X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbecopyheur.c;h=28205e560b7c7e9451fcd8bc3203f7f8965173fa;hb=2adf84106c02caf097c2d6cf1764706bdc437bcc;hp=f21562552acdd9fa3c31b6ba3a924ad8f70c4b03;hpb=3a2cc1d47466d98f2f09a22425d122051a37a5e7;p=libfirm diff --git a/ir/be/becopyheur.c b/ir/be/becopyheur.c index f21562552..28205e560 100644 --- a/ir/be/becopyheur.c +++ b/ir/be/becopyheur.c @@ -23,20 +23,24 @@ #include #endif +#include "debug.h" #include "xmalloc.h" -#include "becopyopt.h" +#include "becopyopt_t.h" #include "becopystat.h" +#include "benodesets.h" #include "bitset.h" +#include "raw_bitset.h" -static firm_dbg_module_t *dbg = NULL; +DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;) -#define SLOTS_PINNED_GLOBAL 256 +#define SEARCH_FREE_COLORS + +#define SLOTS_PINNED_GLOBAL 64 #define SLOTS_CONFLICTS 8 #define SLOTS_CHANGED_NODES 32 -#define MIN(a,b) ((aifg) + return be_ifg_connected(env->ifg, a, b); + else + return values_interfere(env->birg->lv, a, b); +} static int set_cmp_conflict_t(const void *x, const void *y, size_t size) { const conflict_t *xx = x; @@ -101,7 +113,7 @@ static INLINE void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, cons static INLINE int qnode_are_conflicting(const qnode_t *qn, const ir_node *n1, const ir_node *n2) { conflict_t c; /* search for live range interference */ - if (n1!=n2 && nodes_interfere(qn->ou->co->chordal_env, n1, n2)) + if (n1!=n2 && nodes_interfere(qn->ou->co->cenv, n1, n2)) return 1; /* search for recoloring conflicts */ if ((int)n1 < (int)n2) { @@ -124,7 +136,7 @@ static int set_cmp_node_stat_t(const void *x, const void *y, size_t size) { static INLINE node_stat_t *qnode_find_node(const qnode_t *qn, ir_node *irn) { node_stat_t find; find.irn = irn; - return set_find(qn->changed_nodes, &find, sizeof(find), HASH_PTR(irn)); + return set_find(qn->changed_nodes, &find, sizeof(find), nodeset_hash(irn)); } /** @@ -136,7 +148,7 @@ static INLINE node_stat_t *qnode_find_or_insert_node(const qnode_t *qn, ir_node find.irn = irn; find.new_color = NO_COLOR; find.pinned_local = 0; - return set_insert(qn->changed_nodes, &find, sizeof(find), HASH_PTR(irn)); + return set_insert(qn->changed_nodes, &find, sizeof(find), nodeset_hash(irn)); } /** @@ -156,6 +168,7 @@ static INLINE int qnode_get_new_color(const qnode_t *qn, ir_node *irn) { static INLINE void qnode_set_new_color(const qnode_t *qn, ir_node *irn, int color) { node_stat_t *found = qnode_find_or_insert_node(qn, irn); found->new_color = color; + DBG((dbg, LEVEL_3, "\t col(%+F) := %d\n", irn, color)); } /** @@ -178,8 +191,11 @@ static INLINE int qnode_is_pinned_local(const qnode_t *qn, ir_node *irn) { static INLINE void qnode_pin_local(const qnode_t *qn, ir_node *irn) { node_stat_t *found = qnode_find_or_insert_node(qn, irn); found->pinned_local = 1; + if (found->new_color == NO_COLOR) + found->new_color = get_irn_col(qn->ou->co, irn); } + /** * Possible return values of qnode_color_irn() */ @@ -191,144 +207,113 @@ static INLINE void qnode_pin_local(const qnode_t *qn, ir_node *irn) { * Performs virtual re-coloring of node @p n to color @p col. Virtual colors of * other nodes are changed too, as required to preserve correctness. Function is * aware of local and global pinning. Recursive. - * @param irn The node to set the color for - * @param