diff --git a/docs/advanced/pycpp/numpy.rst b/docs/advanced/pycpp/numpy.rst
index d06370611..71123631a 100644
--- a/docs/advanced/pycpp/numpy.rst
+++ b/docs/advanced/pycpp/numpy.rst
@@ -57,10 +57,10 @@ specification.
 
     struct buffer_info {
         void *ptr;
-        size_t itemsize;
+        ssize_t itemsize;
         std::string format;
-        int ndim;
-        std::vector<size_t> shape;
+        ssize_t ndim;
+        std::vector<ssize_t> shape;
         std::vector<ssize_t> strides;
     };
 
@@ -95,8 +95,8 @@ buffer objects (e.g. a NumPy matrix).
                 throw std::runtime_error("Incompatible buffer dimension!");
 
             auto strides = Strides(
-                info.strides[rowMajor ? 0 : 1] / sizeof(Scalar),
-                info.strides[rowMajor ? 1 : 0] / sizeof(Scalar));
+                info.strides[rowMajor ? 0 : 1] / (py::ssize_t)sizeof(Scalar),
+                info.strides[rowMajor ? 1 : 0] / (py::ssize_t)sizeof(Scalar));
 
             auto map = Eigen::Map<Matrix, 0, Strides>(
                 static_cat<Scalar *>(info.ptr), info.shape[0], info.shape[1], strides);
@@ -111,17 +111,14 @@ as follows:
 
     .def_buffer([](Matrix &m) -> py::buffer_info {
         return py::buffer_info(
-            m.data(),                /* Pointer to buffer */
-            sizeof(Scalar),          /* Size of one scalar */
-            /* Python struct-style format descriptor */
-            py::format_descriptor<Scalar>::format(),
-            /* Number of dimensions */
-            2,
-            /* Buffer dimensions */
-            { m.rows(), m.cols() },
-            /* Strides (in bytes) for each index */
+            m.data(),                                /* Pointer to buffer */
+            sizeof(Scalar),                          /* Size of one scalar */
+            py::format_descriptor<Scalar>::format(), /* Python struct-style format descriptor */
+            2,                                       /* Number of dimensions */
+            { m.rows(), m.cols() },                  /* Buffer dimensions */
             { sizeof(Scalar) * (rowMajor ? m.cols() : 1),
               sizeof(Scalar) * (rowMajor ? 1 : m.rows()) }
+                                                     /* Strides (in bytes) for each index */
         );
      })
 
@@ -321,17 +318,17 @@ where ``N`` gives the required dimensionality of the array:
     m.def("sum_3d", [](py::array_t<double> x) {
         auto r = x.unchecked<3>(); // x must have ndim = 3; can be non-writeable
         double sum = 0;
-        for (size_t i = 0; i < r.shape(0); i++)
-            for (size_t j = 0; j < r.shape(1); j++)
-                for (size_t k = 0; k < r.shape(2); k++)
+        for (ssize_t i = 0; i < r.shape(0); i++)
+            for (ssize_t j = 0; j < r.shape(1); j++)
+                for (ssize_t k = 0; k < r.shape(2); k++)
                     sum += r(i, j, k);
         return sum;
     });
     m.def("increment_3d", [](py::array_t<double> x) {
         auto r = x.mutable_unchecked<3>(); // Will throw if ndim != 3 or flags.writeable is false
-        for (size_t i = 0; i < r.shape(0); i++)
-            for (size_t j = 0; j < r.shape(1); j++)
-                for (size_t k = 0; k < r.shape(2); k++)
+        for (ssize_t i = 0; i < r.shape(0); i++)
+            for (ssize_t j = 0; j < r.shape(1); j++)
+                for (ssize_t k = 0; k < r.shape(2); k++)
                     r(i, j, k) += 1.0;
     }, py::arg().noconvert());
 
diff --git a/include/pybind11/buffer_info.h b/include/pybind11/buffer_info.h
index 3090a4242..08877ef30 100644
--- a/include/pybind11/buffer_info.h
+++ b/include/pybind11/buffer_info.h
@@ -15,31 +15,31 @@ NAMESPACE_BEGIN(pybind11)
 
 /// Information record describing a Python buffer object
 struct buffer_info {
-    void *ptr = nullptr;         // Pointer to the underlying storage
-    size_t itemsize = 0;         // Size of individual items in bytes
-    size_t size = 0;             // Total number of entries
-    std::string format;          // For homogeneous buffers, this should be set to format_descriptor<T>::format()
-    size_t ndim = 0;             // Number of dimensions
-    std::vector<size_t> shape;   // Shape of the tensor (1 entry per dimension)
+    void *ptr = nullptr;          // Pointer to the underlying storage
+    ssize_t itemsize = 0;         // Size of individual items in bytes
+    ssize_t size = 0;             // Total number of entries
+    std::string format;           // For homogeneous buffers, this should be set to format_descriptor<T>::format()
+    ssize_t ndim = 0;             // Number of dimensions
+    std::vector<ssize_t> shape;   // Shape of the tensor (1 entry per dimension)
     std::vector<ssize_t> strides; // Number of entries between adjacent entries (for each per dimension)
 
     buffer_info() { }
 
-    buffer_info(void *ptr, size_t itemsize, const std::string &format, size_t ndim,
-                detail::any_container<size_t> shape_in, detail::any_container<ssize_t> strides_in)
+    buffer_info(void *ptr, ssize_t itemsize, const std::string &format, ssize_t ndim,
+                detail::any_container<ssize_t> shape_in, detail::any_container<ssize_t> strides_in)
     : ptr(ptr), itemsize(itemsize), size(1), format(format), ndim(ndim),
       shape(std::move(shape_in)), strides(std::move(strides_in)) {
-        if (ndim != shape.size() || ndim != strides.size())
+        if (ndim != (ssize_t) shape.size() || ndim != (ssize_t) strides.size())
             pybind11_fail("buffer_info: ndim doesn't match shape and/or strides length");
-        for (size_t i = 0; i < ndim; ++i)
+        for (size_t i = 0; i < (size_t) ndim; ++i)
             size *= shape[i];
     }
 
-    buffer_info(void *ptr, size_t itemsize, const std::string &format, size_t size)
+    buffer_info(void *ptr, ssize_t itemsize, const std::string &format, ssize_t size)
     : buffer_info(ptr, itemsize, format, 1, {size}, {itemsize}) { }
 
     explicit buffer_info(Py_buffer *view, bool ownview = true)
-    : buffer_info(view->buf, (size_t) view->itemsize, view->format, (size_t) view->ndim,
+    : buffer_info(view->buf, view->itemsize, view->format, view->ndim,
             {view->shape, view->shape + view->ndim}, {view->strides, view->strides + view->ndim}) {
         this->view = view;
         this->ownview = ownview;
@@ -78,13 +78,13 @@ NAMESPACE_BEGIN(detail)
 
 template <typename T, typename SFINAE = void> struct compare_buffer_info {
     static bool compare(const buffer_info& b) {
-        return b.format == format_descriptor<T>::format() && b.itemsize == sizeof(T);
+        return b.format == format_descriptor<T>::format() && b.itemsize == (ssize_t) sizeof(T);
     }
 };
 
 template <typename T> struct compare_buffer_info<T, detail::enable_if_t<std::is_integral<T>::value>> {
     static bool compare(const buffer_info& b) {
-        return b.itemsize == sizeof(T) && (b.format == format_descriptor<T>::value ||
+        return (size_t) b.itemsize == sizeof(T) && (b.format == format_descriptor<T>::value ||
             ((sizeof(T) == sizeof(long)) && b.format == (std::is_unsigned<T>::value ? "L" : "l")) ||
             ((sizeof(T) == sizeof(size_t)) && b.format == (std::is_unsigned<T>::value ? "N" : "n")));
     }
diff --git a/include/pybind11/class_support.h b/include/pybind11/class_support.h
index 99d1477f9..995f57adf 100644
--- a/include/pybind11/class_support.h
+++ b/include/pybind11/class_support.h
@@ -433,7 +433,7 @@ inline void enable_dynamic_attributes(PyHeapTypeObject *heap_type) {
 #endif
     type->tp_flags |= Py_TPFLAGS_HAVE_GC;
     type->tp_dictoffset = type->tp_basicsize; // place dict at the end
-    type->tp_basicsize += (Py_ssize_t)sizeof(PyObject *); // and allocate enough space for it
+    type->tp_basicsize += (ssize_t)sizeof(PyObject *); // and allocate enough space for it
     type->tp_traverse = pybind11_traverse;
     type->tp_clear = pybind11_clear;
 
