * Collect potetially rematerializable OPs
644 walker_remat_collector(ir_node * irn, void * data)
646 spill_ilp_t *si = data;
648 if(!is_Block(irn) && !is_Phi(irn)) {
649 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
650 get_remats_from_op(si, irn);
655 * Inserts a copy of @p irn before @p pos
658 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
663 bb = is_Block(pos)?pos:get_nodes_block(pos);
664 copy = exact_copy(irn);
665 set_nodes_block(copy, bb);
666 sched_put_before(si, pos, copy);
672 * Inserts a copy of @p irn after @p pos
675 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
680 bb = is_Block(pos)?pos:get_nodes_block(pos);
681 copy = exact_copy(irn);
682 set_nodes_block(copy, bb);
683 sched_put_after(pos, copy);
689 insert_remat_after(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
693 if(can_remat_after(si, remat, pos, live)) {
698 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
700 copy = insert_copy_after(si, remat->op, pos);
702 // ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", remat->value, pos);
703 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
704 op = obstack_alloc(si->obst, sizeof(*op));
706 op->attr.remat.remat = remat;
707 op->attr.remat.pre = 0;
708 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
710 set_irn_link(copy, op);
711 pset_insert_ptr(si->all_possible_remats, copy);
713 proj_copy = insert_copy_after(si, remat->proj, copy);
714 set_irn_n(proj_copy, 0, copy);
715 set_irn_link(proj_copy, op);
716 pset_insert_ptr(si->all_possible_remats, proj_copy);
724 insert_remat_before(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
728 if(can_remat_before(si, remat, pos, live)) {
733 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
735 copy = insert_copy_before(si, remat->op, pos);
737 // ir_snprintf(buf, sizeof(buf), "remat_%N_%N", remat->value, pos);
738 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
739 op = obstack_alloc(si->obst, sizeof(*op));
741 op->attr.remat.remat = remat;
742 op->attr.remat.pre = 1;
743 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(si, pos));
745 set_irn_link(copy, op);
746 pset_insert_ptr(si->all_possible_remats, copy);
748 proj_copy = insert_copy_after(si, remat->proj, copy);
749 set_irn_n(proj_copy, 0, copy);
750 set_irn_link(proj_copy, op);
751 pset_insert_ptr(si->all_possible_remats, proj_copy);
760 * Insert (so far unused) remats into the irg to
761 * recompute the potential liveness of all values
764 walker_remat_insertor(ir_node * bb, void * data)
766 spill_ilp_t *si = data;
767 spill_bb_t *spill_bb;
772 pset *live = pset_new_ptr_default();
774 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
776 live_foreach(bb, li) {
777 ir_node *value = (ir_node *) li->irn;
779 /* add remats at end of block */
780 if (live_is_end(li) && has_reg_class(si, value)) {
781 pset_insert_ptr(live, value);
785 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
786 set_irn_link(bb, spill_bb);
788 irn = sched_last(bb);
789 while(!sched_is_end(irn)) {
795 next = sched_prev(irn);
797 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
799 if(is_Phi(irn) || is_Proj(irn)) {
802 if(has_reg_class(si, irn)) {
803 pset_remove_ptr(live, irn);
806 op = obstack_alloc(si->obst, sizeof(*op));
808 op->attr.live_range.reloads = NULL;
809 op->attr.live_range.ilp = ILP_UNDEF;
810 set_irn_link(irn, op);
816 op = obstack_alloc(si->obst, sizeof(*op));
818 op->attr.live_range.ilp = ILP_UNDEF;
819 op->attr.live_range.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.reloads) * get_irn_arity(irn));
820 memset(op->attr.live_range.reloads, 0xFF, sizeof(*op->attr.live_range.reloads) * get_irn_arity(irn));
821 set_irn_link(irn, op);
823 args = pset_new_ptr_default();
825 /* collect arguments of op */
826 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
827 ir_node *arg = get_irn_n(irn, i);
829 pset_insert_ptr(args, arg);
832 /* set args of op live in epilog */
833 pset_foreach(args, arg) {
834 if(has_reg_class(si, arg)) {
835 pset_insert_ptr(live, arg);
839 /* insert all possible remats after irn */
840 pset_foreach(args, arg) {
841 remat_info_t *remat_info,
845 /* continue if the operand has the wrong reg class
847 if(!has_reg_class(si, arg))
852 query.remats_by_operand = NULL;
853 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
859 /* do not place post remats after jumps */
860 if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
862 if(remat_info->remats_by_operand) {
863 pset_foreach(remat_info->remats_by_operand, remat) {
864 /* do not insert remats producing the same value as one of the operands */
865 if(!pset_find_ptr(args, remat->value)) {
866 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
867 #ifdef REMAT_WHILE_LIVE
868 if(pset_find_ptr(live, remat->value)) {
869 insert_remat_after(si, remat, irn, live);
872 insert_remat_after(si, remat, irn, live);
879 /* delete defined value from live set */
880 if(has_reg_class(si, irn)) {
881 pset_remove_ptr(live, irn);
884 /* insert all possible remats before irn */
885 pset_foreach(args, arg) {
886 remat_info_t *remat_info,
890 /* continue if the operand has the wrong reg class
892 if(!has_reg_class(si, arg))
897 query.remats_by_operand = NULL;
898 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
904 if(remat_info->remats) {
905 pset_foreach(remat_info->remats, remat) {
906 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
907 #ifdef REMAT_WHILE_LIVE
908 if(pset_find_ptr(live, remat->value)) {
909 insert_remat_before(si, remat, irn, live);
912 insert_remat_before(si, remat, irn, live);
922 live_foreach(bb, li) {
923 ir_node *value = (ir_node *) li->irn;
925 /* add remats at end of block */
926 if (live_is_end(li) && has_reg_class(si, value)) {
927 remat_info_t *remat_info,
933 query.remats_by_operand = NULL;
934 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
936 if(remat_info && remat_info->remats) {
937 pset_foreach(remat_info->remats, remat) {
938 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
939 #ifdef REMAT_WHILE_LIVE
940 if(is_live_end(bb, remat->value)) {
941 insert_remat_before(si, remat, bb, NULL); //TODO
944 insert_remat_before(si, remat, bb, NULL);
951 /* add remat2s at beginning of block */
952 if ((live_is_in(li) || is_Phi(value)) && has_reg_class(si, value)) {
953 remat_info_t *remat_info,
959 query.remats_by_operand = NULL;
960 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
962 if(remat_info && remat_info->remats_by_operand) {
963 pset_foreach(remat_info->remats_by_operand, remat) {
964 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at beginning of block %+F\n", remat->op, bb));
965 #ifdef REMAT_WHILE_LIVE
966 if(is_live_in(bb, remat->value)) {
967 insert_remat_after(si, remat, bb, NULL);
970 insert_remat_after(si, remat, bb, NULL);
979 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
982 luke_endwalker(ir_node * bb, void * data)
984 spill_ilp_t *si = (spill_ilp_t*)data;
991 spill_bb_t *spill_bb = get_irn_link(bb);
994 live = pset_new_ptr_default();
995 use_end = pset_new_ptr_default();
997 live_foreach(bb, li) {
998 irn = (ir_node *) li->irn;
999 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1002 pset_insert_ptr(live, irn);
1003 op = get_irn_link(irn);
1004 assert(!op->is_remat);
1008 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1009 /* their reg_out is unimportant because it can always be set */
