[libc-commits] [libc] [llvm] [libc][math][c23] implement `asinpif` function (PR #181511)

via libc-commits libc-commits at lists.llvm.org
Wed Feb 18 10:00:29 PST 2026


================
@@ -0,0 +1,203 @@
+//===-- Implementation header for asinpif -----------------------*- 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_ASINPIF_H
+#define LLVM_LIBC_SRC___SUPPORT_MATH_ASINPIF_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/config.h"
+#include "src/__support/macros/optimization.h"
+#include "src/__support/macros/properties/types.h"
+
+#include "hdr/errno_macros.h"
+#include "hdr/fenv_macros.h"
+#include "src/__support/FPUtil/FEnvImpl.h"
+#include "src/__support/FPUtil/except_value_utils.h"
+
+namespace LIBC_NAMESPACE_DECL {
+namespace math {
+
+LIBC_INLINE constexpr float asinpif(float x) {
+#ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
+  constexpr size_t N_EXCEPTS = 5;
+  constexpr fputil::ExceptValues<float, N_EXCEPTS> ASINPIF_EXCEPTS = {
+      {// (inputs, RZ output, RU offset, RD offset, RN offset)
+       // x = 0x1.e768f6p-122, asinpif(x) = 0x1.364b7ap-123 (RZ)
+       {0x02F3B47B, 0x021B25BD, 1, 0, 0},
+       // x = 0x1.e768f6p-24, asinpif(x) = 0x1.364b7ap-25 (RZ)
+       {0x33F3B47B, 0x331B25BD, 1, 0, 1},
+       // x = 0x1.dddb4ep-19, asinpif(x) = 0x1.303686p-20 (RZ)
+       {0x366EEDA7, 0x35981B43, 1, 0, 1},
+       // x = -0x1.dddb4ep-19, asinpif(x) = -0x1.303686p-20 (RZ)
+       {0xB66EEDA7, 0xB5981B43, 0, 1, 1},
+       // x = -0x1.e768f6p-24, asinpif(x) = -0x1.364b7ap-25 (RZ)
+       {0xB3F3B47B, 0xB31B25BD, 0, 1, 1}}};
+#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
+
+  using FPBits = fputil::FPBits<float>;
+
+  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)) {
+    if (xbits.is_nan()) {
+      if (xbits.is_signaling_nan()) {
+        fputil::raise_except_if_required(FE_INVALID);
+        return FPBits::quiet_nan().get_val();
+      }
+      return x;
+    }
+
+    fputil::raise_except_if_required(FE_INVALID);
+    fputil::set_errno_if_required(EDOM);
+    return FPBits::quiet_nan().get_val();
+  }
+
+#ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
+  auto r = ASINPIF_EXCEPTS.lookup(xbits.uintval());
+  if (LIBC_UNLIKELY(r.has_value()))
+    return r.value();
+#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
+
+  // the coefficients for the polynomial approximation of asin(x)/(pi*x) in the
+  // range [0, 0.5] extracted using Sollya.
+  //
+  // Sollya code:
+  // > prec = 200;
+  // > display = hexadecimal;
+  // > g = asin(x) / (pi * x);
+  // > P = fpminimax(g, [|0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20|],
+  // >              [|D...|], [0, 0.5]);
+  // > for i from 0 to degree(P) do coeff(P, i);
+  // > print("Error:", dirtyinfnorm(P - g, [1e-30; 0.25]));
+  // Error: 0x1.45c281e1cf9b58p-50 ~= 2^−49.652
+  //
+  // Non-zero coefficients (even powers only):
+  constexpr double ASINPI_POLY_COEFFS[] = {
+      0x1.45f306dc9c881p-2,  // x^0
+      0x1.b2995e7b7e756p-5,  // x^2
+      0x1.8723a1d12f828p-6,  // x^4
+      0x1.d1a45564b9545p-7,  // x^6
+      0x1.3ce4ceaa0e1e9p-7,  // x^8
+      0x1.d2c305898ea13p-8,  // x^10
+      0x1.692212e27a5f9p-8,  // x^12
+      0x1.2b22cc744d25bp-8,  // x^14
+      0x1.8427b864479ffp-9,  // x^16
+      0x1.815522d7a2bf1p-8,  // x^18
+      -0x1.f6df98438aef4p-9, // x^20
