X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firgopt.c;h=ae510a8016b6c67038ab74d8a69572265cb24b97;hb=3142e505b37d6e992ebfe599ee69027fea55f75c;hp=b1afa380101c35224acd2480eb07aba800e275f7;hpb=bb9f2e36362333c6635b89f5258171b06c786608;p=libfirm

diff --git a/ir/ir/irgopt.c b/ir/ir/irgopt.c
index b1afa3801..ae510a801 100644
--- a/ir/ir/irgopt.c
+++ b/ir/ir/irgopt.c
@@ -39,6 +39,7 @@
 
 #include "adt/pdeq.h"
 
+#include "irpass_t.h"
 #include "irflag_t.h"
 #include "iredges_t.h"
 #include "irtools.h"
@@ -50,7 +51,8 @@
 /**
  * A wrapper around optimize_inplace_2() to be called from a walker.
  */
-static void optimize_in_place_wrapper (ir_node *n, void *env) {
+static void optimize_in_place_wrapper(ir_node *n, void *env)
+{
 	ir_node *optimized = optimize_in_place_2(n);
 	(void) env;
 
@@ -67,26 +69,27 @@ static void optimize_in_place_wrapper (ir_node *n, void *env) {
  *
  * @note current_ir_graph must be set
  */
-static inline void do_local_optimize(ir_node *n) {
+static inline void do_local_optimize(ir_node *n)
+{
+	ir_graph *irg = get_irn_irg(n);
+
 	/* Handle graph state */
-	assert(get_irg_phase_state(current_ir_graph) != phase_building);
+	assert(get_irg_phase_state(irg) != phase_building);
 
 	if (get_opt_global_cse())
-		set_irg_pinned(current_ir_graph, op_pin_state_floats);
-	set_irg_outs_inconsistent(current_ir_graph);
-	set_irg_doms_inconsistent(current_ir_graph);
-	set_irg_loopinfo_inconsistent(current_ir_graph);
+		set_irg_pinned(irg, op_pin_state_floats);
+	set_irg_doms_inconsistent(irg);
 
 	/* Clean the value_table in irg for the CSE. */
-	del_identities(current_ir_graph->value_table);
-	current_ir_graph->value_table = new_identities();
+	new_identities(irg);
 
 	/* walk over the graph */
 	irg_walk(n, firm_clear_link, optimize_in_place_wrapper, NULL);
 }
 
 /* Applies local optimizations (see iropt.h) to all nodes reachable from node n */
-void local_optimize_node(ir_node *n) {
+void local_optimize_node(ir_node *n)
+{
 	ir_graph *rem = current_ir_graph;
 	current_ir_graph = get_irn_irg(n);
 
@@ -95,48 +98,26 @@ void local_optimize_node(ir_node *n) {
 	current_ir_graph = rem;
 }
 
-/**
- * Block-Walker: uses dominance depth to mark dead blocks.
- */
-static void kill_dead_blocks(ir_node *block, void *env) {
-	(void) env;
-
-	if (get_Block_dom_depth(block) < 0) {
-		/*
-		 * Note that the new dominance code correctly handles
-		 * the End block, i.e. it is always reachable from Start
-		 */
-		set_Block_dead(block);
-	}
-}
-
-/* Applies local optimizations (see iropt.h) to all nodes reachable from node n. */
-void local_optimize_graph(ir_graph *irg) {
-	ir_graph *rem = current_ir_graph;
-	current_ir_graph = irg;
-
-	if (get_irg_dom_state(irg) == dom_consistent)
-		irg_block_walk_graph(irg, NULL, kill_dead_blocks, NULL);
-
-	do_local_optimize(get_irg_end(irg));
-
-	current_ir_graph = rem;
+static void enqueue_node(ir_node *node, pdeq *waitq)
+{
+	if (get_irn_link(node) == waitq)
+		return;
+	pdeq_putr(waitq, node);
+	set_irn_link(node, waitq);
 }
 
 /**
  * Enqueue all users of a node to a wait queue.
  * Handles mode_T nodes.
  */
-static void enqueue_users(ir_node *n, pdeq *waitq) {
+static void enqueue_users(ir_node *n, pdeq *waitq)
+{
 	const ir_edge_t *edge;
 
 	foreach_out_edge(n, edge) {
 		ir_node *succ = get_edge_src_irn(edge);
 
-		if (get_irn_link(succ) != waitq) {
-			pdeq_putr(waitq, succ);
-			set_irn_link(succ, waitq);
-		}
+		enqueue_node(succ, waitq);
 		if (get_irn_mode(succ) == mode_T) {
 		/* A mode_T node has Proj's. Because most optimizations
 			run on the Proj's we have to enqueue them also. */
@@ -145,13 +126,51 @@ static void enqueue_users(ir_node *n, pdeq *waitq) {
 	}
 }
 
