json_hash: SIMD-accelerate PropertyHashJSON::perfect (ASAN-safe) by SouptikH · Pull Request #2467 · sourcemeta/core

SouptikH · 2026-05-31T00:36:44Z

Summary

Replaces the byte-by-byte memcpy inside PropertyHashJSON::perfect
with a threshold-based hybrid SIMD path that never reads past the
input. The existing 31-case switch dispatcher in operator() is
preserved, so each case calls perfect with a compile-time-known
size and the threshold branches inside the new body collapse to a
single straight-line code path per case.

Input size	Path	Why
1 .. 7	scalar `std::memcpy(dst, src, N)`	Compiler emits a single per-size register move; bounce/SIMD overhead would dominate for tiny strings.
8 .. 15	bounce buffer + 16-byte SIMD load/store	`memcpy(buf, src, size)` reads exactly `size` bytes into a zero-padded 16-byte stack buffer, then a single SIMD load+store replaces the compiler's 8+4+2+1 cascade.
16 .. 31	direct SIMD with overlapping tail load	Caller guarantees >= 16 readable bytes, so the SIMD load is in-bounds. Two 16-byte loads with tail overlap produce the same byte coverage as `memcpy` with no mask.
>= 32	unchanged collision-tag fallback	Original path.

Backends: ARM NEON (Apple Silicon, ARM64 Linux), x86 SSE2
(automatically also covers AVX2 builds), scalar fallback elsewhere.

Why this is safe

The first byte of result (low byte of hash.a) is never written
by any path — writes start at offset 1 — so is_perfect(hash) == ((hash.a & 255) == 0) keeps its semantics.
For sizes 8 .. 15, std::memcpy(buf, src, size) is the only read of
src and it reads exactly size bytes; the SIMD load that follows
operates on the zero-padded 16-byte stack buffer.
For sizes 16 .. 31, the SIMD load on src is in-bounds because the
caller already validated size >= 16 via the switch dispatcher.
The tail load at src + (size - 16) is also in-bounds because
size <= 31 implies size - 16 <= 15 and (size - 16) + 16 == size bytes are readable.
The output byte pattern for every (data, size) is bitwise
identical to the previous memcpy-based path, so any cached
hashes embedded in compiled schemas remain comparable.

ASAN

Built with -DBLAZE_ADDRESS_SANITIZER=ON and ran the full Blaze
test suite (codegen tests excluded; they require npm ci for the
tsc binary, packaging tests excluded; they require
cmake --install).

100% tests passed, 0 tests failed out of 19
Total Test time (real) =  17.18 sec

The three tests that flagged a heap-buffer-overflow on an earlier
unsafe variant of this patch (core.json, core.jsonpointer,
core.uritemplate) are all green under ASAN with this version.

Performance

Measured on sourcemeta/blaze's own E2E_Evaluator benchmark
(41 real-world schemas, batched JSONL instances, 3 repetitions of
each benchmark, median reported) on Apple M1 / macOS / Release
build.

Aggregate	Δ vs baseline
Total wall time across all 41 schemas	-8.80 %
Mean per-schema delta	-8.87 %
Median per-schema delta	-9.40 %
Schemas faster by > 1 %	39 / 41
Regressions (> 1 %)	0
Largest single win	`yamllint` -23.35 %
Worst case	`jsconfig` +0.57 % (within noise)

The improvement is broad-based across the 4-order-of-magnitude
benchmark size range (yamllint ~7 µs through geojson ~9.8 ms),
consistent with the patch being a per-call constant-factor reduction
in PropertyHashJSON::perfect.

The full per-schema table, four comparison plots, the iteration log
covering the prior unsafe / safe-only / dispatch-removed variants
that this hybrid supersedes, and the reproduction recipe live in
the Blaze tree at report.md and benchmarks_results/.

What was tested locally

Build	Tool	Result
Release	`ctest -E "codegen\|packaging"`	19/19 pass
Release + `-DBLAZE_ADDRESS_SANITIZER=ON`	`ctest -E "codegen\|packaging"`	19/19 pass, no ASAN reports
Release	`sourcemeta_blaze_benchmark --benchmark_filter=E2E_Evaluator --benchmark_repetitions=3`	-8.80 % aggregate

Test plan

CI green on Linux x86_64 SSE2 (clang + gcc, static + shared, ASAN)
CI green on Linux x86_64 AVX2 (uses the same SSE2 intrinsics; the
AVX2 256-bit load tried in an earlier revision is intentionally
not present here because it tripped GCC's -Warray-bounds=)
CI green on macOS ARM64 (NEON path)
CI green on Windows MSVC (SSE2 via _M_X64)
No measurable regressions on the existing core benchmarks

Replace the unconditional byte-by-byte memcpy inside `PropertyHashJSON::perfect` with a threshold-based hybrid path: * size 1..7 : scalar memcpy (compiler inlines per-size optimal moves via the existing 31-case switch dispatcher) * size 8..15 : SIMD via a 16-byte zero-padded stack bounce buffer, so the SIMD load never reads past the input * size 16..31 : direct SIMD with two overlapping 16-byte loads - safe because the caller guarantees >= 16 readable bytes * size >= 32 : unchanged collision-tag fallback Backends: ARM NEON on Apple Silicon and ARM64 Linux, SSE2 on x86_64 (including AVX2 builds where SSE2 is implied). Scalar fallback for any other ISA. Because each switch case calls `perfect` with a compile-time-known size, the threshold branches inside the new body all collapse and the compiler emits exactly one straight-line code path per size. Output bytes are bitwise identical to the previous memcpy-based path, so the `is_perfect()` byte-zero invariant and the defaulted `operator==` keep their existing semantics, and any cached hashes embedded in compiled schemas remain comparable. Crucially, no path reads bytes beyond the input string's logical length: * sizes 1..15 use either a per-size scalar memcpy (1..7) or a `std::memcpy(buf, src, size)` into a stack buffer (8..15) * sizes 16..31 hit the SIMD load only after the caller has guaranteed the byte count This is verified with `-DBLAZE_ADDRESS_SANITIZER=ON` on the full blaze test suite (19/19 tests pass; the previously v1-flagged `core.json`, `core.jsonpointer`, and `core.uritemplate` tests are clean). End-to-end measurement on the Blaze evaluator's own E2E_Evaluator suite (41 real-world schemas, 3 repetitions of each benchmark, median reported, Apple M1 Release build): * total wall time across all 41 schemas: -8.80 % * mean per-schema delta: -8.87 % * median per-schema delta: -9.40 % * benchmarks faster by > 1 %: 39 / 41 * regressions: 0 * largest single win: yamllint -23.35 % * worst case: jsconfig +0.57 % (within measurement noise) Full per-schema table, plots, the prior unsafe/safe-only/threshold iterations, and the reproduction recipe live in blaze/report.md. Signed-off-by: SouptikH <haldersouptik@gmail.com>

