shylie/llvm-project
1
0
Fork 0
Code Issues Pull Requests Actions 6 Packages Projects Releases Wiki Activity

[libc][math] Refactor sin implementation to header-only in src/__support/math folder. (#175200)

Browse Source 
Part of #147386

in preparation for:
https://discourse.llvm.org/t/rfc-make-clang-builtin-math-functions-constexpr-with-llvm-libc-to-support-c-23-constexpr-math-functions/86450
This commit is contained in:
lntue 2026-01-09 23:45:24 -05:00 committed by GitHub
parent c9faf8da5e
commit 43b8bc4b6f
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
9 changed files with 239 additions and 182 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
libc/shared/math.h
View File

@ -65,5 +65,6 @@
#include "math/ldexpf16.h"
#include "math/rsqrtf.h"
#include "math/rsqrtf16.h"
#include "math/sin.h"
#endif // LLVM_LIBC_SHARED_MATH_H

23
libc/shared/math/sin.h Normal file
View File

@ -0,0 +1,23 @@
//===-- Shared sin function -------------------------------------*- 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_SHARED_MATH_SIN_H
#define LLVM_LIBC_SHARED_MATH_SIN_H
#include "shared/libc_common.h"
#include "src/__support/math/sin.h"
namespace LIBC_NAMESPACE_DECL {
namespace shared {
using math::sin;
} // namespace shared
} // namespace LIBC_NAMESPACE_DECL
#endif // LLVM_LIBC_SHARED_MATH_SIN_H

16
libc/src/__support/math/CMakeLists.txt
View File

@ -1024,3 +1024,19 @@ add_header_library(
.exp10f_utils
libc.src.__support.FPUtil.multiply_add
)
add_header_library(
sin
HDRS
sin.h
DEPENDS
libc.hdr.errno_macros
libc.src.errno.errno
libc.src.__support.FPUtil.double_double
libc.src.__support.FPUtil.dyadic_float
libc.src.__support.FPUtil.fenv_impl
libc.src.__support.FPUtil.fp_bits
libc.src.__support.math.range_reduction_double
libc.src.__support.math.sincos_eval
libc.src.__support.macros.optimization
)

