X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Fconvopt.c;h=55ae04a56c3dd8b8c35ddcd3efe13059bb399b15;hb=48feaf0618330fb060d22d5ef35c94dd7e9556a7;hp=11631c42bee170e653611e7de97e4722666ee27f;hpb=4b3b89d08873a30a8b08fe8507d1796257495e42;p=libfirm

diff --git a/ir/opt/convopt.c b/ir/opt/convopt.c
index 11631c42b..55ae04a56 100644
--- a/ir/opt/convopt.c
+++ b/ir/opt/convopt.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 1995-2007 University of Karlsruhe.  All right reserved.
+ * Copyright (C) 1995-2011 University of Karlsruhe.  All right reserved.
  *
  * This file is part of libFirm.
  *
@@ -21,7 +21,6 @@
  * @file
  * @brief   conv node optimisation
  * @author  Matthias Braun, Christoph Mallon
- * @version $Id$
  *
  * Try to minimize the number of conv nodes by changing modes of operations.
  * The typical example is the following structure:
@@ -36,73 +35,147 @@
  * TODO: * try to optimize cmp modes
  *       * decide when it is useful to move the convs through phis
  */
-#ifdef HAVE_CONFIG_H
 #include "config.h"
-#endif
 
 #include "iroptimize.h"
 
 #include <assert.h>
+#include <stdbool.h>
 #include "debug.h"
 #include "ircons.h"
 #include "irgmod.h"
 #include "irgopt.h"
 #include "irnode_t.h"
+#include "iropt_t.h"
 #include "iredges_t.h"
 #include "irgwalk.h"
 #include "irprintf.h"
+#include "irpass_t.h"
+#include "tv.h"
+#include "vrp.h"
+
+DEBUG_ONLY(static firm_dbg_module_t *dbg;)
 
-DEBUG_ONLY(static firm_dbg_module_t *dbg);
+static inline int imin(int a, int b) { return a < b ? a : b; }
 
-static
-int is_optimizable_node(const ir_node *node)
+static bool is_optimizable_node(const ir_node *node, ir_mode *dest_mode)
 {
-	return
-		is_Add(node) ||
-		is_Sub(node) ||
-		is_Mul(node) ||
-		is_Phi(node);
+	switch (get_irn_opcode(node)) {
+	case iro_Minus:
+	case iro_Phi:
+	case iro_And:
+	case iro_Eor:
+	case iro_Or:
+	case iro_Not:
+		return true;
+	case iro_Add:
+	case iro_Mul:
+	case iro_Sub:
+		if (mode_is_float(get_irn_mode(node)))
+			return false;
+		return true;
+	case iro_Shl: {
+		int modulo_shift = get_mode_modulo_shift(dest_mode);
+		int old_shift    = get_mode_modulo_shift(get_irn_mode(node));
+		/* bail out if modulo shift changes */
+		if (modulo_shift != old_shift)
+			return false;
+		return true;
+	}
+
+	default:
+		return false;
+	}
 }
 
-static tarval* conv_const_tv(const ir_node* cnst, ir_mode* dest_mode)
+static ir_tarval* conv_const_tv(const ir_node* cnst, ir_mode* dest_mode)
 {
 	return tarval_convert_to(get_Const_tarval(cnst), dest_mode);
 }
 
-static
-int get_conv_costs(const ir_node *node, ir_mode *dest_mode)
+static bool is_downconv(ir_mode *src_mode, ir_mode *dest_mode)
+{
+	return ((mode_is_int(src_mode) && mode_is_int(dest_mode))
+		|| (mode_is_float(src_mode) && mode_is_float(dest_mode)))
+		&& get_mode_size_bits(dest_mode) <= get_mode_size_bits(src_mode);
+}
+
+static int get_conv_costs(const ir_node *node, ir_mode *dest_mode)
 {
 	ir_mode *mode = get_irn_mode(node);
-	size_t arity;
-	size_t i;
+	int arity;
+	int i;
 	int costs;
 
 	if (mode == dest_mode)
 		return 0;
 
 	if (is_Const(node)) {
-		/* TODO tarval module is incomplete and can't convert floats to ints */
 		return conv_const_tv(node, dest_mode) == tarval_bad ? 1 : 0;
 	}
 
+	if (is_Conv(node) &&
+			is_downconv(mode, dest_mode) &&
+			get_irn_mode(get_Conv_op(node)) == dest_mode) {
+		return -1;
+	}
+
 	if (get_irn_n_edges(node) > 1) {
 		DB((dbg, LEVEL_3, "multi outs at %+F\n", node));
 		return 1;
 	}
 
