Hi Tomasz,

Thanks for the updated patches.

I re-ran the benchmarks on the updated patches (x86_64, -O2, core
pinned, BABA interleaved, vs trunk):

  scenario (vs trunk)        4 B    16 B    64 B   256 B  1024 B  4096 B
  fmt_to                   1.14x   1.83x   4.14x  17.40x  54.26x 118.79x
  fmt_to_n (notrunc)       1.12x   1.78x   4.90x  16.13x  54.61x 119.57x
  fmt_to_noputarea         1.27x   2.03x   5.78x  19.86x  21.36x  26.49x
  fmt_to_filebuf           1.19x   1.98x   5.26x  12.21x  19.29x  20.71x
  fmt_to_append            1.19x   1.87x   4.40x  13.37x  32.46x  47.73x
  fmt_to_scanner           1.13x   1.09x   1.08x   1.07x   1.08x   1.07x
  print (control)          1.01x   0.99x   0.99x   1.00x   0.97x   0.99x
  string (control)         1.00x   1.10x   1.00x   0.99x   1.04x   1.01x

Perf stat for the primary fmt_to path:

  len    instr%  cycles%  b_IPC  a_IPC
  4      85.8%   87.9%    2.76   2.69
  16     64.5%   54.2%    2.24   2.66
  64     32.8%   19.5%    1.60   2.69
  256    11.7%    5.7%    1.36   2.76
  1024    4.0%    2.3%    1.33   2.37
  4096    1.8%    0.9%    1.33   2.65

Compared with the numbers from Jul 16: 4096B is unchanged (121.22x
-> 118.79x, within noise), 256B improved (10.98x -> 17.40x).


> On Aug 11, 2026, at 16:42, Tomasz Kaminski <[email protected]> wrote:
> 
> 
> 
> On Thu, Jul 16, 2026 at 2:04 PM Anlai Lu <[email protected] <mailto:[email protected]>> 
> wrote:
>> This series replaces per-character sputc() with bulk sputn() and
>> zero-copy put-area writes when formatting to ostreambuf_iterator.
>> Patch 2 adds a _Streambuf_sink that writes directly to a streambuf;
>> patch 3 specializes _Iter_sink to use it; patch 4 uses _Streambuf_sink
>> in std::print; patch 1 adds tests.
>> 
>> Performance:
>> 
>>   fmt_to (ostreambuf_iter)    4 B     16 B    64 B    256 B   1024 B  4096 B
>>     v3:                       0.97x   1.61x   4.15x  10.98x  51.97x 121.22x
>> 
>>   print (uses _Streambuf_sink, per [ostream.formatted.print]):
>>                                4 B     16 B    64 B    256 B   1024 B  4096 B
>>     before (ns/op):           87       114     112     118     190     286
>>     after  (ns/op):           78        78      77      86      82     134
>>     ratio:                   1.12x    1.47x   1.46x   1.37x   2.33x   2.14x
>> 
>>   string / fmt_to_scanner unchanged (controls).
> Hi,
> I have posted an updated version of first 3 patches (see my comments on
> fourth one, explaining why it is non-conforming, and needs a paper to allow
> taht optimization). Would you mind redoing your tests on the updated patches,
> as I do not have access to them? 
>> 
>> Changes in v3
>> ==============
>> (Items marked with [R] were suggested by Tomasz Kaminski.)
>> 
>> - Extracted protected _M_flush(span) from _M_overflow so that both
>>   _M_overflow and _M_finish share the same write path, and the
>>   _Iter_sink specialization reuses it too.  [R]
>> - Added _M_write_failed flag and _M_discarding() override to
>>   _Streambuf_sink, so the format engine stops output early after a
>>   write failure.  [R]
>> - Changed _M_max from __diff_t to size_t with _S_no_limit sentinel.  [R]
>> - Error propagation to the iterator moved to _Iter_sink::_M_finish
>>   (checked just before returning std::move), so _M_overflow no longer
>>   touches _M_out.
>> - Used _Streambuf_sink directly in vprint_nonunicode (patch 4),
>>   replacing the _Str_sink + __ostream_write two-phase approach.
>> - Added _M_set_failed() to ostreambuf_iterator for the iterator's
>>   error propagation path.
>> - Added try/catch in _M_flush to handle sputn I/O exceptions without
>>   letting them propagate as vformat exceptions.  Per
>>   [ostream.formatted.print]/(4.2), vformat exceptions propagate without
>>   badbit; I/O exceptions from sputn are not vformat exceptions and
>>   are converted to _M_write_failed, then reported as badbit per (4.4).
>> - _M_finish() now ref-qualified (&&) on all sinks for consistency;
>>   the print path uses std::move to call it.  [R]
>> - Extracted _M_limit(span) helper in the _Iter_sink specialization for
>>   shared counting/truncation between _M_overflow and _M_finish;
>>   _M_finish now calls _M_flush directly, skipping unnecessary buffer
>>   restoration.  [R]
>> 
>> Not applied
>> ============
>> 1. Destructor flush  [R]
>>    Flush in destruction runs after the error check in print, so it
>>    cannot help report I/O errors.  And no other sink flushes in its
>>    destructor.
>> 
>> 2. _M_bump override  [R]
>>    The default advances _M_next within the current span; commit
>>    happens via _M_flush.  A comment explains this.
>> 
>> Anlai Lu (4):
>>   libstdc++: Add tests for format_to with ostreambuf_iterator
>>   libstdc++: Add _Streambuf_sink for direct streambuf formatting
>>   libstdc++: Specialize _Iter_sink for ostreambuf_iterator
>>   libstdc++: Optimize std::print using _Streambuf_sink
>> 
>>  libstdc++-v3/include/bits/ostream_print.h     |  12 +-
>>  .../include/bits/streambuf_iterator.h         |  11 +
>>  libstdc++-v3/include/std/format               | 199 ++++++++++++++
>>  libstdc++-v3/include/std/streambuf            |   4 +
>>  .../format/functions/format_to_ostreambuf.cc  | 244 ++++++++++++++++++
>>  5 files changed, 463 insertions(+), 7 deletions(-)
>>  create mode 100644 
>> libstdc++-v3/testsuite/std/format/functions/format_to_ostreambuf.cc
>> 
>> base-commit: a0591d86bd2be838e30c132481a0b15a6ba1e112
>> 
>> --
>> 2.34.1

Reply via email to