X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fia32%2Fia32_address_mode.c;h=8acfdbf5a63cdc369ccbc1d02c8acc8c4c569775;hb=bc7b5ee69d084e629590a6977b79a2fab7cd1aa1;hp=4ab4eaeec7f1777c6bbc343944a7b621d94de470;hpb=6acbb624ba344add108d976e579a26b1d5042f67;p=libfirm

diff --git a/ir/be/ia32/ia32_address_mode.c b/ir/be/ia32/ia32_address_mode.c
index 4ab4eaeec..8acfdbf5a 100644
--- a/ir/be/ia32/ia32_address_mode.c
+++ b/ir/be/ia32/ia32_address_mode.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.
  *
@@ -24,9 +24,7 @@
  * @author      Matthias Braun
  * @version     $Id$
  */
-#ifdef HAVE_CONFIG_H
 #include "config.h"
-#endif
 
 #include "ia32_address_mode.h"
 #include "ia32_transform.h"
@@ -45,7 +43,6 @@
 /* gas/ld don't support negative symconsts :-( */
 #undef SUPPORT_NEGATIVE_SYMCONSTS
 
-static be_lv_t  *lv;
 static bitset_t *non_address_mode_nodes;
 
 /**
@@ -67,8 +64,9 @@ static int do_is_immediate(const ir_node *node, int *symconsts, int negate)
 		/* Consts are typically immediates */
 		if (!tarval_is_long(get_Const_tarval(node))) {
 #ifdef DEBUG_libfirm
-			ir_fprintf(stderr, "Optimisation warning tarval of %+F(%+F) is not "
-			           "a long.\n", node, current_ir_graph);
+			ir_fprintf(stderr,
+			           "Optimisation warning tarval of %+F(%+F) is not a long.\n",
+			           node, current_ir_graph);
 #endif
 			return 0;
 		}
@@ -77,28 +75,29 @@ static int do_is_immediate(const ir_node *node, int *symconsts, int negate)
 		/* the first SymConst of a DAG can be fold into an immediate */
 #ifndef SUPPORT_NEGATIVE_SYMCONSTS
 		/* unfortunately the assembler/linker doesn't support -symconst */
-		if(negate)
+		if (negate)
 			return 0;
 #endif
 
-		if(get_SymConst_kind(node) != symconst_addr_ent)
+		if (get_SymConst_kind(node) != symconst_addr_ent)
 			return 0;
 		(*symconsts)++;
-		if(*symconsts > 1)
+		if (*symconsts > 1)
 			return 0;
 
 		return 1;
 	case iro_Add:
 	case iro_Sub:
-		/* Add's and Sub's are typically supported as long as both operands are immediates */
-		if(bitset_is_set(non_address_mode_nodes, get_irn_idx(node)))
+		/* Add's and Sub's are typically supported as long as both operands are
+		 * immediates */
+		if (ia32_is_non_address_mode_node(node))
 			return 0;
 
 		left  = get_binop_left(node);
 		right = get_binop_right(node);
-		if(!do_is_immediate(left, symconsts, negate))
+		if (!do_is_immediate(left, symconsts, negate))
 			return 0;
-		if(!do_is_immediate(right, symconsts, is_Sub(node) ? !negate : negate))
+		if (!do_is_immediate(right, symconsts, is_Sub(node) ? !negate : negate))
 			return 0;
 
