@avx doesn't handle @inbounds

[YingboMa]

MWE:

julia> using LoopVectorization

julia> a = rand(5); foo(a) = @avx for i in 1:10
           @inbounds a[i] = 1
       end
ERROR: LoadError: "Don't know how to handle expression:\n\$(Expr(:inbounds, true))"
Stacktrace:
 [1] push!(::LoopVectorization.LoopSet, ::Expr, ::Int64, ::Int64) at /home/scheme/.julia/packages/LoopVectorization/IESri/src/graphs.jl:550
 [2] add_block!(::LoopVectorization.LoopSet, ::Expr, ::Int64, ::Int64) at /home/scheme/.julia/packages/LoopVectorization/IESri/src/graphs.jl:335
 [3] push!(::LoopVectorization.LoopSet, ::Expr, ::Int64, ::Int64) at /home/scheme/.julia/packages/LoopVectorization/IESri/src/graphs.jl:542
 [4] add_block!(::LoopVectorization.LoopSet, ::Expr, ::Int64, ::Int64) at /home/scheme/.julia/packages/LoopVectorization/IESri/src/graphs.jl:335
 [5] add_loop!(::LoopVectorization.LoopSet, ::Expr, ::Int64) at /home/scheme/.julia/packages/LoopVectorization/IESri/src/graphs.jl:438
 [6] copyto! at /home/scheme/.julia/packages/LoopVectorization/IESri/src/constructors.jl:6 [inlined]
 [7] LoopVectorization.LoopSet(::Expr, ::Symbol) at /home/scheme/.julia/packages/LoopVectorization/IESri/src/constructors.jl:48
 [8] LoopVectorization.LoopSet(::Expr, ::Module) at /home/scheme/.julia/packages/LoopVectorization/IESri/src/constructors.jl:52
 [9] @avx(::LineNumberNode, ::Module, ::Any) at /home/scheme/.julia/packages/LoopVectorization/IESri/src/constructors.jl:93
 [10] eval(::Module, ::Any) at ./boot.jl:331
 [11] eval_user_input(::Any, ::REPL.REPLBackend) at /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.4/REPL/src/REPL.jl:86
 [12] run_backend(::REPL.REPLBackend) at /home/scheme/.julia/packages/Revise/Pcs5V/src/Revise.jl:1073
 [13] top-level scope at none:0
in expression starting at REPL[20]:1

[chriselrod]

Note that @avx does not do any bounds checking, so you should always ensure that a is at least of length 10 beforehand.

julia> a = rand(15);

julia> using LoopVectorization, BenchmarkTools

julia> foo(a) = @avx for i in 1:10
                  a[i] = 1
              end
foo (generic function with 1 method)

julia> @benchmark foo($a)
BenchmarkTools.Trial:
  memory estimate:  0 bytes
  allocs estimate:  0
  --------------
  minimum time:     1.878 ns (0.00% GC)
  median time:      1.896 ns (0.00% GC)
  mean time:        1.899 ns (0.00% GC)
  maximum time:     4.488 ns (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     1000

julia> a'
1×15 LinearAlgebra.Adjoint{Float64,Array{Float64,1}}:
 1.0  1.0  1.0  1.0  1.0  1.0  1.0  1.0  1.0  1.0  0.481493  0.660123  0.615034  0.245557  0.0906999

# julia> @code_llvm debuginfo=:none foo(a)

define nonnull %jl_value_t addrspace(10)* @japi1_foo_2557(%jl_value_t addrspace(10)*, %jl_value_t addrspace(10)**, i32) #0 {
top:
  %3 = alloca %jl_value_t addrspace(10)**, align 8
  store volatile %jl_value_t addrspace(10)** %1, %jl_value_t addrspace(10)*** %3, align 8
  %4 = load %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)** %1, align 8
  %5 = addrspacecast %jl_value_t addrspace(10)* %4 to %jl_value_t addrspace(11)*
  %6 = addrspacecast %jl_value_t addrspace(11)* %5 to %jl_value_t*
  %7 = bitcast %jl_value_t* %6 to i64*
  %8 = load i64, i64* %7, align 8
  %ptr.i2 = inttoptr i64 %8 to <8 x double>*
  store <8 x double> <double 1.000000e+00, double 1.000000e+00, double 1.000000e+00, double 1.000000e+00, double 1.000000e+00, double 1.000000e+00, double 1.000000e+00, double 1.000000e+00>, <8 x double>* %ptr.i2, align 8
  %typptr.i = inttoptr i64 %8 to double*
  %offsetptr.i = getelementptr inbounds double, double* %typptr.i, i64 8
  %ptr.i = bitcast double* %offsetptr.i to <8 x double>*
  call void @llvm.masked.store.v8f64.p0v8f64(<8 x double> <double 1.000000e+00, double 1.000000e+00, double undef, double undef, double undef, double undef, double undef, double undef>, <8 x double>* %ptr.i, i32 8, <8 x i1> <i1 true, i1 true, i1 false, i1 false, i1 false, i1 false, i1 false, i1 false>)
  ret %jl_value_t addrspace(10)* addrspacecast (%jl_value_t* inttoptr (i64 140026789409680 to %jl_value_t*) to %jl_value_t addrspace(10)*)
}

