[libc][math] Refactor cosf implementation to header-only in src/__support/math folder. #152069

bassiounix · 2025-08-05T03:31:15Z

in preparation for: https://discourse.llvm.org/t/rfc-make-clang-builtin-math-functions-constexpr-with-llvm-libc-to-support-c-23-constexpr-math-functions/86450

bassiounix · 2025-08-05T03:31:47Z

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

[libc][math] Refactor cosf implementation to header-only in src/__support/math folder. #152069 👈 (View in Graphite)
[libc][math] Refactor cos implementation to header-only in src/__support/math folder. #151883
[libc][math] Refactor cbrtf implementation to header-only in src/__support/math folder. #151846
[libc][math] Refactor cbrt implementation to header-only in src/__support/math folder. #151837
[libc][math] Refactor atanhf16 implementation to header-only in src/__support/math folder. #151779
[libc][math] Refactor atanhf implementation to header-only in src/__support/math folder. #151399
[libc][math] Refactor atan2f128 implementation to header-only in src/__support/math folder. #151012
[libc][math] Refactor atan2f implementation to header-only in src/__support/math folder. #150993
[libc][math] Refactor atan2 implementation to header-only in src/__support/math folder. #150968
[libc][math] Refactor atanf16 implementation to header-only in src/__support/math folder. #150868
[libc][math] Refactor atanf implementation to header-only in src/__support/math folder. #150854
[libc][math] Refactor atan implementation to header-only in src/__support/math folder. #150852
[libc][math] Refactor asinhf16 implementation to header-only in src/__support/math folder. #150849
[libc][math] Refactor asinhf implementation to header-only in src/__support/math folder. #150843
main

This stack of pull requests is managed by Graphite. Learn more about stacking.

llvmbot · 2025-08-05T03:32:07Z

@llvm/pr-subscribers-libc

Author: Muhammad Bassiouni (bassiounix)

Changes

Part of #147386

in preparation for: https://discourse.llvm.org/t/rfc-make-clang-builtin-math-functions-constexpr-with-llvm-libc-to-support-c-23-constexpr-math-functions/86450

Patch is 27.94 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/152069.diff

16 Files Affected:

(modified) libc/shared/math.h (+1)
(added) libc/shared/math/cosf.h (+23)
(modified) libc/src/__support/math/CMakeLists.txt (+39)
(added) libc/src/__support/math/cosf.h (+152)
(renamed) libc/src/__support/math/range_reduction.h ()
(renamed) libc/src/__support/math/range_reduction_fma.h ()
(renamed) libc/src/__support/math/sincosf_utils.h ()
(modified) libc/src/math/generic/CMakeLists.txt (+10-42)
(modified) libc/src/math/generic/cosf.cpp (+2-131)
(modified) libc/src/math/generic/cospif.cpp (+1-1)
(modified) libc/src/math/generic/sincosf.cpp (+1-1)
(modified) libc/src/math/generic/sinf.cpp (+3-3)
(modified) libc/src/math/generic/sinpif.cpp (+1-1)
(modified) libc/src/math/generic/tanf.cpp (+1-1)
(modified) libc/src/math/generic/tanpif.cpp (+1-1)
(modified) utils/bazel/llvm-project-overlay/libc/BUILD.bazel (+46-38)