col The color to set + * + * If irn == trigger the color @p col must be used. (the first recoloring) + * If irn != trigger an arbitrary free color may be used. If no color is free, @p col is used. + * + * @param irn The node to set the color for + * @param col The color to set * @param trigger The irn that caused the wish to change the color of the irn + * External callers must call with trigger = irn + * * @return CHANGE_SAVE iff setting the color is possible, with all transitive effects. * CHANGE_IMPOSSIBLE iff conflicts with reg-constraintsis occured. * Else the first conflicting ir_node encountered is returned. * - * ASSUMPTION: Assumes that a life range of a single value can't be split into - * several smaller intervals where other values can live in between. - * This should be true in SSA. */ static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const ir_node *trigger) { - ir_node *res; - struct obstack confl_ob; - ir_node **confl, *cn; - int i, irn_col; - const be_chordal_env_t *chordal_env = qn->ou->co->chordal_env; - const arch_env_t *arch_env = chordal_env->arch_env; - const arch_register_class_t *cls = chordal_env->cls; - - DBG((dbg, LEVEL_3, "\t %+F \tcaused col(%+F) \t%2d --> %2d\n", trigger, irn, qnode_get_new_color(qn, irn), col)); - obstack_init(&confl_ob); - irn_col = qnode_get_new_color(qn, irn); + copy_opt_t *co = qn->ou->co; + const be_chordal_env_t *chordal_env = co->cenv; + const arch_register_class_t *cls = co->cls; + const arch_env_t *arch_env = co->aenv; + int irn_col = qnode_get_new_color(qn, irn); + ir_node *sub_res, *curr; + be_ifg_t *ifg = chordal_env->ifg; + void *iter = be_ifg_neighbours_iter_alloca(ifg); + + DBG((dbg, LEVEL_3, "\t %+F \tcaused col(%+F) \t%2d --> %2d\n", trigger, irn, irn_col, col)); + + /* If the target color is already set do nothing */ if (irn_col == col) { - DBG((dbg, LEVEL_4, "\t Already same color.\n")); - goto ret_save; + DBG((dbg, LEVEL_3, "\t %+F same color\n", irn)); + return CHANGE_SAVE; } + + /* If the irn is pinned, changing color is impossible */ if (pset_find_ptr(pinned_global, irn) || qnode_is_pinned_local(qn, irn)) { - res = irn; - goto ret_confl; + DBG((dbg, LEVEL_3, "\t %+F conflicting\n", irn)); + return irn; } - if (!arch_reg_is_allocatable(arch_env, - irn, - arch_pos_make_out(0), - arch_register_for_index(cls, col))) - goto ret_imposs; - - /* get all nodes which would conflict with this change */ - { - struct obstack q; - int in, out; - ir_node *irn_bl; - - irn_bl = get_nodes_block(irn); - - /* first check for a conflicting node which is 'living in' the irns block */ - { - ir_node *n; - pset *live_ins = put_live_in(irn_bl, pset_new_ptr_default()); - for (n = pset_first(live_ins); n; n = pset_next(live_ins)) { - DBG((dbg, LEVEL_4, "Checking %+F which is live-in at the block\n", n)); - if (arch_irn_has_reg_class(arch_env, n, arch_pos_make_out(0), cls) - && n != trigger && qnode_get_new_color(qn, n) == col - && nodes_interfere(chordal_env, irn, n)) { - - DBG((dbg, LEVEL_4, "\t %+F\ttroubles\n", n)); - obstack_ptr_grow(&confl_ob, n); - pset_break(live_ins); - break; - } - } - del_pset(live_ins); + +#ifdef SEARCH_FREE_COLORS + /* If we resolve conflicts (recursive calls) we can use any unused color. + * In case of the first call @p col must be used. + */ + if (irn != trigger) { + bitset_t *free_cols = bitset_alloca(cls->n_regs); + const arch_register_req_t *req; + ir_node *curr; + int free_col; + + /* Get all possible colors */ + bitset_copy(free_cols, co->cenv->ignore_colors); + bitset_flip_all(free_cols); + + /* Exclude colors not assignable to the irn */ + req = arch_get_register_req(arch_env, irn, -1); + if (arch_register_req_is(req, limited)) { + bitset_t *limited = bitset_alloca(cls->n_regs); + rbitset_copy_to_bitset(req->limited, limited); + bitset_and(free_cols, limited); } - /* setup the queue of blocks. */ - obstack_init(&q); - obstack_ptr_grow(&q, irn_bl); - in = 1; - out = 0; - - /* process the queue. The code below checks for every block dominated - * by the irns one, and in which the irn is live, if there are - * conflicting nodes */ - while (out < in) { - ir_node *curr_bl, *sub_bl; - int i, max; - - curr_bl = ((ir_node **)obstack_base(&q))[out++]; - - /* Add to the result all nodes in the block, which have - * the target color and interfere with the irn */ - for (i = 0, max = get_irn_n_outs(curr_bl); i < max; ++i) { - ir_node *n = get_irn_out(curr_bl, i); - DBG((dbg, LEVEL_4, "Checking %+F defined in same block\n", n)); - if (arch_irn_has_reg_class(arch_env, n, arch_pos_make_out(0), cls) - && n != trigger && qnode_get_new_color(qn, n) == col - && nodes_interfere(chordal_env, irn, n)) { - DBG((dbg, LEVEL_4, "\t %+F\ttroubles\n", n)); - obstack_ptr_grow(&confl_ob, n); - } - } + /* Exclude the color of the irn, because it must _change_ its color */ + bitset_clear(free_cols, irn_col); - /* If irn lives out check i-dominated blocks where the irn lives in */ - /* Fill the queue */ - if (is_live_out(curr_bl, irn)) { - dominates_for_each(curr_bl, sub_bl) - if (is_live_in(sub_bl, irn)) { - obstack_ptr_grow(&q, sub_bl); - in++; - } - } + /* Exclude all colors used by adjacent nodes */ + be_ifg_foreach_neighbour(ifg, iter, irn, curr) + bitset_clear(free_cols, qnode_get_new_color(qn, curr)); + + free_col = bitset_next_set(free_cols, 0); + + if (free_col != -1) { + qnode_set_new_color(qn, irn, free_col); + return CHANGE_SAVE; } - obstack_free(&q, NULL); - obstack_ptr_grow(&confl_ob, NULL); - confl = (ir_node **) obstack_finish(&confl_ob); } +#endif /* SEARCH_FREE_COLORS */ - /* process all nodes which would conflict with this change */ - for (i = 0, cn = confl[0]; cn; cn = confl[++i]) { - ir_node *sub_res; + /* If target color is not allocatable changing color is impossible */ + if (!arch_reg_is_allocatable(arch_env, irn, -1, arch_register_for_index(cls, col))) { + DBG((dbg, LEVEL_3, "\t %+F impossible\n", irn)); + return CHANGE_IMPOSSIBLE; + } - /* try to color the conflicting node cn with the color of the irn itself */ - sub_res = qnode_color_irn(qn, cn, irn_col, irn); - if (sub_res != CHANGE_SAVE) { - res = sub_res; - goto ret_confl; + /* + * If we arrive here changing color may be possible, but there may be conflicts. + * Try to color all conflicting nodes 'curr' with the color of the irn itself. + */ + be_ifg_foreach_neighbour(ifg, iter, irn, curr) { + DBG((dbg, LEVEL_3, "\t Confl %+F(%d)\n", curr, qnode_get_new_color(qn, curr))); + if (qnode_get_new_color(qn, curr) == col && curr != trigger) { + sub_res = qnode_color_irn(qn, curr, irn_col, irn); + if (sub_res != CHANGE_SAVE) { + be_ifg_neighbours_break(ifg, iter); + return