xref: /aosp_15_r20/external/llvm-libc/src/__support/FPUtil/generic/FMod.h (revision 71db0c75aadcf003ffe3238005f61d7618a3fead)
1 //===-- Common header for fmod implementations ------------------*- 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_GENERIC_FMOD_H
10 #define LLVM_LIBC_SRC___SUPPORT_FPUTIL_GENERIC_FMOD_H
11 
12 #include "src/__support/CPP/bit.h"
13 #include "src/__support/CPP/limits.h"
14 #include "src/__support/CPP/type_traits.h"
15 #include "src/__support/FPUtil/FEnvImpl.h"
16 #include "src/__support/FPUtil/FPBits.h"
17 #include "src/__support/macros/config.h"
18 #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY
19 
20 namespace LIBC_NAMESPACE_DECL {
21 namespace fputil {
22 namespace generic {
23 
24 //  Objective:
25 //    The  algorithm uses  integer arithmetic  (max uint64_t)  for general case.
26 //    Some common  cases, like  abs(x) < abs(y)  or  abs(x) < 1000 *  abs(y) are
27 //    treated specially to increase  performance.  The part of checking  special
28 //    cases, numbers NaN, INF etc. treated separately.
29 //
30 //  Objective:
31 //    1) FMod definition (https://cplusplus.com/reference/cmath/fmod/):
32 //       fmod = numer - tquot * denom, where tquot is the truncated
33 //       (i.e., rounded towards zero) result of: numer/denom.
34 //    2) FMod with negative x and/or y can be trivially converted to fmod for
35 //       positive x and y. Therefore the algorithm below works only with
36 //       positive numbers.
37 //    3) All positive floating point numbers can be represented as m * 2^e,
38 //       where "m" is positive integer and "e" is signed.
39 //    4) FMod function can be calculated in integer numbers (x > y):
40 //         fmod = m_x * 2^e_x - tquot * m_y * 2^e_y
41 //              = 2^e_y * (m_x * 2^(e_x - e^y) - tquot * m_y).
42 //       All variables in parentheses are unsigned integers.
43 //
44 //  Mathematical background:
45 //    Input x,y in the algorithm is represented (mathematically) like m_x*2^e_x
46 //    and m_y*2^e_y. This is an ambiguous number representation. For example:
47 //      m * 2^e = (2 * m) * 2^(e-1)
48 //    The algorithm uses the facts that
49 //      r = a % b = (a % (N * b)) % b,
50 //      (a * c) % (b * c) = (a % b) * c
51 //    where N is positive  integer number. a, b and c - positive. Let's  adopt
52 //    the formula for representation above.
53 //      a = m_x * 2^e_x, b = m_y * 2^e_y, N = 2^k
54 //      r(k) = a % b = (m_x * 2^e_x) % (2^k * m_y * 2^e_y)
55 //           = 2^(e_y + k) * (m_x * 2^(e_x - e_y - k) % m_y)
56 //      r(k) = m_r * 2^e_r = (m_x % m_y) * 2^(m_y + k)
57 //           = (2^p * (m_x % m_y) * 2^(e_y + k - p))
58 //        m_r = 2^p * (m_x % m_y), e_r = m_y + k - p
59 //
60 //  Algorithm description:
61 //  First, let write x = m_x * 2^e_x and y = m_y * 2^e_y with m_x, m_y, e_x, e_y
62 //  are integers (m_x amd m_y positive).
63 //  Then the naive  implementation of the fmod function with a simple
64 //  for/while loop:
65 //      while (e_x > e_y) {
66 //        m_x *= 2; --e_x; //  m_x * 2^e_x == 2 * m_x * 2^(e_x - 1)
67 //        m_x %= m_y;
68 //      }
69 //  On the other hand, the algorithm exploits the fact that m_x, m_y are the
70 //  mantissas of floating point numbers, which use less bits than the storage
71 //  integers: 24 / 32 for floats and 53 / 64 for doubles, so if in each step of
72 //  the iteration, we can left shift m_x as many bits as the storage integer
73 //  type can hold, the exponent reduction per step will be at least 32 - 24 = 8
74 //  for floats and 64 - 53 = 11 for doubles (double example below):
75 //      while (e_x > e_y) {
76 //        m_x <<= 11; e_x -= 11; //  m_x * 2^e_x == 2^11 * m_x * 2^(e_x - 11)
77 //        m_x %= m_y;
78 //      }
79 //  Some extra improvements are done:
80 //    1) Shift m_y maximum to the right, which can significantly improve
81 //       performance for small integer numbers (y = 3 for example).
82 //       The m_x shift in the loop can be 62 instead of 11 for double.
83 //    2) For some architectures with very slow division, it can be better to
84 //       calculate inverse value ones, and after do multiplication in the loop.
85 //    3) "likely" special cases are treated specially to improve performance.
