[libc-commits] [libc] [llvm] [libc][math] Refactor asinpif16 to Header Only (PR #179021)
Julian Pokrovsky via libc-commits
libc-commits at lists.llvm.org
Fri Jan 30 23:26:35 PST 2026
https://github.com/raventid updated https://github.com/llvm/llvm-project/pull/179021
>From 3a60c6fa01a73e658d2fae3f0d181c538fbddc7a Mon Sep 17 00:00:00 2001
From: raventid <juliankul at gmail.com>
Date: Sat, 31 Jan 2026 15:02:26 +0800
Subject: [PATCH 1/3] [libc][math] Refactor asinpif16 to Header Only
Resolves https://github.com/llvm/llvm-project/issues/178103
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
---
libc/shared/math.h | 1 +
libc/shared/math/asinpif16.h | 28 ++++
libc/src/__support/math/CMakeLists.txt | 14 ++
libc/src/__support/math/asinpif16.h | 138 ++++++++++++++++++
libc/src/math/generic/CMakeLists.txt | 11 +-
libc/src/math/generic/asinpif16.cpp | 113 +-------------
libc/test/shared/CMakeLists.txt | 1 +
libc/test/shared/shared_math_test.cpp | 1 +
.../llvm-project-overlay/libc/BUILD.bazel | 23 +++
9 files changed, 209 insertions(+), 121 deletions(-)
create mode 100644 libc/shared/math/asinpif16.h
create mode 100644 libc/src/__support/math/asinpif16.h
diff --git a/libc/shared/math.h b/libc/shared/math.h
index 452fe3fddc911..bb08568e26d3a 100644
--- a/libc/shared/math.h
+++ b/libc/shared/math.h
@@ -22,6 +22,7 @@
#include "math/asinf16.h"
#include "math/asinhf.h"
#include "math/asinhf16.h"
+#include "math/asinpif16.h"
#include "math/atan.h"
#include "math/atan2.h"
#include "math/atan2f.h"
diff --git a/libc/shared/math/asinpif16.h b/libc/shared/math/asinpif16.h
new file mode 100644
index 0000000000000..b96449731d2cb
--- /dev/null
+++ b/libc/shared/math/asinpif16.h
@@ -0,0 +1,28 @@
+//===-- Shared asinpif16 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_ASINPIF16_H
+#define LLVM_LIBC_SHARED_MATH_ASINPIF16_H
+
+#include "shared/libc_common.h"
+
+#ifdef LIBC_TYPES_HAS_FLOAT16
+
+#include "src/__support/math/asinpif16.h"
+
+namespace LIBC_NAMESPACE_DECL {
+namespace shared {
+
+using math::asinpif16;
+
+} // namespace shared
+} // namespace LIBC_NAMESPACE_DECL
+
+#endif // LIBC_TYPES_HAS_FLOAT16
+
+#endif // LLVM_LIBC_SHARED_MATH_ASINPIF16_H
diff --git a/libc/src/__support/math/CMakeLists.txt b/libc/src/__support/math/CMakeLists.txt
index 0fa48661e58e8..e22c229ff8104 100644
--- a/libc/src/__support/math/CMakeLists.txt
+++ b/libc/src/__support/math/CMakeLists.txt
@@ -332,6 +332,20 @@ add_header_library(
libc.src.__support.macros.optimization
)
+add_header_library(
+ asinpif16
+ HDRS
+ asinpif16.h
+ DEPENDS
+ libc.src.__support.FPUtil.fenv_impl
+ libc.src.__support.FPUtil.fp_bits
+ libc.src.__support.FPUtil.polyeval
+ libc.src.__support.FPUtil.cast
+ libc.src.__support.FPUtil.multiply_add
+ libc.src.__support.FPUtil.sqrt
+ libc.src.__support.macros.optimization
+)
+
add_header_library(
cbrt
HDRS
diff --git a/libc/src/__support/math/asinpif16.h b/libc/src/__support/math/asinpif16.h
new file mode 100644
index 0000000000000..248704d2eb738
--- /dev/null
+++ b/libc/src/__support/math/asinpif16.h
@@ -0,0 +1,138 @@
+//===-- Implementation header for asinpif16 ---------------------*- 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___SUPPORT_MATH_ASINPIF16_H
+#define LLVM_LIBC_SRC___SUPPORT_MATH_ASINPIF16_H
+
+#include "include/llvm-libc-macros/float16-macros.h"
+
+#ifdef LIBC_TYPES_HAS_FLOAT16
+
+#include "src/__support/FPUtil/FEnvImpl.h"