1010 sched_foreach_reverse(bb, irn) {
1014 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1016 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
1017 ir_node *irn_arg = get_irn_n(irn, i);
1018 if(has_reg_class(si, irn_arg)) {
1019 pset_insert_ptr(use_end, irn);
1024 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1025 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1027 spill_bb->ilp = new_set(cmp_spill, 16);
1029 live_foreach(bb, li) {
1030 irn = (ir_node *) li->irn;
1031 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1036 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1038 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1039 spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1040 /* if irn is used at the end of the block, then it is live anyway */
1041 if(!pset_find_ptr(use_end, irn))
1042 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1044 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", bb, irn);
1045 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1047 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", bb, irn);
1048 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE*execution_frequency(si, irn));
1050 spill->reg_in = ILP_UNDEF;
1051 spill->mem_in = ILP_UNDEF;
1060 next_post_remat(const ir_node * irn)
1065 irn = sched_block_first_nonphi(irn);
1067 irn = sched_next_op(irn);
1070 if(sched_is_end(irn))
1073 op = (op_t*)get_irn_link(irn);
1074 if(op->is_remat && !op->attr.remat.pre) {
1083 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1089 ret = sched_block_last_noncf(si, irn);
1090 ret = sched_next(ret);
1091 ret = sched_prev_op(ret);
1093 ret = sched_prev_op(irn);
1096 if(sched_is_end(ret) || is_Phi(ret))
1099 op = (op_t*)get_irn_link(ret);
1100 if(op->is_remat && op->attr.remat.pre) {
1108 * Find a remat of value @p value in the epilog of @p pos
1111 find_post_remat(const ir_node * value, const ir_node * pos)
1113 while((pos = next_post_remat(pos)) != NULL) {
1116 op = get_irn_link(pos);
1117 assert(op->is_remat && !op->attr.remat.pre);
1119 if(op->attr.remat.remat->value == value)
1120 return (ir_node*)pos;
1123 const ir_edge_t *edge;
1124 foreach_out_edge(pos, edge) {
1125 ir_node *proj = get_edge_src_irn(edge);
1126 assert(is_Proj(proj));
1136 * Find a remat of value @p value in the prolog of @p pos
1139 find_pre_remat(const spill_ilp_t * si, const ir_node * value, const ir_node * pos)
1141 while((pos = next_pre_remat(si,pos)) != NULL) {
1144 op = get_irn_link(pos);
1145 assert(op->is_remat && op->attr.remat.pre);
1147 if(op->attr.remat.remat->value == value)
1148 return (ir_node*)pos;
1155 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1157 spill_bb_t *spill_bb = get_irn_link(bb);
1163 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1165 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1167 spill->reg_out = ILP_UNDEF;
1168 spill->reg_in = ILP_UNDEF;
1169 spill->mem_in = ILP_UNDEF;
1171 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1172 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1174 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1175 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE*execution_frequency(si, bb));
1182 * Walk all irg blocks and emit this ILP
1185 luke_blockwalker(ir_node * bb, void * data)
1187 spill_ilp_t *si = (spill_ilp_t*)data;
1193 spill_bb_t *spill_bb = get_irn_link(bb);
1199 live = pset_new_ptr_default();
1201 /* do something at the end of the block */
1203 /* init live values at end of block */
1204 live_foreach(bb, li) {
1205 ir_node *irn = (ir_node *) li->irn;
1207 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1208 pset_insert_ptr(live, irn);
1212 spill_bb->reloads = obstack_alloc(si->obst, pset_count(live) * sizeof(*spill_bb->reloads));
1213 memset(spill_bb->reloads, 0xFF, pset_count(live) * sizeof(*spill_bb->reloads));
1216 live_foreach(bb, li) {
1217 ir_node *irn = (ir_node *) li->irn;
1220 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1221 spill = set_find_spill(spill_bb->ilp, irn);
1224 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1225 spill_bb->reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, bb));
1227 /* reload <= mem_out */
1228 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1229 lpp_set_factor_fast(si->lpp, cst, spill_bb->reloads[i], 1.0);
1230 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
1232 op = get_irn_link(irn);
1233 assert(!op->is_remat);
1235 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1236 op->attr.live_range.ilp = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1237 op->attr.live_range.op = bb;
1239 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1240 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1242 /* reg_out - reload - remat - live_range <= 0 */
1243 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1244 lpp_set_factor_fast(si->lpp, cst, spill_bb->reloads[i], -1.0);
1245 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1246 foreach_pre_remat(si, bb, tmp) {
1247 op_t *remat_op = get_irn_link(tmp);
1248 if(remat_op->attr.remat.remat->value == irn) {
1249 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1256 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1258 sched_foreach_reverse(bb, irn) {
1264 ilp_cst_t check_pre,
1267 set *args = new_set(cmp_keyval, get_irn_arity(irn));
1273 op = get_irn_link(irn);
1275 if(op->is_remat) continue;
1276 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1278 if(has_reg_class(si, irn)) {
1279 assert(pset_find_ptr(live, irn));
1280 pset_remove_ptr(live, irn);
1283 /* init set of irn's arguments */
1284 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
1285 ir_node *irn_arg = get_irn_n(irn, i);
1286 if(has_reg_class(si, irn_arg)) {
1287 set_insert_keyval(args, irn_arg, (void*)i);
1291 /* check the register pressure after the epilog */
1292 ir_snprintf(buf, sizeof(buf), "check_post_remat_%N", irn);
1293 check_post_remat = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1295 /* iterate over L\U */
1296 pset_foreach(live, tmp) {
1297 if(!set_find_keyval(args, tmp)) {
1298 /* if a live value is not used by irn */
1299 tmp_op = get_irn_link(tmp);
1300 // assert(tmp_op->attr.live_range.op != irn);
1301 lpp_set_factor_fast(si->lpp, check_post_remat, tmp_op->attr.live_range.ilp, 1.0);
1304 /* iterate over following remats and remove possibly defined values again from check_post_remat */
1305 foreach_post_remat(irn, tmp) {
1306 op_t *remat_op = get_irn_link(tmp);
1307 const ir_node *value = remat_op->attr.remat.remat->value;
1308 op_t *val_op = get_irn_link(value);
1310 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
1312 /* values that are defined by remats are not counted */
1313 /* TODO assert(val_op->attr.live_range.ilp)) */
1314 if(val_op->attr.live_range.ilp != ILP_UNDEF) {
1315 lpp_set_factor_fast(si->lpp, check_post_remat, val_op->attr.live_range.ilp, 0.0);
1320 /* new live ranges for values from L\U defined by remat2s or used by remats */
1321 pset_foreach(live, tmp) {
1322 ir_node *value = tmp;//remat_op->attr.remat.remat->value;
1323 op_t *value_op = get_irn_link(value);
1325 if(!set_find_keyval(args, value)) {
1326 ilp_var_t prev_lr = ILP_UNDEF;
1330 foreach_post_remat(irn, remat) {
1331 op_t *remat_op = get_irn_link(remat);
1333 /* if value is being rematerialized by this remat */
1334 if(value == remat_op->attr.remat.remat->value) {
