-#ifdef HAVE_CONFIG_H
+/*
+ * 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.
+ */
+
+/**
+ * @file
+ * @brief Backend IRG modification routines.
+ * @author Sebastian Hack, Daniel Grund, Matthias Braun, Christian Wuerdig
+ * @date 04.05.2005
+ *
+ * This file contains the following IRG modifications for be routines:
+ * - insertion of Perm nodes
+ * - empty block elimination
+ * - a simple dead node elimination (set inputs of unreachable nodes to BAD)
+ */
#include "config.h"
-#endif
#include <stdlib.h>
#include "pmap.h"
#include "util.h"
#include "debug.h"
+#include "error.h"
+#include "xmalloc.h"
#include "irflag_t.h"
#include "ircons_t.h"
#include "irnode_t.h"
+#include "ircons_t.h"
#include "irmode_t.h"
#include "irdom_t.h"
#include "iredges_t.h"
+#include "irgraph_t.h"
#include "irgopt.h"
+#include "irgmod.h"
+#include "irprintf.h"
+#include "irgwalk.h"
#include "be_t.h"
+#include "bechordal_t.h"
#include "bearch.h"
-#include "besched_t.h"
+#include "besched.h"
#include "belive_t.h"
-#include "benode_t.h"
-
+#include "benode.h"
+#include "beutil.h"
+#include "beinsn_t.h"
+#include "bessaconstr.h"
+#include "beirg.h"
#include "beirgmod.h"
+#include "bemodule.h"
-#define DBG_MODULE firm_dbg_register("firm.be.irgmod")
-
-struct _dom_front_info_t {
- pmap *df_map;
-};
+DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
-static void compute_df_local(ir_node *bl, void *data)
+static int cmp_node_nr(const void *a, const void *b)
{
- pmap *df_map = ((dom_front_info_t *) data)->df_map;
- ir_node *idom = get_Block_idom(bl);
- pset *df = pmap_get(df_map, bl);
- int i, n;
-
- /*
- * Create a new dom frot set for this node,
- * if none exists.
- */
- if(!df)
- pmap_insert(df_map, bl, pset_new_ptr(16));
-
- for(i = 0, n = get_irn_arity(bl); i < n; ++i) {
-
- /* The predecessor block */
- ir_node *pred = get_nodes_block(get_irn_n(bl, i));
-
- /* The dominance frontier set of the predecessor. */
- pset *df = pmap_get(df_map, pred);
- if(!df) {
- df = pset_new_ptr(16);
- pmap_insert(df_map, pred, df);
- }
-
- assert(df && "dom front set must have been created for this node");
-
- if(pred != idom && bl)
- pset_insert_ptr(df, bl);
- }
+ ir_node **p1 = (ir_node**)a;
+ ir_node **p2 = (ir_node**)b;
+ long n1 = get_irn_node_nr(*p1);
+ long n2 = get_irn_node_nr(*p2);
+ return (n1>n2) - (n1<n2);
}
-static void compute_df_up(ir_node *bl, void *data)
-{
- pmap *df_map = ((dom_front_info_t *) data)->df_map;
- ir_node *y;
-
- for(y = get_Block_dominated_first(bl); y; y = get_Block_dominated_next(y)) {
- ir_node *w;
- pset *df = pmap_get(df_map, y);
+/*
+ ___ _ ____
+ |_ _|_ __ ___ ___ _ __| |_ | _ \ ___ _ __ _ __ ___
+ | || '_ \/ __|/ _ \ '__| __| | |_) / _ \ '__| '_ ` _ \
+ | || | | \__ \ __/ | | |_ | __/ __/ | | | | | | |
+ |___|_| |_|___/\___|_| \__| |_| \___|_| |_| |_| |_|
- for(w = pset_first(df); w; w = pset_next(df))
- if(!block_dominates(bl, w) || bl == w)
- pset_insert_ptr(df, w);
- }
-}
+*/
-dom_front_info_t *be_compute_dominance_frontiers(ir_graph *irg)
+ir_node *insert_Perm_before(ir_graph *irg, const arch_register_class_t *cls,
+ ir_node *pos)
{
- dom_front_info_t *info = malloc(sizeof(*info));
-
- info->df_map = pmap_create();
-
- /*
- * This must be called as a post walker, since the dom front sets
- * of all predecessors must be created when a block is reached.
