[libc-commits] [libc] [llvm] [libc][math] Refactor tanhf16 implementation to header-only in src/__support/math folder. (PR #178645)

Xinlong Chen via libc-commits libc-commits at lists.llvm.org
Thu Jan 29 04:10:09 PST 2026


https://github.com/Xinlong-Chen updated https://github.com/llvm/llvm-project/pull/178645

>From 1c00018096005d4457e616080274d93da75b6439 Mon Sep 17 00:00:00 2001
From: xinlongchen <xinlongchen at tencent.com>
Date: Thu, 29 Jan 2026 20:05:30 +0800
Subject: [PATCH] [libc][math] Refactor tanhf16 implementation to header-only
 in src/__support/math folder.

---
 libc/shared/math.h                            |   1 +
 libc/shared/math/tanhf16.h                    |  23 +++
 libc/src/__support/math/CMakeLists.txt        |  19 ++
 libc/src/__support/math/tanhf16.h             | 163 ++++++++++++++++++
 libc/src/math/generic/CMakeLists.txt          |  12 +-
 libc/src/math/generic/tanhf16.cpp             | 140 +--------------
 libc/test/shared/CMakeLists.txt               |   1 +
 libc/test/shared/shared_math_test.cpp         |   1 +
 .../llvm-project-overlay/libc/BUILD.bazel     |  10 +-
 9 files changed, 220 insertions(+), 150 deletions(-)
 create mode 100644 libc/shared/math/tanhf16.h
 create mode 100644 libc/src/__support/math/tanhf16.h

diff --git a/libc/shared/math.h b/libc/shared/math.h
index 6b1f94fc6cce5..e9e0938ba1ba3 100644
--- a/libc/shared/math.h
+++ b/libc/shared/math.h
@@ -102,5 +102,6 @@
 #include "math/sqrtf16.h"
 #include "math/tan.h"
 #include "math/tanf.h"
+#include "math/tanhf16.h"
 
