[libc-commits] [libc] [libc][math][c23] Add exp2m1f C23 math function (PR #86996)

[via libc-commits]

================
@@ -0,0 +1,144 @@
+//===-- Implementation of exp2m1f 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/exp2m1f.h"
+#include "src/__support/FPUtil/FEnvImpl.h"
+#include "src/__support/FPUtil/FPBits.h"
+#include "src/__support/FPUtil/PolyEval.h"
+#include "src/__support/FPUtil/multiply_add.h"
+#include "src/__support/FPUtil/rounding_mode.h"
+#include "src/__support/common.h"
+#include "src/__support/macros/optimization.h"
+#include "src/__support/macros/properties/cpu_features.h"
+#include "src/errno/libc_errno.h"
+
+#include "explogxf.h"
+
+namespace LIBC_NAMESPACE {
+
+LLVM_LIBC_FUNCTION(float, exp2m1f, (float x)) {
+  using FPBits = fputil::FPBits<float>;
+  FPBits xbits(x);
+
+  uint32_t x_u = xbits.uintval();
+  uint32_t x_abs = x_u & 0x7fff'ffffU;
+
+  // When |x| >= 128, or x is nan, or |x| <= 2^-5
+  if (LIBC_UNLIKELY(x_abs >= 0x4300'0000U || x_abs <= 0x3d00'0000U)) {
+    // |x| <= 2^-5
+    if (x_abs <= 0x3d00'0000) {
+      // TODO: Handle exceptional values.
+
+      // Minimax polynomial generated by Sollya with:
+      // > display = hexadecimal;
+      // > fpminimax((2^x - 1)/x, 5, [|D...|], [-2^-5, 2^-5]);
+      constexpr double COEFFS[] = {
+          0x1.62e42fefa39f3p-1, 0x1.ebfbdff82c57bp-3,  0x1.c6b08d6f2d7aap-5,
+          0x1.3b2ab6fc92f5dp-7, 0x1.5d897cfe27125p-10, 0x1.43090e61e6af1p-13};
+      double xd = x;
+      double xsq = xd * xd;
+      double c0 = fputil::multiply_add(xd, COEFFS[1], COEFFS[0]);
+      double c1 = fputil::multiply_add(xd, COEFFS[3], COEFFS[2]);
+      double c2 = fputil::multiply_add(xd, COEFFS[5], COEFFS[4]);
+      double p = fputil::polyeval(xsq, c0, c1, c2);
+      return static_cast<float>(p * xd);
+    }
+
+    // x >= 128, or x is nan
+    if (xbits.is_pos()) {
+      if (xbits.is_finite()) {
+        int rounding = fputil::quick_get_round();
+        if (rounding == FE_DOWNWARD || rounding == FE_TOWARDZERO)
+          return FPBits::max_normal().get_val();
+
+        fputil::set_errno_if_required(ERANGE);
+        fputil::raise_except_if_required(FE_OVERFLOW);
+      }
+
+      // x >= 128 and 2^x - 1 rounds to +inf, or x is +inf or nan
+      return x + FPBits::inf().get_val();
+    }
+
+    // x <= -25
+    if (x <= -25.0f) {
+      // 2^(-inf) - 1 = -1
+      if (xbits.is_inf())
+        return -1.0f;
+      // 2^nan - 1 = nan
+      if (xbits.is_nan())
+        return x;
+
+      int rounding = fputil::quick_get_round();
+      if (rounding == FE_UPWARD || rounding == FE_TOWARDZERO)
+        return -0x1.ffff'fep-1f; // -1.0f + 0x1.0p-24f
+
+      fputil::set_errno_if_required(ERANGE);
+      fputil::raise_except_if_required(FE_UNDERFLOW);
+      return -1.0f;
+    }
+  }
+
+  // For -25 < x < 128, to compute 2^x, we perform the following range
+  // reduction: find hi, mid, lo such that:
+  //   x = hi + mid + lo, in which:
+  //     hi is an integer,
+  //     0 <= mid * 2^5 < 32 is an integer,
+  //     -2^(-6) <= lo <= 2^(-6).
+  // In particular,
+  //   hi + mid = round(x * 2^5) * 2^(-5).
+  // Then,
+  //   2^x = 2^(hi + mid + lo) = 2^hi * 2^mid * 2^lo.
+  // 2^mid is stored in the lookup table of 32 elements.
+  // 2^lo is computed using a degree-4 minimax polynomial generated by Sollya.
+  // We perform 2^hi * 2^mid by simply add hi to the exponent field of 2^mid.
+
+  // kf = (hi + mid) * 2^5 = round(x * 2^5)
+  float kf;
+  int k;
+#ifdef LIBC_TARGET_CPU_HAS_NEAREST_INT
+  kf = fputil::nearest_integer(x * 32.0f);
+  k = static_cast<int>(kf);
+#else
+  constexpr float HALF[2] = {0.5f, -0.5f};
+  k = static_cast<int>(fputil::multiply_add(x, 32.0f, HALF[x < 0.0f]));
+  kf = static_cast<float>(k);
+#endif // LIBC_TARGET_CPU_HAS_NEAREST_INT
+
+  // dx = lo = x - (hi + mid) = x - kf * 2^(-5)
+  double dx = fputil::multiply_add(-0x1.0p-5f, kf, x);
+
+  // hi = floor(kf * 2^(-4))
+  // exp_hi = shift hi to the exponent field of double precision.
+  int64_t exp_hi =
+      static_cast<int64_t>(static_cast<uint64_t>(k >> ExpBase::MID_BITS)
+                           << fputil::FPBits<double>::FRACTION_LEN);
+  // mh = 2^hi * 2^mid
+  // mh_bits = bit field of mh
+  int64_t mh_bits = ExpBase::EXP_2_MID[k & ExpBase::MID_MASK] + exp_hi;
+  double mh = fputil::FPBits<double>(static_cast<uint64_t>(mh_bits)).get_val();
+
+  // Degree-4 polynomial approximating (2^x - 1)/x generated by Sollya with:
+  // > display = hexadecimal;
+  // > fpminimax((2^x - 1)/x, 4, [|D...|], [-2^-6, 2^-6]);
+  constexpr double COEFFS[5] = {0x1.62e42fefa39efp-1, 0x1.ebfbdff8131c4p-3,
+                                0x1.c6b08d7061695p-5, 0x1.3b2b1bee74b2ap-7,
+                                0x1.5d88091198529p-10};
+  double dx_sq = dx * dx;
+  double c1 = fputil::multiply_add(dx, COEFFS[0], 1.0);
+  double c2 = fputil::multiply_add(dx, COEFFS[2], COEFFS[1]);
+  double c3 = fputil::multiply_add(dx, COEFFS[4], COEFFS[3]);
+  double p = fputil::multiply_add(dx_sq, c3, c2);
+  // 2^x = 2^(hi + mid + lo)
+  //     = 2^(hi + mid) * 2^lo
+  //     ~ mh * (1 + lo * P(lo))
+  //     = mh + (mh*lo) * P(lo)
+  double exp2_lo = fputil::multiply_add(p, dx_sq, c1);
+  return static_cast<float>(fputil::multiply_add(exp2_lo, mh, -1.0));
----------------
overmighty wrote:

Oops. :(

https://github.com/llvm/llvm-project/pull/86996