X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbecopyopt.c;h=e9e0e9cc9bb524c2041d4d221d97577fccd0de5f;hb=1a3b7d363474ab544c13093a2f0b578718d37c7a;hp=065b43d34a6fa396567b288af446caca78d1e5df;hpb=7c767284b8f554996788212eb30709bb949710c4;p=libfirm diff --git a/ir/be/becopyopt.c b/ir/be/becopyopt.c index 065b43d34..e9e0e9cc9 100644 --- a/ir/be/becopyopt.c +++ b/ir/be/becopyopt.c @@ -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" @@ -44,24 +42,23 @@ #include "irprog.h" #include "irloop_t.h" #include "iredges_t.h" -#include "phiclass.h" #include "irbitset.h" #include "irphase_t.h" #include "irprintf_t.h" #include "bemodule.h" -#include "bearch_t.h" -#include "benode_t.h" +#include "bearch.h" +#include "benode.h" #include "beutil.h" -#include "beifg_t.h" +#include "beifg.h" #include "beintlive_t.h" #include "becopyopt_t.h" #include "becopystat.h" #include "belive_t.h" #include "beinsn_t.h" -#include "besched_t.h" +#include "besched.h" #include "bestatevent.h" -#include "beirg_t.h" +#include "beirg.h" #include "error.h" #include "lc_opts.h" @@ -80,7 +77,6 @@ static unsigned dump_flags = 0; static unsigned style_flags = 0; static unsigned do_stats = 0; static cost_fct_t cost_func = co_get_costs_exec_freq; -static unsigned algo = CO_ALGO_HEUR4; static int improve = 1; static const lc_opt_enum_mask_items_t dump_items[] = { @@ -100,16 +96,6 @@ static const lc_opt_enum_mask_items_t style_items[] = { { NULL, 0 } }; -static const lc_opt_enum_mask_items_t algo_items[] = { - { "none", CO_ALGO_NONE }, - { "heur", CO_ALGO_HEUR }, - { "heur2", CO_ALGO_HEUR2 }, - { "heur3", CO_ALGO_HEUR3 }, - { "heur4", CO_ALGO_HEUR4 }, - { "ilp", CO_ALGO_ILP }, - { NULL, 0 } -}; - typedef int (*opt_funcptr)(void); static const lc_opt_enum_func_ptr_items_t cost_func_items[] = { @@ -127,30 +113,31 @@ static lc_opt_enum_mask_var_t style_var = { &style_flags, style_items }; -static lc_opt_enum_mask_var_t algo_var = { - &algo, algo_items -}; - static lc_opt_enum_func_ptr_var_t cost_func_var = { (opt_funcptr*) &cost_func, cost_func_items }; static const lc_opt_table_entry_t options[] = { - LC_OPT_ENT_ENUM_INT ("algo", "select copy optimization algo", &algo_var), LC_OPT_ENT_ENUM_FUNC_PTR ("cost", "select a cost function", &cost_func_var), LC_OPT_ENT_ENUM_MASK ("dump", "dump ifg before or after copy optimization", &dump_var), LC_OPT_ENT_ENUM_MASK ("style", "dump style for ifg dumping", &style_var), LC_OPT_ENT_BOOL ("stats", "dump statistics after each optimization", &do_stats), - LC_OPT_ENT_BOOL ("improve", "run heur3 before if algo can exploit start solutions", &improve), + LC_OPT_ENT_BOOL ("improve", "run heur1 before if algo can exploit start solutions", &improve), LC_OPT_LAST }; -/* Insert additional options registration functions here. */ -extern void be_co_ilp_register_options(lc_opt_entry_t *grp); -extern void be_co2_register_options(lc_opt_entry_t *grp); -extern void be_co3_register_options(lc_opt_entry_t *grp); +static be_module_list_entry_t *copyopts = NULL; +static const co_algo_info *selected_copyopt = NULL; -void be_init_copycoal(void) +void be_register_copyopt(const char *name, co_algo_info *copyopt) +{ + if (selected_copyopt == NULL) + selected_copyopt = copyopt; + be_add_module_to_list(©opts, name, copyopt); +} + +BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyopt); +void be_init_copyopt(void) { lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be"); lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra"); @@ -158,9 +145,25 @@ void be_init_copycoal(void) lc_opt_entry_t *co_grp = lc_opt_get_grp(chordal_grp, "co"); lc_opt_add_table(co_grp, options); + be_add_module_list_opt(co_grp, "algo", "select copy optimization algo", + ©opts, (void**) &selected_copyopt); } -BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copycoal); +static int void_algo(copy_opt_t *co) +{ + (void) co; + return 0; +} + +BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copynone); +void be_init_copynone(void) +{ + static co_algo_info copyheur = { + void_algo, 0 + }; + + be_register_copyopt("none", ©heur); +} #undef QUICK_AND_DIRTY_HACK @@ -168,8 +171,10 @@ static int nodes_interfere(const be_chordal_env_t *env, const ir_node *a, const { if (env->ifg) return be_ifg_connected(env->ifg, a, b); - else - return values_interfere(env->birg, a, b); + else { + be_lv_t *lv = be_get_irg_liveness(env->irg); + return be_values_interfere(lv, a, b); + } } @@ -189,19 +194,18 @@ DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;) copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, cost_fct_t get_costs) { const char *s1, *s2, *s3; - int len; + size_t len; copy_opt_t *co; FIRM_DBG_REGISTER(dbg, "ir.be.copyopt"); 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; - s1 = get_irp_prog_name(); + s1 = get_irp_name(); s2 = get_entity_name(get_irg_entity(co->irg)); s3 = chordal_env->cls->name; len = strlen(s1) + strlen(s2) + strlen(s3) + 5; @@ -211,30 +215,35 @@ copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, cost_fct_t get_costs) return co; } -void free_copy_opt(copy_opt_t *co) { +void free_copy_opt(copy_opt_t *co) +{ xfree(co->name); 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_ignore(irn)) return 0; - reg = arch_get_irn_register(irn); - if (arch_register_type_is(reg, ignore)) - return 0; + if (is_Reg_Phi(irn) || is_Perm_Proj(irn)) + return 1; - req = arch_get_register_req(irn, -1); - if (is_Reg_Phi(irn) || is_Perm_Proj(irn) || is_2addr_code(req)) + req = arch_get_register_req_out(irn); + if (is_2addr_code(req)) return 1; return 0; } -int co_get_costs_loop_depth(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) { +int co_get_costs_loop_depth(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) +{ int cost = 0; ir_loop *loop; ir_node *root_block = get_nodes_block(root); @@ -255,19 +264,22 @@ int co_get_costs_loop_depth(const copy_opt_t *co, ir_node *root, ir_node* arg, i return 1+cost; } -int co_get_costs_exec_freq(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) { +int co_get_costs_exec_freq(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) +{ int res; ir_node *root_bl = get_nodes_block(root); ir_node *copy_bl = is_Phi(root) ? get_Block_cfgpred_block(root_bl, pos) : root_bl; + ir_exec_freq *exec_freq = be_get_irg_exec_freq(co->cenv->irg); (void) arg; - res = get_block_execfreq_ulong(co->cenv->birg->exec_freq, copy_bl); + res = get_block_execfreq_ulong(exec_freq, copy_bl); /* don't allow values smaller than one. */ return res < 1 ? 