[llvm-branch-commits] [libc] [llvm] [libc][math] Refactor cosf implementation to header-only in src/__support/math folder. (PR #152069)
Muhammad Bassiouni via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Mon Aug 4 20:48:43 PDT 2025
https://github.com/bassiounix updated https://github.com/llvm/llvm-project/pull/152069
>From 2ff880d6bf304f919b92ae62cac84e4b9789bcf1 Mon Sep 17 00:00:00 2001
From: bassiounix <muhammad.m.bassiouni at gmail.com>
Date: Tue, 5 Aug 2025 06:30:16 +0300
Subject: [PATCH] [libc][math] Refactor cosf implementation to header-only in
src/__support/math folder.
---
libc/shared/math.h | 1 +
libc/shared/math/cosf.h | 23 +++
libc/src/__support/math/CMakeLists.txt | 39 +++++
libc/src/__support/math/cosf.h | 152 ++++++++++++++++++
.../math}/range_reduction.h | 6 +-
.../math}/range_reduction_fma.h | 6 +-
.../math}/sincosf_utils.h | 6 +-
libc/src/math/generic/CMakeLists.txt | 52 ++----
libc/src/math/generic/cosf.cpp | 133 +--------------
libc/src/math/generic/cospif.cpp | 2 +-
libc/src/math/generic/sincosf.cpp | 2 +-
libc/src/math/generic/sinf.cpp | 6 +-
libc/src/math/generic/sinpif.cpp | 2 +-
libc/src/math/generic/tanf.cpp | 2 +-
libc/src/math/generic/tanpif.cpp | 2 +-
.../llvm-project-overlay/libc/BUILD.bazel | 84 +++++-----
16 files changed, 290 insertions(+), 228 deletions(-)
create mode 100644 libc/shared/math/cosf.h
create mode 100644 libc/src/__support/math/cosf.h
rename libc/src/{math/generic => __support/math}/range_reduction.h (95%)
rename libc/src/{math/generic => __support/math}/range_reduction_fma.h (95%)
rename libc/src/{math/generic => __support/math}/sincosf_utils.h (97%)
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 95%
rename from libc/src/math/generic/range_reduction.h
rename to libc/src/__support/math/range_reduction.h
index 9ea446d89bed5..e3f25e4fda3c1 100644
--- a/libc/src/math/generic/range_reduction.h
+++ b/libc/src/__support/math/range_reduction.h
@@ -6,8 +6,8 @@
//
//===----------------------------------------------------------------------===//
-#ifndef LLVM_LIBC_SRC_MATH_GENERIC_RANGE_REDUCTION_H
-#define LLVM_LIBC_SRC_MATH_GENERIC_RANGE_REDUCTION_H
+#ifndef LIBC_SRC___SUPPORT_MATH_RANGE_REDUCTION_H
+#define LIBC_SRC___SUPPORT_MATH_RANGE_REDUCTION_H
#include "src/__support/FPUtil/FPBits.h"
#include "src/__support/FPUtil/multiply_add.h"
@@ -87,4 +87,4 @@ LIBC_INLINE int64_t large_range_reduction(double x, int x_exp, double &y) {
} // namespace LIBC_NAMESPACE_DECL
-#endif // LLVM_LIBC_SRC_MATH_GENERIC_RANGE_REDUCTION_H
+#endif // 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 95%
rename from libc/src/math/generic/range_reduction_fma.h
rename to libc/src/__support/math/range_reduction_fma.h
index 537d5729dc42c..c06a1d8e5760f 100644
--- a/libc/src/math/generic/range_reduction_fma.h
+++ b/libc/src/__support/math/range_reduction_fma.h
@@ -6,8 +6,8 @@
//
//===----------------------------------------------------------------------===//
-#ifndef LLVM_LIBC_SRC_MATH_GENERIC_RANGE_REDUCTION_FMA_H
-#define LLVM_LIBC_SRC_MATH_GENERIC_RANGE_REDUCTION_FMA_H
+#ifndef LIBC_SRC___SUPPORT_MATH_RANGE_REDUCTION_FMA_H
+#define LIBC_SRC___SUPPORT_MATH_RANGE_REDUCTION_FMA_H
#include "src/__support/FPUtil/FMA.h"
#include "src/__support/FPUtil/FPBits.h"
@@ -89,4 +89,4 @@ LIBC_INLINE int64_t large_range_reduction(double x, int x_exp, double &y) {
} // namespace LIBC_NAMESPACE_DECL
-#endif // LLVM_LIBC_SRC_MATH_GENERIC_RANGE_REDUCTION_FMA_H
+#endif // 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 97%
rename from libc/src/math/generic/sincosf_utils.h
rename to libc/src/__support/math/sincosf_utils.h
index 6eaf820e5c1b0..ed9d9f6bda0eb 100644
--- a/libc/src/math/generic/sincosf_utils.h