@@ -459,18 +459,16 @@ extern "C" inline int pybind11_getbuffer(PyObject *obj, Py_buffer *view, int fla
     view->ndim = 1;
     view->internal = info;
     view->buf = info->ptr;
-    view->itemsize = (Py_ssize_t) info->itemsize;
+    view->itemsize = info->itemsize;
     view->len = view->itemsize;
     for (auto s : info->shape)
-        view->len *= (Py_ssize_t) s;
+        view->len *= s;
     if ((flags & PyBUF_FORMAT) == PyBUF_FORMAT)
         view->format = const_cast<char *>(info->format.c_str());
     if ((flags & PyBUF_STRIDES) == PyBUF_STRIDES) {
         view->ndim = (int) info->ndim;
         view->strides = &info->strides[0];
-        // Next is a pointer cast, let's make sure it's safe.
-        static_assert(sizeof(Py_ssize_t)==sizeof(info->shape[0]), "sizeof(Py_ssize_t) != sizeof(size_t)");
-        view->shape = (Py_ssize_t *) &info->shape[0];
+        view->shape = &info->shape[0];
     }
     Py_INCREF(view->obj);
     return 0;
diff --git a/include/pybind11/eigen.h b/include/pybind11/eigen.h
index 8022a2cbe..b78045520 100644
--- a/include/pybind11/eigen.h
+++ b/include/pybind11/eigen.h
@@ -68,7 +68,7 @@ template <typename T> using is_eigen_other = all_of<
 template <bool EigenRowMajor> struct EigenConformable {
     bool conformable = false;
     EigenIndex rows = 0, cols = 0;
-    EigenDStride stride{0, 0};      // Only valid if negativestridees is false!
+    EigenDStride stride{0, 0};      // Only valid if negativestrides is false!
     bool negativestrides = false;   // If true, do not use stride!
 
     EigenConformable(bool fits = false) : conformable{fits} {}
@@ -207,7 +207,7 @@ template <typename Type_> struct EigenProps {
 // Casts an Eigen type to numpy array.  If given a base, the numpy array references the src data,
 // otherwise it'll make a copy.  writeable lets you turn off the writeable flag for the array.
 template <typename props> handle eigen_array_cast(typename props::Type const &src, handle base = handle(), bool writeable = true) {
-    constexpr size_t elem_size = sizeof(typename props::Scalar);
+    constexpr ssize_t elem_size = sizeof(typename props::Scalar);
     array a;
     if (props::vector)
         a = array({ src.size() }, { elem_size * src.innerStride() }, src.data(), base);
@@ -581,9 +581,9 @@ struct type_caster<Type, enable_if_t<is_eigen_sparse<Type>::value>> {
         object matrix_type = module::import("scipy.sparse").attr(
             rowMajor ? "csr_matrix" : "csc_matrix");
 
-        array data((size_t) src.nonZeros(), src.valuePtr());
-        array outerIndices((size_t) (rowMajor ? src.rows() : src.cols()) + 1, src.outerIndexPtr());
-        array innerIndices((size_t) src.nonZeros(), src.innerIndexPtr());
+        array data(src.nonZeros(), src.valuePtr());
+        array outerIndices((rowMajor ? src.rows() : src.cols()) + 1, src.outerIndexPtr());
+        array innerIndices(src.nonZeros(), src.innerIndexPtr());
 
         return matrix_type(
             std::make_tuple(data, innerIndices, outerIndices),
diff --git a/include/pybind11/numpy.h b/include/pybind11/numpy.h
index 0107366b7..72bb35001 100644
--- a/include/pybind11/numpy.h
+++ b/include/pybind11/numpy.h
@@ -251,10 +251,10 @@ template <typename T> using is_pod_struct = all_of<
     satisfies_none_of<T, std::is_reference, std::is_array, is_std_array, std::is_arithmetic, is_complex, std::is_enum>
 >;
 
-template <size_t Dim = 0, typename Strides> ssize_t byte_offset_unsafe(const Strides &) { return 0; }
-template <size_t Dim = 0, typename Strides, typename... Ix>
-ssize_t byte_offset_unsafe(const Strides &strides, size_t i, Ix... index) {
-    return static_cast<ssize_t>(i) * strides[Dim] + byte_offset_unsafe<Dim + 1>(strides, index...);
+template <ssize_t Dim = 0, typename Strides> ssize_t byte_offset_unsafe(const Strides &) { return 0; }
+template <ssize_t Dim = 0, typename Strides, typename... Ix>
+ssize_t byte_offset_unsafe(const Strides &strides, ssize_t i, Ix... index) {
+    return i * strides[Dim] + byte_offset_unsafe<Dim + 1>(strides, index...);
 }
 
 /** Proxy class providing unsafe, unchecked const access to array data.  This is constructed through
@@ -268,23 +268,23 @@ protected:
     const unsigned char *data_;
     // Storing the shape & strides in local variables (i.e. these arrays) allows the compiler to
     // make large performance gains on big, nested loops, but requires compile-time dimensions
-    conditional_t<Dynamic, const size_t *, std::array<size_t, (size_t) Dims>> shape_;
-    conditional_t<Dynamic, const ssize_t *, std::array<ssize_t, (size_t) Dims>> strides_;
-    const size_t dims_;
+    conditional_t<Dynamic, const ssize_t *, std::array<ssize_t, (size_t) Dims>>
+            shape_, strides_;
+    const ssize_t dims_;
 
     friend class pybind11::array;
     // Constructor for compile-time dimensions:
     template <bool Dyn = Dynamic>
-    unchecked_reference(const void *data, const size_t *shape, const ssize_t *strides, enable_if_t<!Dyn, size_t>)
+    unchecked_reference(const void *data, const ssize_t *shape, const ssize_t *strides, enable_if_t<!Dyn, ssize_t>)
     : data_{reinterpret_cast<const unsigned char *>(data)}, dims_{Dims} {
-        for (size_t i = 0; i < dims_; i++) {
+        for (size_t i = 0; i < (size_t) dims_; i++) {
             shape_[i] = shape[i];
             strides_[i] = strides[i];
         }
     }
     // Constructor for runtime dimensions:
     template <bool Dyn = Dynamic>
-    unchecked_reference(const void *data, const size_t *shape, const ssize_t *strides, enable_if_t<Dyn, size_t> dims)
+    unchecked_reference(const void *data, const ssize_t *shape, const ssize_t *strides, enable_if_t<Dyn, ssize_t> dims)
     : data_{reinterpret_cast<const unsigned char *>(data)}, shape_{shape}, strides_{strides}, dims_{dims} {}
 
 public:
@@ -295,39 +295,39 @@ public:
     template <typename... Ix> const T &operator()(Ix... index) const {
         static_assert(sizeof...(Ix) == Dims || Dynamic,
                 "Invalid number of indices for unchecked array reference");
-        return *reinterpret_cast<const T *>(data_ + byte_offset_unsafe(strides_, size_t(index)...));
+        return *reinterpret_cast<const T *>(data_ + byte_offset_unsafe(strides_, ssize_t(index)...));
     }
     /** Unchecked const reference access to data; this operator only participates if the reference
      * is to a 1-dimensional array.  When present, this is exactly equivalent to `obj(index)`.
      */
-    template <size_t D = Dims, typename = enable_if_t<D == 1 || Dynamic>>
-    const T &operator[](size_t index) const { return operator()(index); }
+    template <ssize_t D = Dims, typename = enable_if_t<D == 1 || Dynamic>>
+    const T &operator[](ssize_t index) const { return operator()(index); }
 