1335 if(cst == ILP_UNDEF) {
1336 /* next_live_range <= prev_live_range + sum remat2s */
1337 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1338 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1339 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1340 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1341 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1342 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1345 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1349 if(cst == ILP_UNDEF) {
1350 foreach_pre_remat(si, irn, remat) {
1354 for (i = 0, n = get_irn_arity(remat); i < n; ++i) {
1355 ir_node *remat_arg = get_irn_n(remat, i);
1357 /* if value is being used by this remat */
1358 if(value == remat_arg) {
1359 /* next_live_range <= prev_live_range */
1360 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1361 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1363 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1364 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1365 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1366 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1369 /* TODO check afterwards whether lr dies after a pre-remat (should not happen) */
1376 if(prev_lr != ILP_UNDEF) {
1377 value_op->attr.live_range.ilp = prev_lr;
1378 value_op->attr.live_range.op = irn;
1383 /* get count of values in my register class defined by irn */
1384 /* also add defined values to check_post_remat; do this before iterating over args */
1385 if(get_irn_mode(irn) == mode_T) {
1386 ir_node *proj = sched_next(irn);
1387 op_t *proj_op = get_irn_link(proj);
1389 while(is_Proj(proj)) {
1390 if(has_reg_class(si, proj)) {
1392 lpp_set_factor_fast(si->lpp, check_post_remat, proj_op->attr.live_range.ilp, 1.0);
1394 proj = sched_next(proj);
1395 proj_op = get_irn_link(proj);
1398 if(has_reg_class(si, irn)) {
1400 lpp_set_factor_fast(si->lpp, check_post_remat, op->attr.live_range.ilp, 1.0);
1403 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
1405 /* count how many regs irn needs for arguments */
1406 k = set_count(args);
1408 /* check the register pressure in the prolog */
1409 /* sum_{L\U} lr <= n - |U| */
1410 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
1411 check_pre = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - k);
1413 /* check the register pressure in the epilog */
1414 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
1415 check_post = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - d);
1417 set_foreach(args, keyval) {
1423 ir_node *arg = keyval->key;
1425 spill = add_to_spill_bb(si, bb, arg);
1427 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
1428 next_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1430 i = (int)keyval->val;
1433 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
1434 op->attr.live_range.reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(si, irn));
1436 /* reload <= mem_out */
1437 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1438 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.reloads[i], 1.0);
1439 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
1441 arg_op = get_irn_link(arg);
1443 /* requirement: arg must be in register for use */
1444 /* reload + remat + live_range == 1 */
1445 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
1446 cst = lpp_add_cst(si->lpp, buf, lpp_equal, 1.0);
1448 lpp_set_factor_fast(si->lpp, cst, next_lr, 1.0);
1449 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.reloads[i], 1.0);
1450 foreach_pre_remat(si, irn, tmp) {
1451 op_t *remat_op = get_irn_link(tmp);
1452 if(remat_op->attr.remat.remat->value == arg) {
1453 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1457 /* the epilog stuff - including post_use, post, post_remat */
1458 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
1459 post_use = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1461 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
1463 /* arg is live throughout epilog if the next live_range is in a register */
1464 if(pset_find_ptr(live, arg)) {
1465 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
1467 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1468 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1469 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
1470 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1472 lpp_set_factor_fast(si->lpp, check_post_remat, arg_op->attr.live_range.ilp, 1.0);
1475 /*forall remat2 which use arg add a similar cst*/
1476 foreach_post_remat(irn, tmp) {
1480 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1481 ir_node *remat_arg = get_irn_n(tmp, i);
1482 op_t *remat_op = get_irn_link(tmp);
1484 if(remat_arg == arg) {
1485 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
1487 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1488 cst = lpp_add_cst(si->lpp, buf, lpp_greater, 0.0);
1489 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1490 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1495 /* new live range begins for each argument */
1496 arg_op->attr.live_range.ilp = next_lr;
1497 arg_op->attr.live_range.op = irn;
1499 pset_insert_ptr(live, arg);
1502 /* start new live ranges for values used by remats */
1503 foreach_pre_remat(si, irn, tmp) {
1507 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1508 ir_node *remat_arg = get_irn_n(tmp, i);
1509 op_t *arg_op = get_irn_link(remat_arg);
1512 if(!has_reg_class(si, remat_arg)) continue;
1514 /* if value is becoming live through use by remat */
1515 if(!pset_find_ptr(live, remat_arg)) {
1516 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1517 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1519 arg_op->attr.live_range.ilp = prev_lr;
1520 arg_op->attr.live_range.op = irn;
1522 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat %+F\n", remat_arg, tmp));
1524 /* TODO ist das hier die richtige Stelle???? */
1525 pset_insert_ptr(live, remat_arg);
1526 add_to_spill_bb(si, bb, remat_arg);
1528 /* TODO check afterwards whether lr dies after a pre-remat (should not happen) */
1532 /* iterate over L\U */
1533 pset_foreach(live, tmp) {
1534 if(!set_find_keyval(args, tmp)) {
1535 /* if a live value is not used by irn */
1536 tmp_op = get_irn_link(tmp);
1537 // assert(tmp_op->attr.live_range.op != irn);
1538 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
1539 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
1543 /* requirements for remats */
1544 foreach_pre_remat(si, irn, tmp) {
1545 op_t *remat_op = get_irn_link(tmp);
1549 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1550 ir_node *remat_arg = get_irn_n(tmp, i);
1551 op_t *arg_op = get_irn_link(remat_arg);
1553 if(!has_reg_class(si, remat_arg)) continue;
1555 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1556 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1557 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1559 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1560 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1562 /* if remat arg is also used by current op then we can use reload placed for this argument */
1563 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
1564 int index = (int)keyval->val;
1566 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.reloads[index], -1.0);
1571 /* requirements for remats2
1573 * TODO unsure if this does the right thing.
