X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firarch.c;h=db787cbc87be2ceb92570ae208df4d1fcabd380a;hb=fd761f39f43ad67aae5217a5d1fdbe9837d766d2;hp=9acb1df15a24c0edcb1f8f823a9b2fcb786f3735;hpb=c6c276b17556a709a907f1f22c78650aee59e526;p=libfirm

diff --git a/ir/ir/irarch.c b/ir/ir/irarch.c
index 9acb1df15..db787cbc8 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) 1995-2011 University of Karlsruhe.  All right reserved.
  *
  * This file is part of libFirm.
  *
@@ -22,16 +22,14 @@
  * @brief   Machine dependent Firm optimizations.
  * @date    28.9.2004
  * @author  Sebastian Hack, Michael Beck
- * @version $Id$
+ *
+ * Implements "Strength 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"
@@ -40,7 +38,7 @@
 #include "iropt_t.h"
 #include "ircons_t.h"
 #include "irgmod.h"
-#include "irvrfy.h"
+#include "irverify.h"
 #include "tv_t.h"
 #include "dbginfo_t.h"
 #include "iropt_dbg.h"
@@ -48,316 +46,577 @@
 #include "irhooks.h"
 #include "ircons.h"
 #include "irarch.h"
+#include "irflag.h"
+#include "be.h"
+#include "error.h"
 
-#undef DEB
+/** The bit mask, which optimizations to apply. */
+static arch_dep_opts_t opts;
 
-#define MAX_BITSTR 64
+void arch_dep_set_opts(arch_dep_opts_t the_opts)
+{
+	opts = the_opts;
+}
 
-/* when we need verifying */
-#ifdef NDEBUG
-# define IRN_VRFY_IRG(res, irg)
-#else
-# define IRN_VRFY_IRG(res, irg)  irn_vrfy_irg(res, irg)
-#endif
+/** check, whether a mode allows a Mulh instruction. */
+static int allow_Mulh(const ir_settings_arch_dep_t *params, ir_mode *mode)
+{
+	if (get_mode_size_bits(mode) > params->max_bits_for_mulh)
+		return 0;
+	return (mode_is_signed(mode) && params->allow_mulhs) || (!mode_is_signed(mode) && params->allow_mulhu);
+}
 
-/** The params got from the factory in arch_dep_init(...). */
-static const ir_settings_arch_dep_t *params = NULL;
+/**
+ * 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 */
+};
 
-/** The bit mask, which optimizations to apply. */
-static arch_dep_opts_t opts;
+/**
+ * The environment for the strength reduction of multiplications.
+ */
+typedef struct mul_env {
+	struct obstack obst;       /**< an obstack for local space. */
+	const ir_settings_arch_dep_t *params;
+	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 */
+	ir_graph       *irg;
+	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;
 
-/* we need this new pseudo op */
-static ir_op *op_Mulh = NULL;
+/**
+ * Some kind of default evaluator. Return the cost of
+ * instructions.
+ */
+static int default_evaluate(insn_kind kind, const ir_mode *mode, ir_tarval *tv)
+{
+	(void) mode;
+	(void) tv;
+
+	if (kind == MUL)
+		return 13;
+	return 1;
+}
 
 /**
- * construct a Mulh: Mulh(a,b) = (a * b) >> w, w is the with in bits of a, b
+ * emit a LEA (or an Add) instruction
  */
-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);
+static instruction *emit_LEA(mul_env *env, instruction *a, instruction *b, unsigned shift)
+{
+	instruction *res = OALLOC(&env->obst, instruction);
+	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;
 }
 
-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);
+/**
+ * emit a SHIFT (or an Add or a Zero) instruction
+ */
+static instruction *emit_SHIFT(mul_env *env, instruction *a, unsigned shift)
+{
+	instruction *res = OALLOC(&env->obst, instruction);
+	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;
 }
 
-void arch_dep_set_opts(arch_dep_opts_t the_opts) {
-	opts = the_opts;
+/**
+ * emit a SUB instruction
+ */
+static instruction *emit_SUB(mul_env *env, instruction *a, instruction *b)
+{
+	instruction *res = OALLOC(&env->obst, instruction);
+	res->kind = SUB;
+	res->in[0] = a;
+	res->in[1] = b;
+	res->shift_count = 0;
+	res->irn = NULL;
+	res->costs = -1;
+	return res;
 }
 
-/** check, whether a mode allows a Mulh instruction. */
-static int allow_Mulh(ir_mode *mode) {
-	if (get_mode_size_bits(mode) > params->max_bits_for_mulh)
-		return 0;
-	return (mode_is_signed(mode) && params->allow_mulhs) || (!mode_is_signed(mode) && params->allow_mulhu);
+/**
+ * emit the ROOT instruction
+ */
+static instruction *emit_ROOT(mul_env *env, ir_node *root_op)
+{
+	instruction *res = OALLOC(&env->obst, instruction);
+	res->kind = ROOT;
+	res->in[0] = NULL;
+	res->in[1] = NULL;
+	res->shift_count = 0;
+	res->irn = root_op;
+	res->costs = 0;
+	return res;
 }
 
