X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;ds=inline;f=ir%2Ftv%2Ffltcalc.c;h=973faf7452ff1b9fcf0e4a0442a08b59cae8e611;hb=b5bfe78cd888b9349d3dffd9e103da389a54ed09;hp=b38d5b916eff6f792d23bbdfcf24049524288f83;hpb=0bb43c316272b3a0e325a4c6eed92fed8f6e3bee;p=libfirm

diff --git a/ir/tv/fltcalc.c b/ir/tv/fltcalc.c
index b38d5b916..973faf745 100644
--- a/ir/tv/fltcalc.c
+++ b/ir/tv/fltcalc.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.
  *
@@ -25,9 +25,7 @@
  * @version  $Id$
  */
 
-#ifdef HAVE_CONFIG_H
-# include "config.h"
-#endif
+#include "config.h"
 
 #include "fltcalc.h"
 #include "strcalc.h"
@@ -41,17 +39,16 @@
 #ifdef HAVE_INTTYPES_H
 # include <inttypes.h>
 #endif
-#ifdef HAVE_STRING_H
-# include <string.h>
-#endif
-#ifdef HAVE_STDLIB_H
-# include <stdlib.h>
-#endif
+#include <string.h>
+#include <stdlib.h>
 #include <stdio.h>
 #include <assert.h>
 
 #include "xmalloc.h"
 
+/** The number of extra precision rounding bits */
+#define ROUNDING_BITS 2
+
 typedef uint32_t UINT32;
 
 #ifdef HAVE_LONG_DOUBLE
@@ -94,29 +91,11 @@ typedef union {
 #endif
 #endif
 
-/**
- * possible float states
- */
-typedef enum {
-	NORMAL,       /**< normal representation, implicit 1 */
-	ZERO,         /**< +/-0 */
-	SUBNORMAL,    /**< denormals, implicit 0 */
-	INF,          /**< +/-oo */
-	NAN,          /**< Not A Number */
-} value_class_t;
-
-/** A descriptor for an IEEE float value. */
-typedef struct {
-	unsigned char exponent_size;    /**< size of exponent in bits */
-	unsigned char mantissa_size;    /**< size of mantissa in bits */
-	value_class_t clss;             /**< state of this float */
-} descriptor_t;
-
 #define CLEAR_BUFFER(buffer) memset(buffer, 0, calc_buffer_size)
 
 /* our floating point value */
 struct _fp_value {
-	descriptor_t desc;
+	ieee_descriptor_t desc;
 	char sign;
 	char value[1];			/* exp[value_size] + mant[value_size] */
 };
@@ -125,8 +104,8 @@ struct _fp_value {
 #define _mant(a) &((a)->value[value_size])
 
 #define _save_result(x) memcpy((x), sc_get_buffer(), value_size)
-#define _shift_right(x, y, b) sc_shr((x), (y), value_size*4, 0, (b))
-#define _shift_left(x, y, b) sc_shl((x), (y), value_size*4, 0, (b))
+#define _shift_right(x, y, res) sc_shr((x), (y), value_size*4, 0, (res))
+#define _shift_left(x, y, res) sc_shl((x), (y), value_size*4, 0, (res))
 
 
 #ifdef FLTCALC_DEBUG
@@ -172,23 +151,24 @@ static void fail_char(const char *str, unsigned int len, int pos) {
 
 /** pack machine-like */
 static void *pack(const fp_value *int_float, void *packed) {
-	char *shift_val;
-	char *temp;
+	char     *shift_val;
+	char     *temp;
 	fp_value *val_buffer;
+	int      pos;
 
-	temp = alloca(value_size);
+	temp      = alloca(value_size);
 	shift_val = alloca(value_size);
 
-	switch (int_float->desc.clss) {
+	switch ((value_class_t)int_float->desc.clss) {
 	case NAN:
 		val_buffer = alloca(calc_buffer_size);
-		fc_get_qnan(int_float->desc.exponent_size, int_float->desc.mantissa_size, val_buffer);
+		fc_get_qnan(&int_float->desc, val_buffer);
 		int_float = val_buffer;
 		break;
 
 	case INF:
 		val_buffer = alloca(calc_buffer_size);
-		fc_get_plusinf(int_float->desc.exponent_size, int_float->desc.mantissa_size, val_buffer);
+		fc_get_plusinf(&int_float->desc, val_buffer);
 		val_buffer->sign = int_float->sign;
 		int_float = val_buffer;
 		break;
@@ -196,29 +176,33 @@ static void *pack(const fp_value *int_float, void *packed) {
 	default:
 		break;
 	}
-	/* pack sign */
+	assert(int_float->desc.explicit_one <= 1);
+
+	/* pack sign: move it to the left after exponent AND mantissa */
 	sc_val_from_ulong(int_float->sign, temp);
 
-	sc_val_from_ulong(int_float->desc.exponent_size + int_float->desc.mantissa_size, NULL);
+	pos = int_float->desc.exponent_size + int_float->desc.mantissa_size + int_float->desc.explicit_one;
+	sc_val_from_ulong(pos, NULL);
 	_shift_left(temp, sc_get_buffer(), packed);
 
-	/* extract exponent */
-	sc_val_from_ulong(int_float->desc.mantissa_size, shift_val);
-
+	/* pack exponent: move it to the left after mantissa */
+	pos = int_float->desc.mantissa_size + int_float->desc.explicit_one;
+	sc_val_from_ulong(pos, shift_val);
 	_shift_left(_exp(int_float), shift_val, temp);
 
+	/* combine sign|exponent */
 	sc_or(temp, packed, packed);
 
 	/* extract mantissa */
-	/* remove 2 rounding bits */
-	sc_val_from_ulong(2, shift_val);
+	/* remove rounding bits */
+	sc_val_from_ulong(ROUNDING_BITS, shift_val);
 	_shift_right(_mant(int_float), shift_val, temp);
 