 		return 1;
@@ -164,8 +163,7 @@ static void eat_immediate(ia32_address_t *addr, ir_node *node, int negate)
 	case iro_SymConst:
 		/* place the entity into the symconst */
 		if (addr->symconst_ent != NULL) {
-			panic("Internal error: more than 1 symconst in address "
-			      "calculation");
+			panic("Internal error: more than 1 symconst in address calculation");
 		}
 		addr->symconst_ent  = get_SymConst_entity(node);
 #ifndef SUPPORT_NEGATIVE_SYMCONSTS
@@ -174,14 +172,14 @@ static void eat_immediate(ia32_address_t *addr, ir_node *node, int negate)
 		addr->symconst_sign = negate;
 		break;
 	case iro_Add:
-		assert(!bitset_is_set(non_address_mode_nodes, get_irn_idx(node)));
+		assert(!ia32_is_non_address_mode_node(node));
 		left  = get_Add_left(node);
 		right = get_Add_right(node);
 		eat_immediate(addr, left, negate);
 		eat_immediate(addr, right, negate);
 		break;
 	case iro_Sub:
-		assert(!bitset_is_set(non_address_mode_nodes, get_irn_idx(node)));
+		assert(!ia32_is_non_address_mode_node(node));
 		left  = get_Sub_left(node);
 		right = get_Sub_right(node);
 		eat_immediate(addr, left, negate);
@@ -197,32 +195,35 @@ static void eat_immediate(ia32_address_t *addr, ir_node *node, int negate)
  *
  * @param addr    the address mode data so far
  * @param node    the node
- * @param force   if set, ignore the marking of node as a non-address-mode node
+ * @param flags   the flags
  *
  * @return the folded node
  */
-static ir_node *eat_immediates(ia32_address_t *addr, ir_node *node, int force)
+static ir_node *eat_immediates(ia32_address_t *addr, ir_node *node,
+                               ia32_create_am_flags_t flags)
 {
-	if(!force && bitset_is_set(non_address_mode_nodes, get_irn_idx(node)))
+	if (!(flags & ia32_create_am_force)     &&
+			ia32_is_non_address_mode_node(node) &&
+			(!(flags & ia32_create_am_double_use) || get_irn_n_edges(node) > 2))
 		return node;
 
-	if(is_Add(node)) {
+	if (is_Add(node)) {
 		ir_node *left  = get_Add_left(node);
 		ir_node *right = get_Add_right(node);
 
-		if(is_immediate(addr, left, 0)) {
+		if (is_immediate(addr, left, 0)) {
 			eat_immediate(addr, left, 0);
 			return eat_immediates(addr, right, 0);
 		}
-		if(is_immediate(addr, right, 0)) {
+		if (is_immediate(addr, right, 0)) {
 			eat_immediate(addr, right, 0);
 			return eat_immediates(addr, left, 0);
 		}
-	} else if(is_Sub(node)) {
+	} else if (is_Sub(node)) {
 		ir_node *left  = get_Sub_left(node);
 		ir_node *right = get_Sub_right(node);
 
-		if(is_immediate(addr, right, 1)) {
+		if (is_immediate(addr, right, 1)) {
 			eat_immediate(addr, right, 1);
 			return eat_immediates(addr, left, 0);
 		}
@@ -244,34 +245,33 @@ static int eat_shl(ia32_address_t *addr, ir_node *node)
 	ir_node *shifted_val;
 	long     val;
 
-	if(is_Shl(node)) {
+	if (is_Shl(node)) {
 		ir_node *right = get_Shl_right(node);
 		tarval  *tv;
 
 		/* we can use shl with 1, 2 or 3 shift */
-		if(!is_Const(right))
+		if (!is_Const(right))
 			return 0;
 		tv = get_Const_tarval(right);
-		if(!tarval_is_long(tv))
+		if (!tarval_is_long(tv))
 			return 0;
 
 		val = get_tarval_long(tv);
-		if(val < 0 || val > 3)
+		if (val < 0 || val > 3)
 			return 0;
-		if(val == 0) {
-			ir_fprintf(stderr, "Optimisation warning: unoptimized Shl(,0) "
-			           "found\n");
+		if (val == 0) {
+			ir_fprintf(stderr, "Optimisation warning: unoptimized Shl(,0) found\n");
 		}
 
 		shifted_val = get_Shl_left(node);
-	} else if(is_Add(node)) {
+	} else if (is_Add(node)) {
 		/* might be an add x, x */
 		ir_node *left  = get_Add_left(node);
 		ir_node *right = get_Add_right(node);
 
-		if(left != right)
+		if (left != right)
 			return 0;
-		if(is_Const(left))
+		if (is_Const(left))
 			return 0;
 
 		val         = 1;
@@ -281,13 +281,13 @@ static int eat_shl(ia32_address_t *addr, ir_node *node)
 	}
 
