X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firarch.c;h=c5ca7e649d2a5f1586416616f3abcf5b8b3f6919;hb=bb9f2e36362333c6635b89f5258171b06c786608;hp=55db0899d505a20fb16b403bd05862fc70643a22;hpb=969f52a3743dd60803f8fd5122aa8c714d17ae32;p=libfirm

diff --git a/ir/ir/irarch.c b/ir/ir/irarch.c
index 55db0899d..c5ca7e649 100644
--- a/ir/ir/irarch.c
+++ b/ir/ir/irarch.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.
  *
@@ -27,9 +27,7 @@
  * Implements "Strenght Reduction of Multiplications by Integer Constants" by Youfeng Wu.
  * Implements Division and Modulo by Consts from "Hackers Delight",
  */
-#ifdef HAVE_CONFIG_H
-# include "config.h"
-#endif
+#include "config.h"
 
 #ifdef HAVE_STDLIB_H
 # include <stdlib.h>
@@ -52,6 +50,7 @@
 #include "ircons.h"
 #include "irarch.h"
 #include "irflag.h"
+#include "error.h"
 
 #undef DEB
 
@@ -229,8 +228,8 @@ static unsigned char *value_to_condensed(mul_env *env, tarval *tv, int *pr) {
  * Calculate the gain when using the generalized complementary technique
  */
 static int calculate_gain(unsigned char *R, int r) {
-	int max_gain = -1;
-	int idx, i;
+	int max_gain = 0;
+	int idx = -1, i;
 	int gain;
 
 	/* the gain for r == 1 */
@@ -244,9 +243,7 @@ static int calculate_gain(unsigned char *R, int r) {
 			idx = i;
 		}
 	}
-	if (max_gain > 0)
-		return idx;
-	return -1;
+	return idx;
 }
 
 /**
@@ -326,11 +323,16 @@ static instruction *decompose_simple_cases(mul_env *env, unsigned char *R, int r
 		assert(r == 2);
 
 		ins = env->root;
+		if (R[1] <= env->max_S) {
+			ins = emit_LEA(env, ins, ins, R[1]);
+			if (R[0] != 0) {
+				ins = emit_SHIFT(env, ins, R[0]);
+			}
+			return ins;
+		}
 		if (R[0] != 0) {
 			ins = emit_SHIFT(env, ins, R[0]);
 		}
-		if (R[1] <= env->max_S)
-			return emit_LEA(env, ins, ins, R[1]);
 
 		ins2 = emit_SHIFT(env, env->root, R[0] + R[1]);
 		return emit_LEA(env, ins, ins2, 0);
@@ -464,7 +466,7 @@ static ir_node *build_graph(mul_env *env, instruction *inst) {
 	case ZERO:
 		return inst->irn = new_r_Const(current_ir_graph, env->blk, env->mode, get_mode_null(env->mode));
 	default:
-		assert(0);
+		panic("Unsupported instruction kind");
 		return NULL;
 	}
 }
@@ -504,16 +506,14 @@ static int evaluate_insn(mul_env *env, instruction *inst) {
 	case ZERO:
 		inst->costs = costs = env->evaluate(inst->kind, NULL);
 		return costs;
-	default:
-		assert(0);
-		return 0;
 	}
+	panic("Unsupported instruction kind");
 }
 
 /**
  * Evaluate the replacement instructions and build a new graph
  * if faster than the Mul.
- * returns the root of the new graph then or irn otherwise.
+ * Returns the root of the new graph then or irn otherwise.
  *
  * @param irn      the Mul operation
  * @param operand  the multiplication operand
@@ -542,7 +542,7 @@ static ir_node *do_decomposition(ir_node *irn, ir_node *operand, tarval *tv) {
 	inst = decompose_mul(&env, R, r, tv);
 
