Merge pull request #1389 from boostorg/1387

Fix cohort handling of zero to match IEEE 754 requirements

Author: mborland

.drone.jsonnet

@@ -259,6 +259,34 @@ local windows_pipeline(name, image, environment, arch = "amd64") =
         "g++-14-multilib",
     ),
 
+    linux_pipeline(
+        "Linux 26.04 GCC 15 32",
+        "cppalliance/droneubuntu2604:1",
+        { TOOLSET: 'gcc', COMPILER: 'g++-15', CXXSTD: '03,11,14,17,20,23', ADDRMD: '32', CXXFLAGS: "-fexcess-precision=fast" },
+        "g++-15-multilib",
+    ),
+
+    linux_pipeline(
+        "Linux 26.04 GCC 15 64",
+        "cppalliance/droneubuntu2604:1",
+        { TOOLSET: 'gcc', COMPILER: 'g++-15', CXXSTD: '03,11,14,17,20,23', ADDRMD: '64', CXXFLAGS: "-fexcess-precision=fast" },
+        "g++-15-multilib",
+    ),
+
+    linux_pipeline(
+        "Linux 26.04 GCC 16 32",
+        "cppalliance/droneubuntu2604:1",
+        { TOOLSET: 'gcc', COMPILER: 'g++-16', CXXSTD: '03,11,14,17,20,23', ADDRMD: '32', CXXFLAGS: "-fexcess-precision=fast" },
+        "g++-16-multilib",
+    ),
+
+    linux_pipeline(
+        "Linux 26.04 GCC 16 64",
+        "cppalliance/droneubuntu2604:1",
+        { TOOLSET: 'gcc', COMPILER: 'g++-16', CXXSTD: '03,11,14,17,20,23', ADDRMD: '64', CXXFLAGS: "-fexcess-precision=fast" },
+        "g++-16-multilib",
+    ),
+
     linux_pipeline(
         "Linux 18.04 Clang 6.0",
         "cppalliance/droneubuntu1804:1",
@@ -316,16 +344,9 @@ local windows_pipeline(name, image, environment, arch = "amd64") =
     ),
 
     linux_pipeline(
-        "Linux 22.04 Clang 14 UBSAN",
+        "Linux 22.04 Clang 14",
         "cppalliance/droneubuntu2204:1",
-        { TOOLSET: 'clang', COMPILER: 'clang++-14', CXXSTD: '03,11,14,17,20,2b' } + ubsan,
-        "clang-14",
-    ),
-
-    linux_pipeline(
-        "Linux 22.04 Clang 14 ASAN",
-        "cppalliance/droneubuntu2204:1",
-        { TOOLSET: 'clang', COMPILER: 'clang++-14', CXXSTD: '03,11,14,17,20,2b' } + asan,
+        { TOOLSET: 'clang', COMPILER: 'clang++-14', CXXSTD: '03,11,14,17,20,2b' },
         "clang-14",
     ),
 
@@ -350,31 +371,55 @@ local windows_pipeline(name, image, environment, arch = "amd64") =
         "cppalliance/droneubuntu2404:1",
         { TOOLSET: 'clang', COMPILER: 'clang++-17', CXXSTD: '03,11,14,17,20,2b' },
         "clang-17",
-        ["deb http://apt.llvm.org/jammy/ llvm-toolchain-jammy-17 main"],
+        ["deb http://apt.llvm.org/noble/ llvm-toolchain-noble-17 main"],
     ),
 
     linux_pipeline(
         "Linux 24.04 Clang 18",
         "cppalliance/droneubuntu2404:1",
         { TOOLSET: 'clang', COMPILER: 'clang++-18', CXXSTD: '03,11,14,17,20,2b' },
         "clang-18",
-        ["deb http://apt.llvm.org/jammy/ llvm-toolchain-jammy-18 main"],
+        ["deb http://apt.llvm.org/noble/ llvm-toolchain-noble-18 main"],
     ),
 
