Re: [PATCH] libstdc++: Make chrono::hh_mm_ss more compact

[Jonathan Wakely <jwakely@redhat.com>]

I've pushed this to trunk. Tested x86_64-linux.

On Mon, 21 Nov 2022 at 20:44, Jonathan Wakely via Libstdc++
<libstdc++@gcc.gnu.org> wrote:
>
> While finishing the time zone support for C++20, I noticed that the
> chrono::hh_mm_ss type is surprisingly large: 40 bytes. That's because
> we use duration<int64_t, ...> for each of the four members, _M_h, _M_m,
> _M_s and _M_ss. This is very wasteful. The patch below reduces it to 16
> bytes (or less for some targets) for most specializations using an
> integral type for the hh_mm_ss<D>::precision::rep type.
>
> Patrick, you added hh_mm_ss, but I assume you just implemented what is
> shown for the exposition-only members in the standard, rather than
> intentionally choosing this 40-byte representation? I don't recall
> discussing it at the time, but I might be forgetting something.
>
> Does the patch below make sense? I could remove the partial
> specialization that use a single byte for the subseconds, or only do
> that on 32-bit targets where alignof(int64_t) is 4. For targets where
> alignof(int64_t) is 8 we're always going to be 16 bytes minimum, so
> maybe there's no point squeezing everything into 12 bytes if we have a
> mandatory 4 bytes of padding.
>
> We could make those members use even less space, something like:
>
> hours _M_h;
> unsigned _M_m : 6;
> unsigned _M_s : 6;
> unsigned _M_is_neg : 1;
> unsigned _M_ss : 51;  // good enough for femtoseconds
>
> This would only require 16 bytes even for hh_mm_ss<femtoseconds>, but
> requires extra instructions to read the values. That seems like the
> wrong trade off. A compact layout for hh_mm_ss<system_clock::duration>
> and hh_mm_ss<high_resolution_clock::duration> seems good enough. Paying
> extra instructions to access the members just to support incredibly high
> precision doesn't make sense to me.
>
> We could also elide the bool _M_is_neg member if _Duration::rep is
> unsigned, but it doesn't seem worth the metaprogramming/maintenance
> cost.
>
> We could consider using duration<uint_least32_t, ratio<3600>> for the
> _M_h member instead of chrono::hours (which uses int64_t). It seems
> unlikely (but not impossible) anybody will want to use hh_mm_ss for
> billions of hours past midnight. If we do that, then using a single byte
> for subseconds (when possible) does make sense, because the whole
> hh_mm_ss would fit in just 8 bytes (4 for hours, 1 for minutes, 1 for
> seconds, 1 for is_neg, and 1 for the subseconds). And then bit-fields
> start to look more appealing. Maybe just for the subseconds and is_neg
> flag, which would be 8 bytes for precision::period >= ratio<1,10'000>:
>
> uint_least32_t _M_h;
> uint_least8_t  _M_m;
> uint_least8_t  _M_s;
> uint_least16_t _M_ss : 15;
> uint_least16_t _M_is_neg : 1;
>
> Has anybody got any other suggestions for optimizing this or other
> chrono types? e.g. I think we should also consider using int32_t as the
> rep for chrono::{days,weeks,years} because a 64-bit type for years is
> unnecessary.
>
> It looks like other implementations don't bother trying to save space
> for chrono::hh_mm_ss, which makes me wonder if there's some reason not
> to do this. For MSVC it's only 32 bytes, which I think is because they
> use a 32-bit type for chrono::hours and chrono::minutes, not because
> they're doing anything special to make chrono::hh_mm_ss smaller.
>
>
> -- >8 --
>
> This uses a single byte for the minutes and seconds members, and places
> the bool member next to those single bytes. This means we do not need 40
> bytes to store a time that can fit in a single 8-byte integer.
>
> When there is no subsecond precision we can do away with the _M_ss
> member altogether. If the subsecond precision is coarse enough, we can
> use a smaller representation for _M_ss, e.g. hh_mm_ss<milliseconds> only
> needs uint_least32_t for _M_ss, and hh_mm_ss<duration<long, ratio<1,10>>
> and hh_mm_ss<duration<int8_t, nano>> each only needs a single byte. In
> the latter case the type can only ever represent up to 255ns anyway, so
> we don't need a larger representation type (in such cases, we could even
> remove the _M_h, _M_m and _M_s members, but it's a very unlikely
> scenario that isn't worth optimizing for).
