[libc] Implement cosf function that is correctly rounded to all rounding modes.

[lntue]

Summary

Implement cosf function that is correctly rounded to all rounding
modes.

Performance benchmark using perf tool from CORE-MATH project

(https://gitlab.inria.fr/core-math/core-math/-/tree/master) on Ryzen 1700:
Before this patch (not correctly rounded):

$ CORE_MATH_PERF_MODE="rdtsc" ./perf.sh cosf
CORE-MATH reciprocal throughput   : 19.043
System LIBC reciprocal throughput : 26.328
LIBC reciprocal throughput        : 30.955

$ CORE_MATH_PERF_MODE="rdtsc" ./perf.sh cosf --latency
GNU libc version: 2.31
GNU libc release: stable
CORE-MATH latency   : 49.995
System LIBC latency : 59.286
LIBC latency        : 60.174

After this patch (correctly rounded):

$ CORE_MATH_PERF_MODE="rdtsc" ./perf.sh cosf
GNU libc version: 2.31
GNU libc release: stable
CORE-MATH reciprocal throughput   : 19.072
System LIBC reciprocal throughput : 26.286
LIBC reciprocal throughput        : 13.631

$ CORE_MATH_PERF_MODE="rdtsc" ./perf.sh cosf --latency
GNU libc version: 2.31
GNU libc release: stable
CORE-MATH latency   : 49.872
System LIBC latency : 59.468
LIBC latency        : 56.119

[orex]

libc/test/src/math/exhaustive/cosf_test.cpp

82

Why the comments below?

[lntue]

Re-enable negative part of the exhaustive test.

libc/test/src/math/exhaustive/cosf_test.cpp

82

Oops, I disabled and forgot to re-enable it.

[zimmermann6]

this looks all good to me:

GNU libc version: 2.33
GNU libc release: release
17.271
25.064
13.555
GNU libc version: 2.33
GNU libc release: release
48.048
58.428
54.403

The first figures are for reciprocal throughput (core-math, glibc, llvm-libc), the second ones are for the latency. Great work!

libc/docs/math.rst

 ============== ================ =============== ======================
 
 Accuracy of Higher Math Functions
 =================================
 
 ============== ================ =============== ======================
 <Func>         <Func_f> (float) <Func> (double) <Func_l> (long double)
 ============== ================ =============== ======================
-cos            0.776 ULPs       large
+cos            |check|          large
 exp            |check|
 exp2           |check|
 expm1          |check|
 fma            |check|          |check|
 hypot          |check|          |check|
 log            |check|
 log10          |check|
 log1p          |check|
* We also use the *perf* tool from the `CORE-MATH <https://core-math.gitlabpages.inria.fr/>`_
   table below, using the system library as reference (such as the `GNU C library <https://www.gnu.org/software/libc/>`_
   on Linux). Fmod performance results obtained with "differential_testing".
 
 +--------------+-------------------------------+-------------------------------+-------------------------------------+---------------------------------------------------------------------+
 | <Func>       | Reciprocal throughput (ns)    | Latency (ns)                  | Testing ranges                      | Testing configuration                                               |
 |              +-----------+-------------------+-----------+-------------------+                                     +------------+-------------------------+--------------+---------------+
 |              | LLVM libc | Reference (glibc) | LLVM libc | Reference (glibc) |                                     | CPU        | OS                      | Compiler     | Special flags |
 +==============+===========+===================+===========+===================+=====================================+============+=========================+==============+===============+
-| cosf         |        37 |                32 |        73 |                72 | :math:`[0, 2\pi]`                   | Ryzen 1700 | Ubuntu 20.04 LTS x86_64 | Clang 12.0.0 |               |
+| cosf         |        14 |                32 |        56 |                59 | :math:`[0, 2\pi]`                   | Ryzen 1700 | Ubuntu 20.04 LTS x86_64 | Clang 12.0.0 |               |
 +--------------+-----------+-------------------+-----------+-------------------+-------------------------------------+------------+-------------------------+--------------+---------------+
 | expf         |         9 |                 7 |        44 |                38 | :math:`[-10, 10]`                   | Ryzen 1700 | Ubuntu 20.04 LTS x86_64 | Clang 12.0.0 | FMA           |
 +--------------+-----------+-------------------+-----------+-------------------+-------------------------------------+------------+-------------------------+--------------+---------------+
 | exp2f        |         8 |                 6 |        35 |                23 | :math:`[-10, 10]`                   | i5-1135G7  | Ubuntu 20.04 LTS x86_64 | Clang 12.0.0 | FMA           |
 +--------------+-----------+-------------------+-----------+-------------------+-------------------------------------+------------+-------------------------+--------------+---------------+
 | expm1f       |         9 |                44 |        42 |               121 | :math:`[-10, 10]`                   | Ryzen 1700 | Ubuntu 20.04 LTS x86_64 | Clang 12.0.0 | FMA           |
 +--------------+-----------+-------------------+-----------+-------------------+-------------------------------------+------------+-------------------------+--------------+---------------+
 | fmodf        |        73 |               263 |        -  |                 - | [MIN_NORMAL, MAX_NORMAL]            | i5 mobile  | Ubuntu 20.04 LTS x86_64 | Clang 12.0.0 |               |