- */
- dom_tree_walk_irg(irg, NULL, compute_df_local, info);
- dom_tree_walk_irg(irg, NULL, compute_df_up, info);
- return info;
+ be_lv_t *lv = be_get_irg_liveness(irg);
+ ir_nodeset_t live;
+
+ ir_node *perm, **nodes;
+ size_t i, n;
+
+ DBG((dbg, LEVEL_1, "Insert Perm before: %+F\n", pos));
+
+ ir_nodeset_init(&live);
+ be_liveness_nodes_live_before(lv, cls, pos, &live);
+
+ n = ir_nodeset_size(&live);
+ if (n == 0) {
+ ir_nodeset_destroy(&live);
+ return NULL;
+ }
+
+ nodes = XMALLOCN(ir_node*, n);
+
+ DBG((dbg, LEVEL_1, "live:\n"));
+ i = 0;
+ foreach_ir_nodeset(&live, irn, iter) {
+ DBG((dbg, LEVEL_1, "\t%+F\n", irn));
+ nodes[i] = irn;
+ i++;
+ }
+ ir_nodeset_destroy(&live);
+ /* make the input order deterministic */
+ qsort(nodes, n, sizeof(nodes[0]), cmp_node_nr);
+
+ ir_node *const bl = get_nodes_block(pos);
+ perm = be_new_Perm(cls, bl, n, nodes);
+ sched_add_before(pos, perm);
+ free(nodes);
+
+ for (i = 0; i < n; ++i) {
+ ir_node *perm_op = get_irn_n(perm, i);
+ be_ssa_construction_env_t senv;
+
+ ir_mode *mode = get_irn_mode(perm_op);
+ ir_node *proj = new_r_Proj(perm, mode, i);
+
+ be_ssa_construction_init(&senv, irg);
+ be_ssa_construction_add_copy(&senv, perm_op);
+ be_ssa_construction_add_copy(&senv, proj);
+ be_ssa_construction_fix_users(&senv, perm_op);
+ be_ssa_construction_update_liveness_phis(&senv, lv);
+ be_liveness_update(lv, perm_op);
+ be_liveness_update(lv, proj);
+ be_ssa_construction_destroy(&senv);
+ }
+
+ return perm;
}
-void be_free_dominance_frontiers(dom_front_info_t *info)
-{
- pmap_entry *ent;
-
- for(ent = pmap_first(info->df_map); ent; ent = pmap_next(info->df_map))
- del_pset(ent->value);
+static int blocks_removed;
- pmap_destroy(info->df_map);
- free(info);
+/**
+ * Post-block-walker: Find blocks containing only one jump and
+ * remove them.
+ */
+static void remove_empty_block(ir_node *block)
+{
+ int i;
+ int arity;
+ ir_node *pred;
+ ir_node *succ_block;
+ ir_node *jump = NULL;
+ ir_graph *irg = get_irn_irg(block);
+ ir_entity *entity;
+
+ if (irn_visited_else_mark(block))
+ return;
+
+ if (get_Block_n_cfgpreds(block) != 1)
+ goto check_preds;
+
+ sched_foreach(block, node) {
+ if (! is_Jmp(node)
+ && !(arch_get_irn_flags(node) & arch_irn_flags_simple_jump))
+ goto check_preds;
+ if (jump != NULL) {
+ /* we should never have 2 jumps in a block */
+ panic("found 2 jumps in a block");
+ }
+ jump = node;
+ }
+
+ if (jump == NULL)
+ goto check_preds;
+
+ entity = get_Block_entity(block);
+ pred = get_Block_cfgpred(block, 0);
+ succ_block = NULL;
+ foreach_out_edge_safe(jump, edge) {
+ int pos = get_edge_src_pos(edge);
+
+ assert(succ_block == NULL);
+ succ_block = get_edge_src_irn(edge);
+ if (get_Block_entity(succ_block) != NULL && entity != NULL) {
+ /*
+ * Currently we can add only one label for a block.
+ * Therefore we cannot combine them if both block already have one.