>
> Except for specializations with a floating-point rep or using higher
> precision than nanoseconds, hh_mm_ss should now fit in 16 bytes, or even
> 12 bytes for x86-32 where alignof(long long) == 4.
>
> libstdc++-v3/ChangeLog:
>
>         * include/std/chrono (chrono::hh_mm_ss): Do not use 64-bit
>         representations for all four duration members. Reorder members.
>         (hh_mm_ss::_S_fractional_width()): Optimize using countr_zero.
>         (hh_mm_ss::hh_mm_ss()): Define as defaulted.
>         (hh_mm_ss::hh_mm_ss(Duration)): Delegate to a new private
>         constructor, instead of calling chrono::abs repeatedly.
>         * testsuite/std/time/hh_mm_ss/1.cc: Check floating-point
>         representations. Check default constructor. Check sizes.
> ---
>  libstdc++-v3/include/std/chrono               | 139 +++++++++++++-----
>  libstdc++-v3/testsuite/std/time/hh_mm_ss/1.cc |  54 ++++++-
>  2 files changed, 158 insertions(+), 35 deletions(-)
>
> diff --git a/libstdc++-v3/include/std/chrono b/libstdc++-v3/include/std/chrono
> index 2468023f6c5..4ee163ba762 100644
> --- a/libstdc++-v3/include/std/chrono
> +++ b/libstdc++-v3/include/std/chrono
> @@ -41,6 +41,7 @@
>  #include <bits/chrono.h>
>
>  #if __cplusplus >= 202002L
> +# include <bit>
>  # include <sstream>
>  # include <string>
>  # include <vector>
> @@ -2262,6 +2263,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
>
>      // HH_MM_SS
>
> +    /// @cond undocumented
>      namespace __detail
>      {
>        consteval long long
> @@ -2273,22 +2275,22 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
>         return __r;
>        }
>      }
> +    /// @endcond
>
> +    /// A duration representing a time since midnight (00h:00min:00s).
>      template<typename _Duration>
>        class hh_mm_ss
>        {
> +       static_assert( __is_duration<_Duration>::value );
> +
>        private:
> -       static constexpr int
> +       static consteval int
>         _S_fractional_width()
>         {
> -         int __multiplicity_2 = 0;
> -         int __multiplicity_5 = 0;
>           auto __den = _Duration::period::den;
> -         while ((__den % 2) == 0)
> -           {
> -             ++__multiplicity_2;
> -             __den /= 2;
> -           }
> +         const int __multiplicity_2 = std::__countr_zero((uintmax_t)__den);
> +         __den >>= __multiplicity_2;
> +         int __multiplicity_5 = 0;
>           while ((__den % 5) == 0)
>             {
>               ++__multiplicity_5;
> @@ -2304,6 +2306,20 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
>           return __width;
>         }
>
> +       constexpr
> +       hh_mm_ss(_Duration __d, bool __is_neg) noexcept
> +       : _M_h (duration_cast<chrono::hours>(__d)),
> +         _M_m (duration_cast<chrono::minutes>(__d - hours())),
> +         _M_s (duration_cast<chrono::seconds>(__d - hours() - minutes())),
> +         _M_is_neg(__is_neg)
> +       {
> +         auto __ss = __d - hours() - minutes() - seconds();
> +         if constexpr (treat_as_floating_point_v<typename precision::rep>)
> +           _M_ss._M_r = __ss.count();
> +         else if constexpr (precision::period::den != 1)
> +           _M_ss._M_r = duration_cast<precision>(__ss).count();
> +       }
> +
>        public:
>         static constexpr unsigned fractional_width = {_S_fractional_width()};
>
> @@ -2312,28 +2328,21 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