 	/* remove leading 1 (or 0 if denormalized) */
-	sc_max_from_bits(int_float->desc.mantissa_size, 0, shift_val); /* all mantissa bits are 1's */
+	sc_max_from_bits(pos, 0, shift_val); /* all mantissa bits are 1's */
 	sc_and(temp, shift_val, temp);
 
-	/* save result */
+	/* combine sign|exponent|mantissa */
 	sc_or(temp, packed, packed);
 
 	return packed;
@@ -235,8 +219,8 @@ static int normalize(const fp_value *in_val, fp_value *out_val, int sticky) {
 	char lsb, guard, round, round_dir = 0;
 	char *temp = alloca(value_size);
 
-	/* +2: save two rounding bits at the end */
-	hsb = 2 + in_val->desc.mantissa_size - sc_get_highest_set_bit(_mant(in_val)) - 1;
+	/* save rounding bits at the end */
+	hsb = ROUNDING_BITS + in_val->desc.mantissa_size - sc_get_highest_set_bit(_mant(in_val)) - 1;
 
 	if (in_val != out_val)   {
 		out_val->sign = in_val->sign;
@@ -246,7 +230,7 @@ static int normalize(const fp_value *in_val, fp_value *out_val, int sticky) {
 	out_val->desc.clss = NORMAL;
 
 	/* mantissa all zeros, so zero exponent (because of explicit one) */
-	if (hsb == 2 + in_val->desc.mantissa_size)   {
+	if (hsb == ROUNDING_BITS + in_val->desc.mantissa_size)   {
 		sc_val_from_ulong(0, _exp(out_val));
 		hsb = -1;
 	}
@@ -295,7 +279,7 @@ static int normalize(const fp_value *in_val, fp_value *out_val, int sticky) {
 	/* perform rounding by adding a value that clears the guard bit and the round bit
 	 * and either causes a carry to round up or not */
 	/* get the last 3 bits of the value */
-	lsb = sc_sub_bits(_mant(out_val), out_val->desc.mantissa_size + 2, 0) & 0x7;
+	lsb = sc_sub_bits(_mant(out_val), out_val->desc.mantissa_size + ROUNDING_BITS, 0) & 0x7;
 	guard = (lsb&0x2)>>1;
 	round = lsb&0x1;
 
@@ -339,7 +323,7 @@ static int normalize(const fp_value *in_val, fp_value *out_val, int sticky) {
 		out_val->desc.clss = ZERO;
 
 	/* check for rounding overflow */
-	hsb = 2 + out_val->desc.mantissa_size - sc_get_highest_set_bit(_mant(out_val)) - 1;
+	hsb = ROUNDING_BITS + out_val->desc.mantissa_size - sc_get_highest_set_bit(_mant(out_val)) - 1;
 	if ((out_val->desc.clss != SUBNORMAL) && (hsb < -1)) {
 		sc_val_from_ulong(1, temp);
 		_shift_right(_mant(out_val), temp, _mant(out_val));
@@ -389,7 +373,7 @@ static int normalize(const fp_value *in_val, fp_value *out_val, int sticky) {
 
 			case FC_TONEGATIVE:
 			case FC_TOZERO:
-				fc_get_max(out_val->desc.exponent_size, out_val->desc.mantissa_size, out_val);
+				fc_get_max(&out_val->desc, out_val);
 			}
 		} else {
 			/* value is negative */
@@ -401,7 +385,7 @@ static int normalize(const fp_value *in_val, fp_value *out_val, int sticky) {
 
 			case FC_TOPOSITIVE:
 			case FC_TOZERO:
-				fc_get_min(out_val->desc.exponent_size, out_val->desc.mantissa_size, out_val);
+				fc_get_min(&out_val->desc, out_val);
 			}
 		}
 	}
@@ -409,16 +393,19 @@ static int normalize(const fp_value *in_val, fp_value *out_val, int sticky) {
 }
 
 /**
- * Operations involving NaN's must return NaN
+ * Operations involving NaN's must return NaN.
+ * They are NOT exact.
  */
 #define handle_NAN(a, b, result) \
 do {                                                      \
   if (a->desc.clss == NAN) {                              \
     if (a != result) memcpy(result, a, calc_buffer_size); \
+    fc_exact = 0;                                         \
     return;                                               \
   }                                                       \
   if (b->desc.clss == NAN) {                              \
     if (b != result) memcpy(result, b, calc_buffer_size); \
+    fc_exact = 0;                                         \
     return;                                               \
   }                                                       \
 }while (0)
@@ -440,14 +427,15 @@ static void _fadd(const fp_value *a, const fp_value *b, fp_value *result) {
 
 	/* make sure result has a descriptor */
 	if (result != a && result != b)
-		memcpy(&result->desc, &a->desc, sizeof(descriptor_t));
+		result->desc = a->desc;
 
 	/* determine if this is an addition or subtraction */
 	sign = a->sign ^ b->sign;
 
 	/* produce NaN on inf - inf */
 	if (sign && (a->desc.clss == INF) && (b->desc.clss == INF)) {
-		fc_get_qnan(a->desc.exponent_size, b->desc.mantissa_size, result);
+		fc_exact = 0;
+		fc_get_qnan(&a->desc, result);
 		return;
 	}
 
@@ -486,12 +474,14 @@ static void _fadd(const fp_value *a, const fp_value *b, fp_value *result) {
 	if (a->desc.clss == ZERO || b->desc.clss == INF) {
 		if (b != result)
 			memcpy(result, b, calc_buffer_size);
+		fc_exact = b->desc.clss == NORMAL;
 		result->sign = res_sign;
 		return;
 	}
 	if (b->desc.clss == ZERO || a->desc.clss == INF) {
 		if (a != result)
 			memcpy(result, a, calc_buffer_size);
+		fc_exact = a->desc.clss == NORMAL;
 		result->sign = res_sign;
 		return;
 	}
@@ -554,15 +544,16 @@ static void _fmul(const fp_value *a, const fp_value *b, fp_value *result) {
 	temp = alloca(value_size);
 
 	if (result != a && result != b)
-		memcpy(&result->desc, &a->desc, sizeof(descriptor_t));
+		result->desc = a->desc;
 
 	result->sign = res_sign = a->sign ^ b->sign;
 
