[PATCH] libstdc++: Optimize C++20 comparison category types
Jonathan Wakely
jwakely@redhat.com
Fri Feb 7 14:23:00 GMT 2020
On 07/02/20 10:04 +0100, Daniel Krügler wrote:
>Am Do., 6. Feb. 2020 um 15:28 Uhr schrieb Jonathan Wakely <jwakely@redhat.com>:
>>
>> On 06/02/20 13:53 +0000, Jonathan Wakely wrote:
>> >On 06/02/20 13:40 +0000, Jonathan Wakely wrote:
>> >>This reduces sizeof(std::partial_ordering) and optimizes conversion and
>> >>comparison operators to avoid conditional branches where possible.
>> >>
>> >> * libsupc++/compare (__cmp_cat::_Ncmp::unordered): Change value to 2.
>> >> (partial_ordering::_M_is_ordered): Remove data member.
>> >> (partial_ordering): Use second bit of _M_value for unordered. Adjust
>> >> comparison operators.
>> >> (weak_ordering::operator partial_ordering): Simplify to remove
>> >> branches.
>> >> (operator<=>(unspecified, weak_ordering)): Likewise.
>> >> (strong_ordering::operator partial_ordering): Likewise.
>> >> (strong_ordering::operator weak_ordering): Likewise.
>> >> (operator<=>(unspecified, strong_ordering)): Likewise.
>> >> * testsuite/18_support/comparisons/categories/partialord.cc: New test.
>> >> * testsuite/18_support/comparisons/categories/strongord.cc: New test.
>> >> * testsuite/18_support/comparisons/categories/weakord.cc: New test.
>> >>
>> >>Tested powerpc64le-linux and x86_64-linux.
>> >>
>> >>This is an ABI change for the partial_ordering type, but that is why I
>> >>think we should do it now, not after GCC 10 is released. The sooner
>> >>the better, before these types are being widely used.
>> >>
>> >>I plan to commit this in the next 12 hours or so, unless there are
>> >>(valid :-) objections.
>> >>
>> >>Thanks to Barry Revzin for pointing out there was room for these
>> >>operators to be improved.
>> >
>> >We could also change the int _M_value data member of all three
>> >comparison category types to be a signed char instead of int. That
>> >would reduce the size further.
>>
>> Or maybe std::int_fast8_t is the right type here.
>
>I like this suggestion.
After discussing it with Jakub, I've decided to just use signed char.
In theory int_fast8_t seems like the right choice, but it depends on
how "fast" is interpreted. Fast for which operations? Is it actually
going to be faster for the simple comparisons to constants and the
bit masks that I use in <compare>?
On a number of operating systems int_fast8_t is always int, even
though byte accesses are no slower on many of the processors whre that
OS runs.
So we decided that unconditionally using a single byte will give us a
small size and alignment, without sacrificing any performance.
Here's what I've tested (powerpc64le-linux and x86_64-linux) and
committed to master.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: patch.txt
Type: text/x-patch
Size: 22287 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/libstdc++/attachments/20200207/49ccd2e2/attachment.bin>
More information about the Libstdc++
mailing list