     /// Pointer access to the data at the given indices.
-    template <typename... Ix> const T *data(Ix... ix) const { return &operator()(size_t(ix)...); }
+    template <typename... Ix> const T *data(Ix... ix) const { return &operator()(ssize_t(ix)...); }
 
     /// Returns the item size, i.e. sizeof(T)
-    constexpr static size_t itemsize() { return sizeof(T); }
+    constexpr static ssize_t itemsize() { return sizeof(T); }
 
     /// Returns the shape (i.e. size) of dimension `dim`
-    size_t shape(size_t dim) const { return shape_[dim]; }
+    ssize_t shape(ssize_t dim) const { return shape_[(size_t) dim]; }
 
     /// Returns the number of dimensions of the array
-    size_t ndim() const { return dims_; }
+    ssize_t ndim() const { return dims_; }
 
     /// Returns the total number of elements in the referenced array, i.e. the product of the shapes
     template <bool Dyn = Dynamic>
-    enable_if_t<!Dyn, size_t> size() const {
-        return std::accumulate(shape_.begin(), shape_.end(), (size_t) 1, std::multiplies<size_t>());
+    enable_if_t<!Dyn, ssize_t> size() const {
+        return std::accumulate(shape_.begin(), shape_.end(), (ssize_t) 1, std::multiplies<ssize_t>());
     }
     template <bool Dyn = Dynamic>
-    enable_if_t<Dyn, size_t> size() const {
-        return std::accumulate(shape_, shape_ + ndim(), (size_t) 1, std::multiplies<size_t>());
+    enable_if_t<Dyn, ssize_t> size() const {
+        return std::accumulate(shape_, shape_ + ndim(), (ssize_t) 1, std::multiplies<ssize_t>());
     }
 
     /// Returns the total number of bytes used by the referenced data.  Note that the actual span in
     /// memory may be larger if the referenced array has non-contiguous strides (e.g. for a slice).
-    size_t nbytes() const {
+    ssize_t nbytes() const {
         return size() * itemsize();
     }
 };
@@ -349,11 +349,11 @@ public:
      * reference is to a 1-dimensional array (or has runtime dimensions).  When present, this is
      * exactly equivalent to `obj(index)`.
      */
-    template <size_t D = Dims, typename = enable_if_t<D == 1 || Dynamic>>
-    T &operator[](size_t index) { return operator()(index); }
+    template <ssize_t D = Dims, typename = enable_if_t<D == 1 || Dynamic>>
+    T &operator[](ssize_t index) { return operator()(index); }
 
     /// Mutable pointer access to the data at the given indices.
-    template <typename... Ix> T *mutable_data(Ix... ix) { return &operator()(size_t(ix)...); }
+    template <typename... Ix> T *mutable_data(Ix... ix) { return &operator()(ssize_t(ix)...); }
 };
 
 template <typename T, ssize_t Dim>
@@ -381,7 +381,7 @@ public:
 
     dtype(const char *format) : dtype(std::string(format)) { }
 
-    dtype(list names, list formats, list offsets, size_t itemsize) {
+    dtype(list names, list formats, list offsets, ssize_t itemsize) {
         dict args;
         args["names"] = names;
         args["formats"] = formats;
@@ -404,8 +404,8 @@ public:
     }
 
     /// Size of the data type in bytes.
-    size_t itemsize() const {
-        return (size_t) detail::array_descriptor_proxy(m_ptr)->elsize;
+    ssize_t itemsize() const {
+        return detail::array_descriptor_proxy(m_ptr)->elsize;
     }
 
     /// Returns true for structured data types.
@@ -425,7 +425,7 @@ private:
         return reinterpret_borrow<object>(obj);
     }
 
-    dtype strip_padding(size_t itemsize) {
+    dtype strip_padding(ssize_t itemsize) {
         // Recursively strip all void fields with empty names that are generated for
         // padding fields (as of NumPy v1.11).
         if (!has_fields())
@@ -501,7 +501,7 @@ public:
             api.PyArray_Type_, descr.release().ptr(), (int) ndim, shape->data(), strides->data(),
             const_cast<void *>(ptr), flags, nullptr));
         if (!tmp)
-            pybind11_fail("NumPy: unable to create array!");
+            throw error_already_set();
         if (ptr) {
             if (base) {
                 api.PyArray_SetBaseObject_(tmp.ptr(), base.inc_ref().ptr());
@@ -528,7 +528,7 @@ public:
         : array(std::move(shape), {}, ptr, base) { }
 
     template <typename T>
-    explicit array(size_t count, const T *ptr, handle base = handle()) : array({count}, {}, ptr, base) { }
+    explicit array(ssize_t count, const T *ptr, handle base = handle()) : array({count}, {}, ptr, base) { }
 
     explicit array(const buffer_info &info)
     : array(pybind11::dtype(info), info.shape, info.strides, info.ptr) { }
@@ -539,23 +539,23 @@ public:
     }
 
     /// Total number of elements
-    size_t size() const {
-        return std::accumulate(shape(), shape() + ndim(), (size_t) 1, std::multiplies<size_t>());
+    ssize_t size() const {
+        return std::accumulate(shape(), shape() + ndim(), (ssize_t) 1, std::multiplies<ssize_t>());
     }
 
     /// Byte size of a single element
-    size_t itemsize() const {
-        return (size_t) detail::array_descriptor_proxy(detail::array_proxy(m_ptr)->descr)->elsize;
+    ssize_t itemsize() const {
+        return detail::array_descriptor_proxy(detail::array_proxy(m_ptr)->descr)->elsize;
     }
 
     /// Total number of bytes
-    size_t nbytes() const {
+    ssize_t nbytes() const {
         return size() * itemsize();
     }
 
     /// Number of dimensions
-    size_t ndim() const {
-        return (size_t) detail::array_proxy(m_ptr)->nd;
+    ssize_t ndim() const {
+        return detail::array_proxy(m_ptr)->nd;
     }
 
     /// Base object
@@ -564,12 +564,12 @@ public:
     }
 
     /// Dimensions of the array
-    const size_t* shape() const {
-        return reinterpret_cast<const size_t *>(detail::array_proxy(m_ptr)->dimensions);
+    const ssize_t* shape() const {
+        return detail::array_proxy(m_ptr)->dimensions;
     }
 
     /// Dimension along a given axis
-    size_t shape(size_t dim) const {
+    ssize_t shape(ssize_t dim) const {
         if (dim >= ndim())
             fail_dim_check(dim, "invalid axis");
         return shape()[dim];
@@ -577,11 +577,11 @@ public:
 