1574 * should insert values into set if they do not become live through remat and
1577 foreach_post_remat(irn, tmp) {
1578 op_t *remat_op = get_irn_link(tmp);
1582 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1583 ir_node *remat_arg = get_irn_n(tmp, i);
1584 op_t *arg_op = get_irn_link(remat_arg);
1586 if(!has_reg_class(si, remat_arg)) continue;
1588 /* only for values in L\U, the others are handled with post_use */
1589 if(!set_find_keyval(args, remat_arg)) {
1590 /* remat <= live_rang(remat_arg) */
1591 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
1592 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1594 /* if value is becoming live through use by remat2 */
1595 if(!pset_find_ptr(live, remat_arg)) {
1598 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1599 lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1601 arg_op->attr.live_range.ilp = lr;
1602 arg_op->attr.live_range.op = irn;
1604 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
1606 pset_insert_ptr(live, remat_arg);
1607 add_to_spill_bb(si, bb, remat_arg);
1610 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1611 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1616 /* iterate over following remats and add them to check_post_remat */
1617 foreach_post_remat(irn, tmp) {
1618 op_t *remat_op = get_irn_link(tmp);
1620 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
1622 lpp_set_factor_fast(si->lpp, check_post_remat, remat_op->attr.remat.ilp, 1.0);
1626 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
1628 pset_foreach(live, tmp) {
1629 assert(has_reg_class(si, tmp));
1632 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
1633 ir_node *arg = get_irn_n(irn, i);
1635 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
1643 /* do something at the beginning of the block */
1645 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
1646 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
1648 pset_foreach(live, irn) {
1649 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
1652 /* construct mem_outs for all values */
1654 set_foreach(spill_bb->ilp, spill) {
1655 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
1656 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1658 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
1659 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
1661 if(pset_find_ptr(live, spill->irn)) {
1662 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
1664 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
1665 spill->mem_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1667 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
1672 /* L\U is empty at bb start */
1673 /* arg is live throughout epilog if it is reg_in into this block */
1675 /* check the register pressure at the beginning of the block
1678 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
1679 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1681 pset_foreach(live, irn) {
1682 spill = set_find_spill(spill_bb->ilp, irn);
1685 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
1686 spill->reg_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1688 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
1690 foreach_post_remat(bb, irn) {
1691 op_t *remat_op = get_irn_link(irn);
1693 DBG((si->dbg, LEVEL_4, "\t next post remat: %+F\n", irn));
1694 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
1696 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1699 /* forall remat2 add requirements */
1700 foreach_post_remat(bb, tmp) {
1704 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1705 ir_node *remat_arg = get_irn_n(tmp, i);
1706 op_t *remat_op = get_irn_link(tmp);
1708 if(!has_reg_class(si, remat_arg)) continue;
1710 spill = set_find_spill(spill_bb->ilp, remat_arg);
1713 /* TODO verify this is placed correctly */
1714 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
1715 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1716 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
1717 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1721 /* mem_in/reg_in for live_in values, especially phis and their arguments */
1722 // if(get_Block_n_cfgpreds(bb) > 1) {
1723 pset_foreach(live, irn) {
1728 spill = set_find_spill(spill_bb->ilp, irn);
1729 assert(spill && spill->irn == irn);
1731 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
1732 for (i = 0, n = get_Phi_n_preds(irn); i < n; ++i) {
1735 ir_node *phi_arg = get_Phi_pred(irn, i);
1736 ir_node *bb_p = get_Block_cfgpred_block(bb, i);
1737 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
1740 /* although the phi is in the right regclass one or more of
1741 * its arguments can be in a different one or at least to
1744 if(has_reg_class(si, phi_arg)) {
1745 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
1746 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1747 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
1748 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1750 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
1751 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
1753 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
1756 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
1757 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
1761 /* else assure the value arrives on all paths in the same resource */
1763 for (i = 0, n = get_Block_n_cfgpreds(bb); i < n; ++i) {
1766 ir_node *bb_p = get_Block_cfgpred_block(bb, i);
1767 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
1770 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
1771 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1772 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
1773 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1775 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
1776 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
1778 spill_p = set_find_spill(spill_bb_p->ilp, irn);
1781 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
1782 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
1788 /* first live ranges from reg_ins */
1789 pset_foreach(live, irn) {
1790 op_t *op = get_irn_link(irn);
1792 spill = set_find_spill(spill_bb->ilp, irn);
1793 assert(spill && spill->irn == irn);
1795 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
1796 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1797 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1798 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
1800 foreach_post_remat(bb, tmp) {
1801 op_t *remat_op = get_irn_link(tmp);
1803 if(remat_op->attr.remat.remat->value == irn) {
1804 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1809 /* walk forward now and compute constraints for placing spills */
1810 /* this must only be done for values that are not defined in this block */
1811 pset_foreach(live, irn) {
1812 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", spill->irn, bb);
1813 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1815 spill = set_find_spill(spill_bb->ilp, irn);
1818 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
1819 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
1821 sched_foreach_op(bb, tmp) {
1822 op_t *op = get_irn_link(tmp);
1824 if(is_Phi(tmp)) continue;
1825 assert(!is_Proj(tmp));
1828 ir_node *value = op->attr.remat.remat->value;
1831 /* only collect remats up to the first use of a value */
1832 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
1838 for (i = 0, n = get_irn_arity(tmp); i < n; ++i) {
1839 ir_node *arg = get_irn_n(tmp, i);
1842 /* if a value is used stop collecting remats */
1848 if(cst == ILP_UNDEF) break;
1853 /* 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)
1854 mem_in(phi) -> not mem_in(orig_value) TODO: how does this depend on a certain predecessor?
