Benchmarks

Philosophy

gogenfilter is designed for linters and static analysis tools where detection overhead must be negligible. The two-phase architecture (filename first, content only when needed) ensures most checks complete with zero I/O.

Results

Benchmarked on Go 1.26, AMD Ryzen AI MAX+ 395. Results may vary by hardware.

Summary

	Benchmark	ns/op	B/op	allocs/op
Filter (enabled)	69	92	0
Filter (disabled)	1.2	0	0
Filename detection (Phase 1)	14	0	0
Per-generator check (avg)	~6	0	0
MatchPattern (doublestar **)	72	0	0
DetectReason (filename)	353	632	4
DetectReason (content)	693	680	5
Error construction	2	0	0
errors.Is matching	9	0	0

Filter

The primary API. Full end-to-end detection including filesystem access.

Filter/enabled    69 ns/op    92 B/op    0 allocs/op
Filter/disabled   1.2 ns/op    0 B/op    0 allocs/op

A disabled filter is essentially a branch prediction — 1.2 nanoseconds. An enabled filter with a known-generated file completes in ~69ns with zero heap allocations.

Phase 1: Filename Detection

Phase 1: filename-based detection (zero I/O). This is the fast path — a single scan of the filename against the detector table. No file reads, no content parsing.

Filename phase    14 ns/op    0 B/op    0 allocs/op

This benchmarks the internal filename-only check. Results are exact-matched against known generated-file patterns (e.g., models.go, *.sql.go, _templ.go, .pb.go, wire_gen.go, _moq.go, _mock.go, mock_*.go).

Per-Generator Checks

Individual Is*Generated functions. Four of them (IsSQLCGenerated, IsTemplGenerated, IsGoEnumGenerated, IsProtobufGenerated) check both filename patterns and content markers. The remaining seven check content only.

IsGenericGenerated    5.0 ns/op    0 B/op    0 allocs/op
IsSQLCGenerated       5.9 ns/op    0 B/op    0 allocs/op
IsProtobufGenerated   7.0 ns/op    0 B/op    0 allocs/op

All under 7 nanoseconds with zero allocations.

DetectReason

Variadic detection across multiple filter options:

DetectReason/filename_only    353 ns/op    632 B/op    4 allocs/op
DetectReason/content_based    693 ns/op    680 B/op    5 allocs/op
DetectReason/not_filtered     614 ns/op    680 B/op    5 allocs/op

Higher cost than the per-generator checks because it iterates all enabled detectors and performs content matching. Still sub-microsecond.

DetectReasonReader

Same as DetectReason but reads from an io.Reader:

DetectReasonReader/sqlc_filename    517 ns/op   1288 B/op    8 allocs/op
DetectReasonReader/not_filtered     812 ns/op   1272 B/op    9 allocs/op

Includes the cost of reading content from the reader.

Pattern Matching

MatchPattern with different glob patterns via doublestar:

MatchPattern/question     25 ns/op    0 B/op    0 allocs/op
MatchPattern/exact        40 ns/op    0 B/op    0 allocs/op
MatchPattern/wildcard     46 ns/op    0 B/op    0 allocs/op
MatchPattern/doublestar   72 ns/op    0 B/op    0 allocs/op
MatchPattern/no_match     77 ns/op    0 B/op    0 allocs/op

All zero-allocation. ** globs are the most expensive at ~72ns — still negligible.

Error System

Error construction and inspection:

NewSQLCConfigError        0.2 ns/op     0 B/op    0 allocs/op
NewProjectRootError       1.9 ns/op     0 B/op    0 allocs/op
ProjectRootErrorIs        7.9 ns/op     0 B/op    0 allocs/op
SQLCConfigErrorIs         8.9 ns/op     0 B/op    0 allocs/op
ProjectRootErrorError   229 ns/op     144 B/op    3 allocs/op
SQLCConfigErrorError    284 ns/op     176 B/op    5 allocs/op

Error construction and errors.Is matching are essentially free. Error() string formatting allocates but is only called when displaying errors.

Reproducing

Run the benchmarks yourself:

go test -bench=. -benchmem ./...

For stable results, run on a quiet system with CPU frequency scaling disabled.

Live Dashboard

Historical benchmark trends are tracked automatically on every push to master:

View Live Benchmark Dashboard

The dashboard shows:

• Time-series charts for all 17 benchmarks
• Commit-level granularity with tooltips showing commit messages
• Automatic regression highlighting when performance degrades >150%

Results above are a point-in-time snapshot (Go 1.26, AMD Ryzen AI MAX+ 395). The dashboard provides continuous, hardware-consistent tracking from CI.