Specialize A * Eye and Eye * A

[cossio]

d = Diagonal(Ones(10))

The following ops fallback to generic (diagonal) matrix multiplies.

d * randn(10, 10)
randn(10, 10) * d

A faster fallback is possible since d is just the identity matrix.

[dlfivefifty]

It already does a faster fallback, lowering to broadcasting:

https://github.com/JuliaLang/julia/blob/85a6990a9c1d49dd5aeaffeb4b38f881dc120823/stdlib/LinearAlgebra/src/diagonal.jl#L279

[cossio]

How's that faster than just returning the other term?

d = Diagonal(Ones(10))
A = randn(10,10)
d * A # should just return A

[dlfivefifty]

No it should return a copy of A, which broadcasting with Ones should do.

though I noticed there’s been changes to diagonal multiple so it’s possible that things are no longer as fast. But adding an override for this will need to be a new minor version