 #endif // LLVM_LIBC_SHARED_MATH_H
diff --git a/libc/shared/math/tanhf16.h b/libc/shared/math/tanhf16.h
new file mode 100644
index 0000000000000..47f3531d94ee1
--- /dev/null
+++ b/libc/shared/math/tanhf16.h
@@ -0,0 +1,23 @@
+//===-- Shared tanhf16 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_TANHF16_H
+#define LLVM_LIBC_SHARED_MATH_TANHF16_H
+
+#include "shared/libc_common.h"
+#include "src/__support/math/tanhf16.h"
+
+namespace LIBC_NAMESPACE_DECL {
+namespace shared {
+
+using math::tanhf16;
+
+} // namespace shared
+} // namespace LIBC_NAMESPACE_DECL
+
+#endif // LLVM_LIBC_SHARED_MATH_TANHF16_H
diff --git a/libc/src/__support/math/CMakeLists.txt b/libc/src/__support/math/CMakeLists.txt
index aad2270340c66..1893eeac13b1d 100644
--- a/libc/src/__support/math/CMakeLists.txt
+++ b/libc/src/__support/math/CMakeLists.txt
@@ -1487,3 +1487,22 @@ add_header_library(
     libc.src.__support.FPUtil.polyeval
     libc.src.__support.macros.optimization
 )
+
+add_header_library(
+  tanhf16
+  HDRS
+    tanhf16.h
+  DEPENDS
+    libc.hdr.fenv_macros
+    libc.src.__support.CPP.array
+    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.nearest_integer
+    libc.src.__support.FPUtil.polyeval
+    libc.src.__support.FPUtil.rounding_mode
+    libc.src.__support.macros.optimization
+    libc.src.__support.math.expxf16_utils
+)
diff --git a/libc/src/__support/math/tanhf16.h b/libc/src/__support/math/tanhf16.h
new file mode 100644
index 0000000000000..4983092e004b1
--- /dev/null
+++ b/libc/src/__support/math/tanhf16.h
@@ -0,0 +1,163 @@
+//===-- Single-precision tanhf16 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
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_LIBC_SRC___SUPPORT_MATH_TANHF16_H
+#define LLVM_LIBC_SRC___SUPPORT_MATH_TANHF16_H
+
+#include "hdr/fenv_macros.h"
+#include "src/__support/CPP/array.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/nearest_integer.h"
+#include "src/__support/FPUtil/rounding_mode.h"
+#include "src/__support/common.h"
+#include "src/__support/macros/config.h"
+#include "src/__support/macros/optimization.h"
+#include "src/__support/math/expxf16_utils.h"
+
+namespace LIBC_NAMESPACE_DECL {
+
+namespace math {
+
+namespace tanhf16_internal {
+#ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
+static constexpr fputil::ExceptValues<float16, 2> TANHF16_EXCEPTS = {{
+    // x = 0x1.f54p+0, tanhf16(x) = 0x1.ecp-1 (RZ)
+    {0x3fd5U, 0x3bb0U, 1U, 0U, 0U},
+    // x = -0x1.f54p+0, tanhf16(x) = -0x1.ecp-1 (RZ)
+    {0xbfd5U, 0xbbb0U, 0U, 1U, 0U},
+}};
+#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS  
+} // namespace tanhf16_internal
+
+LIBC_INLINE static constexpr float16 tanhf16(float16 x) {
+  using namespace tanhf16_internal;
+  using namespace math::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 -2^(-14) <= x <= -2^(-9), or |x| <= 0x1.d2p-4,
+  // or |x| >= atanh(1 - 2^(-11)), or x is NaN.
+  if (LIBC_UNLIKELY(x_abs <= 0x2f48U || x_abs >= 0x4429U)) {
+    // tanh(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 -2^(-14) <= x <= -2^(-9).
+    if (x_u >= 0x8400U && x_u <= 0x9800U) {
+      switch (fputil::quick_get_round()) {
+      case FE_TONEAREST:
+      case FE_DOWNWARD:
+        return x;
+      default:
+        return FPBits(static_cast<uint16_t>(x_u - 1U)).get_val();
+      }
+    }
+
+    // When |x| <= 0x1.d2p-4.
+    if (x_abs <= 0x2f48U) {
+      if (LIBC_UNLIKELY(x_abs == 0))
+        return x;
+
+      float xf = x;
+      float xf_sq = xf * xf;
+      // Degree-7 Taylor expansion generated by Sollya with the following
+      // commands:
+      //   > taylor(tanh(x), 7, 0);
+      //   > display = hexadecimal;
+      //   > // For each coefficient:
+      //   > round(/* put coefficient here */, SG, RN);