     /// Strides of the array
     const ssize_t* strides() const {
-        return reinterpret_cast<const ssize_t *>(detail::array_proxy(m_ptr)->strides);
+        return detail::array_proxy(m_ptr)->strides;
     }
 
     /// Stride along a given axis
-    ssize_t strides(size_t dim) const {
+    ssize_t strides(ssize_t dim) const {
         if (dim >= ndim())
             fail_dim_check(dim, "invalid axis");
         return strides()[dim];
@@ -619,9 +619,9 @@ public:
     /// Byte offset from beginning of the array to a given index (full or partial).
     /// May throw if the index would lead to out of bounds access.
     template<typename... Ix> ssize_t offset_at(Ix... index) const {
-        if (sizeof...(index) > ndim())
+        if ((ssize_t) sizeof...(index) > ndim())
             fail_dim_check(sizeof...(index), "too many indices for an array");
-        return byte_offset(size_t(index)...);
+        return byte_offset(ssize_t(index)...);
     }
 
     ssize_t offset_at() const { return 0; }
@@ -629,7 +629,7 @@ public:
     /// Item count from beginning of the array to a given index (full or partial).
     /// May throw if the index would lead to out of bounds access.
     template<typename... Ix> ssize_t index_at(Ix... index) const {
-        return offset_at(index...) / static_cast<ssize_t>(itemsize());
+        return offset_at(index...) / itemsize();
     }
 
     /** Returns a proxy object that provides access to the array's data without bounds or
@@ -638,7 +638,7 @@ public:
      * and the caller must take care not to access invalid dimensions or dimension indices.
      */
     template <typename T, ssize_t Dims = -1> detail::unchecked_mutable_reference<T, Dims> mutable_unchecked() {
-        if (Dims >= 0 && ndim() != (size_t) Dims)
+        if (Dims >= 0 && ndim() != Dims)
             throw std::domain_error("array has incorrect number of dimensions: " + std::to_string(ndim()) +
                     "; expected " + std::to_string(Dims));
         return detail::unchecked_mutable_reference<T, Dims>(mutable_data(), shape(), strides(), ndim());
@@ -651,7 +651,7 @@ public:
      * invalid dimensions or dimension indices.
      */
     template <typename T, ssize_t Dims = -1> detail::unchecked_reference<T, Dims> unchecked() const {
-        if (Dims >= 0 && ndim() != (size_t) Dims)
+        if (Dims >= 0 && ndim() != Dims)
             throw std::domain_error("array has incorrect number of dimensions: " + std::to_string(ndim()) +
                     "; expected " + std::to_string(Dims));
         return detail::unchecked_reference<T, Dims>(data(), shape(), strides(), ndim());
@@ -690,14 +690,14 @@ public:
 protected:
     template<typename, typename> friend struct detail::npy_format_descriptor;
 
-    void fail_dim_check(size_t dim, const std::string& msg) const {
+    void fail_dim_check(ssize_t dim, const std::string& msg) const {
         throw index_error(msg + ": " + std::to_string(dim) +
                           " (ndim = " + std::to_string(ndim()) + ")");
     }
 
     template<typename... Ix> ssize_t byte_offset(Ix... index) const {
         check_dimensions(index...);
-        return detail::byte_offset_unsafe(strides(), size_t(index)...);
+        return detail::byte_offset_unsafe(strides(), ssize_t(index)...);
     }
 
     void check_writeable() const {
@@ -705,7 +705,7 @@ protected:
             throw std::domain_error("array is not writeable");
     }
 
-    static std::vector<ssize_t> default_strides(const std::vector<size_t>& shape, size_t itemsize) {
+    static std::vector<Py_intptr_t> default_strides(const std::vector<Py_intptr_t>& shape, ssize_t itemsize) {
         auto ndim = shape.size();
         std::vector<ssize_t> strides(ndim);
         if (ndim) {
@@ -718,12 +718,12 @@ protected:
     }
 
     template<typename... Ix> void check_dimensions(Ix... index) const {
-        check_dimensions_impl(size_t(0), shape(), size_t(index)...);
+        check_dimensions_impl(ssize_t(0), shape(), ssize_t(index)...);
     }
 
-    void check_dimensions_impl(size_t, const size_t*) const { }
+    void check_dimensions_impl(ssize_t, const ssize_t*) const { }
 
-    template<typename... Ix> void check_dimensions_impl(size_t axis, const size_t* shape, size_t i, Ix... index) const {
+    template<typename... Ix> void check_dimensions_impl(ssize_t axis, const ssize_t* shape, ssize_t i, Ix... index) const {
         if (i >= *shape) {
             throw index_error(std::string("index ") + std::to_string(i) +
                               " is out of bounds for axis " + std::to_string(axis) +
@@ -772,12 +772,12 @@ public:
     explicit array_t(size_t count, const T *ptr = nullptr, handle base = handle())
         : array({count}, {}, ptr, base) { }
 
-    constexpr size_t itemsize() const {
+    constexpr ssize_t itemsize() const {
         return sizeof(T);
     }
 
     template<typename... Ix> ssize_t index_at(Ix... index) const {
-        return offset_at(index...) / static_cast<ssize_t>(itemsize());
+        return offset_at(index...) / itemsize();
     }
 
     template<typename... Ix> const T* data(Ix... index) const {
@@ -792,14 +792,14 @@ public:
     template<typename... Ix> const T& at(Ix... index) const {
         if (sizeof...(index) != ndim())
             fail_dim_check(sizeof...(index), "index dimension mismatch");
-        return *(static_cast<const T*>(array::data()) + byte_offset(size_t(index)...) / static_cast<ssize_t>(itemsize()));
+        return *(static_cast<const T*>(array::data()) + byte_offset(ssize_t(index)...) / itemsize());
     }
 
     // Mutable reference to element at a given index
     template<typename... Ix> T& mutable_at(Ix... index) {
         if (sizeof...(index) != ndim())
             fail_dim_check(sizeof...(index), "index dimension mismatch");
-        return *(static_cast<T*>(array::mutable_data()) + byte_offset(size_t(index)...) / static_cast<ssize_t>(itemsize()));
+        return *(static_cast<T*>(array::mutable_data()) + byte_offset(ssize_t(index)...) / itemsize());
     }
 
     /** Returns a proxy object that provides access to the array's data without bounds or
@@ -949,16 +949,16 @@ public:
 
 struct field_descriptor {
     const char *name;
-    size_t offset;
-    size_t size;
-    size_t alignment;
+    ssize_t offset;
+    ssize_t size;
+    ssize_t alignment;
     std::string format;
     dtype descr;
 };
 
 inline PYBIND11_NOINLINE void register_structured_dtype(
     const std::initializer_list<field_descriptor>& fields,
-    const std::type_info& tinfo, size_t itemsize,
+    const std::type_info& tinfo, ssize_t itemsize,
     bool (*direct_converter)(PyObject *, void *&)) {
 
     auto& numpy_internals = get_numpy_internals();
@@ -986,7 +986,7 @@ inline PYBIND11_NOINLINE void register_structured_dtype(
     std::vector<field_descriptor> ordered_fields(fields);
     std::sort(ordered_fields.begin(), ordered_fields.end(),
         [](const field_descriptor &a, const field_descriptor &b) { return a.offset < b.offset; });
-    size_t offset = 0;
+    ssize_t offset = 0;
     std::ostringstream oss;
     oss << "T{";
     for (auto& field : ordered_fields) {
@@ -1142,7 +1142,7 @@ public:
 
     common_iterator() : p_ptr(0), m_strides() {}
 
-    common_iterator(void* ptr, const container_type& strides, const std::vector<size_t>& shape)
+    common_iterator(void* ptr, const container_type& strides, const container_type& shape)
         : p_ptr(reinterpret_cast<char*>(ptr)), m_strides(strides.size()) {
         m_strides.back() = static_cast<value_type>(strides.back());
         for (size_type i = m_strides.size() - 1; i != 0; --i) {
@@ -1167,18 +1167,18 @@ private:
 
 template <size_t N> class multi_array_iterator {
 public:
-    using container_type = std::vector<size_t>;
+    using container_type = std::vector<ssize_t>;
 
     multi_array_iterator(const std::array<buffer_info, N> &buffers,
-                         const std::vector<size_t> &shape)
+                         const container_type &shape)
         : m_shape(shape.size()), m_index(shape.size(), 0),
           m_common_iterator() {
 