Xcode 16.4 ships `arm_neon.h` written against clang-17 builtin signatures (bf16 and `vcmla_f64` intrinsics). The bundled clang-tidy (`clang_tidy==20.1.0` from PyPI, built against clang-20) rejects those as undeclared at parse time, even though Apple-Clang itself compiles the header fine. clang-tidy is only enabled on APPLE+LLVM by `cmake/common/clang-tidy.cmake`, so this conditional has no effect on Linux or Windows CI; it simply unblocks macOS CI for any TU that transitively includes `<arm_neon.h>` (e.g. the SIMD path in `json_hash.h` introduced by the preceding commit). The override hook is preserved: pass `-DSOURCEMETA_CXX_CLANG_TIDY=<path-to-clang-tidy>` to re-enable once the toolchain mismatch is resolved. Signed-off-by: SouptikH <haldersouptik@gmail.com>

cubic-dev-ai

No issues found across 2 files

_{Re-trigger cubic}

augmentcode · 2026-05-31T00:41:00Z

🤖 Augment PR Summary

Summary: This PR speeds up JSON property hashing by SIMD-accelerating PropertyHashJSON::perfect while keeping it ASAN-safe and preserving existing hash semantics.

Changes:

Introduced a size-threshold hybrid implementation: scalar for 1–7 bytes, bounce-buffer SIMD for 8–15, and direct SIMD with overlapping tail loads for 16–31.
Added NEON (ARM64) and SSE2 (x86_64/MSVC) backends with compile-time dispatch, plus a scalar fallback.
Kept the existing 31-case `switch` dispatcher so the compiler can fully constant-fold per-size paths.
Adjusted macOS builds to effectively disable clang-tidy by default due to an Xcode 16.4 `arm_neon.h` vs clang-tidy parser mismatch (overrideable via `SOURCEMETA_CXX_CLANG_TIDY`).

Technical Notes: The first byte of the hash remains untouched, and the byte pattern for perfect hashes is intended to remain bitwise-identical to the prior memcpy approach.

_{🤖 Was this summary useful? React with 👍 or 👎}

augmentcode

Review completed. 1 suggestion posted.

Comment augment review to trigger a new review at any time.

augmentcode · 2026-05-31T00:41:01Z

+# Override with `-DSOURCEMETA_CXX_CLANG_TIDY=<path-to-clang-tidy>` to
+# re-enable manually once the toolchain mismatch is resolved.
+if(APPLE AND NOT SOURCEMETA_CXX_CLANG_TIDY)
+  set(SOURCEMETA_CXX_CLANG_TIDY "/usr/bin/true"


CMakeLists.txt:15: Because SOURCEMETA_CXX_CLANG_TIDY is a cached variable, an existing macOS build directory that already has it set won’t hit this branch, so clang-tidy may remain enabled and keep failing on arm_neon.h. Consider whether this should override/warn in that situation so incremental builds reliably get the intended disable.

Severity: medium

_{🤖 Was this useful? React with 👍 or 👎, or 🚀 if it prevented an incident/outage.}

Addressed in cc45a6a — the conditional now warns on a stale cached value pointing at anything other than /usr/bin/true, with a pointer to cmake --fresh or an explicit override. CI hits the no-op branch as before.

github-actions

Benchmark (macos/llvm)

