Aggressive constprop in matvecmul and matmatmul

[jishnub]

This usually eliminates branches, with occasional improvement in type-stability. I've replaced @inline by aggressive constprop in generic_matmatmul, as I felt that this was the objective.

[dkarrasch]

as I felt that this was the objective.

Indeed, it was. Is there any indication that this helps with compilation times or something?

[jishnub]

The impact seems fairly minimal in the simple tests that I've carried out, although this might make JET happier.

[dkarrasch]

Unfortunately, this seems to not help elimination of branches. At least for me, this makes things as in #51812 (comment) worse regarding compile time. If anything, then I would have hoped for a decrease in compile time and elimination of branches, since the runtime of the character checks is negligible.

[jishnub]

Perhaps we may apply it selectively, as it seems to improve type inference in certain cases (e.g. complex-real matvecmul)?

[gdalle]

See also:

• Regression for mul! from 1.9 to 1.10 JuliaSparse/SparseArrays.jl#469

[dkarrasch]

See also:

* [Regression for `mul!` from 1.9 to 1.10 JuliaSparse/SparseArrays.jl#469](https://github.com/JuliaSparse/SparseArrays.jl/issues/469)

This is unrelated to constant propagation.

I'm no longer sure we need these annotations, actually. It seems to suffice to inline some intermediate functions. For instance

julia> using LinearAlgebra

julia> A = zeros(10, 10);

julia> @code_typed mul!(copy(A), A', A)
CodeInfo(
1 ─ %1 = Base.getfield(A, :parent)::Matrix{Float64}
│   %2 = (%1 === B)::Bool
└──      goto JuliaLang/julia#3 if not %2
2 ─ %4 = invoke LinearAlgebra.syrk_wrapper!(C::Matrix{Float64}, 'T'::Char, %1::Matrix{Float64}, $(QuoteNode(LinearAlgebra.MulAddMul{true, true, Bool, Bool}(true, false)))::LinearAlgebra.MulAddMul{true, true, Bool, Bool})::Matrix{Float64}
└──      goto JuliaLang/julia#4
3 ─ %6 = invoke LinearAlgebra.gemm_wrapper!(C::Matrix{Float64}, 'T'::Char, 'N'::Char, %1::Matrix{Float64}, B::Matrix{Float64}, $(QuoteNode(LinearAlgebra.MulAddMul{true, true, Bool, Bool}(true, false)))::LinearAlgebra.MulAddMul{true, true, Bool, Bool})::Matrix{Float64}
└──      goto JuliaLang/julia#4
4 ┄ %8 = φ (#2 => %4, JuliaLang/julia#3 => %6)::Matrix{Float64}
└──      goto JuliaLang/julia#5
5 ─      return %8
) => Matrix{Float64}

That means, the compiler understands that only two branches might be used, but e.g. herk_wrapper! is not. This is already achieved with the @inline annotation of the BLAS-dispatch method of generic_matmatmul!, and it does not improve with aggressive constant propagation! So, before we put more load on the compiler, we should make sure that each and every new annotation has a tangible benefit.

[gdalle]

This is unrelated to constant propagation.

I thought the matmul regression had two sides, a performance side and a type-instability side which may be resolved by constprop? At least that's what @jishnub suggested on Discourse

[dkarrasch]

The regression is JuliaLang/LinearAlgebra.jl#1037. There may be cases where we need to rewrap an unwrapped matrix which unfortunately allocates. But for these cases aggressive constant propagation doesn't seem to help.

[dkarrasch]

Sorry, I clicked the wrong button in my VSCode instance.