86 //
87 //  Simple example:
88 //  The examples below use byte for simplicity.
89 //    1) Shift hy maximum to right without losing bits and increase iy value
90 //       m_y = 0b00101100 e_y = 20 after shift m_y = 0b00001011 e_y = 22.
91 //    2) m_x = m_x % m_y.
92 //    3) Move m_x maximum to left. Note that after (m_x = m_x % m_y) CLZ in m_x
93 //    is not lower than CLZ in m_y. m_x=0b00001001 e_x = 100, m_x=0b10010000,
94 //       e_x = 100-4 = 96.
95 //    4) Repeat (2) until e_x == e_y.
96 //
97 //  Complexity analysis (double):
98 //    Converting x,y to (m_x,e_x),(m_y, e_y): CTZ/shift/AND/OR/if. Loop  count:
99 //      (m_x - m_y) / (64 -  "length of m_y").
100 //      max("length of m_y")  = 53,
101 //      max(e_x - e_y)  = 2048
102 //    Maximum operation is  186. For rare "unrealistic" cases.
103 //
104 //  Special cases (double):
105 //    Supposing  that  case  where |y| > 1e-292 and |x/y|<2000  is  very  common
106 //    special processing is implemented. No m_y alignment, no loop:
107 //      result = (m_x * 2^(e_x - e_y)) % m_y.
108 //    When x and y are both subnormal (rare case but...) the
109 //      result = m_x % m_y.
110 //    Simplified conversion back to double.
111 
112 // Exceptional cases handler according to cppreference.com
113 //    https://en.cppreference.com/w/cpp/numeric/math/fmod
114 // and POSIX standard described in Linux man
115 //   https://man7.org/linux/man-pages/man3/fmod.3p.html
116 // C standard for the function is not full, so not by default (although it can
117 // be implemented in another handler.
118 // Signaling NaN converted to quiet NaN with FE_INVALID exception.
119 //    https://www.open-std.org/JTC1/SC22/WG14/www/docs/n1011.htm
120 template <typename T> struct FModDivisionSimpleHelper {
executeFModDivisionSimpleHelper121   LIBC_INLINE constexpr static T execute(int exp_diff, int sides_zeroes_count,
122                                          T m_x, T m_y) {
123     while (exp_diff > sides_zeroes_count) {
124       exp_diff -= sides_zeroes_count;
125       m_x <<= sides_zeroes_count;
126       m_x %= m_y;
127     }
128     m_x <<= exp_diff;
129     m_x %= m_y;
130     return m_x;
131   }
132 };
133 
134 template <typename T> struct FModDivisionInvMultHelper {
executeFModDivisionInvMultHelper135   LIBC_INLINE constexpr static T execute(int exp_diff, int sides_zeroes_count,
136                                          T m_x, T m_y) {
137     constexpr int LENGTH = sizeof(T) * CHAR_BIT;
138     if (exp_diff > sides_zeroes_count) {
139       T inv_hy = (cpp::numeric_limits<T>::max() / m_y);
140       while (exp_diff > sides_zeroes_count) {
141         exp_diff -= sides_zeroes_count;
142         T hd = (m_x * inv_hy) >> (LENGTH - sides_zeroes_count);
143         m_x <<= sides_zeroes_count;
144         m_x -= hd * m_y;
145         while (LIBC_UNLIKELY(m_x > m_y))
146           m_x -= m_y;
147       }
148       T hd = (m_x * inv_hy) >> (LENGTH - exp_diff);
149       m_x <<= exp_diff;
150       m_x -= hd * m_y;
151       while (LIBC_UNLIKELY(m_x > m_y))
152         m_x -= m_y;
153     } else {
154       m_x <<= exp_diff;
155       m_x %= m_y;
156     }
157     return m_x;
158   }
159 };
160 
161 template <typename T, typename U = typename FPBits<T>::StorageType,
162           typename DivisionHelper = FModDivisionSimpleHelper<U>>
163 class FMod {
164   static_assert(cpp::is_floating_point_v<T> &&
165                     is_unsigned_integral_or_big_int_v<U> &&
166                     (sizeof(U) * CHAR_BIT > FPBits<T>::FRACTION_LEN),
167                 "FMod instantiated with invalid type.");
168 
169 private:
170   using FPB = FPBits<T>;
171   using StorageType = typename FPB::StorageType;
172 
pre_check(T x,T y,T & out)173   LIBC_INLINE static bool pre_check(T x, T y, T &out) {
174     using FPB = fputil::FPBits<T>;
175     const T quiet_nan = FPB::quiet_nan().get_val();
176     FPB sx(x), sy(y);
177     if (LIBC_LIKELY(!sy.is_zero() && !sy.is_inf_or_nan() &&
178                     !sx.is_inf_or_nan()))