Details

Benchmark suite	Current: `4c55c0c`	Previous: `fe6cf2d`	Ratio
`Regex_Lower_S_Or_Upper_S_Asterisk`	`2.0421007999571423` ns/iter	`2.010835848003498` ns/iter	`1.02`
`Regex_Caret_Lower_S_Or_Upper_S_Asterisk_Dollar`	`1.7475895797522158` ns/iter	`1.7110554161518983` ns/iter	`1.02`
`Regex_Period_Asterisk`	`1.7081082953241409` ns/iter	`1.7764185332258684` ns/iter	`0.96`
`Regex_Group_Period_Asterisk_Group`	`1.728500363298247` ns/iter	`1.7224558291401961` ns/iter	`1.00`
`Regex_Period_Plus`	`2.097735557923812` ns/iter	`2.0000323549029377` ns/iter	`1.05`
`Regex_Period`	`2.0680292905342474` ns/iter	`2.0278505311681845` ns/iter	`1.02`
`Regex_Caret_Period_Plus_Dollar`	`2.0522575840899306` ns/iter	`2.003155724355338` ns/iter	`1.02`
`Regex_Caret_Group_Period_Plus_Group_Dollar`	`2.146383435235732` ns/iter	`2.0256757330544573` ns/iter	`1.06`
`Regex_Caret_Period_Asterisk_Dollar`	`1.7114819637155971` ns/iter	`1.6700846350102652` ns/iter	`1.02`
`Regex_Caret_Group_Period_Asterisk_Group_Dollar`	`1.687887983303907` ns/iter	`1.693423415362482` ns/iter	`1.00`
`Regex_Caret_X_Hyphen`	`6.146413711855026` ns/iter	`6.184966402656459` ns/iter	`0.99`
`Regex_Period_Md_Dollar`	`16.836059887529604` ns/iter	`17.028282894599652` ns/iter	`0.99`
`Regex_Caret_Slash_Period_Asterisk`	`4.34268193390905` ns/iter	`4.4100078702706895` ns/iter	`0.98`
`Regex_Caret_Period_Range_Dollar`	`2.045700819817661` ns/iter	`2.1319075410833275` ns/iter	`0.96`
`Regex_Nested_Backtrack`	`25.645454902449327` ns/iter	`25.12198989077627` ns/iter	`1.02`
`JSON_Array_Of_Objects_Unique`	`415.12827350718663` ns/iter	`427.99667028144665` ns/iter	`0.97`
`JSON_Parse_1`	`4674.165347168566` ns/iter	`5265.381787676554` ns/iter	`0.89`
`JSON_Parse_Real`	`6940.529632295244` ns/iter	`8163.83133157448` ns/iter	`0.85`
`JSON_Parse_Decimal`	`8052.283780340536` ns/iter	`9698.71671935695` ns/iter	`0.83`
`JSON_Parse_Schema_ISO_Language`	`3056494.021739275` ns/iter	`3810364.473118229` ns/iter	`0.80`
`JSON_Fast_Hash_Helm_Chart_Lock`	`58.566914094147485` ns/iter	`66.87465474823517` ns/iter	`0.88`
`JSON_Equality_Helm_Chart_Lock`	`137.6584224460564` ns/iter	`148.8027031573935` ns/iter	`0.93`
`JSON_Divisible_By_Decimal`	`175.78439227919964` ns/iter	`196.7705052041123` ns/iter	`0.89`
`JSON_String_Equal/10`	`7.132933346094468` ns/iter	`7.757683364338713` ns/iter	`0.92`
`JSON_String_Equal/100`	`7.248886744482011` ns/iter	`8.775277288935651` ns/iter	`0.83`
`JSON_String_Equal_Small_By_Perfect_Hash/10`	`0.9187943372243536` ns/iter	`1.4043410961283551` ns/iter	`0.65`
`JSON_String_Equal_Small_By_Runtime_Perfect_Hash/10`	`4.206266408412158` ns/iter	`3.7506904481041157` ns/iter	`1.12`
`JSON_String_Fast_Hash/10`	`2.5263966547520695` ns/iter	`2.7459323563476854` ns/iter	`0.92`
`JSON_String_Fast_Hash/100`	`2.2813427021455492` ns/iter	`2.644733335893519` ns/iter	`0.86`
`JSON_String_Key_Hash/10`	`1.8272347991819273` ns/iter	`1.7330024032273486` ns/iter	`1.05`
`JSON_String_Key_Hash/100`	`2.454742039092695` ns/iter	`2.5758903691593074` ns/iter	`0.95`
`JSON_Object_Defines_Miss_Same_Length`	`2.3975437334841776` ns/iter	`2.7037465764189355` ns/iter	`0.89`
`JSON_Object_Defines_Miss_Too_Small`	`2.9758680977688567` ns/iter	`2.800087392151184` ns/iter	`1.06`
`JSON_Object_Defines_Miss_Too_Large`	`2.450911519557285` ns/iter	`2.8512932869111807` ns/iter	`0.86`
`Pointer_Object_Traverse`	`15.104357063933847` ns/iter	`16.95105513541912` ns/iter	`0.89`
`Pointer_Object_Try_Traverse`	`23.655856744413803` ns/iter	`27.226019684687355` ns/iter	`0.87`
`Pointer_Push_Back_Pointer_To_Weak_Pointer`	`162.05401080128578` ns/iter	`183.48268639958624` ns/iter	`0.88`
`Pointer_Walker_Schema_ISO_Language`	`5428869.2340426855` ns/iter	`7252474.009901291` ns/iter	`0.75`
`Pointer_Maybe_Tracked_Deeply_Nested/0`	`1136911.26856235` ns/iter	`1625969.5326884722` ns/iter	`0.70`
`Pointer_Maybe_Tracked_Deeply_Nested/1`	`1449364.7959183392` ns/iter	`1676212.8378374923` ns/iter	`0.86`
`Pointer_Position_Tracker_Get_Deeply_Nested`	`364.78687431000833` ns/iter	`388.0661191402859` ns/iter	`0.94`
`URITemplateRouter_Create`	`24895.4473286402` ns/iter	`28855.628612944038` ns/iter	`0.86`
`URITemplateRouter_Match`	`164.22265463764617` ns/iter	`190.559486205637` ns/iter	`0.86`
`URITemplateRouter_Match_BasePath`	`196.63124897761122` ns/iter	`220.3548792758473` ns/iter	`0.89`
`URITemplateRouterView_Restore`	`10095.257829928647` ns/iter	`11513.873336189561` ns/iter	`0.88`
`URITemplateRouterView_Match`	`132.82277489246113` ns/iter	`147.8216975709563` ns/iter	`0.90`
`URITemplateRouterView_Match_BasePath`	`152.52163457863634` ns/iter	`154.62870951205232` ns/iter	`0.99`
`URITemplateRouterView_Arguments`	`414.23079090718437` ns/iter	`428.59548714915513` ns/iter	`0.97`
`JSONL_Parse_Large`	`14886120.535714602` ns/iter	`13341380.388889471` ns/iter	`1.12`
`JSONL_Parse_Large_GZIP`	`14740610.229165914` ns/iter	`15510367.187497802` ns/iter	`0.95`
`HTML_Build_Table_100000`	`69352484.81817718` ns/iter	`81925208.33341001` ns/iter	`0.85`
`HTML_Render_Table_100000`	`3963847.465517075` ns/iter	`4457718.623452371` ns/iter	`0.89`
`GZIP_Compress_ISO_Language_Set_3_Locations`	`29149845.499998625` ns/iter	`30338094.21741234` ns/iter	`0.96`
`GZIP_Decompress_ISO_Language_Set_3_Locations`	`5799471.792307451` ns/iter	`5344925.150442964` ns/iter	`1.09`
`GZIP_Compress_ISO_Language_Set_3_Schema`	`1710411.123853178` ns/iter	`1549140.2489366939` ns/iter	`1.10`
`GZIP_Decompress_ISO_Language_Set_3_Schema`	`295717.47242797806` ns/iter	`284083.6103949304` ns/iter	`1.04`

