1 //===-- Nearest integer floating-point operations ---------------*- 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_UTILS_FPUTIL_NEAREST_INTEGER_OPERATIONS_H
10 #define LLVM_LIBC_UTILS_FPUTIL_NEAREST_INTEGER_OPERATIONS_H
11 
12 #include "FEnv.h"
13 #include "FPBits.h"
14 
15 #include "utils/CPP/TypeTraits.h"
16 
17 #include <math.h>
18 #if math_errhandling & MATH_ERRNO
19 #include "src/errno/llvmlibc_errno.h"
20 #include <errno.h>
21 #endif
22 
23 namespace __llvm_libc {
24 namespace fputil {
25 
26 template <typename T,
27           cpp::EnableIfType<cpp::IsFloatingPointType<T>::Value, int> = 0>
trunc(T x)28 static inline T trunc(T x) {
29   FPBits<T> bits(x);
30 
31   // If x is infinity or NaN, return it.
32   // If it is zero also we should return it as is, but the logic
33   // later in this function takes care of it. But not doing a zero
34   // check, we improve the run time of non-zero values.
35   if (bits.isInfOrNaN())
36     return x;
37 
38   int exponent = bits.getExponent();
39 
40   // If the exponent is greater than the most negative mantissa
41   // exponent, then x is already an integer.
42   if (exponent >= static_cast<int>(MantissaWidth<T>::value))
43     return x;
44 
45   // If the exponent is such that abs(x) is less than 1, then return 0.
46   if (exponent <= -1) {
47     if (bits.sign)
48       return T(-0.0);
49     else
50       return T(0.0);
51   }
52 
53   int trimSize = MantissaWidth<T>::value - exponent;
54   bits.mantissa = (bits.mantissa >> trimSize) << trimSize;
55   return bits;
56 }
57 
58 template <typename T,
59           cpp::EnableIfType<cpp::IsFloatingPointType<T>::Value, int> = 0>
ceil(T x)60 static inline T ceil(T x) {
61   FPBits<T> bits(x);
62 
63   // If x is infinity NaN or zero, return it.
64   if (bits.isInfOrNaN() || bits.isZero())
65     return x;
66 
67   bool isNeg = bits.sign;
68   int exponent = bits.getExponent();
69 
70   // If the exponent is greater than the most negative mantissa
71   // exponent, then x is already an integer.
72   if (exponent >= static_cast<int>(MantissaWidth<T>::value))
73     return x;
74 
75   if (exponent <= -1) {
76     if (isNeg)
77       return T(-0.0);
78     else
79       return T(1.0);
80   }
81 
82   uint32_t trimSize = MantissaWidth<T>::value - exponent;
83   bits.mantissa = (bits.mantissa >> trimSize) << trimSize;
84   T truncValue = T(bits);
85 
86   // If x is already an integer, return it.
87   if (truncValue == x)
88     return x;
89 
90   // If x is negative, the ceil operation is equivalent to the trunc operation.
91   if (isNeg)
92     return truncValue;
93 
94   return truncValue + T(1.0);
95 }
96 
97 template <typename T,
98           cpp::EnableIfType<cpp::IsFloatingPointType<T>::Value, int> = 0>
floor(T x)99 static inline T floor(T x) {
100   FPBits<T> bits(x);
101   if (bits.sign) {
102     return -ceil(-x);
103   } else {
104     return trunc(x);
105   }
106 }
107 
108 template <typename T,
109           cpp::EnableIfType<cpp::IsFloatingPointType<T>::Value, int> = 0>
round(T x)110 static inline T round(T x) {
111   using UIntType = typename FPBits<T>::UIntType;
112   FPBits<T> bits(x);
113 
114   // If x is infinity NaN or zero, return it.
115   if (bits.isInfOrNaN() || bits.isZero())
116     return x;
117 
118   bool isNeg = bits.sign;
119   int exponent = bits.getExponent();
120 
121   // If the exponent is greater than the most negative mantissa
122   // exponent, then x is already an integer.
123   if (exponent >= static_cast<int>(MantissaWidth<T>::value))
124     return x;
125 
126   if (exponent == -1) {
127     // Absolute value of x is greater than equal to 0.5 but less than 1.
128     if (isNeg)
129       return T(-1.0);
130     else
131       return T(1.0);
132   }
133 
134   if (exponent <= -2) {
135     // Absolute value of x is less than 0.5.
136     if (isNeg)
137       return T(-0.0);
138     else
139       return T(0.0);
140   }
141 
142   uint32_t trimSize = MantissaWidth<T>::value - exponent;
143   bool halfBitSet = bits.mantissa & (UIntType(1) << (trimSize - 1));
144   bits.mantissa = (bits.mantissa >> trimSize) << trimSize;
145   T truncValue = T(bits);
146 
147   // If x is already an integer, return it.
148   if (truncValue == x)
149     return x;
150 
151   if (!halfBitSet) {
152     // Franctional part is less than 0.5 so round value is the
153     // same as the trunc value.
154     return truncValue;
155   } else {
156     return isNeg ? truncValue - T(1.0) : truncValue + T(1.0);
157   }
158 }
159 
160 template <typename T,
161           cpp::EnableIfType<cpp::IsFloatingPointType<T>::Value, int> = 0>
roundUsingCurrentRoundingMode(T x)162 static inline T roundUsingCurrentRoundingMode(T x) {
163   using UIntType = typename FPBits<T>::UIntType;
164   FPBits<T> bits(x);
165 
166   // If x is infinity NaN or zero, return it.