1 : res; } -int co_get_costs_all_one(const copy_opt_t *co, ir_node *root, ir_node *arg, int pos) { +int co_get_costs_all_one(const copy_opt_t *co, ir_node *root, ir_node *arg, int pos) +{ (void) co; (void) root; (void) arg; @@ -290,27 +302,28 @@ int co_get_costs_all_one(const copy_opt_t *co, ir_node *root, ir_node *arg, int * Determines a maximum weighted independent set with respect to * the interference and conflict edges of all nodes in a qnode. */ -static int ou_max_ind_set_costs(unit_t *ou) { +static int ou_max_ind_set_costs(unit_t *ou) +{ be_chordal_env_t *chordal_env = ou->co->cenv; ir_node **safe, **unsafe; int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs; bitset_t *curr; - bitset_pos_t pos; - int max, curr_weight, best_weight = 0; + size_t pos; + int curr_weight, best_weight = 0; /* assign the nodes into two groups. * safe: node has no interference, hence it is in every max stable set. * unsafe: node has an interference */ - safe = alloca((ou->node_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)); + safe = ALLOCAN(ir_node*, ou->node_count - 1); + safe_costs = 0; + safe_count = 0; + unsafe = ALLOCAN(ir_node*, ou->node_count - 1); + unsafe_costs = ALLOCAN(int, ou->node_count - 1); unsafe_count = 0; - for(i=1; inode_count; ++i) { + for (i=1; inode_count; ++i) { int is_safe = 1; - for(o=1; onode_count; ++o) { + for (o=1; onode_count; ++o) { if (i==o) continue; if (nodes_interfere(chordal_env, ou->nodes[i], ou->nodes[o])) { @@ -331,7 +344,7 @@ static int ou_max_ind_set_costs(unit_t *ou) { /* now compute the best set out of the unsafe nodes*/ if (unsafe_count > MIS_HEUR_TRIGGER) { bitset_t *best = bitset_alloca(unsafe_count); - /* Heuristik: Greedy trial and error form index 0 to unsafe_count-1 */ + /* Heuristic: Greedy trial and error form index 0 to unsafe_count-1 */ for (i=0; icls != co->cls) + return; + if (!co_is_optimizable_root(irn)) return; /* Init a new unit */ @@ -404,7 +422,7 @@ static void co_collect_units(ir_node *irn, void *env) { int o, arg_pos; ir_node *arg = get_irn_n(irn, i); - assert(is_curr_reg_class(co, arg) && "Argument not in same register class."); + assert(arch_get_irn_reg_class_out(arg) == co->cls && "Argument not in same register class."); if (arg == irn) continue; if (nodes_interfere(co->cenv, irn, arg)) { @@ -415,26 +433,27 @@ 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)) { - /* Check if arg has occurred at a prior position in the arg/list */ - arg_pos = 0; - for (o=1; onode_count; ++o) { - if (unit->nodes[o] == arg) { - arg_pos = o; - break; - } - } + if (arch_irn_is_ignore(arg)) + continue; - if (!arg_pos) { /* a new argument */ - /* insert node, set costs */ - unit->nodes[unit->node_count] = arg; - unit->costs[unit->node_count] = co->get_costs(co, irn, arg, i); - unit->node_count++; - } else { /* arg has occurred before in same phi */ - /* increase costs for existing arg */ - unit->costs[arg_pos] += co->get_costs(co, irn, arg, i); + /* Check if arg has occurred at a prior position in the arg/list */ + arg_pos = 0; + for (o=1; onode_count; ++o) { + if (unit->nodes[o] == arg) { + arg_pos = o; + break; } } + + if (!arg_pos) { /* a new argument */ + /* insert node, set costs */ + unit->nodes[unit->node_count] = arg; + unit->costs[unit->node_count] = co->get_costs(co, irn, arg, i); + unit->node_count++; + } 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, ir_node*, unit->node_count); unit->costs = XREALLOC(unit->costs, int, unit->node_count); @@ -448,8 +467,6 @@ static void co_collect_units(ir_node *irn, void *env) { 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(irn, -1); - /* Src == Tgt of a 2-addr-code instruction */ if (is_2addr_code(req)) { const unsigned other = req->other_same; @@ -459,10 +476,11 @@ static void co_collect_units(ir_node *irn, void *env) { 