diff --git a/libc/shared/math.h b/libc/shared/math.h
index a5581ed4272a3..0c11640101563 100644
--- a/libc/shared/math.h
+++ b/libc/shared/math.h
@@ -33,6 +33,7 @@
 #include "math/cbrt.h"
 #include "math/cbrtf.h"
 #include "math/cos.h"
+#include "math/cosf.h"
 #include "math/erff.h"
 #include "math/exp.h"
 #include "math/exp10.h"
diff --git a/libc/shared/math/cosf.h b/libc/shared/math/cosf.h
new file mode 100644
index 0000000000000..06182207a82f2
--- /dev/null
+++ b/libc/shared/math/cosf.h
@@ -0,0 +1,23 @@
+//===-- Shared cosf 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_COSF_H
+#define LLVM_LIBC_SHARED_MATH_COSF_H
+
+#include "shared/libc_common.h"
+#include "src/__support/math/cosf.h"
+
+namespace LIBC_NAMESPACE_DECL {
+namespace shared {
+
+using math::cosf;
+
+} // namespace shared
+} // namespace LIBC_NAMESPACE_DECL
+
+#endif // LLVM_LIBC_SHARED_MATH_COSF_H
diff --git a/libc/src/__support/math/CMakeLists.txt b/libc/src/__support/math/CMakeLists.txt
index 24844063fcd24..450d56acafe53 100644
--- a/libc/src/__support/math/CMakeLists.txt
+++ b/libc/src/__support/math/CMakeLists.txt
@@ -374,6 +374,21 @@ add_header_library(
     libc.src.__support.macros.optimization
 )
 
+add_header_library(
+  cosf
+  HDRS
+    cosf.h
+  DEPENDS
+    .sincosf_utils
+    libc.src.errno.errno
+    libc.src.__support.FPUtil.basic_operations
+    libc.src.__support.FPUtil.fenv_impl
+    libc.src.__support.FPUtil.fp_bits
+    libc.src.__support.FPUtil.except_value_utils
+    libc.src.__support.FPUtil.fma
+    libc.src.__support.FPUtil.polyeval
+    libc.src.__support.macros.optimization
+)
 
 add_header_library(
   erff
@@ -649,6 +664,19 @@ add_header_library(
     libc.src.__support.integer_literals
 )
 
