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
27 #ifndef FIRM_TV_FLTCALC_H
28 #define FIRM_TV_FLTCALC_H
31 #include "firm_types.h"
41 /** IEEE-754 Rounding modes. */
43 FC_TONEAREST, /**< if unsure, to the nearest even */
44 FC_TOPOSITIVE, /**< to +oo */
45 FC_TONEGATIVE, /**< to -oo */
46 FC_TOZERO /**< to 0 */
49 #define FC_DEFAULT_PRECISION 64
52 * possible float states
55 FC_NORMAL, /**< normal representation, implicit 1 */
57 FC_SUBNORMAL, /**< denormals, implicit 0 */
59 FC_NAN, /**< Not A Number */
63 typedef struct fp_value fp_value;
66 /** internal buffer access
67 * All functions that accept NULL as return buffer put their result into an
69 * @return fc_get_buffer() returns the pointer to the buffer, fc_get_buffer_length()
70 * returns the size of this buffer
72 const void *fc_get_buffer(void);
73 int fc_get_buffer_length(void);
76 void *fc_val_from_str(const char *str, size_t len, const float_descriptor_t *desc, void *result);
78 /** get the representation of a floating point value
79 * This function tries to builds a representation having the same value as the
80 * float number passed.
81 * If the wished precision is less than the precision of long double the value
82 * built will be rounded. Therefore only an approximation of the passed float
83 * can be expected in this case.
85 * @param l The floating point number to build a representation for
86 * @param desc The floating point descriptor
87 * @param result A buffer to hold the value built. If this is NULL, the internal
88 * accumulator buffer is used. Note that the buffer must be big
89 * enough to hold the value. Use fc_get_buffer_length() to find out
92 * @return The result pointer passed to the function. If this was NULL this returns
93 * a pointer to the internal accumulator buffer
95 fp_value *fc_val_from_ieee754(long double l, const float_descriptor_t *desc,
98 /** retrieve the float value of an internal value
99 * This function casts the internal value to long double and returns a
100 * long double with that value.
101 * This implies that values of higher precision than long double are subject to
102 * rounding, so the returned value might not the same than the actually
105 * @param val The representation of a float value
107 * @return a float value approximating the represented value
109 long double fc_val_to_ieee754(const fp_value *val);
111 /** cast a value to another precision
112 * This function changes the precision of a float representation.
113 * If the new precision is less than the original precision the returned
114 * value might not be the same as the original value.
116 * @param val The value to be casted
117 * @param desc The floating point descriptor
118 * @param result A buffer to hold the value built. If this is NULL, the internal
119 * accumulator buffer is used. Note that the buffer must be big
120 * enough to hold the value. Use fc_get_buffer_length() to find out
122 * @return The result pointer passed to the function. If this was NULL this returns
123 * a pointer to the internal accumulator buffer
125 fp_value *fc_cast(const fp_value *val, const float_descriptor_t *desc, fp_value *result);
128 /** build a special float value
129 * This function builds a representation for a special float value, as indicated by the
132 * @param desc The floating point descriptor
133 * @param result A buffer to hold the value built. If this is NULL, the internal
134 * accumulator buffer is used. Note that the buffer must be big
135 * enough to hold the value. Use fc_get_buffer_length() to find out
137 * @return The result pointer passed to the function. If this was NULL this returns
138 * a pointer to the internal accumulator buffer
140 fp_value *fc_get_min(const float_descriptor_t *desc, fp_value *result);
141 fp_value *fc_get_max(const float_descriptor_t *desc, fp_value *result);
142 fp_value *fc_get_snan(const float_descriptor_t *desc, fp_value *result);
143 fp_value *fc_get_qnan(const float_descriptor_t *desc, fp_value *result);
144 fp_value *fc_get_plusinf(const float_descriptor_t *desc, fp_value *result);
145 fp_value *fc_get_minusinf(const float_descriptor_t *desc, fp_value *result);
148 int fc_is_zero(const fp_value *a);
149 int fc_is_negative(const fp_value *a);
150 int fc_is_inf(const fp_value *a);
151 int fc_is_nan(const fp_value *a);
152 int fc_is_subnormal(const fp_value *a);
154 fp_value *fc_add(const fp_value *a, const fp_value *b, fp_value *result);
155 fp_value *fc_sub(const fp_value *a, const fp_value *b, fp_value *result);
156 fp_value *fc_mul(const fp_value *a, const fp_value *b, fp_value *result);
157 fp_value *fc_div(const fp_value *a, const fp_value *b, fp_value *result);
158 fp_value *fc_neg(const fp_value *a, fp_value *result);
159 fp_value *fc_int(const fp_value *a, fp_value *result);
160 fp_value *fc_rnd(const fp_value *a, fp_value *result);
162 char *fc_print(const fp_value *a, char *buf, int buflen, unsigned base);
164 /** Compare two values
165 * This function compares two values
167 * @param a Value No. 1
168 * @param b Value No. 2
169 * @result The returned value will be one of
173 * 2 if either value is NaN
175 int fc_comp(const fp_value *a, const fp_value *b);
178 * Converts an floating point value into an integer value.