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(co->aenv, o, ignore) && - !nodes_interfere(co->cenv, irn, o)) { - ++count; - } + if (arch_irn_is_ignore(o)) + continue; + if (nodes_interfere(co->cenv, irn, o)) + continue; + ++count; } } @@ -477,7 +495,7 @@ static void co_collect_units(ir_node *irn, void *env) { 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(co->aenv, o, ignore) && + if (!arch_irn_is_ignore(o) && !nodes_interfere(co->cenv, irn, o)) { unit->nodes[k] = o; unit->costs[k] = co->get_costs(co, irn, o, -1); @@ -497,7 +515,7 @@ static void co_collect_units(ir_node *irn, void *env) { struct list_head *tmp; /* Determine the maximum costs this unit can cause: all_nodes_cost */ - for(i=1; inode_count; ++i) { + for (i=1; inode_count; ++i) { unit->sort_key = MAX(unit->sort_key, unit->costs[i]); unit->all_nodes_costs += unit->costs[i]; } @@ -516,7 +534,8 @@ static void co_collect_units(ir_node *irn, void *env) { #ifdef QUICK_AND_DIRTY_HACK -static int compare_ous(const void *k1, const void *k2) { +static int compare_ous(const void *k1, const void *k2) +{ const unit_t *u1 = *((const unit_t **) k1); const unit_t *u2 = *((const unit_t **) k2); int i, o, u1_has_constr, u2_has_constr; @@ -525,7 +544,7 @@ static int compare_ous(const void *k1, const void *k2) { /* Units with constraints come first */ u1_has_constr = 0; for (i=0; inode_count; ++i) { - arch_get_register_req(&req, u1->nodes[i], -1); + arch_get_register_req_out(&req, u1->nodes[i]); if (arch_register_req_is(&req, limited)) { u1_has_constr = 1; break; @@ -534,7 +553,7 @@ static int compare_ous(const void *k1, const void *k2) { u2_has_constr = 0; for (i=0; inode_count; ++i) { - arch_get_register_req(&req, u2->nodes[i], -1); + arch_get_register_req_out(&req, u2->nodes[i]); if (arch_register_req_is(&req, limited)) { u2_has_constr = 1; break; @@ -560,14 +579,15 @@ static int compare_ous(const void *k1, const void *k2) { /** * Sort the ou's according to constraints and their sort_key */ -static void co_sort_units(copy_opt_t *co) { +static void co_sort_units(copy_opt_t *co) +{ int i, count = 0, costs; unit_t *ou, **ous; /* get the number of ous, remove them form the list and fill the array */ list_for_each_entry(unit_t, ou, &co->units, units) count++; - ous = alloca(count * sizeof(*ous)); + ous = ALLOCAN(unit_t, count); costs = co_get_max_copy_costs(co); @@ -596,7 +616,8 @@ static void co_sort_units(copy_opt_t *co) { } #endif -void co_build_ou_structure(copy_opt_t *co) { +void co_build_ou_structure(copy_opt_t *co) +{ DBG((dbg, LEVEL_1, "\tCollecting optimization units\n")); INIT_LIST_HEAD(&co->units); irg_walk_graph(co->irg, co_collect_units, NULL, co); @@ -605,7 +626,8 @@ void co_build_ou_structure(copy_opt_t *co) { #endif } -void co_free_ou_structure(copy_opt_t *co) { +void co_free_ou_structure(copy_opt_t *co) +{ unit_t *curr, *tmp; ASSERT_OU_AVAIL(co); list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) { @@ -618,7 +640,8 @@ void co_free_ou_structure(copy_opt_t *co) { /* co_solve_heuristic() is implemented in becopyheur.c */ -int co_get_max_copy_costs(const copy_opt_t *co) { +int co_get_max_copy_costs(const copy_opt_t *co) +{ int i, res = 0; unit_t *curr; @@ -632,7 +655,8 @@ int co_get_max_copy_costs(const copy_opt_t *co) { return res; } -int co_get_inevit_copy_costs(const copy_opt_t *co) { +int co_get_inevit_copy_costs(const copy_opt_t *co) +{ int res = 0; unit_t *curr; @@ -643,7 +667,8 @@ int