[llvm-branch-commits] [libc] [llvm] [libc][math] Refactor cos implementation to header-only in src/__support/math folder. (PR #151883)
via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Sun Aug 3 12:26:38 PDT 2025
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-libc
Author: Muhammad Bassiouni (bassiounix)
<details>
<summary>Changes</summary>
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
---
Patch is 33.10 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/151883.diff
16 Files Affected:
- (modified) libc/shared/math.h (+1)
- (added) libc/shared/math/cos.h (+23)
- (modified) libc/src/__support/math/CMakeLists.txt (+47)
- (added) libc/src/__support/math/cos.h (+173)
- (renamed) libc/src/__support/math/range_reduction_double_common.h (+14-4)
- (renamed) libc/src/__support/math/range_reduction_double_fma.h (+9-1)
- (renamed) libc/src/__support/math/range_reduction_double_nofma.h (+9-1)
- (renamed) libc/src/__support/math/sincos_eval.h (+6-2)
- (modified) libc/src/math/generic/CMakeLists.txt (+6-44)
- (modified) libc/src/math/generic/cos.cpp (+2-153)
- (modified) libc/src/math/generic/sin.cpp (+8-6)
- (modified) libc/src/math/generic/sincos.cpp (+8-6)
- (modified) libc/src/math/generic/tan.cpp (+4-3)
- (modified) libc/test/shared/CMakeLists.txt (+1)
- (modified) libc/test/shared/shared_math_test.cpp (+2-1)
- (modified) utils/bazel/llvm-project-overlay/libc/BUILD.bazel (+55-46)
``````````diff
diff --git a/libc/shared/math.h b/libc/shared/math.h
index ea645f0afedbc..a5581ed4272a3 100644
--- a/libc/shared/math.h
+++ b/libc/shared/math.h
@@ -32,6 +32,7 @@
#include "math/atanhf16.h"
#include "math/cbrt.h"
#include "math/cbrtf.h"
+#include "math/cos.h"
#include "math/erff.h"
#include "math/exp.h"
#include "math/exp10.h"
diff --git a/libc/shared/math/cos.h b/libc/shared/math/cos.h
new file mode 100644
index 0000000000000..c498550f098b4
--- /dev/null
+++ b/libc/shared/math/cos.h
@@ -0,0 +1,23 @@
+//===-- Shared cos 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_COS_H
+#define LLVM_LIBC_SHARED_MATH_COS_H
+
+#include "shared/libc_common.h"
+#include "src/__support/math/cos.h"
+
+namespace LIBC_NAMESPACE_DECL {
+namespace shared {
+
+using math::cos;
+
+} // namespace shared
+} // namespace LIBC_NAMESPACE_DECL
+
+#endif // LLVM_LIBC_SHARED_MATH_COS_H
diff --git a/libc/src/__support/math/CMakeLists.txt b/libc/src/__support/math/CMakeLists.txt
index fe928a8fadd5e..24844063fcd24 100644
--- a/libc/src/__support/math/CMakeLists.txt
+++ b/libc/src/__support/math/CMakeLists.txt
@@ -357,6 +357,24 @@ add_header_library(
libc.src.__support.macros.optimization
)
+add_header_library(
+ cos
+ HDRS
+ cos.h
+ DEPENDS
+ 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.except_value_utils
+ libc.src.__support.FPUtil.fenv_impl
+ libc.src.__support.FPUtil.fp_bits
+ libc.src.__support.math.range_reduction_double
+ libc.src.__support.macros.optimization
+)
+
+
add_header_library(
erff
HDRS
@@ -613,3 +631,32 @@ add_header_library(
libc.src.__support.macros.optimization
libc.src.__support.macros.properties.cpu_features
)
+
+add_header_library(
+ range_reduction_double
+ HDRS
+ range_reduction_double_common.h
+ range_reduction_double_fma.h
+ range_reduction_double_nofma.h
+ DEPENDS
+ libc.src.__support.FPUtil.double_double
+ libc.src.__support.FPUtil.dyadic_float
+ libc.src.__support.FPUtil.fp_bits
+ libc.src.__support.FPUtil.fma
+ libc.src.__support.FPUtil.multiply_add
+ libc.src.__support.FPUtil.nearest_integer
+ libc.src.__support.common
+ libc.src.__support.integer_literals
+)
+
+add_header_library(
+ sincos_eval
+ HDRS
+ sincos_eval.h
+ DEPENDS
+ libc.src.__support.FPUtil.double_double
+ libc.src.__support.FPUtil.dyadic_float
+ libc.src.__support.FPUtil.multiply_add
+ libc.src.__support.FPUtil.polyeval
+ libc.src.__support.integer_literals
+)
diff --git a/libc/src/__support/math/cos.h b/libc/src/__support/math/cos.h
new file mode 100644
index 0000000000000..0802f9e4f6e49
--- /dev/null
+++ b/libc/src/__support/math/cos.h
@@ -0,0 +1,173 @@
+//===-- Implementation header for cos ---------------------------*- 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_COS_H
+#define LIBC_SRC___SUPPORT_MATH_COS_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/FPUtil/except_value_utils.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 cos(double x) {
+ using namespace range_reduction_double_internal;
+ using DoubleDouble = fputil::DoubleDouble;
+ 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^-27
+ if (LIBC_UNLIKELY(x_e < FPBits::EXP_BIAS - 27)) {
+ // Signed zeros.