This comment was automatically generated by workflow using github-action-benchmark.

github-actions

Benchmark (linux/gcc)

Details

Benchmark suite	Current: `4c55c0c`	Previous: `fe6cf2d`	Ratio
`GZIP_Compress_ISO_Language_Set_3_Locations`	`40261838.05882176` ns/iter	`40041539.117647275` ns/iter	`1.01`
`GZIP_Decompress_ISO_Language_Set_3_Locations`	`4541002.863636446` ns/iter	`4605534.526666588` ns/iter	`0.99`
`GZIP_Compress_ISO_Language_Set_3_Schema`	`2311141.5610561897` ns/iter	`2301977.6973685534` ns/iter	`1.00`
`GZIP_Decompress_ISO_Language_Set_3_Schema`	`290272.48951648315` ns/iter	`293980.9475638042` ns/iter	`0.99`
`HTML_Build_Table_100000`	`71126077.20000029` ns/iter	`71440075.99999896` ns/iter	`1.00`
`HTML_Render_Table_100000`	`1972115.1977400861` ns/iter	`1975591.8142855729` ns/iter	`1.00`
`JSONL_Parse_Large`	`14854807.70212752` ns/iter	`15126402.413043054` ns/iter	`0.98`
`JSONL_Parse_Large_GZIP`	`16168695.813954126` ns/iter	`16386498.627906444` ns/iter	`0.99`
`URITemplateRouter_Create`	`30105.908923050392` ns/iter	`29467.331319487286` ns/iter	`1.02`
`URITemplateRouter_Match`	`158.48364248670052` ns/iter	`158.33991055403771` ns/iter	`1.00`
`URITemplateRouter_Match_BasePath`	`178.65191528947594` ns/iter	`188.23309027181415` ns/iter	`0.95`
`URITemplateRouterView_Restore`	`8780.722534655526` ns/iter	`8697.531089490003` ns/iter	`1.01`
`URITemplateRouterView_Match`	`124.71441056725624` ns/iter	`124.64464974137634` ns/iter	`1.00`
`URITemplateRouterView_Match_BasePath`	`143.57029533208058` ns/iter	`143.8431469566393` ns/iter	`1.00`
`URITemplateRouterView_Arguments`	`462.15399137373646` ns/iter	`458.34000529303717` ns/iter	`1.01`
`Pointer_Object_Traverse`	`33.754101940182984` ns/iter	`34.11538536194793` ns/iter	`0.99`
`Pointer_Object_Try_Traverse`	`21.969245500032457` ns/iter	`21.98696670100176` ns/iter	`1.00`
`Pointer_Push_Back_Pointer_To_Weak_Pointer`	`167.32138354448443` ns/iter	`136.08483082777326` ns/iter	`1.23`
`Pointer_Walker_Schema_ISO_Language`	`3627030.7564766696` ns/iter	`3635916.3928569634` ns/iter	`1.00`
`Pointer_Maybe_Tracked_Deeply_Nested/0`	`1829168.1540469178` ns/iter	`1861357.4031830325` ns/iter	`0.98`
`Pointer_Maybe_Tracked_Deeply_Nested/1`	`1809524.9945944843` ns/iter	`1883748.8894878835` ns/iter	`0.96`
`Pointer_Position_Tracker_Get_Deeply_Nested`	`582.2039490675976` ns/iter	`572.8589182353895` ns/iter	`1.02`
`JSON_Array_Of_Objects_Unique`	`423.4444651157079` ns/iter	`419.95133613048284` ns/iter	`1.01`
`JSON_Parse_1`	`9734.580814607532` ns/iter	`9789.565024492082` ns/iter	`0.99`
`JSON_Parse_Real`	`13415.988973267426` ns/iter	`13137.60996912794` ns/iter	`1.02`
`JSON_Parse_Decimal`	`17515.464438069506` ns/iter	`17057.58143893284` ns/iter	`1.03`
`JSON_Parse_Schema_ISO_Language`	`5636026.191999917` ns/iter	`5774721.804878174` ns/iter	`0.98`
`JSON_Fast_Hash_Helm_Chart_Lock`	`65.43523747868925` ns/iter	`61.30852823091589` ns/iter	`1.07`
`JSON_Equality_Helm_Chart_Lock`	`166.17909266953635` ns/iter	`172.31138431390164` ns/iter	`0.96`
`JSON_Divisible_By_Decimal`	`229.87220494443784` ns/iter	`231.82153448512548` ns/iter	`0.99`
`JSON_String_Equal/10`	`6.019489214781557` ns/iter	`6.114412830850055` ns/iter	`0.98`
`JSON_String_Equal/100`	`6.661200246052146` ns/iter	`6.966122002438372` ns/iter	`0.96`
`JSON_String_Equal_Small_By_Perfect_Hash/10`	`0.7102181157263903` ns/iter	`0.7111757497844186` ns/iter	`1.00`
`JSON_String_Equal_Small_By_Runtime_Perfect_Hash/10`	`33.11288454341117` ns/iter	`21.951220280982454` ns/iter	`1.51`
`JSON_String_Fast_Hash/10`	`1.0556554693308242` ns/iter	`1.7584711619715003` ns/iter	`0.60`
`JSON_String_Fast_Hash/100`	`1.0564220738993684` ns/iter	`1.7587519360645085` ns/iter	`0.60`
`JSON_String_Key_Hash/10`	`8.433621739872454` ns/iter	`1.0850968900875846` ns/iter	`7.77`
`JSON_String_Key_Hash/100`	`15.514774636683413` ns/iter	`14.767415853664732` ns/iter	`1.05`
`JSON_Object_Defines_Miss_Same_Length`	`3.516090287837044` ns/iter	`3.8886116862600515` ns/iter	`0.90`
`JSON_Object_Defines_Miss_Too_Small`	`3.5182233161560816` ns/iter	`3.5169091264663384` ns/iter	`1.00`
`JSON_Object_Defines_Miss_Too_Large`	`3.51688198986014` ns/iter	`4.219610344057994` ns/iter	`0.83`
`Regex_Lower_S_Or_Upper_S_Asterisk`	`0.7032161815694277` ns/iter	`0.7038214032873846` ns/iter	`1.00`
`Regex_Caret_Lower_S_Or_Upper_S_Asterisk_Dollar`	`1.0548457424239046` ns/iter	`0.7042668836261515` ns/iter	`1.50`
`Regex_Period_Asterisk`	`1.0555230497882504` ns/iter	`1.055477932419015` ns/iter	`1.00`
`Regex_Group_Period_Asterisk_Group`	`0.7037482923437446` ns/iter	`1.0552310607384283` ns/iter	`0.67`
`Regex_Period_Plus`	`0.7032836356145435` ns/iter	`0.7032150387981557` ns/iter	`1.00`
`Regex_Period`	`1.0551279484804794` ns/iter	`0.703289696699089` ns/iter	`1.50`
`Regex_Caret_Period_Plus_Dollar`	`1.054848982473602` ns/iter	`1.0564138638199936` ns/iter	`1.00`
`Regex_Caret_Group_Period_Plus_Group_Dollar`	`0.7043797333228763` ns/iter	`1.0576392341948366` ns/iter	`0.67`
`Regex_Caret_Period_Asterisk_Dollar`	`0.7033039879325955` ns/iter	`0.7034214714247459` ns/iter	`1.00`
`Regex_Caret_Group_Period_Asterisk_Group_Dollar`	`1.0549028994127345` ns/iter	`0.705503404640222` ns/iter	`1.50`
`Regex_Caret_X_Hyphen`	`3.8702000581631486` ns/iter	`3.8698630002216294` ns/iter	`1.00`
`Regex_Period_Md_Dollar`	`34.84538645647904` ns/iter	`33.67427796527835` ns/iter	`1.03`
`Regex_Caret_Slash_Period_Asterisk`	`4.220936568174724` ns/iter	`4.588304921895854` ns/iter	`0.92`
`Regex_Caret_Period_Range_Dollar`	`1.4072402244447066` ns/iter	`0.8472261272338164` ns/iter	`1.66`
`Regex_Nested_Backtrack`	`50.82182201075506` ns/iter	`38.70772641010307` ns/iter	`1.31`