181
libc/src/__support/math/sin.h Normal file
View File

@ -0,0 +1,181 @@
//===-- Implementation header for sin ---------------------------*- 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 LIBC_SRC___SUPPORT_MATH_SIN_H
#define LIBC_SRC___SUPPORT_MATH_SIN_H
#include "range_reduction_double_common.h"
#include "sincos_eval.h"
#include "src/__support/FPUtil/FEnvImpl.h"
#include "src/__support/FPUtil/FPBits.h"
#include "src/__support/FPUtil/double_double.h"
#include "src/__support/FPUtil/dyadic_float.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
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
#include "range_reduction_double_fma.h"
#else
#include "range_reduction_double_nofma.h"
#endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE
namespace LIBC_NAMESPACE_DECL {
namespace math {
LIBC_INLINE static constexpr double sin(double x) {
using namespace math::range_reduction_double_internal;
using FPBits = typename fputil::FPBits<double>;
FPBits xbits(x);
uint16_t x_e = xbits.get_biased_exponent();
DoubleDouble y;
unsigned k = 0;
LargeRangeReduction range_reduction_large{};
// |x| < 2^16
if (LIBC_LIKELY(x_e < FPBits::EXP_BIAS + FAST_PASS_EXPONENT)) {
// |x| < 2^-7
if (LIBC_UNLIKELY(x_e < FPBits::EXP_BIAS - 7)) {
// |x| < 2^-26, |sin(x) - x| < ulp(x)/2.
if (LIBC_UNLIKELY(x_e < FPBits::EXP_BIAS - 26)) {
// Signed zeros.
if (LIBC_UNLIKELY(x == 0.0))
return x + x; // Make sure it works with FTZ/DAZ.
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
return fputil::multiply_add(x, -0x1.0p-54, x);
#else
if (LIBC_UNLIKELY(x_e < 4)) {
int rounding_mode = fputil::quick_get_round();
if (rounding_mode == FE_TOWARDZERO ||
(xbits.sign() == Sign::POS && rounding_mode == FE_DOWNWARD) ||
(xbits.sign() == Sign::NEG && rounding_mode == FE_UPWARD))
return FPBits(xbits.uintval() - 1).get_val();
}
return fputil::multiply_add(x, -0x1.0p-54, x);
#endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE
}
// No range reduction needed.
k = 0;
y.lo = 0.0;
y.hi = x;
} else {
// Small range reduction.
k = range_reduction_small(x, y);
}
} else {
// Inf or NaN
if (LIBC_UNLIKELY(x_e > 2 * FPBits::EXP_BIAS)) {
// sin(+-Inf) = NaN
if (xbits.is_signaling_nan()) {
fputil::raise_except_if_required(FE_INVALID);
return FPBits::quiet_nan().get_val();
}
if (xbits.get_mantissa() == 0) {
fputil::set_errno_if_required(EDOM);
fputil::raise_except_if_required(FE_INVALID);
}
return x + FPBits::quiet_nan().get_val();
}
// Large range reduction.
k = range_reduction_large.fast(x, y);
}
DoubleDouble sin_y, cos_y;
[[maybe_unused]] double err =
math::sincos_eval_internal::sincos_eval(y, sin_y, cos_y);
// Look up sin(k * pi/128) and cos(k * pi/128)
#ifdef LIBC_MATH_HAS_SMALL_TABLES
// Memory saving versions. Use 65-entry table.
auto get_idx_dd = [](unsigned kk) -> DoubleDouble {
unsigned idx = (kk & 64) ? 64 - (kk & 63) : (kk & 63);
DoubleDouble ans = SIN_K_PI_OVER_128[idx];
if (kk & 128) {
ans.hi = -ans.hi;
ans.lo = -ans.lo;
}
return ans;
};
DoubleDouble sin_k = get_idx_dd(k);
DoubleDouble cos_k = get_idx_dd(k + 64);
#else
// Fast look up version, but needs 256-entry table.
// cos(k * pi/128) = sin(k * pi/128 + pi/2) = sin((k + 64) * pi/128).
DoubleDouble sin_k = SIN_K_PI_OVER_128[k & 255];
DoubleDouble cos_k = SIN_K_PI_OVER_128[(k + 64) & 255];
#endif
// After range reduction, k = round(x * 128 / pi) and y = x - k * (pi / 128).
// So k is an integer and -pi / 256 <= y <= pi / 256.
// Then sin(x) = sin((k * pi/128 + y)
// = sin(y) * cos(k*pi/128) + cos(y) * sin(k*pi/128)
DoubleDouble sin_k_cos_y = fputil::quick_mult(cos_y, sin_k);
DoubleDouble cos_k_sin_y = fputil::quick_mult(sin_y, cos_k);
DoubleDouble rr = fputil::exact_add<false>(sin_k_cos_y.hi, cos_k_sin_y.hi);
rr.lo += sin_k_cos_y.lo + cos_k_sin_y.lo;
#ifdef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
return rr.hi + rr.lo;
#else
// Accurate test and pass for correctly rounded implementation.
double rlp = rr.lo + err;
double rlm = rr.lo - err;
double r_upper = rr.hi + rlp; // (rr.lo + ERR);
double r_lower = rr.hi + rlm; // (rr.lo - ERR);
// Ziv's rounding test.
if (LIBC_LIKELY(r_upper == r_lower))
return r_upper;
Float128 u_f128, sin_u, cos_u;
if (LIBC_LIKELY(x_e < FPBits::EXP_BIAS + FAST_PASS_EXPONENT))
u_f128 = range_reduction_small_f128(x);
else
u_f128 = range_reduction_large.accurate();
math::sincos_eval_internal::sincos_eval(u_f128, sin_u, cos_u);
auto get_sin_k = [](unsigned kk) -> Float128 {
unsigned idx = (kk & 64) ? 64 - (kk & 63) : (kk & 63);
Float128 ans = SIN_K_PI_OVER_128_F128[idx];
if (kk & 128)
ans.sign = Sign::NEG;
return ans;
};
// cos(k * pi/128) = sin(k * pi/128 + pi/2) = sin((k + 64) * pi/128).
Float128 sin_k_f128 = get_sin_k(k);
Float128 cos_k_f128 = get_sin_k(k + 64);
// sin(x) = sin(k * pi/128 + u)
// = sin(u) * cos(k*pi/128) + cos(u) * sin(k*pi/128)
Float128 r = fputil::quick_add(fputil::quick_mul(sin_k_f128, cos_u),
fputil::quick_mul(cos_k_f128, sin_u));
// TODO: Add assertion if Ziv's accuracy tests fail in debug mode.
// https://github.com/llvm/llvm-project/issues/96452.
return static_cast<double>(r);
#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
}
} // namespace math
} // namespace LIBC_NAMESPACE_DECL
#endif // LIBC_SRC___SUPPORT_MATH_SIN_H

