libc/src/math/generic/tanhf.cpp

   1 //===-- Single-precision tanh 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/tanhf.h"
  10 #include "src/__support/FPUtil/FPBits.h"
  11 #include "src/__support/macros/optimization.h"            // LIBC_UNLIKELY
  12 #include "src/__support/macros/properties/cpu_features.h" // LIBC_TARGET_CPU_HAS_FMA
  13 #include "src/math/generic/explogxf.h"
  14
  15 namespace __llvm_libc {
  16
  17 LLVM_LIBC_FUNCTION(float, tanhf, (float x)) {
  18   using FPBits = typename fputil::FPBits<float>;
  19   FPBits xbits(x);
  20   bool sign = xbits.get_sign();
  21   uint32_t x_abs = xbits.uintval() & FPBits::FloatProp::EXP_MANT_MASK;
  22
  23   // |x| <= 2^-26
  24   if (LIBC_UNLIKELY(x_abs <= 0x3280'0000U)) {
  25     return static_cast<float>(
  26         LIBC_UNLIKELY(x_abs == 0) ? x : (x - 0x1.5555555555555p-2 * x * x * x));
  27   }
  28
  29   // When |x| >= 15, or x is inf or nan
  30   if (LIBC_UNLIKELY(x_abs >= 0x4170'0000U)) {
  31     if (xbits.is_nan())
  32       return x + 1.0f; // sNaN to qNaN + signal
  33
  34     if (xbits.is_inf())
  35       return sign ? -1.0f : 1.0f;
  36
  37     if (sign) {
  38       return -1.0f + opt_barrier(FPBits(FPBits::MIN_NORMAL).get_val());
  39     } else
  40       return 1.0f - opt_barrier(FPBits(FPBits::MIN_NORMAL).get_val());
  41   }
  42
  43   // |x| <= 0.078125
  44   if (LIBC_UNLIKELY(x_abs <= 0x3da0'0000U)) {
  45     double xdbl = x;
  46     double x2 = xdbl * xdbl;
  47     // Pure Taylor series.
  48     double pe = fputil::polyeval(x2, 0.0, -0x1.5555555555555p-2,
  49                                  0x1.1111111111111p-3, -0x1.ba1ba1ba1ba1cp-5,
  50                                  0x1.664f4882c10fap-6, -0x1.226e355e6c23dp-7);
  51     return static_cast<float>(fputil::multiply_add(xdbl, pe, xdbl));
  52   }
  53
  54   if (LIBC_UNLIKELY(xbits.bits == 0x4058'e0a3U)) {
  55     if (fputil::get_round() == FE_DOWNWARD)
  56       return FPBits(0x3f7f'6ad9U).get_val();
  57   }
  58
  59   // Range reduction: e^(2x) = 2^(mid + hi) * e^lo
  60   auto ep = exp_b_range_reduc<ExpBase>(2.0f * x); // exp(2 * x)
  61   double r = ExpBase::powb_lo(ep.lo);
  62   // tanh(x) = (exp(2x) - 1) / (exp(2x) + 1)
  63 #if defined(LIBC_TARGET_CPU_HAS_FMA)
  64   return static_cast<float>(fputil::multiply_add(ep.mh, r, -1.0) /
  65                             fputil::multiply_add(ep.mh, r, 1.0));
  66 #else
  67   double exp_x = ep.mh * r;
  68   return static_cast<float>((exp_x - 1.0) / (exp_x + 1.0));
  69 #endif // LIBC_TARGET_CPU_HAS_FMA
  70 }
  71
  72 } // namespace __llvm_libc