+#include "src/__support/FPUtil/FPBits.h"
+#include "src/__support/FPUtil/PolyEval.h"
+#include "src/__support/FPUtil/cast.h"
+#include "src/__support/FPUtil/multiply_add.h"
+#include "src/__support/FPUtil/sqrt.h"
+#include "src/__support/macros/optimization.h"
+
+namespace LIBC_NAMESPACE_DECL {
+
+namespace math {
+
+LIBC_INLINE static constexpr float16 asinpif16(float16 x) {
+ using FPBits = fputil::FPBits<float16>;
+
+ FPBits xbits(x);
+ bool is_neg = xbits.is_neg();
+ double x_abs = fputil::cast<double>(xbits.abs().get_val());
+
+ auto signed_result = [is_neg](auto r) -> auto { return is_neg ? -r : r; };
+
+ if (LIBC_UNLIKELY(x_abs > 1.0)) {
+ // aspinf16(NaN) = NaN
+ if (xbits.is_nan()) {
+ if (xbits.is_signaling_nan()) {
+ fputil::raise_except_if_required(FE_INVALID);
+ return FPBits::quiet_nan().get_val();
+ }
+ return x;
+ }
+
+ // 1 < |x| <= +/-inf
+ fputil::raise_except_if_required(FE_INVALID);
+ fputil::set_errno_if_required(EDOM);
+
+ return FPBits::quiet_nan().get_val();
+ }
+
+ // the coefficients for the polynomial approximation of asin(x)/pi in the
+ // range [0, 0.5] extracted using python-sympy
+ //
+ // Python code to generate the coefficients:
+ // > from sympy import *
+ // > import math
+ // > x = symbols('x')
+ // > print(series(asin(x)/math.pi, x, 0, 21))
+ //
+ // OUTPUT:
+ //
+ // 0.318309886183791*x + 0.0530516476972984*x**3 + 0.0238732414637843*x**5 +
+ // 0.0142102627760621*x**7 + 0.00967087327815336*x**9 +
+ // 0.00712127941391293*x**11 + 0.00552355646848375*x**13 +
+ // 0.00444514782463692*x**15 + 0.00367705242846804*x**17 +
+ // 0.00310721681820837*x**19 + O(x**21)
+ //
+ // it's very accurate in the range [0, 0.5] and has a maximum error of
+ // 0.0000000000000001 in the range [0, 0.5].
+ constexpr double POLY_COEFFS[] = {
+ 0x1.45f306dc9c889p-2, // x^1
+ 0x1.b2995e7b7b5fdp-5, // x^3
+ 0x1.8723a1d588a36p-6, // x^5
+ 0x1.d1a452f20430dp-7, // x^7
+ 0x1.3ce52a3a09f61p-7, // x^9
+ 0x1.d2b33e303d375p-8, // x^11
+ 0x1.69fde663c674fp-8, // x^13
+ 0x1.235134885f19bp-8, // x^15
+ };
+ // polynomial evaluation using horner's method
+ // work only for |x| in [0, 0.5]
+ auto asinpi_polyeval = [&POLY_COEFFS](double x) -> double {
+ return x * fputil::polyeval(x * x, POLY_COEFFS[0], POLY_COEFFS[1],
+ POLY_COEFFS[2], POLY_COEFFS[3], POLY_COEFFS[4],
+ POLY_COEFFS[5], POLY_COEFFS[6], POLY_COEFFS[7]);
+ };
+
+ // if |x| <= 0.5:
+ if (LIBC_UNLIKELY(x_abs <= 0.5)) {
+ // Use polynomial approximation of asin(x)/pi in the range [0, 0.5]
+ double result = asinpi_polyeval(fputil::cast<double>(x));
+ return fputil::cast<float16>(result);
+ }
+
+ // If |x| > 0.5, we need to use the range reduction method:
+ // y = asin(x) => x = sin(y)
+ // because: sin(a) = cos(pi/2 - a)
+ // therefore:
+ // x = cos(pi/2 - y)
+ // let z = pi/2 - y,
+ // x = cos(z)
+ // because: cos(2a) = 1 - 2 * sin^2(a), z = 2a, a = z/2
+ // therefore:
+ // cos(z) = 1 - 2 * sin^2(z/2)
+ // sin(z/2) = sqrt((1 - cos(z))/2)
+ // sin(z/2) = sqrt((1 - x)/2)
+ // let u = (1 - x)/2
+ // then:
+ // sin(z/2) = sqrt(u)
+ // z/2 = asin(sqrt(u))
+ // z = 2 * asin(sqrt(u))
+ // pi/2 - y = 2 * asin(sqrt(u))
+ // y = pi/2 - 2 * asin(sqrt(u))
+ // y/pi = 1/2 - 2 * asin(sqrt(u))/pi
+ //
+ // Finally, we can write:
+ // asinpi(x) = 1/2 - 2 * asinpi(sqrt(u))
+ // where u = (1 - x) /2
+ // = 0.5 - 0.5 * x
+ // = multiply_add(-0.5, x, 0.5)