-/* 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);
-
-	/* 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;
 
-	if (is_Mul(irn) && mode_is_int(mode)) {
-		ir_node *block   = get_nodes_block(irn);
-		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;
+/**
+ * Returns the condensed representation of the tarval tv
+ */
+static unsigned char *value_to_condensed(mul_env *env, ir_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 = (unsigned char*)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);
 
-		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
-
-			for (p = bitstr; *p != '\0'; p++) {
-				int bit = *p != '0';
-
-				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++;
-
-				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
+	*pr = r;
+	return R;
+}
 
-			/* Go over all recorded one groups. */
-			curr_bit = compr[0];
+/**
+ * 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;
+}
 
-			for(i = 1; i < compr_len; i = end_of_group + 2) {
-				int j, zeros_in_group, ones_in_group;
+/**
+ * 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 = (unsigned char*)obstack_alloc(&env->obst, env->bits);
+	int i, l, j;
+	unsigned char c;
+
+	memset(value, 0, env->bits);
+
+	j = 0;
+	for (i = 0; i < gain; ++i) {
+		j += R[i];
+		value[j] = 1;
+	}
 
-				ones_in_group = compr[i];
-				zeros_in_group = 0;
+	/* negate and propagate 1 */
+	c = 1;
+	for (i = 0; i <= j; ++i) {
+		unsigned char v = !value[i];
 
-				/* 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;
+		value[i] = v ^ c;
+		c = v & c;
+	}
 
-				if(zeros_in_group >= ones_in_group - 1)
-					end_of_group = i;
+	/* 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;
+		}
+	}
 
-#ifdef DEB
-				fprintf(stderr, "  i:%d, eg:%d\n", i, end_of_group);
-#endif
+	*prs = r;
+	return R;
+}
 
-				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;
+/**
+ * creates a tarval from a condensed representation.
+ */
+static ir_tarval *condensed_to_value(mul_env *env, unsigned char *R, int r)
+{
+	ir_tarval *tv  = get_mode_one(env->mode);
+	ir_tarval *res = NULL;
+	for (int i = 0; i < r; ++i) {
+		int j = R[i];
+		if (j) {
+			ir_tarval *t = new_tarval_from_long(j, mode_Iu);
+			tv = tarval_shl(tv, t);
+		}
+		res = res ? tarval_add(res, tv) : tv;
+	}
+	return res;
+}
 
-#ifdef DEB
-					fprintf(stderr, "    j:%d, ones:%d\n", j, curr_ones);
-#endif
+/* forward */
+static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, ir_tarval *N);
 
-					/* 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;
+/*
+ * handle simple cases with up-to 2 bits set
+ */
+static instruction *decompose_simple_cases(mul_env *env, unsigned char *R, int r, ir_tarval *N)
+{
+	instruction *ins, *ins2;
 
-					for(k = 0; k < curr_ones; k++)
-						shift_without_sub[shift_without_sub_pos++] = biased_curr_bit + k;
+	(void) N;
+	if (r == 1) {
+		return emit_SHIFT(env, env->root, R[0]);
+	} else {
+		assert(r == 2);
 
-					curr_bit += curr_ones;
-					biased_curr_bit = curr_bit + 1;
+		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(!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;
+		ins2 = emit_SHIFT(env, env->root, R[0] + R[1]);
+		return emit_LEA(env, ins, ins2, 0);
+	}
+}
 
-					curr_bit += compr[j + 1];
-				}
+/**
+ * Main decompose driver.
+ */
+static instruction *decompose_mul(mul_env *env, unsigned char *R, int r, ir_tarval *N)
+{
+	unsigned i;
+	int gain;
+
+	if (r <= 2)
+		return decompose_simple_cases(env, R, r, N);
+
+	if (env->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 = (unsigned char*)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];
 			}
 
-			{
-				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 achieve 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;
-				}
+			instr1 = decompose_mul(env, R1, r1, NULL);
+			instr2 = decompose_mul(env, R2, r2, NULL);
+			return emit_SUB(env, instr2, instr1);
+		}
+	}
 
-				/* 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
-					goto end;
-				}
+	if (N == NULL)
+		N = condensed_to_value(env, R, r);
 
-				/* 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;
+	for (i = env->max_S; i > 0; --i) {
+		ir_tarval *div_res, *mod_res;
+		ir_tarval *tv = new_tarval_from_long((1 << i) + 1, env->mode);
 