+add_header_library(
+  range_reduction
+  HDRS
+    range_reduction.h
+    range_reduction_fma.h
+  DEPENDS
+    libc.src.__support.FPUtil.fp_bits
+    libc.src.__support.FPUtil.fma
+    libc.src.__support.FPUtil.multiply_add
+    libc.src.__support.FPUtil.nearest_integer
+    libc.src.__support.common
+)
+
 add_header_library(
   sincos_eval
   HDRS
@@ -660,3 +688,14 @@ add_header_library(
     libc.src.__support.FPUtil.polyeval
     libc.src.__support.integer_literals
 )
+
+add_header_library(
+  sincosf_utils
+  HDRS
+    sincosf_utils.h
+  DEPENDS
+    .range_reduction
+    libc.src.__support.FPUtil.fp_bits
+    libc.src.__support.FPUtil.polyeval
+    libc.src.__support.common
+)
diff --git a/libc/src/__support/math/cosf.h b/libc/src/__support/math/cosf.h
new file mode 100644
index 0000000000000..074be0b314637
--- /dev/null
+++ b/libc/src/__support/math/cosf.h
@@ -0,0 +1,152 @@
+//===-- Implementation header for cosf --------------------------*- 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 LIBC_SRC___SUPPORT_MATH_COSF_H
+#define LIBC_SRC___SUPPORT_MATH_COSF_H
+
+#include "sincosf_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/multiply_add.h"
+#include "src/__support/macros/config.h"
+#include "src/__support/macros/optimization.h"            // LIBC_UNLIKELY
+#include "src/__support/macros/properties/cpu_features.h" // LIBC_TARGET_CPU_HAS_FMA
+
+namespace LIBC_NAMESPACE_DECL {
+
+namespace math {
+
+LIBC_INLINE static constexpr float cosf(float x) {
+
+#ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
+  // Exceptional cases for cosf.
+  constexpr size_t N_EXCEPTS = 6;
+
+  constexpr fputil::ExceptValues<float, N_EXCEPTS> COSF_EXCEPTS{{
+      // (inputs, RZ output, RU offset, RD offset, RN offset)
+      // x = 0x1.64a032p43, cos(x) = 0x1.9d4ba4p-1 (RZ)
+      {0x55325019, 0x3f4ea5d2, 1, 0, 0},
+      // x = 0x1.4555p51, cos(x) = 0x1.115d7cp-1 (RZ)
+      {0x5922aa80, 0x3f08aebe, 1, 0, 1},
+      // x = 0x1.48a858p54, cos(x) = 0x1.f48148p-2 (RZ)
+      {0x5aa4542c, 0x3efa40a4, 1, 0, 0},
+      // x = 0x1.3170fp63, cos(x) = 0x1.fe2976p-1 (RZ)
+      {0x5f18b878, 0x3f7f14bb, 1, 0, 0},
+      // x = 0x1.2b9622p67, cos(x) = 0x1.f0285cp-1 (RZ)
+      {0x6115cb11, 0x3f78142e, 1, 0, 1},
+      // x = 0x1.ddebdep120, cos(x) = 0x1.114438p-1 (RZ)
+      {0x7beef5ef, 0x3f08a21c, 1, 0, 0},
+  }};
+#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
+
+  using FPBits = typename fputil::FPBits<float>;
+
+  FPBits xbits(x);
+  xbits.set_sign(Sign::POS);
+
+  uint32_t x_abs = xbits.uintval();
+  double xd = static_cast<double>(xbits.get_val());
+
+  // Range reduction:
+  // For |x| > pi/16, we perform range reduction as follows:
+  // Find k and y such that:
+  //   x = (k + y) * pi/32
+  //   k is an integer
+  //   |y| < 0.5
+  // For small range (|x| < 2^45 when FMA instructions are available, 2^22
+  // otherwise), this is done by performing:
+  //   k = round(x * 32/pi)
+  //   y = x * 32/pi - k
+  // For large range, we will omit all the higher parts of 16/pi such that the
+  // least significant bits of their full products with x are larger than 63,
+  // since cos((k + y + 64*i) * pi/32) = cos(x + i * 2pi) = cos(x).
+  //
+  // When FMA instructions are not available, we store the digits of 32/pi in
+  // chunks of 28-bit precision.  This will make sure that the products:
+  //   x * THIRTYTWO_OVER_PI_28[i] are all exact.
+  // When FMA instructions are available, we simply store the digits of 32/pi in
+  // chunks of doubles (53-bit of precision).
+  // So when multiplying by the largest values of single precision, the
+  // resulting output should be correct up to 2^(-208 + 128) ~ 2^-80.  By the
+  // worst-case analysis of range reduction, |y| >= 2^-38, so this should give
+  // us more than 40 bits of accuracy. For the worst-case estimation of range
+  // reduction, see for instances:
+  //   Elementary Functions by J-M. Muller, Chapter 11,
