X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbecopyopt.c;h=ca32b86b779e7ee16bb16c762e4ac13c14c5eb61;hb=bb9f2e36362333c6635b89f5258171b06c786608;hp=4494fb7223a71c0c62c92056145e76732d27bac8;hpb=f6414b3b350e99ef7f9cbabf07386026eb853625;p=libfirm diff --git a/ir/be/becopyopt.c b/ir/be/becopyopt.c index 4494fb722..ca32b86b7 100644 --- a/ir/be/becopyopt.c +++ b/ir/be/becopyopt.c @@ -1,5 +1,5 @@ /* - * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved. + * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved. * * This file is part of libFirm. * @@ -29,9 +29,7 @@ * - Register-constrained nodes * - Two-address code instructions */ -#ifdef HAVE_CONFIG_H #include "config.h" -#endif #include "execfreq.h" #include "xmalloc.h" @@ -60,15 +58,12 @@ #include "belive_t.h" #include "beinsn_t.h" #include "besched_t.h" -#include "benodesets.h" -#include "bejavacoal.h" #include "bestatevent.h" #include "beirg_t.h" #include "error.h" -#include -#include -#include +#include "lc_opts.h" +#include "lc_opts_enum.h" #define DUMP_BEFORE 1 #define DUMP_AFTER 2 @@ -107,13 +102,9 @@ static const lc_opt_enum_mask_items_t algo_items[] = { { "none", CO_ALGO_NONE }, { "heur", CO_ALGO_HEUR }, { "heur2", CO_ALGO_HEUR2 }, -#ifdef WITH_JVM { "heur3", CO_ALGO_HEUR3 }, -#endif /* WITH_JVM */ { "heur4", CO_ALGO_HEUR4 }, -#ifdef WITH_ILP { "ilp", CO_ALGO_ILP }, -#endif /* WITH_ILP */ { NULL, 0 } }; @@ -201,9 +192,8 @@ copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, cost_fct_t get_costs) FIRM_DBG_REGISTER(dbg, "ir.be.copyopt"); - co = xcalloc(1, sizeof(*co)); + co = XMALLOCZ(copy_opt_t); co->cenv = chordal_env; - co->aenv = chordal_env->birg->main_env->arch_env; co->irg = chordal_env->irg; co->cls = chordal_env->cls; co->get_costs = get_costs; @@ -212,7 +202,7 @@ copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, cost_fct_t get_costs) s2 = get_entity_name(get_irg_entity(co->irg)); s3 = chordal_env->cls->name; len = strlen(s1) + strlen(s2) + strlen(s3) + 5; - co->name = xmalloc(len); + co->name = XMALLOCN(char, len); snprintf(co->name, len, "%s__%s__%s", s1, s2, s3); return co; @@ -223,19 +213,24 @@ void free_copy_opt(copy_opt_t *co) { free(co); } -int co_is_optimizable_root(const copy_opt_t *co, ir_node *irn) { +/** + * Checks if a node is optimizable, viz. has something to do with coalescing + * @param irn The irn to check + */ +static int co_is_optimizable_root(ir_node *irn) +{ const arch_register_req_t *req; const arch_register_t *reg; - if (arch_irn_is(co->aenv, irn, ignore)) + if (arch_irn_is(irn, ignore)) return 0; - reg = arch_get_irn_register(co->aenv, irn); + reg = arch_get_irn_register(irn); if (arch_register_type_is(reg, ignore)) return 0; - req = arch_get_register_req(co->aenv, irn, -1); - if (is_Reg_Phi(irn) || is_Perm_Proj(co->aenv, irn) || is_2addr_code(req)) + req = arch_get_register_req(irn, -1); + if (is_Reg_Phi(irn) || is_Perm_Proj(irn) || is_2addr_code(req)) return 1; return 0; @@ -387,11 +382,11 @@ static void co_collect_units(ir_node *irn, void *env) { if (!is_curr_reg_class(co, irn)) return; - if (!co_is_optimizable_root(co, irn)) + if (!co_is_optimizable_root(irn)) return; /* Init a new unit */ - unit = xcalloc(1, sizeof(*unit)); + unit = XMALLOCZ(unit_t); unit->co = co; unit->node_count = 1; INIT_LIST_HEAD(&unit->queue); @@ -402,8 +397,8 @@ static void co_collect_units(ir_node *irn, void *env) { /* init */ arity = get_irn_arity(irn); - unit->nodes = xmalloc((arity+1) * sizeof(*unit->nodes)); - unit->costs = xmalloc((arity+1) * sizeof(*unit->costs)); + unit->nodes = XMALLOCN(ir_node*, arity + 1); + unit->costs = XMALLOCN(int, arity + 1); unit->nodes[0] = irn; /* fill */ @@ -422,7 +417,7 @@ static void co_collect_units(ir_node *irn, void *env) { /* Else insert the argument of the phi to the members of this ou */ DBG((dbg, LEVEL_1, "\t Member: %+F\n", arg)); - if (! arch_irn_is(co->aenv, arg, ignore)) { + if (!arch_irn_is(arg, ignore)) { /* Check if arg has occurred at a prior position in the arg/list */ arg_pos = 0; for (o=1; onode_count; ++o) { @@ -437,40 +432,61 @@ static void co_collect_units(ir_node *irn, void *env) { unit->nodes[unit->node_count] = arg; unit->costs[unit->node_count] = co->get_costs(co, irn, arg, i); unit->node_count++; - } else { /* arg has occured before in same phi */ + } else { /* arg has occurred before in same phi */ /* increase costs for existing arg */ unit->costs[arg_pos] += co->get_costs(co, irn, arg, i); } } } - unit->nodes = xrealloc(unit->nodes, unit->node_count * sizeof(*unit->nodes)); - unit->costs = xrealloc(unit->costs, unit->node_count * sizeof(*unit->costs)); - } else - - /* Proj of a perm with corresponding arg */ - if (is_Perm_Proj(co->aenv, irn)) { + unit->nodes = XREALLOC(unit->nodes, ir_node*, unit->node_count); + unit->costs = XREALLOC(unit->costs, int, unit->node_count); + } else if (is_Perm_Proj(irn)) { + /* Proj of a perm with corresponding arg */ assert(!nodes_interfere(co->cenv, irn, get_Perm_src(irn))); - unit->nodes = xmalloc(2 * sizeof(*unit->nodes)); - unit->costs = xmalloc(2 * sizeof(*unit->costs)); + unit->nodes = XMALLOCN(ir_node*, 2); + unit->costs = XMALLOCN(int, 2); unit->node_count = 2; unit->nodes[0] = irn; unit->nodes[1] = get_Perm_src(irn); unit->costs[1] = co->get_costs(co, irn, unit->nodes[1], -1); } else { - const arch_register_req_t *req = - arch_get_register_req(co->aenv, irn, -1); + const arch_register_req_t *req = arch_get_register_req(irn, -1); /* Src == Tgt of a 2-addr-code instruction */ if (is_2addr_code(req)) { - ir_node *other = get_irn_n(skip_Proj(irn), req->other_same); - if (!arch_irn_is(co->aenv, other, ignore) && - !nodes_interfere(co->cenv, irn, other)) { - unit->nodes = xmalloc(2 * sizeof(*unit->nodes)); - unit->costs = xmalloc(2 * sizeof(*unit->costs)); - unit->node_count = 2; - unit->nodes[0] = irn; - unit->nodes[1] = other; - unit->costs[1] = co->get_costs(co, irn, other, -1); + const unsigned other = req->other_same; + int count = 0; + int i; + + for (i = 0; (1U << i) <= other; ++i) { + if (other & (1U << i)) { + ir_node *o = get_irn_n(skip_Proj(irn), i); + if (!arch_irn_is(o, ignore) && + !nodes_interfere(co->cenv, irn, o)) { + ++count; + } + } + } + + if (count != 0) { + int k = 0; + ++count; + unit->nodes = XMALLOCN(ir_node*, count); + unit->costs = XMALLOCN(int, count); + unit->node_count = count; + unit->nodes[k++] = irn; + + for (i = 0; 1U << i <= other; ++i) { + if (other & (1U << i)) { + ir_node *o = get_irn_n(skip_Proj(irn), i); + if (!arch_irn_is(o, ignore) && + !nodes_interfere(co->cenv, irn, o)) { + unit->nodes[k] = o; + unit->costs[k] = co->get_costs(co, irn, o, -1); + ++k; + } + } + } } } else { assert(0 && "This is not an optimizable node!"); @@ -507,12 +523,11 @@ static int compare_ous(const void *k1, const void *k2) { const unit_t *u2 = *((const unit_t **) k2); int i, o, u1_has_constr, u2_has_constr; arch_register_req_t req; - const arch_env_t *aenv = u1->co->aenv; /* Units with constraints come first */ u1_has_constr = 0; for (i=0; inode_count; ++i) { - arch_get_register_req(aenv, &req, u1->nodes[i], -1); + arch_get_register_req(&req, u1->nodes[i], -1); if (arch_register_req_is(&req, limited)) { u1_has_constr = 1; break; @@ -521,7 +536,7 @@ static int compare_ous(const void *k1, const void *k2) { u2_has_constr = 0; for (i=0; inode_count; ++i) { - arch_get_register_req(aenv, &req, u2->nodes[i], -1); + arch_get_register_req(&req, u2->nodes[i], -1); if (arch_register_req_is(&req, limited)) { u2_has_constr = 1; break; @@ -637,11 +652,11 @@ int co_get_copy_costs(const copy_opt_t *co) { ASSERT_OU_AVAIL(co); list_for_each_entry(unit_t, curr, &co->units, units) { - int root_col = get_irn_col(co, curr->nodes[0]); + int root_col = get_irn_col(curr->nodes[0]); DBG((dbg, LEVEL_1, " %3d costs for root %+F color %d\n", curr->inevitable_costs, curr->nodes[0], root_col)); res += curr->inevitable_costs; for (i=1; inode_count; ++i) { - int arg_col = get_irn_col(co, curr->nodes[i]); + int arg_col = get_irn_col(curr->nodes[i]); if (root_col != arg_col) { DBG((dbg, LEVEL_1, " %3d for arg %+F color %d\n", curr->costs[i], curr->nodes[i], arg_col)); res += curr->costs[i]; @@ -679,7 +694,7 @@ void co_complete_stats(const copy_opt_t *co, co_complete_stats_t *stat) stat->aff_edges += 1; stat->max_costs += neigh->costs; - if(get_irn_col(co, an->irn) != get_irn_col(co, neigh->irn)) { + if (get_irn_col(an->irn) != get_irn_col(neigh->irn)) { stat->costs += neigh->costs; stat->unsatisfied_edges += 1; } @@ -723,7 +738,7 @@ static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) { new_node.irn = n1; new_node.degree = 0; new_node.neighbours = NULL; - node = set_insert(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn)); + node = set_insert(co->nodes, &new_node, sizeof(new_node), hash_irn(new_node.irn)); for (nbr = node->neighbours; nbr; nbr = nbr->next) if (nbr->irn == n2) { @@ -746,7 +761,7 @@ static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) { nbr->costs += costs; } -static INLINE void add_edges(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) { +static inline void add_edges(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) { if (! be_ifg_connected(co->cenv->ifg, n1, n2)) { add_edge(co, n1, n2, costs); add_edge(co, n2, n1, costs); @@ -758,34 +773,35 @@ static void build_graph_walker(ir_node *irn, void *env) { int pos, max; const arch_register_t *reg; - if (!is_curr_reg_class(co, irn) || arch_irn_is(co->aenv, irn, ignore)) + if (!is_curr_reg_class(co, irn) || arch_irn_is(irn, ignore)) return; - reg = arch_get_irn_register(co->aenv, irn); + reg = arch_get_irn_register(irn); if (arch_register_type_is(reg, ignore)) return; - /* Phis */ - if (is_Reg_Phi(irn)) + if (is_Reg_Phi(irn)) { /* Phis */ for (pos=0, max=get_irn_arity(irn); posget_costs(co, irn, arg, pos)); } - - /* Perms */ - else if (is_Perm_Proj(co->aenv, irn)) { + } else if (is_Perm_Proj(irn)) { /* Perms */ ir_node *arg = get_Perm_src(irn); add_edges(co, irn, arg, co->get_costs(co, irn, arg, 0)); } - - /* 2-address code */ - else { - const arch_register_req_t *req = - arch_get_register_req(co->aenv, irn, -1); + else { /* 2-address code */ + const arch_register_req_t *req = arch_get_register_req(irn, -1); if (is_2addr_code(req)) { - ir_node *other = get_irn_n(skip_Proj(irn), req->other_same); - if (! arch_irn_is(co->aenv, other, ignore)) - add_edges(co, irn, other, co->get_costs(co, irn, other, 0)); + const unsigned other = req->other_same; + int i; + + for (i = 0; 1U << i <= other; ++i) { + if (other & (1U << i)) { + ir_node *other = get_irn_n(skip_Proj(irn), i); + if (!arch_irn_is(other, ignore)) + add_edges(co, irn, other, co->get_costs(co, irn, other, 0)); + } + } } } } @@ -813,21 +829,45 @@ int co_gs_is_optimizable(copy_opt_t *co, ir_node *irn) { ASSERT_GS_AVAIL(co); new_node.irn = irn; - n = set_find(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn)); + n = set_find(co->nodes, &new_node, sizeof(new_node), hash_irn(new_node.irn)); if (n) { return (n->degree > 0); } else return 0; } +static int co_dump_appel_disjoint_constraints(const copy_opt_t *co, ir_node *a, ir_node *b) +{ + ir_node *nodes[] = { a, b }; + bitset_t *constr[] = { NULL, NULL }; + const arch_register_req_t *req; + int j; + + constr[0] = bitset_alloca(co->cls->n_regs); + constr[1] = bitset_alloca(co->cls->n_regs); + + for (j = 0; j < 2; ++j) { + req = arch_get_register_req(nodes[j], BE_OUT_POS(0)); + if(arch_register_req_is(req, limited)) + rbitset_copy_to_bitset(req->limited, constr[j]); + else + bitset_set_all(constr[j]); + + } + + return !bitset_intersect(constr[0], constr[1]); +} + void co_dump_appel_graph(const copy_opt_t *co, FILE *f) { be_ifg_t *ifg = co->cenv->ifg; int *color_map = alloca(co->cls->n_regs * sizeof(color_map[0])); + int *node_map = XMALLOCN(int, get_irg_last_idx(co->irg) + 1); ir_node *irn; void *it, *nit; - int i, n, n_regs; + int n, n_regs; + unsigned i; n_regs = 0; for(i = 0; i < co->cls->n_regs; ++i) { @@ -845,21 +885,21 @@ void co_dump_appel_graph(const copy_opt_t *co, FILE *f) n = n_regs; be_ifg_foreach_node(ifg, it, irn) { - if(!arch_irn_is(co->aenv, irn, ignore)) - set_irn_link(irn, INT_TO_PTR(n++)); + if (!arch_irn_is(irn, ignore)) + node_map[get_irn_idx(irn)] = n++; } fprintf(f, "%d %d\n", n, n_regs); be_ifg_foreach_node(ifg, it, irn) { - if(!arch_irn_is(co->aenv, irn, ignore)) { - int idx = PTR_TO_INT(get_irn_link(irn)); + if (!arch_irn_is(irn, ignore)) { + int idx = node_map[get_irn_idx(irn)]; affinity_node_t *a = get_affinity_info(co, irn); const arch_register_req_t *req; ir_node *adj; - req = arch_get_register_req(co->aenv, irn, BE_OUT_POS(0)); + req = arch_get_register_req(irn, BE_OUT_POS(0)); if(arch_register_req_is(req, limited)) { for(i = 0; i < co->cls->n_regs; ++i) { if(!rbitset_is_set(req->limited, i) && color_map[i] >= 0) @@ -867,10 +907,10 @@ void co_dump_appel_graph(const copy_opt_t *co, FILE *f) } } - be_ifg_foreach_neighbour(ifg, nit, irn, adj) { - if(!arch_irn_is(co->aenv, adj, ignore)) { - int adj_idx = PTR_TO_INT(get_irn_link(adj)); + if (!arch_irn_is(adj, ignore) && + !co_dump_appel_disjoint_constraints(co, irn, adj)) { + int adj_idx = node_map[get_irn_idx(adj)]; if(idx < adj_idx) fprintf(f, "%d %d -1\n", idx, adj_idx); } @@ -880,8 +920,8 @@ void co_dump_appel_graph(const copy_opt_t *co, FILE *f) neighb_t *n; co_gs_foreach_neighb(a, n) { - if(!arch_irn_is(co->aenv, n->irn, ignore)) { - int n_idx = PTR_TO_INT(get_irn_link(n->irn)); + if (!arch_irn_is(n->irn, ignore)) { + int n_idx = node_map[get_irn_idx(n->irn)]; if(idx < n_idx) fprintf(f, "%d %d %d\n", idx, n_idx, (int) n->costs); } @@ -889,331 +929,8 @@ void co_dump_appel_graph(const copy_opt_t *co, FILE *f) } } } -} - -typedef struct _appel_clique_walker_t { - ir_phase ph; - const copy_opt_t *co; - int curr_nr; - int node_count; - FILE *f; - int dumb; - int *color_map; - struct obstack obst; -} appel_clique_walker_t; - -typedef struct _appel_block_info_t { - int *live_end_nr; - int *live_in_nr; - int *phi_nr; - ir_node **live_end; - ir_node **live_in; - ir_node **phi; - int n_live_end; - int n_live_in; - int n_phi; -} appel_block_info_t; - -static int appel_aff_weight(const appel_clique_walker_t *env, ir_node *bl) -{ -#if 0 - double freq = get_block_execfreq(env->co->cenv->execfreq, bl); - int res = (int) freq; - return res == 0 ? 1 : res; -#else - ir_loop *loop = get_irn_loop(bl); - (void) env; - if(loop) { - int d = get_loop_depth(loop); - return 1 + d * d; - } - return 1; -#endif -} - -static void *appel_clique_walker_irn_init(ir_phase *phase, ir_node *irn, void *old) -{ - appel_block_info_t *res = NULL; - (void) old; - - if(is_Block(irn)) { - appel_clique_walker_t *d = (void *) phase; - res = phase_alloc(phase, sizeof(res[0])); - res->phi_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr)); - res->live_end_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr)); - res->live_in_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in_nr)); - res->live_end = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end)); - res->live_in = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in)); - res->phi = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in)); - } - - return res; -} - -typedef struct _insn_list_t { - be_insn_t *insn; - struct list_head list; -} insn_list_t; - -static int appel_get_live_end_nr(appel_clique_walker_t *env, ir_node *bl, ir_node *irn) -{ - appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl); - int i; - - for(i = 0; i < bli->n_live_end; ++i) - if(bli->live_end[i] == irn) - return bli->live_end_nr[i]; - - return -1; -} - -static int appel_dump_clique(appel_clique_walker_t *env, pset *live, ir_node *bl, int curr_nr, int start_nr) -{ - ir_node **live_arr = alloca(env->co->cls->n_regs * sizeof(live_arr[0])); - ir_node *irn; - int n_live; - int j; - - n_live = 0; - foreach_pset(live, irn) - live_arr[n_live++] = irn; - - /* dump the live after clique */ - if(!env->dumb) { - for(j = 0; j < n_live; ++j) { - int k; - - for(k = j + 1; k < n_live; ++k) { - fprintf(env->f, "%d %d -1 ", curr_nr + j, curr_nr + k); - } - fprintf(env->f, "\n"); - } - } - - /* dump the affinities */ - for(j = 0; !env->dumb && j < n_live; ++j) { - ir_node *irn = live_arr[j]; - int old_nr = PTR_TO_INT(get_irn_link(irn)); - - /* if the node was already live in the last insn dump the affinity */ - if(old_nr > start_nr) { - int weight = appel_aff_weight(env, bl); - fprintf(env->f, "%d %d %d\n", old_nr, curr_nr + j, weight); - } - } - - /* set the current numbers into the link field. */ - for(j = 0; j < n_live; ++j) { - ir_node *irn = live_arr[j]; - set_irn_link(irn, INT_TO_PTR(curr_nr + j)); - } - - return curr_nr + n_live; -} - -static void appel_walker(ir_node *bl, void *data) -{ - appel_clique_walker_t *env = data; - appel_block_info_t *bli = phase_get_or_set_irn_data(&env->ph, bl); - struct obstack *obst = &env->obst; - void *base = obstack_base(obst); - pset *live = pset_new_ptr_default(); - be_lv_t *lv = env->co->cenv->birg->lv; - - int n_insns = 0; - int n_nodes = 0; - int start_nr = env->curr_nr; - int curr_nr = start_nr; - - be_insn_env_t insn_env; - int i, j; - ir_node *irn; - be_insn_t **insns; - - insn_env.aenv = env->co->aenv; - insn_env.cls = env->co->cls; - insn_env.obst = obst; - insn_env.ignore_colors = env->co->cenv->ignore_colors; - - /* Guess how many insns will be in this block. */ - sched_foreach(bl, irn) - n_nodes++; - - bli->n_phi = 0; - insns = xmalloc(n_nodes * sizeof(insns[0])); - - /* Put all insns in an array. */ - irn = sched_first(bl); - while(!sched_is_end(irn)) { - be_insn_t *insn; - insn = be_scan_insn(&insn_env, irn); - insns[n_insns++] = insn; - irn = insn->next_insn; - } - - DBG((dbg, LEVEL_2, "%+F\n", bl)); - be_liveness_end_of_block(lv, env->co->aenv, env->co->cls, bl, live); - - /* Generate the bad and ugly. */ - for(i = n_insns - 1; i >= 0; --i) { - be_insn_t *insn = insns[i]; - - /* The first live set has to be saved in the block border set. */ - if(i == n_insns - 1) { - j = 0; - foreach_pset(live, irn) { - bli->live_end[j] = irn; - bli->live_end_nr[j] = curr_nr + j; - ++j; - } - bli->n_live_end = j; - } - - if(!env->dumb) { - for(j = 0; j < insn->use_start; ++j) { - ir_node *op = insn->ops[j].carrier; - bitset_t *adm = insn->ops[j].regs; - int k; - int nr; - - if(!insn->ops[j].has_constraints) - continue; - - nr = 0; - foreach_pset(live, irn) { - if(irn == op) { - pset_break(live); - break; - } - ++nr; - } - - assert(nr < pset_count(live)); - - for(k = 0; k < env->co->cls->n_regs; ++k) { - int mapped_col = env->color_map[k]; - if(mapped_col >= 0 && !bitset_is_set(adm, k) && !bitset_is_set(env->co->cenv->ignore_colors, k)) - fprintf(env->f, "%d %d -1\n", curr_nr + nr, mapped_col); - } - } - } - - /* dump the clique and update the stuff. */ - curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr); - - /* remove all defs. */ - for(j = 0; j < insn->use_start; ++j) - pset_remove_ptr(live, insn->ops[j].carrier); - - if(is_Phi(insn->irn) && arch_irn_consider_in_reg_alloc(env->co->aenv, env->co->cls, insn->irn)) { - bli->phi[bli->n_phi] = insn->irn; - bli->phi_nr[bli->n_phi] = PTR_TO_INT(get_irn_link(insn->irn)); - bli->n_phi++; - } - - /* add all uses */ - else - for(j = insn->use_start; j < insn->n_ops; ++j) - pset_insert_ptr(live, insn->ops[j].carrier); - } - /* print the start clique. */ - curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr); - - i = 0; - foreach_pset(live, irn) { - bli->live_in[i] = irn; - bli->live_in_nr[i] = PTR_TO_INT(get_irn_link(irn)); - ++i; - } - bli->n_live_in = i; - - del_pset(live); - free(insns); - obstack_free(obst, base); - env->curr_nr = curr_nr; -} - -static void appel_inter_block_aff(ir_node *bl, void *data) -{ - appel_clique_walker_t *env = data; - appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl); - - int i, j, n; - - for(i = 0; i < bli->n_live_in; ++i) { - ir_node *irn = bli->live_in[i]; - - for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) { - ir_node *pred = get_Block_cfgpred_block(bl, j); - - int nr = appel_get_live_end_nr(env, pred, irn); - assert(nr >= 0); - fprintf(env->f, "%d %d 1\n", bli->live_in_nr[i], nr); - } - } - - for(i = 0; i < bli->n_phi; ++i) { - ir_node *irn = bli->phi[i]; - - for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) { - ir_node *pred = get_Block_cfgpred_block(bl, j); - ir_node *op = get_irn_n(irn, j); - - int nr = appel_get_live_end_nr(env, pred, op); - assert(nr >= 0); - fprintf(env->f, "%d %d 1\n", bli->phi_nr[i], nr); - } - } - -} - -void co_dump_appel_graph_cliques(const copy_opt_t *co, FILE *f) -{ - int i; - int n_colors; - appel_clique_walker_t env; - bitset_t *adm = bitset_alloca(co->cls->n_regs); - be_lv_t *lv = co->cenv->birg->lv; - - be_liveness_recompute(lv); - obstack_init(&env.obst); - phase_init(&env.ph, "appel_clique_dumper", co->irg, PHASE_DEFAULT_GROWTH, appel_clique_walker_irn_init, NULL); - env.curr_nr = co->cls->n_regs; - env.co = co; - env.f = f; - - bitset_copy(adm, co->cenv->ignore_colors); - bitset_flip_all(adm); - - /* Make color map. */ - env.color_map = alloca(co->cls->n_regs * sizeof(env.color_map[0])); - for(i = 0, n_colors = 0; i < co->cls->n_regs; ++i) { - const arch_register_t *reg = &co->cls->regs[i]; - env.color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_colors++; - } - - env.dumb = 1; - env.curr_nr = n_colors; - irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL); - irg_block_walk_graph(co->irg, appel_walker, NULL, &env); - - fprintf(f, "%d %d\n", env.curr_nr, n_colors); - - /* make the first k nodes interfere */ - for(i = 0; i < n_colors; ++i) { - int j; - for(j = i + 1; j < n_colors; ++j) - fprintf(f, "%d %d -1 ", i, j); - fprintf(f, "\n"); - } - - env.dumb = 0; - env.curr_nr = n_colors; - irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL); - irg_block_walk_graph(co->irg, appel_walker, NULL, &env); - irg_block_walk_graph(co->irg, appel_inter_block_aff, NULL, &env); - obstack_free(&env.obst, NULL); + xfree(node_map); } /* @@ -1277,31 +994,28 @@ static void ifg_dump_graph_attr(FILE *f, void *self) static int ifg_is_dump_node(void *self, ir_node *irn) { - co_ifg_dump_t *cod = self; - return !arch_irn_is(cod->co->aenv, irn, ignore); + (void)self; + return !arch_irn_is(irn, ignore); } static void ifg_dump_node_attr(FILE *f, void *self, ir_node *irn) { co_ifg_dump_t *env = self; - const arch_register_t *reg = arch_get_irn_register(env->co->aenv, irn); + const arch_register_t *reg = arch_get_irn_register(irn); const arch_register_req_t *req; int limited; - req = arch_get_register_req(env->co->aenv, irn, BE_OUT_POS(0)); + req = arch_get_register_req(irn, BE_OUT_POS(0)); limited = arch_register_req_is(req, limited); if(env->flags & CO_IFG_DUMP_LABELS) { ir_fprintf(f, "label=\"%+F", irn); -#if 0 - // TODO fix this... if((env->flags & CO_IFG_DUMP_CONSTR) && limited) { bitset_t *bs = bitset_alloca(env->co->cls->n_regs); - req.limited(req.limited_env, bs); + rbitset_copy_to_bitset(req->limited, bs); ir_fprintf(f, "\\n%B", bs); } -#endif ir_fprintf(f, "\" "); } else { fprintf(f, "label=\"\" shape=point " ); @@ -1320,12 +1034,12 @@ static void ifg_dump_at_end(FILE *file, void *self) affinity_node_t *a; co_gs_foreach_aff_node(env->co, a) { - const arch_register_t *ar = arch_get_irn_register(env->co->aenv, a->irn); + const arch_register_t *ar = arch_get_irn_register(a->irn); unsigned aidx = get_irn_idx(a->irn); neighb_t *n; co_gs_foreach_neighb(a, n) { - const arch_register_t *nr = arch_get_irn_register(env->co->aenv, n->irn); + const arch_register_t *nr = arch_get_irn_register(n->irn); unsigned nidx = get_irn_idx(n->irn); if(aidx < nidx) { @@ -1395,10 +1109,16 @@ static co_algo_info_t algos[] = { { void_algo, "none", 0 }, { co_solve_heuristic, "heur1", 0 }, { co_solve_heuristic_new, "heur2", 0 }, +#ifdef WITH_JVM { co_solve_heuristic_java, "heur3", 0 }, +#else + { NULL, "heur3", 0 }, +#endif { co_solve_heuristic_mst, "heur4", 0 }, #ifdef WITH_ILP { co_solve_ilp2, "ilp", 1 }, +#else + { NULL, "ilp", 1 }, #endif { NULL, "", 0 } }; @@ -1416,8 +1136,19 @@ static FILE *my_open(const be_chordal_env_t *env, const char *prefix, const char { FILE *result; char buf[1024]; + size_t i, n; + char *tu_name; + + n = strlen(env->birg->main_env->cup_name); + tu_name = XMALLOCN(char, n + 1); + strcpy(tu_name, env->birg->main_env->cup_name); + for (i = 0; i < n; ++i) + if (tu_name[i] == '.') + tu_name[i] = '_'; - ir_snprintf(buf, sizeof(buf), "%s%F_%s%s", prefix, env->irg, env->cls->name, suffix); + + ir_snprintf(buf, sizeof(buf), "%s%s_%F_%s%s", prefix, tu_name, env->irg, env->cls->name, suffix); + xfree(tu_name); result = fopen(buf, "wt"); if(result == NULL) { panic("Couldn't open '%s' for writing.", buf); @@ -1428,7 +1159,7 @@ static FILE *my_open(const be_chordal_env_t *env, const char *prefix, const char void co_driver(be_chordal_env_t *cenv) { - lc_timer_t *timer = lc_timer_register("firm.be.copyopt", "runtime"); + ir_timer_t *timer = ir_timer_register("firm.be.copyopt", "runtime"); co_complete_stats_t before, after; copy_opt_t *co; co_algo_t *algo_func; @@ -1454,13 +1185,6 @@ void co_driver(be_chordal_env_t *cenv) be_stat_ev_ull("co_init_costs", before.costs); be_stat_ev_ull("co_init_unsat", before.unsatisfied_edges); - /* Dump the interference graph in Appel's format. */ - if (dump_flags & DUMP_APPEL) { - FILE *f = my_open(cenv, "", ".apl"); - co_dump_appel_graph(co, f); - fclose(f); - } - if (dump_flags & DUMP_BEFORE) { FILE *f = my_open(cenv, "", "-before.dot"); co_dump_ifg_dot(co, f, style_flags); @@ -1492,11 +1216,11 @@ void co_driver(be_chordal_env_t *cenv) algo_func = algos[algo].algo; /* perform actual copy minimization */ - lc_timer_reset_and_start(timer); + ir_timer_reset_and_start(timer); was_optimal = algo_func(co); - lc_timer_stop(timer); + ir_timer_stop(timer); - be_stat_ev("co_time", lc_timer_elapsed_msec(timer)); + be_stat_ev("co_time", ir_timer_elapsed_msec(timer)); be_stat_ev_ull("co_optimal", was_optimal); if (dump_flags & DUMP_AFTER) { @@ -1520,6 +1244,14 @@ void co_driver(be_chordal_env_t *cenv) printf("%10" ULL_FMT " %5s\n", after.costs, "-"); } + /* Dump the interference graph in Appel's format. */ + if (dump_flags & DUMP_APPEL) { + FILE *f = my_open(cenv, "", ".apl"); + fprintf(f, "# %lld %lld\n", after.costs, after.unsatisfied_edges); + co_dump_appel_graph(co, f); + fclose(f); + } + be_stat_ev_ull("co_after_costs", after.costs); be_stat_ev_ull("co_after_unsat", after.unsatisfied_edges);