X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firarch.c;h=387310f1936a8d37867529754a8083f3f307720d;hb=f250d98f949371038744bd320096fa98f4b218fe;hp=041286077512696115fd8afebb422e0b16613b65;hpb=6be9d429e8a25e1a818b9912ca3c038d020c4f8d;p=libfirm

diff --git a/ir/ir/irarch.c b/ir/ir/irarch.c
index 041286077..387310f19 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.
  *
@@ -23,15 +23,13 @@
  * @date    28.9.2004
  * @author  Sebastian Hack, Michael Beck
  * @version $Id$
+ *
+ * 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
-
-#ifdef HAVE_STDLIB_H
-# include <stdlib.h>
-#endif
+#include "config.h"
 
+#include <stdlib.h>
 #include <assert.h>
 
 #include "irnode_t.h"
@@ -48,6 +46,8 @@
 #include "irhooks.h"
 #include "ircons.h"
 #include "irarch.h"
+#include "irflag.h"
+#include "error.h"
 
 #undef DEB
 
@@ -66,40 +66,11 @@ static const ir_settings_arch_dep_t *params = NULL;
 /** The bit mask, which optimizations to apply. */
 static arch_dep_opts_t opts;
 
-/* we need this new pseudo op */
-static ir_op *op_Mulh = NULL;
-
-/**
- * construct a Mulh: Mulh(a,b) = (a * b) >> w, w is the with in bits of a, b
- */
-static ir_node *
-new_rd_Mulh (dbg_info *db, ir_graph *irg, ir_node *block,
-       ir_node *op1, ir_node *op2, ir_mode *mode) {
-	ir_node *in[2];
-	ir_node *res;
-
-	in[0] = op1;
-	in[1] = op2;
-	res = new_ir_node(db, irg, block, op_Mulh, mode, 2, in);
-	res = optimize_node(res);
-	IRN_VRFY_IRG(res, irg);
-	return res;
-}
-
-ir_op *get_op_Mulh(void)  { return op_Mulh; }
-
 void arch_dep_init(arch_dep_params_factory_t factory) {
 	opts = arch_dep_none;
 
 	if (factory != NULL)
 		params = factory();
-
-	if (! op_Mulh) {
-		int mulh_opc = get_next_ir_opcode();
-
-		/* create the Mulh operation */
-		op_Mulh = new_ir_op(mulh_opc, "Mulh",  op_pin_state_floats, irop_flag_commutative, oparity_binary, 0, 0, NULL);
-	}
 }
 
 void arch_dep_set_opts(arch_dep_opts_t the_opts) {
@@ -113,242 +84,517 @@ static int allow_Mulh(ir_mode *mode) {
 	return (mode_is_signed(mode) && params->allow_mulhs) || (!mode_is_signed(mode) && params->allow_mulhu);
 }
 
-/* Replace Muls with Shifts and Add/Subs. */
-ir_node *arch_dep_replace_mul_with_shifts(ir_node *irn) {
-	ir_node *res = irn;
-	ir_mode *mode = get_irn_mode(irn);
+/**
+ * An instruction,
+ */
+typedef struct instruction instruction;
+struct instruction {
+	insn_kind   kind;        /**< the instruction kind */
+	instruction *in[2];      /**< the ins */
+	unsigned    shift_count; /**< shift count for LEA and SHIFT */
+	ir_node     *irn;        /**< the generated node for this instruction if any. */
+	int         costs;       /**< the costs for this instruction */
+};
 
-	/* If the architecture dependent optimizations were not initialized
-	   or this optimization was not enabled. */
-	if (params == NULL || (opts & arch_dep_mul_to_shift) == 0)
-		return irn;
+/**
+ * The environment for the strength reduction of multiplications.
+ */
+typedef struct _mul_env {
+	struct obstack obst;       /**< an obstack for local space. */
+	ir_mode        *mode;      /**< the mode of the multiplication constant */
+	unsigned       bits;       /**< number of bits in the mode */
+	unsigned       max_S;      /**< the maximum LEA shift value. */
+	instruction    *root;      /**< the root of the instruction tree */
+	ir_node        *op;        /**< the operand that is multiplied */
+	ir_node        *blk;       /**< the block where the new graph is built */
+	dbg_info       *dbg;       /**< the debug info for the new graph. */
+	ir_mode        *shf_mode;  /**< the (unsigned) mode for the shift constants */
+	int            fail;       /**< set to 1 if the instruction sequence fails the constraints */
+	int            n_shift;    /**< maximum number of allowed shift instructions */
+
+	evaluate_costs_func evaluate;  /**< the evaluate callback */
+} mul_env;
 