 	/* produce NaN on 0 * inf */
 	if (a->desc.clss == ZERO) {
-		if (b->desc.clss == INF)
-			fc_get_qnan(a->desc.exponent_size, a->desc.mantissa_size, result);
-		else {
+		if (b->desc.clss == INF) {
+			fc_get_qnan(&a->desc, result);
+			fc_exact = 0;
+		} else {
 			if (a != result)
 				memcpy(result, a, calc_buffer_size);
 			result->sign = res_sign;
@@ -570,9 +561,10 @@ static void _fmul(const fp_value *a, const fp_value *b, fp_value *result) {
 		return;
 	}
 	if (b->desc.clss == ZERO) {
-		if (a->desc.clss == INF)
-			fc_get_qnan(a->desc.exponent_size, a->desc.mantissa_size, result);
-		else {
+		if (a->desc.clss == INF) {
+			fc_get_qnan(&a->desc, result);
+			fc_exact = 0;
+		} else {
 			if (b != result)
 				memcpy(result, b, calc_buffer_size);
 			result->sign = res_sign;
@@ -581,12 +573,14 @@ static void _fmul(const fp_value *a, const fp_value *b, fp_value *result) {
 	}
 
 	if (a->desc.clss == INF) {
+		fc_exact = 0;
 		if (a != result)
 			memcpy(result, a, calc_buffer_size);
 		result->sign = res_sign;
 		return;
 	}
 	if (b->desc.clss == INF) {
+		fc_exact = 0;
 		if (b != result)
 			memcpy(result, b, calc_buffer_size);
 		result->sign = res_sign;
@@ -611,8 +605,8 @@ static void _fmul(const fp_value *a, const fp_value *b, fp_value *result) {
 	 * point are the sum of the factors' digits after the radix point. As all
 	 * values are normalized they both have the same amount of these digits,
 	 * which has to be restored by proper shifting
-	 * +2 because of the two rounding bits */
-	sc_val_from_ulong(2 + result->desc.mantissa_size, temp);
+	 * because of the rounding bits */
+	sc_val_from_ulong(ROUNDING_BITS + result->desc.mantissa_size, temp);
 
 	_shift_right(_mant(result), temp, _mant(result));
 	sticky = sc_had_carry();
@@ -637,16 +631,17 @@ static void _fdiv(const fp_value *a, const fp_value *b, fp_value *result) {
 	dividend = alloca(value_size);
 
 	if (result != a && result != b)
-		memcpy(&result->desc, &a->desc, sizeof(descriptor_t));
+		result->desc = a->desc;
 
 	result->sign = res_sign = a->sign ^ b->sign;
 
 	/* produce NAN on 0/0 and inf/inf */
 	if (a->desc.clss == ZERO) {
-		if (b->desc.clss == ZERO)
-			/* 0/0 -> nan */
-			fc_get_qnan(a->desc.exponent_size, a->desc.mantissa_size, result);
-		else {
+		if (b->desc.clss == ZERO) {
+			/* 0/0 -> NaN */
+			fc_get_qnan(&a->desc, result);
+			fc_exact = 0;
+		} else {
 			/* 0/x -> a */
 			if (a != result)
 				memcpy(result, a, calc_buffer_size);
@@ -656,10 +651,11 @@ static void _fdiv(const fp_value *a, const fp_value *b, fp_value *result) {
 	}
 
 	if (b->desc.clss == INF) {
-		if (a->desc.clss == INF)
-			/* inf/inf -> nan */
-			fc_get_qnan(a->desc.exponent_size, a->desc.mantissa_size, result);
-		else {
+		fc_exact = 0;
+		if (a->desc.clss == INF) {
+			/* inf/inf -> NaN */
+			fc_get_qnan(&a->desc, result);
+		} else {
 			/* x/inf -> 0 */
 			sc_val_from_ulong(0, NULL);
 			_save_result(_exp(result));
@@ -670,6 +666,7 @@ static void _fdiv(const fp_value *a, const fp_value *b, fp_value *result) {
 	}
 
 	if (a->desc.clss == INF) {
+		fc_exact = 0;
 		/* inf/x -> inf */
 		if (a != result)
 			memcpy(result, a, calc_buffer_size);
@@ -677,11 +674,12 @@ static void _fdiv(const fp_value *a, const fp_value *b, fp_value *result) {
 		return;
 	}
 	if (b->desc.clss == ZERO) {
+		fc_exact = 0;
 		/* division by zero */
 		if (result->sign)
-			fc_get_minusinf(a->desc.exponent_size, a->desc.mantissa_size, result);
+			fc_get_minusinf(&a->desc, result);
 		else
-			fc_get_plusinf(a->desc.exponent_size, a->desc.mantissa_size, result);
+			fc_get_plusinf(&a->desc, result);
 		return;
 	}
 
@@ -702,7 +700,7 @@ static void _fdiv(const fp_value *a, const fp_value *b, fp_value *result) {
 	 * fit into the integer precision, but due to the rounding bits (which
 	 * are always zero because the values are all normalized) the divisor
 	 * can be shifted right instead to achieve the same result */
-	sc_val_from_ulong(2 + result->desc.mantissa_size, temp);
+	sc_val_from_ulong(ROUNDING_BITS + result->desc.mantissa_size, temp);
 
 	_shift_left(_mant(a), temp, dividend);
 