>                                    chrono::seconds::rep>,
>                      ratio<1, __detail::__pow10(fractional_width)>>;
>
> -       constexpr
> -       hh_mm_ss() noexcept
> -       : hh_mm_ss{_Duration::zero()}
> -       { }
> +       constexpr hh_mm_ss() noexcept = default;
>
>         constexpr explicit
>         hh_mm_ss(_Duration __d) noexcept
> -       : _M_is_neg (__d < _Duration::zero()),
> -         _M_h (duration_cast<chrono::hours>(abs(__d))),
> -         _M_m (duration_cast<chrono::minutes>(abs(__d) - hours())),
> -         _M_s (duration_cast<chrono::seconds>(abs(__d) - hours() - minutes()))
> -       {
> -         if constexpr (treat_as_floating_point_v<typename precision::rep>)
> -           _M_ss = abs(__d) - hours() - minutes() - seconds();
> -         else
> -           _M_ss = duration_cast<precision>(abs(__d) - hours()
> -                                            - minutes() - seconds());
> -       }
> +       : hh_mm_ss(chrono::abs(__d), __d < _Duration::zero())
> +       { }
>
>         constexpr bool
>         is_negative() const noexcept
> -       { return _M_is_neg; }
> +       {
> +         if constexpr (!__is_unsigned)
> +           return _M_is_neg;
> +         else
> +           return false;
> +       }
>
>         constexpr chrono::hours
>         hours() const noexcept
> @@ -2349,7 +2358,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
>
>         constexpr precision
>         subseconds() const noexcept
> -       { return _M_ss; }
> +       { return static_cast<precision>(_M_ss); }
>
>         constexpr explicit
>         operator precision() const noexcept
> @@ -2358,20 +2367,82 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
>         constexpr precision
>         to_duration() const noexcept
>         {
> -         if (_M_is_neg)
> -           return -(_M_h + _M_m + _M_s + _M_ss);
> -         else
> -           return _M_h + _M_m + _M_s + _M_ss;
> +         if constexpr (!__is_unsigned)
> +           if (_M_is_neg)
> +             return -(_M_h + _M_m + _M_s + subseconds());
> +         return _M_h + _M_m + _M_s + subseconds();
>         }
>
>         // TODO: Implement operator<<.
>
>        private:
> -       bool _M_is_neg;
> -       chrono::hours _M_h;
> -       chrono::minutes _M_m;
> -       chrono::seconds _M_s;
> -       precision _M_ss;
> +       static constexpr bool __is_unsigned
> +         = __and_v<is_integral<typename _Duration::rep>,
> +                   is_unsigned<typename _Duration::rep>>;
> +
> +       template<typename _Ratio>
> +         using __byte_duration = duration<unsigned char, _Ratio>;
> +
> +       // The type of the _M_ss member that holds the subsecond precision.
> +       template<typename _Dur>
> +         struct __subseconds
> +         {
> +           typename _Dur::rep _M_r{};
> +
> +           constexpr explicit
> +           operator _Dur() const noexcept
> +           { return _Dur(_M_r); }
> +         };
> +
> +       // An empty class if this precision doesn't need subseconds.
> +       template<typename _Rep, typename _Period>
> +         requires (!treat_as_floating_point_v<_Rep>) && (_Period::den == 1)
> +         struct __subseconds<duration<_Rep, _Period>>
> +         {
> +           constexpr explicit
> +           operator duration<_Rep, _Period>() const noexcept
> +           { return {}; }
> +         };
> +
> +       // True if the maximum constructor argument can be represented in _Tp.