+ if (LIBC_UNLIKELY(x == 0.0))
+ return 1.0;
+
+ // For |x| < 2^-27, |cos(x) - 1| < |x|^2/2 < 2^-54 = ulp(1 - 2^-53)/2.
+ return fputil::round_result_slightly_down(1.0);
+ }
+ // 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)) {
+ if (xbits.is_signaling_nan()) {
+ fputil::raise_except_if_required(FE_INVALID);
+ return FPBits::quiet_nan().get_val();
+ }
+ // cos(+-Inf) = NaN
+ 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 msin_k = get_idx_dd(k + 128);
+ DoubleDouble cos_k = get_idx_dd(k + 64);
+#else
+ // Fast look up version, but needs 256-entry table.
+ // -sin(k * pi/128) = sin((k + 128) * pi/128)
+ // cos(k * pi/128) = sin(k * pi/128 + pi/2) = sin((k + 64) * pi/128).
+ DoubleDouble msin_k = SIN_K_PI_OVER_128[(k + 128) & 255];
+ DoubleDouble cos_k = SIN_K_PI_OVER_128[(k + 64) & 255];
+#endif // LIBC_MATH_HAS_SMALL_TABLES
+
+ // 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 cos(x) = cos((k * pi/128 + y)
+ // = cos(y) * cos(k*pi/128) - sin(y) * sin(k*pi/128)
+ DoubleDouble cos_k_cos_y = fputil::quick_mult(cos_y, cos_k);
+ DoubleDouble msin_k_sin_y = fputil::quick_mult(sin_y, msin_k);
+
+ DoubleDouble rr = fputil::exact_add<false>(cos_k_cos_y.hi, msin_k_sin_y.hi);
+ rr.lo += msin_k_sin_y.lo + cos_k_cos_y.lo;
+
+#ifdef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
+ return rr.hi + rr.lo;
+#else
+ using Float128 = typename fputil::DyadicFloat<128>;
+ 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;
+ };
+
+ // -sin(k * pi/128) = sin((k + 128) * pi/128)
+ // cos(k * pi/128) = sin(k * pi/128 + pi/2) = sin((k + 64) * pi/128).
+ Float128 msin_k_f128 = get_sin_k(k + 128);
+ Float128 cos_k_f128 = get_sin_k(k + 64);
+
+ // cos(x) = cos((k * pi/128 + u)
+ // = cos(u) * cos(k*pi/128) - sin(u) * sin(k*pi/128)
+ Float128 r = fputil::quick_add(fputil::quick_mul(cos_k_f128, cos_u),
+ fputil::quick_mul(msin_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_COS_H
diff --git a/libc/src/math/generic/range_reduction_double_common.h b/libc/src/__support/math/range_reduction_double_common.h
similarity index 98%
rename from libc/src/math/generic/range_reduction_double_common.h
rename to libc/src/__support/math/range_reduction_double_common.h
index a93ee25201813..0cace009abf90 100644
--- a/libc/src/math/generic/range_reduction_double_common.h
+++ b/libc/src/__support/math/range_reduction_double_common.h
@@ -20,6 +20,10 @@
namespace LIBC_NAMESPACE_DECL {
+namespace math {
+
+namespace range_reduction_double_internal {
+
#ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE
static constexpr unsigned SPLIT = fputil::DefaultSplit<double>::VALUE;
#else
@@ -40,7 +44,7 @@ using Float128 = LIBC_NAMESPACE::fputil::DyadicFloat<128>;
// Error bound:
// |(x - k * pi/128) - (u_hi + u_lo)| <= max(ulp(ulp(u_hi)), 2^-119)
// <= 2^-111.
-LIBC_INLINE unsigned range_reduction_small(double x, DoubleDouble &u) {
+LIBC_INLINE static unsigned range_reduction_small(double x, DoubleDouble &u) {
// Values of -pi/128 used for inputs with absolute value <= 2^16.
// The first 3 parts are generated with (53 - 21 = 32)-bit precision, so that
// the product k * MPI_OVER_128[i] is exact.
