/*
- * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
+ * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved.
*
* This file is part of libFirm.
*
* @brief tarval floating point calculations
* @date 2003
* @author Mathias Heil
- * @version $Id$
*/
-
-#ifdef HAVE_CONFIG_H
-# include "config.h"
-#endif
+#include "config.h"
#include "fltcalc.h"
#include "strcalc.h"
+#include "error.h"
-#include <math.h> /* need isnan() and isinf() (will be changed)*/
-/* undef some reused constants defined by math.h */
-#ifdef NAN
-# undef NAN
-#endif
-
-#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 <math.h>
+#include <inttypes.h>
+#include <string.h>
+#include <stdlib.h>
#include <stdio.h>
#include <assert.h>
+#include <stdbool.h>
#include "xmalloc.h"
-typedef uint32_t UINT32;
+/*
+ * portability stuff (why do we even care about the msvc people with their C89?)
+ */
-#ifdef HAVE_LONG_DOUBLE
-#ifdef WORDS_BIGENDIAN
-typedef union {
- struct {
- UINT32 high;
- UINT32 mid;
- UINT32 low;
- } val;
- volatile long double d;
-} value_t;
+
+static long double string_to_long_double(const char *str)
+{
+#if __STDC_VERSION__ >= 199901L || _POSIX_C_SOURCE >= 200112L
+ return strtold(str, NULL);
#else
-typedef union {
- struct {
- UINT32 low;
- UINT32 mid;
- UINT32 high;
- } val;
- volatile long double d;
-} value_t;
+ return strtod(str, NULL);
#endif
+}
+
+static bool my_isnan(long double val)
+{
+#if __STDC_VERSION__ >= 199901L
+ return isnan(val);
#else
-#ifdef WORDS_BIGENDIAN
+ /* hopefully the compiler does not optimize aggressively (=incorrect) */
+ return val != val;
+#endif
+}
+
+static bool my_isinf(long double val)
+{
+#if __STDC_VERSION__ >= 199901L
+ return isinf(val);
+#else
+ /* hopefully the compiler does not optimize aggressively (=incorrect) */
+ return my_isnan(val-val) && !my_isnan(val);
+#endif
+}
+
+/** The number of extra precision rounding bits */
+#define ROUNDING_BITS 2
+
typedef union {
struct {
- UINT32 high;
- UINT32 low;
- } val;
- volatile double d;
-} value_t;
+#ifdef WORDS_BIGENDIAN
+ uint32_t high;
#else
-typedef union {
+ uint32_t low;
+#endif
+ uint32_t mid;
+#ifdef WORDS_BIGENDIAN
+ uint32_t low;
+#else
+ uint32_t high;
+#endif
+ } val_ld12;
struct {
- UINT32 low;
- UINT32 high;
- } val;
- volatile double d;
-} value_t;
+#ifdef WORDS_BIGENDIAN
+ uint32_t high;
+#else
+ uint32_t low;
#endif
+#ifdef WORDS_BIGENDIAN
+ uint32_t low;
+#else
+ uint32_t high;
#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;
+ } val_ld8;
+ volatile long double d;
+} value_t;
#define CLEAR_BUFFER(buffer) memset(buffer, 0, calc_buffer_size)
/* our floating point value */
-struct _fp_value {
- descriptor_t desc;
+struct fp_value {
+ float_descriptor_t desc;
+ unsigned char clss;
char sign;
- char value[1]; /* exp[value_size] + mant[value_size] */
+ char value[1]; /* exp[value_size] + mant[value_size] */
};
#define _exp(a) &((a)->value[0])
#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
# define TRACEPRINTF(x) ((void)0)
#endif
-/** The immediate precision. */
-static unsigned immediate_prec = 0;
-
/** A temporal buffer. */
static fp_value *calc_buffer = NULL;
/** Exact flag. */
static int fc_exact = 1;
-#if 0
-static void fail_char(const char *str, unsigned int len, int pos) {
- if (*(str+pos))
- printf("ERROR: Unexpected character '%c'\n", *(str + pos));
- else
- printf("ERROR: Unexpected end of string\n");
- while (len-- && *str) printf("%c", *str++); printf("\n");
- while (pos--) printf(" "); printf("^\n");
- /* the front end has to to check constant strings */
- exit(-1);
-}
-#endif
-
/** pack machine-like */
-static void *pack(const fp_value *int_float, void *packed) {
- char *shift_val;
- char *temp;
+static void *pack(const fp_value *int_float, void *packed)
+{
+ char *shift_val;
+ char *temp;
fp_value *val_buffer;
+ int pos;
- temp = alloca(value_size);
- shift_val = alloca(value_size);
+ temp = (char*) alloca(value_size);
+ shift_val = (char*) alloca(value_size);
- switch (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);
+ switch ((value_class_t)int_float->clss) {
+ case FC_NAN:
+ val_buffer = (fp_value*) alloca(calc_buffer_size);
+ 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);
+ case FC_INF:
+ val_buffer = (fp_value*) alloca(calc_buffer_size);
+ fc_get_plusinf(&int_float->desc, val_buffer);
val_buffer->sign = int_float->sign;
