Implement alias scopes

[YingboMa]

This PR is based on the original PR #20257 by @Keno.

struct Const{T<:Array}
    a::T
end

@eval Base.getindex(A::Const, i1::Int) = Core.const_arrayref($(Expr(:boundscheck)), A.a, i1)
@eval Base.getindex(A::Const, i1::Int, i2::Int, I::Int...) =  (Base.@_inline_meta; Core.const_arrayref($(Expr(:boundscheck)), A.a, i1, i2, I...))

macro aliasscope(body)
    sym = gensym()
    esc(quote
        $(Expr(:aliasscope))
        $sym = $body
        $(Expr(:popaliasscope))
        $sym
    end)
end

using BenchmarkTools, Test
function micro_ker_const!(AB, Ac::Vector{Float64}, Bc::Vector{Float64}, kc::Int, offSetA::Int, offSetB::Int)
    MR = 8; NR = 6;
    @inbounds @aliasscope for k in 1:kc
        for j in 1:NR, i in 1:MR
            AB[i+(j-1)*MR] = muladd(Const(Ac)[offSetA+i], Const(Bc)[offSetB+j], Const(AB)[i+(j-1)*MR])
        end
        offSetA += MR
        offSetB += NR
    end
    return
end
function micro_ker!(AB, Ac::Vector{Float64}, Bc::Vector{Float64}, kc::Int, offSetA::Int, offSetB::Int)
    MR = 8; NR = 6;
    @inbounds for k in 1:kc
        for j in 1:NR, i in 1:MR
            AB[i+(j-1)*MR] = muladd(Ac[offSetA+i], Bc[offSetB+j], AB[i+(j-1)*MR])
        end
        offSetA += MR
        offSetB += NR
    end
    return
end
AB = zeros(8, 6); Ac = rand(8*32); Bc = rand(32*6);
micro_ker!(AB, Ac, Bc, 32, 0, 0)
@test reshape(Ac, 8, 32) * reshape(Bc, 6, 32)' ≈ reshape(AB, 8, 6)
AB .= 0
micro_ker_const!(AB, Ac, Bc, 32, 0, 0)
@test reshape(Ac, 8, 32) * reshape(Bc, 6, 32)' ≈ reshape(AB, 8, 6)
@btime micro_ker_const!($AB, $Ac, $Bc, 32, 0, 0)
@btime micro_ker!($AB, $Ac, $Bc, 32, 0, 0)

julia> @btime micro_ker_const!($AB, $Ac, $Bc, 32, 0, 0)
  231.211 ns (0 allocations: 0 bytes)

julia> @btime micro_ker!($AB, $Ac, $Bc, 32, 0, 0)
  863.300 ns (0 allocations: 0 bytes)

julia> @code_llvm optimize=false raw=true dump_module=true micro_ker_const!(AB, Ac, Bc, 32, 0, 0)
...
    %34 = load double, double addrspace(13)* %33, align 8, !dbg !40, !tbaa !48, !alias.scope !51
...
    %46 = load double, double addrspace(13)* %45, align 8, !dbg !40, !tbaa !48, !alias.scope !51
...
    %60 = load double, double addrspace(13)* %59, align 8, !dbg !40, !tbaa !48, !alias.scope !51
...
    store double %62, double addrspace(13)* %75, align 8, !dbg !61, !tbaa !48, !noalias !51
...
!51 = !{!52}
!52 = !{!"aliasscope", !53}

CC: @vchuravy

[YingboMa]

Related issues: #19658 #8087 #28126 #29026

[vchuravy]

@maleadt we should probably talk about how we could extend this for CUDAnative. Right now it is based on a compiler intrinsic.

Thanks Yingbo, for hacking on this together!

[vchuravy]

This block and the for-loop the part that Yingbo and I were the most unsure about. We mimicked the + 1 from the linetable construct above.

[vchuravy]

Same with the +1 here.

[vchuravy]

#20257 (comment)

could just call jl_f_arrayref here?

[Keno]

FWIW, I remain unconvinced that this is the correct long-term solution, but I of course recognize the short term utility here, so we should perhaps figure out a way to expose this without boxing ourself in with this as a support API (e.g. perhaps add the intrinsics, but only add the supporting definitions in a package that is explicitly disclaimed as unsupported).

[vchuravy]

I agree with that, this "feels" a bit odd and I wouldn't want us to have to live with it forever. Right now this PR is relatively minimal and doesn't expose anything besides the AST nodes and the intrinsic.
I would love to have this for some stuff Peter and I are working on, but I am fine with declaring the experimental and unsupported.

[StefanKarpinski]

It seems very naively to me like what might be needed is the ability to tell the compiler that this is a good point to check if a set of arrays alias each other and allow the compiler to generate code assuming both positive and negative results. That way you get fast behavior in the good case and correct behavior no matter what.

[vchuravy]

This is what LLVM does under the hood right now. Sometimes it fails because it can't infer the array bounds (so doing it at the Julia level might help) and sometimes you want to avoid the code size increase. We have unalias now so it should be more straightforward to write valid code and not shoot yourself in the foot.

[YingboMa]

The CI error seems unrelated
https://travis-ci.org/JuliaLang/julia/jobs/497227682#L5559

failed to clone from https://github.com/JuliaLang/Example.jl.git, error: GitError(Code:ERROR, Class:Net, curl error: Could not resolve host: github.com

[vchuravy]

@Keno are you okay with merging this as is?

[StefanKarpinski]

Question: is this change purely internal or is it a public API? It doesn't add or change any exports, so I'm assuming it's internal not public. In that case, it seems fine to merge it as an experimental feature to learn how to deal with aliasing better.

[Keno]

I'm ok with this on the understanding that it will definitely go away in the future.

[vchuravy]

Completely internal and all I interested parties are okay with this going away :)