-	if (get_irn_op(irn) == op_Mul && mode_is_int(mode)) {
-		ir_node *block   = get_irn_n(irn, -1);
-		ir_node *left    = get_binop_left(irn);
-		ir_node *right   = get_binop_right(irn);
-		tarval *tv       = NULL;
-		ir_node *operand = NULL;
-
-		/* Look, if one operand is a constant. */
-		if (get_irn_opcode(left) == iro_Const) {
-			tv = get_Const_tarval(left);
-			operand = right;
-		} else if(get_irn_opcode(right) == iro_Const) {
-			tv = get_Const_tarval(right);
-			operand = left;
-		}
+/**
+ * Some kind of default evaluator. Return the cost of
+ * instructions.
+ */
+static int default_evaluate(insn_kind kind, tarval *tv) {
+	(void) tv;
 
-		if (tv != NULL) {
-			int maximum_shifts = params->maximum_shifts;
-			int also_use_subs = params->also_use_subs;
-			int highest_shift_amount = params->highest_shift_amount;
-
-			char *bitstr = get_tarval_bitpattern(tv);
-			char *p;
-			int i, last = 0;
-			int counter = 0;
-			int curr_bit;
-			int compr_len = 0;
-			char compr[MAX_BITSTR];
-
-			int singleton;
-			int end_of_group;
-			int shift_with_sub[MAX_BITSTR] = { 0 };
-			int shift_without_sub[MAX_BITSTR] = { 0 };
-			int shift_with_sub_pos = 0;
-			int shift_without_sub_pos = 0;
-
-#if DEB
-			{
-				long val = get_tarval_long(tv);
-				fprintf(stderr, "Found mul with %ld(%lx) = ", val, val);
-				for(p = bitstr; *p != '\0'; p++)
-					printf("%c", *p);
-				printf("\n");
-			}
-#endif
+	if (kind == MUL)
+		return 13;
+	return 1;
+}
 
-			for(p = bitstr; *p != '\0'; p++) {
-				int bit = *p != '0';
+/**
+ * emit a LEA (or an Add) instruction
+ */
+static instruction *emit_LEA(mul_env *env, instruction *a, instruction *b, unsigned shift) {
+	instruction *res = obstack_alloc(&env->obst, sizeof(*res));
+	res->kind = shift > 0 ? LEA : ADD;
+	res->in[0] = a;
+	res->in[1] = b;
+	res->shift_count = shift;
+	res->irn = NULL;
+	res->costs = -1;
+	return res;
+}
 
-				if (bit != last) {
-					/* The last was 1 we are now at 0 OR
-					 * The last was 0 and we are now at 1 */
-					compr[compr_len++] = counter;
-					counter = 1;
-				} else
-					counter++;
+/**
+ * emit a SHIFT (or an Add or a Zero) instruction
+ */
+static instruction *emit_SHIFT(mul_env *env, instruction *a, unsigned shift) {
+	instruction *res = obstack_alloc(&env->obst, sizeof(*res));
+	if (shift == env->bits) {
+		/* a 2^bits with bits resolution is a zero */
+		res->kind = ZERO;
+		res->in[0] = NULL;
+		res->in[1] = NULL;
+		res->shift_count = 0;
+	} else if (shift != 1) {
+		res->kind = SHIFT;
+		res->in[0] = a;
+		res->in[1] = NULL;
+		res->shift_count = shift;
+	} else {
+		res->kind = ADD;
+		res->in[0] = a;
+		res->in[1] = a;
+		res->shift_count = 0;
+	}
+	res->irn = NULL;
+	res->costs = -1;
+	return res;
+}
 
