This is the mail archive of the
mailing list for the GCC project.
Re: [PATCH] Optimize to_chars
- From: Jonathan Wakely <jwakely at redhat dot com>
- To: Antony Polukhin <antoshkka at gmail dot com>
- Cc: libstdc++ <libstdc++ at gcc dot gnu dot org>, gcc-patches List <gcc-patches at gcc dot gnu dot org>
- Date: Fri, 30 Aug 2019 17:01:26 +0100
- Subject: Re: [PATCH] Optimize to_chars
- References: <CAKqmYPZh+KXP3J_5X3xmz=-D0h013_w201Z=E5e56sfvCtAr1A@mail.gmail.com>
On 30/08/19 17:27 +0300, Antony Polukhin wrote:
Bunch of micro optimizations for std::to_chars:
* For base == 8 replacing the lookup in __digits table with arithmetic
computations leads to a same CPU cycles for a loop (exchanges two
movzx with 3 bit ops https://godbolt.org/z/RTui7m ). However this
saves 129 bytes of data and totally avoids a chance of cache misses on
* For base == 16 replacing the lookup in __digits table with
arithmetic computations leads to a few additional instructions, but
totally avoids a chance of cache misses on __digits (- ~9 cache misses
for worst case) and saves 513 bytes of const data.
* Replacing __first[pos] and __first[pos - 1] with __first and
__first on final iterations saves ~2% of code size.
* Removing trailing '\0' from arrays of digits allows the linker to
merge the symbols (so that "0123456789abcdefghijklmnopqrstuvwxyz" and
"0123456789abcdef" could share the same address). This improves data
locality and reduces binary sizes.
* Using __detail::__to_chars_len_2 instead of a generic
__detail::__to_chars_len makes the operation O(1) instead of O(N). It
also makes the code two times shorter ( https://godbolt.org/z/Peq_PG)
In sum: this significantly reduces the size of a binary (for about
4KBs only for base-8 conversion https://godbolt.org/z/WPKijS ), deals
with latency (CPU cache misses) without changing the iterations count
and without adding costly instructions into the loops.
This is great, thanks.
Have you tried comparing the improved code to libc++'s implementation?
I believe they use precomputed arrays of digits, but they use larger
arrays that allow 4 bytes to be written at once, which is considerably
faster (and those precomputed arrays libe in libc++.so not in the
header). Would we be better off keeping the precomputed arrays and
expanding them to do 4-byte writes?
Since we don't have a patch to do that, I think I'll commit yours. We
can always go back to precomputed arrays later if somebody does that
My only comments are on the changelog:
* include/std/charconv (__detail::__to_chars_8,
__detail::__to_chars_16): Replace array of precomputed digits
When the list of changed functions is split across lines it should be
* include/std/charconv (__detail::__to_chars_8)
(__detail::__to_chars_16): Replace array of precomputed digits
i.e close the parentheses before the line break, and reopen on the
with arithmetic operations to avoid CPU cache misses. Remove
zero termination from array of digits to allow symbol merge with
generic implementation of __detail::__to_chars. Replace final
offsets with constants. Use __detail::__to_chars_len_2 instead
of a generic __detail::__to_chars_len.
* include/std/charconv (__detail::__to_chars): Remove
Don't repeat the asterisk and filename for later changes in the same
(__detail::__to_chars): Remove zero termination from array of digits.
(__detail::__to_chars_2): Leading digit is always '1'.
There's no changelog entry for the changes to __to_chars_len_8 and
Also, please don't include the ChangeLog diff in the patch, because it
just makes it hard to apply the patch (the ChangeLog part will almost
always fail to apply because somebody else will have modified the
ChangeLog file since you created the patch, and so that hunk won't
apply. The ChangeLog text should be sent as a separate (plain text)
attachment, or just in the email body (as you did above).
I'll test this and commit it, thanks!