`add generic syrk/herk` worsened effects of `*(::Matrix, ::Matrix)`

[adienes]

on cab0dc6 :

julia> Base.infer_effects(*, Tuple{Matrix{Int}, Matrix{Int}})
(!c,?e,!n,!t,+s,?m,!u,+o,+r)

but on a32a281 :

julia> Base.infer_effects(*, Tuple{Matrix{Int}, Matrix{Int}})
(!c,!e,!n,!t,!s,!m,!u,+o,!r)

this might be known / intentional, in which case please close this issue. but I am just assuming from the title of #1249 that it did not intend to have an impact on * methods of scalar matrices

[araujoms]

That was definitely not intentional. I don't know anything about these effects, so debugging will be challenging. Note that it wasn't for all scalars, though, I don't get any difference for BlasFloat.

[araujoms]

I've been debugging this, and the culprit seems to be the functions uppercase and wrap, that are called for example in the new syrk_wrapper!:

LinearAlgebra.jl/src/matmul.jl

Lines 817 to 827 in 27b0f13

	Base.@constprop :aggressive function syrk_wrapper!(C::StridedMatrix{T}, tA::AbstractChar, A::StridedVecOrMat{T},
	α::Number, β::Number) where {T<:Number}
	tA_uc = uppercase(tA) # potentially strip a WrapperChar
	aat = (tA_uc == 'N')
	if T <: Union{Real,Complex} && (iszero(β) || issymmetric(C))
	return copytri!(generic_syrk!(C, A, false, aat, α, β), 'U')
	end
	tAt = aat ? 'T' : 'N'
	return _generic_matmatmul!(C, wrap(A, tA), wrap(A, tAt), α, β)
	end

If I remove them from all possible paths the compiler effects get nice again. Is it worth it, though? How bad are these effects?

[adienes]

honestly the effects probably don't matter much here, but they probably will be fighting @constprop :aggressive / make that annotation less useful.

however, I do wonder why uppercase has such bad effects, maybe that can be improved in isolation on Julia' side?

[araujoms]

That's above my skill level. But I don't even know why this uppercase stuff is necessary, looks like a horrible mess. Perhaps @dkarrasch or @stevengj would know?

[dkarrasch]

@jishnub introduced it at some point. Probably because of some Val stuff, I don't remember.

[jishnub]

It was to improve type stability in Symmetric/Hermitian matrix multiplication by propagating constants. While we may explore ditching the constant-propagation in favor of types, this specific case is clearly something that Base may improve upon.

[adienes]

Char effects fixed on Base. wrap is owned by LA.jl so I didn't touch it

[araujoms]

Nice! I'll rewrite the function to not need wrap then, as it is inherently type-unstable.

[dkarrasch]

Nice! I'll rewrite the function to not need wrap then, as it is inherently type-unstable.

I'm very curious how you will do that. Typically, it is used when the most generic path is taken, when you need to combine the wrapper and the storage array back again and when performance is probably not very good anyway.