Re: Optimize UUID parse using SIMD

On Thu, Aug 6, 2026 at 8:07 PM Chao Li <li(dot)evan(dot)chao(at)gmail(dot)com> wrote:
>
>
>
> > On Aug 7, 2026, at 00:09, Masahiko Sawada <sawada(dot)mshk(at)gmail(dot)com> wrote:
> >
> > On Wed, Aug 5, 2026 at 8:40 PM Chao Li <li(dot)evan(dot)chao(at)gmail(dot)com> wrote:
> >>
> >>
> >>
> >>> On Aug 6, 2026, at 08:19, Masahiko Sawada <sawada(dot)mshk(at)gmail(dot)com> wrote:
> >>>
> >>> On Tue, Jun 30, 2026 at 11:03 AM Haibo Yan <tristan(dot)yim(at)gmail(dot)com> wrote:
> >>>>
> >>>> On Tue, Jun 30, 2026 at 10:53 AM Masahiko Sawada <sawada(dot)mshk(at)gmail(dot)com> wrote:
> >>>>>
> >>>>> On Mon, Jun 29, 2026 at 5:53 PM Haibo Yan <tristan(dot)yim(at)gmail(dot)com> wrote:
> >>>>>>
> >>>>>> On Mon, Jun 29, 2026 at 2:55 PM Masahiko Sawada <sawada(dot)mshk(at)gmail(dot)com> wrote:
> >>>>>>>
> >>>>>>> On Sun, Jun 28, 2026 at 7:20 PM Haibo Yan <tristan(dot)yim(at)gmail(dot)com> wrote:
> >>>>>>>>
> >>>>>>>> On Thu, Jun 25, 2026 at 3:16 PM Masahiko Sawada <sawada(dot)mshk(at)gmail(dot)com> wrote:
> >>>>>>>>>
> >>>>>>>>> On Thu, Jun 25, 2026 at 2:31 PM Haibo Yan <tristan(dot)yim(at)gmail(dot)com> wrote:
> >>>>>>>>>>
> >>>>>>>>>>
> >>>>>>>>>>
> >>>>>>>>>> On Thu, Jun 25, 2026 at 11:28 AM Masahiko Sawada <sawada(dot)mshk(at)gmail(dot)com> wrote:
> >>>>>>>>>>>
> >>>>>>>>>>> Hi all,
> >>>>>>>>>>>
> >>>>>>>>>>> I'd like to propose the $subject.
> >>>>>>>>>>>
> >>>>>>>>>>> Since commit ec8719ccbfcd made hex_decode_safe() SIMD-aware, decoding
> >>>>>>>>>>> a run of hex digits is now fast. The attached patch reuses
> >>>>>>>>>>> hex_decode_safe() in the UUID input function to speed up parsing.
> >>>>>>>>>>>
> >>>>>>>>>>> We accept several textual forms of a UUID[1]. The fast path handles
> >>>>>>>>>>> the common ones: 32 hex digits, the canonical 8x-4x-4x-4x-12x form
> >>>>>>>>>>> (where "nx" means n hex digits), and either of those wrapped in
> >>>>>>>>>>> braces. Otherwise, it falls back to the ordinary scalar UUID parse.
> >>>>>>>>>>>
> >>>>>>>>>>> I've benchmarked the parse speed using the following query:
> >>>>>>>>>>>
> >>>>>>>>>>> CREATE TEMP TABLE u AS SELECT gen_random_uuid()::text AS t FROM
> >>>>>>>>>>> generate_series(1, 1000000);
> >>>>>>>>>>> EXPLAIN (ANALYZE, TIMING OFF) SELECT t::uuid FROM u;
> >>>>>>>>>>>
> >>>>>>>>>>> I compared the execution time of the second query, which measures
> >>>>>>>>>>> uuid_in() alone, with/without SIMD optimization. Here are results (the
> >>>>>>>>>>> median of 5 runs):
> >>>>>>>>>>>
> >>>>>>>>>>> HEAD: 208.879 ms
> >>>>>>>>>>> Patched: 40.983 ms
> >>>>>>>>>>>
> >>>>>>>>>>> The improvements look promising to me. But in a realistic pipeline the
> >>>>>>>>>>> parse is a small fraction of the work, so end-to-end gains could be
> >>>>>>>>>>> much smaller.
> >>>>>>>>>>>
> >>>>>>>>>>> Feedback is very welcome.
> >>>>>>>>>>>
> >>>>>>>>>> I may be missing something, but I wonder whether the fast path is relying on
> >>>>>>>>>> slightly different input semantics from the existing UUID parser.
> >>>>>>>>>>
> >>>>>>>>>> In particular, hex_decode_safe() is not a strict “32 hex characters only”
> >>>>>>>>>> decoder. It skips whitespace, which is fine for its existing callers, but I
> >>>>>>>>>> don’t think UUID input should treat whitespace inside the UUID body as
> >>>>>>>>>> ignorable.