         // Manual copy to avoid conversion warning if using std::copy
         for (size_t i = 0; i < shape.size(); ++i)
-            m_shape[i] = static_cast<container_type::value_type>(shape[i]);
+            m_shape[i] = shape[i];
 
-        std::vector<ssize_t> strides(shape.size());
+        container_type strides(shape.size());
         for (size_t i = 0; i < N; ++i)
             init_common_iterator(buffers[i], shape, m_common_iterator[i], strides);
     }
@@ -1205,8 +1205,9 @@ private:
     using common_iter = common_iterator;
 
     void init_common_iterator(const buffer_info &buffer,
-                              const std::vector<size_t> &shape,
-                              common_iter &iterator, std::vector<ssize_t> &strides) {
+                              const container_type &shape,
+                              common_iter &iterator,
+                              container_type &strides) {
         auto buffer_shape_iter = buffer.shape.rbegin();
         auto buffer_strides_iter = buffer.strides.rbegin();
         auto shape_iter = shape.rbegin();
@@ -1245,13 +1246,13 @@ enum class broadcast_trivial { non_trivial, c_trivial, f_trivial };
 // singleton or a full-size, C-contiguous (`c_trivial`) or Fortran-contiguous (`f_trivial`) storage
 // buffer; returns `non_trivial` otherwise.
 template <size_t N>
-broadcast_trivial broadcast(const std::array<buffer_info, N> &buffers, size_t &ndim, std::vector<size_t> &shape) {
-    ndim = std::accumulate(buffers.begin(), buffers.end(), size_t(0), [](size_t res, const buffer_info& buf) {
+broadcast_trivial broadcast(const std::array<buffer_info, N> &buffers, ssize_t &ndim, std::vector<ssize_t> &shape) {
+    ndim = std::accumulate(buffers.begin(), buffers.end(), ssize_t(0), [](ssize_t res, const buffer_info& buf) {
         return std::max(res, buf.ndim);
     });
 
     shape.clear();
-    shape.resize(ndim, 1);
+    shape.resize((size_t) ndim, 1);
 
     // Figure out the output size, and make sure all input arrays conform (i.e. are either size 1 or
     // the full size).
@@ -1286,11 +1287,10 @@ broadcast_trivial broadcast(const std::array<buffer_info, N> &buffers, size_t &n
 
         // Check for C contiguity (but only if previous inputs were also C contiguous)
         if (trivial_broadcast_c) {
-            ssize_t expect_stride = static_cast<ssize_t>(buffers[i].itemsize);
+            ssize_t expect_stride = buffers[i].itemsize;
             auto end = buffers[i].shape.crend();
-            auto shape_iter = buffers[i].shape.crbegin();
-            auto stride_iter = buffers[i].strides.crbegin();
-            for (; trivial_broadcast_c && shape_iter != end; ++shape_iter, ++stride_iter) {
+            for (auto shape_iter = buffers[i].shape.crbegin(), stride_iter = buffers[i].strides.crbegin();
+                    trivial_broadcast_c && shape_iter != end; ++shape_iter, ++stride_iter) {
                 if (expect_stride == *stride_iter)
                     expect_stride *= *shape_iter;
                 else
@@ -1300,11 +1300,10 @@ broadcast_trivial broadcast(const std::array<buffer_info, N> &buffers, size_t &n
 
         // Check for Fortran contiguity (if previous inputs were also F contiguous)
         if (trivial_broadcast_f) {
-            ssize_t expect_stride = static_cast<ssize_t>(buffers[i].itemsize);
+            ssize_t expect_stride = buffers[i].itemsize;
             auto end = buffers[i].shape.cend();
-            auto shape_iter = buffers[i].shape.cbegin();
-            auto stride_iter = buffers[i].strides.cbegin();
-            for (; trivial_broadcast_f && shape_iter != end; ++shape_iter, ++stride_iter) {
+            for (auto shape_iter = buffers[i].shape.cbegin(), stride_iter = buffers[i].strides.cbegin();
+                    trivial_broadcast_f && shape_iter != end; ++shape_iter, ++stride_iter) {
                 if (expect_stride == *stride_iter)
                     expect_stride *= *shape_iter;
                 else
@@ -1336,25 +1335,26 @@ struct vectorize_helper {
         std::array<buffer_info, N> buffers {{ args.request()... }};
 
         /* Determine dimensions parameters of output array */
-        size_t ndim = 0;
-        std::vector<size_t> shape(0);
-        auto trivial = broadcast(buffers, ndim, shape);
+        ssize_t nd = 0;
+        std::vector<ssize_t> shape(0);
+        auto trivial = broadcast(buffers, nd, shape);
+        size_t ndim = (size_t) nd;
 
-        size_t size = 1;
+        ssize_t size = 1;
         std::vector<ssize_t> strides(ndim);
         if (ndim > 0) {
             if (trivial == broadcast_trivial::f_trivial) {
-                strides[0] = static_cast<ssize_t>(sizeof(Return));
+                strides[0] = sizeof(Return);
                 for (size_t i = 1; i < ndim; ++i) {
-                    strides[i] = strides[i - 1] * static_cast<ssize_t>(shape[i - 1]);
+                    strides[i] = strides[i - 1] * shape[i - 1];
                     size *= shape[i - 1];
                 }
                 size *= shape[ndim - 1];
             }
             else {
-                strides[ndim-1] = static_cast<ssize_t>(sizeof(Return));
+                strides[ndim-1] = sizeof(Return);
                 for (size_t i = ndim - 1; i > 0; --i) {
-                    strides[i - 1] = strides[i] * static_cast<ssize_t>(shape[i]);
+                    strides[i - 1] = strides[i] * shape[i];
                     size *= shape[i];
                 }
                 size *= shape[0];
@@ -1372,7 +1372,7 @@ struct vectorize_helper {
         if (trivial == broadcast_trivial::non_trivial) {
             apply_broadcast<Index...>(buffers, buf, index);
         } else {
-            for (size_t i = 0; i < size; ++i)
+            for (ssize_t i = 0; i < size; ++i)
                 output[i] = f((reinterpret_cast<Args *>(buffers[Index].ptr)[buffers[Index].size == 1 ? 0 : i])...);
         }
 