+	if (ir_zero_when_converted(node, dest_mode)) {
+		return -1;
+	}
+
+#if 0 // TODO
+	/* Take the minimum of the conversion costs for Phi predecessors as only one
+	 * branch is actually executed at a time */
+	if (is_Phi(node)) {
+		int i;
+		int arity = get_Phi_n_preds(node);
+		int costs;
+
+		costs = get_conv_costs(get_Phi_pred(node, 0), dest_mode);
+		for (i = 1; i < arity; ++i) {
+			ir_node *pred = get_Phi_pred(node, i);
+			int c = get_conv_costs(pred, dest_mode);
+			if (c < costs) costs = c;
+		}
+
+		return costs;
+	}
+#endif
+
+	if (!is_downconv(mode, dest_mode)) {
+		return 1;
+	}
+
 	if (is_Conv(node)) {
-		return get_conv_costs(get_Conv_op(node), dest_mode) - 1;
+		ir_node *pred      = get_Conv_op(node);
+		ir_mode *pred_mode = get_irn_mode(pred);
+
+		if (smaller_mode(pred_mode, dest_mode)) {
+			return get_conv_costs(get_Conv_op(node), dest_mode) - 1;
+		}
+		if (may_leave_out_middle_conv(pred_mode, mode, dest_mode)) {
+			return 0;
+		} else {
+			return 1;
+		}
 	}
 
-	if (!is_optimizable_node(node)) {
+	if (!is_optimizable_node(node, dest_mode)) {
 		return 1;
 	}
 
 	costs = 0;
-	arity = get_irn_arity(node);
+	// The shift count does not participate in the conv optimisation
+	arity = is_Shl(node) ? 1 : get_irn_arity(node);
 	for (i = 0; i < arity; ++i) {
 		ir_node *pred = get_irn_n(node, i);
-		costs += get_conv_costs(pred, dest_mode);
+		costs += imin(get_conv_costs(pred, dest_mode), 1);
 	}
 
 	return costs;
@@ -111,92 +184,101 @@ int get_conv_costs(const ir_node *node, ir_mode *dest_mode)
 static ir_node *place_conv(ir_node *node, ir_mode *dest_mode)
 {
 	ir_node *block = get_nodes_block(node);
-	ir_node *conv = new_r_Conv(current_ir_graph, block, node, dest_mode);
+	ir_node *conv = new_r_Conv(block, node, dest_mode);
 	return conv;
 }
 
-static
-ir_node *conv_transform(ir_node *node, ir_mode *dest_mode)
+static ir_node *conv_transform(ir_node *node, ir_mode *dest_mode)
 {
-	size_t arity;
-	size_t i;
+	ir_mode  *mode = get_irn_mode(node);
+	ir_graph *irg  = get_irn_irg(node);
+	int       arity;
+	int       conv_arity;
+	int       i;
+	ir_node  *new_node;
+	ir_node **ins;
 
-	if (get_irn_mode(node) == dest_mode)
+	if (mode == dest_mode)
 		return node;
 
 	if (is_Const(node)) {
-		/* TODO tarval module is incomplete and can't convert floats to ints */
-		tarval *tv = conv_const_tv(node, dest_mode);
+		ir_tarval *tv = conv_const_tv(node, dest_mode);
 		if (tv == tarval_bad) {
 			return place_conv(node, dest_mode);
 		} else {
-			return new_Const(dest_mode, tv);
+			return new_r_Const(irg, tv);
 		}
 	}
 
+	if (is_Conv(node) &&
+			is_downconv(mode, dest_mode) &&
+			get_irn_mode(get_Conv_op(node)) == dest_mode) {
+		return get_Conv_op(node);
+	}
+
 	if (get_irn_n_edges(node) > 1) {
 		return place_conv(node, dest_mode);
 	}
 
+	if (!is_downconv(mode, dest_mode)) {
+		return place_conv(node, dest_mode);
+	}
+
 	if (is_Conv(node)) {
-		return conv_transform(get_Conv_op(node), dest_mode);
+		ir_node *pred      = get_Conv_op(node);
+		ir_mode *pred_mode = get_irn_mode(pred);
+
+		if (smaller_mode(pred_mode, dest_mode)) {
+			return conv_transform(get_Conv_op(node), dest_mode);
+		}
+		return place_conv(node, dest_mode);
 	}
 
-	if (!is_optimizable_node(node)) {
+	if (!is_optimizable_node(node, dest_mode)) {
 		return place_conv(node, dest_mode);
 	}
 
