A high-performance JSON library for Go with a drop-in encoding/json API.

Powered by AI + autoresearch automation. I don't know how the fuck it works, but it's fast as hell.

Decode beats bytedance/sonic by ≥ 10 % on all 11 gates (amd64, AVX-512BW): struct workloads cluster at +50–88 %, interface{} at +10–63 %, at a fraction of sonic's memory footprint. ARM64 (Neoverse-N1) struct decode runs +11–14 % over sonic at ~11 % of its resident memory.

Full scorecard — every corpus, encode figures, stdlib and goccy/go-json comparisons — lives in RESULTS.md.

go get github.com/wudi/jsonx

Requires Go 1.22+.

The package surface mirrors encoding/json:

import "github.com/wudi/jsonx"

type User struct {
    ID    int64  `json:"id"`
    Name  string `json:"name"`
    Email string `json:"email"`
}

var u User
err := jsonx.Unmarshal(data, &u)
out, err := jsonx.Marshal(u)

if !jsonx.Valid(data) { ... }

dec := jsonx.NewDecoder(r)
enc := jsonx.NewEncoder(w)

Same struct tags, same interface target types, same byte output up to shortest-representation equivalence.

• linux/amd64, darwin/amd64 — AVX-512BW string scan, whitespace skipper, and digit-emission kernels. Runtime-gated on cpuid.
• linux/arm64, darwin/arm64 — NEON + scalar arm64 kernels.
• Other GOOS/GOARCH — pure-Go SWAR fallback.

Assembly kernels are gated on a runtime CPU-feature check; absent the feature, the library falls through to pure Go without a rebuild.

Highlights (see program.md for the full experiment log):

• Type-specialized decoder plan cache — one compiled function per reflect.Type, struct fields written via unsafe.Add(ptr, offset).
• 8-byte prefix field dispatch replaces FNV-1a on the struct hot path.
• Slab-allocated interface{} boxes collapse hundreds of tiny mallocgc calls into one geometric-grown allocation.
• AVX-512BW string scan (23.8 GB/s on EPYC-Genoa) and whitespace skipper (single VPCMPUB $6 per 64-byte chunk); NEON equivalents via CMHI.
• Clinger + Eisel-Lemire float parsing; Schubfach formatting (BSL-1.0, ~41 % faster than strconv.AppendFloat).
• Encoder with one allocation per call: pooled []byte, pre-quoted struct keys, inlined encodeAny type switch.

go test ./...
cd bench && go test -bench=. -benchtime=3s -count=5 -benchmem

Coverage includes full-corpus round-trip vs encoding/json, dedicated correctness tests for each SIMD kernel, and an unconditional pure-Go fallback_test.go.

jsonx targets server workloads where payloads are large enough (> ~64 bytes) to amortize SIMD costs and structs are reused often enough to amortize plan-cache warmup. For short-lived processes or tiny payloads, encoding/json is often already fast enough. bytedance/sonic remains excellent when its JIT codegen and native structural indexer apply — notably on very large pretty-printed decode.

• Schubfach — Alexander Bolz, Drachennest (BSL-1.0). Go port derived from ByteDance sonic's native/f64toa.c.
• Eisel-Lemire detailedPowersOfTen — ported from Go's strconv.
• Avo — amd64 kernels generated with mmcloughlin/avo.
• Corpora — twitter.json, citm_catalog.json, canada.json from nativejson-benchmark; small.json and formatted synthetics are local.

Prior art: bytedance/sonic, goccy/go-json, encoding/json.

MIT