X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbepeephole.c;h=484a5acab715901240e702e859a8345a01a75266;hb=bed05ed2d3b2ae84849fd2e75603e8834153cd5b;hp=284974f9235ff60aa600d7afe084fdc522aa05f0;hpb=6730cf921d356d992d35526daf57f82af7ab0816;p=libfirm diff --git a/ir/be/bepeephole.c b/ir/be/bepeephole.c index 284974f92..484a5acab 100644 --- a/ir/be/bepeephole.c +++ b/ir/be/bepeephole.c @@ -1,5 +1,5 @@ /* - * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved. + * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved. * * This file is part of libFirm. * @@ -25,6 +25,7 @@ */ #include "config.h" +#include "array_t.h" #include "bepeephole.h" #include "iredges_t.h" @@ -62,7 +63,7 @@ static void clear_reg_value(ir_node *node) if (reg == NULL) { panic("No register assigned at %+F", node); } - if (arch_register_type_is(reg, virtual)) + if (reg->type & arch_register_type_virtual) return; cls = arch_register_get_class(reg); reg_idx = arch_register_get_index(reg); @@ -87,7 +88,7 @@ static void set_reg_value(ir_node *node) if (reg == NULL) { panic("No register assigned at %+F", node); } - if (arch_register_type_is(reg, virtual)) + if (reg->type & arch_register_type_virtual) return; cls = arch_register_get_class(reg); reg_idx = arch_register_get_index(reg); @@ -131,9 +132,9 @@ void be_peephole_new_node(ir_node * nw) /** * must be called from peephole optimisations before a node will be killed * and its users will be redirected to new_node. - * so bepeephole can update it's internal state. + * so bepeephole can update its internal state. * - * Note: killing a node and rewiring os only allowed if new_node produces + * Note: killing a node and rewiring is only allowed if new_node produces * the same registers as old_node. */ static void be_peephole_before_exchange(const ir_node *old_node, @@ -193,9 +194,9 @@ static void process_block(ir_node *block, void *data) (void) data; /* construct initial register assignment */ - n_classes = arch_env_get_n_reg_class(arch_env); + n_classes = arch_env->n_register_classes; for (i = 0; i < n_classes; ++i) { - const arch_register_class_t *cls = arch_env_get_reg_class(arch_env, i); + const arch_register_class_t *cls = &arch_env->register_classes[i]; unsigned n_regs = arch_register_class_n_regs(cls); memset(register_values[i], 0, sizeof(ir_node*) * n_regs); } @@ -233,137 +234,28 @@ static void process_block(ir_node *block, void *data) } } -static void kill_node_and_preds(ir_node *node) -{ - ir_graph *irg = get_irn_irg(node); - int arity, i; - - arity = get_irn_arity(node); - for (i = 0; i < arity; ++i) { - ir_node *pred = get_irn_n(node, i); - - set_irn_n(node, i, new_r_Bad(irg)); - if (get_irn_n_edges(pred) != 0) - continue; - - kill_node_and_preds(pred); - } - - if (!is_Proj(node)) - sched_remove(node); - kill_node(node); -} - -/** - * Walk through the block schedule and skip all barrier nodes. - */ -static void skip_barrier(ir_node *block, ir_graph *irg) -{ - ir_node *irn; - - sched_foreach_reverse(block, irn) { - int arity; - unsigned *used; - unsigned n_used; - const ir_edge_t *edge, *next; - - if (!be_is_Barrier(irn)) - continue; - - /* track which outputs are actually used, as we have to create - * keep nodes for unused outputs */ - arity = get_irn_arity(irn); - rbitset_alloca(used, arity); - - foreach_out_edge_safe(irn, edge, next) { - ir_node *proj = get_edge_src_irn(edge); - int pn; - ir_node *pred; - - if (is_Anchor(proj)) - continue; - - pn = (int) get_Proj_proj(proj); - pred = get_irn_n(irn, pn); - - rbitset_set(used, pn); - - edges_reroute_kind(proj, pred, EDGE_KIND_NORMAL, irg); - edges_reroute_kind(proj, pred, EDGE_KIND_DEP, irg); - } - - /* the barrier also had the effect of a Keep for unused inputs. - * we now have to create an explicit Keep for them */ - n_used = rbitset_popcount(used, arity); - if (n_used < (unsigned) arity) { - int n_in = arity - (int) n_used; - ir_node **in = ALLOCAN(ir_node*, n_in); - int i = 0; - int n = 0; - ir_node *keep; - - for (i = 0; i < arity; ++i) { - if (rbitset_is_set(used, i)) - continue; - assert(n < n_in); - in[n++] = get_irn_n(irn, i); - } - keep = be_new_Barrier(get_nodes_block(irn), n_in, in); - keep_alive(keep); - sched_add_before(irn, keep); - } - - kill_node_and_preds(irn); - break; - } -} - -/** - * Kill the Barrier nodes for better peephole optimization. - */ -static void kill_barriers(ir_graph *irg) -{ - ir_node *end_blk = get_irg_end_block(irg); - ir_node *start_blk = get_irg_start_block(irg); - int i; - - /* skip the barrier on all return blocks */ - for (i = get_Block_n_cfgpreds(end_blk) - 1; i >= 0; --i) { - ir_node *be_ret = get_Block_cfgpred(end_blk, i); - ir_node *ret_blk = get_nodes_block(be_ret); - - if (ret_blk == start_blk) - continue; - - skip_barrier(ret_blk, irg); - } - - /* skip the barrier on the start block */ - start_blk = get_irg_start_block(irg); - skip_barrier(start_blk, irg); -} - /** * Check whether the node has only one user. Explicitly ignore the anchor. */ -static int has_only_one_user(ir_node *node) +bool be_has_only_one_user(ir_node *node) { int n = get_irn_n_edges(node); + int n_users; const ir_edge_t *edge; if (n <= 1) return 1; - if (n > 2) - return 0; - + n_users = 0; foreach_out_edge(node, edge) { ir_node *src = get_edge_src_irn(edge); - if (is_Anchor(src)) - return 1; + /* ignore anchor and keep-alive edges */ + if (is_Anchor(src) || is_End(src)) + continue; + n_users++; } - return 0; + return n_users == 1; } /* @@ -380,7 +272,7 @@ ir_node *be_peephole_IncSP_IncSP(ir_node *node) if (!be_is_IncSP(pred)) return node; - if (!has_only_one_user(pred)) + if (!be_has_only_one_user(pred)) return node; pred_offs = be_get_IncSP_offset(pred); @@ -415,10 +307,6 @@ void be_peephole_opt(ir_graph *irg) unsigned n_classes; unsigned i; - /* barrier nodes are used for register allocations. They hinders - * peephole optimizations, so remove them here. */ - kill_barriers(irg); - /* we sometimes find BadE nodes in float apps like optest_float.c or * kahansum.c for example... */ be_liveness_invalidate(be_get_irg_liveness(irg)); @@ -427,10 +315,10 @@ void be_peephole_opt(ir_graph *irg) arch_env = be_get_irg_arch_env(irg); lv = be_get_irg_liveness(irg); - n_classes = arch_env_get_n_reg_class(arch_env); + n_classes = arch_env->n_register_classes; register_values = XMALLOCN(ir_node**, n_classes); for (i = 0; i < n_classes; ++i) { - const arch_register_class_t *cls = arch_env_get_reg_class(arch_env, i); + const arch_register_class_t *cls = &arch_env->register_classes[i]; unsigned n_regs = arch_register_class_n_regs(cls); register_values[i] = XMALLOCN(ir_node*, n_regs); }