+ */
+ goto check_preds;
+ }
+
+ set_irn_n(succ_block, pos, pred);
+ }
+
+ if (entity != NULL) {
+ /* move the label to the successor block */
+ set_Block_entity(succ_block, entity);
+ }
+
+ /* there can be some non-scheduled Pin nodes left in the block, move them
+ * to the succ block (Pin) or pred block (Sync) */
+ foreach_out_edge_safe(block, edge) {
+ ir_node *const node = get_edge_src_irn(edge);
+
+ if (node == jump)
+ continue;
+ /* we simply kill Pins, because there are some strange interactions
+ * between jump threading, which produce PhiMs with Pins, we simply
+ * kill the pins here, everything is scheduled anyway */
+ if (is_Pin(node)) {
+ exchange(node, get_Pin_op(node));
+ continue;
+ }
+ if (is_Sync(node)) {
+ set_nodes_block(node, get_nodes_block(pred));
+ continue;
+ }
+ if (is_End(node)) { /* End-keep, reroute it to the successor */
+ int pos = get_edge_src_pos(edge);
+ set_irn_n(node, pos, succ_block);
+ continue;
+ }
+ panic("Unexpected node %+F in block %+F with empty schedule", node, block);
+ }
+
+ set_Block_cfgpred(block, 0, new_r_Bad(irg, mode_X));
+ kill_node(jump);
+ blocks_removed = 1;
+
+ /* check predecessor */
+ remove_empty_block(get_nodes_block(pred));
+ return;
+
+check_preds:
+ arity = get_Block_n_cfgpreds(block);
+ for (i = 0; i < arity; ++i) {
+ ir_node *pred = get_Block_cfgpred_block(block, i);
+ remove_empty_block(pred);
+ }
}
-pset *be_get_dominance_frontier(dom_front_info_t *info, ir_node *block)
+/* removes basic blocks that just contain a jump instruction */
+int be_remove_empty_blocks(ir_graph *irg)
{
- return pmap_get(info->df_map, block);
+ ir_node *end;
+ int i, arity;
+
+ blocks_removed = 0;
+
+ ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
+ inc_irg_visited(irg);
+ remove_empty_block(get_irg_end_block(irg));
+ end = get_irg_end(irg);
+ arity = get_irn_arity(end);
+ for (i = 0; i < arity; ++i) {
+ ir_node *pred = get_irn_n(end, i);
+ if (!is_Block(pred))
+ continue;
+ remove_empty_block(pred);
+ }
+ ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
+
+ if (blocks_removed) {
+ /* invalidate analysis info */
+ clear_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE);
+ }
+ return blocks_removed;
}
+//---------------------------------------------------------------------------
+
+typedef struct remove_dead_nodes_env_t_ {
+ bitset_t *reachable;
+ ir_graph *irg;
+ be_lv_t *lv;
+} remove_dead_nodes_env_t;
+
/**
- * Algorithm to place the Phi-Functions.
- * @see Appel, Modern Compiler Implementation in Java, 2nd ed., p. 399ff
- *
- * This function takes an original node and a set of already placed
- * copies of that node called @p copies. It places phi nodes at the
- * iterated dominance frontiers of these copies and puts these phi nodes
- * in the @p copies set, since they are another form of copies of the
- * original value.
- *
- * The rename phase (see below) is responsible for fixing up the usages
- * of the original node.
- *
- * @param orig The original node.
- * @param copies A set contianing nodes representing a copy of the
- * original node. Each node must be inserted into the block's schedule.
- * @param copy_blocks A set in which the blocks are recorded which
- * contain a copy. This is just for efficiency in later phases (see
- * rename).
+ * Post-walker: remember all visited nodes in a bitset.
*/
-static void place_phi_functions(ir_node *orig, pset *copies,
- pset *copy_blocks, dom_front_info_t *df_info)
+static void mark_dead_nodes_walker(ir_node *node, void *data)
{
- int i;
- ir_node *orig_block = get_nodes_block(orig);
- ir_graph *irg = get_irn_irg(orig);
- ir_mode *mode = get_irn_mode(orig);
- pdeq *worklist = new_pdeq();
- pset *phi_blocks = pset_new_ptr(8);
- ir_node **ins = NULL;
- void *it;
- firm_dbg_module_t *dbg = DBG_MODULE;
-
- /*
- * Allocate an array for all blocks where the copies and the original
- * value were defined.
- */
- int n_orig_blocks = pset_count(copy_blocks);
- ir_node **orig_blocks = malloc(n_orig_blocks * sizeof(orig_blocks[0]));
-
- /*
- * Fill the worklist queue and the rest of the orig blocks array.
