llvmcall errors for some Vec sizes

[perrutquist]

Example:

julia> v = Vec(ntuple(i->1.0, 7))
<7 x Float64>[1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]

julia> 1.0*v
ERROR: error compiling *: Failed to parse LLVM Assembly: 
julia: llvmcall:3:26: error: '%0' defined with type '[7 x double]'
%res = fmul <7 x double> %0, %1
                         ^

(For other lengths, such as 6 or 8, there's no error. )

I suspect this has to to with NTuple{N, VecElement{T}} not mapping to a LLVM vector for all values of N.

[vchuravy]

Right I think it turns out that LLVM creates invalid code for vector lengths that are not multiplies of 2 and Julia chooses a different representation for them. Couldn't quickly find where that happens, but I remember investigating this years ago.

[KristofferC]

See discussion in #1.

[eschnett]

So this is Julia applying a work-around for a long-standing bug in LLVM?

Which vector lengths are working in Julia? SIMD should emit an error message for these.

[perrutquist]

From: https://docs.julialang.org/en/v1/base/simd-types/
"A homogeneous tuple of VecElement{T} maps to an LLVM vector type when T is a primitive bits type and the tuple length is in the set {2-6,8-10,16}"

Although the vector length 12 also seems to work, even though it's not in the above-mentioned set.

[perrutquist]

Rather than an error message, how about letting Vec(::NTuple{7,T)) work with an NTuple{8,VecElement{T}} internally?

[fredrikekre]

Which vector lengths are working in Julia?

I investigated that some time ago: https://github.com/KristofferC/Tensors.jl/blob/218188cbfb3d755d5fbdd5c33efc115942b9781a/src/simd.jl#L33-L34. I don't remember which Juila/LLVM version that was for though.

[eschnett]

It seems the respective bug in LLVM was corrected in 2017 (see https://bugs.llvm.org/show_bug.cgi?id=27708). We might just be able to remove that special rule again (see jl_special_vector_alignment in src/datatype.c in Julia). I'm investigating.

[eschnett]

Yes, removing the special case that only allows certain vector sizes seems to work fine now when this special casing is removed. (It was necessary for LLVM 3.8 and earlier.)

[perrutquist]

Wow! After merging JuliaLang/julia#34490 and JuliaLang/julia#34473 SIMD.jl works beautifully! (I had almost given up on using this package due to segfaults and other problems.) Maybe mention these PRs in the readme until they are in a Julia release?

[KristofferC]

This is now fixed upstream. Will be in Julia 1.4.