//===-- Single-precision sin function -------------------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//

#include "src/math/sinf.h"
#include "sincosf_utils.h"
#include "src/__support/FPUtil/BasicOperations.h"
#include "src/__support/FPUtil/FEnvImpl.h"
#include "src/__support/FPUtil/FPBits.h"
#include "src/__support/FPUtil/PolyEval.h"
#include "src/__support/FPUtil/multiply_add.h"
#include "src/__support/FPUtil/rounding_mode.h"
#include "src/__support/common.h"
#include "src/__support/macros/config.h"
#include "src/__support/macros/optimization.h"            // LIBC_UNLIKELY
#include "src/__support/macros/properties/cpu_features.h" // LIBC_TARGET_CPU_HAS_FMA

#if defined(LIBC_TARGET_CPU_HAS_FMA)
#include "range_reduction_fma.h"
#else
#include "range_reduction.h"
#endif

namespace LIBC_NAMESPACE_DECL {

LLVM_LIBC_FUNCTION(float, sinf, (float x)) {
  using FPBits = typename fputil::FPBits<float>;
  FPBits xbits(x);

  uint32_t x_u = xbits.uintval();
  uint32_t x_abs = x_u & 0x7fff'ffffU;
  double xd = static_cast<double>(x);

  // Range reduction:
  // For |x| > pi/32, we perform range reduction as follows:
  // Find k and y such that:
  //   x = (k + y) * pi/32
  //   k is an integer
  //   |y| < 0.5
  // For small range (|x| < 2^45 when FMA instructions are available, 2^22
  // otherwise), this is done by performing:
  //   k = round(x * 32/pi)
  //   y = x * 32/pi - k
  // For large range, we will omit all the higher parts of 32/pi such that the
  // least significant bits of their full products with x are larger than 63,
  // since sin((k + y + 64*i) * pi/32) = sin(x + i * 2pi) = sin(x).
  //
  // When FMA instructions are not available, we store the digits of 32/pi in
  // chunks of 28-bit precision.  This will make sure that the products:
  //   x * THIRTYTWO_OVER_PI_28[i] are all exact.
  // When FMA instructions are available, we simply store the digits of 32/pi in
  // chunks of doubles (53-bit of precision).
  // So when multiplying by the largest values of single precision, the
  // resulting output should be correct up to 2^(-208 + 128) ~ 2^-80.  By the
  // worst-case analysis of range reduction, |y| >= 2^-38, so this should give
  // us more than 40 bits of accuracy. For the worst-case estimation of range
  // reduction, see for instances:
  //   Elementary Functions by J-M. Muller, Chapter 11,
  //   Handbook of Floating-Point Arithmetic by J-M. Muller et. al.,
  //   Chapter 10.2.
  //
  // Once k and y are computed, we then deduce the answer by the sine of sum
  // formula:
  //   sin(x) = sin((k + y)*pi/32)
  //          = sin(y*pi/32) * cos(k*pi/32) + cos(y*pi/32) * sin(k*pi/32)
  // The values of sin(k*pi/32) and cos(k*pi/32) for k = 0..31 are precomputed
  // and stored using a vector of 32 doubles. Sin(y*pi/32) and cos(y*pi/32) are
  // computed using degree-7 and degree-6 minimax polynomials generated by
  // Sollya respectively.

  // |x| <= pi/16
  if (LIBC_UNLIKELY(x_abs <= 0x3e49'0fdbU)) {

    // |x| < 0x1.d12ed2p-12f
    if (LIBC_UNLIKELY(x_abs < 0x39e8'9769U)) {
      if (LIBC_UNLIKELY(x_abs == 0U)) {
        // For signed zeros.
        return x;
      }
      // When |x| < 2^-12, the relative error of the approximation sin(x) ~ x
      // is:
      //   |sin(x) - x| / |sin(x)| < |x^3| / (6|x|)
      //                           = x^2 / 6
      //                           < 2^-25
      //                           < epsilon(1)/2.
      // So the correctly rounded values of sin(x) are:
      //   = x - sign(x)*eps(x) if rounding mode = FE_TOWARDZERO,
      //                        or (rounding mode = FE_UPWARD and x is
      //                        negative),
      //   = x otherwise.
      // To simplify the rounding decision and make it more efficient, we use
      //   fma(x, -2^-25, x) instead.
      // An exhaustive test shows that this formula work correctly for all
      // rounding modes up to |x| < 0x1.c555dep-11f.
      // Note: to use the formula x - 2^-25*x to decide the correct rounding, we
      // do need fma(x, -2^-25, x) to prevent underflow caused by -2^-25*x when
      // |x| < 2^-125. For targets without FMA instructions, we simply use
      // double for intermediate results as it is more efficient than using an
      // emulated version of FMA.
#if defined(LIBC_TARGET_CPU_HAS_FMA)
      return fputil::multiply_add(x, -0x1.0p-25f, x);
#else
      return static_cast<float>(fputil::multiply_add(xd, -0x1.0p-25, xd));
#endif // LIBC_TARGET_CPU_HAS_FMA
    }

    // |x| < pi/16.
    double xsq = xd * xd;

    // Degree-9 polynomial approximation:
    //   sin(x) ~ x + a_3 x^3 + a_5 x^5 + a_7 x^7 + a_9 x^9
    //          = x (1 + a_3 x^2 + ... + a_9 x^8)
    //          = x * P(x^2)
    // generated by Sollya with the following commands:
    // > display = hexadecimal;
    // > Q = fpminimax(sin(x)/x, [|0, 2, 4, 6, 8|], [|1, D...|], [0, pi/16]);
    double result =
        fputil::polyeval(xsq, 1.0, -0x1.55555555554c6p-3, 0x1.1111111085e65p-7,
                         -0x1.a019f70fb4d4fp-13, 0x1.718d179815e74p-19);
    return static_cast<float>(xd * result);
  }

  if (LIBC_UNLIKELY(x_abs == 0x4619'9998U)) { // x = 0x1.33333p13
    float r = -0x1.63f4bap-2f;
    int rounding = fputil::quick_get_round();
    if ((rounding == FE_DOWNWARD && xbits.is_pos()) ||
        (rounding == FE_UPWARD && xbits.is_neg()))
      r = -0x1.63f4bcp-2f;
    return xbits.is_neg() ? -r : r;
  }

  if (LIBC_UNLIKELY(x_abs >= 0x7f80'0000U)) {
    if (x_abs == 0x7f80'0000U) {
      fputil::set_errno_if_required(EDOM);
      fputil::raise_except_if_required(FE_INVALID);
    }
    return x + FPBits::quiet_nan().get_val();
  }

  // Combine the results with the sine of sum formula:
  //   sin(x) = sin((k + y)*pi/32)
  //          = sin(y*pi/32) * cos(k*pi/32) + cos(y*pi/32) * sin(k*pi/32)
  //          = sin_y * cos_k + (1 + cosm1_y) * sin_k
  //          = sin_y * cos_k + (cosm1_y * sin_k + sin_k)
  double sin_k, cos_k, sin_y, cosm1_y;

  sincosf_eval(xd, x_abs, sin_k, cos_k, sin_y, cosm1_y);

  return static_cast<float>(fputil::multiply_add(
      sin_y, cos_k, fputil::multiply_add(cosm1_y, sin_k, sin_k)));
}

} // namespace LIBC_NAMESPACE_DECL