+      return fputil::cast<float16>(
+          xf * fputil::polyeval(xf_sq, 0x1p+0f, -0x1.555556p-2f, 0x1.111112p-3f,
+                                -0x1.ba1ba2p-5f));
+    }
+
+    // tanh(+/-inf) = +/-1
+    if (x_bits.is_inf())
+      return FPBits::one(x_bits.sign()).get_val();
+
+    // When |x| >= atanh(1 - 2^(-11)).
+    fputil::raise_except_if_required(FE_INEXACT);
+
+    int rounding_mode = fputil::quick_get_round();
+    if ((rounding_mode == FE_TONEAREST && x_abs >= 0x4482U) ||
+        (rounding_mode == FE_UPWARD && x_bits.is_pos()) ||
+        (rounding_mode == FE_DOWNWARD && x_bits.is_neg())) {
+      return FPBits::one(x_bits.sign()).get_val();
+    }
+    if (x_bits.is_pos())
+      return fputil::cast<float16>(0x1.ffcp-1);
+    return fputil::cast<float16>(-0x1.ffcp-1);
+  }
+
+#ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
+  if (auto r = TANHF16_EXCEPTS.lookup(x_u); LIBC_UNLIKELY(r.has_value()))
+    return r.value();
+#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
+
+  // For atanh(-1 + 2^(-11)) < x < atanh(1 - 2^(-11)), to compute tanh(x), we
+  // perform the following range reduction: find hi, mid, lo, such that:
+  //   x = (hi + mid) * log(2) * 0.5 + lo, in which
+  //     hi is an integer,
+  //     mid * 2^5 is an integer,
+  //     -2^(-5) <= lo < 2^(-5).
+  // In particular,
+  //   hi + mid = round(x * log2(e) * 2 * 2^5) * 2^(-5).
+  // Then,
+  //   tanh(x) = sinh(x)/cosh(x)
+  //           = (e^x - e^(-x)) / (e^x + e^(-x))
+  //           = (e^(2x) - 1) / (e^(2x) + 1)
+  //           = (2^(hi + mid) * e^(2*lo) - 1) / (2^(hi + mid) * e^(2*lo) + 1)
+  //           = (e^(2*lo) - 2^(-hi - mid)) / (e^(2*lo) + 2^(-hi - mid))
+  // We store 2^(-mid) in the lookup table EXP2_MID_5_BITS, and compute
+  // 2^(-hi - mid) by adding -hi to the exponent field of 2^(-mid).
+  // e^lo is computed using a degree-3 minimax polynomial generated by Sollya.
+
+  float xf = x;
+  float kf = fputil::nearest_integer(xf * (LOG2F_E * 2.0f * 0x1.0p+5f));
+  int x_hi_mid = -static_cast<int>(kf);
+  unsigned x_hi = static_cast<unsigned>(x_hi_mid) >> 5;
+  unsigned x_mid = static_cast<unsigned>(x_hi_mid) & 0x1f;
+  // lo = x - (hi + mid)
+  //    = round(x * log2(e) * 2 * 2^5) * log(2) * 0.5 * (-2^(-5)) + x
+  float lo = fputil::multiply_add(kf, LOGF_2 * 0.5f * -0x1.0p-5f, xf);
+
+  uint32_t exp2_hi_mid_bits =
+      EXP2_MID_5_BITS[x_mid] +
+      static_cast<uint32_t>(x_hi << fputil::FPBits<float>::FRACTION_LEN);
+  // exp2_hi_mid = 2^(-hi - mid)
+  float exp2_hi_mid = fputil::FPBits<float>(exp2_hi_mid_bits).get_val();
+  // Degree-3 minimax polynomial generated by Sollya with the following
+  // commands:
+  //   > display = hexadecimal;
+  //   > P = fpminimax(expm1(2*x)/x, 2, [|SG...|], [-2^-5, 2^-5]);
+  //   > 1 + x * P;
+  float exp_2lo =
+      fputil::polyeval(lo, 0x1p+0f, 0x1p+1f, 0x1.001p+1f, 0x1.555ddep+0f);
+  return fputil::cast<float16>((exp_2lo - exp2_hi_mid) /
+                               (exp_2lo + exp2_hi_mid));
+}
+
+} // namespace math
+
+} // namespace LIBC_NAMESPACE_DECL
+
+#endif // LLVM_LIBC_SRC___SUPPORT_MATH_TANHF16_H
diff --git a/libc/src/math/generic/CMakeLists.txt b/libc/src/math/generic/CMakeLists.txt
index 21c3a861c87d7..ab7e0f3509fcf 100644
--- a/libc/src/math/generic/CMakeLists.txt
+++ b/libc/src/math/generic/CMakeLists.txt
@@ -4141,17 +4141,7 @@ add_entrypoint_object(
     ../tanhf16.h
   DEPENDS
     libc.hdr.fenv_macros
-    libc.src.__support.CPP.array
-    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.nearest_integer
-    libc.src.__support.FPUtil.polyeval
-    libc.src.__support.FPUtil.rounding_mode
-    libc.src.__support.macros.optimization
-    libc.src.__support.math.expxf16_utils
+    libc.src.__support.math.tanhf16
 )
 