sub_res; + } } } - /* if we arrive here all sub changes can be applied, so it's save to change this irn */ -ret_save: - DBG((dbg, LEVEL_3, "\t %+F save\n", irn)); - obstack_free(&confl_ob, NULL); + /* + * If we arrive here, all conflicts were resolved. + * So it is save to change this irn + */ qnode_set_new_color(qn, irn, col); return CHANGE_SAVE; - -ret_imposs: - DBG((dbg, LEVEL_3, "\t %+F impossible\n", irn)); - obstack_free(&confl_ob, NULL); - return CHANGE_IMPOSSIBLE; - -ret_confl: - DBG((dbg, LEVEL_3, "\t %+F conflicting\n", irn)); - obstack_free(&confl_ob, NULL); - return res; } + /** * Tries to set the colors for all members of this queue node; * to the target color qn->color @@ -382,56 +367,96 @@ static int qnode_try_color(const qnode_t *qn) { * the interference and conflict edges of all nodes in a qnode. */ static INLINE void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) { - ir_node **irns; + ir_node **safe, **unsafe; + int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs; + bitset_t *curr, *best; int max, next, pos, curr_weight, best_weight = 0; - bitset_t *best, *curr; - - irns = alloca((ou->node_count-1) * sizeof(*irns)); - best = bitset_alloca(ou->node_count-1); - curr = bitset_alloca(ou->node_count-1); - - /* brute force the best set */ - bitset_set_all(curr); - while ((max = bitset_popcnt(curr)) != 0) { - /* check if curr is a stable set */ - int i, o, is_stable_set = 1; - - /* copy the irns */ - i = 0; - bitset_foreach(curr, pos) - irns[i++] = ou->nodes[1+pos]; - assert(i==max); - - for(i=0; inode_count-1) * sizeof(*safe)); + safe_costs = 0; + safe_count = 0; + unsafe = alloca((ou->node_count-1) * sizeof(*unsafe)); + unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs)); + unsafe_count = 0; + for(i=1; inode_count; ++i) { + int is_safe = 1; + for(o=1; onode_count; ++o) { + if (qnode_are_conflicting(qn, ou->nodes[i], ou->nodes[o])) { + if (i!=o) { + unsafe_costs[unsafe_count] = ou->costs[i]; + unsafe[unsafe_count] = ou->nodes[i]; + ++unsafe_count; } + is_safe = 0; + break; + } + } + if (is_safe) { + safe_costs += ou->costs[i]; + safe[safe_count++] = ou->nodes[i]; + } + } + + - if (is_stable_set) { - /* calc current weigth */ + /* now compute the best set out of the unsafe nodes*/ + best = bitset_alloca(unsafe_count); + + if (unsafe_count > MIS_HEUR_TRIGGER) { + /* Heuristic: Greedy trial and error form index 0 to unsafe_count-1 */ + for (i=0; icosts[1+pos]; + curr_weight += unsafe_costs[pos]; /* any better ? */ if (curr_weight > best_weight) { best_weight = curr_weight; bitset_copy(best, curr); } - } - bitset_minus1(curr); +no_stable_set: + bitset_minus1(curr); + } } /* transfer the best set into the qn */ - qn->mis_size = bitset_popcnt(best)+1; - qn->mis_costs = best_weight; - next = 0; - qn->mis[next++] = ou->nodes[0]; /* the root is alwazs in a max stable set */ + qn->mis_size = 1+safe_count+bitset_popcnt(best); + qn->mis_costs = safe_costs+best_weight; + qn->mis[0] = ou->nodes[0]; /* the root is always in a max stable set */ + next = 1; + for (i=0; imis[next++] = safe[i]; bitset_foreach(best, pos) - qn->mis[next++] = ou->nodes[1+pos]; + qn->mis[next++] = unsafe[pos]; } /** @@ -441,7 +466,7 @@ static INLINE qnode_t *new_qnode(const unit_t *ou, int color) { qnode_t *qn = xmalloc(sizeof(*qn)); qn->ou = ou; qn->color = color; - qn->mis = malloc(ou->node_count * sizeof(*qn->mis)); + qn->mis = xmalloc(ou->node_count * sizeof(*qn->mis)); qn->conflicts = new_set(set_cmp_conflict_t, SLOTS_CONFLICTS); qn->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES); return qn; @@ -492,8 +517,10 @@ static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) { */ static void ou_optimize(unit_t *ou) { int i; - qnode_t *curr, *tmp; - bitset_t *pos_regs = bitset_alloca(ou->co->chordal_env->cls->n_regs); + qnode_t *curr = NULL, *tmp; + const arch_env_t *aenv = ou->co->aenv; + const arch_register_class_t *cls = ou->co->cls; + bitset_t *pos_regs = bitset_alloca(cls->n_regs); DBG((dbg, LEVEL_1, "\tOptimizing unit:\n")); for (i=0; inode_count; ++i) @@ -501,7 +528,15 @@ static void ou_optimize(unit_t *ou) { /* init queue */ INIT_LIST_HEAD(&ou->queue); - arch_get_allocatable_regs(ou->co->chordal_env->arch_env, ou->nodes[0], arch_pos_make_out(0), ou->co->chordal_env->cls, pos_regs); + + arch_get_allocatable_regs(aenv, ou->nodes[0], -1, pos_regs); + + /* exclude ingore colors */ + bitset_andnot(pos_regs, ou->co->cenv->ignore_colors); + + assert(bitset_popcnt(pos_regs) != 0 && "No register is allowed for this node !!?"); + + /* create new qnode */ bitset_foreach(pos_regs, i) ou_insert_qnode(ou, new_qnode(ou, i)); @@ -515,6 +550,7 @@ static void ou_optimize(unit_t *ou) { /* try */ if (qnode_try_color(curr)) break; + /* no success, so re-insert */ del_set(curr->changed_nodes); curr->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES); @@ -524,13 +560,14 @@ static void ou_optimize(unit_t *ou) { /* apply the best found qnode */ if (curr->mis_size >= 2) { node_stat_t *ns; - DBG((dbg, LEVEL_1, "\t Best color: %d Costs: %d/%d\n", curr->color, ou->complete_costs - curr->mis_costs, ou->complete_costs)); + int root_col = qnode_get_new_color(curr, ou->nodes[0]); + DBG((dbg, LEVEL_1, "\t Best color: %d Costs: %d << %d << %d\n", curr->color, ou->min_nodes_costs, ou->all_nodes_costs - curr->mis_costs, ou->all_nodes_costs)); /* globally pin root and all args which have the same color */ pset_insert_ptr(pinned_global, ou->nodes[0]); for (i=1; inode_count; ++i) { ir_node *irn = ou->nodes[i]; int nc = qnode_get_new_color(curr, irn); - if (nc != NO_COLOR && nc == qnode_get_new_color(curr, ou->nodes[0])) + if (nc != NO_COLOR && nc == root_col) pset_insert_ptr(pinned_global, irn); } @@ -542,10 +579,6 @@ static void ou_optimize(unit_t *ou) { set_irn_col(ou->co, ns->irn, ns->new_color); } } - /* - * Enable for checking register allocation after each ou - * be_ra_chordal_check(ou->co->chordal_env); - */ } /* free best qnode (curr) and queue */ @@ -554,13 +587,11 @@ static void ou_optimize(unit_t *ou) { free_qnode(curr); } -void co_heur_opt(copy_opt_t *co) { +int co_solve_heuristic(copy_opt_t *co) { unit_t *curr; - dbg = firm_dbg_register("ir.be.copyoptheur"); - if (!strcmp(co->name, DEBUG_IRG)) - firm_dbg_set_mask(dbg, DEBUG_IRG_LVL_HEUR); - else - firm_dbg_set_mask(dbg, DEBUG_LVL_HEUR); + FIRM_DBG_REGISTER(dbg, "ir.be.copyoptheur"); + + ASSERT_OU_AVAIL(co); pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL); list_for_each_entry(unit_t, curr, &co->units, units) @@ -568,4 +599,5 @@ void co_heur_opt(copy_opt_t *co) { ou_optimize(curr); del_pset(pinned_global); + return 0; }