+
+ double u = fputil::multiply_add(-0.5, x_abs, 0.5);
+ double asinpi_sqrt_u = asinpi_polyeval(fputil::sqrt<double>(u));
+ double result = fputil::multiply_add(-2.0, asinpi_sqrt_u, 0.5);
+
+ return fputil::cast<float16>(signed_result(result));
+}
+
+} // namespace math
+
+} // namespace LIBC_NAMESPACE_DECL
+
+#endif // LIBC_TYPES_HAS_FLOAT16
+
+#endif // LLVM_LIBC_SRC___SUPPORT_MATH_ASINPIF16_H
diff --git a/libc/src/math/generic/CMakeLists.txt b/libc/src/math/generic/CMakeLists.txt
index 2475aa52c3c0e..9c1d20f9bb6ad 100644
--- a/libc/src/math/generic/CMakeLists.txt
+++ b/libc/src/math/generic/CMakeLists.txt
@@ -4213,16 +4213,7 @@ add_entrypoint_object(
HDRS
../asinpif16.h
DEPENDS
- libc.hdr.errno_macros
- libc.hdr.fenv_macros
- libc.src.__support.FPUtil.cast
- libc.src.__support.FPUtil.except_value_utils
- libc.src.__support.FPUtil.fenv_impl
- libc.src.__support.FPUtil.fp_bits
- libc.src.__support.FPUtil.multiply_add
- libc.src.__support.FPUtil.polyeval
- libc.src.__support.FPUtil.sqrt
- libc.src.__support.macros.optimization
+ libc.src.__support.math.asinpif16
)
add_entrypoint_object(
diff --git a/libc/src/math/generic/asinpif16.cpp b/libc/src/math/generic/asinpif16.cpp
index 395f4c4ebc070..bb765075f64e7 100644
--- a/libc/src/math/generic/asinpif16.cpp
+++ b/libc/src/math/generic/asinpif16.cpp
@@ -7,121 +7,12 @@
//===----------------------------------------------------------------------===//
#include "src/math/asinpif16.h"
-#include "hdr/errno_macros.h"
-#include "hdr/fenv_macros.h"
-#include "src/__support/FPUtil/FEnvImpl.h"
-#include "src/__support/FPUtil/FPBits.h"
-#include "src/__support/FPUtil/PolyEval.h"
-#include "src/__support/FPUtil/cast.h"
-#include "src/__support/FPUtil/except_value_utils.h"
-#include "src/__support/FPUtil/multiply_add.h"
-#include "src/__support/FPUtil/sqrt.h"
-#include "src/__support/macros/optimization.h"
+#include "src/__support/math/asinpif16.h"
namespace LIBC_NAMESPACE_DECL {
LLVM_LIBC_FUNCTION(float16, asinpif16, (float16 x)) {
- using FPBits = fputil::FPBits<float16>;
-
- FPBits xbits(x);
- bool is_neg = xbits.is_neg();
- double x_abs = fputil::cast<double>(xbits.abs().get_val());
-
- auto signed_result = [is_neg](auto r) -> auto { return is_neg ? -r : r; };
-
- if (LIBC_UNLIKELY(x_abs > 1.0)) {
- // aspinf16(NaN) = NaN
- if (xbits.is_nan()) {
- if (xbits.is_signaling_nan()) {
- fputil::raise_except_if_required(FE_INVALID);
- return FPBits::quiet_nan().get_val();
- }
- return x;
- }
-
- // 1 < |x| <= +/-inf
- fputil::raise_except_if_required(FE_INVALID);
- fputil::set_errno_if_required(EDOM);
-
- return FPBits::quiet_nan().get_val();
- }
-
- // the coefficients for the polynomial approximation of asin(x)/pi in the
- // range [0, 0.5] extracted using python-sympy
- //
- // Python code to generate the coefficients:
- // > from sympy import *
- // > import math
- // > x = symbols('x')
- // > print(series(asin(x)/math.pi, x, 0, 21))
- //
- // OUTPUT:
- //
- // 0.318309886183791*x + 0.0530516476972984*x**3 + 0.0238732414637843*x**5 +
- // 0.0142102627760621*x**7 + 0.00967087327815336*x**9 +
- // 0.00712127941391293*x**11 + 0.00552355646848375*x**13 +
- // 0.00444514782463692*x**15 + 0.00367705242846804*x**17 +
- // 0.00310721681820837*x**19 + O(x**21)
- //
- // it's very accurate in the range [0, 0.5] and has a maximum error of
- // 0.0000000000000001 in the range [0, 0.5].