1857 /* mem_in of mem-phi has associated costs (but first one is free) */
1858 /* define n_mem_copies as positive integer in each predecessor block,
1859 #mem_in into this block from predecessor block - 1 weighted with SPILL_COST*execfreq(predecessor)
1864 /* ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1865 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1866 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1870 ir_snprintf(buf, sizeof(buf), "ce2_%N_%N_%d", op, irn, i);
1871 cst = lpp_add_cst(si->lpp, buf, lpp_less, n_operands);
1872 lpp_set_factor_fast(si->lpp, cst, op_lr->in_mem_var, -1.0);
1873 op_lr->is_remat_var = lpp_add_var(si->lpp, buf, lpp_binary, COST_REMAT);
1882 process_block(ir_node * block, void *data)
1884 spill_ilp_t *si = data;
1886 int n_regs = arch_register_class_n_regs(si->cls);
1887 int n_preds = get_irn_arity(block);
1891 DBG((si->dbg, LEVEL_3, "\nProcessing %+F\n", block));
1894 * compute all possible remats, sorted by operands *and* resulting
1897 ir_node *new_r_Tuple(ir_graph * irg, ir_node * block, int arity, ir_node * in[]);
1898 ir_node *new_r_Proj(ir_graph * irg, ir_node * block, ir_node * arg, ir_mode * mode, long proj);
1899 void *get_irn_link(const ir_node * node);
1900 void set_irn_link(ir_node * node, void *link);
1902 int block_dominates(const ir_node * a, const ir_node * b);
1905 void compute_doms(ir_graph * irg);
1906 void free_dom(ir_graph * irg);
1910 * nochmal: - laufe in dominatorreihenfolge die bloecke ab -
1911 * sammle dabei moeglichkeiten zur rematerialisierung von werten
1912 * (auch inverse etc.) - mehrere listen: fuer zielwert und fuer
1913 * jeden operanden (get_remats_with_operand) - ignoriere projs
1914 * (und phis?) und andere registerklassen - wenn ich nun auf eine
1915 * op treffe füge ich für alle operanden mögliche remats ein
1916 * (keine reloads und spills) in den schedule direkt vor der op
1917 * ein (in der Hoffnung, dass diese nicht wegoptimiert werden...).
1918 * - dann überspringe ich die folgenden projs - nach den projs
1919 * wird der epilog eingefügt (nur remats, keine spills) -
1920 * lebendigkeit neu berechnen
1922 * -durchlaufe den ganzen mist nochmal und emittiere dabei das ILP:
1925 * datenhaltung: - remat merkt sich die zugehörige ilp_var und op
1926 * - op merkt sich (zugehörige remats) - block merkt sich ilp_vars
1929 * nachbearbeitung: - durchlaufe alle instruktionen, dabei entferne
1930 * nicht benötigte remats und füge benötigte reloads ein - wenn
1931 * spill in Grundblock eingesetzt werden soll durchlaufen wir den
1932 * Block von oben runter und setzen das spill an der letzten
1933 * möglichen Stelle vor der ersten Benutzung/Reload ein. -
1934 * Speicherkopienminimierung: teste Speicherwerte auf Interferenz
1935 * und weise Spillkontexte zu. Sorge bei Phis dafuer, dass gleiche
1936 * Kontexte zusammenfliessen (Operanden und Ergebnis hat gleichen
1939 * nachoptimierung: - laufe wieder von oben nach unten durch jeden
1940 * block - wenn ich ein reload finde, dann schiebe es hoch solange
1941 * es der Registerdruck erlaubt (wenn ich es über sein
1942 * zugehöriges spill schiebe ist das ein bug, oder?)
1952 return fabs(x) < 0.00001;
1957 is_spilled(const spill_ilp_t * si, const live_range_t * lr)
1959 return !is_zero(lpp_get_var_sol(si->lpp, lr->in_mem_var));
1964 is_mem_phi(const ir_node * phi, void *data)
1966 spill_ilp_t *si = data;
1967 // return is_spilled(si, get_use_head(si, phi)->closest_use);
1971 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
1973 spill_ilp_t *si = get_irg_link(current_ir_graph);
1975 if(pset_find_ptr(si->all_possible_remats, n)) {
1976 op_t *op = (op_t*)get_irn_link(n);
1977 assert(op && op->is_remat);
1979 if(op->attr.remat.pre) {
1980 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
1982 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
1992 dump_graph_with_remats(ir_graph * irg, const char * suffix)
1994 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
1995 be_dump(irg, suffix, dump_ir_block_graph_sched);
1996 set_dump_node_vcgattr_hook(NULL);
2000 * Edge hook to dump the schedule edges with annotated register pressure.