diff --git a/include/pybind11/pytypes.h b/include/pybind11/pytypes.h
index 26f5ae409..f1763110e 100644
--- a/include/pybind11/pytypes.h
+++ b/include/pybind11/pytypes.h
@@ -1165,15 +1165,15 @@ public:
         static std::vector<Py_ssize_t> py_strides { };
         static std::vector<Py_ssize_t> py_shape { };
         buf.buf = info.ptr;
-        buf.itemsize = (Py_ssize_t) info.itemsize;
+        buf.itemsize = info.itemsize;
         buf.format = const_cast<char *>(info.format.c_str());
         buf.ndim = (int) info.ndim;
-        buf.len = (Py_ssize_t) info.size;
+        buf.len = info.size;
         py_strides.clear();
         py_shape.clear();
-        for (size_t i = 0; i < info.ndim; ++i) {
-            py_strides.push_back((Py_ssize_t) info.strides[i]);
-            py_shape.push_back((Py_ssize_t) info.shape[i]);
+        for (size_t i = 0; i < (size_t) info.ndim; ++i) {
+            py_strides.push_back(info.strides[i]);
+            py_shape.push_back(info.shape[i]);
         }
         buf.strides = py_strides.data();
         buf.shape = py_shape.data();
diff --git a/include/pybind11/stl_bind.h b/include/pybind11/stl_bind.h
index 7bba090b2..e581b73eb 100644
--- a/include/pybind11/stl_bind.h
+++ b/include/pybind11/stl_bind.h
@@ -345,21 +345,21 @@ vector_buffer(Class_& cl) {
     format_descriptor<T>::format();
 
     cl.def_buffer([](Vector& v) -> buffer_info {
-        return buffer_info(v.data(), sizeof(T), format_descriptor<T>::format(), 1, {v.size()}, {sizeof(T)});
+        return buffer_info(v.data(), static_cast<ssize_t>(sizeof(T)), format_descriptor<T>::format(), 1, {v.size()}, {sizeof(T)});
     });
 
     cl.def("__init__", [](Vector& vec, buffer buf) {
         auto info = buf.request();
-        if (info.ndim != 1 || info.strides[0] <= 0 || info.strides[0] % static_cast<ssize_t>(sizeof(T)))
+        if (info.ndim != 1 || info.strides[0] % static_cast<ssize_t>(sizeof(T)))
             throw type_error("Only valid 1D buffers can be copied to a vector");
-        if (!detail::compare_buffer_info<T>::compare(info) || sizeof(T) != info.itemsize)
+        if (!detail::compare_buffer_info<T>::compare(info) || (ssize_t) sizeof(T) != info.itemsize)
             throw type_error("Format mismatch (Python: " + info.format + " C++: " + format_descriptor<T>::format() + ")");
         new (&vec) Vector();
-        vec.reserve(info.shape[0]);
+        vec.reserve((size_t) info.shape[0]);
         T *p = static_cast<T*>(info.ptr);
-        auto step = info.strides[0] / static_cast<ssize_t>(sizeof(T));
-        T *end = p + static_cast<ssize_t>(info.shape[0]) * step;
-        for (; p < end; p += step)
+        ssize_t step = info.strides[0] / static_cast<ssize_t>(sizeof(T));
+        T *end = p + info.shape[0] * step;
+        for (; p != end; p += step)
             vec.push_back(*p);
     });
 
diff --git a/tests/test_buffers.cpp b/tests/test_buffers.cpp
index d533bd380..6b5b75551 100644
--- a/tests/test_buffers.cpp
+++ b/tests/test_buffers.cpp
@@ -12,16 +12,16 @@
 
 class Matrix {
 public:
-    Matrix(size_t rows, size_t cols) : m_rows(rows), m_cols(cols) {
+    Matrix(ssize_t rows, ssize_t cols) : m_rows(rows), m_cols(cols) {
         print_created(this, std::to_string(m_rows) + "x" + std::to_string(m_cols) + " matrix");
-        m_data = new float[rows*cols];
-        memset(m_data, 0, sizeof(float) * rows * cols);
+        m_data = new float[(size_t) (rows*cols)];
+        memset(m_data, 0, sizeof(float) * (size_t) (rows * cols));
     }
 
     Matrix(const Matrix &s) : m_rows(s.m_rows), m_cols(s.m_cols) {
         print_copy_created(this, std::to_string(m_rows) + "x" + std::to_string(m_cols) + " matrix");
-        m_data = new float[m_rows * m_cols];
-        memcpy(m_data, s.m_data, sizeof(float) * m_rows * m_cols);
+        m_data = new float[(size_t) (m_rows * m_cols)];
+        memcpy(m_data, s.m_data, sizeof(float) * (size_t) (m_rows * m_cols));
     }
 
     Matrix(Matrix &&s) : m_rows(s.m_rows), m_cols(s.m_cols), m_data(s.m_data) {
@@ -41,8 +41,8 @@ public:
         delete[] m_data;
         m_rows = s.m_rows;
         m_cols = s.m_cols;
-        m_data = new float[m_rows * m_cols];
-        memcpy(m_data, s.m_data, sizeof(float) * m_rows * m_cols);
+        m_data = new float[(size_t) (m_rows * m_cols)];
+        memcpy(m_data, s.m_data, sizeof(float) * (size_t) (m_rows * m_cols));
         return *this;
     }
 
@@ -56,47 +56,47 @@ public:
         return *this;
     }
 
-    float operator()(size_t i, size_t j) const {
-        return m_data[i*m_cols + j];
+    float operator()(ssize_t i, ssize_t j) const {
+        return m_data[(size_t) (i*m_cols + j)];
     }
 
-    float &operator()(size_t i, size_t j) {
-        return m_data[i*m_cols + j];
+    float &operator()(ssize_t i, ssize_t j) {
+        return m_data[(size_t) (i*m_cols + j)];
     }
 
     float *data() { return m_data; }
 
-    size_t rows() const { return m_rows; }
-    size_t cols() const { return m_cols; }
+    ssize_t rows() const { return m_rows; }
+    ssize_t cols() const { return m_cols; }
 private:
-    size_t m_rows;
-    size_t m_cols;
+    ssize_t m_rows;
+    ssize_t m_cols;
     float *m_data;
 };
 
 test_initializer buffers([](py::module &m) {
     py::class_<Matrix> mtx(m, "Matrix", py::buffer_protocol());
 
-    mtx.def(py::init<size_t, size_t>())
+    mtx.def(py::init<ssize_t, ssize_t>())
         /// Construct from a buffer
         .def("__init__", [](Matrix &v, py::buffer b) {
             py::buffer_info info = b.request();
             if (info.format != py::format_descriptor<float>::format() || info.ndim != 2)
                 throw std::runtime_error("Incompatible buffer format!");
             new (&v) Matrix(info.shape[0], info.shape[1]);
-            memcpy(v.data(), info.ptr, sizeof(float) * v.rows() * v.cols());
+            memcpy(v.data(), info.ptr, sizeof(float) * (size_t) (v.rows() * v.cols()));
         })
 
        .def("rows", &Matrix::rows)
        .def("cols", &Matrix::cols)
 
