Re: [PATCH] split: add content-defined chunking to --bytes
From:
Leonid Evdokimov
Subject:
Re: [PATCH] split: add content-defined chunking to --bytes
Date:
Sun, 29 Mar 2026 16:37:00 +0300
On Mon, Mar 16, 2026 at 6:41 PM Pádraig Brady <P@draigbrady.com> wrote:
> Newer hardware for example won't have to worry about the JCC issue,
> so that might be left as a distro level / user level setting.
Affected Skylake chips get 15..43% speedup in case of Jcc-workaround
being both needed and implemented. I've found just one case when the
Jcc-workaround penalizes execution time of the aligned loop. It's
BUZHash on Raptor Lake's pinned to Efficient-core. This case gets a
20% slowdown. Performance-core of the same CPU is unaffected.
Longer story below.
I run a benchmark on a dozen x86_64 CPUs from different eras (both
newer and older than Skylake), while copyright assignment paperwork is
going on. That's the CPUs I was able to sample:
- Intel Xeon L5630 10Q1 Nehalem/Westmere
- AMD A4-3310MX 11Q2 AMD A-Series, Llano
- Intel Xeon E3-1245 V2 12Q2 Sandy Bridge/Ivy Bridge
- Intel Celeron N2830 14Q1 Silvermont/Bay Trail
- Intel Xeon E5-2620 v3 14Q3 Haswell
- Intel 6th-gen Core i7-6600U 15Q3 Skylake/Skylake
- Intel Xeon Gold 6133 17Q3 Skylake/Skylake
- Intel 8th-gen Core i5-8250U 17Q3 Skylake/Kaby Lake R
- AMD Ryzen 5 PRO 3400GE 19Q3 Ryzen 3000
- Intel 10th-gen Core i7-10750H 20Q2 Skylake/Comet Lake
- AMD EPYC 7773X 64-Core 21Q1 EPYC 7003
- Intel Xeon Platinum 8592+ 23Q4 Raptor Cove/Emerald Rapids
- Intel 14th-gen Core i7-14700HX 24Q1 Raptor Lake
I've compiled the CDC code with bare -O2 and align-loops=32 /
Jcc-workaround toggled on/off. I used GCC 13, 14, 15, Clang 22 and
inspected resulting binaries with objdump checking if they're actually
affected by Jcc-erratum and under-alignment or not.
The only under-aligned loops were clang22/{buz64,gear32} and
gcc13/{gear32,gear64}. All other lookup loops were already aligned to
the 32-byte boundary.
The opcodes GCC 13 and Clang 22 generate for these tiny loops were
often less performant than ones GCC 15 generates. 35% difference was
not unusual in this micro-benchmark, 100% difference was also
possible: 1cpb with GCC-15 vs. 2cpb with Clang-22 for GearHash on AMD
EPYC 7773X.
I'm attaching a summarized cycles-per-byte table in case someone is
curious about raw numbers.
Lesson learned: tiny loops are extremely sensitive & CC version matters a lot.
So, I think the way to conceptualize this song and dance around CFLAGS
is neither "optimization" nor "Jcc workaround" but "fix for _known_
performance regressions".
E.g. gcc13 GearHash build with Jcc-workaround & default loop alignment
is penalized on EPYC and almost all Skylakes, probably, due to
misalignment.
clang22 BUZHash build with align-loops=32 and no Jcc-workaround is
penalized on affected Intel chips due to unlucky layout triggering the
issue.
So, a single --enable option for every tunable knob makes little sense
to me. I'd suggest --enable-cdc-cflags[=FLAGS] with default value
depending on CC version.
It turns out, as expected, that building a single x86_64 binary that
performs well across a large variety of chips is a non-trivial task
:-)
--
WBRBW, Leonid Evdokimov, https://darkk.net.ru tel:+79816800702
PGP: 6691 DE6B 4CCD C1C1 76A0 0D4A E1F2 A980 7F50 FAB2