179       return false;
180 
181     if (sx.is_nan() || sy.is_nan()) {
182       if (sx.is_signaling_nan() || sy.is_signaling_nan())
183         fputil::raise_except_if_required(FE_INVALID);
184       out = quiet_nan;
185       return true;
186     }
187 
188     if (sx.is_inf() || sy.is_zero()) {
189       fputil::raise_except_if_required(FE_INVALID);
190       fputil::set_errno_if_required(EDOM);
191       out = quiet_nan;
192       return true;
193     }
194 
195     out = x;
196     return true;
197   }
198 
eval_internal(FPB sx,FPB sy)199   LIBC_INLINE static constexpr FPB eval_internal(FPB sx, FPB sy) {
200 
201     if (LIBC_LIKELY(sx.uintval() <= sy.uintval())) {
202       if (sx.uintval() < sy.uintval())
203         return sx;             // |x|<|y| return x
204       return FPB::zero();      // |x|=|y| return 0.0
205     }
206 
207     int e_x = sx.get_biased_exponent();
208     int e_y = sy.get_biased_exponent();
209 
210     // Most common case where |y| is "very normal" and |x/y| < 2^EXP_LEN
211     if (LIBC_LIKELY(e_y > int(FPB::FRACTION_LEN) &&
212                     e_x - e_y <= int(FPB::EXP_LEN))) {
213       StorageType m_x = sx.get_explicit_mantissa();
214       StorageType m_y = sy.get_explicit_mantissa();
215       StorageType d = (e_x == e_y)
216                           ? (m_x - m_y)
217                           : static_cast<StorageType>(m_x << (e_x - e_y)) % m_y;
218       if (d == 0)
219         return FPB::zero();
220       // iy - 1 because of "zero power" for number with power 1
221       return FPB::make_value(d, e_y - 1);
222     }
223     // Both subnormal special case.
224     if (LIBC_UNLIKELY(e_x == 0 && e_y == 0)) {
225       FPB d;
226       d.set_mantissa(sx.uintval() % sy.uintval());
227       return d;
228     }
229 
230     // Note that hx is not subnormal by conditions above.
231     U m_x = static_cast<U>(sx.get_explicit_mantissa());
232     e_x--;
233 
234     U m_y = static_cast<U>(sy.get_explicit_mantissa());
235     constexpr int DEFAULT_LEAD_ZEROS =
236         sizeof(U) * CHAR_BIT - FPB::FRACTION_LEN - 1;
237     int lead_zeros_m_y = DEFAULT_LEAD_ZEROS;
238     if (LIBC_LIKELY(e_y > 0)) {
239       e_y--;
240     } else {
241       m_y = static_cast<U>(sy.get_mantissa());
242       lead_zeros_m_y = cpp::countl_zero(m_y);
243     }
244 
245     // Assume hy != 0
246     int tail_zeros_m_y = cpp::countr_zero(m_y);
247     int sides_zeroes_count = lead_zeros_m_y + tail_zeros_m_y;
248     // n > 0 by conditions above
249     int exp_diff = e_x - e_y;
250     {
251       // Shift hy right until the end or n = 0
252       int right_shift = exp_diff < tail_zeros_m_y ? exp_diff : tail_zeros_m_y;
253       m_y >>= right_shift;
254       exp_diff -= right_shift;
255       e_y += right_shift;
256     }
257 
258     {
259       // Shift hx left until the end or n = 0
260       int left_shift =
261           exp_diff < DEFAULT_LEAD_ZEROS ? exp_diff : DEFAULT_LEAD_ZEROS;
262       m_x <<= left_shift;
263       exp_diff -= left_shift;
264     }
265 
266     m_x %= m_y;
267     if (LIBC_UNLIKELY(m_x == 0))
268       return FPB::zero();
269 
270     if (exp_diff == 0)
271       return FPB::make_value(static_cast<StorageType>(m_x), e_y);
272 
273     // hx next can't be 0, because hx < hy, hy % 2 == 1 hx * 2^i % hy != 0
274     m_x = DivisionHelper::execute(exp_diff, sides_zeroes_count, m_x, m_y);
275     return FPB::make_value(static_cast<StorageType>(m_x), e_y);
276   }
277 
278 public:
eval(T x,T y)279   LIBC_INLINE static T eval(T x, T y) {
280     if (T out; LIBC_UNLIKELY(pre_check(x, y, out)))
281       return out;
282     FPB sx(x), sy(y);
283     Sign sign = sx.sign();
284     sx.set_sign(Sign::POS);
285     sy.set_sign(Sign::POS);
286     FPB result = eval_internal(sx, sy);
287     result.set_sign(sign);
288     return result.get_val();
289   }
290 };
291 
292 } // namespace generic
293 } // namespace fputil
294 } // namespace LIBC_NAMESPACE_DECL
295 
296 #endif // LLVM_LIBC_SRC___SUPPORT_FPUTIL_GENERIC_FMOD_H
297