+++ b/libc/src/__support/math/sincosf_utils.h
@@ -6,8 +6,8 @@
//
//===----------------------------------------------------------------------===//
-#ifndef LLVM_LIBC_SRC_MATH_GENERIC_SINCOSF_UTILS_H
-#define LLVM_LIBC_SRC_MATH_GENERIC_SINCOSF_UTILS_H
+#ifndef LIBC_SRC___SUPPORT_MATH_SINCOSF_UTILS_H
+#define LIBC_SRC___SUPPORT_MATH_SINCOSF_UTILS_H
#include "src/__support/FPUtil/FPBits.h"
#include "src/__support/FPUtil/PolyEval.h"
@@ -122,4 +122,4 @@ LIBC_INLINE void sincospif_eval(double xd, double &sin_k, double &cos_k,
} // namespace LIBC_NAMESPACE_DECL
-#endif // LLVM_LIBC_SRC_MATH_GENERIC_SINCOSF_UTILS_H
+#endif // 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"
#include "src/__support/FPUtil/FEnvImpl.h"
#include "src/__support/FPUtil/FPBits.h"
#include "src/__support/FPUtil/multiply_add.h"
@@ -16,6 +15,7 @@
#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/sinf.cpp b/libc/src/math/generic/sinf.cpp
index 38ea56f5f28c6..a8e634ce2f5ac 100644
--- a/libc/src/math/generic/sinf.cpp
+++ b/libc/src/math/generic/sinf.cpp
@@ -7,7 +7,6 @@
//===----------------------------------------------------------------------===//
#include "src/math/sinf.h"
-#include "sincosf_utils.h"
#include "src/__support/FPUtil/BasicOperations.h"
#include "src/__support/FPUtil/FEnvImpl.h"
#include "src/__support/FPUtil/FPBits.h"
@@ -18,11 +17,12 @@
#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"
#if defined(LIBC_TARGET_CPU_HAS_FMA_DOUBLE)
-#include "range_reduction_fma.h"
+#include "src/__support/math/range_reduction_fma.h"
#else
-#include "range_reduction.h"
+#include "src/__support/math/range_reduction.h"
#endif
namespace LIBC_NAMESPACE_DECL {
diff --git a/libc/src/math/generic/sinpif.cpp b/libc/src/math/generic/sinpif.cpp
index 492689d594d90..f3383f19db60f 100644
--- a/libc/src/math/generic/sinpif.cpp
+++ b/libc/src/math/generic/sinpif.cpp
@@ -7,7 +7,6 @@
//===----------------------------------------------------------------------===//
#include "src/math/sinpif.h"
-#include "sincosf_utils.h"
#include "src/__support/FPUtil/FEnvImpl.h"
#include "src/__support/FPUtil/FPBits.h"
#include "src/__support/FPUtil/PolyEval.h"
@@ -15,6 +14,7 @@
#include "src/__support/common.h"
#include "src/__support/macros/config.h"
#include "src/__support/macros/optimization.h" // LIBC_UNLIKELY
+#include "src/__support/math/sincosf_utils.h"
namespace LIBC_NAMESPACE_DECL {
diff --git a/libc/src/math/generic/tanf.cpp b/libc/src/math/generic/tanf.cpp
index ca5e35dca4c91..a8c557b14808c 100644
--- a/libc/src/math/generic/tanf.cpp
+++ b/libc/src/math/generic/tanf.cpp
@@ -7,7 +7,6 @@
//===----------------------------------------------------------------------===//
#include "src/math/tanf.h"
-#include "sincosf_utils.h"
#include "src/__support/FPUtil/FEnvImpl.h"
#include "src/__support/FPUtil/FPBits.h"
#include "src/__support/FPUtil/PolyEval.h"
@@ -18,6 +17,7 @@
#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/tanpif.cpp b/libc/src/math/generic/tanpif.cpp
index 58d46c9481aa5..b49f3cea4911b 100644
--- a/libc/src/math/generic/tanpif.cpp
+++ b/libc/src/math/generic/tanpif.cpp
@@ -7,7 +7,6 @@
//===----------------------------------------------------------------------===//
#include "src/math/tanpif.h"
-#include "sincosf_utils.h"
#include "src/__support/FPUtil/FEnvImpl.h"
#include "src/__support/FPUtil/FPBits.h"
#include "src/__support/FPUtil/cast.h"
@@ -16,6 +15,7 @@
#include "src/__support/common.h"
#include "src/__support/macros/config.h"
#include "src/__support/macros/optimization.h" // LIBC_UNLIKELY
+#include "src/__support/math/sincosf_utils.h"
namespace LIBC_NAMESPACE_DECL {
diff --git a/utils/bazel/llvm-project-overlay/libc/BUILD.bazel b/utils/bazel/llvm-project-overlay/libc/BUILD.bazel