@@ -267,13 +271,15 @@ struct LargeRangeReduction {
}
#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
+ LIBC_INLINE LargeRangeReduction() = default;
+
private:
// Index of x in the look-up table ONE_TWENTY_EIGHT_OVER_PI.
- unsigned idx;
+ unsigned idx = 0;
// x scaled down by 2^(-16 *(idx - 3))).
- double x_reduced;
+ double x_reduced = 0;
// Parts of (x * 128/pi) mod 1.
- double y_hi, y_lo;
+ double y_hi = 0, y_lo = 0;
DoubleDouble y_mid;
};
@@ -369,6 +375,10 @@ static constexpr Float128 SIN_K_PI_OVER_128_F128[65] = {
};
#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
+} // namespace range_reduction_double_internal
+
+} // namespace math
+
} // namespace LIBC_NAMESPACE_DECL
#endif // LLVM_LIBC_SRC_MATH_GENERIC_RANGE_REDUCTION_DOUBLE_COMMON_H
diff --git a/libc/src/math/generic/range_reduction_double_fma.h b/libc/src/__support/math/range_reduction_double_fma.h
similarity index 98%
rename from libc/src/math/generic/range_reduction_double_fma.h
rename to libc/src/__support/math/range_reduction_double_fma.h
index 160fb2461fe21..156e51befcd04 100644
--- a/libc/src/math/generic/range_reduction_double_fma.h
+++ b/libc/src/__support/math/range_reduction_double_fma.h
@@ -16,10 +16,14 @@
#include "src/__support/common.h"
#include "src/__support/macros/config.h"
#include "src/__support/macros/optimization.h"
-#include "src/math/generic/range_reduction_double_common.h"
+#include "src/__support/math/range_reduction_double_common.h"
namespace LIBC_NAMESPACE_DECL {
+namespace math {
+
+namespace range_reduction_double_internal {
+
using LIBC_NAMESPACE::fputil::DoubleDouble;
LIBC_INLINE unsigned LargeRangeReduction::fast(double x, DoubleDouble &u) {
@@ -341,6 +345,10 @@ LIBC_INLINE constexpr DoubleDouble SIN_K_PI_OVER_128[] = {
#endif // !LIBC_MATH_HAS_SMALL_TABLES
};
+} // namespace range_reduction_double_internal
+
+} // namespace math
+
} // namespace LIBC_NAMESPACE_DECL
#endif // LLVM_LIBC_SRC_MATH_GENERIC_RANGE_REDUCTION_DOUBLE_FMA_H
diff --git a/libc/src/math/generic/range_reduction_double_nofma.h b/libc/src/__support/math/range_reduction_double_nofma.h
similarity index 98%
rename from libc/src/math/generic/range_reduction_double_nofma.h
rename to libc/src/__support/math/range_reduction_double_nofma.h
index 9d13d246ce91f..d89cd72e94005 100644
--- a/libc/src/math/generic/range_reduction_double_nofma.h
+++ b/libc/src/__support/math/range_reduction_double_nofma.h
@@ -16,10 +16,14 @@
#include "src/__support/common.h"
#include "src/__support/macros/config.h"
#include "src/__support/macros/optimization.h"
-#include "src/math/generic/range_reduction_double_common.h"
+#include "src/__support/math/range_reduction_double_common.h"
namespace LIBC_NAMESPACE_DECL {
+namespace math {
+
+namespace range_reduction_double_internal {
+
using fputil::DoubleDouble;
LIBC_INLINE unsigned LargeRangeReduction::fast(double x, DoubleDouble &u) {
@@ -342,6 +346,10 @@ LIBC_INLINE constexpr DoubleDouble SIN_K_PI_OVER_128[] = {
#endif // !LIBC_MATH_HAS_SMALL_TABLES
};
+} // namespace range_reduction_double_internal
+
+} // namespace math
+
} // namespace LIBC_NAMESPACE_DECL
#endif // LLVM_LIBC_SRC_MATH_GENERIC_RANGE_REDUCTION_DOUBLE_NOFMA_H
diff --git a/libc/src/math/generic/sincos_eval.h b/libc/src/__support/math/sincos_eval.h
similarity index 98%
rename from libc/src/math/generic/sincos_eval.h
rename to libc/src/__support/math/sincos_eval.h
index 41a4c75849ff4..fc741af19551f 100644
--- a/libc/src/math/generic/sincos_eval.h
+++ b/libc/src/__support/math/sincos_eval.h
@@ -18,7 +18,9 @@
namespace LIBC_NAMESPACE_DECL {
-namespace generic {
+namespace math {
+
+namespace sincos_eval_internal {
using fputil::DoubleDouble;
using Float128 = fputil::DyadicFloat<128>;
@@ -131,7 +133,9 @@ LIBC_INLINE void sincos_eval(const Float128 &u, Float128 &sin_u,
COS_COEFFS[6]);
}
-} // namespace generic
+} // namespace sincos_eval_internal
+
+} // namespace math