- constexpr double POLY_COEFFS[] = {
- 0x1.45f306dc9c889p-2, // x^1
- 0x1.b2995e7b7b5fdp-5, // x^3
- 0x1.8723a1d588a36p-6, // x^5
- 0x1.d1a452f20430dp-7, // x^7
- 0x1.3ce52a3a09f61p-7, // x^9
- 0x1.d2b33e303d375p-8, // x^11
- 0x1.69fde663c674fp-8, // x^13
- 0x1.235134885f19bp-8, // x^15
- };
- // polynomial evaluation using horner's method
- // work only for |x| in [0, 0.5]
- auto asinpi_polyeval = [&](double x) -> double {
- return x * fputil::polyeval(x * x, POLY_COEFFS[0], POLY_COEFFS[1],
- POLY_COEFFS[2], POLY_COEFFS[3], POLY_COEFFS[4],
- POLY_COEFFS[5], POLY_COEFFS[6], POLY_COEFFS[7]);
- };
-
- // if |x| <= 0.5:
- if (LIBC_UNLIKELY(x_abs <= 0.5)) {
- // Use polynomial approximation of asin(x)/pi in the range [0, 0.5]
- double result = asinpi_polyeval(fputil::cast<double>(x));
- return fputil::cast<float16>(result);
- }
-
- // If |x| > 0.5, we need to use the range reduction method:
- // y = asin(x) => x = sin(y)
- // because: sin(a) = cos(pi/2 - a)
- // therefore:
- // x = cos(pi/2 - y)
- // let z = pi/2 - y,
- // x = cos(z)
- // because: cos(2a) = 1 - 2 * sin^2(a), z = 2a, a = z/2
- // therefore:
- // cos(z) = 1 - 2 * sin^2(z/2)
- // sin(z/2) = sqrt((1 - cos(z))/2)
- // sin(z/2) = sqrt((1 - x)/2)
- // let u = (1 - x)/2
- // then:
- // sin(z/2) = sqrt(u)
- // z/2 = asin(sqrt(u))
- // z = 2 * asin(sqrt(u))
- // pi/2 - y = 2 * asin(sqrt(u))
- // y = pi/2 - 2 * asin(sqrt(u))
- // y/pi = 1/2 - 2 * asin(sqrt(u))/pi
- //
- // Finally, we can write:
- // asinpi(x) = 1/2 - 2 * asinpi(sqrt(u))
- // where u = (1 - x) /2
- // = 0.5 - 0.5 * x
- // = multiply_add(-0.5, x, 0.5)
-
- double u = fputil::multiply_add(-0.5, x_abs, 0.5);
- double asinpi_sqrt_u = asinpi_polyeval(fputil::sqrt<double>(u));
- double result = fputil::multiply_add(-2.0, asinpi_sqrt_u, 0.5);
-
- return fputil::cast<float16>(signed_result(result));
+ return math::asinpif16(x);
}
} // namespace LIBC_NAMESPACE_DECL
diff --git a/libc/test/shared/CMakeLists.txt b/libc/test/shared/CMakeLists.txt
index fb9874ab3ec03..9b78cc2e2bb64 100644
--- a/libc/test/shared/CMakeLists.txt
+++ b/libc/test/shared/CMakeLists.txt
@@ -18,6 +18,7 @@ add_fp_unittest(
libc.src.__support.math.asinf16
libc.src.__support.math.asinhf
libc.src.__support.math.asinhf16
+ libc.src.__support.math.asinpif16
libc.src.__support.math.atan
libc.src.__support.math.atan2
libc.src.__support.math.atan2f
diff --git a/libc/test/shared/shared_math_test.cpp b/libc/test/shared/shared_math_test.cpp
index 92511cc55267e..77ddf2d5d197d 100644
--- a/libc/test/shared/shared_math_test.cpp
+++ b/libc/test/shared/shared_math_test.cpp
@@ -22,6 +22,7 @@ TEST(LlvmLibcSharedMathTest, AllFloat16) {
EXPECT_FP_EQ(0x0p+0f16, LIBC_NAMESPACE::shared::asinf16(0.0f16));
EXPECT_FP_EQ(0x0p+0f16, LIBC_NAMESPACE::shared::asinhf16(0.0f16));
+ EXPECT_FP_EQ(0x0p+0f16, LIBC_NAMESPACE::shared::asinpif16(0.0f16));
EXPECT_FP_EQ(0x0p+0f16, LIBC_NAMESPACE::shared::atanf16(0.0f16));