12
libc/src/math/generic/CMakeLists.txt
View File

@ -363,17 +363,7 @@ add_entrypoint_object(
HDRS
../sin.h
DEPENDS
libc.src.__support.math.range_reduction_double
libc.src.__support.math.sincos_eval
libc.hdr.errno_macros
libc.src.errno.errno
libc.src.__support.FPUtil.double_double
libc.src.__support.FPUtil.dyadic_float
libc.src.__support.FPUtil.fenv_impl
libc.src.__support.FPUtil.fp_bits
libc.src.__support.FPUtil.multiply_add
libc.src.__support.FPUtil.rounding_mode
libc.src.__support.macros.optimization
libc.src.__support.math.sin
)
add_entrypoint_object(

168
libc/src/math/generic/sin.cpp
View File

@ -7,174 +7,10 @@
//===----------------------------------------------------------------------===//
#include "src/math/sin.h"
#include "hdr/errno_macros.h"
#include "src/__support/FPUtil/FEnvImpl.h"
#include "src/__support/FPUtil/FPBits.h"
#include "src/__support/FPUtil/double_double.h"
#include "src/__support/FPUtil/dyadic_float.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
#include "src/__support/math/range_reduction_double_common.h"
#include "src/__support/math/sincos_eval.h"
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
#include "src/__support/math/range_reduction_double_fma.h"
#else
#include "src/__support/math/range_reduction_double_nofma.h"
#endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE
#include "src/__support/math/sin.h"
namespace LIBC_NAMESPACE_DECL {
using DoubleDouble = fputil::DoubleDouble;
using Float128 = typename fputil::DyadicFloat<128>;
LLVM_LIBC_FUNCTION(double, sin, (double x)) {
using namespace math::range_reduction_double_internal;
using FPBits = typename fputil::FPBits<double>;
FPBits xbits(x);
uint16_t x_e = xbits.get_biased_exponent();
DoubleDouble y;
unsigned k;
LargeRangeReduction range_reduction_large{};
// |x| < 2^16
if (LIBC_LIKELY(x_e < FPBits::EXP_BIAS + FAST_PASS_EXPONENT)) {
// |x| < 2^-7
if (LIBC_UNLIKELY(x_e < FPBits::EXP_BIAS - 7)) {
// |x| < 2^-26, |sin(x) - x| < ulp(x)/2.
if (LIBC_UNLIKELY(x_e < FPBits::EXP_BIAS - 26)) {
// Signed zeros.
if (LIBC_UNLIKELY(x == 0.0))
return x + x; // Make sure it works with FTZ/DAZ.
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
return fputil::multiply_add(x, -0x1.0p-54, x);
#else
if (LIBC_UNLIKELY(x_e < 4)) {
int rounding_mode = fputil::quick_get_round();
if (rounding_mode == FE_TOWARDZERO ||
(xbits.sign() == Sign::POS && rounding_mode == FE_DOWNWARD) ||
(xbits.sign() == Sign::NEG && rounding_mode == FE_UPWARD))
return FPBits(xbits.uintval() - 1).get_val();
}
return fputil::multiply_add(x, -0x1.0p-54, x);
#endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE
}
// No range reduction needed.
k = 0;
y.lo = 0.0;
y.hi = x;
} else {
// Small range reduction.
k = range_reduction_small(x, y);
}
} else {
// Inf or NaN
if (LIBC_UNLIKELY(x_e > 2 * FPBits::EXP_BIAS)) {
// sin(+-Inf) = NaN
if (xbits.is_signaling_nan()) {
fputil::raise_except_if_required(FE_INVALID);
return FPBits::quiet_nan().get_val();
}
if (xbits.get_mantissa() == 0) {
fputil::set_errno_if_required(EDOM);
fputil::raise_except_if_required(FE_INVALID);
}
return x + FPBits::quiet_nan().get_val();
}
// Large range reduction.
k = range_reduction_large.fast(x, y);
}
DoubleDouble sin_y, cos_y;
[[maybe_unused]] double err =
math::sincos_eval_internal::sincos_eval(y, sin_y, cos_y);
// Look up sin(k * pi/128) and cos(k * pi/128)
#ifdef LIBC_MATH_HAS_SMALL_TABLES
// Memory saving versions. Use 65-entry table.
auto get_idx_dd = [](unsigned kk) -> DoubleDouble {
unsigned idx = (kk & 64) ? 64 - (kk & 63) : (kk & 63);
DoubleDouble ans = SIN_K_PI_OVER_128[idx];
if (kk & 128) {
ans.hi = -ans.hi;
ans.lo = -ans.lo;
}
return ans;
};
DoubleDouble sin_k = get_idx_dd(k);
DoubleDouble cos_k = get_idx_dd(k + 64);
#else
// Fast look up version, but needs 256-entry table.
// cos(k * pi/128) = sin(k * pi/128 + pi/2) = sin((k + 64) * pi/128).
DoubleDouble sin_k = SIN_K_PI_OVER_128[k & 255];
DoubleDouble cos_k = SIN_K_PI_OVER_128[(k + 64) & 255];
#endif
// After range reduction, k = round(x * 128 / pi) and y = x - k * (pi / 128).
// So k is an integer and -pi / 256 <= y <= pi / 256.
// Then sin(x) = sin((k * pi/128 + y)
// = sin(y) * cos(k*pi/128) + cos(y) * sin(k*pi/128)
DoubleDouble sin_k_cos_y = fputil::quick_mult(cos_y, sin_k);
DoubleDouble cos_k_sin_y = fputil::quick_mult(sin_y, cos_k);
DoubleDouble rr = fputil::exact_add<false>(sin_k_cos_y.hi, cos_k_sin_y.hi);
rr.lo += sin_k_cos_y.lo + cos_k_sin_y.lo;
#ifdef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
return rr.hi + rr.lo;
#else
// Accurate test and pass for correctly rounded implementation.
double rlp = rr.lo + err;
double rlm = rr.lo - err;
double r_upper = rr.hi + rlp; // (rr.lo + ERR);
double r_lower = rr.hi + rlm; // (rr.lo - ERR);
// Ziv's rounding test.
if (LIBC_LIKELY(r_upper == r_lower))
return r_upper;
Float128 u_f128, sin_u, cos_u;
if (LIBC_LIKELY(x_e < FPBits::EXP_BIAS + FAST_PASS_EXPONENT))
u_f128 = range_reduction_small_f128(x);
else
u_f128 = range_reduction_large.accurate();
math::sincos_eval_internal::sincos_eval(u_f128, sin_u, cos_u);
auto get_sin_k = [](unsigned kk) -> Float128 {
unsigned idx = (kk & 64) ? 64 - (kk & 63) : (kk & 63);
Float128 ans = SIN_K_PI_OVER_128_F128[idx];
if (kk & 128)
ans.sign = Sign::NEG;
return ans;
};
// cos(k * pi/128) = sin(k * pi/128 + pi/2) = sin((k + 64) * pi/128).
Float128 sin_k_f128 = get_sin_k(k);
Float128 cos_k_f128 = get_sin_k(k + 64);
// sin(x) = sin(k * pi/128 + u)
// = sin(u) * cos(k*pi/128) + cos(u) * sin(k*pi/128)
Float128 r = fputil::quick_add(fputil::quick_mul(sin_k_f128, cos_u),
fputil::quick_mul(cos_k_f128, sin_u));
// TODO: Add assertion if Ziv's accuracy tests fail in debug mode.
// https://github.com/llvm/llvm-project/issues/96452.
return static_cast<double>(r);
#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
}
LLVM_LIBC_FUNCTION(double, sin, (double x)) { return math::sin(x); }
} // namespace LIBC_NAMESPACE_DECL