} // namespace LIBC_NAMESPACE_DECL
diff --git a/libc/src/math/generic/CMakeLists.txt b/libc/src/math/generic/CMakeLists.txt
index 44c17cf5b0770..b224808e54e45 100644
--- a/libc/src/math/generic/CMakeLists.txt
+++ b/libc/src/math/generic/CMakeLists.txt
@@ -275,23 +275,6 @@ add_header_library(
libc.src.__support.common
)
-add_header_library(
- range_reduction_double
- HDRS
- range_reduction_double_common.h
- range_reduction_double_fma.h
- range_reduction_double_nofma.h
- DEPENDS
- libc.src.__support.FPUtil.double_double
- libc.src.__support.FPUtil.dyadic_float
- libc.src.__support.FPUtil.fp_bits
- libc.src.__support.FPUtil.fma
- libc.src.__support.FPUtil.multiply_add
- libc.src.__support.FPUtil.nearest_integer
- libc.src.__support.common
- libc.src.__support.integer_literals
-)
-
add_header_library(
sincosf_utils
HDRS
@@ -313,18 +296,6 @@ add_header_library(
libc.src.__support.common
)
-add_header_library(
- sincos_eval
- HDRS
- sincos_eval.h
- DEPENDS
- libc.src.__support.FPUtil.double_double
- libc.src.__support.FPUtil.dyadic_float
- libc.src.__support.FPUtil.multiply_add
- libc.src.__support.FPUtil.polyeval
- libc.src.__support.integer_literals
-)
-
add_entrypoint_object(
cos
SRCS
@@ -332,16 +303,7 @@ add_entrypoint_object(
HDRS
../cos.h
DEPENDS
- .range_reduction_double
- .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.except_value_utils
- libc.src.__support.FPUtil.fenv_impl
- libc.src.__support.FPUtil.fp_bits
- libc.src.__support.macros.optimization
+ libc.src.__support.math.cos
)
add_entrypoint_object(
@@ -420,8 +382,8 @@ add_entrypoint_object(
HDRS
../sin.h
DEPENDS
- .range_reduction_double
- .sincos_eval
+ 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
@@ -480,8 +442,8 @@ add_entrypoint_object(
HDRS
../sincos.h
DEPENDS
- .range_reduction_double
- .sincos_eval
+ 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
@@ -553,7 +515,7 @@ add_entrypoint_object(
HDRS
../tan.h
DEPENDS
- .range_reduction_double
+ libc.src.__support.math.range_reduction_double
libc.hdr.errno_macros
libc.src.errno.errno
libc.src.__support.FPUtil.double_double
diff --git a/libc/src/math/generic/cos.cpp b/libc/src/math/generic/cos.cpp
index 5da0f86812a89..aabf3bc7edcb0 100644
--- a/libc/src/math/generic/cos.cpp
+++ b/libc/src/math/generic/cos.cpp
@@ -7,161 +7,10 @@
//===----------------------------------------------------------------------===//
#include "src/math/cos.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/except_value_utils.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/math/generic/range_reduction_double_common.h"
-#include "src/math/generic/sincos_eval.h"
-
-#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
+#include "src/__support/math/cos.h"
namespace LIBC_NAMESPACE_DECL {
-using DoubleDouble = fputil::DoubleDouble;
-using Float128 = typename fputil::DyadicFloat<128>;
-
-LLVM_LIBC_FUNCTION(double, cos, (double x)) {
- 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^-27
- if (LIBC_UNLIKELY(x_e < FPBits::EXP_BIAS - 27)) {
- // Signed zeros.
- if (LIBC_UNLIKELY(x == 0.0))
- return 1.0;
-
- // For |x| < 2^-27, |cos(x) - 1| < |x|^2/2 < 2^-54 = ulp(1 - 2^-53)/2.
- return fputil::round_result_slightly_down(1.0);
- }
- // 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)) {
- if (xbits.is_signaling_nan()) {
- fputil::raise_except_if_required(FE_INVALID);
- return FPBits::quiet_nan().get_val();
- }
- // cos(+-Inf) = NaN
- 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 = generic::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 msin_k = get_idx_dd(k + 128);
- DoubleDouble cos_k = get_idx_dd(k + 64);
-#else
- // Fast look up version, but needs 256-entry table.
- // -sin(k * pi/128) = sin((k + 128) * pi/128)
- // cos(k * pi/128) = sin(k * pi/128 + pi/2) =...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/151883
More information about the llvm-branch-commits
mailing list