From c8e8b20c7da1da0a51d600a2e16a00f0888be052 Mon Sep 17 00:00:00 2001
From: Hendrik Ranocha <mail@ranocha.de>
Date: Fri, 5 Feb 2021 09:58:33 +0100
Subject: [PATCH] specialize copyto! for Q from qr

This fixes some performance problems reported in #38972 and #37102
(multiplying `Q` by a `Diagonal` or `UniformScaling`).
---
 stdlib/LinearAlgebra/src/qr.jl  | 6 ++++++
 stdlib/LinearAlgebra/test/qr.jl | 9 +++++++++
 2 files changed, 15 insertions(+)

diff --git a/stdlib/LinearAlgebra/src/qr.jl b/stdlib/LinearAlgebra/src/qr.jl
index a76577bb63a0d5..a0de58e4c73ffb 100644
--- a/stdlib/LinearAlgebra/src/qr.jl
+++ b/stdlib/LinearAlgebra/src/qr.jl
@@ -533,6 +533,12 @@ function getindex(Q::AbstractQ, i::Integer, j::Integer)
     return dot(x, lmul!(Q, y))
 end
 
+# specialization avoiding the fallback using slow `getindex`
+function copyto!(dest::AbstractMatrix, src::LinearAlgebra.AbstractQ)
+    copyto!(dest, I)
+    lmul!(src, dest)
+end
+
 ## Multiplication by Q
 ### QB
 lmul!(A::QRCompactWYQ{T,S}, B::StridedVecOrMat{T}) where {T<:BlasFloat, S<:StridedMatrix} =
diff --git a/stdlib/LinearAlgebra/test/qr.jl b/stdlib/LinearAlgebra/test/qr.jl
index b4e6d383bb2622..c61dcaf5de43a9 100644
--- a/stdlib/LinearAlgebra/test/qr.jl
+++ b/stdlib/LinearAlgebra/test/qr.jl
@@ -322,4 +322,13 @@ end
     end
 end
 
+@testset "QR factorization of Q" begin
+    for T in (Float32, Float64, ComplexF32, ComplexF64)
+        Q1, R1 = qr(randn(T,5,5))
+        Q2, R2 = qr(Q1)
+        @test Q1 ≈ Q2
+        @test R2 ≈ I
+    end
+end
+
 end # module TestQR