> +       template<typename _Tp>
> +         static constexpr bool __fits
> +           = duration_values<typename _Duration::rep>::max()
> +               <= duration_values<_Tp>::max();
> +
> +       template<typename _Rep, typename _Period>
> +         requires (!treat_as_floating_point_v<_Rep>)
> +           && ratio_less_v<_Period, ratio<1, 1>>
> +           && (ratio_greater_equal_v<_Period, ratio<1, 250>>
> +                 || __fits<unsigned char>)
> +         struct __subseconds<duration<_Rep, _Period>>
> +         {
> +           unsigned char _M_r{};
> +
> +           constexpr explicit
> +           operator duration<_Rep, _Period>() const noexcept
> +           { return duration<_Rep, _Period>(_M_r); }
> +         };
> +
> +       template<typename _Rep, typename _Period>
> +         requires (!treat_as_floating_point_v<_Rep>)
> +           && ratio_less_v<_Period, ratio<1, 250>>
> +           && (ratio_greater_equal_v<_Period, ratio<1, 4'000'000'000>>
> +                 || __fits<uint_least32_t>)
> +         struct __subseconds<duration<_Rep, _Period>>
> +         {
> +           uint_least32_t _M_r{};
> +
> +           constexpr explicit
> +           operator duration<_Rep, _Period>() const noexcept
> +           { return duration<_Rep, _Period>(_M_r); }
> +         };
> +
> +       chrono::hours               _M_h{};
> +       __byte_duration<ratio<60>>  _M_m{};
> +       __byte_duration<ratio<1>>   _M_s{};
> +       bool                        _M_is_neg{};
> +       __subseconds<precision>     _M_ss{};
>        };
>
>      // 12/24 HOURS FUNCTIONS
> diff --git a/libstdc++-v3/testsuite/std/time/hh_mm_ss/1.cc b/libstdc++-v3/testsuite/std/time/hh_mm_ss/1.cc
> index 31dd1b2a8f3..1c63276a5f0 100644
> --- a/libstdc++-v3/testsuite/std/time/hh_mm_ss/1.cc
> +++ b/libstdc++-v3/testsuite/std/time/hh_mm_ss/1.cc
> @@ -59,5 +59,57 @@ constexpr_hh_mm_ss()
>
>    static_assert(seconds{hh_mm_ss{100min}} == 100min);
>
> -  // TODO: treat_as_floating_point_v
> +  // treat_as_floating_point_v
> +  using fseconds = duration<double, ratio<1>>;
> +  constexpr hh_mm_ss<fseconds> fsec{0x123.0004p5s};
> +  static_assert(std::is_same_v<hh_mm_ss<fseconds>::precision, fseconds>);
> +  static_assert(fsec.hours() == 2h);
> +  static_assert(fsec.minutes() == 35min);
> +  static_assert(fsec.seconds() == 12s);
> +  static_assert(fsec.subseconds() == 0x.0004p5s);
> +  static_assert(!fsec.is_negative());
> +
> +  using fminutes = duration<double, ratio<60>>;
> +  constexpr hh_mm_ss<fminutes> fmin{-0x1.23p4min};
> +  static_assert(std::is_same_v<hh_mm_ss<fminutes>::precision, fseconds>);
> +  static_assert(fmin.hours() == 0h);
> +  static_assert(fmin.minutes() == 18min);
> +  static_assert(fmin.seconds() == 11s);
> +  static_assert(fmin.subseconds() == 0.25s);
> +  static_assert(fmin.is_negative());
> +}
> +
> +constexpr void
> +default_construction()
> +{
> +  using namespace std::chrono;
> +
> +  constexpr hh_mm_ss<seconds> s1;
> +  static_assert(s1.to_duration() == s1.to_duration().zero());
> +  constexpr hh_mm_ss<duration<char>> s2;
> +  static_assert(s2.to_duration() == s2.to_duration().zero());
> +  constexpr hh_mm_ss<duration<int, std::centi>> s3;
> +  static_assert(s3.to_duration() == s3.to_duration().zero());
> +  constexpr hh_mm_ss<duration<long long, std::femto>> s4;
> +  static_assert(s4.to_duration() == s4.to_duration().zero());
> +  constexpr hh_mm_ss<duration<double>> s5;
> +  static_assert(s5.to_duration() == s5.to_duration().zero());
> +}
> +
> +constexpr void
> +size()
> +{
> +  using namespace std::chrono;
> +
> +  struct S0 { long long h; char m; char s; bool neg; };
> +  static_assert(sizeof(hh_mm_ss<seconds>) == sizeof(S0));
> +  struct S1 { long long h; char m; char s; bool neg; char ss; };
> +  static_assert(sizeof(hh_mm_ss<duration<int, std::centi>>) == sizeof(S1));
> +  struct S2 { long long h; char m, s; bool neg; int ss; };
> +  static_assert(sizeof(hh_mm_ss<duration<int, std::milli>>) == sizeof(S2));
> +  static_assert(sizeof(hh_mm_ss<duration<int, std::pico>>) == sizeof(S2));
> +  struct S3 { long long h; char m, s; bool neg; long long ss; };
> +  static_assert(sizeof(hh_mm_ss<duration<long long, std::pico>>) == sizeof(S3));
> +  struct S4 { long long h; char m, s; bool neg; double ss; };
> +  static_assert(sizeof(hh_mm_ss<duration<double, std::micro>>) == sizeof(S4));
>  }
> --
> 2.38.1
>