180 int fc_flt2int(const fp_value *a, void *result, ir_mode *dst_mode);
183 * Returns non-zero if the mantissa is zero, i.e. 1.0Exxx
185 int fc_zero_mantissa(const fp_value *value);
188 * Returns the exponent of a value.
190 int fc_get_exponent(const fp_value *value);
193 * Return non-zero if a given value can be converted lossless into another precision.
195 int fc_can_lossless_conv_to(const fp_value *value, const float_descriptor_t *desc);
197 /** Set new rounding mode
198 * This function sets the rounding mode to one of the following, returning
199 * the previously set rounding mode.
200 * FC_TONEAREST (default):
201 * Any unrepresentable value is rounded to the nearest representable
202 * value. If it lies in the middle the value with the least significant
203 * bit of zero is chosen (the even one).
204 * Values too big to represent will round to +/-infinity.
206 * Any unrepresentable value is rounded towards negative infinity.
207 * Positive values too big to represent will round to the biggest
208 * representable value, negative values too small to represent will
209 * round to -infinity.
211 * Any unrepresentable value is rounded towards positive infinity
212 * Negative values too small to represent will round to the biggest
213 * representable value, positive values too big to represent will
214 * round to +infinity.
216 * Any unrepresentable value is rounded towards zero, effectively
217 * chopping off any bits beyond the mantissa size.
218 * Values too big to represent will round to the biggest/smallest
219 * representable value.
221 * These modes correspond to the modes required by the IEEE-754 standard.
223 * @param mode The new rounding mode. Any value other than the four
224 * defined values will have no effect.
225 * @return The previous rounding mode.
227 * @see fc_get_rounding_mode()
228 * @see IEEE754, IEEE854 Floating Point Standard
230 fc_rounding_mode_t fc_set_rounding_mode(fc_rounding_mode_t mode);
232 /** Get the rounding mode
233 * This function retrieves the currently used rounding mode
235 * @return The current rounding mode
236 * @see fc_set_rounding_mode()
238 fc_rounding_mode_t fc_get_rounding_mode(void);
240 /** Get bit representation of a value
241 * This function allows to read a value in encoded form, byte wise.
242 * The value will be packed corresponding to the way used by the IEEE
243 * encoding formats, i.e.
245 * exp_size bits exponent + bias
246 * mant_size bits mantissa, without leading 1
248 * As in IEEE, an exponent of 0 indicates a denormalized number, which
249 * implies a most significant bit of zero instead of one; an exponent
250 * of all ones (2**exp_size - 1) encodes infinity if the mantissa is
251 * all zeros, else Not A Number.
253 * @param val A pointer to the value. If NULL is passed a copy of the
254 * most recent value passed to this function is used, saving the
255 * packing step. This behavior may be changed in the future.
256 * @param num_bit The maximum number of bits to return. Any bit beyond
257 * num_bit will be returned as zero.
258 * @param byte_ofs The byte index to read, 0 is the least significant
260 * @return 8 bits of encoded data
262 unsigned char fc_sub_bits(const fp_value *val, unsigned num_bit, unsigned byte_ofs);
265 * Returns non-zero if the result of the last operation was exact.
267 int fc_is_exact(void);
269 void init_fltcalc(int precision);
270 void finish_fltcalc(void);
272 #endif /* FIRM_TV_FLTCALC_H */