 add_entrypoint_object(
diff --git a/libc/src/math/generic/tanhf16.cpp b/libc/src/math/generic/tanhf16.cpp
index fc0e28b95a834..564f37c79d308 100644
--- a/libc/src/math/generic/tanhf16.cpp
+++ b/libc/src/math/generic/tanhf16.cpp
@@ -7,146 +7,10 @@
 //===----------------------------------------------------------------------===//
 
 #include "src/math/tanhf16.h"
-#include "hdr/fenv_macros.h"
-#include "src/__support/CPP/array.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/nearest_integer.h"
-#include "src/__support/FPUtil/rounding_mode.h"
-#include "src/__support/common.h"
-#include "src/__support/macros/config.h"
-#include "src/__support/macros/optimization.h"
-#include "src/__support/math/expxf16_utils.h"
+#include "src/__support/math/tanhf16.h"
 
 namespace LIBC_NAMESPACE_DECL {
 
-#ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
-static constexpr fputil::ExceptValues<float16, 2> TANHF16_EXCEPTS = {{
-    // x = 0x1.f54p+0, tanhf16(x) = 0x1.ecp-1 (RZ)
-    {0x3fd5U, 0x3bb0U, 1U, 0U, 0U},
-    // x = -0x1.f54p+0, tanhf16(x) = -0x1.ecp-1 (RZ)
-    {0xbfd5U, 0xbbb0U, 0U, 1U, 0U},
-}};
-#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
-
-LLVM_LIBC_FUNCTION(float16, tanhf16, (float16 x)) {
-  using namespace math::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 -2^(-14) <= x <= -2^(-9), or |x| <= 0x1.d2p-4,
-  // or |x| >= atanh(1 - 2^(-11)), or x is NaN.
-  if (LIBC_UNLIKELY(x_abs <= 0x2f48U || x_abs >= 0x4429U)) {
-    // tanh(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 -2^(-14) <= x <= -2^(-9).
-    if (x_u >= 0x8400U && x_u <= 0x9800U) {
-      switch (fputil::quick_get_round()) {
-      case FE_TONEAREST:
-      case FE_DOWNWARD:
-        return x;
-      default:
-        return FPBits(static_cast<uint16_t>(x_u - 1U)).get_val();
-      }
-    }
-
-    // When |x| <= 0x1.d2p-4.
-    if (x_abs <= 0x2f48U) {
-      if (LIBC_UNLIKELY(x_abs == 0))
-        return x;
-
-      float xf = x;
-      float xf_sq = xf * xf;
-      // Degree-7 Taylor expansion generated by Sollya with the following
-      // commands:
-      //   > taylor(tanh(x), 7, 0);
-      //   > display = hexadecimal;
-      //   > // For each coefficient:
-      //   > round(/* put coefficient here */, SG, RN);
-      return fputil::cast<float16>(
-          xf * fputil::polyeval(xf_sq, 0x1p+0f, -0x1.555556p-2f, 0x1.111112p-3f,
-                                -0x1.ba1ba2p-5f));
-    }
-
-    // tanh(+/-inf) = +/-1
-    if (x_bits.is_inf())
-      return FPBits::one(x_bits.sign()).get_val();
-
-    // When |x| >= atanh(1 - 2^(-11)).
-    fputil::raise_except_if_required(FE_INEXACT);
-
-    int rounding_mode = fputil::quick_get_round();
-    if ((rounding_mode == FE_TONEAREST && x_abs >= 0x4482U) ||
-        (rounding_mode == FE_UPWARD && x_bits.is_pos()) ||
-        (rounding_mode == FE_DOWNWARD && x_bits.is_neg())) {
-      return FPBits::one(x_bits.sign()).get_val();
-    }
-    if (x_bits.is_pos())
-      return fputil::cast<float16>(0x1.ffcp-1);
-    return fputil::cast<float16>(-0x1.ffcp-1);
-  }
-
-#ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
-  if (auto r = TANHF16_EXCEPTS.lookup(x_u); LIBC_UNLIKELY(r.has_value()))
-    return r.value();
-#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
-
-  // For atanh(-1 + 2^(-11)) < x < atanh(1 - 2^(-11)), to compute tanh(x), we
-  // perform the following range reduction: find hi, mid, lo, such that:
-  //   x = (hi + mid) * log(2) * 0.5 + lo, in which
-  //     hi is an integer,
-  //     mid * 2^5 is an integer,
-  //     -2^(-5) <= lo < 2^(-5).
-  // In particular,
-  //   hi + mid = round(x * log2(e) * 2 * 2^5) * 2^(-5).
-  // Then,
-  //   tanh(x) = sinh(x)/cosh(x)
-  //           = (e^x - e^(-x)) / (e^x + e^(-x))
-  //           = (e^(2x) - 1) / (e^(2x) + 1)
-  //           = (2^(hi + mid) * e^(2*lo) - 1) / (2^(hi + mid) * e^(2*lo) + 1)
-  //           = (e^(2*lo) - 2^(-hi - mid)) / (e^(2*lo) + 2^(-hi - mid))
-  // We store 2^(-mid) in the lookup table EXP2_MID_5_BITS, and compute
-  // 2^(-hi - mid) by adding -hi to the exponent field of 2^(-mid).
-  // e^lo is computed using a degree-3 minimax polynomial generated by Sollya.
-
-  float xf = x;
-  float kf = fputil::nearest_integer(xf * (LOG2F_E * 2.0f * 0x1.0p+5f));
-  int x_hi_mid = -static_cast<int>(kf);
-  unsigned x_hi = static_cast<unsigned>(x_hi_mid) >> 5;
-  unsigned x_mid = static_cast<unsigned>(x_hi_mid) & 0x1f;
-  // lo = x - (hi + mid)
-  //    = round(x * log2(e) * 2 * 2^5) * log(2) * 0.5 * (-2^(-5)) + x
-  float lo = fputil::multiply_add(kf, LOGF_2 * 0.5f * -0x1.0p-5f, xf);
-
-  uint32_t exp2_hi_mid_bits =
-      EXP2_MID_5_BITS[x_mid] +
-      static_cast<uint32_t>(x_hi << fputil::FPBits<float>::FRACTION_LEN);
-  // exp2_hi_mid = 2^(-hi - mid)
-  float exp2_hi_mid = fputil::FPBits<float>(exp2_hi_mid_bits).get_val();
-  // Degree-3 minimax polynomial generated by Sollya with the following
-  // commands:
-  //   > display = hexadecimal;
-  //   > P = fpminimax(expm1(2*x)/x, 2, [|SG...|], [-2^-5, 2^-5]);
-  //   > 1 + x * P;
-  float exp_2lo =
-      fputil::polyeval(lo, 0x1p+0f, 0x1p+1f, 0x1.001p+1f, 0x1.555ddep+0f);
-  return fputil::cast<float16>((exp_2lo - exp2_hi_mid) /
-                               (exp_2lo + exp2_hi_mid));
-}
+LLVM_LIBC_FUNCTION(float16, tanhf16, (float16 x)) { return math::tanhf16(x);}
 