index 40f7c168e75ab..0f1c5acb9e15c 100644
--- a/utils/bazel/llvm-project-overlay/libc/BUILD.bazel
+++ b/utils/bazel/llvm-project-overlay/libc/BUILD.bazel
@@ -1952,31 +1952,6 @@ libc_support_library(
],
)
-libc_support_library(
- name = "range_reduction",
- hdrs = [
- "src/math/generic/range_reduction.h",
- "src/math/generic/range_reduction_fma.h",
- ],
- deps = [
- ":__support_common",
- ":__support_fputil_fma",
- ":__support_fputil_fp_bits",
- ":__support_fputil_multiply_add",
- ":__support_fputil_nearest_integer",
- ],
-)
-
-libc_support_library(
- name = "sincosf_utils",
- hdrs = ["src/math/generic/sincosf_utils.h"],
- deps = [
- ":__support_fputil_fp_bits",
- ":__support_fputil_polyeval",
- ":range_reduction",
- ],
-)
-
libc_support_library(
name = "sincosf16_utils",
hdrs = ["src/math/generic/sincosf16_utils.h"],
@@ -2402,6 +2377,18 @@ libc_support_library(
],
)
+libc_support_library(
+ name = "__support_math_cosf",
+ hdrs = ["src/__support/math/cosf.h"],
+ deps = [
+ ":__support_fputil_except_value_utils",
+ ":__support_macros_optimization",
+ ":__support_macros_properties_cpu_features",
+ ":__support_sincosf_utils",
+ ":errno",
+ ],
+)
+
libc_support_library(
name = "__support_math_erff",
hdrs = ["src/__support/math/erff.h"],
@@ -2675,6 +2662,21 @@ libc_support_library(
],
)
+libc_support_library(
+ name = "__support_range_reduction",
+ hdrs = [
+ "src/__support/math/range_reduction.h",
+ "src/__support/math/range_reduction_fma.h",
+ ],
+ deps = [
+ ":__support_common",
+ ":__support_fputil_fma",
+ ":__support_fputil_fp_bits",
+ ":__support_fputil_multiply_add",
+ ":__support_fputil_nearest_integer",
+ ],
+)
+
libc_support_library(
name = "__support_sincos_eval",
hdrs = [
@@ -2691,6 +2693,16 @@ libc_support_library(
],
)
+libc_support_library(
+ name = "__support_sincosf_utils",
+ hdrs = ["src/__support/math/sincosf_utils.h"],
+ deps = [
+ ":__support_fputil_fp_bits",
+ ":__support_fputil_polyeval",
+ ":__support_range_reduction",
+ ],
+)
+
############################### complex targets ################################
libc_function(
@@ -3147,11 +3159,7 @@ libc_math_function(
libc_math_function(
name = "cosf",
additional_deps = [
- ":__support_fputil_fma",
- ":__support_macros_optimization",
- ":__support_macros_properties_cpu_features",
- ":sincosf_utils",
- ":errno",
+ ":__support_math_cosf",
],
)
@@ -3196,7 +3204,7 @@ libc_math_function(
":__support_macros_optimization",
":common_constants",
":explogxf",
- ":sincosf_utils",
+ ":__support_sincosf_utils",
],
)
@@ -4273,8 +4281,8 @@ libc_math_function(
":__support_fputil_rounding_mode",
":__support_macros_optimization",
":__support_macros_properties_cpu_features",
- ":range_reduction",
- ":sincosf_utils",
+ ":__support_range_reduction",
+ ":__support_sincosf_utils",
],
)
@@ -4306,7 +4314,7 @@ libc_math_function(
":__support_fputil_rounding_mode",
":__support_macros_optimization",
":__support_macros_properties_cpu_features",
- ":sincosf_utils",
+ ":__support_sincosf_utils",
],
)
@@ -4334,7 +4342,7 @@ libc_math_function(
libc_math_function(
name = "sinpif",
additional_deps = [
- ":sincosf_utils",
+ ":__support_sincosf_utils",
],
)
@@ -4402,8 +4410,8 @@ libc_math_function(
":__support_fputil_polyeval",
":__support_macros_optimization",
":__support_macros_properties_cpu_features",
- ":range_reduction",
- ":sincosf_utils",
+ ":__support_range_reduction",
+ ":__support_sincosf_utils",
],
)
@@ -4441,7 +4449,7 @@ libc_math_function(
libc_math_function(
name = "tanpif",
additional_deps = [
- ":sincosf_utils",
+ ":__support_sincosf_utils",
":hdr_fenv_macros",
":__support_macros_config",
":__support_macros_optimization",
More information about the llvm-branch-commits
mailing list