; julia> @code_native debuginfo=:none foo(a)
        .text
        movq    %rsi, -8(%rsp)
        movq    (%rsi), %rax
        movq    (%rax), %rax
        movabsq $.rodata.cst8, %rcx
        vbroadcastsd    (%rcx), %zmm0
        vmovupd %zmm0, (%rax)
        movb    $3, %cl
        kmovd   %ecx, %k1
        vmovupd %zmm0, 64(%rax) {%k1}
        movabsq $jl_system_image_data, %rax
        vzeroupper
        retq

So the fix here was to ignore the @inbounds.

julia> fooinbounds(a) = @avx for i in 1:10
                  @inbounds a[i] = 1
              end
fooinbounds (generic function with 1 method)

julia> fooinbounds(a); a'
1×15 LinearAlgebra.Adjoint{Float64,Array{Float64,1}}:
 1.0  1.0  1.0  1.0  1.0  1.0  1.0  1.0  1.0  1.0  0.481493  0.660123  0.615034  0.245557  0.0906999

# julia> @code_llvm debuginfo=:none fooinbounds(a)

define nonnull %jl_value_t addrspace(10)* @japi1_fooinbounds_2676(%jl_value_t addrspace(10)*, %jl_value_t addrspace(10)**, i32) #0 {
top:
  %3 = alloca %jl_value_t addrspace(10)**, align 8
  store volatile %jl_value_t addrspace(10)** %1, %jl_value_t addrspace(10)*** %3, align 8
  %4 = load %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)** %1, align 8
  %5 = addrspacecast %jl_value_t addrspace(10)* %4 to %jl_value_t addrspace(11)*
  %6 = addrspacecast %jl_value_t addrspace(11)* %5 to %jl_value_t*
  %7 = bitcast %jl_value_t* %6 to i64*
  %8 = load i64, i64* %7, align 8
  %ptr.i2 = inttoptr i64 %8 to <8 x double>*
  store <8 x double> <double 1.000000e+00, double 1.000000e+00, double 1.000000e+00, double 1.000000e+00, double 1.000000e+00, double 1.000000e+00, double 1.000000e+00, double 1.000000e+00>, <8 x double>* %ptr.i2, align 8
  %typptr.i = inttoptr i64 %8 to double*
  %offsetptr.i = getelementptr inbounds double, double* %typptr.i, i64 8
  %ptr.i = bitcast double* %offsetptr.i to <8 x double>*
  call void @llvm.masked.store.v8f64.p0v8f64(<8 x double> <double 1.000000e+00, double 1.000000e+00, double undef, double undef, double undef, double undef, double undef, double undef>, <8 x double>* %ptr.i, i32 8, <8 x i1> <i1 true, i1 true, i1 false, i1 false, i1 false, i1 false, i1 false, i1 false>)
  ret %jl_value_t addrspace(10)* addrspacecast (%jl_value_t* inttoptr (i64 140026789409680 to %jl_value_t*) to %jl_value_t addrspace(10)*)
}

; julia> @code_native debuginfo=:none fooinbounds(a)
        .text
        movq    %rsi, -8(%rsp)
        movq    (%rsi), %rax
        movq    (%rax), %rax
        movabsq $.rodata.cst8, %rcx
        vbroadcastsd    (%rcx), %zmm0
        vmovupd %zmm0, (%rax)
        movb    $3, %cl
        kmovd   %ecx, %k1
        vmovupd %zmm0, 64(%rax) {%k1}
        movabsq $jl_system_image_data, %rax
        vzeroupper
        retq

I wonder if it'd be better if the second store were a xmm store instead of a masked zmm (or ymm) store.