-					highest_shift_wide = curr > highest_shift_wide ? curr : highest_shift_wide;
-					highest_shift_seq = curr_seq > highest_shift_seq ? curr_seq : highest_shift_seq;
+		div_res = tarval_divmod(N, tv, &mod_res);
+		if (mod_res == get_mode_null(env->mode)) {
+			unsigned char *Rs;
+			int rs;
 
-					last_shift = curr;
-				}
+			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 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
-					goto end;
-				}
+#define IMAX(a,b) ((a) > (b) ? (a) : (b))
 
-				/* If we have subs, we cannot do sequential. */
-				if(1 /* also_use_subs */) {
-					if(n > 0) {
-						ir_node *curr = NULL;
+/**
+ * basic decomposition routine
+ */
+static instruction *basic_decompose_mul(mul_env *env, unsigned char *R, int r, ir_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);
+	}
+}
 
-						i = n - 1;
+/**
+ * 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;
+	ir_graph *irg = env->irg;
+
+	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_r_Const_long(irg, env->shf_mode, inst->shift_count);
+		r = new_rd_Shl(env->dbg, env->blk, r, c, env->mode);
+		return inst->irn = new_rd_Add(env->dbg, env->blk, l, r, env->mode);
+	case SHIFT:
+		l = build_graph(env, inst->in[0]);
+		c = new_r_Const_long(irg, env->shf_mode, inst->shift_count);
+		return inst->irn = new_rd_Shl(env->dbg, 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, 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, env->blk, l, r, env->mode);
+	case ZERO:
+		return inst->irn = new_r_Const(irg, get_mode_null(env->mode));
+	default:
+		panic("Unsupported instruction kind");
+	}
+}
 
-						do {
-							int curr_shift = shifts[i];
-							int sub = curr_shift < 0;
-							int amount = abs(curr_shift) - 1;
-							ir_node *aux = operand;
+/**
+ * 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;
 
-							assert(amount >= 0 && "What is a negative shift??");
+	if (inst->costs >= 0) {
+		/* was already evaluated */
+		return 0;
+	}
 
-							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);
-							}
+	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, env->mode, NULL);
+		inst->costs = costs;
+		return costs;
+	case SHIFT:
+		if (inst->shift_count > env->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, env->mode, NULL);
+		inst->costs = costs;
+		return costs;
+	case ZERO:
+		inst->costs = costs = env->evaluate(inst->kind, env->mode, NULL);
+		return costs;
+	case MUL:
+	case ROOT:
+		break;
+	}
+	panic("Unsupported instruction kind");
+}
 
-							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;
+/**
+ * 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, ir_tarval *tv)
+{
+	mul_env       env;
+	instruction   *inst;
+	unsigned char *R;
+	int           r;
+	ir_node       *res = irn;
+	int           mul_costs;
+
+	obstack_init(&env.obst);
+	env.params   = be_get_backend_param()->dep_param;
+	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  = env.params->maximum_shifts;
+	env.evaluate = env.params->evaluate != NULL ? env.params->evaluate : default_evaluate;
+	env.irg      = get_irn_irg(irn);
+
+	R = value_to_condensed(&env, tv, &r);
+	inst = decompose_mul(&env, R, r, tv);
+
+	/* the paper suggests 70% here */
+	mul_costs = (env.evaluate(MUL, env.mode, 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;
+}
 
-						} while(--i >= 0);
+/* 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);
+	ir_graph  *irg;
+	ir_node   *left;
+	ir_node   *right;
+	ir_node   *operand;
+	ir_tarval *tv;
+	const ir_settings_arch_dep_t *params = be_get_backend_param()->dep_param;
 
-						res = curr;
-					}
-				}
+	/* 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 res;
+
+	assert(is_Mul(irn));
+	if (!mode_is_int(mode))
+		return res;
+
+	/* we should never do the reverse transformations again
+	   (like x+x -> 2*x) */
+	irg = get_irn_irg(irn);
+	add_irg_constraints(irg, IR_GRAPH_CONSTRAINT_ARCH_DEP);
+
+	left    = get_binop_left(irn);
+	right   = get_binop_right(irn);
+	tv      = NULL;
+	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;
+	}
 
-#ifdef DEB
-				{
-					const char *prefix = "";
-					for (i = 0; i < n; ++i) {
-						fprintf(stderr, "%s%d", prefix, shifts[i]);
-						prefix = ", ";
-					}
-					fprintf(stderr, "\n");
-				}
-#endif
+	/* multiplications with 0 are a special case which we leave for
+	 * equivalent_node_Mul because the code here can't handle them */
+	if (tv == get_mode_null(mode))
+		return res;
 
-			}
+	if (tv != NULL) {
+		res = do_decomposition(irn, operand, tv);
 
-end:
-			if(bitstr)
-				free(bitstr);
+		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);
-
 	return res;
 }
 