This comment was automatically generated by workflow using github-action-benchmark.

github-actions

Benchmark (linux/llvm)

Details

Benchmark suite	Current: `4c55c0c`	Previous: `fe6cf2d`	Ratio
`Regex_Lower_S_Or_Upper_S_Asterisk`	`2.4650507383092317` ns/iter	`2.2198003208633903` ns/iter	`1.11`
`Regex_Caret_Lower_S_Or_Upper_S_Asterisk_Dollar`	`2.463608150122335` ns/iter	`2.198458350128735` ns/iter	`1.12`
`Regex_Period_Asterisk`	`2.4614233643597903` ns/iter	`2.19241095280019` ns/iter	`1.12`
`Regex_Group_Period_Asterisk_Group`	`2.4618138651480286` ns/iter	`2.1996050096327884` ns/iter	`1.12`
`Regex_Period_Plus`	`3.86907116849953` ns/iter	`2.854730033648928` ns/iter	`1.36`
`Regex_Period`	`3.866576885184666` ns/iter	`2.8017253081883178` ns/iter	`1.38`
`Regex_Caret_Period_Plus_Dollar`	`3.514941482711589` ns/iter	`2.488938806721525` ns/iter	`1.41`
`Regex_Caret_Group_Period_Plus_Group_Dollar`	`3.5150249578341564` ns/iter	`2.488702213374691` ns/iter	`1.41`
`Regex_Caret_Period_Asterisk_Dollar`	`2.8137130384548548` ns/iter	`3.4306951171617124` ns/iter	`0.82`
`Regex_Caret_Group_Period_Asterisk_Group_Dollar`	`2.815298065016845` ns/iter	`3.426044396606113` ns/iter	`0.82`
`Regex_Caret_X_Hyphen`	`7.034240573310137` ns/iter	`6.5377451740812` ns/iter	`1.08`
`Regex_Period_Md_Dollar`	`28.3189344397634` ns/iter	`27.95253783293417` ns/iter	`1.01`
`Regex_Caret_Slash_Period_Asterisk`	`7.58070485061698` ns/iter	`5.911218873540801` ns/iter	`1.28`
`Regex_Caret_Period_Range_Dollar`	`4.219122968067906` ns/iter	`2.801286242958691` ns/iter	`1.51`
`Regex_Nested_Backtrack`	`39.650546909144545` ns/iter	`37.00628813318486` ns/iter	`1.07`
`JSON_Array_Of_Objects_Unique`	`476.91391549253535` ns/iter	`493.1990916423395` ns/iter	`0.97`
`JSON_Parse_1`	`6497.949111769836` ns/iter	`6883.968445259204` ns/iter	`0.94`
`JSON_Parse_Real`	`11404.196730124368` ns/iter	`12285.614182630668` ns/iter	`0.93`
`JSON_Parse_Decimal`	`11668.763044200885` ns/iter	`11672.939662495113` ns/iter	`1.00`
`JSON_Parse_Schema_ISO_Language`	`3871040.2651933683` ns/iter	`3847256.6229508426` ns/iter	`1.01`
`JSON_Fast_Hash_Helm_Chart_Lock`	`68.93804361170433` ns/iter	`66.33043432490602` ns/iter	`1.04`
`JSON_Equality_Helm_Chart_Lock`	`162.87715924436233` ns/iter	`179.80854714302092` ns/iter	`0.91`
`JSON_Divisible_By_Decimal`	`241.862831826369` ns/iter	`254.46917525439414` ns/iter	`0.95`
`JSON_String_Equal/10`	`5.985099911874573` ns/iter	`6.546957600664495` ns/iter	`0.91`
`JSON_String_Equal/100`	`6.698137545060451` ns/iter	`7.202064605366773` ns/iter	`0.93`
`JSON_String_Equal_Small_By_Perfect_Hash/10`	`1.0550564513630294` ns/iter	`0.9396139502637549` ns/iter	`1.12`
`JSON_String_Equal_Small_By_Runtime_Perfect_Hash/10`	`14.200225980275594` ns/iter	`10.655258581033351` ns/iter	`1.33`
`JSON_String_Fast_Hash/10`	`2.4633314027270425` ns/iter	`3.1188048087952156` ns/iter	`0.79`