int_float = val_buffer;
break;
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;
*
* @return non-zero if result is exact
*/
-static int normalize(const fp_value *in_val, fp_value *out_val, int sticky) {
+static int normalize(const fp_value *in_val, fp_value *out_val, int sticky)
+{
int exact = 1;
int hsb;
char lsb, guard, round, round_dir = 0;
- char *temp = alloca(value_size);
+ char *temp = (char*) 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;
- memcpy(&out_val->desc, &in_val->desc, sizeof(out_val->desc));
+ out_val->desc = in_val->desc;
}
- out_val->desc.clss = NORMAL;
+ out_val->clss = FC_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;
}
/* denormalized means exponent of zero */
sc_val_from_ulong(0, _exp(out_val));
- out_val->desc.clss = SUBNORMAL;
+ out_val->clss = FC_SUBNORMAL;
}
/* 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;
}
/* could have rounded down to zero */
- if (sc_is_zero(_mant(out_val)) && (out_val->desc.clss == SUBNORMAL))
- out_val->desc.clss = ZERO;
+ if (sc_is_zero(_mant(out_val)) && (out_val->clss == FC_SUBNORMAL))
+ out_val->clss = FC_ZERO;
/* check for rounding overflow */
- hsb = 2 + out_val->desc.mantissa_size - sc_get_highest_set_bit(_mant(out_val)) - 1;
- if ((out_val->desc.clss != SUBNORMAL) && (hsb < -1)) {
+ hsb = ROUNDING_BITS + out_val->desc.mantissa_size - sc_get_highest_set_bit(_mant(out_val)) - 1;
+ if ((out_val->clss != FC_SUBNORMAL) && (hsb < -1)) {
sc_val_from_ulong(1, temp);
_shift_right(_mant(out_val), temp, _mant(out_val));
if (exact && sc_had_carry())
exact = 0;
sc_add(_exp(out_val), temp, _exp(out_val));
- } else if ((out_val->desc.clss == SUBNORMAL) && (hsb == -1)) {
+ } else if ((out_val->clss == FC_SUBNORMAL) && (hsb == -1)) {
/* overflow caused the mantissa to be normal again,
* so adapt the exponent accordingly */
sc_val_from_ulong(1, temp);
sc_add(_exp(out_val), temp, _exp(out_val));
- out_val->desc.clss = NORMAL;
+ out_val->clss = FC_NORMAL;
}
/* no further rounding is needed, because rounding overflow means
* the carry of the original rounding was propagated all the way
switch (rounding_mode) {
case FC_TONEAREST:
case FC_TOPOSITIVE:
- out_val->desc.clss = INF;
+ out_val->clss = FC_INF;
break;
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 */
switch (rounding_mode) {
case FC_TONEAREST:
case FC_TONEGATIVE:
- out_val->desc.clss = INF;
+ out_val->clss = FC_INF;
break;
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);
}
}
}
}
/**
- * 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->clss == FC_NAN) { \
if (a != result) memcpy(result, a, calc_buffer_size); \
+ fc_exact = 0; \
return; \
} \
- if (b->desc.clss == NAN) { \
+ if (b->clss == FC_NAN) { \
if (b != result) memcpy(result, b, calc_buffer_size); \
+ fc_exact = 0; \
return; \
} \
}while (0)
/**
* calculate a + b, where a is the value with the bigger exponent
*/
-static void _fadd(const fp_value *a, const fp_value *b, fp_value *result) {
+static void _fadd(const fp_value *a, const fp_value *b, fp_value *result)
+{
char *temp;
char *exp_diff;
/* 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);
+ if (sign && (a->clss == FC_INF) && (b->clss == FC_INF)) {
+ fc_exact = 0;
+ fc_get_qnan(&a->desc, result);
return;
}
- temp = alloca(value_size);
- exp_diff = alloca(value_size);
+ temp = (char*) alloca(value_size);
+ exp_diff = (char*) alloca(value_size);
/* get exponent difference */
sc_sub(_exp(a), _exp(b), exp_diff);
result->sign = res_sign;
/* sign has been taken care of, check for special cases */
- if (a->desc.clss == ZERO || b->desc.clss == INF) {
+ if (a->clss == FC_ZERO || b->clss == FC_INF) {
if (b != result)
memcpy(result, b, calc_buffer_size);
+ fc_exact = b->clss == FC_NORMAL;
result->sign = res_sign;
return;
}
- if (b->desc.clss == ZERO || a->desc.clss == INF) {
+ if (b->clss == FC_ZERO || a->clss == FC_INF) {
if (a != result)
memcpy(result, a, calc_buffer_size);
+ fc_exact = a->clss == FC_NORMAL;
result->sign = res_sign;
return;
}
/* shift the smaller value to the right to align the radix point */
/* subnormals have their radix point shifted to the right,
* take care of this first */
- if ((b->desc.clss == SUBNORMAL) && (a->desc.clss != SUBNORMAL)) {
+ if ((b->clss == FC_SUBNORMAL) && (a->clss != FC_SUBNORMAL)) {
sc_val_from_ulong(1, temp);
sc_sub(exp_diff, temp, exp_diff);
}