     linux_pipeline(
         "Linux 24.04 Clang 19",
         "cppalliance/droneubuntu2404:1",
         { TOOLSET: 'clang', COMPILER: 'clang++-19', CXXSTD: '03,11,14,17,20,2b' },
         "clang-19",
-        ["deb http://apt.llvm.org/jammy/ llvm-toolchain-jammy-19 main"],
+        ["deb http://apt.llvm.org/noble/ llvm-toolchain-noble-19 main"],
     ),
 
     linux_pipeline(
         "Linux 24.04 Clang 20",
         "cppalliance/droneubuntu2404:1",
         { TOOLSET: 'clang', COMPILER: 'clang++-20', CXXSTD: '03,11,14,17,20,2b' },
         "clang-20",
-        ["deb http://apt.llvm.org/jammy/ llvm-toolchain-jammy-20 main"],
+        ["deb http://apt.llvm.org/noble/ llvm-toolchain-noble-20 main"],
+    ),
+
+    linux_pipeline(
+        "Linux 24.04 Clang 21",
+        "cppalliance/droneubuntu2404:1",
+        { TOOLSET: 'clang', COMPILER: 'clang++-21', CXXSTD: '14,17,20,2b' },
+        "clang-21",
+        ["deb http://apt.llvm.org/noble/ llvm-toolchain-noble-21 main"],
+    ),
+
+    linux_pipeline(
+        "Linux 24.04 Clang 21 UBSAN",
+        "cppalliance/droneubuntu2404:1",
+        { TOOLSET: 'clang', COMPILER: 'clang++-21', CXXSTD: '14,17,20,2b' } + ubsan,
+        "clang-21",
+        ["deb http://apt.llvm.org/noble/ llvm-toolchain-noble-21 main"],
+    ),
+
+    linux_pipeline(
+        "Linux 24.04 Clang 21 ASAN",
+        "cppalliance/droneubuntu2404:1",
+        { TOOLSET: 'clang', COMPILER: 'clang++-21', CXXSTD: '14,17,20,2b' } + asan,
+        "clang-21",
+        ["deb http://apt.llvm.org/noble/ llvm-toolchain-noble-21 main"],
     ),
 
     windows_pipeline(

include/boost/decimal/charconv.hpp

@@ -1071,7 +1071,7 @@ BOOST_DECIMAL_CUDA_CONSTEXPR auto to_chars_cohort_preserving_scientific(char* fi
     using unsigned_integer = typename TargetDecimalType::significand_type;
 
     const auto fp = fpclassify(value);
-    if (!(fp == FP_NORMAL || fp == FP_SUBNORMAL))
+    if (!(fp == FP_NORMAL || fp == FP_SUBNORMAL || fp == FP_ZERO))
     {
         // Cohorts are irrelevant for non-finite values
         return to_chars_nonfinite(first, last, value, fp, chars_format::scientific, -1);

include/boost/decimal/decimal128_t.hpp

@@ -725,6 +725,18 @@ BOOST_DECIMAL_CUDA_CONSTEXPR decimal128_t::decimal128_t(T1 coeff, T2 exp, const
 
     if (reduced_coeff == zero)
     {
+        // IEEE 754-2008 3.5.1: zero has a cohort with one representation per exponent.
+        // Clamp the requested exponent to the representable range and encode it; sign was already set.
+        auto zero_biased_exp {biased_exp};
+        if (zero_biased_exp < 0)
+        {
+            zero_biased_exp = 0;
+        }
+        else if (zero_biased_exp > static_cast<int>(detail::d128_max_biased_exponent))
+        {
+            zero_biased_exp = static_cast<int>(detail::d128_max_biased_exponent);
+        }
+        bits_.high |= (static_cast<std::uint64_t>(zero_biased_exp) << detail::d128_not_11_exp_high_word_shift) & detail::d128_not_11_exp_mask;
         return;
     }
 

include/boost/decimal/decimal32_t.hpp

@@ -642,7 +642,18 @@ BOOST_DECIMAL_CUDA_CONSTEXPR decimal32_t::decimal32_t(T1 coeff, T2 exp, const de
 
