X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbecopyopt.c;h=cba1f255958bebfb272267ec230fa1ad8ec00b69;hb=945c6c2ceebef5e41c0486c31f49d2319cacb3da;hp=4b9fcc75400b239150c65a6f384d218b5cff90e8;hpb=8a1c0ad68eaba7cecab7a3abd552681ad893de3e;p=libfirm diff --git a/ir/be/becopyopt.c b/ir/be/becopyopt.c index 4b9fcc754..cba1f2559 100644 --- a/ir/be/becopyopt.c +++ b/ir/be/becopyopt.c @@ -1,20 +1,6 @@ /* - * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved. - * * This file is part of libFirm. - * - * This file may be distributed and/or modified under the terms of the - * GNU General Public License version 2 as published by the Free Software - * Foundation and appearing in the file LICENSE.GPL included in the - * packaging of this file. - * - * Licensees holding valid libFirm Professional Edition licenses may use - * this file in accordance with the libFirm Commercial License. - * Agreement provided with the Software. - * - * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE - * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR - * PURPOSE. + * Copyright (C) 2012 University of Karlsruhe. */ /** @@ -22,7 +8,6 @@ * @brief Copy minimization driver. * @author Daniel Grund * @date 12.04.2005 - * @version $Id$ * * Main file for the optimization reducing the copies needed for: * - Phi coalescing @@ -31,35 +16,37 @@ */ #include "config.h" -#include "execfreq.h" -#include "xmalloc.h" #include "debug.h" -#include "pmap.h" -#include "raw_bitset.h" -#include "irnode.h" +#include "error.h" +#include "execfreq_t.h" +#include "irdump_t.h" +#include "iredges_t.h" #include "irgraph.h" #include "irgwalk.h" -#include "irprog.h" #include "irloop_t.h" -#include "iredges_t.h" -#include "irbitset.h" -#include "irphase_t.h" -#include "irprintf_t.h" +#include "irnode.h" +#include "irprintf.h" +#include "irprog.h" +#include "irtools.h" +#include "pmap.h" +#include "raw_bitset.h" +#include "util.h" +#include "xmalloc.h" -#include "bemodule.h" #include "bearch.h" -#include "benode.h" -#include "beutil.h" -#include "beifg.h" -#include "beintlive_t.h" #include "becopyopt_t.h" #include "becopystat.h" -#include "belive_t.h" +#include "bedump.h" +#include "beifg.h" #include "beinsn_t.h" -#include "besched.h" -#include "bestatevent.h" +#include "beintlive_t.h" #include "beirg.h" -#include "error.h" +#include "belive_t.h" +#include "bemodule.h" +#include "benode.h" +#include "besched.h" +#include "statev_t.h" +#include "beutil.h" #include "lc_opts.h" #include "lc_opts_enum.h" @@ -73,9 +60,23 @@ #define COST_FUNC_LOOP 2 #define COST_FUNC_ALL_ONE 3 +/** + * Flags for dumping the IFG. + */ +enum { + CO_IFG_DUMP_COLORS = 1 << 0, /**< Dump the graph colored. */ + CO_IFG_DUMP_LABELS = 1 << 1, /**< Dump node/edge labels. */ + CO_IFG_DUMP_SHAPE = 1 << 2, /**< Give constrained nodes special shapes. */ + CO_IFG_DUMP_CONSTR = 1 << 3, /**< Dump the node constraints in the label. */ +}; + +static int co_get_costs_loop_depth(const ir_node *root, int pos); +static int co_get_costs_exec_freq(const ir_node *root, int pos); +static int co_get_costs_all_one(const ir_node *root, int pos); + static unsigned dump_flags = 0; -static unsigned style_flags = 0; -static unsigned do_stats = 0; +static unsigned style_flags = CO_IFG_DUMP_COLORS; +static int do_stats = 0; static cost_fct_t cost_func = co_get_costs_exec_freq; static int improve = 1; @@ -97,7 +98,6 @@ static const lc_opt_enum_mask_items_t style_items[] = { }; typedef int (*opt_funcptr)(void); - static const lc_opt_enum_func_ptr_items_t cost_func_items[] = { { "freq", (opt_funcptr) co_get_costs_exec_freq }, { "loop", (opt_funcptr) co_get_costs_loop_depth }, @@ -136,7 +136,7 @@ void be_register_copyopt(const char *name, co_algo_info *copyopt) be_add_module_to_list(©opts, name, copyopt); } -BE_REGISTER_MODULE_CONSTRUCTOR(be_init_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"); @@ -155,7 +155,7 @@ static int void_algo(copy_opt_t *co) return 0; } -BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copynone); +BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copynone) void be_init_copynone(void) { static co_algo_info copyheur = { @@ -167,17 +167,6 @@ void be_init_copynone(void) #undef QUICK_AND_DIRTY_HACK -static int nodes_interfere(const be_chordal_env_t *env, const ir_node *a, const ir_node *b) -{ - if (env->ifg) - return be_ifg_connected(env->ifg, a, b); - else { - be_lv_t *lv = be_get_irg_liveness(env->irg); - return be_values_interfere(lv, a, b); - } -} - - /****************************************************************************** _____ _ / ____| | | @@ -193,31 +182,18 @@ 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; - copy_opt_t *co; - FIRM_DBG_REGISTER(dbg, "ir.be.copyopt"); - co = XMALLOCZ(copy_opt_t); + copy_opt_t *const co = XMALLOCZ(copy_opt_t); co->cenv = chordal_env; co->irg = chordal_env->irg; co->cls = chordal_env->cls; co->get_costs = get_costs; - - 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; - co->name = XMALLOCN(char, len); - snprintf(co->name, len, "%s__%s__%s", s1, s2, s3); - return co; } void free_copy_opt(copy_opt_t *co) { - xfree(co->name); free(co); } @@ -227,62 +203,69 @@ void free_copy_opt(copy_opt_t *co) */ static int co_is_optimizable_root(ir_node *irn) { - const arch_register_req_t *req; - - if (arch_irn_is_ignore(irn)) + arch_register_req_t const *const req = arch_get_irn_register_req(irn); + if (arch_register_req_is(req, ignore)) return 0; if (is_Reg_Phi(irn) || is_Perm_Proj(irn)) return 1; - req = arch_get_register_req_out(irn); - if (is_2addr_code(req)) + if (arch_register_req_is(req, should_be_same)) return 1; return 0; } -int co_get_costs_loop_depth(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) +/** + * Computes the costs of a copy according to loop depth + * @param pos the argument position of arg in the root arguments + * @return Must be >= 0 in all cases. + */ +static int co_get_costs_loop_depth(const ir_node *root, int pos) { - int cost = 0; + ir_node *block = get_nodes_block(root); ir_loop *loop; - ir_node *root_block = get_nodes_block(root); - (void) co; - (void) arg; + int cost; if (is_Phi(root)) { - /* for phis the copies are placed in the corresponding pred-block */ - loop = get_irn_loop(get_Block_cfgpred_block(root_block, pos)); - } else { - /* a perm places the copy in the same block as it resides */ - loop = get_irn_loop(root_block); + block = get_Block_cfgpred_block(block, pos); } + loop = get_irn_loop(block); if (loop) { int d = get_loop_depth(loop); cost = d*d; + } else { + cost = 0; } return 1+cost; } -int co_get_costs_exec_freq(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) +static ir_execfreq_int_factors factors; + +/** + * Computes the costs of a copy according to execution frequency + * @param pos the argument position of arg in the root arguments + * @return Must be >= 0 in all cases. + */ +static int co_get_costs_exec_freq(const ir_node *root, 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(exec_freq, copy_bl); + ir_node *copy_bl + = is_Phi(root) ? get_Block_cfgpred_block(root_bl, pos) : root_bl; + int res = get_block_execfreq_int(&factors, 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) +/** + * All costs equal 1. Using this will reduce the _number_ of copies. + * @param co The copy opt object. + * @return Must be >= 0 in all cases. + */ +static int co_get_costs_all_one(const ir_node *root, int pos) { - (void) co; (void) root; - (void) arg; (void) pos; return 1; } @@ -302,13 +285,11 @@ 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 *const ou, be_lv_t const *const lv) { - 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; - unsigned pos; int curr_weight, best_weight = 0; /* assign the nodes into two groups. @@ -326,7 +307,7 @@ static int ou_max_ind_set_costs(unit_t *ou) for (o=1; onode_count; ++o) { if (i==o) continue; - if (nodes_interfere(chordal_env, ou->nodes[i], ou->nodes[o])) { + if (be_values_interfere(lv, ou->nodes[i], ou->nodes[o])) { unsafe_costs[unsafe_count] = ou->costs[i]; unsafe[unsafe_count] = ou->nodes[i]; ++unsafe_count; @@ -344,12 +325,12 @@ 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)) @@ -403,10 +384,10 @@ static void co_collect_units(ir_node *irn, void *env) /* Init a new unit */ unit = XMALLOCZ(unit_t); - unit->co = co; unit->node_count = 1; INIT_LIST_HEAD(&unit->queue); + be_lv_t *const lv = be_get_irg_liveness(co->irg); /* Phi with some/all of its arguments */ if (is_Reg_Phi(irn)) { int i, arity; @@ -422,11 +403,11 @@ static void co_collect_units(ir_node *irn, void *env) int o, arg_pos; ir_node *arg = get_irn_n(irn, i); - assert(arch_get_irn_reg_class_out(arg) == co->cls && "Argument not in same register class."); + assert(arch_get_irn_reg_class(arg) == co->cls && "Argument not in same register class."); if (arg == irn) continue; - if (nodes_interfere(co->cenv, irn, arg)) { - unit->inevitable_costs += co->get_costs(co, irn, arg, i); + if (be_values_interfere(lv, irn, arg)) { + unit->inevitable_costs += co->get_costs(irn, i); continue; } @@ -448,65 +429,63 @@ static void co_collect_units(ir_node *irn, void *env) 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->costs[unit->node_count] = co->get_costs(irn, 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->costs[arg_pos] += co->get_costs(irn, i); } } 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))); + assert(!be_values_interfere(lv, irn, get_Perm_src(irn))); 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 { + unit->costs[1] = co->get_costs(irn, -1); + } else if (arch_register_req_is(req, should_be_same)) { /* Src == Tgt of a 2-addr-code instruction */ - if (is_2addr_code(req)) { - const unsigned other = req->other_same; - int count = 0; - int i; + 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_ignore(o)) + continue; + if (be_values_interfere(lv, irn, o)) + continue; + ++count; + } + } - for (i = 0; (1U << i) <= other; ++i) { + 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_ignore(o)) - continue; - if (nodes_interfere(co->cenv, irn, o)) - continue; - ++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_ignore(o) && - !nodes_interfere(co->cenv, irn, o)) { - unit->nodes[k] = o; - unit->costs[k] = co->get_costs(co, irn, o, -1); - ++k; - } + if (!arch_irn_is_ignore(o) && + !be_values_interfere(lv, irn, o)) { + unit->nodes[k] = o; + unit->costs[k] = co->get_costs(irn, -1); + ++k; } } } - } else { - assert(0 && "This is not an optimizable node!"); } + } else { + assert(0 && "This is not an optimizable node!"); } /* Insert the new unit at a position according to its costs */ @@ -521,7 +500,7 @@ static void co_collect_units(ir_node *irn, void *env) } /* Determine the minimal costs this unit will cause: min_nodes_costs */ - unit->min_nodes_costs += unit->all_nodes_costs - ou_max_ind_set_costs(unit); + unit->min_nodes_costs += unit->all_nodes_costs - ou_max_ind_set_costs(unit, lv); /* Insert the new ou according to its sort_key */ tmp = &co->units; while (tmp->next != &co->units && list_entry_units(tmp->next)->sort_key > unit->sort_key) @@ -544,7 +523,7 @@ 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_out(&req, u1->nodes[i]); + arch_get_irn_register_req(&req, u1->nodes[i]); if (arch_register_req_is(&req, limited)) { u1_has_constr = 1; break; @@ -553,7 +532,7 @@ static int compare_ous(const void *k1, const void *k2) u2_has_constr = 0; for (i=0; inode_count; ++i) { - arch_get_register_req_out(&req, u2->nodes[i]); + arch_get_irn_register_req(&req, u2->nodes[i]); if (arch_register_req_is(&req, limited)) { u2_has_constr = 1; break; @@ -563,14 +542,6 @@ static int compare_ous(const void *k1, const void *k2) if (u1_has_constr != u2_has_constr) return u2_has_constr - u1_has_constr; - /* Now check, whether the two units are connected */ -#if 0 - for (i=0; inode_count; ++i) - for (o=0; onode_count; ++o) - if (u1->nodes[i] == u2->nodes[o]) - return 0; -#endif - /* After all, the sort key decides. Greater keys come first. */ return u2->sort_key - u1->sort_key; @@ -582,7 +553,7 @@ static int compare_ous(const void *k1, const void *k2) static void co_sort_units(copy_opt_t *co) { int i, count = 0, costs; - unit_t *ou, **ous; + unit_t **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) @@ -628,7 +599,6 @@ void co_build_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) { xfree(curr->nodes); @@ -643,7 +613,6 @@ void co_free_ou_structure(copy_opt_t *co) int co_get_max_copy_costs(const copy_opt_t *co) { int i, res = 0; - unit_t *curr; ASSERT_OU_AVAIL(co); @@ -658,7 +627,6 @@ int co_get_max_copy_costs(const copy_opt_t *co) int co_get_inevit_copy_costs(const copy_opt_t *co) { int res = 0; - unit_t *curr; ASSERT_OU_AVAIL(co); @@ -670,7 +638,6 @@ int co_get_inevit_copy_costs(const copy_opt_t *co) int co_get_copy_costs(const copy_opt_t *co) { int i, res = 0; - unit_t *curr; ASSERT_OU_AVAIL(co); @@ -692,7 +659,6 @@ int co_get_copy_costs(const copy_opt_t *co) int co_get_lower_bound(const copy_opt_t *co) { int res = 0; - unit_t *curr; ASSERT_OU_AVAIL(co); @@ -703,18 +669,17 @@ int co_get_lower_bound(const copy_opt_t *co) void co_complete_stats(const copy_opt_t *co, co_complete_stats_t *stat) { - bitset_t *seen = bitset_irg_malloc(co->irg); - affinity_node_t *an; + bitset_t *seen = bitset_malloc(get_irg_last_idx(co->irg)); memset(stat, 0, sizeof(stat[0])); /* count affinity edges. */ + be_lv_t *const lv = be_get_irg_liveness(co->irg); co_gs_foreach_aff_node(co, an) { - neighb_t *neigh; stat->aff_nodes += 1; - bitset_add_irn(seen, an->irn); + bitset_set(seen, get_irn_idx(an->irn)); co_gs_foreach_neighb(an, neigh) { - if (!bitset_contains_irn(seen, neigh->irn)) { + if (!bitset_is_set(seen, get_irn_idx(neigh->irn))) { stat->aff_edges += 1; stat->max_costs += neigh->costs; @@ -723,7 +688,7 @@ 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 (be_values_interfere(lv, an->irn, neigh->irn)) { stat->aff_int += 1; stat->inevit_costs += neigh->costs; } @@ -748,8 +713,8 @@ 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; + 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); @@ -762,9 +727,8 @@ static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) int allocnew = 1; 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 = set_insert(affinity_node_t, co->nodes, &new_node, sizeof(new_node), hash_irn(new_node.irn)); for (nbr = node->neighbours; nbr; nbr = nbr->next) if (nbr->irn == n2) { @@ -780,7 +744,6 @@ static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) nbr->next = node->neighbours; node->neighbours = nbr; - node->degree++; } /* now nbr points to n1's neighbour-entry of n2 */ @@ -789,7 +752,8 @@ static void add_edge(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)) { + be_lv_t *const lv = be_get_irg_liveness(co->irg); + if (!be_values_interfere(lv, n1, n2)) { add_edge(co, n1, n2, costs); add_edge(co, n2, n1, costs); } @@ -798,34 +762,32 @@ static inline void add_edges(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs static void build_graph_walker(ir_node *irn, void *env) { const arch_register_req_t *req; - copy_opt_t *co = env; + copy_opt_t *co = (copy_opt_t*)env; int pos, max; if (get_irn_mode(irn) == mode_T) return; - req = arch_get_register_req_out(irn); - if (req->cls != co->cls || arch_irn_is_ignore(irn)) + req = arch_get_irn_register_req(irn); + if (req->cls != co->cls || arch_register_req_is(req, ignore)) return; if (is_Reg_Phi(irn)) { /* Phis */ for (pos=0, max=get_irn_arity(irn); posget_costs(co, irn, arg, pos)); + add_edges(co, irn, arg, co->get_costs(irn, pos)); } } 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 */ - if (is_2addr_code(req)) { - const unsigned other = req->other_same; - int i; + add_edges(co, irn, arg, co->get_costs(irn, -1)); + } else if (arch_register_req_is(req, should_be_same)) { + 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_ignore(other)) - add_edges(co, irn, other, co->get_costs(co, irn, other, 0)); - } + 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_ignore(other)) + add_edges(co, irn, other, co->get_costs(irn, -1)); } } } @@ -855,11 +817,8 @@ 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), hash_irn(new_node.irn)); - if (n) { - return (n->degree > 0); - } else - return 0; + n = set_find(affinity_node_t, co->nodes, &new_node, sizeof(new_node), hash_irn(new_node.irn)); + return n && n->neighbours; } static int co_dump_appel_disjoint_constraints(const copy_opt_t *co, ir_node *a, ir_node *b) @@ -872,7 +831,7 @@ static int co_dump_appel_disjoint_constraints(const copy_opt_t *co, ir_node *a, constr[1] = bitset_alloca(co->cls->n_regs); for (j = 0; j < 2; ++j) { - const arch_register_req_t *req = arch_get_register_req_out(nodes[j]); + const arch_register_req_t *req = arch_get_irn_register_req(nodes[j]); if (arch_register_req_is(req, limited)) rbitset_copy_to_bitset(req->limited, constr[j]); else @@ -883,7 +842,14 @@ static int co_dump_appel_disjoint_constraints(const copy_opt_t *co, ir_node *a, return !bitset_intersect(constr[0], constr[1]); } -void co_dump_appel_graph(const copy_opt_t *co, FILE *f) +/** + * Dump the interference graph according to the Appel/George coalescing contest file format. + * See: http://www.cs.princeton.edu/~appel/coalesce/format.html + * @note Requires graph structure. + * @param co The copy opt object. + * @param f A file to dump to. + */ +static void co_dump_appel_graph(const copy_opt_t *co, FILE *f) { be_ifg_t *ifg = co->cenv->ifg; int *color_map = ALLOCAN(int, co->cls->n_regs); @@ -891,8 +857,6 @@ void co_dump_appel_graph(const copy_opt_t *co, FILE *f) ir_graph *irg = co->irg; be_irg_t *birg = be_birg_from_irg(irg); - ir_node *irn; - nodes_iter_t it; neighbours_iter_t nit; int n, n_regs; unsigned i; @@ -913,7 +877,7 @@ void co_dump_appel_graph(const copy_opt_t *co, FILE *f) */ n = n_regs; - be_ifg_foreach_node(ifg, &it, irn) { + be_ifg_foreach_node(ifg, irn) { if (arch_irn_is_ignore(irn)) continue; node_map[get_irn_idx(irn)] = n++; @@ -921,204 +885,46 @@ void co_dump_appel_graph(const copy_opt_t *co, FILE *f) fprintf(f, "%d %d\n", n, n_regs); - 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; - - 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_ignore(adj) && - !