 /**
  * calculated the ld2 of a tarval if tarval is 2^n, else returns -1.
  */
-static int tv_ld2(tarval *tv, int bits) {
+static int tv_ld2(ir_tarval *tv, int bits)
+{
 	int i, k = 0, num;
 
 	for (num = i = 0; i < bits; ++i) {
@@ -386,7 +645,7 @@ 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)
@@ -397,7 +656,7 @@ static int tv_ld2(tarval *tv, int bits) {
 
 /** The result of a the magic() function. */
 struct ms {
-	tarval *M;        /**< magic number */
+	ir_tarval *M;     /**< magic number */
 	int s;            /**< shift amount */
 	int need_add;     /**< an additional add is needed */
 	int need_sub;     /**< an additional sub is needed */
@@ -408,15 +667,16 @@ struct ms {
  *
  * see Hacker's Delight: 10-6 Integer Division by Constants: Incorporation into a Compiler
  */
-static struct ms magic(tarval *d) {
+static struct ms magic(ir_tarval *d)
+{
 	ir_mode *mode   = get_tarval_mode(d);
 	ir_mode *u_mode = find_unsigned_mode(mode);
 	int bits        = get_mode_size_bits(u_mode);
 	int p;
-	tarval *ad, *anc, *delta, *q1, *r1, *q2, *r2, *t;     /* unsigned */
-	pn_Cmp d_cmp, M_cmp;
+	ir_tarval *ad, *anc, *delta, *q1, *r1, *q2, *r2, *t;     /* unsigned */
+	ir_relation d_cmp, M_cmp;
 
-	tarval *bits_minus_1, *two_bits_1;
+	ir_tarval *bits_minus_1, *two_bits_1;
 
 	struct ms mag;
 
@@ -443,7 +703,7 @@ static struct ms magic(tarval *d) {
 		q1 = ADD(q1, q1);                           /* Update q1 = 2^p/|nc| */
 		r1 = ADD(r1, r1);                           /* Update r1 = rem(2^p, |nc|) */
 
-		if (CMP(r1, anc) & pn_Cmp_Ge) {
+		if (CMP(r1, anc) & ir_relation_greater_equal) {
 			q1 = ADD(q1, ONE(u_mode));
 			r1 = SUB(r1, anc);
 		}
@@ -451,17 +711,17 @@ static struct ms magic(tarval *d) {
 		q2 = ADD(q2, q2);                           /* Update q2 = 2^p/|d| */
 		r2 = ADD(r2, r2);                           /* Update r2 = rem(2^p, |d|) */
 
-		if (CMP(r2, ad) & pn_Cmp_Ge) {
+		if (CMP(r2, ad) & ir_relation_greater_equal) {
 			q2 = ADD(q2, ONE(u_mode));
 			r2 = SUB(r2, ad);
 		}
 
 		delta = SUB(ad, r2);
-	} while (CMP(q1, delta) & pn_Cmp_Lt || (CMP(q1, delta) & pn_Cmp_Eq && CMP(r1, ZERO(u_mode)) & pn_Cmp_Eq));
+	} while (CMP(q1, delta) & ir_relation_less || (CMP(q1, delta) & ir_relation_equal && CMP(r1, ZERO(u_mode)) & ir_relation_equal));
 
 	d_cmp = CMP(d, ZERO(mode));
 
-	if (d_cmp & pn_Cmp_Ge)
+	if (d_cmp & ir_relation_greater_equal)
 		mag.M = ADD(CNV(q2, mode), ONE(mode));
 	else
 		mag.M = SUB(ZERO(mode), ADD(CNV(q2, mode), ONE(mode)));
@@ -471,10 +731,10 @@ static struct ms magic(tarval *d) {
 	mag.s = p - bits;
 
 	/* need an add if d > 0 && M < 0 */
-	mag.need_add = d_cmp & pn_Cmp_Gt && M_cmp & pn_Cmp_Lt;
+	mag.need_add = d_cmp & ir_relation_greater && M_cmp & ir_relation_less;
 
 	/* need a sub if d < 0 && M > 0 */
-	mag.need_sub = d_cmp & pn_Cmp_Lt && M_cmp & pn_Cmp_Gt;
+	mag.need_sub = d_cmp & ir_relation_less && M_cmp & ir_relation_greater;
 
 	tarval_set_integer_overflow_mode(rem);
 