     if (reduced_coeff == 0U)
     {
-        // Normalize our handling of zeros
+        // IEEE 754-2008 3.5.1: zero has a cohort with one representation per exponent.
+        // Clamp the requested exponent to the representable range and encode it; sign was already set.
+        auto zero_biased_exp {biased_exp};
+        if (zero_biased_exp < 0)
+        {
+            zero_biased_exp = 0;
+        }
+        else if (zero_biased_exp > static_cast<int>(detail::d32_max_biased_exponent))
+        {
+            zero_biased_exp = static_cast<int>(detail::d32_max_biased_exponent);
+        }
+        bits_ |= (static_cast<std::uint32_t>(zero_biased_exp) << detail::d32_not_11_exp_shift) & detail::d32_not_11_exp_mask;
         return;
     }
 

include/boost/decimal/decimal64_t.hpp

@@ -636,7 +636,18 @@ BOOST_DECIMAL_CUDA_CONSTEXPR decimal64_t::decimal64_t(T1 coeff, T2 exp, const de
 
     if (reduced_coeff == 0U)
     {
-        // Normalize our handling of zeros
+        // IEEE 754-2008 3.5.1: zero has a cohort with one representation per exponent.
+        // Clamp the requested exponent to the representable range and encode it; sign was already set.
+        auto zero_biased_exp {biased_exp};
+        if (zero_biased_exp < 0)
+        {
+            zero_biased_exp = 0;
+        }
+        else if (zero_biased_exp > static_cast<int>(detail::d64_max_biased_exponent))
+        {
+            zero_biased_exp = static_cast<int>(detail::d64_max_biased_exponent);
+        }
+        bits_ |= (static_cast<std::uint64_t>(zero_biased_exp) << detail::d64_not_11_exp_shift) & detail::d64_not_11_exp_mask;
         return;
     }
 

include/boost/decimal/decimal_fast128_t.hpp

@@ -522,13 +522,32 @@ constexpr decimal_fast128_t::decimal_fast128_t(T1 coeff, T2 exp, const detail::c
     const auto is_negative {static_cast<bool>(resultant_sign)};
     sign_ = is_negative;
 
+    // IEEE 754-2008 3.5.1: zero has a cohort with one representation per exponent.
+    // Skip normalization for zero (which would otherwise expand the significand and shift the exponent)
+    // and clamp the requested exponent to the representable range.
+    if (min_coeff == minimum_coefficient_size{0})
+    {
+        significand_ = static_cast<significand_type>(0);
+        auto biased_exp {static_cast<int>(exp) + detail::bias_v<decimal_fast128_t>};
+        if (biased_exp < 0)
+        {
+            biased_exp = 0;
+        }
+        else if (biased_exp > detail::max_biased_exp_v<decimal_fast128_t>)
+        {
+            biased_exp = detail::max_biased_exp_v<decimal_fast128_t>;
+        }
+        exponent_ = static_cast<exponent_type>(biased_exp);
+        return;
+    }
+
     // Normalize the significand in the constructor, so we don't have
     // to calculate the number of digits for operations
     detail::normalize<decimal_fast128_t>(min_coeff, exp, is_negative);
 
     significand_ = static_cast<significand_type>(min_coeff);
 
-    const auto biased_exp {significand_ == 0U ? 0 : exp + detail::bias_v<decimal_fast128_t>};
+    const auto biased_exp {static_cast<int>(exp) + detail::bias_v<decimal_fast128_t>};
 
     if (biased_exp > detail::max_biased_exp_v<decimal_fast128_t>)
     {
@@ -540,10 +559,9 @@ constexpr decimal_fast128_t::decimal_fast128_t(T1 coeff, T2 exp, const detail::c
     }
     else
     {
-        // Flush denorms to zero
+        // Flush denorms to zero, preserving sign per IEEE 754-2008 3.5.1
         significand_ = static_cast<significand_type>(0);
-        exponent_ = static_cast<exponent_type>(detail::bias_v<decimal_fast128_t>);
-        sign_ = false;
+        exponent_ = static_cast<exponent_type>(0);
     }
 }
 

include/boost/decimal/decimal_fast32_t.hpp

@@ -516,12 +516,31 @@ constexpr decimal_fast32_t::decimal_fast32_t(T1 coeff, T2 exp, const detail::con
     const auto is_negative {static_cast<bool>(resultant_sign)};
     sign_ = is_negative;
 