@@ -719,7 +717,7 @@ static void _fdiv(const fp_value *a, const fp_value *b, fp_value *result) {
 }
 
 #if 0
-static void _power_of_ten(int exp, descriptor_t *desc, char *result) {
+static void _power_of_ten(int exp, ieee_descriptor_t *desc, char *result) {
 	char *build;
 	char *temp;
 
@@ -728,12 +726,12 @@ static void _power_of_ten(int exp, descriptor_t *desc, char *result) {
 
 	/* set new descriptor (else result is supposed to already have one) */
 	if (desc != NULL)
-		memcpy(&result->desc, desc, sizeof(descriptor_t));
+		result->desc = *desc;
 
 	build = alloca(value_size);
 	temp = alloca(value_size);
 
-	sc_val_from_ulong((1 << result->desc.exponent_size)/2-1, _exp(result));
+	sc_val_from_ulong((1 << (result->desc.exponent_size - 1)) - 1, _exp(result));
 
 	if (exp > 0) {
 		/* temp is value of ten now */
@@ -747,7 +745,7 @@ static void _power_of_ten(int exp, descriptor_t *desc, char *result) {
 		_save_result(build);
 
 		/* temp is amount of left shift needed to put the value left of the radix point */
-		sc_val_from_ulong(result->desc.mantissa_size + 2, temp);
+		sc_val_from_ulong(result->desc.mantissa_size + ROUNDING_BITS, temp);
 
 		_shift_left(build, temp, _mant(result));
 
@@ -783,7 +781,7 @@ static void _trunc(const fp_value *a, fp_value *result) {
 	temp = alloca(value_size);
 
 	if (a != result)
-		memcpy(&result->desc, &a->desc, sizeof(descriptor_t));
+		result->desc = a->desc;
 
 	exp_bias = (1 << (a->desc.exponent_size - 1)) - 1;
 	exp_val  = sc_val_to_long(_exp(a)) - exp_bias;
@@ -813,9 +811,10 @@ static void _trunc(const fp_value *a, fp_value *result) {
 	/* and the mask and return the result */
 	sc_and(_mant(a), temp, _mant(result));
 
-	if (a != result) memcpy(_exp(result), _exp(a), value_size);
-
-	return;
+	if (a != result) {
+		memcpy(_exp(result), _exp(a), value_size);
+		result->sign = a->sign;
+	}
 }
 
 /********
@@ -829,7 +828,7 @@ int fc_get_buffer_length(void) {
 	return calc_buffer_size;
 }
 
-void *fc_val_from_str(const char *str, unsigned int len, char exp_size, char mant_size, void *result) {
+void *fc_val_from_str(const char *str, unsigned int len, const ieee_descriptor_t *desc, void *result) {
 #if 0
 	enum {
 		START,
@@ -855,9 +854,10 @@ void *fc_val_from_str(const char *str, unsigned int len, char exp_size, char man
 	power_val = alloca(calc_buffer_size);
 	mant_str = alloca((len)?(len):(strlen(str)));
 
-	result->desc.exponent_size = exp_size;
-	result->desc.mantissa_size = mant_size;
-	result->desc.clss = NORMAL;
+	result->desc.exponent_size = desc->exponent_size;
+	result->desc.mantissa_size = desc->mantissa_size;
+	result->desc.explicit_one  = desc->explicit_one;
+	result->desc.clss          = NORMAL;
 
 	old_str = str;
 	pos = 0;
@@ -991,11 +991,11 @@ done:
 	sc_val_from_str(mant_str, strlen(mant_str), _mant(result));
 
 	/* shift to put value left of radix point */
-	sc_val_from_ulong(mant_size + 2, exp_val);
+	sc_val_from_ulong(mant_size + ROUNDING_BITS, exp_val);
 
 	_shift_left(_mant(result), exp_val, _mant(result));
 
-	sc_val_from_ulong((1 << exp_size)/2-1, _exp(result));
+	sc_val_from_ulong((1 << (exp_size - 1)) - 1, _exp(result));
 
 	_normalize(result, result, 0);
 
@@ -1010,34 +1010,44 @@ done:
 	return result;
 #else
 	/* XXX excuse of an implementation to make things work */
-	LLDBL val;
-	fp_value *tmp = alloca(calc_buffer_size);
+	LLDBL             val;
+	fp_value          *tmp = alloca(calc_buffer_size);
+	ieee_descriptor_t tmp_desc;
 	(void) len;
 
-#ifdef HAVE_LONG_DOUBLE
+#if defined(HAVE_LONG_DOUBLE) && !defined(__CYGWIN__)
 	val = strtold(str, NULL);
 	DEBUGPRINTF(("val_from_str(%s)\n", str));
-	fc_val_from_ieee754(val, 15, 64, tmp);
+	tmp_desc.exponent_size = 15;
+	tmp_desc.mantissa_size = 63;
+	tmp_desc.explicit_one  = 1;
+	tmp_desc.clss          = NORMAL;
+	fc_val_from_ieee754(val, &tmp_desc, tmp);
 #else
 	val = strtod(str, NULL);
 	DEBUGPRINTF(("val_from_str(%s)\n", str));
-	fc_val_from_ieee754(val, 11, 52, tmp);
+	tmp_desc.exponent_size = 11;
+	tmp_desc.mantissa_size = 52;
+	tmp_desc.explicit_one  = 0;
+	tmp_desc.clss          = NORMAL;
+	fc_val_from_ieee754(val, &tmp_desc, tmp);
 #endif /* HAVE_LONG_DOUBLE */
-	return fc_cast(tmp, exp_size, mant_size, result);
+	return fc_cast(tmp, desc, result);
 #endif
 }
 