- */
- for(it = pset_first(copies), i = 0; it; it = pset_next(copies)) {
- ir_node *copy_block = get_nodes_block(it);
-
- if(!block_dominates(orig_block, copy_block)) {
- assert(block_dominates(orig_block, copy_block)
- && "The block of the copy must be dominated by the block of the value");
- }
-
- pdeq_putr(worklist, copy_block);
- orig_blocks[i++] = copy_block;
- }
-
- while(!pdeq_empty(worklist)) {
- ir_node *bl = pdeq_getl(worklist);
- ir_node *y;
- pset *df = be_get_dominance_frontier(df_info, bl);
-
- for(y = pset_first(df); y; y = pset_next(df)) {
- int n_preds = get_irn_arity(y);
-
- if(!pset_find_ptr(phi_blocks, y)) {
- ir_node *phi;
- int insert = 1;
-
- /*
- * Set the orig node as the only operand of the
- * phi node.
- */
- ins = realloc(ins, n_preds * sizeof(ins[0]));
- for(i = 0; i < n_preds; ++i)
- ins[i] = orig;
-
- /* Insert phi node */
- phi = new_r_Phi(irg, y, n_preds, ins, mode);
- DBG((dbg, LEVEL_2, " inserting phi %+F with %d args in block %+F\n",
- phi, n_preds, bl));
-
- /*
- * The phi node itself is also a copy of the original
- * value. So put it in the copies set also, so that
- * the rename phase can treat them right.
- */
- pset_insert_ptr(copies, phi);
- pset_insert_ptr(copy_blocks, y);
-
- /* Insert the phi node into the schedule */
- sched_add_before(sched_first(y), phi);
-
- /* Insert the phi node in the phi blocks set. */
- pset_insert_ptr(phi_blocks, y);
-
- /*
- * If orig or a copy of it were not defined in y,
- * add y to the worklist.
- */
- for(i = 0; i < n_orig_blocks; ++i)
- if(orig_blocks[i] == y) {
- insert = 0;
- break;
- }
-
- if(insert)
- pdeq_putr(worklist, y);
-
- }
- }
- }
-
- del_pset(phi_blocks);
- del_pdeq(worklist);
-
- free(orig_blocks);
-
- if(ins)
- free(ins);
+ remove_dead_nodes_env_t *env = (remove_dead_nodes_env_t*) data;
+ bitset_set(env->reachable, get_irn_idx(node));
}
/**
- * Find the copy of the given original node whose value is 'active'
- * at a usage.
- *
- * The usage is given as a node and a position. Initially, the given operand
- * points to a node for which copies were introduced. We have to find
- * the valid copy for this usage. This is done by travering the
- * dominance tree upwards. If the usage is a phi function, we start
- * traversing from the predecessor block which corresponds to the phi
- * usage.
- *
- * @param usage The node which uses the original node.
- * @param pos The number of the argument which corresponds to the
- * original node.
- * @param copy_blocks A set containing all basic block in which copies
- * of the original node are located.
- * @param copies A set containing all node which are copies from the
- * original node.
- * @return The valid copy for usage.
+ * Post-block-walker:
+ * Walk through the schedule of every block and remove all dead nodes from it.
*/
-static ir_node *search_def(ir_node *usage, int pos, pset *copies, pset *copy_blocks)
+static void remove_dead_nodes_walker(ir_node *block, void *data)
{
- ir_node *curr_bl;
- ir_node *start_irn;
-
- curr_bl = get_nodes_block(usage);
-
- /*
- * If the usage is in a phi node, search the copy in the
- * predecessor denoted by pos.
- */
- if(is_Phi(usage)) {
- curr_bl = get_nodes_block(get_irn_n(curr_bl, pos));
- start_irn = sched_last(curr_bl);
- }
-
- else {
- start_irn = sched_prev(usage);
- }
-
- /*
- * Traverse the dominance tree upwards from the
- * predecessor block of the usage.
- */
- while(curr_bl != NULL) {
-
- /*
- * If this block contains a copy, search the block
- * instruction by instruction.
