X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbepeephole.c;h=3649a9152ebf03969c665eeea9604ac1d3da6565;hb=ea40997bb109173b1ba75d01a958a3cf05a771c3;hp=25ec92282dce282b4ad405eb402ba961c12439d7;hpb=d4bbab11371a9c68aa2a716eaa7589b82a28691e;p=libfirm diff --git a/ir/be/bepeephole.c b/ir/be/bepeephole.c index 25ec92282..3649a9152 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. * @@ -23,10 +23,9 @@ * @author Matthias Braun * @version $Id$ */ -#ifdef HAVE_CONFIG_H #include "config.h" -#endif +#include "array_t.h" #include "bepeephole.h" #include "iredges_t.h" @@ -36,11 +35,11 @@ #include "irgmod.h" #include "error.h" -#include "beirg_t.h" +#include "beirg.h" #include "belive_t.h" -#include "bearch_t.h" -#include "benode_t.h" -#include "besched_t.h" +#include "bearch.h" +#include "benode.h" +#include "besched.h" #include "bemodule.h" DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;) @@ -57,21 +56,21 @@ static void clear_reg_value(ir_node *node) unsigned reg_idx; unsigned cls_idx; - if(!mode_is_data(get_irn_mode(node))) + if (!mode_is_data(get_irn_mode(node))) return; reg = arch_get_irn_register(node); - if(reg == NULL) { + 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); cls_idx = arch_register_class_index(cls); //assert(register_values[cls_idx][reg_idx] != NULL); - DBG((dbg, LEVEL_1, "Clear Register %s\n", reg->name)); + DB((dbg, LEVEL_1, "Clear Register %s\n", reg->name)); register_values[cls_idx][reg_idx] = NULL; } @@ -82,27 +81,27 @@ static void set_reg_value(ir_node *node) unsigned reg_idx; unsigned cls_idx; - if(!mode_is_data(get_irn_mode(node))) + if (!mode_is_data(get_irn_mode(node))) return; reg = arch_get_irn_register(node); - if(reg == NULL) { + 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); cls_idx = arch_register_class_index(cls); - DBG((dbg, LEVEL_1, "Set Register %s: %+F\n", reg->name, node)); + DB((dbg, LEVEL_1, "Set Register %s: %+F\n", reg->name, node)); register_values[cls_idx][reg_idx] = node; } static void clear_defs(ir_node *node) { /* clear values defined */ - if(get_irn_mode(node) == mode_T) { + if (get_irn_mode(node) == mode_T) { const ir_edge_t *edge; foreach_out_edge(node, edge) { ir_node *proj = get_edge_src_irn(edge); @@ -119,7 +118,7 @@ static void set_uses(ir_node *node) /* set values used */ arity = get_irn_arity(node); - for(i = 0; i < arity; ++i) { + for (i = 0; i < arity; ++i) { ir_node *in = get_irn_n(node, i); set_reg_value(in); } @@ -133,19 +132,20 @@ 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. */ -void be_peephole_before_exchange(const ir_node *old_node, ir_node *new_node) +static void be_peephole_before_exchange(const ir_node *old_node, + ir_node *new_node) { const arch_register_t *reg; const arch_register_class_t *cls; unsigned reg_idx; unsigned cls_idx; - DBG((dbg, LEVEL_1, "About to exchange and kill %+F with %+F\n", old_node, new_node)); + DB((dbg, LEVEL_1, "About to exchange and kill %+F with %+F\n", old_node, new_node)); if (current_node == old_node) { /* next node to be processed will be killed. Its scheduling predecessor @@ -194,24 +194,24 @@ 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); - for(i = 0; i < n_classes; ++i) { - const arch_register_class_t *cls = arch_env_get_reg_class(arch_env, i); + n_classes = arch_env->n_register_classes; + for (i = 0; i < n_classes; ++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); } assert(lv->nodes && "live sets must be computed"); - DBG((dbg, LEVEL_1, "\nProcessing block %+F (from end)\n", block)); + DB((dbg, LEVEL_1, "\nProcessing block %+F (from end)\n", block)); be_lv_foreach(lv, block, be_lv_state_end, l) { ir_node *node = be_lv_get_irn(lv, block, l); set_reg_value(node); } - DBG((dbg, LEVEL_1, "\nstart processing\n")); + DB((dbg, LEVEL_1, "\nstart processing\n")); /* walk the block from last insn to the first */ current_node = sched_last(block); - for( ; !sched_is_begin(current_node); + for ( ; !sched_is_begin(current_node); current_node = sched_prev(current_node)) { ir_op *op; peephole_opt_func peephole_node; @@ -228,80 +228,38 @@ static void process_block(ir_node *block, void *data) if (peephole_node == NULL) continue; + DB((dbg, LEVEL_2, "optimize %+F\n", current_node)); peephole_node(current_node); assert(!is_Bad(current_node)); } } -static void kill_node_and_preds(ir_node *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_Bad()); - 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. + * Check whether the node has only one user. Explicitly ignore the anchor. */ -static void skip_barrier(ir_node *ret_blk, ir_graph *irg) { - ir_node *irn; - - sched_foreach_reverse(ret_blk, irn) { - const ir_edge_t *edge, *next; - - if (!be_is_Barrier(irn)) +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; + + n_users = 0; + foreach_out_edge(node, edge) { + ir_node *src = get_edge_src_irn(edge); + /* ignore anchor and keep-alive edges */ + if (is_Anchor(src) || is_End(src)) continue; - - foreach_out_edge_safe(irn, edge, next) { - ir_node *proj = get_edge_src_irn(edge); - int pn = (int)get_Proj_proj(proj); - ir_node *pred = get_irn_n(irn, pn); - - edges_reroute_kind(proj, pred, EDGE_KIND_NORMAL, irg); - edges_reroute_kind(proj, pred, EDGE_KIND_DEP, irg); - } - - kill_node_and_preds(irn); - break; + n_users++; } -} - -/** - * 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; - 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); - skip_barrier(ret_blk, irg); - } - - /* skip the barrier on the start block */ - start_blk = get_irg_start_block(irg); - skip_barrier(start_blk, irg); + return n_users == 1; } /* - * Tries to optimize a beIncSp node with it's previous IncSP node. + * Tries to optimize a beIncSP node with its previous IncSP node. * Must be run from a be_peephole_opt() context. */ ir_node *be_peephole_IncSP_IncSP(ir_node *node) @@ -314,28 +272,12 @@ ir_node *be_peephole_IncSP_IncSP(ir_node *node) if (!be_is_IncSP(pred)) return node; - if (get_irn_n_edges(pred) > 1) + if (!be_has_only_one_user(pred)) return node; pred_offs = be_get_IncSP_offset(pred); curr_offs = be_get_IncSP_offset(node); - - if (pred_offs == BE_STACK_FRAME_SIZE_EXPAND) { - if (curr_offs != BE_STACK_FRAME_SIZE_SHRINK) { - return node; - } - offs = 0; - } else if (pred_offs == BE_STACK_FRAME_SIZE_SHRINK) { - if (curr_offs != BE_STACK_FRAME_SIZE_EXPAND) { - return node; - } - offs = 0; - } else if (curr_offs == BE_STACK_FRAME_SIZE_EXPAND || - curr_offs == BE_STACK_FRAME_SIZE_SHRINK) { - return node; - } else { - offs = curr_offs + pred_offs; - } + offs = curr_offs + pred_offs; /* add node offset to pred and remove our IncSP */ be_set_IncSP_offset(pred, offs); @@ -344,43 +286,37 @@ ir_node *be_peephole_IncSP_IncSP(ir_node *node) return pred; } -void be_peephole_opt(be_irg_t *birg) +void be_peephole_opt(ir_graph *irg) { - ir_graph *irg = be_get_birg_irg(birg); - unsigned n_classes; - unsigned i; - - /* barrier nodes are used for register allocations. They hinders - * peephole optimizations, so remove them here. */ - kill_barriers(irg); + unsigned n_classes; + unsigned i; /* we sometimes find BadE nodes in float apps like optest_float.c or * kahansum.c for example... */ - be_liveness_invalidate(birg->lv); - be_liveness_assure_sets(be_assure_liveness(birg)); + be_liveness_invalidate(be_get_irg_liveness(irg)); + be_liveness_assure_sets(be_assure_liveness(irg)); - arch_env = be_get_birg_arch_env(birg); - lv = be_get_birg_liveness(birg); + arch_env = be_get_irg_arch_env(irg); + lv = be_get_irg_liveness(irg); - n_classes = arch_env_get_n_reg_class(arch_env); - register_values = alloca(sizeof(register_values[0]) * n_classes); - for(i = 0; i < n_classes; ++i) { - const arch_register_class_t *cls = arch_env_get_reg_class(arch_env, i); + 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->register_classes[i]; unsigned n_regs = arch_register_class_n_regs(cls); - register_values[i] = alloca(sizeof(ir_node*) * n_regs); + register_values[i] = XMALLOCN(ir_node*, n_regs); } irg_block_walk_graph(irg, process_block, NULL, NULL); -} -void be_peephole_init(void) -{ - clear_irp_opcodes_generic_func(); + for (i = 0; i < n_classes; ++i) { + xfree(register_values[i]); + } + xfree(register_values); } +BE_REGISTER_MODULE_CONSTRUCTOR(be_init_peephole) void be_init_peephole(void) { FIRM_DBG_REGISTER(dbg, "firm.be.peephole"); } - -BE_REGISTER_MODULE_CONSTRUCTOR(be_init_peephole);