Specialize LinearAlgebra.BLAS.dot for strided vectors of floats.

Fixes #37767.

Author: dbrueg

stdlib/LinearAlgebra/src/blas.jl

@@ -434,6 +434,53 @@ for (fname, elty) in ((:cblas_zdotu_sub,:ComplexF64),
     end
 end
 
+@inline function _dot_length_check(x,y)
+    n = length(x)
+    if n != length(y)
+        throw(DimensionMismatch("dot product arguments have lengths $(length(x)) and $(length(y))"))
+    end
+    n
+end
+
+for (elty, f) in ((Float32, :dot), (Float64, :dot),
+                  (ComplexF32, :dotc), (ComplexF64, :dotc),
+                  (ComplexF32, :dotu), (ComplexF64, :dotu))
+    @eval begin
+        function $f(x::DenseArray{$elty}, y::DenseArray{$elty})
+            n = _dot_length_check(x,y)
+            $f(n, x, 1, y, 1)
+        end
+
+        function $f(x::StridedVector{$elty}, y::DenseArray{$elty}) 
+            n = _dot_length_check(x,y)
+            xstride = stride(x,1)
+            ystride = stride(y,1)
+            # see subarray.jl / pointer(x)
+            # (for x a SubArray of a Vector, pointer(x) points past the last element
+            # in the case of negative stride)
+            x_delta = xstride < 0 ? n : 1
+            GC.@preserve x $f(n,pointer(x,x_delta),xstride,y,ystride)
+        end
+
+        function $f(x::DenseArray{$elty}, y::StridedVector{$elty}) 
+            n = _dot_length_check(x,y)
+            xstride = stride(x,1)
+            ystride = stride(y,1)
+            y_delta = ystride < 0 ? n : 1
+            GC.@preserve y $f(n,x,xstride,pointer(y,y_delta),ystride)
+        end
+
+        function $f(x::StridedVector{$elty}, y::StridedVector{$elty})
+            n = _dot_length_check(x,y)
+            xstride = stride(x,1)
+            ystride = stride(y,1)
+            x_delta = xstride < 0 ? n : 1
+            y_delta = ystride < 0 ? n : 1
+            GC.@preserve x y $f(n,pointer(x,x_delta),xstride,pointer(y,y_delta),ystride)
+        end
+    end
+end
+
 function dot(DX::Union{DenseArray{T},AbstractVector{T}}, DY::Union{DenseArray{T},AbstractVector{T}}) where T<:BlasReal
     require_one_based_indexing(DX, DY)
     n = length(DX)

stdlib/LinearAlgebra/test/matmul.jl

@@ -205,6 +205,19 @@ end
     @test *(Asub, adjoint(Asub)) == *(Aref, adjoint(Aref))
 end
 
+@testset "dot product of subarrays of vectors (floats, negative stride, issue #37767)" begin
+    for T in (Float32, Float64, ComplexF32, ComplexF64)
+        a = Vector{T}(3:2:7)
+        b = Vector{T}(1:10)
+        v = view(b,7:-2:3)
+        @test dot(a,Vector(v)) ≈ 67.0
+        @test dot(a,v) ≈ 67.0
+        @test dot(v,a) ≈ 67.0
+        @test dot(Vector(v),Vector(v)) ≈ 83.0
+        @test dot(v,v) ≈ 83.0
+    end
+end
+
 @testset "Complex matrix x real MatOrVec etc (issue #29224)" for T1 in (Float32,Float64)
     for T2 in (Float32,Float64)
         for arg1_real in (true,false)