1 | //===-- A class to store a normalized floating point number -----*- C++ -*-===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | |
9 | #ifndef LLVM_LIBC_SRC___SUPPORT_FPUTIL_NORMALFLOAT_H |
10 | #define LLVM_LIBC_SRC___SUPPORT_FPUTIL_NORMALFLOAT_H |
11 | |
12 | #include "FPBits.h" |
13 | |
14 | #include "src/__support/CPP/type_traits.h" |
15 | #include "src/__support/common.h" |
16 | |
17 | #include <stdint.h> |
18 | |
19 | namespace LIBC_NAMESPACE { |
20 | namespace fputil { |
21 | |
22 | // A class which stores the normalized form of a floating point value. |
23 | // The special IEEE-754 bits patterns of Zero, infinity and NaNs are |
24 | // are not handled by this class. |
25 | // |
26 | // A normalized floating point number is of this form: |
27 | // (-1)*sign * 2^exponent * <mantissa> |
28 | // where <mantissa> is of the form 1.<...>. |
29 | template <typename T> struct NormalFloat { |
30 | static_assert( |
31 | cpp::is_floating_point_v<T>, |
32 | "NormalFloat template parameter has to be a floating point type." ); |
33 | |
34 | using StorageType = typename FPBits<T>::StorageType; |
35 | static constexpr StorageType ONE = |
36 | (StorageType(1) << FPBits<T>::FRACTION_LEN); |
37 | |
38 | // Unbiased exponent value. |
39 | int32_t exponent; |
40 | |
41 | StorageType mantissa; |
42 | // We want |StorageType| to have atleast one bit more than the actual mantissa |
43 | // bit width to accommodate the implicit 1 value. |
44 | static_assert(sizeof(StorageType) * 8 >= FPBits<T>::FRACTION_LEN + 1, |
45 | "Bad type for mantissa in NormalFloat." ); |
46 | |
47 | Sign sign = Sign::POS; |
48 | |
49 | LIBC_INLINE NormalFloat(Sign s, int32_t e, StorageType m) |
50 | : exponent(e), mantissa(m), sign(s) { |
51 | if (mantissa >= ONE) |
52 | return; |
53 | |
54 | unsigned normalization_shift = evaluate_normalization_shift(m: mantissa); |
55 | mantissa = mantissa << normalization_shift; |
56 | exponent -= normalization_shift; |
57 | } |
58 | |
59 | LIBC_INLINE explicit NormalFloat(T x) { init_from_bits(bits: FPBits<T>(x)); } |
60 | |
61 | LIBC_INLINE explicit NormalFloat(FPBits<T> bits) { init_from_bits(bits); } |
62 | |
63 | // Compares this normalized number with another normalized number. |
64 | // Returns -1 is this number is less than |other|, 0 if this number is equal |
65 | // to |other|, and 1 if this number is greater than |other|. |
66 | LIBC_INLINE int cmp(const NormalFloat<T> &other) const { |
67 | const int result = sign.is_neg() ? -1 : 1; |
68 | if (sign != other.sign) |
69 | return result; |
70 | |
71 | if (exponent > other.exponent) { |
72 | return result; |
73 | } else if (exponent == other.exponent) { |
74 | if (mantissa > other.mantissa) |
75 | return result; |
76 | else if (mantissa == other.mantissa) |
77 | return 0; |
78 | else |
79 | return -result; |
80 | } else { |
81 | return -result; |
82 | } |
83 | } |
84 | |
85 | // Returns a new normalized floating point number which is equal in value |
86 | // to this number multiplied by 2^e. That is: |
87 | // new = this * 2^e |
88 | LIBC_INLINE NormalFloat<T> mul2(int e) const { |
89 | NormalFloat<T> result = *this; |
90 | result.exponent += e; |
91 | return result; |
92 | } |
93 | |
94 | LIBC_INLINE operator T() const { |
95 | int biased_exponent = exponent + FPBits<T>::EXP_BIAS; |