/* if subtracting a little more than the represented value or adding a little
* more than the represented value to a negative value this, in addition to the
* still set sticky bit, takes account of the 'little more' */
- char *temp1 = alloca(calc_buffer_size);
+ char *temp1 = (char*) alloca(calc_buffer_size);
sc_val_from_ulong(1, temp1);
sc_add(temp, temp1, temp);
}
/* _normalize expects a 'normal' radix point, adding two subnormals
* results in a subnormal radix point -> shifting before normalizing */
- if ((a->desc.clss == SUBNORMAL) && (b->desc.clss == SUBNORMAL)) {
+ if ((a->clss == FC_SUBNORMAL) && (b->clss == FC_SUBNORMAL)) {
sc_val_from_ulong(1, NULL);
_shift_left(_mant(result), sc_get_buffer(), _mant(result));
}
/**
* calculate a * b
*/
-static void _fmul(const fp_value *a, const fp_value *b, fp_value *result) {
+static void _fmul(const fp_value *a, const fp_value *b, fp_value *result)
+{
int sticky;
char *temp;
char res_sign;
handle_NAN(a, b, result);
- temp = alloca(value_size);
+ temp = (char*) 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 (a->clss == FC_ZERO) {
+ if (b->clss == FC_INF) {
+ fc_get_qnan(&a->desc, result);
+ fc_exact = 0;
+ } else {
if (a != result)
memcpy(result, a, calc_buffer_size);
result->sign = res_sign;
}
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 (b->clss == FC_ZERO) {
+ if (a->clss == FC_INF) {
+ fc_get_qnan(&a->desc, result);
+ fc_exact = 0;
+ } else {
if (b != result)
memcpy(result, b, calc_buffer_size);
result->sign = res_sign;
return;
}
- if (a->desc.clss == INF) {
+ if (a->clss == FC_INF) {
+ fc_exact = 0;
if (a != result)
memcpy(result, a, calc_buffer_size);
result->sign = res_sign;
return;
}
- if (b->desc.clss == INF) {
+ if (b->clss == FC_INF) {
+ fc_exact = 0;
if (b != result)
memcpy(result, b, calc_buffer_size);
result->sign = res_sign;
sc_sub(_exp(result), temp, _exp(result));
/* mixed normal, subnormal values introduce an error of 1, correct it */
- if ((a->desc.clss == SUBNORMAL) ^ (b->desc.clss == SUBNORMAL)) {
+ if ((a->clss == FC_SUBNORMAL) ^ (b->clss == FC_SUBNORMAL)) {
sc_val_from_ulong(1, temp);
sc_add(_exp(result), temp, _exp(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();
/**
* calculate a / b
*/
-static void _fdiv(const fp_value *a, const fp_value *b, fp_value *result) {
+static void _fdiv(const fp_value *a, const fp_value *b, fp_value *result)
+{
int sticky;
char *temp, *dividend;
char res_sign;
handle_NAN(a, b, result);
- temp = alloca(value_size);
- dividend = alloca(value_size);
+ temp = (char*) alloca(value_size);
+ dividend = (char*) 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 {
+ /* produce FC_NAN on 0/0 and inf/inf */
+ if (a->clss == FC_ZERO) {
+ if (b->clss == FC_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);
return;
}
- 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 {
+ if (b->clss == FC_INF) {
+ fc_exact = 0;
+ if (a->clss == FC_INF) {
+ /* inf/inf -> NaN */
+ fc_get_qnan(&a->desc, result);
+ } else {
/* x/inf -> 0 */
sc_val_from_ulong(0, NULL);
_save_result(_exp(result));
_save_result(_mant(result));
- result->desc.clss = ZERO;
+ result->clss = FC_ZERO;
}
return;
}
- if (a->desc.clss == INF) {
+ if (a->clss == FC_INF) {
+ fc_exact = 0;
/* inf/x -> inf */
if (a != result)
memcpy(result, a, calc_buffer_size);
result->sign = res_sign;
return;
}
- if (b->desc.clss == ZERO) {
+ if (b->clss == FC_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;
}
sc_add(_exp(result), temp, _exp(result));
/* mixed normal, subnormal values introduce an error of 1, correct it */
- if ((a->desc.clss == SUBNORMAL) ^ (b->desc.clss == SUBNORMAL)) {
+ if ((a->clss == FC_SUBNORMAL) ^ (b->clss == FC_SUBNORMAL)) {
sc_val_from_ulong(1, temp);
sc_add(_exp(result), temp, _exp(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);
{
- char *divisor = alloca(calc_buffer_size);
+ char *divisor = (char*) alloca(calc_buffer_size);
sc_val_from_ulong(1, divisor);
_shift_right(_mant(b), divisor, divisor);
sc_div(dividend, divisor, _mant(result));
}
#if 0
-static void _power_of_ten(int exp, descriptor_t *desc, char *result) {
+static void _power_of_ten(int exp, float_descriptor_t *desc, char *result)
+{
char *build;
char *temp;
/* 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 */
_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));
*
* This does not clip to any integer range.