+	// We want to create a new node with the right mode
 	arity = get_irn_arity(node);
-	for (i = 0; i < arity; i++) {
+	ins = ALLOCAN(ir_node *, arity);
+
+	// The shift count does not participate in the conv optimisation
+	conv_arity = is_Shl(node) ? 1 : arity;
+	for (i = 0; i < conv_arity; i++) {
 		ir_node *pred = get_irn_n(node, i);
-		ir_node *transformed = conv_transform(pred, dest_mode);
-		set_irn_n(node, i, transformed);
+		ir_node *transformed;
+		if (get_conv_costs(pred, dest_mode) > 0) {
+			transformed = place_conv(pred, dest_mode);
+		} else {
+			transformed = conv_transform(pred, dest_mode);
+		}
+		ins[i] = transformed;
 	}
-	set_irn_mode(node, dest_mode);
-	return node;
-}
 
-static
-int is_downconv(ir_mode *src_mode, ir_mode *dest_mode)
-{
-	return
-		mode_is_int(src_mode) &&
-		mode_is_int(dest_mode) &&
-		get_mode_size_bits(dest_mode) < get_mode_size_bits(src_mode);
-}
+	for (i = conv_arity; i < arity; i++) {
+		ins[i] = get_irn_n(node, i);
+	}
 
-/* TODO, backends (at least ia23) can't handle it at the moment,
-   and it's probably not more efficient on most
-   archs */
-#if 0
-static
-void try_optimize_cmp(ir_node *node)
-{
-	ir_node *left  = get_Cmp_left(node);
-	ir_node *right = get_Cmp_right(node);
-	ir_node *conv  = NULL;
+	new_node = new_ir_node(get_irn_dbg_info(node),
+				irg,
+				get_nodes_block(node),
+				get_irn_op(node),
+				dest_mode,
+				arity,
+				ins);
+	copy_node_attr(irg, node, new_node);
 
-	if(is_downconv
+	return new_node;
 }
-#endif
-
-static char changed;
 
-static
-void conv_opt_walker(ir_node *node, void *data)
+static void conv_opt_walker(ir_node *node, void *data)
 {
 	ir_node *transformed;
 	ir_node *pred;
 	ir_mode *pred_mode;
 	ir_mode *mode;
 	int costs;
-
-#if 0
-	if(is_Cmp(node)) {
-		try_optimize_cmp(node);
-		return;
-	}
-#endif
+	bool *changed = (bool*)data;
 
 	if (!is_Conv(node))
 		return;
@@ -205,35 +287,53 @@ void conv_opt_walker(ir_node *node, void *data)
 	mode      = get_irn_mode(node);
 	pred_mode = get_irn_mode(pred);
 
+	if (mode_is_reference(mode) || mode_is_reference(pred_mode))
+		return;
+
 	if (!is_Phi(pred) && !is_downconv(pred_mode, mode))
 		return;
 
 	/* - 1 for the initial conv */
 	costs = get_conv_costs(pred, mode) - 1;
 	DB((dbg, LEVEL_2, "Costs for %+F -> %+F: %d\n", node, pred, costs));
-	if (costs >= 0) return;
+	if (costs > 0)
+		return;
 
 	transformed = conv_transform(pred, mode);
-	exchange(node, transformed);
-	changed = 1;
+	if (node != transformed) {
+		exchange(node, transformed);
+		*changed = true;
+	}
 }
 
 void conv_opt(ir_graph *irg)
 {
-	char invalidate = 0;
+	bool global_changed = false;
+	bool changed;
 	FIRM_DBG_REGISTER(dbg, "firm.opt.conv");
 
+	assure_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_OUT_EDGES);
+
 	DB((dbg, LEVEL_1, "===> Performing conversion optimization on %+F\n", irg));
 
-	edges_assure(irg);
 	do {
-		changed = 0;
-		irg_walk_graph(irg, NULL, conv_opt_walker, NULL);
+		changed = false;
+		irg_walk_graph(irg, NULL, conv_opt_walker, &changed);
 		local_optimize_graph(irg);
-		invalidate |= changed;
+		global_changed |= changed;
 	} while (changed);
 
-	if (invalidate) {
-		set_irg_outs_inconsistent(irg);
-	}
+	confirm_irg_properties(irg,
+		global_changed ? IR_GRAPH_PROPERTIES_NONE : IR_GRAPH_PROPERTIES_ALL);
+}
+
+/* Creates an ir_graph pass for conv_opt. */
+ir_graph_pass_t *conv_opt_pass(const char *name)
+{
+	ir_graph_pass_t *path = def_graph_pass(name ? name : "conv_opt", conv_opt);
+
+	/* safe to run parallel on all irgs */
+	ir_graph_pass_set_parallel(path, 1);
+
+	return path;
 }