[libc][math] Refactor coshf16 implementation to header-only in src/__support/math folder.

libc/shared/math.h

@@ -36,6 +36,7 @@
 #include "math/cosf.h"
 #include "math/cosf16.h"
 #include "math/coshf.h"
+#include "math/coshf16.h"
 #include "math/erff.h"
 #include "math/exp.h"
 #include "math/exp10.h"

libc/shared/math/coshf16.h

@@ -0,0 +1,28 @@
+//===-- Shared coshf16 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_COSHF16_H
+#define LLVM_LIBC_SHARED_MATH_COSHF16_H
+
+#include "shared/libc_common.h"
+
+#ifdef LIBC_TYPES_HAS_FLOAT16
+
+#include "src/__support/math/coshf16.h"
+
+namespace LIBC_NAMESPACE_DECL {
+namespace shared {
+
+using math::coshf16;
+
+} // namespace shared
+} // namespace LIBC_NAMESPACE_DECL
+
+#endif // LIBC_TYPES_HAS_FLOAT16
+
+#endif // LLVM_LIBC_SHARED_MATH_COSHF16_H

libc/src/__support/math/CMakeLists.txt

@@ -419,6 +419,19 @@ add_header_library(
     libc.src.__support.macros.optimization
 )
 
+add_header_library(
+  coshf16
+  HDRS
+    coshf16.h
+  DEPENDS
+    .expxf16_utils
+    libc.src.__support.FPUtil.except_value_utils
+    libc.src.__support.FPUtil.fenv_impl
+    libc.src.__support.FPUtil.fp_bits
+    libc.src.__support.FPUtil.rounding_mode
+    libc.src.__support.macros.optimization
+)
+
 add_header_library(
   erff
   HDRS
@@ -489,6 +502,21 @@ add_header_library(
     libc.include.llvm-libc-macros.float16_macros
 )
 
+add_header_library(
+  expxf16_utils
+  HDRS
+    expxf16_utils.h
+  DEPENDS
+    libc.hdr.stdint_proxy
+    libc.src.__support.FPUtil.fp_bits
+    libc.src.__support.FPUtil.cast
+    libc.src.__support.FPUtil.multiply_add
+    libc.src.__support.FPUtil.nearest_integer
+    libc.src.__support.macros.attributes
+    libc.src.__support.math.expf16_utils
+    libc.src.__support.math.exp10_float16_constants
+)
+
 add_header_library(
   frexpf128
   HDRS