-				last = bit;
-			}
-			compr[compr_len++] = counter;
-
-#ifdef DEB
-			{
-				const char *prefix = "";
-				for(i = 0; i < compr_len; i++, prefix = ",")
-					fprintf(stderr, "%s%d", prefix, compr[i]);
-				fprintf("\n");
-			}
-#endif
+/**
+ * emit a SUB instruction
+ */
+static instruction *emit_SUB(mul_env *env, instruction *a, instruction *b) {
+	instruction *res = obstack_alloc(&env->obst, sizeof(*res));
+	res->kind = SUB;
+	res->in[0] = a;
+	res->in[1] = b;
+	res->shift_count = 0;
+	res->irn = NULL;
+	res->costs = -1;
+	return res;
+}
 
-			/* Go over all recorded one groups. */
-			curr_bit = compr[0];
+/**
+ * emit the ROOT instruction
+ */
+static instruction *emit_ROOT(mul_env *env, ir_node *root_op) {
+	instruction *res = obstack_alloc(&env->obst, sizeof(*res));
+	res->kind = ROOT;
+	res->in[0] = NULL;
+	res->in[1] = NULL;
+	res->shift_count = 0;
+	res->irn = root_op;
+	res->costs = 0;
+	return res;
+}
 
-			for(i = 1; i < compr_len; i = end_of_group + 2) {
-				int j, zeros_in_group, ones_in_group;
 
-				ones_in_group = compr[i];
-				zeros_in_group = 0;
+/**
+ * Returns the condensed representation of the tarval tv
+ */
+static unsigned char *value_to_condensed(mul_env *env, tarval *tv, int *pr) {
+	ir_mode *mode = get_tarval_mode(tv);
+	int     bits = get_mode_size_bits(mode);
+	char    *bitstr = get_tarval_bitpattern(tv);
+	int     i, l, r;
+	unsigned char *R = obstack_alloc(&env->obst, bits);
+
+	l = r = 0;
+	for (i = 0; bitstr[i] != '\0'; ++i) {
+		if (bitstr[i] == '1') {
+			R[r] = i - l;
+			l = i;
+			++r;
+		}
+	}
+	free(bitstr);
 
-				/* Scan for singular 0s in a sequence. */
-				for(j = i + 1; j < compr_len && compr[j] == 1; j += 2) {
-					zeros_in_group += 1;
-					ones_in_group += (j + 1 < compr_len ? compr[j + 1] : 0);
-				}
-				end_of_group = j - 1;
+	*pr = r;
+	return R;
+}
 
-				if(zeros_in_group >= ones_in_group - 1)
-					end_of_group = i;
+/**
+ * Calculate the gain when using the generalized complementary technique
+ */
+static int calculate_gain(unsigned char *R, int r) {
+	int max_gain = 0;
+	int idx = -1, i;
+	int gain;
+
+	/* the gain for r == 1 */
+	gain = 2 - 3 - R[0];
+	for (i = 2; i < r; ++i) {
+		/* calculate the gain for r from the gain for r-1 */
+		gain += 2 - R[i - 1];
+
+		if (gain > max_gain) {
+			max_gain = gain;
+			idx = i;
+		}
+	}
+	return idx;
+}
 
-#ifdef DEB
-				fprintf(stderr, "  i:%d, eg:%d\n", i, end_of_group);
-#endif
+/**
+ * Calculates the condensed complement of a given (R,r) tuple
+ */
+static unsigned char *complement_condensed(mul_env *env, unsigned char *R, int r, int gain, int *prs) {
+	unsigned char *value = obstack_alloc(&env->obst, env->bits);
+	int i, l, j;
+	unsigned char c;
 
-				singleton = compr[i] == 1 && i == end_of_group;
-				for(j = i; j <= end_of_group; j += 2) {
-					int curr_ones = compr[j];
-					int biased_curr_bit = curr_bit + 1;
-					int k;
+	memset(value, 0, env->bits);
 
-#ifdef DEB
-					fprintf(stderr, "    j:%d, ones:%d\n", j, curr_ones);
-#endif
+	j = 0;
+	for (i = 0; i < gain; ++i) {
+		j += R[i];
+		value[j] = 1;
+	}
 
-					/* If this ones group is a singleton group (it has no
-					   singleton zeros inside. */
-					if(singleton)
-						shift_with_sub[shift_with_sub_pos++] = biased_curr_bit;
-					else if(j == i)
-						shift_with_sub[shift_with_sub_pos++] = -biased_curr_bit;
+	/* negate and propagate 1 */
+	c = 1;
+	for (i = 0; i <= j; ++i) {
+		unsigned char v = !value[i];
 
-					for(k = 0; k < curr_ones; k++)
-						shift_without_sub[shift_without_sub_pos++] = biased_curr_bit + k;
+		value[i] = v ^ c;
+		c = v & c;
+	}
 
-					curr_bit += curr_ones;
-					biased_curr_bit = curr_bit + 1;
+	/* condense it again */
+	l = r = 0;
+	R = value;
+	for (i = 0; i <= j; ++i) {
+		if (value[i] == 1) {
+			R[r] = i - l;
+			l = i;
+			++r;
+		}
+	}
 
-					if(!singleton && j == end_of_group)
-						shift_with_sub[shift_with_sub_pos++] = biased_curr_bit;
-					else if(j != end_of_group)
-						shift_with_sub[shift_with_sub_pos++] = -biased_curr_bit;
+	*prs = r;
+	return R;
+}
 
-					curr_bit += compr[j + 1];
-				}
+/**
+ * creates a tarval from a condensed representation.
+ */
+static tarval *condensed_to_value(mul_env *env, unsigned char *R, int r) {
+	tarval *res, *tv;
+	int i, j;
+
+	j = 0;
+	tv = get_mode_one(env->mode);
+	res = NULL;
+	for (i = 0; i < r; ++i) {
+		j = R[i];
+		if (j) {
+			tarval *t = new_tarval_from_long(j, mode_Iu);
+			tv = tarval_shl(tv, t);
+		}
+		res = res ? tarval_add(res, tv) : tv;
+	}
+	return res;
+}
+
+/* forward */
+static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N);
+
+/*
+ * handle simple cases with up-to 2 bits set
+ */
+static instruction *decompose_simple_cases(mul_env *env, unsigned char *R, int r, tarval *N) {
+	instruction *ins, *ins2;
+
+	(void) N;
+	if (r == 1) {
+		return emit_SHIFT(env, env->root, R[0]);
+	} else {
+		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]);
+		}
 
-			{
-				int *shifts = shift_with_sub;
-				int n = shift_with_sub_pos;
-				int highest_shift_wide = 0;
-				int highest_shift_seq = 0;
-				int last_shift = 0;
-
-				/* If we may not use subs, or we can achive the same with adds,
-				   prefer adds. */
-				if(!also_use_subs || shift_with_sub_pos >= shift_without_sub_pos) {
-					shifts = shift_without_sub;
-					n = shift_without_sub_pos;
-				}
+		ins2 = emit_SHIFT(env, env->root, R[0] + R[1]);
+		return emit_LEA(env, ins, ins2, 0);
+	}
+}
 
-				/* If the number of needed shifts exceeds the given maximum,
-				   use the Mul and exit. */
-				if(n > maximum_shifts) {
-#ifdef DEB
-					fprintf(stderr, "Only allowed %d shifts, but %d are needed\n",
-						maximum_shifts, n);
-#endif
-					return irn;
-				}
+/**
+ * Main decompose driver.
+ */
+static instruction *decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N) {
+	unsigned i;
+	int gain;
+
+	if (r <= 2)
+		return decompose_simple_cases(env, R, r, N);
+
+	if (params->also_use_subs) {
+		gain = calculate_gain(R, r);
+		if (gain > 0) {
+			instruction *instr1, *instr2;
+			unsigned char *R1, *R2;
+			int r1, r2, i, k, j;
+
+			R1 = complement_condensed(env, R, r, gain, &r1);
+			r2 = r - gain + 1;
+			R2 = obstack_alloc(&env->obst, r2);
+
+			k = 1;
+			for (i = 0; i < gain; ++i) {
+				k += R[i];
+			}
+			R2[0] = k;
+			R2[1] = R[gain] - 1;
+			j = 2;
+			if (R2[1] == 0) {
+				/* Two identical bits: normalize */
+				++R2[0];
+				--j;
+				--r2;
+			}
+			for (i = gain + 1; i < r; ++i) {
+				R2[j++] = R[i];
+			}
 
-				/* Compute the highest shift needed for both, the
-				   sequential and wide representations. */
-				for(i = 0; i < n; i++) {
-					int curr = abs(shifts[i]);
-					int curr_seq = curr - last;
+			instr1 = decompose_mul(env, R1, r1, NULL);
+			instr2 = decompose_mul(env, R2, r2, NULL);
+			return emit_SUB(env, instr2, instr1);
+		}
+	}
 