+    // IEEE 754-2008 3.5.1: zero has a cohort with one representation per exponent.
+    // Skip normalization for zero (which would otherwise expand the significand and shift the exponent)
+    // and clamp the requested exponent to the representable range.
+    if (min_coeff == minimum_coefficient_size{0})
+    {
+        significand_ = static_cast<significand_type>(0);
+        auto biased_exp {static_cast<int>(exp) + detail::bias};
+        if (biased_exp < 0)
+        {
+            biased_exp = 0;
+        }
+        else if (biased_exp > detail::max_biased_exp_v<decimal_fast32_t>)
+        {
+            biased_exp = detail::max_biased_exp_v<decimal_fast32_t>;
+        }
+        exponent_ = static_cast<exponent_type>(biased_exp);
+        return;
+    }
+
     // Normalize in the constructor, so we never have to worry about it again
     detail::normalize<decimal_fast32_t>(min_coeff, exp, is_negative);
 
     significand_ = static_cast<significand_type>(min_coeff);
 
-    const auto biased_exp {significand_ == 0U ? 0 : exp + detail::bias};
+    const auto biased_exp {static_cast<int>(exp) + detail::bias};
 
     // decimal32_t exponent holds 8 bits
     if (biased_exp > detail::max_biased_exp_v<decimal_fast32_t>)
@@ -534,10 +553,9 @@ constexpr decimal_fast32_t::decimal_fast32_t(T1 coeff, T2 exp, const detail::con
     }
     else
     {
-        // Flush denorms to zero
+        // Flush denorms to zero, preserving sign per IEEE 754-2008 3.5.1
         significand_ = static_cast<significand_type>(0);
-        exponent_ = static_cast<exponent_type>(detail::bias);
-        sign_ = false;
+        exponent_ = static_cast<exponent_type>(0);
     }
 }
 

include/boost/decimal/decimal_fast64_t.hpp

@@ -525,12 +525,31 @@ constexpr decimal_fast64_t::decimal_fast64_t(T1 coeff, T2 exp, const detail::con
     const auto is_negative {static_cast<bool>(resultant_sign)};
     sign_ = is_negative;
 
+    // IEEE 754-2008 3.5.1: zero has a cohort with one representation per exponent.
+    // Skip normalization for zero (which would otherwise expand the significand and shift the exponent)
+    // and clamp the requested exponent to the representable range.
+    if (min_coeff == minimum_coefficient_size{0})
+    {
+        significand_ = static_cast<significand_type>(0);
+        auto biased_exp {static_cast<int>(exp) + detail::bias_v<decimal64_t>};
+        if (biased_exp < 0)
+        {
+            biased_exp = 0;
+        }
+        else if (biased_exp > detail::max_biased_exp_v<decimal64_t>)
+        {
+            biased_exp = detail::max_biased_exp_v<decimal64_t>;
+        }
+        exponent_ = static_cast<exponent_type>(biased_exp);
+        return;
+    }
+
     // Normalize the value, so we don't have to worry about it with operations
     detail::normalize<decimal_fast64_t>(min_coeff, exp, is_negative);
 
     significand_ = static_cast<significand_type>(min_coeff);
 
-    const auto biased_exp {significand_ == 0U ? 0 : exp + detail::bias_v<decimal64_t>};
+    const auto biased_exp {static_cast<int>(exp) + detail::bias_v<decimal64_t>};
 
     if (biased_exp > detail::max_biased_exp_v<decimal64_t>)
     {
@@ -542,10 +561,9 @@ constexpr decimal_fast64_t::decimal_fast64_t(T1 coeff, T2 exp, const detail::con
     }
     else
     {
-        // Flush denorms to zero
+        // Flush denorms to zero, preserving sign per IEEE 754-2008 3.5.1
         significand_ = static_cast<significand_type>(0);
-        exponent_ = static_cast<exponent_type>(detail::bias_v<decimal64_t>);
-        sign_ = false;
+        exponent_ = static_cast<exponent_type>(0);
     }
 }
 

include/boost/decimal/detail/add_impl.hpp

@@ -42,6 +42,28 @@ BOOST_DECIMAL_CUDA_CONSTEXPR auto add_impl(const T& lhs, const T& rhs) noexcept
     auto lhs_exp {lhs.biased_exponent()};
     auto rhs_exp {rhs.biased_exponent()};
 