@@ -483,7 +743,7 @@ static struct ms magic(tarval *d) {
 
 /** The result of the magicu() function. */
 struct mu {
-	tarval *M;        /**< magic add constant */
+	ir_tarval *M;     /**< magic add constant */
 	int s;            /**< shift amount */
 	int need_add;     /**< add indicator */
 };
@@ -493,12 +753,13 @@ struct mu {
  *
  * see Hacker's Delight: 10-10 Integer Division by Constants: Incorporation into a Compiler (Unsigned)
  */
-static struct mu magicu(tarval *d) {
+static struct mu magicu(ir_tarval *d)
+{
 	ir_mode *mode   = get_tarval_mode(d);
 	int bits        = get_mode_size_bits(mode);
 	int p;
-	tarval *nc, *delta, *q1, *r1, *q2, *r2;
-	tarval *bits_minus_1, *two_bits_1, *seven_ff;
+	ir_tarval *nc, *delta, *q1, *r1, *q2, *r2;
+	ir_tarval *bits_minus_1, *two_bits_1, *seven_ff;
 
 	struct mu magu;
 
@@ -521,7 +782,7 @@ static struct mu magicu(tarval *d) {
 
 	do {
 		++p;
-		if (CMP(r1, SUB(nc, r1)) & pn_Cmp_Ge) {
+		if (CMP(r1, SUB(nc, r1)) & ir_relation_greater_equal) {
 			q1 = ADD(ADD(q1, q1), ONE(mode));
 			r1 = SUB(ADD(r1, r1), nc);
 		}
@@ -530,15 +791,15 @@ static struct mu magicu(tarval *d) {
 			r1 = ADD(r1, r1);
 		}
 
-		if (CMP(ADD(r2, ONE(mode)), SUB(d, r2)) & pn_Cmp_Ge) {
-			if (CMP(q2, seven_ff) & pn_Cmp_Ge)
+		if (CMP(ADD(r2, ONE(mode)), SUB(d, r2)) & ir_relation_greater_equal) {
+			if (CMP(q2, seven_ff) & ir_relation_greater_equal)
 				magu.need_add = 1;
 
 			q2 = ADD(ADD(q2, q2), ONE(mode));
 			r2 = SUB(ADD(ADD(r2, r2), ONE(mode)), d);
 		}
 		else {
-			if (CMP(q2, two_bits_1) & pn_Cmp_Ge)
+			if (CMP(q2, two_bits_1) & ir_relation_greater_equal)
 				magu.need_add = 1;
 
 			q2 = ADD(q2, q2);
@@ -546,7 +807,7 @@ static struct mu magicu(tarval *d) {
 		}
 		delta = SUB(SUB(d, ONE(mode)), r2);
 	} while (p < 2*bits &&
-		(CMP(q1, delta) & pn_Cmp_Lt || (CMP(q1, delta) & pn_Cmp_Eq && CMP(r1, ZERO(mode)) & pn_Cmp_Eq)));
+		(CMP(q1, delta) & ir_relation_less || (CMP(q1, delta) & ir_relation_equal && CMP(r1, ZERO(mode)) & ir_relation_equal)));
 
 	magu.M = ADD(q2, ONE(mode));                       /* Magic number */
 	magu.s = p - bits;                                 /* and shift amount */
@@ -559,159 +820,182 @@ static struct mu magicu(tarval *d) {
 /**
  * Build the Mulh replacement code for n / tv.
  *
- * Note that 'div' might be a mod or DivMod operation as well
+ * Note that 'div' might be a Mod operation as well
  */
-static ir_node *replace_div_by_mulh(ir_node *div, tarval *tv) {
+static ir_node *replace_div_by_mulh(ir_node *div, ir_tarval *tv)
+{
 	dbg_info *dbg  = get_irn_dbg_info(div);
 	ir_node *n     = get_binop_left(div);
-	ir_node *block = get_irn_n(div, -1);
+	ir_node *block = get_nodes_block(div);
 	ir_mode *mode  = get_irn_mode(n);
 	int bits       = get_mode_size_bits(mode);
-	ir_node *q, *t, *c;
+	ir_node *q;
 
 	/* Beware: do not transform bad code */
 	if (is_Bad(n) || is_Bad(block))
 		return div;
 
 	if (mode_is_signed(mode)) {
+		ir_graph *irg = get_irn_irg(div);
 		struct ms mag = magic(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);
+		ir_node *c = new_r_Const(irg, mag.M);
+		ir_node *t;
+		q = new_rd_Mulh(dbg, block, n, c, mode);
 
 		/* do we need an Add or Sub */
 		if (mag.need_add)
-			q = new_rd_Add(dbg, current_ir_graph, block, q, n, mode);
+			q = new_rd_Add(dbg, block, q, n, mode);
 		else if (mag.need_sub)
-			q = new_rd_Sub(dbg, current_ir_graph, block, q, n, mode);
+			q = new_rd_Sub(dbg, block, q, n, mode);
 
 		/* 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_r_Const_long(irg, mode_Iu, mag.s);
+			q = new_rd_Shrs(dbg, block, q, c, mode);
 		}
 