+  //   Handbook of Floating-Point Arithmetic by J-M. Muller et. al.,
+  //   Chapter 10.2.
+  //
+  // Once k and y are computed, we then deduce the answer by the cosine of sum
+  // formula:
+  //   cos(x) = cos((k + y)*pi/32)
+  //          = cos(y*pi/32) * cos(k*pi/32) - sin(y*pi/32) * sin(k*pi/32)
+  // The values of sin(k*pi/32) and cos(k*pi/32) for k = 0..63 are precomputed
+  // and stored using a vector of 32 doubles. Sin(y*pi/32) and cos(y*pi/32) are
+  // computed using degree-7 and degree-6 minimax polynomials generated by
+  // Sollya respectively.
+
+  // |x| < 0x1.0p-12f
+  if (LIBC_UNLIKELY(x_abs < 0x3980'0000U)) {
+    // When |x| < 2^-12, the relative error of the approximation cos(x) ~ 1
+    // is:
+    //   |cos(x) - 1| < |x^2 / 2| = 2^-25 < epsilon(1)/2.
+    // So the correctly rounded values of cos(x) are:
+    //   = 1 - eps(x) if rounding mode = FE_TOWARDZERO or FE_DOWWARD,
+    //   = 1 otherwise.
+    // To simplify the rounding decision and make it more efficient and to
+    // prevent compiler to perform constant folding, we use
+    //   fma(x, -2^-25, 1) instead.
+    // Note: to use the formula 1 - 2^-25*x to decide the correct rounding, we
+    // do need fma(x, -2^-25, 1) to prevent underflow caused by -2^-25*x when
+    // |x| < 2^-125. For targets without FMA instructions, we simply use
+    // double for intermediate results as it is more efficient than using an
+    // emulated version of FMA.
+#if defined(LIBC_TARGET_CPU_HAS_FMA_FLOAT)
+    return fputil::multiply_add(xbits.get_val(), -0x1.0p-25f, 1.0f);
+#else
+    return static_cast<float>(fputil::multiply_add(xd, -0x1.0p-25, 1.0));
+#endif // LIBC_TARGET_CPU_HAS_FMA_FLOAT
+  }
+
+#ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
+  if (auto r = COSF_EXCEPTS.lookup(x_abs); LIBC_UNLIKELY(r.has_value()))
+    return r.value();
+#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
+
+  // x is inf or nan.
+  if (LIBC_UNLIKELY(x_abs >= 0x7f80'0000U)) {
+    if (xbits.is_signaling_nan()) {
+      fputil::raise_except_if_required(FE_INVALID);
+      return FPBits::quiet_nan().get_val();
+    }
+
+    if (x_abs == 0x7f80'0000U) {
+      fputil::set_errno_if_required(EDOM);
+      fputil::raise_except_if_required(FE_INVALID);
+    }
+    return x + FPBits::quiet_nan().get_val();
+  }
+
+  // Combine the results with the sine of sum formula:
+  //   cos(x) = cos((k + y)*pi/32)
+  //          = cos(y*pi/32) * cos(k*pi/32) - sin(y*pi/32) * sin(k*pi/32)
+  //          = cosm1_y * cos_k + sin_y * sin_k
+  //          = (cosm1_y * cos_k + cos_k) + sin_y * sin_k
+  double sin_k = 0, cos_k = 0, sin_y = 0, cosm1_y = 0;
+
+  sincosf_eval(xd, x_abs, sin_k, cos_k, sin_y, cosm1_y);
+
+  return static_cast<float>(fputil::multiply_add(
+      sin_y, -sin_k, fputil::multiply_add(cosm1_y, cos_k, cos_k)));
+}
+
+} // namespace math
+
+} // namespace LIBC_NAMESPACE_DECL
+
+#endif // LIBC_SRC___SUPPORT_MATH_COSF_H
diff --git a/libc/src/math/generic/range_reduction.h b/libc/src/__support/math/range_reduction.h
similarity index 100%
rename from libc/src/math/generic/range_reduction.h
rename to libc/src/__support/math/range_reduction.h
diff --git a/libc/src/math/generic/range_reduction_fma.h b/libc/src/__support/math/range_reduction_fma.h
similarity index 100%
rename from libc/src/math/generic/range_reduction_fma.h
rename to libc/src/__support/math/range_reduction_fma.h
diff --git a/libc/src/math/generic/sincosf_utils.h b/libc/src/__support/math/sincosf_utils.h
similarity index 100%
rename from libc/src/math/generic/sincosf_utils.h
rename to libc/src/__support/math/sincosf_utils.h
diff --git a/libc/src/math/generic/CMakeLists.txt b/libc/src/math/generic/CMakeLists.txt
index b224808e54e45..6f00c93a6e689 100644
--- a/libc/src/math/generic/CMakeLists.txt
+++ b/libc/src/math/generic/CMakeLists.txt
@@ -262,30 +262,6 @@ add_entrypoint_object(
     libc.src.__support.FPUtil.generic.add_sub
 )
 