-fp_value *fc_val_from_ieee754(LLDBL l, char exp_size, char mant_size, fp_value *result) {
-	char *temp;
-	int bias_res, bias_val, mant_val;
+fp_value *fc_val_from_ieee754(LLDBL l, const ieee_descriptor_t *desc, fp_value *result) {
+	char    *temp;
+	int     bias_res, bias_val, mant_val;
 	value_t srcval;
-	UINT32 sign, exponent, mantissa0, mantissa1;
+	char    sign;
+	UINT32  exponent, mantissa0, mantissa1;
 
 	srcval.d = l;
-	bias_res = ((1<<exp_size)/2-1);
+	bias_res = ((1 << (desc->exponent_size - 1)) - 1);
 
 #ifdef HAVE_LONG_DOUBLE
-	mant_val  = 64;
+	mant_val  = 63;
 	bias_val  = 0x3fff;
 	sign      = (srcval.val.high & 0x00008000) != 0;
 	exponent  = (srcval.val.high & 0x00007FFF) ;
@@ -1063,13 +1073,17 @@ fp_value *fc_val_from_ieee754(LLDBL l, char exp_size, char mant_size, fp_value *
 	if (result == NULL) result = calc_buffer;
 	temp = alloca(value_size);
 
-	result->desc.exponent_size = exp_size;
-	result->desc.mantissa_size = mant_size;
+	/* CLEAR the buffer, else some bits might be uninitialized */
+	memset(result, 0, fc_get_buffer_length());
+
+	result->desc.exponent_size = desc->exponent_size;
+	result->desc.mantissa_size = desc->mantissa_size;
+	result->desc.explicit_one  = desc->explicit_one;
 
 	/* extract sign */
 	result->sign = sign;
 
-	/* sign and flag suffice to identify nan or inf, no exponent/mantissa
+	/* sign and flag suffice to identify NaN or inf, no exponent/mantissa
 	 * encoding is needed. the function can return immediately in these cases */
 	if (isnan(l)) {
 		result->desc.clss = NAN;
@@ -1086,23 +1100,16 @@ fp_value *fc_val_from_ieee754(LLDBL l, char exp_size, char mant_size, fp_value *
 	 * this looks more complicated than it is: unbiased input exponent + output bias,
 	 * minus the mantissa difference which is added again later when the output float
 	 * becomes normalized */
-#ifdef HAVE_EXPLICIT_ONE
-	sc_val_from_long((exponent-bias_val+bias_res)-(mant_val-mant_size-1), _exp(result));
-#else
-	sc_val_from_long((exponent-bias_val+bias_res)-(mant_val-mant_size), _exp(result));
-#endif
+	sc_val_from_long((exponent - bias_val + bias_res) - (mant_val - desc->mantissa_size), _exp(result));
 
 	/* build mantissa representation */
-#ifndef HAVE_EXPLICIT_ONE
 	if (exponent != 0) {
 		/* insert the hidden bit */
 		sc_val_from_ulong(1, temp);
-		sc_val_from_ulong(mant_val + 2, NULL);
+		sc_val_from_ulong(mant_val + ROUNDING_BITS, NULL);
 		_shift_left(temp, sc_get_buffer(), NULL);
 	}
-	else
-#endif
-	{
+	else {
 		sc_val_from_ulong(0, NULL);
 	}
 
@@ -1116,7 +1123,7 @@ fp_value *fc_val_from_ieee754(LLDBL l, char exp_size, char mant_size, fp_value *
 
 	/* bits from the lower word */
 	sc_val_from_ulong(mantissa1, temp);
-	sc_val_from_ulong(2, NULL);
+	sc_val_from_ulong(ROUNDING_BITS, NULL);
 	_shift_left(temp, sc_get_buffer(), temp);
 	sc_or(_mant(result), temp, _mant(result));
 
@@ -1145,22 +1152,25 @@ LLDBL fc_val_to_ieee754(const fp_value *val) {
 	UINT32 mantissa0;
 	UINT32 mantissa1;
 
-	value_t buildval;
+	value_t           buildval;
+	ieee_descriptor_t desc;
+	unsigned          mantissa_size;
 
 #ifdef HAVE_LONG_DOUBLE
-	char result_exponent = 15;
-	char result_mantissa = 64;
+	desc.exponent_size = 15;
+	desc.mantissa_size = 63;
+	desc.explicit_one  = 1;
+	desc.clss          = NORMAL;
 #else
-	char result_exponent = 11;
-	char result_mantissa = 52;
+	desc.exponent_size = 11;
+	desc.mantissa_size = 52;
+	desc.explicit_one  = 0;
+	desc.clss          = NORMAL;
 #endif
+	mantissa_size = desc.mantissa_size + desc.explicit_one;
 
 	temp = alloca(calc_buffer_size);
-#ifdef HAVE_EXPLICIT_ONE
-	value = fc_cast(val, result_exponent, result_mantissa-1, temp);
-#else
-	value = fc_cast(val, result_exponent, result_mantissa, temp);
-#endif
+	value = fc_cast(val, &desc, temp);
 
 	sign = value->sign;
 
@@ -1168,17 +1178,17 @@ LLDBL fc_val_to_ieee754(const fp_value *val) {
 	 * lead to wrong results */
 	exponent = sc_val_to_long(_exp(value)) ;
 
-	sc_val_from_ulong(2, NULL);
+	sc_val_from_ulong(ROUNDING_BITS, NULL);
 	_shift_right(_mant(value), sc_get_buffer(), _mant(value));
 
 	mantissa0 = 0;
 	mantissa1 = 0;
 
 	for (byte_offset = 0; byte_offset < 4; byte_offset++)
-		mantissa1 |= sc_sub_bits(_mant(value), result_mantissa, byte_offset) << (byte_offset<<3);
+		mantissa1 |= sc_sub_bits(_mant(value), mantissa_size, byte_offset) << (byte_offset << 3);
 
