[libc-commits] [libc] [libc][math][c23] Implement C23 math function atanpif16 (PR #150400)

[via libc-commits]

================
@@ -0,0 +1,155 @@
+//===-- Half-precision atanpi function ------------------------------------===//
+//
+// 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
+//
+//===----------------------------------------------------------------------===//
+
+#include "src/math/atanpif16.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/multiply_add.h"
+#include "src/__support/FPUtil/sqrt.h"
+
+namespace LIBC_NAMESPACE_DECL {
+
+// Using Python's SymPy library, we can obtain the polynomial approximation of
+// arctan(x)/pi. The steps are as follows:
+//  >>> from sympy import *
+//  >>> import math
+//  >>> x = symbols('x')
+//  >>> print(series(atan(x)/math.pi, x, 0, 17))
+//
+// Output:
+// 0.318309886183791*x - 0.106103295394597*x**3 + 0.0636619772367581*x**5 -
+// 0.0454728408833987*x**7 + 0.0353677651315323*x**9 - 0.0289372623803446*x**11
+// + 0.0244853758602916*x**13 - 0.0212206590789194*x**15 + O(x**17)
+//
+// We will assign this 19-degree Taylor polynomial as g(x). This polynomial
+// approximation is accurate for arctan(x)/pi when |x| is in the range [0, 0.5].
+//
+//
+// To compute arctan(x) for all real x, we divide the domain into the following
+// cases:
+//
+// * Case 1: |x| <= 0.5
+//      In this range, the direct polynomial approximation is used:
+//      arctan(x)/pi = sign(x) * g(|x|)
+//      or equivalently, arctan(x) = sign(x) * pi * g(|x|).
+//
+// * Case 2: 0.5 < |x| <= 1
+//      We use the double-angle identity for the tangent function, specifically:
+//        arctan(x) = 2 * arctan(x / (1 + sqrt(1 + x^2))).
+//      Applying this, we have:
+//        arctan(x)/pi = sign(x) * 2 * arctan(x')/pi,
+//        where x' = |x| / (1 + sqrt(1 + x^2)).
+//        Thus, arctan(x)/pi = sign(x) * 2 * g(x')
+//
+//      When |x| is in (0.5, 1], the value of x' will always fall within the
+//      interval [0.207, 0.414], which is within the accurate range of g(x).
+//
+// * Case 3: |x| > 1
+//      For values of |x| greater than 1, we use the reciprocal transformation
+//      identity:
+//        arctan(x) = pi/2 - arctan(1/x) for x > 0.
+//      For any x (real number), this generalizes to:
+//        arctan(x)/pi = sign(x) * (1/2 - arctan(1/|x|)/pi).
+//      Then, using g(x) for arctan(1/|x|)/pi:
+//        arctan(x)/pi = sign(x) * (1/2 - g(1/|x|)).
+//
+//      Note that if 1/|x| still falls outside the
+//      g(x)'s primary range of accuracy (i.e., if 0.5 < 1/|x| <= 1), the rule
+//      from Case 2 must be applied recursively to 1/|x|.
+
+LLVM_LIBC_FUNCTION(float16, atanpif16, (float16 x)) {
+  using FPBits = fputil::FPBits<float16>;
+
+  FPBits xbits(x);
+  bool is_neg = xbits.is_neg();
+
+  auto signed_result = [is_neg](double r) -> float16 {
+    return fputil::cast<float16>(is_neg ? -r : r);
+  };
+
+  if (LIBC_UNLIKELY(xbits.is_inf_or_nan())) {
----------------
overmighty wrote:

Should `#include "src/__support/macros/optimization.h"` for `LIBC_UNLIKELY`.

https://github.com/llvm/llvm-project/pull/150400