1
libc/test/shared/CMakeLists.txt
View File

@ -61,4 +61,5 @@ add_fp_unittest(
libc.src.__support.math.ldexpf16
libc.src.__support.math.rsqrtf
libc.src.__support.math.rsqrtf16
libc.src.__support.math.sin
)

1
libc/test/shared/shared_math_test.cpp
View File

@ -90,6 +90,7 @@ TEST(LlvmLibcSharedMathTest, AllDouble) {
EXPECT_FP_EQ(0x1p+0, LIBC_NAMESPACE::shared::exp2(0.0));
EXPECT_FP_EQ(0x1p+0, LIBC_NAMESPACE::shared::exp10(0.0));
EXPECT_FP_EQ(0x0p+0, LIBC_NAMESPACE::shared::expm1(0.0));
EXPECT_FP_EQ(0.0, LIBC_NAMESPACE::shared::sin(0.0));
}
#ifdef LIBC_TYPES_HAS_FLOAT128

18
utils/bazel/llvm-project-overlay/libc/BUILD.bazel
View File

@ -3192,6 +3192,18 @@ libc_support_library(
],
)
libc_support_library(
name = "__support_math_sin",
hdrs = ["src/__support/math/sin.h"],
deps = [
":__support_fputil_multiply_add",
":__support_macros_optimization",
":__support_macros_properties_cpu_features",
":__support_range_reduction_double",
":__support_sincos_eval",
],
)
libc_support_library(
name = "__support_math_sinhfcoshf_utils",
hdrs = ["src/__support/math/sinhfcoshf_utils.h"],
@ -4748,11 +4760,7 @@ libc_math_function(name = "setpayloadsigf16")
libc_math_function(
name = "sin",
additional_deps = [
":__support_fputil_multiply_add",
":__support_macros_optimization",
":__support_macros_properties_cpu_features",
":__support_range_reduction_double",
":__support_sincos_eval",
":__support_math_sin",
],
)