 	/* we can only eat a shl if we don't have a scale or index set yet */
-	if(addr->scale != 0 || addr->index != NULL)
+	if (addr->scale != 0 || addr->index != NULL)
 		return 0;
-	if(bitset_is_set(non_address_mode_nodes, get_irn_idx(node)))
+	if (ia32_is_non_address_mode_node(node))
 		return 0;
 
 #ifndef AGGRESSIVE_AM
-	if(get_irn_n_edges(node) > 1)
+	if (get_irn_n_edges(node) > 1)
 		return 0;
 #endif
 
@@ -297,91 +297,95 @@ static int eat_shl(ia32_address_t *addr, ir_node *node)
 }
 
 /* Create an address mode for a given node. */
-void ia32_create_address_mode(ia32_address_t *addr, ir_node *node, int force)
+void ia32_create_address_mode(ia32_address_t *addr, ir_node *node, ia32_create_am_flags_t flags)
 {
 	int      res = 0;
 	ir_node *eat_imms;
 
-	if(is_immediate(addr, node, 0)) {
+	if (is_immediate(addr, node, 0)) {
 		eat_immediate(addr, node, 0);
 		return;
 	}
 
 #ifndef AGGRESSIVE_AM
-	if(!force && get_irn_n_edges(node) > 1) {
+	if (!(flags & ia32_create_am_force) && get_irn_n_edges(node) > 1) {
 		addr->base = node;
 		return;
 	}
 #endif
 
-	if(!force && bitset_is_set(non_address_mode_nodes, get_irn_idx(node))) {
+	if (!(flags & ia32_create_am_force)     &&
+			ia32_is_non_address_mode_node(node) &&
+			(!(flags & ia32_create_am_double_use) || get_irn_n_edges(node) > 2)) {
 		addr->base = node;
 		return;
 	}
 
-	eat_imms = eat_immediates(addr, node, force);
-	if(eat_imms != node) {
-		if(force) {
+	eat_imms = eat_immediates(addr, node, flags);
+	if (eat_imms != node) {
+		if (flags & ia32_create_am_force) {
 			eat_imms = ia32_skip_downconv(eat_imms);
 		}
 
 		res  = 1;
 		node = eat_imms;
 #ifndef AGGRESSIVE_AM
-		if(get_irn_n_edges(node) > 1) {
+		if (get_irn_n_edges(node) > 1) {
 			addr->base = node;
 			return;
 		}
 #endif
-		if(bitset_is_set(non_address_mode_nodes, get_irn_idx(node))) {
+		if (ia32_is_non_address_mode_node(node)) {
 			addr->base = node;
 			return;
 		}
 	}
 
 	/* starting point Add, Sub or Shl, FrameAddr */
-	if(is_Shl(node)) {
+	if (is_Shl(node)) {
 		/* We don't want to eat add x, x as shl here, so only test for real Shl
 		 * instructions, because we want the former as Lea x, x, not Shl x, 1 */
-		if(eat_shl(addr, node))
+		if (eat_shl(addr, node))
 			return;
-	} else if(is_immediate(addr, node, 0)) {
+	} else if (is_immediate(addr, node, 0)) {
 		eat_immediate(addr, node, 0);
 		return;
-	} else if(be_is_FrameAddr(node)) {
+	} else if (be_is_FrameAddr(node)) {
 		assert(addr->base == NULL);
 		assert(addr->frame_entity == NULL);
 		addr->base         = be_get_FrameAddr_frame(node);
 		addr->use_frame    = 1;
 		addr->frame_entity = be_get_FrameAddr_entity(node);
 		return;
-	} else if(is_Add(node)) {
+	} else if (is_Add(node)) {
 		ir_node *left  = get_Add_left(node);
 		ir_node *right = get_Add_right(node);
 
-		if(force) {
+		if (flags & ia32_create_am_force) {
 			left  = ia32_skip_downconv(left);
 			right = ia32_skip_downconv(right);
 		}
 
-		assert(force || !is_immediate(addr, left, 0));
-		assert(force || !is_immediate(addr, right, 0));
+		assert(flags & ia32_create_am_force || !is_immediate(addr, left,  0));
+		assert(flags & ia32_create_am_force || !is_immediate(addr, right, 0));
 