co_get_inevit_copy_costs(const copy_opt_t *co) { return res; } -int co_get_copy_costs(const copy_opt_t *co) { +int co_get_copy_costs(const copy_opt_t *co) +{ int i, res = 0; unit_t *curr; @@ -664,7 +689,8 @@ int co_get_copy_costs(const copy_opt_t *co) { return res; } -int co_get_lower_bound(const copy_opt_t *co) { +int co_get_lower_bound(const copy_opt_t *co) +{ int res = 0; unit_t *curr; @@ -688,7 +714,7 @@ void co_complete_stats(const copy_opt_t *co, co_complete_stats_t *stat) stat->aff_nodes += 1; bitset_add_irn(seen, an->irn); co_gs_foreach_neighb(an, neigh) { - if(!bitset_contains_irn(seen, neigh->irn)) { + if (!bitset_contains_irn(seen, neigh->irn)) { stat->aff_edges += 1; stat->max_costs += neigh->costs; @@ -697,7 +723,7 @@ void co_complete_stats(const copy_opt_t *co, co_complete_stats_t *stat) stat->unsatisfied_edges += 1; } - if(nodes_interfere(co->cenv, an->irn, neigh->irn)) { + if (nodes_interfere(co->cenv, an->irn, neigh->irn)) { stat->aff_int += 1; stat->inevit_costs += neigh->costs; } @@ -720,15 +746,17 @@ void co_complete_stats(const copy_opt_t *co, co_complete_stats_t *stat) |_| |___/ ******************************************************************************/ -static int compare_affinity_node_t(const void *k1, const void *k2, size_t size) { - const affinity_node_t *n1 = k1; - const affinity_node_t *n2 = k2; +static int compare_affinity_node_t(const void *k1, const void *k2, size_t size) +{ + const affinity_node_t *n1 = (const affinity_node_t*)k1; + const affinity_node_t *n2 = (const affinity_node_t*)k2; (void) size; return (n1->irn != n2->irn); } -static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) { +static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) +{ affinity_node_t new_node, *node; neighb_t *nbr; int allocnew = 1; @@ -736,7 +764,7 @@ 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), hash_irn(new_node.irn)); + node = (affinity_node_t*)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 +774,7 @@ static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) { /* if we did not find n2 in n1's neighbourhood insert it */ if (allocnew) { - nbr = obstack_alloc(&co->obst, sizeof(*nbr)); + nbr = OALLOC(&co->obst, neighb_t); nbr->irn = n2; nbr->costs = 0; nbr->next = node->neighbours; @@ -759,23 +787,24 @@ 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); } } -static void build_graph_walker(ir_node *irn, void *env) { - copy_opt_t *co = env; +static void build_graph_walker(ir_node *irn, void *env) +{ + const arch_register_req_t *req; + copy_opt_t *co = (copy_opt_t*)env; int pos, max; - const arch_register_t *reg; - if (!is_curr_reg_class(co, irn) || arch_irn_is(co->aenv, irn, ignore)) + if (get_irn_mode(irn) == mode_T) return; - - reg = arch_get_irn_register(irn); - if (arch_register_type_is(reg, ignore)) + req = arch_get_register_req_out(irn); + if (req->cls != co->cls || arch_irn_is_ignore(irn)) return; if (is_Reg_Phi(irn)) { /* Phis */ @@ -786,9 +815,7 @@ static void build_graph_walker(ir_node *irn, void *env) { } 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)); - } - else { /* 2-address code */ - const arch_register_req_t *req = arch_get_register_req(irn, -1); + } else { /* 2-address code */ if (is_2addr_code(req)) { const unsigned other = req->other_same; int i; @@ -796,7 +823,7 @@ static void build_graph_walker(ir_node *irn, void *env) { 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(co->aenv, other, ignore)) + if (!arch_irn_is_ignore(other)) add_edges(co, irn, other, co->get_costs(co, irn, other, 0)); } } @@ -804,14 +831,16 @@ static void build_graph_walker(ir_node *irn, void *env) { } } -void co_build_graph_structure(copy_opt_t *co) { +void co_build_graph_structure(copy_opt_t *co) +{ obstack_init(&co->obst); co->nodes = new_set(compare_affinity_node_t, 32); irg_walk_graph(co->irg, build_graph_walker, NULL, co); } -void co_free_graph_structure(copy_opt_t *co) { +void co_free_graph_structure(copy_opt_t *co) +{ ASSERT_GS_AVAIL(co); del_set(co->nodes); @@ -819,15 +848,14 @@ void co_free_graph_structure(copy_opt_t *co) { co->nodes = NULL; } -/* co_solve_ilp1() co_solve_ilp2() are implemented in becopyilpX.c */ - -int co_gs_is_optimizable(copy_opt_t *co, ir_node *irn) { +int co_gs_is_optimizable(copy_opt_t *co, ir_node *irn) +{ affinity_node_t new_node, *n; ASSERT_GS_AVAIL(co); new_node.irn = irn; - n = set_find(co->nodes, &new_node, sizeof(new_node), hash_irn(new_node.irn)); + n = (affinity_node_t*)set_find(co->nodes, &new_node, sizeof(new_node), hash_irn(new_node.irn)); if (n) { return (n->degree > 0); } else @@ -838,15 +866,14 @@ static int co_dump_appel_disjoint_constraints(const copy_opt_t *co, ir_node *a, { 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)) + const arch_register_req_t *req = arch_get_register_req_out(nodes[j]); + if (arch_register_req_is(req, limited)) rbitset_copy_to_bitset(req->limited, constr[j]); else bitset_set_all(constr[j]); @@ -858,19 +885,26 @@ static int co_dump_appel_disjoint_constraints(const copy_opt_t *co, ir_node *a, 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); + be_ifg_t *ifg = co->cenv->ifg; + int *color_map = ALLOCAN(int, co->cls->n_regs); + int *node_map = XMALLOCN(int, get_irg_last_idx(co->irg) + 1); + ir_graph *irg = co->irg; + be_irg_t *birg = be_birg_from_irg(irg); ir_node *irn; - void *it, *nit; + nodes_iter_t it; + neighbours_iter_t nit; int n, n_regs; unsigned i; n_regs = 0; - for(i = 0; i < co->cls->n_regs; ++i) { + for (i = 0; i < co->cls->n_regs; ++i) { const arch_register_t *reg = &co->cls->regs[i]; - color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_regs++; + if (rbitset_is_set(birg->allocatable_regs, reg->global_index)) { + color_map[i] = n_regs++; + } else { + color_map[i] = -1; + } } /* @@ -878,48 +912,45 @@ void co_dump_appel_graph(const copy_opt_t *co, FILE *f) * the values below n are the pre-colored register nodes */ - it = be_ifg_nodes_iter_alloca(ifg); - nit = be_ifg_neighbours_iter_alloca(ifg); - n = n_regs; - be_ifg_foreach_node(ifg, it, irn) { - if(!arch_irn_is(co->aenv, irn, ignore)) - node_map[get_irn_idx(irn)] = n++; + be_ifg_foreach_node(ifg, &it, irn) { + if (arch_irn_is_ignore(irn)) + continue; + 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 = node_map[get_irn_idx(irn)]; - affinity_node_t *a = get_affinity_info(co, irn); - - const arch_register_req_t *req; - ir_node *adj; + be_ifg_foreach_node(ifg, &it, irn) { + if (!arch_irn_is_ignore(irn)) { + int idx = node_map[get_irn_idx(irn)]; + affinity_node_t *a = get_affinity_info(co, irn); + const arch_register_req_t *req = arch_get_register_req_out(irn); + ir_node *adj; - 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) + 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) fprintf(f, "%d %d -1\n", color_map[i], idx); } } - be_ifg_foreach_neighbour(ifg, nit, irn, adj) { - if(!arch_irn_is(co->aenv, adj, ignore) && !co_dump_appel_disjoint_constraints(co, irn, adj)) { + be_ifg_foreach_neighbour(ifg, &nit, irn, adj) { + if (!arch_irn_is_ignore(adj) && + !co_dump_appel_disjoint_constraints(co, irn, adj)) { int adj_idx = node_map[get_irn_idx(adj)]; - if(idx < adj_idx) + if (idx < adj_idx) fprintf(f, "%d %d -1\n", idx, adj_idx); } } - if(a) { + if (a) { neighb_t *n; co_gs_foreach_neighb(a, n) { - if(!arch_irn_is(co->aenv, n->irn, ignore)) { + if (!arch_irn_is_ignore(n->irn)) { int n_idx = node_map[get_irn_idx(n->irn)]; - if(idx < n_idx) + if (idx < n_idx) fprintf(f, "%d %d %d\n", idx, n_idx, (int) n->costs); } } @@ -978,7 +1009,7 @@ static const char *get_dot_color_name(size_t col) return col < sizeof(names)/sizeof(names[0]) ? names[col] : "white"; } -typedef struct _co_ifg_dump_t { +typedef struct co_ifg_dump_t { const copy_opt_t *co; unsigned flags; } co_ifg_dump_t; @@ -991,24 +1022,21 @@ 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_ignore(irn); } 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(irn); - const arch_register_req_t *req; - int limited; + co_ifg_dump_t *env = (co_ifg_dump_t*)self; + const arch_register_t *reg = arch_get_irn_register(irn); + const arch_register_req_t *req = arch_get_register_req_out(irn); + int limited = arch_register_req_is(req, limited); - req = arch_get_register_req(irn, BE_OUT_POS(0)); - limited = arch_register_req_is(req, limited); - - if(env->flags & CO_IFG_DUMP_LABELS) { + if (env->flags & CO_IFG_DUMP_LABELS) { ir_fprintf(f, "label=\"%+F", irn); - if((env->flags & CO_IFG_DUMP_CONSTR) && limited) { + if ((env->flags & CO_IFG_DUMP_CONSTR) && limited) { bitset_t *bs = bitset_alloca(env->co->cls->n_regs); rbitset_copy_to_bitset(req->limited, bs); ir_fprintf(f, "\\n%B", bs); @@ -1018,16 +1046,16 @@ static void ifg_dump_node_attr(FILE *f, void *self, ir_node *irn) fprintf(f, "label=\"\" shape=point " ); } - if(env->flags & CO_IFG_DUMP_SHAPE) + if (env->flags & CO_IFG_DUMP_SHAPE) fprintf(f, "shape=%s ", limited ? "diamond" : "ellipse"); - if(env->flags & CO_IFG_DUMP_COLORS) + if (env->flags & CO_IFG_DUMP_COLORS) fprintf(f, "style=filled color=%s ", get_dot_color_name(reg->index)); } static void ifg_dump_at_end(FILE *file, void *self) { - co_ifg_dump_t *env = self; + co_ifg_dump_t *env = (co_ifg_dump_t*)self; affinity_node_t *a; co_gs_foreach_aff_node(env->co, a) { @@ -1039,12 +1067,12 @@ static void ifg_dump_at_end(FILE *file, void *self) const arch_register_t *nr = arch_get_irn_register(n->irn); unsigned nidx = get_irn_idx(n->irn); - if(aidx < nidx) { + if (aidx < nidx) { const char *color = nr == ar ? "blue" : "red"; - fprintf(file, "\tn%d -- n%d [weight=0.01 ", aidx, nidx); - if(env->flags & CO_IFG_DUMP_LABELS) + fprintf(file, "\tn%u -- n%u [weight=0.01 ", aidx, nidx); + if (env->flags & CO_IFG_DUMP_LABELS) fprintf(file, "label=\"%d\" ", n->costs); - if(env->flags & CO_IFG_DUMP_COLORS) + if (env->flags & CO_IFG_DUMP_COLORS) fprintf(file, "color=%s ", color); else fprintf(file, "style=dotted"); @@ -1081,45 +1109,6 @@ void co_solve_park_moon(copy_opt_t *opt) (void) opt; } -static int void_algo(copy_opt_t *co) -{ - (void) co; - return 0; -} - -/* - _ _ _ _ _ - / \ | | __ _ ___ _ __(_) |_| |__ _ __ ___ ___ - / _ \ | |/ _` |/ _ \| '__| | __| '_ \| '_ ` _ \/ __| - / ___ \| | (_| | (_) | | | | |_| | | | | | | | \__ \ - /_/ \_\_|\__, |\___/|_| |_|\__|_| |_|_| |_| |_|___/ - |___/ -*/ - -typedef struct { - co_algo_t *algo; - const