96 | // Max exponent is of the form 0xFF...E. That is why -2 and not -1. |
97 | constexpr int MAX_EXPONENT_VALUE = (1 << FPBits<T>::EXP_LEN) - 2; |
98 | if (biased_exponent > MAX_EXPONENT_VALUE) { |
99 | return FPBits<T>::inf(sign).get_val(); |
100 | } |
101 | |
102 | FPBits<T> result(T(0.0)); |
103 | result.set_sign(sign); |
104 | |
105 | constexpr int SUBNORMAL_EXPONENT = -FPBits<T>::EXP_BIAS + 1; |
106 | if (exponent < SUBNORMAL_EXPONENT) { |
107 | unsigned shift = SUBNORMAL_EXPONENT - exponent; |
108 | // Since exponent > subnormalExponent, shift is strictly greater than |
109 | // zero. |
110 | if (shift <= FPBits<T>::FRACTION_LEN + 1) { |
111 | // Generate a subnormal number. Might lead to loss of precision. |
112 | // We round to nearest and round halfway cases to even. |
113 | const StorageType shift_out_mask = (StorageType(1) << shift) - 1; |
114 | const StorageType shift_out_value = mantissa & shift_out_mask; |
115 | const StorageType halfway_value = StorageType(1) << (shift - 1); |
116 | result.set_biased_exponent(0); |
117 | result.set_mantissa(mantissa >> shift); |
118 | StorageType new_mantissa = result.get_mantissa(); |
119 | if (shift_out_value > halfway_value) { |
120 | new_mantissa += 1; |
121 | } else if (shift_out_value == halfway_value) { |
122 | // Round to even. |
123 | if (result.get_mantissa() & 0x1) |
124 | new_mantissa += 1; |
125 | } |
126 | result.set_mantissa(new_mantissa); |
127 | // Adding 1 to mantissa can lead to overflow. This can only happen if |
128 | // mantissa was all ones (0b111..11). For such a case, we will carry |
129 | // the overflow into the exponent. |
130 | if (new_mantissa == ONE) |
131 | result.set_biased_exponent(1); |
132 | return result.get_val(); |
133 | } else { |
134 | return result.get_val(); |
135 | } |
136 | } |
137 | |
138 | result.set_biased_exponent(exponent + FPBits<T>::EXP_BIAS); |
139 | result.set_mantissa(mantissa); |
140 | return result.get_val(); |
141 | } |
142 | |
143 | private: |
144 | LIBC_INLINE void init_from_bits(FPBits<T> bits) { |
145 | sign = bits.sign(); |
146 | |
147 | if (bits.is_inf_or_nan() || bits.is_zero()) { |
148 | // Ignore special bit patterns. Implementations deal with them separately |
149 | // anyway so this should not be a problem. |
150 | exponent = 0; |
151 | mantissa = 0; |
152 | return; |
153 | } |
154 | |
155 | // Normalize subnormal numbers. |
156 | if (bits.is_subnormal()) { |
157 | unsigned shift = evaluate_normalization_shift(m: bits.get_mantissa()); |
158 | mantissa = StorageType(bits.get_mantissa()) << shift; |
159 | exponent = 1 - FPBits<T>::EXP_BIAS - shift; |
160 | } else { |
161 | exponent = bits.get_biased_exponent() - FPBits<T>::EXP_BIAS; |
162 | mantissa = ONE | bits.get_mantissa(); |
163 | } |
164 | } |
165 | |
166 | LIBC_INLINE unsigned evaluate_normalization_shift(StorageType m) { |
167 | unsigned shift = 0; |
168 | for (; (ONE & m) == 0 && (shift < FPBits<T>::FRACTION_LEN); |
169 | m <<= 1, ++shift) |
170 | ; |
171 | return shift; |
172 | } |
173 | }; |
174 | |
175 | #ifdef LIBC_TYPES_LONG_DOUBLE_IS_X86_FLOAT80 |
176 | template <> |
177 | LIBC_INLINE void |
178 | NormalFloat<long double>::init_from_bits(FPBits<long double> bits) { |
179 | sign = bits.sign(); |
180 | |
181 | if (bits.is_inf_or_nan() || bits.is_zero()) { |
182 | // Ignore special bit patterns. Implementations deal with them separately |
