1 //===-- Single-precision e^x - 1 function ---------------------------------===//
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 #include "src/math/expm1f.h"
10 #include "common_constants.h" // Lookup tables EXP_M1 and EXP_M2.
11 #include "src/__support/FPUtil/BasicOperations.h"
12 #include "src/__support/FPUtil/FEnvImpl.h"
13 #include "src/__support/FPUtil/FMA.h"
14 #include "src/__support/FPUtil/FPBits.h"
15 #include "src/__support/FPUtil/PolyEval.h"
16 #include "src/__support/common.h"
17 
18 #include <errno.h>
19 
20 namespace __llvm_libc {
21 
22 LLVM_LIBC_FUNCTION(float, expm1f, (float x)) {
23   using FPBits = typename fputil::FPBits<float>;
24   FPBits xbits(x);
25 
26   uint32_t x_u = xbits.uintval();
27   uint32_t x_abs = x_u & 0x7fff'ffffU;
28 
29   // Exceptional value
30   if (unlikely(x_u == 0x3e35'bec5U)) { // x = 0x1.6b7d8ap-3f
31     int round_mode = fputil::get_round();
32     if (round_mode == FE_TONEAREST || round_mode == FE_UPWARD)
33       return 0x1.8dbe64p-3f;
34     return 0x1.8dbe62p-3f;
35   }
36 
37   // When |x| > 25*log(2), or nan
38   if (unlikely(x_abs >= 0x418a'a123U)) {
39     // x < log(2^-25)
40     if (xbits.get_sign()) {
41       // exp(-Inf) = 0
42       if (xbits.is_inf())
43         return -1.0f;
44       // exp(nan) = nan
45       if (xbits.is_nan())
46         return x;
47       int round_mode = fputil::get_round();
48       if (round_mode == FE_UPWARD || round_mode == FE_TOWARDZERO)
49         return -0x1.ffff'fep-1f; // -1.0f + 0x1.0p-24f
50       return -1.0f;
51     } else {
52       // x >= 89 or nan
53       if (xbits.uintval() >= 0x42b2'0000) {
54         if (xbits.uintval() < 0x7f80'0000U) {
55           int rounding = fputil::get_round();
56           if (rounding == FE_DOWNWARD || rounding == FE_TOWARDZERO)
57             return static_cast<float>(FPBits(FPBits::MAX_NORMAL));
58 
59           errno = ERANGE;
60         }
61         return x + static_cast<float>(FPBits::inf());
62       }
63     }
64   }
65 
66   // |x| < 2^-4
67   if (x_abs < 0x3d80'0000U) {
68     // |x| < 2^-25
69     if (x_abs < 0x3300'0000U) {
70       // x = -0.0f
71       if (unlikely(xbits.uintval() == 0x8000'0000U))
72         return x;
73       // When |x| < 2^-25, the relative error of the approximation e^x - 1 ~ x
74       // is:
75       //   |(e^x - 1) - x| / |e^x - 1| < |x^2| / |x|
76       //                               = |x|
77       //                               < 2^-25
78       //                               < epsilon(1)/2.
79       // So the correctly rounded values of expm1(x) are:
80       //   = x + eps(x) if rounding mode = FE_UPWARD,
81       //                   or (rounding mode = FE_TOWARDZERO and x is negative),
82       //   = x otherwise.
83       // To simplify the rounding decision and make it more efficient, we use
84       //   fma(x, x, x) ~ x + x^2 instead.
85       return fputil::multiply_add(x, x, x);
86     }
87 
88     // 2^-25 <= |x| < 2^-4
89     double xd = static_cast<double>(x);
90     double xsq = xd * xd;
91     // Degree-8 minimax polynomial generated by Sollya with:
92     // > display = hexadecimal;
93     // > P = fpminimax((expm1(x) - x)/x^2, 6, [|D...|], [-2^-4, 2^-4]);
94     double r =
95         fputil::polyeval(xd, 0x1p-1, 0x1.55555555557ddp-3, 0x1.55555555552fap-5,
96                          0x1.111110fcd58b7p-7, 0x1.6c16c1717660bp-10,
97                          0x1.a0241f0006d62p-13, 0x1.a01e3f8d3c06p-16);
98     return static_cast<float>(fputil::multiply_add(r, xsq, xd));
99   }
100 
101   // For -18 < x < 89, to compute expm1(x), we perform the following range
102   // reduction: find hi, mid, lo such that:
103   //   x = hi + mid + lo, in which
104   //     hi is an integer,
105   //     mid * 2^7 is an integer
106   //     -2^(-8) <= lo < 2^-8.
107   // In particular,
108   //   hi + mid = round(x * 2^7) * 2^(-7).
109   // Then,
110   //   expm1(x) = exp(hi + mid + lo) - 1 = exp(hi) * exp(mid) * exp(lo) - 1.
111   // We store exp(hi) and exp(mid) in the lookup tables EXP_M1 and EXP_M2
112   // respectively.  exp(lo) is computed using a degree-4 minimax polynomial
113   // generated by Sollya.
114 
115   // x_hi = hi + mid.
116   int x_hi = static_cast<int>(x * 0x1.0p7f + (xbits.get_sign() ? -0.5f : 0.5f));
117   // Subtract (hi + mid) from x to get lo.
118   x -= static_cast<float>(x_hi) * 0x1.0p-7f;
119   double xd = static_cast<double>(x);
120   x_hi += 104 << 7;
121   // hi = x_hi >> 7
122   double exp_hi = EXP_M1[x_hi >> 7];
123   // lo = x_hi & 0x0000'007fU;
124   double exp_mid = EXP_M2[x_hi & 0x7f];
125   double exp_hi_mid = exp_hi * exp_mid;
126   // Degree-4 minimax polynomial generated by Sollya with the following
127   // commands:
128   //   > display = hexadecimal;
129   //   > Q = fpminimax(expm1(x)/x, 3, [|D...|], [-2^-8, 2^-8]);
130   //   > Q;
131   double exp_lo =
132       fputil::polyeval(xd, 0x1.0p0, 0x1.ffffffffff777p-1, 0x1.000000000071cp-1,
133                        0x1.555566668e5e7p-3, 0x1.55555555ef243p-5);
134   return static_cast<float>(fputil::multiply_add(exp_hi_mid, exp_lo, -1.0));
135 }
136 
137 } // namespace __llvm_libc
138