2 * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief tarval floating point calculations
24 * @author Mathias Heil
26 #ifndef FIRM_TV_FLTCALC_H
27 #define FIRM_TV_FLTCALC_H
30 #include "firm_types.h"
40 /** IEEE-754 Rounding modes. */
42 FC_TONEAREST, /**< if unsure, to the nearest even */
43 FC_TOPOSITIVE, /**< to +oo */
44 FC_TONEGATIVE, /**< to -oo */
45 FC_TOZERO /**< to 0 */
48 #define FC_DEFAULT_PRECISION 64
51 * possible float states
54 FC_NORMAL, /**< normal representation, implicit 1 */
56 FC_SUBNORMAL, /**< denormals, implicit 0 */
58 FC_NAN, /**< Not A Number */
62 typedef struct fp_value fp_value;
65 /** internal buffer access
66 * All functions that accept NULL as return buffer put their result into an
68 * @return fc_get_buffer() returns the pointer to the buffer, fc_get_buffer_length()
69 * returns the size of this buffer
71 const void *fc_get_buffer(void);
72 int fc_get_buffer_length(void);
75 void *fc_val_from_str(const char *str, size_t len, const float_descriptor_t *desc, void *result);
77 /** get the representation of a floating point value
78 * This function tries to builds a representation having the same value as the
79 * float number passed.
80 * If the wished precision is less than the precision of long double the value
81 * built will be rounded. Therefore only an approximation of the passed float
82 * can be expected in this case.
84 * @param l The floating point number to build a representation for
85 * @param desc The floating point descriptor
86 * @param result A buffer to hold the value built. If this is NULL, the internal
87 * accumulator buffer is used. Note that the buffer must be big
88 * enough to hold the value. Use fc_get_buffer_length() to find out
91 * @return The result pointer passed to the function. If this was NULL this returns
92 * a pointer to the internal accumulator buffer
94 fp_value *fc_val_from_ieee754(long double l, const float_descriptor_t *desc,
97 /** retrieve the float value of an internal value
98 * This function casts the internal value to long double and returns a
99 * long double with that value.
100 * This implies that values of higher precision than long double are subject to
101 * rounding, so the returned value might not the same than the actually
104 * @param val The representation of a float value
106 * @return a float value approximating the represented value
108 long double fc_val_to_ieee754(const fp_value *val);
110 /** cast a value to another precision
111 * This function changes the precision of a float representation.
112 * If the new precision is less than the original precision the returned
113 * value might not be the same as the original value.
115 * @param val The value to be casted
116 * @param desc The floating point descriptor
117 * @param result A buffer to hold the value built. If this is NULL, the internal
118 * accumulator buffer is used. Note that the buffer must be big
119 * enough to hold the value. Use fc_get_buffer_length() to find out
121 * @return The result pointer passed to the function. If this was NULL this returns
122 * a pointer to the internal accumulator buffer
124 fp_value *fc_cast(const fp_value *val, const float_descriptor_t *desc, fp_value *result);
127 /** build a special float value
128 * This function builds a representation for a special float value, as indicated by the
131 * @param desc The floating point descriptor
132 * @param result A buffer to hold the value built. If this is NULL, the internal
133 * accumulator buffer is used. Note that the buffer must be big
134 * enough to hold the value. Use fc_get_buffer_length() to find out
136 * @return The result pointer passed to the function. If this was NULL this returns
137 * a pointer to the internal accumulator buffer
139 fp_value *fc_get_min(const float_descriptor_t *desc, fp_value *result);
140 fp_value *fc_get_max(const float_descriptor_t *desc, fp_value *result);
141 fp_value *fc_get_snan(const float_descriptor_t *desc, fp_value *result);
142 fp_value *fc_get_qnan(const float_descriptor_t *desc, fp_value *result);
143 fp_value *fc_get_plusinf(const float_descriptor_t *desc, fp_value *result);
144 fp_value *fc_get_minusinf(const float_descriptor_t *desc, fp_value *result);
147 int fc_is_zero(const fp_value *a);
148 int fc_is_negative(const fp_value *a);
149 int fc_is_inf(const fp_value *a);
150 int fc_is_nan(const fp_value *a);
151 int fc_is_subnormal(const fp_value *a);
153 fp_value *fc_add(const fp_value *a, const fp_value *b, fp_value *result);
154 fp_value *fc_sub(const fp_value *a, const fp_value *b, fp_value *result);
155 fp_value *fc_mul(const fp_value *a, const fp_value *b, fp_value *result);
156 fp_value *fc_div(const fp_value *a, const fp_value *b, fp_value *result);
157 fp_value *fc_neg(const fp_value *a, fp_value *result);
158 fp_value *fc_int(const fp_value *a, fp_value *result);
159 fp_value *fc_rnd(const fp_value *a, fp_value *result);
161 char *fc_print(const fp_value *a, char *buf, int buflen, unsigned base);
163 /** Compare two values
164 * This function compares two values
166 * @param a Value No. 1
167 * @param b Value No. 2
168 * @result The returned value will be one of
172 * 2 if either value is NaN
174 int fc_comp(const fp_value *a, const fp_value *b);
177 * Converts an floating point value into an integer value.