-add_header_library(
-  range_reduction
-  HDRS
-    range_reduction.h
-    range_reduction_fma.h
-  DEPENDS
-    libc.src.__support.FPUtil.fp_bits
-    libc.src.__support.FPUtil.fma
-    libc.src.__support.FPUtil.multiply_add
-    libc.src.__support.FPUtil.nearest_integer
-    libc.src.__support.common
-)
-
-add_header_library(
-  sincosf_utils
-  HDRS
-    sincosf_utils.h
-  DEPENDS
-    .range_reduction
-    libc.src.__support.FPUtil.fp_bits
-    libc.src.__support.FPUtil.polyeval
-    libc.src.__support.common
-)
-
 add_header_library(
   sincosf16_utils
   HDRS
@@ -313,15 +289,7 @@ add_entrypoint_object(
   HDRS
     ../cosf.h
   DEPENDS
-    .sincosf_utils
-    libc.src.errno.errno
-    libc.src.__support.FPUtil.basic_operations
-    libc.src.__support.FPUtil.fenv_impl
-    libc.src.__support.FPUtil.fp_bits
-    libc.src.__support.FPUtil.except_value_utils
-    libc.src.__support.FPUtil.fma
-    libc.src.__support.FPUtil.polyeval
-    libc.src.__support.macros.optimization
+    libc.src.__support.math.cosf
 )
 