 		/* final */
-		c = new_r_Const_long(current_ir_graph, block, mode_Iu, bits-1);
-		t = new_rd_Shr(dbg, current_ir_graph, block, q, c, mode);
+		c = new_r_Const_long(irg, mode_Iu, bits - 1);
+		t = new_rd_Shr(dbg, block, q, c, mode);
 
-		q = new_rd_Add(dbg, current_ir_graph, block, q, t, mode);
+		q = new_rd_Add(dbg, block, q, t, mode);
 	} else {
 		struct mu mag = magicu(tv);
-		ir_node *c;
+		ir_graph *irg = get_irn_irg(div);
 
 		/* 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);
+		ir_node *c = new_r_Const(irg, mag.M);
+		q = new_rd_Mulh(dbg, 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);
+				ir_node *t = new_rd_Sub(dbg, block, n, q, mode);
 
-				c = new_r_Const(current_ir_graph, block, mode_Iu, get_mode_one(mode_Iu));
-				t = new_rd_Shr(dbg, current_ir_graph, block, t, c, mode);
+				c = new_r_Const(irg, get_mode_one(mode_Iu));
+				t = new_rd_Shr(dbg, block, t, c, mode);
 
-				t = new_rd_Add(dbg, current_ir_graph, block, t, q, mode);
+				t = new_rd_Add(dbg, block, t, q, mode);
 
-				c = new_r_Const_long(current_ir_graph, block, mode_Iu, mag.s-1);
-				q = new_rd_Shr(dbg, current_ir_graph, block, t, c, mode);
+				c = new_r_Const_long(irg, mode_Iu, mag.s - 1);
+				q = new_rd_Shr(dbg, block, t, c, mode);
 			} else {
 				/* use the default scheme */
-				q = new_rd_Add(dbg, current_ir_graph, block, q, n, mode);
+				q = new_rd_Add(dbg, 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);
-			q = new_rd_Shr(dbg, current_ir_graph, block, q, c, mode);
+			c = new_r_Const_long(irg, mode_Iu, mag.s);
+			q = new_rd_Shr(dbg, block, q, c, mode);
 		}
 	}
 	return q;
 }
 
 /* Replace Divs with Shifts and Add/Subs and Mulh. */
-ir_node *arch_dep_replace_div_by_const(ir_node *irn) {
+ir_node *arch_dep_replace_div_by_const(ir_node *irn)
+{
+	const ir_settings_arch_dep_t *params = be_get_backend_param()->dep_param;
 	ir_node *res  = irn;
 
 	/* If the architecture dependent optimizations were not initialized
-	or this optimization was not enabled. */
+		or this optimization was not enabled. */
 	if (params == NULL || (opts & arch_dep_div_by_const) == 0)
 		return irn;
 
-	if (get_irn_opcode(irn) == iro_Div) {
-		ir_node *c = get_Div_right(irn);
-		ir_node *block, *left;
-		ir_mode *mode;
-		tarval *tv, *ntv;
-		dbg_info *dbg;
-		int n, bits;
-		int k, n_flag;
+	if (!is_Div(irn))
+		return irn;
 
-		if (get_irn_op(c) != op_Const)
-			return irn;
+	ir_node *c = get_Div_right(irn);
+	ir_node *block, *left;
+	ir_mode *mode;
+	ir_tarval *tv, *ntv;
+	dbg_info *dbg;
+	int n, bits;
+	int k;
+	int n_flag = 0;
 
-		tv = get_Const_tarval(c);
+	if (! is_Const(c))
+		return irn;
 
-		/* check for division by zero */
-		if (classify_tarval(tv) == TV_CLASSIFY_NULL)
-			return irn;
+	tv = get_Const_tarval(c);
 
-		left  = get_Div_left(irn);
-		mode  = get_irn_mode(left);
-		block = get_irn_n(irn, -1);
-		dbg   = get_irn_dbg_info(irn);
+	/* check for division by zero */
+	if (tarval_is_null(tv))
+		return irn;
 
-		bits = get_mode_size_bits(mode);
-		n    = (bits + 7) / 8;
+	left  = get_Div_left(irn);
+	mode  = get_irn_mode(left);
 
-		k = -1;
-		if (mode_is_signed(mode)) {
-			/* for signed divisions, the algorithm works for a / -2^k by negating the result */
-			ntv = tarval_neg(tv);
-			n_flag = 1;
-			k = tv_ld2(ntv, n);
-		}
+	/* can only handle integer Div's */
+	if (!mode_is_int(mode))
+		return irn;
 
-		if (k < 0) {
-			n_flag = 0;
-			k = tv_ld2(tv, n);
-		}
+	block = get_nodes_block(irn);
+	dbg   = get_irn_dbg_info(irn);
 
-		if (k >= 0) { /* division by 2^k or -2^k */
-			if (mode_is_signed(mode)) {
-				ir_node *k_node;
-				ir_node *curr = left;
+	bits = get_mode_size_bits(mode);
+	n    = (bits + 7) / 8;
 