char *name; - int can_improve_existing; -} co_algo_info_t; - -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 } -}; - /* __ __ _ ____ _ | \/ | __ _(_)_ __ | _ \ _ __(_)_ _____ _ __ @@ -1135,10 +1124,11 @@ static FILE *my_open(const be_chordal_env_t *env, const char *prefix, const char char buf[1024]; size_t i, n; char *tu_name; + const char *cup_name = be_get_irg_main_env(env->irg)->cup_name; - n = strlen(env->birg->main_env->cup_name); + n = strlen(cup_name); tu_name = XMALLOCN(char, n + 1); - strcpy(tu_name, env->birg->main_env->cup_name); + strcpy(tu_name, cup_name); for (i = 0; i < n; ++i) if (tu_name[i] == '.') tu_name[i] = '_'; @@ -1147,7 +1137,7 @@ static FILE *my_open(const be_chordal_env_t *env, const char *prefix, const char 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) { + if (result == NULL) { panic("Couldn't open '%s' for writing.", buf); } @@ -1156,16 +1146,18 @@ static FILE *my_open(const be_chordal_env_t *env, const char *prefix, const char void co_driver(be_chordal_env_t *cenv) { - ir_timer_t *timer = ir_timer_register("firm.be.copyopt", "runtime"); + ir_timer_t *timer = ir_timer_new(); co_complete_stats_t before, after; copy_opt_t *co; - co_algo_t *algo_func; int was_optimal = 0; - if (algo >= CO_ALGO_LAST) + assert(selected_copyopt); + + /* skip copymin if algo is 'none' */ + if (selected_copyopt->copyopt == void_algo) return; - be_liveness_assure_chk(be_get_birg_liveness(cenv->birg)); + be_liveness_assure_chk(be_get_irg_liveness(cenv->irg)); co = new_copy_opt(cenv, cost_func); co_build_ou_structure(co); @@ -1188,37 +1180,26 @@ void co_driver(be_chordal_env_t *cenv) fclose(f); } - /* if the algo can improve results, provide an initial solution with heur3 */ - if (improve && algos[algo].can_improve_existing) { + /* if the algo can improve results, provide an initial solution with heur1 */ + if (improve && selected_copyopt->can_improve_existing) { co_complete_stats_t stats; /* produce a heuristic solution */ -#ifdef WITH_JVM - co_solve_heuristic_java(co); -#else co_solve_heuristic(co); -#endif /* WITH_JVM */ /* do the stats and provide the current costs */ co_complete_stats(co, &stats); be_stat_ev_ull("co_prepare_costs", stats.costs); } -#ifdef WITH_JVM - /* start the JVM here so that it does not tamper the timing. */ - if (algo == CO_ALGO_HEUR3) - be_java_coal_start_jvm(); -#endif /* WITH_JVM */ - - algo_func = algos[algo].algo; - /* perform actual copy minimization */ ir_timer_reset_and_start(timer); - was_optimal = algo_func(co); + was_optimal = selected_copyopt->copyopt(co); ir_timer_stop(timer); be_stat_ev("co_time", ir_timer_elapsed_msec(timer)); be_stat_ev_ull("co_optimal", was_optimal); + ir_timer_free(timer); if (dump_flags & DUMP_AFTER) { FILE *f = my_open(cenv, "", "-after.dot"); @@ -1235,7 +1216,7 @@ void co_driver(be_chordal_env_t *cenv) ir_printf("%30F ", cenv->irg); printf("%10s %10" ULL_FMT "%10" ULL_FMT "%10" ULL_FMT, cenv->cls->name, after.max_costs, before.costs, after.inevit_costs); - if(optimizable_costs > 0) + if (optimizable_costs > 0) printf("%10" ULL_FMT " %5.2f\n", after.costs, (evitable * 100.0) / optimizable_costs); else printf("%10" ULL_FMT " %5s\n", after.costs, "-"); @@ -1244,7 +1225,7 @@ void co_driver(be_chordal_env_t *cenv) /* 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); + fprintf(f, "# %llu %llu\n", after.costs, after.unsatisfied_edges); co_dump_appel_graph(co, f); fclose(f); }