2 * Copyright (C) 1995-2007 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_config.h"
31 #include "firm_types.h"
33 #ifdef HAVE_LONG_DOUBLE
34 /* XXX Set this via autoconf */
35 #define HAVE_EXPLICIT_ONE
36 typedef long double LLDBL;
48 /** IEEE-754 Rounding modes. */
50 FC_TONEAREST, /**< if unsure, to the nearest even */
51 FC_TOPOSITIVE, /**< to +oo */
52 FC_TONEGATIVE, /**< to -oo */
53 FC_TOZERO /**< to 0 */
56 #define FC_DEFAULT_PRECISION 64
58 typedef struct _fp_value fp_value;
61 /** internal buffer access
62 * All functions that accept NULL as return buffer put their result into an
64 * @return fc_get_buffer() returns the pointer to the buffer, fc_get_buffer_length()
65 * returns the size of this buffer
67 const void *fc_get_buffer(void);
68 int fc_get_buffer_length(void);
71 void *fc_val_from_str(const char *str, unsigned int len, char exp_size, char mant_size, void *result);
73 /** get the representation of a floating point value
74 * This function tries to builds a representation having the same value as the
75 * float number passed.
76 * If the wished precision is less than the precision of LLDBL the value built
77 * will be rounded. Therefore only an approximation of the passed float can be
78 * expected in this case.
80 * @param l The floating point number to build a representation for
81 * @param exp_size The number of bits of the new exponent
82 * @param mant_size The number of bits of the new mantissa
83 * @param result A buffer to hold the value built. If this is NULL, the internal
84 * accumulator buffer is used. Note that the buffer must be big
85 * enough to hold the value. Use fc_get_buffer_length() to find out
87 * @return The result pointer passed to the function. If this was NULL this returns
88 * a pointer to the internal accumulator buffer
90 fp_value *fc_val_from_ieee754(LLDBL l, char exp_size, char mant_size, fp_value *result);
92 /** retrieve the float value of an internal value
93 * This function casts the internal value to LLDBL and returns a LLDBL with
95 * This implies that values of higher precision than LLDBL are subject to
96 * rounding, so the returned value might not the same than the actually
99 * @param val The representation of a float value
100 * @return a float value approximating the represented value
102 LLDBL fc_val_to_ieee754(const fp_value *val);
104 /** cast a value to another precision
105 * This function changes the precision of a float representation.
106 * If the new precision is less than the original precision the returned
107 * value might not be the same as the original value.
109 * @param val The value to be casted
110 * @param exp_size The number of bits of the new exponent
111 * @param mant_size The number of bits of the new mantissa
112 * @param result A buffer to hold the value built. If this is NULL, the internal
113 * accumulator buffer is used. Note that the buffer must be big
114 * enough to hold the value. Use fc_get_buffer_length() to find out
116 * @return The result pointer passed to the function. If this was NULL this returns
117 * a pointer to the internal accumulator buffer
119 fp_value *fc_cast(const fp_value *val, char exp_size, char mant_size, fp_value *result);
122 /** build a special float value
123 * This function builds a representation for a special float value, as indicated by the
126 * @param exponent_size The number of bits of exponent of the float type the value
128 * @param mantissa_size The number of bits of mantissa of the float type the value
130 * @param result A buffer to hold the value built. If this is NULL, the internal
131 * accumulator buffer is used. Note that the buffer must be big
132 * enough to hold the value. Use fc_get_buffer_length() to find out
134 * @return The result pointer passed to the function. If this was NULL this returns
135 * a pointer to the internal accumulator buffer
137 fp_value *fc_get_min(unsigned int exponent_size, unsigned int mantissa_size, fp_value *result);
138 fp_value *fc_get_max(unsigned int exponent_size, unsigned int mantissa_size, fp_value *result);
139 fp_value *fc_get_snan(unsigned int exponent_size, unsigned int mantissa_size, fp_value *result);
140 fp_value *fc_get_qnan(unsigned int exponent_size, unsigned int mantissa_size, fp_value *result);
141 fp_value *fc_get_plusinf(unsigned int exponent_size, unsigned int mantissa_size, fp_value *result);
142 fp_value *fc_get_minusinf(unsigned int exponent_size, unsigned int mantissa_size, fp_value *result);
145 int fc_is_zero(const fp_value *a);
146 int fc_is_negative(const fp_value *a);
147 int fc_is_inf(const fp_value *a);
148 int fc_is_nan(const fp_value *a);
149 int fc_is_subnormal(const fp_value *a);
151 fp_value *fc_add(const fp_value *a, const fp_value *b, fp_value *result);
152 fp_value *fc_sub(const fp_value *a, const fp_value *b, fp_value *result);
153 fp_value *fc_mul(const fp_value *a, const fp_value *b, fp_value *result);
154 fp_value *fc_div(const fp_value *a, const fp_value *b, fp_value *result);
155 fp_value *fc_neg(const fp_value *a, fp_value *result);
156 fp_value *fc_int(const fp_value *a, fp_value *result);
157 fp_value *fc_rnd(const fp_value *a, fp_value *result);
159 char *fc_print(const fp_value *a, char *buf, int buflen, unsigned base);
161 /** Compare two values
162 * This function compares two values
164 * @param a Value No. 1
165 * @param b Value No. 2
166 * @result The returned value will be one of
170 * 2 if either value is NaN
172 int fc_comp(const fp_value *a, const fp_value *b);
175 * Converts an floating point value into an integer value.