- */
- if(pset_find_ptr(copy_blocks, curr_bl)) {
- ir_node *irn;
-
- /* Look at each instruction from last to first. */
- for(irn = start_irn; !is_Block(irn); irn = sched_prev(irn)) {
-
- /* Take the first copy we find. */
- if(pset_find_ptr(copies, irn))
- return irn;
- }
- }
-
- /* If were not done yet, look in the immediate dominator */
- curr_bl = get_Block_idom(curr_bl);
- if(curr_bl)
- start_irn = sched_last(curr_bl);
- }
-
- return NULL;
+ remove_dead_nodes_env_t *env = (remove_dead_nodes_env_t*) data;
+ ir_node *node, *next;
+
+ for (node = sched_first(block); ! sched_is_end(node); node = next) {
+ /* get next node now, as after calling sched_remove it will be invalid */
+ next = sched_next(node);
+
+ if (bitset_is_set(env->reachable, get_irn_idx(node)))
+ continue;
+
+ if (env->lv != NULL)
+ be_liveness_remove(env->lv, node);
+ sched_remove(node);
+
+ /* kill projs */
+ if (get_irn_mode(node) == mode_T) {
+ foreach_out_edge_safe(node, edge) {
+ ir_node *proj = get_edge_src_irn(edge);
+ if (!is_Proj(proj))
+ continue;
+ if (env->lv != NULL)
+ be_liveness_remove(env->lv, proj);
+ kill_node(proj);
+ }
+ }
+ kill_node(node);
+ }
}
-static void fix_usages(ir_node *orig, pset *copies, pset *copy_blocks)
+void be_remove_dead_nodes_from_schedule(ir_graph *irg)
{
- int i = 0;
- int n_outs = 0;
- const ir_edge_t *edge;
- firm_dbg_module_t *dbg = DBG_MODULE;
-
- struct {
- ir_node *irn;
- int pos;
- } *outs;
-
- /* Count the number of outs. */
- foreach_out_edge(orig, edge)
- n_outs++;
-
- /*
- * Put all outs into an array.
- * This is neccessary, since the outs would be modified while
- * interating on them what could bring the outs module in trouble.
- */
- DBG((dbg, LEVEL_2, " Users of %+F\n", orig));
- outs = malloc(n_outs * sizeof(outs[0]));
- foreach_out_edge(orig, edge) {
- outs[i].irn = get_edge_src_irn(edge);
- outs[i].pos = get_edge_src_pos(edge);
- i += 1;
- }
-
- /*
- * Search the valid def for each out and set it.
- */
- for(i = 0; i < n_outs; ++i) {
- ir_node *def;
- ir_node *irn = outs[i].irn;
- int pos = outs[i].pos;
-
- def = search_def(irn, pos, copies, copy_blocks);
- DBG((dbg, LEVEL_2, " %+F(%d) -> %+F\n", irn, pos, def));
-
- if(def != NULL)
- set_irn_n(irn, pos, def);
- }
-
- free(outs);
+ remove_dead_nodes_env_t env;
+ env.reachable = bitset_alloca(get_irg_last_idx(irg));
+ env.lv = be_get_irg_liveness(irg);
+ env.irg = irg;
+
+ // mark all reachable nodes
+ irg_walk_graph(irg, mark_dead_nodes_walker, NULL, &env);
+
+ // walk schedule and remove non-marked nodes
+ irg_block_walk_graph(irg, remove_dead_nodes_walker, NULL, &env);
}
-void be_introduce_copies(dom_front_info_t *info, ir_node *orig, int n, ir_node *copy_nodes[])
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_irgmod)
+void be_init_irgmod(void)
{
- pset *copies = pset_new_ptr(2 * n);
- pset *copy_blocks = pset_new_ptr(2 * n);
- int save_optimize = get_optimize();
- int save_normalize = get_opt_normalize();
- firm_dbg_module_t *dbg = DBG_MODULE;
- int i;
-
- firm_dbg_set_mask(dbg, -1);
- DBG((dbg, LEVEL_1, "Introducing following copies of %+F\n", orig));
-
- /* Fill the sets. */
- pset_insert_ptr(copies, orig);
- pset_insert_ptr(copy_blocks, get_nodes_block(orig));
-
- for(i = 0; i < n; ++i) {
- DBG((dbg, LEVEL_1,
- " %+F in block %+F\n", copy_nodes[i], get_nodes_block(copy_nodes[i])));
- pset_insert_ptr(copies, copy_nodes[i]);
- pset_insert_ptr(copy_blocks, get_nodes_block(copy_nodes[i]));
- }
-
- /*
- * Disable optimization so that the phi functions do not
- * disappear.
- */
- set_optimize(0);
- set_opt_normalize(0);
-
- /*
- * Place the phi functions and reroute the usages.
- */
- place_phi_functions(orig, copies, copy_blocks, info);
- fix_usages(orig, copies, copy_blocks);
-
- /* reset the optimizations */
- set_optimize(save_optimize);
- set_opt_normalize(save_normalize);
-
- del_pset(copies);
- del_pset(copy_blocks);
+ FIRM_DBG_REGISTER(dbg, "firm.be.irgmod");
}