> >>>>>>>>>
> >>>>>>>>> Good catch! hex_decode_safe() skips whitespaces so the patch accepts
> >>>>>>>>> the following UUID value, which is bad:
> >>>>>>>>>
> >>>>>>>>> select '019f00b5-7f8a-722f-b707-59f0ed25cd '::uuid;
> >>>>>>>>> uuid
> >>>>>>>>> --------------------------------------
> >>>>>>>>> 019f00b5-7f8a-722f-b707-59f0ed25cd00
> >>>>>>>>> (1 row)
> >>>>>>>>>
> >>>>>>>>>> Also, since hex_decode_safe() returns void, the UUID fast path
> >>>>>>>>>> cannot verify that exactly UUID_LEN bytes were produced.
> >>>>>>>>>
> >>>>>>>>> IIUC hex_decode_safe() does return the output length in bytes. So I
> >>>>>>>>> think we can fallback to the scalar UUID parser if
> >>>>>>>>> esctx.error_occurred is true or if the returned value is not 16.
> >>>>>>>>>
> >>>>>>>>
> >>>>>>>> You’re right, I misread that part. Checking both esctx.error_occurred and
> >>>>>>>> the returned length sounds good to me.
> >>>>>>>>
> >>>>>>>>>>
> >>>>>>>>>> So I think it would be safer either to pre-validate that the 32 source
> >>>>>>>>>> characters are all hex digits before calling hex_decode_safe(), or to use a
> >>>>>>>>>> UUID-specific strict hex decoder for this path. After that, a comment
> >>>>>>>>>> explaining why hex_decode_safe() is safe here would make the invariant much
> >>>>>>>>>> clearer.
> >>>>>>>>>
> >>>>>>>>> IIUC hex_decode_simd_helper() accepts only hex digits so we could
> >>>>>>>>> re-use it for UUID parsing. Let me check if the above idea of using
> >>>>>>>>> the return value works for us first.
> >>>>>>>>>
> >>>>>>>>
> >>>>>>>> That sounds reasonable. My main concern was to keep the fast path’s accepted
> >>>>>>>> input set identical to the scalar UUID parser. Falling back when the decoded
> >>>>>>>> length is not UUID_LEN, together with regression tests for whitespace cases,
> >>>>>>>> should address that.
> >>>>>>>>
> >>>>>>>>>>
> >>>>>>>>>> Could you also add a few regression tests for invalid inputs that contain
> >>>>>>>>>> whitespace inside otherwise fast-path-looking UUID strings? For example:
> >>>>>>>>>>
> >>>>>>>>>> ---------------------------------------------------------------
> >>>>>>>>>>
> >>>>>>>>>> SELECT 'a0eebc99 9c0b4ef8bb6d6bb9bd380a11'::uuid;
> >>>>>>>>>> SELECT 'a0eebc999c0b4ef8bb6d6bb9bd380a1 '::uuid;
> >>>>>>>>>> SELECT '{a0eebc999c0b4ef8bb6d6bb9bd380a1 }'::uuid;
> >>>>>>>>>> SELECT 'a0eebc99-9c0b-4ef8-bb6d-6bb9bd380a1 '::uuid;
> >>>>>>>>>> ---------------------------------------------------------------
> >>>>>>>>>>
> >>>>>>>>>> These should continue to be rejected in the same way as the scalar parser.
> >>>>>>>>>> Regards,
> >>>>>>>>>
> >>>>>>>>> Agreed.
> >>>>>>>>>
> >>>>>>>
> >>>>>>> I've attached the updated patch.
> >>>>>>>
> >>>>>>> Regards,
> >>>>>>>
> >>>>>>> --
> >>>>>>> Masahiko Sawada
> >>>>>>> Amazon Web Services: https://aws.amazon.com
> >>>>>>
> >>>>>> I noticed a few typos in the comments:
> >>>>>>
> >>>>>> src/backend/utils/adt/uuid.c
> >>>>>> line 56: “scalar implmentation” -> “scalar implementation”
> >>>>>> line 109: “swalled” -> “swallowed”
> >>>>>> line 110: “kepping” -> “keeping”
> >>>>>> line 118: “grammer” -> “grammar”
> >>>>>> line 119: “whitespaces” -> “whitespace”
> >>>>>>
> >>>>>> Could you fix them ?
> >>>>>
> >>>>> Oops, I fixed them and rechecked other places.
> >>>>>
> >>>>> I've attached the updated patch.
> >>>>>
> >>>>> Regards,
> >>>>>
> >>>>> --
> >>>>> Masahiko Sawada
> >>>>> Amazon Web Services: https://aws.amazon.com
> >>>>
> >>>> The code looks good to me now. I only noticed one small typo in the
> >>>> commit trailer: Reviwed-by should be Reviewed-by.
> >>>>
> >>>> Otherwise, it looks good. Thank you for fixing these issues.
> >>>>
> >>>
> >>> After spending more time on this patch, I find out two things:
> >>>
> >>> 1. USE_NO_SIMD doesn't work in uuid.c without including port/simd.h.