         /// Bare bones interface
-       .def("__getitem__", [](const Matrix &m, std::pair<size_t, size_t> i) {
+       .def("__getitem__", [](const Matrix &m, std::pair<ssize_t, ssize_t> i) {
             if (i.first >= m.rows() || i.second >= m.cols())
                 throw py::index_error();
             return m(i.first, i.second);
         })
-       .def("__setitem__", [](Matrix &m, std::pair<size_t, size_t> i, float v) {
+       .def("__setitem__", [](Matrix &m, std::pair<ssize_t, ssize_t> i, float v) {
             if (i.first >= m.rows() || i.second >= m.cols())
                 throw py::index_error();
             m(i.first, i.second) = v;
@@ -109,8 +109,8 @@ test_initializer buffers([](py::module &m) {
                 py::format_descriptor<float>::format(), /* Python struct-style format descriptor */
                 2,                                      /* Number of dimensions */
                 { m.rows(), m.cols() },                 /* Buffer dimensions */
-                { static_cast<ssize_t>(sizeof(float) * m.rows()), /* Strides (in bytes) for each index */
-                  static_cast<ssize_t>(sizeof(float)) }
+                { sizeof(float) * size_t(m.rows()),     /* Strides (in bytes) for each index */
+                  sizeof(float) }
             );
         })
         ;
diff --git a/tests/test_eigen.py b/tests/test_eigen.py
index 67524f67b..4548fb0e9 100644
--- a/tests/test_eigen.py
+++ b/tests/test_eigen.py
@@ -186,7 +186,7 @@ def test_negative_stride_from_python(msg):
         double_threer(second_row)
     assert msg(excinfo.value) == """
     double_threer(): incompatible function arguments. The following argument types are supported:
-        1. (numpy.ndarray[float32[1, 3], flags.writeable]) -> arg0: None
+        1. (arg0: numpy.ndarray[float32[1, 3], flags.writeable]) -> None
 
     Invoked with: array([ 5.,  4.,  3.], dtype=float32)
 """
@@ -195,7 +195,7 @@ def test_negative_stride_from_python(msg):
         double_threec(second_col)
     assert msg(excinfo.value) == """
     double_threec(): incompatible function arguments. The following argument types are supported:
-        1. (numpy.ndarray[float32[3, 1], flags.writeable]) -> arg0: None
+        1. (arg0: numpy.ndarray[float32[3, 1], flags.writeable]) -> None
 
     Invoked with: array([ 7.,  4.,  1.], dtype=float32)
 """
diff --git a/tests/test_numpy_array.cpp b/tests/test_numpy_array.cpp
index 2e7004eb5..ea5be7902 100644
--- a/tests/test_numpy_array.cpp
+++ b/tests/test_numpy_array.cpp
@@ -13,53 +13,52 @@
 #include <pybind11/stl.h>
 
 #include <cstdint>
-#include <vector>
 
 using arr = py::array;
 using arr_t = py::array_t<uint16_t, 0>;
 static_assert(std::is_same<arr_t::value_type, uint16_t>::value, "");
 
 template<typename... Ix> arr data(const arr& a, Ix... index) {
-    return arr(a.nbytes() - size_t(a.offset_at(index...)), (const uint8_t *) a.data(index...));
+    return arr(a.nbytes() - a.offset_at(index...), (const uint8_t *) a.data(index...));
 }
 
 template<typename... Ix> arr data_t(const arr_t& a, Ix... index) {
-    return arr(a.size() - size_t(a.index_at(index...)), a.data(index...));
+    return arr(a.size() - a.index_at(index...), a.data(index...));
 }
 
 arr& mutate_data(arr& a) {
     auto ptr = (uint8_t *) a.mutable_data();
-    for (size_t i = 0; i < a.nbytes(); i++)
+    for (ssize_t i = 0; i < a.nbytes(); i++)
         ptr[i] = (uint8_t) (ptr[i] * 2);
     return a;
 }
 
 arr_t& mutate_data_t(arr_t& a) {
     auto ptr = a.mutable_data();
-    for (size_t i = 0; i < a.size(); i++)
+    for (ssize_t i = 0; i < a.size(); i++)
         ptr[i]++;
     return a;
 }
 
 template<typename... Ix> arr& mutate_data(arr& a, Ix... index) {
     auto ptr = (uint8_t *) a.mutable_data(index...);
-    for (size_t i = 0; i < a.nbytes() - size_t(a.offset_at(index...)); i++)
+    for (ssize_t i = 0; i < a.nbytes() - a.offset_at(index...); i++)
         ptr[i] = (uint8_t) (ptr[i] * 2);
     return a;
 }
 
 template<typename... Ix> arr_t& mutate_data_t(arr_t& a, Ix... index) {
     auto ptr = a.mutable_data(index...);
-    for (size_t i = 0; i < a.size() - size_t(a.index_at(index...)); i++)
+    for (ssize_t i = 0; i < a.size() - a.index_at(index...); i++)
         ptr[i]++;
     return a;
 }
 
-template<typename... Ix> py::ssize_t index_at(const arr& a, Ix... idx) { return a.index_at(idx...); }
-template<typename... Ix> py::ssize_t index_at_t(const arr_t& a, Ix... idx) { return a.index_at(idx...); }
-template<typename... Ix> py::ssize_t offset_at(const arr& a, Ix... idx) { return a.offset_at(idx...); }
-template<typename... Ix> py::ssize_t offset_at_t(const arr_t& a, Ix... idx) { return a.offset_at(idx...); }
-template<typename... Ix> py::ssize_t at_t(const arr_t& a, Ix... idx) { return a.at(idx...); }
+template<typename... Ix> ssize_t index_at(const arr& a, Ix... idx) { return a.index_at(idx...); }
+template<typename... Ix> ssize_t index_at_t(const arr_t& a, Ix... idx) { return a.index_at(idx...); }
+template<typename... Ix> ssize_t offset_at(const arr& a, Ix... idx) { return a.offset_at(idx...); }
+template<typename... Ix> ssize_t offset_at_t(const arr_t& a, Ix... idx) { return a.offset_at(idx...); }
+template<typename... Ix> ssize_t at_t(const arr_t& a, Ix... idx) { return a.at(idx...); }
 template<typename... Ix> arr_t& mutate_at_t(arr_t& a, Ix... idx) { a.mutable_at(idx...)++; return a; }
 
 #define def_index_fn(name, type) \
@@ -88,9 +87,9 @@ test_initializer numpy_array([](py::module &m) {
 
     sm.def("ndim", [](const arr& a) { return a.ndim(); });
     sm.def("shape", [](const arr& a) { return arr(a.ndim(), a.shape()); });
-    sm.def("shape", [](const arr& a, size_t dim) { return a.shape(dim); });
+    sm.def("shape", [](const arr& a, ssize_t dim) { return a.shape(dim); });
     sm.def("strides", [](const arr& a) { return arr(a.ndim(), a.strides()); });
-    sm.def("strides", [](const arr& a, size_t dim) { return a.strides(dim); });
+    sm.def("strides", [](const arr& a, ssize_t dim) { return a.strides(dim); });
     sm.def("writeable", [](const arr& a) { return a.writeable(); });
     sm.def("size", [](const arr& a) { return a.size(); });
     sm.def("itemsize", [](const arr& a) { return a.itemsize(); });
@@ -202,33 +201,33 @@ test_initializer numpy_array([](py::module &m) {
 
     sm.def("proxy_add2", [](py::array_t<double> a, double v) {
         auto r = a.mutable_unchecked<2>();
-        for (size_t i = 0; i < r.shape(0); i++)
-            for (size_t j = 0; j < r.shape(1); j++)
+        for (ssize_t i = 0; i < r.shape(0); i++)
+            for (ssize_t j = 0; j < r.shape(1); j++)
                 r(i, j) += v;
     }, py::arg().noconvert(), py::arg());
 