*/
-static void _trunc(const fp_value *a, fp_value *result) {
+static void _trunc(const fp_value *a, fp_value *result)
+{
/*
* When exponent == 0 all bits left of the radix point
* are the integral part of the value. For 15bit exp_size
/* fixme: can be exact */
fc_exact = 0;
- temp = alloca(value_size);
+ temp = (char*) alloca(value_size);
- if (a != result)
- memcpy(&result->desc, &a->desc, sizeof(descriptor_t));
+ if (a != result) {
+ result->desc = a->desc;
+ result->clss = a->clss;
+ }
exp_bias = (1 << (a->desc.exponent_size - 1)) - 1;
exp_val = sc_val_to_long(_exp(a)) - exp_bias;
sc_val_from_ulong(0, NULL);
_save_result(_exp(result));
_save_result(_mant(result));
- result->desc.clss = ZERO;
+ result->clss = FC_ZERO;
return;
}
- if (exp_val > a->desc.mantissa_size) {
+ if (exp_val > (long)a->desc.mantissa_size) {
if (a != result)
memcpy(result, a, calc_buffer_size);
/* 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;
+ }
}
/********
* functions defined in fltcalc.h
********/
-const void *fc_get_buffer(void) {
+const void *fc_get_buffer(void)
+{
return calc_buffer;
}
-int fc_get_buffer_length(void) {
+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) {
-#if 0
- enum {
- START,
- LEFT_OF_DOT,
- RIGHT_OF_DOT,
- EXP_START,
- EXPONENT,
- END
- };
-
- char exp_sign;
- int exp_int, hsb, state;
-
- const char *old_str;
-
- int pos;
- char *mant_str, *exp_val, *power_val;
-
- (void) len;
- if (result == NULL) result = calc_buffer;
-
- exp_val = alloca(value_size);
- 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;
-
- old_str = str;
- pos = 0;
- exp_int = 0;
- state = START;
-
- while (len == 0 || str-old_str < len) {
- switch (state) {
- case START:
- switch (*str) {
- case '+':
- result->sign = 0;
- state = LEFT_OF_DOT;
- str++;
- break;
-
- case '-':
- result->sign = 1;
- state = LEFT_OF_DOT;
- str++;
- break;
-
- case '0': case '1': case '2': case '3': case '4': case '5': case '6': case '7': case '8': case '9':
- result->sign = 0;
- state = LEFT_OF_DOT;
- break;
-
- case '.':
- result->sign = 0;
- state = RIGHT_OF_DOT;
- str++;
- break;
-
- case 'n':
- case 'N':
- case 'i':
- case 'I':
- break;
-
- default:
- fail_char(old_str, len, str - old_str);
- }
- break;
-
- case LEFT_OF_DOT:
- switch (*str) {
- case '0': case '1': case '2': case '3': case '4': case '5': case '6': case '7': case '8': case '9':
- mant_str[pos++] = *(str++);
- break;
-
- case '.':
- state = RIGHT_OF_DOT;
- str++;
- break;
-
- case 'e':
- case 'E':
- state = EXP_START;
- str++;
- break;
-
- case '\0':
- mant_str[pos] = '\0';
- goto done;
-
- default:
- fail_char(old_str, len, str - old_str);
- }
- break;
-
- case RIGHT_OF_DOT:
- switch (*str) {
- case '0': case '1': case '2': case '3': case '4': case '5': case '6': case '7': case '8': case '9':
- mant_str[pos++] = *(str++);
- exp_int++;
- break;
-
- case 'e':
- case 'E':
- state = EXP_START;
- str++;
- break;
+void *fc_val_from_str(const char *str, size_t len,
+ const float_descriptor_t *desc, void *result)
+{
+ char *buffer;
- case '\0':
- mant_str[pos] = '\0';
- goto done;
-
- default:
- fail_char(old_str, len, str - old_str);
- }
- break;
-
- case EXP_START:
- switch (*str) {
- case '-':
- exp_sign = 1;
- /* fall through */
- case '+':
- if (*(str-1) != 'e' && *(str-1) != 'E') fail_char(old_str, len, str - old_str);
- str++;
- break;
-
- case '0': case '1': case '2': case '3': case '4': case '5': case '6': case '7': case '8': case '9':
- mant_str[pos] = '\0';
- pos = 1;
- str++;
- state = EXPONENT;
- break;
-
- default:
- fail_char(old_str, len, str - old_str);
- }
- break;
-
- case EXPONENT:
- switch (*str) {
- case '0': case '1': case '2': case '3': case '4': case '5': case '6': case '7': case '8': case '9':
- pos++;
- str++;
- break;
-
- case '\0': goto done;
-
- default:
- fail_char(old_str, len, str - old_str);
- }
- }
- } /* switch(state) */
-
-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);
-
- _shift_left(_mant(result), exp_val, _mant(result));
-
- sc_val_from_ulong((1 << exp_size)/2-1, _exp(result));
-
- _normalize(result, result, 0);
-
- if (state == EXPONENT) {
- exp_int -= atoi(str-pos);
- }
-
- _power_of_ten(exp_int, &result->desc, power_val);
-
- _fdiv(result, power_val, result);
-
- return result;
-#else
/* XXX excuse of an implementation to make things work */
- LLDBL val;
- fp_value *tmp = alloca(calc_buffer_size);
- (void) len;
+ long double val;
+ fp_value *tmp = (fp_value*) alloca(calc_buffer_size);
+ float_descriptor_t tmp_desc;