 add_entrypoint_object(
@@ -350,7 +318,7 @@ add_entrypoint_object(
   HDRS
     ../cospif.h
   DEPENDS
-    .sincosf_utils
+    libc.src.__support.math.sincosf_utils
     libc.src.__support.FPUtil.fenv_impl
     libc.src.__support.FPUtil.fp_bits
     libc.src.__support.FPUtil.fma
@@ -402,8 +370,8 @@ add_entrypoint_object(
   HDRS
     ../sinf.h
   DEPENDS
-    .range_reduction
-    .sincosf_utils
+    libc.src.__support.math.range_reduction
+    libc.src.__support.math.sincosf_utils
     libc.src.errno.errno
     libc.src.__support.FPUtil.basic_operations
     libc.src.__support.FPUtil.fenv_impl
@@ -462,7 +430,7 @@ add_entrypoint_object(
   HDRS
     ../sinpif.h
   DEPENDS
-    .sincosf_utils
+    libc.src.__support.math.sincosf_utils
     libc.src.__support.FPUtil.fenv_impl
     libc.src.__support.FPUtil.fp_bits
     libc.src.__support.FPUtil.fma
@@ -479,8 +447,8 @@ add_entrypoint_object(
   HDRS
     ../sincosf.h
   DEPENDS
-    .range_reduction
-    .sincosf_utils
+    libc.src.__support.math.range_reduction
+    libc.src.__support.math.sincosf_utils
     libc.src.errno.errno
     libc.src.__support.FPUtil.fenv_impl
     libc.src.__support.FPUtil.fp_bits
@@ -534,8 +502,8 @@ add_entrypoint_object(
   HDRS
     ../tanf.h
   DEPENDS
-    .range_reduction
-    .sincosf_utils
+    libc.src.__support.math.range_reduction
+    libc.src.__support.math.sincosf_utils
     libc.src.errno.errno
     libc.src.__support.FPUtil.fenv_impl
     libc.src.__support.FPUtil.fenv_impl
@@ -573,7 +541,7 @@ add_entrypoint_object(
   HDRS
     ../tanpif.h
   DEPENDS
-    .sincosf_utils
+    libc.src.__support.math.sincosf_utils
     libc.src.__support.FPUtil.except_value_utils
     libc.src.__support.FPUtil.fenv_impl
     libc.src.__support.FPUtil.fp_bits
diff --git a/libc/src/math/generic/cosf.cpp b/libc/src/math/generic/cosf.cpp
index 7cdae09869588..5c23d99d52441 100644
--- a/libc/src/math/generic/cosf.cpp
+++ b/libc/src/math/generic/cosf.cpp
@@ -7,139 +7,10 @@
 //===----------------------------------------------------------------------===//
 
