Re: [PATCH] libstdc++: Optimize integer std::from_chars

[Jonathan Wakely <jwakely@redhat.com>]

On Thu, 14 Apr 2022 at 13:32, Patrick Palka via Libstdc++
<libstdc++@gcc.gnu.org> wrote:
>
> This applies the following optimizations to the integer std::from_chars
> implementation:
>
>   1. Use a lookup table for converting an alphanumeric digit to its
>      base-36 value instead of using a range test (for 0-9) and switch
>      (for a-z and A-Z).  The table is constructed using a C++14
>      constexpr function which doesn't assume a particular character
>      encoding or __CHAR_BIT__ value.  The new conversion function
>      __from_chars_alnum_to_val is templated on whether we care
>      only about the decimal digits, in which case we can perform the
>      conversion with a single subtraction since the digit characters
>      are guaranteed to be contiguous (unlike the letters).
>   2. Generalize __from_chars_binary to handle all power-of-two bases.
>      This function, now named __from_chars_pow2_base, is also templated
>      on whether we care only about the decimal digits in order to speed
>      up digit conversion for base 2, 4 and 8.
>   3. In __from_chars_digit, use
>        static_cast<unsigned char>(__c - '0') < __base
>      instead of
>        '0' <= __c && __c <= ('0' + (__base - 1)).
>      as the digit recognition test (exhaustively verified that the two
>      tests are equivalent).
>   4. In __from_chars_alnum, use a nested loop to consume the rest of the
>      digits in the overflow case (mirroring __from_chars_digit) so that
>      the main loop doesn't have to maintain the __valid overflow flag.
>
> At this point, __from_chars_digit is nearly identical to
> __from_chars_alnum, so this patch combines the two functions, removing
> the former and templatizing the latter according to whether we care only
> about the decimal digits.  Finally,
>
>   5. In __from_chars_alnum, keep track of a lower bound on the number of
>      unused bits in the result and use that to omit the overflow check
>      when it's safe to do so.
>
> In passing this replaces the non-portable function ascii_to_hexit
> used by __floating_from_chars_hex with the new conversion function.
>
> Here are some runtime measurements for a simple 15-line benchmark that
> roundtrips printing/parsing 200 million integers via std::to/from_chars
> (average of 5 runs):
>
>   Base  Before  After (seconds, lower is better)
>      2    9.37   9.37
>      3   12.13  15.79
>      8    3.67   4.15
>     10    3.86   4.90
>     11    5.03   6.84
>     16    2.93   4.14
>     32    2.39   3.85
>     36    3.26   5.22
>
> Testedon x86_64-pc-linux-gnu, does this look OK for trunk?  Also tested
> against libc++'s from_chars tests for good measure.
>
> libstdc++-v3/ChangeLog:
>
>         * include/std/charconv (__from_chars_alnum_to_val_table): Define.
>         (__from_chars_alnum_to_val): Define.
>         (__from_chars_binary): Rename to ...
>         (__from_chars_pow2_base): ... this.  Generalize to handle any
>         power-of-two base using __from_chars_alnum_to_val.
>         (__from_chars_digit): Optimize digit recognition to a single
>         test instead of two tests.  Use [[__unlikely___]] attribute.
>         (__from_chars_alpha_to_num): Remove.
>         (__from_chars_alnum): Use __from_chars_alnum_to_val.  Use a
>         nested loop for the overflow case.
>         (from_chars): Adjust appropriately.
>         * src/c++17/floating_from_chars.cc (ascii_to_hexit): Remove.
>         (__floating_from_chars_hex): Use __from_chars_alnum_to_val
>         to recognize a hex digit instead.
> ---
>  libstdc++-v3/include/std/charconv             | 250 ++++++++----------
