Icedland Iced Versions Save

• AVX-VNNI-INT16
• SHA512
• SM3
• SM4
• TSE

• AMX-COMPLEX

• All enums are now saved as integers, which results in faster json (de)serialization and less strings in the final binary when using a human readable serialization format such as json. This also means we don't need a hashmap to quickly look up values so the optional hashbrown dependency that was used if no_std + serde has now been removed.

• RMPQUERY

• AMX-FP16
• AVX-IFMA
• AVX-NE-CONVERT
• AVX-VNNI-INT8
• CMPCCXADD
• MSRLIST
• PREFETCHITI
• RAO-INT
• WRMSRNS

• The assembler also uses it so its encoding speed should be improved as well

• For reference, Rust: 260MB/s, Python module: 50MB/s, nodejs+wasm module: 46MB/s

• Made CodeAssembler traits public #236 (Credit @Kixiron)
• Make some Instruction/Register fns const #240 (Credit @i509VCB)
• Removed HashMap (hashbrown crate if no_std) dependency from block_encoder feature #248
  • Fixes #247 (can't build with no_std on Windows)
  • This also resulted in better block encoder perf (30s -> 19s when encoding a ton of instructions)
• Removed deprecated Instruction::with_* methods (use with{1,2,3,4,5}() instead)
• Bumped MSRV to 1.54.0 #252

• PyPy wheels uploaded to PyPI (Linux x64, Windows x64) #250
• Minimum supported Python version is now 3.7 #251

• BigInt is supported by default now #258

• Enable decoding of KNC instructions by passing in DecoderOptions::KNC to the Decoder constructor
• Rust: mvex feature must also be enabled to decode KNC instructions

• Enum type annotations can be used without mypy reporting errors, see here
• Added musllinux-x86_64 (Alpine) wheel

• (serde feature): json deserialization (Instruction) is faster (json is still inefficient and slow, though)
• Minimum supported Rust version (MSRV) updated from 1.41.0 to 1.48.0

let mut a = CodeAssembler::new(64)?;
a.xor(ecx, ecx)?;
a.add(byte_ptr(rdx + r15 * 8 + 7), 0x10)?;
assert_eq!(a.instructions().len(), 2);
let bytes = a.assemble(0x1234_5678)?;
assert_eq!(bytes, b"\x31\xC9\x42\x80\x44\xFA\x07\x10");

// old: 
let _ = Instruction::try_with_reg_reg_u32(Code::Imul_r16_rm16_imm16, Register::CX, Register::DX, 0x5AA5)?;
// new: (the '3' suffix means '3 operands')
let _ = Instruction::with3(Code::Imul_r16_rm16_imm16, Register::CX, Register::DX, 0x5AA5)?;
// A number suffix (`u32`, `u64`, `i64`) is sometimes needed to help the compiler:
let _ = Instruction::with2(Code::Mov_r64_imm64, Register::RAX, 0x1234_5678_9ABC_DEF0u64)?;

• Decoding throughput up 25% from ~205 MB/s to ~255 MB/s (Skylake 3.5GHz CPU)
• It's now 3x-9x faster than x86 decoders written in C (and ~2.5x faster than iced-C#)
• Benchmark: see here