+
+ buffer = (char*) alloca(len+1);
+ memcpy(buffer, str, len);
+ buffer[len] = '\0';
+ val = string_to_long_double(buffer);
-#ifdef HAVE_LONG_DOUBLE
- val = strtold(str, NULL);
- DEBUGPRINTF(("val_from_str(%s)\n", str));
- fc_val_from_ieee754(val, 15, 64, tmp);
-#else
- val = strtod(str, NULL);
DEBUGPRINTF(("val_from_str(%s)\n", str));
- fc_val_from_ieee754(val, 11, 52, tmp);
-#endif /* HAVE_LONG_DOUBLE */
- return fc_cast(tmp, exp_size, mant_size, result);
-#endif
+ tmp_desc.exponent_size = 15;
+ tmp_desc.mantissa_size = 63;
+ tmp_desc.explicit_one = 1;
+ fc_val_from_ieee754(val, &tmp_desc, tmp);
+
+ return fc_cast(tmp, desc, (fp_value*) result);
}
-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;
- value_t srcval;
- UINT32 sign, exponent, mantissa0, mantissa1;
+fp_value *fc_val_from_ieee754(long double l, const float_descriptor_t *desc,
+ fp_value *result)
+{
+ char *temp;
+ int bias_res, bias_val, mant_val;
+ value_t srcval;
+ char sign;
+ uint32_t exponent, mantissa0, mantissa1;
+ size_t long_double_size = sizeof(long double);
srcval.d = l;
- bias_res = ((1<<exp_size)/2-1);
-
-#ifdef HAVE_LONG_DOUBLE
- mant_val = 64;
- bias_val = 0x3fff;
- sign = (srcval.val.high & 0x00008000) != 0;
- exponent = (srcval.val.high & 0x00007FFF) ;
- mantissa0 = srcval.val.mid;
- mantissa1 = srcval.val.low;
-#else /* no long double */
- mant_val = 52;
- bias_val = 0x3ff;
- sign = (srcval.val.high & 0x80000000) != 0;
- exponent = (srcval.val.high & 0x7FF00000) >> 20;
- mantissa0 = srcval.val.high & 0x000FFFFF;
- mantissa1 = srcval.val.low;
-#endif
-
-#ifdef HAVE_LONG_DOUBLE
- TRACEPRINTF(("val_from_float(%.8X%.8X%.8X)\n", ((int*)&l)[2], ((int*)&l)[1], ((int*)&l)[0]));/* srcval.val.high, srcval.val.mid, srcval.val.low)); */
- DEBUGPRINTF(("(%d-%.4X-%.8X%.8X)\n", sign, exponent, mantissa0, mantissa1));
-#else
- TRACEPRINTF(("val_from_float(%.8X%.8X)\n", srcval.val.high, srcval.val.low));
- DEBUGPRINTF(("(%d-%.3X-%.5X%.8X)\n", sign, exponent, mantissa0, mantissa1));
-#endif
+ bias_res = ((1 << (desc->exponent_size - 1)) - 1);
+
+ if (long_double_size == 8) {
+ mant_val = 52;
+ bias_val = 0x3ff;
+ sign = (srcval.val_ld8.high & 0x80000000) != 0;
+ exponent = (srcval.val_ld8.high & 0x7FF00000) >> 20;
+ mantissa0 = srcval.val_ld8.high & 0x000FFFFF;
+ mantissa1 = srcval.val_ld8.low;
+ } else {
+ /* we assume an x86-like 80bit representation of the value... */
+ assert(sizeof(long double)==12 || sizeof(long double)==16);
+ mant_val = 63;
+ bias_val = 0x3fff;
+ sign = (srcval.val_ld12.high & 0x00008000) != 0;
+ exponent = (srcval.val_ld12.high & 0x00007FFF) ;
+ mantissa0 = srcval.val_ld12.mid;
+ mantissa1 = srcval.val_ld12.low;
+ }
- if (result == NULL) result = calc_buffer;
- temp = alloca(value_size);
+ if (result == NULL)
+ result = calc_buffer;
+ temp = (char*) alloca(value_size);
- /* CLEAR the buffer */
+ /* CLEAR the buffer, else some bits might be uninitialized */
memset(result, 0, fc_get_buffer_length());
- result->desc.exponent_size = exp_size;
- result->desc.mantissa_size = mant_size;
-
- /* extract sign */
+ result->desc = *desc;
+ result->clss = FC_NORMAL;
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;
+ if (my_isnan(l)) {
+ result->clss = FC_NAN;
TRACEPRINTF(("val_from_float resulted in NAN\n"));
return result;
- }
- else if (isinf(l)) {
- result->desc.clss = INF;
+ } else if (my_isinf(l)) {
+ result->clss = FC_INF;
TRACEPRINTF(("val_from_float resulted in %sINF\n", (result->sign == 1) ? "-" : ""));
return result;
}
* 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);
}
/* 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));
return result;
}
-LLDBL fc_val_to_ieee754(const fp_value *val) {
+long double fc_val_to_ieee754(const fp_value *val)
+{
fp_value *value;
fp_value *temp = NULL;
- int byte_offset;
+ unsigned byte_offset;
- UINT32 sign;
- UINT32 exponent;
- UINT32 mantissa0;
- UINT32 mantissa1;
+ uint32_t sign;
+ uint32_t exponent;
+ uint32_t mantissa0;
+ uint32_t mantissa1;
- value_t buildval;
+ value_t buildval;
+ float_descriptor_t desc;
+ unsigned mantissa_size;
-#ifdef HAVE_LONG_DOUBLE
- char result_exponent = 15;
- char result_mantissa = 64;
-#else
- char result_exponent = 11;
- char result_mantissa = 52;
-#endif
+ size_t long_double_size = sizeof(long double);
- 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
+ if (long_double_size == 8) {
+ desc.exponent_size = 11;
+ desc.mantissa_size = 52;
+ desc.explicit_one = 0;
+ } else {
+ desc.exponent_size = 15;
