Preserve scalar arithmetic for small GenericLU systems

Project.toml

@@ -1,6 +1,6 @@
 name = "LinearSolve"
 uuid = "7ed4a6bd-45f5-4d41-b270-4a48e9bafcae"
-version = "5.10.0"
+version = "5.10.1"
 authors = ["SciML"]
 
 [deps]

src/blocked_lufact.jl

@@ -53,9 +53,7 @@ end
     return kp, amax
 end
 
-# Unblocked factorization for the whole matrix at small sizes. Identical
-# control flow to the scalar `generic_lufact!` RowMaximum path, with hoisted
-# multipliers, `muladd`, and `@simd ivdep` on the column updates.
+# Keep small factorizations identical to the scalar `generic_lufact!` RowMaximum path.
 function _blocked_lu_unblocked!(A::AbstractMatrix{T}, ipiv, m::Int, n::Int) where {T}
     minmn = min(m, n)
     info = 0
@@ -79,8 +77,8 @@ function _blocked_lu_unblocked!(A::AbstractMatrix{T}, ipiv, m::Int, n::Int) wher
         end
         for j in (k + 1):n
             Akj = A[k, j]
-            @simd ivdep for i in (k + 1):m
-                A[i, j] = muladd(-A[i, k], Akj, A[i, j])
+            for i in (k + 1):m
+                A[i, j] -= A[i, k] * Akj
             end
         end
     end

test/Core/blocked_lufact.jl

@@ -41,6 +41,17 @@ end
         @test !endswith(String(m_generic.file), "blocked_lufact.jl")
     end
 
+    @testset "small path matches scalar arithmetic" begin
+        A = randn(MersenneTwister(1234), 4, 4)
+        F = LinearSolve.generic_lufact!(copy(A), RowMaximum(), _ipiv(4); check = false)
+        A_scalar_dispatch = @view copy(A)[:, [1, 2, 3, 4]]
+        F_scalar = LinearSolve.generic_lufact!(
+            A_scalar_dispatch, RowMaximum(), _ipiv(4); check = false
+        )
+        @test F.factors == F_scalar.factors
+        @test F.ipiv == F_scalar.ipiv
+    end
+
     @testset "residual, square, default params ($T)" for T in (Float64, Float32)
         for n in (
                 1, 2, 3, 5, 7, 8, 9, 13, 16, 17, 31, 32, 33, 40, 41, 63, 64, 65,