-					highest_shift_wide = curr > highest_shift_wide ? curr : highest_shift_wide;
-					highest_shift_seq = curr_seq > highest_shift_seq ? curr_seq : highest_shift_seq;
+	if (N == NULL)
+		N = condensed_to_value(env, R, r);
 
-					last_shift = curr;
-				}
+	for (i = env->max_S; i > 0; --i) {
+		tarval *div_res, *mod_res;
+		tarval *tv = new_tarval_from_long((1 << i) + 1, env->mode);
 
-				/* If the highest shift amount is greater than the given limit,
-				   give back the Mul */
-				if(highest_shift_seq > highest_shift_amount) {
-#ifdef DEB
-					fprintf(stderr, "Shift argument %d exceeds maximum %d\n",
-						highest_shift_seq, highest_shift_amount);
-#endif
-					return irn;
-				}
+		div_res = tarval_divmod(N, tv, &mod_res);
+		if (mod_res == get_mode_null(env->mode)) {
+			unsigned char *Rs;
+			int rs;
+
+			Rs = value_to_condensed(env, div_res, &rs);
+			if (rs < r) {
+				instruction *N1 = decompose_mul(env, Rs, rs, div_res);
+				return emit_LEA(env, N1, N1, i);
+			}
+		}
+	}
+	return basic_decompose_mul(env, R, r, N);
+}
 
-				/* If we have subs, we cannot do sequential. */
-				if(1 /* also_use_subs */) {
-					if(n > 0) {
-						ir_node *curr = NULL;
+#define IMAX(a,b) ((a) > (b) ? (a) : (b))
 
-						i = n - 1;
+/**
+ * basic decomposition routine
+ */
+static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, tarval *N) {
+	instruction *Ns;
+	unsigned t;
+
+	if (R[0] == 0) {					/* Case 1 */
+		t = R[1] > IMAX(env->max_S, R[1]);
+		R[1] -= t;
+		Ns = decompose_mul(env, &R[1], r - 1, N);
+		return emit_LEA(env, env->root, Ns, t);
+	} else if (R[0] <= env->max_S) {	/* Case 2 */
+		t = R[0];
+		R[1] += t;
+		Ns = decompose_mul(env, &R[1], r - 1, N);
+		return emit_LEA(env, Ns, env->root, t);
+	} else {
+		t = R[0];
+		R[0] = 0;
+		Ns = decompose_mul(env, R, r, N);
+		return emit_SHIFT(env, Ns, t);
+	}
+}
 
-						do {
-							int curr_shift = shifts[i];
-							int sub = curr_shift < 0;
-							int amount = abs(curr_shift) - 1;
-							ir_node *aux = operand;
+/**
+ * recursive build the graph form the instructions.
+ *
+ * @param env   the environment
+ * @param inst  the instruction
+ */
+static ir_node *build_graph(mul_env *env, instruction *inst) {
+	ir_node *l, *r, *c;
+
+	if (inst->irn)
+		return inst->irn;
+
+	switch (inst->kind) {
+	case LEA:
+		l = build_graph(env, inst->in[0]);
+		r = build_graph(env, inst->in[1]);
+		c = new_Const_long(env->shf_mode, inst->shift_count);
+		r = new_rd_Shl(env->dbg, current_ir_graph, env->blk, r, c, env->mode);
+		return inst->irn = new_rd_Add(env->dbg, current_ir_graph, env->blk, l, r, env->mode);
+	case SHIFT:
+		l = build_graph(env, inst->in[0]);
+		c = new_Const_long(env->shf_mode, inst->shift_count);
+		return inst->irn = new_rd_Shl(env->dbg, current_ir_graph, env->blk, l, c, env->mode);
+	case SUB:
+		l = build_graph(env, inst->in[0]);
+		r = build_graph(env, inst->in[1]);
+		return inst->irn = new_rd_Sub(env->dbg, current_ir_graph, env->blk, l, r, env->mode);
+	case ADD:
+		l = build_graph(env, inst->in[0]);
+		r = build_graph(env, inst->in[1]);
+		return inst->irn = new_rd_Add(env->dbg, current_ir_graph, env->blk, l, r, env->mode);
+	case ZERO:
+		return inst->irn = new_Const(get_mode_null(env->mode));
+	default:
+		panic("Unsupported instruction kind");
+		return NULL;
+	}
+}
 
