Re: [PATCH] Use Boyer-Moore-Horspool for simple LIKE contains patterns

Hi Haibo,

Thanks for the detailed matrix. I reproduced the low-entropy regression
region and agree with your reading: the cost is driven by alphabet size
and mismatch position rather than literal length, so adjusting
LIKE_BMH_MIN_LITERAL_LEN cannot describe the boundary.

I would like to get your thoughts on the general approach first.
I've attached a rough PoC patch just for reference; it still needs
some cleanup, and a proper patch along with the full numbers
will follow later.

Needle-only rule
----------------

I first tried a needle-only check: skip BMH when the trailing bytes of the
literal are periodic. While it handles cases like repeat('a') and certain
repeat('ab') patterns, it does nothing when the periodicity lies in the
haystack rather than the literal (e.g., repeat('abcd') with a 64-byte
literal stays around 15x slower). It also needlessly forces literals like
'%aaaa%' to the generic matcher on ordinary text where BMH would otherwise
win. A preparation-time test on the pattern alone does not seem viable.

Bounded work with resume
------------------------

The approach I am leaning towards is a runtime guard along the lines of
your suggestion, structured as follows:

- like_bmh_search() checks the guard (last) byte first and only counts
inner-loop byte comparisons beyond that guard. The fast path where the
last byte differs has zero accounting overhead, preserving standard BMH
performance.

- When the comparison count exceeds a given threshold (currently
prototyping slen / 2), BMH aborts and reports the offset up to which it
has ruled out matches.

- LikeMatchText() then delegates only the remaining, unsearched suffix
(backed up to a character boundary, tested under UTF-8) to
GenericMatchText().
Because the pattern begins with '%', evaluating the suffix yields the
exact same semantics without rescanning the entire string from the
start.

Preliminary numbers (100,000 rows, best of 5, ms; HEAD / v3 / v3 + bounded
work):

repeat('a',1024) LIKE '%~aaaaaaaaaaaaaaa%' 111 / 552 / 131
repeat('a',1024) LIKE '%~' || 63 x 'a' || '%' 113 / 2007 / 135
repeat('a',1024) LIKE '%aaaaaaaaaaaaaaa~%' 1896 / 353 / 325
English text LIKE '%worst of crimes%' (miss) 162 / 65 / 47

In a microbenchmark, the worst-case late-mismatch penalty drops from
~100x down to ~4x at 1024 bytes (~2.3x at 32 bytes). Early mismatch and
match-present cases remain unaffected or slightly faster thanks to the
guard byte.

A residual 1.2-4x overhead remains in cases where the generic matcher
quickly bails out (e.g., the first pattern byte is absent from the
haystack)
while BMH exhausts its comparison budget before falling back.
Tightening the budget lowers this ceiling, but also trims BMH's advantages
on
benign inputs.

As a note, the attached PoC is an incremental patch on top of v3 rather
than HEAD.

Regards,
Atsushi Ogawa

2026年9月15日(火) 10:24 Haibo Yan <tristan(dot)yim(at)gmail(dot)com>:

> On Fri, Jul 17, 2026 at 3:23 AM Atsushi Ogawa
> <atsushi(dot)ogawa001(at)gmail(dot)com> wrote:
> >
> > Hi Greg,
> >
> > Thanks for the careful review. I have attached a v2 patch.
> >
> > > git grep shows we already use BMH in src/backend/utils/adt/varlena.c
> > > Worth acknowledging that in a code comment somewhere? I didn't see any
> > > obvious advantage to refactoring things out at quick glance, but a
> mention
> > > might be nice.
> >
> > Agreed. I added a comment at the top of like_bmh.c that
> cross-references the
> > existing Boyer-Moore-Horspool implementation in varlena.c and explains
> why I
> > kept the implementations separate. The varlena.c code searches one
> > (haystack, needle) pair with an adaptively sized skip table, whereas the
> LIKE
> > path interprets its internal backslash escapes while extracting the
> literal
> > and caches the prepared search state in FmgrInfo for use across rows. I
> did
> > not find a clean way to share that machinery without introducing more
> coupling
> > than seemed useful.
> >
> > > + * by '%' wildcards. Remove backslash escapes while building the
> search
> > > + * state.
> > >
> > > Slightly off comment. This is for like_bmh_pattern_is_eligible - we
> are not
> > > removing here, just skipping things when we count.
> >
> > Right. I reworded the comment to say that the eligibility check skips
> > backslash escapes while counting the literal length. The escapes are
> removed
> > later, when the search state is built.
> >
> > > if (i + 1 >= plen - 1)
> > >
> > > Worth a comment to explain that we are catching the '%foo\%' case here.
> >
> > Added. The new comment explains that this rejects patterns such as
> > '%foo\%', where the backslash escapes the closing '%' rather than a
> literal
> > byte.
> >
> > > pattern_stable = get_fn_expr_arg_stable(flinfo, 1);
> > >
> > > /*
> > > * ScalarArrayOpExpr invokes the operator once per array element. The
> > > * array expression can be stable while the pattern passed to this
> function
> > > * changes between calls, so it must not use a cached search state.
> > > */
> > > if (flinfo->fn_expr != NULL && IsA(flinfo->fn_expr, ScalarArrayOpExpr))
> > > pattern_stable = false;
> > >
> > > My first thought was to make this an if/else so we don't reclobber, but
> > > seeing how later on we check collation every time, I'm wondering if we
> > > shouldn't just check the pattern as well every time via a memcmp like
> > > regexp.c does in RE_compile_and_cache (and remove that block above).
> > > So we store it verbatim in the like_bmh_init() function with memcpy,
> then
> > > make the check inside like_bmh_match() that looks like this:
> > >
> > > unlikely(collation has changed)
> > >
> > > into:
> > >
> > > unlikely(
> > > collation has changed
> > > OR pattern length has changed
> > > OR pattern itself has changed (e.g. memcmp true)
> > > )
> > >
> > > Also means you could then roll get_fn_expr_arg_stable into that big
> old ||
> > > grouping, and remove pattern_stable entirely.
> >
> > I implemented the suggested verbatim-pattern cache and benchmarked it
> directly
> > against the initial patch's structural-stability design. The test
> scanned two
> > million rows per transaction, with a warmup followed by the median of
> seven
> > pgbench runs of 40 transactions each. The benchmark used an AMD EPYC
> 7763
> > host with 8 vCPUs, GCC 11.4.0, and an -O2 -g build, using a UTF-8
> database
> > with C locale. The results below are median latency per scan:
> >
> > case initial patch memcmp vs.
> initial
> > ------------------------------------ ------------- --------
> -----------
> > constant, 4-byte literal 63.7 ms 65.1 ms
> +2.3%
> > constant, 32-byte literal 48.8 ms 48.4 ms
> -0.8%
> > constant, 4-byte literal, 8-byte input 43.2 ms 44.0 ms
> +1.8%
> > non-constant, fixed value at runtime 92.9 ms 57.0 ms
> -38.6%
> > non-constant, changes on every row 91.6 ms 172.4 ms
> +88.2%
> >
> > The per-row length check and memcmp were therefore not particularly
> expensive
> > for stable constant patterns. The more important tradeoff involved
> > non-constant patterns. When the value remained fixed at runtime, the
> verbatim
> > cache was faster because it could use BMH. When the pattern changed on
> every
> > row, however, it was substantially slower than the initial patch, which
> sends
> > that case to the existing generic matcher. The verbatim variant had to
> repeat
> > the eligibility check and rebuild the 256-entry skip table for every row.
> >
> > I then tested a hybrid of the two approaches. Patterns that
> > get_fn_expr_arg_stable() identifies as a Const or external Param keep the
> > existing comparison-free search state. An eligible non-stable pattern
> stores
> > its verbatim bytes and is revalidated with a length check and memcmp.
> On the
> > first mismatch, the state is changed permanently to the generic marker.
> The
> > mismatching row and all later rows use the existing matcher; the
> eligibility
> > check and skip-table build are never repeated.
