X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbelower.c;h=0ff5504eb4ba5349f9ea058f910aa785396ec14b;hb=b9d45e08e23bcf058fa8f2d9e18dd78e8cccd044;hp=be27b68c9f866f0b97e9c03e63a8374e307ebc3f;hpb=962297fd814612b8086cdd7205db98272e288b89;p=libfirm diff --git a/ir/be/belower.c b/ir/be/belower.c index be27b68c9..0ff5504eb 100644 --- a/ir/be/belower.c +++ b/ir/be/belower.c @@ -14,7 +14,6 @@ #include #include "ircons.h" -#include "ident.h" #include "debug.h" #include "irhooks.h" @@ -23,6 +22,8 @@ #include "benode_t.h" #include "bechordal_t.h" #include "besched_t.h" +#include "bestat.h" +#include "benodesets.h" #include "irgmod.h" #include "iredges_t.h" @@ -36,28 +37,25 @@ #endif #undef is_Perm -#define is_Perm(arch_env, irn) (arch_irn_classify(arch_env, irn) == arch_irn_class_perm) - -/* collect static data about perms */ -typedef struct _perm_stat_t { - const arch_register_class_t *cls; /**< the current register class */ - int *perm_size_ar; /**< the sizes of all perms in an irg */ - int *real_perm_size_ar; /**< the sizes of all perms in an irg */ - int *chain_len_ar; /**< the sizes of all chains for all perms */ - int *cycle_len_ar; /**< the siyes of all cycles for all perms */ - int num_perms; /**< number of all perms */ - int num_real_perms; /**< number of all perms */ - int num_chains; /**< the number of all chains */ - int num_cycles; /**< the number of all cycles */ -} perm_stat_t; +#define is_Perm(arch_env, irn) (arch_irn_class_is(arch_env, irn, perm)) + +/* associates op with it's copy and CopyKeep */ +typedef struct { + ir_node *op; /* an irn which must be different */ + pset *copies; /* all non-spillable copies of this irn */ +} op_copy_assoc_t; + +/* environment for constraints */ +typedef struct { + be_irg_t *birg; + pset *op_set; + struct obstack obst; +} constraint_env_t; /* lowering walker environment */ typedef struct _lower_env_t { be_chordal_env_t *chord_env; unsigned do_copy:1; - unsigned do_stat:1; - unsigned pstat_n:30; - perm_stat_t **pstat; DEBUG_ONLY(firm_dbg_module_t *dbg_module;) } lower_env_t; @@ -86,6 +84,14 @@ typedef struct _perm_cycle_t { perm_type_t type; /**< type (CHAIN or CYCLE) */ } perm_cycle_t; +/* Compare the two operands */ +static int cmp_op_copy_assoc(const void *a, const void *b) { + const op_copy_assoc_t *op1 = a; + const op_copy_assoc_t *op2 = b; + + return op1->op != op2->op; +} + /* Compare the in registers of two register pairs */ static int compare_reg_pair(const void *a, const void *b) { const reg_pair_t *pair_a = a; @@ -264,17 +270,15 @@ static void lower_perm_node(ir_node *irn, void *walk_env) { const arch_register_class_t *reg_class; const arch_env_t *arch_env; lower_env_t *env = walk_env; - perm_stat_t **pstat = env->pstat; - int pstat_idx = -1; - int real_size = 0; - int n, i, pn, do_copy, j, n_ops; + int real_size = 0; + int keep_perm = 0; + int n, i, pn, do_copy, j, n_ops; reg_pair_t *pairs; const ir_edge_t *edge; perm_cycle_t *cycle; ir_node *sched_point, *block, *in[2]; ir_node *arg1, *arg2, *res1, *res2; ir_node *cpyxchg = NULL; - ident *cls_id; DEBUG_ONLY(firm_dbg_module_t *mod;) arch_env = env->chord_env->birg->main_env->arch_env; @@ -296,23 +300,8 @@ static void lower_perm_node(ir_node *irn, void *walk_env) { assert(n == get_irn_n_edges(irn) && "perm's in and out numbers different"); reg_class = arch_get_irn_register(arch_env, get_irn_n(irn, 0))->reg_class; - cls_id = new_id_from_str(reg_class->name); pairs = alloca(n * sizeof(pairs[0])); - if (env->do_stat) { - /* determine index in statistics */ - for (i = 0; i < env->pstat_n; i++) { - if (strcmp(pstat[i]->cls->name, reg_class->name) == 0) { - pstat_idx = i; - break; - } - } - assert(pstat_idx >= 0 && "could not determine class index for statistics"); - - pstat[pstat_idx]->num_perms++; - pstat[pstat_idx]->perm_size_ar[n - 1]++; - } - /* build the list of register pairs (in, out) */ i = 0; foreach_out_edge(irn, edge) { @@ -349,6 +338,7 @@ static void lower_perm_node(ir_node *irn, void *walk_env) { /* reroute the edges from the proj to the argument */ edges_reroute(pairs[i].out_node, pairs[i].in_node, env->chord_env->irg); + set_irn_n(pairs[i].out_node, 0, new_Bad()); pairs[i].checked = 1; } @@ -359,17 +349,13 @@ static void lower_perm_node(ir_node *irn, void *walk_env) { do_copy = 0; } - if (env->do_stat && get_n_checked_pairs(pairs, n) < n) { - pstat[pstat_idx]->num_real_perms++; - pstat[pstat_idx]->real_perm_size_ar[n - 1]++; - real_size = n - get_n_checked_pairs(pairs, n); - } + real_size = n - get_n_checked_pairs(pairs, n); - hook_be_block_stat_perm(cls_id, reg_class->n_regs, irn, block, n, real_size); + be_do_stat_perm(reg_class->name, reg_class->n_regs, irn, block, n, real_size); /* check for cycles and chains */ while (get_n_checked_pairs(pairs, n) < n) { - i = 0; + i = n_ops = 0; /* go to the first not-checked pair */ while (pairs[i].checked) i++; @@ -382,24 +368,14 @@ static void lower_perm_node(ir_node *irn, void *walk_env) { } DB((mod, LEVEL_1, "\n")); - /* statistics */ - if (env->do_stat) { - int n_idx = cycle->n_elems - 1; - if (cycle->type == PERM_CHAIN) { - pstat[pstat_idx]->num_chains++; - pstat[pstat_idx]->chain_len_ar[n_idx]++; - } - else { - pstat[pstat_idx]->num_cycles++; - pstat[pstat_idx]->cycle_len_ar[n_idx]++; - } - } - - /* We don't need to do anything if we have a Perm with two + /* + We don't need to do anything if we have a Perm with two elements which represents a cycle, because those nodes - already represent exchange nodes */ + already represent exchange nodes + */ if (n == 2 && cycle->type == PERM_CYCLE) { free(cycle); + keep_perm = 1; continue; } @@ -413,8 +389,6 @@ static void lower_perm_node(ir_node *irn, void *walk_env) { res1 = get_node_for_register(pairs, n, cycle->elems[i], 1); res2 = get_node_for_register(pairs, n, cycle->elems[i + 1], 1); - - n_ops = 0; /* If we have a cycle and don't copy: we need to create exchange nodes NOTE: An exchange node is a perm node with 2 INs and 2 OUTs @@ -517,20 +491,22 @@ static void lower_perm_node(ir_node *irn, void *walk_env) { /* set the new scheduling point */ sched_point = cpyxchg; } - - if (env->do_stat) { - hook_be_block_stat_permcycle(cls_id, irn, block, cycle->type == PERM_CHAIN, cycle->n_elems, n_ops); - } } + be_do_stat_permcycle(reg_class->name, irn, block, cycle->type == PERM_CHAIN, cycle->n_elems, n_ops); + free((void *) cycle->elems); free(cycle); } - - /* remove the perm from schedule */ - sched_remove(irn); + if (! keep_perm) { + int arity = get_irn_arity(irn); + for(i = 0; i < arity; ++i) { + set_irn_n(irn, i, new_Bad()); + } + sched_remove(irn); + } } @@ -552,25 +528,14 @@ static ir_node *belower_skip_proj(ir_node *irn) { return irn; } -static void fix_in(ir_node *irn, ir_node *old, ir_node *nw) { - int i, n; - - irn = belower_skip_proj(irn); - n = get_irn_arity(irn); - - for (i = 0; i < n; i++) { - if (get_irn_n(irn, i) == old) { - set_irn_n(irn, i, nw); - break; - } - } -} - -static void gen_assure_different_pattern(ir_node *irn, be_irg_t *birg, ir_node *other_different) { - const arch_env_t *arch_env = birg->main_env->arch_env; - ir_node *in[2], *keep, *cpy, *temp; - ir_node *block = get_nodes_block(irn); - const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, other_different, -1); +static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, constraint_env_t *env) { + be_irg_t *birg = env->birg; + pset *op_set = env->op_set; + const arch_env_t *arch_env = birg->main_env->arch_env; + ir_node *block = get_nodes_block(irn); + const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, other_different, -1); + ir_node *in[2], *keep, *cpy; + op_copy_assoc_t key, *entry; FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower"); if (arch_irn_is(arch_env, other_different, ignore) || ! mode_is_datab(get_irn_mode(other_different))) { @@ -583,16 +548,12 @@ static void gen_assure_different_pattern(ir_node *irn, be_irg_t *birg, ir_node * /* in block far far away */ /* The copy is optimized later if not needed */ - temp = new_rd_Unknown(birg->irg, get_irn_mode(other_different)); - cpy = be_new_Copy(cls, birg->irg, block, temp); + cpy = be_new_Copy(cls, birg->irg, block, other_different); be_node_set_flags(cpy, BE_OUT_POS(0), arch_irn_flags_dont_spill); in[0] = irn; in[1] = cpy; - /* Let the irn use the copy instead of the old other_different */ - fix_in(irn, other_different, cpy); - /* Add the Keep resp. CopyKeep and reroute the users */ /* of the other_different irn in case of CopyKeep. */ if (get_n_out_edges(other_different) == 0) { @@ -601,12 +562,36 @@ static void gen_assure_different_pattern(ir_node *irn, be_irg_t *birg, ir_node * else { keep = be_new_CopyKeep_single(cls, birg->irg, block, cpy, irn, get_irn_mode(other_different)); be_node_set_reg_class(keep, 1, cls); - edges_reroute(other_different, keep, birg->irg); } - /* after rerouting: let the copy point to the other_different irn */ + /* let the copy point to the other_different irn */ set_irn_n(cpy, 0, other_different); + /* insert copy and keep into schedule */ + assert(sched_is_scheduled(irn) && "need schedule to assure constraints"); + sched_add_before(belower_skip_proj(irn), cpy); + sched_add_after(irn, keep); + + /* insert the other different and it's copies into the set */ + key.op = other_different; + key.copies = NULL; + entry = pset_find(op_set, &key, nodeset_hash(other_different)); + + if (! entry) { + entry = obstack_alloc(&env->obst, sizeof(*entry)); + entry->copies = pset_new_ptr_default(); + entry->op = other_different; + } + + /* insert copy */ + pset_insert_ptr(entry->copies, cpy); + + /* insert keep in case of CopyKeep */ + if (be_is_CopyKeep(keep)) + pset_insert_ptr(entry->copies, keep); + + pset_insert(op_set, entry, nodeset_hash(other_different)); + DBG((mod, LEVEL_1, "created %+F for %+F to assure should_be_different\n", keep, irn)); } @@ -614,22 +599,20 @@ static void gen_assure_different_pattern(ir_node *irn, be_irg_t *birg, ir_node * * Checks if node has a should_be_different constraint in output * and adds a Keep then to assure the constraint. */ -static void assure_different_constraints(ir_node *irn, be_irg_t *birg) { - const arch_env_t *arch_env = birg->main_env->arch_env; +static void assure_different_constraints(ir_node *irn, constraint_env_t *env) { const arch_register_req_t *req; - arch_register_req_t req_temp; - int i, n; + arch_register_req_t req_temp; - req = arch_get_register_req(arch_env, &req_temp, irn, -1); + req = arch_get_register_req(env->birg->main_env->arch_env, &req_temp, irn, -1); if (req) { if (arch_register_req_is(req, should_be_different)) { - gen_assure_different_pattern(irn, birg, req->other_different); + gen_assure_different_pattern(irn, req->other_different, env); } else if (arch_register_req_is(req, should_be_different_from_all)) { - n = get_irn_arity(belower_skip_proj(irn)); + int i, n = get_irn_arity(belower_skip_proj(irn)); for (i = 0; i < n; i++) { - gen_assure_different_pattern(irn, birg, get_irn_n(belower_skip_proj(irn), i)); + gen_assure_different_pattern(irn, get_irn_n(belower_skip_proj(irn), i), env); } } } @@ -661,7 +644,71 @@ static void assure_constraints_walker(ir_node *irn, void *walk_env) { * @param birg The birg structure containing the irg */ void assure_constraints(be_irg_t *birg) { - irg_walk_blkwise_graph(birg->irg, NULL, assure_constraints_walker, birg); + constraint_env_t cenv; + op_copy_assoc_t *entry; + dom_front_info_t *dom; + ir_node **nodes; + FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower"); + + cenv.birg = birg; + cenv.op_set = new_pset(cmp_op_copy_assoc, 16); + obstack_init(&cenv.obst); + + irg_walk_blkwise_graph(birg->irg, NULL, assure_constraints_walker, &cenv); + + /* introduce copies needs dominance information */ + dom = be_compute_dominance_frontiers(birg->irg); + + /* for all */ + foreach_pset(cenv.op_set, entry) { + int n; + ir_node *cp; + + n = pset_count(entry->copies); + nodes = alloca((n + 1) * sizeof(nodes[0])); + + /* put the node in an array */ + n = 0; + nodes[n++] = entry->op; + DBG((mod, LEVEL_1, "introduce copies for %+F ", entry->op)); + + /* collect all copies */ + foreach_pset(entry->copies, cp) { + nodes[n++] = cp; + DB((mod, LEVEL_1, ", %+F ", cp)); + } + + DB((mod, LEVEL_1, "\n")); + + /* introduce the copies for the operand and it's copies */ + be_ssa_constr(dom, NULL, n, nodes); + + + /* Could be that not all CopyKeeps are really needed, */ + /* so we transform unnecessary ones into Keeps. */ + foreach_pset(entry->copies, cp) { + if (be_is_CopyKeep(cp) && get_irn_n_edges(cp) < 1) { + ir_node *keep; + int n = get_irn_arity(cp); + + keep = be_new_Keep(arch_get_irn_reg_class(birg->main_env->arch_env, cp, -1), + birg->irg, get_nodes_block(cp), n, (ir_node **)&get_irn_in(cp)[1]); + sched_add_before(cp, keep); + + /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */ + while (--n >= -1) + set_irn_n(cp, n, new_Bad()); + sched_remove(cp); + } + } + + del_pset(entry->copies); + } + + be_free_dominance_frontiers(dom); + + del_pset(cenv.op_set); + obstack_free(&cenv.obst, NULL); } @@ -685,60 +732,6 @@ static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env) { return; } -static void lower_print_perm_stat(lower_env_t *env) { - int i, j, total_len_chain, total_len_cycle, total_size_perm, total_size_real_perm; - - printf("=== IRG: %s ===\n", get_entity_name(get_irg_entity(env->chord_env->irg))); - for (i = 0; i < env->pstat_n; i++) { - if (env->pstat[i]->num_perms == 0) - continue; - - printf("CLASS: %s\n", env->pstat[i]->cls->name); - printf("# total perms: %d (size:num -> 1:%d", env->pstat[i]->num_perms, env->pstat[i]->perm_size_ar[0]); - - total_size_perm = env->pstat[i]->perm_size_ar[0]; - for (j = 1; j < env->pstat[i]->cls->n_regs; j++) { - total_size_perm += (j + 1) * env->pstat[i]->perm_size_ar[j]; - printf(", %d:%d", j + 1, env->pstat[i]->perm_size_ar[j]); - } - printf(")\n"); - printf("avg perm size: %.2f\n", env->pstat[i]->num_perms ? (float)total_size_perm / (float)env->pstat[i]->num_perms : 0); - - printf("# real perms: %d (size:num -> 1:%d", env->pstat[i]->num_real_perms, env->pstat[i]->real_perm_size_ar[0]); - - total_size_real_perm = env->pstat[i]->real_perm_size_ar[0]; - for (j = 1; j < env->pstat[i]->cls->n_regs; j++) { - total_size_real_perm += (j + 1) * env->pstat[i]->real_perm_size_ar[j]; - printf(", %d:%d", j + 1, env->pstat[i]->real_perm_size_ar[j]); - } - printf(")\n"); - printf("avg real perm size: %.2f\n", env->pstat[i]->num_real_perms ? (float)total_size_real_perm / (float)env->pstat[i]->num_real_perms : 0); - - printf("# total chains: %d (lenght:num -> 1:%d", env->pstat[i]->num_chains, env->pstat[i]->chain_len_ar[0]); - - total_len_chain = env->pstat[i]->chain_len_ar[0]; - - for (j = 1; j < env->pstat[i]->cls->n_regs; j++) { - total_len_chain += (j + 1) * env->pstat[i]->chain_len_ar[j]; - printf(", %d:%d", j + 1, env->pstat[i]->chain_len_ar[j]); - } - printf(")\n"); - printf("avg chain length: %.2f\n", env->pstat[i]->num_chains ? (float)total_len_chain / (float)env->pstat[i]->num_chains : 0); - printf("avg chains/perm: %.2f\n", env->pstat[i]->num_real_perms ? (float)env->pstat[i]->num_chains / (float)env->pstat[i]->num_real_perms : 0); - - printf("# total cycles: %d (length:num -> 1:%d", env->pstat[i]->num_cycles, env->pstat[i]->cycle_len_ar[0]); - - total_len_cycle = env->pstat[i]->cycle_len_ar[0]; - for (j = 1; j < env->pstat[i]->cls->n_regs; j++) { - total_len_cycle += (j + 1) * env->pstat[i]->cycle_len_ar[j]; - printf(", %d:%d", j + 1, env->pstat[i]->cycle_len_ar[j]); - } - printf(")\n"); - printf("avg cycle length: %.2f\n", env->pstat[i]->num_cycles ? (float)total_len_cycle / (float)env->pstat[i]->num_cycles : 0); - printf("avg cycles/perm: %.2f\n", env->pstat[i]->num_real_perms ? (float)env->pstat[i]->num_cycles / (float)env->pstat[i]->num_real_perms : 0); - } -} - /** * Walks over all blocks in an irg and performs lowering need to be * done after register allocation (e.g. perm lowering). @@ -746,48 +739,14 @@ static void lower_print_perm_stat(lower_env_t *env) { * @param chord_env The chordal environment containing the irg * @param do_copy 1 == resolve cycles with a free reg if available */ -void lower_nodes_after_ra(be_chordal_env_t *chord_env, int do_copy, int do_stat) { +void lower_nodes_after_ra(be_chordal_env_t *chord_env, int do_copy) { lower_env_t env; env.chord_env = chord_env; env.do_copy = do_copy; - env.do_stat = do_stat; FIRM_DBG_REGISTER(env.dbg_module, "firm.be.lower"); - /* if we want statistics: allocate memory for the data and initialize with 0 */ - if (do_stat) { - const arch_isa_t *isa = chord_env->birg->main_env->arch_env->isa; - int i, n = arch_isa_get_n_reg_class(isa); - - env.pstat = alloca(n * sizeof(env.pstat[0])); - env.pstat_n = n; - - for (i = 0; i < n; i++) { - const arch_register_class_t *cls = arch_isa_get_reg_class(isa, i); - int n_regs = cls->n_regs; - - env.pstat[i] = alloca(sizeof(*(env.pstat[0]))); - memset(env.pstat[i], 0, sizeof(*(env.pstat[0]))); - - env.pstat[i]->perm_size_ar = alloca(n_regs * sizeof(env.pstat[i]->perm_size_ar[0])); - env.pstat[i]->real_perm_size_ar = alloca(n_regs * sizeof(env.pstat[i]->real_perm_size_ar[0])); - env.pstat[i]->chain_len_ar = alloca(n_regs * sizeof(env.pstat[i]->chain_len_ar[0])); - env.pstat[i]->cycle_len_ar = alloca(n_regs * sizeof(env.pstat[i]->cycle_len_ar[0])); - - memset(env.pstat[i]->perm_size_ar, 0, n_regs * sizeof(env.pstat[i]->perm_size_ar[0])); - memset(env.pstat[i]->real_perm_size_ar, 0, n_regs * sizeof(env.pstat[i]->real_perm_size_ar[0])); - memset(env.pstat[i]->chain_len_ar, 0, n_regs * sizeof(env.pstat[i]->chain_len_ar[0])); - memset(env.pstat[i]->cycle_len_ar, 0, n_regs * sizeof(env.pstat[i]->cycle_len_ar[0])); - - env.pstat[i]->cls = cls; - } - } - irg_walk_blkwise_graph(chord_env->irg, NULL, lower_nodes_after_ra_walker, &env); - - if (do_stat) { - lower_print_perm_stat(&env); - } } #undef is_Perm