-							assert(amount >= 0 && "What is a negative shift??");
+/**
+ * Calculate the costs for the given instruction sequence.
+ * Note that additional costs due to higher register pressure are NOT evaluated yet
+ */
+static int evaluate_insn(mul_env *env, instruction *inst) {
+	int costs;
 
-							if (amount != 0) {
-								ir_node *cnst = new_r_Const_long(current_ir_graph, block, mode_Iu, amount);
-								aux = new_r_Shl(current_ir_graph, block, operand, cnst, mode);
-							}
+	if (inst->costs >= 0) {
+		/* was already evaluated */
+		return 0;
+	}
 
-							if (curr) {
-								if (sub)
-									curr = new_r_Sub(current_ir_graph, block, curr, aux, mode);
-								else
-									curr = new_r_Add(current_ir_graph, block, curr, aux, mode);
-							} else
-								curr = aux;
+	switch (inst->kind) {
+	case LEA:
+	case SUB:
+	case ADD:
+		costs  = evaluate_insn(env, inst->in[0]);
+		costs += evaluate_insn(env, inst->in[1]);
+		costs += env->evaluate(inst->kind, NULL);
+		inst->costs = costs;
+		return costs;
+	case SHIFT:
+		if (inst->shift_count > params->highest_shift_amount)
+			env->fail = 1;
+		if (env->n_shift <= 0)
+			env->fail = 1;
+		else
+			--env->n_shift;
+		costs  = evaluate_insn(env, inst->in[0]);
+		costs += env->evaluate(inst->kind, NULL);
+		inst->costs = costs;
+		return costs;
+	case ZERO:
+		inst->costs = costs = env->evaluate(inst->kind, NULL);
+		return costs;
+	case MUL:
+	case ROOT:
+		break;
+	}
+	panic("Unsupported instruction kind");
+}
 
-						} while(--i >= 0);
+/**
+ * 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.
+ *
+ * @param irn      the Mul operation
+ * @param operand  the multiplication operand
+ * @param tv       the multiplication constant
+ *
+ * @return the new graph
+ */
+static ir_node *do_decomposition(ir_node *irn, ir_node *operand, tarval *tv) {
+	mul_env       env;
+	instruction   *inst;
+	unsigned char *R;
+	int           r;
+	ir_node       *res = irn;
+	int           mul_costs;
+
+	obstack_init(&env.obst);
+	env.mode     = get_tarval_mode(tv);
+	env.bits     = (unsigned)get_mode_size_bits(env.mode);
+	env.max_S    = 3;
+	env.root     = emit_ROOT(&env, operand);
+	env.fail     = 0;
+	env.n_shift  = params->maximum_shifts;
+	env.evaluate = params->evaluate != NULL ? params->evaluate : default_evaluate;
+
+	R = value_to_condensed(&env, tv, &r);
+	inst = decompose_mul(&env, R, r, tv);
+
+	/* the paper suggests 70% here */
+	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);
+		env.dbg      = get_irn_dbg_info(irn);
+		env.shf_mode = find_unsigned_mode(env.mode);
+		if (env.shf_mode == NULL)
+			env.shf_mode = mode_Iu;
+
+		res = build_graph(&env, inst);
+	}
+	obstack_free(&env.obst, NULL);
+	return res;
+}
 
-						res = curr;
-					}
-				}
+/* Replace Muls with Shifts and Add/Subs. */
+ir_node *arch_dep_replace_mul_with_shifts(ir_node *irn) {
+	ir_node *res = irn;
+	ir_mode *mode = get_irn_mode(irn);
 
-#ifdef DEB
-				{
-					const char *prefix = "";
-					for (i = 0; i < n; ++i) {
-						fprintf(stderr, "%s%d", prefix, shifts[i]);
-						prefix = ", ";
-					}
-					fprintf(stderr, "\n");
-				}
-#endif
+	/* If the architecture dependent optimizations were not initialized
+	   or this optimization was not enabled. */
+	if (params == NULL || (opts & arch_dep_mul_to_shift) == 0)
+		return irn;
 