 #include "src/math/cosf.h"
-#include "sincosf_utils.h"
-#include "src/__support/FPUtil/BasicOperations.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/multiply_add.h"
-#include "src/__support/common.h"
-#include "src/__support/macros/config.h"
-#include "src/__support/macros/optimization.h"            // LIBC_UNLIKELY
-#include "src/__support/macros/properties/cpu_features.h" // LIBC_TARGET_CPU_HAS_FMA
+#include "src/__support/math/cosf.h"
 
 namespace LIBC_NAMESPACE_DECL {
 
-#ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
-// Exceptional cases for cosf.
-static constexpr size_t N_EXCEPTS = 6;
-
-static constexpr fputil::ExceptValues<float, N_EXCEPTS> COSF_EXCEPTS{{
-    // (inputs, RZ output, RU offset, RD offset, RN offset)
-    // x = 0x1.64a032p43, cos(x) = 0x1.9d4ba4p-1 (RZ)
-    {0x55325019, 0x3f4ea5d2, 1, 0, 0},
-    // x = 0x1.4555p51, cos(x) = 0x1.115d7cp-1 (RZ)
-    {0x5922aa80, 0x3f08aebe, 1, 0, 1},
-    // x = 0x1.48a858p54, cos(x) = 0x1.f48148p-2 (RZ)
-    {0x5aa4542c, 0x3efa40a4, 1, 0, 0},
-    // x = 0x1.3170fp63, cos(x) = 0x1.fe2976p-1 (RZ)
-    {0x5f18b878, 0x3f7f14bb, 1, 0, 0},
-    // x = 0x1.2b9622p67, cos(x) = 0x1.f0285cp-1 (RZ)
-    {0x6115cb11, 0x3f78142e, 1, 0, 1},
-    // x = 0x1.ddebdep120, cos(x) = 0x1.114438p-1 (RZ)
-    {0x7beef5ef, 0x3f08a21c, 1, 0, 0},
-}};
-#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
-
-LLVM_LIBC_FUNCTION(float, cosf, (float x)) {
-  using FPBits = typename fputil::FPBits<float>;
-
-  FPBits xbits(x);
-  xbits.set_sign(Sign::POS);
-
-  uint32_t x_abs = xbits.uintval();
-  double xd = static_cast<double>(xbits.get_val());
-
-  // Range reduction:
-  // For |x| > pi/16, we perform range reduction as follows:
-  // Find k and y such that:
-  //   x = (k + y) * pi/32
-  //   k is an integer
-  //   |y| < 0.5
-  // For small range (|x| < 2^45 when FMA instructions are available, 2^22
-  // otherwise), this is done by performing:
-  //   k = round(x * 32/pi)
-  //   y = x * 32/pi - k
-  // For large range, we will omit all the higher parts of 16/pi such that the
-  // least significant bits of their full products with x are larger than 63,
-  // since cos((k + y + 64*i) * pi/32) = cos(x + i * 2pi) = cos(x).
-  //
-  // When FMA instructions are not available, we store the digits of 32/pi in
-  // chunks of 28-bit precision.  This will make sure that the products:
-  //   x * THIRTYTWO_OVER_PI_28[i] are all exact.
-  // When FMA instructions are available, we simply store the digits of 32/pi in
-  // chunks of doubles (53-bit of precision).
-  // So when multiplying by the largest values of single precision, the
-  // resulting output should be correct up to 2^(-208 + 128) ~ 2^-80.  By the
-  // worst-case analysis of range reduction, |y| >= 2^-38, so this should give
-  // us more than 40 bits of accuracy. For the worst-case estimation of range
-  // reduction, see for instances:
-  //   Elementary Functions by J-M. Muller, Chapter 11,
-  //   Handbook of Floating-Point Arithmetic by J-M. Muller et. al.,
-  //   Chapter 10.2.
-  //
-  // Once k and y are computed, we then deduce the answer by the cosine of sum
-  // formula:
-  //   cos(x) = cos((k + y)*pi/32)
-  //          = cos(y*pi/32) * cos(k*pi/32) - sin(y*pi/32) * sin(k*pi/32)
-  // The values of sin(k*pi/32) and cos(k*pi/32) for k = 0..63 are precomputed
-  // and stored using a vector of 32 doubles. Sin(y*pi/32) and cos(y*pi/32) are
-  // computed using degree-7 and degree-6 minimax polynomials generated by
-  // Sollya respectively.
-
-  // |x| < 0x1.0p-12f
-  if (LIBC_UNLIKELY(x_abs < 0x3980'0000U)) {
-    // When |x| < 2^-12, the relative error of the approximation cos(x) ~ 1
-    // is:
-    //   |cos(x) - 1| < |x^2 / 2| = 2^-25 < epsilon(1)/2.
-    // So the correctly rounded values of cos(x) are:
-    //   = 1 - eps(x) if rounding mode = FE_TOWARDZERO or FE_DOWWARD,
-    //   = 1 otherwise.
-    // To simplify the rounding decision and make it more efficient and to
-    // prevent compiler to perform constant folding, we use
-    //   fma(x, -2^-25, 1) instead.
-    // Note: to use the formula 1 - 2^-25*x to decide the correct rounding, we
-    // do need fma(x, -2^-25, 1) to prevent underflow caused by -2^-25*x when
-    // |x| < 2^-125. For targets without FMA instructions, we simply use
-    // double for intermediate results as it is more efficient than using an
-    // emulated version of FMA.
-#if defined(LIBC_TARGET_CPU_HAS_FMA_FLOAT)
-    return fputil::multiply_add(xbits.get_val(), -0x1.0p-25f, 1.0f);
-#else
-    return static_cast<float>(fputil::multiply_add(xd, -0x1.0p-25, 1.0));
-#endif // LIBC_TARGET_CPU_HAS_FMA_FLOAT
-  }
-
-#ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
-  if (auto r = COSF_EXCEPTS.lookup(x_abs); LIBC_UNLIKELY(r.has_value()))
-    return r.value();
-#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
-
-  // x is inf or nan.
-  if (LIBC_UNLIKELY(x_abs >= 0x7f80'0000U)) {
-    if (xbits.is_signaling_nan()) {
-      fputil::raise_except_if_required(FE_INVALID);
-      return FPBits::quiet_nan().get_val();
-    }
-
-    if (x_abs == 0x7f80'0000U) {
-      fputil::set_errno_if_required(EDOM);
-      fputil::raise_except_if_required(FE_INVALID);
-    }
-    return x + FPBits::quiet_nan().get_val();
-  }
-
-  // Combine the results with the sine of sum formula:
-  //   cos(x) = cos((k + y)*pi/32)
-  //          = cos(y*pi/32) * cos(k*pi/32) - sin(y*pi/32) * sin(k*pi/32)
-  //          = cosm1_y * cos_k + sin_y * sin_k
-  //          = (cosm1_y * cos_k + cos_k) + sin_y * sin_k
-  double sin_k, cos_k, sin_y, cosm1_y;
-
-  sincosf_eval(xd, x_abs, sin_k, cos_k, sin_y, cosm1_y);
-
-  return static_cast<float>(fputil::multiply_add(
-      sin_y, -sin_k, fputil::multiply_add(cosm1_y, cos_k, cos_k)));
-}
+LLVM_LIBC_FUNCTION(float, cosf, (float x)) { return math::cosf(x); }
 
 } // namespace LIBC_NAMESPACE_DECL
diff --git a/libc/src/math/generic/cospif.cpp b/libc/src/math/generic/cospif.cpp
index 5b6880f853b26..6b524a2cace7e 100644
--- a/libc/src/math/generic/cospif.cpp
+++ b/libc/src/math/generic/cospif.cpp
@@ -7,7 +7,6 @@
 //===----------------------------------------------------------------------===//
 