+      0x1.4b50c2eb13708p-7   // x^22
+  };
+  // polynomial evaluation using horner's method
+  // work only for |x| in [0, 0.5]
+  // Returns v * P(v2) where P(v2) = c0 + c1*v2 + c2*v2^2 + ...
+  auto asinpi_polyeval = [&](double v, double v2) -> double {
+    return v * fputil::polyeval(v2, ASINPI_POLY_COEFFS[0],
+                                ASINPI_POLY_COEFFS[1], ASINPI_POLY_COEFFS[2],
+                                ASINPI_POLY_COEFFS[3], ASINPI_POLY_COEFFS[4],
+                                ASINPI_POLY_COEFFS[5], ASINPI_POLY_COEFFS[6],
+                                ASINPI_POLY_COEFFS[7], ASINPI_POLY_COEFFS[8],
+                                ASINPI_POLY_COEFFS[9], ASINPI_POLY_COEFFS[10],
+                                ASINPI_POLY_COEFFS[11]);
+  };
+
+  // Returns P(v2) - c0 = c1*v2 + c2*v2^2 + ...
+  // This is the "tail" of the polynomial, used to avoid cancellation
+  // in the range reduction path.
+  auto asinpi_polyeval_tail = [&](double v2) -> double {
+    return v2 * fputil::polyeval(v2, ASINPI_POLY_COEFFS[1],
+                                 ASINPI_POLY_COEFFS[2], ASINPI_POLY_COEFFS[3],
+                                 ASINPI_POLY_COEFFS[4], ASINPI_POLY_COEFFS[5],
+                                 ASINPI_POLY_COEFFS[6], ASINPI_POLY_COEFFS[7],
+                                 ASINPI_POLY_COEFFS[8], ASINPI_POLY_COEFFS[9],
+                                 ASINPI_POLY_COEFFS[10],
+                                 ASINPI_POLY_COEFFS[11]);
+  };
+
+  // if |x| <= 0.5:
+  if (LIBC_UNLIKELY(x_abs <= 0.5)) {
+    double x_d = fputil::cast<double>(x);
+    double result = asinpi_polyeval(x_d, x_d * x_d);
+    return fputil::cast<float>(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)
+  //
+  // asinpi(x) = 0.5 - 2 * sqrt(u) * P(u)
+  //
+  // To avoid cancellation when |x| is near 0.5 (where 2*sqrt(u)*P(u) ~ 0.5),
+  // we split P(u) into its leading term c0 and the tail:
+  //   P(u) = c0 + tail(u)
+  //
+  // We further split the constant 1/pi into high and low parts for precision:
+  //   1/pi = ONE_OVER_PI_HI + ONE_OVER_PI_LO
+  //
+  // And rewrite the expression as:
+  //   0.5 - 2*sqrt(u) * (1/pi + (c0 - 1/pi) + tail(u))
+  //
+  // The term (0.5 - 2*sqrt(u)*ONE_OVER_PI_HI) is computed exactly using FMA.
+  // The remaining small terms are added separately:
+  //   - 2*sqrt(u) * ONE_OVER_PI_LO
+  //   - 2*sqrt(u) * (c0 - 1/pi)      [absorbed into tail sum]
+  //   - 2*sqrt(u) * tail(u)
+
+  constexpr double ONE_OVER_PI_HI = 0x1.45f306dc9c883p-2;
+  constexpr double ONE_OVER_PI_LO = -0x1.6b01ec5417056p-56;
+  // Verify: ONE_OVER_PI_HI + ONE_OVER_PI_LO ≈ 1/pi to ~106 bits
+
+  // DELTA_C0 = c0 - ONE_OVER_PI_HI (difference between Sollya c0 and 1/pi hi)
+  constexpr double DELTA_C0 = ASINPI_POLY_COEFFS[0] - ONE_OVER_PI_HI;
+
+  double u = fputil::multiply_add(-0.5, x_abs, 0.5);
+  double sqrt_u = fputil::sqrt<double>(u);
+
+  // compute the tail: P(u) - c0
+  double tail = asinpi_polyeval_tail(u);
+
+  double neg2_sqrt_u = -2.0 * sqrt_u;
+  double result_hi = fputil::multiply_add(neg2_sqrt_u, ONE_OVER_PI_HI, 0.5);
+  double result = result_hi + neg2_sqrt_u * (ONE_OVER_PI_LO + DELTA_C0 + tail);
----------------
lntue wrote:

use `multiply_add` for the outer part

https://github.com/llvm/llvm-project/pull/181511


More information about the libc-commits mailing list