Vectorise indexing of contiguous PyArrays

Branch on strides[1] == sizeof(R) so LLVM can version loops on the
contiguous case. No change to the type parameters.

Author: Beforerr

CHANGELOG.md

@@ -1,6 +1,7 @@
 # Changelog
 
 ## Unreleased
+* Faster indexing into contiguous `PyArray`s.
 * Bug fixes.
 
 ## 0.9.36 (2026-09-18)

src/Wrap/PyArray.jl

@@ -672,11 +672,17 @@ end
 end
 
 pyarray_offset(x::PyArray{T,N,M,true}, i::Int) where {T,N,M} =
-    N == 0 ? 0 : (i - 1) * x.strides[1]
-pyarray_offset(x::PyArray{T,1,M,true}, i::Int) where {T,M} = (i - 1) .* x.strides[1]
-pyarray_offset(x::PyArray{T,N}, i::Vararg{Int,N}) where {T,N} = sum((i .- 1) .* x.strides)
+    N == 0 ? 0 : pyarray_offset1(x, i)
+pyarray_offset(x::PyArray{T,1,M,true}, i::Int) where {T,M} = pyarray_offset1(x, i)
+pyarray_offset(x::PyArray{T,N}, i::Vararg{Int,N}) where {T,N} =
+    pyarray_offset1(x, i[1]) + sum((Base.tail(i) .- 1) .* Base.tail(x.strides); init = 0)
 pyarray_offset(x::PyArray{T,0}) where {T} = 0
 
+# The type can't say the first stride is the element size (strided arrays can be linear too).
+# Branching on it lets LLVM version loops on the contiguous case and vectorise them.
+pyarray_offset1(x::PyArray{T,N,M,L,R}, i::Int) where {T,N,M,L,R} =
+    x.strides[1] == sizeof(R) ? (i - 1) * sizeof(R) : (i - 1) * x.strides[1]
+
 function pyarray_load(::Type{T}, p::Ptr{R}) where {T,R}
     if R == T
         unsafe_load(p)

test/Wrap.jl

@@ -99,6 +99,23 @@
         )
         @test_throws Exception PyArray(nd; array = false, buffer = true)
     end
+    @testset "linear with unit and non-unit first stride" begin
+        tb = pyimport("_testbuffer")
+        nd = tb.ndarray(
+            pylist(1:24),
+            shape = pylist([4, 6]),
+            format = "i",
+            flags = tb.ND_FORTRAN,
+        )
+        e = reshape(1:24, 4, 6)
+        for (v, ev) in
+            ((nd, e), (nd[pyslice(nothing, nothing, 2), pyslice(nothing)], e[1:2:end, :]))
+            a = PyArray(v; array = false)
+            @test a isa PyArray{Cint,2,false,true,Cint}
+            @test [a[i, j] for i in axes(a, 1), j in axes(a, 2)] == ev
+            @test [a[i] for i in eachindex(IndexLinear(), a)] == vec(ev)
+        end
+    end
 end
 
 @testitem "PyDict" begin