2003 sched_pressure_edge_hook(FILE *F, ir_node *irn)
2005 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
2006 ir_node *prev = sched_prev(irn);
2007 fprintf(F, "edge:{sourcename:\"");
2009 fprintf(F, "\" targetname:\"");
2011 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
2012 fprintf(F, "\" color:magenta}\n");
2018 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
2020 DUMP_NODE_EDGE_FUNC old = get_dump_node_edge_hook();
2022 dump_consts_local(0);
2023 set_dump_node_edge_hook(sched_pressure_edge_hook);
2024 dump_ir_block_graph(irg, suffix);
2025 set_dump_node_edge_hook(old);
2029 walker_pressure_annotator(ir_node * bb, void * data)
2031 spill_ilp_t *si = data;
2036 pset *live = pset_new_ptr_default();
2039 live_foreach(bb, li) {
2040 irn = (ir_node *) li->irn;
2042 if (live_is_end(li) && has_reg_class(si, irn)) {
2043 pset_insert_ptr(live, irn);
2047 set_irn_link(bb, (void*)pset_count(live));
2049 sched_foreach_reverse(bb, irn) {
2051 set_irn_link(irn, (void*)pset_count(live));
2055 if(has_reg_class(si, irn)) {
2056 pset_remove_ptr(live, irn);
2057 if(is_Proj(irn)) ++projs;
2060 if(!is_Proj(irn)) projs = 0;
2062 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
2063 ir_node *arg = get_irn_n(irn, i);
2065 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
2067 set_irn_link(irn, (void*)pset_count(live)+projs);
2074 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
2076 be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
2081 connect_all_remats_with_keep(spill_ilp_t * si)
2089 n_remats = pset_count(si->all_possible_remats);
2091 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
2094 pset_foreach(si->all_possible_remats, irn) {
2099 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
2101 obstack_free(si->obst, ins);
2106 /** insert a spill at an arbitrary position */
2107 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert, ir_node *ctx)
2109 ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
2110 ir_graph *irg = get_irn_irg(bl);
2111 ir_node *frame = get_irg_frame(irg);
2115 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
2116 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
2118 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn, ctx);
2121 * search the right insertion point. a spill of a phi cannot be put
2122 * directly after the phi, if there are some phis behind the one which
2123 * is spilled. Also, a spill of a Proj must be after all Projs of the
2126 * Here's one special case:
2127 * If the spill is in the start block, the spill must be after the frame
2128 * pointer is set up. This is done by setting insert to the end of the block
2129 * which is its default initialization (see above).
2132 if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
2135 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
2138 sched_add_after(insert, spill);
2143 delete_remat(spill_ilp_t * si, ir_node * remat) {
2146 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2148 sched_remove(remat);
2150 /* kill links to operands */
2151 for (i = -1, n = get_irn_arity(remat); i < n; ++i) {
2152 set_irn_n(remat, i, bad);
2157 clean_remat_info(spill_ilp_t * si)
2162 remat_info_t *remat_info;
2163 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2165 set_foreach(si->remat_info, remat_info) {
2166 if(!remat_info->remats) continue;
2168 pset_foreach(remat_info->remats, remat)
2170 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
2171 set_irn_n(remat->proj, -1, bad);
2172 set_irn_n(remat->proj, 0, bad);
2175 if(get_irn_n_edges(remat->op) == 0) {
2176 for (i = -1, n = get_irn_arity(remat->op); i < n; ++i) {
2177 set_irn_n(remat->op, i, bad);
2182 if(remat_info->remats) del_pset(remat_info->remats);
2183 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
2188 delete_unnecessary_remats(spill_ilp_t * si)
2192 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2196 ir_node *end = get_irg_end(si->chordal_env->irg);
2199 for (i = 0, n = get_irn_arity(si->keep); i < n; ++i) {
2200 ir_node *keep_arg = get_irn_n(si->keep, i);
2201 op_t *arg_op = get_irn_link(keep_arg);
2204 assert(arg_op->is_remat);
2206 name = si->lpp->vars[arg_op->attr.remat.ilp];
2208 if(is_zero(name->value)) {
2209 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
2210 /* TODO check whether reload is preferred over remat (could be bug) */
2211 delete_remat(si, keep_arg);
2213 if(arg_op->attr.remat.pre) {
2214 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
2216 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
2220 set_irn_n(si->keep, i, bad);
2223 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
2224 ir_node *end_arg = get_End_keepalive(end, i);
2226 if(end_arg != si->keep) {
2227 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
2230 keeps = obstack_finish(si->obst);
2231 set_End_keepalives(end, n-1, keeps);
2232 obstack_free(si->obst, keeps);
2235 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
2240 pset_foreach(si->all_possible_remats, remat) {
2241 op_t *remat_op = get_irn_link(remat);
2242 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
2244 if(is_zero(name->value)) {
2245 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
2246 /* TODO check whether reload is preferred over remat (could be bug) */
2247 delete_remat(si, remat);
2249 if(remat_op->attr.remat.pre) {
2250 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
2252 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
2260 * @param before The node after which the spill will be placed in the schedule
2262 /* TODO set context properly */
2264 insert_spill(spill_ilp_t * si, const ir_node * irn, const ir_node * value, const ir_node * before)
2268 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
2270 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
2272 spill = be_spill2(arch_env, irn, before, irn);
2274 defs = set_insert_def(si->values, value);
2277 /* enter into the linked list */
2278 set_irn_link(spill, defs->spills);
2279 defs->spills = spill;
2281 #ifdef KEEPALIVE_SPILLS
2290 * Add remat to list of defs, destroys link field!
