The way to use AES.

[sunoru]

AES encryption is an important part of the RNGs for secure use. There are some methods to use AES:

1. Implement the encryption algorithm in pure Julia. There are also existing packages such as AES.jl (but this is not registered), and Nettle.jl (though this also provides many other functions).
2. Use the mature OpenSSL. This may also have two different ways: to wrap it by ourselves or use the existing ones (for example, Crypto.jl, but it is not completed so doesn't guarantee the security)...

It would be better if we can directly use AES-NI on the most kinds of CPU, so I'm also wondering how will llvm do with this? Anyway, OpenSSL should have been already optimized about such things.

[sunoru]

When trying to use the AES intrinsics (as https://github.com/frasercrmck/llvm-leg/blob/master/test/CodeGen/X86/aes_intrinsics.ll), I found for now the llvmcall only accepts array ABI such as [2 x i64] (JuliaLang/julia@a93e9a1), but to use the intrinsics it's necessary to pass <2 x i64>

julia> function aesenc(x::Tuple{UInt64, UInt64}, y::Tuple{UInt64, UInt64})
           Base.llvmcall(("""declare <2 x i64> @llvm.x86.aesni.aesenc(<2 x i64>, <2 x i64>)""","""
               %3 = call <2 x i64> @llvm.x86.aesni.aesenc(<2 x i64> %0, <2 x i64> %1)
           ret <2 x i64> %3"""), Tuple{UInt64, UInt64}, Tuple{Tuple{UInt64, UInt64}, Tuple{UInt64, UInt64}}, x, y)
       end
aesenc (generic function with 1 method)

julia> aesenc((1 % UInt64, 1 % UInt64), (1 % UInt64, 1 % UInt64))
ERROR: error compiling aesenc: Failed to parse LLVM Assembly: 
julia: llvmcall:5:58: error: '%0' defined with type '[2 x i64]'
    %3 = call <2 x i64> @llvm.x86.aesni.aesenc(<2 x i64> %0, <2 x i64> %1)
                                                         ^

 in eval(::Module, ::Any) at ./boot.jl:231
 in macro expansion at ./REPL.jl:92 [inlined]
 in (::Base.REPL.##1#2{Base.REPL.REPLBackend})() at ./event.jl:46

[simonbyrne]

Sorry we haven't had a chance to chat recently, I've been travelling a lot. The distinction here is between what LLVM terms vectors and arrays. From what I understand (which may well be wrong), vectors are what can be used with SIMD instructions, whereas arrays are more general structs.

In 0.5 you can use VecElement introduced by JuliaLang/julia#15244 inside Tuples:

julia> function aesenc(x, y)
                  Base.llvmcall(("""declare <2 x i64> @llvm.x86.aesni.aesenc(<2 x i64>, <2 x i64>)""","""
                      %3 = call <2 x i64> @llvm.x86.aesni.aesenc(<2 x i64> %0, <2 x i64> %1)
                  ret <2 x i64> %3"""), NTuple{2,VecElement{UInt64}}, Tuple{NTuple{2,VecElement{UInt64}},NTuple{2,VecElement{UInt64}}}, x, y)
              end
aesenc (generic function with 1 method)

julia> aesenc((VecElement(1 % UInt64), VecElement(1 % UInt64)), (VecElement(1 % UInt64), VecElement(1 % UInt64)))
(VecElement{UInt64}(0x63636363427c7c5c),VecElement{UInt64}(0x63636363427c7c5c))

julia> @code_native aesenc((VecElement(1 % UInt64), VecElement(1 % UInt64)), (VecElement(1 % UInt64), VecElement(1 % UInt64)))
    .section    __TEXT,__text,regular,pure_instructions
Filename: REPL[1]
    pushq   %rbp
    movq    %rsp, %rbp
Source line: 2
    vaesenc %xmm1, %xmm0, %xmm0
    popq    %rbp
    retq
    nopl    (%rax,%rax)

[sunoru]

Thanks, that's what I need.