libc/src/__support/math/coshf16.h

@@ -0,0 +1,124 @@
+//===-- Implementation header for coshf16 -----------------------*- 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_COSHF16_H
+#define LLVM_LIBC_SRC___SUPPORT_MATH_COSHF16_H
+
+#include "include/llvm-libc-macros/float16-macros.h"
+
+#ifdef LIBC_TYPES_HAS_FLOAT16
+
+#include "expxf16_utils.h"
+#include "src/__support/FPUtil/FEnvImpl.h"
+#include "src/__support/FPUtil/FPBits.h"
+#include "src/__support/FPUtil/except_value_utils.h"
+#include "src/__support/FPUtil/rounding_mode.h"
+#include "src/__support/macros/config.h"
+#include "src/__support/macros/optimization.h"
+
+namespace LIBC_NAMESPACE_DECL {
+
+namespace math {
+
+LIBC_INLINE static constexpr float16 coshf16(float16 x) {
+
+#ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
+  constexpr fputil::ExceptValues<float16, 9> COSHF16_EXCEPTS_POS = {{
+      // x = 0x1.6ap-5, coshf16(x) = 0x1p+0 (RZ)
+      {0x29a8U, 0x3c00U, 1U, 0U, 1U},
+      // x = 0x1.8c4p+0, coshf16(x) = 0x1.3a8p+1 (RZ)
+      {0x3e31U, 0x40eaU, 1U, 0U, 0U},
+      // x = 0x1.994p+0, coshf16(x) = 0x1.498p+1 (RZ)
+      {0x3e65U, 0x4126U, 1U, 0U, 0U},
+      // x = 0x1.b6p+0, coshf16(x) = 0x1.6d8p+1 (RZ)
+      {0x3ed8U, 0x41b6U, 1U, 0U, 1U},
+      // x = 0x1.aap+1, coshf16(x) = 0x1.be8p+3 (RZ)
+      {0x42a8U, 0x4afaU, 1U, 0U, 1U},
+      // x = 0x1.cc4p+1, coshf16(x) = 0x1.23cp+4 (RZ)
+      {0x4331U, 0x4c8fU, 1U, 0U, 0U},
+      // x = 0x1.288p+2, coshf16(x) = 0x1.9b4p+5 (RZ)
+      {0x44a2U, 0x526dU, 1U, 0U, 0U},
+      // x = 0x1.958p+2, coshf16(x) = 0x1.1a4p+8 (RZ)
+      {0x4656U, 0x5c69U, 1U, 0U, 0U},
+      // x = 0x1.5fp+3, coshf16(x) = 0x1.c54p+14 (RZ)
+      {0x497cU, 0x7715U, 1U, 0U, 1U},
+  }};
+
+  constexpr fputil::ExceptValues<float16, 6> COSHF16_EXCEPTS_NEG = {{
+      // x = -0x1.6ap-5, coshf16(x) = 0x1p+0 (RZ)
+      {0xa9a8U, 0x3c00U, 1U, 0U, 1U},
+      // x = -0x1.b6p+0, coshf16(x) = 0x1.6d8p+1 (RZ)
+      {0xbed8U, 0x41b6U, 1U, 0U, 1U},
+      // x = -0x1.288p+2, coshf16(x) = 0x1.9b4p+5 (RZ)
+      {0xc4a2U, 0x526dU, 1U, 0U, 0U},
+      // x = -0x1.5fp+3, coshf16(x) = 0x1.c54p+14 (RZ)
+      {0xc97cU, 0x7715U, 1U, 0U, 1U},
+      // x = -0x1.8c4p+0, coshf16(x) = 0x1.3a8p+1 (RZ)
+      {0xbe31U, 0x40eaU, 1U, 0U, 0U},
+      // x = -0x1.994p+0, coshf16(x) = 0x1.498p+1 (RZ)
+      {0xbe65U, 0x4126U, 1U, 0U, 0U},
+  }};
+#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
+
+  using namespace expxf16_internal;
+  using FPBits = fputil::FPBits<float16>;
+  FPBits x_bits(x);
+
+  uint16_t x_u = x_bits.uintval();
+  uint16_t x_abs = x_u & 0x7fffU;
+
+  // When |x| >= acosh(2^16), or x is NaN.
+  if (LIBC_UNLIKELY(x_abs >= 0x49e5U)) {
+    // cosh(NaN) = NaN
+    if (x_bits.is_nan()) {
+      if (x_bits.is_signaling_nan()) {
+        fputil::raise_except_if_required(FE_INVALID);
+        return FPBits::quiet_nan().get_val();
+      }
+
+      return x;
+    }
+
+    // When |x| >= acosh(2^16).
+    if (x_abs >= 0x49e5U) {
+      // cosh(+/-inf) = +inf
+      if (x_bits.is_inf())
+        return FPBits::inf().get_val();
+
+      switch (fputil::quick_get_round()) {
+      case FE_TONEAREST:
+      case FE_UPWARD:
+        fputil::set_errno_if_required(ERANGE);
+        fputil::raise_except_if_required(FE_OVERFLOW | FE_INEXACT);
+        return FPBits::inf().get_val();
+      default:
+        return FPBits::max_normal().get_val();
+      }
+    }
+  }
+
+#ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
+  if (x_bits.is_pos()) {
+    if (auto r = COSHF16_EXCEPTS_POS.lookup(x_u); LIBC_UNLIKELY(r.has_value()))
+      return r.value();
+  } else {
+    if (auto r = COSHF16_EXCEPTS_NEG.lookup(x_u); LIBC_UNLIKELY(r.has_value()))
+      return r.value();
+  }
+#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
+
+  return eval_sinh_or_cosh</*IsSinh=*/false>(x);
+}
+
+} // namespace math
+
+} // namespace LIBC_NAMESPACE_DECL
+
+#endif // LIBC_TYPES_HAS_FLOAT16
+
+#endif // LLVM_LIBC_SRC___SUPPORT_MATH_COSHF16_H