-	for (; (byte_offset<<3) < result_mantissa; byte_offset++)
-		mantissa0 |= sc_sub_bits(_mant(value), result_mantissa, byte_offset) << ((byte_offset-4)<<3);
+	for (; (byte_offset<<3) < desc.mantissa_size; byte_offset++)
+		mantissa0 |= sc_sub_bits(_mant(value), mantissa_size, byte_offset) << ((byte_offset - 4) << 3);
 
 #ifdef HAVE_LONG_DOUBLE
 	buildval.val.high = sign << 15;
@@ -1197,14 +1207,16 @@ LLDBL fc_val_to_ieee754(const fp_value *val) {
 	return buildval.d;
 }
 
-fp_value *fc_cast(const fp_value *value, char exp_size, char mant_size, fp_value *result) {
+fp_value *fc_cast(const fp_value *value, const ieee_descriptor_t *desc, fp_value *result) {
 	char *temp;
 	int exp_offset, val_bias, res_bias;
 
 	if (result == NULL) result = calc_buffer;
 	temp = alloca(value_size);
 
-	if (value->desc.exponent_size == exp_size && value->desc.mantissa_size == mant_size) {
+	if (value->desc.exponent_size == desc->exponent_size &&
+		value->desc.mantissa_size == desc->mantissa_size &&
+		value->desc.explicit_one  == desc->explicit_one) {
 		if (value != result)
 			memcpy(result, value, calc_buffer_size);
 		return result;
@@ -1212,15 +1224,16 @@ fp_value *fc_cast(const fp_value *value, char exp_size, char mant_size, fp_value
 
 	if (value->desc.clss == NAN) {
 		if (sc_get_highest_set_bit(_mant(value)) == value->desc.mantissa_size + 1)
-			return fc_get_qnan(exp_size, mant_size, result);
+			return fc_get_qnan(desc, result);
 		else
-			return fc_get_snan(exp_size, mant_size, result);
+			return fc_get_snan(desc, result);
 	}
 
 	/* set the descriptor of the new value */
-	result->desc.exponent_size = exp_size;
-	result->desc.mantissa_size = mant_size;
-	result->desc.clss = value->desc.clss;
+	result->desc.exponent_size = desc->exponent_size;
+	result->desc.mantissa_size = desc->mantissa_size;
+	result->desc.explicit_one  = desc->explicit_one;
+	result->desc.clss          = value->desc.clss;
 
 	result->sign = value->sign;
 
@@ -1228,9 +1241,9 @@ fp_value *fc_cast(const fp_value *value, char exp_size, char mant_size, fp_value
 	 * this would change the exponent, which is unwanted. So calculate this
 	 * offset and add it */
 	val_bias = (1 << (value->desc.exponent_size - 1)) - 1;
-	res_bias = (1 << (exp_size - 1)) - 1;
+	res_bias = (1 << (desc->exponent_size - 1)) - 1;
 
-	exp_offset = (res_bias - val_bias) - (value->desc.mantissa_size - mant_size);
+	exp_offset = (res_bias - val_bias) - (value->desc.mantissa_size - desc->mantissa_size);
 	sc_val_from_long(exp_offset, temp);
 	sc_add(_exp(value), temp, _exp(result));
 
@@ -1249,43 +1262,45 @@ fp_value *fc_cast(const fp_value *value, char exp_size, char mant_size, fp_value
 	return result;
 }
 
-fp_value *fc_get_max(unsigned int exponent_size, unsigned int mantissa_size, fp_value *result) {
+fp_value *fc_get_max(const ieee_descriptor_t *desc, fp_value *result) {
 	if (result == NULL) result = calc_buffer;
 
-	result->desc.exponent_size = exponent_size;
-	result->desc.mantissa_size = mantissa_size;
-	result->desc.clss = NORMAL;
+	result->desc.exponent_size = desc->exponent_size;
+	result->desc.mantissa_size = desc->mantissa_size;
+	result->desc.explicit_one  = desc->explicit_one;
+	result->desc.clss          = NORMAL;
 
 	result->sign = 0;
 
-	sc_val_from_ulong((1<<exponent_size) - 2, _exp(result));
+	sc_val_from_ulong((1 << desc->exponent_size) - 2, _exp(result));
 
-	sc_max_from_bits(mantissa_size + 1, 0, _mant(result));
-	sc_val_from_ulong(2, NULL);
+	sc_max_from_bits(desc->mantissa_size + 1, 0, _mant(result));
+	sc_val_from_ulong(ROUNDING_BITS, NULL);
 	_shift_left(_mant(result), sc_get_buffer(), _mant(result));
 
 	return result;
 }
 
-fp_value *fc_get_min(unsigned int exponent_size, unsigned int mantissa_size, fp_value *result) {
+fp_value *fc_get_min(const ieee_descriptor_t *desc, fp_value *result) {
 	if (result == NULL) result = calc_buffer;
 
-	fc_get_max(exponent_size, mantissa_size, result);
+	fc_get_max(desc, result);
 	result->sign = 1;
 
 	return result;
 }
 
-fp_value *fc_get_snan(unsigned int exponent_size, unsigned int mantissa_size, fp_value *result) {
+fp_value *fc_get_snan(const ieee_descriptor_t *desc, fp_value *result) {
 	if (result == NULL) result = calc_buffer;
 
-	result->desc.exponent_size = exponent_size;
-	result->desc.mantissa_size = mantissa_size;
-	result->desc.clss = NAN;
+	result->desc.exponent_size = desc->exponent_size;
+	result->desc.mantissa_size = desc->mantissa_size;
+	result->desc.explicit_one  = desc->explicit_one;
+	result->desc.clss          = NAN;
 
 	result->sign = 0;
 
-	sc_val_from_ulong((1<<exponent_size)-1, _exp(result));
+	sc_val_from_ulong((1 << desc->exponent_size) - 1, _exp(result));
 
