X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;ds=sidebyside;f=ir%2Fbe%2Fbepeephole.c;h=79700d134979a7f27372b6bde3fa5ad8271aa3a7;hb=871e6821cfb5583cdf3c6207b9f28fada1f93868;hp=cc726c92fbfbcacef1ba05bc6d5ac947d7351497;hpb=ea75e9d38674b468f602a0699fb64b9c01254797;p=libfirm diff --git a/ir/be/bepeephole.c b/ir/be/bepeephole.c index cc726c92f..79700d134 100644 --- a/ir/be/bepeephole.c +++ b/ir/be/bepeephole.c @@ -69,7 +69,7 @@ static void clear_reg_value(ir_node *node) 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; } @@ -93,7 +93,7 @@ static void set_reg_value(ir_node *node) 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; } @@ -144,7 +144,7 @@ static void be_peephole_before_exchange(const ir_node *old_node, 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 @@ -201,12 +201,12 @@ static void process_block(ir_node *block, void *data) } 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); @@ -227,6 +227,7 @@ 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)); } @@ -255,16 +256,24 @@ static void kill_node_and_preds(ir_node *node) /** * Walk through the block schedule and skip all barrier nodes. */ -static void skip_barrier(ir_node *ret_blk, ir_graph *irg) +static void skip_barrier(ir_node *block, ir_graph *irg) { ir_node *irn; - sched_foreach_reverse(ret_blk, 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; @@ -276,10 +285,33 @@ static void skip_barrier(ir_node *ret_blk, ir_graph *irg) 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; } @@ -291,7 +323,7 @@ static void skip_barrier(ir_node *ret_blk, ir_graph *irg) static void kill_barriers(ir_graph *irg) { ir_node *end_blk = get_irg_end_block(irg); - ir_node *start_blk; + ir_node *start_blk = get_irg_start_block(irg); int i; /* skip the barrier on all return blocks */ @@ -299,6 +331,9 @@ static void kill_barriers(ir_graph *irg) 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); } @@ -374,11 +409,10 @@ 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; + unsigned n_classes; + unsigned i; /* barrier nodes are used for register allocations. They hinders * peephole optimizations, so remove them here. */ @@ -386,21 +420,26 @@ void be_peephole_opt(be_irg_t *birg) /* 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 = ALLOCAN(ir_node**, n_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); unsigned n_regs = arch_register_class_n_regs(cls); - register_values[i] = ALLOCAN(ir_node*, n_regs); + register_values[i] = XMALLOCN(ir_node*, n_regs); } irg_block_walk_graph(irg, process_block, NULL, NULL); + + for (i = 0; i < n_classes; ++i) { + xfree(register_values[i]); + } + xfree(register_values); } BE_REGISTER_MODULE_CONSTRUCTOR(be_init_peephole);