 #include "src/math/cospif.h"
-#include "sincosf_utils.h"
 #include "src/__support/FPUtil/FEnvImpl.h"
 #include "src/__support/FPUtil/FPBits.h"
 #include "src/__support/FPUtil/multiply_add.h"
@@ -15,6 +14,7 @@
 #include "src/__support/macros/config.h"
 #include "src/__support/macros/optimization.h"            // LIBC_UNLIKELY
 #include "src/__support/macros/properties/cpu_features.h" // LIBC_TARGET_CPU_HAS_FMA
+#include "src/__support/math/sincosf_utils.h"
 
 namespace LIBC_NAMESPACE_DECL {
 
diff --git a/libc/src/math/generic/sincosf.cpp b/libc/src/math/generic/sincosf.cpp
index 9c7bf181e485e..5179c985a2240 100644
--- a/libc/src/math/generic/sincosf.cpp
+++ b/libc/src/math/generic/sincosf.cpp
@@ -7,7 +7,6 @@
 //===----------------------------------------------------------------------===//
 
 #include "src/math/sincosf.h"
-#include "sincosf_utils.h"
 #includ...
[truncated]

…port/math folder.

bassiounix requested a review from lntue August 5, 2025 03:31

bassiounix added bazel "Peripheral" support tier build system: utils/bazel libc labels Aug 5, 2025 — with Graphite App

bassiounix marked this pull request as ready for review August 5, 2025 03:32

bassiounix requested review from rupprecht, keith and aaronmondal as code owners August 5, 2025 03:32

bassiounix mentioned this pull request Aug 3, 2025

[libc][math] Tracking Issue - Refactor Higher Math Functions to Header Only #147386

Open

bassiounix removed request for keith, rupprecht and aaronmondal August 5, 2025 03:32

bassiounix force-pushed the users/bassiounix/spr/08-05-_libc_math_refactor_cosf_implementation_to_header-only_in_src___support_math_folder branch 2 times, most recently from 01e33d3 to 171a818 Compare August 5, 2025 03:37

bassiounix force-pushed the users/bassiounix/spr/08-03-_libc_math_refactor_cos_implementation_to_header-only_in_src___support_math_folder branch 2 times, most recently from a66f719 to 37848d7 Compare August 5, 2025 03:46

[libc][math] Refactor cosf implementation to header-only in src/__sup…

2ff880d

…port/math folder.

bassiounix force-pushed the users/bassiounix/spr/08-05-_libc_math_refactor_cosf_implementation_to_header-only_in_src___support_math_folder branch from 171a818 to 2ff880d Compare August 5, 2025 03:48

bassiounix force-pushed the users/bassiounix/spr/08-03-_libc_math_refactor_cos_implementation_to_header-only_in_src___support_math_folder branch from 37848d7 to 14a8e8a Compare August 5, 2025 03:48

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[libc][math] Refactor cosf implementation to header-only in src/__support/math folder. #152069

[libc][math] Refactor cosf implementation to header-only in src/__support/math folder. #152069

Uh oh!

bassiounix commented Aug 5, 2025 •

edited

Loading

Uh oh!

bassiounix commented Aug 5, 2025

Uh oh!

llvmbot commented Aug 5, 2025

Uh oh!

Uh oh!

[libc][math] Refactor cosf implementation to header-only in src/__support/math folder. #152069

Are you sure you want to change the base?

[libc][math] Refactor cosf implementation to header-only in src/__support/math folder. #152069

Uh oh!

Conversation

bassiounix commented Aug 5, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

bassiounix commented Aug 5, 2025

Uh oh!

llvmbot commented Aug 5, 2025

Uh oh!

Uh oh!

bassiounix commented Aug 5, 2025 •

edited

Loading