 	/* signaling NaN has non-zero mantissa with msb not set */
 	sc_val_from_ulong(1, _mant(result));
@@ -1293,46 +1308,48 @@ fp_value *fc_get_snan(unsigned int exponent_size, unsigned int mantissa_size, fp
 	return result;
 }
 
-fp_value *fc_get_qnan(unsigned int exponent_size, unsigned int mantissa_size, fp_value *result) {
+fp_value *fc_get_qnan(const ieee_descriptor_t *desc, fp_value *result) {
 	if (result == NULL) result = calc_buffer;
 
-	result->desc.exponent_size = exponent_size;
-	result->desc.mantissa_size = mantissa_size;
-	result->desc.clss = NAN;
+	result->desc.exponent_size = desc->exponent_size;
+	result->desc.mantissa_size = desc->mantissa_size;
+	result->desc.explicit_one  = desc->explicit_one;
+	result->desc.clss          = NAN;
 
 	result->sign = 0;
 
-	sc_val_from_ulong((1<<exponent_size)-1, _exp(result));
+	sc_val_from_ulong((1 << desc->exponent_size) - 1, _exp(result));
 
 	/* quiet NaN has the msb of the mantissa set, so shift one there */
 	sc_val_from_ulong(1, _mant(result));
 	/* mantissa_size >+< 1 because of two extra rounding bits */
-	sc_val_from_ulong(mantissa_size + 1, NULL);
+	sc_val_from_ulong(desc->mantissa_size + 1, NULL);
 	_shift_left(_mant(result), sc_get_buffer(), _mant(result));
 
 	return result;
 }
 
-fp_value *fc_get_plusinf(unsigned int exponent_size, unsigned int mantissa_size, fp_value *result) {
+fp_value *fc_get_plusinf(const ieee_descriptor_t *desc, fp_value *result) {
 	if (result == NULL) result = calc_buffer;
 
-	result->desc.exponent_size = exponent_size;
-	result->desc.mantissa_size = mantissa_size;
-	result->desc.clss = NORMAL;
+	result->desc.exponent_size = desc->exponent_size;
+	result->desc.mantissa_size = desc->mantissa_size;
+	result->desc.explicit_one  = desc->explicit_one;
+	result->desc.clss          = INF;
 
 	result->sign = 0;
 
-	sc_val_from_ulong((1<<exponent_size)-1, _exp(result));
+	sc_val_from_ulong((1 << desc->exponent_size) - 1, _exp(result));
 
 	sc_val_from_ulong(0, _mant(result));
 
 	return result;
 }
 
-fp_value *fc_get_minusinf(unsigned int exponent_size, unsigned int mantissa_size, fp_value *result) {
+fp_value *fc_get_minusinf(const ieee_descriptor_t *desc, fp_value *result) {
 	if (result == NULL) result = calc_buffer;
 
-	fc_get_plusinf(exponent_size, mantissa_size, result);
+	fc_get_plusinf(desc, result);
 	result->sign = 1;
 
 	return result;
@@ -1407,38 +1424,37 @@ int fc_is_subnormal(const fp_value *a) {
 
 char *fc_print(const fp_value *val, char *buf, int buflen, unsigned base) {
 	char *mul_1;
+	LLDBL flt_val;
 
 	mul_1 = alloca(calc_buffer_size);
 
 	switch (base) {
 	case FC_DEC:
-		switch (val->desc.clss) {
+		switch ((value_class_t)val->desc.clss) {
 		case INF:
-			if (buflen >= 8 + val->sign) sprintf(buf, "%sINFINITY", val->sign ? "-":"");
-			else snprintf(buf, buflen, "%sINF", val->sign ? "-":NULL);
+			snprintf(buf, buflen, "%cINF", val->sign ? '-' : '+');
 			break;
 		case NAN:
-			snprintf(buf, buflen, "NAN");
+			snprintf(buf, buflen, "NaN");
 			break;
 		case ZERO:
 			snprintf(buf, buflen, "0.0");
 			break;
 		default:
-			/* XXX to be implemented */
+			flt_val = fc_val_to_ieee754(val);
 #ifdef HAVE_LONG_DOUBLE
 			/* XXX 30 is arbitrary */
-			snprintf(buf, buflen, "%.30LE", fc_val_to_ieee754(val));
+			snprintf(buf, buflen, "%.30LE", flt_val);
 #else
-			snprintf(buf, buflen, "%.18E", fc_val_to_ieee754(val));
+			snprintf(buf, buflen, "%.18E", flt_val);
 #endif
 		}
 		break;
 
 	case FC_HEX:
-		switch (val->desc.clss) {
+		switch ((value_class_t)val->desc.clss) {
 		case INF:
-			if (buflen >= 8+val->sign) sprintf(buf, "%sINFINITY", val->sign?"-":"");
-			else snprintf(buf, buflen, "%sINF", val->sign?"-":NULL);
+			snprintf(buf, buflen, "%cINF", val->sign ? '-' : '+');
 			break;
 		case NAN:
 			snprintf(buf, buflen, "NAN");
@@ -1447,10 +1463,11 @@ char *fc_print(const fp_value *val, char *buf, int buflen, unsigned base) {
 			snprintf(buf, buflen, "0.0");
 			break;
 		default:
+			flt_val = fc_val_to_ieee754(val);
 #ifdef HAVE_LONG_DOUBLE
-			snprintf(buf, buflen, "%LA", fc_val_to_ieee754(val));
+			snprintf(buf, buflen, "%LA", flt_val);
 #else
-			snprintf(buf, buflen, "%A", fc_val_to_ieee754(val));
+			snprintf(buf, buflen, "%A", flt_val);
 #endif
 		}
 		break;
@@ -1468,7 +1485,7 @@ unsigned char fc_sub_bits(const fp_value *value, unsigned num_bits, unsigned byt
 	/* this is used to cache the packed version of the value */
 	static char *packed_value = NULL;
 