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); - } - } + be_ifg_foreach_node(ifg, irn) { + arch_register_req_t const *const req = arch_get_irn_register_req(irn); + if (arch_register_req_is(req, ignore)) + continue; - if (a) { - neighb_t *n; + int idx = node_map[get_irn_idx(irn)]; + affinity_node_t *a = get_affinity_info(co, irn); - co_gs_foreach_neighb(a, n) { - if (!arch_irn_is_ignore(n->irn)) { - 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); - } - } + 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); } } - } - - xfree(node_map); -} - -/* - ___ _____ ____ ____ ___ _____ ____ _ - |_ _| ___/ ___| | _ \ / _ \_ _| | _ \ _ _ _ __ ___ _ __ (_)_ __ __ _ - | || |_ | | _ | | | | | | || | | | | | | | | '_ ` _ \| '_ \| | '_ \ / _` | - | || _|| |_| | | |_| | |_| || | | |_| | |_| | | | | | | |_) | | | | | (_| | - |___|_| \____| |____/ \___/ |_| |____/ \__,_|_| |_| |_| .__/|_|_| |_|\__, | - |_| |___/ -*/ - -static const char *get_dot_color_name(size_t col) -{ - static const char *names[] = { - "blue", - "red", - "green", - "yellow", - "cyan", - "magenta", - "orange", - "chocolate", - "beige", - "navy", - "darkgreen", - "darkred", - "lightPink", - "chartreuse", - "lightskyblue", - "linen", - "pink", - "lightslateblue", - "mintcream", - "red", - "darkolivegreen", - "mediumblue", - "mistyrose", - "salmon", - "darkseagreen", - "mediumslateblue" - "moccasin", - "tomato", - "forestgreen", - "darkturquoise", - "palevioletred" - }; - - return col < sizeof(names)/sizeof(names[0]) ? names[col] : "white"; -} -typedef struct co_ifg_dump_t { - const copy_opt_t *co; - unsigned flags; -} co_ifg_dump_t; - -static void ifg_dump_graph_attr(FILE *f, void *self) -{ - (void) self; - fprintf(f, "overlap=scale"); -} - -static int ifg_is_dump_node(void *self, ir_node *irn) -{ - (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 = arch_get_register_req_out(irn); - int limited = arch_register_req_is(req, limited); - - if (env->flags & CO_IFG_DUMP_LABELS) { - ir_fprintf(f, "label=\"%+F", irn); - - 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); + 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) + fprintf(f, "%d %d -1\n", idx, adj_idx); + } } - ir_fprintf(f, "\" "); - } else { - fprintf(f, "label=\"\" shape=point " ); - } - - if (env->flags & CO_IFG_DUMP_SHAPE) - fprintf(f, "shape=%s ", limited ? "diamond" : "ellipse"); - 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; - affinity_node_t *a; - - co_gs_foreach_aff_node(env->co, a) { - 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(n->irn); - unsigned nidx = get_irn_idx(n->irn); - - if (aidx < nidx) { - const char *color = nr == ar ? "blue" : "red"; - 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) - fprintf(file, "color=%s ", color); - else - fprintf(file, "style=dotted"); - fprintf(file, "];\n"); + if (a) { + co_gs_foreach_neighb(a, n) { + if (!arch_irn_is_ignore(n->irn)) { + 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); + } } } } -} - - -static be_ifg_dump_dot_cb_t ifg_dot_cb = { - ifg_is_dump_node, - ifg_dump_graph_attr, - ifg_dump_node_attr, - NULL, - NULL, - ifg_dump_at_end -}; - - - -void co_dump_ifg_dot(const copy_opt_t *co, FILE *f, unsigned flags) -{ - co_ifg_dump_t cod; - cod.co = co; - cod.flags = flags; - be_ifg_dump_dot(co->cenv->ifg, co->irg, f, &ifg_dot_cb, &cod); -} - - -void co_solve_park_moon(copy_opt_t *opt) -{ - (void) opt; + xfree(node_map); } -/* - __ __ _ ____ _ - | \/ | __ _(_)_ __ | _ \ _ __(_)_ _____ _ __ - | |\/| |/ _` | | '_ \ | | | | '__| \ \ / / _ \ '__| - | | | | (_| | | | | | | |_| | | | |\ V / __/ | - |_| |_|\__,_|_|_| |_| |____/|_| |_| \_/ \___|_| - -*/ - -static FILE *my_open(const be_chordal_env_t *env, const char *prefix, const char *suffix) +static FILE *my_open(const be_chordal_env_t *env, const char *prefix, + const char *suffix) { FILE *result; char buf[1024]; @@ -1157,7 +963,7 @@ void co_driver(be_chordal_env_t *cenv) if (selected_copyopt->copyopt == void_algo) return; - be_liveness_assure_chk(be_get_irg_liveness(cenv->irg)); + be_assure_live_chk(cenv->irg); co = new_copy_opt(cenv, cost_func); co_build_ou_structure(co); @@ -1165,18 +971,18 @@ void co_driver(be_chordal_env_t *cenv) co_complete_stats(co, &before); - be_stat_ev_ull("co_aff_nodes", before.aff_nodes); - be_stat_ev_ull("co_aff_edges", before.aff_edges); - be_stat_ev_ull("co_max_costs", before.max_costs); - be_stat_ev_ull("co_inevit_costs", before.inevit_costs); - be_stat_ev_ull("co_aff_int", before.aff_int); + stat_ev_ull("co_aff_nodes", before.aff_nodes); + stat_ev_ull("co_aff_edges", before.aff_edges); + stat_ev_ull("co_max_costs", before.max_costs); + stat_ev_ull("co_inevit_costs", before.inevit_costs); + stat_ev_ull("co_aff_int", before.aff_int); - be_stat_ev_ull("co_init_costs", before.costs); - be_stat_ev_ull("co_init_unsat", before.unsatisfied_edges); + stat_ev_ull("co_init_costs", before.costs); + stat_ev_ull("co_init_unsat", before.unsatisfied_edges); if (dump_flags & DUMP_BEFORE) { - FILE *f = my_open(cenv, "", "-before.dot"); - co_dump_ifg_dot(co, f, style_flags); + FILE *f = my_open(cenv, "", "-before.vcg"); + be_dump_ifg_co(f, co, style_flags & CO_IFG_DUMP_LABELS, style_flags & CO_IFG_DUMP_COLORS); fclose(f); } @@ -1189,7 +995,7 @@ void co_driver(be_chordal_env_t *cenv) /* do the stats and provide the current costs */ co_complete_stats(co, &stats); - be_stat_ev_ull("co_prepare_costs", stats.costs); + stat_ev_ull("co_prepare_costs", stats.costs); } /* perform actual copy minimization */ @@ -1197,29 +1003,29 @@ void co_driver(be_chordal_env_t *cenv) 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); + stat_ev_dbl("co_time", ir_timer_elapsed_msec(timer)); + stat_ev_ull("co_optimal", was_optimal); ir_timer_free(timer); if (dump_flags & DUMP_AFTER) { - FILE *f = my_open(cenv, "", "-after.dot"); - co_dump_ifg_dot(co, f, style_flags); + FILE *f = my_open(cenv, "", "-after.vcg"); + be_dump_ifg_co(f, co, style_flags & CO_IFG_DUMP_LABELS, style_flags & CO_IFG_DUMP_COLORS); fclose(f); } co_complete_stats(co, &after); if (do_stats) { - ulong64 optimizable_costs = after.max_costs - after.inevit_costs; - ulong64 evitable = after.costs - after.inevit_costs; + unsigned long long optimizable_costs = after.max_costs - after.inevit_costs; + unsigned long long evitable = after.costs - after.inevit_costs; 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); + printf("%10s %10llu%10llu%10llu", cenv->cls->name, after.max_costs, before.costs, after.inevit_costs); if (optimizable_costs > 0) - printf("%10" ULL_FMT " %5.2f\n", after.costs, (evitable * 100.0) / optimizable_costs); + printf("%10llu %5.2f\n", after.costs, (evitable * 100.0) / optimizable_costs); else - printf("%10" ULL_FMT " %5s\n", after.costs, "-"); + printf("%10llu %5s\n", after.costs, "-"); } /* Dump the interference graph in Appel's format. */ @@ -1230,8 +1036,8 @@ void co_driver(be_chordal_env_t *cenv) fclose(f); } - be_stat_ev_ull("co_after_costs", after.costs); - be_stat_ev_ull("co_after_unsat", after.unsatisfied_edges); + stat_ev_ull("co_after_costs", after.costs); + stat_ev_ull("co_after_unsat", after.unsatisfied_edges); co_free_graph_structure(co); co_free_ou_structure(co);