> >>> But including port/simd.h seems wrong as it doesn't use any SIMD
> >>> support functions.
> >>>
> >>> 2. hex_decode_safe() is faster than the current UUID parse
> >>> (isxdigit()+strtoul() approach) even without SIMD. I've created a
> >>> small benchmark test tool (attached as 0002 patch, not intended to be
> >>> pushed into the core), and measures UUID parsing performance of three
> >>> approaches: 'scalar' is the current string_to_uuid() that uses
> >>> isxdigit()+strtoul()), 'simd' uses hex_decode_safe() with SIMD, and
> >>> 'nosimd' uses hex_decode_safe() without SIMD, with different shapes of
> >>> UUIDs. Here are results:
> >>>
> >>> =# select path, shape, n_inputs, best_ms::numeric(10,3) from
> >>> uuid_parse_bench(100000, 5);
> >>> path | shape | n_inputs | best_ms
> >>> --------+------------------+----------+---------
> >>> scalar | canonical | 100000 | 22.661
> >>> simd | canonical | 100000 | 1.400
> >>> nosimd | canonical | 100000 | 1.652
> >>> scalar | bare32 | 100000 | 15.932
> >>> simd | bare32 | 100000 | 0.471
> >>> nosimd | bare32 | 100000 | 1.110
> >>> scalar | braced_canonical | 100000 | 17.330
> >>> simd | braced_canonical | 100000 | 1.088
> >>> nosimd | braced_canonical | 100000 | 1.314
> >>> scalar | braced_bare32 | 100000 | 15.942
> >>> simd | braced_bare32 | 100000 | 0.488
> >>> nosimd | braced_bare32 | 100000 | 1.141
> >>> scalar | dashed4 | 100000 | 16.185
> >>> simd | dashed4 | 100000 | 16.493
> >>> nosimd | dashed4 | 100000 | 16.403
> >>> scalar | invalid_hex | 100000 | 0.199
> >>> simd | invalid_hex | 100000 | 1.150
> >>> nosimd | invalid_hex | 100000 | 0.385
> >>> (18 rows)
> >>>
> >>> Each of shape means:
> >>> - 'canonical': 8x-4x-4x-4x-12x, what uuid_out() emits
> >>> - 'bare32': 32 contiguous hex digits
> >>> - 'braced_canonical': {8x-4x-4x-4x-12x}
> >>> - 'braced_bare32': {32 hex digits}
> >>> - 'dashed4': dash after every group of 4
> >>> - 'invalid_hdx': canonical but with a invalid digit
> >>>
> >>> 'nosimd' is 10x~ faster than 'scalar' in most cases. All paths are
> >>> mostly the same in 'dashed4' and 'invalid_hex' cases because 'simd'
> >>> and 'nosimd' fall back to the 'scalar' case. According to these
> >>> results, my conclusion is that we can use hex_decode_safe() for
> >>> canonical forms and 32 contiguous hex forms anyway, and let
> >>> hex_decode_safe() choose whether to use SIMD. We would win in either
> >>> case. We still use the current scalar approach for uncommon UUID forms
> >>> and error reporting purposes.
> >>>
> >>> Regards,
> >>>
> >>> --
> >>> Masahiko Sawada
> >>> Amazon Web Services: https://aws.amazon.com
> >>> <v4-0001-Optimize-UUID-parse-using-SIMD.patch><v4-0002-uuid_parse_bench-module.patch>
> >>
> >> A few comments on v4.
> >>
> >> 1 - 0001
> >> ```
> >> +static void
> >> +string_to_uuid(const char *source, pg_uuid_t *uuid, Node *escontext)
> >> +{
> >> + const char *body = source;
> >> + size_t len = strlen(source);
> >> ```
> >>
> >> I think it would be better to avoid strlen(). The old code processes at most UUID_LEN (16) byte pairs, so it does not need to scan arbitrarily far on malformed input. So, maybe we could use something like strnlen(source, 39) instead.
> >
> > While strnlen(source, 39) works there, 39 is a magic number and it's
> > tied to the current format check logic. What is the benefit of using
> > strnlen(source, 39) instead? I'm not sure it warrants having the magic
> > number.
>
> It doesn't have to be exactly 39; 1024 (long enough) would also work, or perhaps something based on UUID_LEN, such as UUID_LEN * 3. I think the main point is to avoid unbounded scanning on malformed input.
>
> The old code did not have this issue because it only examined as much input as needed based on UUID_LEN. The new fast path starts to use strlen(), so this would be a new risk introduced by the optimization.

I don't think the scan can be really unbounded. string_to_uuid()
receives a cstring, so by the time it is called the caller has already
walked or copied the whole string to produce it. So unless the
unbounded scan can be reached in some path I have overlooked, I'd
prefer to keep strlen() here. Happy to change it if you still think it
is worth it.

Regards,

--
Masahiko Sawada
Amazon Web Services: https://aws.amazon.com