+	set_arch_dep_running(1);
+	{
+		if (is_Mul(irn) && mode_is_int(mode)) {
+			ir_node *left    = get_binop_left(irn);
+			ir_node *right   = get_binop_right(irn);
+			tarval *tv       = NULL;
+			ir_node *operand = NULL;
+
+			/* Look, if one operand is a constant. */
+			if (is_Const(left)) {
+				tv = get_Const_tarval(left);
+				operand = right;
+			} else if (is_Const(right)) {
+				tv = get_Const_tarval(right);
+				operand = left;
 			}
 
-			if(bitstr)
-				free(bitstr);
-		}
+			if (tv != NULL) {
+				res = do_decomposition(irn, operand, tv);
 
+				if (res != irn) {
+					hook_arch_dep_replace_mul_with_shifts(irn);
+					exchange(irn, res);
+				}
+			}
+		}
 	}
-
-	if (res != irn)
-		hook_arch_dep_replace_mul_with_shifts(irn);
+	//set_arch_dep_running(0);
 
 	return res;
 }
@@ -385,7 +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)
@@ -576,7 +822,7 @@ static ir_node *replace_div_by_mulh(ir_node *div, tarval *tv) {
 		struct ms mag = magic(tv);
 
 		/* generate the Mulh instruction */
-		c = new_r_Const(current_ir_graph, block, mode, mag.M);
+		c = new_Const(mag.M);
 		q = new_rd_Mulh(dbg, current_ir_graph, block, n, c, mode);
 
 		/* do we need an Add or Sub */
@@ -587,12 +833,12 @@ static ir_node *replace_div_by_mulh(ir_node *div, tarval *tv) {
 
 		/* Do we need the shift */
 		if (mag.s > 0) {
-			c = new_r_Const_long(current_ir_graph, block, mode_Iu, mag.s);
-			q    = new_rd_Shrs(dbg, current_ir_graph, block, q, c, mode);
+			c = new_Const_long(mode_Iu, mag.s);
+			q = new_rd_Shrs(dbg, current_ir_graph, block, q, c, mode);
 		}
 
 		/* final */
-		c = new_r_Const_long(current_ir_graph, block, mode_Iu, bits-1);
+		c = new_Const_long(mode_Iu, bits - 1);
 		t = new_rd_Shr(dbg, current_ir_graph, block, q, c, mode);
 
 		q = new_rd_Add(dbg, current_ir_graph, block, q, t, mode);
@@ -601,27 +847,27 @@ static ir_node *replace_div_by_mulh(ir_node *div, tarval *tv) {
 		ir_node *c;
 
 		/* 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);
+		c = new_Const(mag.M);
+		q = new_rd_Mulh(dbg, current_ir_graph, block, n, c, mode);
 
 		if (mag.need_add) {
 			if (mag.s > 0) {
 				/* use the GM scheme */
 				t = new_rd_Sub(dbg, current_ir_graph, block, n, q, mode);
 
-				c = new_r_Const(current_ir_graph, block, mode_Iu, get_mode_one(mode_Iu));
+				c = new_Const(get_mode_one(mode_Iu));
 				t = new_rd_Shr(dbg, current_ir_graph, block, t, c, mode);
 
 				t = new_rd_Add(dbg, current_ir_graph, block, t, q, mode);
 
-				c = new_r_Const_long(current_ir_graph, block, mode_Iu, mag.s-1);
+				c = new_Const_long(mode_Iu, mag.s - 1);
 				q = new_rd_Shr(dbg, current_ir_graph, block, t, c, mode);
 			} else {
 				/* use the default scheme */
 				q = new_rd_Add(dbg, current_ir_graph, block, q, n, mode);
 			}
 		} else if (mag.s > 0) { /* default scheme, shift needed */
-			c = new_r_Const_long(current_ir_graph, block, mode_Iu, mag.s);
+			c = new_Const_long(mode_Iu, mag.s);
 			q = new_rd_Shr(dbg, current_ir_graph, block, q, c, mode);
 		}
 	}
@@ -637,7 +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;
@@ -646,13 +892,13 @@ 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);
 
 		/* check for division by zero */
-		if (classify_tarval(tv) == TV_CLASSIFY_NULL)
+		if (tarval_is_null(tv))
 			return irn;
 
 		left  = get_Div_left(irn);
@@ -681,29 +927,34 @@ 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_Const_long(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_Const_long(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);
+				k_node = new_Const_long(mode_Iu, k);
 				res    = new_rd_Shrs(dbg, current_ir_graph, block, curr, k_node, mode);
 
 				if (n_flag) { /* negate the result */
 					ir_node *k_node;
 