-				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 = -1;
+	if (mode_is_signed(mode)) {
+		/* for signed divisions, the algorithm works for a / -2^k by negating the result */
+		ntv = tarval_neg(tv);
+		n_flag = 1;
+		k = tv_ld2(ntv, n);
+	}
 
-				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);
+	if (k < 0) {
+		n_flag = 0;
+		k = tv_ld2(tv, n);
+	}
 
-				curr   = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode);
+	if (k > 0) { /* division by 2^k or -2^k */
+		ir_graph *irg = get_irn_irg(irn);
+		if (mode_is_signed(mode)) {
+			ir_node *k_node;
+			ir_node *curr = 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);
+			/* create the correction code for signed values only if there might be a remainder */
+			if (! get_Div_no_remainder(irn)) {
+				if (k != 1) {
+					k_node = new_r_Const_long(irg, mode_Iu, k - 1);
+					curr   = new_rd_Shrs(dbg, block, left, k_node, mode);
+				}
 
-				if (n_flag) { /* negate the result */
-					ir_node *k_node;
+				k_node = new_r_Const_long(irg, mode_Iu, bits - k);
+				curr   = new_rd_Shr(dbg, block, curr, k_node, mode);
+				/* curr is now 2^(k-1) in case left <  0
+				 *          or       0 in case left >= 0
+				 *
+				 * For an example, where this fixup is necessary consider -3 / 2,
+				 * which should compute to -1,
+				 * but simply shifting right by one computes -2.
+				 */
+
+				curr   = new_rd_Add(dbg, block, left, curr, mode);
+			}
 
-					k_node = new_r_Const(current_ir_graph, block, mode, 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(irg, mode_Iu, k);
+			res    = new_rd_Shrs(dbg, block, curr, k_node, mode);
 
-				k_node = new_r_Const_long(current_ir_graph, block, mode_Iu, k);
-				res    = new_rd_Shr(dbg, current_ir_graph, block, left, k_node, mode);
+			if (n_flag) { /* negate the result */
+				k_node = new_r_Const(irg, get_mode_null(mode));
+				res = new_rd_Sub(dbg, block, k_node, res, mode);
 			}
-		} else {
-			/* other constant */
-			if (allow_Mulh(mode))
-				res = replace_div_by_mulh(irn, tv);
+		} else {      /* unsigned case */
+			ir_node *k_node;
+
+			k_node = new_r_Const_long(irg, mode_Iu, k);
+			res    = new_rd_Shr(dbg, block, left, k_node, mode);
 		}
+	} else if (k != 0) {
+		/* other constant */
+		if (allow_Mulh(params, mode))
+			res = replace_div_by_mulh(irn, tv);
+	} else { /* k == 0  i.e. division by 1 */
+		res = left;
 	}
 
 	if (res != irn)
@@ -721,7 +1005,9 @@ ir_node *arch_dep_replace_div_by_const(ir_node *irn) {
 }
 
 /* Replace Mods with Shifts and Add/Subs and Mulh. */
-ir_node *arch_dep_replace_mod_by_const(ir_node *irn) {
+ir_node *arch_dep_replace_mod_by_const(ir_node *irn)
+{
+	const ir_settings_arch_dep_t *params = be_get_backend_param()->dep_param;
 	ir_node *res  = irn;
 
 	/* If the architecture dependent optimizations were not initialized
@@ -729,27 +1015,27 @@ 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;
-		tarval *tv, *ntv;
+		ir_tarval *tv, *ntv;
 		dbg_info *dbg;
 		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);
 		mode  = get_irn_mode(left);
-		block = get_irn_n(irn, -1);
+		block = get_nodes_block(irn);
 		dbg   = get_irn_dbg_info(irn);
 		bits = get_mode_size_bits(mode);
 		n    = (bits + 7) / 8;
@@ -765,7 +1051,14 @@ ir_node *arch_dep_replace_mod_by_const(ir_node *irn) {
 			k = tv_ld2(tv, n);
 		}
 
-		if (k >= 0) {
+		/* k == 0  i.e. modulo by 1 */
+		if (k == 0) {
+			ir_graph *irg = get_irn_irg(irn);
+
+			res = new_r_Const(irg, get_mode_null(mode));
+		}
+		else if (k > 0) {
+			ir_graph *irg = get_irn_irg(irn);
 			/* division by 2^k or -2^k:
 			 * we use "modulus" here, so x % y == x % -y that's why is no difference between the case 2^k and -2^k
 			 */
@@ -774,35 +1067,35 @@ 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);
-					curr   = new_rd_Shrs(dbg, current_ir_graph, block, left, k_node, mode);
+					k_node = new_r_Const_long(irg, mode_Iu, k - 1);
+					curr   = new_rd_Shrs(dbg, 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(irg, mode_Iu, bits - k);
+				curr   = new_rd_Shr(dbg, block, curr, k_node, mode);
 