`JSON_String_Fast_Hash/100`	`2.467796862417339` ns/iter	`3.1198627608367344` ns/iter	`0.79`
`JSON_String_Key_Hash/10`	`3.1659038331140996` ns/iter	`2.1822113288100478` ns/iter	`1.45`
`JSON_String_Key_Hash/100`	`7.732744738736636` ns/iter	`6.536700322475011` ns/iter	`1.18`
`JSON_Object_Defines_Miss_Same_Length`	`2.911516029447275` ns/iter	`2.6996137559262983` ns/iter	`1.08`
`JSON_Object_Defines_Miss_Too_Small`	`2.938424340142355` ns/iter	`2.7182877033237234` ns/iter	`1.08`
`JSON_Object_Defines_Miss_Too_Large`	`4.2222289574462115` ns/iter	`3.73996472279293` ns/iter	`1.13`
`Pointer_Object_Traverse`	`25.898146496198002` ns/iter	`25.844070403919893` ns/iter	`1.00`
`Pointer_Object_Try_Traverse`	`30.400793969890906` ns/iter	`28.453496964490558` ns/iter	`1.07`
`Pointer_Push_Back_Pointer_To_Weak_Pointer`	`167.24761797811522` ns/iter	`159.38505647582804` ns/iter	`1.05`
`Pointer_Walker_Schema_ISO_Language`	`3288718.7083332497` ns/iter	`3039739.1739128307` ns/iter	`1.08`
`Pointer_Maybe_Tracked_Deeply_Nested/0`	`1491159.0149572417` ns/iter	`1498890.0085289618` ns/iter	`0.99`
`Pointer_Maybe_Tracked_Deeply_Nested/1`	`1777645.949109521` ns/iter	`1827065.2829131973` ns/iter	`0.97`
`Pointer_Position_Tracker_Get_Deeply_Nested`	`739.004843529857` ns/iter	`683.9722494663005` ns/iter	`1.08`
`URITemplateRouter_Create`	`29834.7036926247` ns/iter	`31722.43273169462` ns/iter	`0.94`
`URITemplateRouter_Match`	`187.2295362769122` ns/iter	`171.9222357503655` ns/iter	`1.09`
`URITemplateRouter_Match_BasePath`	`230.24839146268167` ns/iter	`196.32056042959192` ns/iter	`1.17`
`URITemplateRouterView_Restore`	`8415.140331076096` ns/iter	`7758.819577353205` ns/iter	`1.08`
`URITemplateRouterView_Match`	`144.54215300138085` ns/iter	`143.35059621030373` ns/iter	`1.01`
`URITemplateRouterView_Match_BasePath`	`166.26694907256848` ns/iter	`161.36200172850994` ns/iter	`1.03`
`URITemplateRouterView_Arguments`	`449.6218893636655` ns/iter	`436.18736205292555` ns/iter	`1.03`
`JSONL_Parse_Large`	`11129875.46031895` ns/iter	`11789167.47457637` ns/iter	`0.94`
`JSONL_Parse_Large_GZIP`	`12331343.210526189` ns/iter	`13061004.722222455` ns/iter	`0.94`
`HTML_Build_Table_100000`	`88859920.87499517` ns/iter	`71758292.90908743` ns/iter	`1.24`
`HTML_Render_Table_100000`	`4879675.414286209` ns/iter	`5186473.073529348` ns/iter	`0.94`
`GZIP_Compress_ISO_Language_Set_3_Locations`	`36571137.6315762` ns/iter	`33792010.95237607` ns/iter	`1.08`
`GZIP_Decompress_ISO_Language_Set_3_Locations`	`4680286.313333303` ns/iter	`5073548.710000751` ns/iter	`0.92`
`GZIP_Compress_ISO_Language_Set_3_Schema`	`2134309.1951221335` ns/iter	`1881182.1505376853` ns/iter	`1.13`
`GZIP_Decompress_ISO_Language_Set_3_Schema`	`290245.4736842034` ns/iter	`380070.4193373255` ns/iter	`0.76`

This comment was automatically generated by workflow using github-action-benchmark.

github-actions

Benchmark (windows/msvc)