2293 insert_remat(spill_ilp_t * si, ir_node * remat)
2296 op_t *remat_op = get_irn_link(remat);
2298 assert(remat_op->is_remat);
2300 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
2303 /* enter into the linked list */
2304 set_irn_link(remat, defs->remats);
2305 defs->remats = remat;
2309 collect_spills(spill_ilp_t * si, ir_node * value, pset * spills, pset * visited)
2314 defs = set_find_def(si->values, value);
2316 if(defs && defs->spills) {
2317 for(next = defs->spills; next; next = get_irn_link(next)) {
2318 pset_insert_ptr(spills, next);
2320 } else if (is_Phi(value)) {
2322 if(!pset_find_ptr(visited, value)) {
2326 pset_insert_ptr(visited, value);
2327 for(i=0, n=get_irn_arity(value); i<n; ++i) {
2328 ir_node *arg = get_irn_n(value, i);
2330 collect_spills(si, arg, spills, visited);
2334 // assert(0 && "Phi operand not spilled");
2339 get_spills_for_value(spill_ilp_t * si, ir_node * value)
2341 pset *spills = pset_new_ptr_default();
2342 pset *visited = pset_new_ptr_default();
2344 collect_spills(si, value, spills, visited);
2351 * Add reload before operation and add to list of defs
2354 insert_reload(spill_ilp_t * si, const ir_node * value, const ir_node * after)
2359 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
2361 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
2363 defs = set_find_def(si->values, value);
2364 /* get a spill of this value */
2365 if((!defs || !defs->spills) && is_Phi(value)) {
2368 spills = get_spills_for_value(si, value);
2370 spill = pset_first(spills);
2374 defs = set_insert_def(si->values, value);
2377 spill = defs->spills;
2379 assert(spill && "no spill placed before reload");
2381 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
2383 /* enter into the linked list */
2384 set_irn_link(reload, defs->remats);
2385 defs->remats = reload;
2392 walker_spill_placer(ir_node * bb, void * data) {
2393 spill_ilp_t *si = (spill_ilp_t*)data;
2395 spill_bb_t *spill_bb = get_irn_link(bb);
2396 pset *spills_to_do = pset_new_ptr_default();
2399 set_foreach(spill_bb->ilp, spill) {
2402 assert(spill->spill > 0);
2404 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2405 name = si->lpp->vars[spill->mem_in];
2406 if(!is_zero(name->value)) {
2407 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F\n", spill->irn));
2411 name = si->lpp->vars[spill->spill];
2412 if(!is_zero(name->value)) {
2413 if(spill->reg_in > 0) {
2414 name = si->lpp->vars[spill->reg_in];
2415 if(!is_zero(name->value)) {
2416 insert_spill(si, spill->irn, spill->irn, bb);
2420 pset_insert_ptr(spills_to_do, spill->irn);
2423 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
2426 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
2427 op_t *op = get_irn_link(irn);
2429 if(be_is_Spill(irn)) continue;
2432 /* TODO fix this if we want to support remats with more than two nodes */
2433 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
2434 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
2436 insert_spill(si, irn, op->attr.remat.remat->value, irn);
2439 if(pset_find_ptr(spills_to_do, irn)) {
2440 pset_remove_ptr(spills_to_do, irn);
2442 insert_spill(si, irn, irn, irn);
2448 assert(pset_count(spills_to_do) == 0);
2450 /* afterwards free data in block */
2451 del_pset(spills_to_do);
2455 walker_reload_placer(ir_node * bb, void * data) {
2456 spill_ilp_t *si = (spill_ilp_t*)data;
2458 spill_bb_t *spill_bb = get_irn_link(bb);
2462 sched_foreach_reverse(bb, irn) {
2463 op_t *op = get_irn_link(irn);
2465 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
2466 if(is_Phi(irn)) break;
2469 if(get_irn_mode(irn) != mode_T) {
2470 insert_remat(si, irn);
2475 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
2476 ir_node *arg = get_irn_n(irn, i);
2478 if(op->attr.live_range.reloads && op->attr.live_range.reloads[i] != ILP_UNDEF) {
2481 name = si->lpp->vars[op->attr.live_range.reloads[i]];
2482 if(!is_zero(name->value)) {
2484 ir_node *insert_pos = irn;
2485 ir_node *prev = sched_prev(insert_pos);
2486 op_t *prev_op = get_irn_link(prev);
2488 /* insert reload before pre-remats */
2489 while(!sched_is_end(prev) && !be_is_Reload(prev) && !be_is_Spill(prev)
2490 && prev_op->is_remat && prev_op->attr.remat.pre) {
2493 prev = sched_prev(insert_pos);
2494 prev_op = get_irn_link(prev);
2497 reload = insert_reload(si, arg, insert_pos);
2499 set_irn_n(irn, i, reload);
2502 #ifdef KEEPALIVE_SPILLS
2512 live_foreach(bb, li) {
2513 ir_node *irn = (ir_node *) li->irn;
2515 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
2518 name = si->lpp->vars[spill_bb->reloads[i]];
2519 if(!is_zero(name->value)) {
2521 ir_node *insert_pos = bb;
2522 ir_node *prev = sched_prev(insert_pos);
2523 op_t *prev_op = get_irn_link(prev);
2525 /* insert reload before pre-remats */
2526 while(!sched_is_end(prev) && !be_is_Reload(prev) && !be_is_Spill(prev)
2527 && prev_op->is_remat && prev_op->attr.remat.pre) {
2530 prev = sched_prev(insert_pos);
2531 prev_op = get_irn_link(prev);
2534 reload = insert_reload(si, irn, insert_pos);
2536 #ifdef KEEPALIVE_SPILLS
2544 del_set(spill_bb->ilp);
2548 walker_collect_used(ir_node * irn, void * data)
2550 lc_bitset_t *used = data;
2552 lc_bitset_set(used, get_irn_idx(irn));
2556 walker_kill_unused(ir_node * bb, void * data)
2558 lc_bitset_t *used = data;
2559 const ir_node *bad = get_irg_bad(get_irn_irg(bb));
2563 for(irn=sched_first(bb); !sched_is_end(irn);) {
2564 ir_node *next = sched_next(irn);
2568 if(!lc_bitset_is_set(used, get_irn_idx(irn))) {
2569 assert(!be_is_Spill(irn) && !be_is_Reload(irn) && "something is fishy, spill or remat is unused");
2573 set_nodes_block(irn, bad);
2574 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
2575 ir_node *arg = get_irn_n(irn, i);
2577 set_irn_n(irn, i, bad);
2585 kill_all_unused_values_in_schedule(spill_ilp_t * si)
2587 lc_bitset_t *used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
2589 irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, used);
2590 irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, used);
2592 lc_bitset_free(used);
2596 print_irn_pset(pset * p)
2600 pset_foreach(p, irn) {
2601 ir_printf("%+F\n", irn);
2606 rewire_uses(spill_ilp_t * si)
2608 dom_front_info_t *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
2611 /* then fix uses of spills */
2612 set_foreach(si->values, defs) {
2615 ir_node *next = defs->remats;
2619 reloads = pset_new_ptr_default();
2622 if(be_is_Reload(next)) {
2623 pset_insert_ptr(reloads, next);
2627 next = get_irn_link(next);
2630 spills = get_spills_for_value(si, defs->value);
