Performance regression for (*)(A, D::Diagonal) when A is the adjoint of a sparse matrix

[cschwarzbach]

Matrix-matrix multiplication with types A <: Adjoint{T, SparseMatrixCSC{T, N}} and B <: Diagonal is significantly slower in Julia 1.11 than in Julia 1.10. The MWE below executes in O(n) time with Julia 1.10, and in O(n²) time with Julia 1.11.

The performance regression appears to have been introduced in PR JuliaLang/julia#52389 that removed specialized methods.

MWE

using LinearAlgebra
using SparseArrays
using BenchmarkTools

n = 100;
A = adjoint(sparse(Float64, I, n, n));
B = Diagonal(ones(n));

print("Julia ", VERSION, ", n = ", n, ":")
@btime *($A, $B)

Output:

Julia 1.10.8, n = 100:  531.160 ns (5 allocations: 2.59 KiB)
Julia 1.11.3, n = 100:  65.927 μs (10 allocations: 8.53 KiB)

Julia 1.10.8, n = 1000:  4.626 μs (5 allocations: 23.88 KiB)
Julia 1.11.3, n = 1000:  6.240 ms (13 allocations: 57.38 KiB)

Julia 1.10.8, n = 10000:  43.303 μs (6 allocations: 234.73 KiB)
Julia 1.11.3, n = 10000:  623.509 ms (13 allocations: 666.56 KiB)

Julia 1.10.8, n = 100000:  450.953 μs (6 allocations: 2.29 MiB)
Julia 1.11.3, n = 100000:  62.992 s (13 allocations: 5.01 MiB)

These results are on a Linux x86-64 platform; similar results were observed on MacOS x86-64.