libc/src/math/generic/CMakeLists.txt

 )
 
 add_entrypoint_object(
   cosf
   SRCS
     cosf.cpp
   HDRS
     ../cosf.h
+    range_reduction.h
+    range_reduction_fma.h
   DEPENDS
-    .sincosf_utils
+    .common_constants
     libc.include.math
     libc.src.errno.errno
+    libc.src.__support.FPUtil.fputil
+    libc.src.__support.FPUtil.fma
+    libc.src.__support.FPUtil.multiply_add
+    libc.src.__support.FPUtil.nearest_integer
+    libc.src.__support.FPUtil.polyeval
   COMPILE_OPTIONS
     -O3
 )
 
 add_entrypoint_object(
   sinf
   SRCS
     sinf.cpp

libc/src/math/generic/cosf.cpp

 //===-- Single-precision cos 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/cosf.h"
-#include "math_utils.h"
-#include "sincosf_utils.h"
-
+#include "common_constants.h"
+#include "src/__support/FPUtil/BasicOperations.h"
+#include "src/__support/FPUtil/FEnvImpl.h"
+#include "src/__support/FPUtil/FPBits.h"
+#include "src/__support/FPUtil/PolyEval.h"
+#include "src/__support/FPUtil/except_value_utils.h"
+#include "src/__support/FPUtil/multiply_add.h"
 #include "src/__support/common.h"
-#include <math.h>
-
-#include <stdint.h>
-
-namespace __llvm_libc {
-
-// Fast cosf implementation. Worst-case ULP is 0.5607, maximum relative
-// error is 0.5303 * 2^-23. A single-step range reduction is used for
-// small values. Large inputs have their range reduced using fast integer
-// arithmetic.
-LLVM_LIBC_FUNCTION(float, cosf, (float y)) {
-  double x = y;
-  double s;
-  int n;
-  const sincos_t *p = &SINCOSF_TABLE[0];
-
-  if (abstop12(y) < abstop12(PIO4)) {
-    double x2 = x * x;
-
-    if (unlikely(abstop12(y) < abstop12(as_float(0x39800000))))
-      return 1.0f;
 
-    return sinf_poly(x, x2, p, 1);
-  } else if (likely(abstop12(y) < abstop12(120.0f))) {
-    x = reduce_fast(x, p, &n);
+#include <errno.h>
 
-    // Setup the signs for sin and cos.
-    s = p->sign[n & 3];
+#if defined(LIBC_TARGET_HAS_FMA)
+#include "range_reduction_fma.h"
+// using namespace __llvm_libc::fma;
+using __llvm_libc::fma::FAST_PASS_BOUND;
+using __llvm_libc::fma::large_range_reduction;
+using __llvm_libc::fma::small_range_reduction;
+#else
+#include "range_reduction.h"
+// using namespace __llvm_libc::generic;
+using __llvm_libc::generic::FAST_PASS_BOUND;
+using __llvm_libc::generic::large_range_reduction;
+using __llvm_libc::generic::small_range_reduction;
+#endif
 
-    if (n & 2)
-      p = &SINCOSF_TABLE[1];
+namespace __llvm_libc {
 
-    return sinf_poly(x * s, x * x, p, n ^ 1);
-  } else if (abstop12(y) < abstop12(INFINITY)) {
-    uint32_t xi = as_uint32_bits(y);
-    int sign = xi >> 31;
+// Exceptional cases for cosf.
+static constexpr int COSF_EXCEPTS = 6;
 
-    x = reduce_large(xi, &n);
+static constexpr fputil::ExceptionalValues<float, COSF_EXCEPTS> CosfExcepts{
+    /* inputs */ {
+        0x55325019, // x = 0x1.64a032p43
+        0x5922aa80, // x = 0x1.4555p51
+        0x5aa4542c, // x = 0x1.48a858p54
+        0x5f18b878, // x = 0x1.3170fp63
+        0x6115cb11, // x = 0x1.2b9622p67
+        0x7beef5ef, // x = 0x1.ddebdep120
+    },
+    /* outputs (RZ, RU offset, RD offset, RN offset) */
+    {
+        {0x3f4ea5d2, 1, 0, 0}, // x = 0x1.64a032p43, cos(x) = 0x1.9d4ba4p-1 (RZ)
+        {0x3f08aebe, 1, 0, 1}, // x = 0x1.4555p51, cos(x) = 0x1.115d7cp-1 (RZ)
+        {0x3efa40a4, 1, 0, 0}, // x = 0x1.48a858p54, cos(x) = 0x1.f48148p-2 (RZ)
+        {0x3f7f14bb, 1, 0, 0}, // x = 0x1.3170fp63, cos(x) = 0x1.fe2976p-1 (RZ)
+        {0x3f78142e, 1, 0, 1}, // x = 0x1.2b9622p67, cos(x) = 0x1.f0285cp-1 (RZ)
+        {0x3f08a21c, 1, 0,
+         0}, // x = 0x1.ddebdep120, cos(x) = 0x1.114438p-1 (RZ)
+    }};
+
+LLVM_LIBC_FUNCTION(float, cosf, (float x)) {
+  using FPBits = typename fputil::FPBits<float>;
+  FPBits xbits(x);
+  xbits.set_sign(false);
+
+  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/16
+  //   k is an integer
+  //   |y| < 0.5
+  // For small range (|x| < 2^46 when FMA instructions are available, 2^22
+  // otherwise), this is done by performing:
+  //   k = round(x * 16/pi)
+  //   y = x * 16/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 31,
+  // since cos((k + y + 32*i) * pi/16) = cos(x + i * 2pi) = cos(x).
+  //
+  // When FMA instructions are not available, we store the digits of 16/pi in
+  // chunks of 28-bit precision.  This will make sure that the products:
+  //   x * SIXTEEN_OVER_PI_28[i] are all exact.
+  // When FMA instructions are available, we simply store the digits of 16/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/16)
+  //          = cos(y*pi/16) * cos(k*pi/16) - sin(y*pi/16) * sin(k*pi/16)
+  // The values of sin(k*pi/16) and cos(k*pi/16) for k = 0..31 are precomputed
+  // and stored using a vector of 32 doubles. Sin(y*pi/16) and cos(y*pi/16) are
+  // computed using degree-7 and degree-8 minimax polynomials generated by
+  // Sollya respectively.
+
+  // |x| < 0x1.0p-12f
+  if (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_HAS_FMA)
+    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_HAS_FMA
+  }
 
-    // Setup signs for sin and cos - include original sign.
-    s = p->sign[(n + sign) & 3];
+  using ExceptChecker = typename fputil::ExceptionChecker<float, COSF_EXCEPTS>;
+  {
+    float result;
+    if (ExceptChecker::check_odd_func(CosfExcepts, x_abs, false, result))
+      return result;
+  }
 
-    if ((n + sign) & 2)
-      p = &SINCOSF_TABLE[1];
+  // TODO(lntue): refactor range reduction and most of polynomial approximation
+  // to share between sinf, cosf, and sincosf.
+  int k;
+  double y;
+
+  if (likely(x_abs < FAST_PASS_BOUND)) {
+    k = small_range_reduction(xd, y);
+  } else {
+    // x is inf or nan.
+    if (unlikely(x_abs >= 0x7f80'0000U)) {
+      if (x_abs == 0x7f80'0000U) {
+        errno = EDOM;
+        fputil::set_except(FE_INVALID);
+      }
+      return x +
+             FPBits::build_nan(1 << (fputil::MantissaWidth<float>::VALUE - 1));
+    }
 
-    return sinf_poly(x * s, x * x, p, n ^ 1);
+    k = large_range_reduction(xd, xbits.get_exponent(), y);
   }
 
-  return invalid(y);
+  // After range reduction, k = round(x * 16 / pi) and y = (x * 16 / pi) - k.
+  // So k is an integer and -0.5 <= y <= 0.5.
+  // Then cos(x) = cos((k + y)*pi/16)
+  //             = cos(y*pi/16) * cos(k*pi/16) - sin(y*pi/16) * sin(k*pi/16)
+
+  double ysq = y * y;
+
+  // Degree-6 minimax even polynomial for sin(y*pi/16)/y generated by Sollya
+  // with:
+  // > Q = fpminimax(sin(y*pi/16)/y, [|0, 2, 4, 6|], [|D...|], [0, 0.5]);
+  double sin_y =
+      y * fputil::polyeval(ysq, 0x1.921fb54442d17p-3, -0x1.4abbce6256adp-10,
+                           0x1.466bc5a5ac6b3p-19, -0x1.32bdcb4207562p-29);
+  // Degree-8 minimax even polynomial for cos(y*pi/16) generated by Sollya with:
+  // > P = fpminimax(cos(x*pi/16), [|0, 2, 4, 6, 8|], [|1, D...|], [0, 0.5]);
+  // Note that cosm1_y = cos(y*pi/16) - 1.
+  double cosm1_y =
+      ysq * fputil::polyeval(ysq, -0x1.3bd3cc9be45dcp-6, 0x1.03c1f081b08ap-14,
+                             -0x1.55d3c6fb0fb6ep-24, 0x1.e1d3d60f58873p-35);
+
+  double sin_k = -SIN_K_PI_OVER_16[k & 31];
+  // cos(k * pi/16) = sin(k * pi/16 + pi/2) = sin((k + 8) * pi/16).
+  // cos_k = y * cos(k * pi/16)
+  double cos_k = SIN_K_PI_OVER_16[(k + 8) & 31];
+
+  // Combine the results with the sine of sum formula:
+  //   cos(x) = cos((k + y)*pi/16)
+  //          = cos(y*pi/16) * cos(k*pi/16) - sin(y*pi/16) * sin(k*pi/16)
+  //          = cosm1_y * cos_k + sin_y * sin_k
+  //          = (cosm1_y * cos_k + cos_k) + sin_y * sin_k
+  return fputil::multiply_add(sin_y, sin_k,
+                              fputil::multiply_add(cosm1_y, cos_k, cos_k));
 }
 
 } // namespace __llvm_libc

libc/test/src/math/cosf_test.cpp

 
 TEST(LlvmLibcCosfTest, InFloatRange) {
   constexpr uint32_t COUNT = 1000000;
   constexpr uint32_t STEP = UINT32_MAX / COUNT;
   for (uint32_t i = 0, v = 0; i <= COUNT; ++i, v += STEP) {
     float x = float(FPBits(v));
     if (isnan(x) || isinf(x))
       continue;
-    ASSERT_MPFR_MATCH(mpfr::Operation::Cos, x, __llvm_libc::cosf(x), 1.0);
+    ASSERT_MPFR_MATCH_ALL_ROUNDING(mpfr::Operation::Cos, x,
+                                   __llvm_libc::cosf(x), 0.5);
   }
 }
 
-// For small values, cos(x) is 1.
-TEST(LlvmLibcCosfTest, SmallValues) {
-  float x = float(FPBits(0x17800000U));
-  float result = __llvm_libc::cosf(x);
-  EXPECT_MPFR_MATCH(mpfr::Operation::Cos, x, result, 1.0);
-  EXPECT_FP_EQ(1.0f, result);
-
-  x = float(FPBits(0x0040000U));
-  result = __llvm_libc::cosf(x);
-  EXPECT_MPFR_MATCH(mpfr::Operation::Cos, x, result, 1.0);
-  EXPECT_FP_EQ(1.0f, result);
+TEST(LlvmLibcCosfTest, SpecificBitPatterns) {
+  constexpr int N = 42;
+  constexpr uint32_t INPUTS[N] = {
+      0x3f06'0a92U, // x = pi/6
+      0x3f3a'dc51U, // x = 0x1.75b8a2p-1f
+      0x3f49'0fdbU, // x = pi/4
+      0x3f86'0a92U, // x = pi/3
+      0x3fa7'832aU, // x = 0x1.4f0654p+0f
+      0x3fc9'0fdbU, // x = pi/2
+      0x4017'1973U, // x = 0x1.2e32e6p+1f
+      0x4049'0fdbU, // x = pi
+      0x4096'cbe4U, // x = 0x1.2d97c8p+2f
+      0x40c9'0fdbU, // x = 2*pi
+      0x433b'7490U, // x = 0x1.76e92p+7f
+      0x437c'e5f1U, // x = 0x1.f9cbe2p+7f
+      0x4619'9998U, // x = 0x1.33333p+13f
+      0x474d'246fU, // x = 0x1.9a48dep+15f
+      0x4afd'ece4U, // x = 0x1.fbd9c8p+22f
+      0x4c23'32e9U, // x = 0x1.4665d2p+25f
+      0x50a3'e87fU, // x = 0x1.47d0fep+34f
+      0x5239'47f6U, // x = 0x1.728fecp+37f
+      0x53b1'46a6U, // x = 0x1.628d4cp+40f
+      0x5532'5019U, // x = 0x1.64a032p+43f
+      0x55ca'fb2aU, // x = 0x1.95f654p+44f
+      0x588e'f060U, // x = 0x1.1de0cp+50f
+      0x5922'aa80U, // x = 0x1.4555p+51f
+      0x5aa4'542cU, // x = 0x1.48a858p+54f
+      0x5c07'bcd0U, // x = 0x1.0f79ap+57f
+      0x5ebc'fddeU, // x = 0x1.79fbbcp+62f
+      0x5f18'b878U, // x = 0x1.3170fp+63f
+      0x5fa6'eba7U, // x = 0x1.4dd74ep+64f
+      0x6115'cb11U, // x = 0x1.2b9622p+67f
+      0x61a4'0b40U, // x = 0x1.48168p+68f
+      0x6386'134eU, // x = 0x1.0c269cp+72f
+      0x6589'8498U, // x = 0x1.13093p+76f
+      0x6600'0001U, // x = 0x1.000002p+77f
+      0x664e'46e4U, // x = 0x1.9c8dc8p+77f
+      0x66b0'14aaU, // x = 0x1.602954p+78f
+      0x67a9'242bU, // x = 0x1.524856p+80f
+      0x6a19'76f1U, // x = 0x1.32ede2p+85f
+      0x6c55'da58U, // x = 0x1.abb4bp+89f
+      0x6f79'be45U, // x = 0x1.f37c8ap+95f
+      0x7276'69d4U, // x = 0x1.ecd3a8p+101f
+      0x7758'4625U, // x = 0x1.b08c4ap+111f
+      0x7bee'f5efU, // x = 0x1.ddebdep+120f
+  };
+
+  for (int i = 0; i < N; ++i) {
+    float x = float(FPBits(INPUTS[i]));
+    EXPECT_MPFR_MATCH_ALL_ROUNDING(mpfr::Operation::Cos, x,
+                                   __llvm_libc::cosf(x), 0.5);
+    EXPECT_MPFR_MATCH_ALL_ROUNDING(mpfr::Operation::Cos, -x,
+                                   __llvm_libc::cosf(-x), 0.5);
+  }
 }
 
 // SDCOMP-26094: check cosf in the cases for which the range reducer
 // returns values furthest beyond its nominal upper bound of pi/4.
 TEST(LlvmLibcCosfTest, SDCOMP_26094) {
   for (uint32_t v : SDCOMP26094_VALUES) {
     float x = float(FPBits(v));
-    ASSERT_MPFR_MATCH(mpfr::Operation::Cos, x, __llvm_libc::cosf(x), 1.0);
+    ASSERT_MPFR_MATCH(mpfr::Operation::Cos, x, __llvm_libc::cosf(x), 0.5);
   }
 }

libc/test/src/math/exhaustive/CMakeLists.txt

DEPENDS
     libc.src.math.sinf
     libc.src.__support.FPUtil.fputil
   LINK_LIBRARIES
     -lpthread
 )
 
 add_fp_unittest(
   cosf_test
+  NO_RUN_POSTBUILD
   NEED_MPFR
   SUITE
     libc_math_exhaustive_tests
   SRCS
     cosf_test.cpp
   DEPENDS
+    .exhaustive_test
     libc.include.math
     libc.src.math.cosf
     libc.src.__support.FPUtil.fputil
+  LINK_LIBRARIES
+    -lpthread
 )
 
 add_fp_unittest(
   expf_test
   NO_RUN_POSTBUILD
   NEED_MPFR
   SUITE
     libc_math_exhaustive_tests

libc/test/src/math/exhaustive/cosf_test.cpp

 //===-- Exhaustive test for cosf ------------------------------------------===//
 //
 // 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 "exhaustive_test.h"
 #include "src/__support/FPUtil/FPBits.h"
 #include "src/math/cosf.h"
 #include "utils/MPFRWrapper/MPFRUtils.h"
-#include <math.h>
+#include "utils/UnitTest/FPMatcher.h"
+
+#include <thread>
 
 using FPBits = __llvm_libc::fputil::FPBits<float>;
 
 namespace mpfr = __llvm_libc::testing::mpfr;
 
-TEST(LlvmLibccosffExhaustiveTest, AllValues) {
-  uint32_t bits = 0;
+struct LlvmLibcCosfExhaustiveTest : public LlvmLibcExhaustiveTest<uint32_t> {
+  bool check(uint32_t start, uint32_t stop,
+             mpfr::RoundingMode rounding) override {
+    mpfr::ForceRoundingMode r(rounding);
+    uint32_t bits = start;
+    bool result = true;
     do {
       FPBits xbits(bits);
       float x = float(xbits);
-    ASSERT_MPFR_MATCH(mpfr::Operation::Cos, x, __llvm_libc::cosf(x), 1.0);
-  } while (bits++ < 0xffff'ffffU);
+      bool r = EXPECT_MPFR_MATCH(mpfr::Operation::Cos, x, __llvm_libc::cosf(x),
+                                 0.5, rounding);
+      result &= r;
+    } while (++bits < stop);
+    return result;
+  }
+};
+
+// Range: [0, +Inf);
+static constexpr uint32_t POS_START = 0x0000'0000U;
+static constexpr uint32_t POS_STOP = 0x7f80'0000U;
+
+TEST_F(LlvmLibcCosfExhaustiveTest, PostiveRangeRoundNearestTieToEven) {
+  test_full_range(POS_START, POS_STOP, mpfr::RoundingMode::Nearest);
+}
+
+TEST_F(LlvmLibcCosfExhaustiveTest, PostiveRangeRoundUp) {
+  test_full_range(POS_START, POS_STOP, mpfr::RoundingMode::Upward);
+}
+
+TEST_F(LlvmLibcCosfExhaustiveTest, PostiveRangeRoundDown) {
+  test_full_range(POS_START, POS_STOP, mpfr::RoundingMode::Downward);
+}
+
+TEST_F(LlvmLibcCosfExhaustiveTest, PostiveRangeRoundTowardZero) {
+  test_full_range(POS_START, POS_STOP, mpfr::RoundingMode::TowardZero);
+}
+
+// Range: (-Inf, 0];
+static constexpr uint32_t NEG_START = 0x8000'0000U;
+static constexpr uint32_t NEG_STOP = 0xff80'0000U;
+
+TEST_F(LlvmLibcCosfExhaustiveTest, NegativeRangeRoundNearestTieToEven) {
+  test_full_range(NEG_START, NEG_STOP, mpfr::RoundingMode::Nearest);
+}
+
+TEST_F(LlvmLibcCosfExhaustiveTest, NegativeRangeRoundUp) {
+  test_full_range(NEG_START, NEG_STOP, mpfr::RoundingMode::Upward);
+}
+
+TEST_F(LlvmLibcCosfExhaustiveTest, NegativeRangeRoundDown) {
+  test_full_range(NEG_START, NEG_STOP, mpfr::RoundingMode::Downward);
+}
+
+TEST_F(LlvmLibcCosfExhaustiveTest, NegativeRangeRoundTowardZero) {
+  test_full_range(NEG_START, NEG_STOP, mpfr::RoundingMode::TowardZero);
 }

utils/bazel/llvm-project-overlay/libc/BUILD.bazel

 
 libc_math_function(name = "fmaxf")
 
 libc_math_function(name = "fmaxl")
 
 libc_math_function(
     name = "cosf",
     additional_deps = [
-        ":math_utils",
-        ":sincosf_utils",
+        ":__support_fputil_fma",
+        ":__support_fputil_multiply_add",
+        ":__support_fputil_polyeval",
+        ":common_constants",
+        ":range_reduction",
     ],
 )
 
 libc_math_function(
     name = "sincosf",
     additional_deps = [
         ":math_utils",
         ":sincosf_utils",