183 | // anyway so this should not be a problem. |
184 | exponent = 0; |
185 | mantissa = 0; |
186 | return; |
187 | } |
188 | |
189 | if (bits.is_subnormal()) { |
190 | if (bits.get_implicit_bit() == 0) { |
191 | // Since we ignore zero value, the mantissa in this case is non-zero. |
192 | int normalization_shift = |
193 | evaluate_normalization_shift(m: bits.get_mantissa()); |
194 | exponent = -16382 - normalization_shift; |
195 | mantissa = (bits.get_mantissa() << normalization_shift); |
196 | } else { |
197 | exponent = -16382; |
198 | mantissa = ONE | bits.get_mantissa(); |
199 | } |
200 | } else { |
201 | if (bits.get_implicit_bit() == 0) { |
202 | // Invalid number so just store 0 similar to a NaN. |
203 | exponent = 0; |
204 | mantissa = 0; |
205 | } else { |
206 | exponent = bits.get_biased_exponent() - 16383; |
207 | mantissa = ONE | bits.get_mantissa(); |
208 | } |
209 | } |
210 | } |
211 | |
212 | template <> LIBC_INLINE NormalFloat<long double>::operator long double() const { |
213 | using LDBits = FPBits<long double>; |
214 | int biased_exponent = exponent + LDBits::EXP_BIAS; |
215 | // Max exponent is of the form 0xFF...E. That is why -2 and not -1. |
216 | constexpr int MAX_EXPONENT_VALUE = (1 << LDBits::EXP_LEN) - 2; |
217 | if (biased_exponent > MAX_EXPONENT_VALUE) { |
218 | return LDBits::inf(sign).get_val(); |
219 | } |
220 | |
221 | FPBits<long double> result(0.0l); |
222 | result.set_sign(sign); |
223 | |
224 | constexpr int SUBNORMAL_EXPONENT = -LDBits::EXP_BIAS + 1; |
225 | if (exponent < SUBNORMAL_EXPONENT) { |
226 | unsigned shift = SUBNORMAL_EXPONENT - exponent; |
227 | if (shift <= LDBits::FRACTION_LEN + 1) { |
228 | // Generate a subnormal number. Might lead to loss of precision. |
229 | // We round to nearest and round halfway cases to even. |
230 | const StorageType shift_out_mask = (StorageType(1) << shift) - 1; |
231 | const StorageType shift_out_value = mantissa & shift_out_mask; |
232 | const StorageType halfway_value = StorageType(1) << (shift - 1); |
233 | result.set_biased_exponent(0); |
234 | result.set_mantissa(mantissa >> shift); |
235 | StorageType new_mantissa = result.get_mantissa(); |
236 | if (shift_out_value > halfway_value) { |
237 | new_mantissa += 1; |
238 | } else if (shift_out_value == halfway_value) { |
239 | // Round to even. |
240 | if (result.get_mantissa() & 0x1) |
241 | new_mantissa += 1; |
242 | } |
243 | result.set_mantissa(new_mantissa); |
244 | // Adding 1 to mantissa can lead to overflow. This can only happen if |
245 | // mantissa was all ones (0b111..11). For such a case, we will carry |
246 | // the overflow into the exponent and set the implicit bit to 1. |
247 | if (new_mantissa == ONE) { |
248 | result.set_biased_exponent(1); |
249 | result.set_implicit_bit(1); |
250 | } else { |
251 | result.set_implicit_bit(0); |
252 | } |
253 | return result.get_val(); |
254 | } else { |
255 | return result.get_val(); |
256 | } |
257 | } |
258 | |
259 | result.set_biased_exponent(biased_exponent); |
260 | result.set_mantissa(mantissa); |
261 | result.set_implicit_bit(1); |
262 | return result.get_val(); |
263 | } |
264 | #endif // LIBC_TYPES_LONG_DOUBLE_IS_X86_FLOAT80 |
265 | |
266 | } // namespace fputil |
267 | } // namespace LIBC_NAMESPACE |
268 | |
269 | #endif // LLVM_LIBC_SRC___SUPPORT_FPUTIL_NORMALFLOAT_H |
270 | |