-					k_node = new_r_Const(current_ir_graph, block, mode, get_mode_null(mode));
+					k_node = new_Const(get_mode_null(mode));
 					res = new_rd_Sub(dbg, current_ir_graph, block, k_node, res, mode);
 				}
 			} else {      /* unsigned case */
 				ir_node *k_node;
 
-				k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k);
+				k_node = new_Const_long(mode_Iu, k);
 				res    = new_rd_Shr(dbg, current_ir_graph, block, left, k_node, mode);
 			}
 		} else {
@@ -728,7 +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;
@@ -737,13 +988,13 @@ 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);
 
 		/* check for division by zero */
-		if (classify_tarval(tv) == TV_CLASSIFY_NULL)
+		if (tarval_is_null(tv))
 			return irn;
 
 		left  = get_Mod_left(irn);
@@ -773,23 +1024,23 @@ ir_node *arch_dep_replace_mod_by_const(ir_node *irn) {
 				ir_node *curr = left;
 
 				if (k != 1) {
-					k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k - 1);
+					k_node = new_Const_long(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);
+				k_node = new_Const_long(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);
 
-				k_node = new_r_Const_long(current_ir_graph, block, mode, (-1) << k);
+				k_node = new_Const_long(mode, (-1) << k);
 				curr   = new_rd_And(dbg, current_ir_graph, block, curr, k_node, mode);
 
 				res    = new_rd_Sub(dbg, current_ir_graph, block, left, curr, mode);
 			} else {      /* unsigned case */
 				ir_node *k_node;
 
-				k_node = new_r_Const_long(current_ir_graph, block, mode, (1 << k) - 1);
+				k_node = new_Const_long(mode, (1 << k) - 1);
 				res    = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode);
 			}
 		} else {
@@ -822,7 +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;
@@ -831,13 +1082,13 @@ 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);
 
 		/* check for division by zero */
-		if (classify_tarval(tv) == TV_CLASSIFY_NULL)
+		if (tarval_is_null(tv))
 			return;
 
 		left  = get_DivMod_left(irn);
@@ -867,37 +1118,37 @@ void arch_dep_replace_divmod_by_const(ir_node **div, ir_node **mod, ir_node *irn
 				ir_node *curr = left;
 
 				if (k != 1) {
-					k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k - 1);
+					k_node = new_Const_long(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);
+				k_node = new_Const_long(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);
 
-				c_k    = new_r_Const_long(current_ir_graph, block, mode_Iu, k);
+				c_k    = new_Const_long(mode_Iu, k);
 
 				*div   = new_rd_Shrs(dbg, current_ir_graph, block, curr, c_k, mode);
 
 				if (n_flag) { /* negate the div result */
 					ir_node *k_node;
 
-					k_node = new_r_Const(current_ir_graph, block, mode, get_mode_null(mode));
+					k_node = new_Const(get_mode_null(mode));
 					*div = new_rd_Sub(dbg, current_ir_graph, block, k_node, *div, mode);
 				}
 
-				k_node = new_r_Const_long(current_ir_graph, block, mode, (-1) << k);
+				k_node = new_Const_long(mode, (-1) << k);
 				curr   = new_rd_And(dbg, current_ir_graph, block, curr, k_node, mode);
 
 				*mod   = new_rd_Sub(dbg, current_ir_graph, block, left, curr, mode);
 			} else {      /* unsigned case */
 				ir_node *k_node;
 
-				k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k);
+				k_node = new_Const_long(mode_Iu, k);
 				*div   = new_rd_Shr(dbg, current_ir_graph, block, left, k_node, mode);
 
-				k_node = new_r_Const_long(current_ir_graph, block, mode, (1 << k) - 1);
+				k_node = new_Const_long(mode, (1 << k) - 1);
 				*mod   = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode);
 			}
 		} else {
@@ -925,6 +1176,7 @@ static const ir_settings_arch_dep_t default_params = {
 	1,  /* also use subs */
 	4,  /* maximum shifts */
 	31, /* maximum shift amount */
+	default_evaluate,  /* default evaluator */
 
 	0,  /* allow Mulhs */
 	0,  /* allow Mulus */