2631 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));
2632 if(pset_count(spills) > 1) {
2633 assert(pset_count(reloads) > 0);
2634 // print_irn_pset(spills);
2635 // print_irn_pset(reloads);
2636 be_ssa_constr_set_uses(dfi, spills, reloads);
2644 /* first fix uses of remats and reloads */
2645 set_foreach(si->values, defs) {
2647 ir_node *next = defs->remats;
2650 nodes = pset_new_ptr_default();
2651 pset_insert_ptr(nodes, defs->value);
2654 pset_insert_ptr(nodes, next);
2655 next = get_irn_link(next);
2658 if(pset_count(nodes) > 1) {
2659 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
2660 be_ssa_constr_set(dfi, nodes);
2667 // remove_unused_defs(si);
2669 be_free_dominance_frontiers(dfi);
2673 writeback_results(spill_ilp_t * si)
2675 /* walk through the graph and collect all spills, reloads and remats for a value */
2677 si->values = new_set(cmp_defs, 4096);
2679 DBG((si->dbg, LEVEL_1, "Applying results\n"));
2680 delete_unnecessary_remats(si);
2681 irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
2682 irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
2684 /* clean the remat info! there are still back-edges leading there! */
2685 clean_remat_info(si);
2689 del_set(si->values);
2693 get_n_regs(spill_ilp_t * si)
2695 int arch_n_regs = arch_register_class_n_regs(si->cls);
2699 for(i=0; i<arch_n_regs; i++) {
2700 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
2705 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
2710 #define enqueue(r) if(reloads[ins_pos]) put(); \
2711 assert(reloads[ins_pos]==NULL); \
2712 reload[ins_pos] = (r); \
2713 ins_pos = (ins_pos+1)%si->n_regs; \
2715 #define put() if(reloads[del_pos]) sched_put_before(irn, reloads[del_pos]); \
2716 reloads[del_pos] = NULL; \
2717 del_pos = (del_pos+1)%si->n_regs; \
2721 walker_reload_mover(ir_node * bb, void * data)
2723 spill_ilp_t *si = data;
2726 sched_foreach(bb, irn) {
2727 if(be_is_Reload(irn) && has_reg_class(si, irn)) {
2728 ir_node *reload = irn;
2731 /* move reload upwards */
2733 int pressure = (int)get_irn_link(reload);
2734 if(pressure < si->n_regs) {
2735 irn = sched_prev(reload);
2736 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
2737 sched_remove(reload);
2738 pressure = (int)get_irn_link(irn);
2740 while(pressure < si->n_regs) {
2741 if(sched_is_end(irn) || (be_is_Reload(irn) && has_reg_class(si, irn))) break;
2743 set_irn_link(irn, (void*)(pressure+1));
2744 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
2745 irn = sched_prev(irn);
2747 pressure = (int)get_irn_link(irn);
2750 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
2751 sched_put_after(irn, reload);
2752 // sched_add_after(irn, reload);
2753 // pressure = (int)get_irn_link(sched_next(reload));
2754 // set_irn_link(reload, (void*)(pressure-1));
2761 move_reloads_upward(spill_ilp_t * si)
2763 irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
2767 be_spill_remat(const be_chordal_env_t * chordal_env)
2769 char problem_name[256];
2770 char dump_suffix[256];
2771 char dump_suffix2[256];
2772 char dump_suffix3[256];
2773 struct obstack obst;
2776 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
2777 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
2778 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
2779 ir_snprintf(dump_suffix2, sizeof(dump_suffix3), "-%s-reloads_moved", chordal_env->cls->name);
2781 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
2782 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
2784 obstack_init(&obst);
2785 si.chordal_env = chordal_env;
2787 si.senv = be_new_spill_env(chordal_env, is_mem_phi, &si);
2788 si.cls = chordal_env->cls;
2789 si.lpp = new_lpp(problem_name, lpp_minimize);
2790 si.remat_info = new_set(cmp_remat_info, 4096);
2791 si.all_possible_remats = pset_new_ptr_default();
2792 si.execfreqs = compute_execfreq(chordal_env->irg);
2796 si.n_regs = get_n_regs(&si);
2798 set_irg_link(chordal_env->irg, &si);
2799 compute_doms(chordal_env->irg);
2801 #ifdef COLLECT_REMATS
2802 // irg_block_walk_graph(chordal_env->irg, process_block, NULL, &si);
2803 /* collect remats */
2804 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
2805 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
2808 /* insert possible remats */
2809 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
2810 irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
2811 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
2814 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
2815 connect_all_remats_with_keep(&si);
2816 /* dump graph with inserted remats */
2817 dump_graph_with_remats(chordal_env->irg, dump_suffix);
2821 /* recompute liveness */
2822 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
2823 be_liveness(chordal_env->irg);
2827 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
2828 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
2829 irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
2831 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
2832 irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
2839 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
2840 if ((f = fopen(buf, "wt")) != NULL) {
2841 lpp_dump_plain(si.lpp, f);
2848 DBG((si.dbg, LEVEL_1, "\tSolving %F\n", chordal_env->irg));
2849 lpp_set_time_limit(si.lpp, ILP_TIMEOUT);
2852 lpp_solve_cplex(si.lpp);
2854 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
2856 assert(lpp_is_sol_valid(si.lpp)
2857 && "solution of ILP must be valid");
2859 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, si.lpp->objval));
2861 #ifdef DUMP_SOLUTION
2866 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
2867 if ((f = fopen(buf, "wt")) != NULL) {
2869 for (i = 0; i < si.lpp->var_next; ++i) {
2870 lpp_name_t *name = si.lpp->vars[i];
2871 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
2878 writeback_results(&si);
2882 kill_all_unused_values_in_schedule(&si);
2884 be_liveness(chordal_env->irg);
2885 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
2887 dump_pressure_graph(&si, dump_suffix2);
2889 // TODO fix temporarily exceeded regpressure due to remat2s
2891 // TODO insert copys to fix interferences in memory
2893 // move reloads upwards
2894 move_reloads_upward(&si);
2895 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
2896 dump_pressure_graph(&si, dump_suffix3);
2898 free_dom(chordal_env->irg);
2899 del_pset(si.all_possible_remats);
2900 del_set(si.execfreqs);
2902 obstack_free(&obst, NULL);
2906 #else /* WITH_ILP */
2909 only_that_you_can_compile_without_WITH_ILP_defined(void)
2913 #endif /* WITH_ILP */