-		if(eat_shl(addr, left)) {
+		if (eat_shl(addr, left)) {
 			left = NULL;
-		} else if(eat_shl(addr, right)) {
+		} else if (eat_shl(addr, right)) {
 			right = NULL;
 		}
-		if(left != NULL && be_is_FrameAddr(left)
-				&& !bitset_is_set(non_address_mode_nodes, get_irn_idx(left))) {
+		if (left != NULL          &&
+				be_is_FrameAddr(left) &&
+				!ia32_is_non_address_mode_node(left)) {
 			assert(addr->base == NULL);
 			assert(addr->frame_entity == NULL);
 			addr->base         = be_get_FrameAddr_frame(left);
 			addr->use_frame    = 1;
 			addr->frame_entity = be_get_FrameAddr_entity(left);
 			left               = NULL;
-		} else if(right != NULL && be_is_FrameAddr(right)
-				&& !bitset_is_set(non_address_mode_nodes, get_irn_idx(right))) {
+		} else if (right != NULL   &&
+				be_is_FrameAddr(right) &&
+				!ia32_is_non_address_mode_node(right)) {
 			assert(addr->base == NULL);
 			assert(addr->frame_entity == NULL);
 			addr->base         = be_get_FrameAddr_frame(right);
@@ -390,8 +394,8 @@ void ia32_create_address_mode(ia32_address_t *addr, ir_node *node, int force)
 			right              = NULL;
 		}
 
-		if(left != NULL) {
-			if(addr->base != NULL) {
+		if (left != NULL) {
+			if (addr->base != NULL) {
 				assert(addr->index == NULL && addr->scale == 0);
 				assert(right == NULL);
 				addr->index = left;
@@ -399,8 +403,8 @@ void ia32_create_address_mode(ia32_address_t *addr, ir_node *node, int force)
 				addr->base = left;
 			}
 		}
-		if(right != NULL) {
-			if(addr->base == NULL) {
+		if (right != NULL) {
+			if (addr->base == NULL) {
 				addr->base = right;
 			} else {
 				assert(addr->index == NULL && addr->scale == 0);
@@ -418,48 +422,66 @@ void ia32_mark_non_am(ir_node *node)
 	bitset_set(non_address_mode_nodes, get_irn_idx(node));
 }
 
+int ia32_is_non_address_mode_node(ir_node const *node)
+{
+	return bitset_is_set(non_address_mode_nodes, get_irn_idx(node));
+}
+
+static int value_last_used_here(be_lv_t *lv, ir_node *here, ir_node *value)
+{
+	ir_node         *block = get_nodes_block(here);
+	const ir_edge_t *edge;
+
+	/* If the value is live end it is for sure it does not die here */
+	if (be_is_live_end(lv, block, value)) return 0;
+
+	/* if multiple nodes in this block use the value, then we cannot decide
+	 * whether the value will die here (because there is no schedule yet).
+	 * Assume it does not die in this case. */
+	foreach_out_edge(value, edge) {
+		ir_node *user = get_edge_src_irn(edge);
+		if (user != here && get_nodes_block(user) == block) {
+			return 0;
+		}
+	}
+
+	return 1;
+}
+
 /**
  * Walker: mark those nodes that cannot be part of an address mode because
- * there value must be access through an register
+ * their value must be accessed through a register
  */
 static void mark_non_address_nodes(ir_node *node, void *env)
 {
-	int i, arity;
-	ir_node *ptr;
-	ir_node *mem;
+	be_lv_t *lv = env;
+	int      arity;
+	int      i;
 	ir_node *val;
-	ir_node *block;
 	ir_node *left;
 	ir_node *right;
 	ir_mode *mode;
-	const ir_edge_t *edge;
-	(void) env;
 
 	mode = get_irn_mode(node);
-	if(!mode_is_int(mode) && !mode_is_reference(mode) && mode != mode_b)
+	if (!mode_is_int(mode) && !mode_is_reference(mode) && mode != mode_b)
 		return;
 