EXPECT_FP_EQ(0x0p+0f16, LIBC_NAMESPACE::shared::atanhf16(0.0f16));
EXPECT_FP_EQ(0x1p+0f16, LIBC_NAMESPACE::shared::cosf16(0.0f16));
diff --git a/utils/bazel/llvm-project-overlay/libc/BUILD.bazel b/utils/bazel/llvm-project-overlay/libc/BUILD.bazel
index 87b2a392287a6..aa0707f688d57 100644
--- a/utils/bazel/llvm-project-overlay/libc/BUILD.bazel
+++ b/utils/bazel/llvm-project-overlay/libc/BUILD.bazel
@@ -3634,6 +3634,21 @@ libc_support_library(
],
)
+libc_support_library(
+ name = "__support_math_asinpif16",
+ hdrs = ["src/__support/math/asinpif16.h"],
+ deps = [
+ ":__support_fputil_cast",
+ ":__support_fputil_fenv_impl",
+ ":__support_fputil_fp_bits",
+ ":__support_fputil_multiply_add",
+ ":__support_fputil_polyeval",
+ ":__support_fputil_sqrt",
+ ":__support_macros_optimization",
+ ":__support_macros_properties_types",
+ ],
+)
+
libc_support_library(
name = "__support_math_tan",
hdrs = ["src/__support/math/tan.h"],
@@ -5292,6 +5307,14 @@ libc_math_function(
],
)
+libc_math_function(
+ name = "asinpif16",
+ additional_deps = [
+ ":__support_math_asinpif16",
+ ":errno",
+ ]
+)
+
libc_math_function(
name = "sqrt",
additional_deps = [
>From 17a8c9d7ebda15077e97153906f26f0b6a87ae91 Mon Sep 17 00:00:00 2001
From: raventid <juliankul at gmail.com>
Date: Sat, 31 Jan 2026 15:25:23 +0800
Subject: [PATCH 2/3] remove unused dependency from Bazel build
---
utils/bazel/llvm-project-overlay/libc/BUILD.bazel | 1 -
1 file changed, 1 deletion(-)
diff --git a/utils/bazel/llvm-project-overlay/libc/BUILD.bazel b/utils/bazel/llvm-project-overlay/libc/BUILD.bazel
index aa0707f688d57..585977990301d 100644
--- a/utils/bazel/llvm-project-overlay/libc/BUILD.bazel
+++ b/utils/bazel/llvm-project-overlay/libc/BUILD.bazel
@@ -3645,7 +3645,6 @@ libc_support_library(
":__support_fputil_polyeval",
":__support_fputil_sqrt",
":__support_macros_optimization",
- ":__support_macros_properties_types",
],
)
>From 4402ca797d6063b800acb598c5a5da63ff4dafb5 Mon Sep 17 00:00:00 2001
From: raventid <juliankul at gmail.com>
Date: Sat, 31 Jan 2026 15:26:20 +0800
Subject: [PATCH 3/3] do not capture POLY_COEFFS lambda
---
libc/src/__support/math/asinpif16.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/libc/src/__support/math/asinpif16.h b/libc/src/__support/math/asinpif16.h
index 248704d2eb738..8e1d6afac2c29 100644
--- a/libc/src/__support/math/asinpif16.h
+++ b/libc/src/__support/math/asinpif16.h
@@ -82,7 +82,7 @@ LIBC_INLINE static constexpr float16 asinpif16(float16 x) {
};
// polynomial evaluation using horner's method
// work only for |x| in [0, 0.5]
- auto asinpi_polyeval = [&POLY_COEFFS](double x) -> double {
+ auto asinpi_polyeval = [](double x) -> double {
return x * fputil::polyeval(x * x, POLY_COEFFS[0], POLY_COEFFS[1],
POLY_COEFFS[2], POLY_COEFFS[3], POLY_COEFFS[4],
POLY_COEFFS[5], POLY_COEFFS[6], POLY_COEFFS[7]);
More information about the libc-commits
mailing list