+    // IEEE 754-2008 3.5.1: zero has a cohort with one representation per exponent.
+    // The exponent-comparison logic below assumes both operands have full precision
+    // (which expand_significand normalizes), but a cohorted zero has no precision to
+    // expand and its exponent can be arbitrarily large or small. Short-circuit so the
+    // result is the non-zero operand (or zero with a sensible cohort if both are zero).
+    if (big_lhs == 0U && big_rhs == 0U)
+    {
+        // IEEE 754-2008 6.1: preferred quantum for the sum of two zeros is min(exp_x, exp_y).
+        // IEEE 754-2008 6.3: sum of opposite-sign zeros is +0 in default rounding.
+        const auto result_exp {lhs_exp < rhs_exp ? lhs_exp : rhs_exp};
+        const bool result_sign {lhs.isneg() && rhs.isneg()};
+        return ReturnType{lhs.full_significand(), result_exp, result_sign};
+    }
+    if (big_lhs == 0U)
+    {
+        return ReturnType{rhs.full_significand(), rhs.biased_exponent(), rhs.isneg()};
+    }
+    if (big_rhs == 0U)
+    {
+        return ReturnType{lhs.full_significand(), lhs.biased_exponent(), lhs.isneg()};
+    }
+
     // Align to larger exponent
     if (lhs_exp != rhs_exp)
     {
@@ -212,6 +234,27 @@ BOOST_DECIMAL_CUDA_CONSTEXPR auto d128_add_impl_new(const T& lhs, const T& rhs)
     promoted_sig_type promoted_lhs {big_lhs};
     promoted_sig_type promoted_rhs {big_rhs};
 
+    // IEEE 754-2008 3.5.1: zero has a cohort with one representation per exponent.
+    // The exponent-comparison alignment below mishandles a cohorted zero whose exp
+    // is far from the non-zero operand's exp. Short-circuit so the result is the
+    // non-zero operand (or zero with a sensible cohort if both are zero).
+    if (big_lhs == typename T::significand_type{0} && big_rhs == typename T::significand_type{0})
+    {
+        // IEEE 754-2008 6.1: preferred quantum for the sum of two zeros is min(exp_x, exp_y).
+        // IEEE 754-2008 6.3: sum of opposite-sign zeros is +0 in default rounding.
+        const auto result_exp {lhs_exp < rhs_exp ? lhs_exp : rhs_exp};
+        const bool result_sign {lhs.isneg() && rhs.isneg()};
+        return ReturnType{lhs.full_significand(), result_exp, result_sign};
+    }
+    if (big_lhs == typename T::significand_type{0})
+    {
+        return ReturnType{rhs.full_significand(), rhs.biased_exponent(), rhs.isneg()};
+    }
+    if (big_rhs == typename T::significand_type{0})
+    {
+        return ReturnType{lhs.full_significand(), lhs.biased_exponent(), lhs.isneg()};
+    }
+
     // Align to larger exponent
     if (lhs_exp != rhs_exp)
     {

include/boost/decimal/detail/comparison.hpp

@@ -62,6 +62,14 @@ BOOST_DECIMAL_FORCE_INLINE BOOST_DECIMAL_CUDA_CONSTEXPR auto equality_impl(Decim
         return (lhs_sig == 0U && rhs_sig == 0U);
     }
 
+    // Step 4b: Same-sign zeros from any cohort compare equal regardless of their exponents
+    // (IEEE 754-2008 3.5.1). Without this, two zeros with delta_exp greater than the type's
+    // precision would fall through to the early-return below and wrongly compare unequal.
+    if (lhs_sig == 0U && rhs_sig == 0U)
+    {
+        return true;
+    }
+
     // Step 5: Check the exponents
     // If the difference is greater than we can represent in the significand than we can assume they are different
     const auto lhs_exp {lhs_components.exp};