+ desc.mantissa_size = 63;
+ desc.explicit_one = 1;
+ }
+ mantissa_size = desc.mantissa_size + desc.explicit_one;
+
+ temp = (fp_value*) alloca(calc_buffer_size);
+ value = fc_cast(val, &desc, temp);
sign = value->sign;
* 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);
-
- for (; (byte_offset<<3) < result_mantissa; byte_offset++)
- mantissa0 |= sc_sub_bits(_mant(value), result_mantissa, byte_offset) << ((byte_offset-4)<<3);
-
-#ifdef HAVE_LONG_DOUBLE
- buildval.val.high = sign << 15;
- buildval.val.high |= exponent;
- buildval.val.mid = mantissa0;
- buildval.val.low = mantissa1;
-#else /* no long double */
- mantissa0 &= 0x000FFFFF; /* get rid of garbage */
- buildval.val.high = sign << 31;
- buildval.val.high |= exponent << 20;
- buildval.val.high |= mantissa0;
- buildval.val.low = mantissa1;
-#endif
+ mantissa1 |= sc_sub_bits(_mant(value), mantissa_size, byte_offset) << (byte_offset << 3);
+
+ for (; (byte_offset<<3) < desc.mantissa_size; byte_offset++)
+ mantissa0 |= sc_sub_bits(_mant(value), mantissa_size, byte_offset) << ((byte_offset - 4) << 3);
+
+ if (long_double_size == 8) {
+ mantissa0 &= 0x000FFFFF; /* get rid of garbage */
+ buildval.val_ld8.high = sign << 31;
+ buildval.val_ld8.high |= exponent << 20;
+ buildval.val_ld8.high |= mantissa0;
+ buildval.val_ld8.low = mantissa1;
+ } else {
+ buildval.val_ld12.high = sign << 15;
+ buildval.val_ld12.high |= exponent;
+ buildval.val_ld12.mid = mantissa0;
+ buildval.val_ld12.low = mantissa1;
+ }
TRACEPRINTF(("val_to_float: %d-%x-%x%x\n", sign, exponent, mantissa0, mantissa1));
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 float_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);
+ temp = (char*) 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;
}
- if (value->desc.clss == NAN) {
+ if (value->clss == FC_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);
+ }
+ else if (value->clss == FC_INF) {
+ if (value->sign == 0)
+ return fc_get_plusinf(desc, result);
+ else
+ return fc_get_minusinf(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 = *desc;
+ result->clss = value->clss;
result->sign = value->sign;
/* when the mantissa sizes differ normalizing has to shift to align it.
* 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));
/* _normalize expects normalized radix point */
- if (value->desc.clss == SUBNORMAL) {
+ if (value->clss == FC_SUBNORMAL) {
sc_val_from_ulong(1, NULL);
_shift_left(_mant(value), sc_get_buffer(), _mant(result));
} else if (value != result) {
return result;
}
-fp_value *fc_get_max(unsigned int exponent_size, unsigned int mantissa_size, fp_value *result) {
+fp_value *fc_get_max(const float_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 = *desc;
+ result->clss = FC_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 float_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 float_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 = *desc;
+ result->clss = FC_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));
return result;
}
-fp_value *fc_get_qnan(unsigned int exponent_size, unsigned int mantissa_size, fp_value *result) {
+fp_value *fc_get_qnan(const float_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 = *desc;
+ result->clss = FC_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) {
- if (result == NULL) result = calc_buffer;
+fp_value *fc_get_plusinf(const float_descriptor_t *desc, fp_value *result)
+{
+ char *mant;
- result->desc.exponent_size = exponent_size;
- result->desc.mantissa_size = mantissa_size;
- result->desc.clss = NORMAL;
+ if (result == NULL) result = calc_buffer;
+ result->desc = *desc;
+ result->clss = FC_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));
+ mant = _mant(result);
+ sc_val_from_ulong(0, mant);
+ if (desc->explicit_one) {
+ sc_set_bit_at(mant, result->desc.mantissa_size + ROUNDING_BITS);
+ }
return result;
}
-fp_value *fc_get_minusinf(unsigned int exponent_size, unsigned int mantissa_size, fp_value *result) {
+fp_value *fc_get_minusinf(const float_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;
}
-int fc_comp(const fp_value *val_a, const fp_value *val_b) {
+int fc_comp(const fp_value *val_a, const fp_value *val_b)
+{
int mul = 1;
/*
* Unordered if NaN or equal
*/
if (val_a == val_b)
- return val_a->desc.clss == NAN ? 2 : 0;
+ return val_a->clss == FC_NAN ? 2 : 0;
/* unordered if one is a NaN */
- if (val_a->desc.clss == NAN || val_b->desc.clss == NAN)
+ if (val_a->clss == FC_NAN || val_b->clss == FC_NAN)
return 2;