     sm.def("proxy_init3", [](double start) {
         py::array_t<double, py::array::c_style> a({ 3, 3, 3 });
         auto r = a.mutable_unchecked<3>();
-        for (size_t i = 0; i < r.shape(0); i++)
-        for (size_t j = 0; j < r.shape(1); j++)
-        for (size_t k = 0; k < r.shape(2); k++)
+        for (ssize_t i = 0; i < r.shape(0); i++)
+        for (ssize_t j = 0; j < r.shape(1); j++)
+        for (ssize_t k = 0; k < r.shape(2); k++)
             r(i, j, k) = start++;
         return a;
     });
     sm.def("proxy_init3F", [](double start) {
         py::array_t<double, py::array::f_style> a({ 3, 3, 3 });
         auto r = a.mutable_unchecked<3>();
-        for (size_t k = 0; k < r.shape(2); k++)
-        for (size_t j = 0; j < r.shape(1); j++)
-        for (size_t i = 0; i < r.shape(0); i++)
+        for (ssize_t k = 0; k < r.shape(2); k++)
+        for (ssize_t j = 0; j < r.shape(1); j++)
+        for (ssize_t i = 0; i < r.shape(0); i++)
             r(i, j, k) = start++;
         return a;
     });
     sm.def("proxy_squared_L2_norm", [](py::array_t<double> a) {
         auto r = a.unchecked<1>();
         double sumsq = 0;
-        for (size_t i = 0; i < r.shape(0); i++)
+        for (ssize_t i = 0; i < r.shape(0); i++)
             sumsq += r[i] * r(i); // Either notation works for a 1D array
         return sumsq;
     });
@@ -243,17 +242,17 @@ test_initializer numpy_array([](py::module &m) {
     sm.def("proxy_add2_dyn", [](py::array_t<double> a, double v) {
         auto r = a.mutable_unchecked();
         if (r.ndim() != 2) throw std::domain_error("error: ndim != 2");
-        for (size_t i = 0; i < r.shape(0); i++)
-            for (size_t j = 0; j < r.shape(1); j++)
+        for (ssize_t i = 0; i < r.shape(0); i++)
+            for (ssize_t j = 0; j < r.shape(1); j++)
                 r(i, j) += v;
     }, py::arg().noconvert(), py::arg());
     sm.def("proxy_init3_dyn", [](double start) {
         py::array_t<double, py::array::c_style> a({ 3, 3, 3 });
         auto r = a.mutable_unchecked();
         if (r.ndim() != 3) throw std::domain_error("error: ndim != 3");
-        for (size_t i = 0; i < r.shape(0); i++)
-        for (size_t j = 0; j < r.shape(1); j++)
-        for (size_t k = 0; k < r.shape(2); k++)
+        for (ssize_t i = 0; i < r.shape(0); i++)
+        for (ssize_t j = 0; j < r.shape(1); j++)
+        for (ssize_t k = 0; k < r.shape(2); k++)
             r(i, j, k) = start++;
         return a;
     });
@@ -269,6 +268,9 @@ test_initializer numpy_array([](py::module &m) {
     sm.def("array_fail_test", []() { return py::array(py::object()); });
     sm.def("array_t_fail_test", []() { return py::array_t<double>(py::object()); });
 
+    // Make sure the error from numpy is being passed through:
+    sm.def("array_fail_test_negative_size", []() { int c = 0; return py::array(-1, &c); });
+
     // Issue (unnumbered; reported in #788): regression: initializer lists can be ambiguous
     sm.def("array_initializer_list", []() { return py::array_t<float>(1); }); // { 1 } also works, but clang warns about it
     sm.def("array_initializer_list", []() { return py::array_t<float>({ 1, 2 }); });
@@ -277,7 +279,7 @@ test_initializer numpy_array([](py::module &m) {
 
     // reshape array to 2D without changing size
     sm.def("array_reshape2", [](py::array_t<double> a) {
-        const size_t dim_sz = (size_t)std::sqrt(a.size());
+        const ssize_t dim_sz = (ssize_t)std::sqrt(a.size());
         if (dim_sz * dim_sz != a.size())
             throw std::domain_error("array_reshape2: input array total size is not a squared integer");
         a.resize({dim_sz, dim_sz});
diff --git a/tests/test_numpy_array.py b/tests/test_numpy_array.py
index 90f9b8028..90fa142d8 100644
--- a/tests/test_numpy_array.py
+++ b/tests/test_numpy_array.py
@@ -384,7 +384,8 @@ def test_array_unchecked_dyn_dims(msg):
 
 
 def test_array_failure():
-    from pybind11_tests.array import array_fail_test, array_t_fail_test
+    from pybind11_tests.array import (array_fail_test, array_t_fail_test,
+                                      array_fail_test_negative_size)
 
     with pytest.raises(ValueError) as excinfo:
         array_fail_test()
@@ -394,6 +395,10 @@ def test_array_failure():
         array_t_fail_test()
     assert str(excinfo.value) == 'cannot create a pybind11::array_t from a nullptr'
 
+    with pytest.raises(ValueError) as excinfo:
+        array_fail_test_negative_size()
+    assert str(excinfo.value) == 'negative dimensions are not allowed'
+
 
 def test_array_resize(msg):
     from pybind11_tests.array import (array_reshape2, array_resize3)
diff --git a/tests/test_numpy_dtypes.cpp b/tests/test_numpy_dtypes.cpp
index 4fc797c8c..a8ba3d87f 100644
--- a/tests/test_numpy_dtypes.cpp
+++ b/tests/test_numpy_dtypes.cpp
@@ -149,7 +149,7 @@ py::array_t<StringStruct, 0> create_string_array(bool non_empty) {
     if (non_empty) {
         auto req = arr.request();
         auto ptr = static_cast<StringStruct*>(req.ptr);
-        for (size_t i = 0; i < req.size * req.itemsize; i++)
+        for (ssize_t i = 0; i < req.size * req.itemsize; i++)
             static_cast<char*>(req.ptr)[i] = 0;
         ptr[1].a[0] = 'a'; ptr[1].b[0] = 'a';
         ptr[2].a[0] = 'a'; ptr[2].b[0] = 'a';
@@ -178,7 +178,7 @@ py::list print_recarray(py::array_t<S, 0> arr) {
     const auto req = arr.request();
     const auto ptr = static_cast<S*>(req.ptr);
     auto l = py::list();
-    for (size_t i = 0; i < req.size; i++) {
+    for (ssize_t i = 0; i < req.size; i++) {
         std::stringstream ss;
         ss << ptr[i];
         l.append(py::str(ss.str()));
@@ -225,7 +225,7 @@ py::array_t<int32_t, 0> test_array_ctors(int i) {
     using arr_t = py::array_t<int32_t, 0>;
 
     std::vector<int32_t> data { 1, 2, 3, 4, 5, 6 };
-    std::vector<size_t> shape { 3, 2 };
+    std::vector<ssize_t> shape { 3, 2 };
     std::vector<ssize_t> strides { 8, 4 };
 
     auto ptr = data.data();
diff --git a/tests/test_numpy_vectorize.cpp b/tests/test_numpy_vectorize.cpp
index 8e951c6e1..34197469c 100644
--- a/tests/test_numpy_vectorize.cpp
+++ b/tests/test_numpy_vectorize.cpp
@@ -50,8 +50,8 @@ test_initializer numpy_vectorize([](py::module &m) {
                 py::array_t<float, py::array::forcecast> arg2,
                 py::array_t<double, py::array::forcecast> arg3
                 ) {
-        size_t ndim;
-        std::vector<size_t> shape;
+        ssize_t ndim;
+        std::vector<ssize_t> shape;
         std::array<py::buffer_info, 3> buffers {{ arg1.request(), arg2.request(), arg3.request() }};
         return py::detail::broadcast(buffers, ndim, shape);
     });