 } // namespace LIBC_NAMESPACE_DECL
diff --git a/libc/test/shared/CMakeLists.txt b/libc/test/shared/CMakeLists.txt
index b0a5ad730cbb2..98f7909547fa6 100644
--- a/libc/test/shared/CMakeLists.txt
+++ b/libc/test/shared/CMakeLists.txt
@@ -99,4 +99,5 @@ add_fp_unittest(
     libc.src.__support.math.sqrt
     libc.src.__support.math.tan
     libc.src.__support.math.tanf
+    libc.src.__support.math.tanhf16
 )
diff --git a/libc/test/shared/shared_math_test.cpp b/libc/test/shared/shared_math_test.cpp
index a21863e23c3f9..cce8be020bf91 100644
--- a/libc/test/shared/shared_math_test.cpp
+++ b/libc/test/shared/shared_math_test.cpp
@@ -100,6 +100,7 @@ TEST(LlvmLibcSharedMathTest, AllFloat) {
   EXPECT_FP_EQ(0x0p+0f, LIBC_NAMESPACE::shared::sinpif(0.0f));
   EXPECT_FP_EQ(0.0f, LIBC_NAMESPACE::shared::sinf(0.0f));
   EXPECT_FP_EQ(0.0f, LIBC_NAMESPACE::shared::tanf(0.0f));
+  EXPECT_FP_EQ(0.0f, LIBC_NAMESPACE::shared::tanhf16(0.0f16));
 }
 
 TEST(LlvmLibcSharedMathTest, AllDouble) {
diff --git a/utils/bazel/llvm-project-overlay/libc/BUILD.bazel b/utils/bazel/llvm-project-overlay/libc/BUILD.bazel
index a313ba48889d8..85f472b6aabdd 100644
--- a/utils/bazel/llvm-project-overlay/libc/BUILD.bazel
+++ b/utils/bazel/llvm-project-overlay/libc/BUILD.bazel
@@ -3710,6 +3710,14 @@ libc_support_library(
     ],
 )
 
+libc_support_library(
+    name = "__support_math_tanhf16",
+    hdrs = ["src/__support/math/tanhf16.h"],
+    deps = [
+        ":__support_math_expxf16_utils",
+    ],
+)
+
 ############################### complex targets ################################
 
 libc_function(
@@ -5420,7 +5428,7 @@ libc_math_function(
 libc_math_function(
     name = "tanhf16",
     additional_deps = [
-        ":__support_math_expxf16_utils",
+        ":__support_math_tanhf16",
     ],
 )
 



More information about the libc-commits mailing list