/* zero is equal independent of sign */
- if ((val_a->desc.clss == ZERO) && (val_b->desc.clss == ZERO))
+ if ((val_a->clss == FC_ZERO) && (val_b->clss == FC_ZERO))
return 0;
/* different signs make compare easy */
mul = val_a->sign ? -1 : 1;
/* both infinity means equality */
- if ((val_a->desc.clss == INF) && (val_b->desc.clss == INF))
+ if ((val_a->clss == FC_INF) && (val_b->clss == FC_INF))
return 0;
/* infinity is bigger than the rest */
- if (val_a->desc.clss == INF)
+ if (val_a->clss == FC_INF)
return 1 * mul;
- if (val_b->desc.clss == INF)
+ if (val_b->clss == FC_INF)
return -1 * mul;
/* check first exponent, that mantissa if equal */
}
}
-int fc_is_zero(const fp_value *a) {
- return a->desc.clss == ZERO;
+int fc_is_zero(const fp_value *a)
+{
+ return a->clss == FC_ZERO;
}
-int fc_is_negative(const fp_value *a) {
+int fc_is_negative(const fp_value *a)
+{
return a->sign;
}
-int fc_is_inf(const fp_value *a) {
- return a->desc.clss == INF;
+int fc_is_inf(const fp_value *a)
+{
+ return a->clss == FC_INF;
}
-int fc_is_nan(const fp_value *a) {
- return a->desc.clss == NAN;
+int fc_is_nan(const fp_value *a)
+{
+ return a->clss == FC_NAN;
}
-int fc_is_subnormal(const fp_value *a) {
- return a->desc.clss == SUBNORMAL;
+int fc_is_subnormal(const fp_value *a)
+{
+ return a->clss == FC_SUBNORMAL;
}
-char *fc_print(const fp_value *val, char *buf, int buflen, unsigned base) {
+char *fc_print(const fp_value *val, char *buf, int buflen, unsigned base)
+{
char *mul_1;
+ long double flt_val;
- mul_1 = alloca(calc_buffer_size);
+ mul_1 = (char*) alloca(calc_buffer_size);
switch (base) {
case FC_DEC:
- switch (val->desc.clss) {
- case INF:
- if (buflen >= 8 + val->sign) sprintf(buf, "%sINFINITY", val->sign ? "-":"");
- else snprintf(buf, buflen, "%sINF", val->sign ? "-":NULL);
+ switch ((value_class_t)val->clss) {
+ case FC_INF:
+ snprintf(buf, buflen, "%cINF", val->sign ? '-' : '+');
break;
- case NAN:
- snprintf(buf, buflen, "NAN");
+ case FC_NAN:
+ snprintf(buf, buflen, "NaN");
break;
- case ZERO:
+ case FC_ZERO:
snprintf(buf, buflen, "0.0");
break;
default:
- /* XXX to be implemented */
-#ifdef HAVE_LONG_DOUBLE
+ flt_val = fc_val_to_ieee754(val);
/* XXX 30 is arbitrary */
- snprintf(buf, buflen, "%.30LE", fc_val_to_ieee754(val));
-#else
- snprintf(buf, buflen, "%.18E", fc_val_to_ieee754(val));
-#endif
+ snprintf(buf, buflen, "%.30LE", flt_val);
}
break;
case FC_HEX:
- switch (val->desc.clss) {
- case INF:
- if (buflen >= 8+val->sign) sprintf(buf, "%sINFINITY", val->sign?"-":"");
- else snprintf(buf, buflen, "%sINF", val->sign?"-":NULL);
+ switch ((value_class_t)val->clss) {
+ case FC_INF:
+ snprintf(buf, buflen, "%cINF", val->sign ? '-' : '+');
break;
- case NAN:
- snprintf(buf, buflen, "NAN");
+ case FC_NAN:
+ snprintf(buf, buflen, "NaN");
break;
- case ZERO:
+ case FC_ZERO:
snprintf(buf, buflen, "0.0");
break;
default:
-#ifdef HAVE_LONG_DOUBLE
- snprintf(buf, buflen, "%LA", fc_val_to_ieee754(val));
-#else
- snprintf(buf, buflen, "%A", fc_val_to_ieee754(val));
-#endif
+ flt_val = fc_val_to_ieee754(val);
+ snprintf(buf, buflen, "%LA", flt_val);
}
break;
return buf;
}
-unsigned char fc_sub_bits(const fp_value *value, unsigned num_bits, unsigned byte_ofs) {
+unsigned char fc_sub_bits(const fp_value *value, unsigned num_bits, unsigned byte_ofs)
+{
/* 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);
return sc_sub_bits(packed_value, num_bits, byte_ofs);
}
-int fc_zero_mantissa(const fp_value *value) {
- return sc_get_lowest_set_bit(_mant(value)) == 2 + value->desc.mantissa_size;
+/* 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)) == ROUNDING_BITS + value->desc.mantissa_size;
}
-int fc_get_exponent(const fp_value *value) {
+/* 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 float_descriptor_t *desc)
+{
+ int v;
+ int exp_bias;
+
+ /* handle some special cases first */
+ switch (value->clss) {
+ case FC_ZERO:
+ case FC_INF:
+ case FC_NAN:
+ return 1;
+ default:
+ break;
+ }
-fc_rounding_mode_t fc_set_rounding_mode(fc_rounding_mode_t mode) {
+ /* 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 <= (int)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)
rounding_mode = mode;
return rounding_mode;
}
-fc_rounding_mode_t fc_get_rounding_mode(void) {
+fc_rounding_mode_t fc_get_rounding_mode(void)
+{
return rounding_mode;
}
-void init_fltcalc(int precision) {
+void init_fltcalc(int precision)
+{
if (calc_buffer == NULL) {
/* 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);