> >
> > ScalarArrayOpExpr still has to be classified as non-stable, since its
> array
> > expression can be a Const while the operator receives a different
> element on
> > each call. It now uses the same revalidation path and falls back
> permanently
> > if the elements differ.
> >
> > I reran the comparison on aarch64 using two clean build trees based on
> the
> > same source revision and configured with the same options. Both servers
> used
> > the same data directory. The table contained two million 32-byte
> strings, a
> > fixed pattern column, and an alternating pattern column. Parallel query
> was
> > disabled, each server was warmed before measurement, and the server
> order was
> > alternated in ABBA order. The figures below are medians of 16 EXPLAIN
> > (ANALYZE, TIMING OFF) runs:
> >
> > case initial patch hybrid vs. initial
> > -------------------------------- ------------- -------- -----------
> > constant pattern 194.1 ms 185.6 ms -4.4%
> > non-constant, fixed at runtime 299.8 ms 199.9 ms -33.3%
> > non-constant, changes every row 303.4 ms 304.9 ms +0.5%
> > generic fallback control 288.5 ms 289.5 ms +0.3%
> >
> > The constant-pattern difference appears to be a compiler-dependent
> code-layout
> > effect rather than a benefit of the hybrid design, so I do not interpret
> it as
> > a general speedup. More importantly, the runtime-fixed case captures the
> > benefit of the verbatim cache, while the row-varying case tracks the
> generic
> > fallback control instead of rebuilding the 256-entry skip table for
> every row.
> >
> > The attached v2 patch uses this hybrid design. Thus the common stable
> path
> > does not pay a memcmp, runtime-fixed non-constant values can use BMH,
> and a
> > pattern that is observed to vary falls back without any rebuild penalty.
> >
> > > Hm...that collation test and message is already caught and done by
> > > GenericMatchText, so you could throw !OidIsValid(collation) into that
> ||
> > > group as well, and remove the ereport section entirely. It then falls
> > > through later to GenericMatchText, which complains about the collation
> > > there.
> >
> > Done. The invalid-collation case is now included in the rejection group
> and
> > falls through to GenericMatchText. I removed the duplicate ereport
> block from
> > like_bmh.c.
> >
> > > It did have one test failure:
> > >
> > > @@ -151,8 +151,8 @@
> > > p | matched
> > > --------+---------
> > > %abcd% | t
> > > - %b%e% | f
> > > %b_d% | t
> > > + %b%e% | f
> > > %wxyz% | f
> > > (4 rows)
> > >
> > > I think it's from the "Row-varying patterns must use the generic
> matcher."
> > > test.
> >
> > Thanks for catching this. This was a locale-dependent sort-order issue
> in the
> > test, not a matcher failure. The query now uses ORDER BY p COLLATE "C".
> >
> > I retested the revised patch against PostgreSQL HEAD 0348090: all 246
> core
> > regression tests passed, including like_bmh, and all four contrib/pg_trgm
> > tests passed.
> >
> > Thanks,
> > Atsushi Ogawa
> >
> > 2026年7月15日(水) 3:29 Greg Sabino Mullane <htamfids(at)gmail(dot)com>:
> >>
> >> Great idea, love seeing the speedups! Also appreciate the background,
> detailed explanation, and benchmarks. Quick code review:
> >>
> >> git grep shows we already use BMH in src/backend/utils/adt/varlena.c
> >> Worth acknowledging that in a code comment somewhere? I didn't see any
> obvious advantage to refactoring things out at quick glance, but a mention
> might be nice.
> >>
> >> > + * by '%' wildcards. Remove backslash escapes while building the
> search state.
> >>
> >> Slightly off comment. This is for like_bmh_pattern_is_eligible - we are
> not removing here, just skipping things when we count.
> >>
> >> > if (i + 1 >= plen - 1)
> >>
> >> Worth a comment to explain that we are catching the '%foo\%' case here.
> >>
> >>
> >> > pattern_stable = get_fn_expr_arg_stable(flinfo, 1);
> >> >
> >> > /*
> >> > * ScalarArrayOpExpr invokes the operator once per array element. The
> >> > * array expression can be stable while the pattern passed to this
> function
> >> > * changes between calls, so it must not use a cached search state.
> >> > */
> >> > if (flinfo->fn_expr != NULL && IsA(flinfo->fn_expr,
> ScalarArrayOpExpr))
> >> > pattern_stable = false;
> >>
> >> My first thought was to make this an if/else so we don't reclobber, but
> seeing how later on we check collation every time, I'm wondering if we
> shouldn't just check the pattern as well every time via a memcmp like
> regexp.c does in RE_compile_and_cache (and remove that block above). So we
> store it verbatim in the like_bmh_init() function with memcpy, then make
> the check inside like_bmh_match() that looks like this:
> >>
> >> unlikely(collation has changed)
> >>
> >> into:
> >>
> >> unlikely(
> >> collation has changed
> >> OR pattern length has changed
> >> OR pattern itself has changed (e.g. memcmp true)
> >> )
> >>
> >> Also means you could then roll get_fn_expr_arg_stable into that big old