-	switch(get_irn_opcode(node)) {
+	switch (get_irn_opcode(node)) {
 	case iro_Load:
-		ptr = get_Load_ptr(node);
-		mem = get_Load_mem(node);
-
-		bitset_set(non_address_mode_nodes, get_irn_idx(mem));
+		/* Nothing to do. especially do not mark the pointer, because we want to
+		 * turn it into AM. */
 		break;
 
 	case iro_Store:
+		/* Do not mark the pointer, because we want to turn it into AM. */
 		val = get_Store_value(node);
-		ptr = get_Store_ptr(node);
-		mem = get_Store_mem(node);
-
-		bitset_set(non_address_mode_nodes, get_irn_idx(val));
-		bitset_set(non_address_mode_nodes, get_irn_idx(mem));
+		ia32_mark_non_am(val);
 		break;
 
 	case iro_Shl:
 	case iro_Add:
 		/* only 1 user: AM folding is always beneficial */
-		if(get_irn_n_edges(node) <= 1)
+		if (get_irn_n_edges(node) <= 1)
 			break;
 
 		/* for adds and shls with multiple users we use this heuristic:
@@ -468,40 +490,30 @@ static void mark_non_address_nodes(ir_node *node, void *env)
 		 * pressure. Otherwise we fold them in aggressively in the hope, that
 		 * the node itself doesn't exist anymore and we were able to save the
 		 * register for the result */
-		block = get_nodes_block(node);
 		left  = get_binop_left(node);
 		right = get_binop_right(node);
 
-		/* live end: we won't save a register by AM folding */
-		if(be_is_live_end(lv, block, left) || be_is_live_end(lv, block, right))
+		/* Fold AM if any of the two operands does not die here.  This duplicates
+		 * an addition and has the same register pressure for the case that only
+		 * one operand dies, but is faster (on Pentium 4).
+		 * && instead of || only folds AM if both operands do not die here */
+		if (!value_last_used_here(lv, node, left) ||
+				!value_last_used_here(lv, node, right)) {
 			return;
-
-		/* if multiple nodes in this block use left/right values, then we
-		 * can't really decide wether the values will die after node.
-		 * We use aggressive mode then, since it's usually just multiple address
-		 * calculations. */
-		foreach_out_edge(left, edge) {
-			ir_node *user = get_edge_src_irn(edge);
-			if(user != node && get_nodes_block(user) == block)
-				return;
-		}
-		foreach_out_edge(right, edge) {
-			ir_node *user = get_edge_src_irn(edge);
-			if(user != node && get_nodes_block(user) == block)
-				return;
 		}
 
-		/* noone-else in this block is using left/right so we'll reduce register
-		 * pressure if we don't fold the node */
-		bitset_set(non_address_mode_nodes, get_irn_idx(node));
+		/* At least one of left and right are not used by anyone else, so it is
+		 * beneficial for the register pressure (if both are unused otherwise,
+		 * else neutral) and ALU use to not fold AM. */
+		ia32_mark_non_am(node);
 		break;
 
 	default:
 		arity = get_irn_arity(node);
 
-		for(i = 0; i < arity; ++i) {
+		for (i = 0; i < arity; ++i) {
 			ir_node *in = get_irn_n(node, i);
-			bitset_set(non_address_mode_nodes, get_irn_idx(in));
+			ia32_mark_non_am(in);
 		}
 		break;
 	}
@@ -510,11 +522,11 @@ static void mark_non_address_nodes(ir_node *node, void *env)
 void ia32_calculate_non_address_mode_nodes(be_irg_t *birg)
 {
 	ir_graph *irg = be_get_birg_irg(birg);
+	be_lv_t  *lv  = be_assure_liveness(birg);
 
-	lv                     = be_assure_liveness(birg);
 	non_address_mode_nodes = bitset_malloc(get_irg_last_idx(irg));
 
-	irg_walk_graph(irg, NULL, mark_non_address_nodes, NULL);
+	irg_walk_graph(irg, NULL, mark_non_address_nodes, lv);
 }
 
 void ia32_free_non_address_mode_nodes(void)