+/**
+ * Block-Walker: uses dominance depth to mark dead blocks.
+ */
+static void find_unreachable_blocks(ir_node *block, void *env)
+{
+	pdeq *waitq = (pdeq*) env;
+
+	if (get_Block_dom_depth(block) < 0) {
+		ir_graph *irg = get_irn_irg(block);
+		ir_node  *end = get_irg_end(irg);
+
+		const ir_edge_t *edge;
+		foreach_block_succ(block, edge) {
+			const ir_edge_t *edge2;
+			ir_node *succ_block = get_edge_src_irn(edge);
+			enqueue_node(succ_block, waitq);
+			foreach_out_edge(succ_block, edge2) {
+				ir_node *succ = get_edge_src_irn(edge2);
+				if (is_Phi(succ))
+					enqueue_node(succ, waitq);
+			}
+		}
+		enqueue_node(end, waitq);
+	}
+}
+
+/* Applies local optimizations (see iropt.h) to all nodes reachable from node n. */
+void local_optimize_graph(ir_graph *irg)
+{
+	ir_graph *rem = current_ir_graph;
+	current_ir_graph = irg;
+
+	do_local_optimize(get_irg_end(irg));
+
+	current_ir_graph = rem;
+}
+
 /**
  * Data flow optimization walker.
- * Optimizes all nodes and enqueue it's users
+ * Optimizes all nodes and enqueue its users
  * if done.
  */
-static void opt_walker(ir_node *n, void *env) {
-	pdeq *waitq = env;
+static void opt_walker(ir_node *n, void *env)
+{
+	pdeq *waitq = (pdeq*)env;
 	ir_node *optimized;
 
 	optimized = optimize_in_place_2(n);
@@ -163,71 +182,156 @@ static void opt_walker(ir_node *n, void *env) {
 	}
 }
 
+/**
+ * Return the number of non-Bad predecessors of the given node.
+ */
+static int count_non_bads(ir_node *node)
+{
+	int arity = get_irn_arity(node);
+	int count = 0;
+	int i;
+	for (i = 0; i < arity; ++i) {
+		if (!is_Bad(get_irn_n(node, i)))
+			++count;
+	}
+	return count;
+}
+
+/**
+ * Block-walker, remove Bad block predecessors and shorten Phis.
+ * Phi links must be uptodate.
+ */
+static void block_remove_bads(ir_node *block, void *env)
+{
+	int *changed = (int *)env;
+	int i, j;
+	ir_node **new_in, *new_block, *phi;
+	const int max = get_irn_arity(block);
+	const int new_max = count_non_bads(block);
+	assert(max >= new_max);
+
+	if (is_Bad(block) || max == new_max)
+		return;
+
+	new_in = ALLOCAN(ir_node*, new_max);
+	*changed = 1;
+
+	assert(get_Block_dom_depth(block) >= 0);
+
+	/* 1. Create a new block without Bad inputs */
+	for (i = j = 0; i < max; ++i) {
+		ir_node *block_pred = get_irn_n(block, i);
+		if (!is_Bad(block_pred)) {
+			new_in[j++] = block_pred;
+		}
+	}
+	assert(j == new_max);
+
+	/* If the end block is unreachable, it might have zero predecessors. */
+	if (new_max == 0) {
+		ir_node *end_block = get_irg_end_block(get_irn_irg(block));
+		if (block == end_block) {
+			set_irn_in(block, new_max, new_in);
+			return;
+		}
+	}
+
+	new_block = new_r_Block(get_irn_irg(block), new_max, new_in);
+
+	/* 2. Remove inputs on Phis, where the block input is Bad. */
+	phi = get_Block_phis(block);
+	if (phi != NULL) {
+		do {
+			ir_node *next = get_Phi_next(phi);
+			if (get_irn_arity(phi) != new_max) {
+				ir_node *new_phi;
+
+				for (i = j = 0; i < max; ++i) {
+					ir_node *block_pred = get_irn_n(block, i);
+
+					if (!is_Bad(block_pred)) {
+						ir_node *pred = get_irn_n(phi, i);
+						new_in[j++] = pred;
+					}
+				}
+				assert(j == new_max);
+
+				new_phi = new_r_Phi(new_block, new_max, new_in, get_irn_mode(phi));
+				exchange(phi, new_phi);
+			}
+			phi = next;
+		} while (phi != NULL);
+	}
+
+	exchange(block, new_block);
+}
+
+/* Remove Bad nodes from Phi and Block inputs.
+ *
+ * Precondition: No unreachable code.
+ * Postcondition: No Bad nodes.
+ */
+static int remove_Bads(ir_graph *irg)
+{
+	int changed = 0;
+	/* build phi list per block */
+	irg_walk_graph(irg, firm_clear_block_phis, firm_collect_block_phis, NULL);
+
+	/* actually remove Bads */
+	irg_block_walk_graph(irg, NULL, block_remove_bads, (void *)&changed);
+
+	return changed;
+}
+
 /* Applies local optimizations to all nodes in the graph until fixpoint. */
-int optimize_graph_df(ir_graph *irg) {
+int optimize_graph_df(ir_graph *irg)
+{
 	pdeq     *waitq = new_pdeq();
 	ir_graph *rem = current_ir_graph;
 	ir_node  *end;
-	int      i, state, n_ka, changed;
+	int      state, changed;
 
 	current_ir_graph = irg;
 
 	state = edges_assure(irg);
 
