Impact of https://github.com/ArchRobison/julia/tree/adr/simdtuple changes

benchmark.jl

T=NTuple{4,Float32}

zero(::Type{T}) = (0.0f0,0.0f0,0.0f0,0.f0)

function add( a::T, b::T )
    (a[1]+b[1], a[2]+b[2], a[3]+b[3], a[4]+b[4])
end

function mul( a::T, b::T )
    (a[1]*b[1], a[2]*b[2], a[3]*b[3], a[4]*b[4])
end

function muladd( a::T, b::T, c::T )
    @fastmath add(a,mul(b,c))
end

function vec( a::T, x::Vector{T}, y::Vector{T} )
    @inbounds for i=1:length(x)
        y[i] = muladd(y[i],a,x[i])
    end
    y
end

function f()
    u = fill( (0.0f0,0.0f0,0.0f0,0.f0), 1000)
    v = fill( (0.0f0,0.0f0,0.0f0,0.f0), 1000)
    @time for i=1:100000
        vec(zero(T), v, u)
    end
end

code_native(vec,(T,Vector{T},Vector{T}))

f()
f()
f()

output_with_simdtuple.txt

L64:L176:L196:  .text
        pushq   %rax
        movq    %fs:40, %rax
        movq    %rax, (%rsp)
        movq    8(%rdi), %rax
        cmpq    $1, %rax
        jl      L176
        movq    (%rsi), %rcx
        movq    (%rdi), %rdx
        vmovshdup       %xmm0, %xmm8    # xmm8 = xmm0[1,1,3,3]
        vpermilpd       $1, %xmm0, %xmm9 # xmm9 = xmm0[1,0]
        vpermilps       $231, %xmm0, %xmm3 # xmm3 = xmm0[3,1,2,3]
        nopw    %cs:(%rax,%rax)
        vmovups (%rcx), %xmm4
        vmovups (%rdx), %xmm5
        vmulss  %xmm5, %xmm0, %xmm6
        vmovshdup       %xmm5, %xmm7    # xmm7 = xmm5[1,1,3,3]
        vmulss  %xmm7, %xmm8, %xmm7
        vpermilpd       $1, %xmm5, %xmm1 # xmm1 = xmm5[1,0]
        vmulss  %xmm1, %xmm9, %xmm1
        vpermilps       $231, %xmm5, %xmm5 # xmm5 = xmm5[3,1,2,3]
        vmulss  %xmm5, %xmm3, %xmm5
        vaddss  %xmm6, %xmm4, %xmm6
        vmovshdup       %xmm4, %xmm2    # xmm2 = xmm4[1,1,3,3]
        vaddss  %xmm7, %xmm2, %xmm2
        vpermilpd       $1, %xmm4, %xmm7 # xmm7 = xmm4[1,0]
        vaddss  %xmm1, %xmm7, %xmm1
        vpermilps       $231, %xmm4, %xmm4 # xmm4 = xmm4[3,1,2,3]
        vaddss  %xmm5, %xmm4, %xmm4
        vinsertps       $16, %xmm2, %xmm6, %xmm2 # xmm2 = xmm6[0],xmm2[0],xmm6[2,3]
        vinsertps       $32, %xmm1, %xmm2, %xmm1 # xmm1 = xmm2[0,1],xmm1[0],xmm2[3]
        vinsertps       $48, %xmm4, %xmm1, %xmm1 # xmm1 = xmm1[0,1,2],xmm4[0]
        vmovups %xmm1, (%rcx)
        addq    $-1, %rax
        addq    $16, %rdx
        addq    $16, %rcx
        cmpq    $0, %rax
        jne     L64
        movq    %fs:40, %rax
        cmpq    (%rsp), %rax
        jne     L196
        movq    %rsi, %rax
        popq    %rdx
        retq
        movabsq $__stack_chk_fail, %rax
        callq   *%rax
        pushq   %rax
        movq    (%rsi), %rax
        vmovaps (%rax), %xmm0
        movq    8(%rsi), %rdi
        movq    16(%rsi), %rsi
        movabsq $vec, %rax
        callq   *%rax
        popq    %rdx
        retq
  0.317651 seconds (100.03 k allocations: 3.054 MB)
  0.297012 seconds (100.00 k allocations: 3.052 MB)
  0.297075 seconds (100.00 k allocations: 3.052 MB)

