X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Fconvopt.c;h=43ac2d90aa2ae1ed236eb84ad6d10965b6f841e0;hb=a619ce99e40de4eb4481a590970a881e9f24627a;hp=f7f8ad036773a4c989e45259f35b406de9f2de7d;hpb=18886c01c78fb5fa3eaf1586fd4bbaa7b202460b;p=libfirm

diff --git a/ir/opt/convopt.c b/ir/opt/convopt.c
index f7f8ad036..43ac2d90a 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) 1995-2008 University of Karlsruhe.  All right reserved.
  *
  * This file is part of libFirm.
  *
@@ -21,7 +21,7 @@
  * @file
  * @brief   conv node optimisation
  * @author  Matthias Braun, Christoph Mallon
- * @version $Id: condeval.c 13543 2007-04-29 19:29:02Z beck $
+ * @version $Id$
  *
  * Try to minimize the number of conv nodes by changing modes of operations.
  * The typical example is the following structure:
@@ -32,168 +32,290 @@
  *          Add Is            gets transformed to           |
  *            |
  *         Conv Hs
+ *
+ * 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 "iredges_t.h"
 #include "irgwalk.h"
 #include "irprintf.h"
+#include "irpass_t.h"
 
 DEBUG_ONLY(static firm_dbg_module_t *dbg);
 
-static
-int is_optimizable_node(const ir_node *node)
+static inline int imin(int a, int b) { return a < b ? a : b; }
+
+static bool is_optimizable_node(const ir_node *node)
 {
-	if(is_Const(node)) {
-		ir_mode *mode = get_irn_mode(node);
-		/* tarval module is incomplete and can't convert floats to ints */
-		if(!mode_is_int(mode))
-			return 0;
-		return 1;
+	switch (get_irn_opcode(node)) {
+	case iro_Add:
+	case iro_And:
+	case iro_Eor:
+	case iro_Minus:
+	case iro_Mul:
+	case iro_Not:
+	case iro_Or:
+	case iro_Phi:
+	case iro_Shl:
+	case iro_Sub:
+		return true;
+	default:
+		return false;
 	}
-	return is_Add(node) || is_Sub(node) || is_Mul(node);
 }
 
-static
-int get_conv_costs(const ir_node *node, ir_mode *dest_mode)
+static tarval* conv_const_tv(const ir_node* cnst, ir_mode* dest_mode)
+{
+	return tarval_convert_to(get_Const_tarval(cnst), dest_mode);
+}
+
+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);
+}
+
+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 costs;
 
-	if(mode == dest_mode)
+	if (mode == dest_mode)
 		return 0;
 
-	/* TODO... */
-	if(!is_Const(node) && get_irn_n_edges(node) > 1) {
+	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(is_Conv(node)) {
-		return get_conv_costs(get_Conv_op(node), dest_mode) - 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)) {
+		size_t i;
+		size_t 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_optimizable_node(node)) {
-		int i;
-		int arity = get_irn_arity(node);
-		int costs = 0;
+	if (!is_downconv(mode, dest_mode)) {
+		return 1;
+	}
 
-		for(i = 0; i < arity; ++i) {
-			ir_node *pred = get_irn_n(node, i);
-			costs += get_conv_costs(pred, dest_mode);
+	if (is_Conv(node)) {
+		ir_node *pred      = get_Conv_op(node);
+		ir_mode *pred_mode = get_irn_mode(pred);
+
+		if (!values_in_mode(dest_mode, pred_mode)) {
+			return 1;
 		}
+		return get_conv_costs(get_Conv_op(node), dest_mode) - 1;
+	}
 
-		return costs;
+	if (!is_optimizable_node(node)) {
+		return 1;
+	}
+
+	costs = 0;
+	// 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 += imin(get_conv_costs(pred, dest_mode), 1);
 	}
 
-	return 1;
+	return costs;
 }
 
-static
-ir_node *conv_transform(ir_node *node, ir_mode *dest_mode)
+static ir_node *place_conv(ir_node *node, ir_mode *dest_mode)
 {
-	size_t arity;
-	size_t i;
+	ir_node *block = get_nodes_block(node);
+	ir_node *conv = new_r_Conv(block, node, dest_mode);
+	return conv;
+}
 