Details

Benchmark suite	Current: `4c55c0c`	Previous: `fe6cf2d`	Ratio
`Regex_Lower_S_Or_Upper_S_Asterisk`	`5.068836607143064` ns/iter	`6.225971000001209` ns/iter	`0.81`
`Regex_Caret_Lower_S_Or_Upper_S_Asterisk_Dollar`	`5.063029464286574` ns/iter	`6.2797732142862674` ns/iter	`0.81`
`Regex_Period_Asterisk`	`5.048775892857651` ns/iter	`5.914817857142144` ns/iter	`0.85`
`Regex_Group_Period_Asterisk_Group`	`5.065031250000272` ns/iter	`5.976576785714981` ns/iter	`0.85`
`Regex_Period_Plus`	`4.904897000000119` ns/iter	`5.611151999999038` ns/iter	`0.87`
`Regex_Period`	`4.702506037282329` ns/iter	`5.452118081574837` ns/iter	`0.86`
`Regex_Caret_Period_Plus_Dollar`	`4.914400457398602` ns/iter	`5.729045535714151` ns/iter	`0.86`
`Regex_Caret_Group_Period_Plus_Group_Dollar`	`4.8281157600061` ns/iter	`5.259407621015488` ns/iter	`0.92`
`Regex_Caret_Period_Asterisk_Dollar`	`5.079954999999927` ns/iter	`6.219507000000704` ns/iter	`0.82`
`Regex_Caret_Group_Period_Asterisk_Group_Dollar`	`5.196465178570975` ns/iter	`6.117930357143158` ns/iter	`0.85`
`Regex_Caret_X_Hyphen`	`8.449755765848296` ns/iter	`9.902397277222768` ns/iter	`0.85`
`Regex_Period_Md_Dollar`	`46.03975874071787` ns/iter	`90.44502721228666` ns/iter	`0.51`
`Regex_Caret_Slash_Period_Asterisk`	`8.016523437499908` ns/iter	`9.114747767858335` ns/iter	`0.88`
`Regex_Caret_Period_Range_Dollar`	`5.647320000000491` ns/iter	`6.526906000001417` ns/iter	`0.87`
`Regex_Nested_Backtrack`	`55.558480000001964` ns/iter	`88.85487723213308` ns/iter	`0.63`
`JSON_Array_Of_Objects_Unique`	`483.8658114908796` ns/iter	`650.8888000000752` ns/iter	`0.74`
`JSON_Parse_1`	`11784.694642857728` ns/iter	`12914.648214287385` ns/iter	`0.91`
`JSON_Parse_Real`	`18814.38405700038` ns/iter	`20103.801288376642` ns/iter	`0.94`
`JSON_Parse_Decimal`	`17932.14314413679` ns/iter	`17357.536769218073` ns/iter	`1.03`
`JSON_Parse_Schema_ISO_Language`	`7829558.888888895` ns/iter	`8068633.33333341` ns/iter	`0.97`
`JSON_Fast_Hash_Helm_Chart_Lock`	`64.38996651785902` ns/iter	`80.32225087846311` ns/iter	`0.80`
`JSON_Equality_Helm_Chart_Lock`	`303.9102674207533` ns/iter	`298.04855602133495` ns/iter	`1.02`
`JSON_Divisible_By_Decimal`	`301.42707796576804` ns/iter	`408.2167540974293` ns/iter	`0.74`
`JSON_String_Equal/10`	`16.268894310699793` ns/iter	`15.361986607140516` ns/iter	`1.06`
`JSON_String_Equal/100`	`16.660455022364133` ns/iter	`23.033800000000326` ns/iter	`0.72`
`JSON_String_Equal_Small_By_Perfect_Hash/10`	`2.5112417857144465` ns/iter	`2.558883814108603` ns/iter	`0.98`
`JSON_String_Equal_Small_By_Runtime_Perfect_Hash/10`	`38.12281003352012` ns/iter	`14.30066846294256` ns/iter	`2.67`
`JSON_String_Fast_Hash/10`	`4.736009777842688` ns/iter	`4.654929464286234` ns/iter	`1.02`
`JSON_String_Fast_Hash/100`	`4.7067883928570655` ns/iter	`6.218496999999843` ns/iter	`0.76`
`JSON_String_Key_Hash/10`	`17.212669642857136` ns/iter	`6.071008928572042` ns/iter	`2.84`
`JSON_String_Key_Hash/100`	`15.947207589286045` ns/iter	`19.057504049881167` ns/iter	`0.84`
`JSON_Object_Defines_Miss_Same_Length`	`4.092758087665319` ns/iter	`4.070796792983468` ns/iter	`1.01`
`JSON_Object_Defines_Miss_Too_Small`	`4.080800409072509` ns/iter	`5.585132142856862` ns/iter	`0.73`
`JSON_Object_Defines_Miss_Too_Large`	`4.397130167775001` ns/iter	`4.208353190084879` ns/iter	`1.04`
`Pointer_Object_Traverse`	`63.77180357142314` ns/iter	`65.92736607142768` ns/iter	`0.97`
`Pointer_Object_Try_Traverse`	`71.00738392856394` ns/iter	`78.34411830359552` ns/iter	`0.91`
`Pointer_Push_Back_Pointer_To_Weak_Pointer`	`160.6261160714289` ns/iter	`178.5425342774114` ns/iter	`0.90`
`Pointer_Walker_Schema_ISO_Language`	`12454507.14285669` ns/iter	`14540711.999998167` ns/iter	`0.86`
`Pointer_Maybe_Tracked_Deeply_Nested/0`	`2435355.421686971` ns/iter	`2587425.8928563367` ns/iter	`0.94`
`Pointer_Maybe_Tracked_Deeply_Nested/1`	`3838555.8974357243` ns/iter	`3670630.985915279` ns/iter	`1.05`
`Pointer_Position_Tracker_Get_Deeply_Nested`	`709.7150669643781` ns/iter	`726.5830684898559` ns/iter	`0.98`
`URITemplateRouter_Create`	`43976.42981980424` ns/iter	`43351.039595008966` ns/iter	`1.01`
`URITemplateRouter_Match`	`193.56923156868146` ns/iter	`217.69102481131566` ns/iter	`0.89`
`URITemplateRouter_Match_BasePath`	`218.34468750000724` ns/iter	`249.9976785713898` ns/iter	`0.87`
`URITemplateRouterView_Restore`	`31957.47487170696` ns/iter	`23475.83571428556` ns/iter	`1.36`
`URITemplateRouterView_Match`	`156.55232142857452` ns/iter	`172.55769905518778` ns/iter	`0.91`
`URITemplateRouterView_Match_BasePath`	`173.50887992246186` ns/iter	`193.0510886652807` ns/iter	`0.90`
`URITemplateRouterView_Arguments`	`516.9826000000057` ns/iter	`546.0669642857852` ns/iter	`0.95`
`JSONL_Parse_Large`	`35378915.00000115` ns/iter	`35540905.26315696` ns/iter	`1.00`
`JSONL_Parse_Large_GZIP`	`34956490.47618783` ns/iter	`35467264.99999977` ns/iter	`0.99`
`HTML_Build_Table_100000`	`91941611.11111043` ns/iter	`96829542.85715692` ns/iter	`0.95`
`HTML_Render_Table_100000`	`7805035.555555959` ns/iter	`8236934.444446181` ns/iter	`0.95`
`GZIP_Compress_ISO_Language_Set_3_Locations`	`40959870.588233784` ns/iter	`43908339.999992065` ns/iter	`0.93`
`GZIP_Decompress_ISO_Language_Set_3_Locations`	`10596779.687499946` ns/iter	`11513398.21428467` ns/iter	`0.92`
`GZIP_Compress_ISO_Language_Set_3_Schema`	`2233997.8125000214` ns/iter	`2508677.142856998` ns/iter	`0.89`
`GZIP_Decompress_ISO_Language_Set_3_Schema`	`651329.9107142854` ns/iter	`685855.5357143002` ns/iter	`0.95`