>  libstdc++-v3/src/c++17/floating_from_chars.cc |  18 +-
>  2 files changed, 105 insertions(+), 163 deletions(-)
>
> diff --git a/libstdc++-v3/include/std/charconv b/libstdc++-v3/include/std/charconv
> index 2ce9c7d4cb9..5e44459749a 100644
> --- a/libstdc++-v3/include/std/charconv
> +++ b/libstdc++-v3/include/std/charconv
> @@ -407,176 +407,127 @@ namespace __detail
>        return true;
>      }
>
> -  /// std::from_chars implementation for integers in base 2.
> -  template<typename _Tp>
> +  // Construct and return a lookup table that maps 0-9, A-Z and a-z to the
> +  // corresponding corresponding base-36 value and maps all other characters
> +  // to 127.
> +  constexpr auto
> +  __from_chars_alnum_to_val_table()
> +  {
> +    constexpr unsigned char __lower_letters[]
> +      = { 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j',
> +         'k', 'l', 'm', 'n', 'o', 'p', 'q', 'r', 's', 't',
> +         'u', 'v', 'w', 'x', 'y', 'z' };
> +    constexpr unsigned char __upper_letters[]
> +      = { 'A', 'B', 'C', 'D', 'E', 'F', 'G', 'H', 'I', 'J',
> +         'K', 'L', 'M', 'N', 'O', 'P', 'Q', 'R', 'S', 'T',
> +         'U', 'V', 'W', 'X', 'Y', 'Z' };
> +    struct { unsigned char __data[1u << __CHAR_BIT__] = {}; } __table;
> +    for (auto& __entry : __table.__data)
> +      __entry = 127;
> +    for (int __i = 0; __i < 10; ++__i)
> +      __table.__data['0' + __i] = __i;
> +    for (int __i = 0; __i < 26; ++__i)
> +      {
> +       __table.__data[__lower_letters[__i]] = 10 + __i;
> +       __table.__data[__upper_letters[__i]] = 10 + __i;
> +      }
> +    return __table;
> +  }
> +
> +  /// If _DecOnly is true: if the character is a decimal digit, then
> +  /// return its corresponding base-10 value, otherwise return a value >= 127.
> +  /// If _DecOnly is false: if the character is an alphanumeric digit, then
> +  /// return its corresponding base-36 value, otherwise return a value >= 127.
> +  template<bool _DecOnly>
> +    unsigned char
> +    __from_chars_alnum_to_val(unsigned char __c)
> +    {
> +      if _GLIBCXX17_CONSTEXPR (_DecOnly)
> +       return __c - '0';
> +      else
> +       {
> +         static constexpr auto __table = __from_chars_alnum_to_val_table();
> +         return __table.__data[__c];
> +       }
> +    }
> +
> +  /// std::from_chars implementation for integers in a power-of-two base.
> +  /// If _DecOnly is true, then we may assume __base is at most 8.
> +  template<bool _DecOnly, typename _Tp>
>      bool
> -    __from_chars_binary(const char*& __first, const char* __last, _Tp& __val)
> +    __from_chars_pow2_base(const char*& __first, const char* __last, _Tp& __val,
> +                          int __base)
>      {
>        static_assert(is_integral<_Tp>::value, "implementation bug");
>        static_assert(is_unsigned<_Tp>::value, "implementation bug");
>
> +      __glibcxx_assert((__base & (__base - 1)) == 0);
> +      __glibcxx_assert(_DecOnly ? __base <= 8 : __base <= 32);

I think we should remove these assertions. They can only fail if we
messed up in libstdc++, so users should not have to pay to assert that
the implementation is correct. We should save assertions for things
the user might get wrong.

If you want, leave them there but commented out. That serves as a
reminder of the preconditions, and we can uncomment them while
refactoring the code, to ensure we don't break something later. Maybe
we should have a separate kind of assertion, something like
__glibcxx_impl_assert, which would be enabled by
_GLIBCXX_IMPL_ASSERTIONS that we can enable for the testsuite. But
that would be something to add in stage 1.

Apart from that, this looks great, please push to trunk, thanks!