libc/src/math/generic/expxf16.h → libc/src/__support/math/expxf16_utils.h

@@ -21,7 +21,11 @@
 
 namespace LIBC_NAMESPACE_DECL {
 
-LIBC_INLINE ExpRangeReduction exp2_range_reduction(float16 x) {
+namespace math {
+
+namespace expxf16_internal {
+
+LIBC_INLINE static constexpr ExpRangeReduction exp2_range_reduction(float16 x) {
   // For -25 < x < 16, to compute 2^x, we perform the following range reduction:
   // find hi, mid, lo, such that:
   //   x = hi + mid + lo, in which
@@ -117,7 +121,8 @@ static constexpr cpp::array<uint32_t, 32> EXP2_MID_5_BITS = {
 // The main point of these formulas is that the expensive part of calculating
 // the polynomials approximating lower parts of e^x and e^(-x) is shared and
 // only done once.
-template <bool IsSinh> LIBC_INLINE float16 eval_sinh_or_cosh(float16 x) {
+template <bool IsSinh>
+LIBC_INLINE static constexpr float16 eval_sinh_or_cosh(float16 x) {
   float xf = x;
   float kf = fputil::nearest_integer(xf * (LOG2F_E * 0x1.0p+5f));
   int x_hi_mid_p = static_cast<int>(kf);
@@ -174,7 +179,7 @@ template <bool IsSinh> LIBC_INLINE float16 eval_sinh_or_cosh(float16 x) {
 // Generated by Sollya with the following commands:
 //   > display = hexadecimal;
 //   > for i from 0 to 31 do print(round(log(1 + i * 2^-5), SG, RN));
-constexpr cpp::array<float, 32> LOGF_F = {
+static constexpr cpp::array<float, 32> LOGF_F = {
     0x0p+0f,        0x1.f829bp-6f,  0x1.f0a30cp-5f, 0x1.6f0d28p-4f,
     0x1.e27076p-4f, 0x1.29553p-3f,  0x1.5ff308p-3f, 0x1.9525aap-3f,
     0x1.c8ff7cp-3f, 0x1.fb9186p-3f, 0x1.1675cap-2f, 0x1.2e8e2cp-2f,
@@ -188,7 +193,7 @@ constexpr cpp::array<float, 32> LOGF_F = {
 // Generated by Sollya with the following commands:
 //   > display = hexadecimal;
 //   > for i from 0 to 31 do print(round(log2(1 + i * 2^-5), SG, RN));
-constexpr cpp::array<float, 32> LOG2F_F = {
+static constexpr cpp::array<float, 32> LOG2F_F = {
     0x0p+0f,        0x1.6bad38p-5f, 0x1.663f7p-4f,  0x1.08c588p-3f,
     0x1.5c01a4p-3f, 0x1.acf5e2p-3f, 0x1.fbc16cp-3f, 0x1.24407ap-2f,
     0x1.49a784p-2f, 0x1.6e221cp-2f, 0x1.91bba8p-2f, 0x1.b47ecp-2f,
@@ -202,7 +207,7 @@ constexpr cpp::array<float, 32> LOG2F_F = {
 // Generated by Sollya with the following commands:
 //   > display = hexadecimal;
 //   > for i from 0 to 31 do print(round(log10(1 + i * 2^-5), SG, RN));
-constexpr cpp::array<float, 32> LOG10F_F = {
+static constexpr cpp::array<float, 32> LOG10F_F = {
     0x0p+0f,        0x1.b5e908p-7f, 0x1.af5f92p-6f, 0x1.3ed11ap-5f,
     0x1.a30a9ep-5f, 0x1.02428cp-4f, 0x1.31b306p-4f, 0x1.5fe804p-4f,
     0x1.8cf184p-4f, 0x1.b8de4ep-4f, 0x1.e3bc1ap-4f, 0x1.06cbd6p-3f,
@@ -216,7 +221,7 @@ constexpr cpp::array<float, 32> LOG10F_F = {
 // Generated by Sollya with the following commands:
 //   > display = hexadecimal;
 //   > for i from 0 to 31 do print(round(1 / (1 + i * 2^-5), SG, RN));
-constexpr cpp::array<float, 32> ONE_OVER_F_F = {
+static constexpr cpp::array<float, 32> ONE_OVER_F_F = {
     0x1p+0f,        0x1.f07c2p-1f,  0x1.e1e1e2p-1f, 0x1.d41d42p-1f,
     0x1.c71c72p-1f, 0x1.bacf92p-1f, 0x1.af286cp-1f, 0x1.a41a42p-1f,
     0x1.99999ap-1f, 0x1.8f9c18p-1f, 0x1.861862p-1f, 0x1.7d05f4p-1f,
@@ -227,6 +232,10 @@ constexpr cpp::array<float, 32> ONE_OVER_F_F = {
     0x1.111112p-1f, 0x1.0c9714p-1f, 0x1.08421p-1f,  0x1.041042p-1f,
 };
 
+} // namespace expxf16_internal
+
+} // namespace math
+
 } // namespace LIBC_NAMESPACE_DECL
 
 #endif // LLVM_LIBC_SRC_MATH_GENERIC_EXPXF16_H