2 * Copyright (C) 1995-2008 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
30 #include "firm_types.h"
32 #ifdef HAVE_LONG_DOUBLE
33 /* XXX Set this via autoconf */
34 #define HAVE_EXPLICIT_ONE
35 typedef long double LLDBL;
47 /** IEEE-754 Rounding modes. */
49 FC_TONEAREST, /**< if unsure, to the nearest even */
50 FC_TOPOSITIVE, /**< to +oo */
51 FC_TONEGATIVE, /**< to -oo */
52 FC_TOZERO /**< to 0 */
55 #define FC_DEFAULT_PRECISION 64
58 * possible float states
61 NORMAL, /**< normal representation, implicit 1 */
63 SUBNORMAL, /**< denormals, implicit 0 */
65 NAN, /**< Not A Number */
69 * A descriptor for an IEEE float value.
71 typedef struct ieee_descriptor_t {
72 unsigned char exponent_size; /**< size of exponent in bits */
73 unsigned char mantissa_size; /**< size of mantissa in bits */
74 unsigned char explicit_one; /**< set if the leading one is explicit */
75 unsigned char clss; /**< state of this float */
79 typedef struct _fp_value fp_value;
82 /** internal buffer access
83 * All functions that accept NULL as return buffer put their result into an
85 * @return fc_get_buffer() returns the pointer to the buffer, fc_get_buffer_length()
86 * returns the size of this buffer
88 const void *fc_get_buffer(void);
89 int fc_get_buffer_length(void);
92 void *fc_val_from_str(const char *str, unsigned int len, const ieee_descriptor_t *desc, void *result);
94 /** get the representation of a floating point value
95 * This function tries to builds a representation having the same value as the
96 * float number passed.
97 * If the wished precision is less than the precision of LLDBL the value built
98 * will be rounded. Therefore only an approximation of the passed float can be
99 * expected in this case.
101 * @param l The floating point number to build a representation for
102 * @param desc The floating point descriptor
103 * @param result A buffer to hold the value built. If this is NULL, the internal
104 * accumulator buffer is used. Note that the buffer must be big
105 * enough to hold the value. Use fc_get_buffer_length() to find out
108 * @return The result pointer passed to the function. If this was NULL this returns
109 * a pointer to the internal accumulator buffer
111 fp_value *fc_val_from_ieee754(LLDBL l, const ieee_descriptor_t *desc, fp_value *result);
113 /** retrieve the float value of an internal value
114 * This function casts the internal value to LLDBL and returns a LLDBL with
116 * This implies that values of higher precision than LLDBL are subject to
117 * rounding, so the returned value might not the same than the actually
120 * @param val The representation of a float value
122 * @return a float value approximating the represented value
124 LLDBL fc_val_to_ieee754(const fp_value *val);
126 /** cast a value to another precision
127 * This function changes the precision of a float representation.
128 * If the new precision is less than the original precision the returned
129 * value might not be the same as the original value.
131 * @param val The value to be casted
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_cast(const fp_value *val, const ieee_descriptor_t *desc, fp_value *result);
143 /** build a special float value
144 * This function builds a representation for a special float value, as indicated by the
147 * @param desc The floating point descriptor
148 * @param result A buffer to hold the value built. If this is NULL, the internal
149 * accumulator buffer is used. Note that the buffer must be big
150 * enough to hold the value. Use fc_get_buffer_length() to find out
152 * @return The result pointer passed to the function. If this was NULL this returns
153 * a pointer to the internal accumulator buffer
155 fp_value *fc_get_min(const ieee_descriptor_t *desc, fp_value *result);
156 fp_value *fc_get_max(const ieee_descriptor_t *desc, fp_value *result);
157 fp_value *fc_get_snan(const ieee_descriptor_t *desc, fp_value *result);
158 fp_value *fc_get_qnan(const ieee_descriptor_t *desc, fp_value *result);
159 fp_value *fc_get_plusinf(const ieee_descriptor_t *desc, fp_value *result);
160 fp_value *fc_get_minusinf(const ieee_descriptor_t *desc, fp_value *result);
163 int fc_is_zero(const fp_value *a);
164 int fc_is_negative(const fp_value *a);
165 int fc_is_inf(const fp_value *a);
166 int fc_is_nan(const fp_value *a);
167 int fc_is_subnormal(const fp_value *a);
169 fp_value *fc_add(const fp_value *a, const fp_value *b, fp_value *result);
170 fp_value *fc_sub(const fp_value *a, const fp_value *b, fp_value *result);
171 fp_value *fc_mul(const fp_value *a, const fp_value *b, fp_value *result);
172 fp_value *fc_div(const fp_value *a, const fp_value *b, fp_value *result);
173 fp_value *fc_neg(const fp_value *a, fp_value *result);
174 fp_value *fc_int(const fp_value *a, fp_value *result);
175 fp_value *fc_rnd(const fp_value *a, fp_value *result);
177 char *fc_print(const fp_value *a, char *buf, int buflen, unsigned base);
179 /** Compare two values
180 * This function compares two values
182 * @param a Value No. 1
183 * @param b Value No. 2
184 * @result The returned value will be one of
188 * 2 if either value is NaN
190 int fc_comp(const fp_value *a, const fp_value *b);
193 * Converts an floating point value into an integer value.