177 int fc_flt2int(const fp_value *a, void *result, ir_mode *dst_mode);
180 * Returns non-zero if the mantissa is zero, i.e. 1.0Exxx
182 int fc_zero_mantissa(const fp_value *value);
185 * Returns the exponent of a value.
187 int fc_get_exponent(const fp_value *value);
189 /** Set new rounding mode
190 * This function sets the rounding mode to one of the following, returning
191 * the previously set rounding mode.
192 * FC_TONEAREST (default):
193 * Any unrepresentable value is rounded to the nearest representable
194 * value. If it lies in the middle the value with the least significant
195 * bit of zero is chosen (the even one).
196 * Values too big to represent will round to +/-infinity.
198 * Any unrepresentable value is rounded towards negative infinity.
199 * Positive values too big to represent will round to the biggest
200 * representable value, negative values too small to represent will
201 * round to -infinity.
203 * Any unrepresentable value is rounded towards positive infinity
204 * Negative values too small to represent will round to the biggest
205 * representable value, positive values too big to represent will
206 * round to +infinity.
208 * Any unrepresentable value is rounded towards zero, effectively
209 * chopping off any bits beyond the mantissa size.
210 * Values too big to represent will round to the biggest/smallest
211 * representable value.
213 * These modes correspond to the modes required by the IEEE-754 standard.
215 * @param mode The new rounding mode. Any value other than the four
216 * defined values will have no effect.
217 * @return The previous rounding mode.
219 * @see fc_get_rounding_mode()
220 * @see IEEE754, IEEE854 Floating Point Standard
222 fc_rounding_mode_t fc_set_rounding_mode(fc_rounding_mode_t mode);
224 /** Get the rounding mode
225 * This function retrieves the currently used rounding mode
227 * @return The current rounding mode
228 * @see fc_set_rounding_mode()
230 fc_rounding_mode_t fc_get_rounding_mode(void);
232 /** Get bit representation of a value
233 * This function allows to read a value in encoded form, byte wise.
234 * The value will be packed corresponding to the way used by the IEEE
235 * encoding formats, i.e.
237 * exp_size bits exponent + bias
238 * mant_size bits mantissa, without leading 1
240 * As in IEEE, an exponent of 0 indicates a denormalized number, which
241 * implies a most significant bit of zero instead of one; an exponent
242 * of all ones (2**exp_size - 1) encodes infinity if the mantissa is
243 * all zeros, else Not A Number.
245 * @param val A pointer to the value. If NULL is passed a copy of the
246 * most recent value passed to this function is used, saving the
247 * packing step. This behavior may be changed in the future.
248 * @param num_bit The maximum number of bits to return. Any bit beyond
249 * num_bit will be returned as zero.
250 * @param byte_ofs The byte index to read, 0 is the least significant
252 * @return 8 bits of encoded data
254 unsigned char fc_sub_bits(const fp_value *val, unsigned num_bit, unsigned byte_ofs);
257 * Set the immediate precision for IEEE-754 results. Set this to
258 * 0 to get the same precision as the operands.
259 * For x87 compatibility, set this to 80.
261 * @return the old setting
263 unsigned fc_set_immediate_precision(unsigned bits);
266 * Returns non-zero if the result of the last operation was exact.
268 int fc_is_exact(void);
270 void init_fltcalc(int precision);
271 void finish_fltcalc(void);
273 #endif /* FIRM_TV_FLTCALC_H */