179 int fc_flt2int(const fp_value *a, void *result, ir_mode *dst_mode);
182 * Returns non-zero if the mantissa is zero, i.e. 1.0Exxx
184 int fc_zero_mantissa(const fp_value *value);
187 * Returns the exponent of a value.
189 int fc_get_exponent(const fp_value *value);
192 * Return non-zero if a given value can be converted lossless into another precision.
194 int fc_can_lossless_conv_to(const fp_value *value, const float_descriptor_t *desc);
196 /** Set new rounding mode
197 * This function sets the rounding mode to one of the following, returning
198 * the previously set rounding mode.
199 * FC_TONEAREST (default):
200 * Any unrepresentable value is rounded to the nearest representable
201 * value. If it lies in the middle the value with the least significant
202 * bit of zero is chosen (the even one).
203 * Values too big to represent will round to +/-infinity.
205 * Any unrepresentable value is rounded towards negative infinity.
206 * Positive values too big to represent will round to the biggest
207 * representable value, negative values too small to represent will
208 * round to -infinity.
210 * Any unrepresentable value is rounded towards positive infinity
211 * Negative values too small to represent will round to the biggest
212 * representable value, positive values too big to represent will
213 * round to +infinity.
215 * Any unrepresentable value is rounded towards zero, effectively
216 * chopping off any bits beyond the mantissa size.
217 * Values too big to represent will round to the biggest/smallest
218 * representable value.
220 * These modes correspond to the modes required by the IEEE-754 standard.
222 * @param mode The new rounding mode. Any value other than the four
223 * defined values will have no effect.
224 * @return The previous rounding mode.
226 * @see fc_get_rounding_mode()
227 * @see IEEE754, IEEE854 Floating Point Standard
229 fc_rounding_mode_t fc_set_rounding_mode(fc_rounding_mode_t mode);
231 /** Get the rounding mode
232 * This function retrieves the currently used rounding mode
234 * @return The current rounding mode
235 * @see fc_set_rounding_mode()
237 fc_rounding_mode_t fc_get_rounding_mode(void);
239 /** Get bit representation of a value
240 * This function allows to read a value in encoded form, byte wise.
241 * The value will be packed corresponding to the way used by the IEEE
242 * encoding formats, i.e.
244 * exp_size bits exponent + bias
245 * mant_size bits mantissa, without leading 1
247 * As in IEEE, an exponent of 0 indicates a denormalized number, which
248 * implies a most significant bit of zero instead of one; an exponent
249 * of all ones (2**exp_size - 1) encodes infinity if the mantissa is
250 * all zeros, else Not A Number.
252 * @param val A pointer to the value. If NULL is passed a copy of the
253 * most recent value passed to this function is used, saving the
254 * packing step. This behavior may be changed in the future.
255 * @param num_bit The maximum number of bits to return. Any bit beyond
256 * num_bit will be returned as zero.
257 * @param byte_ofs The byte index to read, 0 is the least significant
259 * @return 8 bits of encoded data
261 unsigned char fc_sub_bits(const fp_value *val, unsigned num_bit, unsigned byte_ofs);
264 * Returns non-zero if the result of the last operation was exact.
266 int fc_is_exact(void);
268 void init_fltcalc(int precision);
269 void finish_fltcalc(void);
271 #endif /* FIRM_TV_FLTCALC_H */