 	/* the paper suggests 70% here */
-	mul_costs = (env.evaluate(MUL, tv) * 7) / 10;
+	mul_costs = (env.evaluate(MUL, tv) * 7 + 5) / 10;
 	if (evaluate_insn(&env, inst) <= mul_costs && !env.fail) {
 		env.op       = operand;
 		env.blk      = get_nodes_block(irn);
@@ -594,7 +594,7 @@ ir_node *arch_dep_replace_mul_with_shifts(ir_node *irn) {
 			}
 		}
 	}
-	set_arch_dep_running(0);
+	//set_arch_dep_running(0);
 
 	return res;
 }
@@ -631,7 +631,7 @@ static int tv_ld2(tarval *tv, int bits) {
 #define SHL(a, b) tarval_shl(a, b)
 #define SHR(a, b) tarval_shr(a, b)
 #define ADD(a, b) tarval_add(a, b)
-#define SUB(a, b) tarval_sub(a, b)
+#define SUB(a, b) tarval_sub(a, b, NULL)
 #define MUL(a, b) tarval_mul(a, b)
 #define DIV(a, b) tarval_div(a, b)
 #define MOD(a, b) tarval_mod(a, b)
@@ -848,7 +848,7 @@ static ir_node *replace_div_by_mulh(ir_node *div, tarval *tv) {
 
 		/* generate the Mulh instruction */
 		c = new_r_Const(current_ir_graph, block, mode, mag.M);
-		q    = new_rd_Mulh(dbg, current_ir_graph, block, n, c, mode);
+		q = new_rd_Mulh(dbg, current_ir_graph, block, n, c, mode);
 
 		if (mag.need_add) {
 			if (mag.s > 0) {
@@ -883,7 +883,7 @@ ir_node *arch_dep_replace_div_by_const(ir_node *irn) {
 	if (params == NULL || (opts & arch_dep_div_by_const) == 0)
 		return irn;
 
-	if (get_irn_opcode(irn) == iro_Div) {
+	if (is_Div(irn)) {
 		ir_node *c = get_Div_right(irn);
 		ir_node *block, *left;
 		ir_mode *mode;
@@ -892,7 +892,7 @@ ir_node *arch_dep_replace_div_by_const(ir_node *irn) {
 		int n, bits;
 		int k, n_flag;
 
-		if (get_irn_op(c) != op_Const)
+		if (! is_Const(c))
 			return irn;
 
 		tv = get_Const_tarval(c);
@@ -927,15 +927,20 @@ ir_node *arch_dep_replace_div_by_const(ir_node *irn) {
 				ir_node *k_node;
 				ir_node *curr = left;
 
-				if (k != 1) {
-					k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k - 1);
-					curr   = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode);
-				}
+				/* create the correction code for signed values only if there might be a remainder */
+				if (! is_Div_remainderless(irn)) {
+					if (k != 1) {
+						k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k - 1);
+						curr   = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode);
+					}
 
-				k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, bits - k);
-				curr   = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode);
+					k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, bits - k);
+					curr   = new_rd_Shr(dbg, current_ir_graph, block, curr, k_node, mode);
 
-				curr   = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode);
+					curr   = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode);
+				} else {
+					k_node = left;
+				}
 
 				k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k);
 				res    = new_rd_Shrs(dbg, current_ir_graph, block, curr, k_node, mode);
@@ -974,7 +979,7 @@ ir_node *arch_dep_replace_mod_by_const(ir_node *irn) {
 	if (params == NULL || (opts & arch_dep_mod_by_const) == 0)
 		return irn;
 
-	if (get_irn_opcode(irn) == iro_Mod) {
+	if (is_Mod(irn)) {
 		ir_node *c = get_Mod_right(irn);
 		ir_node *block, *left;
 		ir_mode *mode;
@@ -983,7 +988,7 @@ ir_node *arch_dep_replace_mod_by_const(ir_node *irn) {
 		int n, bits;
 		int k;
 
-		if (get_irn_op(c) != op_Const)
+		if (! is_Const(c))
 			return irn;
 
 		tv = get_Const_tarval(c);
@@ -1068,7 +1073,7 @@ void arch_dep_replace_divmod_by_const(ir_node **div, ir_node **mod, ir_node *irn
 		((opts & (arch_dep_div_by_const|arch_dep_mod_by_const)) != (arch_dep_div_by_const|arch_dep_mod_by_const)))
 		return;
 
-	if (get_irn_opcode(irn) == iro_DivMod) {
+	if (is_DivMod(irn)) {
 		ir_node *c = get_DivMod_right(irn);
 		ir_node *block, *left;
 		ir_mode *mode;
@@ -1077,7 +1082,7 @@ void arch_dep_replace_divmod_by_const(ir_node **div, ir_node **mod, ir_node *irn
 		int n, bits;
 		int k, n_flag;
 
-		if (get_irn_op(c) != op_Const)
+		if (! is_Const(c))
 			return;
 
 		tv = get_Const_tarval(c);