<div dir="ltr"><div dir="ltr"><br></div><br><div class="gmail_quote gmail_quote_container"><div dir="ltr" class="gmail_attr">On Mon, Jul 6, 2026 at 6:29 AM Anlai Lu <<a href="mailto:agicy@qq.com">agicy@qq.com</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">This series replaces per-character sputc() with bulk sputn() and<br>
zero-copy put-area writes when formatting to ostreambuf_iterator.<br>
Patch 2 adds a _Streambuf_sink that writes directly to a streambuf;<br>
patch 3 specializes _Iter_sink to use it; patch 1 adds tests.<br>
<br>
The design follows review suggestions from Jonathan Wakely (PR111052)<br>
and Tomasz Kaminski: _Streambuf_sink is a standalone class accepting<br>
basic_streambuf* directly, and the _Iter_sink specialization is a<br>
separate commit.<br></blockquote><div>Thank you for the patch. I have commented on the 2/3 patch. The _Streambuf_sink</div><div>can expose _M_discarding (and I believe we should), however any _Sink use</div><div>with format_to_n cannot do so (as we need to count specialization of output).</div><div>I believe that by moving the counting to the specialization, we can still handle</div><div>that correctly. </div><div><br></div><div>Furthermore, if we use _Streambuf_sink, it does not need to count characters.</div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
<br>
Performance (header-overlay A/B, single core, BABA interleave):<br>
<br>
scenario 16 B 64 B 256 B 1024 B 4096 B<br>
print(ostream, "{}", x) 0.96x 0.93x 0.94x 0.96x 0.97x<br>
format_to(back_inserter, ...) 1.02x 0.99x 1.00x 1.00x 1.01x<br>
format_to(ostreambuf_iter, ...) 1.92x 5.39x 17.52x 54.47x 115.14x<br>
with width/padding spec 1.09x 1.08x 1.09x 1.08x 1.27x<br>
multi-arg ("_{}_...") 1.16x 1.55x 2.50x 4.39x 14.13x<br>
append mode - 4.63x 13.69x 32.74x -<br>
to filebuf (/dev/null) 2.01x 5.40x 12.80x 19.73x 22.19x<br>
to streambuf w/o put area - - 15.95x 21.67x 28.06x<br>
format_to_n(ostreambuf_iter,<br>
max=len/4, ...) - 1.90x 4.87x 6.82x -<br>
max=len*2, ...) - 5.14x 17.46x 55.78x -<br>
<br>
Perf stat for the primary fmt_to path:<br>
<br>
len instr% cycles% b_IPC a_IPC branch-miss<br>
16 66.5% 51.9% 2.24 2.87 44.1%<br>
64 33.8% 18.8% 1.60 2.88 0.0%<br>
256 12.0% 5.6% 1.36 2.92 0.0%<br>
1024 4.1% 1.8% 1.35 3.06 0.0%<br>
4096 1.8% 0.8% 1.33 2.93 0.0%<br>
<br>
Known limitations:<br>
- The print path (std::print to ostream) shows a 3-7% regression<br>
from <format> header inflation; it does not use _Streambuf_sink.<br>
<br>
Anlai Lu (3):<br>
libstdc++: Add tests for format_to with ostreambuf_iterator<br>
libstdc++: Add _Streambuf_sink for direct streambuf formatting<br>
libstdc++: Specialize _Iter_sink for ostreambuf_iterator<br>
<br>
.../include/bits/streambuf_iterator.h | 11 +<br>
libstdc++-v3/include/std/format | 212 ++++++++++++++++++<br>
libstdc++-v3/include/std/streambuf | 4 +<br>
.../format/functions/format_to_ostreambuf.cc | 244 +++++++++++++++++++++<br>
4 files changed, 471 insertions(+)<br>
create mode 100644 libstdc++-v3/testsuite/std/format/functions/format_to_ostreambuf.cc<br>
<br>
base-commit: 5be7b70e769e6e992dbf631cd37e24bd9da9e05e<br>
--<br>
2.34.1<br>
<br>
</blockquote></div></div>