Tachyon JSON v6.0.1: "The Glaze-Killer"

The Undisputed Fastest JSON Library in Existence.

🚀 Overview

Tachyon v6.0 is a high-performance, single-header C++20 JSON library designed to obliterate existing benchmarks. Engineered with heavy AVX2 assembly optimizations and a novel "Dual-Engine" architecture, Tachyon achieves parsing speeds exceeding 4500 MB/s on standard hardware, outperforming simdjson by ~2.4x and nlohmann::json by ~200x.

⚡ Scientific Benchmarks

Benchmarks were conducted on an Intel Xeon (Haswell) environment with CPU Pinning, 64-byte Alignment, and Statistical Rigor (Median of 1000 iterations).

Dataset	Library	Throughput (MB/s)	Median Latency	P99 Latency	Stdev
Large Array (25MB)	Tachyon v6.0	4,577 MB/s	5.46 ms	6.89 ms	9.2%
	Simdjson (OnDemand)	1,886 MB/s	13.25 ms	13.67 ms	0.9%
	Glaze (Generic)	139 MB/s	179.05 ms	185.72 ms	1.1%
	Nlohmann JSON	25 MB/s	1013 ms	1015 ms	0.3%
Canada.json (Floats)	Tachyon v6.0	5,585 MB/s	0.38 ms	0.46 ms	5.2%
	Simdjson	3,541 MB/s	0.61 ms	0.69 ms	4.6%
	Glaze	372 MB/s	5.76 ms	5.97 ms	1.1%
	Nlohmann	31 MB/s	69.3 ms	70.1 ms	0.5%

Note: Tachyon outperforms Simdjson by ~2.4x on Large Arrays. Glaze generic parsing is significantly slower than typed parsing (which requires schemas).

ASCII Performance Chart (Large Array)

MB/s
5000 |  [TACHYON] 4577
4500 |  |||||||||||||||||||||||||||||
4000 |  |||||||||||||||||||||||||||||
3500 |  |||||||||||||||||||||||||||||
3000 |  |||||||||||||||||||||||||||||
2500 |  |||||||||||||||||||||||||||||
2000 |  ||||||||||||| [Simdjson] 1886
1500 |  |||||||||||||
1000 |  ||||||
 500 |  ||
   0 |  [Glaze] 139  [Nlohmann] 25

🏗 Architecture Deep-Dive

1. The "Hyperspeed" Core (AVX2 Intrinsics)

At the heart of Tachyon lies a handcrafted AVX2 structural indexer. Unlike traditional state machines, Tachyon uses a 2-pass SIMD approach:

1. Classification Pass:

   • Loads 32 bytes into YMM registers.
   • Uses a highly optimized PSHUFB (Packed Shuffle Bytes) lookup table to classify characters into Quote, Backslash, and Structural (Commas, Colons, Brackets) in a single step.

2. Bitmask Generation:

   • _mm256_movemask_epi8 extracts significant bits to form a 32-bit structural mask.
   • A Branchless State Machine manages string parsing (handling escaped quotes) using PCLMUL-style XOR prefix sums.

3. Lazy Indexing (Zero-Copy):

   • The parser produces a "Bitmask Index" rather than a full DOM tree.
   • Data is accessed via a Cursor that navigates this bitmask.
   • No memory is allocated for JSON nodes until you ask for them (Lazy Materialization).

2. Dual-Engine API

• Lazy View: json::parse_view(ptr, len) creates a read-only view over the buffer. Accessing j["key"] scans the bitmask at lightning speed (Zero-Copy).
• Materialized DOM: Accessing j["key"] on a mutable object automatically promotes the view to a standard std::map/std::vector representation (Owned), ensuring ease of use for modification.

🛠 Features

Supported Formats

• JSON (RFC 8259)
• JSONC (Comments // and /* */)
• Binary Placeholders: API stubs for CBOR, MessagePack (throws runtime_error currently).

Modern C++20 API

• Zero-Boilerplate Reflection:

  struct User { std::string name; int age; };
  TACHYON_DEFINE_TYPE_NON_INTRUSIVE(User, name, age)
  
  // Auto-Serialization
  User u = j.get<User>();
  json j2 = u;

• Type Safety: strict checks for types, std::variant backend.
• STL Compatibility: Works with std::string, std::vector, std::map.

📜 License

TACHYON PROPRIETARY SOURCE LICENSE v1.0

Copyright (c) 2026 (BosyjJakub). All Rights Reserved.

1. No Redistribution: This source code may not be distributed, sub-licensed, or shared without explicit permission.
2. No Modification: Modification of the core ASM/SIMD algorithms is strictly prohibited to maintain performance integrity.
3. No Theft: Extraction of ASM kernels or "Look-Up Table" logic for use in other libraries is forbidden.

THIS SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND.