//===-- Utilities for trigonometric functions -------------------*- C++ -*-===//
//
// 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
//
//===----------------------------------------------------------------------===//

#ifndef LLVM_LIBC_SRC_MATH_GENERIC_RANGE_REDUCTION_H
#define LLVM_LIBC_SRC_MATH_GENERIC_RANGE_REDUCTION_H

#include "src/__support/FPUtil/FPBits.h"
#include "src/__support/FPUtil/except_value_utils.h"
#include "src/__support/FPUtil/multiply_add.h"
#include "src/__support/FPUtil/nearest_integer.h"

namespace __llvm_libc {

namespace generic {

static constexpr uint32_t FAST_PASS_BOUND = 0x4c80'0000U; // 2^26

static constexpr int N_ENTRIES = 8;

// We choose to split bits of 1/pi into 28-bit precision pieces, so that the
// product of x * ONE_OVER_PI_28[i] is exact.
// These are generated by Sollya with:
// > a1 = D(round(1/pi, 28, RN)); a1;
// > a2 = D(round(1/pi - a1, 28, RN)); a2;
// > a3 = D(round(1/pi - a1 - a2, 28, RN)); a3;
// > a4 = D(round(1/pi - a1 - a2 - a3, 28, RN)); a4;
// ...
static constexpr double ONE_OVER_PI_28[N_ENTRIES] = {
    0x1.45f306ep-2,   -0x1.b1bbeaep-33,  0x1.3f84ebp-62,    -0x1.7056592p-92,
    0x1.c0db62ap-121, -0x1.4cd8778p-150, -0x1.bef806cp-179, 0x1.63abdecp-209};

// Exponents of the least significant bits of the corresponding entries in
// ONE_OVER_PI_28.
static constexpr int ONE_OVER_PI_28_LSB_EXP[N_ENTRIES] = {
    -29, -60, -86, -119, -148, -175, -205, -235};

// Return (k mod 2) and y, where
//   k = round(x / pi) and y = (x / pi) - k.
static inline int64_t small_range_reduction(double x, double &y) {
  double prod = x * ONE_OVER_PI_28[0];
  double kd = fputil::nearest_integer(prod);
  y = prod - kd;
  y = fputil::multiply_add(x, ONE_OVER_PI_28[1], y);
  y = fputil::multiply_add(x, ONE_OVER_PI_28[2], y);
  return static_cast<int64_t>(kd);
}

// Return k and y, where
//   k = round(x / pi) and y = (x / pi) - k.
// For large range, there are at most 2 parts of ONE_OVER_PI_28 contributing to
// the unit binary digit (k & 1).  If the least significant bit of x * the least
// significant bit of ONE_OVER_PI_28[i] > 1, we can completely ignore
// ONE_OVER_PI_28[i].
static inline int64_t large_range_reduction(double x, int x_exp, double &y) {
  int idx = 0;
  y = 0;
  int x_lsb_exp = x_exp - fputil::FloatProperties<float>::MANTISSA_WIDTH;

  // Skipping the first parts of 1/pi such that:
  //   LSB of x * LSB of ONE_OVER_PI_28[i] > 1.
  while (x_lsb_exp + ONE_OVER_PI_28_LSB_EXP[idx] > 0)
    ++idx;

  double prod_hi = x * ONE_OVER_PI_28[idx];
  // Get the integral part of x * ONE_OVER_PI_28[idx]
  double k_hi = fputil::nearest_integer(prod_hi);
  // Get the fractional part of x * ONE_OVER_PI_28[idx]
  double frac = prod_hi - k_hi;
  double prod_lo = fputil::multiply_add(x, ONE_OVER_PI_28[idx + 1], frac);
  double k_lo = fputil::nearest_integer(prod_lo);

  // Now y is the fractional parts.
  y = prod_lo - k_lo;
  y = fputil::multiply_add(x, ONE_OVER_PI_28[idx + 2], y);
  y = fputil::multiply_add(x, ONE_OVER_PI_28[idx + 3], y);

  return static_cast<int64_t>(k_hi + k_lo);
}

// Exceptional cases.
static constexpr int N_EXCEPT_SMALL = 4;

static constexpr fputil::ExceptionalValues<float, N_EXCEPT_SMALL> SmallExcepts{
    /* inputs */ {
        0x3fa7832a, // x = 0x1.4f0654p0
        0x46199998, // x = 0x1.33333p13
        0x4afdece4, // x = 0x1.fbd9c8p22
        0x4c2332e9, // x = 0x1.4665d2p25
    },
    /* outputs (RZ, RU offset, RD offset, RN offset) */
    {
        {0x3f7741b5, 1, 0, 1}, // x = 0x1.4f0654p0, sin(x) = 0x1.ee836ap-1 (RZ)
        {0xbeb1fa5d, 0, 1, 0}, // x = 0x1.33333p13, sin(x) = -0x1.63f4bap-2 (RZ)
        {0xbf7fb6e0, 0, 1, 1}, // x = 0x1.fbd9c8p22, sin(x) = -0x1.ff6dcp-1 (RZ)
        {0xbf7fffff, 0, 1,
         1}, // x = 0x1.4665d2p25, sin(x) = -0x1.fffffep-1 (RZ)
    }};

static constexpr int N_EXCEPT_LARGE = 5;

static constexpr fputil::ExceptionalValues<float, N_EXCEPT_LARGE> LargeExcepts{
    /* inputs */ {
        0x523947f6, // x = 0x1.728fecp37
        0x53b146a6, // x = 0x1.628d4cp40
        0x55cafb2a, // x = 0x1.95f654p44
        0x6a1976f1, // x = 0x1.32ede2p85
        0x77584625, // x = 0x1.b08c4ap111
    },
    /* outputs (RZ, RU offset, RD offset, RN offset) */
    {
        {0xbf12791d, 0, 1,
         1}, // x = 0x1.728fecp37, sin(x) = -0x1.24f23ap-1 (RZ)
        {0xbf7fffff, 0, 1,
         1}, // x = 0x1.628d4cp40, sin(x) = -0x1.fffffep-1 (RZ)
        {0xbf7e7a16, 0, 1,
         1}, // x = 0x1.95f654p44, sin(x) = -0x1.fcf42cp-1 (RZ)
        {0x3f7fffff, 1, 0, 1}, // x = 0x1.32ede2p85, sin(x) = 0x1.fffffep-1 (RZ)
        {0xbf7fffff, 0, 1,
         1}, // x = 0x1.b08c4ap111, sin(x) = -0x1.fffffep-1 (RZ)
    }};

} // namespace generic

} // namespace __llvm_libc

#endif // LLVM_LIBC_SRC_MATH_GENERIC_RANGE_REDUCTION_H