> || grouping, and remove pattern_stable entirely.
> >>
> >> Hm...that collation test and message is already caught and done by
> GenericMatchText, so you could throw !OidIsValid(collation) into that ||
> group as well, and remove the ereport section entirely. It then falls
> through later to GenericMatchText, which complains about the collation
> there.
> >>
> >> Anyway, the patch compiled cleanly against d15a6bc2 (Tue Jul 14
> 10:28:04 2026 +0200)
> >>
> >> It did have one test failure:
> >>
> >> @@ -151,8 +151,8 @@
> >> p | matched
> >> --------+---------
> >> %abcd% | t
> >> - %b%e% | f
> >> %b_d% | t
> >> + %b%e% | f
> >> %wxyz% | f
> >> (4 rows)
> >>
> >> I think it's from the "Row-varying patterns must use the generic
> matcher." test.
> >>
> >>
> >> Cheers,
> >> Greg
> >>
>
> Hi Ogawa-san,
>
> I did some more testing of the BMH fast path, specifically to check
> whether the
> repetitive-input regression I mentioned is just one adversarial
> construction or
> part of a broader pattern.
>
> I ran a matrix varying haystack structure, literal length, mismatch
> position,
> and haystack length, with the existing LIKE matcher and the patch's BMH
> search
> in the same binary. The overall result is actually quite favorable to BMH:
> it
> wins most of the tested cases, often by a large margin. However, there is
> also a
> fairly well-defined regression region on low-entropy inputs when the
> backwards
> comparison fails late.
>
> A few representative numbers are:
>
> haystack literal length existing LIKE
> BMH ratio
> repeat('a', 1024) 16 2.14 ms
> 20.78 ms 9.7x
> repeat('a', 1024) 64 2.13 ms
> 66.93 ms 31.4x
> repeat('ab', ...), 1024 64 2.12 ms
> 33.24 ms 15.7x
> repeat('abcd', ...), 1024 64 2.12 ms
> 16.70 ms 7.9x
> random alphabet=4, 1024 4 2.24 ms 7.41 ms
> 3.3x
> English text, 1024 16 2.23 ms
> 1.40 ms 0.63x
>
> The important part seems to be the combination of low effective alphabet
> size
> and mismatch position, rather than literal length by itself.
>
> For example, against `repeat('a', 1024)`, a literal shaped roughly as
>
> ~aaaaaaaaaaaaaaa
>
> causes Horspool to compare almost the whole literal backwards before
> failing,
> while the skip for `a` is only one byte. For a 16-byte literal I counted
> 1009
> candidate alignments and 16 comparisons per alignment. The existing LIKE
> matcher
> has almost the opposite behavior here: its first literal byte (`~`) is
> absent
> from the haystack, so it rejects candidates very cheaply.
>
> Mismatch position changes the result dramatically. With the same 1024-byte
> repeated-`a` input and a 16-byte literal I measured approximately:
>
> immediate mismatch: BMH / existing LIKE = 0.04x
> middle mismatch: = 0.34x
> late mismatch: = 9.7x
>
> So BMH can be much faster or much slower on very similar inputs.
>
> This also means that increasing `LIKE_BMH_MIN_LITERAL_LEN` does not appear
> to
> address the issue. The worst measured regression actually increased with
> literal
> length:
>
> 4 bytes 4.3x
> 8 bytes 5.1x
> 16 bytes 9.9x
> 32 bytes 18.5x
> 64 bytes 33.7x
>
> The regression is not universal. In this test set, English text and random
> data
> over medium/large alphabets did not show >2x regressions, and for literals
> >= 8
> bytes they did not show meaningful regressions at all. So I would describe
> this
> as a narrow but systematic low-entropy case rather than a general
> problem with BMH.
>
> I also tried looking for a cheap needle-only rule that could avoid the
> bad cases.
> There are some useful signals in the skip table, but they have substantial
> false
> positives. More fundamentally, the same byte-identical literal can be a
> large
> win or a large loss depending only on the haystack/match position, so a
> preparation-time test based only on the literal cannot completely solve
> this.
>
> This seems related to the concerns raised in the earlier BMH/LIKE
> discussions:
>
>
> https://www.postgresql.org/message-id/CALkFZpcbipVJO%3DxVvNQMZ7uLUgHzBn65GdjtBHdeb47QV4XzLw%40mail.gmail.com
>
> and Tom Lane's later discussion here:
>
> https://www.postgresql.org/message-id/3811203.1675907383%40sss.pgh.pa.us
>
> There is also the recent related thread here:
>
>
> https://www.postgresql.org/message-id/flat/88272f23-19b4-493d-bdd7-258218b74881%40gmail.com
>
> Given that this is a performance optimization, I think it would be useful
> to
> decide explicitly how much regression on this class of inputs is
> acceptable, or
> whether some bounded-work fallback would make sense. A runtime guard might
> be
> more promising than a needle-only eligibility rule, since it could notice
> that
> the search is doing unusually large amounts of work without requiring a
> separate
> scan of the haystack.
>
> I don't think these results argue against using BMH in general — in the
> same
> matrix it was substantially faster in most cases — but they do suggest that
> the current literal-length threshold alone doesn't describe the
> profitability
> boundary.
>
> Regards,
> Haibo
>