-				curr   = new_rd_Add(dbg, current_ir_graph, block, left, curr, mode);
+				curr   = new_rd_Add(dbg, block, left, curr, mode);
 
-				k_node = new_r_Const_long(current_ir_graph, block, mode, (-1) << k);
-				curr   = new_rd_And(dbg, current_ir_graph, block, curr, k_node, mode);
+				k_node = new_r_Const_long(irg, mode, (-1) << k);
+				curr   = new_rd_And(dbg, block, curr, k_node, mode);
 
-				res    = new_rd_Sub(dbg, current_ir_graph, block, left, curr, mode);
+				res    = new_rd_Sub(dbg, 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);
-				res    = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode);
+				k_node = new_r_Const_long(irg, mode, (1 << k) - 1);
+				res    = new_rd_And(dbg, block, left, k_node, mode);
 			}
 		} else {
 			/* other constant */
-			if (allow_Mulh(mode)) {
+			if (allow_Mulh(params, mode)) {
 				res = replace_div_by_mulh(irn, tv);
 
-				res = new_rd_Mul(dbg, current_ir_graph, block, res, c, mode);
+				res = new_rd_Mul(dbg, block, res, c, mode);
 
 				/* res = arch_dep_mul_to_shift(res); */
 
-				res = new_rd_Sub(dbg, current_ir_graph, block, left, res, mode);
+				res = new_rd_Sub(dbg, block, left, res, mode);
 			}
 		}
 	}
@@ -812,127 +1105,3 @@ ir_node *arch_dep_replace_mod_by_const(ir_node *irn) {
 
 	return res;
 }
-
-/* Replace DivMods with Shifts and Add/Subs and Mulh. */
-void arch_dep_replace_divmod_by_const(ir_node **div, ir_node **mod, ir_node *irn) {
-	*div = *mod = NULL;
-
-	/* If the architecture dependent optimizations were not initialized
-	   or this optimization was not enabled. */
-	if (params == NULL ||
-		((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) {
-		ir_node *c = get_DivMod_right(irn);
-		ir_node *block, *left;
-		ir_mode *mode;
-		tarval *tv, *ntv;
-		dbg_info *dbg;
-		int n, bits;
-		int k, n_flag;
-
-		if (get_irn_op(c) != op_Const)
-			return;
-
-		tv = get_Const_tarval(c);
-
-		/* check for division by zero */
-		if (classify_tarval(tv) == TV_CLASSIFY_NULL)
-			return;
-
-		left  = get_DivMod_left(irn);
-		mode  = get_irn_mode(left);
-		block = get_irn_n(irn, -1);
-		dbg   = get_irn_dbg_info(irn);
-
-		bits = get_mode_size_bits(mode);
-		n    = (bits + 7) / 8;
-
-		k = -1;
-		if (mode_is_signed(mode)) {
-			/* for signed divisions, the algorithm works for a / -2^k by negating the result */
-			ntv = tarval_neg(tv);
-			n_flag = 1;
-			k = tv_ld2(ntv, n);
-		}
-
-		if (k < 0) {
-			n_flag = 0;
-			k = tv_ld2(tv, n);
-		}
-
-		if (k >= 0) { /* division by 2^k or -2^k */
-			if (mode_is_signed(mode)) {
-				ir_node *k_node, *c_k;
-				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);
-				}
-
-				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);
-
-				c_k    = new_r_Const_long(current_ir_graph, block, 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));
-					*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);
-				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);
-				*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);
-				*mod   = new_rd_And(dbg, current_ir_graph, block, left, k_node, mode);
-			}
-		} else {
-			/* other constant */
-			if (allow_Mulh(mode)) {
-				ir_node *t;
-
-				*div = replace_div_by_mulh(irn, tv);
-
-				t    = new_rd_Mul(dbg, current_ir_graph, block, *div, c, mode);
-
-				/* t = arch_dep_mul_to_shift(t); */
-
-				*mod = new_rd_Sub(dbg, current_ir_graph, block, left, t, mode);
-			}
-		}
-	}
-
-	if (*div)
-		hook_arch_dep_replace_division_by_const(irn);
-}
-
-
-static const ir_settings_arch_dep_t default_params = {
-	1,  /* also use subs */
-	4,  /* maximum shifts */
-	31, /* maximum shift amount */
-
-	0,  /* allow Mulhs */
-	0,  /* allow Mulus */
-	32  /* Mulh allowed up to 32 bit */
-};
-
-/* A default parameter factory for testing purposes. */
-const ir_settings_arch_dep_t *arch_dep_default_factory(void) {
-	return &default_params;
-}