-	if (get_opt_global_cse())
-		set_irg_pinned(current_ir_graph, op_pin_state_floats);
-
 	/* Clean the value_table in irg for the CSE. */
-	del_identities(irg->value_table);
-	irg->value_table = new_identities();
+	new_identities(irg);
+
+	if (get_opt_global_cse()) {
+		set_irg_pinned(irg, op_pin_state_floats);
+	}
 
-	if (get_irg_dom_state(irg) == dom_consistent)
-		irg_block_walk_graph(irg, NULL, kill_dead_blocks, NULL);
+	/* The following enables unreachable code elimination (=Blocks may be
+	 * Bad). */
+	set_irg_state(irg, IR_GRAPH_STATE_BAD_BLOCK);
 
 	/* invalidate info */
-	set_irg_outs_inconsistent(irg);
 	set_irg_doms_inconsistent(irg);
-	set_irg_loopinfo_inconsistent(irg);
 
 	ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
 
-	end  = get_irg_end(irg);
-	n_ka = get_End_n_keepalives(end);
+	/* Calculate dominance so we can kill unreachable code */
+	assure_doms(irg);
 
 	/* walk over the graph, but don't touch keep-alives */
-	irg_walk(get_irg_end_block(irg), NULL, opt_walker, waitq);
-
-	/*
-	 * Optimize keep-alives by removing superfluous ones.
-	 * Beware: the last transformation might add new keep-alives
-	 * that keep blocks that are where visited! So, check only the
-	 * "old" keep-alives, not the new ones!
-	 *
-	 * FIXME: it might be better to completely remove this
-	 * optimization here ...
-	 */
-	for (i = n_ka - 1; i >= 0; --i) {
-		ir_node *ka = get_End_keepalive(end, i);
-
-		if (irn_visited(ka) && !is_irn_keep(ka)) {
-			/* this node can be regularly visited, no need to keep it */
-			set_End_keepalive(end, i, get_irg_bad(irg));
-		}
-	}
-	/* now walk again and visit all not yet visited nodes */
-	set_irg_visited(current_ir_graph, get_irg_visited(irg) - 1);
-	irg_walk(get_irg_end(irg), NULL, opt_walker, waitq);
+	irg_walk_graph(irg, NULL, opt_walker, waitq);
 
 	/* any optimized nodes are stored in the wait queue,
 	 * so if it's not empty, the graph has been changed */
 	changed = !pdeq_empty(waitq);
 
-	/* finish the wait queue */
-	while (! pdeq_empty(waitq)) {
-		ir_node *n = pdeq_getl(waitq);
-		if (! is_Bad(n))
+	while (!pdeq_empty(waitq)) {
+		/* finish the wait queue */
+		while (! pdeq_empty(waitq)) {
+			ir_node *n = (ir_node*)pdeq_getl(waitq);
 			opt_walker(n, waitq);
+		}
+		/* Calculate dominance so we can kill unreachable code */
+		compute_doms(irg);
+		irg_block_walk_graph(irg, NULL, find_unreachable_blocks, waitq);
 	}
+	set_irg_doms_inconsistent(irg);
 
 	del_pdeq(waitq);
 
@@ -236,6 +340,23 @@ int optimize_graph_df(ir_graph *irg) {
 	if (! state)
 		edges_deactivate(irg);
 
+	if (remove_Bads(irg)) {
+		edges_deactivate(irg);
+	}
+
+	/* Finally kill BAD and doublets from the keep alives.
+	   Doing this AFTER edges where deactivated saves cycles */
+	end = get_irg_end(irg);
+	remove_End_Bads_and_doublets(end);
+
+	clear_irg_state(irg, IR_GRAPH_STATE_BAD_BLOCK);
+
 	current_ir_graph = rem;
 	return changed;
 }
+
+/* Creates an ir_graph pass for optimize_graph_df. */
+ir_graph_pass_t *optimize_graph_df_pass(const char *name)
+{
+	return def_graph_pass_ret(name ? name : "optimize_graph_df", optimize_graph_df);
+}  /* optimize_graph_df_pass */