167   if (bits.isInfOrNaN() || bits.isZero())
168     return x;
169 
170   bool isNeg = bits.sign;
171   int exponent = bits.getExponent();
172   int roundingMode = getRound();
173 
174   // If the exponent is greater than the most negative mantissa
175   // exponent, then x is already an integer.
176   if (exponent >= static_cast<int>(MantissaWidth<T>::value))
177     return x;
178 
179   if (exponent <= -1) {
180     switch (roundingMode) {
181     case FE_DOWNWARD:
182       return isNeg ? T(-1.0) : T(0.0);
183     case FE_UPWARD:
184       return isNeg ? T(-0.0) : T(1.0);
185     case FE_TOWARDZERO:
186       return isNeg ? T(-0.0) : T(0.0);
187     case FE_TONEAREST:
188       if (exponent <= -2 || bits.mantissa == 0)
189         return isNeg ? T(-0.0) : T(0.0); // abs(x) <= 0.5
190       else
191         return isNeg ? T(-1.0) : T(1.0); // abs(x) > 0.5
192     default:
193       __builtin_unreachable();
194     }
195   }
196 
197   uint32_t trimSize = MantissaWidth<T>::value - exponent;
198   FPBits<T> newBits = bits;
199   newBits.mantissa = (bits.mantissa >> trimSize) << trimSize;
200   T truncValue = T(newBits);
201 
202   // If x is already an integer, return it.
203   if (truncValue == x)
204     return x;
205 
206   UIntType trimValue = bits.mantissa & ((UIntType(1) << trimSize) - 1);
207   UIntType halfValue = (UIntType(1) << (trimSize - 1));
208   // If exponent is 0, trimSize will be equal to the mantissa width, and
209   // truncIsOdd` will not be correct. So, we handle it as a special case
210   // below.
211   UIntType truncIsOdd = newBits.mantissa & (UIntType(1) << trimSize);
212 
213   switch (roundingMode) {
214   case FE_DOWNWARD:
215     return isNeg ? truncValue - T(1.0) : truncValue;
216   case FE_UPWARD:
217     return isNeg ? truncValue : truncValue + T(1.0);
218   case FE_TOWARDZERO:
219     return truncValue;
220   case FE_TONEAREST:
221     if (trimValue > halfValue) {
222       return isNeg ? truncValue - T(1.0) : truncValue + T(1.0);
223     } else if (trimValue == halfValue) {
224       if (exponent == 0)
225         return isNeg ? T(-2.0) : T(2.0);
226       if (truncIsOdd)
227         return isNeg ? truncValue - T(1.0) : truncValue + T(1.0);
228       else
229         return truncValue;
230     } else {
231       return truncValue;
232     }
233   default:
234     __builtin_unreachable();
235   }
236 }
237 
238 namespace internal {
239 
240 template <typename F, typename I,
241           cpp::EnableIfType<cpp::IsFloatingPointType<F>::Value &&
242                                 cpp::IsIntegral<I>::Value,
243                             int> = 0>
roundedFloatToSignedInteger(F x)244 static inline I roundedFloatToSignedInteger(F x) {
245   constexpr I IntegerMin = (I(1) << (sizeof(I) * 8 - 1));
246   constexpr I IntegerMax = -(IntegerMin + 1);
247   FPBits<F> bits(x);
248   auto setDomainErrorAndRaiseInvalid = []() {
249 #if math_errhandling & MATH_ERRNO
250     llvmlibc_errno = EDOM;
251 #endif
252 #if math_errhandling & MATH_ERREXCEPT
253     raiseExcept(FE_INVALID);
254 #endif
255   };
256 
257   if (bits.isInfOrNaN()) {
258     setDomainErrorAndRaiseInvalid();
259     return bits.sign ? IntegerMin : IntegerMax;
260   }
261 
262   int exponent = bits.getExponent();
263   constexpr int exponentLimit = sizeof(I) * 8 - 1;
264   if (exponent > exponentLimit) {
265     setDomainErrorAndRaiseInvalid();
266     return bits.sign ? IntegerMin : IntegerMax;
267   } else if (exponent == exponentLimit) {
268     if (bits.sign == 0 || bits.mantissa != 0) {
269       setDomainErrorAndRaiseInvalid();
270       return bits.sign ? IntegerMin : IntegerMax;
271     }
272     // If the control reaches here, then it means that the rounded
273     // value is the most negative number for the signed integer type I.
274   }
275 
276   // For all other cases, if `x` can fit in the integer type `I`,
277   // we just return `x`. Implicit conversion will convert the
278   // floating point value to the exact integer value.
279   return x;
280 }
281 
282 } // namespace internal
283 
284 template <typename F, typename I,
285           cpp::EnableIfType<cpp::IsFloatingPointType<F>::Value &&
286                                 cpp::IsIntegral<I>::Value,
287                             int> = 0>
roundToSignedInteger(F x)288 static inline I roundToSignedInteger(F x) {
289   return internal::roundedFloatToSignedInteger<F, I>(round(x));
290 }
291 
292 template <typename F, typename I,
293           cpp::EnableIfType<cpp::IsFloatingPointType<F>::Value &&
294                                 cpp::IsIntegral<I>::Value,
295                             int> = 0>
roundToSignedIntegerUsingCurrentRoundingMode(F x)296 static inline I roundToSignedIntegerUsingCurrentRoundingMode(F x) {
297   return internal::roundedFloatToSignedInteger<F, I>(
298       roundUsingCurrentRoundingMode(x));
299 }
300 
301 } // namespace fputil
302 } // namespace __llvm_libc
303 
304 #endif // LLVM_LIBC_UTILS_FPUTIL_NEAREST_INTEGER_OPERATIONS_H
305