-	if (get_irn_mode(node) == dest_mode)
+static ir_node *conv_transform(ir_node *node, ir_mode *dest_mode)
+{
+	ir_mode *mode = get_irn_mode(node);
+	size_t   arity;
+	size_t   i;
+
+	if (mode == dest_mode)
 		return node;
 
 	if (is_Const(node)) {
-		tarval *tv = tarval_convert_to(get_Const_tarval(node), dest_mode);
-		assert(get_tarval_mode(tv) == dest_mode);
-		return new_Const(dest_mode, tv);
+		/* TODO tarval module is incomplete and can't convert floats to ints */
+		tarval *tv = conv_const_tv(node, dest_mode);
+		if (tv == tarval_bad) {
+			return place_conv(node, dest_mode);
+		} else {
+			return new_Const(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 (!is_optimizable_node(node) || get_irn_n_edges(node) > 1) {
-		ir_node *block = get_nodes_block(node);
-		ir_node *conv = new_r_Conv(current_ir_graph, block, node, dest_mode);
-		return conv;
+	if (get_irn_n_edges(node) > 1) {
+		return place_conv(node, dest_mode);
 	}
 
-	if (is_Conv(node))
+	if (!is_downconv(mode, dest_mode)) {
+		return place_conv(node, dest_mode);
+	}
+
+	if (is_Conv(node)) {
+		ir_node *pred      = get_Conv_op(node);
+		ir_mode *pred_mode = get_irn_mode(pred);
+
+		if (!values_in_mode(dest_mode, pred_mode)) {
+			return place_conv(node, dest_mode);
+		}
 		return conv_transform(get_Conv_op(node), dest_mode);
+	}
+
+	if (!is_optimizable_node(node)) {
+		return place_conv(node, dest_mode);
+	}
 
-	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);
-		ir_node *transformed = conv_transform(pred, dest_mode);
+		ir_node *transformed;
+		if (get_conv_costs(pred, dest_mode) > 0) {
+			transformed = place_conv(pred, dest_mode);
+		} else {
+			transformed = conv_transform(pred, dest_mode);
+		}
 		set_irn_n(node, i, transformed);
 	}
 	set_irn_mode(node, dest_mode);
 	return node;
 }
 
-static
-int is_downconv(ir_mode *src_mode, ir_mode *dest_mode)
-{
-	if(!mode_is_int(src_mode) || !mode_is_int(dest_mode))
-		return 0;
-	if(get_mode_size_bits(dest_mode) >= get_mode_size_bits(src_mode))
-		return 0;
-
-	return 1;
-}
-
-/* TODO, backends can't handle and it's probably not more efficient on most
-   archs */
+/* TODO, backends (at least ia32) 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)
+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;
 
-	if(is_downconv
+	if (is_downconv
 }
 #endif
 
-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;
+	bool *changed = data;
 
 #if 0
-	if(is_Cmp(node)) {
+	if (is_Cmp(node)) {
 		try_optimize_cmp(node);
 		return;
 	}
 #endif
 
-	if(!is_Conv(node))
+	if (!is_Conv(node))
 		return;
 
 	pred      = get_Conv_op(node);
 	mode      = get_irn_mode(node);
 	pred_mode = get_irn_mode(pred);
 
-	if(!is_downconv(pred_mode, mode))
+	if (mode_is_reference(mode) || mode_is_reference(pred_mode))
+		return;
+
+	if (!is_Phi(pred) && !is_downconv(pred_mode, mode))
 		return;
 
-	costs = get_conv_costs(pred, mode);
+	/* - 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);
+	if (node != transformed) {
+		exchange(node, transformed);
+		*changed = true;
+	}
 }
 
-void conv_opt(ir_graph *irg)
+int conv_opt(ir_graph *irg)
 {
+	bool changed;
+	bool invalidate = false;
 	FIRM_DBG_REGISTER(dbg, "firm.opt.conv");
 
 	DB((dbg, LEVEL_1, "===> Performing conversion optimization on %+F\n", irg));
 
 	edges_assure(irg);
-	irg_walk_graph(irg, conv_opt_walker, NULL, NULL);
+	do {
+		changed = false;
+		irg_walk_graph(irg, NULL, conv_opt_walker, &changed);
+		local_optimize_graph(irg);
+		invalidate |= changed;
+	} while (changed);
+
+	if (invalidate) {
+		set_irg_outs_inconsistent(irg);
+	}
+	return invalidate;
+}
+
+/* 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_ret(name ? name : "conv_opt", conv_opt);
+
+	/* safe to run parallel on all irgs */
+	ir_graph_pass_set_parallel(path, 1);
+
+	return path;
 }