output_with_simdtuple_and-O.txt

L32:L62:L82:    .text
        pushq   %rax
        movq    %fs:40, %rax
        movq    %rax, (%rsp)
        movq    8(%rdi), %rax
        cmpq    $1, %rax
        jl      L62
        movq    (%rsi), %rcx
        movq    (%rdi), %rdx
        nop
        vmulps  (%rdx), %xmm0, %xmm1
        vaddps  (%rcx), %xmm1, %xmm1
        vmovups %xmm1, (%rcx)
        addq    $-1, %rax
        addq    $16, %rdx
        addq    $16, %rcx
        cmpq    $0, %rax
        jne     L32
        movq    %fs:40, %rax
        cmpq    (%rsp), %rax
        jne     L82
        movq    %rsi, %rax
        popq    %rdx
        retq
        movabsq $__stack_chk_fail, %rax
        callq   *%rax
        nop
        pushq   %rax
        movq    (%rsi), %rax
        vmovaps (%rax), %xmm0
        movq    8(%rsi), %rdi
        movq    16(%rsi), %rsi
        movabsq $vec, %rax
        callq   *%rax
        popq    %rdx
        retq
  0.111175 seconds (100.03 k allocations: 3.054 MB)
  0.084908 seconds (100.00 k allocations: 3.052 MB)
  0.087529 seconds (100.00 k allocations: 3.052 MB)

output_without_changes.txt

L48:L123:       .text
        movq    8(%rsi), %rax
        cmpq    $1, %rax
        jl      L123
        movq    (%rdx), %rcx
        movq    (%rsi), %rsi
        vmovss  (%rdi), %xmm0           # xmm0 = mem[0],zero,zero,zero
        vmovss  4(%rdi), %xmm1          # xmm1 = mem[0],zero,zero,zero
        vmovss  8(%rdi), %xmm2          # xmm2 = mem[0],zero,zero,zero
        vmovss  12(%rdi), %xmm3         # xmm3 = mem[0],zero,zero,zero
        nopw    %cs:(%rax,%rax)
        vmulss  (%rsi), %xmm0, %xmm4
        vmulss  4(%rsi), %xmm1, %xmm5
        vmulss  8(%rsi), %xmm2, %xmm6
        vmulss  12(%rsi), %xmm3, %xmm7
        vaddss  (%rcx), %xmm4, %xmm4
        vaddss  4(%rcx), %xmm5, %xmm5
        vaddss  8(%rcx), %xmm6, %xmm6
        vaddss  12(%rcx), %xmm7, %xmm7
        vmovss  %xmm7, 12(%rcx)
        vmovss  %xmm6, 8(%rcx)
        vmovss  %xmm5, 4(%rcx)
        vmovss  %xmm4, (%rcx)
        addq    $-1, %rax
        addq    $16, %rsi
        addq    $16, %rcx
        cmpq    $0, %rax
        jne     L48
        movq    %rdx, %rax
        retq
        nop
        pushq   %rax
        movq    (%rsi), %rdi
        movq    8(%rsi), %rax
        movq    16(%rsi), %rdx
        movabsq $vec, %rcx
        movq    %rax, %rsi
        callq   *%rcx
        popq    %rdx
        retq
  0.210977 seconds (100.03 k allocations: 3.054 MB)
  0.189340 seconds (100.00 k allocations: 3.052 MB)
  0.188498 seconds (100.00 k allocations: 3.052 MB)