Re: [PATCH v4 08/14] libfdt: Speed up fdt_find_string_len_() with memchr()

From: David Laight

Date: Wed Nov 18 2026 - 17:15:48 EST


On Wed, 18 Nov 2026 12:39:59 -0800
Doug Anderson <dianders@xxxxxxxxxxxx> wrote:

> Hi David,
>
> On Wed, Nov 18, 2026 at 11:15 AM David Laight
> <david.laight.linux@xxxxxxxxx> wrote:
> >
> > > 2. Every stored string ends with `'\0'`, whereas `s[0]` only appears in
> > > a smaller fraction of bytes. On a real upstream Qualcomm base DTB's
> > > `dt_strings` block (855 strings in 11,259 bytes, average 13.2 bytes
> > > per string), `'\0'` occurs on 7.6% of bytes while `s[0]` averages
> > > 3.9%. Replaying those 855 insertions makes 362k `memchr()` calls when
> > > scanning for `'\0'` vs. 187k `memchr()` calls when scanning for
> > > `s[0]`.
> >
> > You need to count the number of times that a letter that is the first
> > character of a string you are going to lookup appears in the entire buffer.
> > There might only be one 'j' in the entire buffer - but you never look for it.
> > (But I might still be wrong)
>
> That 3.9% (`187,315` `memchr()` calls across ~4.77M scanned bytes) was
> actually measured directly by replaying those 855 real lookups (`s[0]`
> of each string being inserted), not an unweighted 1/26 average across
> the alphabet. Even though common prefixes like `'p'`, `'u'`, `'q'`, and
> `'s'` get looked up a lot, 26.6% of the `dt_strings` block (`2,139`
> bytes of `'_'`, `'-'`, `'0'-'9'`, and `'#'`, plus `855` NUL bytes) are
> characters that almost never appear as `s[0]`, which pulls the
> lookup-weighted hit rate back down to 3.9%.
>
> Intuitively, `'\0'` hits once per string (`1 / L`), so `s[0]` wins as
> long as average string length `L` stays below the ~22-26 byte crossover
> -- and even on plain non-`-@` DTBs without `/__symbols__`, property
> names average ~12-14 bytes (`'\0'` still made 1.6x-2.2x more `memchr()`
> calls than `s[0]` across plain DTB `fdt_sw` builds, `/chosen` fixups,
> and `.dtbo` overlay application).
>
> > Have you tried an open coded scans instead of calling memchr() and memcmp()?
> > With an average of 13 bytes per string (26 for your memchr()) the code loop
> > might be faster.
> > But beware, gcc is likely to convert the code back to a call to memcmp()
> > - it certainly does for strlen and memcpy.
> >
> > Oh - and if this code runs in kernel the timings will be significantly
> > different to glibc - it uses some magic avx instructions for strlen()
> > and memchr() is effectively the same.
> > The kernel functions are likely to be out of line byte loops.
>
> Out of curiosity I timed all of those across 200 replays of the same
> 855-string (11,259-byte) Qualcomm `dt_strings` block with `gcc 15.2` on
> x86_64 (`gcc -O2` kept the open-coded loops inline and didn't turn them
> back into `memcmp()`):

Once you are doing timings, adding a few more doesn't take long.

> * glibc `memchr` + `!p[slen]` + glibc `memcmp`:
> 0.25 s (`-O0`), 0.23 s (`-Os`), 0.22 s (`-O2`)
> * glibc `memchr` + `!p[slen]` + open-coded inner compare:
> 0.23 s (`-O0`), 0.21 s (`-Os`), 0.21 s (`-O2`)

So that one is a marginal win.

> * Fully open-coded (`p[0] == s[0] && !p[slen]` + inner byte loop):
> 0.91 s (`-O0`), 0.62 s (`-Os`), 0.40 s (`-O2`)
> * Out-of-line byte-loop `memchr` + `!p[slen]` + byte-loop `memcmp`:
> 0.86 s (`-O0`), 0.90 s (`-Os`), 0.50 s (`-O2`)
> * Original per-byte out-of-line byte-loop `memcmp`:
> 3.31 s (`-O0`), 1.87 s (`-Os`), 1.86 s (`-O2`)
>
> Even with a ~25-byte average stride, glibc's AVX2 `memchr()` handles
> almost an entire stride in a single 32-byte vector load + compare, so it
> beats the open-coded scalar loop by ~1.9x at `-O2` (and ~2.7x at `-Os`).

Using avx2 is 'cheating' ...
Someone at Intel realised that strlen() does matter and threw some silicon at it.
They finally did the same for integer divide (but I think AMD did it first)
and did a divide that generates move than one bit per clock.

> Open-coding just the inner compare barely moves the needle since
> `!p[slen]` already cut `memcmp()` down to 14.3k calls across the 855
> insertions.
>
> On the kernel / bootloader side, `fdt_find_string_len_()` is only called
> by `fdt_sw` and `fdt_rw` when adding new property names to `dt_strings`
> -- normal read-only DTB parsing (`fdt_ro.c`) never calls it. And even
> with a naive out-of-line byte-loop `memchr()`, `memchr()` + `!p[slen]` +
> `memcmp()` is still 3.7x faster at `-O2` (`1.86 s -> 0.50 s`) and 2.1x
> faster at `-Os` (`1.87 s -> 0.90 s`) than the old code, so keeping the
> simpler `memchr()` + `!p[slen]` + `memcmp()` seems like the right
> tradeoff.

Seems reasonable.

David

>
> -Doug