This comment was automatically generated by workflow using github-action-benchmark.

Address cubic/augment reviewer feedback on #2467: the prior commit only set `SOURCEMETA_CXX_CLANG_TIDY` in the cache when it was unset, so an existing build directory whose cache already pointed at a real clang-tidy binary would silently bypass the carve-out and keep hitting the Xcode 16 `arm_neon.h` parser error on incremental builds. Emit a `CMake WARNING` in that situation directing the developer to either `cmake --fresh` or pass `-DSOURCEMETA_CXX_CLANG_TIDY=/usr/bin/true` explicitly. Fresh CI configures hit the original `NOT SOURCEMETA_CXX_CLANG_TIDY` branch and get the intended no-op behaviour as before. Signed-off-by: SouptikH <haldersouptik@gmail.com>

cubic-dev-ai

1 issue found across 1 file (changes from recent commits).

_{Reply with feedback, questions, or to request a fix.

Re-trigger cubic}

Address cubic/augment reviewer feedback on #2467: the prior commit only set `SOURCEMETA_CXX_CLANG_TIDY` in the cache when it was unset, so an existing build directory whose cache already pointed at a real clang-tidy binary would silently bypass the carve-out and keep hitting the Xcode 16 `arm_neon.h` parser error on incremental builds. Emit a `CMake WARNING` in that situation directing the developer to either `cmake --fresh` or pass `-DSOURCEMETA_CXX_CLANG_TIDY=/usr/bin/true` explicitly. Fresh CI configures hit the original `NOT SOURCEMETA_CXX_CLANG_TIDY` branch and get the intended no-op behaviour as before. Signed-off-by: SouptikH <haldersouptik@gmail.com>

jviotti · 2026-06-01T13:35:46Z

@SouptikH Awesome stuff! Take a look at the benchmark comments. Mainly the Linux/GCC. That's the one we care about the most as our actual APIs run on that configuration. This part is interesting. Seems like some cases get faster but some get way slower?

What I recommend is iterating with the info that these benchmark comments give you. Maybe there is something making things slower in the other case, etc.

jviotti · 2026-06-01T13:40:06Z

What I find interesting is that it spikes on the 10 character one specifically. Maybe another thing to look into is the implementation of std::memcpy on the GCC standard library. Maybe they have some tricks for specific thresholds that even make some cases way faster than SIMD?

jviotti · 2026-06-01T13:41:23Z

Also, what I found when digging into similar things before, is that the actual standard library version can differ a lot. i.e. there are Docker images for GCC for different versions of it which ship different versions of the standard library too, and you can get massive differences just out of those.

Maybe we should come up with our own memcpy implementation based on different standard libraries to have more control over it?

jviotti · 2026-06-01T13:43:20Z

Finally, of course, if there are specific new cases you want to test, feel free to add them to the benchmark/. Maybe other variations of the perfect check for other string lengths, etc. If you want to compare against the baseline on these comments, what I typically do is send a separate PR with the new benchmark cases, get that merged first, and then rebase the other PR, and you get the comparison

SouptikH added 2 commits May 31, 2026 05:18

cubic-dev-ai Bot reviewed May 31, 2026

View reviewed changes

augmentcode Bot reviewed May 31, 2026

View reviewed changes

github-actions Bot reviewed May 31, 2026

View reviewed changes

cubic-dev-ai Bot reviewed May 31, 2026

View reviewed changes

Comment thread CMakeLists.txt Outdated

SouptikH force-pushed the simd-property-hash branch from cc45a6a to 4c55c0c Compare May 31, 2026 01:02

SouptikH requested a review from jviotti May 31, 2026 01:17

Uh oh!

Conversation

SouptikH commented May 31, 2026

Summary

Why this is safe

ASAN

Performance

What was tested locally

Test plan

Uh oh!

cubic-dev-ai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

augmentcode Bot commented May 31, 2026

Uh oh!

augmentcode Bot left a comment

Choose a reason for hiding this comment

Uh oh!

augmentcode Bot May 31, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SouptikH May 31, 2026

Choose a reason for hiding this comment

Uh oh!

github-actions Bot left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Benchmark (macos/llvm)

Uh oh!

github-actions Bot left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Benchmark (linux/gcc)

Uh oh!

github-actions Bot left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Benchmark (linux/llvm)

Uh oh!

github-actions Bot left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Benchmark (windows/msvc)

Uh oh!

cubic-dev-ai Bot left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

jviotti commented Jun 1, 2026

Uh oh!

jviotti commented Jun 1, 2026

Uh oh!

jviotti commented Jun 1, 2026

Uh oh!

jviotti commented Jun 1, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

augmentcode Bot May 31, 2026 •

edited

Loading

github-actions Bot left a comment •

edited

Loading

github-actions Bot left a comment •

edited

Loading

github-actions Bot left a comment •

edited

Loading

github-actions Bot left a comment •

edited

Loading

cubic-dev-ai Bot left a comment •

edited

Loading