195 int fc_flt2int(const fp_value *a, void *result, ir_mode *dst_mode);
198 * Returns non-zero if the mantissa is zero, i.e. 1.0Exxx
200 int fc_zero_mantissa(const fp_value *value);
203 * Returns the exponent of a value.
205 int fc_get_exponent(const fp_value *value);
208 * Return non-zero if a given value can be converted lossless into another precision.
210 int fc_can_lossless_conv_to(const fp_value *value, const ieee_descriptor_t *desc);
212 /** Set new rounding mode
213 * This function sets the rounding mode to one of the following, returning
214 * the previously set rounding mode.
215 * FC_TONEAREST (default):
216 * Any unrepresentable value is rounded to the nearest representable
217 * value. If it lies in the middle the value with the least significant
218 * bit of zero is chosen (the even one).
219 * Values too big to represent will round to +/-infinity.
221 * Any unrepresentable value is rounded towards negative infinity.
222 * Positive values too big to represent will round to the biggest
223 * representable value, negative values too small to represent will
224 * round to -infinity.
226 * Any unrepresentable value is rounded towards positive infinity
227 * Negative values too small to represent will round to the biggest
228 * representable value, positive values too big to represent will
229 * round to +infinity.
231 * Any unrepresentable value is rounded towards zero, effectively
232 * chopping off any bits beyond the mantissa size.
233 * Values too big to represent will round to the biggest/smallest
234 * representable value.
236 * These modes correspond to the modes required by the IEEE-754 standard.
238 * @param mode The new rounding mode. Any value other than the four
239 * defined values will have no effect.
240 * @return The previous rounding mode.
242 * @see fc_get_rounding_mode()
243 * @see IEEE754, IEEE854 Floating Point Standard
245 fc_rounding_mode_t fc_set_rounding_mode(fc_rounding_mode_t mode);
247 /** Get the rounding mode
248 * This function retrieves the currently used rounding mode
250 * @return The current rounding mode
251 * @see fc_set_rounding_mode()
253 fc_rounding_mode_t fc_get_rounding_mode(void);
255 /** Get bit representation of a value
256 * This function allows to read a value in encoded form, byte wise.
257 * The value will be packed corresponding to the way used by the IEEE
258 * encoding formats, i.e.
260 * exp_size bits exponent + bias
261 * mant_size bits mantissa, without leading 1
263 * As in IEEE, an exponent of 0 indicates a denormalized number, which
264 * implies a most significant bit of zero instead of one; an exponent
265 * of all ones (2**exp_size - 1) encodes infinity if the mantissa is
266 * all zeros, else Not A Number.
268 * @param val A pointer to the value. If NULL is passed a copy of the
269 * most recent value passed to this function is used, saving the
270 * packing step. This behavior may be changed in the future.
271 * @param num_bit The maximum number of bits to return. Any bit beyond
272 * num_bit will be returned as zero.
273 * @param byte_ofs The byte index to read, 0 is the least significant
275 * @return 8 bits of encoded data
277 unsigned char fc_sub_bits(const fp_value *val, unsigned num_bit, unsigned byte_ofs);
280 * Set the immediate precision for IEEE-754 results. Set this to
281 * 0 to get the same precision as the operands.
282 * For x87 compatibility, set this to 80.
284 * @return the old setting
286 unsigned fc_set_immediate_precision(unsigned bits);
289 * Returns non-zero if the result of the last operation was exact.
291 int fc_is_exact(void);
293 void init_fltcalc(int precision);
294 void finish_fltcalc(void);
296 #endif /* FIRM_TV_FLTCALC_H */