-	if (packed_value == NULL) packed_value = xmalloc(value_size);
+	if (packed_value == NULL) packed_value = XMALLOCN(char, value_size);
 
 	if (value != NULL)
 		pack(value, packed_value);
@@ -1476,15 +1493,43 @@ unsigned char fc_sub_bits(const fp_value *value, unsigned num_bits, unsigned byt
 	return sc_sub_bits(packed_value, num_bits, byte_ofs);
 }
 
+/* Returns non-zero if the mantissa is zero, i.e. 1.0Exxx */
 int fc_zero_mantissa(const fp_value *value) {
-	return sc_get_lowest_set_bit(_mant(value)) == 2 + value->desc.mantissa_size;
+	return sc_get_lowest_set_bit(_mant(value)) == ROUNDING_BITS + value->desc.mantissa_size;
 }
 
+/* Returns the exponent of a value. */
 int fc_get_exponent(const fp_value *value) {
 	int exp_bias = (1 << (value->desc.exponent_size - 1)) - 1;
 	return sc_val_to_long(_exp(value)) - exp_bias;
 }
 
+/* Return non-zero if a given value can be converted lossless into another precision */
+int fc_can_lossless_conv_to(const fp_value *value, const ieee_descriptor_t *desc) {
+	int v;
+	int exp_bias;
+
+	/* handle some special cases first */
+	switch (value->desc.clss) {
+	case ZERO:
+	case INF:
+	case NAN:
+		return 1;
+	default:
+		break;
+	}
+
+	/* check if the exponent can be encoded: note, 0 and all ones are reserved for the exponent */
+	exp_bias = (1 << (desc->exponent_size - 1)) - 1;
+	v = fc_get_exponent(value) + exp_bias;
+	if (0 < v && v < (1 << desc->exponent_size) - 1) {
+		/* exponent can be encoded, now check the mantissa */
+		v = value->desc.mantissa_size + ROUNDING_BITS - sc_get_lowest_set_bit(_mant(value));
+		return v < desc->mantissa_size;
+	}
+	return 0;
+}
+
 
 fc_rounding_mode_t fc_set_rounding_mode(fc_rounding_mode_t mode) {
 	if (mode == FC_TONEAREST || mode == FC_TOPOSITIVE || mode == FC_TONEGATIVE || mode == FC_TOZERO)
@@ -1502,11 +1547,11 @@ void init_fltcalc(int precision) {
 		/* does nothing if already init */
 		if (precision == 0) precision = FC_DEFAULT_PRECISION;
 
-		init_strcalc(precision + 4);
+		init_strcalc(precision + 2 + ROUNDING_BITS);
 
-		/* needs additionally two bits to round, a bit as explicit 1., and one for
+		/* needs additionally rounding bits, one bit as explicit 1., and one for
 		 * addition overflow */
-		max_precision = sc_get_precision() - 4;
+		max_precision = sc_get_precision() - (2 + ROUNDING_BITS);
 		if (max_precision < precision)
 			printf("WARNING: not enough precision available, using %d\n", max_precision);
 
@@ -1627,6 +1672,7 @@ fp_value *fc_int(const fp_value *a, fp_value *result) {
 fp_value *fc_rnd(const fp_value *a, fp_value *result) {
 	if (result == NULL) result = calc_buffer;
 
+	(void) a;
 	TRACEPRINTF(("%s ", fc_print(a, buffer, sizeof(buffer), FC_PACKED)));
 	TRACEPRINTF(("rounded to integer "));
 
@@ -1636,6 +1682,68 @@ fp_value *fc_rnd(const fp_value *a, fp_value *result) {
 	return result;
 }
 
+/*
+ * convert a floating point value into an sc value ...
+ */
+int fc_flt2int(const fp_value *a, void *result, ir_mode *dst_mode) {
+	if (a->desc.clss == NORMAL) {
+		int exp_bias = (1 << (a->desc.exponent_size - 1)) - 1;
+		int exp_val  = sc_val_to_long(_exp(a)) - exp_bias;
+		int shift, highest;
+		int mantissa_size;
+
+		if (a->sign && !mode_is_signed(dst_mode)) {
+			/* FIXME: for now we cannot convert this */
+			return 0;
+		}
+
+		assert(exp_val >= 0 && "floating point value not integral before fc_flt2int() call");
+		mantissa_size = a->desc.mantissa_size + ROUNDING_BITS;
+		shift         = exp_val - mantissa_size;
+
+		mantissa_size += a->desc.explicit_one;
+		if (shift > 0) {
+			sc_shlI(_mant(a),  shift, mantissa_size, 0, result);
+		} else {
+			sc_shrI(_mant(a), -shift, mantissa_size, 0, result);
+		}
+
+		/* check for overflow */
+		highest = sc_get_highest_set_bit(result);
+
+		if (mode_is_signed(dst_mode)) {
+			if (highest == sc_get_lowest_set_bit(result)) {
+				/* need extra test for MIN_INT */
+				if (highest >= (int) get_mode_size_bits(dst_mode)) {
+					/* FIXME: handle overflow */
+					return 0;
+				}
+			} else {
+				if (highest >= (int) get_mode_size_bits(dst_mode) - 1) {
+					/* FIXME: handle overflow */
+					return 0;
+				}
+			}
+		} else {
+			if (highest >= (int) get_mode_size_bits(dst_mode)) {
+				/* FIXME: handle overflow */
+				return 0;
+			}
+		}
+
+		if (a->sign)
+			sc_neg(result, result);
+
+		return 1;
+	}
+	else if (a->desc.clss == ZERO) {
+		sc_zero(result);
+		return 1;
+	}
+	return 0;
+}
+
+
 unsigned fc_set_immediate_precision(unsigned bits) {
 	unsigned old = immediate_prec;