value_size = sc_get_buffer_length();
calc_buffer_size = sizeof(fp_value) + 2*value_size - 1;
- calc_buffer = xmalloc(calc_buffer_size);
+ calc_buffer = (fp_value*) xmalloc(calc_buffer_size);
memset(calc_buffer, 0, calc_buffer_size);
DEBUGPRINTF(("init fltcalc:\n\tVALUE_SIZE = %d\ntCALC_BUFFER_SIZE = %d\n\tcalc_buffer = %p\n\n", value_size, calc_buffer_size, calc_buffer));
-#ifdef HAVE_LONG_DOUBLE
- DEBUGPRINTF(("\tUsing long double (1-15-64) interface\n"));
-#else
- DEBUGPRINTF(("\tUsing double (1-11-52) interface\n"));
-#endif
-#ifdef WORDS_BIGENDIAN
- DEBUGPRINTF(("\tWord order is big endian\n\n"));
-#else
- DEBUGPRINTF(("\tWord order is little endian\n\n"));
-#endif
}
}
-void finish_fltcalc (void) {
+void finish_fltcalc (void)
+{
free(calc_buffer); calc_buffer = NULL;
}
#endif
/* definition of interface functions */
-fp_value *fc_add(const fp_value *a, const fp_value *b, fp_value *result) {
+fp_value *fc_add(const fp_value *a, const fp_value *b, fp_value *result)
+{
if (result == NULL) result = calc_buffer;
TRACEPRINTF(("%s ", fc_print(a, buffer, sizeof(buffer), FC_PACKED)));
return result;
}
-fp_value *fc_sub(const fp_value *a, const fp_value *b, fp_value *result) {
+fp_value *fc_sub(const fp_value *a, const fp_value *b, fp_value *result)
+{
fp_value *temp;
if (result == NULL) result = calc_buffer;
TRACEPRINTF(("%s ", fc_print(a, buffer, sizeof(buffer), FC_PACKED)));
TRACEPRINTF(("- %s ", fc_print(b, buffer, sizeof(buffer), FC_PACKED)));
- temp = alloca(calc_buffer_size);
+ temp = (fp_value*) alloca(calc_buffer_size);
memcpy(temp, b, calc_buffer_size);
temp->sign = !b->sign;
if (sc_comp(_exp(a), _exp(temp)) == -1)
return result;
}
-fp_value *fc_mul(const fp_value *a, const fp_value *b, fp_value *result) {
+fp_value *fc_mul(const fp_value *a, const fp_value *b, fp_value *result)
+{
if (result == NULL) result = calc_buffer;
TRACEPRINTF(("%s ", fc_print(a, buffer, sizeof(buffer), FC_PACKED)));
return result;
}
-fp_value *fc_div(const fp_value *a, const fp_value *b, fp_value *result) {
+fp_value *fc_div(const fp_value *a, const fp_value *b, fp_value *result)
+{
if (result == NULL) result = calc_buffer;
TRACEPRINTF(("%s ", fc_print(a, buffer, sizeof(buffer), FC_PACKED)));
return result;
}
-fp_value *fc_neg(const fp_value *a, fp_value *result) {
+fp_value *fc_neg(const fp_value *a, fp_value *result)
+{
if (result == NULL) result = calc_buffer;
TRACEPRINTF(("- %s ", fc_print(a, buffer, sizeof(buffer), FC_PACKED)));
return result;
}
-fp_value *fc_int(const fp_value *a, fp_value *result) {
+fp_value *fc_int(const fp_value *a, fp_value *result)
+{
if (result == NULL) result = calc_buffer;
TRACEPRINTF(("%s ", fc_print(a, buffer, sizeof(buffer), FC_PACKED)));
return result;
}
-fp_value *fc_rnd(const fp_value *a, fp_value *result) {
- if (result == NULL) result = calc_buffer;
-
- (void) a;
+fp_value *fc_rnd(const fp_value *a, fp_value *result)
+{
+ (void)a;
+ (void)result;
TRACEPRINTF(("%s ", fc_print(a, buffer, sizeof(buffer), FC_PACKED)));
TRACEPRINTF(("rounded to integer "));
- assert(!"fc_rnd() not yet implemented");
-
- TRACEPRINTF(("= %s\n", fc_print(result, buffer, sizeof(buffer), FC_PACKED)));
- return result;
+ panic("not yet implemented");
}
/*
* 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 fc_flt2int(const fp_value *a, void *result, ir_mode *dst_mode)
+{
+ if (a->clss == FC_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;
+ int tgt_bits;
if (a->sign && !mode_is_signed(dst_mode)) {
/* FIXME: for now we cannot convert this */
return 0;
}
+ tgt_bits = get_mode_size_bits(dst_mode);
+ if (mode_is_signed(dst_mode))
+ --tgt_bits;
+
assert(exp_val >= 0 && "floating point value not integral before fc_flt2int() call");
- shift = exp_val - a->desc.mantissa_size - 2;
+ mantissa_size = a->desc.mantissa_size + ROUNDING_BITS;
+ shift = exp_val - mantissa_size;
+ if (tgt_bits < mantissa_size + 1)
+ tgt_bits = mantissa_size + 1;
if (shift > 0) {
- sc_shlI(_mant(a), shift, 64, 0, result);
+ sc_shlI(_mant(a), shift, tgt_bits, 0, result);
} else {
- sc_shrI(_mant(a), -shift, 64, 0, result);
+ sc_shrI(_mant(a), -shift, tgt_bits, 0, result);
}
/* check for overflow */
sc_neg(result, result);
return 1;
- }
- else if (a->desc.clss == ZERO) {
+ } else if (a->clss == FC_ZERO) {
sc_zero(result);
return 1;
}
return 0;
}
-
-unsigned fc_set_immediate_precision(unsigned bits) {
- unsigned old = immediate_prec;
-
- immediate_prec = bits;
- return old;
-}
-
-int fc_is_exact(void) {
+int fc_is_exact(void)
+{
return fc_exact;
}